lib/ringbuffer/backend_internal.h

   1 #ifndef _LIB_RING_BUFFER_BACKEND_INTERNAL_H
   2 #define _LIB_RING_BUFFER_BACKEND_INTERNAL_H
   3
   4 /*
   5  * lib/ringbuffer/backend_internal.h
   6  *
   7  * Ring buffer backend (internal helpers).
   8  *
   9  * Copyright (C) 2008-2012 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 #include <wrapper/ringbuffer/config.h>
  27 #include <wrapper/ringbuffer/backend_types.h>
  28 #include <wrapper/ringbuffer/frontend_types.h>
  29 #include <linux/string.h>
  30 #include <linux/uaccess.h>
  31
  32 /* Ring buffer backend API presented to the frontend */
  33
  34 /* Ring buffer and channel backend create/free */
  35
  36 int lib_ring_buffer_backend_create(struct lib_ring_buffer_backend *bufb,
  37                                    struct channel_backend *chan, int cpu);
  38 void channel_backend_unregister_notifiers(struct channel_backend *chanb);
  39 void lib_ring_buffer_backend_free(struct lib_ring_buffer_backend *bufb);
  40 int channel_backend_init(struct channel_backend *chanb,
  41                          const char *name,
  42                          const struct lib_ring_buffer_config *config,
  43                          void *priv, size_t subbuf_size,
  44                          size_t num_subbuf);
  45 void channel_backend_free(struct channel_backend *chanb);
  46
  47 void lib_ring_buffer_backend_reset(struct lib_ring_buffer_backend *bufb);
  48 void channel_backend_reset(struct channel_backend *chanb);
  49
  50 int lib_ring_buffer_backend_init(void);
  51 void lib_ring_buffer_backend_exit(void);
  52
  53 extern void _lib_ring_buffer_write(struct lib_ring_buffer_backend *bufb,
  54                                    size_t offset, const void *src, size_t len,
  55                                    size_t pagecpy);
  56 extern void _lib_ring_buffer_memset(struct lib_ring_buffer_backend *bufb,
  57                                     size_t offset, int c, size_t len,
  58                                     size_t pagecpy);
  59 extern void _lib_ring_buffer_strcpy(struct lib_ring_buffer_backend *bufb,
  60                                    size_t offset, const char *src, size_t len,
  61                                    size_t pagecpy, int pad);
  62 extern void _lib_ring_buffer_copy_from_user_inatomic(struct lib_ring_buffer_backend *bufb,
  63                                             size_t offset, const void *src,
  64                                             size_t len, size_t pagecpy);
  65 extern void _lib_ring_buffer_strcpy_from_user_inatomic(struct lib_ring_buffer_backend *bufb,
  66                 size_t offset, const char __user *src, size_t len,
  67                 size_t pagecpy, int pad);
  68
  69 /*
  70  * Subbuffer ID bits for overwrite mode. Need to fit within a single word to be
  71  * exchanged atomically.
  72  *
  73  * Top half word, except lowest bit, belongs to "offset", which is used to keep
  74  * to count the produced buffers.  For overwrite mode, this provides the
  75  * consumer with the capacity to read subbuffers in order, handling the
  76  * situation where producers would write up to 2^15 buffers (or 2^31 for 64-bit
  77  * systems) concurrently with a single execution of get_subbuf (between offset
  78  * sampling and subbuffer ID exchange).
  79  */
  80
  81 #define HALF_ULONG_BITS         (BITS_PER_LONG >> 1)
  82
  83 #define SB_ID_OFFSET_SHIFT      (HALF_ULONG_BITS + 1)
  84 #define SB_ID_OFFSET_COUNT      (1UL << SB_ID_OFFSET_SHIFT)
  85 #define SB_ID_OFFSET_MASK       (~(SB_ID_OFFSET_COUNT - 1))
  86 /*
  87  * Lowest bit of top word half belongs to noref. Used only for overwrite mode.
  88  */
  89 #define SB_ID_NOREF_SHIFT       (SB_ID_OFFSET_SHIFT - 1)
  90 #define SB_ID_NOREF_COUNT       (1UL << SB_ID_NOREF_SHIFT)
  91 #define SB_ID_NOREF_MASK        SB_ID_NOREF_COUNT
  92 /*
  93  * In overwrite mode: lowest half of word is used for index.
  94  * Limit of 2^16 subbuffers per buffer on 32-bit, 2^32 on 64-bit.
  95  * In producer-consumer mode: whole word used for index.
  96  */
  97 #define SB_ID_INDEX_SHIFT       0
  98 #define SB_ID_INDEX_COUNT       (1UL << SB_ID_INDEX_SHIFT)
  99 #define SB_ID_INDEX_MASK        (SB_ID_NOREF_COUNT - 1)
 100
 101 /*
 102  * Construct the subbuffer id from offset, index and noref. Use only the index
 103  * for producer-consumer mode (offset and noref are only used in overwrite
 104  * mode).
 105  */
 106 static inline
 107 unsigned long subbuffer_id(const struct lib_ring_buffer_config *config,
 108                            unsigned long offset, unsigned long noref,
 109                            unsigned long index)
 110 {
 111         if (config->mode == RING_BUFFER_OVERWRITE)
 112                 return (offset << SB_ID_OFFSET_SHIFT)
 113                        | (noref << SB_ID_NOREF_SHIFT)
 114                        | index;
 115         else
 116                 return index;
 117 }
 118
 119 /*
 120  * Compare offset with the offset contained within id. Return 1 if the offset
 121  * bits are identical, else 0.
 122  */
 123 static inline
 124 int subbuffer_id_compare_offset(const struct lib_ring_buffer_config *config,
 125                                 unsigned long id, unsigned long offset)
 126 {
 127         return (id & SB_ID_OFFSET_MASK) == (offset << SB_ID_OFFSET_SHIFT);
 128 }
 129
 130 static inline
 131 unsigned long subbuffer_id_get_index(const struct lib_ring_buffer_config *config,
 132                                      unsigned long id)
 133 {
 134         if (config->mode == RING_BUFFER_OVERWRITE)
 135                 return id & SB_ID_INDEX_MASK;
 136         else
 137                 return id;
 138 }
 139
 140 static inline
 141 unsigned long subbuffer_id_is_noref(const struct lib_ring_buffer_config *config,
 142                                     unsigned long id)
 143 {
 144         if (config->mode == RING_BUFFER_OVERWRITE)
 145                 return !!(id & SB_ID_NOREF_MASK);
 146         else
 147                 return 1;
 148 }
 149
 150 /*
 151  * Only used by reader on subbuffer ID it has exclusive access to. No volatile
 152  * needed.
 153  */
 154 static inline
 155 void subbuffer_id_set_noref(const struct lib_ring_buffer_config *config,
 156                             unsigned long *id)
 157 {
 158         if (config->mode == RING_BUFFER_OVERWRITE)
 159                 *id |= SB_ID_NOREF_MASK;
 160 }
 161
 162 static inline
 163 void subbuffer_id_set_noref_offset(const struct lib_ring_buffer_config *config,
 164                                    unsigned long *id, unsigned long offset)
 165 {
 166         unsigned long tmp;
 167
 168         if (config->mode == RING_BUFFER_OVERWRITE) {
 169                 tmp = *id;
 170                 tmp &= ~SB_ID_OFFSET_MASK;
 171                 tmp |= offset << SB_ID_OFFSET_SHIFT;
 172                 tmp |= SB_ID_NOREF_MASK;
 173                 /* Volatile store, read concurrently by readers. */
 174                 ACCESS_ONCE(*id) = tmp;
 175         }
 176 }
 177
 178 /* No volatile access, since already used locally */
 179 static inline
 180 void subbuffer_id_clear_noref(const struct lib_ring_buffer_config *config,
 181                               unsigned long *id)
 182 {
 183         if (config->mode == RING_BUFFER_OVERWRITE)
 184                 *id &= ~SB_ID_NOREF_MASK;
 185 }
 186
 187 /*
 188  * For overwrite mode, cap the number of subbuffers per buffer to:
 189  * 2^16 on 32-bit architectures
 190  * 2^32 on 64-bit architectures
 191  * This is required to fit in the index part of the ID. Return 0 on success,
 192  * -EPERM on failure.
 193  */
 194 static inline
 195 int subbuffer_id_check_index(const struct lib_ring_buffer_config *config,
 196                              unsigned long num_subbuf)
 197 {
 198         if (config->mode == RING_BUFFER_OVERWRITE)
 199                 return (num_subbuf > (1UL << HALF_ULONG_BITS)) ? -EPERM : 0;
 200         else
 201                 return 0;
 202 }
 203
 204 /*
 205  * The ring buffer can count events recorded and overwritten per buffer,
 206  * but it is disabled by default due to its performance overhead.
 207  */
 208 #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
 209 static inline
 210 void subbuffer_count_record(const struct lib_ring_buffer_config *config,
 211                             struct lib_ring_buffer_backend *bufb,
 212                             unsigned long idx)
 213 {
 214         unsigned long sb_bindex;
 215
 216         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 217         v_inc(config, &bufb->array[sb_bindex]->records_commit);
 218 }
 219 #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
 220 static inline
 221 void subbuffer_count_record(const struct lib_ring_buffer_config *config,
 222                             struct lib_ring_buffer_backend *bufb,
 223                             unsigned long idx)
 224 {
 225 }
 226 #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
 227
 228 /*
 229  * Reader has exclusive subbuffer access for record consumption. No need to
 230  * perform the decrement atomically.
 231  */
 232 static inline
 233 void subbuffer_consume_record(const struct lib_ring_buffer_config *config,
 234                               struct lib_ring_buffer_backend *bufb)
 235 {
 236         unsigned long sb_bindex;
 237
 238         sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
 239         CHAN_WARN_ON(bufb->chan,
 240                      !v_read(config, &bufb->array[sb_bindex]->records_unread));
 241         /* Non-atomic decrement protected by exclusive subbuffer access */
 242         _v_dec(config, &bufb->array[sb_bindex]->records_unread);
 243         v_inc(config, &bufb->records_read);
 244 }
 245
 246 static inline
 247 unsigned long subbuffer_get_records_count(
 248                                 const struct lib_ring_buffer_config *config,
 249                                 struct lib_ring_buffer_backend *bufb,
 250                                 unsigned long idx)
 251 {
 252         unsigned long sb_bindex;
 253
 254         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 255         return v_read(config, &bufb->array[sb_bindex]->records_commit);
 256 }
 257
 258 /*
 259  * Must be executed at subbuffer delivery when the writer has _exclusive_
 260  * subbuffer access. See lib_ring_buffer_check_deliver() for details.
 261  * lib_ring_buffer_get_records_count() must be called to get the records
 262  * count before this function, because it resets the records_commit
 263  * count.
 264  */
 265 static inline
 266 unsigned long subbuffer_count_records_overrun(
 267                                 const struct lib_ring_buffer_config *config,
 268                                 struct lib_ring_buffer_backend *bufb,
 269                                 unsigned long idx)
 270 {
 271         struct lib_ring_buffer_backend_pages *pages;
 272         unsigned long overruns, sb_bindex;
 273
 274         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 275         pages = bufb->array[sb_bindex];
 276         overruns = v_read(config, &pages->records_unread);
 277         v_set(config, &pages->records_unread,
 278               v_read(config, &pages->records_commit));
 279         v_set(config, &pages->records_commit, 0);
 280
 281         return overruns;
 282 }
 283
 284 static inline
 285 void subbuffer_set_data_size(const struct lib_ring_buffer_config *config,
 286                              struct lib_ring_buffer_backend *bufb,
 287                              unsigned long idx,
 288                              unsigned long data_size)
 289 {
 290         struct lib_ring_buffer_backend_pages *pages;
 291         unsigned long sb_bindex;
 292
 293         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 294         pages = bufb->array[sb_bindex];
 295         pages->data_size = data_size;
 296 }
 297
 298 static inline
 299 unsigned long subbuffer_get_read_data_size(
 300                                 const struct lib_ring_buffer_config *config,
 301                                 struct lib_ring_buffer_backend *bufb)
 302 {
 303         struct lib_ring_buffer_backend_pages *pages;
 304         unsigned long sb_bindex;
 305
 306         sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
 307         pages = bufb->array[sb_bindex];
 308         return pages->data_size;
 309 }
 310
 311 static inline
 312 unsigned long subbuffer_get_data_size(
 313                                 const struct lib_ring_buffer_config *config,
 314                                 struct lib_ring_buffer_backend *bufb,
 315                                 unsigned long idx)
 316 {
 317         struct lib_ring_buffer_backend_pages *pages;
 318         unsigned long sb_bindex;
 319
 320         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 321         pages = bufb->array[sb_bindex];
 322         return pages->data_size;
 323 }
 324
 325 static inline
 326 void subbuffer_inc_packet_count(const struct lib_ring_buffer_config *config,
 327                                 struct lib_ring_buffer_backend *bufb,
 328                                 unsigned long idx)
 329 {
 330         bufb->buf_cnt[idx].seq_cnt++;
 331 }
 332
 333 /**
 334  * lib_ring_buffer_clear_noref - Clear the noref subbuffer flag, called by
 335  *                               writer.
 336  */
 337 static inline
 338 void lib_ring_buffer_clear_noref(const struct lib_ring_buffer_config *config,
 339                                  struct lib_ring_buffer_backend *bufb,
 340                                  unsigned long idx)
 341 {
 342         unsigned long id, new_id;
 343
 344         if (config->mode != RING_BUFFER_OVERWRITE)
 345                 return;
 346
 347         /*
 348          * Performing a volatile access to read the sb_pages, because we want to
 349          * read a coherent version of the pointer and the associated noref flag.
 350          */
 351         id = ACCESS_ONCE(bufb->buf_wsb[idx].id);
 352         for (;;) {
 353                 /* This check is called on the fast path for each record. */
 354                 if (likely(!subbuffer_id_is_noref(config, id))) {
 355                         /*
 356                          * Store after load dependency ordering the writes to
 357                          * the subbuffer after load and test of the noref flag
 358                          * matches the memory barrier implied by the cmpxchg()
 359                          * in update_read_sb_index().
 360                          */
 361                         return; /* Already writing to this buffer */
 362                 }
 363                 new_id = id;
 364                 subbuffer_id_clear_noref(config, &new_id);
 365                 new_id = cmpxchg(&bufb->buf_wsb[idx].id, id, new_id);
 366                 if (likely(new_id == id))
 367                         break;
 368                 id = new_id;
 369         }
 370 }
 371
 372 /**
 373  * lib_ring_buffer_set_noref_offset - Set the noref subbuffer flag and offset,
 374  *                                    called by writer.
 375  */
 376 static inline
 377 void lib_ring_buffer_set_noref_offset(const struct lib_ring_buffer_config *config,
 378                                       struct lib_ring_buffer_backend *bufb,
 379                                       unsigned long idx, unsigned long offset)
 380 {
 381         if (config->mode != RING_BUFFER_OVERWRITE)
 382                 return;
 383
 384         /*
 385          * Because ring_buffer_set_noref() is only called by a single thread
 386          * (the one which updated the cc_sb value), there are no concurrent
 387          * updates to take care of: other writers have not updated cc_sb, so
 388          * they cannot set the noref flag, and concurrent readers cannot modify
 389          * the pointer because the noref flag is not set yet.
 390          * The smp_wmb() in ring_buffer_commit() takes care of ordering writes
 391          * to the subbuffer before this set noref operation.
 392          * subbuffer_set_noref() uses a volatile store to deal with concurrent
 393          * readers of the noref flag.
 394          */
 395         CHAN_WARN_ON(bufb->chan,
 396                      subbuffer_id_is_noref(config, bufb->buf_wsb[idx].id));
 397         /*
 398          * Memory barrier that ensures counter stores are ordered before set
 399          * noref and offset.
 400          */
 401         smp_mb();
 402         subbuffer_id_set_noref_offset(config, &bufb->buf_wsb[idx].id, offset);
 403 }
 404
 405 /**
 406  * update_read_sb_index - Read-side subbuffer index update.
 407  */
 408 static inline
 409 int update_read_sb_index(const struct lib_ring_buffer_config *config,
 410                          struct lib_ring_buffer_backend *bufb,
 411                          struct channel_backend *chanb,
 412                          unsigned long consumed_idx,
 413                          unsigned long consumed_count)
 414 {
 415         unsigned long old_id, new_id;
 416
 417         if (config->mode == RING_BUFFER_OVERWRITE) {
 418                 /*
 419                  * Exchange the target writer subbuffer with our own unused
 420                  * subbuffer. No need to use ACCESS_ONCE() here to read the
 421                  * old_wpage, because the value read will be confirmed by the
 422                  * following cmpxchg().
 423                  */
 424                 old_id = bufb->buf_wsb[consumed_idx].id;
 425                 if (unlikely(!subbuffer_id_is_noref(config, old_id)))
 426                         return -EAGAIN;
 427                 /*
 428                  * Make sure the offset count we are expecting matches the one
 429                  * indicated by the writer.
 430                  */
 431                 if (unlikely(!subbuffer_id_compare_offset(config, old_id,
 432                                                           consumed_count)))
 433                         return -EAGAIN;
 434                 CHAN_WARN_ON(bufb->chan,
 435                              !subbuffer_id_is_noref(config, bufb->buf_rsb.id));
 436                 subbuffer_id_set_noref_offset(config, &bufb->buf_rsb.id,
 437                                               consumed_count);
 438                 new_id = cmpxchg(&bufb->buf_wsb[consumed_idx].id, old_id,
 439                                  bufb->buf_rsb.id);
 440                 if (unlikely(old_id != new_id))
 441                         return -EAGAIN;
 442                 bufb->buf_rsb.id = new_id;
 443         } else {
 444                 /* No page exchange, use the writer page directly */
 445                 bufb->buf_rsb.id = bufb->buf_wsb[consumed_idx].id;
 446         }
 447         return 0;
 448 }
 449
 450 /*
 451  * Use the architecture-specific memcpy implementation for constant-sized
 452  * inputs, but rely on an inline memcpy for length statically unknown.
 453  * The function call to memcpy is just way too expensive for a fast path.
 454  */
 455 #define lib_ring_buffer_do_copy(config, dest, src, len)         \
 456 do {                                                            \
 457         size_t __len = (len);                                   \
 458         if (__builtin_constant_p(len))                          \
 459                 memcpy(dest, src, __len);                       \
 460         else                                                    \
 461                 inline_memcpy(dest, src, __len);                \
 462 } while (0)
 463
 464 /*
 465  * We use __copy_from_user_inatomic to copy userspace data since we already
 466  * did the access_ok for the whole range.
 467  *
 468  * Return 0 if OK, nonzero on error.
 469  */
 470 static inline
 471 unsigned long lib_ring_buffer_do_copy_from_user_inatomic(void *dest,
 472                                                 const void __user *src,
 473                                                 unsigned long len)
 474 {
 475         return __copy_from_user_inatomic(dest, src, len);
 476 }
 477
 478 /*
 479  * write len bytes to dest with c
 480  */
 481 static inline
 482 void lib_ring_buffer_do_memset(char *dest, int c,
 483         unsigned long len)
 484 {
 485         unsigned long i;
 486
 487         for (i = 0; i < len; i++)
 488                 dest[i] = c;
 489 }
 490
 491 #endif /* _LIB_RING_BUFFER_BACKEND_INTERNAL_H */