2 * QEMU Enhanced Disk Format
4 * Copyright IBM, Corp. 2010
7 * Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
8 * Anthony Liguori <aliguori@us.ibm.com>
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-timer.h"
19 #include "migration.h"
21 static void qed_aio_cancel(BlockDriverAIOCB *blockacb)
23 QEDAIOCB *acb = (QEDAIOCB *)blockacb;
24 bool finished = false;
26 /* Wait for the request to finish */
27 acb->finished = &finished;
33 static AIOPool qed_aio_pool = {
34 .aiocb_size = sizeof(QEDAIOCB),
35 .cancel = qed_aio_cancel,
38 static int bdrv_qed_probe(const uint8_t *buf, int buf_size,
41 const QEDHeader *header = (const QEDHeader *)buf;
43 if (buf_size < sizeof(*header)) {
46 if (le32_to_cpu(header->magic) != QED_MAGIC) {
53 * Check whether an image format is raw
55 * @fmt: Backing file format, may be NULL
57 static bool qed_fmt_is_raw(const char *fmt)
59 return fmt && strcmp(fmt, "raw") == 0;
62 static void qed_header_le_to_cpu(const QEDHeader *le, QEDHeader *cpu)
64 cpu->magic = le32_to_cpu(le->magic);
65 cpu->cluster_size = le32_to_cpu(le->cluster_size);
66 cpu->table_size = le32_to_cpu(le->table_size);
67 cpu->header_size = le32_to_cpu(le->header_size);
68 cpu->features = le64_to_cpu(le->features);
69 cpu->compat_features = le64_to_cpu(le->compat_features);
70 cpu->autoclear_features = le64_to_cpu(le->autoclear_features);
71 cpu->l1_table_offset = le64_to_cpu(le->l1_table_offset);
72 cpu->image_size = le64_to_cpu(le->image_size);
73 cpu->backing_filename_offset = le32_to_cpu(le->backing_filename_offset);
74 cpu->backing_filename_size = le32_to_cpu(le->backing_filename_size);
77 static void qed_header_cpu_to_le(const QEDHeader *cpu, QEDHeader *le)
79 le->magic = cpu_to_le32(cpu->magic);
80 le->cluster_size = cpu_to_le32(cpu->cluster_size);
81 le->table_size = cpu_to_le32(cpu->table_size);
82 le->header_size = cpu_to_le32(cpu->header_size);
83 le->features = cpu_to_le64(cpu->features);
84 le->compat_features = cpu_to_le64(cpu->compat_features);
85 le->autoclear_features = cpu_to_le64(cpu->autoclear_features);
86 le->l1_table_offset = cpu_to_le64(cpu->l1_table_offset);
87 le->image_size = cpu_to_le64(cpu->image_size);
88 le->backing_filename_offset = cpu_to_le32(cpu->backing_filename_offset);
89 le->backing_filename_size = cpu_to_le32(cpu->backing_filename_size);
92 static int qed_write_header_sync(BDRVQEDState *s)
97 qed_header_cpu_to_le(&s->header, &le);
98 ret = bdrv_pwrite(s->bs->file, 0, &le, sizeof(le));
99 if (ret != sizeof(le)) {
114 static void qed_write_header_cb(void *opaque, int ret)
116 QEDWriteHeaderCB *write_header_cb = opaque;
118 qemu_vfree(write_header_cb->buf);
119 gencb_complete(write_header_cb, ret);
122 static void qed_write_header_read_cb(void *opaque, int ret)
124 QEDWriteHeaderCB *write_header_cb = opaque;
125 BDRVQEDState *s = write_header_cb->s;
128 qed_write_header_cb(write_header_cb, ret);
133 qed_header_cpu_to_le(&s->header, (QEDHeader *)write_header_cb->buf);
135 bdrv_aio_writev(s->bs->file, 0, &write_header_cb->qiov,
136 write_header_cb->nsectors, qed_write_header_cb,
141 * Update header in-place (does not rewrite backing filename or other strings)
143 * This function only updates known header fields in-place and does not affect
144 * extra data after the QED header.
146 static void qed_write_header(BDRVQEDState *s, BlockDriverCompletionFunc cb,
149 /* We must write full sectors for O_DIRECT but cannot necessarily generate
150 * the data following the header if an unrecognized compat feature is
151 * active. Therefore, first read the sectors containing the header, update
152 * them, and write back.
155 int nsectors = (sizeof(QEDHeader) + BDRV_SECTOR_SIZE - 1) /
157 size_t len = nsectors * BDRV_SECTOR_SIZE;
158 QEDWriteHeaderCB *write_header_cb = gencb_alloc(sizeof(*write_header_cb),
161 write_header_cb->s = s;
162 write_header_cb->nsectors = nsectors;
163 write_header_cb->buf = qemu_blockalign(s->bs, len);
164 write_header_cb->iov.iov_base = write_header_cb->buf;
165 write_header_cb->iov.iov_len = len;
166 qemu_iovec_init_external(&write_header_cb->qiov, &write_header_cb->iov, 1);
168 bdrv_aio_readv(s->bs->file, 0, &write_header_cb->qiov, nsectors,
169 qed_write_header_read_cb, write_header_cb);
172 static uint64_t qed_max_image_size(uint32_t cluster_size, uint32_t table_size)
174 uint64_t table_entries;
177 table_entries = (table_size * cluster_size) / sizeof(uint64_t);
178 l2_size = table_entries * cluster_size;
180 return l2_size * table_entries;
183 static bool qed_is_cluster_size_valid(uint32_t cluster_size)
185 if (cluster_size < QED_MIN_CLUSTER_SIZE ||
186 cluster_size > QED_MAX_CLUSTER_SIZE) {
189 if (cluster_size & (cluster_size - 1)) {
190 return false; /* not power of 2 */
195 static bool qed_is_table_size_valid(uint32_t table_size)
197 if (table_size < QED_MIN_TABLE_SIZE ||
198 table_size > QED_MAX_TABLE_SIZE) {
201 if (table_size & (table_size - 1)) {
202 return false; /* not power of 2 */
207 static bool qed_is_image_size_valid(uint64_t image_size, uint32_t cluster_size,
210 if (image_size % BDRV_SECTOR_SIZE != 0) {
211 return false; /* not multiple of sector size */
213 if (image_size > qed_max_image_size(cluster_size, table_size)) {
214 return false; /* image is too large */
220 * Read a string of known length from the image file
223 * @offset: File offset to start of string, in bytes
224 * @n: String length in bytes
225 * @buf: Destination buffer
226 * @buflen: Destination buffer length in bytes
227 * @ret: 0 on success, -errno on failure
229 * The string is NUL-terminated.
231 static int qed_read_string(BlockDriverState *file, uint64_t offset, size_t n,
232 char *buf, size_t buflen)
238 ret = bdrv_pread(file, offset, buf, n);
247 * Allocate new clusters
250 * @n: Number of contiguous clusters to allocate
251 * @ret: Offset of first allocated cluster
253 * This function only produces the offset where the new clusters should be
254 * written. It updates BDRVQEDState but does not make any changes to the image
257 static uint64_t qed_alloc_clusters(BDRVQEDState *s, unsigned int n)
259 uint64_t offset = s->file_size;
260 s->file_size += n * s->header.cluster_size;
264 QEDTable *qed_alloc_table(BDRVQEDState *s)
266 /* Honor O_DIRECT memory alignment requirements */
267 return qemu_blockalign(s->bs,
268 s->header.cluster_size * s->header.table_size);
272 * Allocate a new zeroed L2 table
274 static CachedL2Table *qed_new_l2_table(BDRVQEDState *s)
276 CachedL2Table *l2_table = qed_alloc_l2_cache_entry(&s->l2_cache);
278 l2_table->table = qed_alloc_table(s);
279 l2_table->offset = qed_alloc_clusters(s, s->header.table_size);
281 memset(l2_table->table->offsets, 0,
282 s->header.cluster_size * s->header.table_size);
286 static void qed_aio_next_io(void *opaque, int ret);
288 static void qed_plug_allocating_write_reqs(BDRVQEDState *s)
290 assert(!s->allocating_write_reqs_plugged);
292 s->allocating_write_reqs_plugged = true;
295 static void qed_unplug_allocating_write_reqs(BDRVQEDState *s)
299 assert(s->allocating_write_reqs_plugged);
301 s->allocating_write_reqs_plugged = false;
303 acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs);
305 qed_aio_next_io(acb, 0);
309 static void qed_finish_clear_need_check(void *opaque, int ret)
314 static void qed_flush_after_clear_need_check(void *opaque, int ret)
316 BDRVQEDState *s = opaque;
318 bdrv_aio_flush(s->bs, qed_finish_clear_need_check, s);
320 /* No need to wait until flush completes */
321 qed_unplug_allocating_write_reqs(s);
324 static void qed_clear_need_check(void *opaque, int ret)
326 BDRVQEDState *s = opaque;
329 qed_unplug_allocating_write_reqs(s);
333 s->header.features &= ~QED_F_NEED_CHECK;
334 qed_write_header(s, qed_flush_after_clear_need_check, s);
337 static void qed_need_check_timer_cb(void *opaque)
339 BDRVQEDState *s = opaque;
341 /* The timer should only fire when allocating writes have drained */
342 assert(!QSIMPLEQ_FIRST(&s->allocating_write_reqs));
344 trace_qed_need_check_timer_cb(s);
346 qed_plug_allocating_write_reqs(s);
348 /* Ensure writes are on disk before clearing flag */
349 bdrv_aio_flush(s->bs, qed_clear_need_check, s);
352 static void qed_start_need_check_timer(BDRVQEDState *s)
354 trace_qed_start_need_check_timer(s);
356 /* Use vm_clock so we don't alter the image file while suspended for
359 qemu_mod_timer(s->need_check_timer, qemu_get_clock_ns(vm_clock) +
360 get_ticks_per_sec() * QED_NEED_CHECK_TIMEOUT);
363 /* It's okay to call this multiple times or when no timer is started */
364 static void qed_cancel_need_check_timer(BDRVQEDState *s)
366 trace_qed_cancel_need_check_timer(s);
367 qemu_del_timer(s->need_check_timer);
370 static void bdrv_qed_rebind(BlockDriverState *bs)
372 BDRVQEDState *s = bs->opaque;
376 static int bdrv_qed_open(BlockDriverState *bs, int flags)
378 BDRVQEDState *s = bs->opaque;
384 QSIMPLEQ_INIT(&s->allocating_write_reqs);
386 ret = bdrv_pread(bs->file, 0, &le_header, sizeof(le_header));
390 qed_header_le_to_cpu(&le_header, &s->header);
392 if (s->header.magic != QED_MAGIC) {
395 if (s->header.features & ~QED_FEATURE_MASK) {
396 /* image uses unsupported feature bits */
398 snprintf(buf, sizeof(buf), "%" PRIx64,
399 s->header.features & ~QED_FEATURE_MASK);
400 qerror_report(QERR_UNKNOWN_BLOCK_FORMAT_FEATURE,
401 bs->device_name, "QED", buf);
404 if (!qed_is_cluster_size_valid(s->header.cluster_size)) {
408 /* Round down file size to the last cluster */
409 file_size = bdrv_getlength(bs->file);
413 s->file_size = qed_start_of_cluster(s, file_size);
415 if (!qed_is_table_size_valid(s->header.table_size)) {
418 if (!qed_is_image_size_valid(s->header.image_size,
419 s->header.cluster_size,
420 s->header.table_size)) {
423 if (!qed_check_table_offset(s, s->header.l1_table_offset)) {
427 s->table_nelems = (s->header.cluster_size * s->header.table_size) /
429 s->l2_shift = ffs(s->header.cluster_size) - 1;
430 s->l2_mask = s->table_nelems - 1;
431 s->l1_shift = s->l2_shift + ffs(s->table_nelems) - 1;
433 if ((s->header.features & QED_F_BACKING_FILE)) {
434 if ((uint64_t)s->header.backing_filename_offset +
435 s->header.backing_filename_size >
436 s->header.cluster_size * s->header.header_size) {
440 ret = qed_read_string(bs->file, s->header.backing_filename_offset,
441 s->header.backing_filename_size, bs->backing_file,
442 sizeof(bs->backing_file));
447 if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) {
448 pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw");
452 /* Reset unknown autoclear feature bits. This is a backwards
453 * compatibility mechanism that allows images to be opened by older
454 * programs, which "knock out" unknown feature bits. When an image is
455 * opened by a newer program again it can detect that the autoclear
456 * feature is no longer valid.
458 if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 &&
459 !bdrv_is_read_only(bs->file) && !(flags & BDRV_O_INCOMING)) {
460 s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK;
462 ret = qed_write_header_sync(s);
467 /* From here on only known autoclear feature bits are valid */
468 bdrv_flush(bs->file);
471 s->l1_table = qed_alloc_table(s);
472 qed_init_l2_cache(&s->l2_cache);
474 ret = qed_read_l1_table_sync(s);
479 /* If image was not closed cleanly, check consistency */
480 if (s->header.features & QED_F_NEED_CHECK) {
481 /* Read-only images cannot be fixed. There is no risk of corruption
482 * since write operations are not possible. Therefore, allow
483 * potentially inconsistent images to be opened read-only. This can
484 * aid data recovery from an otherwise inconsistent image.
486 if (!bdrv_is_read_only(bs->file) &&
487 !(flags & BDRV_O_INCOMING)) {
488 BdrvCheckResult result = {0};
490 ret = qed_check(s, &result, true);
494 if (!result.corruptions && !result.check_errors) {
495 /* Ensure fixes reach storage before clearing check bit */
498 s->header.features &= ~QED_F_NEED_CHECK;
499 qed_write_header_sync(s);
504 s->need_check_timer = qemu_new_timer_ns(vm_clock,
505 qed_need_check_timer_cb, s);
509 qed_free_l2_cache(&s->l2_cache);
510 qemu_vfree(s->l1_table);
515 static void bdrv_qed_close(BlockDriverState *bs)
517 BDRVQEDState *s = bs->opaque;
519 qed_cancel_need_check_timer(s);
520 qemu_free_timer(s->need_check_timer);
522 /* Ensure writes reach stable storage */
523 bdrv_flush(bs->file);
525 /* Clean shutdown, no check required on next open */
526 if (s->header.features & QED_F_NEED_CHECK) {
527 s->header.features &= ~QED_F_NEED_CHECK;
528 qed_write_header_sync(s);
531 qed_free_l2_cache(&s->l2_cache);
532 qemu_vfree(s->l1_table);
535 static int qed_create(const char *filename, uint32_t cluster_size,
536 uint64_t image_size, uint32_t table_size,
537 const char *backing_file, const char *backing_fmt)
541 .cluster_size = cluster_size,
542 .table_size = table_size,
545 .compat_features = 0,
546 .l1_table_offset = cluster_size,
547 .image_size = image_size,
550 uint8_t *l1_table = NULL;
551 size_t l1_size = header.cluster_size * header.table_size;
553 BlockDriverState *bs = NULL;
555 ret = bdrv_create_file(filename, NULL);
560 ret = bdrv_file_open(&bs, filename, BDRV_O_RDWR | BDRV_O_CACHE_WB);
565 /* File must start empty and grow, check truncate is supported */
566 ret = bdrv_truncate(bs, 0);
572 header.features |= QED_F_BACKING_FILE;
573 header.backing_filename_offset = sizeof(le_header);
574 header.backing_filename_size = strlen(backing_file);
576 if (qed_fmt_is_raw(backing_fmt)) {
577 header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
581 qed_header_cpu_to_le(&header, &le_header);
582 ret = bdrv_pwrite(bs, 0, &le_header, sizeof(le_header));
586 ret = bdrv_pwrite(bs, sizeof(le_header), backing_file,
587 header.backing_filename_size);
592 l1_table = g_malloc0(l1_size);
593 ret = bdrv_pwrite(bs, header.l1_table_offset, l1_table, l1_size);
598 ret = 0; /* success */
605 static int bdrv_qed_create(const char *filename, QEMUOptionParameter *options)
607 uint64_t image_size = 0;
608 uint32_t cluster_size = QED_DEFAULT_CLUSTER_SIZE;
609 uint32_t table_size = QED_DEFAULT_TABLE_SIZE;
610 const char *backing_file = NULL;
611 const char *backing_fmt = NULL;
613 while (options && options->name) {
614 if (!strcmp(options->name, BLOCK_OPT_SIZE)) {
615 image_size = options->value.n;
616 } else if (!strcmp(options->name, BLOCK_OPT_BACKING_FILE)) {
617 backing_file = options->value.s;
618 } else if (!strcmp(options->name, BLOCK_OPT_BACKING_FMT)) {
619 backing_fmt = options->value.s;
620 } else if (!strcmp(options->name, BLOCK_OPT_CLUSTER_SIZE)) {
621 if (options->value.n) {
622 cluster_size = options->value.n;
624 } else if (!strcmp(options->name, BLOCK_OPT_TABLE_SIZE)) {
625 if (options->value.n) {
626 table_size = options->value.n;
632 if (!qed_is_cluster_size_valid(cluster_size)) {
633 fprintf(stderr, "QED cluster size must be within range [%u, %u] and power of 2\n",
634 QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE);
637 if (!qed_is_table_size_valid(table_size)) {
638 fprintf(stderr, "QED table size must be within range [%u, %u] and power of 2\n",
639 QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE);
642 if (!qed_is_image_size_valid(image_size, cluster_size, table_size)) {
643 fprintf(stderr, "QED image size must be a non-zero multiple of "
644 "cluster size and less than %" PRIu64 " bytes\n",
645 qed_max_image_size(cluster_size, table_size));
649 return qed_create(filename, cluster_size, image_size, table_size,
650 backing_file, backing_fmt);
659 static void qed_is_allocated_cb(void *opaque, int ret, uint64_t offset, size_t len)
661 QEDIsAllocatedCB *cb = opaque;
662 *cb->pnum = len / BDRV_SECTOR_SIZE;
663 cb->is_allocated = (ret == QED_CLUSTER_FOUND || ret == QED_CLUSTER_ZERO);
665 qemu_coroutine_enter(cb->co, NULL);
669 static int coroutine_fn bdrv_qed_co_is_allocated(BlockDriverState *bs,
671 int nb_sectors, int *pnum)
673 BDRVQEDState *s = bs->opaque;
674 uint64_t pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE;
675 size_t len = (size_t)nb_sectors * BDRV_SECTOR_SIZE;
676 QEDIsAllocatedCB cb = {
680 QEDRequest request = { .l2_table = NULL };
682 qed_find_cluster(s, &request, pos, len, qed_is_allocated_cb, &cb);
684 /* Now sleep if the callback wasn't invoked immediately */
685 while (cb.is_allocated == -1) {
686 cb.co = qemu_coroutine_self();
687 qemu_coroutine_yield();
690 qed_unref_l2_cache_entry(request.l2_table);
692 return cb.is_allocated;
695 static int bdrv_qed_make_empty(BlockDriverState *bs)
700 static BDRVQEDState *acb_to_s(QEDAIOCB *acb)
702 return acb->common.bs->opaque;
706 * Read from the backing file or zero-fill if no backing file
709 * @pos: Byte position in device
710 * @qiov: Destination I/O vector
711 * @cb: Completion function
712 * @opaque: User data for completion function
714 * This function reads qiov->size bytes starting at pos from the backing file.
715 * If there is no backing file then zeroes are read.
717 static void qed_read_backing_file(BDRVQEDState *s, uint64_t pos,
719 BlockDriverCompletionFunc *cb, void *opaque)
721 uint64_t backing_length = 0;
724 /* If there is a backing file, get its length. Treat the absence of a
725 * backing file like a zero length backing file.
727 if (s->bs->backing_hd) {
728 int64_t l = bdrv_getlength(s->bs->backing_hd);
736 /* Zero all sectors if reading beyond the end of the backing file */
737 if (pos >= backing_length ||
738 pos + qiov->size > backing_length) {
739 qemu_iovec_memset(qiov, 0, qiov->size);
742 /* Complete now if there are no backing file sectors to read */
743 if (pos >= backing_length) {
748 /* If the read straddles the end of the backing file, shorten it */
749 size = MIN((uint64_t)backing_length - pos, qiov->size);
751 BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING);
752 bdrv_aio_readv(s->bs->backing_hd, pos / BDRV_SECTOR_SIZE,
753 qiov, size / BDRV_SECTOR_SIZE, cb, opaque);
762 } CopyFromBackingFileCB;
764 static void qed_copy_from_backing_file_cb(void *opaque, int ret)
766 CopyFromBackingFileCB *copy_cb = opaque;
767 qemu_vfree(copy_cb->iov.iov_base);
768 gencb_complete(©_cb->gencb, ret);
771 static void qed_copy_from_backing_file_write(void *opaque, int ret)
773 CopyFromBackingFileCB *copy_cb = opaque;
774 BDRVQEDState *s = copy_cb->s;
777 qed_copy_from_backing_file_cb(copy_cb, ret);
781 BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE);
782 bdrv_aio_writev(s->bs->file, copy_cb->offset / BDRV_SECTOR_SIZE,
783 ©_cb->qiov, copy_cb->qiov.size / BDRV_SECTOR_SIZE,
784 qed_copy_from_backing_file_cb, copy_cb);
788 * Copy data from backing file into the image
791 * @pos: Byte position in device
792 * @len: Number of bytes
793 * @offset: Byte offset in image file
794 * @cb: Completion function
795 * @opaque: User data for completion function
797 static void qed_copy_from_backing_file(BDRVQEDState *s, uint64_t pos,
798 uint64_t len, uint64_t offset,
799 BlockDriverCompletionFunc *cb,
802 CopyFromBackingFileCB *copy_cb;
804 /* Skip copy entirely if there is no work to do */
810 copy_cb = gencb_alloc(sizeof(*copy_cb), cb, opaque);
812 copy_cb->offset = offset;
813 copy_cb->iov.iov_base = qemu_blockalign(s->bs, len);
814 copy_cb->iov.iov_len = len;
815 qemu_iovec_init_external(©_cb->qiov, ©_cb->iov, 1);
817 qed_read_backing_file(s, pos, ©_cb->qiov,
818 qed_copy_from_backing_file_write, copy_cb);
822 * Link one or more contiguous clusters into a table
826 * @index: First cluster index
827 * @n: Number of contiguous clusters
828 * @cluster: First cluster offset
830 * The cluster offset may be an allocated byte offset in the image file, the
831 * zero cluster marker, or the unallocated cluster marker.
833 static void qed_update_l2_table(BDRVQEDState *s, QEDTable *table, int index,
834 unsigned int n, uint64_t cluster)
837 for (i = index; i < index + n; i++) {
838 table->offsets[i] = cluster;
839 if (!qed_offset_is_unalloc_cluster(cluster) &&
840 !qed_offset_is_zero_cluster(cluster)) {
841 cluster += s->header.cluster_size;
846 static void qed_aio_complete_bh(void *opaque)
848 QEDAIOCB *acb = opaque;
849 BlockDriverCompletionFunc *cb = acb->common.cb;
850 void *user_opaque = acb->common.opaque;
851 int ret = acb->bh_ret;
852 bool *finished = acb->finished;
854 qemu_bh_delete(acb->bh);
855 qemu_aio_release(acb);
857 /* Invoke callback */
858 cb(user_opaque, ret);
860 /* Signal cancel completion */
866 static void qed_aio_complete(QEDAIOCB *acb, int ret)
868 BDRVQEDState *s = acb_to_s(acb);
870 trace_qed_aio_complete(s, acb, ret);
873 qemu_iovec_destroy(&acb->cur_qiov);
874 qed_unref_l2_cache_entry(acb->request.l2_table);
876 /* Free the buffer we may have allocated for zero writes */
877 if (acb->flags & QED_AIOCB_ZERO) {
878 qemu_vfree(acb->qiov->iov[0].iov_base);
879 acb->qiov->iov[0].iov_base = NULL;
882 /* Arrange for a bh to invoke the completion function */
884 acb->bh = qemu_bh_new(qed_aio_complete_bh, acb);
885 qemu_bh_schedule(acb->bh);
887 /* Start next allocating write request waiting behind this one. Note that
888 * requests enqueue themselves when they first hit an unallocated cluster
889 * but they wait until the entire request is finished before waking up the
890 * next request in the queue. This ensures that we don't cycle through
891 * requests multiple times but rather finish one at a time completely.
893 if (acb == QSIMPLEQ_FIRST(&s->allocating_write_reqs)) {
894 QSIMPLEQ_REMOVE_HEAD(&s->allocating_write_reqs, next);
895 acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs);
897 qed_aio_next_io(acb, 0);
898 } else if (s->header.features & QED_F_NEED_CHECK) {
899 qed_start_need_check_timer(s);
905 * Commit the current L2 table to the cache
907 static void qed_commit_l2_update(void *opaque, int ret)
909 QEDAIOCB *acb = opaque;
910 BDRVQEDState *s = acb_to_s(acb);
911 CachedL2Table *l2_table = acb->request.l2_table;
912 uint64_t l2_offset = l2_table->offset;
914 qed_commit_l2_cache_entry(&s->l2_cache, l2_table);
916 /* This is guaranteed to succeed because we just committed the entry to the
919 acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset);
920 assert(acb->request.l2_table != NULL);
922 qed_aio_next_io(opaque, ret);
926 * Update L1 table with new L2 table offset and write it out
928 static void qed_aio_write_l1_update(void *opaque, int ret)
930 QEDAIOCB *acb = opaque;
931 BDRVQEDState *s = acb_to_s(acb);
935 qed_aio_complete(acb, ret);
939 index = qed_l1_index(s, acb->cur_pos);
940 s->l1_table->offsets[index] = acb->request.l2_table->offset;
942 qed_write_l1_table(s, index, 1, qed_commit_l2_update, acb);
946 * Update L2 table with new cluster offsets and write them out
948 static void qed_aio_write_l2_update(QEDAIOCB *acb, int ret, uint64_t offset)
950 BDRVQEDState *s = acb_to_s(acb);
951 bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1;
959 qed_unref_l2_cache_entry(acb->request.l2_table);
960 acb->request.l2_table = qed_new_l2_table(s);
963 index = qed_l2_index(s, acb->cur_pos);
964 qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters,
968 /* Write out the whole new L2 table */
969 qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true,
970 qed_aio_write_l1_update, acb);
972 /* Write out only the updated part of the L2 table */
973 qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters, false,
974 qed_aio_next_io, acb);
979 qed_aio_complete(acb, ret);
982 static void qed_aio_write_l2_update_cb(void *opaque, int ret)
984 QEDAIOCB *acb = opaque;
985 qed_aio_write_l2_update(acb, ret, acb->cur_cluster);
989 * Flush new data clusters before updating the L2 table
991 * This flush is necessary when a backing file is in use. A crash during an
992 * allocating write could result in empty clusters in the image. If the write
993 * only touched a subregion of the cluster, then backing image sectors have
994 * been lost in the untouched region. The solution is to flush after writing a
995 * new data cluster and before updating the L2 table.
997 static void qed_aio_write_flush_before_l2_update(void *opaque, int ret)
999 QEDAIOCB *acb = opaque;
1000 BDRVQEDState *s = acb_to_s(acb);
1002 if (!bdrv_aio_flush(s->bs->file, qed_aio_write_l2_update_cb, opaque)) {
1003 qed_aio_complete(acb, -EIO);
1008 * Write data to the image file
1010 static void qed_aio_write_main(void *opaque, int ret)
1012 QEDAIOCB *acb = opaque;
1013 BDRVQEDState *s = acb_to_s(acb);
1014 uint64_t offset = acb->cur_cluster +
1015 qed_offset_into_cluster(s, acb->cur_pos);
1016 BlockDriverCompletionFunc *next_fn;
1018 trace_qed_aio_write_main(s, acb, ret, offset, acb->cur_qiov.size);
1021 qed_aio_complete(acb, ret);
1025 if (acb->find_cluster_ret == QED_CLUSTER_FOUND) {
1026 next_fn = qed_aio_next_io;
1028 if (s->bs->backing_hd) {
1029 next_fn = qed_aio_write_flush_before_l2_update;
1031 next_fn = qed_aio_write_l2_update_cb;
1035 BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO);
1036 bdrv_aio_writev(s->bs->file, offset / BDRV_SECTOR_SIZE,
1037 &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE,
1042 * Populate back untouched region of new data cluster
1044 static void qed_aio_write_postfill(void *opaque, int ret)
1046 QEDAIOCB *acb = opaque;
1047 BDRVQEDState *s = acb_to_s(acb);
1048 uint64_t start = acb->cur_pos + acb->cur_qiov.size;
1050 qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start;
1051 uint64_t offset = acb->cur_cluster +
1052 qed_offset_into_cluster(s, acb->cur_pos) +
1056 qed_aio_complete(acb, ret);
1060 trace_qed_aio_write_postfill(s, acb, start, len, offset);
1061 qed_copy_from_backing_file(s, start, len, offset,
1062 qed_aio_write_main, acb);
1066 * Populate front untouched region of new data cluster
1068 static void qed_aio_write_prefill(void *opaque, int ret)
1070 QEDAIOCB *acb = opaque;
1071 BDRVQEDState *s = acb_to_s(acb);
1072 uint64_t start = qed_start_of_cluster(s, acb->cur_pos);
1073 uint64_t len = qed_offset_into_cluster(s, acb->cur_pos);
1075 trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster);
1076 qed_copy_from_backing_file(s, start, len, acb->cur_cluster,
1077 qed_aio_write_postfill, acb);
1081 * Check if the QED_F_NEED_CHECK bit should be set during allocating write
1083 static bool qed_should_set_need_check(BDRVQEDState *s)
1085 /* The flush before L2 update path ensures consistency */
1086 if (s->bs->backing_hd) {
1090 return !(s->header.features & QED_F_NEED_CHECK);
1093 static void qed_aio_write_zero_cluster(void *opaque, int ret)
1095 QEDAIOCB *acb = opaque;
1098 qed_aio_complete(acb, ret);
1102 qed_aio_write_l2_update(acb, 0, 1);
1106 * Write new data cluster
1108 * @acb: Write request
1109 * @len: Length in bytes
1111 * This path is taken when writing to previously unallocated clusters.
1113 static void qed_aio_write_alloc(QEDAIOCB *acb, size_t len)
1115 BDRVQEDState *s = acb_to_s(acb);
1116 BlockDriverCompletionFunc *cb;
1118 /* Cancel timer when the first allocating request comes in */
1119 if (QSIMPLEQ_EMPTY(&s->allocating_write_reqs)) {
1120 qed_cancel_need_check_timer(s);
1123 /* Freeze this request if another allocating write is in progress */
1124 if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs)) {
1125 QSIMPLEQ_INSERT_TAIL(&s->allocating_write_reqs, acb, next);
1127 if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs) ||
1128 s->allocating_write_reqs_plugged) {
1129 return; /* wait for existing request to finish */
1132 acb->cur_nclusters = qed_bytes_to_clusters(s,
1133 qed_offset_into_cluster(s, acb->cur_pos) + len);
1134 qemu_iovec_copy(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1136 if (acb->flags & QED_AIOCB_ZERO) {
1137 /* Skip ahead if the clusters are already zero */
1138 if (acb->find_cluster_ret == QED_CLUSTER_ZERO) {
1139 qed_aio_next_io(acb, 0);
1143 cb = qed_aio_write_zero_cluster;
1145 cb = qed_aio_write_prefill;
1146 acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters);
1149 if (qed_should_set_need_check(s)) {
1150 s->header.features |= QED_F_NEED_CHECK;
1151 qed_write_header(s, cb, acb);
1158 * Write data cluster in place
1160 * @acb: Write request
1161 * @offset: Cluster offset in bytes
1162 * @len: Length in bytes
1164 * This path is taken when writing to already allocated clusters.
1166 static void qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset, size_t len)
1168 /* Allocate buffer for zero writes */
1169 if (acb->flags & QED_AIOCB_ZERO) {
1170 struct iovec *iov = acb->qiov->iov;
1172 if (!iov->iov_base) {
1173 iov->iov_base = qemu_blockalign(acb->common.bs, iov->iov_len);
1174 memset(iov->iov_base, 0, iov->iov_len);
1178 /* Calculate the I/O vector */
1179 acb->cur_cluster = offset;
1180 qemu_iovec_copy(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1182 /* Do the actual write */
1183 qed_aio_write_main(acb, 0);
1187 * Write data cluster
1189 * @opaque: Write request
1190 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1,
1192 * @offset: Cluster offset in bytes
1193 * @len: Length in bytes
1195 * Callback from qed_find_cluster().
1197 static void qed_aio_write_data(void *opaque, int ret,
1198 uint64_t offset, size_t len)
1200 QEDAIOCB *acb = opaque;
1202 trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len);
1204 acb->find_cluster_ret = ret;
1207 case QED_CLUSTER_FOUND:
1208 qed_aio_write_inplace(acb, offset, len);
1211 case QED_CLUSTER_L2:
1212 case QED_CLUSTER_L1:
1213 case QED_CLUSTER_ZERO:
1214 qed_aio_write_alloc(acb, len);
1218 qed_aio_complete(acb, ret);
1226 * @opaque: Read request
1227 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1,
1229 * @offset: Cluster offset in bytes
1230 * @len: Length in bytes
1232 * Callback from qed_find_cluster().
1234 static void qed_aio_read_data(void *opaque, int ret,
1235 uint64_t offset, size_t len)
1237 QEDAIOCB *acb = opaque;
1238 BDRVQEDState *s = acb_to_s(acb);
1239 BlockDriverState *bs = acb->common.bs;
1241 /* Adjust offset into cluster */
1242 offset += qed_offset_into_cluster(s, acb->cur_pos);
1244 trace_qed_aio_read_data(s, acb, ret, offset, len);
1250 qemu_iovec_copy(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1252 /* Handle zero cluster and backing file reads */
1253 if (ret == QED_CLUSTER_ZERO) {
1254 qemu_iovec_memset(&acb->cur_qiov, 0, acb->cur_qiov.size);
1255 qed_aio_next_io(acb, 0);
1257 } else if (ret != QED_CLUSTER_FOUND) {
1258 qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov,
1259 qed_aio_next_io, acb);
1263 BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
1264 bdrv_aio_readv(bs->file, offset / BDRV_SECTOR_SIZE,
1265 &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE,
1266 qed_aio_next_io, acb);
1270 qed_aio_complete(acb, ret);
1274 * Begin next I/O or complete the request
1276 static void qed_aio_next_io(void *opaque, int ret)
1278 QEDAIOCB *acb = opaque;
1279 BDRVQEDState *s = acb_to_s(acb);
1280 QEDFindClusterFunc *io_fn = (acb->flags & QED_AIOCB_WRITE) ?
1281 qed_aio_write_data : qed_aio_read_data;
1283 trace_qed_aio_next_io(s, acb, ret, acb->cur_pos + acb->cur_qiov.size);
1285 /* Handle I/O error */
1287 qed_aio_complete(acb, ret);
1291 acb->qiov_offset += acb->cur_qiov.size;
1292 acb->cur_pos += acb->cur_qiov.size;
1293 qemu_iovec_reset(&acb->cur_qiov);
1295 /* Complete request */
1296 if (acb->cur_pos >= acb->end_pos) {
1297 qed_aio_complete(acb, 0);
1301 /* Find next cluster and start I/O */
1302 qed_find_cluster(s, &acb->request,
1303 acb->cur_pos, acb->end_pos - acb->cur_pos,
1307 static BlockDriverAIOCB *qed_aio_setup(BlockDriverState *bs,
1309 QEMUIOVector *qiov, int nb_sectors,
1310 BlockDriverCompletionFunc *cb,
1311 void *opaque, int flags)
1313 QEDAIOCB *acb = qemu_aio_get(&qed_aio_pool, bs, cb, opaque);
1315 trace_qed_aio_setup(bs->opaque, acb, sector_num, nb_sectors,
1319 acb->finished = NULL;
1321 acb->qiov_offset = 0;
1322 acb->cur_pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE;
1323 acb->end_pos = acb->cur_pos + nb_sectors * BDRV_SECTOR_SIZE;
1324 acb->request.l2_table = NULL;
1325 qemu_iovec_init(&acb->cur_qiov, qiov->niov);
1328 qed_aio_next_io(acb, 0);
1329 return &acb->common;
1332 static BlockDriverAIOCB *bdrv_qed_aio_readv(BlockDriverState *bs,
1334 QEMUIOVector *qiov, int nb_sectors,
1335 BlockDriverCompletionFunc *cb,
1338 return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb, opaque, 0);
1341 static BlockDriverAIOCB *bdrv_qed_aio_writev(BlockDriverState *bs,
1343 QEMUIOVector *qiov, int nb_sectors,
1344 BlockDriverCompletionFunc *cb,
1347 return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb,
1348 opaque, QED_AIOCB_WRITE);
1357 static void coroutine_fn qed_co_write_zeroes_cb(void *opaque, int ret)
1359 QEDWriteZeroesCB *cb = opaque;
1364 qemu_coroutine_enter(cb->co, NULL);
1368 static int coroutine_fn bdrv_qed_co_write_zeroes(BlockDriverState *bs,
1372 BlockDriverAIOCB *blockacb;
1373 QEDWriteZeroesCB cb = { .done = false };
1377 /* Zero writes start without an I/O buffer. If a buffer becomes necessary
1378 * then it will be allocated during request processing.
1380 iov.iov_base = NULL,
1381 iov.iov_len = nb_sectors * BDRV_SECTOR_SIZE,
1383 qemu_iovec_init_external(&qiov, &iov, 1);
1384 blockacb = qed_aio_setup(bs, sector_num, &qiov, nb_sectors,
1385 qed_co_write_zeroes_cb, &cb,
1386 QED_AIOCB_WRITE | QED_AIOCB_ZERO);
1391 cb.co = qemu_coroutine_self();
1392 qemu_coroutine_yield();
1398 static int bdrv_qed_truncate(BlockDriverState *bs, int64_t offset)
1400 BDRVQEDState *s = bs->opaque;
1401 uint64_t old_image_size;
1404 if (!qed_is_image_size_valid(offset, s->header.cluster_size,
1405 s->header.table_size)) {
1409 /* Shrinking is currently not supported */
1410 if ((uint64_t)offset < s->header.image_size) {
1414 old_image_size = s->header.image_size;
1415 s->header.image_size = offset;
1416 ret = qed_write_header_sync(s);
1418 s->header.image_size = old_image_size;
1423 static int64_t bdrv_qed_getlength(BlockDriverState *bs)
1425 BDRVQEDState *s = bs->opaque;
1426 return s->header.image_size;
1429 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1431 BDRVQEDState *s = bs->opaque;
1433 memset(bdi, 0, sizeof(*bdi));
1434 bdi->cluster_size = s->header.cluster_size;
1435 bdi->is_dirty = s->header.features & QED_F_NEED_CHECK;
1439 static int bdrv_qed_change_backing_file(BlockDriverState *bs,
1440 const char *backing_file,
1441 const char *backing_fmt)
1443 BDRVQEDState *s = bs->opaque;
1444 QEDHeader new_header, le_header;
1446 size_t buffer_len, backing_file_len;
1449 /* Refuse to set backing filename if unknown compat feature bits are
1450 * active. If the image uses an unknown compat feature then we may not
1451 * know the layout of data following the header structure and cannot safely
1454 if (backing_file && (s->header.compat_features &
1455 ~QED_COMPAT_FEATURE_MASK)) {
1459 memcpy(&new_header, &s->header, sizeof(new_header));
1461 new_header.features &= ~(QED_F_BACKING_FILE |
1462 QED_F_BACKING_FORMAT_NO_PROBE);
1464 /* Adjust feature flags */
1466 new_header.features |= QED_F_BACKING_FILE;
1468 if (qed_fmt_is_raw(backing_fmt)) {
1469 new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
1473 /* Calculate new header size */
1474 backing_file_len = 0;
1477 backing_file_len = strlen(backing_file);
1480 buffer_len = sizeof(new_header);
1481 new_header.backing_filename_offset = buffer_len;
1482 new_header.backing_filename_size = backing_file_len;
1483 buffer_len += backing_file_len;
1485 /* Make sure we can rewrite header without failing */
1486 if (buffer_len > new_header.header_size * new_header.cluster_size) {
1490 /* Prepare new header */
1491 buffer = g_malloc(buffer_len);
1493 qed_header_cpu_to_le(&new_header, &le_header);
1494 memcpy(buffer, &le_header, sizeof(le_header));
1495 buffer_len = sizeof(le_header);
1498 memcpy(buffer + buffer_len, backing_file, backing_file_len);
1499 buffer_len += backing_file_len;
1502 /* Write new header */
1503 ret = bdrv_pwrite_sync(bs->file, 0, buffer, buffer_len);
1506 memcpy(&s->header, &new_header, sizeof(new_header));
1511 static void bdrv_qed_invalidate_cache(BlockDriverState *bs)
1513 BDRVQEDState *s = bs->opaque;
1516 memset(s, 0, sizeof(BDRVQEDState));
1517 bdrv_qed_open(bs, bs->open_flags);
1520 static int bdrv_qed_check(BlockDriverState *bs, BdrvCheckResult *result,
1523 BDRVQEDState *s = bs->opaque;
1525 return qed_check(s, result, !!fix);
1528 static QEMUOptionParameter qed_create_options[] = {
1530 .name = BLOCK_OPT_SIZE,
1532 .help = "Virtual disk size (in bytes)"
1534 .name = BLOCK_OPT_BACKING_FILE,
1536 .help = "File name of a base image"
1538 .name = BLOCK_OPT_BACKING_FMT,
1540 .help = "Image format of the base image"
1542 .name = BLOCK_OPT_CLUSTER_SIZE,
1544 .help = "Cluster size (in bytes)",
1545 .value = { .n = QED_DEFAULT_CLUSTER_SIZE },
1547 .name = BLOCK_OPT_TABLE_SIZE,
1549 .help = "L1/L2 table size (in clusters)"
1551 { /* end of list */ }
1554 static BlockDriver bdrv_qed = {
1555 .format_name = "qed",
1556 .instance_size = sizeof(BDRVQEDState),
1557 .create_options = qed_create_options,
1559 .bdrv_probe = bdrv_qed_probe,
1560 .bdrv_rebind = bdrv_qed_rebind,
1561 .bdrv_open = bdrv_qed_open,
1562 .bdrv_close = bdrv_qed_close,
1563 .bdrv_create = bdrv_qed_create,
1564 .bdrv_co_is_allocated = bdrv_qed_co_is_allocated,
1565 .bdrv_make_empty = bdrv_qed_make_empty,
1566 .bdrv_aio_readv = bdrv_qed_aio_readv,
1567 .bdrv_aio_writev = bdrv_qed_aio_writev,
1568 .bdrv_co_write_zeroes = bdrv_qed_co_write_zeroes,
1569 .bdrv_truncate = bdrv_qed_truncate,
1570 .bdrv_getlength = bdrv_qed_getlength,
1571 .bdrv_get_info = bdrv_qed_get_info,
1572 .bdrv_change_backing_file = bdrv_qed_change_backing_file,
1573 .bdrv_invalidate_cache = bdrv_qed_invalidate_cache,
1574 .bdrv_check = bdrv_qed_check,
1577 static void bdrv_qed_init(void)
1579 bdrv_register(&bdrv_qed);
1582 block_init(bdrv_qed_init);