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 /* Buffer write helpers */
 160
 161 static inline
 162 void lib_ring_buffer_reserve_push_reader(struct lib_ring_buffer *buf,
 163                                          struct channel *chan,
 164                                          unsigned long offset)
 165 {
 166         unsigned long consumed_old, consumed_new;
 167
 168         do {
 169                 consumed_old = atomic_long_read(&buf->consumed);
 170                 /*
 171                  * If buffer is in overwrite mode, push the reader consumed
 172                  * count if the write position has reached it and we are not
 173                  * at the first iteration (don't push the reader farther than
 174                  * the writer). This operation can be done concurrently by many
 175                  * writers in the same buffer, the writer being at the farthest
 176                  * write position sub-buffer index in the buffer being the one
 177                  * which will win this loop.
 178                  */
 179                 if (unlikely(subbuf_trunc(offset, chan)
 180                               - subbuf_trunc(consumed_old, chan)
 181                              >= chan->backend.buf_size))
 182                         consumed_new = subbuf_align(consumed_old, chan);
 183                 else
 184                         return;
 185         } while (unlikely(atomic_long_cmpxchg(&buf->consumed, consumed_old,
 186                                               consumed_new) != consumed_old));
 187 }
 188
 189 static inline
 190 void lib_ring_buffer_vmcore_check_deliver(const struct lib_ring_buffer_config *config,
 191                                           struct lib_ring_buffer *buf,
 192                                           unsigned long commit_count,
 193                                           unsigned long idx)
 194 {
 195         if (config->oops == RING_BUFFER_OOPS_CONSISTENCY)
 196                 v_set(config, &buf->commit_hot[idx].seq, commit_count);
 197 }
 198
 199 static inline
 200 int lib_ring_buffer_poll_deliver(const struct lib_ring_buffer_config *config,
 201                                  struct lib_ring_buffer *buf,
 202                                  struct channel *chan)
 203 {
 204         unsigned long consumed_old, consumed_idx, commit_count, write_offset;
 205
 206         consumed_old = atomic_long_read(&buf->consumed);
 207         consumed_idx = subbuf_index(consumed_old, chan);
 208         commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb);
 209         /*
 210          * No memory barrier here, since we are only interested
 211          * in a statistically correct polling result. The next poll will
 212          * get the data is we are racing. The mb() that ensures correct
 213          * memory order is in get_subbuf.
 214          */
 215         write_offset = v_read(config, &buf->offset);
 216
 217         /*
 218          * Check that the subbuffer we are trying to consume has been
 219          * already fully committed.
 220          */
 221
 222         if (((commit_count - chan->backend.subbuf_size)
 223              & chan->commit_count_mask)
 224             - (buf_trunc(consumed_old, chan)
 225                >> chan->backend.num_subbuf_order)
 226             != 0)
 227                 return 0;
 228
 229         /*
 230          * Check that we are not about to read the same subbuffer in
 231          * which the writer head is.
 232          */
 233         if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan)
 234             == 0)
 235                 return 0;
 236
 237         return 1;
 238
 239 }
 240
 241 static inline
 242 int lib_ring_buffer_pending_data(const struct lib_ring_buffer_config *config,
 243                                  struct lib_ring_buffer *buf,
 244                                  struct channel *chan)
 245 {
 246         return !!subbuf_offset(v_read(config, &buf->offset), chan);
 247 }
 248
 249 static inline
 250 unsigned long lib_ring_buffer_get_data_size(const struct lib_ring_buffer_config *config,
 251                                             struct lib_ring_buffer *buf,
 252                                             unsigned long idx)
 253 {
 254         return subbuffer_get_data_size(config, &buf->backend, idx);
 255 }
 256
 257 /*
 258  * Check if all space reservation in a buffer have been committed. This helps
 259  * knowing if an execution context is nested (for per-cpu buffers only).
 260  * This is a very specific ftrace use-case, so we keep this as "internal" API.
 261  */
 262 static inline
 263 int lib_ring_buffer_reserve_committed(const struct lib_ring_buffer_config *config,
 264                                       struct lib_ring_buffer *buf,
 265                                       struct channel *chan)
 266 {
 267         unsigned long offset, idx, commit_count;
 268
 269         CHAN_WARN_ON(chan, config->alloc != RING_BUFFER_ALLOC_PER_CPU);
 270         CHAN_WARN_ON(chan, config->sync != RING_BUFFER_SYNC_PER_CPU);
 271
 272         /*
 273          * Read offset and commit count in a loop so they are both read
 274          * atomically wrt interrupts. By deal with interrupt concurrency by
 275          * restarting both reads if the offset has been pushed. Note that given
 276          * we only have to deal with interrupt concurrency here, an interrupt
 277          * modifying the commit count will also modify "offset", so it is safe
 278          * to only check for offset modifications.
 279          */
 280         do {
 281                 offset = v_read(config, &buf->offset);
 282                 idx = subbuf_index(offset, chan);
 283                 commit_count = v_read(config, &buf->commit_hot[idx].cc);
 284         } while (offset != v_read(config, &buf->offset));
 285
 286         return ((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
 287                      - (commit_count & chan->commit_count_mask) == 0);
 288 }
 289
 290 static inline
 291 void lib_ring_buffer_check_deliver(const struct lib_ring_buffer_config *config,
 292                                    struct lib_ring_buffer *buf,
 293                                    struct channel *chan,
 294                                    unsigned long offset,
 295                                    unsigned long commit_count,
 296                                    unsigned long idx)
 297 {
 298         unsigned long old_commit_count = commit_count
 299                                          - chan->backend.subbuf_size;
 300         u64 tsc;
 301
 302         /* Check if all commits have been done */
 303         if (unlikely((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
 304                      - (old_commit_count & chan->commit_count_mask) == 0)) {
 305                 /*
 306                  * If we succeeded at updating cc_sb below, we are the subbuffer
 307                  * writer delivering the subbuffer. Deals with concurrent
 308                  * updates of the "cc" value without adding a add_return atomic
 309                  * operation to the fast path.
 310                  *
 311                  * We are doing the delivery in two steps:
 312                  * - First, we cmpxchg() cc_sb to the new value
 313                  *   old_commit_count + 1. This ensures that we are the only
 314                  *   subbuffer user successfully filling the subbuffer, but we
 315                  *   do _not_ set the cc_sb value to "commit_count" yet.
 316                  *   Therefore, other writers that would wrap around the ring
 317                  *   buffer and try to start writing to our subbuffer would
 318                  *   have to drop records, because it would appear as
 319                  *   non-filled.
 320                  *   We therefore have exclusive access to the subbuffer control
 321                  *   structures.  This mutual exclusion with other writers is
 322                  *   crucially important to perform record overruns count in
 323                  *   flight recorder mode locklessly.
 324                  * - When we are ready to release the subbuffer (either for
 325                  *   reading or for overrun by other writers), we simply set the
 326                  *   cc_sb value to "commit_count" and perform delivery.
 327                  *
 328                  * The subbuffer size is least 2 bytes (minimum size: 1 page).
 329                  * This guarantees that old_commit_count + 1 != commit_count.
 330                  */
 331                 if (likely(v_cmpxchg(config, &buf->commit_cold[idx].cc_sb,
 332                                          old_commit_count, old_commit_count + 1)
 333                            == old_commit_count)) {
 334                         /*
 335                          * Start of exclusive subbuffer access. We are
 336                          * guaranteed to be the last writer in this subbuffer
 337                          * and any other writer trying to access this subbuffer
 338                          * in this state is required to drop records.
 339                          */
 340                         tsc = config->cb.ring_buffer_clock_read(chan);
 341                         v_add(config,
 342                               subbuffer_get_records_count(config,
 343                                                           &buf->backend, idx),
 344                               &buf->records_count);
 345                         v_add(config,
 346                               subbuffer_count_records_overrun(config,
 347                                                               &buf->backend,
 348                                                               idx),
 349                               &buf->records_overrun);
 350                         config->cb.buffer_end(buf, tsc, idx,
 351                                               lib_ring_buffer_get_data_size(config,
 352                                                                         buf,
 353                                                                         idx));
 354
 355                         /*
 356                          * Set noref flag and offset for this subbuffer id.
 357                          * Contains a memory barrier that ensures counter stores
 358                          * are ordered before set noref and offset.
 359                          */
 360                         lib_ring_buffer_set_noref_offset(config, &buf->backend, idx,
 361                                                          buf_trunc_val(offset, chan));
 362
 363                         /*
 364                          * Order set_noref and record counter updates before the
 365                          * end of subbuffer exclusive access. Orders with
 366                          * respect to writers coming into the subbuffer after
 367                          * wrap around, and also order wrt concurrent readers.
 368                          */
 369                         smp_mb();
 370                         /* End of exclusive subbuffer access */
 371                         v_set(config, &buf->commit_cold[idx].cc_sb,
 372                               commit_count);
 373                         lib_ring_buffer_vmcore_check_deliver(config, buf,
 374                                                          commit_count, idx);
 375
 376                         /*
 377                          * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
 378                          */
 379                         if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER
 380                             && atomic_long_read(&buf->active_readers)
 381                             && lib_ring_buffer_poll_deliver(config, buf, chan)) {
 382                                 wake_up_interruptible(&buf->read_wait);
 383                                 wake_up_interruptible(&chan->read_wait);
 384                         }
 385
 386                 }
 387         }
 388 }
 389
 390 /*
 391  * lib_ring_buffer_write_commit_counter
 392  *
 393  * For flight recording. must be called after commit.
 394  * This function increments the subbuffer's commit_seq counter each time the
 395  * commit count reaches back the reserve offset (modulo subbuffer size). It is
 396  * useful for crash dump.
 397  */
 398 static inline
 399 void lib_ring_buffer_write_commit_counter(const struct lib_ring_buffer_config *config,
 400                                           struct lib_ring_buffer *buf,
 401                                           struct channel *chan,
 402                                           unsigned long idx,
 403                                           unsigned long buf_offset,
 404                                           unsigned long commit_count,
 405                                           size_t slot_size)
 406 {
 407         unsigned long offset, commit_seq_old;
 408
 409         if (config->oops != RING_BUFFER_OOPS_CONSISTENCY)
 410                 return;
 411
 412         offset = buf_offset + slot_size;
 413
 414         /*
 415          * subbuf_offset includes commit_count_mask. We can simply
 416          * compare the offsets within the subbuffer without caring about
 417          * buffer full/empty mismatch because offset is never zero here
 418          * (subbuffer header and record headers have non-zero length).
 419          */
 420         if (unlikely(subbuf_offset(offset - commit_count, chan)))
 421                 return;
 422
 423         commit_seq_old = v_read(config, &buf->commit_hot[idx].seq);
 424         while ((long) (commit_seq_old - commit_count) < 0)
 425                 commit_seq_old = v_cmpxchg(config, &buf->commit_hot[idx].seq,
 426                                            commit_seq_old, commit_count);
 427 }
 428
 429 extern int lib_ring_buffer_create(struct lib_ring_buffer *buf,
 430                                   struct channel_backend *chanb, int cpu);
 431 extern void lib_ring_buffer_free(struct lib_ring_buffer *buf);
 432
 433 /* Keep track of trap nesting inside ring buffer code */
 434 DECLARE_PER_CPU(unsigned int, lib_ring_buffer_nesting);
 435
 436 #endif /* _LIB_RING_BUFFER_FRONTEND_INTERNAL_H */