lib/ringbuffer/backend_internal.h

   1 /* SPDX-License-Identifier: (GPL-2.0-only OR LGPL-2.1-only)
   2  *
   3  * lib/ringbuffer/backend_internal.h
   4  *
   5  * Ring buffer backend (internal helpers).
   6  *
   7  * Copyright (C) 2008-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   8  */
   9
  10 #ifndef _LIB_RING_BUFFER_BACKEND_INTERNAL_H
  11 #define _LIB_RING_BUFFER_BACKEND_INTERNAL_H
  12
  13 #include <wrapper/compiler.h>
  14 #include <wrapper/ringbuffer/config.h>
  15 #include <wrapper/ringbuffer/backend_types.h>
  16 #include <wrapper/ringbuffer/frontend_types.h>
  17 #include <linux/string.h>
  18 #include <linux/uaccess.h>
  19
  20 /* Ring buffer backend API presented to the frontend */
  21
  22 /* Ring buffer and channel backend create/free */
  23
  24 int lib_ring_buffer_backend_create(struct lib_ring_buffer_backend *bufb,
  25                                    struct channel_backend *chan, int cpu);
  26 void channel_backend_unregister_notifiers(struct channel_backend *chanb);
  27 void lib_ring_buffer_backend_free(struct lib_ring_buffer_backend *bufb);
  28 int channel_backend_init(struct channel_backend *chanb,
  29                          const char *name,
  30                          const struct lib_ring_buffer_config *config,
  31                          void *priv, size_t subbuf_size,
  32                          size_t num_subbuf);
  33 void channel_backend_free(struct channel_backend *chanb);
  34
  35 void lib_ring_buffer_backend_reset(struct lib_ring_buffer_backend *bufb);
  36 void channel_backend_reset(struct channel_backend *chanb);
  37
  38 int lib_ring_buffer_backend_init(void);
  39 void lib_ring_buffer_backend_exit(void);
  40
  41 extern void _lib_ring_buffer_write(struct lib_ring_buffer_backend *bufb,
  42                                    size_t offset, const void *src, size_t len,
  43                                    size_t pagecpy);
  44 extern void _lib_ring_buffer_memset(struct lib_ring_buffer_backend *bufb,
  45                                     size_t offset, int c, size_t len,
  46                                     size_t pagecpy);
  47 extern void _lib_ring_buffer_strcpy(struct lib_ring_buffer_backend *bufb,
  48                                    size_t offset, const char *src, size_t len,
  49                                    size_t pagecpy, int pad);
  50 extern void _lib_ring_buffer_copy_from_user_inatomic(struct lib_ring_buffer_backend *bufb,
  51                                             size_t offset, const void *src,
  52                                             size_t len, size_t pagecpy);
  53 extern void _lib_ring_buffer_strcpy_from_user_inatomic(struct lib_ring_buffer_backend *bufb,
  54                 size_t offset, const char __user *src, size_t len,
  55                 size_t pagecpy, int pad);
  56
  57 /*
  58  * Subbuffer ID bits for overwrite mode. Need to fit within a single word to be
  59  * exchanged atomically.
  60  *
  61  * Top half word, except lowest bit, belongs to "offset", which is used to keep
  62  * to count the produced buffers.  For overwrite mode, this provides the
  63  * consumer with the capacity to read subbuffers in order, handling the
  64  * situation where producers would write up to 2^15 buffers (or 2^31 for 64-bit
  65  * systems) concurrently with a single execution of get_subbuf (between offset
  66  * sampling and subbuffer ID exchange).
  67  */
  68
  69 #define HALF_ULONG_BITS         (BITS_PER_LONG >> 1)
  70
  71 #define SB_ID_OFFSET_SHIFT      (HALF_ULONG_BITS + 1)
  72 #define SB_ID_OFFSET_COUNT      (1UL << SB_ID_OFFSET_SHIFT)
  73 #define SB_ID_OFFSET_MASK       (~(SB_ID_OFFSET_COUNT - 1))
  74 /*
  75  * Lowest bit of top word half belongs to noref. Used only for overwrite mode.
  76  */
  77 #define SB_ID_NOREF_SHIFT       (SB_ID_OFFSET_SHIFT - 1)
  78 #define SB_ID_NOREF_COUNT       (1UL << SB_ID_NOREF_SHIFT)
  79 #define SB_ID_NOREF_MASK        SB_ID_NOREF_COUNT
  80 /*
  81  * In overwrite mode: lowest half of word is used for index.
  82  * Limit of 2^16 subbuffers per buffer on 32-bit, 2^32 on 64-bit.
  83  * In producer-consumer mode: whole word used for index.
  84  */
  85 #define SB_ID_INDEX_SHIFT       0
  86 #define SB_ID_INDEX_COUNT       (1UL << SB_ID_INDEX_SHIFT)
  87 #define SB_ID_INDEX_MASK        (SB_ID_NOREF_COUNT - 1)
  88
  89 /*
  90  * Construct the subbuffer id from offset, index and noref. Use only the index
  91  * for producer-consumer mode (offset and noref are only used in overwrite
  92  * mode).
  93  */
  94 static inline
  95 unsigned long subbuffer_id(const struct lib_ring_buffer_config *config,
  96                            unsigned long offset, unsigned long noref,
  97                            unsigned long index)
  98 {
  99         if (config->mode == RING_BUFFER_OVERWRITE)
 100                 return (offset << SB_ID_OFFSET_SHIFT)
 101                        | (noref << SB_ID_NOREF_SHIFT)
 102                        | index;
 103         else
 104                 return index;
 105 }
 106
 107 /*
 108  * Compare offset with the offset contained within id. Return 1 if the offset
 109  * bits are identical, else 0.
 110  */
 111 static inline
 112 int subbuffer_id_compare_offset(const struct lib_ring_buffer_config *config,
 113                                 unsigned long id, unsigned long offset)
 114 {
 115         return (id & SB_ID_OFFSET_MASK) == (offset << SB_ID_OFFSET_SHIFT);
 116 }
 117
 118 static inline
 119 unsigned long subbuffer_id_get_index(const struct lib_ring_buffer_config *config,
 120                                      unsigned long id)
 121 {
 122         if (config->mode == RING_BUFFER_OVERWRITE)
 123                 return id & SB_ID_INDEX_MASK;
 124         else
 125                 return id;
 126 }
 127
 128 static inline
 129 unsigned long subbuffer_id_is_noref(const struct lib_ring_buffer_config *config,
 130                                     unsigned long id)
 131 {
 132         if (config->mode == RING_BUFFER_OVERWRITE)
 133                 return !!(id & SB_ID_NOREF_MASK);
 134         else
 135                 return 1;
 136 }
 137
 138 /*
 139  * Only used by reader on subbuffer ID it has exclusive access to. No volatile
 140  * needed.
 141  */
 142 static inline
 143 void subbuffer_id_set_noref(const struct lib_ring_buffer_config *config,
 144                             unsigned long *id)
 145 {
 146         if (config->mode == RING_BUFFER_OVERWRITE)
 147                 *id |= SB_ID_NOREF_MASK;
 148 }
 149
 150 static inline
 151 void subbuffer_id_set_noref_offset(const struct lib_ring_buffer_config *config,
 152                                    unsigned long *id, unsigned long offset)
 153 {
 154         unsigned long tmp;
 155
 156         if (config->mode == RING_BUFFER_OVERWRITE) {
 157                 tmp = *id;
 158                 tmp &= ~SB_ID_OFFSET_MASK;
 159                 tmp |= offset << SB_ID_OFFSET_SHIFT;
 160                 tmp |= SB_ID_NOREF_MASK;
 161                 /* Volatile store, read concurrently by readers. */
 162                 WRITE_ONCE(*id, tmp);
 163         }
 164 }
 165
 166 /* No volatile access, since already used locally */
 167 static inline
 168 void subbuffer_id_clear_noref(const struct lib_ring_buffer_config *config,
 169                               unsigned long *id)
 170 {
 171         if (config->mode == RING_BUFFER_OVERWRITE)
 172                 *id &= ~SB_ID_NOREF_MASK;
 173 }
 174
 175 /*
 176  * For overwrite mode, cap the number of subbuffers per buffer to:
 177  * 2^16 on 32-bit architectures
 178  * 2^32 on 64-bit architectures
 179  * This is required to fit in the index part of the ID. Return 0 on success,
 180  * -EPERM on failure.
 181  */
 182 static inline
 183 int subbuffer_id_check_index(const struct lib_ring_buffer_config *config,
 184                              unsigned long num_subbuf)
 185 {
 186         if (config->mode == RING_BUFFER_OVERWRITE)
 187                 return (num_subbuf > (1UL << HALF_ULONG_BITS)) ? -EPERM : 0;
 188         else
 189                 return 0;
 190 }
 191
 192 static inline
 193 void lib_ring_buffer_backend_get_pages(const struct lib_ring_buffer_config *config,
 194                         struct lib_ring_buffer_ctx *ctx,
 195                         struct lib_ring_buffer_backend_pages **backend_pages)
 196 {
 197         struct lib_ring_buffer_backend *bufb = &ctx->buf->backend;
 198         struct channel_backend *chanb = &ctx->chan->backend;
 199         size_t sbidx, offset = ctx->buf_offset;
 200         unsigned long sb_bindex, id;
 201         struct lib_ring_buffer_backend_pages *rpages;
 202
 203         offset &= chanb->buf_size - 1;
 204         sbidx = offset >> chanb->subbuf_size_order;
 205         id = bufb->buf_wsb[sbidx].id;
 206         sb_bindex = subbuffer_id_get_index(config, id);
 207         rpages = bufb->array[sb_bindex];
 208         CHAN_WARN_ON(ctx->chan,
 209                      config->mode == RING_BUFFER_OVERWRITE
 210                      && subbuffer_id_is_noref(config, id));
 211         *backend_pages = rpages;
 212 }
 213
 214 /* Get backend pages from cache. */
 215 static inline
 216 struct lib_ring_buffer_backend_pages *
 217         lib_ring_buffer_get_backend_pages_from_ctx(const struct lib_ring_buffer_config *config,
 218                 struct lib_ring_buffer_ctx *ctx)
 219 {
 220         return ctx->backend_pages;
 221 }
 222
 223 /*
 224  * The ring buffer can count events recorded and overwritten per buffer,
 225  * but it is disabled by default due to its performance overhead.
 226  */
 227 #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
 228 static inline
 229 void subbuffer_count_record(const struct lib_ring_buffer_config *config,
 230                             struct lib_ring_buffer_backend *bufb,
 231                             unsigned long idx)
 232 {
 233         unsigned long sb_bindex;
 234
 235         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 236         v_inc(config, &bufb->array[sb_bindex]->records_commit);
 237 }
 238 #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
 239 static inline
 240 void subbuffer_count_record(const struct lib_ring_buffer_config *config,
 241                             struct lib_ring_buffer_backend *bufb,
 242                             unsigned long idx)
 243 {
 244 }
 245 #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
 246
 247 /*
 248  * Reader has exclusive subbuffer access for record consumption. No need to
 249  * perform the decrement atomically.
 250  */
 251 static inline
 252 void subbuffer_consume_record(const struct lib_ring_buffer_config *config,
 253                               struct lib_ring_buffer_backend *bufb)
 254 {
 255         unsigned long sb_bindex;
 256
 257         sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
 258         CHAN_WARN_ON(bufb->chan,
 259                      !v_read(config, &bufb->array[sb_bindex]->records_unread));
 260         /* Non-atomic decrement protected by exclusive subbuffer access */
 261         _v_dec(config, &bufb->array[sb_bindex]->records_unread);
 262         v_inc(config, &bufb->records_read);
 263 }
 264
 265 static inline
 266 unsigned long subbuffer_get_records_count(
 267                                 const struct lib_ring_buffer_config *config,
 268                                 struct lib_ring_buffer_backend *bufb,
 269                                 unsigned long idx)
 270 {
 271         unsigned long sb_bindex;
 272
 273         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 274         return v_read(config, &bufb->array[sb_bindex]->records_commit);
 275 }
 276
 277 /*
 278  * Must be executed at subbuffer delivery when the writer has _exclusive_
 279  * subbuffer access. See lib_ring_buffer_check_deliver() for details.
 280  * lib_ring_buffer_get_records_count() must be called to get the records
 281  * count before this function, because it resets the records_commit
 282  * count.
 283  */
 284 static inline
 285 unsigned long subbuffer_count_records_overrun(
 286                                 const struct lib_ring_buffer_config *config,
 287                                 struct lib_ring_buffer_backend *bufb,
 288                                 unsigned long idx)
 289 {
 290         struct lib_ring_buffer_backend_pages *pages;
 291         unsigned long overruns, sb_bindex;
 292
 293         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 294         pages = bufb->array[sb_bindex];
 295         overruns = v_read(config, &pages->records_unread);
 296         v_set(config, &pages->records_unread,
 297               v_read(config, &pages->records_commit));
 298         v_set(config, &pages->records_commit, 0);
 299
 300         return overruns;
 301 }
 302
 303 static inline
 304 void subbuffer_set_data_size(const struct lib_ring_buffer_config *config,
 305                              struct lib_ring_buffer_backend *bufb,
 306                              unsigned long idx,
 307                              unsigned long data_size)
 308 {
 309         struct lib_ring_buffer_backend_pages *pages;
 310         unsigned long sb_bindex;
 311
 312         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 313         pages = bufb->array[sb_bindex];
 314         pages->data_size = data_size;
 315 }
 316
 317 static inline
 318 unsigned long subbuffer_get_read_data_size(
 319                                 const struct lib_ring_buffer_config *config,
 320                                 struct lib_ring_buffer_backend *bufb)
 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_rsb.id);
 326         pages = bufb->array[sb_bindex];
 327         return pages->data_size;
 328 }
 329
 330 static inline
 331 unsigned long subbuffer_get_data_size(
 332                                 const struct lib_ring_buffer_config *config,
 333                                 struct lib_ring_buffer_backend *bufb,
 334                                 unsigned long idx)
 335 {
 336         struct lib_ring_buffer_backend_pages *pages;
 337         unsigned long sb_bindex;
 338
 339         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 340         pages = bufb->array[sb_bindex];
 341         return pages->data_size;
 342 }
 343
 344 static inline
 345 void subbuffer_inc_packet_count(const struct lib_ring_buffer_config *config,
 346                                 struct lib_ring_buffer_backend *bufb,
 347                                 unsigned long idx)
 348 {
 349         bufb->buf_cnt[idx].seq_cnt++;
 350 }
 351
 352 /**
 353  * lib_ring_buffer_clear_noref - Clear the noref subbuffer flag, called by
 354  *                               writer.
 355  */
 356 static inline
 357 void lib_ring_buffer_clear_noref(const struct lib_ring_buffer_config *config,
 358                                  struct lib_ring_buffer_backend *bufb,
 359                                  unsigned long idx)
 360 {
 361         unsigned long id, new_id;
 362
 363         if (config->mode != RING_BUFFER_OVERWRITE)
 364                 return;
 365
 366         /*
 367          * Performing a volatile access to read the sb_pages, because we want to
 368          * read a coherent version of the pointer and the associated noref flag.
 369          */
 370         id = READ_ONCE(bufb->buf_wsb[idx].id);
 371         for (;;) {
 372                 /* This check is called on the fast path for each record. */
 373                 if (likely(!subbuffer_id_is_noref(config, id))) {
 374                         /*
 375                          * Store after load dependency ordering the writes to
 376                          * the subbuffer after load and test of the noref flag
 377                          * matches the memory barrier implied by the cmpxchg()
 378                          * in update_read_sb_index().
 379                          */
 380                         return; /* Already writing to this buffer */
 381                 }
 382                 new_id = id;
 383                 subbuffer_id_clear_noref(config, &new_id);
 384                 new_id = cmpxchg(&bufb->buf_wsb[idx].id, id, new_id);
 385                 if (likely(new_id == id))
 386                         break;
 387                 id = new_id;
 388         }
 389 }
 390
 391 /**
 392  * lib_ring_buffer_set_noref_offset - Set the noref subbuffer flag and offset,
 393  *                                    called by writer.
 394  */
 395 static inline
 396 void lib_ring_buffer_set_noref_offset(const struct lib_ring_buffer_config *config,
 397                                       struct lib_ring_buffer_backend *bufb,
 398                                       unsigned long idx, unsigned long offset)
 399 {
 400         if (config->mode != RING_BUFFER_OVERWRITE)
 401                 return;
 402
 403         /*
 404          * Because ring_buffer_set_noref() is only called by a single thread
 405          * (the one which updated the cc_sb value), there are no concurrent
 406          * updates to take care of: other writers have not updated cc_sb, so
 407          * they cannot set the noref flag, and concurrent readers cannot modify
 408          * the pointer because the noref flag is not set yet.
 409          * The smp_wmb() in ring_buffer_commit() takes care of ordering writes
 410          * to the subbuffer before this set noref operation.
 411          * subbuffer_set_noref() uses a volatile store to deal with concurrent
 412          * readers of the noref flag.
 413          */
 414         CHAN_WARN_ON(bufb->chan,
 415                      subbuffer_id_is_noref(config, bufb->buf_wsb[idx].id));
 416         /*
 417          * Memory barrier that ensures counter stores are ordered before set
 418          * noref and offset.
 419          */
 420         smp_mb();
 421         subbuffer_id_set_noref_offset(config, &bufb->buf_wsb[idx].id, offset);
 422 }
 423
 424 /**
 425  * update_read_sb_index - Read-side subbuffer index update.
 426  */
 427 static inline
 428 int update_read_sb_index(const struct lib_ring_buffer_config *config,
 429                          struct lib_ring_buffer_backend *bufb,
 430                          struct channel_backend *chanb,
 431                          unsigned long consumed_idx,
 432                          unsigned long consumed_count)
 433 {
 434         unsigned long old_id, new_id;
 435
 436         if (config->mode == RING_BUFFER_OVERWRITE) {
 437                 /*
 438                  * Exchange the target writer subbuffer with our own unused
 439                  * subbuffer. No need to use READ_ONCE() here to read the
 440                  * old_wpage, because the value read will be confirmed by the
 441                  * following cmpxchg().
 442                  */
 443                 old_id = bufb->buf_wsb[consumed_idx].id;
 444                 if (unlikely(!subbuffer_id_is_noref(config, old_id)))
 445                         return -EAGAIN;
 446                 /*
 447                  * Make sure the offset count we are expecting matches the one
 448                  * indicated by the writer.
 449                  */
 450                 if (unlikely(!subbuffer_id_compare_offset(config, old_id,
 451                                                           consumed_count)))
 452                         return -EAGAIN;
 453                 CHAN_WARN_ON(bufb->chan,
 454                              !subbuffer_id_is_noref(config, bufb->buf_rsb.id));
 455                 subbuffer_id_set_noref_offset(config, &bufb->buf_rsb.id,
 456                                               consumed_count);
 457                 new_id = cmpxchg(&bufb->buf_wsb[consumed_idx].id, old_id,
 458                                  bufb->buf_rsb.id);
 459                 if (unlikely(old_id != new_id))
 460                         return -EAGAIN;
 461                 bufb->buf_rsb.id = new_id;
 462         } else {
 463                 /* No page exchange, use the writer page directly */
 464                 bufb->buf_rsb.id = bufb->buf_wsb[consumed_idx].id;
 465         }
 466         return 0;
 467 }
 468
 469 static inline __attribute__((always_inline))
 470 void lttng_inline_memcpy(void *dest, const void *src,
 471                 unsigned long len)
 472 {
 473         switch (len) {
 474         case 1:
 475                 *(uint8_t *) dest = *(const uint8_t *) src;
 476                 break;
 477         case 2:
 478                 *(uint16_t *) dest = *(const uint16_t *) src;
 479                 break;
 480         case 4:
 481                 *(uint32_t *) dest = *(const uint32_t *) src;
 482                 break;
 483         case 8:
 484                 *(uint64_t *) dest = *(const uint64_t *) src;
 485                 break;
 486         default:
 487                 inline_memcpy(dest, src, len);
 488         }
 489 }
 490
 491 /*
 492  * Use the architecture-specific memcpy implementation for constant-sized
 493  * inputs, but rely on an inline memcpy for length statically unknown.
 494  * The function call to memcpy is just way too expensive for a fast path.
 495  */
 496 #define lib_ring_buffer_do_copy(config, dest, src, len)         \
 497 do {                                                            \
 498         size_t __len = (len);                                   \
 499         if (__builtin_constant_p(len))                          \
 500                 memcpy(dest, src, __len);                       \
 501         else                                                    \
 502                 lttng_inline_memcpy(dest, src, __len);          \
 503 } while (0)
 504
 505 /*
 506  * We use __copy_from_user_inatomic to copy userspace data since we already
 507  * did the access_ok for the whole range.
 508  *
 509  * Return 0 if OK, nonzero on error.
 510  */
 511 static inline
 512 unsigned long lib_ring_buffer_do_copy_from_user_inatomic(void *dest,
 513                                                 const void __user *src,
 514                                                 unsigned long len)
 515 {
 516         return __copy_from_user_inatomic(dest, src, len);
 517 }
 518
 519 /*
 520  * write len bytes to dest with c
 521  */
 522 static inline
 523 void lib_ring_buffer_do_memset(char *dest, int c,
 524         unsigned long len)
 525 {
 526         unsigned long i;
 527
 528         for (i = 0; i < len; i++)
 529                 dest[i] = c;
 530 }
 531
 532 #endif /* _LIB_RING_BUFFER_BACKEND_INTERNAL_H */