libringbuffer/frontend_internal.h

   1 #ifndef _LINUX_RING_BUFFER_FRONTEND_INTERNAL_H
   2 #define _LINUX_RING_BUFFER_FRONTEND_INTERNAL_H
   3
   4 /*
   5  * linux/ringbuffer/frontend_internal.h
   6  *
   7  * (C) Copyright 2005-2010 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   8  *
   9  * Ring Buffer Library Synchronization Header (internal helpers).
  10  *
  11  * Author:
  12  *      Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
  13  *
  14  * See ring_buffer_frontend.c for more information on wait-free algorithms.
  15  *
  16  * Dual LGPL v2.1/GPL v2 license.
  17  */
  18
  19 #include <urcu/compiler.h>
  20
  21 #include <ust/ringbuffer-config.h>
  22 #include "backend_types.h"
  23 #include "frontend_types.h"
  24 #include "shm.h"
  25
  26 /* Buffer offset macros */
  27
  28 /* buf_trunc mask selects only the buffer number. */
  29 static inline
  30 unsigned long buf_trunc(unsigned long offset, struct channel *chan)
  31 {
  32         return offset & ~(chan->backend.buf_size - 1);
  33
  34 }
  35
  36 /* Select the buffer number value (counter). */
  37 static inline
  38 unsigned long buf_trunc_val(unsigned long offset, struct channel *chan)
  39 {
  40         return buf_trunc(offset, chan) >> chan->backend.buf_size_order;
  41 }
  42
  43 /* buf_offset mask selects only the offset within the current buffer. */
  44 static inline
  45 unsigned long buf_offset(unsigned long offset, struct channel *chan)
  46 {
  47         return offset & (chan->backend.buf_size - 1);
  48 }
  49
  50 /* subbuf_offset mask selects the offset within the current subbuffer. */
  51 static inline
  52 unsigned long subbuf_offset(unsigned long offset, struct channel *chan)
  53 {
  54         return offset & (chan->backend.subbuf_size - 1);
  55 }
  56
  57 /* subbuf_trunc mask selects the subbuffer number. */
  58 static inline
  59 unsigned long subbuf_trunc(unsigned long offset, struct channel *chan)
  60 {
  61         return offset & ~(chan->backend.subbuf_size - 1);
  62 }
  63
  64 /* subbuf_align aligns the offset to the next subbuffer. */
  65 static inline
  66 unsigned long subbuf_align(unsigned long offset, struct channel *chan)
  67 {
  68         return (offset + chan->backend.subbuf_size)
  69                & ~(chan->backend.subbuf_size - 1);
  70 }
  71
  72 /* subbuf_index returns the index of the current subbuffer within the buffer. */
  73 static inline
  74 unsigned long subbuf_index(unsigned long offset, struct channel *chan)
  75 {
  76         return buf_offset(offset, chan) >> chan->backend.subbuf_size_order;
  77 }
  78
  79 /*
  80  * Last TSC comparison functions. Check if the current TSC overflows tsc_bits
  81  * bits from the last TSC read. When overflows are detected, the full 64-bit
  82  * timestamp counter should be written in the record header. Reads and writes
  83  * last_tsc atomically.
  84  */
  85
  86 #if (CAA_BITS_PER_LONG == 32)
  87 static inline
  88 void save_last_tsc(const struct lib_ring_buffer_config *config,
  89                    struct lib_ring_buffer *buf, u64 tsc)
  90 {
  91         if (config->tsc_bits == 0 || config->tsc_bits == 64)
  92                 return;
  93
  94         /*
  95          * Ensure the compiler performs this update in a single instruction.
  96          */
  97         v_set(config, &buf->last_tsc, (unsigned long)(tsc >> config->tsc_bits));
  98 }
  99
 100 static inline
 101 int last_tsc_overflow(const struct lib_ring_buffer_config *config,
 102                       struct lib_ring_buffer *buf, u64 tsc)
 103 {
 104         unsigned long tsc_shifted;
 105
 106         if (config->tsc_bits == 0 || config->tsc_bits == 64)
 107                 return 0;
 108
 109         tsc_shifted = (unsigned long)(tsc >> config->tsc_bits);
 110         if (unlikely(tsc_shifted
 111                      - (unsigned long)v_read(config, &buf->last_tsc)))
 112                 return 1;
 113         else
 114                 return 0;
 115 }
 116 #else
 117 static inline
 118 void save_last_tsc(const struct lib_ring_buffer_config *config,
 119                    struct lib_ring_buffer *buf, u64 tsc)
 120 {
 121         if (config->tsc_bits == 0 || config->tsc_bits == 64)
 122                 return;
 123
 124         v_set(config, &buf->last_tsc, (unsigned long)tsc);
 125 }
 126
 127 static inline
 128 int last_tsc_overflow(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 0;
 133
 134         if (unlikely((tsc - v_read(config, &buf->last_tsc))
 135                      >> config->tsc_bits))
 136                 return 1;
 137         else
 138                 return 0;
 139 }
 140 #endif
 141
 142 extern
 143 int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx);
 144
 145 extern
 146 void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf,
 147                                  enum switch_mode mode,
 148                                  struct shm_handle *handle);
 149
 150 /* Buffer write helpers */
 151
 152 static inline
 153 void lib_ring_buffer_reserve_push_reader(struct lib_ring_buffer *buf,
 154                                          struct channel *chan,
 155                                          unsigned long offset)
 156 {
 157         unsigned long consumed_old, consumed_new;
 158
 159         do {
 160                 consumed_old = uatomic_read(&buf->consumed);
 161                 /*
 162                  * If buffer is in overwrite mode, push the reader consumed
 163                  * count if the write position has reached it and we are not
 164                  * at the first iteration (don't push the reader farther than
 165                  * the writer). This operation can be done concurrently by many
 166                  * writers in the same buffer, the writer being at the farthest
 167                  * write position sub-buffer index in the buffer being the one
 168                  * which will win this loop.
 169                  */
 170                 if (unlikely(subbuf_trunc(offset, chan)
 171                               - subbuf_trunc(consumed_old, chan)
 172                              >= chan->backend.buf_size))
 173                         consumed_new = subbuf_align(consumed_old, chan);
 174                 else
 175                         return;
 176         } while (unlikely(uatomic_cmpxchg(&buf->consumed, consumed_old,
 177                                               consumed_new) != consumed_old));
 178 }
 179
 180 static inline
 181 void lib_ring_buffer_vmcore_check_deliver(const struct lib_ring_buffer_config *config,
 182                                           struct lib_ring_buffer *buf,
 183                                           unsigned long commit_count,
 184                                           unsigned long idx,
 185                                           struct shm_handle *handle)
 186 {
 187         if (config->oops == RING_BUFFER_OOPS_CONSISTENCY)
 188                 v_set(config, &shmp_index(handle, buf->commit_hot, idx)->seq, commit_count);
 189 }
 190
 191 static inline
 192 int lib_ring_buffer_poll_deliver(const struct lib_ring_buffer_config *config,
 193                                  struct lib_ring_buffer *buf,
 194                                  struct channel *chan,
 195                                  struct shm_handle *handle)
 196 {
 197         unsigned long consumed_old, consumed_idx, commit_count, write_offset;
 198
 199         consumed_old = uatomic_read(&buf->consumed);
 200         consumed_idx = subbuf_index(consumed_old, chan);
 201         commit_count = v_read(config, &shmp_index(handle, buf->commit_cold, consumed_idx)->cc_sb);
 202         /*
 203          * No memory barrier here, since we are only interested
 204          * in a statistically correct polling result. The next poll will
 205          * get the data is we are racing. The mb() that ensures correct
 206          * memory order is in get_subbuf.
 207          */
 208         write_offset = v_read(config, &buf->offset);
 209
 210         /*
 211          * Check that the subbuffer we are trying to consume has been
 212          * already fully committed.
 213          */
 214
 215         if (((commit_count - chan->backend.subbuf_size)
 216              & chan->commit_count_mask)
 217             - (buf_trunc(consumed_old, chan)
 218                >> chan->backend.num_subbuf_order)
 219             != 0)
 220                 return 0;
 221
 222         /*
 223          * Check that we are not about to read the same subbuffer in
 224          * which the writer head is.
 225          */
 226         if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan)
 227             == 0)
 228                 return 0;
 229
 230         return 1;
 231
 232 }
 233
 234 static inline
 235 int lib_ring_buffer_pending_data(const struct lib_ring_buffer_config *config,
 236                                  struct lib_ring_buffer *buf,
 237                                  struct channel *chan)
 238 {
 239         return !!subbuf_offset(v_read(config, &buf->offset), chan);
 240 }
 241
 242 static inline
 243 unsigned long lib_ring_buffer_get_data_size(const struct lib_ring_buffer_config *config,
 244                                             struct lib_ring_buffer *buf,
 245                                             unsigned long idx,
 246                                             struct shm_handle *handle)
 247 {
 248         return subbuffer_get_data_size(config, &buf->backend, idx, handle);
 249 }
 250
 251 /*
 252  * Check if all space reservation in a buffer have been committed. This helps
 253  * knowing if an execution context is nested (for per-cpu buffers only).
 254  * This is a very specific ftrace use-case, so we keep this as "internal" API.
 255  */
 256 static inline
 257 int lib_ring_buffer_reserve_committed(const struct lib_ring_buffer_config *config,
 258                                       struct lib_ring_buffer *buf,
 259                                       struct channel *chan,
 260                                       struct shm_handle *handle)
 261 {
 262         unsigned long offset, idx, commit_count;
 263
 264         CHAN_WARN_ON(chan, config->alloc != RING_BUFFER_ALLOC_PER_CPU);
 265         CHAN_WARN_ON(chan, config->sync != RING_BUFFER_SYNC_PER_CPU);
 266
 267         /*
 268          * Read offset and commit count in a loop so they are both read
 269          * atomically wrt interrupts. By deal with interrupt concurrency by
 270          * restarting both reads if the offset has been pushed. Note that given
 271          * we only have to deal with interrupt concurrency here, an interrupt
 272          * modifying the commit count will also modify "offset", so it is safe
 273          * to only check for offset modifications.
 274          */
 275         do {
 276                 offset = v_read(config, &buf->offset);
 277                 idx = subbuf_index(offset, chan);
 278                 commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, idx)->cc);
 279         } while (offset != v_read(config, &buf->offset));
 280
 281         return ((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
 282                      - (commit_count & chan->commit_count_mask) == 0);
 283 }
 284
 285 static inline
 286 void lib_ring_buffer_check_deliver(const struct lib_ring_buffer_config *config,
 287                                    struct lib_ring_buffer *buf,
 288                                    struct channel *chan,
 289                                    unsigned long offset,
 290                                    unsigned long commit_count,
 291                                    unsigned long idx,
 292                                    struct shm_handle *handle)
 293 {
 294         unsigned long old_commit_count = commit_count
 295                                          - chan->backend.subbuf_size;
 296         u64 tsc;
 297
 298         /* Check if all commits have been done */
 299         if (unlikely((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
 300                      - (old_commit_count & chan->commit_count_mask) == 0)) {
 301                 /*
 302                  * If we succeeded at updating cc_sb below, we are the subbuffer
 303                  * writer delivering the subbuffer. Deals with concurrent
 304                  * updates of the "cc" value without adding a add_return atomic
 305                  * operation to the fast path.
 306                  *
 307                  * We are doing the delivery in two steps:
 308                  * - First, we cmpxchg() cc_sb to the new value
 309                  *   old_commit_count + 1. This ensures that we are the only
 310                  *   subbuffer user successfully filling the subbuffer, but we
 311                  *   do _not_ set the cc_sb value to "commit_count" yet.
 312                  *   Therefore, other writers that would wrap around the ring
 313                  *   buffer and try to start writing to our subbuffer would
 314                  *   have to drop records, because it would appear as
 315                  *   non-filled.
 316                  *   We therefore have exclusive access to the subbuffer control
 317                  *   structures.  This mutual exclusion with other writers is
 318                  *   crucially important to perform record overruns count in
 319                  *   flight recorder mode locklessly.
 320                  * - When we are ready to release the subbuffer (either for
 321                  *   reading or for overrun by other writers), we simply set the
 322                  *   cc_sb value to "commit_count" and perform delivery.
 323                  *
 324                  * The subbuffer size is least 2 bytes (minimum size: 1 page).
 325                  * This guarantees that old_commit_count + 1 != commit_count.
 326                  */
 327                 if (likely(v_cmpxchg(config, &shmp_index(handle, buf->commit_cold, idx)->cc_sb,
 328                                          old_commit_count, old_commit_count + 1)
 329                            == old_commit_count)) {
 330                         /*
 331                          * Start of exclusive subbuffer access. We are
 332                          * guaranteed to be the last writer in this subbuffer
 333                          * and any other writer trying to access this subbuffer
 334                          * in this state is required to drop records.
 335                          */
 336                         tsc = config->cb.ring_buffer_clock_read(chan);
 337                         v_add(config,
 338                               subbuffer_get_records_count(config,
 339                                                           &buf->backend,
 340                                                           idx, handle),
 341                               &buf->records_count);
 342                         v_add(config,
 343                               subbuffer_count_records_overrun(config,
 344                                                               &buf->backend,
 345                                                               idx, handle),
 346                               &buf->records_overrun);
 347                         config->cb.buffer_end(buf, tsc, idx,
 348                                               lib_ring_buffer_get_data_size(config,
 349                                                                         buf,
 350                                                                         idx,
 351                                                                         handle),
 352                                               handle);
 353
 354                         /*
 355                          * Set noref flag and offset for this subbuffer id.
 356                          * Contains a memory barrier that ensures counter stores
 357                          * are ordered before set noref and offset.
 358                          */
 359                         lib_ring_buffer_set_noref_offset(config, &buf->backend, idx,
 360                                                          buf_trunc_val(offset, chan), handle);
 361
 362                         /*
 363                          * Order set_noref and record counter updates before the
 364                          * end of subbuffer exclusive access. Orders with
 365                          * respect to writers coming into the subbuffer after
 366                          * wrap around, and also order wrt concurrent readers.
 367                          */
 368                         cmm_smp_mb();
 369                         /* End of exclusive subbuffer access */
 370                         v_set(config, &shmp_index(handle, buf->commit_cold, idx)->cc_sb,
 371                               commit_count);
 372                         lib_ring_buffer_vmcore_check_deliver(config, buf,
 373                                                  commit_count, idx, handle);
 374
 375                         /*
 376                          * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
 377                          */
 378                         if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER
 379                             && (uatomic_read(&buf->active_readers)
 380                                 || uatomic_read(&buf->active_shadow_readers))
 381                             && lib_ring_buffer_poll_deliver(config, buf, chan, handle)) {
 382                                 int wakeup_fd = shm_get_wakeup_fd(handle, &buf->self._ref);
 383
 384                                 if (wakeup_fd >= 0) {
 385                                         int ret;
 386                                         /*
 387                                          * Wake-up the other end by
 388                                          * writing a null byte in the
 389                                          * pipe (non-blocking).
 390                                          * Important note: Because
 391                                          * writing into the pipe is
 392                                          * non-blocking (and therefore
 393                                          * we allow dropping wakeup
 394                                          * data, as long as there is
 395                                          * wakeup data present in the
 396                                          * pipe buffer to wake up the
 397                                          * consumer), the consumer
 398                                          * should perform the following
 399                                          * sequence for waiting:
 400                                          * 1) empty the pipe (reads).
 401                                          * 2) check if there is data in
 402                                          *    the buffer.
 403                                          * 3) wait on the pipe (poll).
 404                                          */
 405                                         do {
 406                                                 ret = write(wakeup_fd, "", 1);
 407                                         } while (ret == -1L && errno == EINTR);
 408                                 }
 409                         }
 410
 411                 }
 412         }
 413 }
 414
 415 /*
 416  * lib_ring_buffer_write_commit_counter
 417  *
 418  * For flight recording. must be called after commit.
 419  * This function increments the subbuffer's commit_seq counter each time the
 420  * commit count reaches back the reserve offset (modulo subbuffer size). It is
 421  * useful for crash dump.
 422  */
 423 static inline
 424 void lib_ring_buffer_write_commit_counter(const struct lib_ring_buffer_config *config,
 425                                           struct lib_ring_buffer *buf,
 426                                           struct channel *chan,
 427                                           unsigned long idx,
 428                                           unsigned long buf_offset,
 429                                           unsigned long commit_count,
 430                                           size_t slot_size,
 431                                           struct shm_handle *handle)
 432 {
 433         unsigned long offset, commit_seq_old;
 434
 435         if (config->oops != RING_BUFFER_OOPS_CONSISTENCY)
 436                 return;
 437
 438         offset = buf_offset + slot_size;
 439
 440         /*
 441          * subbuf_offset includes commit_count_mask. We can simply
 442          * compare the offsets within the subbuffer without caring about
 443          * buffer full/empty mismatch because offset is never zero here
 444          * (subbuffer header and record headers have non-zero length).
 445          */
 446         if (unlikely(subbuf_offset(offset - commit_count, chan)))
 447                 return;
 448
 449         commit_seq_old = v_read(config, &shmp_index(handle, buf->commit_hot, idx)->seq);
 450         while ((long) (commit_seq_old - commit_count) < 0)
 451                 commit_seq_old = v_cmpxchg(config, &shmp_index(handle, buf->commit_hot, idx)->seq,
 452                                            commit_seq_old, commit_count);
 453 }
 454
 455 extern int lib_ring_buffer_create(struct lib_ring_buffer *buf,
 456                                   struct channel_backend *chanb, int cpu,
 457                                   struct shm_handle *handle,
 458                                   struct shm_object *shmobj);
 459 extern void lib_ring_buffer_free(struct lib_ring_buffer *buf,
 460                                  struct shm_handle *handle);
 461
 462 /* Keep track of trap nesting inside ring buffer code */
 463 extern __thread unsigned int lib_ring_buffer_nesting;
 464
 465 #endif /* _LINUX_RING_BUFFER_FRONTEND_INTERNAL_H */