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