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 extern void _lib_ring_buffer_memset(struct lttng_kernel_ring_buffer_backend *bufb,
  45                                     size_t offset, int c, size_t len);
  46 extern void _lib_ring_buffer_strcpy(struct lttng_kernel_ring_buffer_backend *bufb,
  47                                    size_t offset, const char *src, size_t len, int pad);
  48 extern void _lib_ring_buffer_pstrcpy(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_copy_from_user_inatomic(struct lttng_kernel_ring_buffer_backend *bufb,
  51                                             size_t offset, const void __user *src, size_t len);
  52 extern void _lib_ring_buffer_strcpy_from_user_inatomic(struct lttng_kernel_ring_buffer_backend *bufb,
  53                 size_t offset, const char __user *src, size_t len, int pad);
  54 extern void _lib_ring_buffer_pstrcpy_from_user_inatomic(struct lttng_kernel_ring_buffer_backend *bufb,
  55                                    size_t offset, const char __user *src, size_t len, 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 lttng_kernel_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 lttng_kernel_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 lttng_kernel_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 lttng_kernel_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 lttng_kernel_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 lttng_kernel_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 lttng_kernel_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 lttng_kernel_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 lttng_kernel_ring_buffer_config *config,
 194                         struct lttng_kernel_ring_buffer_ctx *ctx,
 195                         struct lttng_kernel_ring_buffer_backend_pages **backend_pages)
 196 {
 197         struct lttng_kernel_ring_buffer_backend *bufb = &ctx->priv.buf->backend;
 198         struct channel_backend *chanb = &ctx->priv.chan->backend;
 199         size_t sbidx, offset = ctx->priv.buf_offset;
 200         unsigned long sb_bindex, id;
 201         struct lttng_kernel_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->priv.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 lttng_kernel_ring_buffer_backend_pages *
 217         lib_ring_buffer_get_backend_pages_from_ctx(const struct lttng_kernel_ring_buffer_config *config,
 218                 struct lttng_kernel_ring_buffer_ctx *ctx)
 219 {
 220         return ctx->priv.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 lttng_kernel_ring_buffer_config *config,
 230                             struct lttng_kernel_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
 239 /*
 240  * Reader has exclusive subbuffer access for record consumption. No need to
 241  * perform the decrement atomically.
 242  */
 243 static inline
 244 void subbuffer_consume_record(const struct lttng_kernel_ring_buffer_config *config,
 245                               struct lttng_kernel_ring_buffer_backend *bufb)
 246 {
 247         unsigned long sb_bindex;
 248
 249         sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
 250         CHAN_WARN_ON(bufb->chan,
 251                      !v_read(config, &bufb->array[sb_bindex]->records_unread));
 252         /* Non-atomic decrement protected by exclusive subbuffer access */
 253         _v_dec(config, &bufb->array[sb_bindex]->records_unread);
 254         v_inc(config, &bufb->records_read);
 255 }
 256 #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
 257 static inline
 258 void subbuffer_count_record(const struct lttng_kernel_ring_buffer_config *config,
 259                             struct lttng_kernel_ring_buffer_backend *bufb,
 260                             unsigned long idx)
 261 {
 262 }
 263 static inline
 264 void subbuffer_consume_record(const struct lttng_kernel_ring_buffer_config *config,
 265                               struct lttng_kernel_ring_buffer_backend *bufb)
 266 {
 267 }
 268 #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
 269
 270 static inline
 271 unsigned long subbuffer_get_records_count(
 272                                 const struct lttng_kernel_ring_buffer_config *config,
 273                                 struct lttng_kernel_ring_buffer_backend *bufb,
 274                                 unsigned long idx)
 275 {
 276         unsigned long sb_bindex;
 277
 278         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 279         return v_read(config, &bufb->array[sb_bindex]->records_commit);
 280 }
 281
 282 /*
 283  * Must be executed at subbuffer delivery when the writer has _exclusive_
 284  * subbuffer access. See lib_ring_buffer_check_deliver() for details.
 285  * lib_ring_buffer_get_records_count() must be called to get the records
 286  * count before this function, because it resets the records_commit
 287  * count.
 288  */
 289 static inline
 290 unsigned long subbuffer_count_records_overrun(
 291                                 const struct lttng_kernel_ring_buffer_config *config,
 292                                 struct lttng_kernel_ring_buffer_backend *bufb,
 293                                 unsigned long idx)
 294 {
 295         struct lttng_kernel_ring_buffer_backend_pages *pages;
 296         unsigned long overruns, sb_bindex;
 297
 298         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 299         pages = bufb->array[sb_bindex];
 300         overruns = v_read(config, &pages->records_unread);
 301         v_set(config, &pages->records_unread,
 302               v_read(config, &pages->records_commit));
 303         v_set(config, &pages->records_commit, 0);
 304
 305         return overruns;
 306 }
 307
 308 static inline
 309 void subbuffer_set_data_size(const struct lttng_kernel_ring_buffer_config *config,
 310                              struct lttng_kernel_ring_buffer_backend *bufb,
 311                              unsigned long idx,
 312                              unsigned long data_size)
 313 {
 314         struct lttng_kernel_ring_buffer_backend_pages *pages;
 315         unsigned long sb_bindex;
 316
 317         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 318         pages = bufb->array[sb_bindex];
 319         pages->data_size = data_size;
 320 }
 321
 322 static inline
 323 unsigned long subbuffer_get_read_data_size(
 324                                 const struct lttng_kernel_ring_buffer_config *config,
 325                                 struct lttng_kernel_ring_buffer_backend *bufb)
 326 {
 327         struct lttng_kernel_ring_buffer_backend_pages *pages;
 328         unsigned long sb_bindex;
 329
 330         sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
 331         pages = bufb->array[sb_bindex];
 332         return pages->data_size;
 333 }
 334
 335 static inline
 336 unsigned long subbuffer_get_data_size(
 337                                 const struct lttng_kernel_ring_buffer_config *config,
 338                                 struct lttng_kernel_ring_buffer_backend *bufb,
 339                                 unsigned long idx)
 340 {
 341         struct lttng_kernel_ring_buffer_backend_pages *pages;
 342         unsigned long sb_bindex;
 343
 344         sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
 345         pages = bufb->array[sb_bindex];
 346         return pages->data_size;
 347 }
 348
 349 static inline
 350 void subbuffer_inc_packet_count(const struct lttng_kernel_ring_buffer_config *config,
 351                                 struct lttng_kernel_ring_buffer_backend *bufb,
 352                                 unsigned long idx)
 353 {
 354         bufb->buf_cnt[idx].seq_cnt++;
 355 }
 356
 357 /**
 358  * lib_ring_buffer_clear_noref - Clear the noref subbuffer flag, called by
 359  *                               writer.
 360  */
 361 static inline
 362 void lib_ring_buffer_clear_noref(const struct lttng_kernel_ring_buffer_config *config,
 363                                  struct lttng_kernel_ring_buffer_backend *bufb,
 364                                  unsigned long idx)
 365 {
 366         unsigned long id, new_id;
 367
 368         if (config->mode != RING_BUFFER_OVERWRITE)
 369                 return;
 370
 371         /*
 372          * Performing a volatile access to read the sb_pages, because we want to
 373          * read a coherent version of the pointer and the associated noref flag.
 374          */
 375         id = LTTNG_READ_ONCE(bufb->buf_wsb[idx].id);
 376         for (;;) {
 377                 /* This check is called on the fast path for each record. */
 378                 if (likely(!subbuffer_id_is_noref(config, id))) {
 379                         /*
 380                          * Store after load dependency ordering the writes to
 381                          * the subbuffer after load and test of the noref flag
 382                          * matches the memory barrier implied by the cmpxchg()
 383                          * in update_read_sb_index().
 384                          */
 385                         return; /* Already writing to this buffer */
 386                 }
 387                 new_id = id;
 388                 subbuffer_id_clear_noref(config, &new_id);
 389                 new_id = cmpxchg(&bufb->buf_wsb[idx].id, id, new_id);
 390                 if (likely(new_id == id))
 391                         break;
 392                 id = new_id;
 393         }
 394 }
 395
 396 /**
 397  * lib_ring_buffer_set_noref_offset - Set the noref subbuffer flag and offset,
 398  *                                    called by writer.
 399  */
 400 static inline
 401 void lib_ring_buffer_set_noref_offset(const struct lttng_kernel_ring_buffer_config *config,
 402                                       struct lttng_kernel_ring_buffer_backend *bufb,
 403                                       unsigned long idx, unsigned long offset)
 404 {
 405         if (config->mode != RING_BUFFER_OVERWRITE)
 406                 return;
 407
 408         /*
 409          * Because ring_buffer_set_noref() is only called by a single thread
 410          * (the one which updated the cc_sb value), there are no concurrent
 411          * updates to take care of: other writers have not updated cc_sb, so
 412          * they cannot set the noref flag, and concurrent readers cannot modify
 413          * the pointer because the noref flag is not set yet.
 414          * The smp_wmb() in ring_buffer_commit() takes care of ordering writes
 415          * to the subbuffer before this set noref operation.
 416          * subbuffer_set_noref() uses a volatile store to deal with concurrent
 417          * readers of the noref flag.
 418          */
 419         CHAN_WARN_ON(bufb->chan,
 420                      subbuffer_id_is_noref(config, bufb->buf_wsb[idx].id));
 421         /*
 422          * Memory barrier that ensures counter stores are ordered before set
 423          * noref and offset.
 424          */
 425         smp_mb();
 426         subbuffer_id_set_noref_offset(config, &bufb->buf_wsb[idx].id, offset);
 427 }
 428
 429 /**
 430  * update_read_sb_index - Read-side subbuffer index update.
 431  */
 432 static inline
 433 int update_read_sb_index(const struct lttng_kernel_ring_buffer_config *config,
 434                          struct lttng_kernel_ring_buffer_backend *bufb,
 435                          struct channel_backend *chanb,
 436                          unsigned long consumed_idx,
 437                          unsigned long consumed_count)
 438 {
 439         unsigned long old_id, new_id;
 440
 441         if (config->mode == RING_BUFFER_OVERWRITE) {
 442                 /*
 443                  * Exchange the target writer subbuffer with our own unused
 444                  * subbuffer. No need to use READ_ONCE() here to read the
 445                  * old_wpage, because the value read will be confirmed by the
 446                  * following cmpxchg().
 447                  */
 448                 old_id = bufb->buf_wsb[consumed_idx].id;
 449                 if (unlikely(!subbuffer_id_is_noref(config, old_id)))
 450                         return -EAGAIN;
 451                 /*
 452                  * Make sure the offset count we are expecting matches the one
 453                  * indicated by the writer.
 454                  */
 455                 if (unlikely(!subbuffer_id_compare_offset(config, old_id,
 456                                                           consumed_count)))
 457                         return -EAGAIN;
 458                 CHAN_WARN_ON(bufb->chan,
 459                              !subbuffer_id_is_noref(config, bufb->buf_rsb.id));
 460                 subbuffer_id_set_noref_offset(config, &bufb->buf_rsb.id,
 461                                               consumed_count);
 462                 new_id = cmpxchg(&bufb->buf_wsb[consumed_idx].id, old_id,
 463                                  bufb->buf_rsb.id);
 464                 if (unlikely(old_id != new_id))
 465                         return -EAGAIN;
 466                 bufb->buf_rsb.id = new_id;
 467         } else {
 468                 /* No page exchange, use the writer page directly */
 469                 bufb->buf_rsb.id = bufb->buf_wsb[consumed_idx].id;
 470         }
 471         return 0;
 472 }
 473
 474 static inline __attribute__((always_inline))
 475 void lttng_inline_memcpy(void *dest, const void *src,
 476                 unsigned long len)
 477 {
 478         switch (len) {
 479         case 1:
 480                 *(uint8_t *) dest = *(const uint8_t *) src;
 481                 break;
 482         case 2:
 483                 *(uint16_t *) dest = *(const uint16_t *) src;
 484                 break;
 485         case 4:
 486                 *(uint32_t *) dest = *(const uint32_t *) src;
 487                 break;
 488         case 8:
 489                 *(uint64_t *) dest = *(const uint64_t *) src;
 490                 break;
 491         default:
 492                 inline_memcpy(dest, src, len);
 493         }
 494 }
 495
 496 /*
 497  * Use the architecture-specific memcpy implementation for constant-sized
 498  * inputs, but rely on an inline memcpy for length statically unknown.
 499  * The function call to memcpy is just way too expensive for a fast path.
 500  */
 501 #define lib_ring_buffer_do_copy(config, dest, src, len)         \
 502 do {                                                            \
 503         size_t __len = (len);                                   \
 504         if (__builtin_constant_p(len))                          \
 505                 memcpy(dest, src, __len);                       \
 506         else                                                    \
 507                 lttng_inline_memcpy(dest, src, __len);          \
 508 } while (0)
 509
 510 /*
 511  * We use __copy_from_user_inatomic to copy userspace data since we already
 512  * did the access_ok for the whole range.
 513  *
 514  * Return 0 if OK, nonzero on error.
 515  */
 516 static inline
 517 unsigned long lib_ring_buffer_do_copy_from_user_inatomic(void *dest,
 518                                                 const void __user *src,
 519                                                 unsigned long len)
 520 {
 521         return __copy_from_user_inatomic(dest, src, len);
 522 }
 523
 524 /*
 525  * write len bytes to dest with c
 526  */
 527 static inline
 528 void lib_ring_buffer_do_memset(char *dest, int c,
 529         unsigned long len)
 530 {
 531         unsigned long i;
 532
 533         for (i = 0; i < len; i++)
 534                 dest[i] = c;
 535 }
 536
 537 #endif /* _LIB_RING_BUFFER_BACKEND_INTERNAL_H */