lib/ringbuffer/frontend_internal.h

   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 extern
 162 void lib_ring_buffer_switch_remote_empty(struct lib_ring_buffer *buf);
 163
 164 /* Buffer write helpers */
 165
 166 static inline
 167 void lib_ring_buffer_reserve_push_reader(struct lib_ring_buffer *buf,
 168                                          struct channel *chan,
 169                                          unsigned long offset)
 170 {
 171         unsigned long consumed_old, consumed_new;
 172
 173         do {
 174                 consumed_old = atomic_long_read(&buf->consumed);
 175                 /*
 176                  * If buffer is in overwrite mode, push the reader consumed
 177                  * count if the write position has reached it and we are not
 178                  * at the first iteration (don't push the reader farther than
 179                  * the writer). This operation can be done concurrently by many
 180                  * writers in the same buffer, the writer being at the farthest
 181                  * write position sub-buffer index in the buffer being the one
 182                  * which will win this loop.
 183                  */
 184                 if (unlikely(subbuf_trunc(offset, chan)
 185                               - subbuf_trunc(consumed_old, chan)
 186                              >= chan->backend.buf_size))
 187                         consumed_new = subbuf_align(consumed_old, chan);
 188                 else
 189                         return;
 190         } while (unlikely(atomic_long_cmpxchg(&buf->consumed, consumed_old,
 191                                               consumed_new) != consumed_old));
 192 }
 193
 194 static inline
 195 void lib_ring_buffer_vmcore_check_deliver(const struct lib_ring_buffer_config *config,
 196                                           struct lib_ring_buffer *buf,
 197                                           unsigned long commit_count,
 198                                           unsigned long idx)
 199 {
 200         if (config->oops == RING_BUFFER_OOPS_CONSISTENCY)
 201                 v_set(config, &buf->commit_hot[idx].seq, commit_count);
 202 }
 203
 204 static inline
 205 int lib_ring_buffer_poll_deliver(const struct lib_ring_buffer_config *config,
 206                                  struct lib_ring_buffer *buf,
 207                                  struct channel *chan)
 208 {
 209         unsigned long consumed_old, consumed_idx, commit_count, write_offset;
 210
 211         consumed_old = atomic_long_read(&buf->consumed);
 212         consumed_idx = subbuf_index(consumed_old, chan);
 213         commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb);
 214         /*
 215          * No memory barrier here, since we are only interested
 216          * in a statistically correct polling result. The next poll will
 217          * get the data is we are racing. The mb() that ensures correct
 218          * memory order is in get_subbuf.
 219          */
 220         write_offset = v_read(config, &buf->offset);
 221
 222         /*
 223          * Check that the subbuffer we are trying to consume has been
 224          * already fully committed.
 225          */
 226
 227         if (((commit_count - chan->backend.subbuf_size)
 228              & chan->commit_count_mask)
 229             - (buf_trunc(consumed_old, chan)
 230                >> chan->backend.num_subbuf_order)
 231             != 0)
 232                 return 0;
 233
 234         /*
 235          * Check that we are not about to read the same subbuffer in
 236          * which the writer head is.
 237          */
 238         if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan)
 239             == 0)
 240                 return 0;
 241
 242         return 1;
 243
 244 }
 245
 246 static inline
 247 int lib_ring_buffer_pending_data(const struct lib_ring_buffer_config *config,
 248                                  struct lib_ring_buffer *buf,
 249                                  struct channel *chan)
 250 {
 251         return !!subbuf_offset(v_read(config, &buf->offset), chan);
 252 }
 253
 254 static inline
 255 unsigned long lib_ring_buffer_get_data_size(const struct lib_ring_buffer_config *config,
 256                                             struct lib_ring_buffer *buf,
 257                                             unsigned long idx)
 258 {
 259         return subbuffer_get_data_size(config, &buf->backend, idx);
 260 }
 261
 262 /*
 263  * Check if all space reservation in a buffer have been committed. This helps
 264  * knowing if an execution context is nested (for per-cpu buffers only).
 265  * This is a very specific ftrace use-case, so we keep this as "internal" API.
 266  */
 267 static inline
 268 int lib_ring_buffer_reserve_committed(const struct lib_ring_buffer_config *config,
 269                                       struct lib_ring_buffer *buf,
 270                                       struct channel *chan)
 271 {
 272         unsigned long offset, idx, commit_count;
 273
 274         CHAN_WARN_ON(chan, config->alloc != RING_BUFFER_ALLOC_PER_CPU);
 275         CHAN_WARN_ON(chan, config->sync != RING_BUFFER_SYNC_PER_CPU);
 276
 277         /*
 278          * Read offset and commit count in a loop so they are both read
 279          * atomically wrt interrupts. By deal with interrupt concurrency by
 280          * restarting both reads if the offset has been pushed. Note that given
 281          * we only have to deal with interrupt concurrency here, an interrupt
 282          * modifying the commit count will also modify "offset", so it is safe
 283          * to only check for offset modifications.
 284          */
 285         do {
 286                 offset = v_read(config, &buf->offset);
 287                 idx = subbuf_index(offset, chan);
 288                 commit_count = v_read(config, &buf->commit_hot[idx].cc);
 289         } while (offset != v_read(config, &buf->offset));
 290
 291         return ((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
 292                      - (commit_count & chan->commit_count_mask) == 0);
 293 }
 294
 295 /*
 296  * Receive end of subbuffer TSC as parameter. It has been read in the
 297  * space reservation loop of either reserve or switch, which ensures it
 298  * progresses monotonically with event records in the buffer. Therefore,
 299  * it ensures that the end timestamp of a subbuffer is <= begin
 300  * timestamp of the following subbuffers.
 301  */
 302 static inline
 303 void lib_ring_buffer_check_deliver(const struct lib_ring_buffer_config *config,
 304                                    struct lib_ring_buffer *buf,
 305                                    struct channel *chan,
 306                                    unsigned long offset,
 307                                    unsigned long commit_count,
 308                                    unsigned long idx,
 309                                    u64 tsc)
 310 {
 311         unsigned long old_commit_count = commit_count
 312                                          - chan->backend.subbuf_size;
 313
 314         /* Check if all commits have been done */
 315         if (unlikely((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
 316                      - (old_commit_count & chan->commit_count_mask) == 0)) {
 317                 /*
 318                  * If we succeeded at updating cc_sb below, we are the subbuffer
 319                  * writer delivering the subbuffer. Deals with concurrent
 320                  * updates of the "cc" value without adding a add_return atomic
 321                  * operation to the fast path.
 322                  *
 323                  * We are doing the delivery in two steps:
 324                  * - First, we cmpxchg() cc_sb to the new value
 325                  *   old_commit_count + 1. This ensures that we are the only
 326                  *   subbuffer user successfully filling the subbuffer, but we
 327                  *   do _not_ set the cc_sb value to "commit_count" yet.
 328                  *   Therefore, other writers that would wrap around the ring
 329                  *   buffer and try to start writing to our subbuffer would
 330                  *   have to drop records, because it would appear as
 331                  *   non-filled.
 332                  *   We therefore have exclusive access to the subbuffer control
 333                  *   structures.  This mutual exclusion with other writers is
 334                  *   crucially important to perform record overruns count in
 335                  *   flight recorder mode locklessly.
 336                  * - When we are ready to release the subbuffer (either for
 337                  *   reading or for overrun by other writers), we simply set the
 338                  *   cc_sb value to "commit_count" and perform delivery.
 339                  *
 340                  * The subbuffer size is least 2 bytes (minimum size: 1 page).
 341                  * This guarantees that old_commit_count + 1 != commit_count.
 342                  */
 343
 344                 /*
 345                  * Order prior updates to reserve count prior to the
 346                  * commit_cold cc_sb update.
 347                  */
 348                 smp_wmb();
 349                 if (likely(v_cmpxchg(config, &buf->commit_cold[idx].cc_sb,
 350                                          old_commit_count, old_commit_count + 1)
 351                            == old_commit_count)) {
 352                         /*
 353                          * Start of exclusive subbuffer access. We are
 354                          * guaranteed to be the last writer in this subbuffer
 355                          * and any other writer trying to access this subbuffer
 356                          * in this state is required to drop records.
 357                          */
 358                         v_add(config,
 359                               subbuffer_get_records_count(config,
 360                                                           &buf->backend, idx),
 361                               &buf->records_count);
 362                         v_add(config,
 363                               subbuffer_count_records_overrun(config,
 364                                                               &buf->backend,
 365                                                               idx),
 366                               &buf->records_overrun);
 367                         config->cb.buffer_end(buf, tsc, idx,
 368                                               lib_ring_buffer_get_data_size(config,
 369                                                                         buf,
 370                                                                         idx));
 371
 372                         /*
 373                          * Increment the packet counter while we have exclusive
 374                          * access.
 375                          */
 376                         subbuffer_inc_packet_count(config, &buf->backend, idx);
 377
 378                         /*
 379                          * Set noref flag and offset for this subbuffer id.
 380                          * Contains a memory barrier that ensures counter stores
 381                          * are ordered before set noref and offset.
 382                          */
 383                         lib_ring_buffer_set_noref_offset(config, &buf->backend, idx,
 384                                                          buf_trunc_val(offset, chan));
 385
 386                         /*
 387                          * Order set_noref and record counter updates before the
 388                          * end of subbuffer exclusive access. Orders with
 389                          * respect to writers coming into the subbuffer after
 390                          * wrap around, and also order wrt concurrent readers.
 391                          */
 392                         smp_mb();
 393                         /* End of exclusive subbuffer access */
 394                         v_set(config, &buf->commit_cold[idx].cc_sb,
 395                               commit_count);
 396                         /*
 397                          * Order later updates to reserve count after
 398                          * the commit_cold cc_sb update.
 399                          */
 400                         smp_wmb();
 401                         lib_ring_buffer_vmcore_check_deliver(config, buf,
 402                                                          commit_count, idx);
 403
 404                         /*
 405                          * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
 406                          */
 407                         if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER
 408                             && atomic_long_read(&buf->active_readers)
 409                             && lib_ring_buffer_poll_deliver(config, buf, chan)) {
 410                                 wake_up_interruptible(&buf->read_wait);
 411                                 wake_up_interruptible(&chan->read_wait);
 412                         }
 413
 414                 }
 415         }
 416 }
 417
 418 /*
 419  * lib_ring_buffer_write_commit_counter
 420  *
 421  * For flight recording. must be called after commit.
 422  * This function increments the subbuffer's commit_seq counter each time the
 423  * commit count reaches back the reserve offset (modulo subbuffer size). It is
 424  * useful for crash dump.
 425  */
 426 static inline
 427 void lib_ring_buffer_write_commit_counter(const struct lib_ring_buffer_config *config,
 428                                           struct lib_ring_buffer *buf,
 429                                           struct channel *chan,
 430                                           unsigned long idx,
 431                                           unsigned long buf_offset,
 432                                           unsigned long commit_count)
 433 {
 434         unsigned long commit_seq_old;
 435
 436         if (config->oops != RING_BUFFER_OOPS_CONSISTENCY)
 437                 return;
 438
 439         /*
 440          * subbuf_offset includes commit_count_mask. We can simply
 441          * compare the offsets within the subbuffer without caring about
 442          * buffer full/empty mismatch because offset is never zero here
 443          * (subbuffer header and record headers have non-zero length).
 444          */
 445         if (unlikely(subbuf_offset(buf_offset - commit_count, chan)))
 446                 return;
 447
 448         commit_seq_old = v_read(config, &buf->commit_hot[idx].seq);
 449         while ((long) (commit_seq_old - commit_count) < 0)
 450                 commit_seq_old = v_cmpxchg(config, &buf->commit_hot[idx].seq,
 451                                            commit_seq_old, commit_count);
 452 }
 453
 454 extern int lib_ring_buffer_create(struct lib_ring_buffer *buf,
 455                                   struct channel_backend *chanb, int cpu);
 456 extern void lib_ring_buffer_free(struct lib_ring_buffer *buf);
 457
 458 /* Keep track of trap nesting inside ring buffer code */
 459 DECLARE_PER_CPU(unsigned int, lib_ring_buffer_nesting);
 460
 461 #endif /* _LIB_RING_BUFFER_FRONTEND_INTERNAL_H */