1 /* 2 * Quorum Block filter 3 * 4 * Copyright (C) 2012-2014 Nodalink, EURL. 5 * 6 * Author: 7 * Benoît Canet <benoit.canet@irqsave.net> 8 * 9 * Based on the design and code of blkverify.c (Copyright (C) 2010 IBM, Corp) 10 * and blkmirror.c (Copyright (C) 2011 Red Hat, Inc). 11 * 12 * This work is licensed under the terms of the GNU GPL, version 2 or later. 13 * See the COPYING file in the top-level directory. 14 */ 15 16 #include "qemu/osdep.h" 17 #include "qemu/cutils.h" 18 #include "qemu/module.h" 19 #include "qemu/option.h" 20 #include "block/block_int.h" 21 #include "block/coroutines.h" 22 #include "block/qdict.h" 23 #include "qapi/error.h" 24 #include "qapi/qapi-events-block.h" 25 #include "qapi/qmp/qdict.h" 26 #include "qapi/qmp/qerror.h" 27 #include "qapi/qmp/qlist.h" 28 #include "qapi/qmp/qstring.h" 29 #include "crypto/hash.h" 30 31 #define HASH_LENGTH 32 32 33 #define INDEXSTR_LEN 32 34 35 #define QUORUM_OPT_VOTE_THRESHOLD "vote-threshold" 36 #define QUORUM_OPT_BLKVERIFY "blkverify" 37 #define QUORUM_OPT_REWRITE "rewrite-corrupted" 38 #define QUORUM_OPT_READ_PATTERN "read-pattern" 39 40 /* This union holds a vote hash value */ 41 typedef union QuorumVoteValue { 42 uint8_t h[HASH_LENGTH]; /* SHA-256 hash */ 43 int64_t l; /* simpler 64 bits hash */ 44 } QuorumVoteValue; 45 46 /* A vote item */ 47 typedef struct QuorumVoteItem { 48 int index; 49 QLIST_ENTRY(QuorumVoteItem) next; 50 } QuorumVoteItem; 51 52 /* this structure is a vote version. A version is the set of votes sharing the 53 * same vote value. 54 * The set of votes will be tracked with the items field and its cardinality is 55 * vote_count. 56 */ 57 typedef struct QuorumVoteVersion { 58 QuorumVoteValue value; 59 int index; 60 int vote_count; 61 QLIST_HEAD(, QuorumVoteItem) items; 62 QLIST_ENTRY(QuorumVoteVersion) next; 63 } QuorumVoteVersion; 64 65 /* this structure holds a group of vote versions together */ 66 typedef struct QuorumVotes { 67 QLIST_HEAD(, QuorumVoteVersion) vote_list; 68 bool (*compare)(QuorumVoteValue *a, QuorumVoteValue *b); 69 } QuorumVotes; 70 71 /* the following structure holds the state of one quorum instance */ 72 typedef struct BDRVQuorumState { 73 BdrvChild **children; /* children BlockDriverStates */ 74 int num_children; /* children count */ 75 unsigned next_child_index; /* the index of the next child that should 76 * be added 77 */ 78 int threshold; /* if less than threshold children reads gave the 79 * same result a quorum error occurs. 80 */ 81 bool is_blkverify; /* true if the driver is in blkverify mode 82 * Writes are mirrored on two children devices. 83 * On reads the two children devices' contents are 84 * compared and if a difference is spotted its 85 * location is printed and the code aborts. 86 * It is useful to debug other block drivers by 87 * comparing them with a reference one. 88 */ 89 bool rewrite_corrupted;/* true if the driver must rewrite-on-read corrupted 90 * block if Quorum is reached. 91 */ 92 93 QuorumReadPattern read_pattern; 94 } BDRVQuorumState; 95 96 typedef struct QuorumAIOCB QuorumAIOCB; 97 98 /* Quorum will create one instance of the following structure per operation it 99 * performs on its children. 100 * So for each read/write operation coming from the upper layer there will be 101 * $children_count QuorumChildRequest. 102 */ 103 typedef struct QuorumChildRequest { 104 BlockDriverState *bs; 105 QEMUIOVector qiov; 106 uint8_t *buf; 107 int ret; 108 QuorumAIOCB *parent; 109 } QuorumChildRequest; 110 111 /* Quorum will use the following structure to track progress of each read/write 112 * operation received by the upper layer. 113 * This structure hold pointers to the QuorumChildRequest structures instances 114 * used to do operations on each children and track overall progress. 115 */ 116 struct QuorumAIOCB { 117 BlockDriverState *bs; 118 Coroutine *co; 119 120 /* Request metadata */ 121 uint64_t offset; 122 uint64_t bytes; 123 int flags; 124 125 QEMUIOVector *qiov; /* calling IOV */ 126 127 QuorumChildRequest *qcrs; /* individual child requests */ 128 int count; /* number of completed AIOCB */ 129 int success_count; /* number of successfully completed AIOCB */ 130 131 int rewrite_count; /* number of replica to rewrite: count down to 132 * zero once writes are fired 133 */ 134 135 QuorumVotes votes; 136 137 bool is_read; 138 int vote_ret; 139 int children_read; /* how many children have been read from */ 140 }; 141 142 typedef struct QuorumCo { 143 QuorumAIOCB *acb; 144 int idx; 145 } QuorumCo; 146 147 static void quorum_aio_finalize(QuorumAIOCB *acb) 148 { 149 g_free(acb->qcrs); 150 g_free(acb); 151 } 152 153 static bool quorum_sha256_compare(QuorumVoteValue *a, QuorumVoteValue *b) 154 { 155 return !memcmp(a->h, b->h, HASH_LENGTH); 156 } 157 158 static bool quorum_64bits_compare(QuorumVoteValue *a, QuorumVoteValue *b) 159 { 160 return a->l == b->l; 161 } 162 163 static QuorumAIOCB *quorum_aio_get(BlockDriverState *bs, 164 QEMUIOVector *qiov, 165 uint64_t offset, 166 uint64_t bytes, 167 int flags) 168 { 169 BDRVQuorumState *s = bs->opaque; 170 QuorumAIOCB *acb = g_new(QuorumAIOCB, 1); 171 int i; 172 173 *acb = (QuorumAIOCB) { 174 .co = qemu_coroutine_self(), 175 .bs = bs, 176 .offset = offset, 177 .bytes = bytes, 178 .flags = flags, 179 .qiov = qiov, 180 .votes.compare = quorum_sha256_compare, 181 .votes.vote_list = QLIST_HEAD_INITIALIZER(acb.votes.vote_list), 182 }; 183 184 acb->qcrs = g_new0(QuorumChildRequest, s->num_children); 185 for (i = 0; i < s->num_children; i++) { 186 acb->qcrs[i].buf = NULL; 187 acb->qcrs[i].ret = 0; 188 acb->qcrs[i].parent = acb; 189 } 190 191 return acb; 192 } 193 194 static void quorum_report_bad(QuorumOpType type, uint64_t offset, 195 uint64_t bytes, char *node_name, int ret) 196 { 197 const char *msg = NULL; 198 int64_t start_sector = offset / BDRV_SECTOR_SIZE; 199 int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE); 200 201 if (ret < 0) { 202 msg = strerror(-ret); 203 } 204 205 qapi_event_send_quorum_report_bad(type, !!msg, msg, node_name, start_sector, 206 end_sector - start_sector); 207 } 208 209 static void quorum_report_failure(QuorumAIOCB *acb) 210 { 211 const char *reference = bdrv_get_device_or_node_name(acb->bs); 212 int64_t start_sector = acb->offset / BDRV_SECTOR_SIZE; 213 int64_t end_sector = DIV_ROUND_UP(acb->offset + acb->bytes, 214 BDRV_SECTOR_SIZE); 215 216 qapi_event_send_quorum_failure(reference, start_sector, 217 end_sector - start_sector); 218 } 219 220 static int quorum_vote_error(QuorumAIOCB *acb); 221 222 static bool quorum_has_too_much_io_failed(QuorumAIOCB *acb) 223 { 224 BDRVQuorumState *s = acb->bs->opaque; 225 226 if (acb->success_count < s->threshold) { 227 acb->vote_ret = quorum_vote_error(acb); 228 quorum_report_failure(acb); 229 return true; 230 } 231 232 return false; 233 } 234 235 static int read_fifo_child(QuorumAIOCB *acb); 236 237 static void quorum_copy_qiov(QEMUIOVector *dest, QEMUIOVector *source) 238 { 239 int i; 240 assert(dest->niov == source->niov); 241 assert(dest->size == source->size); 242 for (i = 0; i < source->niov; i++) { 243 assert(dest->iov[i].iov_len == source->iov[i].iov_len); 244 memcpy(dest->iov[i].iov_base, 245 source->iov[i].iov_base, 246 source->iov[i].iov_len); 247 } 248 } 249 250 static void quorum_report_bad_acb(QuorumChildRequest *sacb, int ret) 251 { 252 QuorumAIOCB *acb = sacb->parent; 253 QuorumOpType type = acb->is_read ? QUORUM_OP_TYPE_READ : QUORUM_OP_TYPE_WRITE; 254 quorum_report_bad(type, acb->offset, acb->bytes, sacb->bs->node_name, ret); 255 } 256 257 static void quorum_report_bad_versions(BDRVQuorumState *s, 258 QuorumAIOCB *acb, 259 QuorumVoteValue *value) 260 { 261 QuorumVoteVersion *version; 262 QuorumVoteItem *item; 263 264 QLIST_FOREACH(version, &acb->votes.vote_list, next) { 265 if (acb->votes.compare(&version->value, value)) { 266 continue; 267 } 268 QLIST_FOREACH(item, &version->items, next) { 269 quorum_report_bad(QUORUM_OP_TYPE_READ, acb->offset, acb->bytes, 270 s->children[item->index]->bs->node_name, 0); 271 } 272 } 273 } 274 275 static void quorum_rewrite_entry(void *opaque) 276 { 277 QuorumCo *co = opaque; 278 QuorumAIOCB *acb = co->acb; 279 BDRVQuorumState *s = acb->bs->opaque; 280 281 /* Ignore any errors, it's just a correction attempt for already 282 * corrupted data. 283 * Mask out BDRV_REQ_WRITE_UNCHANGED because this overwrites the 284 * area with different data from the other children. */ 285 bdrv_co_pwritev(s->children[co->idx], acb->offset, acb->bytes, 286 acb->qiov, acb->flags & ~BDRV_REQ_WRITE_UNCHANGED); 287 288 /* Wake up the caller after the last rewrite */ 289 acb->rewrite_count--; 290 if (!acb->rewrite_count) { 291 qemu_coroutine_enter_if_inactive(acb->co); 292 } 293 } 294 295 static bool quorum_rewrite_bad_versions(QuorumAIOCB *acb, 296 QuorumVoteValue *value) 297 { 298 QuorumVoteVersion *version; 299 QuorumVoteItem *item; 300 int count = 0; 301 302 /* first count the number of bad versions: done first to avoid concurrency 303 * issues. 304 */ 305 QLIST_FOREACH(version, &acb->votes.vote_list, next) { 306 if (acb->votes.compare(&version->value, value)) { 307 continue; 308 } 309 QLIST_FOREACH(item, &version->items, next) { 310 count++; 311 } 312 } 313 314 /* quorum_rewrite_entry will count down this to zero */ 315 acb->rewrite_count = count; 316 317 /* now fire the correcting rewrites */ 318 QLIST_FOREACH(version, &acb->votes.vote_list, next) { 319 if (acb->votes.compare(&version->value, value)) { 320 continue; 321 } 322 QLIST_FOREACH(item, &version->items, next) { 323 Coroutine *co; 324 QuorumCo data = { 325 .acb = acb, 326 .idx = item->index, 327 }; 328 329 co = qemu_coroutine_create(quorum_rewrite_entry, &data); 330 qemu_coroutine_enter(co); 331 } 332 } 333 334 /* return true if any rewrite is done else false */ 335 return count; 336 } 337 338 static void quorum_count_vote(QuorumVotes *votes, 339 QuorumVoteValue *value, 340 int index) 341 { 342 QuorumVoteVersion *v = NULL, *version = NULL; 343 QuorumVoteItem *item; 344 345 /* look if we have something with this hash */ 346 QLIST_FOREACH(v, &votes->vote_list, next) { 347 if (votes->compare(&v->value, value)) { 348 version = v; 349 break; 350 } 351 } 352 353 /* It's a version not yet in the list add it */ 354 if (!version) { 355 version = g_new0(QuorumVoteVersion, 1); 356 QLIST_INIT(&version->items); 357 memcpy(&version->value, value, sizeof(version->value)); 358 version->index = index; 359 version->vote_count = 0; 360 QLIST_INSERT_HEAD(&votes->vote_list, version, next); 361 } 362 363 version->vote_count++; 364 365 item = g_new0(QuorumVoteItem, 1); 366 item->index = index; 367 QLIST_INSERT_HEAD(&version->items, item, next); 368 } 369 370 static void quorum_free_vote_list(QuorumVotes *votes) 371 { 372 QuorumVoteVersion *version, *next_version; 373 QuorumVoteItem *item, *next_item; 374 375 QLIST_FOREACH_SAFE(version, &votes->vote_list, next, next_version) { 376 QLIST_REMOVE(version, next); 377 QLIST_FOREACH_SAFE(item, &version->items, next, next_item) { 378 QLIST_REMOVE(item, next); 379 g_free(item); 380 } 381 g_free(version); 382 } 383 } 384 385 static int quorum_compute_hash(QuorumAIOCB *acb, int i, QuorumVoteValue *hash) 386 { 387 QEMUIOVector *qiov = &acb->qcrs[i].qiov; 388 size_t len = sizeof(hash->h); 389 uint8_t *data = hash->h; 390 391 /* XXX - would be nice if we could pass in the Error ** 392 * and propagate that back, but this quorum code is 393 * restricted to just errno values currently */ 394 if (qcrypto_hash_bytesv(QCRYPTO_HASH_ALG_SHA256, 395 qiov->iov, qiov->niov, 396 &data, &len, 397 NULL) < 0) { 398 return -EINVAL; 399 } 400 401 return 0; 402 } 403 404 static QuorumVoteVersion *quorum_get_vote_winner(QuorumVotes *votes) 405 { 406 int max = 0; 407 QuorumVoteVersion *candidate, *winner = NULL; 408 409 QLIST_FOREACH(candidate, &votes->vote_list, next) { 410 if (candidate->vote_count > max) { 411 max = candidate->vote_count; 412 winner = candidate; 413 } 414 } 415 416 return winner; 417 } 418 419 /* qemu_iovec_compare is handy for blkverify mode because it returns the first 420 * differing byte location. Yet it is handcoded to compare vectors one byte 421 * after another so it does not benefit from the libc SIMD optimizations. 422 * quorum_iovec_compare is written for speed and should be used in the non 423 * blkverify mode of quorum. 424 */ 425 static bool quorum_iovec_compare(QEMUIOVector *a, QEMUIOVector *b) 426 { 427 int i; 428 int result; 429 430 assert(a->niov == b->niov); 431 for (i = 0; i < a->niov; i++) { 432 assert(a->iov[i].iov_len == b->iov[i].iov_len); 433 result = memcmp(a->iov[i].iov_base, 434 b->iov[i].iov_base, 435 a->iov[i].iov_len); 436 if (result) { 437 return false; 438 } 439 } 440 441 return true; 442 } 443 444 static bool quorum_compare(QuorumAIOCB *acb, QEMUIOVector *a, QEMUIOVector *b) 445 { 446 BDRVQuorumState *s = acb->bs->opaque; 447 ssize_t offset; 448 449 /* This driver will replace blkverify in this particular case */ 450 if (s->is_blkverify) { 451 offset = qemu_iovec_compare(a, b); 452 if (offset != -1) { 453 fprintf(stderr, "quorum: offset=%" PRIu64 " bytes=%" PRIu64 454 " contents mismatch at offset %" PRIu64 "\n", 455 acb->offset, acb->bytes, acb->offset + offset); 456 exit(1); 457 } 458 return true; 459 } 460 461 return quorum_iovec_compare(a, b); 462 } 463 464 /* Do a vote to get the error code */ 465 static int quorum_vote_error(QuorumAIOCB *acb) 466 { 467 BDRVQuorumState *s = acb->bs->opaque; 468 QuorumVoteVersion *winner = NULL; 469 QuorumVotes error_votes; 470 QuorumVoteValue result_value; 471 int i, ret = 0; 472 bool error = false; 473 474 QLIST_INIT(&error_votes.vote_list); 475 error_votes.compare = quorum_64bits_compare; 476 477 for (i = 0; i < s->num_children; i++) { 478 ret = acb->qcrs[i].ret; 479 if (ret) { 480 error = true; 481 result_value.l = ret; 482 quorum_count_vote(&error_votes, &result_value, i); 483 } 484 } 485 486 if (error) { 487 winner = quorum_get_vote_winner(&error_votes); 488 ret = winner->value.l; 489 } 490 491 quorum_free_vote_list(&error_votes); 492 493 return ret; 494 } 495 496 static void quorum_vote(QuorumAIOCB *acb) 497 { 498 bool quorum = true; 499 int i, j, ret; 500 QuorumVoteValue hash; 501 BDRVQuorumState *s = acb->bs->opaque; 502 QuorumVoteVersion *winner; 503 504 if (quorum_has_too_much_io_failed(acb)) { 505 return; 506 } 507 508 /* get the index of the first successful read */ 509 for (i = 0; i < s->num_children; i++) { 510 if (!acb->qcrs[i].ret) { 511 break; 512 } 513 } 514 515 assert(i < s->num_children); 516 517 /* compare this read with all other successful reads stopping at quorum 518 * failure 519 */ 520 for (j = i + 1; j < s->num_children; j++) { 521 if (acb->qcrs[j].ret) { 522 continue; 523 } 524 quorum = quorum_compare(acb, &acb->qcrs[i].qiov, &acb->qcrs[j].qiov); 525 if (!quorum) { 526 break; 527 } 528 } 529 530 /* Every successful read agrees */ 531 if (quorum) { 532 quorum_copy_qiov(acb->qiov, &acb->qcrs[i].qiov); 533 return; 534 } 535 536 /* compute hashes for each successful read, also store indexes */ 537 for (i = 0; i < s->num_children; i++) { 538 if (acb->qcrs[i].ret) { 539 continue; 540 } 541 ret = quorum_compute_hash(acb, i, &hash); 542 /* if ever the hash computation failed */ 543 if (ret < 0) { 544 acb->vote_ret = ret; 545 goto free_exit; 546 } 547 quorum_count_vote(&acb->votes, &hash, i); 548 } 549 550 /* vote to select the most represented version */ 551 winner = quorum_get_vote_winner(&acb->votes); 552 553 /* if the winner count is smaller than threshold the read fails */ 554 if (winner->vote_count < s->threshold) { 555 quorum_report_failure(acb); 556 acb->vote_ret = -EIO; 557 goto free_exit; 558 } 559 560 /* we have a winner: copy it */ 561 quorum_copy_qiov(acb->qiov, &acb->qcrs[winner->index].qiov); 562 563 /* some versions are bad print them */ 564 quorum_report_bad_versions(s, acb, &winner->value); 565 566 /* corruption correction is enabled */ 567 if (s->rewrite_corrupted) { 568 quorum_rewrite_bad_versions(acb, &winner->value); 569 } 570 571 free_exit: 572 /* free lists */ 573 quorum_free_vote_list(&acb->votes); 574 } 575 576 static void read_quorum_children_entry(void *opaque) 577 { 578 QuorumCo *co = opaque; 579 QuorumAIOCB *acb = co->acb; 580 BDRVQuorumState *s = acb->bs->opaque; 581 int i = co->idx; 582 QuorumChildRequest *sacb = &acb->qcrs[i]; 583 584 sacb->bs = s->children[i]->bs; 585 sacb->ret = bdrv_co_preadv(s->children[i], acb->offset, acb->bytes, 586 &acb->qcrs[i].qiov, 0); 587 588 if (sacb->ret == 0) { 589 acb->success_count++; 590 } else { 591 quorum_report_bad_acb(sacb, sacb->ret); 592 } 593 594 acb->count++; 595 assert(acb->count <= s->num_children); 596 assert(acb->success_count <= s->num_children); 597 598 /* Wake up the caller after the last read */ 599 if (acb->count == s->num_children) { 600 qemu_coroutine_enter_if_inactive(acb->co); 601 } 602 } 603 604 static int read_quorum_children(QuorumAIOCB *acb) 605 { 606 BDRVQuorumState *s = acb->bs->opaque; 607 int i; 608 609 acb->children_read = s->num_children; 610 for (i = 0; i < s->num_children; i++) { 611 acb->qcrs[i].buf = qemu_blockalign(s->children[i]->bs, acb->qiov->size); 612 qemu_iovec_init(&acb->qcrs[i].qiov, acb->qiov->niov); 613 qemu_iovec_clone(&acb->qcrs[i].qiov, acb->qiov, acb->qcrs[i].buf); 614 } 615 616 for (i = 0; i < s->num_children; i++) { 617 Coroutine *co; 618 QuorumCo data = { 619 .acb = acb, 620 .idx = i, 621 }; 622 623 co = qemu_coroutine_create(read_quorum_children_entry, &data); 624 qemu_coroutine_enter(co); 625 } 626 627 while (acb->count < s->num_children) { 628 qemu_coroutine_yield(); 629 } 630 631 /* Do the vote on read */ 632 quorum_vote(acb); 633 for (i = 0; i < s->num_children; i++) { 634 qemu_vfree(acb->qcrs[i].buf); 635 qemu_iovec_destroy(&acb->qcrs[i].qiov); 636 } 637 638 while (acb->rewrite_count) { 639 qemu_coroutine_yield(); 640 } 641 642 return acb->vote_ret; 643 } 644 645 static int read_fifo_child(QuorumAIOCB *acb) 646 { 647 BDRVQuorumState *s = acb->bs->opaque; 648 int n, ret; 649 650 /* We try to read the next child in FIFO order if we failed to read */ 651 do { 652 n = acb->children_read++; 653 acb->qcrs[n].bs = s->children[n]->bs; 654 ret = bdrv_co_preadv(s->children[n], acb->offset, acb->bytes, 655 acb->qiov, 0); 656 if (ret < 0) { 657 quorum_report_bad_acb(&acb->qcrs[n], ret); 658 } 659 } while (ret < 0 && acb->children_read < s->num_children); 660 661 /* FIXME: rewrite failed children if acb->children_read > 1? */ 662 663 return ret; 664 } 665 666 static int quorum_co_preadv(BlockDriverState *bs, uint64_t offset, 667 uint64_t bytes, QEMUIOVector *qiov, int flags) 668 { 669 BDRVQuorumState *s = bs->opaque; 670 QuorumAIOCB *acb = quorum_aio_get(bs, qiov, offset, bytes, flags); 671 int ret; 672 673 acb->is_read = true; 674 acb->children_read = 0; 675 676 if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) { 677 ret = read_quorum_children(acb); 678 } else { 679 ret = read_fifo_child(acb); 680 } 681 quorum_aio_finalize(acb); 682 683 return ret; 684 } 685 686 static void write_quorum_entry(void *opaque) 687 { 688 QuorumCo *co = opaque; 689 QuorumAIOCB *acb = co->acb; 690 BDRVQuorumState *s = acb->bs->opaque; 691 int i = co->idx; 692 QuorumChildRequest *sacb = &acb->qcrs[i]; 693 694 sacb->bs = s->children[i]->bs; 695 if (acb->flags & BDRV_REQ_ZERO_WRITE) { 696 sacb->ret = bdrv_co_pwrite_zeroes(s->children[i], acb->offset, 697 acb->bytes, acb->flags); 698 } else { 699 sacb->ret = bdrv_co_pwritev(s->children[i], acb->offset, acb->bytes, 700 acb->qiov, acb->flags); 701 } 702 if (sacb->ret == 0) { 703 acb->success_count++; 704 } else { 705 quorum_report_bad_acb(sacb, sacb->ret); 706 } 707 acb->count++; 708 assert(acb->count <= s->num_children); 709 assert(acb->success_count <= s->num_children); 710 711 /* Wake up the caller after the last write */ 712 if (acb->count == s->num_children) { 713 qemu_coroutine_enter_if_inactive(acb->co); 714 } 715 } 716 717 static int quorum_co_pwritev(BlockDriverState *bs, uint64_t offset, 718 uint64_t bytes, QEMUIOVector *qiov, int flags) 719 { 720 BDRVQuorumState *s = bs->opaque; 721 QuorumAIOCB *acb = quorum_aio_get(bs, qiov, offset, bytes, flags); 722 int i, ret; 723 724 for (i = 0; i < s->num_children; i++) { 725 Coroutine *co; 726 QuorumCo data = { 727 .acb = acb, 728 .idx = i, 729 }; 730 731 co = qemu_coroutine_create(write_quorum_entry, &data); 732 qemu_coroutine_enter(co); 733 } 734 735 while (acb->count < s->num_children) { 736 qemu_coroutine_yield(); 737 } 738 739 quorum_has_too_much_io_failed(acb); 740 741 ret = acb->vote_ret; 742 quorum_aio_finalize(acb); 743 744 return ret; 745 } 746 747 static int quorum_co_pwrite_zeroes(BlockDriverState *bs, int64_t offset, 748 int bytes, BdrvRequestFlags flags) 749 750 { 751 return quorum_co_pwritev(bs, offset, bytes, NULL, 752 flags | BDRV_REQ_ZERO_WRITE); 753 } 754 755 static int64_t quorum_getlength(BlockDriverState *bs) 756 { 757 BDRVQuorumState *s = bs->opaque; 758 int64_t result; 759 int i; 760 761 /* check that all file have the same length */ 762 result = bdrv_getlength(s->children[0]->bs); 763 if (result < 0) { 764 return result; 765 } 766 for (i = 1; i < s->num_children; i++) { 767 int64_t value = bdrv_getlength(s->children[i]->bs); 768 if (value < 0) { 769 return value; 770 } 771 if (value != result) { 772 return -EIO; 773 } 774 } 775 776 return result; 777 } 778 779 static coroutine_fn int quorum_co_flush(BlockDriverState *bs) 780 { 781 BDRVQuorumState *s = bs->opaque; 782 QuorumVoteVersion *winner = NULL; 783 QuorumVotes error_votes; 784 QuorumVoteValue result_value; 785 int i; 786 int result = 0; 787 int success_count = 0; 788 789 QLIST_INIT(&error_votes.vote_list); 790 error_votes.compare = quorum_64bits_compare; 791 792 for (i = 0; i < s->num_children; i++) { 793 result = bdrv_co_flush(s->children[i]->bs); 794 if (result) { 795 quorum_report_bad(QUORUM_OP_TYPE_FLUSH, 0, 0, 796 s->children[i]->bs->node_name, result); 797 result_value.l = result; 798 quorum_count_vote(&error_votes, &result_value, i); 799 } else { 800 success_count++; 801 } 802 } 803 804 if (success_count >= s->threshold) { 805 result = 0; 806 } else { 807 winner = quorum_get_vote_winner(&error_votes); 808 result = winner->value.l; 809 } 810 quorum_free_vote_list(&error_votes); 811 812 return result; 813 } 814 815 static bool quorum_recurse_can_replace(BlockDriverState *bs, 816 BlockDriverState *to_replace) 817 { 818 BDRVQuorumState *s = bs->opaque; 819 int i; 820 821 for (i = 0; i < s->num_children; i++) { 822 /* 823 * We have no idea whether our children show the same data as 824 * this node (@bs). It is actually highly likely that 825 * @to_replace does not, because replacing a broken child is 826 * one of the main use cases here. 827 * 828 * We do know that the new BDS will match @bs, so replacing 829 * any of our children by it will be safe. It cannot change 830 * the data this quorum node presents to its parents. 831 * 832 * However, replacing @to_replace by @bs in any of our 833 * children's chains may change visible data somewhere in 834 * there. We therefore cannot recurse down those chains with 835 * bdrv_recurse_can_replace(). 836 * (More formally, bdrv_recurse_can_replace() requires that 837 * @to_replace will be replaced by something matching the @bs 838 * passed to it. We cannot guarantee that.) 839 * 840 * Thus, we can only check whether any of our immediate 841 * children matches @to_replace. 842 * 843 * (In the future, we might add a function to recurse down a 844 * chain that checks that nothing there cares about a change 845 * in data from the respective child in question. For 846 * example, most filters do not care when their child's data 847 * suddenly changes, as long as their parents do not care.) 848 */ 849 if (s->children[i]->bs == to_replace) { 850 /* 851 * We now have to ensure that there is no other parent 852 * that cares about replacing this child by a node with 853 * potentially different data. 854 * We do so by checking whether there are any other parents 855 * at all, which is stricter than necessary, but also very 856 * simple. (We may decide to implement something more 857 * complex and permissive when there is an actual need for 858 * it.) 859 */ 860 return QLIST_FIRST(&to_replace->parents) == s->children[i] && 861 QLIST_NEXT(s->children[i], next_parent) == NULL; 862 } 863 } 864 865 return false; 866 } 867 868 static int quorum_valid_threshold(int threshold, int num_children, Error **errp) 869 { 870 871 if (threshold < 1) { 872 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, 873 "vote-threshold", "a value >= 1"); 874 return -ERANGE; 875 } 876 877 if (threshold > num_children) { 878 error_setg(errp, "threshold may not exceed children count"); 879 return -ERANGE; 880 } 881 882 return 0; 883 } 884 885 static QemuOptsList quorum_runtime_opts = { 886 .name = "quorum", 887 .head = QTAILQ_HEAD_INITIALIZER(quorum_runtime_opts.head), 888 .desc = { 889 { 890 .name = QUORUM_OPT_VOTE_THRESHOLD, 891 .type = QEMU_OPT_NUMBER, 892 .help = "The number of vote needed for reaching quorum", 893 }, 894 { 895 .name = QUORUM_OPT_BLKVERIFY, 896 .type = QEMU_OPT_BOOL, 897 .help = "Trigger block verify mode if set", 898 }, 899 { 900 .name = QUORUM_OPT_REWRITE, 901 .type = QEMU_OPT_BOOL, 902 .help = "Rewrite corrupted block on read quorum", 903 }, 904 { 905 .name = QUORUM_OPT_READ_PATTERN, 906 .type = QEMU_OPT_STRING, 907 .help = "Allowed pattern: quorum, fifo. Quorum is default", 908 }, 909 { /* end of list */ } 910 }, 911 }; 912 913 static void quorum_refresh_flags(BlockDriverState *bs) 914 { 915 BDRVQuorumState *s = bs->opaque; 916 int i; 917 918 bs->supported_zero_flags = 919 BDRV_REQ_FUA | BDRV_REQ_MAY_UNMAP | BDRV_REQ_NO_FALLBACK; 920 921 for (i = 0; i < s->num_children; i++) { 922 bs->supported_zero_flags &= s->children[i]->bs->supported_zero_flags; 923 } 924 925 bs->supported_zero_flags |= BDRV_REQ_WRITE_UNCHANGED; 926 } 927 928 static int quorum_open(BlockDriverState *bs, QDict *options, int flags, 929 Error **errp) 930 { 931 BDRVQuorumState *s = bs->opaque; 932 Error *local_err = NULL; 933 QemuOpts *opts = NULL; 934 const char *pattern_str; 935 bool *opened; 936 int i; 937 int ret = 0; 938 939 qdict_flatten(options); 940 941 /* count how many different children are present */ 942 s->num_children = qdict_array_entries(options, "children."); 943 if (s->num_children < 0) { 944 error_setg(errp, "Option children is not a valid array"); 945 ret = -EINVAL; 946 goto exit; 947 } 948 if (s->num_children < 1) { 949 error_setg(errp, "Number of provided children must be 1 or more"); 950 ret = -EINVAL; 951 goto exit; 952 } 953 954 opts = qemu_opts_create(&quorum_runtime_opts, NULL, 0, &error_abort); 955 if (!qemu_opts_absorb_qdict(opts, options, errp)) { 956 ret = -EINVAL; 957 goto exit; 958 } 959 960 s->threshold = qemu_opt_get_number(opts, QUORUM_OPT_VOTE_THRESHOLD, 0); 961 /* and validate it against s->num_children */ 962 ret = quorum_valid_threshold(s->threshold, s->num_children, errp); 963 if (ret < 0) { 964 goto exit; 965 } 966 967 pattern_str = qemu_opt_get(opts, QUORUM_OPT_READ_PATTERN); 968 if (!pattern_str) { 969 ret = QUORUM_READ_PATTERN_QUORUM; 970 } else { 971 ret = qapi_enum_parse(&QuorumReadPattern_lookup, pattern_str, 972 -EINVAL, NULL); 973 } 974 if (ret < 0) { 975 error_setg(errp, "Please set read-pattern as fifo or quorum"); 976 goto exit; 977 } 978 s->read_pattern = ret; 979 980 if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) { 981 s->is_blkverify = qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false); 982 if (s->is_blkverify && (s->num_children != 2 || s->threshold != 2)) { 983 error_setg(errp, "blkverify=on can only be set if there are " 984 "exactly two files and vote-threshold is 2"); 985 ret = -EINVAL; 986 goto exit; 987 } 988 989 s->rewrite_corrupted = qemu_opt_get_bool(opts, QUORUM_OPT_REWRITE, 990 false); 991 if (s->rewrite_corrupted && s->is_blkverify) { 992 error_setg(errp, 993 "rewrite-corrupted=on cannot be used with blkverify=on"); 994 ret = -EINVAL; 995 goto exit; 996 } 997 } 998 999 /* allocate the children array */ 1000 s->children = g_new0(BdrvChild *, s->num_children); 1001 opened = g_new0(bool, s->num_children); 1002 1003 for (i = 0; i < s->num_children; i++) { 1004 char indexstr[INDEXSTR_LEN]; 1005 ret = snprintf(indexstr, INDEXSTR_LEN, "children.%d", i); 1006 assert(ret < INDEXSTR_LEN); 1007 1008 s->children[i] = bdrv_open_child(NULL, options, indexstr, bs, 1009 &child_of_bds, BDRV_CHILD_DATA, false, 1010 &local_err); 1011 if (local_err) { 1012 error_propagate(errp, local_err); 1013 ret = -EINVAL; 1014 goto close_exit; 1015 } 1016 1017 opened[i] = true; 1018 } 1019 s->next_child_index = s->num_children; 1020 1021 bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED; 1022 quorum_refresh_flags(bs); 1023 1024 g_free(opened); 1025 goto exit; 1026 1027 close_exit: 1028 /* cleanup on error */ 1029 for (i = 0; i < s->num_children; i++) { 1030 if (!opened[i]) { 1031 continue; 1032 } 1033 bdrv_unref_child(bs, s->children[i]); 1034 } 1035 g_free(s->children); 1036 g_free(opened); 1037 exit: 1038 qemu_opts_del(opts); 1039 return ret; 1040 } 1041 1042 static void quorum_close(BlockDriverState *bs) 1043 { 1044 BDRVQuorumState *s = bs->opaque; 1045 int i; 1046 1047 for (i = 0; i < s->num_children; i++) { 1048 bdrv_unref_child(bs, s->children[i]); 1049 } 1050 1051 g_free(s->children); 1052 } 1053 1054 static void quorum_add_child(BlockDriverState *bs, BlockDriverState *child_bs, 1055 Error **errp) 1056 { 1057 BDRVQuorumState *s = bs->opaque; 1058 BdrvChild *child; 1059 char indexstr[INDEXSTR_LEN]; 1060 int ret; 1061 1062 if (s->is_blkverify) { 1063 error_setg(errp, "Cannot add a child to a quorum in blkverify mode"); 1064 return; 1065 } 1066 1067 assert(s->num_children <= INT_MAX / sizeof(BdrvChild *)); 1068 if (s->num_children == INT_MAX / sizeof(BdrvChild *) || 1069 s->next_child_index == UINT_MAX) { 1070 error_setg(errp, "Too many children"); 1071 return; 1072 } 1073 1074 ret = snprintf(indexstr, INDEXSTR_LEN, "children.%u", s->next_child_index); 1075 if (ret < 0 || ret >= INDEXSTR_LEN) { 1076 error_setg(errp, "cannot generate child name"); 1077 return; 1078 } 1079 s->next_child_index++; 1080 1081 bdrv_drained_begin(bs); 1082 1083 /* We can safely add the child now */ 1084 bdrv_ref(child_bs); 1085 1086 child = bdrv_attach_child(bs, child_bs, indexstr, &child_of_bds, 1087 BDRV_CHILD_DATA, errp); 1088 if (child == NULL) { 1089 s->next_child_index--; 1090 goto out; 1091 } 1092 s->children = g_renew(BdrvChild *, s->children, s->num_children + 1); 1093 s->children[s->num_children++] = child; 1094 quorum_refresh_flags(bs); 1095 1096 out: 1097 bdrv_drained_end(bs); 1098 } 1099 1100 static void quorum_del_child(BlockDriverState *bs, BdrvChild *child, 1101 Error **errp) 1102 { 1103 BDRVQuorumState *s = bs->opaque; 1104 char indexstr[INDEXSTR_LEN]; 1105 int i; 1106 1107 for (i = 0; i < s->num_children; i++) { 1108 if (s->children[i] == child) { 1109 break; 1110 } 1111 } 1112 1113 /* we have checked it in bdrv_del_child() */ 1114 assert(i < s->num_children); 1115 1116 if (s->num_children <= s->threshold) { 1117 error_setg(errp, 1118 "The number of children cannot be lower than the vote threshold %d", 1119 s->threshold); 1120 return; 1121 } 1122 1123 /* We know now that num_children > threshold, so blkverify must be false */ 1124 assert(!s->is_blkverify); 1125 1126 snprintf(indexstr, INDEXSTR_LEN, "children.%u", s->next_child_index - 1); 1127 if (!strncmp(child->name, indexstr, INDEXSTR_LEN)) { 1128 s->next_child_index--; 1129 } 1130 1131 bdrv_drained_begin(bs); 1132 1133 /* We can safely remove this child now */ 1134 memmove(&s->children[i], &s->children[i + 1], 1135 (s->num_children - i - 1) * sizeof(BdrvChild *)); 1136 s->children = g_renew(BdrvChild *, s->children, --s->num_children); 1137 bdrv_unref_child(bs, child); 1138 1139 quorum_refresh_flags(bs); 1140 bdrv_drained_end(bs); 1141 } 1142 1143 static void quorum_gather_child_options(BlockDriverState *bs, QDict *target, 1144 bool backing_overridden) 1145 { 1146 BDRVQuorumState *s = bs->opaque; 1147 QList *children_list; 1148 int i; 1149 1150 /* 1151 * The generic implementation for gathering child options in 1152 * bdrv_refresh_filename() would use the names of the children 1153 * as specified for bdrv_open_child() or bdrv_attach_child(), 1154 * which is "children.%u" with %u being a value 1155 * (s->next_child_index) that is incremented each time a new child 1156 * is added (and never decremented). Since children can be 1157 * deleted at runtime, there may be gaps in that enumeration. 1158 * When creating a new quorum BDS and specifying the children for 1159 * it through runtime options, the enumeration used there may not 1160 * have any gaps, though. 1161 * 1162 * Therefore, we have to create a new gap-less enumeration here 1163 * (which we can achieve by simply putting all of the children's 1164 * full_open_options into a QList). 1165 * 1166 * XXX: Note that there are issues with the current child option 1167 * structure quorum uses (such as the fact that children do 1168 * not really have unique permanent names). Therefore, this 1169 * is going to have to change in the future and ideally we 1170 * want quorum to be covered by the generic implementation. 1171 */ 1172 1173 children_list = qlist_new(); 1174 qdict_put(target, "children", children_list); 1175 1176 for (i = 0; i < s->num_children; i++) { 1177 qlist_append(children_list, 1178 qobject_ref(s->children[i]->bs->full_open_options)); 1179 } 1180 } 1181 1182 static char *quorum_dirname(BlockDriverState *bs, Error **errp) 1183 { 1184 /* In general, there are multiple BDSs with different dirnames below this 1185 * one; so there is no unique dirname we could return (unless all are equal 1186 * by chance, or there is only one). Therefore, to be consistent, just 1187 * always return NULL. */ 1188 error_setg(errp, "Cannot generate a base directory for quorum nodes"); 1189 return NULL; 1190 } 1191 1192 static void quorum_child_perm(BlockDriverState *bs, BdrvChild *c, 1193 BdrvChildRole role, 1194 BlockReopenQueue *reopen_queue, 1195 uint64_t perm, uint64_t shared, 1196 uint64_t *nperm, uint64_t *nshared) 1197 { 1198 BDRVQuorumState *s = bs->opaque; 1199 1200 *nperm = perm & DEFAULT_PERM_PASSTHROUGH; 1201 if (s->rewrite_corrupted) { 1202 *nperm |= BLK_PERM_WRITE; 1203 } 1204 1205 /* 1206 * We cannot share RESIZE or WRITE, as this would make the 1207 * children differ from each other. 1208 */ 1209 *nshared = (shared & (BLK_PERM_CONSISTENT_READ | 1210 BLK_PERM_WRITE_UNCHANGED)) 1211 | DEFAULT_PERM_UNCHANGED; 1212 } 1213 1214 /* 1215 * Each one of the children can report different status flags even 1216 * when they contain the same data, so what this function does is 1217 * return BDRV_BLOCK_ZERO if *all* children agree that a certain 1218 * region contains zeroes, and BDRV_BLOCK_DATA otherwise. 1219 */ 1220 static int coroutine_fn quorum_co_block_status(BlockDriverState *bs, 1221 bool want_zero, 1222 int64_t offset, int64_t count, 1223 int64_t *pnum, int64_t *map, 1224 BlockDriverState **file) 1225 { 1226 BDRVQuorumState *s = bs->opaque; 1227 int i, ret; 1228 int64_t pnum_zero = count; 1229 int64_t pnum_data = 0; 1230 1231 for (i = 0; i < s->num_children; i++) { 1232 int64_t bytes; 1233 ret = bdrv_co_common_block_status_above(s->children[i]->bs, NULL, false, 1234 want_zero, offset, count, 1235 &bytes, NULL, NULL, NULL); 1236 if (ret < 0) { 1237 quorum_report_bad(QUORUM_OP_TYPE_READ, offset, count, 1238 s->children[i]->bs->node_name, ret); 1239 pnum_data = count; 1240 break; 1241 } 1242 /* 1243 * Even if all children agree about whether there are zeroes 1244 * or not at @offset they might disagree on the size, so use 1245 * the smallest when reporting BDRV_BLOCK_ZERO and the largest 1246 * when reporting BDRV_BLOCK_DATA. 1247 */ 1248 if (ret & BDRV_BLOCK_ZERO) { 1249 pnum_zero = MIN(pnum_zero, bytes); 1250 } else { 1251 pnum_data = MAX(pnum_data, bytes); 1252 } 1253 } 1254 1255 if (pnum_data) { 1256 *pnum = pnum_data; 1257 return BDRV_BLOCK_DATA; 1258 } else { 1259 *pnum = pnum_zero; 1260 return BDRV_BLOCK_ZERO; 1261 } 1262 } 1263 1264 static const char *const quorum_strong_runtime_opts[] = { 1265 QUORUM_OPT_VOTE_THRESHOLD, 1266 QUORUM_OPT_BLKVERIFY, 1267 QUORUM_OPT_REWRITE, 1268 QUORUM_OPT_READ_PATTERN, 1269 1270 NULL 1271 }; 1272 1273 static BlockDriver bdrv_quorum = { 1274 .format_name = "quorum", 1275 1276 .instance_size = sizeof(BDRVQuorumState), 1277 1278 .bdrv_open = quorum_open, 1279 .bdrv_close = quorum_close, 1280 .bdrv_gather_child_options = quorum_gather_child_options, 1281 .bdrv_dirname = quorum_dirname, 1282 .bdrv_co_block_status = quorum_co_block_status, 1283 1284 .bdrv_co_flush_to_disk = quorum_co_flush, 1285 1286 .bdrv_getlength = quorum_getlength, 1287 1288 .bdrv_co_preadv = quorum_co_preadv, 1289 .bdrv_co_pwritev = quorum_co_pwritev, 1290 .bdrv_co_pwrite_zeroes = quorum_co_pwrite_zeroes, 1291 1292 .bdrv_add_child = quorum_add_child, 1293 .bdrv_del_child = quorum_del_child, 1294 1295 .bdrv_child_perm = quorum_child_perm, 1296 1297 .bdrv_recurse_can_replace = quorum_recurse_can_replace, 1298 1299 .strong_runtime_opts = quorum_strong_runtime_opts, 1300 }; 1301 1302 static void bdrv_quorum_init(void) 1303 { 1304 if (!qcrypto_hash_supports(QCRYPTO_HASH_ALG_SHA256)) { 1305 /* SHA256 hash support is required for quorum device */ 1306 return; 1307 } 1308 bdrv_register(&bdrv_quorum); 1309 } 1310 1311 block_init(bdrv_quorum_init); 1312