2 * QEMU Enhanced Disk Format
4 * Copyright IBM, Corp. 2010
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.
15 #include "qemu/osdep.h"
16 #include "qapi/error.h"
17 #include "qemu/timer.h"
18 #include "qemu/bswap.h"
21 #include "qapi/qmp/qerror.h"
22 #include "sysemu/block-backend.h"
24 static int bdrv_qed_probe(const uint8_t *buf, int buf_size,
27 const QEDHeader *header = (const QEDHeader *)buf;
29 if (buf_size < sizeof(*header)) {
32 if (le32_to_cpu(header->magic) != QED_MAGIC) {
39 * Check whether an image format is raw
41 * @fmt: Backing file format, may be NULL
43 static bool qed_fmt_is_raw(const char *fmt)
45 return fmt && strcmp(fmt, "raw") == 0;
48 static void qed_header_le_to_cpu(const QEDHeader *le, QEDHeader *cpu)
50 cpu->magic = le32_to_cpu(le->magic);
51 cpu->cluster_size = le32_to_cpu(le->cluster_size);
52 cpu->table_size = le32_to_cpu(le->table_size);
53 cpu->header_size = le32_to_cpu(le->header_size);
54 cpu->features = le64_to_cpu(le->features);
55 cpu->compat_features = le64_to_cpu(le->compat_features);
56 cpu->autoclear_features = le64_to_cpu(le->autoclear_features);
57 cpu->l1_table_offset = le64_to_cpu(le->l1_table_offset);
58 cpu->image_size = le64_to_cpu(le->image_size);
59 cpu->backing_filename_offset = le32_to_cpu(le->backing_filename_offset);
60 cpu->backing_filename_size = le32_to_cpu(le->backing_filename_size);
63 static void qed_header_cpu_to_le(const QEDHeader *cpu, QEDHeader *le)
65 le->magic = cpu_to_le32(cpu->magic);
66 le->cluster_size = cpu_to_le32(cpu->cluster_size);
67 le->table_size = cpu_to_le32(cpu->table_size);
68 le->header_size = cpu_to_le32(cpu->header_size);
69 le->features = cpu_to_le64(cpu->features);
70 le->compat_features = cpu_to_le64(cpu->compat_features);
71 le->autoclear_features = cpu_to_le64(cpu->autoclear_features);
72 le->l1_table_offset = cpu_to_le64(cpu->l1_table_offset);
73 le->image_size = cpu_to_le64(cpu->image_size);
74 le->backing_filename_offset = cpu_to_le32(cpu->backing_filename_offset);
75 le->backing_filename_size = cpu_to_le32(cpu->backing_filename_size);
78 int qed_write_header_sync(BDRVQEDState *s)
83 qed_header_cpu_to_le(&s->header, &le);
84 ret = bdrv_pwrite(s->bs->file, 0, &le, sizeof(le));
85 if (ret != sizeof(le)) {
92 * Update header in-place (does not rewrite backing filename or other strings)
94 * This function only updates known header fields in-place and does not affect
95 * extra data after the QED header.
97 static int qed_write_header(BDRVQEDState *s)
99 /* We must write full sectors for O_DIRECT but cannot necessarily generate
100 * the data following the header if an unrecognized compat feature is
101 * active. Therefore, first read the sectors containing the header, update
102 * them, and write back.
105 int nsectors = DIV_ROUND_UP(sizeof(QEDHeader), BDRV_SECTOR_SIZE);
106 size_t len = nsectors * BDRV_SECTOR_SIZE;
112 buf = qemu_blockalign(s->bs, len);
113 iov = (struct iovec) {
117 qemu_iovec_init_external(&qiov, &iov, 1);
119 ret = bdrv_preadv(s->bs->file, 0, &qiov);
125 qed_header_cpu_to_le(&s->header, (QEDHeader *) buf);
127 ret = bdrv_pwritev(s->bs->file, 0, &qiov);
138 static uint64_t qed_max_image_size(uint32_t cluster_size, uint32_t table_size)
140 uint64_t table_entries;
143 table_entries = (table_size * cluster_size) / sizeof(uint64_t);
144 l2_size = table_entries * cluster_size;
146 return l2_size * table_entries;
149 static bool qed_is_cluster_size_valid(uint32_t cluster_size)
151 if (cluster_size < QED_MIN_CLUSTER_SIZE ||
152 cluster_size > QED_MAX_CLUSTER_SIZE) {
155 if (cluster_size & (cluster_size - 1)) {
156 return false; /* not power of 2 */
161 static bool qed_is_table_size_valid(uint32_t table_size)
163 if (table_size < QED_MIN_TABLE_SIZE ||
164 table_size > QED_MAX_TABLE_SIZE) {
167 if (table_size & (table_size - 1)) {
168 return false; /* not power of 2 */
173 static bool qed_is_image_size_valid(uint64_t image_size, uint32_t cluster_size,
176 if (image_size % BDRV_SECTOR_SIZE != 0) {
177 return false; /* not multiple of sector size */
179 if (image_size > qed_max_image_size(cluster_size, table_size)) {
180 return false; /* image is too large */
186 * Read a string of known length from the image file
189 * @offset: File offset to start of string, in bytes
190 * @n: String length in bytes
191 * @buf: Destination buffer
192 * @buflen: Destination buffer length in bytes
193 * @ret: 0 on success, -errno on failure
195 * The string is NUL-terminated.
197 static int qed_read_string(BdrvChild *file, uint64_t offset, size_t n,
198 char *buf, size_t buflen)
204 ret = bdrv_pread(file, offset, buf, n);
213 * Allocate new clusters
216 * @n: Number of contiguous clusters to allocate
217 * @ret: Offset of first allocated cluster
219 * This function only produces the offset where the new clusters should be
220 * written. It updates BDRVQEDState but does not make any changes to the image
223 static uint64_t qed_alloc_clusters(BDRVQEDState *s, unsigned int n)
225 uint64_t offset = s->file_size;
226 s->file_size += n * s->header.cluster_size;
230 QEDTable *qed_alloc_table(BDRVQEDState *s)
232 /* Honor O_DIRECT memory alignment requirements */
233 return qemu_blockalign(s->bs,
234 s->header.cluster_size * s->header.table_size);
238 * Allocate a new zeroed L2 table
240 static CachedL2Table *qed_new_l2_table(BDRVQEDState *s)
242 CachedL2Table *l2_table = qed_alloc_l2_cache_entry(&s->l2_cache);
244 l2_table->table = qed_alloc_table(s);
245 l2_table->offset = qed_alloc_clusters(s, s->header.table_size);
247 memset(l2_table->table->offsets, 0,
248 s->header.cluster_size * s->header.table_size);
252 static void qed_plug_allocating_write_reqs(BDRVQEDState *s)
254 assert(!s->allocating_write_reqs_plugged);
256 s->allocating_write_reqs_plugged = true;
259 static void qed_unplug_allocating_write_reqs(BDRVQEDState *s)
261 assert(s->allocating_write_reqs_plugged);
263 s->allocating_write_reqs_plugged = false;
264 qemu_co_enter_next(&s->allocating_write_reqs);
267 static void qed_clear_need_check(void *opaque, int ret)
269 BDRVQEDState *s = opaque;
272 qed_unplug_allocating_write_reqs(s);
276 s->header.features &= ~QED_F_NEED_CHECK;
277 ret = qed_write_header(s);
280 qed_unplug_allocating_write_reqs(s);
282 ret = bdrv_flush(s->bs);
286 static void qed_need_check_timer_cb(void *opaque)
288 BDRVQEDState *s = opaque;
290 /* The timer should only fire when allocating writes have drained */
291 assert(!s->allocating_acb);
293 trace_qed_need_check_timer_cb(s);
296 qed_plug_allocating_write_reqs(s);
298 /* Ensure writes are on disk before clearing flag */
299 bdrv_aio_flush(s->bs->file->bs, qed_clear_need_check, s);
303 void qed_acquire(BDRVQEDState *s)
305 aio_context_acquire(bdrv_get_aio_context(s->bs));
308 void qed_release(BDRVQEDState *s)
310 aio_context_release(bdrv_get_aio_context(s->bs));
313 static void qed_start_need_check_timer(BDRVQEDState *s)
315 trace_qed_start_need_check_timer(s);
317 /* Use QEMU_CLOCK_VIRTUAL so we don't alter the image file while suspended for
320 timer_mod(s->need_check_timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
321 NANOSECONDS_PER_SECOND * QED_NEED_CHECK_TIMEOUT);
324 /* It's okay to call this multiple times or when no timer is started */
325 static void qed_cancel_need_check_timer(BDRVQEDState *s)
327 trace_qed_cancel_need_check_timer(s);
328 timer_del(s->need_check_timer);
331 static void bdrv_qed_detach_aio_context(BlockDriverState *bs)
333 BDRVQEDState *s = bs->opaque;
335 qed_cancel_need_check_timer(s);
336 timer_free(s->need_check_timer);
339 static void bdrv_qed_attach_aio_context(BlockDriverState *bs,
340 AioContext *new_context)
342 BDRVQEDState *s = bs->opaque;
344 s->need_check_timer = aio_timer_new(new_context,
345 QEMU_CLOCK_VIRTUAL, SCALE_NS,
346 qed_need_check_timer_cb, s);
347 if (s->header.features & QED_F_NEED_CHECK) {
348 qed_start_need_check_timer(s);
352 static void bdrv_qed_drain(BlockDriverState *bs)
354 BDRVQEDState *s = bs->opaque;
356 /* Fire the timer immediately in order to start doing I/O as soon as the
359 if (s->need_check_timer && timer_pending(s->need_check_timer)) {
360 qed_cancel_need_check_timer(s);
361 qed_need_check_timer_cb(s);
365 static int bdrv_qed_do_open(BlockDriverState *bs, QDict *options, int flags,
368 BDRVQEDState *s = bs->opaque;
374 qemu_co_queue_init(&s->allocating_write_reqs);
376 ret = bdrv_pread(bs->file, 0, &le_header, sizeof(le_header));
380 qed_header_le_to_cpu(&le_header, &s->header);
382 if (s->header.magic != QED_MAGIC) {
383 error_setg(errp, "Image not in QED format");
386 if (s->header.features & ~QED_FEATURE_MASK) {
387 /* image uses unsupported feature bits */
388 error_setg(errp, "Unsupported QED features: %" PRIx64,
389 s->header.features & ~QED_FEATURE_MASK);
392 if (!qed_is_cluster_size_valid(s->header.cluster_size)) {
396 /* Round down file size to the last cluster */
397 file_size = bdrv_getlength(bs->file->bs);
401 s->file_size = qed_start_of_cluster(s, file_size);
403 if (!qed_is_table_size_valid(s->header.table_size)) {
406 if (!qed_is_image_size_valid(s->header.image_size,
407 s->header.cluster_size,
408 s->header.table_size)) {
411 if (!qed_check_table_offset(s, s->header.l1_table_offset)) {
415 s->table_nelems = (s->header.cluster_size * s->header.table_size) /
417 s->l2_shift = ctz32(s->header.cluster_size);
418 s->l2_mask = s->table_nelems - 1;
419 s->l1_shift = s->l2_shift + ctz32(s->table_nelems);
421 /* Header size calculation must not overflow uint32_t */
422 if (s->header.header_size > UINT32_MAX / s->header.cluster_size) {
426 if ((s->header.features & QED_F_BACKING_FILE)) {
427 if ((uint64_t)s->header.backing_filename_offset +
428 s->header.backing_filename_size >
429 s->header.cluster_size * s->header.header_size) {
433 ret = qed_read_string(bs->file, s->header.backing_filename_offset,
434 s->header.backing_filename_size, bs->backing_file,
435 sizeof(bs->backing_file));
440 if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) {
441 pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw");
445 /* Reset unknown autoclear feature bits. This is a backwards
446 * compatibility mechanism that allows images to be opened by older
447 * programs, which "knock out" unknown feature bits. When an image is
448 * opened by a newer program again it can detect that the autoclear
449 * feature is no longer valid.
451 if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 &&
452 !bdrv_is_read_only(bs->file->bs) && !(flags & BDRV_O_INACTIVE)) {
453 s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK;
455 ret = qed_write_header_sync(s);
460 /* From here on only known autoclear feature bits are valid */
461 bdrv_flush(bs->file->bs);
464 s->l1_table = qed_alloc_table(s);
465 qed_init_l2_cache(&s->l2_cache);
467 ret = qed_read_l1_table_sync(s);
472 /* If image was not closed cleanly, check consistency */
473 if (!(flags & BDRV_O_CHECK) && (s->header.features & QED_F_NEED_CHECK)) {
474 /* Read-only images cannot be fixed. There is no risk of corruption
475 * since write operations are not possible. Therefore, allow
476 * potentially inconsistent images to be opened read-only. This can
477 * aid data recovery from an otherwise inconsistent image.
479 if (!bdrv_is_read_only(bs->file->bs) &&
480 !(flags & BDRV_O_INACTIVE)) {
481 BdrvCheckResult result = {0};
483 ret = qed_check(s, &result, true);
490 bdrv_qed_attach_aio_context(bs, bdrv_get_aio_context(bs));
494 qed_free_l2_cache(&s->l2_cache);
495 qemu_vfree(s->l1_table);
500 static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags,
503 bs->file = bdrv_open_child(NULL, options, "file", bs, &child_file,
509 return bdrv_qed_do_open(bs, options, flags, errp);
512 static void bdrv_qed_refresh_limits(BlockDriverState *bs, Error **errp)
514 BDRVQEDState *s = bs->opaque;
516 bs->bl.pwrite_zeroes_alignment = s->header.cluster_size;
519 /* We have nothing to do for QED reopen, stubs just return
521 static int bdrv_qed_reopen_prepare(BDRVReopenState *state,
522 BlockReopenQueue *queue, Error **errp)
527 static void bdrv_qed_close(BlockDriverState *bs)
529 BDRVQEDState *s = bs->opaque;
531 bdrv_qed_detach_aio_context(bs);
533 /* Ensure writes reach stable storage */
534 bdrv_flush(bs->file->bs);
536 /* Clean shutdown, no check required on next open */
537 if (s->header.features & QED_F_NEED_CHECK) {
538 s->header.features &= ~QED_F_NEED_CHECK;
539 qed_write_header_sync(s);
542 qed_free_l2_cache(&s->l2_cache);
543 qemu_vfree(s->l1_table);
546 static int qed_create(const char *filename, uint32_t cluster_size,
547 uint64_t image_size, uint32_t table_size,
548 const char *backing_file, const char *backing_fmt,
549 QemuOpts *opts, Error **errp)
553 .cluster_size = cluster_size,
554 .table_size = table_size,
557 .compat_features = 0,
558 .l1_table_offset = cluster_size,
559 .image_size = image_size,
562 uint8_t *l1_table = NULL;
563 size_t l1_size = header.cluster_size * header.table_size;
564 Error *local_err = NULL;
568 ret = bdrv_create_file(filename, opts, &local_err);
570 error_propagate(errp, local_err);
574 blk = blk_new_open(filename, NULL, NULL,
575 BDRV_O_RDWR | BDRV_O_RESIZE | BDRV_O_PROTOCOL,
578 error_propagate(errp, local_err);
582 blk_set_allow_write_beyond_eof(blk, true);
584 /* File must start empty and grow, check truncate is supported */
585 ret = blk_truncate(blk, 0, errp);
591 header.features |= QED_F_BACKING_FILE;
592 header.backing_filename_offset = sizeof(le_header);
593 header.backing_filename_size = strlen(backing_file);
595 if (qed_fmt_is_raw(backing_fmt)) {
596 header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
600 qed_header_cpu_to_le(&header, &le_header);
601 ret = blk_pwrite(blk, 0, &le_header, sizeof(le_header), 0);
605 ret = blk_pwrite(blk, sizeof(le_header), backing_file,
606 header.backing_filename_size, 0);
611 l1_table = g_malloc0(l1_size);
612 ret = blk_pwrite(blk, header.l1_table_offset, l1_table, l1_size, 0);
617 ret = 0; /* success */
624 static int bdrv_qed_create(const char *filename, QemuOpts *opts, Error **errp)
626 uint64_t image_size = 0;
627 uint32_t cluster_size = QED_DEFAULT_CLUSTER_SIZE;
628 uint32_t table_size = QED_DEFAULT_TABLE_SIZE;
629 char *backing_file = NULL;
630 char *backing_fmt = NULL;
633 image_size = ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
635 backing_file = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FILE);
636 backing_fmt = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FMT);
637 cluster_size = qemu_opt_get_size_del(opts,
638 BLOCK_OPT_CLUSTER_SIZE,
639 QED_DEFAULT_CLUSTER_SIZE);
640 table_size = qemu_opt_get_size_del(opts, BLOCK_OPT_TABLE_SIZE,
641 QED_DEFAULT_TABLE_SIZE);
643 if (!qed_is_cluster_size_valid(cluster_size)) {
644 error_setg(errp, "QED cluster size must be within range [%u, %u] "
646 QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE);
650 if (!qed_is_table_size_valid(table_size)) {
651 error_setg(errp, "QED table size must be within range [%u, %u] "
653 QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE);
657 if (!qed_is_image_size_valid(image_size, cluster_size, table_size)) {
658 error_setg(errp, "QED image size must be a non-zero multiple of "
659 "cluster size and less than %" PRIu64 " bytes",
660 qed_max_image_size(cluster_size, table_size));
665 ret = qed_create(filename, cluster_size, image_size, table_size,
666 backing_file, backing_fmt, opts, errp);
669 g_free(backing_file);
675 BlockDriverState *bs;
680 BlockDriverState **file;
683 static void qed_is_allocated_cb(void *opaque, int ret, uint64_t offset, size_t len)
685 QEDIsAllocatedCB *cb = opaque;
686 BDRVQEDState *s = cb->bs->opaque;
687 *cb->pnum = len / BDRV_SECTOR_SIZE;
689 case QED_CLUSTER_FOUND:
690 offset |= qed_offset_into_cluster(s, cb->pos);
691 cb->status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID | offset;
692 *cb->file = cb->bs->file->bs;
694 case QED_CLUSTER_ZERO:
695 cb->status = BDRV_BLOCK_ZERO;
712 static int64_t coroutine_fn bdrv_qed_co_get_block_status(BlockDriverState *bs,
714 int nb_sectors, int *pnum,
715 BlockDriverState **file)
717 BDRVQEDState *s = bs->opaque;
718 size_t len = (size_t)nb_sectors * BDRV_SECTOR_SIZE;
719 QEDIsAllocatedCB cb = {
721 .pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE,
722 .status = BDRV_BLOCK_OFFSET_MASK,
726 QEDRequest request = { .l2_table = NULL };
730 ret = qed_find_cluster(s, &request, cb.pos, &len, &offset);
731 qed_is_allocated_cb(&cb, ret, offset, len);
733 /* The callback was invoked immediately */
734 assert(cb.status != BDRV_BLOCK_OFFSET_MASK);
736 qed_unref_l2_cache_entry(request.l2_table);
741 static BDRVQEDState *acb_to_s(QEDAIOCB *acb)
743 return acb->bs->opaque;
747 * Read from the backing file or zero-fill if no backing file
750 * @pos: Byte position in device
751 * @qiov: Destination I/O vector
752 * @backing_qiov: Possibly shortened copy of qiov, to be allocated here
753 * @cb: Completion function
754 * @opaque: User data for completion function
756 * This function reads qiov->size bytes starting at pos from the backing file.
757 * If there is no backing file then zeroes are read.
759 static int qed_read_backing_file(BDRVQEDState *s, uint64_t pos,
761 QEMUIOVector **backing_qiov)
763 uint64_t backing_length = 0;
767 /* If there is a backing file, get its length. Treat the absence of a
768 * backing file like a zero length backing file.
770 if (s->bs->backing) {
771 int64_t l = bdrv_getlength(s->bs->backing->bs);
778 /* Zero all sectors if reading beyond the end of the backing file */
779 if (pos >= backing_length ||
780 pos + qiov->size > backing_length) {
781 qemu_iovec_memset(qiov, 0, 0, qiov->size);
784 /* Complete now if there are no backing file sectors to read */
785 if (pos >= backing_length) {
789 /* If the read straddles the end of the backing file, shorten it */
790 size = MIN((uint64_t)backing_length - pos, qiov->size);
792 assert(*backing_qiov == NULL);
793 *backing_qiov = g_new(QEMUIOVector, 1);
794 qemu_iovec_init(*backing_qiov, qiov->niov);
795 qemu_iovec_concat(*backing_qiov, qiov, 0, size);
797 BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO);
798 ret = bdrv_preadv(s->bs->backing, pos, *backing_qiov);
806 * Copy data from backing file into the image
809 * @pos: Byte position in device
810 * @len: Number of bytes
811 * @offset: Byte offset in image file
813 static int qed_copy_from_backing_file(BDRVQEDState *s, uint64_t pos,
814 uint64_t len, uint64_t offset)
817 QEMUIOVector *backing_qiov = NULL;
821 /* Skip copy entirely if there is no work to do */
826 iov = (struct iovec) {
827 .iov_base = qemu_blockalign(s->bs, len),
830 qemu_iovec_init_external(&qiov, &iov, 1);
832 ret = qed_read_backing_file(s, pos, &qiov, &backing_qiov);
835 qemu_iovec_destroy(backing_qiov);
836 g_free(backing_qiov);
844 BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE);
845 ret = bdrv_pwritev(s->bs->file, offset, &qiov);
851 qemu_vfree(iov.iov_base);
856 * Link one or more contiguous clusters into a table
860 * @index: First cluster index
861 * @n: Number of contiguous clusters
862 * @cluster: First cluster offset
864 * The cluster offset may be an allocated byte offset in the image file, the
865 * zero cluster marker, or the unallocated cluster marker.
867 static void qed_update_l2_table(BDRVQEDState *s, QEDTable *table, int index,
868 unsigned int n, uint64_t cluster)
871 for (i = index; i < index + n; i++) {
872 table->offsets[i] = cluster;
873 if (!qed_offset_is_unalloc_cluster(cluster) &&
874 !qed_offset_is_zero_cluster(cluster)) {
875 cluster += s->header.cluster_size;
880 static void qed_aio_complete(QEDAIOCB *acb)
882 BDRVQEDState *s = acb_to_s(acb);
885 qemu_iovec_destroy(&acb->cur_qiov);
886 qed_unref_l2_cache_entry(acb->request.l2_table);
888 /* Free the buffer we may have allocated for zero writes */
889 if (acb->flags & QED_AIOCB_ZERO) {
890 qemu_vfree(acb->qiov->iov[0].iov_base);
891 acb->qiov->iov[0].iov_base = NULL;
894 /* Start next allocating write request waiting behind this one. Note that
895 * requests enqueue themselves when they first hit an unallocated cluster
896 * but they wait until the entire request is finished before waking up the
897 * next request in the queue. This ensures that we don't cycle through
898 * requests multiple times but rather finish one at a time completely.
900 if (acb == s->allocating_acb) {
901 s->allocating_acb = NULL;
902 if (!qemu_co_queue_empty(&s->allocating_write_reqs)) {
903 qemu_co_enter_next(&s->allocating_write_reqs);
904 } else if (s->header.features & QED_F_NEED_CHECK) {
905 qed_start_need_check_timer(s);
911 * Update L1 table with new L2 table offset and write it out
913 static int qed_aio_write_l1_update(QEDAIOCB *acb)
915 BDRVQEDState *s = acb_to_s(acb);
916 CachedL2Table *l2_table = acb->request.l2_table;
917 uint64_t l2_offset = l2_table->offset;
920 index = qed_l1_index(s, acb->cur_pos);
921 s->l1_table->offsets[index] = l2_table->offset;
923 ret = qed_write_l1_table(s, index, 1);
925 /* Commit the current L2 table to the cache */
926 qed_commit_l2_cache_entry(&s->l2_cache, l2_table);
928 /* This is guaranteed to succeed because we just committed the entry to the
931 acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset);
932 assert(acb->request.l2_table != NULL);
939 * Update L2 table with new cluster offsets and write them out
941 static int qed_aio_write_l2_update(QEDAIOCB *acb, uint64_t offset)
943 BDRVQEDState *s = acb_to_s(acb);
944 bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1;
948 qed_unref_l2_cache_entry(acb->request.l2_table);
949 acb->request.l2_table = qed_new_l2_table(s);
952 index = qed_l2_index(s, acb->cur_pos);
953 qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters,
957 /* Write out the whole new L2 table */
958 ret = qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true);
962 return qed_aio_write_l1_update(acb);
964 /* Write out only the updated part of the L2 table */
965 ret = qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters,
975 * Write data to the image file
977 static int qed_aio_write_main(QEDAIOCB *acb)
979 BDRVQEDState *s = acb_to_s(acb);
980 uint64_t offset = acb->cur_cluster +
981 qed_offset_into_cluster(s, acb->cur_pos);
984 trace_qed_aio_write_main(s, acb, 0, offset, acb->cur_qiov.size);
986 BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO);
987 ret = bdrv_pwritev(s->bs->file, offset, &acb->cur_qiov);
992 if (acb->find_cluster_ret != QED_CLUSTER_FOUND) {
993 if (s->bs->backing) {
995 * Flush new data clusters before updating the L2 table
997 * This flush is necessary when a backing file is in use. A crash
998 * during an allocating write could result in empty clusters in the
999 * image. If the write only touched a subregion of the cluster,
1000 * then backing image sectors have been lost in the untouched
1001 * region. The solution is to flush after writing a new data
1002 * cluster and before updating the L2 table.
1004 ret = bdrv_flush(s->bs->file->bs);
1009 ret = qed_aio_write_l2_update(acb, acb->cur_cluster);
1018 * Populate untouched regions of new data cluster
1020 static int qed_aio_write_cow(QEDAIOCB *acb)
1022 BDRVQEDState *s = acb_to_s(acb);
1023 uint64_t start, len, offset;
1026 /* Populate front untouched region of new data cluster */
1027 start = qed_start_of_cluster(s, acb->cur_pos);
1028 len = qed_offset_into_cluster(s, acb->cur_pos);
1030 trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster);
1031 ret = qed_copy_from_backing_file(s, start, len, acb->cur_cluster);
1036 /* Populate back untouched region of new data cluster */
1037 start = acb->cur_pos + acb->cur_qiov.size;
1038 len = qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start;
1039 offset = acb->cur_cluster +
1040 qed_offset_into_cluster(s, acb->cur_pos) +
1043 trace_qed_aio_write_postfill(s, acb, start, len, offset);
1044 ret = qed_copy_from_backing_file(s, start, len, offset);
1049 return qed_aio_write_main(acb);
1053 * Check if the QED_F_NEED_CHECK bit should be set during allocating write
1055 static bool qed_should_set_need_check(BDRVQEDState *s)
1057 /* The flush before L2 update path ensures consistency */
1058 if (s->bs->backing) {
1062 return !(s->header.features & QED_F_NEED_CHECK);
1066 * Write new data cluster
1068 * @acb: Write request
1069 * @len: Length in bytes
1071 * This path is taken when writing to previously unallocated clusters.
1073 static int qed_aio_write_alloc(QEDAIOCB *acb, size_t len)
1075 BDRVQEDState *s = acb_to_s(acb);
1078 /* Cancel timer when the first allocating request comes in */
1079 if (s->allocating_acb == NULL) {
1080 qed_cancel_need_check_timer(s);
1083 /* Freeze this request if another allocating write is in progress */
1084 if (s->allocating_acb != acb || s->allocating_write_reqs_plugged) {
1085 if (s->allocating_acb != NULL) {
1086 qemu_co_queue_wait(&s->allocating_write_reqs, NULL);
1087 assert(s->allocating_acb == NULL);
1089 s->allocating_acb = acb;
1090 return -EAGAIN; /* start over with looking up table entries */
1093 acb->cur_nclusters = qed_bytes_to_clusters(s,
1094 qed_offset_into_cluster(s, acb->cur_pos) + len);
1095 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1097 if (acb->flags & QED_AIOCB_ZERO) {
1098 /* Skip ahead if the clusters are already zero */
1099 if (acb->find_cluster_ret == QED_CLUSTER_ZERO) {
1103 acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters);
1106 if (qed_should_set_need_check(s)) {
1107 s->header.features |= QED_F_NEED_CHECK;
1108 ret = qed_write_header(s);
1114 if (acb->flags & QED_AIOCB_ZERO) {
1115 ret = qed_aio_write_l2_update(acb, 1);
1117 ret = qed_aio_write_cow(acb);
1126 * Write data cluster in place
1128 * @acb: Write request
1129 * @offset: Cluster offset in bytes
1130 * @len: Length in bytes
1132 * This path is taken when writing to already allocated clusters.
1134 static int qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset, size_t len)
1136 /* Allocate buffer for zero writes */
1137 if (acb->flags & QED_AIOCB_ZERO) {
1138 struct iovec *iov = acb->qiov->iov;
1140 if (!iov->iov_base) {
1141 iov->iov_base = qemu_try_blockalign(acb->bs, iov->iov_len);
1142 if (iov->iov_base == NULL) {
1145 memset(iov->iov_base, 0, iov->iov_len);
1149 /* Calculate the I/O vector */
1150 acb->cur_cluster = offset;
1151 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1153 /* Do the actual write */
1154 return qed_aio_write_main(acb);
1158 * Write data cluster
1160 * @opaque: Write request
1161 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1162 * @offset: Cluster offset in bytes
1163 * @len: Length in bytes
1165 static int qed_aio_write_data(void *opaque, int ret,
1166 uint64_t offset, size_t len)
1168 QEDAIOCB *acb = opaque;
1170 trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len);
1172 acb->find_cluster_ret = ret;
1175 case QED_CLUSTER_FOUND:
1176 return qed_aio_write_inplace(acb, offset, len);
1178 case QED_CLUSTER_L2:
1179 case QED_CLUSTER_L1:
1180 case QED_CLUSTER_ZERO:
1181 return qed_aio_write_alloc(acb, len);
1184 g_assert_not_reached();
1191 * @opaque: Read request
1192 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1193 * @offset: Cluster offset in bytes
1194 * @len: Length in bytes
1196 static int qed_aio_read_data(void *opaque, int ret, uint64_t offset, size_t len)
1198 QEDAIOCB *acb = opaque;
1199 BDRVQEDState *s = acb_to_s(acb);
1200 BlockDriverState *bs = acb->bs;
1202 /* Adjust offset into cluster */
1203 offset += qed_offset_into_cluster(s, acb->cur_pos);
1205 trace_qed_aio_read_data(s, acb, ret, offset, len);
1207 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1209 /* Handle zero cluster and backing file reads */
1210 if (ret == QED_CLUSTER_ZERO) {
1211 qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size);
1213 } else if (ret != QED_CLUSTER_FOUND) {
1214 return qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov,
1215 &acb->backing_qiov);
1218 BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
1219 ret = bdrv_preadv(bs->file, offset, &acb->cur_qiov);
1227 * Begin next I/O or complete the request
1229 static int qed_aio_next_io(QEDAIOCB *acb)
1231 BDRVQEDState *s = acb_to_s(acb);
1237 trace_qed_aio_next_io(s, acb, 0, acb->cur_pos + acb->cur_qiov.size);
1239 if (acb->backing_qiov) {
1240 qemu_iovec_destroy(acb->backing_qiov);
1241 g_free(acb->backing_qiov);
1242 acb->backing_qiov = NULL;
1245 acb->qiov_offset += acb->cur_qiov.size;
1246 acb->cur_pos += acb->cur_qiov.size;
1247 qemu_iovec_reset(&acb->cur_qiov);
1249 /* Complete request */
1250 if (acb->cur_pos >= acb->end_pos) {
1255 /* Find next cluster and start I/O */
1256 len = acb->end_pos - acb->cur_pos;
1257 ret = qed_find_cluster(s, &acb->request, acb->cur_pos, &len, &offset);
1262 if (acb->flags & QED_AIOCB_WRITE) {
1263 ret = qed_aio_write_data(acb, ret, offset, len);
1265 ret = qed_aio_read_data(acb, ret, offset, len);
1268 if (ret < 0 && ret != -EAGAIN) {
1273 trace_qed_aio_complete(s, acb, ret);
1274 qed_aio_complete(acb);
1278 static int coroutine_fn qed_co_request(BlockDriverState *bs, int64_t sector_num,
1279 QEMUIOVector *qiov, int nb_sectors,
1284 .cur_pos = (uint64_t) sector_num * BDRV_SECTOR_SIZE,
1285 .end_pos = (sector_num + nb_sectors) * BDRV_SECTOR_SIZE,
1289 qemu_iovec_init(&acb.cur_qiov, qiov->niov);
1291 trace_qed_aio_setup(bs->opaque, &acb, sector_num, nb_sectors, NULL, flags);
1294 return qed_aio_next_io(&acb);
1297 static int coroutine_fn bdrv_qed_co_readv(BlockDriverState *bs,
1298 int64_t sector_num, int nb_sectors,
1301 return qed_co_request(bs, sector_num, qiov, nb_sectors, 0);
1304 static int coroutine_fn bdrv_qed_co_writev(BlockDriverState *bs,
1305 int64_t sector_num, int nb_sectors,
1308 return qed_co_request(bs, sector_num, qiov, nb_sectors, QED_AIOCB_WRITE);
1311 static int coroutine_fn bdrv_qed_co_pwrite_zeroes(BlockDriverState *bs,
1314 BdrvRequestFlags flags)
1316 BDRVQEDState *s = bs->opaque;
1320 /* Fall back if the request is not aligned */
1321 if (qed_offset_into_cluster(s, offset) ||
1322 qed_offset_into_cluster(s, count)) {
1326 /* Zero writes start without an I/O buffer. If a buffer becomes necessary
1327 * then it will be allocated during request processing.
1329 iov.iov_base = NULL;
1330 iov.iov_len = count;
1332 qemu_iovec_init_external(&qiov, &iov, 1);
1333 return qed_co_request(bs, offset >> BDRV_SECTOR_BITS, &qiov,
1334 count >> BDRV_SECTOR_BITS,
1335 QED_AIOCB_WRITE | QED_AIOCB_ZERO);
1338 static int bdrv_qed_truncate(BlockDriverState *bs, int64_t offset, Error **errp)
1340 BDRVQEDState *s = bs->opaque;
1341 uint64_t old_image_size;
1344 if (!qed_is_image_size_valid(offset, s->header.cluster_size,
1345 s->header.table_size)) {
1346 error_setg(errp, "Invalid image size specified");
1350 if ((uint64_t)offset < s->header.image_size) {
1351 error_setg(errp, "Shrinking images is currently not supported");
1355 old_image_size = s->header.image_size;
1356 s->header.image_size = offset;
1357 ret = qed_write_header_sync(s);
1359 s->header.image_size = old_image_size;
1360 error_setg_errno(errp, -ret, "Failed to update the image size");
1365 static int64_t bdrv_qed_getlength(BlockDriverState *bs)
1367 BDRVQEDState *s = bs->opaque;
1368 return s->header.image_size;
1371 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1373 BDRVQEDState *s = bs->opaque;
1375 memset(bdi, 0, sizeof(*bdi));
1376 bdi->cluster_size = s->header.cluster_size;
1377 bdi->is_dirty = s->header.features & QED_F_NEED_CHECK;
1378 bdi->unallocated_blocks_are_zero = true;
1379 bdi->can_write_zeroes_with_unmap = true;
1383 static int bdrv_qed_change_backing_file(BlockDriverState *bs,
1384 const char *backing_file,
1385 const char *backing_fmt)
1387 BDRVQEDState *s = bs->opaque;
1388 QEDHeader new_header, le_header;
1390 size_t buffer_len, backing_file_len;
1393 /* Refuse to set backing filename if unknown compat feature bits are
1394 * active. If the image uses an unknown compat feature then we may not
1395 * know the layout of data following the header structure and cannot safely
1398 if (backing_file && (s->header.compat_features &
1399 ~QED_COMPAT_FEATURE_MASK)) {
1403 memcpy(&new_header, &s->header, sizeof(new_header));
1405 new_header.features &= ~(QED_F_BACKING_FILE |
1406 QED_F_BACKING_FORMAT_NO_PROBE);
1408 /* Adjust feature flags */
1410 new_header.features |= QED_F_BACKING_FILE;
1412 if (qed_fmt_is_raw(backing_fmt)) {
1413 new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
1417 /* Calculate new header size */
1418 backing_file_len = 0;
1421 backing_file_len = strlen(backing_file);
1424 buffer_len = sizeof(new_header);
1425 new_header.backing_filename_offset = buffer_len;
1426 new_header.backing_filename_size = backing_file_len;
1427 buffer_len += backing_file_len;
1429 /* Make sure we can rewrite header without failing */
1430 if (buffer_len > new_header.header_size * new_header.cluster_size) {
1434 /* Prepare new header */
1435 buffer = g_malloc(buffer_len);
1437 qed_header_cpu_to_le(&new_header, &le_header);
1438 memcpy(buffer, &le_header, sizeof(le_header));
1439 buffer_len = sizeof(le_header);
1442 memcpy(buffer + buffer_len, backing_file, backing_file_len);
1443 buffer_len += backing_file_len;
1446 /* Write new header */
1447 ret = bdrv_pwrite_sync(bs->file, 0, buffer, buffer_len);
1450 memcpy(&s->header, &new_header, sizeof(new_header));
1455 static void bdrv_qed_invalidate_cache(BlockDriverState *bs, Error **errp)
1457 BDRVQEDState *s = bs->opaque;
1458 Error *local_err = NULL;
1463 memset(s, 0, sizeof(BDRVQEDState));
1464 ret = bdrv_qed_do_open(bs, NULL, bs->open_flags, &local_err);
1466 error_propagate(errp, local_err);
1467 error_prepend(errp, "Could not reopen qed layer: ");
1469 } else if (ret < 0) {
1470 error_setg_errno(errp, -ret, "Could not reopen qed layer");
1475 static int bdrv_qed_check(BlockDriverState *bs, BdrvCheckResult *result,
1478 BDRVQEDState *s = bs->opaque;
1480 return qed_check(s, result, !!fix);
1483 static QemuOptsList qed_create_opts = {
1484 .name = "qed-create-opts",
1485 .head = QTAILQ_HEAD_INITIALIZER(qed_create_opts.head),
1488 .name = BLOCK_OPT_SIZE,
1489 .type = QEMU_OPT_SIZE,
1490 .help = "Virtual disk size"
1493 .name = BLOCK_OPT_BACKING_FILE,
1494 .type = QEMU_OPT_STRING,
1495 .help = "File name of a base image"
1498 .name = BLOCK_OPT_BACKING_FMT,
1499 .type = QEMU_OPT_STRING,
1500 .help = "Image format of the base image"
1503 .name = BLOCK_OPT_CLUSTER_SIZE,
1504 .type = QEMU_OPT_SIZE,
1505 .help = "Cluster size (in bytes)",
1506 .def_value_str = stringify(QED_DEFAULT_CLUSTER_SIZE)
1509 .name = BLOCK_OPT_TABLE_SIZE,
1510 .type = QEMU_OPT_SIZE,
1511 .help = "L1/L2 table size (in clusters)"
1513 { /* end of list */ }
1517 static BlockDriver bdrv_qed = {
1518 .format_name = "qed",
1519 .instance_size = sizeof(BDRVQEDState),
1520 .create_opts = &qed_create_opts,
1521 .supports_backing = true,
1523 .bdrv_probe = bdrv_qed_probe,
1524 .bdrv_open = bdrv_qed_open,
1525 .bdrv_close = bdrv_qed_close,
1526 .bdrv_reopen_prepare = bdrv_qed_reopen_prepare,
1527 .bdrv_child_perm = bdrv_format_default_perms,
1528 .bdrv_create = bdrv_qed_create,
1529 .bdrv_has_zero_init = bdrv_has_zero_init_1,
1530 .bdrv_co_get_block_status = bdrv_qed_co_get_block_status,
1531 .bdrv_co_readv = bdrv_qed_co_readv,
1532 .bdrv_co_writev = bdrv_qed_co_writev,
1533 .bdrv_co_pwrite_zeroes = bdrv_qed_co_pwrite_zeroes,
1534 .bdrv_truncate = bdrv_qed_truncate,
1535 .bdrv_getlength = bdrv_qed_getlength,
1536 .bdrv_get_info = bdrv_qed_get_info,
1537 .bdrv_refresh_limits = bdrv_qed_refresh_limits,
1538 .bdrv_change_backing_file = bdrv_qed_change_backing_file,
1539 .bdrv_invalidate_cache = bdrv_qed_invalidate_cache,
1540 .bdrv_check = bdrv_qed_check,
1541 .bdrv_detach_aio_context = bdrv_qed_detach_aio_context,
1542 .bdrv_attach_aio_context = bdrv_qed_attach_aio_context,
1543 .bdrv_drain = bdrv_qed_drain,
1546 static void bdrv_qed_init(void)
1548 bdrv_register(&bdrv_qed);
1551 block_init(bdrv_qed_init);