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 "config.h"
  22 #include "backend_types.h"
  23 #include "frontend_types.h"
  24
  25 /* Buffer offset macros */
  26
  27 /* buf_trunc mask selects only the buffer number. */
  28 static inline
  29 unsigned long buf_trunc(unsigned long offset, struct channel *chan)
  30 {
  31         return offset & ~(chan->backend.buf_size - 1);
  32
  33 }
  34
  35 /* Select the buffer number value (counter). */
  36 static inline
  37 unsigned long buf_trunc_val(unsigned long offset, struct channel *chan)
  38 {
  39         return buf_trunc(offset, chan) >> chan->backend.buf_size_order;
  40 }
  41
  42 /* buf_offset mask selects only the offset within the current buffer. */
  43 static inline
  44 unsigned long buf_offset(unsigned long offset, struct channel *chan)
  45 {
  46         return offset & (chan->backend.buf_size - 1);
  47 }
  48
  49 /* subbuf_offset mask selects the offset within the current subbuffer. */
  50 static inline
  51 unsigned long subbuf_offset(unsigned long offset, struct channel *chan)
  52 {
  53         return offset & (chan->backend.subbuf_size - 1);
  54 }
  55
  56 /* subbuf_trunc mask selects the subbuffer number. */
  57 static inline
  58 unsigned long subbuf_trunc(unsigned long offset, struct channel *chan)
  59 {
  60         return offset & ~(chan->backend.subbuf_size - 1);
  61 }
  62
  63 /* subbuf_align aligns the offset to the next subbuffer. */
  64 static inline
  65 unsigned long subbuf_align(unsigned long offset, struct channel *chan)
  66 {
  67         return (offset + chan->backend.subbuf_size)
  68                & ~(chan->backend.subbuf_size - 1);
  69 }
  70
  71 /* subbuf_index returns the index of the current subbuffer within the buffer. */
  72 static inline
  73 unsigned long subbuf_index(unsigned long offset, struct channel *chan)
  74 {
  75         return buf_offset(offset, chan) >> chan->backend.subbuf_size_order;
  76 }
  77
  78 /*
  79  * Last TSC comparison functions. Check if the current TSC overflows tsc_bits
  80  * bits from the last TSC read. When overflows are detected, the full 64-bit
  81  * timestamp counter should be written in the record header. Reads and writes
  82  * last_tsc atomically.
  83  */
  84
  85 #if (CAA_BITS_PER_LONG == 32)
  86 static inline
  87 void save_last_tsc(const struct lib_ring_buffer_config *config,
  88                    struct lib_ring_buffer *buf, u64 tsc)
  89 {
  90         if (config->tsc_bits == 0 || config->tsc_bits == 64)
  91                 return;
  92
  93         /*
  94          * Ensure the compiler performs this update in a single instruction.
  95          */
  96         v_set(config, &buf->last_tsc, (unsigned long)(tsc >> config->tsc_bits));
  97 }
  98
  99 static inline
 100 int last_tsc_overflow(const struct lib_ring_buffer_config *config,
 101                       struct lib_ring_buffer *buf, u64 tsc)
 102 {
 103         unsigned long tsc_shifted;
 104
 105         if (config->tsc_bits == 0 || config->tsc_bits == 64)
 106                 return 0;
 107
 108         tsc_shifted = (unsigned long)(tsc >> config->tsc_bits);
 109         if (unlikely(tsc_shifted
 110                      - (unsigned long)v_read(config, &buf->last_tsc)))
 111                 return 1;
 112         else
 113                 return 0;
 114 }
 115 #else
 116 static inline
 117 void save_last_tsc(const struct lib_ring_buffer_config *config,
 118                    struct lib_ring_buffer *buf, u64 tsc)
 119 {
 120         if (config->tsc_bits == 0 || config->tsc_bits == 64)
 121                 return;
 122
 123         v_set(config, &buf->last_tsc, (unsigned long)tsc);
 124 }
 125
 126 static inline
 127 int last_tsc_overflow(const struct lib_ring_buffer_config *config,
 128                       struct lib_ring_buffer *buf, u64 tsc)
 129 {
 130         if (config->tsc_bits == 0 || config->tsc_bits == 64)
 131                 return 0;
 132
 133         if (unlikely((tsc - v_read(config, &buf->last_tsc))
 134                      >> config->tsc_bits))
 135                 return 1;
 136         else
 137                 return 0;
 138 }
 139 #endif
 140
 141 extern
 142 int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx);
 143
 144 extern
 145 void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf,
 146                                  enum switch_mode mode);
 147
 148 /* Buffer write helpers */
 149
 150 static inline
 151 void lib_ring_buffer_reserve_push_reader(struct lib_ring_buffer *buf,
 152                                          struct channel *chan,
 153                                          unsigned long offset)
 154 {
 155         unsigned long consumed_old, consumed_new;
 156
 157         do {
 158                 consumed_old = uatomic_read(&buf->consumed);
 159                 /*
 160                  * If buffer is in overwrite mode, push the reader consumed
 161                  * count if the write position has reached it and we are not
 162                  * at the first iteration (don't push the reader farther than
 163                  * the writer). This operation can be done concurrently by many
 164                  * writers in the same buffer, the writer being at the farthest
 165                  * write position sub-buffer index in the buffer being the one
 166                  * which will win this loop.
 167                  */
 168                 if (unlikely(subbuf_trunc(offset, chan)
 169                               - subbuf_trunc(consumed_old, chan)
 170                              >= chan->backend.buf_size))
 171                         consumed_new = subbuf_align(consumed_old, chan);
 172                 else
 173                         return;
 174         } while (unlikely(uatomic_cmpxchg(&buf->consumed, consumed_old,
 175                                               consumed_new) != consumed_old));
 176 }
 177
 178 static inline
 179 void lib_ring_buffer_vmcore_check_deliver(const struct lib_ring_buffer_config *config,
 180                                           struct lib_ring_buffer *buf,
 181                                           unsigned long commit_count,
 182                                           unsigned long idx)
 183 {
 184         if (config->oops == RING_BUFFER_OOPS_CONSISTENCY)
 185                 v_set(config, &buf->commit_hot[idx].seq, commit_count);
 186 }
 187
 188 static inline
 189 int lib_ring_buffer_poll_deliver(const struct lib_ring_buffer_config *config,
 190                                  struct lib_ring_buffer *buf,
 191                                  struct channel *chan)
 192 {
 193         unsigned long consumed_old, consumed_idx, commit_count, write_offset;
 194
 195         consumed_old = uatomic_read(&buf->consumed);
 196         consumed_idx = subbuf_index(consumed_old, chan);
 197         commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb);
 198         /*
 199          * No memory barrier here, since we are only interested
 200          * in a statistically correct polling result. The next poll will
 201          * get the data is we are racing. The mb() that ensures correct
 202          * memory order is in get_subbuf.
 203          */
 204         write_offset = v_read(config, &buf->offset);
 205
 206         /*
 207          * Check that the subbuffer we are trying to consume has been
 208          * already fully committed.
 209          */
 210
 211         if (((commit_count - chan->backend.subbuf_size)
 212              & chan->commit_count_mask)
 213             - (buf_trunc(consumed_old, chan)
 214                >> chan->backend.num_subbuf_order)
 215             != 0)
 216                 return 0;
 217
 218         /*
 219          * Check that we are not about to read the same subbuffer in
 220          * which the writer head is.
 221          */
 222         if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan)
 223             == 0)
 224                 return 0;
 225
 226         return 1;
 227
 228 }
 229
 230 static inline
 231 int lib_ring_buffer_pending_data(const struct lib_ring_buffer_config *config,
 232                                  struct lib_ring_buffer *buf,
 233                                  struct channel *chan)
 234 {
 235         return !!subbuf_offset(v_read(config, &buf->offset), chan);
 236 }
 237
 238 static inline
 239 unsigned long lib_ring_buffer_get_data_size(const struct lib_ring_buffer_config *config,
 240                                             struct lib_ring_buffer *buf,
 241                                             unsigned long idx)
 242 {
 243         return subbuffer_get_data_size(config, &buf->backend, idx);
 244 }
 245
 246 /*
 247  * Check if all space reservation in a buffer have been committed. This helps
 248  * knowing if an execution context is nested (for per-cpu buffers only).
 249  * This is a very specific ftrace use-case, so we keep this as "internal" API.
 250  */
 251 static inline
 252 int lib_ring_buffer_reserve_committed(const struct lib_ring_buffer_config *config,
 253                                       struct lib_ring_buffer *buf,
 254                                       struct channel *chan)
 255 {
 256         unsigned long offset, idx, commit_count;
 257
 258         CHAN_WARN_ON(chan, config->alloc != RING_BUFFER_ALLOC_PER_CPU);
 259         CHAN_WARN_ON(chan, config->sync != RING_BUFFER_SYNC_PER_CPU);
 260
 261         /*
 262          * Read offset and commit count in a loop so they are both read
 263          * atomically wrt interrupts. By deal with interrupt concurrency by
 264          * restarting both reads if the offset has been pushed. Note that given
 265          * we only have to deal with interrupt concurrency here, an interrupt
 266          * modifying the commit count will also modify "offset", so it is safe
 267          * to only check for offset modifications.
 268          */
 269         do {
 270                 offset = v_read(config, &buf->offset);
 271                 idx = subbuf_index(offset, chan);
 272                 commit_count = v_read(config, &buf->commit_hot[idx].cc);
 273         } while (offset != v_read(config, &buf->offset));
 274
 275         return ((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
 276                      - (commit_count & chan->commit_count_mask) == 0);
 277 }
 278
 279 static inline
 280 void lib_ring_buffer_check_deliver(const struct lib_ring_buffer_config *config,
 281                                    struct lib_ring_buffer *buf,
 282                                    struct channel *chan,
 283                                    unsigned long offset,
 284                                    unsigned long commit_count,
 285                                    unsigned long idx)
 286 {
 287         unsigned long old_commit_count = commit_count
 288                                          - chan->backend.subbuf_size;
 289         u64 tsc;
 290
 291         /* Check if all commits have been done */
 292         if (unlikely((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
 293                      - (old_commit_count & chan->commit_count_mask) == 0)) {
 294                 /*
 295                  * If we succeeded at updating cc_sb below, we are the subbuffer
 296                  * writer delivering the subbuffer. Deals with concurrent
 297                  * updates of the "cc" value without adding a add_return atomic
 298                  * operation to the fast path.
 299                  *
 300                  * We are doing the delivery in two steps:
 301                  * - First, we cmpxchg() cc_sb to the new value
 302                  *   old_commit_count + 1. This ensures that we are the only
 303                  *   subbuffer user successfully filling the subbuffer, but we
 304                  *   do _not_ set the cc_sb value to "commit_count" yet.
 305                  *   Therefore, other writers that would wrap around the ring
 306                  *   buffer and try to start writing to our subbuffer would
 307                  *   have to drop records, because it would appear as
 308                  *   non-filled.
 309                  *   We therefore have exclusive access to the subbuffer control
 310                  *   structures.  This mutual exclusion with other writers is
 311                  *   crucially important to perform record overruns count in
 312                  *   flight recorder mode locklessly.
 313                  * - When we are ready to release the subbuffer (either for
 314                  *   reading or for overrun by other writers), we simply set the
 315                  *   cc_sb value to "commit_count" and perform delivery.
 316                  *
 317                  * The subbuffer size is least 2 bytes (minimum size: 1 page).
 318                  * This guarantees that old_commit_count + 1 != commit_count.
 319                  */
 320                 if (likely(v_cmpxchg(config, &buf->commit_cold[idx].cc_sb,
 321                                          old_commit_count, old_commit_count + 1)
 322                            == old_commit_count)) {
 323                         /*
 324                          * Start of exclusive subbuffer access. We are
 325                          * guaranteed to be the last writer in this subbuffer
 326                          * and any other writer trying to access this subbuffer
 327                          * in this state is required to drop records.
 328                          */
 329                         tsc = config->cb.ring_buffer_clock_read(chan);
 330                         v_add(config,
 331                               subbuffer_get_records_count(config,
 332                                                           &buf->backend, idx),
 333                               &buf->records_count);
 334                         v_add(config,
 335                               subbuffer_count_records_overrun(config,
 336                                                               &buf->backend,
 337                                                               idx),
 338                               &buf->records_overrun);
 339                         config->cb.buffer_end(buf, tsc, idx,
 340                                               lib_ring_buffer_get_data_size(config,
 341                                                                         buf,
 342                                                                         idx));
 343
 344                         /*
 345                          * Set noref flag and offset for this subbuffer id.
 346                          * Contains a memory barrier that ensures counter stores
 347                          * are ordered before set noref and offset.
 348                          */
 349                         lib_ring_buffer_set_noref_offset(config, &buf->backend, idx,
 350                                                          buf_trunc_val(offset, chan));
 351
 352                         /*
 353                          * Order set_noref and record counter updates before the
 354                          * end of subbuffer exclusive access. Orders with
 355                          * respect to writers coming into the subbuffer after
 356                          * wrap around, and also order wrt concurrent readers.
 357                          */
 358                         cmm_smp_mb();
 359                         /* End of exclusive subbuffer access */
 360                         v_set(config, &buf->commit_cold[idx].cc_sb,
 361                               commit_count);
 362                         lib_ring_buffer_vmcore_check_deliver(config, buf,
 363                                                          commit_count, idx);
 364
 365                         /*
 366                          * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
 367                          */
 368                         if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER
 369                             && uatomic_read(&buf->active_readers)
 370                             && lib_ring_buffer_poll_deliver(config, buf, chan)) {
 371                                 //wake_up_interruptible(&buf->read_wait);
 372                                 //wake_up_interruptible(&chan->read_wait);
 373                         }
 374
 375                 }
 376         }
 377 }
 378
 379 /*
 380  * lib_ring_buffer_write_commit_counter
 381  *
 382  * For flight recording. must be called after commit.
 383  * This function increments the subbuffer's commit_seq counter each time the
 384  * commit count reaches back the reserve offset (modulo subbuffer size). It is
 385  * useful for crash dump.
 386  */
 387 static inline
 388 void lib_ring_buffer_write_commit_counter(const struct lib_ring_buffer_config *config,
 389                                           struct lib_ring_buffer *buf,
 390                                           struct channel *chan,
 391                                           unsigned long idx,
 392                                           unsigned long buf_offset,
 393                                           unsigned long commit_count,
 394                                           size_t slot_size)
 395 {
 396         unsigned long offset, commit_seq_old;
 397
 398         if (config->oops != RING_BUFFER_OOPS_CONSISTENCY)
 399                 return;
 400
 401         offset = buf_offset + slot_size;
 402
 403         /*
 404          * subbuf_offset includes commit_count_mask. We can simply
 405          * compare the offsets within the subbuffer without caring about
 406          * buffer full/empty mismatch because offset is never zero here
 407          * (subbuffer header and record headers have non-zero length).
 408          */
 409         if (unlikely(subbuf_offset(offset - commit_count, chan)))
 410                 return;
 411
 412         commit_seq_old = v_read(config, &buf->commit_hot[idx].seq);
 413         while ((long) (commit_seq_old - commit_count) < 0)
 414                 commit_seq_old = v_cmpxchg(config, &buf->commit_hot[idx].seq,
 415                                            commit_seq_old, commit_count);
 416 }
 417
 418 extern int lib_ring_buffer_create(struct lib_ring_buffer *buf,
 419                                   struct channel_backend *chanb, int cpu);
 420 extern void lib_ring_buffer_free(struct lib_ring_buffer *buf);
 421
 422 /* Keep track of trap nesting inside ring buffer code */
 423 extern __thread unsigned int lib_ring_buffer_nesting;
 424
 425 #endif /* _LINUX_RING_BUFFER_FRONTEND_INTERNAL_H */