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 static inline
 205 void lib_ring_buffer_backend_get_pages(const struct lib_ring_buffer_config *config,
 206                         struct lib_ring_buffer_ctx *ctx,
 207                         struct lib_ring_buffer_backend_pages **backend_pages)
 208 {
 209         struct lib_ring_buffer_backend *bufb = &ctx->buf->backend;
 210         struct channel_backend *chanb = &ctx->chan->backend;
 211         size_t sbidx, offset = ctx->buf_offset;
 212         unsigned long sb_bindex, id;
 213         struct lib_ring_buffer_backend_pages *rpages;
 214
 215         offset &= chanb->buf_size - 1;
 216         sbidx = offset >> chanb->subbuf_size_order;
 217         id = bufb->buf_wsb[sbidx].id;
 218         sb_bindex = subbuffer_id_get_index(config, id);
 219         rpages = bufb->array[sb_bindex];
 220         CHAN_WARN_ON(ctx->chan,
 221                      config->mode == RING_BUFFER_OVERWRITE
 222                      && subbuffer_id_is_noref(config, id));
 223         *backend_pages = rpages;
 224 }
 225
 226 /* Get backend pages from cache. */
 227 static inline
 228 struct lib_ring_buffer_backend_pages *
 229         lib_ring_buffer_get_backend_pages_from_ctx(const struct lib_ring_buffer_config *config,
 230                 struct lib_ring_buffer_ctx *ctx)
 231 {
 232         return ctx->backend_pages;
 233 }
 234
 235 /*
 236  * The ring buffer can count events recorded and overwritten per buffer,
 237  * but it is disabled by default due to its performance overhead.
 238  */
 239 #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
 240 static inline
 241 void subbuffer_count_record(const struct lib_ring_buffer_config *config,
 242                             struct lib_ring_buffer_backend *bufb,
 243                             unsigned long idx)
 244 {
 245         unsigned long sb_bindex;
 246
 247         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 248         v_inc(config, &bufb->array[sb_bindex]->records_commit);
 249 }
 250 #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
 251 static inline
 252 void subbuffer_count_record(const struct lib_ring_buffer_config *config,
 253                             struct lib_ring_buffer_backend *bufb,
 254                             unsigned long idx)
 255 {
 256 }
 257 #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
 258
 259 /*
 260  * Reader has exclusive subbuffer access for record consumption. No need to
 261  * perform the decrement atomically.
 262  */
 263 static inline
 264 void subbuffer_consume_record(const struct lib_ring_buffer_config *config,
 265                               struct lib_ring_buffer_backend *bufb)
 266 {
 267         unsigned long sb_bindex;
 268
 269         sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
 270         CHAN_WARN_ON(bufb->chan,
 271                      !v_read(config, &bufb->array[sb_bindex]->records_unread));
 272         /* Non-atomic decrement protected by exclusive subbuffer access */
 273         _v_dec(config, &bufb->array[sb_bindex]->records_unread);
 274         v_inc(config, &bufb->records_read);
 275 }
 276
 277 static inline
 278 unsigned long subbuffer_get_records_count(
 279                                 const struct lib_ring_buffer_config *config,
 280                                 struct lib_ring_buffer_backend *bufb,
 281                                 unsigned long idx)
 282 {
 283         unsigned long sb_bindex;
 284
 285         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 286         return v_read(config, &bufb->array[sb_bindex]->records_commit);
 287 }
 288
 289 /*
 290  * Must be executed at subbuffer delivery when the writer has _exclusive_
 291  * subbuffer access. See lib_ring_buffer_check_deliver() for details.
 292  * lib_ring_buffer_get_records_count() must be called to get the records
 293  * count before this function, because it resets the records_commit
 294  * count.
 295  */
 296 static inline
 297 unsigned long subbuffer_count_records_overrun(
 298                                 const struct lib_ring_buffer_config *config,
 299                                 struct lib_ring_buffer_backend *bufb,
 300                                 unsigned long idx)
 301 {
 302         struct lib_ring_buffer_backend_pages *pages;
 303         unsigned long overruns, sb_bindex;
 304
 305         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 306         pages = bufb->array[sb_bindex];
 307         overruns = v_read(config, &pages->records_unread);
 308         v_set(config, &pages->records_unread,
 309               v_read(config, &pages->records_commit));
 310         v_set(config, &pages->records_commit, 0);
 311
 312         return overruns;
 313 }
 314
 315 static inline
 316 void subbuffer_set_data_size(const struct lib_ring_buffer_config *config,
 317                              struct lib_ring_buffer_backend *bufb,
 318                              unsigned long idx,
 319                              unsigned long data_size)
 320 {
 321         struct lib_ring_buffer_backend_pages *pages;
 322         unsigned long sb_bindex;
 323
 324         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 325         pages = bufb->array[sb_bindex];
 326         pages->data_size = data_size;
 327 }
 328
 329 static inline
 330 unsigned long subbuffer_get_read_data_size(
 331                                 const struct lib_ring_buffer_config *config,
 332                                 struct lib_ring_buffer_backend *bufb)
 333 {
 334         struct lib_ring_buffer_backend_pages *pages;
 335         unsigned long sb_bindex;
 336
 337         sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
 338         pages = bufb->array[sb_bindex];
 339         return pages->data_size;
 340 }
 341
 342 static inline
 343 unsigned long subbuffer_get_data_size(
 344                                 const struct lib_ring_buffer_config *config,
 345                                 struct lib_ring_buffer_backend *bufb,
 346                                 unsigned long idx)
 347 {
 348         struct lib_ring_buffer_backend_pages *pages;
 349         unsigned long sb_bindex;
 350
 351         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 352         pages = bufb->array[sb_bindex];
 353         return pages->data_size;
 354 }
 355
 356 static inline
 357 void subbuffer_inc_packet_count(const struct lib_ring_buffer_config *config,
 358                                 struct lib_ring_buffer_backend *bufb,
 359                                 unsigned long idx)
 360 {
 361         bufb->buf_cnt[idx].seq_cnt++;
 362 }
 363
 364 /**
 365  * lib_ring_buffer_clear_noref - Clear the noref subbuffer flag, called by
 366  *                               writer.
 367  */
 368 static inline
 369 void lib_ring_buffer_clear_noref(const struct lib_ring_buffer_config *config,
 370                                  struct lib_ring_buffer_backend *bufb,
 371                                  unsigned long idx)
 372 {
 373         unsigned long id, new_id;
 374
 375         if (config->mode != RING_BUFFER_OVERWRITE)
 376                 return;
 377
 378         /*
 379          * Performing a volatile access to read the sb_pages, because we want to
 380          * read a coherent version of the pointer and the associated noref flag.
 381          */
 382         id = ACCESS_ONCE(bufb->buf_wsb[idx].id);
 383         for (;;) {
 384                 /* This check is called on the fast path for each record. */
 385                 if (likely(!subbuffer_id_is_noref(config, id))) {
 386                         /*
 387                          * Store after load dependency ordering the writes to
 388                          * the subbuffer after load and test of the noref flag
 389                          * matches the memory barrier implied by the cmpxchg()
 390                          * in update_read_sb_index().
 391                          */
 392                         return; /* Already writing to this buffer */
 393                 }
 394                 new_id = id;
 395                 subbuffer_id_clear_noref(config, &new_id);
 396                 new_id = cmpxchg(&bufb->buf_wsb[idx].id, id, new_id);
 397                 if (likely(new_id == id))
 398                         break;
 399                 id = new_id;
 400         }
 401 }
 402
 403 /**
 404  * lib_ring_buffer_set_noref_offset - Set the noref subbuffer flag and offset,
 405  *                                    called by writer.
 406  */
 407 static inline
 408 void lib_ring_buffer_set_noref_offset(const struct lib_ring_buffer_config *config,
 409                                       struct lib_ring_buffer_backend *bufb,
 410                                       unsigned long idx, unsigned long offset)
 411 {
 412         if (config->mode != RING_BUFFER_OVERWRITE)
 413                 return;
 414
 415         /*
 416          * Because ring_buffer_set_noref() is only called by a single thread
 417          * (the one which updated the cc_sb value), there are no concurrent
 418          * updates to take care of: other writers have not updated cc_sb, so
 419          * they cannot set the noref flag, and concurrent readers cannot modify
 420          * the pointer because the noref flag is not set yet.
 421          * The smp_wmb() in ring_buffer_commit() takes care of ordering writes
 422          * to the subbuffer before this set noref operation.
 423          * subbuffer_set_noref() uses a volatile store to deal with concurrent
 424          * readers of the noref flag.
 425          */
 426         CHAN_WARN_ON(bufb->chan,
 427                      subbuffer_id_is_noref(config, bufb->buf_wsb[idx].id));
 428         /*
 429          * Memory barrier that ensures counter stores are ordered before set
 430          * noref and offset.
 431          */
 432         smp_mb();
 433         subbuffer_id_set_noref_offset(config, &bufb->buf_wsb[idx].id, offset);
 434 }
 435
 436 /**
 437  * update_read_sb_index - Read-side subbuffer index update.
 438  */
 439 static inline
 440 int update_read_sb_index(const struct lib_ring_buffer_config *config,
 441                          struct lib_ring_buffer_backend *bufb,
 442                          struct channel_backend *chanb,
 443                          unsigned long consumed_idx,
 444                          unsigned long consumed_count)
 445 {
 446         unsigned long old_id, new_id;
 447
 448         if (config->mode == RING_BUFFER_OVERWRITE) {
 449                 /*
 450                  * Exchange the target writer subbuffer with our own unused
 451                  * subbuffer. No need to use ACCESS_ONCE() here to read the
 452                  * old_wpage, because the value read will be confirmed by the
 453                  * following cmpxchg().
 454                  */
 455                 old_id = bufb->buf_wsb[consumed_idx].id;
 456                 if (unlikely(!subbuffer_id_is_noref(config, old_id)))
 457                         return -EAGAIN;
 458                 /*
 459                  * Make sure the offset count we are expecting matches the one
 460                  * indicated by the writer.
 461                  */
 462                 if (unlikely(!subbuffer_id_compare_offset(config, old_id,
 463                                                           consumed_count)))
 464                         return -EAGAIN;
 465                 CHAN_WARN_ON(bufb->chan,
 466                              !subbuffer_id_is_noref(config, bufb->buf_rsb.id));
 467                 subbuffer_id_set_noref_offset(config, &bufb->buf_rsb.id,
 468                                               consumed_count);
 469                 new_id = cmpxchg(&bufb->buf_wsb[consumed_idx].id, old_id,
 470                                  bufb->buf_rsb.id);
 471                 if (unlikely(old_id != new_id))
 472                         return -EAGAIN;
 473                 bufb->buf_rsb.id = new_id;
 474         } else {
 475                 /* No page exchange, use the writer page directly */
 476                 bufb->buf_rsb.id = bufb->buf_wsb[consumed_idx].id;
 477         }
 478         return 0;
 479 }
 480
 481 static inline __attribute__((always_inline))
 482 void lttng_inline_memcpy(void *dest, const void *src,
 483                 unsigned long len)
 484 {
 485         switch (len) {
 486         case 1:
 487                 *(uint8_t *) dest = *(const uint8_t *) src;
 488                 break;
 489         case 2:
 490                 *(uint16_t *) dest = *(const uint16_t *) src;
 491                 break;
 492         case 4:
 493                 *(uint32_t *) dest = *(const uint32_t *) src;
 494                 break;
 495         case 8:
 496                 *(uint64_t *) dest = *(const uint64_t *) src;
 497                 break;
 498         default:
 499                 inline_memcpy(dest, src, len);
 500         }
 501 }
 502
 503 /*
 504  * Use the architecture-specific memcpy implementation for constant-sized
 505  * inputs, but rely on an inline memcpy for length statically unknown.
 506  * The function call to memcpy is just way too expensive for a fast path.
 507  */
 508 #define lib_ring_buffer_do_copy(config, dest, src, len)         \
 509 do {                                                            \
 510         size_t __len = (len);                                   \
 511         if (__builtin_constant_p(len))                          \
 512                 memcpy(dest, src, __len);                       \
 513         else                                                    \
 514                 lttng_inline_memcpy(dest, src, __len);          \
 515 } while (0)
 516
 517 /*
 518  * We use __copy_from_user_inatomic to copy userspace data since we already
 519  * did the access_ok for the whole range.
 520  *
 521  * Return 0 if OK, nonzero on error.
 522  */
 523 static inline
 524 unsigned long lib_ring_buffer_do_copy_from_user_inatomic(void *dest,
 525                                                 const void __user *src,
 526                                                 unsigned long len)
 527 {
 528         return __copy_from_user_inatomic(dest, src, len);
 529 }
 530
 531 /*
 532  * write len bytes to dest with c
 533  */
 534 static inline
 535 void lib_ring_buffer_do_memset(char *dest, int c,
 536         unsigned long len)
 537 {
 538         unsigned long i;
 539
 540         for (i = 0; i < len; i++)
 541                 dest[i] = c;
 542 }
 543
 544 #endif /* _LIB_RING_BUFFER_BACKEND_INTERNAL_H */