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