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/compiler.h>
  27 #include <wrapper/ringbuffer/config.h>
  28 #include <wrapper/ringbuffer/backend_types.h>
  29 #include <wrapper/ringbuffer/frontend_types.h>
  30 #include <linux/string.h>
  31 #include <linux/uaccess.h>
  32
  33 /* Ring buffer backend API presented to the frontend */
  34
  35 /* Ring buffer and channel backend create/free */
  36
  37 int lib_ring_buffer_backend_create(struct lib_ring_buffer_backend *bufb,
  38                                    struct channel_backend *chan, int cpu);
  39 void channel_backend_unregister_notifiers(struct channel_backend *chanb);
  40 void lib_ring_buffer_backend_free(struct lib_ring_buffer_backend *bufb);
  41 int channel_backend_init(struct channel_backend *chanb,
  42                          const char *name,
  43                          const struct lib_ring_buffer_config *config,
  44                          void *priv, size_t subbuf_size,
  45                          size_t num_subbuf);
  46 void channel_backend_free(struct channel_backend *chanb);
  47
  48 void lib_ring_buffer_backend_reset(struct lib_ring_buffer_backend *bufb);
  49 void channel_backend_reset(struct channel_backend *chanb);
  50
  51 int lib_ring_buffer_backend_init(void);
  52 void lib_ring_buffer_backend_exit(void);
  53
  54 extern void _lib_ring_buffer_write(struct lib_ring_buffer_backend *bufb,
  55                                    size_t offset, const void *src, size_t len,
  56                                    size_t pagecpy);
  57 extern void _lib_ring_buffer_memset(struct lib_ring_buffer_backend *bufb,
  58                                     size_t offset, int c, size_t len,
  59                                     size_t pagecpy);
  60 extern void _lib_ring_buffer_strcpy(struct lib_ring_buffer_backend *bufb,
  61                                    size_t offset, const char *src, size_t len,
  62                                    size_t pagecpy, int pad);
  63 extern void _lib_ring_buffer_copy_from_user_inatomic(struct lib_ring_buffer_backend *bufb,
  64                                             size_t offset, const void *src,
  65                                             size_t len, size_t pagecpy);
  66 extern void _lib_ring_buffer_strcpy_from_user_inatomic(struct lib_ring_buffer_backend *bufb,
  67                 size_t offset, const char __user *src, size_t len,
  68                 size_t pagecpy, int pad);
  69
  70 /*
  71  * Subbuffer ID bits for overwrite mode. Need to fit within a single word to be
  72  * exchanged atomically.
  73  *
  74  * Top half word, except lowest bit, belongs to "offset", which is used to keep
  75  * to count the produced buffers.  For overwrite mode, this provides the
  76  * consumer with the capacity to read subbuffers in order, handling the
  77  * situation where producers would write up to 2^15 buffers (or 2^31 for 64-bit
  78  * systems) concurrently with a single execution of get_subbuf (between offset
  79  * sampling and subbuffer ID exchange).
  80  */
  81
  82 #define HALF_ULONG_BITS         (BITS_PER_LONG >> 1)
  83
  84 #define SB_ID_OFFSET_SHIFT      (HALF_ULONG_BITS + 1)
  85 #define SB_ID_OFFSET_COUNT      (1UL << SB_ID_OFFSET_SHIFT)
  86 #define SB_ID_OFFSET_MASK       (~(SB_ID_OFFSET_COUNT - 1))
  87 /*
  88  * Lowest bit of top word half belongs to noref. Used only for overwrite mode.
  89  */
  90 #define SB_ID_NOREF_SHIFT       (SB_ID_OFFSET_SHIFT - 1)
  91 #define SB_ID_NOREF_COUNT       (1UL << SB_ID_NOREF_SHIFT)
  92 #define SB_ID_NOREF_MASK        SB_ID_NOREF_COUNT
  93 /*
  94  * In overwrite mode: lowest half of word is used for index.
  95  * Limit of 2^16 subbuffers per buffer on 32-bit, 2^32 on 64-bit.
  96  * In producer-consumer mode: whole word used for index.
  97  */
  98 #define SB_ID_INDEX_SHIFT       0
  99 #define SB_ID_INDEX_COUNT       (1UL << SB_ID_INDEX_SHIFT)
 100 #define SB_ID_INDEX_MASK        (SB_ID_NOREF_COUNT - 1)
 101
 102 /*
 103  * Construct the subbuffer id from offset, index and noref. Use only the index
 104  * for producer-consumer mode (offset and noref are only used in overwrite
 105  * mode).
 106  */
 107 static inline
 108 unsigned long subbuffer_id(const struct lib_ring_buffer_config *config,
 109                            unsigned long offset, unsigned long noref,
 110                            unsigned long index)
 111 {
 112         if (config->mode == RING_BUFFER_OVERWRITE)
 113                 return (offset << SB_ID_OFFSET_SHIFT)
 114                        | (noref << SB_ID_NOREF_SHIFT)
 115                        | index;
 116         else
 117                 return index;
 118 }
 119
 120 /*
 121  * Compare offset with the offset contained within id. Return 1 if the offset
 122  * bits are identical, else 0.
 123  */
 124 static inline
 125 int subbuffer_id_compare_offset(const struct lib_ring_buffer_config *config,
 126                                 unsigned long id, unsigned long offset)
 127 {
 128         return (id & SB_ID_OFFSET_MASK) == (offset << SB_ID_OFFSET_SHIFT);
 129 }
 130
 131 static inline
 132 unsigned long subbuffer_id_get_index(const struct lib_ring_buffer_config *config,
 133                                      unsigned long id)
 134 {
 135         if (config->mode == RING_BUFFER_OVERWRITE)
 136                 return id & SB_ID_INDEX_MASK;
 137         else
 138                 return id;
 139 }
 140
 141 static inline
 142 unsigned long subbuffer_id_is_noref(const struct lib_ring_buffer_config *config,
 143                                     unsigned long id)
 144 {
 145         if (config->mode == RING_BUFFER_OVERWRITE)
 146                 return !!(id & SB_ID_NOREF_MASK);
 147         else
 148                 return 1;
 149 }
 150
 151 /*
 152  * Only used by reader on subbuffer ID it has exclusive access to. No volatile
 153  * needed.
 154  */
 155 static inline
 156 void subbuffer_id_set_noref(const struct lib_ring_buffer_config *config,
 157                             unsigned long *id)
 158 {
 159         if (config->mode == RING_BUFFER_OVERWRITE)
 160                 *id |= SB_ID_NOREF_MASK;
 161 }
 162
 163 static inline
 164 void subbuffer_id_set_noref_offset(const struct lib_ring_buffer_config *config,
 165                                    unsigned long *id, unsigned long offset)
 166 {
 167         unsigned long tmp;
 168
 169         if (config->mode == RING_BUFFER_OVERWRITE) {
 170                 tmp = *id;
 171                 tmp &= ~SB_ID_OFFSET_MASK;
 172                 tmp |= offset << SB_ID_OFFSET_SHIFT;
 173                 tmp |= SB_ID_NOREF_MASK;
 174                 /* Volatile store, read concurrently by readers. */
 175                 WRITE_ONCE(*id, tmp);
 176         }
 177 }
 178
 179 /* No volatile access, since already used locally */
 180 static inline
 181 void subbuffer_id_clear_noref(const struct lib_ring_buffer_config *config,
 182                               unsigned long *id)
 183 {
 184         if (config->mode == RING_BUFFER_OVERWRITE)
 185                 *id &= ~SB_ID_NOREF_MASK;
 186 }
 187
 188 /*
 189  * For overwrite mode, cap the number of subbuffers per buffer to:
 190  * 2^16 on 32-bit architectures
 191  * 2^32 on 64-bit architectures
 192  * This is required to fit in the index part of the ID. Return 0 on success,
 193  * -EPERM on failure.
 194  */
 195 static inline
 196 int subbuffer_id_check_index(const struct lib_ring_buffer_config *config,
 197                              unsigned long num_subbuf)
 198 {
 199         if (config->mode == RING_BUFFER_OVERWRITE)
 200                 return (num_subbuf > (1UL << HALF_ULONG_BITS)) ? -EPERM : 0;
 201         else
 202                 return 0;
 203 }
 204
 205 static inline
 206 void lib_ring_buffer_backend_get_pages(const struct lib_ring_buffer_config *config,
 207                         struct lib_ring_buffer_ctx *ctx,
 208                         struct lib_ring_buffer_backend_pages **backend_pages)
 209 {
 210         struct lib_ring_buffer_backend *bufb = &ctx->buf->backend;
 211         struct channel_backend *chanb = &ctx->chan->backend;
 212         size_t sbidx, offset = ctx->buf_offset;
 213         unsigned long sb_bindex, id;
 214         struct lib_ring_buffer_backend_pages *rpages;
 215
 216         offset &= chanb->buf_size - 1;
 217         sbidx = offset >> chanb->subbuf_size_order;
 218         id = bufb->buf_wsb[sbidx].id;
 219         sb_bindex = subbuffer_id_get_index(config, id);
 220         rpages = bufb->array[sb_bindex];
 221         CHAN_WARN_ON(ctx->chan,
 222                      config->mode == RING_BUFFER_OVERWRITE
 223                      && subbuffer_id_is_noref(config, id));
 224         *backend_pages = rpages;
 225 }
 226
 227 /* Get backend pages from cache. */
 228 static inline
 229 struct lib_ring_buffer_backend_pages *
 230         lib_ring_buffer_get_backend_pages_from_ctx(const struct lib_ring_buffer_config *config,
 231                 struct lib_ring_buffer_ctx *ctx)
 232 {
 233         return ctx->backend_pages;
 234 }
 235
 236 /*
 237  * The ring buffer can count events recorded and overwritten per buffer,
 238  * but it is disabled by default due to its performance overhead.
 239  */
 240 #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
 241 static inline
 242 void subbuffer_count_record(const struct lib_ring_buffer_config *config,
 243                             struct lib_ring_buffer_backend *bufb,
 244                             unsigned long idx)
 245 {
 246         unsigned long sb_bindex;
 247
 248         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 249         v_inc(config, &bufb->array[sb_bindex]->records_commit);
 250 }
 251 #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
 252 static inline
 253 void subbuffer_count_record(const struct lib_ring_buffer_config *config,
 254                             struct lib_ring_buffer_backend *bufb,
 255                             unsigned long idx)
 256 {
 257 }
 258 #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
 259
 260 /*
 261  * Reader has exclusive subbuffer access for record consumption. No need to
 262  * perform the decrement atomically.
 263  */
 264 static inline
 265 void subbuffer_consume_record(const struct lib_ring_buffer_config *config,
 266                               struct lib_ring_buffer_backend *bufb)
 267 {
 268         unsigned long sb_bindex;
 269
 270         sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
 271         CHAN_WARN_ON(bufb->chan,
 272                      !v_read(config, &bufb->array[sb_bindex]->records_unread));
 273         /* Non-atomic decrement protected by exclusive subbuffer access */
 274         _v_dec(config, &bufb->array[sb_bindex]->records_unread);
 275         v_inc(config, &bufb->records_read);
 276 }
 277
 278 static inline
 279 unsigned long subbuffer_get_records_count(
 280                                 const struct lib_ring_buffer_config *config,
 281                                 struct lib_ring_buffer_backend *bufb,
 282                                 unsigned long idx)
 283 {
 284         unsigned long sb_bindex;
 285
 286         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 287         return v_read(config, &bufb->array[sb_bindex]->records_commit);
 288 }
 289
 290 /*
 291  * Must be executed at subbuffer delivery when the writer has _exclusive_
 292  * subbuffer access. See lib_ring_buffer_check_deliver() for details.
 293  * lib_ring_buffer_get_records_count() must be called to get the records
 294  * count before this function, because it resets the records_commit
 295  * count.
 296  */
 297 static inline
 298 unsigned long subbuffer_count_records_overrun(
 299                                 const struct lib_ring_buffer_config *config,
 300                                 struct lib_ring_buffer_backend *bufb,
 301                                 unsigned long idx)
 302 {
 303         struct lib_ring_buffer_backend_pages *pages;
 304         unsigned long overruns, sb_bindex;
 305
 306         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 307         pages = bufb->array[sb_bindex];
 308         overruns = v_read(config, &pages->records_unread);
 309         v_set(config, &pages->records_unread,
 310               v_read(config, &pages->records_commit));
 311         v_set(config, &pages->records_commit, 0);
 312
 313         return overruns;
 314 }
 315
 316 static inline
 317 void subbuffer_set_data_size(const struct lib_ring_buffer_config *config,
 318                              struct lib_ring_buffer_backend *bufb,
 319                              unsigned long idx,
 320                              unsigned long data_size)
 321 {
 322         struct lib_ring_buffer_backend_pages *pages;
 323         unsigned long sb_bindex;
 324
 325         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 326         pages = bufb->array[sb_bindex];
 327         pages->data_size = data_size;
 328 }
 329
 330 static inline
 331 unsigned long subbuffer_get_read_data_size(
 332                                 const struct lib_ring_buffer_config *config,
 333                                 struct lib_ring_buffer_backend *bufb)
 334 {
 335         struct lib_ring_buffer_backend_pages *pages;
 336         unsigned long sb_bindex;
 337
 338         sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
 339         pages = bufb->array[sb_bindex];
 340         return pages->data_size;
 341 }
 342
 343 static inline
 344 unsigned long subbuffer_get_data_size(
 345                                 const struct lib_ring_buffer_config *config,
 346                                 struct lib_ring_buffer_backend *bufb,
 347                                 unsigned long idx)
 348 {
 349         struct lib_ring_buffer_backend_pages *pages;
 350         unsigned long sb_bindex;
 351
 352         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 353         pages = bufb->array[sb_bindex];
 354         return pages->data_size;
 355 }
 356
 357 static inline
 358 void subbuffer_inc_packet_count(const struct lib_ring_buffer_config *config,
 359                                 struct lib_ring_buffer_backend *bufb,
 360                                 unsigned long idx)
 361 {
 362         bufb->buf_cnt[idx].seq_cnt++;
 363 }
 364
 365 /**
 366  * lib_ring_buffer_clear_noref - Clear the noref subbuffer flag, called by
 367  *                               writer.
 368  */
 369 static inline
 370 void lib_ring_buffer_clear_noref(const struct lib_ring_buffer_config *config,
 371                                  struct lib_ring_buffer_backend *bufb,
 372                                  unsigned long idx)
 373 {
 374         unsigned long id, new_id;
 375
 376         if (config->mode != RING_BUFFER_OVERWRITE)
 377                 return;
 378
 379         /*
 380          * Performing a volatile access to read the sb_pages, because we want to
 381          * read a coherent version of the pointer and the associated noref flag.
 382          */
 383         id = READ_ONCE(bufb->buf_wsb[idx].id);
 384         for (;;) {
 385                 /* This check is called on the fast path for each record. */
 386                 if (likely(!subbuffer_id_is_noref(config, id))) {
 387                         /*
 388                          * Store after load dependency ordering the writes to
 389                          * the subbuffer after load and test of the noref flag
 390                          * matches the memory barrier implied by the cmpxchg()
 391                          * in update_read_sb_index().
 392                          */
 393                         return; /* Already writing to this buffer */
 394                 }
 395                 new_id = id;
 396                 subbuffer_id_clear_noref(config, &new_id);
 397                 new_id = cmpxchg(&bufb->buf_wsb[idx].id, id, new_id);
 398                 if (likely(new_id == id))
 399                         break;
 400                 id = new_id;
 401         }
 402 }
 403
 404 /**
 405  * lib_ring_buffer_set_noref_offset - Set the noref subbuffer flag and offset,
 406  *                                    called by writer.
 407  */
 408 static inline
 409 void lib_ring_buffer_set_noref_offset(const struct lib_ring_buffer_config *config,
 410                                       struct lib_ring_buffer_backend *bufb,
 411                                       unsigned long idx, unsigned long offset)
 412 {
 413         if (config->mode != RING_BUFFER_OVERWRITE)
 414                 return;
 415
 416         /*
 417          * Because ring_buffer_set_noref() is only called by a single thread
 418          * (the one which updated the cc_sb value), there are no concurrent
 419          * updates to take care of: other writers have not updated cc_sb, so
 420          * they cannot set the noref flag, and concurrent readers cannot modify
 421          * the pointer because the noref flag is not set yet.
 422          * The smp_wmb() in ring_buffer_commit() takes care of ordering writes
 423          * to the subbuffer before this set noref operation.
 424          * subbuffer_set_noref() uses a volatile store to deal with concurrent
 425          * readers of the noref flag.
 426          */
 427         CHAN_WARN_ON(bufb->chan,
 428                      subbuffer_id_is_noref(config, bufb->buf_wsb[idx].id));
 429         /*
 430          * Memory barrier that ensures counter stores are ordered before set
 431          * noref and offset.
 432          */
 433         smp_mb();
 434         subbuffer_id_set_noref_offset(config, &bufb->buf_wsb[idx].id, offset);
 435 }
 436
 437 /**
 438  * update_read_sb_index - Read-side subbuffer index update.
 439  */
 440 static inline
 441 int update_read_sb_index(const struct lib_ring_buffer_config *config,
 442                          struct lib_ring_buffer_backend *bufb,
 443                          struct channel_backend *chanb,
 444                          unsigned long consumed_idx,
 445                          unsigned long consumed_count)
 446 {
 447         unsigned long old_id, new_id;
 448
 449         if (config->mode == RING_BUFFER_OVERWRITE) {
 450                 /*
 451                  * Exchange the target writer subbuffer with our own unused
 452                  * subbuffer. No need to use READ_ONCE() here to read the
 453                  * old_wpage, because the value read will be confirmed by the
 454                  * following cmpxchg().
 455                  */
 456                 old_id = bufb->buf_wsb[consumed_idx].id;
 457                 if (unlikely(!subbuffer_id_is_noref(config, old_id)))
 458                         return -EAGAIN;
 459                 /*
 460                  * Make sure the offset count we are expecting matches the one
 461                  * indicated by the writer.
 462                  */
 463                 if (unlikely(!subbuffer_id_compare_offset(config, old_id,
 464                                                           consumed_count)))
 465                         return -EAGAIN;
 466                 CHAN_WARN_ON(bufb->chan,
 467                              !subbuffer_id_is_noref(config, bufb->buf_rsb.id));
 468                 subbuffer_id_set_noref_offset(config, &bufb->buf_rsb.id,
 469                                               consumed_count);
 470                 new_id = cmpxchg(&bufb->buf_wsb[consumed_idx].id, old_id,
 471                                  bufb->buf_rsb.id);
 472                 if (unlikely(old_id != new_id))
 473                         return -EAGAIN;
 474                 bufb->buf_rsb.id = new_id;
 475         } else {
 476                 /* No page exchange, use the writer page directly */
 477                 bufb->buf_rsb.id = bufb->buf_wsb[consumed_idx].id;
 478         }
 479         return 0;
 480 }
 481
 482 static inline __attribute__((always_inline))
 483 void lttng_inline_memcpy(void *dest, const void *src,
 484                 unsigned long len)
 485 {
 486         switch (len) {
 487         case 1:
 488                 *(uint8_t *) dest = *(const uint8_t *) src;
 489                 break;
 490         case 2:
 491                 *(uint16_t *) dest = *(const uint16_t *) src;
 492                 break;
 493         case 4:
 494                 *(uint32_t *) dest = *(const uint32_t *) src;
 495                 break;
 496         case 8:
 497                 *(uint64_t *) dest = *(const uint64_t *) src;
 498                 break;
 499         default:
 500                 inline_memcpy(dest, src, len);
 501         }
 502 }
 503
 504 /*
 505  * Use the architecture-specific memcpy implementation for constant-sized
 506  * inputs, but rely on an inline memcpy for length statically unknown.
 507  * The function call to memcpy is just way too expensive for a fast path.
 508  */
 509 #define lib_ring_buffer_do_copy(config, dest, src, len)         \
 510 do {                                                            \
 511         size_t __len = (len);                                   \
 512         if (__builtin_constant_p(len))                          \
 513                 memcpy(dest, src, __len);                       \
 514         else                                                    \
 515                 lttng_inline_memcpy(dest, src, __len);          \
 516 } while (0)
 517
 518 /*
 519  * We use __copy_from_user_inatomic to copy userspace data since we already
 520  * did the access_ok for the whole range.
 521  *
 522  * Return 0 if OK, nonzero on error.
 523  */
 524 static inline
 525 unsigned long lib_ring_buffer_do_copy_from_user_inatomic(void *dest,
 526                                                 const void __user *src,
 527                                                 unsigned long len)
 528 {
 529         return __copy_from_user_inatomic(dest, src, len);
 530 }
 531
 532 /*
 533  * write len bytes to dest with c
 534  */
 535 static inline
 536 void lib_ring_buffer_do_memset(char *dest, int c,
 537         unsigned long len)
 538 {
 539         unsigned long i;
 540
 541         for (i = 0; i < len; i++)
 542                 dest[i] = c;
 543 }
 544
 545 #endif /* _LIB_RING_BUFFER_BACKEND_INTERNAL_H */