| 1 | #ifndef _LINUX_RING_BUFFER_BACKEND_INTERNAL_H |
| 2 | #define _LINUX_RING_BUFFER_BACKEND_INTERNAL_H |
| 3 | |
| 4 | /* |
| 5 | * linux/ringbuffer/backend_internal.h |
| 6 | * |
| 7 | * Copyright (C) 2008-2010 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
| 8 | * |
| 9 | * Ring buffer backend (internal helpers). |
| 10 | * |
| 11 | * Dual LGPL v2.1/GPL v2 license. |
| 12 | */ |
| 13 | |
| 14 | #include <unistd.h> |
| 15 | #include <urcu/compiler.h> |
| 16 | |
| 17 | #include "config.h" |
| 18 | #include "backend_types.h" |
| 19 | #include "frontend_types.h" |
| 20 | #include "shm.h" |
| 21 | |
| 22 | /* Ring buffer backend API presented to the frontend */ |
| 23 | |
| 24 | /* Ring buffer and channel backend create/free */ |
| 25 | |
| 26 | int lib_ring_buffer_backend_create(struct lib_ring_buffer_backend *bufb, |
| 27 | struct channel_backend *chan, int cpu, |
| 28 | struct shm_header *shm_header); |
| 29 | void channel_backend_unregister_notifiers(struct channel_backend *chanb); |
| 30 | void lib_ring_buffer_backend_free(struct lib_ring_buffer_backend *bufb); |
| 31 | int channel_backend_init(struct channel_backend *chanb, |
| 32 | const char *name, |
| 33 | const struct lib_ring_buffer_config *config, |
| 34 | void *priv, size_t subbuf_size, |
| 35 | size_t num_subbuf, struct shm_header *shm_header); |
| 36 | void channel_backend_free(struct channel_backend *chanb); |
| 37 | |
| 38 | void lib_ring_buffer_backend_reset(struct lib_ring_buffer_backend *bufb); |
| 39 | void channel_backend_reset(struct channel_backend *chanb); |
| 40 | |
| 41 | int lib_ring_buffer_backend_init(void); |
| 42 | void lib_ring_buffer_backend_exit(void); |
| 43 | |
| 44 | extern void _lib_ring_buffer_write(struct lib_ring_buffer_backend *bufb, |
| 45 | size_t offset, const void *src, size_t len, |
| 46 | ssize_t pagecpy); |
| 47 | |
| 48 | /* |
| 49 | * Subbuffer ID bits for overwrite mode. Need to fit within a single word to be |
| 50 | * exchanged atomically. |
| 51 | * |
| 52 | * Top half word, except lowest bit, belongs to "offset", which is used to keep |
| 53 | * to count the produced buffers. For overwrite mode, this provides the |
| 54 | * consumer with the capacity to read subbuffers in order, handling the |
| 55 | * situation where producers would write up to 2^15 buffers (or 2^31 for 64-bit |
| 56 | * systems) concurrently with a single execution of get_subbuf (between offset |
| 57 | * sampling and subbuffer ID exchange). |
| 58 | */ |
| 59 | |
| 60 | #define HALF_ULONG_BITS (CAA_BITS_PER_LONG >> 1) |
| 61 | |
| 62 | #define SB_ID_OFFSET_SHIFT (HALF_ULONG_BITS + 1) |
| 63 | #define SB_ID_OFFSET_COUNT (1UL << SB_ID_OFFSET_SHIFT) |
| 64 | #define SB_ID_OFFSET_MASK (~(SB_ID_OFFSET_COUNT - 1)) |
| 65 | /* |
| 66 | * Lowest bit of top word half belongs to noref. Used only for overwrite mode. |
| 67 | */ |
| 68 | #define SB_ID_NOREF_SHIFT (SB_ID_OFFSET_SHIFT - 1) |
| 69 | #define SB_ID_NOREF_COUNT (1UL << SB_ID_NOREF_SHIFT) |
| 70 | #define SB_ID_NOREF_MASK SB_ID_NOREF_COUNT |
| 71 | /* |
| 72 | * In overwrite mode: lowest half of word is used for index. |
| 73 | * Limit of 2^16 subbuffers per buffer on 32-bit, 2^32 on 64-bit. |
| 74 | * In producer-consumer mode: whole word used for index. |
| 75 | */ |
| 76 | #define SB_ID_INDEX_SHIFT 0 |
| 77 | #define SB_ID_INDEX_COUNT (1UL << SB_ID_INDEX_SHIFT) |
| 78 | #define SB_ID_INDEX_MASK (SB_ID_NOREF_COUNT - 1) |
| 79 | |
| 80 | /* |
| 81 | * Construct the subbuffer id from offset, index and noref. Use only the index |
| 82 | * for producer-consumer mode (offset and noref are only used in overwrite |
| 83 | * mode). |
| 84 | */ |
| 85 | static inline |
| 86 | unsigned long subbuffer_id(const struct lib_ring_buffer_config *config, |
| 87 | unsigned long offset, unsigned long noref, |
| 88 | unsigned long index) |
| 89 | { |
| 90 | if (config->mode == RING_BUFFER_OVERWRITE) |
| 91 | return (offset << SB_ID_OFFSET_SHIFT) |
| 92 | | (noref << SB_ID_NOREF_SHIFT) |
| 93 | | index; |
| 94 | else |
| 95 | return index; |
| 96 | } |
| 97 | |
| 98 | /* |
| 99 | * Compare offset with the offset contained within id. Return 1 if the offset |
| 100 | * bits are identical, else 0. |
| 101 | */ |
| 102 | static inline |
| 103 | int subbuffer_id_compare_offset(const struct lib_ring_buffer_config *config, |
| 104 | unsigned long id, unsigned long offset) |
| 105 | { |
| 106 | return (id & SB_ID_OFFSET_MASK) == (offset << SB_ID_OFFSET_SHIFT); |
| 107 | } |
| 108 | |
| 109 | static inline |
| 110 | unsigned long subbuffer_id_get_index(const struct lib_ring_buffer_config *config, |
| 111 | unsigned long id) |
| 112 | { |
| 113 | if (config->mode == RING_BUFFER_OVERWRITE) |
| 114 | return id & SB_ID_INDEX_MASK; |
| 115 | else |
| 116 | return id; |
| 117 | } |
| 118 | |
| 119 | static inline |
| 120 | unsigned long subbuffer_id_is_noref(const struct lib_ring_buffer_config *config, |
| 121 | unsigned long id) |
| 122 | { |
| 123 | if (config->mode == RING_BUFFER_OVERWRITE) |
| 124 | return !!(id & SB_ID_NOREF_MASK); |
| 125 | else |
| 126 | return 1; |
| 127 | } |
| 128 | |
| 129 | /* |
| 130 | * Only used by reader on subbuffer ID it has exclusive access to. No volatile |
| 131 | * needed. |
| 132 | */ |
| 133 | static inline |
| 134 | void subbuffer_id_set_noref(const struct lib_ring_buffer_config *config, |
| 135 | unsigned long *id) |
| 136 | { |
| 137 | if (config->mode == RING_BUFFER_OVERWRITE) |
| 138 | *id |= SB_ID_NOREF_MASK; |
| 139 | } |
| 140 | |
| 141 | static inline |
| 142 | void subbuffer_id_set_noref_offset(const struct lib_ring_buffer_config *config, |
| 143 | unsigned long *id, unsigned long offset) |
| 144 | { |
| 145 | unsigned long tmp; |
| 146 | |
| 147 | if (config->mode == RING_BUFFER_OVERWRITE) { |
| 148 | tmp = *id; |
| 149 | tmp &= ~SB_ID_OFFSET_MASK; |
| 150 | tmp |= offset << SB_ID_OFFSET_SHIFT; |
| 151 | tmp |= SB_ID_NOREF_MASK; |
| 152 | /* Volatile store, read concurrently by readers. */ |
| 153 | CMM_ACCESS_ONCE(*id) = tmp; |
| 154 | } |
| 155 | } |
| 156 | |
| 157 | /* No volatile access, since already used locally */ |
| 158 | static inline |
| 159 | void subbuffer_id_clear_noref(const struct lib_ring_buffer_config *config, |
| 160 | unsigned long *id) |
| 161 | { |
| 162 | if (config->mode == RING_BUFFER_OVERWRITE) |
| 163 | *id &= ~SB_ID_NOREF_MASK; |
| 164 | } |
| 165 | |
| 166 | /* |
| 167 | * For overwrite mode, cap the number of subbuffers per buffer to: |
| 168 | * 2^16 on 32-bit architectures |
| 169 | * 2^32 on 64-bit architectures |
| 170 | * This is required to fit in the index part of the ID. Return 0 on success, |
| 171 | * -EPERM on failure. |
| 172 | */ |
| 173 | static inline |
| 174 | int subbuffer_id_check_index(const struct lib_ring_buffer_config *config, |
| 175 | unsigned long num_subbuf) |
| 176 | { |
| 177 | if (config->mode == RING_BUFFER_OVERWRITE) |
| 178 | return (num_subbuf > (1UL << HALF_ULONG_BITS)) ? -EPERM : 0; |
| 179 | else |
| 180 | return 0; |
| 181 | } |
| 182 | |
| 183 | static inline |
| 184 | void subbuffer_count_record(const struct lib_ring_buffer_config *config, |
| 185 | struct lib_ring_buffer_backend *bufb, |
| 186 | unsigned long idx) |
| 187 | { |
| 188 | unsigned long sb_bindex; |
| 189 | |
| 190 | sb_bindex = subbuffer_id_get_index(config, shmp(bufb->buf_wsb)[idx].id); |
| 191 | v_inc(config, &shmp(bufb->array)[sb_bindex]->records_commit); |
| 192 | } |
| 193 | |
| 194 | /* |
| 195 | * Reader has exclusive subbuffer access for record consumption. No need to |
| 196 | * perform the decrement atomically. |
| 197 | */ |
| 198 | static inline |
| 199 | void subbuffer_consume_record(const struct lib_ring_buffer_config *config, |
| 200 | struct lib_ring_buffer_backend *bufb) |
| 201 | { |
| 202 | unsigned long sb_bindex; |
| 203 | |
| 204 | sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); |
| 205 | CHAN_WARN_ON(bufb->chan, |
| 206 | !v_read(config, &shmp(bufb->array)[sb_bindex]->records_unread)); |
| 207 | /* Non-atomic decrement protected by exclusive subbuffer access */ |
| 208 | _v_dec(config, &shmp(bufb->array)[sb_bindex]->records_unread); |
| 209 | v_inc(config, &bufb->records_read); |
| 210 | } |
| 211 | |
| 212 | static inline |
| 213 | unsigned long subbuffer_get_records_count( |
| 214 | const struct lib_ring_buffer_config *config, |
| 215 | struct lib_ring_buffer_backend *bufb, |
| 216 | unsigned long idx) |
| 217 | { |
| 218 | unsigned long sb_bindex; |
| 219 | |
| 220 | sb_bindex = subbuffer_id_get_index(config, shmp(bufb->buf_wsb)[idx].id); |
| 221 | return v_read(config, &shmp(bufb->array)[sb_bindex]->records_commit); |
| 222 | } |
| 223 | |
| 224 | /* |
| 225 | * Must be executed at subbuffer delivery when the writer has _exclusive_ |
| 226 | * subbuffer access. See ring_buffer_check_deliver() for details. |
| 227 | * ring_buffer_get_records_count() must be called to get the records count |
| 228 | * before this function, because it resets the records_commit count. |
| 229 | */ |
| 230 | static inline |
| 231 | unsigned long subbuffer_count_records_overrun( |
| 232 | const struct lib_ring_buffer_config *config, |
| 233 | struct lib_ring_buffer_backend *bufb, |
| 234 | unsigned long idx) |
| 235 | { |
| 236 | struct lib_ring_buffer_backend_pages *pages; |
| 237 | unsigned long overruns, sb_bindex; |
| 238 | |
| 239 | sb_bindex = subbuffer_id_get_index(config, shmp(bufb->buf_wsb)[idx].id); |
| 240 | pages = shmp(bufb->array)[sb_bindex]; |
| 241 | overruns = v_read(config, &pages->records_unread); |
| 242 | v_set(config, &pages->records_unread, |
| 243 | v_read(config, &pages->records_commit)); |
| 244 | v_set(config, &pages->records_commit, 0); |
| 245 | |
| 246 | return overruns; |
| 247 | } |
| 248 | |
| 249 | static inline |
| 250 | void subbuffer_set_data_size(const struct lib_ring_buffer_config *config, |
| 251 | struct lib_ring_buffer_backend *bufb, |
| 252 | unsigned long idx, |
| 253 | unsigned long data_size) |
| 254 | { |
| 255 | struct lib_ring_buffer_backend_pages *pages; |
| 256 | unsigned long sb_bindex; |
| 257 | |
| 258 | sb_bindex = subbuffer_id_get_index(config, shmp(bufb->buf_wsb)[idx].id); |
| 259 | pages = shmp(bufb->array)[sb_bindex]; |
| 260 | pages->data_size = data_size; |
| 261 | } |
| 262 | |
| 263 | static inline |
| 264 | unsigned long subbuffer_get_read_data_size( |
| 265 | const struct lib_ring_buffer_config *config, |
| 266 | struct lib_ring_buffer_backend *bufb) |
| 267 | { |
| 268 | struct lib_ring_buffer_backend_pages *pages; |
| 269 | unsigned long sb_bindex; |
| 270 | |
| 271 | sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); |
| 272 | pages = shmp(bufb->array)[sb_bindex]; |
| 273 | return pages->data_size; |
| 274 | } |
| 275 | |
| 276 | static inline |
| 277 | unsigned long subbuffer_get_data_size( |
| 278 | const struct lib_ring_buffer_config *config, |
| 279 | struct lib_ring_buffer_backend *bufb, |
| 280 | unsigned long idx) |
| 281 | { |
| 282 | struct lib_ring_buffer_backend_pages *pages; |
| 283 | unsigned long sb_bindex; |
| 284 | |
| 285 | sb_bindex = subbuffer_id_get_index(config, shmp(bufb->buf_wsb)[idx].id); |
| 286 | pages = shmp(bufb->array)[sb_bindex]; |
| 287 | return pages->data_size; |
| 288 | } |
| 289 | |
| 290 | /** |
| 291 | * lib_ring_buffer_clear_noref - Clear the noref subbuffer flag, called by |
| 292 | * writer. |
| 293 | */ |
| 294 | static inline |
| 295 | void lib_ring_buffer_clear_noref(const struct lib_ring_buffer_config *config, |
| 296 | struct lib_ring_buffer_backend *bufb, |
| 297 | unsigned long idx) |
| 298 | { |
| 299 | unsigned long id, new_id; |
| 300 | |
| 301 | if (config->mode != RING_BUFFER_OVERWRITE) |
| 302 | return; |
| 303 | |
| 304 | /* |
| 305 | * Performing a volatile access to read the sb_pages, because we want to |
| 306 | * read a coherent version of the pointer and the associated noref flag. |
| 307 | */ |
| 308 | id = CMM_ACCESS_ONCE(shmp(bufb->buf_wsb)[idx].id); |
| 309 | for (;;) { |
| 310 | /* This check is called on the fast path for each record. */ |
| 311 | if (likely(!subbuffer_id_is_noref(config, id))) { |
| 312 | /* |
| 313 | * Store after load dependency ordering the writes to |
| 314 | * the subbuffer after load and test of the noref flag |
| 315 | * matches the memory barrier implied by the cmpxchg() |
| 316 | * in update_read_sb_index(). |
| 317 | */ |
| 318 | return; /* Already writing to this buffer */ |
| 319 | } |
| 320 | new_id = id; |
| 321 | subbuffer_id_clear_noref(config, &new_id); |
| 322 | new_id = uatomic_cmpxchg(&shmp(bufb->buf_wsb)[idx].id, id, new_id); |
| 323 | if (likely(new_id == id)) |
| 324 | break; |
| 325 | id = new_id; |
| 326 | } |
| 327 | } |
| 328 | |
| 329 | /** |
| 330 | * lib_ring_buffer_set_noref_offset - Set the noref subbuffer flag and offset, |
| 331 | * called by writer. |
| 332 | */ |
| 333 | static inline |
| 334 | void lib_ring_buffer_set_noref_offset(const struct lib_ring_buffer_config *config, |
| 335 | struct lib_ring_buffer_backend *bufb, |
| 336 | unsigned long idx, unsigned long offset) |
| 337 | { |
| 338 | if (config->mode != RING_BUFFER_OVERWRITE) |
| 339 | return; |
| 340 | |
| 341 | /* |
| 342 | * Because ring_buffer_set_noref() is only called by a single thread |
| 343 | * (the one which updated the cc_sb value), there are no concurrent |
| 344 | * updates to take care of: other writers have not updated cc_sb, so |
| 345 | * they cannot set the noref flag, and concurrent readers cannot modify |
| 346 | * the pointer because the noref flag is not set yet. |
| 347 | * The smp_wmb() in ring_buffer_commit() takes care of ordering writes |
| 348 | * to the subbuffer before this set noref operation. |
| 349 | * subbuffer_set_noref() uses a volatile store to deal with concurrent |
| 350 | * readers of the noref flag. |
| 351 | */ |
| 352 | CHAN_WARN_ON(bufb->chan, |
| 353 | subbuffer_id_is_noref(config, shmp(bufb->buf_wsb)[idx].id)); |
| 354 | /* |
| 355 | * Memory barrier that ensures counter stores are ordered before set |
| 356 | * noref and offset. |
| 357 | */ |
| 358 | cmm_smp_mb(); |
| 359 | subbuffer_id_set_noref_offset(config, &shmp(bufb->buf_wsb)[idx].id, offset); |
| 360 | } |
| 361 | |
| 362 | /** |
| 363 | * update_read_sb_index - Read-side subbuffer index update. |
| 364 | */ |
| 365 | static inline |
| 366 | int update_read_sb_index(const struct lib_ring_buffer_config *config, |
| 367 | struct lib_ring_buffer_backend *bufb, |
| 368 | struct channel_backend *chanb, |
| 369 | unsigned long consumed_idx, |
| 370 | unsigned long consumed_count) |
| 371 | { |
| 372 | unsigned long old_id, new_id; |
| 373 | |
| 374 | if (config->mode == RING_BUFFER_OVERWRITE) { |
| 375 | /* |
| 376 | * Exchange the target writer subbuffer with our own unused |
| 377 | * subbuffer. No need to use CMM_ACCESS_ONCE() here to read the |
| 378 | * old_wpage, because the value read will be confirmed by the |
| 379 | * following cmpxchg(). |
| 380 | */ |
| 381 | old_id = shmp(bufb->buf_wsb)[consumed_idx].id; |
| 382 | if (unlikely(!subbuffer_id_is_noref(config, old_id))) |
| 383 | return -EAGAIN; |
| 384 | /* |
| 385 | * Make sure the offset count we are expecting matches the one |
| 386 | * indicated by the writer. |
| 387 | */ |
| 388 | if (unlikely(!subbuffer_id_compare_offset(config, old_id, |
| 389 | consumed_count))) |
| 390 | return -EAGAIN; |
| 391 | CHAN_WARN_ON(bufb->chan, |
| 392 | !subbuffer_id_is_noref(config, bufb->buf_rsb.id)); |
| 393 | subbuffer_id_set_noref_offset(config, &bufb->buf_rsb.id, |
| 394 | consumed_count); |
| 395 | new_id = uatomic_cmpxchg(&shmp(bufb->buf_wsb)[consumed_idx].id, old_id, |
| 396 | bufb->buf_rsb.id); |
| 397 | if (unlikely(old_id != new_id)) |
| 398 | return -EAGAIN; |
| 399 | bufb->buf_rsb.id = new_id; |
| 400 | } else { |
| 401 | /* No page exchange, use the writer page directly */ |
| 402 | bufb->buf_rsb.id = shmp(bufb->buf_wsb)[consumed_idx].id; |
| 403 | } |
| 404 | return 0; |
| 405 | } |
| 406 | |
| 407 | /* |
| 408 | * Use the architecture-specific memcpy implementation for constant-sized |
| 409 | * inputs, but rely on an inline memcpy for length statically unknown. |
| 410 | * The function call to memcpy is just way too expensive for a fast path. |
| 411 | */ |
| 412 | #define lib_ring_buffer_do_copy(config, dest, src, len) \ |
| 413 | do { \ |
| 414 | size_t __len = (len); \ |
| 415 | if (__builtin_constant_p(len)) \ |
| 416 | memcpy(dest, src, __len); \ |
| 417 | else \ |
| 418 | inline_memcpy(dest, src, __len); \ |
| 419 | } while (0) |
| 420 | |
| 421 | #endif /* _LINUX_RING_BUFFER_BACKEND_INTERNAL_H */ |