1 /* 2 * QEMU Enhanced Disk Format 3 * 4 * Copyright IBM, Corp. 2010 5 * 6 * Authors: 7 * Stefan Hajnoczi <stefanha@linux.vnet.ibm.com> 8 * Anthony Liguori <aliguori@us.ibm.com> 9 * 10 * This work is licensed under the terms of the GNU LGPL, version 2 or later. 11 * See the COPYING.LIB file in the top-level directory. 12 * 13 */ 14 15 #include "qemu/timer.h" 16 #include "trace.h" 17 #include "qed.h" 18 #include "qapi/qmp/qerror.h" 19 #include "migration/migration.h" 20 21 static void qed_aio_cancel(BlockDriverAIOCB *blockacb) 22 { 23 QEDAIOCB *acb = (QEDAIOCB *)blockacb; 24 bool finished = false; 25 26 /* Wait for the request to finish */ 27 acb->finished = &finished; 28 while (!finished) { 29 qemu_aio_wait(); 30 } 31 } 32 33 static const AIOCBInfo qed_aiocb_info = { 34 .aiocb_size = sizeof(QEDAIOCB), 35 .cancel = qed_aio_cancel, 36 }; 37 38 static int bdrv_qed_probe(const uint8_t *buf, int buf_size, 39 const char *filename) 40 { 41 const QEDHeader *header = (const QEDHeader *)buf; 42 43 if (buf_size < sizeof(*header)) { 44 return 0; 45 } 46 if (le32_to_cpu(header->magic) != QED_MAGIC) { 47 return 0; 48 } 49 return 100; 50 } 51 52 /** 53 * Check whether an image format is raw 54 * 55 * @fmt: Backing file format, may be NULL 56 */ 57 static bool qed_fmt_is_raw(const char *fmt) 58 { 59 return fmt && strcmp(fmt, "raw") == 0; 60 } 61 62 static void qed_header_le_to_cpu(const QEDHeader *le, QEDHeader *cpu) 63 { 64 cpu->magic = le32_to_cpu(le->magic); 65 cpu->cluster_size = le32_to_cpu(le->cluster_size); 66 cpu->table_size = le32_to_cpu(le->table_size); 67 cpu->header_size = le32_to_cpu(le->header_size); 68 cpu->features = le64_to_cpu(le->features); 69 cpu->compat_features = le64_to_cpu(le->compat_features); 70 cpu->autoclear_features = le64_to_cpu(le->autoclear_features); 71 cpu->l1_table_offset = le64_to_cpu(le->l1_table_offset); 72 cpu->image_size = le64_to_cpu(le->image_size); 73 cpu->backing_filename_offset = le32_to_cpu(le->backing_filename_offset); 74 cpu->backing_filename_size = le32_to_cpu(le->backing_filename_size); 75 } 76 77 static void qed_header_cpu_to_le(const QEDHeader *cpu, QEDHeader *le) 78 { 79 le->magic = cpu_to_le32(cpu->magic); 80 le->cluster_size = cpu_to_le32(cpu->cluster_size); 81 le->table_size = cpu_to_le32(cpu->table_size); 82 le->header_size = cpu_to_le32(cpu->header_size); 83 le->features = cpu_to_le64(cpu->features); 84 le->compat_features = cpu_to_le64(cpu->compat_features); 85 le->autoclear_features = cpu_to_le64(cpu->autoclear_features); 86 le->l1_table_offset = cpu_to_le64(cpu->l1_table_offset); 87 le->image_size = cpu_to_le64(cpu->image_size); 88 le->backing_filename_offset = cpu_to_le32(cpu->backing_filename_offset); 89 le->backing_filename_size = cpu_to_le32(cpu->backing_filename_size); 90 } 91 92 int qed_write_header_sync(BDRVQEDState *s) 93 { 94 QEDHeader le; 95 int ret; 96 97 qed_header_cpu_to_le(&s->header, &le); 98 ret = bdrv_pwrite(s->bs->file, 0, &le, sizeof(le)); 99 if (ret != sizeof(le)) { 100 return ret; 101 } 102 return 0; 103 } 104 105 typedef struct { 106 GenericCB gencb; 107 BDRVQEDState *s; 108 struct iovec iov; 109 QEMUIOVector qiov; 110 int nsectors; 111 uint8_t *buf; 112 } QEDWriteHeaderCB; 113 114 static void qed_write_header_cb(void *opaque, int ret) 115 { 116 QEDWriteHeaderCB *write_header_cb = opaque; 117 118 qemu_vfree(write_header_cb->buf); 119 gencb_complete(write_header_cb, ret); 120 } 121 122 static void qed_write_header_read_cb(void *opaque, int ret) 123 { 124 QEDWriteHeaderCB *write_header_cb = opaque; 125 BDRVQEDState *s = write_header_cb->s; 126 127 if (ret) { 128 qed_write_header_cb(write_header_cb, ret); 129 return; 130 } 131 132 /* Update header */ 133 qed_header_cpu_to_le(&s->header, (QEDHeader *)write_header_cb->buf); 134 135 bdrv_aio_writev(s->bs->file, 0, &write_header_cb->qiov, 136 write_header_cb->nsectors, qed_write_header_cb, 137 write_header_cb); 138 } 139 140 /** 141 * Update header in-place (does not rewrite backing filename or other strings) 142 * 143 * This function only updates known header fields in-place and does not affect 144 * extra data after the QED header. 145 */ 146 static void qed_write_header(BDRVQEDState *s, BlockDriverCompletionFunc cb, 147 void *opaque) 148 { 149 /* We must write full sectors for O_DIRECT but cannot necessarily generate 150 * the data following the header if an unrecognized compat feature is 151 * active. Therefore, first read the sectors containing the header, update 152 * them, and write back. 153 */ 154 155 int nsectors = (sizeof(QEDHeader) + BDRV_SECTOR_SIZE - 1) / 156 BDRV_SECTOR_SIZE; 157 size_t len = nsectors * BDRV_SECTOR_SIZE; 158 QEDWriteHeaderCB *write_header_cb = gencb_alloc(sizeof(*write_header_cb), 159 cb, opaque); 160 161 write_header_cb->s = s; 162 write_header_cb->nsectors = nsectors; 163 write_header_cb->buf = qemu_blockalign(s->bs, len); 164 write_header_cb->iov.iov_base = write_header_cb->buf; 165 write_header_cb->iov.iov_len = len; 166 qemu_iovec_init_external(&write_header_cb->qiov, &write_header_cb->iov, 1); 167 168 bdrv_aio_readv(s->bs->file, 0, &write_header_cb->qiov, nsectors, 169 qed_write_header_read_cb, write_header_cb); 170 } 171 172 static uint64_t qed_max_image_size(uint32_t cluster_size, uint32_t table_size) 173 { 174 uint64_t table_entries; 175 uint64_t l2_size; 176 177 table_entries = (table_size * cluster_size) / sizeof(uint64_t); 178 l2_size = table_entries * cluster_size; 179 180 return l2_size * table_entries; 181 } 182 183 static bool qed_is_cluster_size_valid(uint32_t cluster_size) 184 { 185 if (cluster_size < QED_MIN_CLUSTER_SIZE || 186 cluster_size > QED_MAX_CLUSTER_SIZE) { 187 return false; 188 } 189 if (cluster_size & (cluster_size - 1)) { 190 return false; /* not power of 2 */ 191 } 192 return true; 193 } 194 195 static bool qed_is_table_size_valid(uint32_t table_size) 196 { 197 if (table_size < QED_MIN_TABLE_SIZE || 198 table_size > QED_MAX_TABLE_SIZE) { 199 return false; 200 } 201 if (table_size & (table_size - 1)) { 202 return false; /* not power of 2 */ 203 } 204 return true; 205 } 206 207 static bool qed_is_image_size_valid(uint64_t image_size, uint32_t cluster_size, 208 uint32_t table_size) 209 { 210 if (image_size % BDRV_SECTOR_SIZE != 0) { 211 return false; /* not multiple of sector size */ 212 } 213 if (image_size > qed_max_image_size(cluster_size, table_size)) { 214 return false; /* image is too large */ 215 } 216 return true; 217 } 218 219 /** 220 * Read a string of known length from the image file 221 * 222 * @file: Image file 223 * @offset: File offset to start of string, in bytes 224 * @n: String length in bytes 225 * @buf: Destination buffer 226 * @buflen: Destination buffer length in bytes 227 * @ret: 0 on success, -errno on failure 228 * 229 * The string is NUL-terminated. 230 */ 231 static int qed_read_string(BlockDriverState *file, uint64_t offset, size_t n, 232 char *buf, size_t buflen) 233 { 234 int ret; 235 if (n >= buflen) { 236 return -EINVAL; 237 } 238 ret = bdrv_pread(file, offset, buf, n); 239 if (ret < 0) { 240 return ret; 241 } 242 buf[n] = '\0'; 243 return 0; 244 } 245 246 /** 247 * Allocate new clusters 248 * 249 * @s: QED state 250 * @n: Number of contiguous clusters to allocate 251 * @ret: Offset of first allocated cluster 252 * 253 * This function only produces the offset where the new clusters should be 254 * written. It updates BDRVQEDState but does not make any changes to the image 255 * file. 256 */ 257 static uint64_t qed_alloc_clusters(BDRVQEDState *s, unsigned int n) 258 { 259 uint64_t offset = s->file_size; 260 s->file_size += n * s->header.cluster_size; 261 return offset; 262 } 263 264 QEDTable *qed_alloc_table(BDRVQEDState *s) 265 { 266 /* Honor O_DIRECT memory alignment requirements */ 267 return qemu_blockalign(s->bs, 268 s->header.cluster_size * s->header.table_size); 269 } 270 271 /** 272 * Allocate a new zeroed L2 table 273 */ 274 static CachedL2Table *qed_new_l2_table(BDRVQEDState *s) 275 { 276 CachedL2Table *l2_table = qed_alloc_l2_cache_entry(&s->l2_cache); 277 278 l2_table->table = qed_alloc_table(s); 279 l2_table->offset = qed_alloc_clusters(s, s->header.table_size); 280 281 memset(l2_table->table->offsets, 0, 282 s->header.cluster_size * s->header.table_size); 283 return l2_table; 284 } 285 286 static void qed_aio_next_io(void *opaque, int ret); 287 288 static void qed_plug_allocating_write_reqs(BDRVQEDState *s) 289 { 290 assert(!s->allocating_write_reqs_plugged); 291 292 s->allocating_write_reqs_plugged = true; 293 } 294 295 static void qed_unplug_allocating_write_reqs(BDRVQEDState *s) 296 { 297 QEDAIOCB *acb; 298 299 assert(s->allocating_write_reqs_plugged); 300 301 s->allocating_write_reqs_plugged = false; 302 303 acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs); 304 if (acb) { 305 qed_aio_next_io(acb, 0); 306 } 307 } 308 309 static void qed_finish_clear_need_check(void *opaque, int ret) 310 { 311 /* Do nothing */ 312 } 313 314 static void qed_flush_after_clear_need_check(void *opaque, int ret) 315 { 316 BDRVQEDState *s = opaque; 317 318 bdrv_aio_flush(s->bs, qed_finish_clear_need_check, s); 319 320 /* No need to wait until flush completes */ 321 qed_unplug_allocating_write_reqs(s); 322 } 323 324 static void qed_clear_need_check(void *opaque, int ret) 325 { 326 BDRVQEDState *s = opaque; 327 328 if (ret) { 329 qed_unplug_allocating_write_reqs(s); 330 return; 331 } 332 333 s->header.features &= ~QED_F_NEED_CHECK; 334 qed_write_header(s, qed_flush_after_clear_need_check, s); 335 } 336 337 static void qed_need_check_timer_cb(void *opaque) 338 { 339 BDRVQEDState *s = opaque; 340 341 /* The timer should only fire when allocating writes have drained */ 342 assert(!QSIMPLEQ_FIRST(&s->allocating_write_reqs)); 343 344 trace_qed_need_check_timer_cb(s); 345 346 qed_plug_allocating_write_reqs(s); 347 348 /* Ensure writes are on disk before clearing flag */ 349 bdrv_aio_flush(s->bs, qed_clear_need_check, s); 350 } 351 352 static void qed_start_need_check_timer(BDRVQEDState *s) 353 { 354 trace_qed_start_need_check_timer(s); 355 356 /* Use QEMU_CLOCK_VIRTUAL so we don't alter the image file while suspended for 357 * migration. 358 */ 359 timer_mod(s->need_check_timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + 360 get_ticks_per_sec() * QED_NEED_CHECK_TIMEOUT); 361 } 362 363 /* It's okay to call this multiple times or when no timer is started */ 364 static void qed_cancel_need_check_timer(BDRVQEDState *s) 365 { 366 trace_qed_cancel_need_check_timer(s); 367 timer_del(s->need_check_timer); 368 } 369 370 static void bdrv_qed_rebind(BlockDriverState *bs) 371 { 372 BDRVQEDState *s = bs->opaque; 373 s->bs = bs; 374 } 375 376 static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags, 377 Error **errp) 378 { 379 BDRVQEDState *s = bs->opaque; 380 QEDHeader le_header; 381 int64_t file_size; 382 int ret; 383 384 s->bs = bs; 385 QSIMPLEQ_INIT(&s->allocating_write_reqs); 386 387 ret = bdrv_pread(bs->file, 0, &le_header, sizeof(le_header)); 388 if (ret < 0) { 389 return ret; 390 } 391 qed_header_le_to_cpu(&le_header, &s->header); 392 393 if (s->header.magic != QED_MAGIC) { 394 return -EMEDIUMTYPE; 395 } 396 if (s->header.features & ~QED_FEATURE_MASK) { 397 /* image uses unsupported feature bits */ 398 char buf[64]; 399 snprintf(buf, sizeof(buf), "%" PRIx64, 400 s->header.features & ~QED_FEATURE_MASK); 401 qerror_report(QERR_UNKNOWN_BLOCK_FORMAT_FEATURE, 402 bs->device_name, "QED", buf); 403 return -ENOTSUP; 404 } 405 if (!qed_is_cluster_size_valid(s->header.cluster_size)) { 406 return -EINVAL; 407 } 408 409 /* Round down file size to the last cluster */ 410 file_size = bdrv_getlength(bs->file); 411 if (file_size < 0) { 412 return file_size; 413 } 414 s->file_size = qed_start_of_cluster(s, file_size); 415 416 if (!qed_is_table_size_valid(s->header.table_size)) { 417 return -EINVAL; 418 } 419 if (!qed_is_image_size_valid(s->header.image_size, 420 s->header.cluster_size, 421 s->header.table_size)) { 422 return -EINVAL; 423 } 424 if (!qed_check_table_offset(s, s->header.l1_table_offset)) { 425 return -EINVAL; 426 } 427 428 s->table_nelems = (s->header.cluster_size * s->header.table_size) / 429 sizeof(uint64_t); 430 s->l2_shift = ffs(s->header.cluster_size) - 1; 431 s->l2_mask = s->table_nelems - 1; 432 s->l1_shift = s->l2_shift + ffs(s->table_nelems) - 1; 433 434 if ((s->header.features & QED_F_BACKING_FILE)) { 435 if ((uint64_t)s->header.backing_filename_offset + 436 s->header.backing_filename_size > 437 s->header.cluster_size * s->header.header_size) { 438 return -EINVAL; 439 } 440 441 ret = qed_read_string(bs->file, s->header.backing_filename_offset, 442 s->header.backing_filename_size, bs->backing_file, 443 sizeof(bs->backing_file)); 444 if (ret < 0) { 445 return ret; 446 } 447 448 if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) { 449 pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw"); 450 } 451 } 452 453 /* Reset unknown autoclear feature bits. This is a backwards 454 * compatibility mechanism that allows images to be opened by older 455 * programs, which "knock out" unknown feature bits. When an image is 456 * opened by a newer program again it can detect that the autoclear 457 * feature is no longer valid. 458 */ 459 if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 && 460 !bdrv_is_read_only(bs->file) && !(flags & BDRV_O_INCOMING)) { 461 s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK; 462 463 ret = qed_write_header_sync(s); 464 if (ret) { 465 return ret; 466 } 467 468 /* From here on only known autoclear feature bits are valid */ 469 bdrv_flush(bs->file); 470 } 471 472 s->l1_table = qed_alloc_table(s); 473 qed_init_l2_cache(&s->l2_cache); 474 475 ret = qed_read_l1_table_sync(s); 476 if (ret) { 477 goto out; 478 } 479 480 /* If image was not closed cleanly, check consistency */ 481 if (!(flags & BDRV_O_CHECK) && (s->header.features & QED_F_NEED_CHECK)) { 482 /* Read-only images cannot be fixed. There is no risk of corruption 483 * since write operations are not possible. Therefore, allow 484 * potentially inconsistent images to be opened read-only. This can 485 * aid data recovery from an otherwise inconsistent image. 486 */ 487 if (!bdrv_is_read_only(bs->file) && 488 !(flags & BDRV_O_INCOMING)) { 489 BdrvCheckResult result = {0}; 490 491 ret = qed_check(s, &result, true); 492 if (ret) { 493 goto out; 494 } 495 } 496 } 497 498 s->need_check_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL, 499 qed_need_check_timer_cb, s); 500 501 out: 502 if (ret) { 503 qed_free_l2_cache(&s->l2_cache); 504 qemu_vfree(s->l1_table); 505 } 506 return ret; 507 } 508 509 static int bdrv_qed_refresh_limits(BlockDriverState *bs) 510 { 511 BDRVQEDState *s = bs->opaque; 512 513 bs->bl.write_zeroes_alignment = s->header.cluster_size >> BDRV_SECTOR_BITS; 514 515 return 0; 516 } 517 518 /* We have nothing to do for QED reopen, stubs just return 519 * success */ 520 static int bdrv_qed_reopen_prepare(BDRVReopenState *state, 521 BlockReopenQueue *queue, Error **errp) 522 { 523 return 0; 524 } 525 526 static void bdrv_qed_close(BlockDriverState *bs) 527 { 528 BDRVQEDState *s = bs->opaque; 529 530 qed_cancel_need_check_timer(s); 531 timer_free(s->need_check_timer); 532 533 /* Ensure writes reach stable storage */ 534 bdrv_flush(bs->file); 535 536 /* Clean shutdown, no check required on next open */ 537 if (s->header.features & QED_F_NEED_CHECK) { 538 s->header.features &= ~QED_F_NEED_CHECK; 539 qed_write_header_sync(s); 540 } 541 542 qed_free_l2_cache(&s->l2_cache); 543 qemu_vfree(s->l1_table); 544 } 545 546 static int qed_create(const char *filename, uint32_t cluster_size, 547 uint64_t image_size, uint32_t table_size, 548 const char *backing_file, const char *backing_fmt) 549 { 550 QEDHeader header = { 551 .magic = QED_MAGIC, 552 .cluster_size = cluster_size, 553 .table_size = table_size, 554 .header_size = 1, 555 .features = 0, 556 .compat_features = 0, 557 .l1_table_offset = cluster_size, 558 .image_size = image_size, 559 }; 560 QEDHeader le_header; 561 uint8_t *l1_table = NULL; 562 size_t l1_size = header.cluster_size * header.table_size; 563 Error *local_err = NULL; 564 int ret = 0; 565 BlockDriverState *bs = NULL; 566 567 ret = bdrv_create_file(filename, NULL, &local_err); 568 if (ret < 0) { 569 qerror_report_err(local_err); 570 error_free(local_err); 571 return ret; 572 } 573 574 ret = bdrv_file_open(&bs, filename, NULL, NULL, 575 BDRV_O_RDWR | BDRV_O_CACHE_WB, &local_err); 576 if (ret < 0) { 577 qerror_report_err(local_err); 578 error_free(local_err); 579 return ret; 580 } 581 582 /* File must start empty and grow, check truncate is supported */ 583 ret = bdrv_truncate(bs, 0); 584 if (ret < 0) { 585 goto out; 586 } 587 588 if (backing_file) { 589 header.features |= QED_F_BACKING_FILE; 590 header.backing_filename_offset = sizeof(le_header); 591 header.backing_filename_size = strlen(backing_file); 592 593 if (qed_fmt_is_raw(backing_fmt)) { 594 header.features |= QED_F_BACKING_FORMAT_NO_PROBE; 595 } 596 } 597 598 qed_header_cpu_to_le(&header, &le_header); 599 ret = bdrv_pwrite(bs, 0, &le_header, sizeof(le_header)); 600 if (ret < 0) { 601 goto out; 602 } 603 ret = bdrv_pwrite(bs, sizeof(le_header), backing_file, 604 header.backing_filename_size); 605 if (ret < 0) { 606 goto out; 607 } 608 609 l1_table = g_malloc0(l1_size); 610 ret = bdrv_pwrite(bs, header.l1_table_offset, l1_table, l1_size); 611 if (ret < 0) { 612 goto out; 613 } 614 615 ret = 0; /* success */ 616 out: 617 g_free(l1_table); 618 bdrv_unref(bs); 619 return ret; 620 } 621 622 static int bdrv_qed_create(const char *filename, QEMUOptionParameter *options, 623 Error **errp) 624 { 625 uint64_t image_size = 0; 626 uint32_t cluster_size = QED_DEFAULT_CLUSTER_SIZE; 627 uint32_t table_size = QED_DEFAULT_TABLE_SIZE; 628 const char *backing_file = NULL; 629 const char *backing_fmt = NULL; 630 631 while (options && options->name) { 632 if (!strcmp(options->name, BLOCK_OPT_SIZE)) { 633 image_size = options->value.n; 634 } else if (!strcmp(options->name, BLOCK_OPT_BACKING_FILE)) { 635 backing_file = options->value.s; 636 } else if (!strcmp(options->name, BLOCK_OPT_BACKING_FMT)) { 637 backing_fmt = options->value.s; 638 } else if (!strcmp(options->name, BLOCK_OPT_CLUSTER_SIZE)) { 639 if (options->value.n) { 640 cluster_size = options->value.n; 641 } 642 } else if (!strcmp(options->name, BLOCK_OPT_TABLE_SIZE)) { 643 if (options->value.n) { 644 table_size = options->value.n; 645 } 646 } 647 options++; 648 } 649 650 if (!qed_is_cluster_size_valid(cluster_size)) { 651 fprintf(stderr, "QED cluster size must be within range [%u, %u] and power of 2\n", 652 QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE); 653 return -EINVAL; 654 } 655 if (!qed_is_table_size_valid(table_size)) { 656 fprintf(stderr, "QED table size must be within range [%u, %u] and power of 2\n", 657 QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE); 658 return -EINVAL; 659 } 660 if (!qed_is_image_size_valid(image_size, cluster_size, table_size)) { 661 fprintf(stderr, "QED image size must be a non-zero multiple of " 662 "cluster size and less than %" PRIu64 " bytes\n", 663 qed_max_image_size(cluster_size, table_size)); 664 return -EINVAL; 665 } 666 667 return qed_create(filename, cluster_size, image_size, table_size, 668 backing_file, backing_fmt); 669 } 670 671 typedef struct { 672 BlockDriverState *bs; 673 Coroutine *co; 674 uint64_t pos; 675 int64_t status; 676 int *pnum; 677 } QEDIsAllocatedCB; 678 679 static void qed_is_allocated_cb(void *opaque, int ret, uint64_t offset, size_t len) 680 { 681 QEDIsAllocatedCB *cb = opaque; 682 BDRVQEDState *s = cb->bs->opaque; 683 *cb->pnum = len / BDRV_SECTOR_SIZE; 684 switch (ret) { 685 case QED_CLUSTER_FOUND: 686 offset |= qed_offset_into_cluster(s, cb->pos); 687 cb->status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID | offset; 688 break; 689 case QED_CLUSTER_ZERO: 690 cb->status = BDRV_BLOCK_ZERO; 691 break; 692 case QED_CLUSTER_L2: 693 case QED_CLUSTER_L1: 694 cb->status = 0; 695 break; 696 default: 697 assert(ret < 0); 698 cb->status = ret; 699 break; 700 } 701 702 if (cb->co) { 703 qemu_coroutine_enter(cb->co, NULL); 704 } 705 } 706 707 static int64_t coroutine_fn bdrv_qed_co_get_block_status(BlockDriverState *bs, 708 int64_t sector_num, 709 int nb_sectors, int *pnum) 710 { 711 BDRVQEDState *s = bs->opaque; 712 size_t len = (size_t)nb_sectors * BDRV_SECTOR_SIZE; 713 QEDIsAllocatedCB cb = { 714 .bs = bs, 715 .pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE, 716 .status = BDRV_BLOCK_OFFSET_MASK, 717 .pnum = pnum, 718 }; 719 QEDRequest request = { .l2_table = NULL }; 720 721 qed_find_cluster(s, &request, cb.pos, len, qed_is_allocated_cb, &cb); 722 723 /* Now sleep if the callback wasn't invoked immediately */ 724 while (cb.status == BDRV_BLOCK_OFFSET_MASK) { 725 cb.co = qemu_coroutine_self(); 726 qemu_coroutine_yield(); 727 } 728 729 qed_unref_l2_cache_entry(request.l2_table); 730 731 return cb.status; 732 } 733 734 static BDRVQEDState *acb_to_s(QEDAIOCB *acb) 735 { 736 return acb->common.bs->opaque; 737 } 738 739 /** 740 * Read from the backing file or zero-fill if no backing file 741 * 742 * @s: QED state 743 * @pos: Byte position in device 744 * @qiov: Destination I/O vector 745 * @cb: Completion function 746 * @opaque: User data for completion function 747 * 748 * This function reads qiov->size bytes starting at pos from the backing file. 749 * If there is no backing file then zeroes are read. 750 */ 751 static void qed_read_backing_file(BDRVQEDState *s, uint64_t pos, 752 QEMUIOVector *qiov, 753 BlockDriverCompletionFunc *cb, void *opaque) 754 { 755 uint64_t backing_length = 0; 756 size_t size; 757 758 /* If there is a backing file, get its length. Treat the absence of a 759 * backing file like a zero length backing file. 760 */ 761 if (s->bs->backing_hd) { 762 int64_t l = bdrv_getlength(s->bs->backing_hd); 763 if (l < 0) { 764 cb(opaque, l); 765 return; 766 } 767 backing_length = l; 768 } 769 770 /* Zero all sectors if reading beyond the end of the backing file */ 771 if (pos >= backing_length || 772 pos + qiov->size > backing_length) { 773 qemu_iovec_memset(qiov, 0, 0, qiov->size); 774 } 775 776 /* Complete now if there are no backing file sectors to read */ 777 if (pos >= backing_length) { 778 cb(opaque, 0); 779 return; 780 } 781 782 /* If the read straddles the end of the backing file, shorten it */ 783 size = MIN((uint64_t)backing_length - pos, qiov->size); 784 785 BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO); 786 bdrv_aio_readv(s->bs->backing_hd, pos / BDRV_SECTOR_SIZE, 787 qiov, size / BDRV_SECTOR_SIZE, cb, opaque); 788 } 789 790 typedef struct { 791 GenericCB gencb; 792 BDRVQEDState *s; 793 QEMUIOVector qiov; 794 struct iovec iov; 795 uint64_t offset; 796 } CopyFromBackingFileCB; 797 798 static void qed_copy_from_backing_file_cb(void *opaque, int ret) 799 { 800 CopyFromBackingFileCB *copy_cb = opaque; 801 qemu_vfree(copy_cb->iov.iov_base); 802 gencb_complete(©_cb->gencb, ret); 803 } 804 805 static void qed_copy_from_backing_file_write(void *opaque, int ret) 806 { 807 CopyFromBackingFileCB *copy_cb = opaque; 808 BDRVQEDState *s = copy_cb->s; 809 810 if (ret) { 811 qed_copy_from_backing_file_cb(copy_cb, ret); 812 return; 813 } 814 815 BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE); 816 bdrv_aio_writev(s->bs->file, copy_cb->offset / BDRV_SECTOR_SIZE, 817 ©_cb->qiov, copy_cb->qiov.size / BDRV_SECTOR_SIZE, 818 qed_copy_from_backing_file_cb, copy_cb); 819 } 820 821 /** 822 * Copy data from backing file into the image 823 * 824 * @s: QED state 825 * @pos: Byte position in device 826 * @len: Number of bytes 827 * @offset: Byte offset in image file 828 * @cb: Completion function 829 * @opaque: User data for completion function 830 */ 831 static void qed_copy_from_backing_file(BDRVQEDState *s, uint64_t pos, 832 uint64_t len, uint64_t offset, 833 BlockDriverCompletionFunc *cb, 834 void *opaque) 835 { 836 CopyFromBackingFileCB *copy_cb; 837 838 /* Skip copy entirely if there is no work to do */ 839 if (len == 0) { 840 cb(opaque, 0); 841 return; 842 } 843 844 copy_cb = gencb_alloc(sizeof(*copy_cb), cb, opaque); 845 copy_cb->s = s; 846 copy_cb->offset = offset; 847 copy_cb->iov.iov_base = qemu_blockalign(s->bs, len); 848 copy_cb->iov.iov_len = len; 849 qemu_iovec_init_external(©_cb->qiov, ©_cb->iov, 1); 850 851 qed_read_backing_file(s, pos, ©_cb->qiov, 852 qed_copy_from_backing_file_write, copy_cb); 853 } 854 855 /** 856 * Link one or more contiguous clusters into a table 857 * 858 * @s: QED state 859 * @table: L2 table 860 * @index: First cluster index 861 * @n: Number of contiguous clusters 862 * @cluster: First cluster offset 863 * 864 * The cluster offset may be an allocated byte offset in the image file, the 865 * zero cluster marker, or the unallocated cluster marker. 866 */ 867 static void qed_update_l2_table(BDRVQEDState *s, QEDTable *table, int index, 868 unsigned int n, uint64_t cluster) 869 { 870 int i; 871 for (i = index; i < index + n; i++) { 872 table->offsets[i] = cluster; 873 if (!qed_offset_is_unalloc_cluster(cluster) && 874 !qed_offset_is_zero_cluster(cluster)) { 875 cluster += s->header.cluster_size; 876 } 877 } 878 } 879 880 static void qed_aio_complete_bh(void *opaque) 881 { 882 QEDAIOCB *acb = opaque; 883 BlockDriverCompletionFunc *cb = acb->common.cb; 884 void *user_opaque = acb->common.opaque; 885 int ret = acb->bh_ret; 886 bool *finished = acb->finished; 887 888 qemu_bh_delete(acb->bh); 889 qemu_aio_release(acb); 890 891 /* Invoke callback */ 892 cb(user_opaque, ret); 893 894 /* Signal cancel completion */ 895 if (finished) { 896 *finished = true; 897 } 898 } 899 900 static void qed_aio_complete(QEDAIOCB *acb, int ret) 901 { 902 BDRVQEDState *s = acb_to_s(acb); 903 904 trace_qed_aio_complete(s, acb, ret); 905 906 /* Free resources */ 907 qemu_iovec_destroy(&acb->cur_qiov); 908 qed_unref_l2_cache_entry(acb->request.l2_table); 909 910 /* Free the buffer we may have allocated for zero writes */ 911 if (acb->flags & QED_AIOCB_ZERO) { 912 qemu_vfree(acb->qiov->iov[0].iov_base); 913 acb->qiov->iov[0].iov_base = NULL; 914 } 915 916 /* Arrange for a bh to invoke the completion function */ 917 acb->bh_ret = ret; 918 acb->bh = qemu_bh_new(qed_aio_complete_bh, acb); 919 qemu_bh_schedule(acb->bh); 920 921 /* Start next allocating write request waiting behind this one. Note that 922 * requests enqueue themselves when they first hit an unallocated cluster 923 * but they wait until the entire request is finished before waking up the 924 * next request in the queue. This ensures that we don't cycle through 925 * requests multiple times but rather finish one at a time completely. 926 */ 927 if (acb == QSIMPLEQ_FIRST(&s->allocating_write_reqs)) { 928 QSIMPLEQ_REMOVE_HEAD(&s->allocating_write_reqs, next); 929 acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs); 930 if (acb) { 931 qed_aio_next_io(acb, 0); 932 } else if (s->header.features & QED_F_NEED_CHECK) { 933 qed_start_need_check_timer(s); 934 } 935 } 936 } 937 938 /** 939 * Commit the current L2 table to the cache 940 */ 941 static void qed_commit_l2_update(void *opaque, int ret) 942 { 943 QEDAIOCB *acb = opaque; 944 BDRVQEDState *s = acb_to_s(acb); 945 CachedL2Table *l2_table = acb->request.l2_table; 946 uint64_t l2_offset = l2_table->offset; 947 948 qed_commit_l2_cache_entry(&s->l2_cache, l2_table); 949 950 /* This is guaranteed to succeed because we just committed the entry to the 951 * cache. 952 */ 953 acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset); 954 assert(acb->request.l2_table != NULL); 955 956 qed_aio_next_io(opaque, ret); 957 } 958 959 /** 960 * Update L1 table with new L2 table offset and write it out 961 */ 962 static void qed_aio_write_l1_update(void *opaque, int ret) 963 { 964 QEDAIOCB *acb = opaque; 965 BDRVQEDState *s = acb_to_s(acb); 966 int index; 967 968 if (ret) { 969 qed_aio_complete(acb, ret); 970 return; 971 } 972 973 index = qed_l1_index(s, acb->cur_pos); 974 s->l1_table->offsets[index] = acb->request.l2_table->offset; 975 976 qed_write_l1_table(s, index, 1, qed_commit_l2_update, acb); 977 } 978 979 /** 980 * Update L2 table with new cluster offsets and write them out 981 */ 982 static void qed_aio_write_l2_update(QEDAIOCB *acb, int ret, uint64_t offset) 983 { 984 BDRVQEDState *s = acb_to_s(acb); 985 bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1; 986 int index; 987 988 if (ret) { 989 goto err; 990 } 991 992 if (need_alloc) { 993 qed_unref_l2_cache_entry(acb->request.l2_table); 994 acb->request.l2_table = qed_new_l2_table(s); 995 } 996 997 index = qed_l2_index(s, acb->cur_pos); 998 qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters, 999 offset); 1000 1001 if (need_alloc) { 1002 /* Write out the whole new L2 table */ 1003 qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true, 1004 qed_aio_write_l1_update, acb); 1005 } else { 1006 /* Write out only the updated part of the L2 table */ 1007 qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters, false, 1008 qed_aio_next_io, acb); 1009 } 1010 return; 1011 1012 err: 1013 qed_aio_complete(acb, ret); 1014 } 1015 1016 static void qed_aio_write_l2_update_cb(void *opaque, int ret) 1017 { 1018 QEDAIOCB *acb = opaque; 1019 qed_aio_write_l2_update(acb, ret, acb->cur_cluster); 1020 } 1021 1022 /** 1023 * Flush new data clusters before updating the L2 table 1024 * 1025 * This flush is necessary when a backing file is in use. A crash during an 1026 * allocating write could result in empty clusters in the image. If the write 1027 * only touched a subregion of the cluster, then backing image sectors have 1028 * been lost in the untouched region. The solution is to flush after writing a 1029 * new data cluster and before updating the L2 table. 1030 */ 1031 static void qed_aio_write_flush_before_l2_update(void *opaque, int ret) 1032 { 1033 QEDAIOCB *acb = opaque; 1034 BDRVQEDState *s = acb_to_s(acb); 1035 1036 if (!bdrv_aio_flush(s->bs->file, qed_aio_write_l2_update_cb, opaque)) { 1037 qed_aio_complete(acb, -EIO); 1038 } 1039 } 1040 1041 /** 1042 * Write data to the image file 1043 */ 1044 static void qed_aio_write_main(void *opaque, int ret) 1045 { 1046 QEDAIOCB *acb = opaque; 1047 BDRVQEDState *s = acb_to_s(acb); 1048 uint64_t offset = acb->cur_cluster + 1049 qed_offset_into_cluster(s, acb->cur_pos); 1050 BlockDriverCompletionFunc *next_fn; 1051 1052 trace_qed_aio_write_main(s, acb, ret, offset, acb->cur_qiov.size); 1053 1054 if (ret) { 1055 qed_aio_complete(acb, ret); 1056 return; 1057 } 1058 1059 if (acb->find_cluster_ret == QED_CLUSTER_FOUND) { 1060 next_fn = qed_aio_next_io; 1061 } else { 1062 if (s->bs->backing_hd) { 1063 next_fn = qed_aio_write_flush_before_l2_update; 1064 } else { 1065 next_fn = qed_aio_write_l2_update_cb; 1066 } 1067 } 1068 1069 BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO); 1070 bdrv_aio_writev(s->bs->file, offset / BDRV_SECTOR_SIZE, 1071 &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE, 1072 next_fn, acb); 1073 } 1074 1075 /** 1076 * Populate back untouched region of new data cluster 1077 */ 1078 static void qed_aio_write_postfill(void *opaque, int ret) 1079 { 1080 QEDAIOCB *acb = opaque; 1081 BDRVQEDState *s = acb_to_s(acb); 1082 uint64_t start = acb->cur_pos + acb->cur_qiov.size; 1083 uint64_t len = 1084 qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start; 1085 uint64_t offset = acb->cur_cluster + 1086 qed_offset_into_cluster(s, acb->cur_pos) + 1087 acb->cur_qiov.size; 1088 1089 if (ret) { 1090 qed_aio_complete(acb, ret); 1091 return; 1092 } 1093 1094 trace_qed_aio_write_postfill(s, acb, start, len, offset); 1095 qed_copy_from_backing_file(s, start, len, offset, 1096 qed_aio_write_main, acb); 1097 } 1098 1099 /** 1100 * Populate front untouched region of new data cluster 1101 */ 1102 static void qed_aio_write_prefill(void *opaque, int ret) 1103 { 1104 QEDAIOCB *acb = opaque; 1105 BDRVQEDState *s = acb_to_s(acb); 1106 uint64_t start = qed_start_of_cluster(s, acb->cur_pos); 1107 uint64_t len = qed_offset_into_cluster(s, acb->cur_pos); 1108 1109 trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster); 1110 qed_copy_from_backing_file(s, start, len, acb->cur_cluster, 1111 qed_aio_write_postfill, acb); 1112 } 1113 1114 /** 1115 * Check if the QED_F_NEED_CHECK bit should be set during allocating write 1116 */ 1117 static bool qed_should_set_need_check(BDRVQEDState *s) 1118 { 1119 /* The flush before L2 update path ensures consistency */ 1120 if (s->bs->backing_hd) { 1121 return false; 1122 } 1123 1124 return !(s->header.features & QED_F_NEED_CHECK); 1125 } 1126 1127 static void qed_aio_write_zero_cluster(void *opaque, int ret) 1128 { 1129 QEDAIOCB *acb = opaque; 1130 1131 if (ret) { 1132 qed_aio_complete(acb, ret); 1133 return; 1134 } 1135 1136 qed_aio_write_l2_update(acb, 0, 1); 1137 } 1138 1139 /** 1140 * Write new data cluster 1141 * 1142 * @acb: Write request 1143 * @len: Length in bytes 1144 * 1145 * This path is taken when writing to previously unallocated clusters. 1146 */ 1147 static void qed_aio_write_alloc(QEDAIOCB *acb, size_t len) 1148 { 1149 BDRVQEDState *s = acb_to_s(acb); 1150 BlockDriverCompletionFunc *cb; 1151 1152 /* Cancel timer when the first allocating request comes in */ 1153 if (QSIMPLEQ_EMPTY(&s->allocating_write_reqs)) { 1154 qed_cancel_need_check_timer(s); 1155 } 1156 1157 /* Freeze this request if another allocating write is in progress */ 1158 if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs)) { 1159 QSIMPLEQ_INSERT_TAIL(&s->allocating_write_reqs, acb, next); 1160 } 1161 if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs) || 1162 s->allocating_write_reqs_plugged) { 1163 return; /* wait for existing request to finish */ 1164 } 1165 1166 acb->cur_nclusters = qed_bytes_to_clusters(s, 1167 qed_offset_into_cluster(s, acb->cur_pos) + len); 1168 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len); 1169 1170 if (acb->flags & QED_AIOCB_ZERO) { 1171 /* Skip ahead if the clusters are already zero */ 1172 if (acb->find_cluster_ret == QED_CLUSTER_ZERO) { 1173 qed_aio_next_io(acb, 0); 1174 return; 1175 } 1176 1177 cb = qed_aio_write_zero_cluster; 1178 } else { 1179 cb = qed_aio_write_prefill; 1180 acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters); 1181 } 1182 1183 if (qed_should_set_need_check(s)) { 1184 s->header.features |= QED_F_NEED_CHECK; 1185 qed_write_header(s, cb, acb); 1186 } else { 1187 cb(acb, 0); 1188 } 1189 } 1190 1191 /** 1192 * Write data cluster in place 1193 * 1194 * @acb: Write request 1195 * @offset: Cluster offset in bytes 1196 * @len: Length in bytes 1197 * 1198 * This path is taken when writing to already allocated clusters. 1199 */ 1200 static void qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset, size_t len) 1201 { 1202 /* Allocate buffer for zero writes */ 1203 if (acb->flags & QED_AIOCB_ZERO) { 1204 struct iovec *iov = acb->qiov->iov; 1205 1206 if (!iov->iov_base) { 1207 iov->iov_base = qemu_blockalign(acb->common.bs, iov->iov_len); 1208 memset(iov->iov_base, 0, iov->iov_len); 1209 } 1210 } 1211 1212 /* Calculate the I/O vector */ 1213 acb->cur_cluster = offset; 1214 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len); 1215 1216 /* Do the actual write */ 1217 qed_aio_write_main(acb, 0); 1218 } 1219 1220 /** 1221 * Write data cluster 1222 * 1223 * @opaque: Write request 1224 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1, 1225 * or -errno 1226 * @offset: Cluster offset in bytes 1227 * @len: Length in bytes 1228 * 1229 * Callback from qed_find_cluster(). 1230 */ 1231 static void qed_aio_write_data(void *opaque, int ret, 1232 uint64_t offset, size_t len) 1233 { 1234 QEDAIOCB *acb = opaque; 1235 1236 trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len); 1237 1238 acb->find_cluster_ret = ret; 1239 1240 switch (ret) { 1241 case QED_CLUSTER_FOUND: 1242 qed_aio_write_inplace(acb, offset, len); 1243 break; 1244 1245 case QED_CLUSTER_L2: 1246 case QED_CLUSTER_L1: 1247 case QED_CLUSTER_ZERO: 1248 qed_aio_write_alloc(acb, len); 1249 break; 1250 1251 default: 1252 qed_aio_complete(acb, ret); 1253 break; 1254 } 1255 } 1256 1257 /** 1258 * Read data cluster 1259 * 1260 * @opaque: Read request 1261 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1, 1262 * or -errno 1263 * @offset: Cluster offset in bytes 1264 * @len: Length in bytes 1265 * 1266 * Callback from qed_find_cluster(). 1267 */ 1268 static void qed_aio_read_data(void *opaque, int ret, 1269 uint64_t offset, size_t len) 1270 { 1271 QEDAIOCB *acb = opaque; 1272 BDRVQEDState *s = acb_to_s(acb); 1273 BlockDriverState *bs = acb->common.bs; 1274 1275 /* Adjust offset into cluster */ 1276 offset += qed_offset_into_cluster(s, acb->cur_pos); 1277 1278 trace_qed_aio_read_data(s, acb, ret, offset, len); 1279 1280 if (ret < 0) { 1281 goto err; 1282 } 1283 1284 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len); 1285 1286 /* Handle zero cluster and backing file reads */ 1287 if (ret == QED_CLUSTER_ZERO) { 1288 qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size); 1289 qed_aio_next_io(acb, 0); 1290 return; 1291 } else if (ret != QED_CLUSTER_FOUND) { 1292 qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov, 1293 qed_aio_next_io, acb); 1294 return; 1295 } 1296 1297 BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO); 1298 bdrv_aio_readv(bs->file, offset / BDRV_SECTOR_SIZE, 1299 &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE, 1300 qed_aio_next_io, acb); 1301 return; 1302 1303 err: 1304 qed_aio_complete(acb, ret); 1305 } 1306 1307 /** 1308 * Begin next I/O or complete the request 1309 */ 1310 static void qed_aio_next_io(void *opaque, int ret) 1311 { 1312 QEDAIOCB *acb = opaque; 1313 BDRVQEDState *s = acb_to_s(acb); 1314 QEDFindClusterFunc *io_fn = (acb->flags & QED_AIOCB_WRITE) ? 1315 qed_aio_write_data : qed_aio_read_data; 1316 1317 trace_qed_aio_next_io(s, acb, ret, acb->cur_pos + acb->cur_qiov.size); 1318 1319 /* Handle I/O error */ 1320 if (ret) { 1321 qed_aio_complete(acb, ret); 1322 return; 1323 } 1324 1325 acb->qiov_offset += acb->cur_qiov.size; 1326 acb->cur_pos += acb->cur_qiov.size; 1327 qemu_iovec_reset(&acb->cur_qiov); 1328 1329 /* Complete request */ 1330 if (acb->cur_pos >= acb->end_pos) { 1331 qed_aio_complete(acb, 0); 1332 return; 1333 } 1334 1335 /* Find next cluster and start I/O */ 1336 qed_find_cluster(s, &acb->request, 1337 acb->cur_pos, acb->end_pos - acb->cur_pos, 1338 io_fn, acb); 1339 } 1340 1341 static BlockDriverAIOCB *qed_aio_setup(BlockDriverState *bs, 1342 int64_t sector_num, 1343 QEMUIOVector *qiov, int nb_sectors, 1344 BlockDriverCompletionFunc *cb, 1345 void *opaque, int flags) 1346 { 1347 QEDAIOCB *acb = qemu_aio_get(&qed_aiocb_info, bs, cb, opaque); 1348 1349 trace_qed_aio_setup(bs->opaque, acb, sector_num, nb_sectors, 1350 opaque, flags); 1351 1352 acb->flags = flags; 1353 acb->finished = NULL; 1354 acb->qiov = qiov; 1355 acb->qiov_offset = 0; 1356 acb->cur_pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE; 1357 acb->end_pos = acb->cur_pos + nb_sectors * BDRV_SECTOR_SIZE; 1358 acb->request.l2_table = NULL; 1359 qemu_iovec_init(&acb->cur_qiov, qiov->niov); 1360 1361 /* Start request */ 1362 qed_aio_next_io(acb, 0); 1363 return &acb->common; 1364 } 1365 1366 static BlockDriverAIOCB *bdrv_qed_aio_readv(BlockDriverState *bs, 1367 int64_t sector_num, 1368 QEMUIOVector *qiov, int nb_sectors, 1369 BlockDriverCompletionFunc *cb, 1370 void *opaque) 1371 { 1372 return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb, opaque, 0); 1373 } 1374 1375 static BlockDriverAIOCB *bdrv_qed_aio_writev(BlockDriverState *bs, 1376 int64_t sector_num, 1377 QEMUIOVector *qiov, int nb_sectors, 1378 BlockDriverCompletionFunc *cb, 1379 void *opaque) 1380 { 1381 return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb, 1382 opaque, QED_AIOCB_WRITE); 1383 } 1384 1385 typedef struct { 1386 Coroutine *co; 1387 int ret; 1388 bool done; 1389 } QEDWriteZeroesCB; 1390 1391 static void coroutine_fn qed_co_write_zeroes_cb(void *opaque, int ret) 1392 { 1393 QEDWriteZeroesCB *cb = opaque; 1394 1395 cb->done = true; 1396 cb->ret = ret; 1397 if (cb->co) { 1398 qemu_coroutine_enter(cb->co, NULL); 1399 } 1400 } 1401 1402 static int coroutine_fn bdrv_qed_co_write_zeroes(BlockDriverState *bs, 1403 int64_t sector_num, 1404 int nb_sectors, 1405 BdrvRequestFlags flags) 1406 { 1407 BlockDriverAIOCB *blockacb; 1408 BDRVQEDState *s = bs->opaque; 1409 QEDWriteZeroesCB cb = { .done = false }; 1410 QEMUIOVector qiov; 1411 struct iovec iov; 1412 1413 /* Refuse if there are untouched backing file sectors */ 1414 if (bs->backing_hd) { 1415 if (qed_offset_into_cluster(s, sector_num * BDRV_SECTOR_SIZE) != 0) { 1416 return -ENOTSUP; 1417 } 1418 if (qed_offset_into_cluster(s, nb_sectors * BDRV_SECTOR_SIZE) != 0) { 1419 return -ENOTSUP; 1420 } 1421 } 1422 1423 /* Zero writes start without an I/O buffer. If a buffer becomes necessary 1424 * then it will be allocated during request processing. 1425 */ 1426 iov.iov_base = NULL, 1427 iov.iov_len = nb_sectors * BDRV_SECTOR_SIZE, 1428 1429 qemu_iovec_init_external(&qiov, &iov, 1); 1430 blockacb = qed_aio_setup(bs, sector_num, &qiov, nb_sectors, 1431 qed_co_write_zeroes_cb, &cb, 1432 QED_AIOCB_WRITE | QED_AIOCB_ZERO); 1433 if (!blockacb) { 1434 return -EIO; 1435 } 1436 if (!cb.done) { 1437 cb.co = qemu_coroutine_self(); 1438 qemu_coroutine_yield(); 1439 } 1440 assert(cb.done); 1441 return cb.ret; 1442 } 1443 1444 static int bdrv_qed_truncate(BlockDriverState *bs, int64_t offset) 1445 { 1446 BDRVQEDState *s = bs->opaque; 1447 uint64_t old_image_size; 1448 int ret; 1449 1450 if (!qed_is_image_size_valid(offset, s->header.cluster_size, 1451 s->header.table_size)) { 1452 return -EINVAL; 1453 } 1454 1455 /* Shrinking is currently not supported */ 1456 if ((uint64_t)offset < s->header.image_size) { 1457 return -ENOTSUP; 1458 } 1459 1460 old_image_size = s->header.image_size; 1461 s->header.image_size = offset; 1462 ret = qed_write_header_sync(s); 1463 if (ret < 0) { 1464 s->header.image_size = old_image_size; 1465 } 1466 return ret; 1467 } 1468 1469 static int64_t bdrv_qed_getlength(BlockDriverState *bs) 1470 { 1471 BDRVQEDState *s = bs->opaque; 1472 return s->header.image_size; 1473 } 1474 1475 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi) 1476 { 1477 BDRVQEDState *s = bs->opaque; 1478 1479 memset(bdi, 0, sizeof(*bdi)); 1480 bdi->cluster_size = s->header.cluster_size; 1481 bdi->is_dirty = s->header.features & QED_F_NEED_CHECK; 1482 bdi->unallocated_blocks_are_zero = true; 1483 bdi->can_write_zeroes_with_unmap = true; 1484 return 0; 1485 } 1486 1487 static int bdrv_qed_change_backing_file(BlockDriverState *bs, 1488 const char *backing_file, 1489 const char *backing_fmt) 1490 { 1491 BDRVQEDState *s = bs->opaque; 1492 QEDHeader new_header, le_header; 1493 void *buffer; 1494 size_t buffer_len, backing_file_len; 1495 int ret; 1496 1497 /* Refuse to set backing filename if unknown compat feature bits are 1498 * active. If the image uses an unknown compat feature then we may not 1499 * know the layout of data following the header structure and cannot safely 1500 * add a new string. 1501 */ 1502 if (backing_file && (s->header.compat_features & 1503 ~QED_COMPAT_FEATURE_MASK)) { 1504 return -ENOTSUP; 1505 } 1506 1507 memcpy(&new_header, &s->header, sizeof(new_header)); 1508 1509 new_header.features &= ~(QED_F_BACKING_FILE | 1510 QED_F_BACKING_FORMAT_NO_PROBE); 1511 1512 /* Adjust feature flags */ 1513 if (backing_file) { 1514 new_header.features |= QED_F_BACKING_FILE; 1515 1516 if (qed_fmt_is_raw(backing_fmt)) { 1517 new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE; 1518 } 1519 } 1520 1521 /* Calculate new header size */ 1522 backing_file_len = 0; 1523 1524 if (backing_file) { 1525 backing_file_len = strlen(backing_file); 1526 } 1527 1528 buffer_len = sizeof(new_header); 1529 new_header.backing_filename_offset = buffer_len; 1530 new_header.backing_filename_size = backing_file_len; 1531 buffer_len += backing_file_len; 1532 1533 /* Make sure we can rewrite header without failing */ 1534 if (buffer_len > new_header.header_size * new_header.cluster_size) { 1535 return -ENOSPC; 1536 } 1537 1538 /* Prepare new header */ 1539 buffer = g_malloc(buffer_len); 1540 1541 qed_header_cpu_to_le(&new_header, &le_header); 1542 memcpy(buffer, &le_header, sizeof(le_header)); 1543 buffer_len = sizeof(le_header); 1544 1545 if (backing_file) { 1546 memcpy(buffer + buffer_len, backing_file, backing_file_len); 1547 buffer_len += backing_file_len; 1548 } 1549 1550 /* Write new header */ 1551 ret = bdrv_pwrite_sync(bs->file, 0, buffer, buffer_len); 1552 g_free(buffer); 1553 if (ret == 0) { 1554 memcpy(&s->header, &new_header, sizeof(new_header)); 1555 } 1556 return ret; 1557 } 1558 1559 static void bdrv_qed_invalidate_cache(BlockDriverState *bs) 1560 { 1561 BDRVQEDState *s = bs->opaque; 1562 1563 bdrv_qed_close(bs); 1564 memset(s, 0, sizeof(BDRVQEDState)); 1565 bdrv_qed_open(bs, NULL, bs->open_flags, NULL); 1566 } 1567 1568 static int bdrv_qed_check(BlockDriverState *bs, BdrvCheckResult *result, 1569 BdrvCheckMode fix) 1570 { 1571 BDRVQEDState *s = bs->opaque; 1572 1573 return qed_check(s, result, !!fix); 1574 } 1575 1576 static QEMUOptionParameter qed_create_options[] = { 1577 { 1578 .name = BLOCK_OPT_SIZE, 1579 .type = OPT_SIZE, 1580 .help = "Virtual disk size (in bytes)" 1581 }, { 1582 .name = BLOCK_OPT_BACKING_FILE, 1583 .type = OPT_STRING, 1584 .help = "File name of a base image" 1585 }, { 1586 .name = BLOCK_OPT_BACKING_FMT, 1587 .type = OPT_STRING, 1588 .help = "Image format of the base image" 1589 }, { 1590 .name = BLOCK_OPT_CLUSTER_SIZE, 1591 .type = OPT_SIZE, 1592 .help = "Cluster size (in bytes)", 1593 .value = { .n = QED_DEFAULT_CLUSTER_SIZE }, 1594 }, { 1595 .name = BLOCK_OPT_TABLE_SIZE, 1596 .type = OPT_SIZE, 1597 .help = "L1/L2 table size (in clusters)" 1598 }, 1599 { /* end of list */ } 1600 }; 1601 1602 static BlockDriver bdrv_qed = { 1603 .format_name = "qed", 1604 .instance_size = sizeof(BDRVQEDState), 1605 .create_options = qed_create_options, 1606 1607 .bdrv_probe = bdrv_qed_probe, 1608 .bdrv_rebind = bdrv_qed_rebind, 1609 .bdrv_open = bdrv_qed_open, 1610 .bdrv_close = bdrv_qed_close, 1611 .bdrv_reopen_prepare = bdrv_qed_reopen_prepare, 1612 .bdrv_create = bdrv_qed_create, 1613 .bdrv_has_zero_init = bdrv_has_zero_init_1, 1614 .bdrv_co_get_block_status = bdrv_qed_co_get_block_status, 1615 .bdrv_aio_readv = bdrv_qed_aio_readv, 1616 .bdrv_aio_writev = bdrv_qed_aio_writev, 1617 .bdrv_co_write_zeroes = bdrv_qed_co_write_zeroes, 1618 .bdrv_truncate = bdrv_qed_truncate, 1619 .bdrv_getlength = bdrv_qed_getlength, 1620 .bdrv_get_info = bdrv_qed_get_info, 1621 .bdrv_refresh_limits = bdrv_qed_refresh_limits, 1622 .bdrv_change_backing_file = bdrv_qed_change_backing_file, 1623 .bdrv_invalidate_cache = bdrv_qed_invalidate_cache, 1624 .bdrv_check = bdrv_qed_check, 1625 }; 1626 1627 static void bdrv_qed_init(void) 1628 { 1629 bdrv_register(&bdrv_qed); 1630 } 1631 1632 block_init(bdrv_qed_init); 1633