1 /* 2 * QEMU Enhanced Disk Format 3 * 4 * Copyright IBM, Corp. 2010 5 * 6 * Authors: 7 * Stefan Hajnoczi <stefanha@linux.vnet.ibm.com> 8 * Anthony Liguori <aliguori@us.ibm.com> 9 * 10 * This work is licensed under the terms of the GNU LGPL, version 2 or later. 11 * See the COPYING.LIB file in the top-level directory. 12 * 13 */ 14 15 #include "qemu/timer.h" 16 #include "trace.h" 17 #include "qed.h" 18 #include "qapi/qmp/qerror.h" 19 #include "migration/migration.h" 20 21 static const AIOCBInfo qed_aiocb_info = { 22 .aiocb_size = sizeof(QEDAIOCB), 23 }; 24 25 static int bdrv_qed_probe(const uint8_t *buf, int buf_size, 26 const char *filename) 27 { 28 const QEDHeader *header = (const QEDHeader *)buf; 29 30 if (buf_size < sizeof(*header)) { 31 return 0; 32 } 33 if (le32_to_cpu(header->magic) != QED_MAGIC) { 34 return 0; 35 } 36 return 100; 37 } 38 39 /** 40 * Check whether an image format is raw 41 * 42 * @fmt: Backing file format, may be NULL 43 */ 44 static bool qed_fmt_is_raw(const char *fmt) 45 { 46 return fmt && strcmp(fmt, "raw") == 0; 47 } 48 49 static void qed_header_le_to_cpu(const QEDHeader *le, QEDHeader *cpu) 50 { 51 cpu->magic = le32_to_cpu(le->magic); 52 cpu->cluster_size = le32_to_cpu(le->cluster_size); 53 cpu->table_size = le32_to_cpu(le->table_size); 54 cpu->header_size = le32_to_cpu(le->header_size); 55 cpu->features = le64_to_cpu(le->features); 56 cpu->compat_features = le64_to_cpu(le->compat_features); 57 cpu->autoclear_features = le64_to_cpu(le->autoclear_features); 58 cpu->l1_table_offset = le64_to_cpu(le->l1_table_offset); 59 cpu->image_size = le64_to_cpu(le->image_size); 60 cpu->backing_filename_offset = le32_to_cpu(le->backing_filename_offset); 61 cpu->backing_filename_size = le32_to_cpu(le->backing_filename_size); 62 } 63 64 static void qed_header_cpu_to_le(const QEDHeader *cpu, QEDHeader *le) 65 { 66 le->magic = cpu_to_le32(cpu->magic); 67 le->cluster_size = cpu_to_le32(cpu->cluster_size); 68 le->table_size = cpu_to_le32(cpu->table_size); 69 le->header_size = cpu_to_le32(cpu->header_size); 70 le->features = cpu_to_le64(cpu->features); 71 le->compat_features = cpu_to_le64(cpu->compat_features); 72 le->autoclear_features = cpu_to_le64(cpu->autoclear_features); 73 le->l1_table_offset = cpu_to_le64(cpu->l1_table_offset); 74 le->image_size = cpu_to_le64(cpu->image_size); 75 le->backing_filename_offset = cpu_to_le32(cpu->backing_filename_offset); 76 le->backing_filename_size = cpu_to_le32(cpu->backing_filename_size); 77 } 78 79 int qed_write_header_sync(BDRVQEDState *s) 80 { 81 QEDHeader le; 82 int ret; 83 84 qed_header_cpu_to_le(&s->header, &le); 85 ret = bdrv_pwrite(s->bs->file, 0, &le, sizeof(le)); 86 if (ret != sizeof(le)) { 87 return ret; 88 } 89 return 0; 90 } 91 92 typedef struct { 93 GenericCB gencb; 94 BDRVQEDState *s; 95 struct iovec iov; 96 QEMUIOVector qiov; 97 int nsectors; 98 uint8_t *buf; 99 } QEDWriteHeaderCB; 100 101 static void qed_write_header_cb(void *opaque, int ret) 102 { 103 QEDWriteHeaderCB *write_header_cb = opaque; 104 105 qemu_vfree(write_header_cb->buf); 106 gencb_complete(write_header_cb, ret); 107 } 108 109 static void qed_write_header_read_cb(void *opaque, int ret) 110 { 111 QEDWriteHeaderCB *write_header_cb = opaque; 112 BDRVQEDState *s = write_header_cb->s; 113 114 if (ret) { 115 qed_write_header_cb(write_header_cb, ret); 116 return; 117 } 118 119 /* Update header */ 120 qed_header_cpu_to_le(&s->header, (QEDHeader *)write_header_cb->buf); 121 122 bdrv_aio_writev(s->bs->file, 0, &write_header_cb->qiov, 123 write_header_cb->nsectors, qed_write_header_cb, 124 write_header_cb); 125 } 126 127 /** 128 * Update header in-place (does not rewrite backing filename or other strings) 129 * 130 * This function only updates known header fields in-place and does not affect 131 * extra data after the QED header. 132 */ 133 static void qed_write_header(BDRVQEDState *s, BlockCompletionFunc cb, 134 void *opaque) 135 { 136 /* We must write full sectors for O_DIRECT but cannot necessarily generate 137 * the data following the header if an unrecognized compat feature is 138 * active. Therefore, first read the sectors containing the header, update 139 * them, and write back. 140 */ 141 142 int nsectors = (sizeof(QEDHeader) + BDRV_SECTOR_SIZE - 1) / 143 BDRV_SECTOR_SIZE; 144 size_t len = nsectors * BDRV_SECTOR_SIZE; 145 QEDWriteHeaderCB *write_header_cb = gencb_alloc(sizeof(*write_header_cb), 146 cb, opaque); 147 148 write_header_cb->s = s; 149 write_header_cb->nsectors = nsectors; 150 write_header_cb->buf = qemu_blockalign(s->bs, len); 151 write_header_cb->iov.iov_base = write_header_cb->buf; 152 write_header_cb->iov.iov_len = len; 153 qemu_iovec_init_external(&write_header_cb->qiov, &write_header_cb->iov, 1); 154 155 bdrv_aio_readv(s->bs->file, 0, &write_header_cb->qiov, nsectors, 156 qed_write_header_read_cb, write_header_cb); 157 } 158 159 static uint64_t qed_max_image_size(uint32_t cluster_size, uint32_t table_size) 160 { 161 uint64_t table_entries; 162 uint64_t l2_size; 163 164 table_entries = (table_size * cluster_size) / sizeof(uint64_t); 165 l2_size = table_entries * cluster_size; 166 167 return l2_size * table_entries; 168 } 169 170 static bool qed_is_cluster_size_valid(uint32_t cluster_size) 171 { 172 if (cluster_size < QED_MIN_CLUSTER_SIZE || 173 cluster_size > QED_MAX_CLUSTER_SIZE) { 174 return false; 175 } 176 if (cluster_size & (cluster_size - 1)) { 177 return false; /* not power of 2 */ 178 } 179 return true; 180 } 181 182 static bool qed_is_table_size_valid(uint32_t table_size) 183 { 184 if (table_size < QED_MIN_TABLE_SIZE || 185 table_size > QED_MAX_TABLE_SIZE) { 186 return false; 187 } 188 if (table_size & (table_size - 1)) { 189 return false; /* not power of 2 */ 190 } 191 return true; 192 } 193 194 static bool qed_is_image_size_valid(uint64_t image_size, uint32_t cluster_size, 195 uint32_t table_size) 196 { 197 if (image_size % BDRV_SECTOR_SIZE != 0) { 198 return false; /* not multiple of sector size */ 199 } 200 if (image_size > qed_max_image_size(cluster_size, table_size)) { 201 return false; /* image is too large */ 202 } 203 return true; 204 } 205 206 /** 207 * Read a string of known length from the image file 208 * 209 * @file: Image file 210 * @offset: File offset to start of string, in bytes 211 * @n: String length in bytes 212 * @buf: Destination buffer 213 * @buflen: Destination buffer length in bytes 214 * @ret: 0 on success, -errno on failure 215 * 216 * The string is NUL-terminated. 217 */ 218 static int qed_read_string(BlockDriverState *file, uint64_t offset, size_t n, 219 char *buf, size_t buflen) 220 { 221 int ret; 222 if (n >= buflen) { 223 return -EINVAL; 224 } 225 ret = bdrv_pread(file, offset, buf, n); 226 if (ret < 0) { 227 return ret; 228 } 229 buf[n] = '\0'; 230 return 0; 231 } 232 233 /** 234 * Allocate new clusters 235 * 236 * @s: QED state 237 * @n: Number of contiguous clusters to allocate 238 * @ret: Offset of first allocated cluster 239 * 240 * This function only produces the offset where the new clusters should be 241 * written. It updates BDRVQEDState but does not make any changes to the image 242 * file. 243 */ 244 static uint64_t qed_alloc_clusters(BDRVQEDState *s, unsigned int n) 245 { 246 uint64_t offset = s->file_size; 247 s->file_size += n * s->header.cluster_size; 248 return offset; 249 } 250 251 QEDTable *qed_alloc_table(BDRVQEDState *s) 252 { 253 /* Honor O_DIRECT memory alignment requirements */ 254 return qemu_blockalign(s->bs, 255 s->header.cluster_size * s->header.table_size); 256 } 257 258 /** 259 * Allocate a new zeroed L2 table 260 */ 261 static CachedL2Table *qed_new_l2_table(BDRVQEDState *s) 262 { 263 CachedL2Table *l2_table = qed_alloc_l2_cache_entry(&s->l2_cache); 264 265 l2_table->table = qed_alloc_table(s); 266 l2_table->offset = qed_alloc_clusters(s, s->header.table_size); 267 268 memset(l2_table->table->offsets, 0, 269 s->header.cluster_size * s->header.table_size); 270 return l2_table; 271 } 272 273 static void qed_aio_next_io(void *opaque, int ret); 274 275 static void qed_plug_allocating_write_reqs(BDRVQEDState *s) 276 { 277 assert(!s->allocating_write_reqs_plugged); 278 279 s->allocating_write_reqs_plugged = true; 280 } 281 282 static void qed_unplug_allocating_write_reqs(BDRVQEDState *s) 283 { 284 QEDAIOCB *acb; 285 286 assert(s->allocating_write_reqs_plugged); 287 288 s->allocating_write_reqs_plugged = false; 289 290 acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs); 291 if (acb) { 292 qed_aio_next_io(acb, 0); 293 } 294 } 295 296 static void qed_finish_clear_need_check(void *opaque, int ret) 297 { 298 /* Do nothing */ 299 } 300 301 static void qed_flush_after_clear_need_check(void *opaque, int ret) 302 { 303 BDRVQEDState *s = opaque; 304 305 bdrv_aio_flush(s->bs, qed_finish_clear_need_check, s); 306 307 /* No need to wait until flush completes */ 308 qed_unplug_allocating_write_reqs(s); 309 } 310 311 static void qed_clear_need_check(void *opaque, int ret) 312 { 313 BDRVQEDState *s = opaque; 314 315 if (ret) { 316 qed_unplug_allocating_write_reqs(s); 317 return; 318 } 319 320 s->header.features &= ~QED_F_NEED_CHECK; 321 qed_write_header(s, qed_flush_after_clear_need_check, s); 322 } 323 324 static void qed_need_check_timer_cb(void *opaque) 325 { 326 BDRVQEDState *s = opaque; 327 328 /* The timer should only fire when allocating writes have drained */ 329 assert(!QSIMPLEQ_FIRST(&s->allocating_write_reqs)); 330 331 trace_qed_need_check_timer_cb(s); 332 333 qed_plug_allocating_write_reqs(s); 334 335 /* Ensure writes are on disk before clearing flag */ 336 bdrv_aio_flush(s->bs, qed_clear_need_check, s); 337 } 338 339 static void qed_start_need_check_timer(BDRVQEDState *s) 340 { 341 trace_qed_start_need_check_timer(s); 342 343 /* Use QEMU_CLOCK_VIRTUAL so we don't alter the image file while suspended for 344 * migration. 345 */ 346 timer_mod(s->need_check_timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + 347 get_ticks_per_sec() * QED_NEED_CHECK_TIMEOUT); 348 } 349 350 /* It's okay to call this multiple times or when no timer is started */ 351 static void qed_cancel_need_check_timer(BDRVQEDState *s) 352 { 353 trace_qed_cancel_need_check_timer(s); 354 timer_del(s->need_check_timer); 355 } 356 357 static void bdrv_qed_rebind(BlockDriverState *bs) 358 { 359 BDRVQEDState *s = bs->opaque; 360 s->bs = bs; 361 } 362 363 static void bdrv_qed_detach_aio_context(BlockDriverState *bs) 364 { 365 BDRVQEDState *s = bs->opaque; 366 367 qed_cancel_need_check_timer(s); 368 timer_free(s->need_check_timer); 369 } 370 371 static void bdrv_qed_attach_aio_context(BlockDriverState *bs, 372 AioContext *new_context) 373 { 374 BDRVQEDState *s = bs->opaque; 375 376 s->need_check_timer = aio_timer_new(new_context, 377 QEMU_CLOCK_VIRTUAL, SCALE_NS, 378 qed_need_check_timer_cb, s); 379 if (s->header.features & QED_F_NEED_CHECK) { 380 qed_start_need_check_timer(s); 381 } 382 } 383 384 static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags, 385 Error **errp) 386 { 387 BDRVQEDState *s = bs->opaque; 388 QEDHeader le_header; 389 int64_t file_size; 390 int ret; 391 392 s->bs = bs; 393 QSIMPLEQ_INIT(&s->allocating_write_reqs); 394 395 ret = bdrv_pread(bs->file, 0, &le_header, sizeof(le_header)); 396 if (ret < 0) { 397 return ret; 398 } 399 qed_header_le_to_cpu(&le_header, &s->header); 400 401 if (s->header.magic != QED_MAGIC) { 402 error_setg(errp, "Image not in QED format"); 403 return -EINVAL; 404 } 405 if (s->header.features & ~QED_FEATURE_MASK) { 406 /* image uses unsupported feature bits */ 407 char buf[64]; 408 snprintf(buf, sizeof(buf), "%" PRIx64, 409 s->header.features & ~QED_FEATURE_MASK); 410 error_set(errp, QERR_UNKNOWN_BLOCK_FORMAT_FEATURE, 411 bdrv_get_device_name(bs), "QED", buf); 412 return -ENOTSUP; 413 } 414 if (!qed_is_cluster_size_valid(s->header.cluster_size)) { 415 return -EINVAL; 416 } 417 418 /* Round down file size to the last cluster */ 419 file_size = bdrv_getlength(bs->file); 420 if (file_size < 0) { 421 return file_size; 422 } 423 s->file_size = qed_start_of_cluster(s, file_size); 424 425 if (!qed_is_table_size_valid(s->header.table_size)) { 426 return -EINVAL; 427 } 428 if (!qed_is_image_size_valid(s->header.image_size, 429 s->header.cluster_size, 430 s->header.table_size)) { 431 return -EINVAL; 432 } 433 if (!qed_check_table_offset(s, s->header.l1_table_offset)) { 434 return -EINVAL; 435 } 436 437 s->table_nelems = (s->header.cluster_size * s->header.table_size) / 438 sizeof(uint64_t); 439 s->l2_shift = ctz32(s->header.cluster_size); 440 s->l2_mask = s->table_nelems - 1; 441 s->l1_shift = s->l2_shift + ctz32(s->table_nelems); 442 443 /* Header size calculation must not overflow uint32_t */ 444 if (s->header.header_size > UINT32_MAX / s->header.cluster_size) { 445 return -EINVAL; 446 } 447 448 if ((s->header.features & QED_F_BACKING_FILE)) { 449 if ((uint64_t)s->header.backing_filename_offset + 450 s->header.backing_filename_size > 451 s->header.cluster_size * s->header.header_size) { 452 return -EINVAL; 453 } 454 455 ret = qed_read_string(bs->file, s->header.backing_filename_offset, 456 s->header.backing_filename_size, bs->backing_file, 457 sizeof(bs->backing_file)); 458 if (ret < 0) { 459 return ret; 460 } 461 462 if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) { 463 pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw"); 464 } 465 } 466 467 /* Reset unknown autoclear feature bits. This is a backwards 468 * compatibility mechanism that allows images to be opened by older 469 * programs, which "knock out" unknown feature bits. When an image is 470 * opened by a newer program again it can detect that the autoclear 471 * feature is no longer valid. 472 */ 473 if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 && 474 !bdrv_is_read_only(bs->file) && !(flags & BDRV_O_INCOMING)) { 475 s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK; 476 477 ret = qed_write_header_sync(s); 478 if (ret) { 479 return ret; 480 } 481 482 /* From here on only known autoclear feature bits are valid */ 483 bdrv_flush(bs->file); 484 } 485 486 s->l1_table = qed_alloc_table(s); 487 qed_init_l2_cache(&s->l2_cache); 488 489 ret = qed_read_l1_table_sync(s); 490 if (ret) { 491 goto out; 492 } 493 494 /* If image was not closed cleanly, check consistency */ 495 if (!(flags & BDRV_O_CHECK) && (s->header.features & QED_F_NEED_CHECK)) { 496 /* Read-only images cannot be fixed. There is no risk of corruption 497 * since write operations are not possible. Therefore, allow 498 * potentially inconsistent images to be opened read-only. This can 499 * aid data recovery from an otherwise inconsistent image. 500 */ 501 if (!bdrv_is_read_only(bs->file) && 502 !(flags & BDRV_O_INCOMING)) { 503 BdrvCheckResult result = {0}; 504 505 ret = qed_check(s, &result, true); 506 if (ret) { 507 goto out; 508 } 509 } 510 } 511 512 bdrv_qed_attach_aio_context(bs, bdrv_get_aio_context(bs)); 513 514 out: 515 if (ret) { 516 qed_free_l2_cache(&s->l2_cache); 517 qemu_vfree(s->l1_table); 518 } 519 return ret; 520 } 521 522 static void bdrv_qed_refresh_limits(BlockDriverState *bs, Error **errp) 523 { 524 BDRVQEDState *s = bs->opaque; 525 526 bs->bl.write_zeroes_alignment = s->header.cluster_size >> BDRV_SECTOR_BITS; 527 } 528 529 /* We have nothing to do for QED reopen, stubs just return 530 * success */ 531 static int bdrv_qed_reopen_prepare(BDRVReopenState *state, 532 BlockReopenQueue *queue, Error **errp) 533 { 534 return 0; 535 } 536 537 static void bdrv_qed_close(BlockDriverState *bs) 538 { 539 BDRVQEDState *s = bs->opaque; 540 541 bdrv_qed_detach_aio_context(bs); 542 543 /* Ensure writes reach stable storage */ 544 bdrv_flush(bs->file); 545 546 /* Clean shutdown, no check required on next open */ 547 if (s->header.features & QED_F_NEED_CHECK) { 548 s->header.features &= ~QED_F_NEED_CHECK; 549 qed_write_header_sync(s); 550 } 551 552 qed_free_l2_cache(&s->l2_cache); 553 qemu_vfree(s->l1_table); 554 } 555 556 static int qed_create(const char *filename, uint32_t cluster_size, 557 uint64_t image_size, uint32_t table_size, 558 const char *backing_file, const char *backing_fmt, 559 QemuOpts *opts, Error **errp) 560 { 561 QEDHeader header = { 562 .magic = QED_MAGIC, 563 .cluster_size = cluster_size, 564 .table_size = table_size, 565 .header_size = 1, 566 .features = 0, 567 .compat_features = 0, 568 .l1_table_offset = cluster_size, 569 .image_size = image_size, 570 }; 571 QEDHeader le_header; 572 uint8_t *l1_table = NULL; 573 size_t l1_size = header.cluster_size * header.table_size; 574 Error *local_err = NULL; 575 int ret = 0; 576 BlockDriverState *bs; 577 578 ret = bdrv_create_file(filename, opts, &local_err); 579 if (ret < 0) { 580 error_propagate(errp, local_err); 581 return ret; 582 } 583 584 bs = NULL; 585 ret = bdrv_open(&bs, filename, NULL, NULL, 586 BDRV_O_RDWR | BDRV_O_CACHE_WB | BDRV_O_PROTOCOL, NULL, 587 &local_err); 588 if (ret < 0) { 589 error_propagate(errp, local_err); 590 return ret; 591 } 592 593 /* File must start empty and grow, check truncate is supported */ 594 ret = bdrv_truncate(bs, 0); 595 if (ret < 0) { 596 goto out; 597 } 598 599 if (backing_file) { 600 header.features |= QED_F_BACKING_FILE; 601 header.backing_filename_offset = sizeof(le_header); 602 header.backing_filename_size = strlen(backing_file); 603 604 if (qed_fmt_is_raw(backing_fmt)) { 605 header.features |= QED_F_BACKING_FORMAT_NO_PROBE; 606 } 607 } 608 609 qed_header_cpu_to_le(&header, &le_header); 610 ret = bdrv_pwrite(bs, 0, &le_header, sizeof(le_header)); 611 if (ret < 0) { 612 goto out; 613 } 614 ret = bdrv_pwrite(bs, sizeof(le_header), backing_file, 615 header.backing_filename_size); 616 if (ret < 0) { 617 goto out; 618 } 619 620 l1_table = g_malloc0(l1_size); 621 ret = bdrv_pwrite(bs, header.l1_table_offset, l1_table, l1_size); 622 if (ret < 0) { 623 goto out; 624 } 625 626 ret = 0; /* success */ 627 out: 628 g_free(l1_table); 629 bdrv_unref(bs); 630 return ret; 631 } 632 633 static int bdrv_qed_create(const char *filename, QemuOpts *opts, Error **errp) 634 { 635 uint64_t image_size = 0; 636 uint32_t cluster_size = QED_DEFAULT_CLUSTER_SIZE; 637 uint32_t table_size = QED_DEFAULT_TABLE_SIZE; 638 char *backing_file = NULL; 639 char *backing_fmt = NULL; 640 int ret; 641 642 image_size = ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0), 643 BDRV_SECTOR_SIZE); 644 backing_file = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FILE); 645 backing_fmt = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FMT); 646 cluster_size = qemu_opt_get_size_del(opts, 647 BLOCK_OPT_CLUSTER_SIZE, 648 QED_DEFAULT_CLUSTER_SIZE); 649 table_size = qemu_opt_get_size_del(opts, BLOCK_OPT_TABLE_SIZE, 650 QED_DEFAULT_TABLE_SIZE); 651 652 if (!qed_is_cluster_size_valid(cluster_size)) { 653 error_setg(errp, "QED cluster size must be within range [%u, %u] " 654 "and power of 2", 655 QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE); 656 ret = -EINVAL; 657 goto finish; 658 } 659 if (!qed_is_table_size_valid(table_size)) { 660 error_setg(errp, "QED table size must be within range [%u, %u] " 661 "and power of 2", 662 QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE); 663 ret = -EINVAL; 664 goto finish; 665 } 666 if (!qed_is_image_size_valid(image_size, cluster_size, table_size)) { 667 error_setg(errp, "QED image size must be a non-zero multiple of " 668 "cluster size and less than %" PRIu64 " bytes", 669 qed_max_image_size(cluster_size, table_size)); 670 ret = -EINVAL; 671 goto finish; 672 } 673 674 ret = qed_create(filename, cluster_size, image_size, table_size, 675 backing_file, backing_fmt, opts, errp); 676 677 finish: 678 g_free(backing_file); 679 g_free(backing_fmt); 680 return ret; 681 } 682 683 typedef struct { 684 BlockDriverState *bs; 685 Coroutine *co; 686 uint64_t pos; 687 int64_t status; 688 int *pnum; 689 } QEDIsAllocatedCB; 690 691 static void qed_is_allocated_cb(void *opaque, int ret, uint64_t offset, size_t len) 692 { 693 QEDIsAllocatedCB *cb = opaque; 694 BDRVQEDState *s = cb->bs->opaque; 695 *cb->pnum = len / BDRV_SECTOR_SIZE; 696 switch (ret) { 697 case QED_CLUSTER_FOUND: 698 offset |= qed_offset_into_cluster(s, cb->pos); 699 cb->status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID | offset; 700 break; 701 case QED_CLUSTER_ZERO: 702 cb->status = BDRV_BLOCK_ZERO; 703 break; 704 case QED_CLUSTER_L2: 705 case QED_CLUSTER_L1: 706 cb->status = 0; 707 break; 708 default: 709 assert(ret < 0); 710 cb->status = ret; 711 break; 712 } 713 714 if (cb->co) { 715 qemu_coroutine_enter(cb->co, NULL); 716 } 717 } 718 719 static int64_t coroutine_fn bdrv_qed_co_get_block_status(BlockDriverState *bs, 720 int64_t sector_num, 721 int nb_sectors, int *pnum) 722 { 723 BDRVQEDState *s = bs->opaque; 724 size_t len = (size_t)nb_sectors * BDRV_SECTOR_SIZE; 725 QEDIsAllocatedCB cb = { 726 .bs = bs, 727 .pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE, 728 .status = BDRV_BLOCK_OFFSET_MASK, 729 .pnum = pnum, 730 }; 731 QEDRequest request = { .l2_table = NULL }; 732 733 qed_find_cluster(s, &request, cb.pos, len, qed_is_allocated_cb, &cb); 734 735 /* Now sleep if the callback wasn't invoked immediately */ 736 while (cb.status == BDRV_BLOCK_OFFSET_MASK) { 737 cb.co = qemu_coroutine_self(); 738 qemu_coroutine_yield(); 739 } 740 741 qed_unref_l2_cache_entry(request.l2_table); 742 743 return cb.status; 744 } 745 746 static BDRVQEDState *acb_to_s(QEDAIOCB *acb) 747 { 748 return acb->common.bs->opaque; 749 } 750 751 /** 752 * Read from the backing file or zero-fill if no backing file 753 * 754 * @s: QED state 755 * @pos: Byte position in device 756 * @qiov: Destination I/O vector 757 * @backing_qiov: Possibly shortened copy of qiov, to be allocated here 758 * @cb: Completion function 759 * @opaque: User data for completion function 760 * 761 * This function reads qiov->size bytes starting at pos from the backing file. 762 * If there is no backing file then zeroes are read. 763 */ 764 static void qed_read_backing_file(BDRVQEDState *s, uint64_t pos, 765 QEMUIOVector *qiov, 766 QEMUIOVector **backing_qiov, 767 BlockCompletionFunc *cb, void *opaque) 768 { 769 uint64_t backing_length = 0; 770 size_t size; 771 772 /* If there is a backing file, get its length. Treat the absence of a 773 * backing file like a zero length backing file. 774 */ 775 if (s->bs->backing_hd) { 776 int64_t l = bdrv_getlength(s->bs->backing_hd); 777 if (l < 0) { 778 cb(opaque, l); 779 return; 780 } 781 backing_length = l; 782 } 783 784 /* Zero all sectors if reading beyond the end of the backing file */ 785 if (pos >= backing_length || 786 pos + qiov->size > backing_length) { 787 qemu_iovec_memset(qiov, 0, 0, qiov->size); 788 } 789 790 /* Complete now if there are no backing file sectors to read */ 791 if (pos >= backing_length) { 792 cb(opaque, 0); 793 return; 794 } 795 796 /* If the read straddles the end of the backing file, shorten it */ 797 size = MIN((uint64_t)backing_length - pos, qiov->size); 798 799 assert(*backing_qiov == NULL); 800 *backing_qiov = g_new(QEMUIOVector, 1); 801 qemu_iovec_init(*backing_qiov, qiov->niov); 802 qemu_iovec_concat(*backing_qiov, qiov, 0, size); 803 804 BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO); 805 bdrv_aio_readv(s->bs->backing_hd, pos / BDRV_SECTOR_SIZE, 806 *backing_qiov, size / BDRV_SECTOR_SIZE, cb, opaque); 807 } 808 809 typedef struct { 810 GenericCB gencb; 811 BDRVQEDState *s; 812 QEMUIOVector qiov; 813 QEMUIOVector *backing_qiov; 814 struct iovec iov; 815 uint64_t offset; 816 } CopyFromBackingFileCB; 817 818 static void qed_copy_from_backing_file_cb(void *opaque, int ret) 819 { 820 CopyFromBackingFileCB *copy_cb = opaque; 821 qemu_vfree(copy_cb->iov.iov_base); 822 gencb_complete(©_cb->gencb, ret); 823 } 824 825 static void qed_copy_from_backing_file_write(void *opaque, int ret) 826 { 827 CopyFromBackingFileCB *copy_cb = opaque; 828 BDRVQEDState *s = copy_cb->s; 829 830 if (copy_cb->backing_qiov) { 831 qemu_iovec_destroy(copy_cb->backing_qiov); 832 g_free(copy_cb->backing_qiov); 833 copy_cb->backing_qiov = NULL; 834 } 835 836 if (ret) { 837 qed_copy_from_backing_file_cb(copy_cb, ret); 838 return; 839 } 840 841 BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE); 842 bdrv_aio_writev(s->bs->file, copy_cb->offset / BDRV_SECTOR_SIZE, 843 ©_cb->qiov, copy_cb->qiov.size / BDRV_SECTOR_SIZE, 844 qed_copy_from_backing_file_cb, copy_cb); 845 } 846 847 /** 848 * Copy data from backing file into the image 849 * 850 * @s: QED state 851 * @pos: Byte position in device 852 * @len: Number of bytes 853 * @offset: Byte offset in image file 854 * @cb: Completion function 855 * @opaque: User data for completion function 856 */ 857 static void qed_copy_from_backing_file(BDRVQEDState *s, uint64_t pos, 858 uint64_t len, uint64_t offset, 859 BlockCompletionFunc *cb, 860 void *opaque) 861 { 862 CopyFromBackingFileCB *copy_cb; 863 864 /* Skip copy entirely if there is no work to do */ 865 if (len == 0) { 866 cb(opaque, 0); 867 return; 868 } 869 870 copy_cb = gencb_alloc(sizeof(*copy_cb), cb, opaque); 871 copy_cb->s = s; 872 copy_cb->offset = offset; 873 copy_cb->backing_qiov = NULL; 874 copy_cb->iov.iov_base = qemu_blockalign(s->bs, len); 875 copy_cb->iov.iov_len = len; 876 qemu_iovec_init_external(©_cb->qiov, ©_cb->iov, 1); 877 878 qed_read_backing_file(s, pos, ©_cb->qiov, ©_cb->backing_qiov, 879 qed_copy_from_backing_file_write, copy_cb); 880 } 881 882 /** 883 * Link one or more contiguous clusters into a table 884 * 885 * @s: QED state 886 * @table: L2 table 887 * @index: First cluster index 888 * @n: Number of contiguous clusters 889 * @cluster: First cluster offset 890 * 891 * The cluster offset may be an allocated byte offset in the image file, the 892 * zero cluster marker, or the unallocated cluster marker. 893 */ 894 static void qed_update_l2_table(BDRVQEDState *s, QEDTable *table, int index, 895 unsigned int n, uint64_t cluster) 896 { 897 int i; 898 for (i = index; i < index + n; i++) { 899 table->offsets[i] = cluster; 900 if (!qed_offset_is_unalloc_cluster(cluster) && 901 !qed_offset_is_zero_cluster(cluster)) { 902 cluster += s->header.cluster_size; 903 } 904 } 905 } 906 907 static void qed_aio_complete_bh(void *opaque) 908 { 909 QEDAIOCB *acb = opaque; 910 BlockCompletionFunc *cb = acb->common.cb; 911 void *user_opaque = acb->common.opaque; 912 int ret = acb->bh_ret; 913 914 qemu_bh_delete(acb->bh); 915 qemu_aio_unref(acb); 916 917 /* Invoke callback */ 918 cb(user_opaque, ret); 919 } 920 921 static void qed_aio_complete(QEDAIOCB *acb, int ret) 922 { 923 BDRVQEDState *s = acb_to_s(acb); 924 925 trace_qed_aio_complete(s, acb, ret); 926 927 /* Free resources */ 928 qemu_iovec_destroy(&acb->cur_qiov); 929 qed_unref_l2_cache_entry(acb->request.l2_table); 930 931 /* Free the buffer we may have allocated for zero writes */ 932 if (acb->flags & QED_AIOCB_ZERO) { 933 qemu_vfree(acb->qiov->iov[0].iov_base); 934 acb->qiov->iov[0].iov_base = NULL; 935 } 936 937 /* Arrange for a bh to invoke the completion function */ 938 acb->bh_ret = ret; 939 acb->bh = aio_bh_new(bdrv_get_aio_context(acb->common.bs), 940 qed_aio_complete_bh, acb); 941 qemu_bh_schedule(acb->bh); 942 943 /* Start next allocating write request waiting behind this one. Note that 944 * requests enqueue themselves when they first hit an unallocated cluster 945 * but they wait until the entire request is finished before waking up the 946 * next request in the queue. This ensures that we don't cycle through 947 * requests multiple times but rather finish one at a time completely. 948 */ 949 if (acb == QSIMPLEQ_FIRST(&s->allocating_write_reqs)) { 950 QSIMPLEQ_REMOVE_HEAD(&s->allocating_write_reqs, next); 951 acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs); 952 if (acb) { 953 qed_aio_next_io(acb, 0); 954 } else if (s->header.features & QED_F_NEED_CHECK) { 955 qed_start_need_check_timer(s); 956 } 957 } 958 } 959 960 /** 961 * Commit the current L2 table to the cache 962 */ 963 static void qed_commit_l2_update(void *opaque, int ret) 964 { 965 QEDAIOCB *acb = opaque; 966 BDRVQEDState *s = acb_to_s(acb); 967 CachedL2Table *l2_table = acb->request.l2_table; 968 uint64_t l2_offset = l2_table->offset; 969 970 qed_commit_l2_cache_entry(&s->l2_cache, l2_table); 971 972 /* This is guaranteed to succeed because we just committed the entry to the 973 * cache. 974 */ 975 acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset); 976 assert(acb->request.l2_table != NULL); 977 978 qed_aio_next_io(opaque, ret); 979 } 980 981 /** 982 * Update L1 table with new L2 table offset and write it out 983 */ 984 static void qed_aio_write_l1_update(void *opaque, int ret) 985 { 986 QEDAIOCB *acb = opaque; 987 BDRVQEDState *s = acb_to_s(acb); 988 int index; 989 990 if (ret) { 991 qed_aio_complete(acb, ret); 992 return; 993 } 994 995 index = qed_l1_index(s, acb->cur_pos); 996 s->l1_table->offsets[index] = acb->request.l2_table->offset; 997 998 qed_write_l1_table(s, index, 1, qed_commit_l2_update, acb); 999 } 1000 1001 /** 1002 * Update L2 table with new cluster offsets and write them out 1003 */ 1004 static void qed_aio_write_l2_update(QEDAIOCB *acb, int ret, uint64_t offset) 1005 { 1006 BDRVQEDState *s = acb_to_s(acb); 1007 bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1; 1008 int index; 1009 1010 if (ret) { 1011 goto err; 1012 } 1013 1014 if (need_alloc) { 1015 qed_unref_l2_cache_entry(acb->request.l2_table); 1016 acb->request.l2_table = qed_new_l2_table(s); 1017 } 1018 1019 index = qed_l2_index(s, acb->cur_pos); 1020 qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters, 1021 offset); 1022 1023 if (need_alloc) { 1024 /* Write out the whole new L2 table */ 1025 qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true, 1026 qed_aio_write_l1_update, acb); 1027 } else { 1028 /* Write out only the updated part of the L2 table */ 1029 qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters, false, 1030 qed_aio_next_io, acb); 1031 } 1032 return; 1033 1034 err: 1035 qed_aio_complete(acb, ret); 1036 } 1037 1038 static void qed_aio_write_l2_update_cb(void *opaque, int ret) 1039 { 1040 QEDAIOCB *acb = opaque; 1041 qed_aio_write_l2_update(acb, ret, acb->cur_cluster); 1042 } 1043 1044 /** 1045 * Flush new data clusters before updating the L2 table 1046 * 1047 * This flush is necessary when a backing file is in use. A crash during an 1048 * allocating write could result in empty clusters in the image. If the write 1049 * only touched a subregion of the cluster, then backing image sectors have 1050 * been lost in the untouched region. The solution is to flush after writing a 1051 * new data cluster and before updating the L2 table. 1052 */ 1053 static void qed_aio_write_flush_before_l2_update(void *opaque, int ret) 1054 { 1055 QEDAIOCB *acb = opaque; 1056 BDRVQEDState *s = acb_to_s(acb); 1057 1058 if (!bdrv_aio_flush(s->bs->file, qed_aio_write_l2_update_cb, opaque)) { 1059 qed_aio_complete(acb, -EIO); 1060 } 1061 } 1062 1063 /** 1064 * Write data to the image file 1065 */ 1066 static void qed_aio_write_main(void *opaque, int ret) 1067 { 1068 QEDAIOCB *acb = opaque; 1069 BDRVQEDState *s = acb_to_s(acb); 1070 uint64_t offset = acb->cur_cluster + 1071 qed_offset_into_cluster(s, acb->cur_pos); 1072 BlockCompletionFunc *next_fn; 1073 1074 trace_qed_aio_write_main(s, acb, ret, offset, acb->cur_qiov.size); 1075 1076 if (ret) { 1077 qed_aio_complete(acb, ret); 1078 return; 1079 } 1080 1081 if (acb->find_cluster_ret == QED_CLUSTER_FOUND) { 1082 next_fn = qed_aio_next_io; 1083 } else { 1084 if (s->bs->backing_hd) { 1085 next_fn = qed_aio_write_flush_before_l2_update; 1086 } else { 1087 next_fn = qed_aio_write_l2_update_cb; 1088 } 1089 } 1090 1091 BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO); 1092 bdrv_aio_writev(s->bs->file, offset / BDRV_SECTOR_SIZE, 1093 &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE, 1094 next_fn, acb); 1095 } 1096 1097 /** 1098 * Populate back untouched region of new data cluster 1099 */ 1100 static void qed_aio_write_postfill(void *opaque, int ret) 1101 { 1102 QEDAIOCB *acb = opaque; 1103 BDRVQEDState *s = acb_to_s(acb); 1104 uint64_t start = acb->cur_pos + acb->cur_qiov.size; 1105 uint64_t len = 1106 qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start; 1107 uint64_t offset = acb->cur_cluster + 1108 qed_offset_into_cluster(s, acb->cur_pos) + 1109 acb->cur_qiov.size; 1110 1111 if (ret) { 1112 qed_aio_complete(acb, ret); 1113 return; 1114 } 1115 1116 trace_qed_aio_write_postfill(s, acb, start, len, offset); 1117 qed_copy_from_backing_file(s, start, len, offset, 1118 qed_aio_write_main, acb); 1119 } 1120 1121 /** 1122 * Populate front untouched region of new data cluster 1123 */ 1124 static void qed_aio_write_prefill(void *opaque, int ret) 1125 { 1126 QEDAIOCB *acb = opaque; 1127 BDRVQEDState *s = acb_to_s(acb); 1128 uint64_t start = qed_start_of_cluster(s, acb->cur_pos); 1129 uint64_t len = qed_offset_into_cluster(s, acb->cur_pos); 1130 1131 trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster); 1132 qed_copy_from_backing_file(s, start, len, acb->cur_cluster, 1133 qed_aio_write_postfill, acb); 1134 } 1135 1136 /** 1137 * Check if the QED_F_NEED_CHECK bit should be set during allocating write 1138 */ 1139 static bool qed_should_set_need_check(BDRVQEDState *s) 1140 { 1141 /* The flush before L2 update path ensures consistency */ 1142 if (s->bs->backing_hd) { 1143 return false; 1144 } 1145 1146 return !(s->header.features & QED_F_NEED_CHECK); 1147 } 1148 1149 static void qed_aio_write_zero_cluster(void *opaque, int ret) 1150 { 1151 QEDAIOCB *acb = opaque; 1152 1153 if (ret) { 1154 qed_aio_complete(acb, ret); 1155 return; 1156 } 1157 1158 qed_aio_write_l2_update(acb, 0, 1); 1159 } 1160 1161 /** 1162 * Write new data cluster 1163 * 1164 * @acb: Write request 1165 * @len: Length in bytes 1166 * 1167 * This path is taken when writing to previously unallocated clusters. 1168 */ 1169 static void qed_aio_write_alloc(QEDAIOCB *acb, size_t len) 1170 { 1171 BDRVQEDState *s = acb_to_s(acb); 1172 BlockCompletionFunc *cb; 1173 1174 /* Cancel timer when the first allocating request comes in */ 1175 if (QSIMPLEQ_EMPTY(&s->allocating_write_reqs)) { 1176 qed_cancel_need_check_timer(s); 1177 } 1178 1179 /* Freeze this request if another allocating write is in progress */ 1180 if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs)) { 1181 QSIMPLEQ_INSERT_TAIL(&s->allocating_write_reqs, acb, next); 1182 } 1183 if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs) || 1184 s->allocating_write_reqs_plugged) { 1185 return; /* wait for existing request to finish */ 1186 } 1187 1188 acb->cur_nclusters = qed_bytes_to_clusters(s, 1189 qed_offset_into_cluster(s, acb->cur_pos) + len); 1190 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len); 1191 1192 if (acb->flags & QED_AIOCB_ZERO) { 1193 /* Skip ahead if the clusters are already zero */ 1194 if (acb->find_cluster_ret == QED_CLUSTER_ZERO) { 1195 qed_aio_next_io(acb, 0); 1196 return; 1197 } 1198 1199 cb = qed_aio_write_zero_cluster; 1200 } else { 1201 cb = qed_aio_write_prefill; 1202 acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters); 1203 } 1204 1205 if (qed_should_set_need_check(s)) { 1206 s->header.features |= QED_F_NEED_CHECK; 1207 qed_write_header(s, cb, acb); 1208 } else { 1209 cb(acb, 0); 1210 } 1211 } 1212 1213 /** 1214 * Write data cluster in place 1215 * 1216 * @acb: Write request 1217 * @offset: Cluster offset in bytes 1218 * @len: Length in bytes 1219 * 1220 * This path is taken when writing to already allocated clusters. 1221 */ 1222 static void qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset, size_t len) 1223 { 1224 /* Allocate buffer for zero writes */ 1225 if (acb->flags & QED_AIOCB_ZERO) { 1226 struct iovec *iov = acb->qiov->iov; 1227 1228 if (!iov->iov_base) { 1229 iov->iov_base = qemu_try_blockalign(acb->common.bs, iov->iov_len); 1230 if (iov->iov_base == NULL) { 1231 qed_aio_complete(acb, -ENOMEM); 1232 return; 1233 } 1234 memset(iov->iov_base, 0, iov->iov_len); 1235 } 1236 } 1237 1238 /* Calculate the I/O vector */ 1239 acb->cur_cluster = offset; 1240 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len); 1241 1242 /* Do the actual write */ 1243 qed_aio_write_main(acb, 0); 1244 } 1245 1246 /** 1247 * Write data cluster 1248 * 1249 * @opaque: Write request 1250 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1, 1251 * or -errno 1252 * @offset: Cluster offset in bytes 1253 * @len: Length in bytes 1254 * 1255 * Callback from qed_find_cluster(). 1256 */ 1257 static void qed_aio_write_data(void *opaque, int ret, 1258 uint64_t offset, size_t len) 1259 { 1260 QEDAIOCB *acb = opaque; 1261 1262 trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len); 1263 1264 acb->find_cluster_ret = ret; 1265 1266 switch (ret) { 1267 case QED_CLUSTER_FOUND: 1268 qed_aio_write_inplace(acb, offset, len); 1269 break; 1270 1271 case QED_CLUSTER_L2: 1272 case QED_CLUSTER_L1: 1273 case QED_CLUSTER_ZERO: 1274 qed_aio_write_alloc(acb, len); 1275 break; 1276 1277 default: 1278 qed_aio_complete(acb, ret); 1279 break; 1280 } 1281 } 1282 1283 /** 1284 * Read data cluster 1285 * 1286 * @opaque: Read request 1287 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1, 1288 * or -errno 1289 * @offset: Cluster offset in bytes 1290 * @len: Length in bytes 1291 * 1292 * Callback from qed_find_cluster(). 1293 */ 1294 static void qed_aio_read_data(void *opaque, int ret, 1295 uint64_t offset, size_t len) 1296 { 1297 QEDAIOCB *acb = opaque; 1298 BDRVQEDState *s = acb_to_s(acb); 1299 BlockDriverState *bs = acb->common.bs; 1300 1301 /* Adjust offset into cluster */ 1302 offset += qed_offset_into_cluster(s, acb->cur_pos); 1303 1304 trace_qed_aio_read_data(s, acb, ret, offset, len); 1305 1306 if (ret < 0) { 1307 goto err; 1308 } 1309 1310 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len); 1311 1312 /* Handle zero cluster and backing file reads */ 1313 if (ret == QED_CLUSTER_ZERO) { 1314 qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size); 1315 qed_aio_next_io(acb, 0); 1316 return; 1317 } else if (ret != QED_CLUSTER_FOUND) { 1318 qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov, 1319 &acb->backing_qiov, qed_aio_next_io, acb); 1320 return; 1321 } 1322 1323 BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO); 1324 bdrv_aio_readv(bs->file, offset / BDRV_SECTOR_SIZE, 1325 &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE, 1326 qed_aio_next_io, acb); 1327 return; 1328 1329 err: 1330 qed_aio_complete(acb, ret); 1331 } 1332 1333 /** 1334 * Begin next I/O or complete the request 1335 */ 1336 static void qed_aio_next_io(void *opaque, int ret) 1337 { 1338 QEDAIOCB *acb = opaque; 1339 BDRVQEDState *s = acb_to_s(acb); 1340 QEDFindClusterFunc *io_fn = (acb->flags & QED_AIOCB_WRITE) ? 1341 qed_aio_write_data : qed_aio_read_data; 1342 1343 trace_qed_aio_next_io(s, acb, ret, acb->cur_pos + acb->cur_qiov.size); 1344 1345 if (acb->backing_qiov) { 1346 qemu_iovec_destroy(acb->backing_qiov); 1347 g_free(acb->backing_qiov); 1348 acb->backing_qiov = NULL; 1349 } 1350 1351 /* Handle I/O error */ 1352 if (ret) { 1353 qed_aio_complete(acb, ret); 1354 return; 1355 } 1356 1357 acb->qiov_offset += acb->cur_qiov.size; 1358 acb->cur_pos += acb->cur_qiov.size; 1359 qemu_iovec_reset(&acb->cur_qiov); 1360 1361 /* Complete request */ 1362 if (acb->cur_pos >= acb->end_pos) { 1363 qed_aio_complete(acb, 0); 1364 return; 1365 } 1366 1367 /* Find next cluster and start I/O */ 1368 qed_find_cluster(s, &acb->request, 1369 acb->cur_pos, acb->end_pos - acb->cur_pos, 1370 io_fn, acb); 1371 } 1372 1373 static BlockAIOCB *qed_aio_setup(BlockDriverState *bs, 1374 int64_t sector_num, 1375 QEMUIOVector *qiov, int nb_sectors, 1376 BlockCompletionFunc *cb, 1377 void *opaque, int flags) 1378 { 1379 QEDAIOCB *acb = qemu_aio_get(&qed_aiocb_info, bs, cb, opaque); 1380 1381 trace_qed_aio_setup(bs->opaque, acb, sector_num, nb_sectors, 1382 opaque, flags); 1383 1384 acb->flags = flags; 1385 acb->qiov = qiov; 1386 acb->qiov_offset = 0; 1387 acb->cur_pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE; 1388 acb->end_pos = acb->cur_pos + nb_sectors * BDRV_SECTOR_SIZE; 1389 acb->backing_qiov = NULL; 1390 acb->request.l2_table = NULL; 1391 qemu_iovec_init(&acb->cur_qiov, qiov->niov); 1392 1393 /* Start request */ 1394 qed_aio_next_io(acb, 0); 1395 return &acb->common; 1396 } 1397 1398 static BlockAIOCB *bdrv_qed_aio_readv(BlockDriverState *bs, 1399 int64_t sector_num, 1400 QEMUIOVector *qiov, int nb_sectors, 1401 BlockCompletionFunc *cb, 1402 void *opaque) 1403 { 1404 return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb, opaque, 0); 1405 } 1406 1407 static BlockAIOCB *bdrv_qed_aio_writev(BlockDriverState *bs, 1408 int64_t sector_num, 1409 QEMUIOVector *qiov, int nb_sectors, 1410 BlockCompletionFunc *cb, 1411 void *opaque) 1412 { 1413 return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb, 1414 opaque, QED_AIOCB_WRITE); 1415 } 1416 1417 typedef struct { 1418 Coroutine *co; 1419 int ret; 1420 bool done; 1421 } QEDWriteZeroesCB; 1422 1423 static void coroutine_fn qed_co_write_zeroes_cb(void *opaque, int ret) 1424 { 1425 QEDWriteZeroesCB *cb = opaque; 1426 1427 cb->done = true; 1428 cb->ret = ret; 1429 if (cb->co) { 1430 qemu_coroutine_enter(cb->co, NULL); 1431 } 1432 } 1433 1434 static int coroutine_fn bdrv_qed_co_write_zeroes(BlockDriverState *bs, 1435 int64_t sector_num, 1436 int nb_sectors, 1437 BdrvRequestFlags flags) 1438 { 1439 BlockAIOCB *blockacb; 1440 BDRVQEDState *s = bs->opaque; 1441 QEDWriteZeroesCB cb = { .done = false }; 1442 QEMUIOVector qiov; 1443 struct iovec iov; 1444 1445 /* Refuse if there are untouched backing file sectors */ 1446 if (bs->backing_hd) { 1447 if (qed_offset_into_cluster(s, sector_num * BDRV_SECTOR_SIZE) != 0) { 1448 return -ENOTSUP; 1449 } 1450 if (qed_offset_into_cluster(s, nb_sectors * BDRV_SECTOR_SIZE) != 0) { 1451 return -ENOTSUP; 1452 } 1453 } 1454 1455 /* Zero writes start without an I/O buffer. If a buffer becomes necessary 1456 * then it will be allocated during request processing. 1457 */ 1458 iov.iov_base = NULL, 1459 iov.iov_len = nb_sectors * BDRV_SECTOR_SIZE, 1460 1461 qemu_iovec_init_external(&qiov, &iov, 1); 1462 blockacb = qed_aio_setup(bs, sector_num, &qiov, nb_sectors, 1463 qed_co_write_zeroes_cb, &cb, 1464 QED_AIOCB_WRITE | QED_AIOCB_ZERO); 1465 if (!blockacb) { 1466 return -EIO; 1467 } 1468 if (!cb.done) { 1469 cb.co = qemu_coroutine_self(); 1470 qemu_coroutine_yield(); 1471 } 1472 assert(cb.done); 1473 return cb.ret; 1474 } 1475 1476 static int bdrv_qed_truncate(BlockDriverState *bs, int64_t offset) 1477 { 1478 BDRVQEDState *s = bs->opaque; 1479 uint64_t old_image_size; 1480 int ret; 1481 1482 if (!qed_is_image_size_valid(offset, s->header.cluster_size, 1483 s->header.table_size)) { 1484 return -EINVAL; 1485 } 1486 1487 /* Shrinking is currently not supported */ 1488 if ((uint64_t)offset < s->header.image_size) { 1489 return -ENOTSUP; 1490 } 1491 1492 old_image_size = s->header.image_size; 1493 s->header.image_size = offset; 1494 ret = qed_write_header_sync(s); 1495 if (ret < 0) { 1496 s->header.image_size = old_image_size; 1497 } 1498 return ret; 1499 } 1500 1501 static int64_t bdrv_qed_getlength(BlockDriverState *bs) 1502 { 1503 BDRVQEDState *s = bs->opaque; 1504 return s->header.image_size; 1505 } 1506 1507 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi) 1508 { 1509 BDRVQEDState *s = bs->opaque; 1510 1511 memset(bdi, 0, sizeof(*bdi)); 1512 bdi->cluster_size = s->header.cluster_size; 1513 bdi->is_dirty = s->header.features & QED_F_NEED_CHECK; 1514 bdi->unallocated_blocks_are_zero = true; 1515 bdi->can_write_zeroes_with_unmap = true; 1516 return 0; 1517 } 1518 1519 static int bdrv_qed_change_backing_file(BlockDriverState *bs, 1520 const char *backing_file, 1521 const char *backing_fmt) 1522 { 1523 BDRVQEDState *s = bs->opaque; 1524 QEDHeader new_header, le_header; 1525 void *buffer; 1526 size_t buffer_len, backing_file_len; 1527 int ret; 1528 1529 /* Refuse to set backing filename if unknown compat feature bits are 1530 * active. If the image uses an unknown compat feature then we may not 1531 * know the layout of data following the header structure and cannot safely 1532 * add a new string. 1533 */ 1534 if (backing_file && (s->header.compat_features & 1535 ~QED_COMPAT_FEATURE_MASK)) { 1536 return -ENOTSUP; 1537 } 1538 1539 memcpy(&new_header, &s->header, sizeof(new_header)); 1540 1541 new_header.features &= ~(QED_F_BACKING_FILE | 1542 QED_F_BACKING_FORMAT_NO_PROBE); 1543 1544 /* Adjust feature flags */ 1545 if (backing_file) { 1546 new_header.features |= QED_F_BACKING_FILE; 1547 1548 if (qed_fmt_is_raw(backing_fmt)) { 1549 new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE; 1550 } 1551 } 1552 1553 /* Calculate new header size */ 1554 backing_file_len = 0; 1555 1556 if (backing_file) { 1557 backing_file_len = strlen(backing_file); 1558 } 1559 1560 buffer_len = sizeof(new_header); 1561 new_header.backing_filename_offset = buffer_len; 1562 new_header.backing_filename_size = backing_file_len; 1563 buffer_len += backing_file_len; 1564 1565 /* Make sure we can rewrite header without failing */ 1566 if (buffer_len > new_header.header_size * new_header.cluster_size) { 1567 return -ENOSPC; 1568 } 1569 1570 /* Prepare new header */ 1571 buffer = g_malloc(buffer_len); 1572 1573 qed_header_cpu_to_le(&new_header, &le_header); 1574 memcpy(buffer, &le_header, sizeof(le_header)); 1575 buffer_len = sizeof(le_header); 1576 1577 if (backing_file) { 1578 memcpy(buffer + buffer_len, backing_file, backing_file_len); 1579 buffer_len += backing_file_len; 1580 } 1581 1582 /* Write new header */ 1583 ret = bdrv_pwrite_sync(bs->file, 0, buffer, buffer_len); 1584 g_free(buffer); 1585 if (ret == 0) { 1586 memcpy(&s->header, &new_header, sizeof(new_header)); 1587 } 1588 return ret; 1589 } 1590 1591 static void bdrv_qed_invalidate_cache(BlockDriverState *bs, Error **errp) 1592 { 1593 BDRVQEDState *s = bs->opaque; 1594 Error *local_err = NULL; 1595 int ret; 1596 1597 bdrv_qed_close(bs); 1598 1599 bdrv_invalidate_cache(bs->file, &local_err); 1600 if (local_err) { 1601 error_propagate(errp, local_err); 1602 return; 1603 } 1604 1605 memset(s, 0, sizeof(BDRVQEDState)); 1606 ret = bdrv_qed_open(bs, NULL, bs->open_flags, &local_err); 1607 if (local_err) { 1608 error_setg(errp, "Could not reopen qed layer: %s", 1609 error_get_pretty(local_err)); 1610 error_free(local_err); 1611 return; 1612 } else if (ret < 0) { 1613 error_setg_errno(errp, -ret, "Could not reopen qed layer"); 1614 return; 1615 } 1616 } 1617 1618 static int bdrv_qed_check(BlockDriverState *bs, BdrvCheckResult *result, 1619 BdrvCheckMode fix) 1620 { 1621 BDRVQEDState *s = bs->opaque; 1622 1623 return qed_check(s, result, !!fix); 1624 } 1625 1626 static QemuOptsList qed_create_opts = { 1627 .name = "qed-create-opts", 1628 .head = QTAILQ_HEAD_INITIALIZER(qed_create_opts.head), 1629 .desc = { 1630 { 1631 .name = BLOCK_OPT_SIZE, 1632 .type = QEMU_OPT_SIZE, 1633 .help = "Virtual disk size" 1634 }, 1635 { 1636 .name = BLOCK_OPT_BACKING_FILE, 1637 .type = QEMU_OPT_STRING, 1638 .help = "File name of a base image" 1639 }, 1640 { 1641 .name = BLOCK_OPT_BACKING_FMT, 1642 .type = QEMU_OPT_STRING, 1643 .help = "Image format of the base image" 1644 }, 1645 { 1646 .name = BLOCK_OPT_CLUSTER_SIZE, 1647 .type = QEMU_OPT_SIZE, 1648 .help = "Cluster size (in bytes)", 1649 .def_value_str = stringify(QED_DEFAULT_CLUSTER_SIZE) 1650 }, 1651 { 1652 .name = BLOCK_OPT_TABLE_SIZE, 1653 .type = QEMU_OPT_SIZE, 1654 .help = "L1/L2 table size (in clusters)" 1655 }, 1656 { /* end of list */ } 1657 } 1658 }; 1659 1660 static BlockDriver bdrv_qed = { 1661 .format_name = "qed", 1662 .instance_size = sizeof(BDRVQEDState), 1663 .create_opts = &qed_create_opts, 1664 .supports_backing = true, 1665 1666 .bdrv_probe = bdrv_qed_probe, 1667 .bdrv_rebind = bdrv_qed_rebind, 1668 .bdrv_open = bdrv_qed_open, 1669 .bdrv_close = bdrv_qed_close, 1670 .bdrv_reopen_prepare = bdrv_qed_reopen_prepare, 1671 .bdrv_create = bdrv_qed_create, 1672 .bdrv_has_zero_init = bdrv_has_zero_init_1, 1673 .bdrv_co_get_block_status = bdrv_qed_co_get_block_status, 1674 .bdrv_aio_readv = bdrv_qed_aio_readv, 1675 .bdrv_aio_writev = bdrv_qed_aio_writev, 1676 .bdrv_co_write_zeroes = bdrv_qed_co_write_zeroes, 1677 .bdrv_truncate = bdrv_qed_truncate, 1678 .bdrv_getlength = bdrv_qed_getlength, 1679 .bdrv_get_info = bdrv_qed_get_info, 1680 .bdrv_refresh_limits = bdrv_qed_refresh_limits, 1681 .bdrv_change_backing_file = bdrv_qed_change_backing_file, 1682 .bdrv_invalidate_cache = bdrv_qed_invalidate_cache, 1683 .bdrv_check = bdrv_qed_check, 1684 .bdrv_detach_aio_context = bdrv_qed_detach_aio_context, 1685 .bdrv_attach_aio_context = bdrv_qed_attach_aio_context, 1686 }; 1687 1688 static void bdrv_qed_init(void) 1689 { 1690 bdrv_register(&bdrv_qed); 1691 } 1692 1693 block_init(bdrv_qed_init); 1694