libringbuffer/backend_internal.h

   1 #ifndef _LTTNG_RING_BUFFER_BACKEND_INTERNAL_H
   2 #define _LTTNG_RING_BUFFER_BACKEND_INTERNAL_H
   3
   4 /*
   5  * libringbuffer/backend_internal.h
   6  *
   7  * Ring buffer backend (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
  26 #include <unistd.h>
  27 #include <urcu/compiler.h>
  28
  29 #include <lttng/ringbuffer-config.h>
  30 #include "backend_types.h"
  31 #include "frontend_types.h"
  32 #include "shm.h"
  33
  34 /* Ring buffer backend API presented to the frontend */
  35
  36 /* Ring buffer and channel backend create/free */
  37
  38 int lib_ring_buffer_backend_create(struct lttng_ust_lib_ring_buffer_backend *bufb,
  39                                    struct channel_backend *chan, int cpu,
  40                                    struct lttng_ust_shm_handle *handle,
  41                                    struct shm_object *shmobj);
  42 void channel_backend_unregister_notifiers(struct channel_backend *chanb);
  43 void lib_ring_buffer_backend_free(struct lttng_ust_lib_ring_buffer_backend *bufb);
  44 int channel_backend_init(struct channel_backend *chanb,
  45                          const char *name,
  46                          const struct lttng_ust_lib_ring_buffer_config *config,
  47                          size_t subbuf_size,
  48                          size_t num_subbuf, struct lttng_ust_shm_handle *handle,
  49                          const int *stream_fds);
  50 void channel_backend_free(struct channel_backend *chanb,
  51                           struct lttng_ust_shm_handle *handle);
  52
  53 void lib_ring_buffer_backend_reset(struct lttng_ust_lib_ring_buffer_backend *bufb,
  54                                    struct lttng_ust_shm_handle *handle);
  55 void channel_backend_reset(struct channel_backend *chanb);
  56
  57 int lib_ring_buffer_backend_init(void);
  58 void lib_ring_buffer_backend_exit(void);
  59
  60 extern void _lib_ring_buffer_write(struct lttng_ust_lib_ring_buffer_backend *bufb,
  61                                    size_t offset, const void *src, size_t len,
  62                                    ssize_t pagecpy);
  63
  64 /*
  65  * Subbuffer ID bits for overwrite mode. Need to fit within a single word to be
  66  * exchanged atomically.
  67  *
  68  * Top half word, except lowest bit, belongs to "offset", which is used to keep
  69  * to count the produced buffers.  For overwrite mode, this provides the
  70  * consumer with the capacity to read subbuffers in order, handling the
  71  * situation where producers would write up to 2^15 buffers (or 2^31 for 64-bit
  72  * systems) concurrently with a single execution of get_subbuf (between offset
  73  * sampling and subbuffer ID exchange).
  74  */
  75
  76 #define HALF_ULONG_BITS         (CAA_BITS_PER_LONG >> 1)
  77
  78 #define SB_ID_OFFSET_SHIFT      (HALF_ULONG_BITS + 1)
  79 #define SB_ID_OFFSET_COUNT      (1UL << SB_ID_OFFSET_SHIFT)
  80 #define SB_ID_OFFSET_MASK       (~(SB_ID_OFFSET_COUNT - 1))
  81 /*
  82  * Lowest bit of top word half belongs to noref. Used only for overwrite mode.
  83  */
  84 #define SB_ID_NOREF_SHIFT       (SB_ID_OFFSET_SHIFT - 1)
  85 #define SB_ID_NOREF_COUNT       (1UL << SB_ID_NOREF_SHIFT)
  86 #define SB_ID_NOREF_MASK        SB_ID_NOREF_COUNT
  87 /*
  88  * In overwrite mode: lowest half of word is used for index.
  89  * Limit of 2^16 subbuffers per buffer on 32-bit, 2^32 on 64-bit.
  90  * In producer-consumer mode: whole word used for index.
  91  */
  92 #define SB_ID_INDEX_SHIFT       0
  93 #define SB_ID_INDEX_COUNT       (1UL << SB_ID_INDEX_SHIFT)
  94 #define SB_ID_INDEX_MASK        (SB_ID_NOREF_COUNT - 1)
  95
  96 /*
  97  * Construct the subbuffer id from offset, index and noref. Use only the index
  98  * for producer-consumer mode (offset and noref are only used in overwrite
  99  * mode).
 100  */
 101 static inline
 102 unsigned long subbuffer_id(const struct lttng_ust_lib_ring_buffer_config *config,
 103                            unsigned long offset, unsigned long noref,
 104                            unsigned long index)
 105 {
 106         if (config->mode == RING_BUFFER_OVERWRITE)
 107                 return (offset << SB_ID_OFFSET_SHIFT)
 108                        | (noref << SB_ID_NOREF_SHIFT)
 109                        | index;
 110         else
 111                 return index;
 112 }
 113
 114 /*
 115  * Compare offset with the offset contained within id. Return 1 if the offset
 116  * bits are identical, else 0.
 117  */
 118 static inline
 119 int subbuffer_id_compare_offset(const struct lttng_ust_lib_ring_buffer_config *config,
 120                                 unsigned long id, unsigned long offset)
 121 {
 122         return (id & SB_ID_OFFSET_MASK) == (offset << SB_ID_OFFSET_SHIFT);
 123 }
 124
 125 static inline
 126 unsigned long subbuffer_id_get_index(const struct lttng_ust_lib_ring_buffer_config *config,
 127                                      unsigned long id)
 128 {
 129         if (config->mode == RING_BUFFER_OVERWRITE)
 130                 return id & SB_ID_INDEX_MASK;
 131         else
 132                 return id;
 133 }
 134
 135 static inline
 136 unsigned long subbuffer_id_is_noref(const struct lttng_ust_lib_ring_buffer_config *config,
 137                                     unsigned long id)
 138 {
 139         if (config->mode == RING_BUFFER_OVERWRITE)
 140                 return !!(id & SB_ID_NOREF_MASK);
 141         else
 142                 return 1;
 143 }
 144
 145 /*
 146  * Only used by reader on subbuffer ID it has exclusive access to. No volatile
 147  * needed.
 148  */
 149 static inline
 150 void subbuffer_id_set_noref(const struct lttng_ust_lib_ring_buffer_config *config,
 151                             unsigned long *id)
 152 {
 153         if (config->mode == RING_BUFFER_OVERWRITE)
 154                 *id |= SB_ID_NOREF_MASK;
 155 }
 156
 157 static inline
 158 void subbuffer_id_set_noref_offset(const struct lttng_ust_lib_ring_buffer_config *config,
 159                                    unsigned long *id, unsigned long offset)
 160 {
 161         unsigned long tmp;
 162
 163         if (config->mode == RING_BUFFER_OVERWRITE) {
 164                 tmp = *id;
 165                 tmp &= ~SB_ID_OFFSET_MASK;
 166                 tmp |= offset << SB_ID_OFFSET_SHIFT;
 167                 tmp |= SB_ID_NOREF_MASK;
 168                 /* Volatile store, read concurrently by readers. */
 169                 CMM_ACCESS_ONCE(*id) = tmp;
 170         }
 171 }
 172
 173 /* No volatile access, since already used locally */
 174 static inline
 175 void subbuffer_id_clear_noref(const struct lttng_ust_lib_ring_buffer_config *config,
 176                               unsigned long *id)
 177 {
 178         if (config->mode == RING_BUFFER_OVERWRITE)
 179                 *id &= ~SB_ID_NOREF_MASK;
 180 }
 181
 182 /*
 183  * For overwrite mode, cap the number of subbuffers per buffer to:
 184  * 2^16 on 32-bit architectures
 185  * 2^32 on 64-bit architectures
 186  * This is required to fit in the index part of the ID. Return 0 on success,
 187  * -EPERM on failure.
 188  */
 189 static inline
 190 int subbuffer_id_check_index(const struct lttng_ust_lib_ring_buffer_config *config,
 191                              unsigned long num_subbuf)
 192 {
 193         if (config->mode == RING_BUFFER_OVERWRITE)
 194                 return (num_subbuf > (1UL << HALF_ULONG_BITS)) ? -EPERM : 0;
 195         else
 196                 return 0;
 197 }
 198
 199 /*
 200  * The ring buffer can count events recorded and overwritten per buffer,
 201  * but it is disabled by default due to its performance overhead.
 202  */
 203 #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
 204 static inline
 205 void subbuffer_count_record(const struct lttng_ust_lib_ring_buffer_config *config,
 206                             struct lttng_ust_lib_ring_buffer_backend *bufb,
 207                             unsigned long idx, struct lttng_ust_shm_handle *handle)
 208 {
 209         unsigned long sb_bindex;
 210
 211         sb_bindex = subbuffer_id_get_index(config, shmp_index(handle, bufb->buf_wsb, idx)->id);
 212         v_inc(config, &shmp(handle, shmp_index(handle, bufb->array, sb_bindex)->shmp)->records_commit);
 213 }
 214 #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
 215 static inline
 216 void subbuffer_count_record(const struct lttng_ust_lib_ring_buffer_config *config,
 217                             struct lttng_ust_lib_ring_buffer_backend *bufb,
 218                             unsigned long idx, struct lttng_ust_shm_handle *handle)
 219 {
 220 }
 221 #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
 222
 223 /*
 224  * Reader has exclusive subbuffer access for record consumption. No need to
 225  * perform the decrement atomically.
 226  */
 227 static inline
 228 void subbuffer_consume_record(const struct lttng_ust_lib_ring_buffer_config *config,
 229                               struct lttng_ust_lib_ring_buffer_backend *bufb,
 230                               struct lttng_ust_shm_handle *handle)
 231 {
 232         unsigned long sb_bindex;
 233
 234         sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
 235         CHAN_WARN_ON(shmp(handle, bufb->chan),
 236                      !v_read(config, &shmp(handle, shmp_index(handle, bufb->array, sb_bindex)->shmp)->records_unread));
 237         /* Non-atomic decrement protected by exclusive subbuffer access */
 238         _v_dec(config, &shmp(handle, shmp_index(handle, bufb->array, sb_bindex)->shmp)->records_unread);
 239         v_inc(config, &bufb->records_read);
 240 }
 241
 242 static inline
 243 unsigned long subbuffer_get_records_count(
 244                                 const struct lttng_ust_lib_ring_buffer_config *config,
 245                                 struct lttng_ust_lib_ring_buffer_backend *bufb,
 246                                 unsigned long idx,
 247                                 struct lttng_ust_shm_handle *handle)
 248 {
 249         unsigned long sb_bindex;
 250
 251         sb_bindex = subbuffer_id_get_index(config, shmp_index(handle, bufb->buf_wsb, idx)->id);
 252         return v_read(config, &shmp(handle, shmp_index(handle, bufb->array, sb_bindex)->shmp)->records_commit);
 253 }
 254
 255 /*
 256  * Must be executed at subbuffer delivery when the writer has _exclusive_
 257  * subbuffer access. See lib_ring_buffer_check_deliver() for details.
 258  * lib_ring_buffer_get_records_count() must be called to get the records
 259  * count before this function, because it resets the records_commit
 260  * count.
 261  */
 262 static inline
 263 unsigned long subbuffer_count_records_overrun(
 264                                 const struct lttng_ust_lib_ring_buffer_config *config,
 265                                 struct lttng_ust_lib_ring_buffer_backend *bufb,
 266                                 unsigned long idx,
 267                                 struct lttng_ust_shm_handle *handle)
 268 {
 269         struct lttng_ust_lib_ring_buffer_backend_pages_shmp *pages;
 270         unsigned long overruns, sb_bindex;
 271
 272         sb_bindex = subbuffer_id_get_index(config, shmp_index(handle, bufb->buf_wsb, idx)->id);
 273         pages = shmp_index(handle, bufb->array, sb_bindex);
 274         overruns = v_read(config, &shmp(handle, pages->shmp)->records_unread);
 275         v_set(config, &shmp(handle, pages->shmp)->records_unread,
 276               v_read(config, &shmp(handle, pages->shmp)->records_commit));
 277         v_set(config, &shmp(handle, pages->shmp)->records_commit, 0);
 278
 279         return overruns;
 280 }
 281
 282 static inline
 283 void subbuffer_set_data_size(const struct lttng_ust_lib_ring_buffer_config *config,
 284                              struct lttng_ust_lib_ring_buffer_backend *bufb,
 285                              unsigned long idx,
 286                              unsigned long data_size,
 287                              struct lttng_ust_shm_handle *handle)
 288 {
 289         struct lttng_ust_lib_ring_buffer_backend_pages_shmp *pages;
 290         unsigned long sb_bindex;
 291
 292         sb_bindex = subbuffer_id_get_index(config, shmp_index(handle, bufb->buf_wsb, idx)->id);
 293         pages = shmp_index(handle, bufb->array, sb_bindex);
 294         shmp(handle, pages->shmp)->data_size = data_size;
 295 }
 296
 297 static inline
 298 unsigned long subbuffer_get_read_data_size(
 299                                 const struct lttng_ust_lib_ring_buffer_config *config,
 300                                 struct lttng_ust_lib_ring_buffer_backend *bufb,
 301                                 struct lttng_ust_shm_handle *handle)
 302 {
 303         struct lttng_ust_lib_ring_buffer_backend_pages_shmp *pages;
 304         unsigned long sb_bindex;
 305
 306         sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
 307         pages = shmp_index(handle, bufb->array, sb_bindex);
 308         return shmp(handle, pages->shmp)->data_size;
 309 }
 310
 311 static inline
 312 unsigned long subbuffer_get_data_size(
 313                                 const struct lttng_ust_lib_ring_buffer_config *config,
 314                                 struct lttng_ust_lib_ring_buffer_backend *bufb,
 315                                 unsigned long idx,
 316                                 struct lttng_ust_shm_handle *handle)
 317 {
 318         struct lttng_ust_lib_ring_buffer_backend_pages_shmp *pages;
 319         unsigned long sb_bindex;
 320
 321         sb_bindex = subbuffer_id_get_index(config, shmp_index(handle, bufb->buf_wsb, idx)->id);
 322         pages = shmp_index(handle, bufb->array, sb_bindex);
 323         return shmp(handle, pages->shmp)->data_size;
 324 }
 325
 326 static inline
 327 void subbuffer_inc_packet_count(const struct lttng_ust_lib_ring_buffer_config *config,
 328                 struct lttng_ust_lib_ring_buffer_backend *bufb,
 329                 unsigned long idx, struct lttng_ust_shm_handle *handle)
 330 {
 331         shmp_index(handle, bufb->buf_cnt, idx)->seq_cnt++;
 332 }
 333
 334 /**
 335  * lib_ring_buffer_clear_noref - Clear the noref subbuffer flag, called by
 336  *                               writer.
 337  */
 338 static inline
 339 void lib_ring_buffer_clear_noref(const struct lttng_ust_lib_ring_buffer_config *config,
 340                                  struct lttng_ust_lib_ring_buffer_backend *bufb,
 341                                  unsigned long idx,
 342                                  struct lttng_ust_shm_handle *handle)
 343 {
 344         unsigned long id, new_id;
 345
 346         if (config->mode != RING_BUFFER_OVERWRITE)
 347                 return;
 348
 349         /*
 350          * Performing a volatile access to read the sb_pages, because we want to
 351          * read a coherent version of the pointer and the associated noref flag.
 352          */
 353         id = CMM_ACCESS_ONCE(shmp_index(handle, bufb->buf_wsb, idx)->id);
 354         for (;;) {
 355                 /* This check is called on the fast path for each record. */
 356                 if (caa_likely(!subbuffer_id_is_noref(config, id))) {
 357                         /*
 358                          * Store after load dependency ordering the writes to
 359                          * the subbuffer after load and test of the noref flag
 360                          * matches the memory barrier implied by the cmpxchg()
 361                          * in update_read_sb_index().
 362                          */
 363                         return; /* Already writing to this buffer */
 364                 }
 365                 new_id = id;
 366                 subbuffer_id_clear_noref(config, &new_id);
 367                 new_id = uatomic_cmpxchg(&shmp_index(handle, bufb->buf_wsb, idx)->id, id, new_id);
 368                 if (caa_likely(new_id == id))
 369                         break;
 370                 id = new_id;
 371         }
 372 }
 373
 374 /**
 375  * lib_ring_buffer_set_noref_offset - Set the noref subbuffer flag and offset,
 376  *                                    called by writer.
 377  */
 378 static inline
 379 void lib_ring_buffer_set_noref_offset(const struct lttng_ust_lib_ring_buffer_config *config,
 380                                       struct lttng_ust_lib_ring_buffer_backend *bufb,
 381                                       unsigned long idx, unsigned long offset,
 382                                       struct lttng_ust_shm_handle *handle)
 383 {
 384         if (config->mode != RING_BUFFER_OVERWRITE)
 385                 return;
 386
 387         /*
 388          * Because ring_buffer_set_noref() is only called by a single thread
 389          * (the one which updated the cc_sb value), there are no concurrent
 390          * updates to take care of: other writers have not updated cc_sb, so
 391          * they cannot set the noref flag, and concurrent readers cannot modify
 392          * the pointer because the noref flag is not set yet.
 393          * The smp_wmb() in ring_buffer_commit() takes care of ordering writes
 394          * to the subbuffer before this set noref operation.
 395          * subbuffer_set_noref() uses a volatile store to deal with concurrent
 396          * readers of the noref flag.
 397          */
 398         CHAN_WARN_ON(shmp(handle, bufb->chan),
 399                      subbuffer_id_is_noref(config, shmp_index(handle, bufb->buf_wsb, idx)->id));
 400         /*
 401          * Memory barrier that ensures counter stores are ordered before set
 402          * noref and offset.
 403          */
 404         cmm_smp_mb();
 405         subbuffer_id_set_noref_offset(config, &shmp_index(handle, bufb->buf_wsb, idx)->id, offset);
 406 }
 407
 408 /**
 409  * update_read_sb_index - Read-side subbuffer index update.
 410  */
 411 static inline
 412 int update_read_sb_index(const struct lttng_ust_lib_ring_buffer_config *config,
 413                          struct lttng_ust_lib_ring_buffer_backend *bufb,
 414                          struct channel_backend *chanb,
 415                          unsigned long consumed_idx,
 416                          unsigned long consumed_count,
 417                          struct lttng_ust_shm_handle *handle)
 418 {
 419         unsigned long old_id, new_id;
 420
 421         if (config->mode == RING_BUFFER_OVERWRITE) {
 422                 /*
 423                  * Exchange the target writer subbuffer with our own unused
 424                  * subbuffer. No need to use CMM_ACCESS_ONCE() here to read the
 425                  * old_wpage, because the value read will be confirmed by the
 426                  * following cmpxchg().
 427                  */
 428                 old_id = shmp_index(handle, bufb->buf_wsb, consumed_idx)->id;
 429                 if (caa_unlikely(!subbuffer_id_is_noref(config, old_id)))
 430                         return -EAGAIN;
 431                 /*
 432                  * Make sure the offset count we are expecting matches the one
 433                  * indicated by the writer.
 434                  */
 435                 if (caa_unlikely(!subbuffer_id_compare_offset(config, old_id,
 436                                                           consumed_count)))
 437                         return -EAGAIN;
 438                 CHAN_WARN_ON(shmp(handle, bufb->chan),
 439                              !subbuffer_id_is_noref(config, bufb->buf_rsb.id));
 440                 subbuffer_id_set_noref_offset(config, &bufb->buf_rsb.id,
 441                                               consumed_count);
 442                 new_id = uatomic_cmpxchg(&shmp_index(handle, bufb->buf_wsb, consumed_idx)->id, old_id,
 443                                  bufb->buf_rsb.id);
 444                 if (caa_unlikely(old_id != new_id))
 445                         return -EAGAIN;
 446                 bufb->buf_rsb.id = new_id;
 447         } else {
 448                 /* No page exchange, use the writer page directly */
 449                 bufb->buf_rsb.id = shmp_index(handle, bufb->buf_wsb, consumed_idx)->id;
 450         }
 451         return 0;
 452 }
 453
 454 #ifndef inline_memcpy
 455 #define inline_memcpy(dest, src, n)     memcpy(dest, src, n)
 456 #endif
 457
 458 /*
 459  * Use the architecture-specific memcpy implementation for constant-sized
 460  * inputs, but rely on an inline memcpy for length statically unknown.
 461  * The function call to memcpy is just way too expensive for a fast path.
 462  */
 463 #define lib_ring_buffer_do_copy(config, dest, src, len)         \
 464 do {                                                            \
 465         size_t __len = (len);                                   \
 466         if (__builtin_constant_p(len))                          \
 467                 memcpy(dest, src, __len);                       \
 468         else                                                    \
 469                 inline_memcpy(dest, src, __len);                \
 470 } while (0)
 471
 472 /*
 473  * write len bytes to dest with c
 474  */
 475 static inline
 476 void lib_ring_buffer_do_memset(char *dest, int c, unsigned long len)
 477 {
 478         unsigned long i;
 479
 480         for (i = 0; i < len; i++)
 481                 dest[i] = c;
 482 }
 483
 484 /* arch-agnostic implementation */
 485
 486 static inline int lttng_ust_fls(unsigned int x)
 487 {
 488         int r = 32;
 489
 490         if (!x)
 491                 return 0;
 492         if (!(x & 0xFFFF0000U)) {
 493                 x <<= 16;
 494                 r -= 16;
 495         }
 496         if (!(x & 0xFF000000U)) {
 497                 x <<= 8;
 498                 r -= 8;
 499         }
 500         if (!(x & 0xF0000000U)) {
 501                 x <<= 4;
 502                 r -= 4;
 503         }
 504         if (!(x & 0xC0000000U)) {
 505                 x <<= 2;
 506                 r -= 2;
 507         }
 508         if (!(x & 0x80000000U)) {
 509                 /* No need to bit shift on last operation */
 510                 r -= 1;
 511         }
 512         return r;
 513 }
 514
 515 static inline int get_count_order(unsigned int count)
 516 {
 517         int order;
 518
 519         order = lttng_ust_fls(count) - 1;
 520         if (count & (count - 1))
 521                 order++;
 522         return order;
 523 }
 524
 525 #endif /* _LTTNG_RING_BUFFER_BACKEND_INTERNAL_H */