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 "qapi/error.h" 17 #include "qemu/timer.h" 18 #include "qemu/bswap.h" 19 #include "trace.h" 20 #include "qed.h" 21 #include "qapi/qmp/qerror.h" 22 #include "migration/migration.h" 23 #include "sysemu/block-backend.h" 24 25 static const AIOCBInfo qed_aiocb_info = { 26 .aiocb_size = sizeof(QEDAIOCB), 27 }; 28 29 static int bdrv_qed_probe(const uint8_t *buf, int buf_size, 30 const char *filename) 31 { 32 const QEDHeader *header = (const QEDHeader *)buf; 33 34 if (buf_size < sizeof(*header)) { 35 return 0; 36 } 37 if (le32_to_cpu(header->magic) != QED_MAGIC) { 38 return 0; 39 } 40 return 100; 41 } 42 43 /** 44 * Check whether an image format is raw 45 * 46 * @fmt: Backing file format, may be NULL 47 */ 48 static bool qed_fmt_is_raw(const char *fmt) 49 { 50 return fmt && strcmp(fmt, "raw") == 0; 51 } 52 53 static void qed_header_le_to_cpu(const QEDHeader *le, QEDHeader *cpu) 54 { 55 cpu->magic = le32_to_cpu(le->magic); 56 cpu->cluster_size = le32_to_cpu(le->cluster_size); 57 cpu->table_size = le32_to_cpu(le->table_size); 58 cpu->header_size = le32_to_cpu(le->header_size); 59 cpu->features = le64_to_cpu(le->features); 60 cpu->compat_features = le64_to_cpu(le->compat_features); 61 cpu->autoclear_features = le64_to_cpu(le->autoclear_features); 62 cpu->l1_table_offset = le64_to_cpu(le->l1_table_offset); 63 cpu->image_size = le64_to_cpu(le->image_size); 64 cpu->backing_filename_offset = le32_to_cpu(le->backing_filename_offset); 65 cpu->backing_filename_size = le32_to_cpu(le->backing_filename_size); 66 } 67 68 static void qed_header_cpu_to_le(const QEDHeader *cpu, QEDHeader *le) 69 { 70 le->magic = cpu_to_le32(cpu->magic); 71 le->cluster_size = cpu_to_le32(cpu->cluster_size); 72 le->table_size = cpu_to_le32(cpu->table_size); 73 le->header_size = cpu_to_le32(cpu->header_size); 74 le->features = cpu_to_le64(cpu->features); 75 le->compat_features = cpu_to_le64(cpu->compat_features); 76 le->autoclear_features = cpu_to_le64(cpu->autoclear_features); 77 le->l1_table_offset = cpu_to_le64(cpu->l1_table_offset); 78 le->image_size = cpu_to_le64(cpu->image_size); 79 le->backing_filename_offset = cpu_to_le32(cpu->backing_filename_offset); 80 le->backing_filename_size = cpu_to_le32(cpu->backing_filename_size); 81 } 82 83 int qed_write_header_sync(BDRVQEDState *s) 84 { 85 QEDHeader le; 86 int ret; 87 88 qed_header_cpu_to_le(&s->header, &le); 89 ret = bdrv_pwrite(s->bs->file, 0, &le, sizeof(le)); 90 if (ret != sizeof(le)) { 91 return ret; 92 } 93 return 0; 94 } 95 96 typedef struct { 97 GenericCB gencb; 98 BDRVQEDState *s; 99 struct iovec iov; 100 QEMUIOVector qiov; 101 int nsectors; 102 uint8_t *buf; 103 } QEDWriteHeaderCB; 104 105 static void qed_write_header_cb(void *opaque, int ret) 106 { 107 QEDWriteHeaderCB *write_header_cb = opaque; 108 109 qemu_vfree(write_header_cb->buf); 110 gencb_complete(write_header_cb, ret); 111 } 112 113 static void qed_write_header_read_cb(void *opaque, int ret) 114 { 115 QEDWriteHeaderCB *write_header_cb = opaque; 116 BDRVQEDState *s = write_header_cb->s; 117 118 if (ret) { 119 qed_write_header_cb(write_header_cb, ret); 120 return; 121 } 122 123 /* Update header */ 124 qed_header_cpu_to_le(&s->header, (QEDHeader *)write_header_cb->buf); 125 126 bdrv_aio_writev(s->bs->file, 0, &write_header_cb->qiov, 127 write_header_cb->nsectors, qed_write_header_cb, 128 write_header_cb); 129 } 130 131 /** 132 * Update header in-place (does not rewrite backing filename or other strings) 133 * 134 * This function only updates known header fields in-place and does not affect 135 * extra data after the QED header. 136 */ 137 static void qed_write_header(BDRVQEDState *s, BlockCompletionFunc cb, 138 void *opaque) 139 { 140 /* We must write full sectors for O_DIRECT but cannot necessarily generate 141 * the data following the header if an unrecognized compat feature is 142 * active. Therefore, first read the sectors containing the header, update 143 * them, and write back. 144 */ 145 146 int nsectors = DIV_ROUND_UP(sizeof(QEDHeader), BDRV_SECTOR_SIZE); 147 size_t len = nsectors * BDRV_SECTOR_SIZE; 148 QEDWriteHeaderCB *write_header_cb = gencb_alloc(sizeof(*write_header_cb), 149 cb, opaque); 150 151 write_header_cb->s = s; 152 write_header_cb->nsectors = nsectors; 153 write_header_cb->buf = qemu_blockalign(s->bs, len); 154 write_header_cb->iov.iov_base = write_header_cb->buf; 155 write_header_cb->iov.iov_len = len; 156 qemu_iovec_init_external(&write_header_cb->qiov, &write_header_cb->iov, 1); 157 158 bdrv_aio_readv(s->bs->file, 0, &write_header_cb->qiov, nsectors, 159 qed_write_header_read_cb, write_header_cb); 160 } 161 162 static uint64_t qed_max_image_size(uint32_t cluster_size, uint32_t table_size) 163 { 164 uint64_t table_entries; 165 uint64_t l2_size; 166 167 table_entries = (table_size * cluster_size) / sizeof(uint64_t); 168 l2_size = table_entries * cluster_size; 169 170 return l2_size * table_entries; 171 } 172 173 static bool qed_is_cluster_size_valid(uint32_t cluster_size) 174 { 175 if (cluster_size < QED_MIN_CLUSTER_SIZE || 176 cluster_size > QED_MAX_CLUSTER_SIZE) { 177 return false; 178 } 179 if (cluster_size & (cluster_size - 1)) { 180 return false; /* not power of 2 */ 181 } 182 return true; 183 } 184 185 static bool qed_is_table_size_valid(uint32_t table_size) 186 { 187 if (table_size < QED_MIN_TABLE_SIZE || 188 table_size > QED_MAX_TABLE_SIZE) { 189 return false; 190 } 191 if (table_size & (table_size - 1)) { 192 return false; /* not power of 2 */ 193 } 194 return true; 195 } 196 197 static bool qed_is_image_size_valid(uint64_t image_size, uint32_t cluster_size, 198 uint32_t table_size) 199 { 200 if (image_size % BDRV_SECTOR_SIZE != 0) { 201 return false; /* not multiple of sector size */ 202 } 203 if (image_size > qed_max_image_size(cluster_size, table_size)) { 204 return false; /* image is too large */ 205 } 206 return true; 207 } 208 209 /** 210 * Read a string of known length from the image file 211 * 212 * @file: Image file 213 * @offset: File offset to start of string, in bytes 214 * @n: String length in bytes 215 * @buf: Destination buffer 216 * @buflen: Destination buffer length in bytes 217 * @ret: 0 on success, -errno on failure 218 * 219 * The string is NUL-terminated. 220 */ 221 static int qed_read_string(BdrvChild *file, uint64_t offset, size_t n, 222 char *buf, size_t buflen) 223 { 224 int ret; 225 if (n >= buflen) { 226 return -EINVAL; 227 } 228 ret = bdrv_pread(file, offset, buf, n); 229 if (ret < 0) { 230 return ret; 231 } 232 buf[n] = '\0'; 233 return 0; 234 } 235 236 /** 237 * Allocate new clusters 238 * 239 * @s: QED state 240 * @n: Number of contiguous clusters to allocate 241 * @ret: Offset of first allocated cluster 242 * 243 * This function only produces the offset where the new clusters should be 244 * written. It updates BDRVQEDState but does not make any changes to the image 245 * file. 246 */ 247 static uint64_t qed_alloc_clusters(BDRVQEDState *s, unsigned int n) 248 { 249 uint64_t offset = s->file_size; 250 s->file_size += n * s->header.cluster_size; 251 return offset; 252 } 253 254 QEDTable *qed_alloc_table(BDRVQEDState *s) 255 { 256 /* Honor O_DIRECT memory alignment requirements */ 257 return qemu_blockalign(s->bs, 258 s->header.cluster_size * s->header.table_size); 259 } 260 261 /** 262 * Allocate a new zeroed L2 table 263 */ 264 static CachedL2Table *qed_new_l2_table(BDRVQEDState *s) 265 { 266 CachedL2Table *l2_table = qed_alloc_l2_cache_entry(&s->l2_cache); 267 268 l2_table->table = qed_alloc_table(s); 269 l2_table->offset = qed_alloc_clusters(s, s->header.table_size); 270 271 memset(l2_table->table->offsets, 0, 272 s->header.cluster_size * s->header.table_size); 273 return l2_table; 274 } 275 276 static void qed_aio_next_io(QEDAIOCB *acb, int ret); 277 278 static void qed_aio_start_io(QEDAIOCB *acb) 279 { 280 qed_aio_next_io(acb, 0); 281 } 282 283 static void qed_aio_next_io_cb(void *opaque, int ret) 284 { 285 QEDAIOCB *acb = opaque; 286 287 qed_aio_next_io(acb, ret); 288 } 289 290 static void qed_plug_allocating_write_reqs(BDRVQEDState *s) 291 { 292 assert(!s->allocating_write_reqs_plugged); 293 294 s->allocating_write_reqs_plugged = true; 295 } 296 297 static void qed_unplug_allocating_write_reqs(BDRVQEDState *s) 298 { 299 QEDAIOCB *acb; 300 301 assert(s->allocating_write_reqs_plugged); 302 303 s->allocating_write_reqs_plugged = false; 304 305 acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs); 306 if (acb) { 307 qed_aio_start_io(acb); 308 } 309 } 310 311 static void qed_finish_clear_need_check(void *opaque, int ret) 312 { 313 /* Do nothing */ 314 } 315 316 static void qed_flush_after_clear_need_check(void *opaque, int ret) 317 { 318 BDRVQEDState *s = opaque; 319 320 bdrv_aio_flush(s->bs, qed_finish_clear_need_check, s); 321 322 /* No need to wait until flush completes */ 323 qed_unplug_allocating_write_reqs(s); 324 } 325 326 static void qed_clear_need_check(void *opaque, int ret) 327 { 328 BDRVQEDState *s = opaque; 329 330 if (ret) { 331 qed_unplug_allocating_write_reqs(s); 332 return; 333 } 334 335 s->header.features &= ~QED_F_NEED_CHECK; 336 qed_write_header(s, qed_flush_after_clear_need_check, s); 337 } 338 339 static void qed_need_check_timer_cb(void *opaque) 340 { 341 BDRVQEDState *s = opaque; 342 343 /* The timer should only fire when allocating writes have drained */ 344 assert(!QSIMPLEQ_FIRST(&s->allocating_write_reqs)); 345 346 trace_qed_need_check_timer_cb(s); 347 348 qed_acquire(s); 349 qed_plug_allocating_write_reqs(s); 350 351 /* Ensure writes are on disk before clearing flag */ 352 bdrv_aio_flush(s->bs->file->bs, qed_clear_need_check, s); 353 qed_release(s); 354 } 355 356 void qed_acquire(BDRVQEDState *s) 357 { 358 aio_context_acquire(bdrv_get_aio_context(s->bs)); 359 } 360 361 void qed_release(BDRVQEDState *s) 362 { 363 aio_context_release(bdrv_get_aio_context(s->bs)); 364 } 365 366 static void qed_start_need_check_timer(BDRVQEDState *s) 367 { 368 trace_qed_start_need_check_timer(s); 369 370 /* Use QEMU_CLOCK_VIRTUAL so we don't alter the image file while suspended for 371 * migration. 372 */ 373 timer_mod(s->need_check_timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + 374 NANOSECONDS_PER_SECOND * QED_NEED_CHECK_TIMEOUT); 375 } 376 377 /* It's okay to call this multiple times or when no timer is started */ 378 static void qed_cancel_need_check_timer(BDRVQEDState *s) 379 { 380 trace_qed_cancel_need_check_timer(s); 381 timer_del(s->need_check_timer); 382 } 383 384 static void bdrv_qed_detach_aio_context(BlockDriverState *bs) 385 { 386 BDRVQEDState *s = bs->opaque; 387 388 qed_cancel_need_check_timer(s); 389 timer_free(s->need_check_timer); 390 } 391 392 static void bdrv_qed_attach_aio_context(BlockDriverState *bs, 393 AioContext *new_context) 394 { 395 BDRVQEDState *s = bs->opaque; 396 397 s->need_check_timer = aio_timer_new(new_context, 398 QEMU_CLOCK_VIRTUAL, SCALE_NS, 399 qed_need_check_timer_cb, s); 400 if (s->header.features & QED_F_NEED_CHECK) { 401 qed_start_need_check_timer(s); 402 } 403 } 404 405 static void bdrv_qed_drain(BlockDriverState *bs) 406 { 407 BDRVQEDState *s = bs->opaque; 408 409 /* Fire the timer immediately in order to start doing I/O as soon as the 410 * header is flushed. 411 */ 412 if (s->need_check_timer && timer_pending(s->need_check_timer)) { 413 qed_cancel_need_check_timer(s); 414 qed_need_check_timer_cb(s); 415 } 416 } 417 418 static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags, 419 Error **errp) 420 { 421 BDRVQEDState *s = bs->opaque; 422 QEDHeader le_header; 423 int64_t file_size; 424 int ret; 425 426 s->bs = bs; 427 QSIMPLEQ_INIT(&s->allocating_write_reqs); 428 429 ret = bdrv_pread(bs->file, 0, &le_header, sizeof(le_header)); 430 if (ret < 0) { 431 return ret; 432 } 433 qed_header_le_to_cpu(&le_header, &s->header); 434 435 if (s->header.magic != QED_MAGIC) { 436 error_setg(errp, "Image not in QED format"); 437 return -EINVAL; 438 } 439 if (s->header.features & ~QED_FEATURE_MASK) { 440 /* image uses unsupported feature bits */ 441 error_setg(errp, "Unsupported QED features: %" PRIx64, 442 s->header.features & ~QED_FEATURE_MASK); 443 return -ENOTSUP; 444 } 445 if (!qed_is_cluster_size_valid(s->header.cluster_size)) { 446 return -EINVAL; 447 } 448 449 /* Round down file size to the last cluster */ 450 file_size = bdrv_getlength(bs->file->bs); 451 if (file_size < 0) { 452 return file_size; 453 } 454 s->file_size = qed_start_of_cluster(s, file_size); 455 456 if (!qed_is_table_size_valid(s->header.table_size)) { 457 return -EINVAL; 458 } 459 if (!qed_is_image_size_valid(s->header.image_size, 460 s->header.cluster_size, 461 s->header.table_size)) { 462 return -EINVAL; 463 } 464 if (!qed_check_table_offset(s, s->header.l1_table_offset)) { 465 return -EINVAL; 466 } 467 468 s->table_nelems = (s->header.cluster_size * s->header.table_size) / 469 sizeof(uint64_t); 470 s->l2_shift = ctz32(s->header.cluster_size); 471 s->l2_mask = s->table_nelems - 1; 472 s->l1_shift = s->l2_shift + ctz32(s->table_nelems); 473 474 /* Header size calculation must not overflow uint32_t */ 475 if (s->header.header_size > UINT32_MAX / s->header.cluster_size) { 476 return -EINVAL; 477 } 478 479 if ((s->header.features & QED_F_BACKING_FILE)) { 480 if ((uint64_t)s->header.backing_filename_offset + 481 s->header.backing_filename_size > 482 s->header.cluster_size * s->header.header_size) { 483 return -EINVAL; 484 } 485 486 ret = qed_read_string(bs->file, s->header.backing_filename_offset, 487 s->header.backing_filename_size, bs->backing_file, 488 sizeof(bs->backing_file)); 489 if (ret < 0) { 490 return ret; 491 } 492 493 if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) { 494 pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw"); 495 } 496 } 497 498 /* Reset unknown autoclear feature bits. This is a backwards 499 * compatibility mechanism that allows images to be opened by older 500 * programs, which "knock out" unknown feature bits. When an image is 501 * opened by a newer program again it can detect that the autoclear 502 * feature is no longer valid. 503 */ 504 if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 && 505 !bdrv_is_read_only(bs->file->bs) && !(flags & BDRV_O_INACTIVE)) { 506 s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK; 507 508 ret = qed_write_header_sync(s); 509 if (ret) { 510 return ret; 511 } 512 513 /* From here on only known autoclear feature bits are valid */ 514 bdrv_flush(bs->file->bs); 515 } 516 517 s->l1_table = qed_alloc_table(s); 518 qed_init_l2_cache(&s->l2_cache); 519 520 ret = qed_read_l1_table_sync(s); 521 if (ret) { 522 goto out; 523 } 524 525 /* If image was not closed cleanly, check consistency */ 526 if (!(flags & BDRV_O_CHECK) && (s->header.features & QED_F_NEED_CHECK)) { 527 /* Read-only images cannot be fixed. There is no risk of corruption 528 * since write operations are not possible. Therefore, allow 529 * potentially inconsistent images to be opened read-only. This can 530 * aid data recovery from an otherwise inconsistent image. 531 */ 532 if (!bdrv_is_read_only(bs->file->bs) && 533 !(flags & BDRV_O_INACTIVE)) { 534 BdrvCheckResult result = {0}; 535 536 ret = qed_check(s, &result, true); 537 if (ret) { 538 goto out; 539 } 540 } 541 } 542 543 bdrv_qed_attach_aio_context(bs, bdrv_get_aio_context(bs)); 544 545 out: 546 if (ret) { 547 qed_free_l2_cache(&s->l2_cache); 548 qemu_vfree(s->l1_table); 549 } 550 return ret; 551 } 552 553 static void bdrv_qed_refresh_limits(BlockDriverState *bs, Error **errp) 554 { 555 BDRVQEDState *s = bs->opaque; 556 557 bs->bl.pwrite_zeroes_alignment = s->header.cluster_size; 558 } 559 560 /* We have nothing to do for QED reopen, stubs just return 561 * success */ 562 static int bdrv_qed_reopen_prepare(BDRVReopenState *state, 563 BlockReopenQueue *queue, Error **errp) 564 { 565 return 0; 566 } 567 568 static void bdrv_qed_close(BlockDriverState *bs) 569 { 570 BDRVQEDState *s = bs->opaque; 571 572 bdrv_qed_detach_aio_context(bs); 573 574 /* Ensure writes reach stable storage */ 575 bdrv_flush(bs->file->bs); 576 577 /* Clean shutdown, no check required on next open */ 578 if (s->header.features & QED_F_NEED_CHECK) { 579 s->header.features &= ~QED_F_NEED_CHECK; 580 qed_write_header_sync(s); 581 } 582 583 qed_free_l2_cache(&s->l2_cache); 584 qemu_vfree(s->l1_table); 585 } 586 587 static int qed_create(const char *filename, uint32_t cluster_size, 588 uint64_t image_size, uint32_t table_size, 589 const char *backing_file, const char *backing_fmt, 590 QemuOpts *opts, Error **errp) 591 { 592 QEDHeader header = { 593 .magic = QED_MAGIC, 594 .cluster_size = cluster_size, 595 .table_size = table_size, 596 .header_size = 1, 597 .features = 0, 598 .compat_features = 0, 599 .l1_table_offset = cluster_size, 600 .image_size = image_size, 601 }; 602 QEDHeader le_header; 603 uint8_t *l1_table = NULL; 604 size_t l1_size = header.cluster_size * header.table_size; 605 Error *local_err = NULL; 606 int ret = 0; 607 BlockBackend *blk; 608 609 ret = bdrv_create_file(filename, opts, &local_err); 610 if (ret < 0) { 611 error_propagate(errp, local_err); 612 return ret; 613 } 614 615 blk = blk_new_open(filename, NULL, NULL, 616 BDRV_O_RDWR | BDRV_O_PROTOCOL, &local_err); 617 if (blk == NULL) { 618 error_propagate(errp, local_err); 619 return -EIO; 620 } 621 622 blk_set_allow_write_beyond_eof(blk, true); 623 624 /* File must start empty and grow, check truncate is supported */ 625 ret = blk_truncate(blk, 0); 626 if (ret < 0) { 627 goto out; 628 } 629 630 if (backing_file) { 631 header.features |= QED_F_BACKING_FILE; 632 header.backing_filename_offset = sizeof(le_header); 633 header.backing_filename_size = strlen(backing_file); 634 635 if (qed_fmt_is_raw(backing_fmt)) { 636 header.features |= QED_F_BACKING_FORMAT_NO_PROBE; 637 } 638 } 639 640 qed_header_cpu_to_le(&header, &le_header); 641 ret = blk_pwrite(blk, 0, &le_header, sizeof(le_header), 0); 642 if (ret < 0) { 643 goto out; 644 } 645 ret = blk_pwrite(blk, sizeof(le_header), backing_file, 646 header.backing_filename_size, 0); 647 if (ret < 0) { 648 goto out; 649 } 650 651 l1_table = g_malloc0(l1_size); 652 ret = blk_pwrite(blk, header.l1_table_offset, l1_table, l1_size, 0); 653 if (ret < 0) { 654 goto out; 655 } 656 657 ret = 0; /* success */ 658 out: 659 g_free(l1_table); 660 blk_unref(blk); 661 return ret; 662 } 663 664 static int bdrv_qed_create(const char *filename, QemuOpts *opts, Error **errp) 665 { 666 uint64_t image_size = 0; 667 uint32_t cluster_size = QED_DEFAULT_CLUSTER_SIZE; 668 uint32_t table_size = QED_DEFAULT_TABLE_SIZE; 669 char *backing_file = NULL; 670 char *backing_fmt = NULL; 671 int ret; 672 673 image_size = ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0), 674 BDRV_SECTOR_SIZE); 675 backing_file = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FILE); 676 backing_fmt = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FMT); 677 cluster_size = qemu_opt_get_size_del(opts, 678 BLOCK_OPT_CLUSTER_SIZE, 679 QED_DEFAULT_CLUSTER_SIZE); 680 table_size = qemu_opt_get_size_del(opts, BLOCK_OPT_TABLE_SIZE, 681 QED_DEFAULT_TABLE_SIZE); 682 683 if (!qed_is_cluster_size_valid(cluster_size)) { 684 error_setg(errp, "QED cluster size must be within range [%u, %u] " 685 "and power of 2", 686 QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE); 687 ret = -EINVAL; 688 goto finish; 689 } 690 if (!qed_is_table_size_valid(table_size)) { 691 error_setg(errp, "QED table size must be within range [%u, %u] " 692 "and power of 2", 693 QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE); 694 ret = -EINVAL; 695 goto finish; 696 } 697 if (!qed_is_image_size_valid(image_size, cluster_size, table_size)) { 698 error_setg(errp, "QED image size must be a non-zero multiple of " 699 "cluster size and less than %" PRIu64 " bytes", 700 qed_max_image_size(cluster_size, table_size)); 701 ret = -EINVAL; 702 goto finish; 703 } 704 705 ret = qed_create(filename, cluster_size, image_size, table_size, 706 backing_file, backing_fmt, opts, errp); 707 708 finish: 709 g_free(backing_file); 710 g_free(backing_fmt); 711 return ret; 712 } 713 714 typedef struct { 715 BlockDriverState *bs; 716 Coroutine *co; 717 uint64_t pos; 718 int64_t status; 719 int *pnum; 720 BlockDriverState **file; 721 } QEDIsAllocatedCB; 722 723 static void qed_is_allocated_cb(void *opaque, int ret, uint64_t offset, size_t len) 724 { 725 QEDIsAllocatedCB *cb = opaque; 726 BDRVQEDState *s = cb->bs->opaque; 727 *cb->pnum = len / BDRV_SECTOR_SIZE; 728 switch (ret) { 729 case QED_CLUSTER_FOUND: 730 offset |= qed_offset_into_cluster(s, cb->pos); 731 cb->status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID | offset; 732 *cb->file = cb->bs->file->bs; 733 break; 734 case QED_CLUSTER_ZERO: 735 cb->status = BDRV_BLOCK_ZERO; 736 break; 737 case QED_CLUSTER_L2: 738 case QED_CLUSTER_L1: 739 cb->status = 0; 740 break; 741 default: 742 assert(ret < 0); 743 cb->status = ret; 744 break; 745 } 746 747 if (cb->co) { 748 aio_co_wake(cb->co); 749 } 750 } 751 752 static int64_t coroutine_fn bdrv_qed_co_get_block_status(BlockDriverState *bs, 753 int64_t sector_num, 754 int nb_sectors, int *pnum, 755 BlockDriverState **file) 756 { 757 BDRVQEDState *s = bs->opaque; 758 size_t len = (size_t)nb_sectors * BDRV_SECTOR_SIZE; 759 QEDIsAllocatedCB cb = { 760 .bs = bs, 761 .pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE, 762 .status = BDRV_BLOCK_OFFSET_MASK, 763 .pnum = pnum, 764 .file = file, 765 }; 766 QEDRequest request = { .l2_table = NULL }; 767 768 qed_find_cluster(s, &request, cb.pos, len, qed_is_allocated_cb, &cb); 769 770 /* Now sleep if the callback wasn't invoked immediately */ 771 while (cb.status == BDRV_BLOCK_OFFSET_MASK) { 772 cb.co = qemu_coroutine_self(); 773 qemu_coroutine_yield(); 774 } 775 776 qed_unref_l2_cache_entry(request.l2_table); 777 778 return cb.status; 779 } 780 781 static BDRVQEDState *acb_to_s(QEDAIOCB *acb) 782 { 783 return acb->common.bs->opaque; 784 } 785 786 /** 787 * Read from the backing file or zero-fill if no backing file 788 * 789 * @s: QED state 790 * @pos: Byte position in device 791 * @qiov: Destination I/O vector 792 * @backing_qiov: Possibly shortened copy of qiov, to be allocated here 793 * @cb: Completion function 794 * @opaque: User data for completion function 795 * 796 * This function reads qiov->size bytes starting at pos from the backing file. 797 * If there is no backing file then zeroes are read. 798 */ 799 static void qed_read_backing_file(BDRVQEDState *s, uint64_t pos, 800 QEMUIOVector *qiov, 801 QEMUIOVector **backing_qiov, 802 BlockCompletionFunc *cb, void *opaque) 803 { 804 uint64_t backing_length = 0; 805 size_t size; 806 807 /* If there is a backing file, get its length. Treat the absence of a 808 * backing file like a zero length backing file. 809 */ 810 if (s->bs->backing) { 811 int64_t l = bdrv_getlength(s->bs->backing->bs); 812 if (l < 0) { 813 cb(opaque, l); 814 return; 815 } 816 backing_length = l; 817 } 818 819 /* Zero all sectors if reading beyond the end of the backing file */ 820 if (pos >= backing_length || 821 pos + qiov->size > backing_length) { 822 qemu_iovec_memset(qiov, 0, 0, qiov->size); 823 } 824 825 /* Complete now if there are no backing file sectors to read */ 826 if (pos >= backing_length) { 827 cb(opaque, 0); 828 return; 829 } 830 831 /* If the read straddles the end of the backing file, shorten it */ 832 size = MIN((uint64_t)backing_length - pos, qiov->size); 833 834 assert(*backing_qiov == NULL); 835 *backing_qiov = g_new(QEMUIOVector, 1); 836 qemu_iovec_init(*backing_qiov, qiov->niov); 837 qemu_iovec_concat(*backing_qiov, qiov, 0, size); 838 839 BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO); 840 bdrv_aio_readv(s->bs->backing, pos / BDRV_SECTOR_SIZE, 841 *backing_qiov, size / BDRV_SECTOR_SIZE, cb, opaque); 842 } 843 844 typedef struct { 845 GenericCB gencb; 846 BDRVQEDState *s; 847 QEMUIOVector qiov; 848 QEMUIOVector *backing_qiov; 849 struct iovec iov; 850 uint64_t offset; 851 } CopyFromBackingFileCB; 852 853 static void qed_copy_from_backing_file_cb(void *opaque, int ret) 854 { 855 CopyFromBackingFileCB *copy_cb = opaque; 856 qemu_vfree(copy_cb->iov.iov_base); 857 gencb_complete(©_cb->gencb, ret); 858 } 859 860 static void qed_copy_from_backing_file_write(void *opaque, int ret) 861 { 862 CopyFromBackingFileCB *copy_cb = opaque; 863 BDRVQEDState *s = copy_cb->s; 864 865 if (copy_cb->backing_qiov) { 866 qemu_iovec_destroy(copy_cb->backing_qiov); 867 g_free(copy_cb->backing_qiov); 868 copy_cb->backing_qiov = NULL; 869 } 870 871 if (ret) { 872 qed_copy_from_backing_file_cb(copy_cb, ret); 873 return; 874 } 875 876 BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE); 877 bdrv_aio_writev(s->bs->file, copy_cb->offset / BDRV_SECTOR_SIZE, 878 ©_cb->qiov, copy_cb->qiov.size / BDRV_SECTOR_SIZE, 879 qed_copy_from_backing_file_cb, copy_cb); 880 } 881 882 /** 883 * Copy data from backing file into the image 884 * 885 * @s: QED state 886 * @pos: Byte position in device 887 * @len: Number of bytes 888 * @offset: Byte offset in image file 889 * @cb: Completion function 890 * @opaque: User data for completion function 891 */ 892 static void qed_copy_from_backing_file(BDRVQEDState *s, uint64_t pos, 893 uint64_t len, uint64_t offset, 894 BlockCompletionFunc *cb, 895 void *opaque) 896 { 897 CopyFromBackingFileCB *copy_cb; 898 899 /* Skip copy entirely if there is no work to do */ 900 if (len == 0) { 901 cb(opaque, 0); 902 return; 903 } 904 905 copy_cb = gencb_alloc(sizeof(*copy_cb), cb, opaque); 906 copy_cb->s = s; 907 copy_cb->offset = offset; 908 copy_cb->backing_qiov = NULL; 909 copy_cb->iov.iov_base = qemu_blockalign(s->bs, len); 910 copy_cb->iov.iov_len = len; 911 qemu_iovec_init_external(©_cb->qiov, ©_cb->iov, 1); 912 913 qed_read_backing_file(s, pos, ©_cb->qiov, ©_cb->backing_qiov, 914 qed_copy_from_backing_file_write, copy_cb); 915 } 916 917 /** 918 * Link one or more contiguous clusters into a table 919 * 920 * @s: QED state 921 * @table: L2 table 922 * @index: First cluster index 923 * @n: Number of contiguous clusters 924 * @cluster: First cluster offset 925 * 926 * The cluster offset may be an allocated byte offset in the image file, the 927 * zero cluster marker, or the unallocated cluster marker. 928 */ 929 static void qed_update_l2_table(BDRVQEDState *s, QEDTable *table, int index, 930 unsigned int n, uint64_t cluster) 931 { 932 int i; 933 for (i = index; i < index + n; i++) { 934 table->offsets[i] = cluster; 935 if (!qed_offset_is_unalloc_cluster(cluster) && 936 !qed_offset_is_zero_cluster(cluster)) { 937 cluster += s->header.cluster_size; 938 } 939 } 940 } 941 942 static void qed_aio_complete_bh(void *opaque) 943 { 944 QEDAIOCB *acb = opaque; 945 BDRVQEDState *s = acb_to_s(acb); 946 BlockCompletionFunc *cb = acb->common.cb; 947 void *user_opaque = acb->common.opaque; 948 int ret = acb->bh_ret; 949 950 qemu_aio_unref(acb); 951 952 /* Invoke callback */ 953 qed_acquire(s); 954 cb(user_opaque, ret); 955 qed_release(s); 956 } 957 958 static void qed_aio_complete(QEDAIOCB *acb, int ret) 959 { 960 BDRVQEDState *s = acb_to_s(acb); 961 962 trace_qed_aio_complete(s, acb, ret); 963 964 /* Free resources */ 965 qemu_iovec_destroy(&acb->cur_qiov); 966 qed_unref_l2_cache_entry(acb->request.l2_table); 967 968 /* Free the buffer we may have allocated for zero writes */ 969 if (acb->flags & QED_AIOCB_ZERO) { 970 qemu_vfree(acb->qiov->iov[0].iov_base); 971 acb->qiov->iov[0].iov_base = NULL; 972 } 973 974 /* Arrange for a bh to invoke the completion function */ 975 acb->bh_ret = ret; 976 aio_bh_schedule_oneshot(bdrv_get_aio_context(acb->common.bs), 977 qed_aio_complete_bh, acb); 978 979 /* Start next allocating write request waiting behind this one. Note that 980 * requests enqueue themselves when they first hit an unallocated cluster 981 * but they wait until the entire request is finished before waking up the 982 * next request in the queue. This ensures that we don't cycle through 983 * requests multiple times but rather finish one at a time completely. 984 */ 985 if (acb == QSIMPLEQ_FIRST(&s->allocating_write_reqs)) { 986 QSIMPLEQ_REMOVE_HEAD(&s->allocating_write_reqs, next); 987 acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs); 988 if (acb) { 989 qed_aio_start_io(acb); 990 } else if (s->header.features & QED_F_NEED_CHECK) { 991 qed_start_need_check_timer(s); 992 } 993 } 994 } 995 996 /** 997 * Commit the current L2 table to the cache 998 */ 999 static void qed_commit_l2_update(void *opaque, int ret) 1000 { 1001 QEDAIOCB *acb = opaque; 1002 BDRVQEDState *s = acb_to_s(acb); 1003 CachedL2Table *l2_table = acb->request.l2_table; 1004 uint64_t l2_offset = l2_table->offset; 1005 1006 qed_commit_l2_cache_entry(&s->l2_cache, l2_table); 1007 1008 /* This is guaranteed to succeed because we just committed the entry to the 1009 * cache. 1010 */ 1011 acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset); 1012 assert(acb->request.l2_table != NULL); 1013 1014 qed_aio_next_io(acb, ret); 1015 } 1016 1017 /** 1018 * Update L1 table with new L2 table offset and write it out 1019 */ 1020 static void qed_aio_write_l1_update(void *opaque, int ret) 1021 { 1022 QEDAIOCB *acb = opaque; 1023 BDRVQEDState *s = acb_to_s(acb); 1024 int index; 1025 1026 if (ret) { 1027 qed_aio_complete(acb, ret); 1028 return; 1029 } 1030 1031 index = qed_l1_index(s, acb->cur_pos); 1032 s->l1_table->offsets[index] = acb->request.l2_table->offset; 1033 1034 qed_write_l1_table(s, index, 1, qed_commit_l2_update, acb); 1035 } 1036 1037 /** 1038 * Update L2 table with new cluster offsets and write them out 1039 */ 1040 static void qed_aio_write_l2_update(QEDAIOCB *acb, int ret, uint64_t offset) 1041 { 1042 BDRVQEDState *s = acb_to_s(acb); 1043 bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1; 1044 int index; 1045 1046 if (ret) { 1047 goto err; 1048 } 1049 1050 if (need_alloc) { 1051 qed_unref_l2_cache_entry(acb->request.l2_table); 1052 acb->request.l2_table = qed_new_l2_table(s); 1053 } 1054 1055 index = qed_l2_index(s, acb->cur_pos); 1056 qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters, 1057 offset); 1058 1059 if (need_alloc) { 1060 /* Write out the whole new L2 table */ 1061 qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true, 1062 qed_aio_write_l1_update, acb); 1063 } else { 1064 /* Write out only the updated part of the L2 table */ 1065 qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters, false, 1066 qed_aio_next_io_cb, acb); 1067 } 1068 return; 1069 1070 err: 1071 qed_aio_complete(acb, ret); 1072 } 1073 1074 static void qed_aio_write_l2_update_cb(void *opaque, int ret) 1075 { 1076 QEDAIOCB *acb = opaque; 1077 qed_aio_write_l2_update(acb, ret, acb->cur_cluster); 1078 } 1079 1080 /** 1081 * Flush new data clusters before updating the L2 table 1082 * 1083 * This flush is necessary when a backing file is in use. A crash during an 1084 * allocating write could result in empty clusters in the image. If the write 1085 * only touched a subregion of the cluster, then backing image sectors have 1086 * been lost in the untouched region. The solution is to flush after writing a 1087 * new data cluster and before updating the L2 table. 1088 */ 1089 static void qed_aio_write_flush_before_l2_update(void *opaque, int ret) 1090 { 1091 QEDAIOCB *acb = opaque; 1092 BDRVQEDState *s = acb_to_s(acb); 1093 1094 if (!bdrv_aio_flush(s->bs->file->bs, qed_aio_write_l2_update_cb, opaque)) { 1095 qed_aio_complete(acb, -EIO); 1096 } 1097 } 1098 1099 /** 1100 * Write data to the image file 1101 */ 1102 static void qed_aio_write_main(void *opaque, int ret) 1103 { 1104 QEDAIOCB *acb = opaque; 1105 BDRVQEDState *s = acb_to_s(acb); 1106 uint64_t offset = acb->cur_cluster + 1107 qed_offset_into_cluster(s, acb->cur_pos); 1108 BlockCompletionFunc *next_fn; 1109 1110 trace_qed_aio_write_main(s, acb, ret, offset, acb->cur_qiov.size); 1111 1112 if (ret) { 1113 qed_aio_complete(acb, ret); 1114 return; 1115 } 1116 1117 if (acb->find_cluster_ret == QED_CLUSTER_FOUND) { 1118 next_fn = qed_aio_next_io_cb; 1119 } else { 1120 if (s->bs->backing) { 1121 next_fn = qed_aio_write_flush_before_l2_update; 1122 } else { 1123 next_fn = qed_aio_write_l2_update_cb; 1124 } 1125 } 1126 1127 BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO); 1128 bdrv_aio_writev(s->bs->file, offset / BDRV_SECTOR_SIZE, 1129 &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE, 1130 next_fn, acb); 1131 } 1132 1133 /** 1134 * Populate back untouched region of new data cluster 1135 */ 1136 static void qed_aio_write_postfill(void *opaque, int ret) 1137 { 1138 QEDAIOCB *acb = opaque; 1139 BDRVQEDState *s = acb_to_s(acb); 1140 uint64_t start = acb->cur_pos + acb->cur_qiov.size; 1141 uint64_t len = 1142 qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start; 1143 uint64_t offset = acb->cur_cluster + 1144 qed_offset_into_cluster(s, acb->cur_pos) + 1145 acb->cur_qiov.size; 1146 1147 if (ret) { 1148 qed_aio_complete(acb, ret); 1149 return; 1150 } 1151 1152 trace_qed_aio_write_postfill(s, acb, start, len, offset); 1153 qed_copy_from_backing_file(s, start, len, offset, 1154 qed_aio_write_main, acb); 1155 } 1156 1157 /** 1158 * Populate front untouched region of new data cluster 1159 */ 1160 static void qed_aio_write_prefill(void *opaque, int ret) 1161 { 1162 QEDAIOCB *acb = opaque; 1163 BDRVQEDState *s = acb_to_s(acb); 1164 uint64_t start = qed_start_of_cluster(s, acb->cur_pos); 1165 uint64_t len = qed_offset_into_cluster(s, acb->cur_pos); 1166 1167 trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster); 1168 qed_copy_from_backing_file(s, start, len, acb->cur_cluster, 1169 qed_aio_write_postfill, acb); 1170 } 1171 1172 /** 1173 * Check if the QED_F_NEED_CHECK bit should be set during allocating write 1174 */ 1175 static bool qed_should_set_need_check(BDRVQEDState *s) 1176 { 1177 /* The flush before L2 update path ensures consistency */ 1178 if (s->bs->backing) { 1179 return false; 1180 } 1181 1182 return !(s->header.features & QED_F_NEED_CHECK); 1183 } 1184 1185 static void qed_aio_write_zero_cluster(void *opaque, int ret) 1186 { 1187 QEDAIOCB *acb = opaque; 1188 1189 if (ret) { 1190 qed_aio_complete(acb, ret); 1191 return; 1192 } 1193 1194 qed_aio_write_l2_update(acb, 0, 1); 1195 } 1196 1197 /** 1198 * Write new data cluster 1199 * 1200 * @acb: Write request 1201 * @len: Length in bytes 1202 * 1203 * This path is taken when writing to previously unallocated clusters. 1204 */ 1205 static void qed_aio_write_alloc(QEDAIOCB *acb, size_t len) 1206 { 1207 BDRVQEDState *s = acb_to_s(acb); 1208 BlockCompletionFunc *cb; 1209 1210 /* Cancel timer when the first allocating request comes in */ 1211 if (QSIMPLEQ_EMPTY(&s->allocating_write_reqs)) { 1212 qed_cancel_need_check_timer(s); 1213 } 1214 1215 /* Freeze this request if another allocating write is in progress */ 1216 if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs)) { 1217 QSIMPLEQ_INSERT_TAIL(&s->allocating_write_reqs, acb, next); 1218 } 1219 if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs) || 1220 s->allocating_write_reqs_plugged) { 1221 return; /* wait for existing request to finish */ 1222 } 1223 1224 acb->cur_nclusters = qed_bytes_to_clusters(s, 1225 qed_offset_into_cluster(s, acb->cur_pos) + len); 1226 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len); 1227 1228 if (acb->flags & QED_AIOCB_ZERO) { 1229 /* Skip ahead if the clusters are already zero */ 1230 if (acb->find_cluster_ret == QED_CLUSTER_ZERO) { 1231 qed_aio_start_io(acb); 1232 return; 1233 } 1234 1235 cb = qed_aio_write_zero_cluster; 1236 } else { 1237 cb = qed_aio_write_prefill; 1238 acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters); 1239 } 1240 1241 if (qed_should_set_need_check(s)) { 1242 s->header.features |= QED_F_NEED_CHECK; 1243 qed_write_header(s, cb, acb); 1244 } else { 1245 cb(acb, 0); 1246 } 1247 } 1248 1249 /** 1250 * Write data cluster in place 1251 * 1252 * @acb: Write request 1253 * @offset: Cluster offset in bytes 1254 * @len: Length in bytes 1255 * 1256 * This path is taken when writing to already allocated clusters. 1257 */ 1258 static void qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset, size_t len) 1259 { 1260 /* Allocate buffer for zero writes */ 1261 if (acb->flags & QED_AIOCB_ZERO) { 1262 struct iovec *iov = acb->qiov->iov; 1263 1264 if (!iov->iov_base) { 1265 iov->iov_base = qemu_try_blockalign(acb->common.bs, iov->iov_len); 1266 if (iov->iov_base == NULL) { 1267 qed_aio_complete(acb, -ENOMEM); 1268 return; 1269 } 1270 memset(iov->iov_base, 0, iov->iov_len); 1271 } 1272 } 1273 1274 /* Calculate the I/O vector */ 1275 acb->cur_cluster = offset; 1276 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len); 1277 1278 /* Do the actual write */ 1279 qed_aio_write_main(acb, 0); 1280 } 1281 1282 /** 1283 * Write data cluster 1284 * 1285 * @opaque: Write request 1286 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1, 1287 * or -errno 1288 * @offset: Cluster offset in bytes 1289 * @len: Length in bytes 1290 * 1291 * Callback from qed_find_cluster(). 1292 */ 1293 static void qed_aio_write_data(void *opaque, int ret, 1294 uint64_t offset, size_t len) 1295 { 1296 QEDAIOCB *acb = opaque; 1297 1298 trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len); 1299 1300 acb->find_cluster_ret = ret; 1301 1302 switch (ret) { 1303 case QED_CLUSTER_FOUND: 1304 qed_aio_write_inplace(acb, offset, len); 1305 break; 1306 1307 case QED_CLUSTER_L2: 1308 case QED_CLUSTER_L1: 1309 case QED_CLUSTER_ZERO: 1310 qed_aio_write_alloc(acb, len); 1311 break; 1312 1313 default: 1314 qed_aio_complete(acb, ret); 1315 break; 1316 } 1317 } 1318 1319 /** 1320 * Read data cluster 1321 * 1322 * @opaque: Read request 1323 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1, 1324 * or -errno 1325 * @offset: Cluster offset in bytes 1326 * @len: Length in bytes 1327 * 1328 * Callback from qed_find_cluster(). 1329 */ 1330 static void qed_aio_read_data(void *opaque, int ret, 1331 uint64_t offset, size_t len) 1332 { 1333 QEDAIOCB *acb = opaque; 1334 BDRVQEDState *s = acb_to_s(acb); 1335 BlockDriverState *bs = acb->common.bs; 1336 1337 /* Adjust offset into cluster */ 1338 offset += qed_offset_into_cluster(s, acb->cur_pos); 1339 1340 trace_qed_aio_read_data(s, acb, ret, offset, len); 1341 1342 if (ret < 0) { 1343 goto err; 1344 } 1345 1346 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len); 1347 1348 /* Handle zero cluster and backing file reads */ 1349 if (ret == QED_CLUSTER_ZERO) { 1350 qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size); 1351 qed_aio_start_io(acb); 1352 return; 1353 } else if (ret != QED_CLUSTER_FOUND) { 1354 qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov, 1355 &acb->backing_qiov, qed_aio_next_io_cb, acb); 1356 return; 1357 } 1358 1359 BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO); 1360 bdrv_aio_readv(bs->file, offset / BDRV_SECTOR_SIZE, 1361 &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE, 1362 qed_aio_next_io_cb, acb); 1363 return; 1364 1365 err: 1366 qed_aio_complete(acb, ret); 1367 } 1368 1369 /** 1370 * Begin next I/O or complete the request 1371 */ 1372 static void qed_aio_next_io(QEDAIOCB *acb, int ret) 1373 { 1374 BDRVQEDState *s = acb_to_s(acb); 1375 QEDFindClusterFunc *io_fn = (acb->flags & QED_AIOCB_WRITE) ? 1376 qed_aio_write_data : qed_aio_read_data; 1377 1378 trace_qed_aio_next_io(s, acb, ret, acb->cur_pos + acb->cur_qiov.size); 1379 1380 if (acb->backing_qiov) { 1381 qemu_iovec_destroy(acb->backing_qiov); 1382 g_free(acb->backing_qiov); 1383 acb->backing_qiov = NULL; 1384 } 1385 1386 /* Handle I/O error */ 1387 if (ret) { 1388 qed_aio_complete(acb, ret); 1389 return; 1390 } 1391 1392 acb->qiov_offset += acb->cur_qiov.size; 1393 acb->cur_pos += acb->cur_qiov.size; 1394 qemu_iovec_reset(&acb->cur_qiov); 1395 1396 /* Complete request */ 1397 if (acb->cur_pos >= acb->end_pos) { 1398 qed_aio_complete(acb, 0); 1399 return; 1400 } 1401 1402 /* Find next cluster and start I/O */ 1403 qed_find_cluster(s, &acb->request, 1404 acb->cur_pos, acb->end_pos - acb->cur_pos, 1405 io_fn, acb); 1406 } 1407 1408 static BlockAIOCB *qed_aio_setup(BlockDriverState *bs, 1409 int64_t sector_num, 1410 QEMUIOVector *qiov, int nb_sectors, 1411 BlockCompletionFunc *cb, 1412 void *opaque, int flags) 1413 { 1414 QEDAIOCB *acb = qemu_aio_get(&qed_aiocb_info, bs, cb, opaque); 1415 1416 trace_qed_aio_setup(bs->opaque, acb, sector_num, nb_sectors, 1417 opaque, flags); 1418 1419 acb->flags = flags; 1420 acb->qiov = qiov; 1421 acb->qiov_offset = 0; 1422 acb->cur_pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE; 1423 acb->end_pos = acb->cur_pos + nb_sectors * BDRV_SECTOR_SIZE; 1424 acb->backing_qiov = NULL; 1425 acb->request.l2_table = NULL; 1426 qemu_iovec_init(&acb->cur_qiov, qiov->niov); 1427 1428 /* Start request */ 1429 qed_aio_start_io(acb); 1430 return &acb->common; 1431 } 1432 1433 static BlockAIOCB *bdrv_qed_aio_readv(BlockDriverState *bs, 1434 int64_t sector_num, 1435 QEMUIOVector *qiov, int nb_sectors, 1436 BlockCompletionFunc *cb, 1437 void *opaque) 1438 { 1439 return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb, opaque, 0); 1440 } 1441 1442 static BlockAIOCB *bdrv_qed_aio_writev(BlockDriverState *bs, 1443 int64_t sector_num, 1444 QEMUIOVector *qiov, int nb_sectors, 1445 BlockCompletionFunc *cb, 1446 void *opaque) 1447 { 1448 return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb, 1449 opaque, QED_AIOCB_WRITE); 1450 } 1451 1452 typedef struct { 1453 Coroutine *co; 1454 int ret; 1455 bool done; 1456 } QEDWriteZeroesCB; 1457 1458 static void coroutine_fn qed_co_pwrite_zeroes_cb(void *opaque, int ret) 1459 { 1460 QEDWriteZeroesCB *cb = opaque; 1461 1462 cb->done = true; 1463 cb->ret = ret; 1464 if (cb->co) { 1465 aio_co_wake(cb->co); 1466 } 1467 } 1468 1469 static int coroutine_fn bdrv_qed_co_pwrite_zeroes(BlockDriverState *bs, 1470 int64_t offset, 1471 int count, 1472 BdrvRequestFlags flags) 1473 { 1474 BlockAIOCB *blockacb; 1475 BDRVQEDState *s = bs->opaque; 1476 QEDWriteZeroesCB cb = { .done = false }; 1477 QEMUIOVector qiov; 1478 struct iovec iov; 1479 1480 /* Fall back if the request is not aligned */ 1481 if (qed_offset_into_cluster(s, offset) || 1482 qed_offset_into_cluster(s, count)) { 1483 return -ENOTSUP; 1484 } 1485 1486 /* Zero writes start without an I/O buffer. If a buffer becomes necessary 1487 * then it will be allocated during request processing. 1488 */ 1489 iov.iov_base = NULL; 1490 iov.iov_len = count; 1491 1492 qemu_iovec_init_external(&qiov, &iov, 1); 1493 blockacb = qed_aio_setup(bs, offset >> BDRV_SECTOR_BITS, &qiov, 1494 count >> BDRV_SECTOR_BITS, 1495 qed_co_pwrite_zeroes_cb, &cb, 1496 QED_AIOCB_WRITE | QED_AIOCB_ZERO); 1497 if (!blockacb) { 1498 return -EIO; 1499 } 1500 if (!cb.done) { 1501 cb.co = qemu_coroutine_self(); 1502 qemu_coroutine_yield(); 1503 } 1504 assert(cb.done); 1505 return cb.ret; 1506 } 1507 1508 static int bdrv_qed_truncate(BlockDriverState *bs, int64_t offset) 1509 { 1510 BDRVQEDState *s = bs->opaque; 1511 uint64_t old_image_size; 1512 int ret; 1513 1514 if (!qed_is_image_size_valid(offset, s->header.cluster_size, 1515 s->header.table_size)) { 1516 return -EINVAL; 1517 } 1518 1519 /* Shrinking is currently not supported */ 1520 if ((uint64_t)offset < s->header.image_size) { 1521 return -ENOTSUP; 1522 } 1523 1524 old_image_size = s->header.image_size; 1525 s->header.image_size = offset; 1526 ret = qed_write_header_sync(s); 1527 if (ret < 0) { 1528 s->header.image_size = old_image_size; 1529 } 1530 return ret; 1531 } 1532 1533 static int64_t bdrv_qed_getlength(BlockDriverState *bs) 1534 { 1535 BDRVQEDState *s = bs->opaque; 1536 return s->header.image_size; 1537 } 1538 1539 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi) 1540 { 1541 BDRVQEDState *s = bs->opaque; 1542 1543 memset(bdi, 0, sizeof(*bdi)); 1544 bdi->cluster_size = s->header.cluster_size; 1545 bdi->is_dirty = s->header.features & QED_F_NEED_CHECK; 1546 bdi->unallocated_blocks_are_zero = true; 1547 bdi->can_write_zeroes_with_unmap = true; 1548 return 0; 1549 } 1550 1551 static int bdrv_qed_change_backing_file(BlockDriverState *bs, 1552 const char *backing_file, 1553 const char *backing_fmt) 1554 { 1555 BDRVQEDState *s = bs->opaque; 1556 QEDHeader new_header, le_header; 1557 void *buffer; 1558 size_t buffer_len, backing_file_len; 1559 int ret; 1560 1561 /* Refuse to set backing filename if unknown compat feature bits are 1562 * active. If the image uses an unknown compat feature then we may not 1563 * know the layout of data following the header structure and cannot safely 1564 * add a new string. 1565 */ 1566 if (backing_file && (s->header.compat_features & 1567 ~QED_COMPAT_FEATURE_MASK)) { 1568 return -ENOTSUP; 1569 } 1570 1571 memcpy(&new_header, &s->header, sizeof(new_header)); 1572 1573 new_header.features &= ~(QED_F_BACKING_FILE | 1574 QED_F_BACKING_FORMAT_NO_PROBE); 1575 1576 /* Adjust feature flags */ 1577 if (backing_file) { 1578 new_header.features |= QED_F_BACKING_FILE; 1579 1580 if (qed_fmt_is_raw(backing_fmt)) { 1581 new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE; 1582 } 1583 } 1584 1585 /* Calculate new header size */ 1586 backing_file_len = 0; 1587 1588 if (backing_file) { 1589 backing_file_len = strlen(backing_file); 1590 } 1591 1592 buffer_len = sizeof(new_header); 1593 new_header.backing_filename_offset = buffer_len; 1594 new_header.backing_filename_size = backing_file_len; 1595 buffer_len += backing_file_len; 1596 1597 /* Make sure we can rewrite header without failing */ 1598 if (buffer_len > new_header.header_size * new_header.cluster_size) { 1599 return -ENOSPC; 1600 } 1601 1602 /* Prepare new header */ 1603 buffer = g_malloc(buffer_len); 1604 1605 qed_header_cpu_to_le(&new_header, &le_header); 1606 memcpy(buffer, &le_header, sizeof(le_header)); 1607 buffer_len = sizeof(le_header); 1608 1609 if (backing_file) { 1610 memcpy(buffer + buffer_len, backing_file, backing_file_len); 1611 buffer_len += backing_file_len; 1612 } 1613 1614 /* Write new header */ 1615 ret = bdrv_pwrite_sync(bs->file, 0, buffer, buffer_len); 1616 g_free(buffer); 1617 if (ret == 0) { 1618 memcpy(&s->header, &new_header, sizeof(new_header)); 1619 } 1620 return ret; 1621 } 1622 1623 static void bdrv_qed_invalidate_cache(BlockDriverState *bs, Error **errp) 1624 { 1625 BDRVQEDState *s = bs->opaque; 1626 Error *local_err = NULL; 1627 int ret; 1628 1629 bdrv_qed_close(bs); 1630 1631 memset(s, 0, sizeof(BDRVQEDState)); 1632 ret = bdrv_qed_open(bs, NULL, bs->open_flags, &local_err); 1633 if (local_err) { 1634 error_propagate(errp, local_err); 1635 error_prepend(errp, "Could not reopen qed layer: "); 1636 return; 1637 } else if (ret < 0) { 1638 error_setg_errno(errp, -ret, "Could not reopen qed layer"); 1639 return; 1640 } 1641 } 1642 1643 static int bdrv_qed_check(BlockDriverState *bs, BdrvCheckResult *result, 1644 BdrvCheckMode fix) 1645 { 1646 BDRVQEDState *s = bs->opaque; 1647 1648 return qed_check(s, result, !!fix); 1649 } 1650 1651 static QemuOptsList qed_create_opts = { 1652 .name = "qed-create-opts", 1653 .head = QTAILQ_HEAD_INITIALIZER(qed_create_opts.head), 1654 .desc = { 1655 { 1656 .name = BLOCK_OPT_SIZE, 1657 .type = QEMU_OPT_SIZE, 1658 .help = "Virtual disk size" 1659 }, 1660 { 1661 .name = BLOCK_OPT_BACKING_FILE, 1662 .type = QEMU_OPT_STRING, 1663 .help = "File name of a base image" 1664 }, 1665 { 1666 .name = BLOCK_OPT_BACKING_FMT, 1667 .type = QEMU_OPT_STRING, 1668 .help = "Image format of the base image" 1669 }, 1670 { 1671 .name = BLOCK_OPT_CLUSTER_SIZE, 1672 .type = QEMU_OPT_SIZE, 1673 .help = "Cluster size (in bytes)", 1674 .def_value_str = stringify(QED_DEFAULT_CLUSTER_SIZE) 1675 }, 1676 { 1677 .name = BLOCK_OPT_TABLE_SIZE, 1678 .type = QEMU_OPT_SIZE, 1679 .help = "L1/L2 table size (in clusters)" 1680 }, 1681 { /* end of list */ } 1682 } 1683 }; 1684 1685 static BlockDriver bdrv_qed = { 1686 .format_name = "qed", 1687 .instance_size = sizeof(BDRVQEDState), 1688 .create_opts = &qed_create_opts, 1689 .supports_backing = true, 1690 1691 .bdrv_probe = bdrv_qed_probe, 1692 .bdrv_open = bdrv_qed_open, 1693 .bdrv_close = bdrv_qed_close, 1694 .bdrv_reopen_prepare = bdrv_qed_reopen_prepare, 1695 .bdrv_create = bdrv_qed_create, 1696 .bdrv_has_zero_init = bdrv_has_zero_init_1, 1697 .bdrv_co_get_block_status = bdrv_qed_co_get_block_status, 1698 .bdrv_aio_readv = bdrv_qed_aio_readv, 1699 .bdrv_aio_writev = bdrv_qed_aio_writev, 1700 .bdrv_co_pwrite_zeroes = bdrv_qed_co_pwrite_zeroes, 1701 .bdrv_truncate = bdrv_qed_truncate, 1702 .bdrv_getlength = bdrv_qed_getlength, 1703 .bdrv_get_info = bdrv_qed_get_info, 1704 .bdrv_refresh_limits = bdrv_qed_refresh_limits, 1705 .bdrv_change_backing_file = bdrv_qed_change_backing_file, 1706 .bdrv_invalidate_cache = bdrv_qed_invalidate_cache, 1707 .bdrv_check = bdrv_qed_check, 1708 .bdrv_detach_aio_context = bdrv_qed_detach_aio_context, 1709 .bdrv_attach_aio_context = bdrv_qed_attach_aio_context, 1710 .bdrv_drain = bdrv_qed_drain, 1711 }; 1712 1713 static void bdrv_qed_init(void) 1714 { 1715 bdrv_register(&bdrv_qed); 1716 } 1717 1718 block_init(bdrv_qed_init); 1719