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 "block/qdict.h"
17 #include "qapi/error.h"
18 #include "qemu/timer.h"
19 #include "qemu/bswap.h"
20 #include "qemu/option.h"
23 #include "sysemu/block-backend.h"
24 #include "qapi/qmp/qdict.h"
25 #include "qapi/qobject-input-visitor.h"
26 #include "qapi/qapi-visit-block-core.h"
28 static QemuOptsList qed_create_opts;
30 static int bdrv_qed_probe(const uint8_t *buf, int buf_size,
33 const QEDHeader *header = (const QEDHeader *)buf;
35 if (buf_size < sizeof(*header)) {
38 if (le32_to_cpu(header->magic) != QED_MAGIC) {
45 * Check whether an image format is raw
47 * @fmt: Backing file format, may be NULL
49 static bool qed_fmt_is_raw(const char *fmt)
51 return fmt && strcmp(fmt, "raw") == 0;
54 static void qed_header_le_to_cpu(const QEDHeader *le, QEDHeader *cpu)
56 cpu->magic = le32_to_cpu(le->magic);
57 cpu->cluster_size = le32_to_cpu(le->cluster_size);
58 cpu->table_size = le32_to_cpu(le->table_size);
59 cpu->header_size = le32_to_cpu(le->header_size);
60 cpu->features = le64_to_cpu(le->features);
61 cpu->compat_features = le64_to_cpu(le->compat_features);
62 cpu->autoclear_features = le64_to_cpu(le->autoclear_features);
63 cpu->l1_table_offset = le64_to_cpu(le->l1_table_offset);
64 cpu->image_size = le64_to_cpu(le->image_size);
65 cpu->backing_filename_offset = le32_to_cpu(le->backing_filename_offset);
66 cpu->backing_filename_size = le32_to_cpu(le->backing_filename_size);
69 static void qed_header_cpu_to_le(const QEDHeader *cpu, QEDHeader *le)
71 le->magic = cpu_to_le32(cpu->magic);
72 le->cluster_size = cpu_to_le32(cpu->cluster_size);
73 le->table_size = cpu_to_le32(cpu->table_size);
74 le->header_size = cpu_to_le32(cpu->header_size);
75 le->features = cpu_to_le64(cpu->features);
76 le->compat_features = cpu_to_le64(cpu->compat_features);
77 le->autoclear_features = cpu_to_le64(cpu->autoclear_features);
78 le->l1_table_offset = cpu_to_le64(cpu->l1_table_offset);
79 le->image_size = cpu_to_le64(cpu->image_size);
80 le->backing_filename_offset = cpu_to_le32(cpu->backing_filename_offset);
81 le->backing_filename_size = cpu_to_le32(cpu->backing_filename_size);
84 int qed_write_header_sync(BDRVQEDState *s)
89 qed_header_cpu_to_le(&s->header, &le);
90 ret = bdrv_pwrite(s->bs->file, 0, &le, sizeof(le));
91 if (ret != sizeof(le)) {
98 * Update header in-place (does not rewrite backing filename or other strings)
100 * This function only updates known header fields in-place and does not affect
101 * extra data after the QED header.
103 * No new allocating reqs can start while this function runs.
105 static int coroutine_fn qed_write_header(BDRVQEDState *s)
107 /* We must write full sectors for O_DIRECT but cannot necessarily generate
108 * the data following the header if an unrecognized compat feature is
109 * active. Therefore, first read the sectors containing the header, update
110 * them, and write back.
113 int nsectors = DIV_ROUND_UP(sizeof(QEDHeader), BDRV_SECTOR_SIZE);
114 size_t len = nsectors * BDRV_SECTOR_SIZE;
120 assert(s->allocating_acb || s->allocating_write_reqs_plugged);
122 buf = qemu_blockalign(s->bs, len);
123 iov = (struct iovec) {
127 qemu_iovec_init_external(&qiov, &iov, 1);
129 ret = bdrv_co_preadv(s->bs->file, 0, qiov.size, &qiov, 0);
135 qed_header_cpu_to_le(&s->header, (QEDHeader *) buf);
137 ret = bdrv_co_pwritev(s->bs->file, 0, qiov.size, &qiov, 0);
148 static uint64_t qed_max_image_size(uint32_t cluster_size, uint32_t table_size)
150 uint64_t table_entries;
153 table_entries = (table_size * cluster_size) / sizeof(uint64_t);
154 l2_size = table_entries * cluster_size;
156 return l2_size * table_entries;
159 static bool qed_is_cluster_size_valid(uint32_t cluster_size)
161 if (cluster_size < QED_MIN_CLUSTER_SIZE ||
162 cluster_size > QED_MAX_CLUSTER_SIZE) {
165 if (cluster_size & (cluster_size - 1)) {
166 return false; /* not power of 2 */
171 static bool qed_is_table_size_valid(uint32_t table_size)
173 if (table_size < QED_MIN_TABLE_SIZE ||
174 table_size > QED_MAX_TABLE_SIZE) {
177 if (table_size & (table_size - 1)) {
178 return false; /* not power of 2 */
183 static bool qed_is_image_size_valid(uint64_t image_size, uint32_t cluster_size,
186 if (image_size % BDRV_SECTOR_SIZE != 0) {
187 return false; /* not multiple of sector size */
189 if (image_size > qed_max_image_size(cluster_size, table_size)) {
190 return false; /* image is too large */
196 * Read a string of known length from the image file
199 * @offset: File offset to start of string, in bytes
200 * @n: String length in bytes
201 * @buf: Destination buffer
202 * @buflen: Destination buffer length in bytes
203 * @ret: 0 on success, -errno on failure
205 * The string is NUL-terminated.
207 static int qed_read_string(BdrvChild *file, uint64_t offset, size_t n,
208 char *buf, size_t buflen)
214 ret = bdrv_pread(file, offset, buf, n);
223 * Allocate new clusters
226 * @n: Number of contiguous clusters to allocate
227 * @ret: Offset of first allocated cluster
229 * This function only produces the offset where the new clusters should be
230 * written. It updates BDRVQEDState but does not make any changes to the image
233 * Called with table_lock held.
235 static uint64_t qed_alloc_clusters(BDRVQEDState *s, unsigned int n)
237 uint64_t offset = s->file_size;
238 s->file_size += n * s->header.cluster_size;
242 QEDTable *qed_alloc_table(BDRVQEDState *s)
244 /* Honor O_DIRECT memory alignment requirements */
245 return qemu_blockalign(s->bs,
246 s->header.cluster_size * s->header.table_size);
250 * Allocate a new zeroed L2 table
252 * Called with table_lock held.
254 static CachedL2Table *qed_new_l2_table(BDRVQEDState *s)
256 CachedL2Table *l2_table = qed_alloc_l2_cache_entry(&s->l2_cache);
258 l2_table->table = qed_alloc_table(s);
259 l2_table->offset = qed_alloc_clusters(s, s->header.table_size);
261 memset(l2_table->table->offsets, 0,
262 s->header.cluster_size * s->header.table_size);
266 static bool qed_plug_allocating_write_reqs(BDRVQEDState *s)
268 qemu_co_mutex_lock(&s->table_lock);
270 /* No reentrancy is allowed. */
271 assert(!s->allocating_write_reqs_plugged);
272 if (s->allocating_acb != NULL) {
273 /* Another allocating write came concurrently. This cannot happen
274 * from bdrv_qed_co_drain_begin, but it can happen when the timer runs.
276 qemu_co_mutex_unlock(&s->table_lock);
280 s->allocating_write_reqs_plugged = true;
281 qemu_co_mutex_unlock(&s->table_lock);
285 static void qed_unplug_allocating_write_reqs(BDRVQEDState *s)
287 qemu_co_mutex_lock(&s->table_lock);
288 assert(s->allocating_write_reqs_plugged);
289 s->allocating_write_reqs_plugged = false;
290 qemu_co_queue_next(&s->allocating_write_reqs);
291 qemu_co_mutex_unlock(&s->table_lock);
294 static void coroutine_fn qed_need_check_timer_entry(void *opaque)
296 BDRVQEDState *s = opaque;
299 trace_qed_need_check_timer_cb(s);
301 if (!qed_plug_allocating_write_reqs(s)) {
305 /* Ensure writes are on disk before clearing flag */
306 ret = bdrv_co_flush(s->bs->file->bs);
308 qed_unplug_allocating_write_reqs(s);
312 s->header.features &= ~QED_F_NEED_CHECK;
313 ret = qed_write_header(s);
316 qed_unplug_allocating_write_reqs(s);
318 ret = bdrv_co_flush(s->bs);
322 static void qed_need_check_timer_cb(void *opaque)
324 Coroutine *co = qemu_coroutine_create(qed_need_check_timer_entry, opaque);
325 qemu_coroutine_enter(co);
328 static void qed_start_need_check_timer(BDRVQEDState *s)
330 trace_qed_start_need_check_timer(s);
332 /* Use QEMU_CLOCK_VIRTUAL so we don't alter the image file while suspended for
335 timer_mod(s->need_check_timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
336 NANOSECONDS_PER_SECOND * QED_NEED_CHECK_TIMEOUT);
339 /* It's okay to call this multiple times or when no timer is started */
340 static void qed_cancel_need_check_timer(BDRVQEDState *s)
342 trace_qed_cancel_need_check_timer(s);
343 timer_del(s->need_check_timer);
346 static void bdrv_qed_detach_aio_context(BlockDriverState *bs)
348 BDRVQEDState *s = bs->opaque;
350 qed_cancel_need_check_timer(s);
351 timer_free(s->need_check_timer);
354 static void bdrv_qed_attach_aio_context(BlockDriverState *bs,
355 AioContext *new_context)
357 BDRVQEDState *s = bs->opaque;
359 s->need_check_timer = aio_timer_new(new_context,
360 QEMU_CLOCK_VIRTUAL, SCALE_NS,
361 qed_need_check_timer_cb, s);
362 if (s->header.features & QED_F_NEED_CHECK) {
363 qed_start_need_check_timer(s);
367 static void coroutine_fn bdrv_qed_co_drain_begin(BlockDriverState *bs)
369 BDRVQEDState *s = bs->opaque;
371 /* Fire the timer immediately in order to start doing I/O as soon as the
374 if (s->need_check_timer && timer_pending(s->need_check_timer)) {
375 qed_cancel_need_check_timer(s);
376 qed_need_check_timer_entry(s);
380 static void bdrv_qed_init_state(BlockDriverState *bs)
382 BDRVQEDState *s = bs->opaque;
384 memset(s, 0, sizeof(BDRVQEDState));
386 qemu_co_mutex_init(&s->table_lock);
387 qemu_co_queue_init(&s->allocating_write_reqs);
390 /* Called with table_lock held. */
391 static int coroutine_fn bdrv_qed_do_open(BlockDriverState *bs, QDict *options,
392 int flags, Error **errp)
394 BDRVQEDState *s = bs->opaque;
399 ret = bdrv_pread(bs->file, 0, &le_header, sizeof(le_header));
403 qed_header_le_to_cpu(&le_header, &s->header);
405 if (s->header.magic != QED_MAGIC) {
406 error_setg(errp, "Image not in QED format");
409 if (s->header.features & ~QED_FEATURE_MASK) {
410 /* image uses unsupported feature bits */
411 error_setg(errp, "Unsupported QED features: %" PRIx64,
412 s->header.features & ~QED_FEATURE_MASK);
415 if (!qed_is_cluster_size_valid(s->header.cluster_size)) {
419 /* Round down file size to the last cluster */
420 file_size = bdrv_getlength(bs->file->bs);
424 s->file_size = qed_start_of_cluster(s, file_size);
426 if (!qed_is_table_size_valid(s->header.table_size)) {
429 if (!qed_is_image_size_valid(s->header.image_size,
430 s->header.cluster_size,
431 s->header.table_size)) {
434 if (!qed_check_table_offset(s, s->header.l1_table_offset)) {
438 s->table_nelems = (s->header.cluster_size * s->header.table_size) /
440 s->l2_shift = ctz32(s->header.cluster_size);
441 s->l2_mask = s->table_nelems - 1;
442 s->l1_shift = s->l2_shift + ctz32(s->table_nelems);
444 /* Header size calculation must not overflow uint32_t */
445 if (s->header.header_size > UINT32_MAX / s->header.cluster_size) {
449 if ((s->header.features & QED_F_BACKING_FILE)) {
450 if ((uint64_t)s->header.backing_filename_offset +
451 s->header.backing_filename_size >
452 s->header.cluster_size * s->header.header_size) {
456 ret = qed_read_string(bs->file, s->header.backing_filename_offset,
457 s->header.backing_filename_size,
458 bs->auto_backing_file,
459 sizeof(bs->auto_backing_file));
463 pstrcpy(bs->backing_file, sizeof(bs->backing_file),
464 bs->auto_backing_file);
466 if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) {
467 pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw");
471 /* Reset unknown autoclear feature bits. This is a backwards
472 * compatibility mechanism that allows images to be opened by older
473 * programs, which "knock out" unknown feature bits. When an image is
474 * opened by a newer program again it can detect that the autoclear
475 * feature is no longer valid.
477 if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 &&
478 !bdrv_is_read_only(bs->file->bs) && !(flags & BDRV_O_INACTIVE)) {
479 s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK;
481 ret = qed_write_header_sync(s);
486 /* From here on only known autoclear feature bits are valid */
487 bdrv_flush(bs->file->bs);
490 s->l1_table = qed_alloc_table(s);
491 qed_init_l2_cache(&s->l2_cache);
493 ret = qed_read_l1_table_sync(s);
498 /* If image was not closed cleanly, check consistency */
499 if (!(flags & BDRV_O_CHECK) && (s->header.features & QED_F_NEED_CHECK)) {
500 /* Read-only images cannot be fixed. There is no risk of corruption
501 * since write operations are not possible. Therefore, allow
502 * potentially inconsistent images to be opened read-only. This can
503 * aid data recovery from an otherwise inconsistent image.
505 if (!bdrv_is_read_only(bs->file->bs) &&
506 !(flags & BDRV_O_INACTIVE)) {
507 BdrvCheckResult result = {0};
509 ret = qed_check(s, &result, true);
516 bdrv_qed_attach_aio_context(bs, bdrv_get_aio_context(bs));
520 qed_free_l2_cache(&s->l2_cache);
521 qemu_vfree(s->l1_table);
526 typedef struct QEDOpenCo {
527 BlockDriverState *bs;
534 static void coroutine_fn bdrv_qed_open_entry(void *opaque)
536 QEDOpenCo *qoc = opaque;
537 BDRVQEDState *s = qoc->bs->opaque;
539 qemu_co_mutex_lock(&s->table_lock);
540 qoc->ret = bdrv_qed_do_open(qoc->bs, qoc->options, qoc->flags, qoc->errp);
541 qemu_co_mutex_unlock(&s->table_lock);
544 static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags,
555 bs->file = bdrv_open_child(NULL, options, "file", bs, &child_file,
561 bdrv_qed_init_state(bs);
562 if (qemu_in_coroutine()) {
563 bdrv_qed_open_entry(&qoc);
565 assert(qemu_get_current_aio_context() == qemu_get_aio_context());
566 qemu_coroutine_enter(qemu_coroutine_create(bdrv_qed_open_entry, &qoc));
567 BDRV_POLL_WHILE(bs, qoc.ret == -EINPROGRESS);
569 BDRV_POLL_WHILE(bs, qoc.ret == -EINPROGRESS);
573 static void bdrv_qed_refresh_limits(BlockDriverState *bs, Error **errp)
575 BDRVQEDState *s = bs->opaque;
577 bs->bl.pwrite_zeroes_alignment = s->header.cluster_size;
580 /* We have nothing to do for QED reopen, stubs just return
582 static int bdrv_qed_reopen_prepare(BDRVReopenState *state,
583 BlockReopenQueue *queue, Error **errp)
588 static void bdrv_qed_close(BlockDriverState *bs)
590 BDRVQEDState *s = bs->opaque;
592 bdrv_qed_detach_aio_context(bs);
594 /* Ensure writes reach stable storage */
595 bdrv_flush(bs->file->bs);
597 /* Clean shutdown, no check required on next open */
598 if (s->header.features & QED_F_NEED_CHECK) {
599 s->header.features &= ~QED_F_NEED_CHECK;
600 qed_write_header_sync(s);
603 qed_free_l2_cache(&s->l2_cache);
604 qemu_vfree(s->l1_table);
607 static int coroutine_fn bdrv_qed_co_create(BlockdevCreateOptions *opts,
610 BlockdevCreateOptionsQed *qed_opts;
611 BlockBackend *blk = NULL;
612 BlockDriverState *bs = NULL;
616 uint8_t *l1_table = NULL;
620 assert(opts->driver == BLOCKDEV_DRIVER_QED);
621 qed_opts = &opts->u.qed;
623 /* Validate options and set default values */
624 if (!qed_opts->has_cluster_size) {
625 qed_opts->cluster_size = QED_DEFAULT_CLUSTER_SIZE;
627 if (!qed_opts->has_table_size) {
628 qed_opts->table_size = QED_DEFAULT_TABLE_SIZE;
631 if (!qed_is_cluster_size_valid(qed_opts->cluster_size)) {
632 error_setg(errp, "QED cluster size must be within range [%u, %u] "
634 QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE);
637 if (!qed_is_table_size_valid(qed_opts->table_size)) {
638 error_setg(errp, "QED table size must be within range [%u, %u] "
640 QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE);
643 if (!qed_is_image_size_valid(qed_opts->size, qed_opts->cluster_size,
644 qed_opts->table_size))
646 error_setg(errp, "QED image size must be a non-zero multiple of "
647 "cluster size and less than %" PRIu64 " bytes",
648 qed_max_image_size(qed_opts->cluster_size,
649 qed_opts->table_size));
653 /* Create BlockBackend to write to the image */
654 bs = bdrv_open_blockdev_ref(qed_opts->file, errp);
659 blk = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
660 ret = blk_insert_bs(blk, bs, errp);
664 blk_set_allow_write_beyond_eof(blk, true);
666 /* Prepare image format */
667 header = (QEDHeader) {
669 .cluster_size = qed_opts->cluster_size,
670 .table_size = qed_opts->table_size,
673 .compat_features = 0,
674 .l1_table_offset = qed_opts->cluster_size,
675 .image_size = qed_opts->size,
678 l1_size = header.cluster_size * header.table_size;
680 /* File must start empty and grow, check truncate is supported */
681 ret = blk_truncate(blk, 0, PREALLOC_MODE_OFF, errp);
686 if (qed_opts->has_backing_file) {
687 header.features |= QED_F_BACKING_FILE;
688 header.backing_filename_offset = sizeof(le_header);
689 header.backing_filename_size = strlen(qed_opts->backing_file);
691 if (qed_opts->has_backing_fmt) {
692 const char *backing_fmt = BlockdevDriver_str(qed_opts->backing_fmt);
693 if (qed_fmt_is_raw(backing_fmt)) {
694 header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
699 qed_header_cpu_to_le(&header, &le_header);
700 ret = blk_pwrite(blk, 0, &le_header, sizeof(le_header), 0);
704 ret = blk_pwrite(blk, sizeof(le_header), qed_opts->backing_file,
705 header.backing_filename_size, 0);
710 l1_table = g_malloc0(l1_size);
711 ret = blk_pwrite(blk, header.l1_table_offset, l1_table, l1_size, 0);
716 ret = 0; /* success */
724 static int coroutine_fn bdrv_qed_co_create_opts(const char *filename,
728 BlockdevCreateOptions *create_options = NULL;
731 BlockDriverState *bs = NULL;
732 Error *local_err = NULL;
735 static const QDictRenames opt_renames[] = {
736 { BLOCK_OPT_BACKING_FILE, "backing-file" },
737 { BLOCK_OPT_BACKING_FMT, "backing-fmt" },
738 { BLOCK_OPT_CLUSTER_SIZE, "cluster-size" },
739 { BLOCK_OPT_TABLE_SIZE, "table-size" },
743 /* Parse options and convert legacy syntax */
744 qdict = qemu_opts_to_qdict_filtered(opts, NULL, &qed_create_opts, true);
746 if (!qdict_rename_keys(qdict, opt_renames, errp)) {
751 /* Create and open the file (protocol layer) */
752 ret = bdrv_create_file(filename, opts, &local_err);
754 error_propagate(errp, local_err);
758 bs = bdrv_open(filename, NULL, NULL,
759 BDRV_O_RDWR | BDRV_O_RESIZE | BDRV_O_PROTOCOL, errp);
765 /* Now get the QAPI type BlockdevCreateOptions */
766 qdict_put_str(qdict, "driver", "qed");
767 qdict_put_str(qdict, "file", bs->node_name);
769 v = qobject_input_visitor_new_flat_confused(qdict, errp);
775 visit_type_BlockdevCreateOptions(v, NULL, &create_options, &local_err);
779 error_propagate(errp, local_err);
784 /* Silently round up size */
785 assert(create_options->driver == BLOCKDEV_DRIVER_QED);
786 create_options->u.qed.size =
787 ROUND_UP(create_options->u.qed.size, BDRV_SECTOR_SIZE);
789 /* Create the qed image (format layer) */
790 ret = bdrv_qed_co_create(create_options, errp);
793 qobject_unref(qdict);
795 qapi_free_BlockdevCreateOptions(create_options);
799 static int coroutine_fn bdrv_qed_co_block_status(BlockDriverState *bs,
801 int64_t pos, int64_t bytes,
802 int64_t *pnum, int64_t *map,
803 BlockDriverState **file)
805 BDRVQEDState *s = bs->opaque;
806 size_t len = MIN(bytes, SIZE_MAX);
808 QEDRequest request = { .l2_table = NULL };
812 qemu_co_mutex_lock(&s->table_lock);
813 ret = qed_find_cluster(s, &request, pos, &len, &offset);
817 case QED_CLUSTER_FOUND:
818 *map = offset | qed_offset_into_cluster(s, pos);
819 status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
820 *file = bs->file->bs;
822 case QED_CLUSTER_ZERO:
823 status = BDRV_BLOCK_ZERO;
835 qed_unref_l2_cache_entry(request.l2_table);
836 qemu_co_mutex_unlock(&s->table_lock);
841 static BDRVQEDState *acb_to_s(QEDAIOCB *acb)
843 return acb->bs->opaque;
847 * Read from the backing file or zero-fill if no backing file
850 * @pos: Byte position in device
851 * @qiov: Destination I/O vector
852 * @backing_qiov: Possibly shortened copy of qiov, to be allocated here
853 * @cb: Completion function
854 * @opaque: User data for completion function
856 * This function reads qiov->size bytes starting at pos from the backing file.
857 * If there is no backing file then zeroes are read.
859 static int coroutine_fn qed_read_backing_file(BDRVQEDState *s, uint64_t pos,
861 QEMUIOVector **backing_qiov)
863 uint64_t backing_length = 0;
867 /* If there is a backing file, get its length. Treat the absence of a
868 * backing file like a zero length backing file.
870 if (s->bs->backing) {
871 int64_t l = bdrv_getlength(s->bs->backing->bs);
878 /* Zero all sectors if reading beyond the end of the backing file */
879 if (pos >= backing_length ||
880 pos + qiov->size > backing_length) {
881 qemu_iovec_memset(qiov, 0, 0, qiov->size);
884 /* Complete now if there are no backing file sectors to read */
885 if (pos >= backing_length) {
889 /* If the read straddles the end of the backing file, shorten it */
890 size = MIN((uint64_t)backing_length - pos, qiov->size);
892 assert(*backing_qiov == NULL);
893 *backing_qiov = g_new(QEMUIOVector, 1);
894 qemu_iovec_init(*backing_qiov, qiov->niov);
895 qemu_iovec_concat(*backing_qiov, qiov, 0, size);
897 BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO);
898 ret = bdrv_co_preadv(s->bs->backing, pos, size, *backing_qiov, 0);
906 * Copy data from backing file into the image
909 * @pos: Byte position in device
910 * @len: Number of bytes
911 * @offset: Byte offset in image file
913 static int coroutine_fn qed_copy_from_backing_file(BDRVQEDState *s,
914 uint64_t pos, uint64_t len,
918 QEMUIOVector *backing_qiov = NULL;
922 /* Skip copy entirely if there is no work to do */
927 iov = (struct iovec) {
928 .iov_base = qemu_blockalign(s->bs, len),
931 qemu_iovec_init_external(&qiov, &iov, 1);
933 ret = qed_read_backing_file(s, pos, &qiov, &backing_qiov);
936 qemu_iovec_destroy(backing_qiov);
937 g_free(backing_qiov);
945 BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE);
946 ret = bdrv_co_pwritev(s->bs->file, offset, qiov.size, &qiov, 0);
952 qemu_vfree(iov.iov_base);
957 * Link one or more contiguous clusters into a table
961 * @index: First cluster index
962 * @n: Number of contiguous clusters
963 * @cluster: First cluster offset
965 * The cluster offset may be an allocated byte offset in the image file, the
966 * zero cluster marker, or the unallocated cluster marker.
968 * Called with table_lock held.
970 static void coroutine_fn qed_update_l2_table(BDRVQEDState *s, QEDTable *table,
971 int index, unsigned int n,
975 for (i = index; i < index + n; i++) {
976 table->offsets[i] = cluster;
977 if (!qed_offset_is_unalloc_cluster(cluster) &&
978 !qed_offset_is_zero_cluster(cluster)) {
979 cluster += s->header.cluster_size;
984 /* Called with table_lock held. */
985 static void coroutine_fn qed_aio_complete(QEDAIOCB *acb)
987 BDRVQEDState *s = acb_to_s(acb);
990 qemu_iovec_destroy(&acb->cur_qiov);
991 qed_unref_l2_cache_entry(acb->request.l2_table);
993 /* Free the buffer we may have allocated for zero writes */
994 if (acb->flags & QED_AIOCB_ZERO) {
995 qemu_vfree(acb->qiov->iov[0].iov_base);
996 acb->qiov->iov[0].iov_base = NULL;
999 /* Start next allocating write request waiting behind this one. Note that
1000 * requests enqueue themselves when they first hit an unallocated cluster
1001 * but they wait until the entire request is finished before waking up the
1002 * next request in the queue. This ensures that we don't cycle through
1003 * requests multiple times but rather finish one at a time completely.
1005 if (acb == s->allocating_acb) {
1006 s->allocating_acb = NULL;
1007 if (!qemu_co_queue_empty(&s->allocating_write_reqs)) {
1008 qemu_co_queue_next(&s->allocating_write_reqs);
1009 } else if (s->header.features & QED_F_NEED_CHECK) {
1010 qed_start_need_check_timer(s);
1016 * Update L1 table with new L2 table offset and write it out
1018 * Called with table_lock held.
1020 static int coroutine_fn qed_aio_write_l1_update(QEDAIOCB *acb)
1022 BDRVQEDState *s = acb_to_s(acb);
1023 CachedL2Table *l2_table = acb->request.l2_table;
1024 uint64_t l2_offset = l2_table->offset;
1027 index = qed_l1_index(s, acb->cur_pos);
1028 s->l1_table->offsets[index] = l2_table->offset;
1030 ret = qed_write_l1_table(s, index, 1);
1032 /* Commit the current L2 table to the cache */
1033 qed_commit_l2_cache_entry(&s->l2_cache, l2_table);
1035 /* This is guaranteed to succeed because we just committed the entry to the
1038 acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset);
1039 assert(acb->request.l2_table != NULL);
1046 * Update L2 table with new cluster offsets and write them out
1048 * Called with table_lock held.
1050 static int coroutine_fn qed_aio_write_l2_update(QEDAIOCB *acb, uint64_t offset)
1052 BDRVQEDState *s = acb_to_s(acb);
1053 bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1;
1057 qed_unref_l2_cache_entry(acb->request.l2_table);
1058 acb->request.l2_table = qed_new_l2_table(s);
1061 index = qed_l2_index(s, acb->cur_pos);
1062 qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters,
1066 /* Write out the whole new L2 table */
1067 ret = qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true);
1071 return qed_aio_write_l1_update(acb);
1073 /* Write out only the updated part of the L2 table */
1074 ret = qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters,
1084 * Write data to the image file
1086 * Called with table_lock *not* held.
1088 static int coroutine_fn qed_aio_write_main(QEDAIOCB *acb)
1090 BDRVQEDState *s = acb_to_s(acb);
1091 uint64_t offset = acb->cur_cluster +
1092 qed_offset_into_cluster(s, acb->cur_pos);
1094 trace_qed_aio_write_main(s, acb, 0, offset, acb->cur_qiov.size);
1096 BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO);
1097 return bdrv_co_pwritev(s->bs->file, offset, acb->cur_qiov.size,
1102 * Populate untouched regions of new data cluster
1104 * Called with table_lock held.
1106 static int coroutine_fn qed_aio_write_cow(QEDAIOCB *acb)
1108 BDRVQEDState *s = acb_to_s(acb);
1109 uint64_t start, len, offset;
1112 qemu_co_mutex_unlock(&s->table_lock);
1114 /* Populate front untouched region of new data cluster */
1115 start = qed_start_of_cluster(s, acb->cur_pos);
1116 len = qed_offset_into_cluster(s, acb->cur_pos);
1118 trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster);
1119 ret = qed_copy_from_backing_file(s, start, len, acb->cur_cluster);
1124 /* Populate back untouched region of new data cluster */
1125 start = acb->cur_pos + acb->cur_qiov.size;
1126 len = qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start;
1127 offset = acb->cur_cluster +
1128 qed_offset_into_cluster(s, acb->cur_pos) +
1131 trace_qed_aio_write_postfill(s, acb, start, len, offset);
1132 ret = qed_copy_from_backing_file(s, start, len, offset);
1137 ret = qed_aio_write_main(acb);
1142 if (s->bs->backing) {
1144 * Flush new data clusters before updating the L2 table
1146 * This flush is necessary when a backing file is in use. A crash
1147 * during an allocating write could result in empty clusters in the
1148 * image. If the write only touched a subregion of the cluster,
1149 * then backing image sectors have been lost in the untouched
1150 * region. The solution is to flush after writing a new data
1151 * cluster and before updating the L2 table.
1153 ret = bdrv_co_flush(s->bs->file->bs);
1157 qemu_co_mutex_lock(&s->table_lock);
1162 * Check if the QED_F_NEED_CHECK bit should be set during allocating write
1164 static bool qed_should_set_need_check(BDRVQEDState *s)
1166 /* The flush before L2 update path ensures consistency */
1167 if (s->bs->backing) {
1171 return !(s->header.features & QED_F_NEED_CHECK);
1175 * Write new data cluster
1177 * @acb: Write request
1178 * @len: Length in bytes
1180 * This path is taken when writing to previously unallocated clusters.
1182 * Called with table_lock held.
1184 static int coroutine_fn qed_aio_write_alloc(QEDAIOCB *acb, size_t len)
1186 BDRVQEDState *s = acb_to_s(acb);
1189 /* Cancel timer when the first allocating request comes in */
1190 if (s->allocating_acb == NULL) {
1191 qed_cancel_need_check_timer(s);
1194 /* Freeze this request if another allocating write is in progress */
1195 if (s->allocating_acb != acb || s->allocating_write_reqs_plugged) {
1196 if (s->allocating_acb != NULL) {
1197 qemu_co_queue_wait(&s->allocating_write_reqs, &s->table_lock);
1198 assert(s->allocating_acb == NULL);
1200 s->allocating_acb = acb;
1201 return -EAGAIN; /* start over with looking up table entries */
1204 acb->cur_nclusters = qed_bytes_to_clusters(s,
1205 qed_offset_into_cluster(s, acb->cur_pos) + len);
1206 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1208 if (acb->flags & QED_AIOCB_ZERO) {
1209 /* Skip ahead if the clusters are already zero */
1210 if (acb->find_cluster_ret == QED_CLUSTER_ZERO) {
1213 acb->cur_cluster = 1;
1215 acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters);
1218 if (qed_should_set_need_check(s)) {
1219 s->header.features |= QED_F_NEED_CHECK;
1220 ret = qed_write_header(s);
1226 if (!(acb->flags & QED_AIOCB_ZERO)) {
1227 ret = qed_aio_write_cow(acb);
1233 return qed_aio_write_l2_update(acb, acb->cur_cluster);
1237 * Write data cluster in place
1239 * @acb: Write request
1240 * @offset: Cluster offset in bytes
1241 * @len: Length in bytes
1243 * This path is taken when writing to already allocated clusters.
1245 * Called with table_lock held.
1247 static int coroutine_fn qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset,
1250 BDRVQEDState *s = acb_to_s(acb);
1253 qemu_co_mutex_unlock(&s->table_lock);
1255 /* Allocate buffer for zero writes */
1256 if (acb->flags & QED_AIOCB_ZERO) {
1257 struct iovec *iov = acb->qiov->iov;
1259 if (!iov->iov_base) {
1260 iov->iov_base = qemu_try_blockalign(acb->bs, iov->iov_len);
1261 if (iov->iov_base == NULL) {
1265 memset(iov->iov_base, 0, iov->iov_len);
1269 /* Calculate the I/O vector */
1270 acb->cur_cluster = offset;
1271 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1273 /* Do the actual write. */
1274 r = qed_aio_write_main(acb);
1276 qemu_co_mutex_lock(&s->table_lock);
1281 * Write data cluster
1283 * @opaque: Write request
1284 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1285 * @offset: Cluster offset in bytes
1286 * @len: Length in bytes
1288 * Called with table_lock held.
1290 static int coroutine_fn qed_aio_write_data(void *opaque, int ret,
1291 uint64_t offset, size_t len)
1293 QEDAIOCB *acb = opaque;
1295 trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len);
1297 acb->find_cluster_ret = ret;
1300 case QED_CLUSTER_FOUND:
1301 return qed_aio_write_inplace(acb, offset, len);
1303 case QED_CLUSTER_L2:
1304 case QED_CLUSTER_L1:
1305 case QED_CLUSTER_ZERO:
1306 return qed_aio_write_alloc(acb, len);
1309 g_assert_not_reached();
1316 * @opaque: Read request
1317 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1318 * @offset: Cluster offset in bytes
1319 * @len: Length in bytes
1321 * Called with table_lock held.
1323 static int coroutine_fn qed_aio_read_data(void *opaque, int ret,
1324 uint64_t offset, size_t len)
1326 QEDAIOCB *acb = opaque;
1327 BDRVQEDState *s = acb_to_s(acb);
1328 BlockDriverState *bs = acb->bs;
1331 qemu_co_mutex_unlock(&s->table_lock);
1333 /* Adjust offset into cluster */
1334 offset += qed_offset_into_cluster(s, acb->cur_pos);
1336 trace_qed_aio_read_data(s, acb, ret, offset, len);
1338 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1340 /* Handle zero cluster and backing file reads, otherwise read
1341 * data cluster directly.
1343 if (ret == QED_CLUSTER_ZERO) {
1344 qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size);
1346 } else if (ret != QED_CLUSTER_FOUND) {
1347 r = qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov,
1348 &acb->backing_qiov);
1350 BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
1351 r = bdrv_co_preadv(bs->file, offset, acb->cur_qiov.size,
1355 qemu_co_mutex_lock(&s->table_lock);
1360 * Begin next I/O or complete the request
1362 static int coroutine_fn qed_aio_next_io(QEDAIOCB *acb)
1364 BDRVQEDState *s = acb_to_s(acb);
1369 qemu_co_mutex_lock(&s->table_lock);
1371 trace_qed_aio_next_io(s, acb, 0, acb->cur_pos + acb->cur_qiov.size);
1373 if (acb->backing_qiov) {
1374 qemu_iovec_destroy(acb->backing_qiov);
1375 g_free(acb->backing_qiov);
1376 acb->backing_qiov = NULL;
1379 acb->qiov_offset += acb->cur_qiov.size;
1380 acb->cur_pos += acb->cur_qiov.size;
1381 qemu_iovec_reset(&acb->cur_qiov);
1383 /* Complete request */
1384 if (acb->cur_pos >= acb->end_pos) {
1389 /* Find next cluster and start I/O */
1390 len = acb->end_pos - acb->cur_pos;
1391 ret = qed_find_cluster(s, &acb->request, acb->cur_pos, &len, &offset);
1396 if (acb->flags & QED_AIOCB_WRITE) {
1397 ret = qed_aio_write_data(acb, ret, offset, len);
1399 ret = qed_aio_read_data(acb, ret, offset, len);
1402 if (ret < 0 && ret != -EAGAIN) {
1407 trace_qed_aio_complete(s, acb, ret);
1408 qed_aio_complete(acb);
1409 qemu_co_mutex_unlock(&s->table_lock);
1413 static int coroutine_fn qed_co_request(BlockDriverState *bs, int64_t sector_num,
1414 QEMUIOVector *qiov, int nb_sectors,
1419 .cur_pos = (uint64_t) sector_num * BDRV_SECTOR_SIZE,
1420 .end_pos = (sector_num + nb_sectors) * BDRV_SECTOR_SIZE,
1424 qemu_iovec_init(&acb.cur_qiov, qiov->niov);
1426 trace_qed_aio_setup(bs->opaque, &acb, sector_num, nb_sectors, NULL, flags);
1429 return qed_aio_next_io(&acb);
1432 static int coroutine_fn bdrv_qed_co_readv(BlockDriverState *bs,
1433 int64_t sector_num, int nb_sectors,
1436 return qed_co_request(bs, sector_num, qiov, nb_sectors, 0);
1439 static int coroutine_fn bdrv_qed_co_writev(BlockDriverState *bs,
1440 int64_t sector_num, int nb_sectors,
1441 QEMUIOVector *qiov, int flags)
1444 return qed_co_request(bs, sector_num, qiov, nb_sectors, QED_AIOCB_WRITE);
1447 static int coroutine_fn bdrv_qed_co_pwrite_zeroes(BlockDriverState *bs,
1450 BdrvRequestFlags flags)
1452 BDRVQEDState *s = bs->opaque;
1456 /* Fall back if the request is not aligned */
1457 if (qed_offset_into_cluster(s, offset) ||
1458 qed_offset_into_cluster(s, bytes)) {
1462 /* Zero writes start without an I/O buffer. If a buffer becomes necessary
1463 * then it will be allocated during request processing.
1465 iov.iov_base = NULL;
1466 iov.iov_len = bytes;
1468 qemu_iovec_init_external(&qiov, &iov, 1);
1469 return qed_co_request(bs, offset >> BDRV_SECTOR_BITS, &qiov,
1470 bytes >> BDRV_SECTOR_BITS,
1471 QED_AIOCB_WRITE | QED_AIOCB_ZERO);
1474 static int coroutine_fn bdrv_qed_co_truncate(BlockDriverState *bs,
1476 PreallocMode prealloc,
1479 BDRVQEDState *s = bs->opaque;
1480 uint64_t old_image_size;
1483 if (prealloc != PREALLOC_MODE_OFF) {
1484 error_setg(errp, "Unsupported preallocation mode '%s'",
1485 PreallocMode_str(prealloc));
1489 if (!qed_is_image_size_valid(offset, s->header.cluster_size,
1490 s->header.table_size)) {
1491 error_setg(errp, "Invalid image size specified");
1495 if ((uint64_t)offset < s->header.image_size) {
1496 error_setg(errp, "Shrinking images is currently not supported");
1500 old_image_size = s->header.image_size;
1501 s->header.image_size = offset;
1502 ret = qed_write_header_sync(s);
1504 s->header.image_size = old_image_size;
1505 error_setg_errno(errp, -ret, "Failed to update the image size");
1510 static int64_t bdrv_qed_getlength(BlockDriverState *bs)
1512 BDRVQEDState *s = bs->opaque;
1513 return s->header.image_size;
1516 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1518 BDRVQEDState *s = bs->opaque;
1520 memset(bdi, 0, sizeof(*bdi));
1521 bdi->cluster_size = s->header.cluster_size;
1522 bdi->is_dirty = s->header.features & QED_F_NEED_CHECK;
1523 bdi->unallocated_blocks_are_zero = true;
1527 static int bdrv_qed_change_backing_file(BlockDriverState *bs,
1528 const char *backing_file,
1529 const char *backing_fmt)
1531 BDRVQEDState *s = bs->opaque;
1532 QEDHeader new_header, le_header;
1534 size_t buffer_len, backing_file_len;
1537 /* Refuse to set backing filename if unknown compat feature bits are
1538 * active. If the image uses an unknown compat feature then we may not
1539 * know the layout of data following the header structure and cannot safely
1542 if (backing_file && (s->header.compat_features &
1543 ~QED_COMPAT_FEATURE_MASK)) {
1547 memcpy(&new_header, &s->header, sizeof(new_header));
1549 new_header.features &= ~(QED_F_BACKING_FILE |
1550 QED_F_BACKING_FORMAT_NO_PROBE);
1552 /* Adjust feature flags */
1554 new_header.features |= QED_F_BACKING_FILE;
1556 if (qed_fmt_is_raw(backing_fmt)) {
1557 new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
1561 /* Calculate new header size */
1562 backing_file_len = 0;
1565 backing_file_len = strlen(backing_file);
1568 buffer_len = sizeof(new_header);
1569 new_header.backing_filename_offset = buffer_len;
1570 new_header.backing_filename_size = backing_file_len;
1571 buffer_len += backing_file_len;
1573 /* Make sure we can rewrite header without failing */
1574 if (buffer_len > new_header.header_size * new_header.cluster_size) {
1578 /* Prepare new header */
1579 buffer = g_malloc(buffer_len);
1581 qed_header_cpu_to_le(&new_header, &le_header);
1582 memcpy(buffer, &le_header, sizeof(le_header));
1583 buffer_len = sizeof(le_header);
1586 memcpy(buffer + buffer_len, backing_file, backing_file_len);
1587 buffer_len += backing_file_len;
1590 /* Write new header */
1591 ret = bdrv_pwrite_sync(bs->file, 0, buffer, buffer_len);
1594 memcpy(&s->header, &new_header, sizeof(new_header));
1599 static void coroutine_fn bdrv_qed_co_invalidate_cache(BlockDriverState *bs,
1602 BDRVQEDState *s = bs->opaque;
1603 Error *local_err = NULL;
1608 bdrv_qed_init_state(bs);
1609 qemu_co_mutex_lock(&s->table_lock);
1610 ret = bdrv_qed_do_open(bs, NULL, bs->open_flags, &local_err);
1611 qemu_co_mutex_unlock(&s->table_lock);
1613 error_propagate_prepend(errp, local_err,
1614 "Could not reopen qed layer: ");
1616 } else if (ret < 0) {
1617 error_setg_errno(errp, -ret, "Could not reopen qed layer");
1622 static int bdrv_qed_co_check(BlockDriverState *bs, BdrvCheckResult *result,
1625 BDRVQEDState *s = bs->opaque;
1628 qemu_co_mutex_lock(&s->table_lock);
1629 ret = qed_check(s, result, !!fix);
1630 qemu_co_mutex_unlock(&s->table_lock);
1635 static QemuOptsList qed_create_opts = {
1636 .name = "qed-create-opts",
1637 .head = QTAILQ_HEAD_INITIALIZER(qed_create_opts.head),
1640 .name = BLOCK_OPT_SIZE,
1641 .type = QEMU_OPT_SIZE,
1642 .help = "Virtual disk size"
1645 .name = BLOCK_OPT_BACKING_FILE,
1646 .type = QEMU_OPT_STRING,
1647 .help = "File name of a base image"
1650 .name = BLOCK_OPT_BACKING_FMT,
1651 .type = QEMU_OPT_STRING,
1652 .help = "Image format of the base image"
1655 .name = BLOCK_OPT_CLUSTER_SIZE,
1656 .type = QEMU_OPT_SIZE,
1657 .help = "Cluster size (in bytes)",
1658 .def_value_str = stringify(QED_DEFAULT_CLUSTER_SIZE)
1661 .name = BLOCK_OPT_TABLE_SIZE,
1662 .type = QEMU_OPT_SIZE,
1663 .help = "L1/L2 table size (in clusters)"
1665 { /* end of list */ }
1669 static BlockDriver bdrv_qed = {
1670 .format_name = "qed",
1671 .instance_size = sizeof(BDRVQEDState),
1672 .create_opts = &qed_create_opts,
1673 .supports_backing = true,
1675 .bdrv_probe = bdrv_qed_probe,
1676 .bdrv_open = bdrv_qed_open,
1677 .bdrv_close = bdrv_qed_close,
1678 .bdrv_reopen_prepare = bdrv_qed_reopen_prepare,
1679 .bdrv_child_perm = bdrv_format_default_perms,
1680 .bdrv_co_create = bdrv_qed_co_create,
1681 .bdrv_co_create_opts = bdrv_qed_co_create_opts,
1682 .bdrv_has_zero_init = bdrv_has_zero_init_1,
1683 .bdrv_co_block_status = bdrv_qed_co_block_status,
1684 .bdrv_co_readv = bdrv_qed_co_readv,
1685 .bdrv_co_writev = bdrv_qed_co_writev,
1686 .bdrv_co_pwrite_zeroes = bdrv_qed_co_pwrite_zeroes,
1687 .bdrv_co_truncate = bdrv_qed_co_truncate,
1688 .bdrv_getlength = bdrv_qed_getlength,
1689 .bdrv_get_info = bdrv_qed_get_info,
1690 .bdrv_refresh_limits = bdrv_qed_refresh_limits,
1691 .bdrv_change_backing_file = bdrv_qed_change_backing_file,
1692 .bdrv_co_invalidate_cache = bdrv_qed_co_invalidate_cache,
1693 .bdrv_co_check = bdrv_qed_co_check,
1694 .bdrv_detach_aio_context = bdrv_qed_detach_aio_context,
1695 .bdrv_attach_aio_context = bdrv_qed_attach_aio_context,
1696 .bdrv_co_drain_begin = bdrv_qed_co_drain_begin,
1699 static void bdrv_qed_init(void)
1701 bdrv_register(&bdrv_qed);
1704 block_init(bdrv_qed_init);