1 // SPDX-License-Identifier: GPL-2.0 2 3 #include <linux/kernel.h> 4 #include <linux/irqflags.h> 5 #include <linux/string.h> 6 #include <linux/errno.h> 7 #include <linux/bug.h> 8 #include "printk_ringbuffer.h" 9 #include "internal.h" 10 11 /** 12 * DOC: printk_ringbuffer overview 13 * 14 * Data Structure 15 * -------------- 16 * The printk_ringbuffer is made up of 3 internal ringbuffers: 17 * 18 * desc_ring 19 * A ring of descriptors and their meta data (such as sequence number, 20 * timestamp, loglevel, etc.) as well as internal state information about 21 * the record and logical positions specifying where in the other 22 * ringbuffer the text strings are located. 23 * 24 * text_data_ring 25 * A ring of data blocks. A data block consists of an unsigned long 26 * integer (ID) that maps to a desc_ring index followed by the text 27 * string of the record. 28 * 29 * The internal state information of a descriptor is the key element to allow 30 * readers and writers to locklessly synchronize access to the data. 31 * 32 * Implementation 33 * -------------- 34 * 35 * Descriptor Ring 36 * ~~~~~~~~~~~~~~~ 37 * The descriptor ring is an array of descriptors. A descriptor contains 38 * essential meta data to track the data of a printk record using 39 * blk_lpos structs pointing to associated text data blocks (see 40 * "Data Rings" below). Each descriptor is assigned an ID that maps 41 * directly to index values of the descriptor array and has a state. The ID 42 * and the state are bitwise combined into a single descriptor field named 43 * @state_var, allowing ID and state to be synchronously and atomically 44 * updated. 45 * 46 * Descriptors have four states: 47 * 48 * reserved 49 * A writer is modifying the record. 50 * 51 * committed 52 * The record and all its data are written. A writer can reopen the 53 * descriptor (transitioning it back to reserved), but in the committed 54 * state the data is consistent. 55 * 56 * finalized 57 * The record and all its data are complete and available for reading. A 58 * writer cannot reopen the descriptor. 59 * 60 * reusable 61 * The record exists, but its text and/or meta data may no longer be 62 * available. 63 * 64 * Querying the @state_var of a record requires providing the ID of the 65 * descriptor to query. This can yield a possible fifth (pseudo) state: 66 * 67 * miss 68 * The descriptor being queried has an unexpected ID. 69 * 70 * The descriptor ring has a @tail_id that contains the ID of the oldest 71 * descriptor and @head_id that contains the ID of the newest descriptor. 72 * 73 * When a new descriptor should be created (and the ring is full), the tail 74 * descriptor is invalidated by first transitioning to the reusable state and 75 * then invalidating all tail data blocks up to and including the data blocks 76 * associated with the tail descriptor (for the text ring). Then 77 * @tail_id is advanced, followed by advancing @head_id. And finally the 78 * @state_var of the new descriptor is initialized to the new ID and reserved 79 * state. 80 * 81 * The @tail_id can only be advanced if the new @tail_id would be in the 82 * committed or reusable queried state. This makes it possible that a valid 83 * sequence number of the tail is always available. 84 * 85 * Descriptor Finalization 86 * ~~~~~~~~~~~~~~~~~~~~~~~ 87 * When a writer calls the commit function prb_commit(), record data is 88 * fully stored and is consistent within the ringbuffer. However, a writer can 89 * reopen that record, claiming exclusive access (as with prb_reserve()), and 90 * modify that record. When finished, the writer must again commit the record. 91 * 92 * In order for a record to be made available to readers (and also become 93 * recyclable for writers), it must be finalized. A finalized record cannot be 94 * reopened and can never become "unfinalized". Record finalization can occur 95 * in three different scenarios: 96 * 97 * 1) A writer can simultaneously commit and finalize its record by calling 98 * prb_final_commit() instead of prb_commit(). 99 * 100 * 2) When a new record is reserved and the previous record has been 101 * committed via prb_commit(), that previous record is automatically 102 * finalized. 103 * 104 * 3) When a record is committed via prb_commit() and a newer record 105 * already exists, the record being committed is automatically finalized. 106 * 107 * Data Ring 108 * ~~~~~~~~~ 109 * The text data ring is a byte array composed of data blocks. Data blocks are 110 * referenced by blk_lpos structs that point to the logical position of the 111 * beginning of a data block and the beginning of the next adjacent data 112 * block. Logical positions are mapped directly to index values of the byte 113 * array ringbuffer. 114 * 115 * Each data block consists of an ID followed by the writer data. The ID is 116 * the identifier of a descriptor that is associated with the data block. A 117 * given data block is considered valid if all of the following conditions 118 * are met: 119 * 120 * 1) The descriptor associated with the data block is in the committed 121 * or finalized queried state. 122 * 123 * 2) The blk_lpos struct within the descriptor associated with the data 124 * block references back to the same data block. 125 * 126 * 3) The data block is within the head/tail logical position range. 127 * 128 * If the writer data of a data block would extend beyond the end of the 129 * byte array, only the ID of the data block is stored at the logical 130 * position and the full data block (ID and writer data) is stored at the 131 * beginning of the byte array. The referencing blk_lpos will point to the 132 * ID before the wrap and the next data block will be at the logical 133 * position adjacent the full data block after the wrap. 134 * 135 * Data rings have a @tail_lpos that points to the beginning of the oldest 136 * data block and a @head_lpos that points to the logical position of the 137 * next (not yet existing) data block. 138 * 139 * When a new data block should be created (and the ring is full), tail data 140 * blocks will first be invalidated by putting their associated descriptors 141 * into the reusable state and then pushing the @tail_lpos forward beyond 142 * them. Then the @head_lpos is pushed forward and is associated with a new 143 * descriptor. If a data block is not valid, the @tail_lpos cannot be 144 * advanced beyond it. 145 * 146 * Info Array 147 * ~~~~~~~~~~ 148 * The general meta data of printk records are stored in printk_info structs, 149 * stored in an array with the same number of elements as the descriptor ring. 150 * Each info corresponds to the descriptor of the same index in the 151 * descriptor ring. Info validity is confirmed by evaluating the corresponding 152 * descriptor before and after loading the info. 153 * 154 * Usage 155 * ----- 156 * Here are some simple examples demonstrating writers and readers. For the 157 * examples a global ringbuffer (test_rb) is available (which is not the 158 * actual ringbuffer used by printk):: 159 * 160 * DEFINE_PRINTKRB(test_rb, 15, 5); 161 * 162 * This ringbuffer allows up to 32768 records (2 ^ 15) and has a size of 163 * 1 MiB (2 ^ (15 + 5)) for text data. 164 * 165 * Sample writer code:: 166 * 167 * const char *textstr = "message text"; 168 * struct prb_reserved_entry e; 169 * struct printk_record r; 170 * 171 * // specify how much to allocate 172 * prb_rec_init_wr(&r, strlen(textstr) + 1); 173 * 174 * if (prb_reserve(&e, &test_rb, &r)) { 175 * snprintf(r.text_buf, r.text_buf_size, "%s", textstr); 176 * 177 * r.info->text_len = strlen(textstr); 178 * r.info->ts_nsec = local_clock(); 179 * r.info->caller_id = printk_caller_id(); 180 * 181 * // commit and finalize the record 182 * prb_final_commit(&e); 183 * } 184 * 185 * Note that additional writer functions are available to extend a record 186 * after it has been committed but not yet finalized. This can be done as 187 * long as no new records have been reserved and the caller is the same. 188 * 189 * Sample writer code (record extending):: 190 * 191 * // alternate rest of previous example 192 * 193 * r.info->text_len = strlen(textstr); 194 * r.info->ts_nsec = local_clock(); 195 * r.info->caller_id = printk_caller_id(); 196 * 197 * // commit the record (but do not finalize yet) 198 * prb_commit(&e); 199 * } 200 * 201 * ... 202 * 203 * // specify additional 5 bytes text space to extend 204 * prb_rec_init_wr(&r, 5); 205 * 206 * // try to extend, but only if it does not exceed 32 bytes 207 * if (prb_reserve_in_last(&e, &test_rb, &r, printk_caller_id(), 32)) { 208 * snprintf(&r.text_buf[r.info->text_len], 209 * r.text_buf_size - r.info->text_len, "hello"); 210 * 211 * r.info->text_len += 5; 212 * 213 * // commit and finalize the record 214 * prb_final_commit(&e); 215 * } 216 * 217 * Sample reader code:: 218 * 219 * struct printk_info info; 220 * struct printk_record r; 221 * char text_buf[32]; 222 * u64 seq; 223 * 224 * prb_rec_init_rd(&r, &info, &text_buf[0], sizeof(text_buf)); 225 * 226 * prb_for_each_record(0, &test_rb, &seq, &r) { 227 * if (info.seq != seq) 228 * pr_warn("lost %llu records\n", info.seq - seq); 229 * 230 * if (info.text_len > r.text_buf_size) { 231 * pr_warn("record %llu text truncated\n", info.seq); 232 * text_buf[r.text_buf_size - 1] = 0; 233 * } 234 * 235 * pr_info("%llu: %llu: %s\n", info.seq, info.ts_nsec, 236 * &text_buf[0]); 237 * } 238 * 239 * Note that additional less convenient reader functions are available to 240 * allow complex record access. 241 * 242 * ABA Issues 243 * ~~~~~~~~~~ 244 * To help avoid ABA issues, descriptors are referenced by IDs (array index 245 * values combined with tagged bits counting array wraps) and data blocks are 246 * referenced by logical positions (array index values combined with tagged 247 * bits counting array wraps). However, on 32-bit systems the number of 248 * tagged bits is relatively small such that an ABA incident is (at least 249 * theoretically) possible. For example, if 4 million maximally sized (1KiB) 250 * printk messages were to occur in NMI context on a 32-bit system, the 251 * interrupted context would not be able to recognize that the 32-bit integer 252 * completely wrapped and thus represents a different data block than the one 253 * the interrupted context expects. 254 * 255 * To help combat this possibility, additional state checking is performed 256 * (such as using cmpxchg() even though set() would suffice). These extra 257 * checks are commented as such and will hopefully catch any ABA issue that 258 * a 32-bit system might experience. 259 * 260 * Memory Barriers 261 * ~~~~~~~~~~~~~~~ 262 * Multiple memory barriers are used. To simplify proving correctness and 263 * generating litmus tests, lines of code related to memory barriers 264 * (loads, stores, and the associated memory barriers) are labeled:: 265 * 266 * LMM(function:letter) 267 * 268 * Comments reference the labels using only the "function:letter" part. 269 * 270 * The memory barrier pairs and their ordering are: 271 * 272 * desc_reserve:D / desc_reserve:B 273 * push descriptor tail (id), then push descriptor head (id) 274 * 275 * desc_reserve:D / data_push_tail:B 276 * push data tail (lpos), then set new descriptor reserved (state) 277 * 278 * desc_reserve:D / desc_push_tail:C 279 * push descriptor tail (id), then set new descriptor reserved (state) 280 * 281 * desc_reserve:D / prb_first_seq:C 282 * push descriptor tail (id), then set new descriptor reserved (state) 283 * 284 * desc_reserve:F / desc_read:D 285 * set new descriptor id and reserved (state), then allow writer changes 286 * 287 * data_alloc:A (or data_realloc:A) / desc_read:D 288 * set old descriptor reusable (state), then modify new data block area 289 * 290 * data_alloc:A (or data_realloc:A) / data_push_tail:B 291 * push data tail (lpos), then modify new data block area 292 * 293 * _prb_commit:B / desc_read:B 294 * store writer changes, then set new descriptor committed (state) 295 * 296 * desc_reopen_last:A / _prb_commit:B 297 * set descriptor reserved (state), then read descriptor data 298 * 299 * _prb_commit:B / desc_reserve:D 300 * set new descriptor committed (state), then check descriptor head (id) 301 * 302 * data_push_tail:D / data_push_tail:A 303 * set descriptor reusable (state), then push data tail (lpos) 304 * 305 * desc_push_tail:B / desc_reserve:D 306 * set descriptor reusable (state), then push descriptor tail (id) 307 * 308 * desc_update_last_finalized:A / desc_last_finalized_seq:A 309 * store finalized record, then set new highest finalized sequence number 310 */ 311 312 #define DATA_SIZE(data_ring) _DATA_SIZE((data_ring)->size_bits) 313 #define DATA_SIZE_MASK(data_ring) (DATA_SIZE(data_ring) - 1) 314 315 #define DESCS_COUNT(desc_ring) _DESCS_COUNT((desc_ring)->count_bits) 316 #define DESCS_COUNT_MASK(desc_ring) (DESCS_COUNT(desc_ring) - 1) 317 318 /* Determine the data array index from a logical position. */ 319 #define DATA_INDEX(data_ring, lpos) ((lpos) & DATA_SIZE_MASK(data_ring)) 320 321 /* Determine the desc array index from an ID or sequence number. */ 322 #define DESC_INDEX(desc_ring, n) ((n) & DESCS_COUNT_MASK(desc_ring)) 323 324 /* Determine how many times the data array has wrapped. */ 325 #define DATA_WRAPS(data_ring, lpos) ((lpos) >> (data_ring)->size_bits) 326 327 /* Determine if a logical position refers to a data-less block. */ 328 #define LPOS_DATALESS(lpos) ((lpos) & 1UL) 329 #define BLK_DATALESS(blk) (LPOS_DATALESS((blk)->begin) && \ 330 LPOS_DATALESS((blk)->next)) 331 332 /* Get the logical position at index 0 of the current wrap. */ 333 #define DATA_THIS_WRAP_START_LPOS(data_ring, lpos) \ 334 ((lpos) & ~DATA_SIZE_MASK(data_ring)) 335 336 /* Get the ID for the same index of the previous wrap as the given ID. */ 337 #define DESC_ID_PREV_WRAP(desc_ring, id) \ 338 DESC_ID((id) - DESCS_COUNT(desc_ring)) 339 340 /* 341 * A data block: mapped directly to the beginning of the data block area 342 * specified as a logical position within the data ring. 343 * 344 * @id: the ID of the associated descriptor 345 * @data: the writer data 346 * 347 * Note that the size of a data block is only known by its associated 348 * descriptor. 349 */ 350 struct prb_data_block { 351 unsigned long id; 352 char data[]; 353 }; 354 355 /* 356 * Return the descriptor associated with @n. @n can be either a 357 * descriptor ID or a sequence number. 358 */ 359 static struct prb_desc *to_desc(struct prb_desc_ring *desc_ring, u64 n) 360 { 361 return &desc_ring->descs[DESC_INDEX(desc_ring, n)]; 362 } 363 364 /* 365 * Return the printk_info associated with @n. @n can be either a 366 * descriptor ID or a sequence number. 367 */ 368 static struct printk_info *to_info(struct prb_desc_ring *desc_ring, u64 n) 369 { 370 return &desc_ring->infos[DESC_INDEX(desc_ring, n)]; 371 } 372 373 static struct prb_data_block *to_block(struct prb_data_ring *data_ring, 374 unsigned long begin_lpos) 375 { 376 return (void *)&data_ring->data[DATA_INDEX(data_ring, begin_lpos)]; 377 } 378 379 /* 380 * Increase the data size to account for data block meta data plus any 381 * padding so that the adjacent data block is aligned on the ID size. 382 */ 383 static unsigned int to_blk_size(unsigned int size) 384 { 385 struct prb_data_block *db = NULL; 386 387 size += sizeof(*db); 388 size = ALIGN(size, sizeof(db->id)); 389 return size; 390 } 391 392 /* 393 * Sanity checker for reserve size. The ringbuffer code assumes that a data 394 * block does not exceed the maximum possible size that could fit within the 395 * ringbuffer. This function provides that basic size check so that the 396 * assumption is safe. 397 */ 398 static bool data_check_size(struct prb_data_ring *data_ring, unsigned int size) 399 { 400 struct prb_data_block *db = NULL; 401 402 if (size == 0) 403 return true; 404 405 /* 406 * Ensure the alignment padded size could possibly fit in the data 407 * array. The largest possible data block must still leave room for 408 * at least the ID of the next block. 409 */ 410 size = to_blk_size(size); 411 if (size > DATA_SIZE(data_ring) - sizeof(db->id)) 412 return false; 413 414 return true; 415 } 416 417 /* Query the state of a descriptor. */ 418 static enum desc_state get_desc_state(unsigned long id, 419 unsigned long state_val) 420 { 421 if (id != DESC_ID(state_val)) 422 return desc_miss; 423 424 return DESC_STATE(state_val); 425 } 426 427 /* 428 * Get a copy of a specified descriptor and return its queried state. If the 429 * descriptor is in an inconsistent state (miss or reserved), the caller can 430 * only expect the descriptor's @state_var field to be valid. 431 * 432 * The sequence number and caller_id can be optionally retrieved. Like all 433 * non-state_var data, they are only valid if the descriptor is in a 434 * consistent state. 435 */ 436 static enum desc_state desc_read(struct prb_desc_ring *desc_ring, 437 unsigned long id, struct prb_desc *desc_out, 438 u64 *seq_out, u32 *caller_id_out) 439 { 440 struct printk_info *info = to_info(desc_ring, id); 441 struct prb_desc *desc = to_desc(desc_ring, id); 442 atomic_long_t *state_var = &desc->state_var; 443 enum desc_state d_state; 444 unsigned long state_val; 445 446 /* Check the descriptor state. */ 447 state_val = atomic_long_read(state_var); /* LMM(desc_read:A) */ 448 d_state = get_desc_state(id, state_val); 449 if (d_state == desc_miss || d_state == desc_reserved) { 450 /* 451 * The descriptor is in an inconsistent state. Set at least 452 * @state_var so that the caller can see the details of 453 * the inconsistent state. 454 */ 455 goto out; 456 } 457 458 /* 459 * Guarantee the state is loaded before copying the descriptor 460 * content. This avoids copying obsolete descriptor content that might 461 * not apply to the descriptor state. This pairs with _prb_commit:B. 462 * 463 * Memory barrier involvement: 464 * 465 * If desc_read:A reads from _prb_commit:B, then desc_read:C reads 466 * from _prb_commit:A. 467 * 468 * Relies on: 469 * 470 * WMB from _prb_commit:A to _prb_commit:B 471 * matching 472 * RMB from desc_read:A to desc_read:C 473 */ 474 smp_rmb(); /* LMM(desc_read:B) */ 475 476 /* 477 * Copy the descriptor data. The data is not valid until the 478 * state has been re-checked. A memcpy() for all of @desc 479 * cannot be used because of the atomic_t @state_var field. 480 */ 481 if (desc_out) { 482 memcpy(&desc_out->text_blk_lpos, &desc->text_blk_lpos, 483 sizeof(desc_out->text_blk_lpos)); /* LMM(desc_read:C) */ 484 } 485 if (seq_out) 486 *seq_out = info->seq; /* also part of desc_read:C */ 487 if (caller_id_out) 488 *caller_id_out = info->caller_id; /* also part of desc_read:C */ 489 490 /* 491 * 1. Guarantee the descriptor content is loaded before re-checking 492 * the state. This avoids reading an obsolete descriptor state 493 * that may not apply to the copied content. This pairs with 494 * desc_reserve:F. 495 * 496 * Memory barrier involvement: 497 * 498 * If desc_read:C reads from desc_reserve:G, then desc_read:E 499 * reads from desc_reserve:F. 500 * 501 * Relies on: 502 * 503 * WMB from desc_reserve:F to desc_reserve:G 504 * matching 505 * RMB from desc_read:C to desc_read:E 506 * 507 * 2. Guarantee the record data is loaded before re-checking the 508 * state. This avoids reading an obsolete descriptor state that may 509 * not apply to the copied data. This pairs with data_alloc:A and 510 * data_realloc:A. 511 * 512 * Memory barrier involvement: 513 * 514 * If copy_data:A reads from data_alloc:B, then desc_read:E 515 * reads from desc_make_reusable:A. 516 * 517 * Relies on: 518 * 519 * MB from desc_make_reusable:A to data_alloc:B 520 * matching 521 * RMB from desc_read:C to desc_read:E 522 * 523 * Note: desc_make_reusable:A and data_alloc:B can be different 524 * CPUs. However, the data_alloc:B CPU (which performs the 525 * full memory barrier) must have previously seen 526 * desc_make_reusable:A. 527 */ 528 smp_rmb(); /* LMM(desc_read:D) */ 529 530 /* 531 * The data has been copied. Return the current descriptor state, 532 * which may have changed since the load above. 533 */ 534 state_val = atomic_long_read(state_var); /* LMM(desc_read:E) */ 535 d_state = get_desc_state(id, state_val); 536 out: 537 if (desc_out) 538 atomic_long_set(&desc_out->state_var, state_val); 539 return d_state; 540 } 541 542 /* 543 * Take a specified descriptor out of the finalized state by attempting 544 * the transition from finalized to reusable. Either this context or some 545 * other context will have been successful. 546 */ 547 static void desc_make_reusable(struct prb_desc_ring *desc_ring, 548 unsigned long id) 549 { 550 unsigned long val_finalized = DESC_SV(id, desc_finalized); 551 unsigned long val_reusable = DESC_SV(id, desc_reusable); 552 struct prb_desc *desc = to_desc(desc_ring, id); 553 atomic_long_t *state_var = &desc->state_var; 554 555 atomic_long_cmpxchg_relaxed(state_var, val_finalized, 556 val_reusable); /* LMM(desc_make_reusable:A) */ 557 } 558 559 /* 560 * Given the text data ring, put the associated descriptor of each 561 * data block from @lpos_begin until @lpos_end into the reusable state. 562 * 563 * If there is any problem making the associated descriptor reusable, either 564 * the descriptor has not yet been finalized or another writer context has 565 * already pushed the tail lpos past the problematic data block. Regardless, 566 * on error the caller can re-load the tail lpos to determine the situation. 567 */ 568 static bool data_make_reusable(struct printk_ringbuffer *rb, 569 unsigned long lpos_begin, 570 unsigned long lpos_end, 571 unsigned long *lpos_out) 572 { 573 574 struct prb_data_ring *data_ring = &rb->text_data_ring; 575 struct prb_desc_ring *desc_ring = &rb->desc_ring; 576 struct prb_data_block *blk; 577 enum desc_state d_state; 578 struct prb_desc desc; 579 struct prb_data_blk_lpos *blk_lpos = &desc.text_blk_lpos; 580 unsigned long id; 581 582 /* Loop until @lpos_begin has advanced to or beyond @lpos_end. */ 583 while ((lpos_end - lpos_begin) - 1 < DATA_SIZE(data_ring)) { 584 blk = to_block(data_ring, lpos_begin); 585 586 /* 587 * Load the block ID from the data block. This is a data race 588 * against a writer that may have newly reserved this data 589 * area. If the loaded value matches a valid descriptor ID, 590 * the blk_lpos of that descriptor will be checked to make 591 * sure it points back to this data block. If the check fails, 592 * the data area has been recycled by another writer. 593 */ 594 id = blk->id; /* LMM(data_make_reusable:A) */ 595 596 d_state = desc_read(desc_ring, id, &desc, 597 NULL, NULL); /* LMM(data_make_reusable:B) */ 598 599 switch (d_state) { 600 case desc_miss: 601 case desc_reserved: 602 case desc_committed: 603 return false; 604 case desc_finalized: 605 /* 606 * This data block is invalid if the descriptor 607 * does not point back to it. 608 */ 609 if (blk_lpos->begin != lpos_begin) 610 return false; 611 desc_make_reusable(desc_ring, id); 612 break; 613 case desc_reusable: 614 /* 615 * This data block is invalid if the descriptor 616 * does not point back to it. 617 */ 618 if (blk_lpos->begin != lpos_begin) 619 return false; 620 break; 621 } 622 623 /* Advance @lpos_begin to the next data block. */ 624 lpos_begin = blk_lpos->next; 625 } 626 627 *lpos_out = lpos_begin; 628 return true; 629 } 630 631 /* 632 * Advance the data ring tail to at least @lpos. This function puts 633 * descriptors into the reusable state if the tail is pushed beyond 634 * their associated data block. 635 */ 636 static bool data_push_tail(struct printk_ringbuffer *rb, unsigned long lpos) 637 { 638 struct prb_data_ring *data_ring = &rb->text_data_ring; 639 unsigned long tail_lpos_new; 640 unsigned long tail_lpos; 641 unsigned long next_lpos; 642 643 /* If @lpos is from a data-less block, there is nothing to do. */ 644 if (LPOS_DATALESS(lpos)) 645 return true; 646 647 /* 648 * Any descriptor states that have transitioned to reusable due to the 649 * data tail being pushed to this loaded value will be visible to this 650 * CPU. This pairs with data_push_tail:D. 651 * 652 * Memory barrier involvement: 653 * 654 * If data_push_tail:A reads from data_push_tail:D, then this CPU can 655 * see desc_make_reusable:A. 656 * 657 * Relies on: 658 * 659 * MB from desc_make_reusable:A to data_push_tail:D 660 * matches 661 * READFROM from data_push_tail:D to data_push_tail:A 662 * thus 663 * READFROM from desc_make_reusable:A to this CPU 664 */ 665 tail_lpos = atomic_long_read(&data_ring->tail_lpos); /* LMM(data_push_tail:A) */ 666 667 /* 668 * Loop until the tail lpos is at or beyond @lpos. This condition 669 * may already be satisfied, resulting in no full memory barrier 670 * from data_push_tail:D being performed. However, since this CPU 671 * sees the new tail lpos, any descriptor states that transitioned to 672 * the reusable state must already be visible. 673 */ 674 while ((lpos - tail_lpos) - 1 < DATA_SIZE(data_ring)) { 675 /* 676 * Make all descriptors reusable that are associated with 677 * data blocks before @lpos. 678 */ 679 if (!data_make_reusable(rb, tail_lpos, lpos, &next_lpos)) { 680 /* 681 * 1. Guarantee the block ID loaded in 682 * data_make_reusable() is performed before 683 * reloading the tail lpos. The failed 684 * data_make_reusable() may be due to a newly 685 * recycled data area causing the tail lpos to 686 * have been previously pushed. This pairs with 687 * data_alloc:A and data_realloc:A. 688 * 689 * Memory barrier involvement: 690 * 691 * If data_make_reusable:A reads from data_alloc:B, 692 * then data_push_tail:C reads from 693 * data_push_tail:D. 694 * 695 * Relies on: 696 * 697 * MB from data_push_tail:D to data_alloc:B 698 * matching 699 * RMB from data_make_reusable:A to 700 * data_push_tail:C 701 * 702 * Note: data_push_tail:D and data_alloc:B can be 703 * different CPUs. However, the data_alloc:B 704 * CPU (which performs the full memory 705 * barrier) must have previously seen 706 * data_push_tail:D. 707 * 708 * 2. Guarantee the descriptor state loaded in 709 * data_make_reusable() is performed before 710 * reloading the tail lpos. The failed 711 * data_make_reusable() may be due to a newly 712 * recycled descriptor causing the tail lpos to 713 * have been previously pushed. This pairs with 714 * desc_reserve:D. 715 * 716 * Memory barrier involvement: 717 * 718 * If data_make_reusable:B reads from 719 * desc_reserve:F, then data_push_tail:C reads 720 * from data_push_tail:D. 721 * 722 * Relies on: 723 * 724 * MB from data_push_tail:D to desc_reserve:F 725 * matching 726 * RMB from data_make_reusable:B to 727 * data_push_tail:C 728 * 729 * Note: data_push_tail:D and desc_reserve:F can 730 * be different CPUs. However, the 731 * desc_reserve:F CPU (which performs the 732 * full memory barrier) must have previously 733 * seen data_push_tail:D. 734 */ 735 smp_rmb(); /* LMM(data_push_tail:B) */ 736 737 tail_lpos_new = atomic_long_read(&data_ring->tail_lpos 738 ); /* LMM(data_push_tail:C) */ 739 if (tail_lpos_new == tail_lpos) 740 return false; 741 742 /* Another CPU pushed the tail. Try again. */ 743 tail_lpos = tail_lpos_new; 744 continue; 745 } 746 747 /* 748 * Guarantee any descriptor states that have transitioned to 749 * reusable are stored before pushing the tail lpos. A full 750 * memory barrier is needed since other CPUs may have made 751 * the descriptor states reusable. This pairs with 752 * data_push_tail:A. 753 */ 754 if (atomic_long_try_cmpxchg(&data_ring->tail_lpos, &tail_lpos, 755 next_lpos)) { /* LMM(data_push_tail:D) */ 756 break; 757 } 758 } 759 760 return true; 761 } 762 763 /* 764 * Advance the desc ring tail. This function advances the tail by one 765 * descriptor, thus invalidating the oldest descriptor. Before advancing 766 * the tail, the tail descriptor is made reusable and all data blocks up to 767 * and including the descriptor's data block are invalidated (i.e. the data 768 * ring tail is pushed past the data block of the descriptor being made 769 * reusable). 770 */ 771 static bool desc_push_tail(struct printk_ringbuffer *rb, 772 unsigned long tail_id) 773 { 774 struct prb_desc_ring *desc_ring = &rb->desc_ring; 775 enum desc_state d_state; 776 struct prb_desc desc; 777 778 d_state = desc_read(desc_ring, tail_id, &desc, NULL, NULL); 779 780 switch (d_state) { 781 case desc_miss: 782 /* 783 * If the ID is exactly 1 wrap behind the expected, it is 784 * in the process of being reserved by another writer and 785 * must be considered reserved. 786 */ 787 if (DESC_ID(atomic_long_read(&desc.state_var)) == 788 DESC_ID_PREV_WRAP(desc_ring, tail_id)) { 789 return false; 790 } 791 792 /* 793 * The ID has changed. Another writer must have pushed the 794 * tail and recycled the descriptor already. Success is 795 * returned because the caller is only interested in the 796 * specified tail being pushed, which it was. 797 */ 798 return true; 799 case desc_reserved: 800 case desc_committed: 801 return false; 802 case desc_finalized: 803 desc_make_reusable(desc_ring, tail_id); 804 break; 805 case desc_reusable: 806 break; 807 } 808 809 /* 810 * Data blocks must be invalidated before their associated 811 * descriptor can be made available for recycling. Invalidating 812 * them later is not possible because there is no way to trust 813 * data blocks once their associated descriptor is gone. 814 */ 815 816 if (!data_push_tail(rb, desc.text_blk_lpos.next)) 817 return false; 818 819 /* 820 * Check the next descriptor after @tail_id before pushing the tail 821 * to it because the tail must always be in a finalized or reusable 822 * state. The implementation of prb_first_seq() relies on this. 823 * 824 * A successful read implies that the next descriptor is less than or 825 * equal to @head_id so there is no risk of pushing the tail past the 826 * head. 827 */ 828 d_state = desc_read(desc_ring, DESC_ID(tail_id + 1), &desc, 829 NULL, NULL); /* LMM(desc_push_tail:A) */ 830 831 if (d_state == desc_finalized || d_state == desc_reusable) { 832 /* 833 * Guarantee any descriptor states that have transitioned to 834 * reusable are stored before pushing the tail ID. This allows 835 * verifying the recycled descriptor state. A full memory 836 * barrier is needed since other CPUs may have made the 837 * descriptor states reusable. This pairs with desc_reserve:D. 838 */ 839 atomic_long_cmpxchg(&desc_ring->tail_id, tail_id, 840 DESC_ID(tail_id + 1)); /* LMM(desc_push_tail:B) */ 841 } else { 842 /* 843 * Guarantee the last state load from desc_read() is before 844 * reloading @tail_id in order to see a new tail ID in the 845 * case that the descriptor has been recycled. This pairs 846 * with desc_reserve:D. 847 * 848 * Memory barrier involvement: 849 * 850 * If desc_push_tail:A reads from desc_reserve:F, then 851 * desc_push_tail:D reads from desc_push_tail:B. 852 * 853 * Relies on: 854 * 855 * MB from desc_push_tail:B to desc_reserve:F 856 * matching 857 * RMB from desc_push_tail:A to desc_push_tail:D 858 * 859 * Note: desc_push_tail:B and desc_reserve:F can be different 860 * CPUs. However, the desc_reserve:F CPU (which performs 861 * the full memory barrier) must have previously seen 862 * desc_push_tail:B. 863 */ 864 smp_rmb(); /* LMM(desc_push_tail:C) */ 865 866 /* 867 * Re-check the tail ID. The descriptor following @tail_id is 868 * not in an allowed tail state. But if the tail has since 869 * been moved by another CPU, then it does not matter. 870 */ 871 if (atomic_long_read(&desc_ring->tail_id) == tail_id) /* LMM(desc_push_tail:D) */ 872 return false; 873 } 874 875 return true; 876 } 877 878 /* Reserve a new descriptor, invalidating the oldest if necessary. */ 879 static bool desc_reserve(struct printk_ringbuffer *rb, unsigned long *id_out) 880 { 881 struct prb_desc_ring *desc_ring = &rb->desc_ring; 882 unsigned long prev_state_val; 883 unsigned long id_prev_wrap; 884 struct prb_desc *desc; 885 unsigned long head_id; 886 unsigned long id; 887 888 head_id = atomic_long_read(&desc_ring->head_id); /* LMM(desc_reserve:A) */ 889 890 do { 891 id = DESC_ID(head_id + 1); 892 id_prev_wrap = DESC_ID_PREV_WRAP(desc_ring, id); 893 894 /* 895 * Guarantee the head ID is read before reading the tail ID. 896 * Since the tail ID is updated before the head ID, this 897 * guarantees that @id_prev_wrap is never ahead of the tail 898 * ID. This pairs with desc_reserve:D. 899 * 900 * Memory barrier involvement: 901 * 902 * If desc_reserve:A reads from desc_reserve:D, then 903 * desc_reserve:C reads from desc_push_tail:B. 904 * 905 * Relies on: 906 * 907 * MB from desc_push_tail:B to desc_reserve:D 908 * matching 909 * RMB from desc_reserve:A to desc_reserve:C 910 * 911 * Note: desc_push_tail:B and desc_reserve:D can be different 912 * CPUs. However, the desc_reserve:D CPU (which performs 913 * the full memory barrier) must have previously seen 914 * desc_push_tail:B. 915 */ 916 smp_rmb(); /* LMM(desc_reserve:B) */ 917 918 if (id_prev_wrap == atomic_long_read(&desc_ring->tail_id 919 )) { /* LMM(desc_reserve:C) */ 920 /* 921 * Make space for the new descriptor by 922 * advancing the tail. 923 */ 924 if (!desc_push_tail(rb, id_prev_wrap)) 925 return false; 926 } 927 928 /* 929 * 1. Guarantee the tail ID is read before validating the 930 * recycled descriptor state. A read memory barrier is 931 * sufficient for this. This pairs with desc_push_tail:B. 932 * 933 * Memory barrier involvement: 934 * 935 * If desc_reserve:C reads from desc_push_tail:B, then 936 * desc_reserve:E reads from desc_make_reusable:A. 937 * 938 * Relies on: 939 * 940 * MB from desc_make_reusable:A to desc_push_tail:B 941 * matching 942 * RMB from desc_reserve:C to desc_reserve:E 943 * 944 * Note: desc_make_reusable:A and desc_push_tail:B can be 945 * different CPUs. However, the desc_push_tail:B CPU 946 * (which performs the full memory barrier) must have 947 * previously seen desc_make_reusable:A. 948 * 949 * 2. Guarantee the tail ID is stored before storing the head 950 * ID. This pairs with desc_reserve:B. 951 * 952 * 3. Guarantee any data ring tail changes are stored before 953 * recycling the descriptor. Data ring tail changes can 954 * happen via desc_push_tail()->data_push_tail(). A full 955 * memory barrier is needed since another CPU may have 956 * pushed the data ring tails. This pairs with 957 * data_push_tail:B. 958 * 959 * 4. Guarantee a new tail ID is stored before recycling the 960 * descriptor. A full memory barrier is needed since 961 * another CPU may have pushed the tail ID. This pairs 962 * with desc_push_tail:C and this also pairs with 963 * prb_first_seq:C. 964 * 965 * 5. Guarantee the head ID is stored before trying to 966 * finalize the previous descriptor. This pairs with 967 * _prb_commit:B. 968 */ 969 } while (!atomic_long_try_cmpxchg(&desc_ring->head_id, &head_id, 970 id)); /* LMM(desc_reserve:D) */ 971 972 desc = to_desc(desc_ring, id); 973 974 /* 975 * If the descriptor has been recycled, verify the old state val. 976 * See "ABA Issues" about why this verification is performed. 977 */ 978 prev_state_val = atomic_long_read(&desc->state_var); /* LMM(desc_reserve:E) */ 979 if (prev_state_val && 980 get_desc_state(id_prev_wrap, prev_state_val) != desc_reusable) { 981 WARN_ON_ONCE(1); 982 return false; 983 } 984 985 /* 986 * Assign the descriptor a new ID and set its state to reserved. 987 * See "ABA Issues" about why cmpxchg() instead of set() is used. 988 * 989 * Guarantee the new descriptor ID and state is stored before making 990 * any other changes. A write memory barrier is sufficient for this. 991 * This pairs with desc_read:D. 992 */ 993 if (!atomic_long_try_cmpxchg(&desc->state_var, &prev_state_val, 994 DESC_SV(id, desc_reserved))) { /* LMM(desc_reserve:F) */ 995 WARN_ON_ONCE(1); 996 return false; 997 } 998 999 /* Now data in @desc can be modified: LMM(desc_reserve:G) */ 1000 1001 *id_out = id; 1002 return true; 1003 } 1004 1005 /* Determine the end of a data block. */ 1006 static unsigned long get_next_lpos(struct prb_data_ring *data_ring, 1007 unsigned long lpos, unsigned int size) 1008 { 1009 unsigned long begin_lpos; 1010 unsigned long next_lpos; 1011 1012 begin_lpos = lpos; 1013 next_lpos = lpos + size; 1014 1015 /* First check if the data block does not wrap. */ 1016 if (DATA_WRAPS(data_ring, begin_lpos) == DATA_WRAPS(data_ring, next_lpos)) 1017 return next_lpos; 1018 1019 /* Wrapping data blocks store their data at the beginning. */ 1020 return (DATA_THIS_WRAP_START_LPOS(data_ring, next_lpos) + size); 1021 } 1022 1023 /* 1024 * Allocate a new data block, invalidating the oldest data block(s) 1025 * if necessary. This function also associates the data block with 1026 * a specified descriptor. 1027 */ 1028 static char *data_alloc(struct printk_ringbuffer *rb, unsigned int size, 1029 struct prb_data_blk_lpos *blk_lpos, unsigned long id) 1030 { 1031 struct prb_data_ring *data_ring = &rb->text_data_ring; 1032 struct prb_data_block *blk; 1033 unsigned long begin_lpos; 1034 unsigned long next_lpos; 1035 1036 if (size == 0) { 1037 /* 1038 * Data blocks are not created for empty lines. Instead, the 1039 * reader will recognize these special lpos values and handle 1040 * it appropriately. 1041 */ 1042 blk_lpos->begin = EMPTY_LINE_LPOS; 1043 blk_lpos->next = EMPTY_LINE_LPOS; 1044 return NULL; 1045 } 1046 1047 size = to_blk_size(size); 1048 1049 begin_lpos = atomic_long_read(&data_ring->head_lpos); 1050 1051 do { 1052 next_lpos = get_next_lpos(data_ring, begin_lpos, size); 1053 1054 if (!data_push_tail(rb, next_lpos - DATA_SIZE(data_ring))) { 1055 /* Failed to allocate, specify a data-less block. */ 1056 blk_lpos->begin = FAILED_LPOS; 1057 blk_lpos->next = FAILED_LPOS; 1058 return NULL; 1059 } 1060 1061 /* 1062 * 1. Guarantee any descriptor states that have transitioned 1063 * to reusable are stored before modifying the newly 1064 * allocated data area. A full memory barrier is needed 1065 * since other CPUs may have made the descriptor states 1066 * reusable. See data_push_tail:A about why the reusable 1067 * states are visible. This pairs with desc_read:D. 1068 * 1069 * 2. Guarantee any updated tail lpos is stored before 1070 * modifying the newly allocated data area. Another CPU may 1071 * be in data_make_reusable() and is reading a block ID 1072 * from this area. data_make_reusable() can handle reading 1073 * a garbage block ID value, but then it must be able to 1074 * load a new tail lpos. A full memory barrier is needed 1075 * since other CPUs may have updated the tail lpos. This 1076 * pairs with data_push_tail:B. 1077 */ 1078 } while (!atomic_long_try_cmpxchg(&data_ring->head_lpos, &begin_lpos, 1079 next_lpos)); /* LMM(data_alloc:A) */ 1080 1081 blk = to_block(data_ring, begin_lpos); 1082 blk->id = id; /* LMM(data_alloc:B) */ 1083 1084 if (DATA_WRAPS(data_ring, begin_lpos) != DATA_WRAPS(data_ring, next_lpos)) { 1085 /* Wrapping data blocks store their data at the beginning. */ 1086 blk = to_block(data_ring, 0); 1087 1088 /* 1089 * Store the ID on the wrapped block for consistency. 1090 * The printk_ringbuffer does not actually use it. 1091 */ 1092 blk->id = id; 1093 } 1094 1095 blk_lpos->begin = begin_lpos; 1096 blk_lpos->next = next_lpos; 1097 1098 return &blk->data[0]; 1099 } 1100 1101 /* 1102 * Try to resize an existing data block associated with the descriptor 1103 * specified by @id. If the resized data block should become wrapped, it 1104 * copies the old data to the new data block. If @size yields a data block 1105 * with the same or less size, the data block is left as is. 1106 * 1107 * Fail if this is not the last allocated data block or if there is not 1108 * enough space or it is not possible make enough space. 1109 * 1110 * Return a pointer to the beginning of the entire data buffer or NULL on 1111 * failure. 1112 */ 1113 static char *data_realloc(struct printk_ringbuffer *rb, unsigned int size, 1114 struct prb_data_blk_lpos *blk_lpos, unsigned long id) 1115 { 1116 struct prb_data_ring *data_ring = &rb->text_data_ring; 1117 struct prb_data_block *blk; 1118 unsigned long head_lpos; 1119 unsigned long next_lpos; 1120 bool wrapped; 1121 1122 /* Reallocation only works if @blk_lpos is the newest data block. */ 1123 head_lpos = atomic_long_read(&data_ring->head_lpos); 1124 if (head_lpos != blk_lpos->next) 1125 return NULL; 1126 1127 /* Keep track if @blk_lpos was a wrapping data block. */ 1128 wrapped = (DATA_WRAPS(data_ring, blk_lpos->begin) != DATA_WRAPS(data_ring, blk_lpos->next)); 1129 1130 size = to_blk_size(size); 1131 1132 next_lpos = get_next_lpos(data_ring, blk_lpos->begin, size); 1133 1134 /* If the data block does not increase, there is nothing to do. */ 1135 if (head_lpos - next_lpos < DATA_SIZE(data_ring)) { 1136 if (wrapped) 1137 blk = to_block(data_ring, 0); 1138 else 1139 blk = to_block(data_ring, blk_lpos->begin); 1140 return &blk->data[0]; 1141 } 1142 1143 if (!data_push_tail(rb, next_lpos - DATA_SIZE(data_ring))) 1144 return NULL; 1145 1146 /* The memory barrier involvement is the same as data_alloc:A. */ 1147 if (!atomic_long_try_cmpxchg(&data_ring->head_lpos, &head_lpos, 1148 next_lpos)) { /* LMM(data_realloc:A) */ 1149 return NULL; 1150 } 1151 1152 blk = to_block(data_ring, blk_lpos->begin); 1153 1154 if (DATA_WRAPS(data_ring, blk_lpos->begin) != DATA_WRAPS(data_ring, next_lpos)) { 1155 struct prb_data_block *old_blk = blk; 1156 1157 /* Wrapping data blocks store their data at the beginning. */ 1158 blk = to_block(data_ring, 0); 1159 1160 /* 1161 * Store the ID on the wrapped block for consistency. 1162 * The printk_ringbuffer does not actually use it. 1163 */ 1164 blk->id = id; 1165 1166 if (!wrapped) { 1167 /* 1168 * Since the allocated space is now in the newly 1169 * created wrapping data block, copy the content 1170 * from the old data block. 1171 */ 1172 memcpy(&blk->data[0], &old_blk->data[0], 1173 (blk_lpos->next - blk_lpos->begin) - sizeof(blk->id)); 1174 } 1175 } 1176 1177 blk_lpos->next = next_lpos; 1178 1179 return &blk->data[0]; 1180 } 1181 1182 /* Return the number of bytes used by a data block. */ 1183 static unsigned int space_used(struct prb_data_ring *data_ring, 1184 struct prb_data_blk_lpos *blk_lpos) 1185 { 1186 /* Data-less blocks take no space. */ 1187 if (BLK_DATALESS(blk_lpos)) 1188 return 0; 1189 1190 if (DATA_WRAPS(data_ring, blk_lpos->begin) == DATA_WRAPS(data_ring, blk_lpos->next)) { 1191 /* Data block does not wrap. */ 1192 return (DATA_INDEX(data_ring, blk_lpos->next) - 1193 DATA_INDEX(data_ring, blk_lpos->begin)); 1194 } 1195 1196 /* 1197 * For wrapping data blocks, the trailing (wasted) space is 1198 * also counted. 1199 */ 1200 return (DATA_INDEX(data_ring, blk_lpos->next) + 1201 DATA_SIZE(data_ring) - DATA_INDEX(data_ring, blk_lpos->begin)); 1202 } 1203 1204 /* 1205 * Given @blk_lpos, return a pointer to the writer data from the data block 1206 * and calculate the size of the data part. A NULL pointer is returned if 1207 * @blk_lpos specifies values that could never be legal. 1208 * 1209 * This function (used by readers) performs strict validation on the lpos 1210 * values to possibly detect bugs in the writer code. A WARN_ON_ONCE() is 1211 * triggered if an internal error is detected. 1212 */ 1213 static const char *get_data(struct prb_data_ring *data_ring, 1214 struct prb_data_blk_lpos *blk_lpos, 1215 unsigned int *data_size) 1216 { 1217 struct prb_data_block *db; 1218 1219 /* Data-less data block description. */ 1220 if (BLK_DATALESS(blk_lpos)) { 1221 /* 1222 * Records that are just empty lines are also valid, even 1223 * though they do not have a data block. For such records 1224 * explicitly return empty string data to signify success. 1225 */ 1226 if (blk_lpos->begin == EMPTY_LINE_LPOS && 1227 blk_lpos->next == EMPTY_LINE_LPOS) { 1228 *data_size = 0; 1229 return ""; 1230 } 1231 1232 /* Data lost, invalid, or otherwise unavailable. */ 1233 return NULL; 1234 } 1235 1236 /* Regular data block: @begin less than @next and in same wrap. */ 1237 if (DATA_WRAPS(data_ring, blk_lpos->begin) == DATA_WRAPS(data_ring, blk_lpos->next) && 1238 blk_lpos->begin < blk_lpos->next) { 1239 db = to_block(data_ring, blk_lpos->begin); 1240 *data_size = blk_lpos->next - blk_lpos->begin; 1241 1242 /* Wrapping data block: @begin is one wrap behind @next. */ 1243 } else if (DATA_WRAPS(data_ring, blk_lpos->begin + DATA_SIZE(data_ring)) == 1244 DATA_WRAPS(data_ring, blk_lpos->next)) { 1245 db = to_block(data_ring, 0); 1246 *data_size = DATA_INDEX(data_ring, blk_lpos->next); 1247 1248 /* Illegal block description. */ 1249 } else { 1250 WARN_ON_ONCE(1); 1251 return NULL; 1252 } 1253 1254 /* A valid data block will always be aligned to the ID size. */ 1255 if (WARN_ON_ONCE(blk_lpos->begin != ALIGN(blk_lpos->begin, sizeof(db->id))) || 1256 WARN_ON_ONCE(blk_lpos->next != ALIGN(blk_lpos->next, sizeof(db->id)))) { 1257 return NULL; 1258 } 1259 1260 /* A valid data block will always have at least an ID. */ 1261 if (WARN_ON_ONCE(*data_size < sizeof(db->id))) 1262 return NULL; 1263 1264 /* Subtract block ID space from size to reflect data size. */ 1265 *data_size -= sizeof(db->id); 1266 1267 return &db->data[0]; 1268 } 1269 1270 /* 1271 * Attempt to transition the newest descriptor from committed back to reserved 1272 * so that the record can be modified by a writer again. This is only possible 1273 * if the descriptor is not yet finalized and the provided @caller_id matches. 1274 */ 1275 static struct prb_desc *desc_reopen_last(struct prb_desc_ring *desc_ring, 1276 u32 caller_id, unsigned long *id_out) 1277 { 1278 unsigned long prev_state_val; 1279 enum desc_state d_state; 1280 struct prb_desc desc; 1281 struct prb_desc *d; 1282 unsigned long id; 1283 u32 cid; 1284 1285 id = atomic_long_read(&desc_ring->head_id); 1286 1287 /* 1288 * To reduce unnecessarily reopening, first check if the descriptor 1289 * state and caller ID are correct. 1290 */ 1291 d_state = desc_read(desc_ring, id, &desc, NULL, &cid); 1292 if (d_state != desc_committed || cid != caller_id) 1293 return NULL; 1294 1295 d = to_desc(desc_ring, id); 1296 1297 prev_state_val = DESC_SV(id, desc_committed); 1298 1299 /* 1300 * Guarantee the reserved state is stored before reading any 1301 * record data. A full memory barrier is needed because @state_var 1302 * modification is followed by reading. This pairs with _prb_commit:B. 1303 * 1304 * Memory barrier involvement: 1305 * 1306 * If desc_reopen_last:A reads from _prb_commit:B, then 1307 * prb_reserve_in_last:A reads from _prb_commit:A. 1308 * 1309 * Relies on: 1310 * 1311 * WMB from _prb_commit:A to _prb_commit:B 1312 * matching 1313 * MB If desc_reopen_last:A to prb_reserve_in_last:A 1314 */ 1315 if (!atomic_long_try_cmpxchg(&d->state_var, &prev_state_val, 1316 DESC_SV(id, desc_reserved))) { /* LMM(desc_reopen_last:A) */ 1317 return NULL; 1318 } 1319 1320 *id_out = id; 1321 return d; 1322 } 1323 1324 /** 1325 * prb_reserve_in_last() - Re-reserve and extend the space in the ringbuffer 1326 * used by the newest record. 1327 * 1328 * @e: The entry structure to setup. 1329 * @rb: The ringbuffer to re-reserve and extend data in. 1330 * @r: The record structure to allocate buffers for. 1331 * @caller_id: The caller ID of the caller (reserving writer). 1332 * @max_size: Fail if the extended size would be greater than this. 1333 * 1334 * This is the public function available to writers to re-reserve and extend 1335 * data. 1336 * 1337 * The writer specifies the text size to extend (not the new total size) by 1338 * setting the @text_buf_size field of @r. To ensure proper initialization 1339 * of @r, prb_rec_init_wr() should be used. 1340 * 1341 * This function will fail if @caller_id does not match the caller ID of the 1342 * newest record. In that case the caller must reserve new data using 1343 * prb_reserve(). 1344 * 1345 * Context: Any context. Disables local interrupts on success. 1346 * Return: true if text data could be extended, otherwise false. 1347 * 1348 * On success: 1349 * 1350 * - @r->text_buf points to the beginning of the entire text buffer. 1351 * 1352 * - @r->text_buf_size is set to the new total size of the buffer. 1353 * 1354 * - @r->info is not touched so that @r->info->text_len could be used 1355 * to append the text. 1356 * 1357 * - prb_record_text_space() can be used on @e to query the new 1358 * actually used space. 1359 * 1360 * Important: All @r->info fields will already be set with the current values 1361 * for the record. I.e. @r->info->text_len will be less than 1362 * @text_buf_size. Writers can use @r->info->text_len to know 1363 * where concatenation begins and writers should update 1364 * @r->info->text_len after concatenating. 1365 */ 1366 bool prb_reserve_in_last(struct prb_reserved_entry *e, struct printk_ringbuffer *rb, 1367 struct printk_record *r, u32 caller_id, unsigned int max_size) 1368 { 1369 struct prb_desc_ring *desc_ring = &rb->desc_ring; 1370 struct printk_info *info; 1371 unsigned int data_size; 1372 struct prb_desc *d; 1373 unsigned long id; 1374 1375 local_irq_save(e->irqflags); 1376 1377 /* Transition the newest descriptor back to the reserved state. */ 1378 d = desc_reopen_last(desc_ring, caller_id, &id); 1379 if (!d) { 1380 local_irq_restore(e->irqflags); 1381 goto fail_reopen; 1382 } 1383 1384 /* Now the writer has exclusive access: LMM(prb_reserve_in_last:A) */ 1385 1386 info = to_info(desc_ring, id); 1387 1388 /* 1389 * Set the @e fields here so that prb_commit() can be used if 1390 * anything fails from now on. 1391 */ 1392 e->rb = rb; 1393 e->id = id; 1394 1395 /* 1396 * desc_reopen_last() checked the caller_id, but there was no 1397 * exclusive access at that point. The descriptor may have 1398 * changed since then. 1399 */ 1400 if (caller_id != info->caller_id) 1401 goto fail; 1402 1403 if (BLK_DATALESS(&d->text_blk_lpos)) { 1404 if (WARN_ON_ONCE(info->text_len != 0)) { 1405 pr_warn_once("wrong text_len value (%hu, expecting 0)\n", 1406 info->text_len); 1407 info->text_len = 0; 1408 } 1409 1410 if (!data_check_size(&rb->text_data_ring, r->text_buf_size)) 1411 goto fail; 1412 1413 if (r->text_buf_size > max_size) 1414 goto fail; 1415 1416 r->text_buf = data_alloc(rb, r->text_buf_size, 1417 &d->text_blk_lpos, id); 1418 } else { 1419 if (!get_data(&rb->text_data_ring, &d->text_blk_lpos, &data_size)) 1420 goto fail; 1421 1422 /* 1423 * Increase the buffer size to include the original size. If 1424 * the meta data (@text_len) is not sane, use the full data 1425 * block size. 1426 */ 1427 if (WARN_ON_ONCE(info->text_len > data_size)) { 1428 pr_warn_once("wrong text_len value (%hu, expecting <=%u)\n", 1429 info->text_len, data_size); 1430 info->text_len = data_size; 1431 } 1432 r->text_buf_size += info->text_len; 1433 1434 if (!data_check_size(&rb->text_data_ring, r->text_buf_size)) 1435 goto fail; 1436 1437 if (r->text_buf_size > max_size) 1438 goto fail; 1439 1440 r->text_buf = data_realloc(rb, r->text_buf_size, 1441 &d->text_blk_lpos, id); 1442 } 1443 if (r->text_buf_size && !r->text_buf) 1444 goto fail; 1445 1446 r->info = info; 1447 1448 e->text_space = space_used(&rb->text_data_ring, &d->text_blk_lpos); 1449 1450 return true; 1451 fail: 1452 prb_commit(e); 1453 /* prb_commit() re-enabled interrupts. */ 1454 fail_reopen: 1455 /* Make it clear to the caller that the re-reserve failed. */ 1456 memset(r, 0, sizeof(*r)); 1457 return false; 1458 } 1459 1460 /* 1461 * @last_finalized_seq value guarantees that all records up to and including 1462 * this sequence number are finalized and can be read. The only exception are 1463 * too old records which have already been overwritten. 1464 * 1465 * It is also guaranteed that @last_finalized_seq only increases. 1466 * 1467 * Be aware that finalized records following non-finalized records are not 1468 * reported because they are not yet available to the reader. For example, 1469 * a new record stored via printk() will not be available to a printer if 1470 * it follows a record that has not been finalized yet. However, once that 1471 * non-finalized record becomes finalized, @last_finalized_seq will be 1472 * appropriately updated and the full set of finalized records will be 1473 * available to the printer. And since each printk() caller will either 1474 * directly print or trigger deferred printing of all available unprinted 1475 * records, all printk() messages will get printed. 1476 */ 1477 static u64 desc_last_finalized_seq(struct printk_ringbuffer *rb) 1478 { 1479 struct prb_desc_ring *desc_ring = &rb->desc_ring; 1480 unsigned long ulseq; 1481 1482 /* 1483 * Guarantee the sequence number is loaded before loading the 1484 * associated record in order to guarantee that the record can be 1485 * seen by this CPU. This pairs with desc_update_last_finalized:A. 1486 */ 1487 ulseq = atomic_long_read_acquire(&desc_ring->last_finalized_seq 1488 ); /* LMM(desc_last_finalized_seq:A) */ 1489 1490 return __ulseq_to_u64seq(rb, ulseq); 1491 } 1492 1493 static bool _prb_read_valid(struct printk_ringbuffer *rb, u64 *seq, 1494 struct printk_record *r, unsigned int *line_count); 1495 1496 /* 1497 * Check if there are records directly following @last_finalized_seq that are 1498 * finalized. If so, update @last_finalized_seq to the latest of these 1499 * records. It is not allowed to skip over records that are not yet finalized. 1500 */ 1501 static void desc_update_last_finalized(struct printk_ringbuffer *rb) 1502 { 1503 struct prb_desc_ring *desc_ring = &rb->desc_ring; 1504 u64 old_seq = desc_last_finalized_seq(rb); 1505 unsigned long oldval; 1506 unsigned long newval; 1507 u64 finalized_seq; 1508 u64 try_seq; 1509 1510 try_again: 1511 finalized_seq = old_seq; 1512 try_seq = finalized_seq + 1; 1513 1514 /* Try to find later finalized records. */ 1515 while (_prb_read_valid(rb, &try_seq, NULL, NULL)) { 1516 finalized_seq = try_seq; 1517 try_seq++; 1518 } 1519 1520 /* No update needed if no later finalized record was found. */ 1521 if (finalized_seq == old_seq) 1522 return; 1523 1524 oldval = __u64seq_to_ulseq(old_seq); 1525 newval = __u64seq_to_ulseq(finalized_seq); 1526 1527 /* 1528 * Set the sequence number of a later finalized record that has been 1529 * seen. 1530 * 1531 * Guarantee the record data is visible to other CPUs before storing 1532 * its sequence number. This pairs with desc_last_finalized_seq:A. 1533 * 1534 * Memory barrier involvement: 1535 * 1536 * If desc_last_finalized_seq:A reads from 1537 * desc_update_last_finalized:A, then desc_read:A reads from 1538 * _prb_commit:B. 1539 * 1540 * Relies on: 1541 * 1542 * RELEASE from _prb_commit:B to desc_update_last_finalized:A 1543 * matching 1544 * ACQUIRE from desc_last_finalized_seq:A to desc_read:A 1545 * 1546 * Note: _prb_commit:B and desc_update_last_finalized:A can be 1547 * different CPUs. However, the desc_update_last_finalized:A 1548 * CPU (which performs the release) must have previously seen 1549 * _prb_commit:B. 1550 */ 1551 if (!atomic_long_try_cmpxchg_release(&desc_ring->last_finalized_seq, 1552 &oldval, newval)) { /* LMM(desc_update_last_finalized:A) */ 1553 old_seq = __ulseq_to_u64seq(rb, oldval); 1554 goto try_again; 1555 } 1556 } 1557 1558 /* 1559 * Attempt to finalize a specified descriptor. If this fails, the descriptor 1560 * is either already final or it will finalize itself when the writer commits. 1561 */ 1562 static void desc_make_final(struct printk_ringbuffer *rb, unsigned long id) 1563 { 1564 struct prb_desc_ring *desc_ring = &rb->desc_ring; 1565 unsigned long prev_state_val = DESC_SV(id, desc_committed); 1566 struct prb_desc *d = to_desc(desc_ring, id); 1567 1568 if (atomic_long_try_cmpxchg_relaxed(&d->state_var, &prev_state_val, 1569 DESC_SV(id, desc_finalized))) { /* LMM(desc_make_final:A) */ 1570 desc_update_last_finalized(rb); 1571 } 1572 } 1573 1574 /** 1575 * prb_reserve() - Reserve space in the ringbuffer. 1576 * 1577 * @e: The entry structure to setup. 1578 * @rb: The ringbuffer to reserve data in. 1579 * @r: The record structure to allocate buffers for. 1580 * 1581 * This is the public function available to writers to reserve data. 1582 * 1583 * The writer specifies the text size to reserve by setting the 1584 * @text_buf_size field of @r. To ensure proper initialization of @r, 1585 * prb_rec_init_wr() should be used. 1586 * 1587 * Context: Any context. Disables local interrupts on success. 1588 * Return: true if at least text data could be allocated, otherwise false. 1589 * 1590 * On success, the fields @info and @text_buf of @r will be set by this 1591 * function and should be filled in by the writer before committing. Also 1592 * on success, prb_record_text_space() can be used on @e to query the actual 1593 * space used for the text data block. 1594 * 1595 * Important: @info->text_len needs to be set correctly by the writer in 1596 * order for data to be readable and/or extended. Its value 1597 * is initialized to 0. 1598 */ 1599 bool prb_reserve(struct prb_reserved_entry *e, struct printk_ringbuffer *rb, 1600 struct printk_record *r) 1601 { 1602 struct prb_desc_ring *desc_ring = &rb->desc_ring; 1603 struct printk_info *info; 1604 struct prb_desc *d; 1605 unsigned long id; 1606 u64 seq; 1607 1608 if (!data_check_size(&rb->text_data_ring, r->text_buf_size)) 1609 goto fail; 1610 1611 /* 1612 * Descriptors in the reserved state act as blockers to all further 1613 * reservations once the desc_ring has fully wrapped. Disable 1614 * interrupts during the reserve/commit window in order to minimize 1615 * the likelihood of this happening. 1616 */ 1617 local_irq_save(e->irqflags); 1618 1619 if (!desc_reserve(rb, &id)) { 1620 /* Descriptor reservation failures are tracked. */ 1621 atomic_long_inc(&rb->fail); 1622 local_irq_restore(e->irqflags); 1623 goto fail; 1624 } 1625 1626 d = to_desc(desc_ring, id); 1627 info = to_info(desc_ring, id); 1628 1629 /* 1630 * All @info fields (except @seq) are cleared and must be filled in 1631 * by the writer. Save @seq before clearing because it is used to 1632 * determine the new sequence number. 1633 */ 1634 seq = info->seq; 1635 memset(info, 0, sizeof(*info)); 1636 1637 /* 1638 * Set the @e fields here so that prb_commit() can be used if 1639 * text data allocation fails. 1640 */ 1641 e->rb = rb; 1642 e->id = id; 1643 1644 /* 1645 * Initialize the sequence number if it has "never been set". 1646 * Otherwise just increment it by a full wrap. 1647 * 1648 * @seq is considered "never been set" if it has a value of 0, 1649 * _except_ for @infos[0], which was specially setup by the ringbuffer 1650 * initializer and therefore is always considered as set. 1651 * 1652 * See the "Bootstrap" comment block in printk_ringbuffer.h for 1653 * details about how the initializer bootstraps the descriptors. 1654 */ 1655 if (seq == 0 && DESC_INDEX(desc_ring, id) != 0) 1656 info->seq = DESC_INDEX(desc_ring, id); 1657 else 1658 info->seq = seq + DESCS_COUNT(desc_ring); 1659 1660 /* 1661 * New data is about to be reserved. Once that happens, previous 1662 * descriptors are no longer able to be extended. Finalize the 1663 * previous descriptor now so that it can be made available to 1664 * readers. (For seq==0 there is no previous descriptor.) 1665 */ 1666 if (info->seq > 0) 1667 desc_make_final(rb, DESC_ID(id - 1)); 1668 1669 r->text_buf = data_alloc(rb, r->text_buf_size, &d->text_blk_lpos, id); 1670 /* If text data allocation fails, a data-less record is committed. */ 1671 if (r->text_buf_size && !r->text_buf) { 1672 prb_commit(e); 1673 /* prb_commit() re-enabled interrupts. */ 1674 goto fail; 1675 } 1676 1677 r->info = info; 1678 1679 /* Record full text space used by record. */ 1680 e->text_space = space_used(&rb->text_data_ring, &d->text_blk_lpos); 1681 1682 return true; 1683 fail: 1684 /* Make it clear to the caller that the reserve failed. */ 1685 memset(r, 0, sizeof(*r)); 1686 return false; 1687 } 1688 1689 /* Commit the data (possibly finalizing it) and restore interrupts. */ 1690 static void _prb_commit(struct prb_reserved_entry *e, unsigned long state_val) 1691 { 1692 struct prb_desc_ring *desc_ring = &e->rb->desc_ring; 1693 struct prb_desc *d = to_desc(desc_ring, e->id); 1694 unsigned long prev_state_val = DESC_SV(e->id, desc_reserved); 1695 1696 /* Now the writer has finished all writing: LMM(_prb_commit:A) */ 1697 1698 /* 1699 * Set the descriptor as committed. See "ABA Issues" about why 1700 * cmpxchg() instead of set() is used. 1701 * 1702 * 1 Guarantee all record data is stored before the descriptor state 1703 * is stored as committed. A write memory barrier is sufficient 1704 * for this. This pairs with desc_read:B and desc_reopen_last:A. 1705 * 1706 * 2. Guarantee the descriptor state is stored as committed before 1707 * re-checking the head ID in order to possibly finalize this 1708 * descriptor. This pairs with desc_reserve:D. 1709 * 1710 * Memory barrier involvement: 1711 * 1712 * If prb_commit:A reads from desc_reserve:D, then 1713 * desc_make_final:A reads from _prb_commit:B. 1714 * 1715 * Relies on: 1716 * 1717 * MB _prb_commit:B to prb_commit:A 1718 * matching 1719 * MB desc_reserve:D to desc_make_final:A 1720 */ 1721 if (!atomic_long_try_cmpxchg(&d->state_var, &prev_state_val, 1722 DESC_SV(e->id, state_val))) { /* LMM(_prb_commit:B) */ 1723 WARN_ON_ONCE(1); 1724 } 1725 1726 /* Restore interrupts, the reserve/commit window is finished. */ 1727 local_irq_restore(e->irqflags); 1728 } 1729 1730 /** 1731 * prb_commit() - Commit (previously reserved) data to the ringbuffer. 1732 * 1733 * @e: The entry containing the reserved data information. 1734 * 1735 * This is the public function available to writers to commit data. 1736 * 1737 * Note that the data is not yet available to readers until it is finalized. 1738 * Finalizing happens automatically when space for the next record is 1739 * reserved. 1740 * 1741 * See prb_final_commit() for a version of this function that finalizes 1742 * immediately. 1743 * 1744 * Context: Any context. Enables local interrupts. 1745 */ 1746 void prb_commit(struct prb_reserved_entry *e) 1747 { 1748 struct prb_desc_ring *desc_ring = &e->rb->desc_ring; 1749 unsigned long head_id; 1750 1751 _prb_commit(e, desc_committed); 1752 1753 /* 1754 * If this descriptor is no longer the head (i.e. a new record has 1755 * been allocated), extending the data for this record is no longer 1756 * allowed and therefore it must be finalized. 1757 */ 1758 head_id = atomic_long_read(&desc_ring->head_id); /* LMM(prb_commit:A) */ 1759 if (head_id != e->id) 1760 desc_make_final(e->rb, e->id); 1761 } 1762 1763 /** 1764 * prb_final_commit() - Commit and finalize (previously reserved) data to 1765 * the ringbuffer. 1766 * 1767 * @e: The entry containing the reserved data information. 1768 * 1769 * This is the public function available to writers to commit+finalize data. 1770 * 1771 * By finalizing, the data is made immediately available to readers. 1772 * 1773 * This function should only be used if there are no intentions of extending 1774 * this data using prb_reserve_in_last(). 1775 * 1776 * Context: Any context. Enables local interrupts. 1777 */ 1778 void prb_final_commit(struct prb_reserved_entry *e) 1779 { 1780 _prb_commit(e, desc_finalized); 1781 1782 desc_update_last_finalized(e->rb); 1783 } 1784 1785 /* 1786 * Count the number of lines in provided text. All text has at least 1 line 1787 * (even if @text_size is 0). Each '\n' processed is counted as an additional 1788 * line. 1789 */ 1790 static unsigned int count_lines(const char *text, unsigned int text_size) 1791 { 1792 unsigned int next_size = text_size; 1793 unsigned int line_count = 1; 1794 const char *next = text; 1795 1796 while (next_size) { 1797 next = memchr(next, '\n', next_size); 1798 if (!next) 1799 break; 1800 line_count++; 1801 next++; 1802 next_size = text_size - (next - text); 1803 } 1804 1805 return line_count; 1806 } 1807 1808 /* 1809 * Given @blk_lpos, copy an expected @len of data into the provided buffer. 1810 * If @line_count is provided, count the number of lines in the data. 1811 * 1812 * This function (used by readers) performs strict validation on the data 1813 * size to possibly detect bugs in the writer code. A WARN_ON_ONCE() is 1814 * triggered if an internal error is detected. 1815 */ 1816 static bool copy_data(struct prb_data_ring *data_ring, 1817 struct prb_data_blk_lpos *blk_lpos, u16 len, char *buf, 1818 unsigned int buf_size, unsigned int *line_count) 1819 { 1820 unsigned int data_size; 1821 const char *data; 1822 1823 /* Caller might not want any data. */ 1824 if ((!buf || !buf_size) && !line_count) 1825 return true; 1826 1827 data = get_data(data_ring, blk_lpos, &data_size); 1828 if (!data) 1829 return false; 1830 1831 /* 1832 * Actual cannot be less than expected. It can be more than expected 1833 * because of the trailing alignment padding. 1834 * 1835 * Note that invalid @len values can occur because the caller loads 1836 * the value during an allowed data race. 1837 */ 1838 if (data_size < (unsigned int)len) 1839 return false; 1840 1841 /* Caller interested in the line count? */ 1842 if (line_count) 1843 *line_count = count_lines(data, len); 1844 1845 /* Caller interested in the data content? */ 1846 if (!buf || !buf_size) 1847 return true; 1848 1849 data_size = min_t(unsigned int, buf_size, len); 1850 1851 memcpy(&buf[0], data, data_size); /* LMM(copy_data:A) */ 1852 return true; 1853 } 1854 1855 /* 1856 * This is an extended version of desc_read(). It gets a copy of a specified 1857 * descriptor. However, it also verifies that the record is finalized and has 1858 * the sequence number @seq. On success, 0 is returned. 1859 * 1860 * Error return values: 1861 * -EINVAL: A finalized record with sequence number @seq does not exist. 1862 * -ENOENT: A finalized record with sequence number @seq exists, but its data 1863 * is not available. This is a valid record, so readers should 1864 * continue with the next record. 1865 */ 1866 static int desc_read_finalized_seq(struct prb_desc_ring *desc_ring, 1867 unsigned long id, u64 seq, 1868 struct prb_desc *desc_out) 1869 { 1870 struct prb_data_blk_lpos *blk_lpos = &desc_out->text_blk_lpos; 1871 enum desc_state d_state; 1872 u64 s; 1873 1874 d_state = desc_read(desc_ring, id, desc_out, &s, NULL); 1875 1876 /* 1877 * An unexpected @id (desc_miss) or @seq mismatch means the record 1878 * does not exist. A descriptor in the reserved or committed state 1879 * means the record does not yet exist for the reader. 1880 */ 1881 if (d_state == desc_miss || 1882 d_state == desc_reserved || 1883 d_state == desc_committed || 1884 s != seq) { 1885 return -EINVAL; 1886 } 1887 1888 /* 1889 * A descriptor in the reusable state may no longer have its data 1890 * available; report it as existing but with lost data. Or the record 1891 * may actually be a record with lost data. 1892 */ 1893 if (d_state == desc_reusable || 1894 (blk_lpos->begin == FAILED_LPOS && blk_lpos->next == FAILED_LPOS)) { 1895 return -ENOENT; 1896 } 1897 1898 return 0; 1899 } 1900 1901 /* 1902 * Copy the ringbuffer data from the record with @seq to the provided 1903 * @r buffer. On success, 0 is returned. 1904 * 1905 * See desc_read_finalized_seq() for error return values. 1906 */ 1907 static int prb_read(struct printk_ringbuffer *rb, u64 seq, 1908 struct printk_record *r, unsigned int *line_count) 1909 { 1910 struct prb_desc_ring *desc_ring = &rb->desc_ring; 1911 struct printk_info *info = to_info(desc_ring, seq); 1912 struct prb_desc *rdesc = to_desc(desc_ring, seq); 1913 atomic_long_t *state_var = &rdesc->state_var; 1914 struct prb_desc desc; 1915 unsigned long id; 1916 int err; 1917 1918 /* Extract the ID, used to specify the descriptor to read. */ 1919 id = DESC_ID(atomic_long_read(state_var)); 1920 1921 /* Get a local copy of the correct descriptor (if available). */ 1922 err = desc_read_finalized_seq(desc_ring, id, seq, &desc); 1923 1924 /* 1925 * If @r is NULL, the caller is only interested in the availability 1926 * of the record. 1927 */ 1928 if (err || !r) 1929 return err; 1930 1931 /* If requested, copy meta data. */ 1932 if (r->info) 1933 memcpy(r->info, info, sizeof(*(r->info))); 1934 1935 /* Copy text data. If it fails, this is a data-less record. */ 1936 if (!copy_data(&rb->text_data_ring, &desc.text_blk_lpos, info->text_len, 1937 r->text_buf, r->text_buf_size, line_count)) { 1938 return -ENOENT; 1939 } 1940 1941 /* Ensure the record is still finalized and has the same @seq. */ 1942 return desc_read_finalized_seq(desc_ring, id, seq, &desc); 1943 } 1944 1945 /* Get the sequence number of the tail descriptor. */ 1946 u64 prb_first_seq(struct printk_ringbuffer *rb) 1947 { 1948 struct prb_desc_ring *desc_ring = &rb->desc_ring; 1949 enum desc_state d_state; 1950 struct prb_desc desc; 1951 unsigned long id; 1952 u64 seq; 1953 1954 for (;;) { 1955 id = atomic_long_read(&rb->desc_ring.tail_id); /* LMM(prb_first_seq:A) */ 1956 1957 d_state = desc_read(desc_ring, id, &desc, &seq, NULL); /* LMM(prb_first_seq:B) */ 1958 1959 /* 1960 * This loop will not be infinite because the tail is 1961 * _always_ in the finalized or reusable state. 1962 */ 1963 if (d_state == desc_finalized || d_state == desc_reusable) 1964 break; 1965 1966 /* 1967 * Guarantee the last state load from desc_read() is before 1968 * reloading @tail_id in order to see a new tail in the case 1969 * that the descriptor has been recycled. This pairs with 1970 * desc_reserve:D. 1971 * 1972 * Memory barrier involvement: 1973 * 1974 * If prb_first_seq:B reads from desc_reserve:F, then 1975 * prb_first_seq:A reads from desc_push_tail:B. 1976 * 1977 * Relies on: 1978 * 1979 * MB from desc_push_tail:B to desc_reserve:F 1980 * matching 1981 * RMB prb_first_seq:B to prb_first_seq:A 1982 */ 1983 smp_rmb(); /* LMM(prb_first_seq:C) */ 1984 } 1985 1986 return seq; 1987 } 1988 1989 /** 1990 * prb_next_reserve_seq() - Get the sequence number after the most recently 1991 * reserved record. 1992 * 1993 * @rb: The ringbuffer to get the sequence number from. 1994 * 1995 * This is the public function available to readers to see what sequence 1996 * number will be assigned to the next reserved record. 1997 * 1998 * Note that depending on the situation, this value can be equal to or 1999 * higher than the sequence number returned by prb_next_seq(). 2000 * 2001 * Context: Any context. 2002 * Return: The sequence number that will be assigned to the next record 2003 * reserved. 2004 */ 2005 u64 prb_next_reserve_seq(struct printk_ringbuffer *rb) 2006 { 2007 struct prb_desc_ring *desc_ring = &rb->desc_ring; 2008 unsigned long last_finalized_id; 2009 atomic_long_t *state_var; 2010 u64 last_finalized_seq; 2011 unsigned long head_id; 2012 struct prb_desc desc; 2013 unsigned long diff; 2014 struct prb_desc *d; 2015 int err; 2016 2017 /* 2018 * It may not be possible to read a sequence number for @head_id. 2019 * So the ID of @last_finailzed_seq is used to calculate what the 2020 * sequence number of @head_id will be. 2021 */ 2022 2023 try_again: 2024 last_finalized_seq = desc_last_finalized_seq(rb); 2025 2026 /* 2027 * @head_id is loaded after @last_finalized_seq to ensure that 2028 * it points to the record with @last_finalized_seq or newer. 2029 * 2030 * Memory barrier involvement: 2031 * 2032 * If desc_last_finalized_seq:A reads from 2033 * desc_update_last_finalized:A, then 2034 * prb_next_reserve_seq:A reads from desc_reserve:D. 2035 * 2036 * Relies on: 2037 * 2038 * RELEASE from desc_reserve:D to desc_update_last_finalized:A 2039 * matching 2040 * ACQUIRE from desc_last_finalized_seq:A to prb_next_reserve_seq:A 2041 * 2042 * Note: desc_reserve:D and desc_update_last_finalized:A can be 2043 * different CPUs. However, the desc_update_last_finalized:A CPU 2044 * (which performs the release) must have previously seen 2045 * desc_read:C, which implies desc_reserve:D can be seen. 2046 */ 2047 head_id = atomic_long_read(&desc_ring->head_id); /* LMM(prb_next_reserve_seq:A) */ 2048 2049 d = to_desc(desc_ring, last_finalized_seq); 2050 state_var = &d->state_var; 2051 2052 /* Extract the ID, used to specify the descriptor to read. */ 2053 last_finalized_id = DESC_ID(atomic_long_read(state_var)); 2054 2055 /* Ensure @last_finalized_id is correct. */ 2056 err = desc_read_finalized_seq(desc_ring, last_finalized_id, last_finalized_seq, &desc); 2057 2058 if (err == -EINVAL) { 2059 if (last_finalized_seq == 0) { 2060 /* 2061 * No record has been finalized or even reserved yet. 2062 * 2063 * The @head_id is initialized such that the first 2064 * increment will yield the first record (seq=0). 2065 * Handle it separately to avoid a negative @diff 2066 * below. 2067 */ 2068 if (head_id == DESC0_ID(desc_ring->count_bits)) 2069 return 0; 2070 2071 /* 2072 * One or more descriptors are already reserved. Use 2073 * the descriptor ID of the first one (@seq=0) for 2074 * the @diff below. 2075 */ 2076 last_finalized_id = DESC0_ID(desc_ring->count_bits) + 1; 2077 } else { 2078 /* Record must have been overwritten. Try again. */ 2079 goto try_again; 2080 } 2081 } 2082 2083 /* Diff of known descriptor IDs to compute related sequence numbers. */ 2084 diff = head_id - last_finalized_id; 2085 2086 /* 2087 * @head_id points to the most recently reserved record, but this 2088 * function returns the sequence number that will be assigned to the 2089 * next (not yet reserved) record. Thus +1 is needed. 2090 */ 2091 return (last_finalized_seq + diff + 1); 2092 } 2093 2094 /* 2095 * Non-blocking read of a record. 2096 * 2097 * On success @seq is updated to the record that was read and (if provided) 2098 * @r and @line_count will contain the read/calculated data. 2099 * 2100 * On failure @seq is updated to a record that is not yet available to the 2101 * reader, but it will be the next record available to the reader. 2102 * 2103 * Note: When the current CPU is in panic, this function will skip over any 2104 * non-existent/non-finalized records in order to allow the panic CPU 2105 * to print any and all records that have been finalized. 2106 */ 2107 static bool _prb_read_valid(struct printk_ringbuffer *rb, u64 *seq, 2108 struct printk_record *r, unsigned int *line_count) 2109 { 2110 u64 tail_seq; 2111 int err; 2112 2113 while ((err = prb_read(rb, *seq, r, line_count))) { 2114 tail_seq = prb_first_seq(rb); 2115 2116 if (*seq < tail_seq) { 2117 /* 2118 * Behind the tail. Catch up and try again. This 2119 * can happen for -ENOENT and -EINVAL cases. 2120 */ 2121 *seq = tail_seq; 2122 2123 } else if (err == -ENOENT) { 2124 /* Record exists, but the data was lost. Skip. */ 2125 (*seq)++; 2126 2127 } else { 2128 /* 2129 * Non-existent/non-finalized record. Must stop. 2130 * 2131 * For panic situations it cannot be expected that 2132 * non-finalized records will become finalized. But 2133 * there may be other finalized records beyond that 2134 * need to be printed for a panic situation. If this 2135 * is the panic CPU, skip this 2136 * non-existent/non-finalized record unless it is 2137 * at or beyond the head, in which case it is not 2138 * possible to continue. 2139 * 2140 * Note that new messages printed on panic CPU are 2141 * finalized when we are here. The only exception 2142 * might be the last message without trailing newline. 2143 * But it would have the sequence number returned 2144 * by "prb_next_reserve_seq() - 1". 2145 */ 2146 if (this_cpu_in_panic() && ((*seq + 1) < prb_next_reserve_seq(rb))) 2147 (*seq)++; 2148 else 2149 return false; 2150 } 2151 } 2152 2153 return true; 2154 } 2155 2156 /** 2157 * prb_read_valid() - Non-blocking read of a requested record or (if gone) 2158 * the next available record. 2159 * 2160 * @rb: The ringbuffer to read from. 2161 * @seq: The sequence number of the record to read. 2162 * @r: A record data buffer to store the read record to. 2163 * 2164 * This is the public function available to readers to read a record. 2165 * 2166 * The reader provides the @info and @text_buf buffers of @r to be 2167 * filled in. Any of the buffer pointers can be set to NULL if the reader 2168 * is not interested in that data. To ensure proper initialization of @r, 2169 * prb_rec_init_rd() should be used. 2170 * 2171 * Context: Any context. 2172 * Return: true if a record was read, otherwise false. 2173 * 2174 * On success, the reader must check r->info.seq to see which record was 2175 * actually read. This allows the reader to detect dropped records. 2176 * 2177 * Failure means @seq refers to a record not yet available to the reader. 2178 */ 2179 bool prb_read_valid(struct printk_ringbuffer *rb, u64 seq, 2180 struct printk_record *r) 2181 { 2182 return _prb_read_valid(rb, &seq, r, NULL); 2183 } 2184 2185 /** 2186 * prb_read_valid_info() - Non-blocking read of meta data for a requested 2187 * record or (if gone) the next available record. 2188 * 2189 * @rb: The ringbuffer to read from. 2190 * @seq: The sequence number of the record to read. 2191 * @info: A buffer to store the read record meta data to. 2192 * @line_count: A buffer to store the number of lines in the record text. 2193 * 2194 * This is the public function available to readers to read only the 2195 * meta data of a record. 2196 * 2197 * The reader provides the @info, @line_count buffers to be filled in. 2198 * Either of the buffer pointers can be set to NULL if the reader is not 2199 * interested in that data. 2200 * 2201 * Context: Any context. 2202 * Return: true if a record's meta data was read, otherwise false. 2203 * 2204 * On success, the reader must check info->seq to see which record meta data 2205 * was actually read. This allows the reader to detect dropped records. 2206 * 2207 * Failure means @seq refers to a record not yet available to the reader. 2208 */ 2209 bool prb_read_valid_info(struct printk_ringbuffer *rb, u64 seq, 2210 struct printk_info *info, unsigned int *line_count) 2211 { 2212 struct printk_record r; 2213 2214 prb_rec_init_rd(&r, info, NULL, 0); 2215 2216 return _prb_read_valid(rb, &seq, &r, line_count); 2217 } 2218 2219 /** 2220 * prb_first_valid_seq() - Get the sequence number of the oldest available 2221 * record. 2222 * 2223 * @rb: The ringbuffer to get the sequence number from. 2224 * 2225 * This is the public function available to readers to see what the 2226 * first/oldest valid sequence number is. 2227 * 2228 * This provides readers a starting point to begin iterating the ringbuffer. 2229 * 2230 * Context: Any context. 2231 * Return: The sequence number of the first/oldest record or, if the 2232 * ringbuffer is empty, 0 is returned. 2233 */ 2234 u64 prb_first_valid_seq(struct printk_ringbuffer *rb) 2235 { 2236 u64 seq = 0; 2237 2238 if (!_prb_read_valid(rb, &seq, NULL, NULL)) 2239 return 0; 2240 2241 return seq; 2242 } 2243 2244 /** 2245 * prb_next_seq() - Get the sequence number after the last available record. 2246 * 2247 * @rb: The ringbuffer to get the sequence number from. 2248 * 2249 * This is the public function available to readers to see what the next 2250 * newest sequence number available to readers will be. 2251 * 2252 * This provides readers a sequence number to jump to if all currently 2253 * available records should be skipped. It is guaranteed that all records 2254 * previous to the returned value have been finalized and are (or were) 2255 * available to the reader. 2256 * 2257 * Context: Any context. 2258 * Return: The sequence number of the next newest (not yet available) record 2259 * for readers. 2260 */ 2261 u64 prb_next_seq(struct printk_ringbuffer *rb) 2262 { 2263 u64 seq; 2264 2265 seq = desc_last_finalized_seq(rb); 2266 2267 /* 2268 * Begin searching after the last finalized record. 2269 * 2270 * On 0, the search must begin at 0 because of hack#2 2271 * of the bootstrapping phase it is not known if a 2272 * record at index 0 exists. 2273 */ 2274 if (seq != 0) 2275 seq++; 2276 2277 /* 2278 * The information about the last finalized @seq might be inaccurate. 2279 * Search forward to find the current one. 2280 */ 2281 while (_prb_read_valid(rb, &seq, NULL, NULL)) 2282 seq++; 2283 2284 return seq; 2285 } 2286 2287 /** 2288 * prb_init() - Initialize a ringbuffer to use provided external buffers. 2289 * 2290 * @rb: The ringbuffer to initialize. 2291 * @text_buf: The data buffer for text data. 2292 * @textbits: The size of @text_buf as a power-of-2 value. 2293 * @descs: The descriptor buffer for ringbuffer records. 2294 * @descbits: The count of @descs items as a power-of-2 value. 2295 * @infos: The printk_info buffer for ringbuffer records. 2296 * 2297 * This is the public function available to writers to setup a ringbuffer 2298 * during runtime using provided buffers. 2299 * 2300 * This must match the initialization of DEFINE_PRINTKRB(). 2301 * 2302 * Context: Any context. 2303 */ 2304 void prb_init(struct printk_ringbuffer *rb, 2305 char *text_buf, unsigned int textbits, 2306 struct prb_desc *descs, unsigned int descbits, 2307 struct printk_info *infos) 2308 { 2309 memset(descs, 0, _DESCS_COUNT(descbits) * sizeof(descs[0])); 2310 memset(infos, 0, _DESCS_COUNT(descbits) * sizeof(infos[0])); 2311 2312 rb->desc_ring.count_bits = descbits; 2313 rb->desc_ring.descs = descs; 2314 rb->desc_ring.infos = infos; 2315 atomic_long_set(&rb->desc_ring.head_id, DESC0_ID(descbits)); 2316 atomic_long_set(&rb->desc_ring.tail_id, DESC0_ID(descbits)); 2317 atomic_long_set(&rb->desc_ring.last_finalized_seq, 0); 2318 2319 rb->text_data_ring.size_bits = textbits; 2320 rb->text_data_ring.data = text_buf; 2321 atomic_long_set(&rb->text_data_ring.head_lpos, BLK0_LPOS(textbits)); 2322 atomic_long_set(&rb->text_data_ring.tail_lpos, BLK0_LPOS(textbits)); 2323 2324 atomic_long_set(&rb->fail, 0); 2325 2326 atomic_long_set(&(descs[_DESCS_COUNT(descbits) - 1].state_var), DESC0_SV(descbits)); 2327 descs[_DESCS_COUNT(descbits) - 1].text_blk_lpos.begin = FAILED_LPOS; 2328 descs[_DESCS_COUNT(descbits) - 1].text_blk_lpos.next = FAILED_LPOS; 2329 2330 infos[0].seq = -(u64)_DESCS_COUNT(descbits); 2331 infos[_DESCS_COUNT(descbits) - 1].seq = 0; 2332 } 2333 2334 /** 2335 * prb_record_text_space() - Query the full actual used ringbuffer space for 2336 * the text data of a reserved entry. 2337 * 2338 * @e: The successfully reserved entry to query. 2339 * 2340 * This is the public function available to writers to see how much actual 2341 * space is used in the ringbuffer to store the text data of the specified 2342 * entry. 2343 * 2344 * This function is only valid if @e has been successfully reserved using 2345 * prb_reserve(). 2346 * 2347 * Context: Any context. 2348 * Return: The size in bytes used by the text data of the associated record. 2349 */ 2350 unsigned int prb_record_text_space(struct prb_reserved_entry *e) 2351 { 2352 return e->text_space; 2353 } 2354