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/osdep.h" 16 #include "qemu/timer.h" 17 #include "trace.h" 18 #include "qed.h" 19 #include "qapi/qmp/qerror.h" 20 #include "migration/migration.h" 21 22 static const AIOCBInfo qed_aiocb_info = { 23 .aiocb_size = sizeof(QEDAIOCB), 24 }; 25 26 static int bdrv_qed_probe(const uint8_t *buf, int buf_size, 27 const char *filename) 28 { 29 const QEDHeader *header = (const QEDHeader *)buf; 30 31 if (buf_size < sizeof(*header)) { 32 return 0; 33 } 34 if (le32_to_cpu(header->magic) != QED_MAGIC) { 35 return 0; 36 } 37 return 100; 38 } 39 40 /** 41 * Check whether an image format is raw 42 * 43 * @fmt: Backing file format, may be NULL 44 */ 45 static bool qed_fmt_is_raw(const char *fmt) 46 { 47 return fmt && strcmp(fmt, "raw") == 0; 48 } 49 50 static void qed_header_le_to_cpu(const QEDHeader *le, QEDHeader *cpu) 51 { 52 cpu->magic = le32_to_cpu(le->magic); 53 cpu->cluster_size = le32_to_cpu(le->cluster_size); 54 cpu->table_size = le32_to_cpu(le->table_size); 55 cpu->header_size = le32_to_cpu(le->header_size); 56 cpu->features = le64_to_cpu(le->features); 57 cpu->compat_features = le64_to_cpu(le->compat_features); 58 cpu->autoclear_features = le64_to_cpu(le->autoclear_features); 59 cpu->l1_table_offset = le64_to_cpu(le->l1_table_offset); 60 cpu->image_size = le64_to_cpu(le->image_size); 61 cpu->backing_filename_offset = le32_to_cpu(le->backing_filename_offset); 62 cpu->backing_filename_size = le32_to_cpu(le->backing_filename_size); 63 } 64 65 static void qed_header_cpu_to_le(const QEDHeader *cpu, QEDHeader *le) 66 { 67 le->magic = cpu_to_le32(cpu->magic); 68 le->cluster_size = cpu_to_le32(cpu->cluster_size); 69 le->table_size = cpu_to_le32(cpu->table_size); 70 le->header_size = cpu_to_le32(cpu->header_size); 71 le->features = cpu_to_le64(cpu->features); 72 le->compat_features = cpu_to_le64(cpu->compat_features); 73 le->autoclear_features = cpu_to_le64(cpu->autoclear_features); 74 le->l1_table_offset = cpu_to_le64(cpu->l1_table_offset); 75 le->image_size = cpu_to_le64(cpu->image_size); 76 le->backing_filename_offset = cpu_to_le32(cpu->backing_filename_offset); 77 le->backing_filename_size = cpu_to_le32(cpu->backing_filename_size); 78 } 79 80 int qed_write_header_sync(BDRVQEDState *s) 81 { 82 QEDHeader le; 83 int ret; 84 85 qed_header_cpu_to_le(&s->header, &le); 86 ret = bdrv_pwrite(s->bs->file->bs, 0, &le, sizeof(le)); 87 if (ret != sizeof(le)) { 88 return ret; 89 } 90 return 0; 91 } 92 93 typedef struct { 94 GenericCB gencb; 95 BDRVQEDState *s; 96 struct iovec iov; 97 QEMUIOVector qiov; 98 int nsectors; 99 uint8_t *buf; 100 } QEDWriteHeaderCB; 101 102 static void qed_write_header_cb(void *opaque, int ret) 103 { 104 QEDWriteHeaderCB *write_header_cb = opaque; 105 106 qemu_vfree(write_header_cb->buf); 107 gencb_complete(write_header_cb, ret); 108 } 109 110 static void qed_write_header_read_cb(void *opaque, int ret) 111 { 112 QEDWriteHeaderCB *write_header_cb = opaque; 113 BDRVQEDState *s = write_header_cb->s; 114 115 if (ret) { 116 qed_write_header_cb(write_header_cb, ret); 117 return; 118 } 119 120 /* Update header */ 121 qed_header_cpu_to_le(&s->header, (QEDHeader *)write_header_cb->buf); 122 123 bdrv_aio_writev(s->bs->file->bs, 0, &write_header_cb->qiov, 124 write_header_cb->nsectors, qed_write_header_cb, 125 write_header_cb); 126 } 127 128 /** 129 * Update header in-place (does not rewrite backing filename or other strings) 130 * 131 * This function only updates known header fields in-place and does not affect 132 * extra data after the QED header. 133 */ 134 static void qed_write_header(BDRVQEDState *s, BlockCompletionFunc cb, 135 void *opaque) 136 { 137 /* We must write full sectors for O_DIRECT but cannot necessarily generate 138 * the data following the header if an unrecognized compat feature is 139 * active. Therefore, first read the sectors containing the header, update 140 * them, and write back. 141 */ 142 143 int nsectors = (sizeof(QEDHeader) + BDRV_SECTOR_SIZE - 1) / 144 BDRV_SECTOR_SIZE; 145 size_t len = nsectors * BDRV_SECTOR_SIZE; 146 QEDWriteHeaderCB *write_header_cb = gencb_alloc(sizeof(*write_header_cb), 147 cb, opaque); 148 149 write_header_cb->s = s; 150 write_header_cb->nsectors = nsectors; 151 write_header_cb->buf = qemu_blockalign(s->bs, len); 152 write_header_cb->iov.iov_base = write_header_cb->buf; 153 write_header_cb->iov.iov_len = len; 154 qemu_iovec_init_external(&write_header_cb->qiov, &write_header_cb->iov, 1); 155 156 bdrv_aio_readv(s->bs->file->bs, 0, &write_header_cb->qiov, nsectors, 157 qed_write_header_read_cb, write_header_cb); 158 } 159 160 static uint64_t qed_max_image_size(uint32_t cluster_size, uint32_t table_size) 161 { 162 uint64_t table_entries; 163 uint64_t l2_size; 164 165 table_entries = (table_size * cluster_size) / sizeof(uint64_t); 166 l2_size = table_entries * cluster_size; 167 168 return l2_size * table_entries; 169 } 170 171 static bool qed_is_cluster_size_valid(uint32_t cluster_size) 172 { 173 if (cluster_size < QED_MIN_CLUSTER_SIZE || 174 cluster_size > QED_MAX_CLUSTER_SIZE) { 175 return false; 176 } 177 if (cluster_size & (cluster_size - 1)) { 178 return false; /* not power of 2 */ 179 } 180 return true; 181 } 182 183 static bool qed_is_table_size_valid(uint32_t table_size) 184 { 185 if (table_size < QED_MIN_TABLE_SIZE || 186 table_size > QED_MAX_TABLE_SIZE) { 187 return false; 188 } 189 if (table_size & (table_size - 1)) { 190 return false; /* not power of 2 */ 191 } 192 return true; 193 } 194 195 static bool qed_is_image_size_valid(uint64_t image_size, uint32_t cluster_size, 196 uint32_t table_size) 197 { 198 if (image_size % BDRV_SECTOR_SIZE != 0) { 199 return false; /* not multiple of sector size */ 200 } 201 if (image_size > qed_max_image_size(cluster_size, table_size)) { 202 return false; /* image is too large */ 203 } 204 return true; 205 } 206 207 /** 208 * Read a string of known length from the image file 209 * 210 * @file: Image file 211 * @offset: File offset to start of string, in bytes 212 * @n: String length in bytes 213 * @buf: Destination buffer 214 * @buflen: Destination buffer length in bytes 215 * @ret: 0 on success, -errno on failure 216 * 217 * The string is NUL-terminated. 218 */ 219 static int qed_read_string(BlockDriverState *file, uint64_t offset, size_t n, 220 char *buf, size_t buflen) 221 { 222 int ret; 223 if (n >= buflen) { 224 return -EINVAL; 225 } 226 ret = bdrv_pread(file, offset, buf, n); 227 if (ret < 0) { 228 return ret; 229 } 230 buf[n] = '\0'; 231 return 0; 232 } 233 234 /** 235 * Allocate new clusters 236 * 237 * @s: QED state 238 * @n: Number of contiguous clusters to allocate 239 * @ret: Offset of first allocated cluster 240 * 241 * This function only produces the offset where the new clusters should be 242 * written. It updates BDRVQEDState but does not make any changes to the image 243 * file. 244 */ 245 static uint64_t qed_alloc_clusters(BDRVQEDState *s, unsigned int n) 246 { 247 uint64_t offset = s->file_size; 248 s->file_size += n * s->header.cluster_size; 249 return offset; 250 } 251 252 QEDTable *qed_alloc_table(BDRVQEDState *s) 253 { 254 /* Honor O_DIRECT memory alignment requirements */ 255 return qemu_blockalign(s->bs, 256 s->header.cluster_size * s->header.table_size); 257 } 258 259 /** 260 * Allocate a new zeroed L2 table 261 */ 262 static CachedL2Table *qed_new_l2_table(BDRVQEDState *s) 263 { 264 CachedL2Table *l2_table = qed_alloc_l2_cache_entry(&s->l2_cache); 265 266 l2_table->table = qed_alloc_table(s); 267 l2_table->offset = qed_alloc_clusters(s, s->header.table_size); 268 269 memset(l2_table->table->offsets, 0, 270 s->header.cluster_size * s->header.table_size); 271 return l2_table; 272 } 273 274 static void qed_aio_next_io(void *opaque, int ret); 275 276 static void qed_plug_allocating_write_reqs(BDRVQEDState *s) 277 { 278 assert(!s->allocating_write_reqs_plugged); 279 280 s->allocating_write_reqs_plugged = true; 281 } 282 283 static void qed_unplug_allocating_write_reqs(BDRVQEDState *s) 284 { 285 QEDAIOCB *acb; 286 287 assert(s->allocating_write_reqs_plugged); 288 289 s->allocating_write_reqs_plugged = false; 290 291 acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs); 292 if (acb) { 293 qed_aio_next_io(acb, 0); 294 } 295 } 296 297 static void qed_finish_clear_need_check(void *opaque, int ret) 298 { 299 /* Do nothing */ 300 } 301 302 static void qed_flush_after_clear_need_check(void *opaque, int ret) 303 { 304 BDRVQEDState *s = opaque; 305 306 bdrv_aio_flush(s->bs, qed_finish_clear_need_check, s); 307 308 /* No need to wait until flush completes */ 309 qed_unplug_allocating_write_reqs(s); 310 } 311 312 static void qed_clear_need_check(void *opaque, int ret) 313 { 314 BDRVQEDState *s = opaque; 315 316 if (ret) { 317 qed_unplug_allocating_write_reqs(s); 318 return; 319 } 320 321 s->header.features &= ~QED_F_NEED_CHECK; 322 qed_write_header(s, qed_flush_after_clear_need_check, s); 323 } 324 325 static void qed_need_check_timer_cb(void *opaque) 326 { 327 BDRVQEDState *s = opaque; 328 329 /* The timer should only fire when allocating writes have drained */ 330 assert(!QSIMPLEQ_FIRST(&s->allocating_write_reqs)); 331 332 trace_qed_need_check_timer_cb(s); 333 334 qed_plug_allocating_write_reqs(s); 335 336 /* Ensure writes are on disk before clearing flag */ 337 bdrv_aio_flush(s->bs, qed_clear_need_check, s); 338 } 339 340 static void qed_start_need_check_timer(BDRVQEDState *s) 341 { 342 trace_qed_start_need_check_timer(s); 343 344 /* Use QEMU_CLOCK_VIRTUAL so we don't alter the image file while suspended for 345 * migration. 346 */ 347 timer_mod(s->need_check_timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + 348 get_ticks_per_sec() * QED_NEED_CHECK_TIMEOUT); 349 } 350 351 /* It's okay to call this multiple times or when no timer is started */ 352 static void qed_cancel_need_check_timer(BDRVQEDState *s) 353 { 354 trace_qed_cancel_need_check_timer(s); 355 timer_del(s->need_check_timer); 356 } 357 358 static void bdrv_qed_detach_aio_context(BlockDriverState *bs) 359 { 360 BDRVQEDState *s = bs->opaque; 361 362 qed_cancel_need_check_timer(s); 363 timer_free(s->need_check_timer); 364 } 365 366 static void bdrv_qed_attach_aio_context(BlockDriverState *bs, 367 AioContext *new_context) 368 { 369 BDRVQEDState *s = bs->opaque; 370 371 s->need_check_timer = aio_timer_new(new_context, 372 QEMU_CLOCK_VIRTUAL, SCALE_NS, 373 qed_need_check_timer_cb, s); 374 if (s->header.features & QED_F_NEED_CHECK) { 375 qed_start_need_check_timer(s); 376 } 377 } 378 379 static void bdrv_qed_drain(BlockDriverState *bs) 380 { 381 BDRVQEDState *s = bs->opaque; 382 383 /* Cancel timer and start doing I/O that were meant to happen as if it 384 * fired, that way we get bdrv_drain() taking care of the ongoing requests 385 * correctly. */ 386 qed_cancel_need_check_timer(s); 387 qed_plug_allocating_write_reqs(s); 388 bdrv_aio_flush(s->bs, qed_clear_need_check, s); 389 } 390 391 static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags, 392 Error **errp) 393 { 394 BDRVQEDState *s = bs->opaque; 395 QEDHeader le_header; 396 int64_t file_size; 397 int ret; 398 399 s->bs = bs; 400 QSIMPLEQ_INIT(&s->allocating_write_reqs); 401 402 ret = bdrv_pread(bs->file->bs, 0, &le_header, sizeof(le_header)); 403 if (ret < 0) { 404 return ret; 405 } 406 qed_header_le_to_cpu(&le_header, &s->header); 407 408 if (s->header.magic != QED_MAGIC) { 409 error_setg(errp, "Image not in QED format"); 410 return -EINVAL; 411 } 412 if (s->header.features & ~QED_FEATURE_MASK) { 413 /* image uses unsupported feature bits */ 414 char buf[64]; 415 snprintf(buf, sizeof(buf), "%" PRIx64, 416 s->header.features & ~QED_FEATURE_MASK); 417 error_setg(errp, QERR_UNKNOWN_BLOCK_FORMAT_FEATURE, 418 bdrv_get_device_or_node_name(bs), "QED", buf); 419 return -ENOTSUP; 420 } 421 if (!qed_is_cluster_size_valid(s->header.cluster_size)) { 422 return -EINVAL; 423 } 424 425 /* Round down file size to the last cluster */ 426 file_size = bdrv_getlength(bs->file->bs); 427 if (file_size < 0) { 428 return file_size; 429 } 430 s->file_size = qed_start_of_cluster(s, file_size); 431 432 if (!qed_is_table_size_valid(s->header.table_size)) { 433 return -EINVAL; 434 } 435 if (!qed_is_image_size_valid(s->header.image_size, 436 s->header.cluster_size, 437 s->header.table_size)) { 438 return -EINVAL; 439 } 440 if (!qed_check_table_offset(s, s->header.l1_table_offset)) { 441 return -EINVAL; 442 } 443 444 s->table_nelems = (s->header.cluster_size * s->header.table_size) / 445 sizeof(uint64_t); 446 s->l2_shift = ctz32(s->header.cluster_size); 447 s->l2_mask = s->table_nelems - 1; 448 s->l1_shift = s->l2_shift + ctz32(s->table_nelems); 449 450 /* Header size calculation must not overflow uint32_t */ 451 if (s->header.header_size > UINT32_MAX / s->header.cluster_size) { 452 return -EINVAL; 453 } 454 455 if ((s->header.features & QED_F_BACKING_FILE)) { 456 if ((uint64_t)s->header.backing_filename_offset + 457 s->header.backing_filename_size > 458 s->header.cluster_size * s->header.header_size) { 459 return -EINVAL; 460 } 461 462 ret = qed_read_string(bs->file->bs, s->header.backing_filename_offset, 463 s->header.backing_filename_size, bs->backing_file, 464 sizeof(bs->backing_file)); 465 if (ret < 0) { 466 return ret; 467 } 468 469 if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) { 470 pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw"); 471 } 472 } 473 474 /* Reset unknown autoclear feature bits. This is a backwards 475 * compatibility mechanism that allows images to be opened by older 476 * programs, which "knock out" unknown feature bits. When an image is 477 * opened by a newer program again it can detect that the autoclear 478 * feature is no longer valid. 479 */ 480 if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 && 481 !bdrv_is_read_only(bs->file->bs) && !(flags & BDRV_O_INACTIVE)) { 482 s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK; 483 484 ret = qed_write_header_sync(s); 485 if (ret) { 486 return ret; 487 } 488 489 /* From here on only known autoclear feature bits are valid */ 490 bdrv_flush(bs->file->bs); 491 } 492 493 s->l1_table = qed_alloc_table(s); 494 qed_init_l2_cache(&s->l2_cache); 495 496 ret = qed_read_l1_table_sync(s); 497 if (ret) { 498 goto out; 499 } 500 501 /* If image was not closed cleanly, check consistency */ 502 if (!(flags & BDRV_O_CHECK) && (s->header.features & QED_F_NEED_CHECK)) { 503 /* Read-only images cannot be fixed. There is no risk of corruption 504 * since write operations are not possible. Therefore, allow 505 * potentially inconsistent images to be opened read-only. This can 506 * aid data recovery from an otherwise inconsistent image. 507 */ 508 if (!bdrv_is_read_only(bs->file->bs) && 509 !(flags & BDRV_O_INACTIVE)) { 510 BdrvCheckResult result = {0}; 511 512 ret = qed_check(s, &result, true); 513 if (ret) { 514 goto out; 515 } 516 } 517 } 518 519 bdrv_qed_attach_aio_context(bs, bdrv_get_aio_context(bs)); 520 521 out: 522 if (ret) { 523 qed_free_l2_cache(&s->l2_cache); 524 qemu_vfree(s->l1_table); 525 } 526 return ret; 527 } 528 529 static void bdrv_qed_refresh_limits(BlockDriverState *bs, Error **errp) 530 { 531 BDRVQEDState *s = bs->opaque; 532 533 bs->bl.write_zeroes_alignment = s->header.cluster_size >> BDRV_SECTOR_BITS; 534 } 535 536 /* We have nothing to do for QED reopen, stubs just return 537 * success */ 538 static int bdrv_qed_reopen_prepare(BDRVReopenState *state, 539 BlockReopenQueue *queue, Error **errp) 540 { 541 return 0; 542 } 543 544 static void bdrv_qed_close(BlockDriverState *bs) 545 { 546 BDRVQEDState *s = bs->opaque; 547 548 bdrv_qed_detach_aio_context(bs); 549 550 /* Ensure writes reach stable storage */ 551 bdrv_flush(bs->file->bs); 552 553 /* Clean shutdown, no check required on next open */ 554 if (s->header.features & QED_F_NEED_CHECK) { 555 s->header.features &= ~QED_F_NEED_CHECK; 556 qed_write_header_sync(s); 557 } 558 559 qed_free_l2_cache(&s->l2_cache); 560 qemu_vfree(s->l1_table); 561 } 562 563 static int qed_create(const char *filename, uint32_t cluster_size, 564 uint64_t image_size, uint32_t table_size, 565 const char *backing_file, const char *backing_fmt, 566 QemuOpts *opts, Error **errp) 567 { 568 QEDHeader header = { 569 .magic = QED_MAGIC, 570 .cluster_size = cluster_size, 571 .table_size = table_size, 572 .header_size = 1, 573 .features = 0, 574 .compat_features = 0, 575 .l1_table_offset = cluster_size, 576 .image_size = image_size, 577 }; 578 QEDHeader le_header; 579 uint8_t *l1_table = NULL; 580 size_t l1_size = header.cluster_size * header.table_size; 581 Error *local_err = NULL; 582 int ret = 0; 583 BlockDriverState *bs; 584 585 ret = bdrv_create_file(filename, opts, &local_err); 586 if (ret < 0) { 587 error_propagate(errp, local_err); 588 return ret; 589 } 590 591 bs = NULL; 592 ret = bdrv_open(&bs, filename, NULL, NULL, 593 BDRV_O_RDWR | BDRV_O_CACHE_WB | BDRV_O_PROTOCOL, 594 &local_err); 595 if (ret < 0) { 596 error_propagate(errp, local_err); 597 return ret; 598 } 599 600 /* File must start empty and grow, check truncate is supported */ 601 ret = bdrv_truncate(bs, 0); 602 if (ret < 0) { 603 goto out; 604 } 605 606 if (backing_file) { 607 header.features |= QED_F_BACKING_FILE; 608 header.backing_filename_offset = sizeof(le_header); 609 header.backing_filename_size = strlen(backing_file); 610 611 if (qed_fmt_is_raw(backing_fmt)) { 612 header.features |= QED_F_BACKING_FORMAT_NO_PROBE; 613 } 614 } 615 616 qed_header_cpu_to_le(&header, &le_header); 617 ret = bdrv_pwrite(bs, 0, &le_header, sizeof(le_header)); 618 if (ret < 0) { 619 goto out; 620 } 621 ret = bdrv_pwrite(bs, sizeof(le_header), backing_file, 622 header.backing_filename_size); 623 if (ret < 0) { 624 goto out; 625 } 626 627 l1_table = g_malloc0(l1_size); 628 ret = bdrv_pwrite(bs, header.l1_table_offset, l1_table, l1_size); 629 if (ret < 0) { 630 goto out; 631 } 632 633 ret = 0; /* success */ 634 out: 635 g_free(l1_table); 636 bdrv_unref(bs); 637 return ret; 638 } 639 640 static int bdrv_qed_create(const char *filename, QemuOpts *opts, Error **errp) 641 { 642 uint64_t image_size = 0; 643 uint32_t cluster_size = QED_DEFAULT_CLUSTER_SIZE; 644 uint32_t table_size = QED_DEFAULT_TABLE_SIZE; 645 char *backing_file = NULL; 646 char *backing_fmt = NULL; 647 int ret; 648 649 image_size = ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0), 650 BDRV_SECTOR_SIZE); 651 backing_file = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FILE); 652 backing_fmt = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FMT); 653 cluster_size = qemu_opt_get_size_del(opts, 654 BLOCK_OPT_CLUSTER_SIZE, 655 QED_DEFAULT_CLUSTER_SIZE); 656 table_size = qemu_opt_get_size_del(opts, BLOCK_OPT_TABLE_SIZE, 657 QED_DEFAULT_TABLE_SIZE); 658 659 if (!qed_is_cluster_size_valid(cluster_size)) { 660 error_setg(errp, "QED cluster size must be within range [%u, %u] " 661 "and power of 2", 662 QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE); 663 ret = -EINVAL; 664 goto finish; 665 } 666 if (!qed_is_table_size_valid(table_size)) { 667 error_setg(errp, "QED table size must be within range [%u, %u] " 668 "and power of 2", 669 QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE); 670 ret = -EINVAL; 671 goto finish; 672 } 673 if (!qed_is_image_size_valid(image_size, cluster_size, table_size)) { 674 error_setg(errp, "QED image size must be a non-zero multiple of " 675 "cluster size and less than %" PRIu64 " bytes", 676 qed_max_image_size(cluster_size, table_size)); 677 ret = -EINVAL; 678 goto finish; 679 } 680 681 ret = qed_create(filename, cluster_size, image_size, table_size, 682 backing_file, backing_fmt, opts, errp); 683 684 finish: 685 g_free(backing_file); 686 g_free(backing_fmt); 687 return ret; 688 } 689 690 typedef struct { 691 BlockDriverState *bs; 692 Coroutine *co; 693 uint64_t pos; 694 int64_t status; 695 int *pnum; 696 BlockDriverState **file; 697 } QEDIsAllocatedCB; 698 699 static void qed_is_allocated_cb(void *opaque, int ret, uint64_t offset, size_t len) 700 { 701 QEDIsAllocatedCB *cb = opaque; 702 BDRVQEDState *s = cb->bs->opaque; 703 *cb->pnum = len / BDRV_SECTOR_SIZE; 704 switch (ret) { 705 case QED_CLUSTER_FOUND: 706 offset |= qed_offset_into_cluster(s, cb->pos); 707 cb->status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID | offset; 708 *cb->file = cb->bs->file->bs; 709 break; 710 case QED_CLUSTER_ZERO: 711 cb->status = BDRV_BLOCK_ZERO; 712 break; 713 case QED_CLUSTER_L2: 714 case QED_CLUSTER_L1: 715 cb->status = 0; 716 break; 717 default: 718 assert(ret < 0); 719 cb->status = ret; 720 break; 721 } 722 723 if (cb->co) { 724 qemu_coroutine_enter(cb->co, NULL); 725 } 726 } 727 728 static int64_t coroutine_fn bdrv_qed_co_get_block_status(BlockDriverState *bs, 729 int64_t sector_num, 730 int nb_sectors, int *pnum, 731 BlockDriverState **file) 732 { 733 BDRVQEDState *s = bs->opaque; 734 size_t len = (size_t)nb_sectors * BDRV_SECTOR_SIZE; 735 QEDIsAllocatedCB cb = { 736 .bs = bs, 737 .pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE, 738 .status = BDRV_BLOCK_OFFSET_MASK, 739 .pnum = pnum, 740 .file = file, 741 }; 742 QEDRequest request = { .l2_table = NULL }; 743 744 qed_find_cluster(s, &request, cb.pos, len, qed_is_allocated_cb, &cb); 745 746 /* Now sleep if the callback wasn't invoked immediately */ 747 while (cb.status == BDRV_BLOCK_OFFSET_MASK) { 748 cb.co = qemu_coroutine_self(); 749 qemu_coroutine_yield(); 750 } 751 752 qed_unref_l2_cache_entry(request.l2_table); 753 754 return cb.status; 755 } 756 757 static BDRVQEDState *acb_to_s(QEDAIOCB *acb) 758 { 759 return acb->common.bs->opaque; 760 } 761 762 /** 763 * Read from the backing file or zero-fill if no backing file 764 * 765 * @s: QED state 766 * @pos: Byte position in device 767 * @qiov: Destination I/O vector 768 * @backing_qiov: Possibly shortened copy of qiov, to be allocated here 769 * @cb: Completion function 770 * @opaque: User data for completion function 771 * 772 * This function reads qiov->size bytes starting at pos from the backing file. 773 * If there is no backing file then zeroes are read. 774 */ 775 static void qed_read_backing_file(BDRVQEDState *s, uint64_t pos, 776 QEMUIOVector *qiov, 777 QEMUIOVector **backing_qiov, 778 BlockCompletionFunc *cb, void *opaque) 779 { 780 uint64_t backing_length = 0; 781 size_t size; 782 783 /* If there is a backing file, get its length. Treat the absence of a 784 * backing file like a zero length backing file. 785 */ 786 if (s->bs->backing) { 787 int64_t l = bdrv_getlength(s->bs->backing->bs); 788 if (l < 0) { 789 cb(opaque, l); 790 return; 791 } 792 backing_length = l; 793 } 794 795 /* Zero all sectors if reading beyond the end of the backing file */ 796 if (pos >= backing_length || 797 pos + qiov->size > backing_length) { 798 qemu_iovec_memset(qiov, 0, 0, qiov->size); 799 } 800 801 /* Complete now if there are no backing file sectors to read */ 802 if (pos >= backing_length) { 803 cb(opaque, 0); 804 return; 805 } 806 807 /* If the read straddles the end of the backing file, shorten it */ 808 size = MIN((uint64_t)backing_length - pos, qiov->size); 809 810 assert(*backing_qiov == NULL); 811 *backing_qiov = g_new(QEMUIOVector, 1); 812 qemu_iovec_init(*backing_qiov, qiov->niov); 813 qemu_iovec_concat(*backing_qiov, qiov, 0, size); 814 815 BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO); 816 bdrv_aio_readv(s->bs->backing->bs, pos / BDRV_SECTOR_SIZE, 817 *backing_qiov, size / BDRV_SECTOR_SIZE, cb, opaque); 818 } 819 820 typedef struct { 821 GenericCB gencb; 822 BDRVQEDState *s; 823 QEMUIOVector qiov; 824 QEMUIOVector *backing_qiov; 825 struct iovec iov; 826 uint64_t offset; 827 } CopyFromBackingFileCB; 828 829 static void qed_copy_from_backing_file_cb(void *opaque, int ret) 830 { 831 CopyFromBackingFileCB *copy_cb = opaque; 832 qemu_vfree(copy_cb->iov.iov_base); 833 gencb_complete(©_cb->gencb, ret); 834 } 835 836 static void qed_copy_from_backing_file_write(void *opaque, int ret) 837 { 838 CopyFromBackingFileCB *copy_cb = opaque; 839 BDRVQEDState *s = copy_cb->s; 840 841 if (copy_cb->backing_qiov) { 842 qemu_iovec_destroy(copy_cb->backing_qiov); 843 g_free(copy_cb->backing_qiov); 844 copy_cb->backing_qiov = NULL; 845 } 846 847 if (ret) { 848 qed_copy_from_backing_file_cb(copy_cb, ret); 849 return; 850 } 851 852 BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE); 853 bdrv_aio_writev(s->bs->file->bs, copy_cb->offset / BDRV_SECTOR_SIZE, 854 ©_cb->qiov, copy_cb->qiov.size / BDRV_SECTOR_SIZE, 855 qed_copy_from_backing_file_cb, copy_cb); 856 } 857 858 /** 859 * Copy data from backing file into the image 860 * 861 * @s: QED state 862 * @pos: Byte position in device 863 * @len: Number of bytes 864 * @offset: Byte offset in image file 865 * @cb: Completion function 866 * @opaque: User data for completion function 867 */ 868 static void qed_copy_from_backing_file(BDRVQEDState *s, uint64_t pos, 869 uint64_t len, uint64_t offset, 870 BlockCompletionFunc *cb, 871 void *opaque) 872 { 873 CopyFromBackingFileCB *copy_cb; 874 875 /* Skip copy entirely if there is no work to do */ 876 if (len == 0) { 877 cb(opaque, 0); 878 return; 879 } 880 881 copy_cb = gencb_alloc(sizeof(*copy_cb), cb, opaque); 882 copy_cb->s = s; 883 copy_cb->offset = offset; 884 copy_cb->backing_qiov = NULL; 885 copy_cb->iov.iov_base = qemu_blockalign(s->bs, len); 886 copy_cb->iov.iov_len = len; 887 qemu_iovec_init_external(©_cb->qiov, ©_cb->iov, 1); 888 889 qed_read_backing_file(s, pos, ©_cb->qiov, ©_cb->backing_qiov, 890 qed_copy_from_backing_file_write, copy_cb); 891 } 892 893 /** 894 * Link one or more contiguous clusters into a table 895 * 896 * @s: QED state 897 * @table: L2 table 898 * @index: First cluster index 899 * @n: Number of contiguous clusters 900 * @cluster: First cluster offset 901 * 902 * The cluster offset may be an allocated byte offset in the image file, the 903 * zero cluster marker, or the unallocated cluster marker. 904 */ 905 static void qed_update_l2_table(BDRVQEDState *s, QEDTable *table, int index, 906 unsigned int n, uint64_t cluster) 907 { 908 int i; 909 for (i = index; i < index + n; i++) { 910 table->offsets[i] = cluster; 911 if (!qed_offset_is_unalloc_cluster(cluster) && 912 !qed_offset_is_zero_cluster(cluster)) { 913 cluster += s->header.cluster_size; 914 } 915 } 916 } 917 918 static void qed_aio_complete_bh(void *opaque) 919 { 920 QEDAIOCB *acb = opaque; 921 BlockCompletionFunc *cb = acb->common.cb; 922 void *user_opaque = acb->common.opaque; 923 int ret = acb->bh_ret; 924 925 qemu_bh_delete(acb->bh); 926 qemu_aio_unref(acb); 927 928 /* Invoke callback */ 929 cb(user_opaque, ret); 930 } 931 932 static void qed_aio_complete(QEDAIOCB *acb, int ret) 933 { 934 BDRVQEDState *s = acb_to_s(acb); 935 936 trace_qed_aio_complete(s, acb, ret); 937 938 /* Free resources */ 939 qemu_iovec_destroy(&acb->cur_qiov); 940 qed_unref_l2_cache_entry(acb->request.l2_table); 941 942 /* Free the buffer we may have allocated for zero writes */ 943 if (acb->flags & QED_AIOCB_ZERO) { 944 qemu_vfree(acb->qiov->iov[0].iov_base); 945 acb->qiov->iov[0].iov_base = NULL; 946 } 947 948 /* Arrange for a bh to invoke the completion function */ 949 acb->bh_ret = ret; 950 acb->bh = aio_bh_new(bdrv_get_aio_context(acb->common.bs), 951 qed_aio_complete_bh, acb); 952 qemu_bh_schedule(acb->bh); 953 954 /* Start next allocating write request waiting behind this one. Note that 955 * requests enqueue themselves when they first hit an unallocated cluster 956 * but they wait until the entire request is finished before waking up the 957 * next request in the queue. This ensures that we don't cycle through 958 * requests multiple times but rather finish one at a time completely. 959 */ 960 if (acb == QSIMPLEQ_FIRST(&s->allocating_write_reqs)) { 961 QSIMPLEQ_REMOVE_HEAD(&s->allocating_write_reqs, next); 962 acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs); 963 if (acb) { 964 qed_aio_next_io(acb, 0); 965 } else if (s->header.features & QED_F_NEED_CHECK) { 966 qed_start_need_check_timer(s); 967 } 968 } 969 } 970 971 /** 972 * Commit the current L2 table to the cache 973 */ 974 static void qed_commit_l2_update(void *opaque, int ret) 975 { 976 QEDAIOCB *acb = opaque; 977 BDRVQEDState *s = acb_to_s(acb); 978 CachedL2Table *l2_table = acb->request.l2_table; 979 uint64_t l2_offset = l2_table->offset; 980 981 qed_commit_l2_cache_entry(&s->l2_cache, l2_table); 982 983 /* This is guaranteed to succeed because we just committed the entry to the 984 * cache. 985 */ 986 acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset); 987 assert(acb->request.l2_table != NULL); 988 989 qed_aio_next_io(opaque, ret); 990 } 991 992 /** 993 * Update L1 table with new L2 table offset and write it out 994 */ 995 static void qed_aio_write_l1_update(void *opaque, int ret) 996 { 997 QEDAIOCB *acb = opaque; 998 BDRVQEDState *s = acb_to_s(acb); 999 int index; 1000 1001 if (ret) { 1002 qed_aio_complete(acb, ret); 1003 return; 1004 } 1005 1006 index = qed_l1_index(s, acb->cur_pos); 1007 s->l1_table->offsets[index] = acb->request.l2_table->offset; 1008 1009 qed_write_l1_table(s, index, 1, qed_commit_l2_update, acb); 1010 } 1011 1012 /** 1013 * Update L2 table with new cluster offsets and write them out 1014 */ 1015 static void qed_aio_write_l2_update(QEDAIOCB *acb, int ret, uint64_t offset) 1016 { 1017 BDRVQEDState *s = acb_to_s(acb); 1018 bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1; 1019 int index; 1020 1021 if (ret) { 1022 goto err; 1023 } 1024 1025 if (need_alloc) { 1026 qed_unref_l2_cache_entry(acb->request.l2_table); 1027 acb->request.l2_table = qed_new_l2_table(s); 1028 } 1029 1030 index = qed_l2_index(s, acb->cur_pos); 1031 qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters, 1032 offset); 1033 1034 if (need_alloc) { 1035 /* Write out the whole new L2 table */ 1036 qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true, 1037 qed_aio_write_l1_update, acb); 1038 } else { 1039 /* Write out only the updated part of the L2 table */ 1040 qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters, false, 1041 qed_aio_next_io, acb); 1042 } 1043 return; 1044 1045 err: 1046 qed_aio_complete(acb, ret); 1047 } 1048 1049 static void qed_aio_write_l2_update_cb(void *opaque, int ret) 1050 { 1051 QEDAIOCB *acb = opaque; 1052 qed_aio_write_l2_update(acb, ret, acb->cur_cluster); 1053 } 1054 1055 /** 1056 * Flush new data clusters before updating the L2 table 1057 * 1058 * This flush is necessary when a backing file is in use. A crash during an 1059 * allocating write could result in empty clusters in the image. If the write 1060 * only touched a subregion of the cluster, then backing image sectors have 1061 * been lost in the untouched region. The solution is to flush after writing a 1062 * new data cluster and before updating the L2 table. 1063 */ 1064 static void qed_aio_write_flush_before_l2_update(void *opaque, int ret) 1065 { 1066 QEDAIOCB *acb = opaque; 1067 BDRVQEDState *s = acb_to_s(acb); 1068 1069 if (!bdrv_aio_flush(s->bs->file->bs, qed_aio_write_l2_update_cb, opaque)) { 1070 qed_aio_complete(acb, -EIO); 1071 } 1072 } 1073 1074 /** 1075 * Write data to the image file 1076 */ 1077 static void qed_aio_write_main(void *opaque, int ret) 1078 { 1079 QEDAIOCB *acb = opaque; 1080 BDRVQEDState *s = acb_to_s(acb); 1081 uint64_t offset = acb->cur_cluster + 1082 qed_offset_into_cluster(s, acb->cur_pos); 1083 BlockCompletionFunc *next_fn; 1084 1085 trace_qed_aio_write_main(s, acb, ret, offset, acb->cur_qiov.size); 1086 1087 if (ret) { 1088 qed_aio_complete(acb, ret); 1089 return; 1090 } 1091 1092 if (acb->find_cluster_ret == QED_CLUSTER_FOUND) { 1093 next_fn = qed_aio_next_io; 1094 } else { 1095 if (s->bs->backing) { 1096 next_fn = qed_aio_write_flush_before_l2_update; 1097 } else { 1098 next_fn = qed_aio_write_l2_update_cb; 1099 } 1100 } 1101 1102 BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO); 1103 bdrv_aio_writev(s->bs->file->bs, offset / BDRV_SECTOR_SIZE, 1104 &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE, 1105 next_fn, acb); 1106 } 1107 1108 /** 1109 * Populate back untouched region of new data cluster 1110 */ 1111 static void qed_aio_write_postfill(void *opaque, int ret) 1112 { 1113 QEDAIOCB *acb = opaque; 1114 BDRVQEDState *s = acb_to_s(acb); 1115 uint64_t start = acb->cur_pos + acb->cur_qiov.size; 1116 uint64_t len = 1117 qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start; 1118 uint64_t offset = acb->cur_cluster + 1119 qed_offset_into_cluster(s, acb->cur_pos) + 1120 acb->cur_qiov.size; 1121 1122 if (ret) { 1123 qed_aio_complete(acb, ret); 1124 return; 1125 } 1126 1127 trace_qed_aio_write_postfill(s, acb, start, len, offset); 1128 qed_copy_from_backing_file(s, start, len, offset, 1129 qed_aio_write_main, acb); 1130 } 1131 1132 /** 1133 * Populate front untouched region of new data cluster 1134 */ 1135 static void qed_aio_write_prefill(void *opaque, int ret) 1136 { 1137 QEDAIOCB *acb = opaque; 1138 BDRVQEDState *s = acb_to_s(acb); 1139 uint64_t start = qed_start_of_cluster(s, acb->cur_pos); 1140 uint64_t len = qed_offset_into_cluster(s, acb->cur_pos); 1141 1142 trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster); 1143 qed_copy_from_backing_file(s, start, len, acb->cur_cluster, 1144 qed_aio_write_postfill, acb); 1145 } 1146 1147 /** 1148 * Check if the QED_F_NEED_CHECK bit should be set during allocating write 1149 */ 1150 static bool qed_should_set_need_check(BDRVQEDState *s) 1151 { 1152 /* The flush before L2 update path ensures consistency */ 1153 if (s->bs->backing) { 1154 return false; 1155 } 1156 1157 return !(s->header.features & QED_F_NEED_CHECK); 1158 } 1159 1160 static void qed_aio_write_zero_cluster(void *opaque, int ret) 1161 { 1162 QEDAIOCB *acb = opaque; 1163 1164 if (ret) { 1165 qed_aio_complete(acb, ret); 1166 return; 1167 } 1168 1169 qed_aio_write_l2_update(acb, 0, 1); 1170 } 1171 1172 /** 1173 * Write new data cluster 1174 * 1175 * @acb: Write request 1176 * @len: Length in bytes 1177 * 1178 * This path is taken when writing to previously unallocated clusters. 1179 */ 1180 static void qed_aio_write_alloc(QEDAIOCB *acb, size_t len) 1181 { 1182 BDRVQEDState *s = acb_to_s(acb); 1183 BlockCompletionFunc *cb; 1184 1185 /* Cancel timer when the first allocating request comes in */ 1186 if (QSIMPLEQ_EMPTY(&s->allocating_write_reqs)) { 1187 qed_cancel_need_check_timer(s); 1188 } 1189 1190 /* Freeze this request if another allocating write is in progress */ 1191 if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs)) { 1192 QSIMPLEQ_INSERT_TAIL(&s->allocating_write_reqs, acb, next); 1193 } 1194 if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs) || 1195 s->allocating_write_reqs_plugged) { 1196 return; /* wait for existing request to finish */ 1197 } 1198 1199 acb->cur_nclusters = qed_bytes_to_clusters(s, 1200 qed_offset_into_cluster(s, acb->cur_pos) + len); 1201 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len); 1202 1203 if (acb->flags & QED_AIOCB_ZERO) { 1204 /* Skip ahead if the clusters are already zero */ 1205 if (acb->find_cluster_ret == QED_CLUSTER_ZERO) { 1206 qed_aio_next_io(acb, 0); 1207 return; 1208 } 1209 1210 cb = qed_aio_write_zero_cluster; 1211 } else { 1212 cb = qed_aio_write_prefill; 1213 acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters); 1214 } 1215 1216 if (qed_should_set_need_check(s)) { 1217 s->header.features |= QED_F_NEED_CHECK; 1218 qed_write_header(s, cb, acb); 1219 } else { 1220 cb(acb, 0); 1221 } 1222 } 1223 1224 /** 1225 * Write data cluster in place 1226 * 1227 * @acb: Write request 1228 * @offset: Cluster offset in bytes 1229 * @len: Length in bytes 1230 * 1231 * This path is taken when writing to already allocated clusters. 1232 */ 1233 static void qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset, size_t len) 1234 { 1235 /* Allocate buffer for zero writes */ 1236 if (acb->flags & QED_AIOCB_ZERO) { 1237 struct iovec *iov = acb->qiov->iov; 1238 1239 if (!iov->iov_base) { 1240 iov->iov_base = qemu_try_blockalign(acb->common.bs, iov->iov_len); 1241 if (iov->iov_base == NULL) { 1242 qed_aio_complete(acb, -ENOMEM); 1243 return; 1244 } 1245 memset(iov->iov_base, 0, iov->iov_len); 1246 } 1247 } 1248 1249 /* Calculate the I/O vector */ 1250 acb->cur_cluster = offset; 1251 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len); 1252 1253 /* Do the actual write */ 1254 qed_aio_write_main(acb, 0); 1255 } 1256 1257 /** 1258 * Write data cluster 1259 * 1260 * @opaque: Write 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_write_data(void *opaque, int ret, 1269 uint64_t offset, size_t len) 1270 { 1271 QEDAIOCB *acb = opaque; 1272 1273 trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len); 1274 1275 acb->find_cluster_ret = ret; 1276 1277 switch (ret) { 1278 case QED_CLUSTER_FOUND: 1279 qed_aio_write_inplace(acb, offset, len); 1280 break; 1281 1282 case QED_CLUSTER_L2: 1283 case QED_CLUSTER_L1: 1284 case QED_CLUSTER_ZERO: 1285 qed_aio_write_alloc(acb, len); 1286 break; 1287 1288 default: 1289 qed_aio_complete(acb, ret); 1290 break; 1291 } 1292 } 1293 1294 /** 1295 * Read data cluster 1296 * 1297 * @opaque: Read request 1298 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1, 1299 * or -errno 1300 * @offset: Cluster offset in bytes 1301 * @len: Length in bytes 1302 * 1303 * Callback from qed_find_cluster(). 1304 */ 1305 static void qed_aio_read_data(void *opaque, int ret, 1306 uint64_t offset, size_t len) 1307 { 1308 QEDAIOCB *acb = opaque; 1309 BDRVQEDState *s = acb_to_s(acb); 1310 BlockDriverState *bs = acb->common.bs; 1311 1312 /* Adjust offset into cluster */ 1313 offset += qed_offset_into_cluster(s, acb->cur_pos); 1314 1315 trace_qed_aio_read_data(s, acb, ret, offset, len); 1316 1317 if (ret < 0) { 1318 goto err; 1319 } 1320 1321 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len); 1322 1323 /* Handle zero cluster and backing file reads */ 1324 if (ret == QED_CLUSTER_ZERO) { 1325 qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size); 1326 qed_aio_next_io(acb, 0); 1327 return; 1328 } else if (ret != QED_CLUSTER_FOUND) { 1329 qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov, 1330 &acb->backing_qiov, qed_aio_next_io, acb); 1331 return; 1332 } 1333 1334 BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO); 1335 bdrv_aio_readv(bs->file->bs, offset / BDRV_SECTOR_SIZE, 1336 &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE, 1337 qed_aio_next_io, acb); 1338 return; 1339 1340 err: 1341 qed_aio_complete(acb, ret); 1342 } 1343 1344 /** 1345 * Begin next I/O or complete the request 1346 */ 1347 static void qed_aio_next_io(void *opaque, int ret) 1348 { 1349 QEDAIOCB *acb = opaque; 1350 BDRVQEDState *s = acb_to_s(acb); 1351 QEDFindClusterFunc *io_fn = (acb->flags & QED_AIOCB_WRITE) ? 1352 qed_aio_write_data : qed_aio_read_data; 1353 1354 trace_qed_aio_next_io(s, acb, ret, acb->cur_pos + acb->cur_qiov.size); 1355 1356 if (acb->backing_qiov) { 1357 qemu_iovec_destroy(acb->backing_qiov); 1358 g_free(acb->backing_qiov); 1359 acb->backing_qiov = NULL; 1360 } 1361 1362 /* Handle I/O error */ 1363 if (ret) { 1364 qed_aio_complete(acb, ret); 1365 return; 1366 } 1367 1368 acb->qiov_offset += acb->cur_qiov.size; 1369 acb->cur_pos += acb->cur_qiov.size; 1370 qemu_iovec_reset(&acb->cur_qiov); 1371 1372 /* Complete request */ 1373 if (acb->cur_pos >= acb->end_pos) { 1374 qed_aio_complete(acb, 0); 1375 return; 1376 } 1377 1378 /* Find next cluster and start I/O */ 1379 qed_find_cluster(s, &acb->request, 1380 acb->cur_pos, acb->end_pos - acb->cur_pos, 1381 io_fn, acb); 1382 } 1383 1384 static BlockAIOCB *qed_aio_setup(BlockDriverState *bs, 1385 int64_t sector_num, 1386 QEMUIOVector *qiov, int nb_sectors, 1387 BlockCompletionFunc *cb, 1388 void *opaque, int flags) 1389 { 1390 QEDAIOCB *acb = qemu_aio_get(&qed_aiocb_info, bs, cb, opaque); 1391 1392 trace_qed_aio_setup(bs->opaque, acb, sector_num, nb_sectors, 1393 opaque, flags); 1394 1395 acb->flags = flags; 1396 acb->qiov = qiov; 1397 acb->qiov_offset = 0; 1398 acb->cur_pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE; 1399 acb->end_pos = acb->cur_pos + nb_sectors * BDRV_SECTOR_SIZE; 1400 acb->backing_qiov = NULL; 1401 acb->request.l2_table = NULL; 1402 qemu_iovec_init(&acb->cur_qiov, qiov->niov); 1403 1404 /* Start request */ 1405 qed_aio_next_io(acb, 0); 1406 return &acb->common; 1407 } 1408 1409 static BlockAIOCB *bdrv_qed_aio_readv(BlockDriverState *bs, 1410 int64_t sector_num, 1411 QEMUIOVector *qiov, int nb_sectors, 1412 BlockCompletionFunc *cb, 1413 void *opaque) 1414 { 1415 return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb, opaque, 0); 1416 } 1417 1418 static BlockAIOCB *bdrv_qed_aio_writev(BlockDriverState *bs, 1419 int64_t sector_num, 1420 QEMUIOVector *qiov, int nb_sectors, 1421 BlockCompletionFunc *cb, 1422 void *opaque) 1423 { 1424 return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb, 1425 opaque, QED_AIOCB_WRITE); 1426 } 1427 1428 typedef struct { 1429 Coroutine *co; 1430 int ret; 1431 bool done; 1432 } QEDWriteZeroesCB; 1433 1434 static void coroutine_fn qed_co_write_zeroes_cb(void *opaque, int ret) 1435 { 1436 QEDWriteZeroesCB *cb = opaque; 1437 1438 cb->done = true; 1439 cb->ret = ret; 1440 if (cb->co) { 1441 qemu_coroutine_enter(cb->co, NULL); 1442 } 1443 } 1444 1445 static int coroutine_fn bdrv_qed_co_write_zeroes(BlockDriverState *bs, 1446 int64_t sector_num, 1447 int nb_sectors, 1448 BdrvRequestFlags flags) 1449 { 1450 BlockAIOCB *blockacb; 1451 BDRVQEDState *s = bs->opaque; 1452 QEDWriteZeroesCB cb = { .done = false }; 1453 QEMUIOVector qiov; 1454 struct iovec iov; 1455 1456 /* Refuse if there are untouched backing file sectors */ 1457 if (bs->backing) { 1458 if (qed_offset_into_cluster(s, sector_num * BDRV_SECTOR_SIZE) != 0) { 1459 return -ENOTSUP; 1460 } 1461 if (qed_offset_into_cluster(s, nb_sectors * BDRV_SECTOR_SIZE) != 0) { 1462 return -ENOTSUP; 1463 } 1464 } 1465 1466 /* Zero writes start without an I/O buffer. If a buffer becomes necessary 1467 * then it will be allocated during request processing. 1468 */ 1469 iov.iov_base = NULL, 1470 iov.iov_len = nb_sectors * BDRV_SECTOR_SIZE, 1471 1472 qemu_iovec_init_external(&qiov, &iov, 1); 1473 blockacb = qed_aio_setup(bs, sector_num, &qiov, nb_sectors, 1474 qed_co_write_zeroes_cb, &cb, 1475 QED_AIOCB_WRITE | QED_AIOCB_ZERO); 1476 if (!blockacb) { 1477 return -EIO; 1478 } 1479 if (!cb.done) { 1480 cb.co = qemu_coroutine_self(); 1481 qemu_coroutine_yield(); 1482 } 1483 assert(cb.done); 1484 return cb.ret; 1485 } 1486 1487 static int bdrv_qed_truncate(BlockDriverState *bs, int64_t offset) 1488 { 1489 BDRVQEDState *s = bs->opaque; 1490 uint64_t old_image_size; 1491 int ret; 1492 1493 if (!qed_is_image_size_valid(offset, s->header.cluster_size, 1494 s->header.table_size)) { 1495 return -EINVAL; 1496 } 1497 1498 /* Shrinking is currently not supported */ 1499 if ((uint64_t)offset < s->header.image_size) { 1500 return -ENOTSUP; 1501 } 1502 1503 old_image_size = s->header.image_size; 1504 s->header.image_size = offset; 1505 ret = qed_write_header_sync(s); 1506 if (ret < 0) { 1507 s->header.image_size = old_image_size; 1508 } 1509 return ret; 1510 } 1511 1512 static int64_t bdrv_qed_getlength(BlockDriverState *bs) 1513 { 1514 BDRVQEDState *s = bs->opaque; 1515 return s->header.image_size; 1516 } 1517 1518 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi) 1519 { 1520 BDRVQEDState *s = bs->opaque; 1521 1522 memset(bdi, 0, sizeof(*bdi)); 1523 bdi->cluster_size = s->header.cluster_size; 1524 bdi->is_dirty = s->header.features & QED_F_NEED_CHECK; 1525 bdi->unallocated_blocks_are_zero = true; 1526 bdi->can_write_zeroes_with_unmap = true; 1527 return 0; 1528 } 1529 1530 static int bdrv_qed_change_backing_file(BlockDriverState *bs, 1531 const char *backing_file, 1532 const char *backing_fmt) 1533 { 1534 BDRVQEDState *s = bs->opaque; 1535 QEDHeader new_header, le_header; 1536 void *buffer; 1537 size_t buffer_len, backing_file_len; 1538 int ret; 1539 1540 /* Refuse to set backing filename if unknown compat feature bits are 1541 * active. If the image uses an unknown compat feature then we may not 1542 * know the layout of data following the header structure and cannot safely 1543 * add a new string. 1544 */ 1545 if (backing_file && (s->header.compat_features & 1546 ~QED_COMPAT_FEATURE_MASK)) { 1547 return -ENOTSUP; 1548 } 1549 1550 memcpy(&new_header, &s->header, sizeof(new_header)); 1551 1552 new_header.features &= ~(QED_F_BACKING_FILE | 1553 QED_F_BACKING_FORMAT_NO_PROBE); 1554 1555 /* Adjust feature flags */ 1556 if (backing_file) { 1557 new_header.features |= QED_F_BACKING_FILE; 1558 1559 if (qed_fmt_is_raw(backing_fmt)) { 1560 new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE; 1561 } 1562 } 1563 1564 /* Calculate new header size */ 1565 backing_file_len = 0; 1566 1567 if (backing_file) { 1568 backing_file_len = strlen(backing_file); 1569 } 1570 1571 buffer_len = sizeof(new_header); 1572 new_header.backing_filename_offset = buffer_len; 1573 new_header.backing_filename_size = backing_file_len; 1574 buffer_len += backing_file_len; 1575 1576 /* Make sure we can rewrite header without failing */ 1577 if (buffer_len > new_header.header_size * new_header.cluster_size) { 1578 return -ENOSPC; 1579 } 1580 1581 /* Prepare new header */ 1582 buffer = g_malloc(buffer_len); 1583 1584 qed_header_cpu_to_le(&new_header, &le_header); 1585 memcpy(buffer, &le_header, sizeof(le_header)); 1586 buffer_len = sizeof(le_header); 1587 1588 if (backing_file) { 1589 memcpy(buffer + buffer_len, backing_file, backing_file_len); 1590 buffer_len += backing_file_len; 1591 } 1592 1593 /* Write new header */ 1594 ret = bdrv_pwrite_sync(bs->file->bs, 0, buffer, buffer_len); 1595 g_free(buffer); 1596 if (ret == 0) { 1597 memcpy(&s->header, &new_header, sizeof(new_header)); 1598 } 1599 return ret; 1600 } 1601 1602 static void bdrv_qed_invalidate_cache(BlockDriverState *bs, Error **errp) 1603 { 1604 BDRVQEDState *s = bs->opaque; 1605 Error *local_err = NULL; 1606 int ret; 1607 1608 bdrv_qed_close(bs); 1609 1610 bdrv_invalidate_cache(bs->file->bs, &local_err); 1611 if (local_err) { 1612 error_propagate(errp, local_err); 1613 return; 1614 } 1615 1616 memset(s, 0, sizeof(BDRVQEDState)); 1617 ret = bdrv_qed_open(bs, NULL, bs->open_flags, &local_err); 1618 if (local_err) { 1619 error_propagate(errp, local_err); 1620 error_prepend(errp, "Could not reopen qed layer: "); 1621 return; 1622 } else if (ret < 0) { 1623 error_setg_errno(errp, -ret, "Could not reopen qed layer"); 1624 return; 1625 } 1626 } 1627 1628 static int bdrv_qed_check(BlockDriverState *bs, BdrvCheckResult *result, 1629 BdrvCheckMode fix) 1630 { 1631 BDRVQEDState *s = bs->opaque; 1632 1633 return qed_check(s, result, !!fix); 1634 } 1635 1636 static QemuOptsList qed_create_opts = { 1637 .name = "qed-create-opts", 1638 .head = QTAILQ_HEAD_INITIALIZER(qed_create_opts.head), 1639 .desc = { 1640 { 1641 .name = BLOCK_OPT_SIZE, 1642 .type = QEMU_OPT_SIZE, 1643 .help = "Virtual disk size" 1644 }, 1645 { 1646 .name = BLOCK_OPT_BACKING_FILE, 1647 .type = QEMU_OPT_STRING, 1648 .help = "File name of a base image" 1649 }, 1650 { 1651 .name = BLOCK_OPT_BACKING_FMT, 1652 .type = QEMU_OPT_STRING, 1653 .help = "Image format of the base image" 1654 }, 1655 { 1656 .name = BLOCK_OPT_CLUSTER_SIZE, 1657 .type = QEMU_OPT_SIZE, 1658 .help = "Cluster size (in bytes)", 1659 .def_value_str = stringify(QED_DEFAULT_CLUSTER_SIZE) 1660 }, 1661 { 1662 .name = BLOCK_OPT_TABLE_SIZE, 1663 .type = QEMU_OPT_SIZE, 1664 .help = "L1/L2 table size (in clusters)" 1665 }, 1666 { /* end of list */ } 1667 } 1668 }; 1669 1670 static BlockDriver bdrv_qed = { 1671 .format_name = "qed", 1672 .instance_size = sizeof(BDRVQEDState), 1673 .create_opts = &qed_create_opts, 1674 .supports_backing = true, 1675 1676 .bdrv_probe = bdrv_qed_probe, 1677 .bdrv_open = bdrv_qed_open, 1678 .bdrv_close = bdrv_qed_close, 1679 .bdrv_reopen_prepare = bdrv_qed_reopen_prepare, 1680 .bdrv_create = bdrv_qed_create, 1681 .bdrv_has_zero_init = bdrv_has_zero_init_1, 1682 .bdrv_co_get_block_status = bdrv_qed_co_get_block_status, 1683 .bdrv_aio_readv = bdrv_qed_aio_readv, 1684 .bdrv_aio_writev = bdrv_qed_aio_writev, 1685 .bdrv_co_write_zeroes = bdrv_qed_co_write_zeroes, 1686 .bdrv_truncate = bdrv_qed_truncate, 1687 .bdrv_getlength = bdrv_qed_getlength, 1688 .bdrv_get_info = bdrv_qed_get_info, 1689 .bdrv_refresh_limits = bdrv_qed_refresh_limits, 1690 .bdrv_change_backing_file = bdrv_qed_change_backing_file, 1691 .bdrv_invalidate_cache = bdrv_qed_invalidate_cache, 1692 .bdrv_check = bdrv_qed_check, 1693 .bdrv_detach_aio_context = bdrv_qed_detach_aio_context, 1694 .bdrv_attach_aio_context = bdrv_qed_attach_aio_context, 1695 .bdrv_drain = bdrv_qed_drain, 1696 }; 1697 1698 static void bdrv_qed_init(void) 1699 { 1700 bdrv_register(&bdrv_qed); 1701 } 1702 1703 block_init(bdrv_qed_init); 1704