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 error_setg(errp, "Image not in QED format"); 395 return -EINVAL; 396 } 397 if (s->header.features & ~QED_FEATURE_MASK) { 398 /* image uses unsupported feature bits */ 399 char buf[64]; 400 snprintf(buf, sizeof(buf), "%" PRIx64, 401 s->header.features & ~QED_FEATURE_MASK); 402 error_set(errp, QERR_UNKNOWN_BLOCK_FORMAT_FEATURE, 403 bs->device_name, "QED", buf); 404 return -ENOTSUP; 405 } 406 if (!qed_is_cluster_size_valid(s->header.cluster_size)) { 407 return -EINVAL; 408 } 409 410 /* Round down file size to the last cluster */ 411 file_size = bdrv_getlength(bs->file); 412 if (file_size < 0) { 413 return file_size; 414 } 415 s->file_size = qed_start_of_cluster(s, file_size); 416 417 if (!qed_is_table_size_valid(s->header.table_size)) { 418 return -EINVAL; 419 } 420 if (!qed_is_image_size_valid(s->header.image_size, 421 s->header.cluster_size, 422 s->header.table_size)) { 423 return -EINVAL; 424 } 425 if (!qed_check_table_offset(s, s->header.l1_table_offset)) { 426 return -EINVAL; 427 } 428 429 s->table_nelems = (s->header.cluster_size * s->header.table_size) / 430 sizeof(uint64_t); 431 s->l2_shift = ffs(s->header.cluster_size) - 1; 432 s->l2_mask = s->table_nelems - 1; 433 s->l1_shift = s->l2_shift + ffs(s->table_nelems) - 1; 434 435 if ((s->header.features & QED_F_BACKING_FILE)) { 436 if ((uint64_t)s->header.backing_filename_offset + 437 s->header.backing_filename_size > 438 s->header.cluster_size * s->header.header_size) { 439 return -EINVAL; 440 } 441 442 ret = qed_read_string(bs->file, s->header.backing_filename_offset, 443 s->header.backing_filename_size, bs->backing_file, 444 sizeof(bs->backing_file)); 445 if (ret < 0) { 446 return ret; 447 } 448 449 if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) { 450 pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw"); 451 } 452 } 453 454 /* Reset unknown autoclear feature bits. This is a backwards 455 * compatibility mechanism that allows images to be opened by older 456 * programs, which "knock out" unknown feature bits. When an image is 457 * opened by a newer program again it can detect that the autoclear 458 * feature is no longer valid. 459 */ 460 if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 && 461 !bdrv_is_read_only(bs->file) && !(flags & BDRV_O_INCOMING)) { 462 s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK; 463 464 ret = qed_write_header_sync(s); 465 if (ret) { 466 return ret; 467 } 468 469 /* From here on only known autoclear feature bits are valid */ 470 bdrv_flush(bs->file); 471 } 472 473 s->l1_table = qed_alloc_table(s); 474 qed_init_l2_cache(&s->l2_cache); 475 476 ret = qed_read_l1_table_sync(s); 477 if (ret) { 478 goto out; 479 } 480 481 /* If image was not closed cleanly, check consistency */ 482 if (!(flags & BDRV_O_CHECK) && (s->header.features & QED_F_NEED_CHECK)) { 483 /* Read-only images cannot be fixed. There is no risk of corruption 484 * since write operations are not possible. Therefore, allow 485 * potentially inconsistent images to be opened read-only. This can 486 * aid data recovery from an otherwise inconsistent image. 487 */ 488 if (!bdrv_is_read_only(bs->file) && 489 !(flags & BDRV_O_INCOMING)) { 490 BdrvCheckResult result = {0}; 491 492 ret = qed_check(s, &result, true); 493 if (ret) { 494 goto out; 495 } 496 } 497 } 498 499 s->need_check_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL, 500 qed_need_check_timer_cb, s); 501 502 out: 503 if (ret) { 504 qed_free_l2_cache(&s->l2_cache); 505 qemu_vfree(s->l1_table); 506 } 507 return ret; 508 } 509 510 static int bdrv_qed_refresh_limits(BlockDriverState *bs) 511 { 512 BDRVQEDState *s = bs->opaque; 513 514 bs->bl.write_zeroes_alignment = s->header.cluster_size >> BDRV_SECTOR_BITS; 515 516 return 0; 517 } 518 519 /* We have nothing to do for QED reopen, stubs just return 520 * success */ 521 static int bdrv_qed_reopen_prepare(BDRVReopenState *state, 522 BlockReopenQueue *queue, Error **errp) 523 { 524 return 0; 525 } 526 527 static void bdrv_qed_close(BlockDriverState *bs) 528 { 529 BDRVQEDState *s = bs->opaque; 530 531 qed_cancel_need_check_timer(s); 532 timer_free(s->need_check_timer); 533 534 /* Ensure writes reach stable storage */ 535 bdrv_flush(bs->file); 536 537 /* Clean shutdown, no check required on next open */ 538 if (s->header.features & QED_F_NEED_CHECK) { 539 s->header.features &= ~QED_F_NEED_CHECK; 540 qed_write_header_sync(s); 541 } 542 543 qed_free_l2_cache(&s->l2_cache); 544 qemu_vfree(s->l1_table); 545 } 546 547 static int qed_create(const char *filename, uint32_t cluster_size, 548 uint64_t image_size, uint32_t table_size, 549 const char *backing_file, const char *backing_fmt, 550 Error **errp) 551 { 552 QEDHeader header = { 553 .magic = QED_MAGIC, 554 .cluster_size = cluster_size, 555 .table_size = table_size, 556 .header_size = 1, 557 .features = 0, 558 .compat_features = 0, 559 .l1_table_offset = cluster_size, 560 .image_size = image_size, 561 }; 562 QEDHeader le_header; 563 uint8_t *l1_table = NULL; 564 size_t l1_size = header.cluster_size * header.table_size; 565 Error *local_err = NULL; 566 int ret = 0; 567 BlockDriverState *bs; 568 569 ret = bdrv_create_file(filename, NULL, &local_err); 570 if (ret < 0) { 571 error_propagate(errp, local_err); 572 return ret; 573 } 574 575 bs = NULL; 576 ret = bdrv_open(&bs, filename, NULL, NULL, 577 BDRV_O_RDWR | BDRV_O_CACHE_WB | BDRV_O_PROTOCOL, NULL, 578 &local_err); 579 if (ret < 0) { 580 error_propagate(errp, local_err); 581 return ret; 582 } 583 584 /* File must start empty and grow, check truncate is supported */ 585 ret = bdrv_truncate(bs, 0); 586 if (ret < 0) { 587 goto out; 588 } 589 590 if (backing_file) { 591 header.features |= QED_F_BACKING_FILE; 592 header.backing_filename_offset = sizeof(le_header); 593 header.backing_filename_size = strlen(backing_file); 594 595 if (qed_fmt_is_raw(backing_fmt)) { 596 header.features |= QED_F_BACKING_FORMAT_NO_PROBE; 597 } 598 } 599 600 qed_header_cpu_to_le(&header, &le_header); 601 ret = bdrv_pwrite(bs, 0, &le_header, sizeof(le_header)); 602 if (ret < 0) { 603 goto out; 604 } 605 ret = bdrv_pwrite(bs, sizeof(le_header), backing_file, 606 header.backing_filename_size); 607 if (ret < 0) { 608 goto out; 609 } 610 611 l1_table = g_malloc0(l1_size); 612 ret = bdrv_pwrite(bs, header.l1_table_offset, l1_table, l1_size); 613 if (ret < 0) { 614 goto out; 615 } 616 617 ret = 0; /* success */ 618 out: 619 g_free(l1_table); 620 bdrv_unref(bs); 621 return ret; 622 } 623 624 static int bdrv_qed_create(const char *filename, QEMUOptionParameter *options, 625 Error **errp) 626 { 627 uint64_t image_size = 0; 628 uint32_t cluster_size = QED_DEFAULT_CLUSTER_SIZE; 629 uint32_t table_size = QED_DEFAULT_TABLE_SIZE; 630 const char *backing_file = NULL; 631 const char *backing_fmt = NULL; 632 633 while (options && options->name) { 634 if (!strcmp(options->name, BLOCK_OPT_SIZE)) { 635 image_size = options->value.n; 636 } else if (!strcmp(options->name, BLOCK_OPT_BACKING_FILE)) { 637 backing_file = options->value.s; 638 } else if (!strcmp(options->name, BLOCK_OPT_BACKING_FMT)) { 639 backing_fmt = options->value.s; 640 } else if (!strcmp(options->name, BLOCK_OPT_CLUSTER_SIZE)) { 641 if (options->value.n) { 642 cluster_size = options->value.n; 643 } 644 } else if (!strcmp(options->name, BLOCK_OPT_TABLE_SIZE)) { 645 if (options->value.n) { 646 table_size = options->value.n; 647 } 648 } 649 options++; 650 } 651 652 if (!qed_is_cluster_size_valid(cluster_size)) { 653 error_setg(errp, "QED cluster size must be within range [%u, %u] " 654 "and power of 2", 655 QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE); 656 return -EINVAL; 657 } 658 if (!qed_is_table_size_valid(table_size)) { 659 error_setg(errp, "QED table size must be within range [%u, %u] " 660 "and power of 2", 661 QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE); 662 return -EINVAL; 663 } 664 if (!qed_is_image_size_valid(image_size, cluster_size, table_size)) { 665 error_setg(errp, "QED image size must be a non-zero multiple of " 666 "cluster size and less than %" PRIu64 " bytes", 667 qed_max_image_size(cluster_size, table_size)); 668 return -EINVAL; 669 } 670 671 return qed_create(filename, cluster_size, image_size, table_size, 672 backing_file, backing_fmt, errp); 673 } 674 675 typedef struct { 676 BlockDriverState *bs; 677 Coroutine *co; 678 uint64_t pos; 679 int64_t status; 680 int *pnum; 681 } QEDIsAllocatedCB; 682 683 static void qed_is_allocated_cb(void *opaque, int ret, uint64_t offset, size_t len) 684 { 685 QEDIsAllocatedCB *cb = opaque; 686 BDRVQEDState *s = cb->bs->opaque; 687 *cb->pnum = len / BDRV_SECTOR_SIZE; 688 switch (ret) { 689 case QED_CLUSTER_FOUND: 690 offset |= qed_offset_into_cluster(s, cb->pos); 691 cb->status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID | offset; 692 break; 693 case QED_CLUSTER_ZERO: 694 cb->status = BDRV_BLOCK_ZERO; 695 break; 696 case QED_CLUSTER_L2: 697 case QED_CLUSTER_L1: 698 cb->status = 0; 699 break; 700 default: 701 assert(ret < 0); 702 cb->status = ret; 703 break; 704 } 705 706 if (cb->co) { 707 qemu_coroutine_enter(cb->co, NULL); 708 } 709 } 710 711 static int64_t coroutine_fn bdrv_qed_co_get_block_status(BlockDriverState *bs, 712 int64_t sector_num, 713 int nb_sectors, int *pnum) 714 { 715 BDRVQEDState *s = bs->opaque; 716 size_t len = (size_t)nb_sectors * BDRV_SECTOR_SIZE; 717 QEDIsAllocatedCB cb = { 718 .bs = bs, 719 .pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE, 720 .status = BDRV_BLOCK_OFFSET_MASK, 721 .pnum = pnum, 722 }; 723 QEDRequest request = { .l2_table = NULL }; 724 725 qed_find_cluster(s, &request, cb.pos, len, qed_is_allocated_cb, &cb); 726 727 /* Now sleep if the callback wasn't invoked immediately */ 728 while (cb.status == BDRV_BLOCK_OFFSET_MASK) { 729 cb.co = qemu_coroutine_self(); 730 qemu_coroutine_yield(); 731 } 732 733 qed_unref_l2_cache_entry(request.l2_table); 734 735 return cb.status; 736 } 737 738 static BDRVQEDState *acb_to_s(QEDAIOCB *acb) 739 { 740 return acb->common.bs->opaque; 741 } 742 743 /** 744 * Read from the backing file or zero-fill if no backing file 745 * 746 * @s: QED state 747 * @pos: Byte position in device 748 * @qiov: Destination I/O vector 749 * @cb: Completion function 750 * @opaque: User data for completion function 751 * 752 * This function reads qiov->size bytes starting at pos from the backing file. 753 * If there is no backing file then zeroes are read. 754 */ 755 static void qed_read_backing_file(BDRVQEDState *s, uint64_t pos, 756 QEMUIOVector *qiov, 757 BlockDriverCompletionFunc *cb, void *opaque) 758 { 759 uint64_t backing_length = 0; 760 size_t size; 761 762 /* If there is a backing file, get its length. Treat the absence of a 763 * backing file like a zero length backing file. 764 */ 765 if (s->bs->backing_hd) { 766 int64_t l = bdrv_getlength(s->bs->backing_hd); 767 if (l < 0) { 768 cb(opaque, l); 769 return; 770 } 771 backing_length = l; 772 } 773 774 /* Zero all sectors if reading beyond the end of the backing file */ 775 if (pos >= backing_length || 776 pos + qiov->size > backing_length) { 777 qemu_iovec_memset(qiov, 0, 0, qiov->size); 778 } 779 780 /* Complete now if there are no backing file sectors to read */ 781 if (pos >= backing_length) { 782 cb(opaque, 0); 783 return; 784 } 785 786 /* If the read straddles the end of the backing file, shorten it */ 787 size = MIN((uint64_t)backing_length - pos, qiov->size); 788 789 BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO); 790 bdrv_aio_readv(s->bs->backing_hd, pos / BDRV_SECTOR_SIZE, 791 qiov, size / BDRV_SECTOR_SIZE, cb, opaque); 792 } 793 794 typedef struct { 795 GenericCB gencb; 796 BDRVQEDState *s; 797 QEMUIOVector qiov; 798 struct iovec iov; 799 uint64_t offset; 800 } CopyFromBackingFileCB; 801 802 static void qed_copy_from_backing_file_cb(void *opaque, int ret) 803 { 804 CopyFromBackingFileCB *copy_cb = opaque; 805 qemu_vfree(copy_cb->iov.iov_base); 806 gencb_complete(©_cb->gencb, ret); 807 } 808 809 static void qed_copy_from_backing_file_write(void *opaque, int ret) 810 { 811 CopyFromBackingFileCB *copy_cb = opaque; 812 BDRVQEDState *s = copy_cb->s; 813 814 if (ret) { 815 qed_copy_from_backing_file_cb(copy_cb, ret); 816 return; 817 } 818 819 BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE); 820 bdrv_aio_writev(s->bs->file, copy_cb->offset / BDRV_SECTOR_SIZE, 821 ©_cb->qiov, copy_cb->qiov.size / BDRV_SECTOR_SIZE, 822 qed_copy_from_backing_file_cb, copy_cb); 823 } 824 825 /** 826 * Copy data from backing file into the image 827 * 828 * @s: QED state 829 * @pos: Byte position in device 830 * @len: Number of bytes 831 * @offset: Byte offset in image file 832 * @cb: Completion function 833 * @opaque: User data for completion function 834 */ 835 static void qed_copy_from_backing_file(BDRVQEDState *s, uint64_t pos, 836 uint64_t len, uint64_t offset, 837 BlockDriverCompletionFunc *cb, 838 void *opaque) 839 { 840 CopyFromBackingFileCB *copy_cb; 841 842 /* Skip copy entirely if there is no work to do */ 843 if (len == 0) { 844 cb(opaque, 0); 845 return; 846 } 847 848 copy_cb = gencb_alloc(sizeof(*copy_cb), cb, opaque); 849 copy_cb->s = s; 850 copy_cb->offset = offset; 851 copy_cb->iov.iov_base = qemu_blockalign(s->bs, len); 852 copy_cb->iov.iov_len = len; 853 qemu_iovec_init_external(©_cb->qiov, ©_cb->iov, 1); 854 855 qed_read_backing_file(s, pos, ©_cb->qiov, 856 qed_copy_from_backing_file_write, copy_cb); 857 } 858 859 /** 860 * Link one or more contiguous clusters into a table 861 * 862 * @s: QED state 863 * @table: L2 table 864 * @index: First cluster index 865 * @n: Number of contiguous clusters 866 * @cluster: First cluster offset 867 * 868 * The cluster offset may be an allocated byte offset in the image file, the 869 * zero cluster marker, or the unallocated cluster marker. 870 */ 871 static void qed_update_l2_table(BDRVQEDState *s, QEDTable *table, int index, 872 unsigned int n, uint64_t cluster) 873 { 874 int i; 875 for (i = index; i < index + n; i++) { 876 table->offsets[i] = cluster; 877 if (!qed_offset_is_unalloc_cluster(cluster) && 878 !qed_offset_is_zero_cluster(cluster)) { 879 cluster += s->header.cluster_size; 880 } 881 } 882 } 883 884 static void qed_aio_complete_bh(void *opaque) 885 { 886 QEDAIOCB *acb = opaque; 887 BlockDriverCompletionFunc *cb = acb->common.cb; 888 void *user_opaque = acb->common.opaque; 889 int ret = acb->bh_ret; 890 bool *finished = acb->finished; 891 892 qemu_bh_delete(acb->bh); 893 qemu_aio_release(acb); 894 895 /* Invoke callback */ 896 cb(user_opaque, ret); 897 898 /* Signal cancel completion */ 899 if (finished) { 900 *finished = true; 901 } 902 } 903 904 static void qed_aio_complete(QEDAIOCB *acb, int ret) 905 { 906 BDRVQEDState *s = acb_to_s(acb); 907 908 trace_qed_aio_complete(s, acb, ret); 909 910 /* Free resources */ 911 qemu_iovec_destroy(&acb->cur_qiov); 912 qed_unref_l2_cache_entry(acb->request.l2_table); 913 914 /* Free the buffer we may have allocated for zero writes */ 915 if (acb->flags & QED_AIOCB_ZERO) { 916 qemu_vfree(acb->qiov->iov[0].iov_base); 917 acb->qiov->iov[0].iov_base = NULL; 918 } 919 920 /* Arrange for a bh to invoke the completion function */ 921 acb->bh_ret = ret; 922 acb->bh = qemu_bh_new(qed_aio_complete_bh, acb); 923 qemu_bh_schedule(acb->bh); 924 925 /* Start next allocating write request waiting behind this one. Note that 926 * requests enqueue themselves when they first hit an unallocated cluster 927 * but they wait until the entire request is finished before waking up the 928 * next request in the queue. This ensures that we don't cycle through 929 * requests multiple times but rather finish one at a time completely. 930 */ 931 if (acb == QSIMPLEQ_FIRST(&s->allocating_write_reqs)) { 932 QSIMPLEQ_REMOVE_HEAD(&s->allocating_write_reqs, next); 933 acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs); 934 if (acb) { 935 qed_aio_next_io(acb, 0); 936 } else if (s->header.features & QED_F_NEED_CHECK) { 937 qed_start_need_check_timer(s); 938 } 939 } 940 } 941 942 /** 943 * Commit the current L2 table to the cache 944 */ 945 static void qed_commit_l2_update(void *opaque, int ret) 946 { 947 QEDAIOCB *acb = opaque; 948 BDRVQEDState *s = acb_to_s(acb); 949 CachedL2Table *l2_table = acb->request.l2_table; 950 uint64_t l2_offset = l2_table->offset; 951 952 qed_commit_l2_cache_entry(&s->l2_cache, l2_table); 953 954 /* This is guaranteed to succeed because we just committed the entry to the 955 * cache. 956 */ 957 acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset); 958 assert(acb->request.l2_table != NULL); 959 960 qed_aio_next_io(opaque, ret); 961 } 962 963 /** 964 * Update L1 table with new L2 table offset and write it out 965 */ 966 static void qed_aio_write_l1_update(void *opaque, int ret) 967 { 968 QEDAIOCB *acb = opaque; 969 BDRVQEDState *s = acb_to_s(acb); 970 int index; 971 972 if (ret) { 973 qed_aio_complete(acb, ret); 974 return; 975 } 976 977 index = qed_l1_index(s, acb->cur_pos); 978 s->l1_table->offsets[index] = acb->request.l2_table->offset; 979 980 qed_write_l1_table(s, index, 1, qed_commit_l2_update, acb); 981 } 982 983 /** 984 * Update L2 table with new cluster offsets and write them out 985 */ 986 static void qed_aio_write_l2_update(QEDAIOCB *acb, int ret, uint64_t offset) 987 { 988 BDRVQEDState *s = acb_to_s(acb); 989 bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1; 990 int index; 991 992 if (ret) { 993 goto err; 994 } 995 996 if (need_alloc) { 997 qed_unref_l2_cache_entry(acb->request.l2_table); 998 acb->request.l2_table = qed_new_l2_table(s); 999 } 1000 1001 index = qed_l2_index(s, acb->cur_pos); 1002 qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters, 1003 offset); 1004 1005 if (need_alloc) { 1006 /* Write out the whole new L2 table */ 1007 qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true, 1008 qed_aio_write_l1_update, acb); 1009 } else { 1010 /* Write out only the updated part of the L2 table */ 1011 qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters, false, 1012 qed_aio_next_io, acb); 1013 } 1014 return; 1015 1016 err: 1017 qed_aio_complete(acb, ret); 1018 } 1019 1020 static void qed_aio_write_l2_update_cb(void *opaque, int ret) 1021 { 1022 QEDAIOCB *acb = opaque; 1023 qed_aio_write_l2_update(acb, ret, acb->cur_cluster); 1024 } 1025 1026 /** 1027 * Flush new data clusters before updating the L2 table 1028 * 1029 * This flush is necessary when a backing file is in use. A crash during an 1030 * allocating write could result in empty clusters in the image. If the write 1031 * only touched a subregion of the cluster, then backing image sectors have 1032 * been lost in the untouched region. The solution is to flush after writing a 1033 * new data cluster and before updating the L2 table. 1034 */ 1035 static void qed_aio_write_flush_before_l2_update(void *opaque, int ret) 1036 { 1037 QEDAIOCB *acb = opaque; 1038 BDRVQEDState *s = acb_to_s(acb); 1039 1040 if (!bdrv_aio_flush(s->bs->file, qed_aio_write_l2_update_cb, opaque)) { 1041 qed_aio_complete(acb, -EIO); 1042 } 1043 } 1044 1045 /** 1046 * Write data to the image file 1047 */ 1048 static void qed_aio_write_main(void *opaque, int ret) 1049 { 1050 QEDAIOCB *acb = opaque; 1051 BDRVQEDState *s = acb_to_s(acb); 1052 uint64_t offset = acb->cur_cluster + 1053 qed_offset_into_cluster(s, acb->cur_pos); 1054 BlockDriverCompletionFunc *next_fn; 1055 1056 trace_qed_aio_write_main(s, acb, ret, offset, acb->cur_qiov.size); 1057 1058 if (ret) { 1059 qed_aio_complete(acb, ret); 1060 return; 1061 } 1062 1063 if (acb->find_cluster_ret == QED_CLUSTER_FOUND) { 1064 next_fn = qed_aio_next_io; 1065 } else { 1066 if (s->bs->backing_hd) { 1067 next_fn = qed_aio_write_flush_before_l2_update; 1068 } else { 1069 next_fn = qed_aio_write_l2_update_cb; 1070 } 1071 } 1072 1073 BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO); 1074 bdrv_aio_writev(s->bs->file, offset / BDRV_SECTOR_SIZE, 1075 &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE, 1076 next_fn, acb); 1077 } 1078 1079 /** 1080 * Populate back untouched region of new data cluster 1081 */ 1082 static void qed_aio_write_postfill(void *opaque, int ret) 1083 { 1084 QEDAIOCB *acb = opaque; 1085 BDRVQEDState *s = acb_to_s(acb); 1086 uint64_t start = acb->cur_pos + acb->cur_qiov.size; 1087 uint64_t len = 1088 qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start; 1089 uint64_t offset = acb->cur_cluster + 1090 qed_offset_into_cluster(s, acb->cur_pos) + 1091 acb->cur_qiov.size; 1092 1093 if (ret) { 1094 qed_aio_complete(acb, ret); 1095 return; 1096 } 1097 1098 trace_qed_aio_write_postfill(s, acb, start, len, offset); 1099 qed_copy_from_backing_file(s, start, len, offset, 1100 qed_aio_write_main, acb); 1101 } 1102 1103 /** 1104 * Populate front untouched region of new data cluster 1105 */ 1106 static void qed_aio_write_prefill(void *opaque, int ret) 1107 { 1108 QEDAIOCB *acb = opaque; 1109 BDRVQEDState *s = acb_to_s(acb); 1110 uint64_t start = qed_start_of_cluster(s, acb->cur_pos); 1111 uint64_t len = qed_offset_into_cluster(s, acb->cur_pos); 1112 1113 trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster); 1114 qed_copy_from_backing_file(s, start, len, acb->cur_cluster, 1115 qed_aio_write_postfill, acb); 1116 } 1117 1118 /** 1119 * Check if the QED_F_NEED_CHECK bit should be set during allocating write 1120 */ 1121 static bool qed_should_set_need_check(BDRVQEDState *s) 1122 { 1123 /* The flush before L2 update path ensures consistency */ 1124 if (s->bs->backing_hd) { 1125 return false; 1126 } 1127 1128 return !(s->header.features & QED_F_NEED_CHECK); 1129 } 1130 1131 static void qed_aio_write_zero_cluster(void *opaque, int ret) 1132 { 1133 QEDAIOCB *acb = opaque; 1134 1135 if (ret) { 1136 qed_aio_complete(acb, ret); 1137 return; 1138 } 1139 1140 qed_aio_write_l2_update(acb, 0, 1); 1141 } 1142 1143 /** 1144 * Write new data cluster 1145 * 1146 * @acb: Write request 1147 * @len: Length in bytes 1148 * 1149 * This path is taken when writing to previously unallocated clusters. 1150 */ 1151 static void qed_aio_write_alloc(QEDAIOCB *acb, size_t len) 1152 { 1153 BDRVQEDState *s = acb_to_s(acb); 1154 BlockDriverCompletionFunc *cb; 1155 1156 /* Cancel timer when the first allocating request comes in */ 1157 if (QSIMPLEQ_EMPTY(&s->allocating_write_reqs)) { 1158 qed_cancel_need_check_timer(s); 1159 } 1160 1161 /* Freeze this request if another allocating write is in progress */ 1162 if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs)) { 1163 QSIMPLEQ_INSERT_TAIL(&s->allocating_write_reqs, acb, next); 1164 } 1165 if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs) || 1166 s->allocating_write_reqs_plugged) { 1167 return; /* wait for existing request to finish */ 1168 } 1169 1170 acb->cur_nclusters = qed_bytes_to_clusters(s, 1171 qed_offset_into_cluster(s, acb->cur_pos) + len); 1172 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len); 1173 1174 if (acb->flags & QED_AIOCB_ZERO) { 1175 /* Skip ahead if the clusters are already zero */ 1176 if (acb->find_cluster_ret == QED_CLUSTER_ZERO) { 1177 qed_aio_next_io(acb, 0); 1178 return; 1179 } 1180 1181 cb = qed_aio_write_zero_cluster; 1182 } else { 1183 cb = qed_aio_write_prefill; 1184 acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters); 1185 } 1186 1187 if (qed_should_set_need_check(s)) { 1188 s->header.features |= QED_F_NEED_CHECK; 1189 qed_write_header(s, cb, acb); 1190 } else { 1191 cb(acb, 0); 1192 } 1193 } 1194 1195 /** 1196 * Write data cluster in place 1197 * 1198 * @acb: Write request 1199 * @offset: Cluster offset in bytes 1200 * @len: Length in bytes 1201 * 1202 * This path is taken when writing to already allocated clusters. 1203 */ 1204 static void qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset, size_t len) 1205 { 1206 /* Allocate buffer for zero writes */ 1207 if (acb->flags & QED_AIOCB_ZERO) { 1208 struct iovec *iov = acb->qiov->iov; 1209 1210 if (!iov->iov_base) { 1211 iov->iov_base = qemu_blockalign(acb->common.bs, iov->iov_len); 1212 memset(iov->iov_base, 0, iov->iov_len); 1213 } 1214 } 1215 1216 /* Calculate the I/O vector */ 1217 acb->cur_cluster = offset; 1218 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len); 1219 1220 /* Do the actual write */ 1221 qed_aio_write_main(acb, 0); 1222 } 1223 1224 /** 1225 * Write data cluster 1226 * 1227 * @opaque: Write request 1228 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1, 1229 * or -errno 1230 * @offset: Cluster offset in bytes 1231 * @len: Length in bytes 1232 * 1233 * Callback from qed_find_cluster(). 1234 */ 1235 static void qed_aio_write_data(void *opaque, int ret, 1236 uint64_t offset, size_t len) 1237 { 1238 QEDAIOCB *acb = opaque; 1239 1240 trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len); 1241 1242 acb->find_cluster_ret = ret; 1243 1244 switch (ret) { 1245 case QED_CLUSTER_FOUND: 1246 qed_aio_write_inplace(acb, offset, len); 1247 break; 1248 1249 case QED_CLUSTER_L2: 1250 case QED_CLUSTER_L1: 1251 case QED_CLUSTER_ZERO: 1252 qed_aio_write_alloc(acb, len); 1253 break; 1254 1255 default: 1256 qed_aio_complete(acb, ret); 1257 break; 1258 } 1259 } 1260 1261 /** 1262 * Read data cluster 1263 * 1264 * @opaque: Read request 1265 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1, 1266 * or -errno 1267 * @offset: Cluster offset in bytes 1268 * @len: Length in bytes 1269 * 1270 * Callback from qed_find_cluster(). 1271 */ 1272 static void qed_aio_read_data(void *opaque, int ret, 1273 uint64_t offset, size_t len) 1274 { 1275 QEDAIOCB *acb = opaque; 1276 BDRVQEDState *s = acb_to_s(acb); 1277 BlockDriverState *bs = acb->common.bs; 1278 1279 /* Adjust offset into cluster */ 1280 offset += qed_offset_into_cluster(s, acb->cur_pos); 1281 1282 trace_qed_aio_read_data(s, acb, ret, offset, len); 1283 1284 if (ret < 0) { 1285 goto err; 1286 } 1287 1288 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len); 1289 1290 /* Handle zero cluster and backing file reads */ 1291 if (ret == QED_CLUSTER_ZERO) { 1292 qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size); 1293 qed_aio_next_io(acb, 0); 1294 return; 1295 } else if (ret != QED_CLUSTER_FOUND) { 1296 qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov, 1297 qed_aio_next_io, acb); 1298 return; 1299 } 1300 1301 BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO); 1302 bdrv_aio_readv(bs->file, offset / BDRV_SECTOR_SIZE, 1303 &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE, 1304 qed_aio_next_io, acb); 1305 return; 1306 1307 err: 1308 qed_aio_complete(acb, ret); 1309 } 1310 1311 /** 1312 * Begin next I/O or complete the request 1313 */ 1314 static void qed_aio_next_io(void *opaque, int ret) 1315 { 1316 QEDAIOCB *acb = opaque; 1317 BDRVQEDState *s = acb_to_s(acb); 1318 QEDFindClusterFunc *io_fn = (acb->flags & QED_AIOCB_WRITE) ? 1319 qed_aio_write_data : qed_aio_read_data; 1320 1321 trace_qed_aio_next_io(s, acb, ret, acb->cur_pos + acb->cur_qiov.size); 1322 1323 /* Handle I/O error */ 1324 if (ret) { 1325 qed_aio_complete(acb, ret); 1326 return; 1327 } 1328 1329 acb->qiov_offset += acb->cur_qiov.size; 1330 acb->cur_pos += acb->cur_qiov.size; 1331 qemu_iovec_reset(&acb->cur_qiov); 1332 1333 /* Complete request */ 1334 if (acb->cur_pos >= acb->end_pos) { 1335 qed_aio_complete(acb, 0); 1336 return; 1337 } 1338 1339 /* Find next cluster and start I/O */ 1340 qed_find_cluster(s, &acb->request, 1341 acb->cur_pos, acb->end_pos - acb->cur_pos, 1342 io_fn, acb); 1343 } 1344 1345 static BlockDriverAIOCB *qed_aio_setup(BlockDriverState *bs, 1346 int64_t sector_num, 1347 QEMUIOVector *qiov, int nb_sectors, 1348 BlockDriverCompletionFunc *cb, 1349 void *opaque, int flags) 1350 { 1351 QEDAIOCB *acb = qemu_aio_get(&qed_aiocb_info, bs, cb, opaque); 1352 1353 trace_qed_aio_setup(bs->opaque, acb, sector_num, nb_sectors, 1354 opaque, flags); 1355 1356 acb->flags = flags; 1357 acb->finished = NULL; 1358 acb->qiov = qiov; 1359 acb->qiov_offset = 0; 1360 acb->cur_pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE; 1361 acb->end_pos = acb->cur_pos + nb_sectors * BDRV_SECTOR_SIZE; 1362 acb->request.l2_table = NULL; 1363 qemu_iovec_init(&acb->cur_qiov, qiov->niov); 1364 1365 /* Start request */ 1366 qed_aio_next_io(acb, 0); 1367 return &acb->common; 1368 } 1369 1370 static BlockDriverAIOCB *bdrv_qed_aio_readv(BlockDriverState *bs, 1371 int64_t sector_num, 1372 QEMUIOVector *qiov, int nb_sectors, 1373 BlockDriverCompletionFunc *cb, 1374 void *opaque) 1375 { 1376 return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb, opaque, 0); 1377 } 1378 1379 static BlockDriverAIOCB *bdrv_qed_aio_writev(BlockDriverState *bs, 1380 int64_t sector_num, 1381 QEMUIOVector *qiov, int nb_sectors, 1382 BlockDriverCompletionFunc *cb, 1383 void *opaque) 1384 { 1385 return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb, 1386 opaque, QED_AIOCB_WRITE); 1387 } 1388 1389 typedef struct { 1390 Coroutine *co; 1391 int ret; 1392 bool done; 1393 } QEDWriteZeroesCB; 1394 1395 static void coroutine_fn qed_co_write_zeroes_cb(void *opaque, int ret) 1396 { 1397 QEDWriteZeroesCB *cb = opaque; 1398 1399 cb->done = true; 1400 cb->ret = ret; 1401 if (cb->co) { 1402 qemu_coroutine_enter(cb->co, NULL); 1403 } 1404 } 1405 1406 static int coroutine_fn bdrv_qed_co_write_zeroes(BlockDriverState *bs, 1407 int64_t sector_num, 1408 int nb_sectors, 1409 BdrvRequestFlags flags) 1410 { 1411 BlockDriverAIOCB *blockacb; 1412 BDRVQEDState *s = bs->opaque; 1413 QEDWriteZeroesCB cb = { .done = false }; 1414 QEMUIOVector qiov; 1415 struct iovec iov; 1416 1417 /* Refuse if there are untouched backing file sectors */ 1418 if (bs->backing_hd) { 1419 if (qed_offset_into_cluster(s, sector_num * BDRV_SECTOR_SIZE) != 0) { 1420 return -ENOTSUP; 1421 } 1422 if (qed_offset_into_cluster(s, nb_sectors * BDRV_SECTOR_SIZE) != 0) { 1423 return -ENOTSUP; 1424 } 1425 } 1426 1427 /* Zero writes start without an I/O buffer. If a buffer becomes necessary 1428 * then it will be allocated during request processing. 1429 */ 1430 iov.iov_base = NULL, 1431 iov.iov_len = nb_sectors * BDRV_SECTOR_SIZE, 1432 1433 qemu_iovec_init_external(&qiov, &iov, 1); 1434 blockacb = qed_aio_setup(bs, sector_num, &qiov, nb_sectors, 1435 qed_co_write_zeroes_cb, &cb, 1436 QED_AIOCB_WRITE | QED_AIOCB_ZERO); 1437 if (!blockacb) { 1438 return -EIO; 1439 } 1440 if (!cb.done) { 1441 cb.co = qemu_coroutine_self(); 1442 qemu_coroutine_yield(); 1443 } 1444 assert(cb.done); 1445 return cb.ret; 1446 } 1447 1448 static int bdrv_qed_truncate(BlockDriverState *bs, int64_t offset) 1449 { 1450 BDRVQEDState *s = bs->opaque; 1451 uint64_t old_image_size; 1452 int ret; 1453 1454 if (!qed_is_image_size_valid(offset, s->header.cluster_size, 1455 s->header.table_size)) { 1456 return -EINVAL; 1457 } 1458 1459 /* Shrinking is currently not supported */ 1460 if ((uint64_t)offset < s->header.image_size) { 1461 return -ENOTSUP; 1462 } 1463 1464 old_image_size = s->header.image_size; 1465 s->header.image_size = offset; 1466 ret = qed_write_header_sync(s); 1467 if (ret < 0) { 1468 s->header.image_size = old_image_size; 1469 } 1470 return ret; 1471 } 1472 1473 static int64_t bdrv_qed_getlength(BlockDriverState *bs) 1474 { 1475 BDRVQEDState *s = bs->opaque; 1476 return s->header.image_size; 1477 } 1478 1479 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi) 1480 { 1481 BDRVQEDState *s = bs->opaque; 1482 1483 memset(bdi, 0, sizeof(*bdi)); 1484 bdi->cluster_size = s->header.cluster_size; 1485 bdi->is_dirty = s->header.features & QED_F_NEED_CHECK; 1486 bdi->unallocated_blocks_are_zero = true; 1487 bdi->can_write_zeroes_with_unmap = true; 1488 return 0; 1489 } 1490 1491 static int bdrv_qed_change_backing_file(BlockDriverState *bs, 1492 const char *backing_file, 1493 const char *backing_fmt) 1494 { 1495 BDRVQEDState *s = bs->opaque; 1496 QEDHeader new_header, le_header; 1497 void *buffer; 1498 size_t buffer_len, backing_file_len; 1499 int ret; 1500 1501 /* Refuse to set backing filename if unknown compat feature bits are 1502 * active. If the image uses an unknown compat feature then we may not 1503 * know the layout of data following the header structure and cannot safely 1504 * add a new string. 1505 */ 1506 if (backing_file && (s->header.compat_features & 1507 ~QED_COMPAT_FEATURE_MASK)) { 1508 return -ENOTSUP; 1509 } 1510 1511 memcpy(&new_header, &s->header, sizeof(new_header)); 1512 1513 new_header.features &= ~(QED_F_BACKING_FILE | 1514 QED_F_BACKING_FORMAT_NO_PROBE); 1515 1516 /* Adjust feature flags */ 1517 if (backing_file) { 1518 new_header.features |= QED_F_BACKING_FILE; 1519 1520 if (qed_fmt_is_raw(backing_fmt)) { 1521 new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE; 1522 } 1523 } 1524 1525 /* Calculate new header size */ 1526 backing_file_len = 0; 1527 1528 if (backing_file) { 1529 backing_file_len = strlen(backing_file); 1530 } 1531 1532 buffer_len = sizeof(new_header); 1533 new_header.backing_filename_offset = buffer_len; 1534 new_header.backing_filename_size = backing_file_len; 1535 buffer_len += backing_file_len; 1536 1537 /* Make sure we can rewrite header without failing */ 1538 if (buffer_len > new_header.header_size * new_header.cluster_size) { 1539 return -ENOSPC; 1540 } 1541 1542 /* Prepare new header */ 1543 buffer = g_malloc(buffer_len); 1544 1545 qed_header_cpu_to_le(&new_header, &le_header); 1546 memcpy(buffer, &le_header, sizeof(le_header)); 1547 buffer_len = sizeof(le_header); 1548 1549 if (backing_file) { 1550 memcpy(buffer + buffer_len, backing_file, backing_file_len); 1551 buffer_len += backing_file_len; 1552 } 1553 1554 /* Write new header */ 1555 ret = bdrv_pwrite_sync(bs->file, 0, buffer, buffer_len); 1556 g_free(buffer); 1557 if (ret == 0) { 1558 memcpy(&s->header, &new_header, sizeof(new_header)); 1559 } 1560 return ret; 1561 } 1562 1563 static void bdrv_qed_invalidate_cache(BlockDriverState *bs, Error **errp) 1564 { 1565 BDRVQEDState *s = bs->opaque; 1566 Error *local_err = NULL; 1567 int ret; 1568 1569 bdrv_qed_close(bs); 1570 1571 bdrv_invalidate_cache(bs->file, &local_err); 1572 if (local_err) { 1573 error_propagate(errp, local_err); 1574 return; 1575 } 1576 1577 memset(s, 0, sizeof(BDRVQEDState)); 1578 ret = bdrv_qed_open(bs, NULL, bs->open_flags, &local_err); 1579 if (local_err) { 1580 error_setg(errp, "Could not reopen qed layer: %s", 1581 error_get_pretty(local_err)); 1582 error_free(local_err); 1583 return; 1584 } else if (ret < 0) { 1585 error_setg_errno(errp, -ret, "Could not reopen qed layer"); 1586 return; 1587 } 1588 } 1589 1590 static int bdrv_qed_check(BlockDriverState *bs, BdrvCheckResult *result, 1591 BdrvCheckMode fix) 1592 { 1593 BDRVQEDState *s = bs->opaque; 1594 1595 return qed_check(s, result, !!fix); 1596 } 1597 1598 static QEMUOptionParameter qed_create_options[] = { 1599 { 1600 .name = BLOCK_OPT_SIZE, 1601 .type = OPT_SIZE, 1602 .help = "Virtual disk size (in bytes)" 1603 }, { 1604 .name = BLOCK_OPT_BACKING_FILE, 1605 .type = OPT_STRING, 1606 .help = "File name of a base image" 1607 }, { 1608 .name = BLOCK_OPT_BACKING_FMT, 1609 .type = OPT_STRING, 1610 .help = "Image format of the base image" 1611 }, { 1612 .name = BLOCK_OPT_CLUSTER_SIZE, 1613 .type = OPT_SIZE, 1614 .help = "Cluster size (in bytes)", 1615 .value = { .n = QED_DEFAULT_CLUSTER_SIZE }, 1616 }, { 1617 .name = BLOCK_OPT_TABLE_SIZE, 1618 .type = OPT_SIZE, 1619 .help = "L1/L2 table size (in clusters)" 1620 }, 1621 { /* end of list */ } 1622 }; 1623 1624 static BlockDriver bdrv_qed = { 1625 .format_name = "qed", 1626 .instance_size = sizeof(BDRVQEDState), 1627 .create_options = qed_create_options, 1628 1629 .bdrv_probe = bdrv_qed_probe, 1630 .bdrv_rebind = bdrv_qed_rebind, 1631 .bdrv_open = bdrv_qed_open, 1632 .bdrv_close = bdrv_qed_close, 1633 .bdrv_reopen_prepare = bdrv_qed_reopen_prepare, 1634 .bdrv_create = bdrv_qed_create, 1635 .bdrv_has_zero_init = bdrv_has_zero_init_1, 1636 .bdrv_co_get_block_status = bdrv_qed_co_get_block_status, 1637 .bdrv_aio_readv = bdrv_qed_aio_readv, 1638 .bdrv_aio_writev = bdrv_qed_aio_writev, 1639 .bdrv_co_write_zeroes = bdrv_qed_co_write_zeroes, 1640 .bdrv_truncate = bdrv_qed_truncate, 1641 .bdrv_getlength = bdrv_qed_getlength, 1642 .bdrv_get_info = bdrv_qed_get_info, 1643 .bdrv_refresh_limits = bdrv_qed_refresh_limits, 1644 .bdrv_change_backing_file = bdrv_qed_change_backing_file, 1645 .bdrv_invalidate_cache = bdrv_qed_invalidate_cache, 1646 .bdrv_check = bdrv_qed_check, 1647 }; 1648 1649 static void bdrv_qed_init(void) 1650 { 1651 bdrv_register(&bdrv_qed); 1652 } 1653 1654 block_init(bdrv_qed_init); 1655