include/ringbuffer/backend_internal.h

   1 /* SPDX-License-Identifier: (GPL-2.0-only OR LGPL-2.1-only)
   2  *
   3  * 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/inline_memcpy.h>
  15 #include <ringbuffer/config.h>
  16 #include <ringbuffer/backend_types.h>
  17 #include <ringbuffer/frontend_types.h>
  18 #include <linux/string.h>
  19 #include <linux/uaccess.h>
  20
  21 /* Ring buffer backend API presented to the frontend */
  22
  23 /* Ring buffer and channel backend create/free */
  24
  25 int lib_ring_buffer_backend_create(struct lttng_kernel_ring_buffer_backend *bufb,
  26                                    struct channel_backend *chan, int cpu);
  27 void channel_backend_unregister_notifiers(struct channel_backend *chanb);
  28 void lib_ring_buffer_backend_free(struct lttng_kernel_ring_buffer_backend *bufb);
  29 int channel_backend_init(struct channel_backend *chanb,
  30                          const char *name,
  31                          const struct lttng_kernel_ring_buffer_config *config,
  32                          void *priv, size_t subbuf_size,
  33                          size_t num_subbuf);
  34 void channel_backend_free(struct channel_backend *chanb);
  35
  36 void lib_ring_buffer_backend_reset(struct lttng_kernel_ring_buffer_backend *bufb);
  37 void channel_backend_reset(struct channel_backend *chanb);
  38
  39 int lib_ring_buffer_backend_init(void);
  40 void lib_ring_buffer_backend_exit(void);
  41
  42 extern void _lib_ring_buffer_write(struct lttng_kernel_ring_buffer_backend *bufb,
  43                                    size_t offset, const void *src, size_t len,
  44                                    size_t pagecpy);
  45 extern void _lib_ring_buffer_memset(struct lttng_kernel_ring_buffer_backend *bufb,
  46                                     size_t offset, int c, size_t len,
  47                                     size_t pagecpy);
  48 extern void _lib_ring_buffer_strcpy(struct lttng_kernel_ring_buffer_backend *bufb,
  49                                    size_t offset, const char *src, size_t len, int pad);
  50 extern void _lib_ring_buffer_pstrcpy(struct lttng_kernel_ring_buffer_backend *bufb,
  51                                    size_t offset, const char *src, size_t len, int pad);
  52 extern void _lib_ring_buffer_copy_from_user_inatomic(struct lttng_kernel_ring_buffer_backend *bufb,
  53                                             size_t offset, const void *src,
  54                                             size_t len, size_t pagecpy);
  55 extern void _lib_ring_buffer_strcpy_from_user_inatomic(struct lttng_kernel_ring_buffer_backend *bufb,
  56                 size_t offset, const char __user *src, size_t len, int pad);
  57 extern void _lib_ring_buffer_pstrcpy_from_user_inatomic(struct lttng_kernel_ring_buffer_backend *bufb,
  58                                    size_t offset, const char __user *src, size_t len, int pad);
  59
  60 /*
  61  * Subbuffer ID bits for overwrite mode. Need to fit within a single word to be
  62  * exchanged atomically.
  63  *
  64  * Top half word, except lowest bit, belongs to "offset", which is used to keep
  65  * to count the produced buffers.  For overwrite mode, this provides the
  66  * consumer with the capacity to read subbuffers in order, handling the
  67  * situation where producers would write up to 2^15 buffers (or 2^31 for 64-bit
  68  * systems) concurrently with a single execution of get_subbuf (between offset
  69  * sampling and subbuffer ID exchange).
  70  */
  71
  72 #define HALF_ULONG_BITS         (BITS_PER_LONG >> 1)
  73
  74 #define SB_ID_OFFSET_SHIFT      (HALF_ULONG_BITS + 1)
  75 #define SB_ID_OFFSET_COUNT      (1UL << SB_ID_OFFSET_SHIFT)
  76 #define SB_ID_OFFSET_MASK       (~(SB_ID_OFFSET_COUNT - 1))
  77 /*
  78  * Lowest bit of top word half belongs to noref. Used only for overwrite mode.
  79  */
  80 #define SB_ID_NOREF_SHIFT       (SB_ID_OFFSET_SHIFT - 1)
  81 #define SB_ID_NOREF_COUNT       (1UL << SB_ID_NOREF_SHIFT)
  82 #define SB_ID_NOREF_MASK        SB_ID_NOREF_COUNT
  83 /*
  84  * In overwrite mode: lowest half of word is used for index.
  85  * Limit of 2^16 subbuffers per buffer on 32-bit, 2^32 on 64-bit.
  86  * In producer-consumer mode: whole word used for index.
  87  */
  88 #define SB_ID_INDEX_SHIFT       0
  89 #define SB_ID_INDEX_COUNT       (1UL << SB_ID_INDEX_SHIFT)
  90 #define SB_ID_INDEX_MASK        (SB_ID_NOREF_COUNT - 1)
  91
  92 /*
  93  * Construct the subbuffer id from offset, index and noref. Use only the index
  94  * for producer-consumer mode (offset and noref are only used in overwrite
  95  * mode).
  96  */
  97 static inline
  98 unsigned long subbuffer_id(const struct lttng_kernel_ring_buffer_config *config,
  99                            unsigned long offset, unsigned long noref,
 100                            unsigned long index)
 101 {
 102         if (config->mode == RING_BUFFER_OVERWRITE)
 103                 return (offset << SB_ID_OFFSET_SHIFT)
 104                        | (noref << SB_ID_NOREF_SHIFT)
 105                        | index;
 106         else
 107                 return index;
 108 }
 109
 110 /*
 111  * Compare offset with the offset contained within id. Return 1 if the offset
 112  * bits are identical, else 0.
 113  */
 114 static inline
 115 int subbuffer_id_compare_offset(const struct lttng_kernel_ring_buffer_config *config,
 116                                 unsigned long id, unsigned long offset)
 117 {
 118         return (id & SB_ID_OFFSET_MASK) == (offset << SB_ID_OFFSET_SHIFT);
 119 }
 120
 121 static inline
 122 unsigned long subbuffer_id_get_index(const struct lttng_kernel_ring_buffer_config *config,
 123                                      unsigned long id)
 124 {
 125         if (config->mode == RING_BUFFER_OVERWRITE)
 126                 return id & SB_ID_INDEX_MASK;
 127         else
 128                 return id;
 129 }
 130
 131 static inline
 132 unsigned long subbuffer_id_is_noref(const struct lttng_kernel_ring_buffer_config *config,
 133                                     unsigned long id)
 134 {
 135         if (config->mode == RING_BUFFER_OVERWRITE)
 136                 return !!(id & SB_ID_NOREF_MASK);
 137         else
 138                 return 1;
 139 }
 140
 141 /*
 142  * Only used by reader on subbuffer ID it has exclusive access to. No volatile
 143  * needed.
 144  */
 145 static inline
 146 void subbuffer_id_set_noref(const struct lttng_kernel_ring_buffer_config *config,
 147                             unsigned long *id)
 148 {
 149         if (config->mode == RING_BUFFER_OVERWRITE)
 150                 *id |= SB_ID_NOREF_MASK;
 151 }
 152
 153 static inline
 154 void subbuffer_id_set_noref_offset(const struct lttng_kernel_ring_buffer_config *config,
 155                                    unsigned long *id, unsigned long offset)
 156 {
 157         unsigned long tmp;
 158
 159         if (config->mode == RING_BUFFER_OVERWRITE) {
 160                 tmp = *id;
 161                 tmp &= ~SB_ID_OFFSET_MASK;
 162                 tmp |= offset << SB_ID_OFFSET_SHIFT;
 163                 tmp |= SB_ID_NOREF_MASK;
 164                 /* Volatile store, read concurrently by readers. */
 165                 WRITE_ONCE(*id, tmp);
 166         }
 167 }
 168
 169 /* No volatile access, since already used locally */
 170 static inline
 171 void subbuffer_id_clear_noref(const struct lttng_kernel_ring_buffer_config *config,
 172                               unsigned long *id)
 173 {
 174         if (config->mode == RING_BUFFER_OVERWRITE)
 175                 *id &= ~SB_ID_NOREF_MASK;
 176 }
 177
 178 /*
 179  * For overwrite mode, cap the number of subbuffers per buffer to:
 180  * 2^16 on 32-bit architectures
 181  * 2^32 on 64-bit architectures
 182  * This is required to fit in the index part of the ID. Return 0 on success,
 183  * -EPERM on failure.
 184  */
 185 static inline
 186 int subbuffer_id_check_index(const struct lttng_kernel_ring_buffer_config *config,
 187                              unsigned long num_subbuf)
 188 {
 189         if (config->mode == RING_BUFFER_OVERWRITE)
 190                 return (num_subbuf > (1UL << HALF_ULONG_BITS)) ? -EPERM : 0;
 191         else
 192                 return 0;
 193 }
 194
 195 static inline
 196 void lib_ring_buffer_backend_get_pages(const struct lttng_kernel_ring_buffer_config *config,
 197                         struct lttng_kernel_ring_buffer_ctx *ctx,
 198                         struct lttng_kernel_ring_buffer_backend_pages **backend_pages)
 199 {
 200         struct lttng_kernel_ring_buffer_backend *bufb = &ctx->priv.buf->backend;
 201         struct channel_backend *chanb = &ctx->priv.chan->backend;
 202         size_t sbidx, offset = ctx->priv.buf_offset;
 203         unsigned long sb_bindex, id;
 204         struct lttng_kernel_ring_buffer_backend_pages *rpages;
 205
 206         offset &= chanb->buf_size - 1;
 207         sbidx = offset >> chanb->subbuf_size_order;
 208         id = bufb->buf_wsb[sbidx].id;
 209         sb_bindex = subbuffer_id_get_index(config, id);
 210         rpages = bufb->array[sb_bindex];
 211         CHAN_WARN_ON(ctx->priv.chan,
 212                      config->mode == RING_BUFFER_OVERWRITE
 213                      && subbuffer_id_is_noref(config, id));
 214         *backend_pages = rpages;
 215 }
 216
 217 /* Get backend pages from cache. */
 218 static inline
 219 struct lttng_kernel_ring_buffer_backend_pages *
 220         lib_ring_buffer_get_backend_pages_from_ctx(const struct lttng_kernel_ring_buffer_config *config,
 221                 struct lttng_kernel_ring_buffer_ctx *ctx)
 222 {
 223         return ctx->priv.backend_pages;
 224 }
 225
 226 /*
 227  * The ring buffer can count events recorded and overwritten per buffer,
 228  * but it is disabled by default due to its performance overhead.
 229  */
 230 #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
 231 static inline
 232 void subbuffer_count_record(const struct lttng_kernel_ring_buffer_config *config,
 233                             struct lttng_kernel_ring_buffer_backend *bufb,
 234                             unsigned long idx)
 235 {
 236         unsigned long sb_bindex;
 237
 238         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 239         v_inc(config, &bufb->array[sb_bindex]->records_commit);
 240 }
 241
 242 /*
 243  * Reader has exclusive subbuffer access for record consumption. No need to
 244  * perform the decrement atomically.
 245  */
 246 static inline
 247 void subbuffer_consume_record(const struct lttng_kernel_ring_buffer_config *config,
 248                               struct lttng_kernel_ring_buffer_backend *bufb)
 249 {
 250         unsigned long sb_bindex;
 251
 252         sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
 253         CHAN_WARN_ON(bufb->chan,
 254                      !v_read(config, &bufb->array[sb_bindex]->records_unread));
 255         /* Non-atomic decrement protected by exclusive subbuffer access */
 256         _v_dec(config, &bufb->array[sb_bindex]->records_unread);
 257         v_inc(config, &bufb->records_read);
 258 }
 259 #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
 260 static inline
 261 void subbuffer_count_record(const struct lttng_kernel_ring_buffer_config *config,
 262                             struct lttng_kernel_ring_buffer_backend *bufb,
 263                             unsigned long idx)
 264 {
 265 }
 266 static inline
 267 void subbuffer_consume_record(const struct lttng_kernel_ring_buffer_config *config,
 268                               struct lttng_kernel_ring_buffer_backend *bufb)
 269 {
 270 }
 271 #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
 272
 273 static inline
 274 unsigned long subbuffer_get_records_count(
 275                                 const struct lttng_kernel_ring_buffer_config *config,
 276                                 struct lttng_kernel_ring_buffer_backend *bufb,
 277                                 unsigned long idx)
 278 {
 279         unsigned long sb_bindex;
 280
 281         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 282         return v_read(config, &bufb->array[sb_bindex]->records_commit);
 283 }
 284
 285 /*
 286  * Must be executed at subbuffer delivery when the writer has _exclusive_
 287  * subbuffer access. See lib_ring_buffer_check_deliver() for details.
 288  * lib_ring_buffer_get_records_count() must be called to get the records
 289  * count before this function, because it resets the records_commit
 290  * count.
 291  */
 292 static inline
 293 unsigned long subbuffer_count_records_overrun(
 294                                 const struct lttng_kernel_ring_buffer_config *config,
 295                                 struct lttng_kernel_ring_buffer_backend *bufb,
 296                                 unsigned long idx)
 297 {
 298         struct lttng_kernel_ring_buffer_backend_pages *pages;
 299         unsigned long overruns, sb_bindex;
 300
 301         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 302         pages = bufb->array[sb_bindex];
 303         overruns = v_read(config, &pages->records_unread);
 304         v_set(config, &pages->records_unread,
 305               v_read(config, &pages->records_commit));
 306         v_set(config, &pages->records_commit, 0);
 307
 308         return overruns;
 309 }
 310
 311 static inline
 312 void subbuffer_set_data_size(const struct lttng_kernel_ring_buffer_config *config,
 313                              struct lttng_kernel_ring_buffer_backend *bufb,
 314                              unsigned long idx,
 315                              unsigned long data_size)
 316 {
 317         struct lttng_kernel_ring_buffer_backend_pages *pages;
 318         unsigned long sb_bindex;
 319
 320         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 321         pages = bufb->array[sb_bindex];
 322         pages->data_size = data_size;
 323 }
 324
 325 static inline
 326 unsigned long subbuffer_get_read_data_size(
 327                                 const struct lttng_kernel_ring_buffer_config *config,
 328                                 struct lttng_kernel_ring_buffer_backend *bufb)
 329 {
 330         struct lttng_kernel_ring_buffer_backend_pages *pages;
 331         unsigned long sb_bindex;
 332
 333         sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
 334         pages = bufb->array[sb_bindex];
 335         return pages->data_size;
 336 }
 337
 338 static inline
 339 unsigned long subbuffer_get_data_size(
 340                                 const struct lttng_kernel_ring_buffer_config *config,
 341                                 struct lttng_kernel_ring_buffer_backend *bufb,
 342                                 unsigned long idx)
 343 {
 344         struct lttng_kernel_ring_buffer_backend_pages *pages;
 345         unsigned long sb_bindex;
 346
 347         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 348         pages = bufb->array[sb_bindex];
 349         return pages->data_size;
 350 }
 351
 352 static inline
 353 void subbuffer_inc_packet_count(const struct lttng_kernel_ring_buffer_config *config,
 354                                 struct lttng_kernel_ring_buffer_backend *bufb,
 355                                 unsigned long idx)
 356 {
 357         bufb->buf_cnt[idx].seq_cnt++;
 358 }
 359
 360 /**
 361  * lib_ring_buffer_clear_noref - Clear the noref subbuffer flag, called by
 362  *                               writer.
 363  */
 364 static inline
 365 void lib_ring_buffer_clear_noref(const struct lttng_kernel_ring_buffer_config *config,
 366                                  struct lttng_kernel_ring_buffer_backend *bufb,
 367                                  unsigned long idx)
 368 {
 369         unsigned long id, new_id;
 370
 371         if (config->mode != RING_BUFFER_OVERWRITE)
 372                 return;
 373
 374         /*
 375          * Performing a volatile access to read the sb_pages, because we want to
 376          * read a coherent version of the pointer and the associated noref flag.
 377          */
 378         id = LTTNG_READ_ONCE(bufb->buf_wsb[idx].id);
 379         for (;;) {
 380                 /* This check is called on the fast path for each record. */
 381                 if (likely(!subbuffer_id_is_noref(config, id))) {
 382                         /*
 383                          * Store after load dependency ordering the writes to
 384                          * the subbuffer after load and test of the noref flag
 385                          * matches the memory barrier implied by the cmpxchg()
 386                          * in update_read_sb_index().
 387                          */
 388                         return; /* Already writing to this buffer */
 389                 }
 390                 new_id = id;
 391                 subbuffer_id_clear_noref(config, &new_id);
 392                 new_id = cmpxchg(&bufb->buf_wsb[idx].id, id, new_id);
 393                 if (likely(new_id == id))
 394                         break;
 395                 id = new_id;
 396         }
 397 }
 398
 399 /**
 400  * lib_ring_buffer_set_noref_offset - Set the noref subbuffer flag and offset,
 401  *                                    called by writer.
 402  */
 403 static inline
 404 void lib_ring_buffer_set_noref_offset(const struct lttng_kernel_ring_buffer_config *config,
 405                                       struct lttng_kernel_ring_buffer_backend *bufb,
 406                                       unsigned long idx, unsigned long offset)
 407 {
 408         if (config->mode != RING_BUFFER_OVERWRITE)
 409                 return;
 410
 411         /*
 412          * Because ring_buffer_set_noref() is only called by a single thread
 413          * (the one which updated the cc_sb value), there are no concurrent
 414          * updates to take care of: other writers have not updated cc_sb, so
 415          * they cannot set the noref flag, and concurrent readers cannot modify
 416          * the pointer because the noref flag is not set yet.
 417          * The smp_wmb() in ring_buffer_commit() takes care of ordering writes
 418          * to the subbuffer before this set noref operation.
 419          * subbuffer_set_noref() uses a volatile store to deal with concurrent
 420          * readers of the noref flag.
 421          */
 422         CHAN_WARN_ON(bufb->chan,
 423                      subbuffer_id_is_noref(config, bufb->buf_wsb[idx].id));
 424         /*
 425          * Memory barrier that ensures counter stores are ordered before set
 426          * noref and offset.
 427          */
 428         smp_mb();
 429         subbuffer_id_set_noref_offset(config, &bufb->buf_wsb[idx].id, offset);
 430 }
 431
 432 /**
 433  * update_read_sb_index - Read-side subbuffer index update.
 434  */
 435 static inline
 436 int update_read_sb_index(const struct lttng_kernel_ring_buffer_config *config,
 437                          struct lttng_kernel_ring_buffer_backend *bufb,
 438                          struct channel_backend *chanb,
 439                          unsigned long consumed_idx,
 440                          unsigned long consumed_count)
 441 {
 442         unsigned long old_id, new_id;
 443
 444         if (config->mode == RING_BUFFER_OVERWRITE) {
 445                 /*
 446                  * Exchange the target writer subbuffer with our own unused
 447                  * subbuffer. No need to use READ_ONCE() here to read the
 448                  * old_wpage, because the value read will be confirmed by the
 449                  * following cmpxchg().
 450                  */
 451                 old_id = bufb->buf_wsb[consumed_idx].id;
 452                 if (unlikely(!subbuffer_id_is_noref(config, old_id)))
 453                         return -EAGAIN;
 454                 /*
 455                  * Make sure the offset count we are expecting matches the one
 456                  * indicated by the writer.
 457                  */
 458                 if (unlikely(!subbuffer_id_compare_offset(config, old_id,
 459                                                           consumed_count)))
 460                         return -EAGAIN;
 461                 CHAN_WARN_ON(bufb->chan,
 462                              !subbuffer_id_is_noref(config, bufb->buf_rsb.id));
 463                 subbuffer_id_set_noref_offset(config, &bufb->buf_rsb.id,
 464                                               consumed_count);
 465                 new_id = cmpxchg(&bufb->buf_wsb[consumed_idx].id, old_id,
 466                                  bufb->buf_rsb.id);
 467                 if (unlikely(old_id != new_id))
 468                         return -EAGAIN;
 469                 bufb->buf_rsb.id = new_id;
 470         } else {
 471                 /* No page exchange, use the writer page directly */
 472                 bufb->buf_rsb.id = bufb->buf_wsb[consumed_idx].id;
 473         }
 474         return 0;
 475 }
 476
 477 static inline __attribute__((always_inline))
 478 void lttng_inline_memcpy(void *dest, const void *src,
 479                 unsigned long len)
 480 {
 481         switch (len) {
 482         case 1:
 483                 *(uint8_t *) dest = *(const uint8_t *) src;
 484                 break;
 485         case 2:
 486                 *(uint16_t *) dest = *(const uint16_t *) src;
 487                 break;
 488         case 4:
 489                 *(uint32_t *) dest = *(const uint32_t *) src;
 490                 break;
 491         case 8:
 492                 *(uint64_t *) dest = *(const uint64_t *) src;
 493                 break;
 494         default:
 495                 inline_memcpy(dest, src, len);
 496         }
 497 }
 498
 499 /*
 500  * Use the architecture-specific memcpy implementation for constant-sized
 501  * inputs, but rely on an inline memcpy for length statically unknown.
 502  * The function call to memcpy is just way too expensive for a fast path.
 503  */
 504 #define lib_ring_buffer_do_copy(config, dest, src, len)         \
 505 do {                                                            \
 506         size_t __len = (len);                                   \
 507         if (__builtin_constant_p(len))                          \
 508                 memcpy(dest, src, __len);                       \
 509         else                                                    \
 510                 lttng_inline_memcpy(dest, src, __len);          \
 511 } while (0)
 512
 513 /*
 514  * We use __copy_from_user_inatomic to copy userspace data since we already
 515  * did the access_ok for the whole range.
 516  *
 517  * Return 0 if OK, nonzero on error.
 518  */
 519 static inline
 520 unsigned long lib_ring_buffer_do_copy_from_user_inatomic(void *dest,
 521                                                 const void __user *src,
 522                                                 unsigned long len)
 523 {
 524         return __copy_from_user_inatomic(dest, src, len);
 525 }
 526
 527 /*
 528  * write len bytes to dest with c
 529  */
 530 static inline
 531 void lib_ring_buffer_do_memset(char *dest, int c,
 532         unsigned long len)
 533 {
 534         unsigned long i;
 535
 536         for (i = 0; i < len; i++)
 537                 dest[i] = c;
 538 }
 539
 540 #endif /* _LIB_RING_BUFFER_BACKEND_INTERNAL_H */