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/osdep.h"
16 #include "block/qdict.h"
17 #include "qapi/error.h"
18 #include "qemu/timer.h"
19 #include "qemu/bswap.h"
20 #include "qemu/main-loop.h"
21 #include "qemu/module.h"
22 #include "qemu/option.h"
25 #include "sysemu/block-backend.h"
26 #include "qapi/qmp/qdict.h"
27 #include "qapi/qobject-input-visitor.h"
28 #include "qapi/qapi-visit-block-core.h"
30 static QemuOptsList qed_create_opts;
32 static int bdrv_qed_probe(const uint8_t *buf, int buf_size,
35 const QEDHeader *header = (const QEDHeader *)buf;
37 if (buf_size < sizeof(*header)) {
40 if (le32_to_cpu(header->magic) != QED_MAGIC) {
47 * Check whether an image format is raw
49 * @fmt: Backing file format, may be NULL
51 static bool qed_fmt_is_raw(const char *fmt)
53 return fmt && strcmp(fmt, "raw") == 0;
56 static void qed_header_le_to_cpu(const QEDHeader *le, QEDHeader *cpu)
58 cpu->magic = le32_to_cpu(le->magic);
59 cpu->cluster_size = le32_to_cpu(le->cluster_size);
60 cpu->table_size = le32_to_cpu(le->table_size);
61 cpu->header_size = le32_to_cpu(le->header_size);
62 cpu->features = le64_to_cpu(le->features);
63 cpu->compat_features = le64_to_cpu(le->compat_features);
64 cpu->autoclear_features = le64_to_cpu(le->autoclear_features);
65 cpu->l1_table_offset = le64_to_cpu(le->l1_table_offset);
66 cpu->image_size = le64_to_cpu(le->image_size);
67 cpu->backing_filename_offset = le32_to_cpu(le->backing_filename_offset);
68 cpu->backing_filename_size = le32_to_cpu(le->backing_filename_size);
71 static void qed_header_cpu_to_le(const QEDHeader *cpu, QEDHeader *le)
73 le->magic = cpu_to_le32(cpu->magic);
74 le->cluster_size = cpu_to_le32(cpu->cluster_size);
75 le->table_size = cpu_to_le32(cpu->table_size);
76 le->header_size = cpu_to_le32(cpu->header_size);
77 le->features = cpu_to_le64(cpu->features);
78 le->compat_features = cpu_to_le64(cpu->compat_features);
79 le->autoclear_features = cpu_to_le64(cpu->autoclear_features);
80 le->l1_table_offset = cpu_to_le64(cpu->l1_table_offset);
81 le->image_size = cpu_to_le64(cpu->image_size);
82 le->backing_filename_offset = cpu_to_le32(cpu->backing_filename_offset);
83 le->backing_filename_size = cpu_to_le32(cpu->backing_filename_size);
86 int qed_write_header_sync(BDRVQEDState *s)
91 qed_header_cpu_to_le(&s->header, &le);
92 ret = bdrv_pwrite(s->bs->file, 0, &le, sizeof(le));
93 if (ret != sizeof(le)) {
100 * Update header in-place (does not rewrite backing filename or other strings)
102 * This function only updates known header fields in-place and does not affect
103 * extra data after the QED header.
105 * No new allocating reqs can start while this function runs.
107 static int coroutine_fn qed_write_header(BDRVQEDState *s)
109 /* We must write full sectors for O_DIRECT but cannot necessarily generate
110 * the data following the header if an unrecognized compat feature is
111 * active. Therefore, first read the sectors containing the header, update
112 * them, and write back.
115 int nsectors = DIV_ROUND_UP(sizeof(QEDHeader), BDRV_SECTOR_SIZE);
116 size_t len = nsectors * BDRV_SECTOR_SIZE;
120 assert(s->allocating_acb || s->allocating_write_reqs_plugged);
122 buf = qemu_blockalign(s->bs, len);
124 ret = bdrv_co_pread(s->bs->file, 0, len, buf, 0);
130 qed_header_cpu_to_le(&s->header, (QEDHeader *) buf);
132 ret = bdrv_co_pwrite(s->bs->file, 0, len, buf, 0);
143 static uint64_t qed_max_image_size(uint32_t cluster_size, uint32_t table_size)
145 uint64_t table_entries;
148 table_entries = (table_size * cluster_size) / sizeof(uint64_t);
149 l2_size = table_entries * cluster_size;
151 return l2_size * table_entries;
154 static bool qed_is_cluster_size_valid(uint32_t cluster_size)
156 if (cluster_size < QED_MIN_CLUSTER_SIZE ||
157 cluster_size > QED_MAX_CLUSTER_SIZE) {
160 if (cluster_size & (cluster_size - 1)) {
161 return false; /* not power of 2 */
166 static bool qed_is_table_size_valid(uint32_t table_size)
168 if (table_size < QED_MIN_TABLE_SIZE ||
169 table_size > QED_MAX_TABLE_SIZE) {
172 if (table_size & (table_size - 1)) {
173 return false; /* not power of 2 */
178 static bool qed_is_image_size_valid(uint64_t image_size, uint32_t cluster_size,
181 if (image_size % BDRV_SECTOR_SIZE != 0) {
182 return false; /* not multiple of sector size */
184 if (image_size > qed_max_image_size(cluster_size, table_size)) {
185 return false; /* image is too large */
191 * Read a string of known length from the image file
194 * @offset: File offset to start of string, in bytes
195 * @n: String length in bytes
196 * @buf: Destination buffer
197 * @buflen: Destination buffer length in bytes
198 * @ret: 0 on success, -errno on failure
200 * The string is NUL-terminated.
202 static int qed_read_string(BdrvChild *file, uint64_t offset, size_t n,
203 char *buf, size_t buflen)
209 ret = bdrv_pread(file, offset, buf, n);
218 * Allocate new clusters
221 * @n: Number of contiguous clusters to allocate
222 * @ret: Offset of first allocated cluster
224 * This function only produces the offset where the new clusters should be
225 * written. It updates BDRVQEDState but does not make any changes to the image
228 * Called with table_lock held.
230 static uint64_t qed_alloc_clusters(BDRVQEDState *s, unsigned int n)
232 uint64_t offset = s->file_size;
233 s->file_size += n * s->header.cluster_size;
237 QEDTable *qed_alloc_table(BDRVQEDState *s)
239 /* Honor O_DIRECT memory alignment requirements */
240 return qemu_blockalign(s->bs,
241 s->header.cluster_size * s->header.table_size);
245 * Allocate a new zeroed L2 table
247 * Called with table_lock held.
249 static CachedL2Table *qed_new_l2_table(BDRVQEDState *s)
251 CachedL2Table *l2_table = qed_alloc_l2_cache_entry(&s->l2_cache);
253 l2_table->table = qed_alloc_table(s);
254 l2_table->offset = qed_alloc_clusters(s, s->header.table_size);
256 memset(l2_table->table->offsets, 0,
257 s->header.cluster_size * s->header.table_size);
261 static bool qed_plug_allocating_write_reqs(BDRVQEDState *s)
263 qemu_co_mutex_lock(&s->table_lock);
265 /* No reentrancy is allowed. */
266 assert(!s->allocating_write_reqs_plugged);
267 if (s->allocating_acb != NULL) {
268 /* Another allocating write came concurrently. This cannot happen
269 * from bdrv_qed_co_drain_begin, but it can happen when the timer runs.
271 qemu_co_mutex_unlock(&s->table_lock);
275 s->allocating_write_reqs_plugged = true;
276 qemu_co_mutex_unlock(&s->table_lock);
280 static void qed_unplug_allocating_write_reqs(BDRVQEDState *s)
282 qemu_co_mutex_lock(&s->table_lock);
283 assert(s->allocating_write_reqs_plugged);
284 s->allocating_write_reqs_plugged = false;
285 qemu_co_queue_next(&s->allocating_write_reqs);
286 qemu_co_mutex_unlock(&s->table_lock);
289 static void coroutine_fn qed_need_check_timer_entry(void *opaque)
291 BDRVQEDState *s = opaque;
294 trace_qed_need_check_timer_cb(s);
296 if (!qed_plug_allocating_write_reqs(s)) {
300 /* Ensure writes are on disk before clearing flag */
301 ret = bdrv_co_flush(s->bs->file->bs);
303 qed_unplug_allocating_write_reqs(s);
307 s->header.features &= ~QED_F_NEED_CHECK;
308 ret = qed_write_header(s);
311 qed_unplug_allocating_write_reqs(s);
313 ret = bdrv_co_flush(s->bs);
317 static void qed_need_check_timer_cb(void *opaque)
319 Coroutine *co = qemu_coroutine_create(qed_need_check_timer_entry, opaque);
320 qemu_coroutine_enter(co);
323 static void qed_start_need_check_timer(BDRVQEDState *s)
325 trace_qed_start_need_check_timer(s);
327 /* Use QEMU_CLOCK_VIRTUAL so we don't alter the image file while suspended for
330 timer_mod(s->need_check_timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
331 NANOSECONDS_PER_SECOND * QED_NEED_CHECK_TIMEOUT);
334 /* It's okay to call this multiple times or when no timer is started */
335 static void qed_cancel_need_check_timer(BDRVQEDState *s)
337 trace_qed_cancel_need_check_timer(s);
338 timer_del(s->need_check_timer);
341 static void bdrv_qed_detach_aio_context(BlockDriverState *bs)
343 BDRVQEDState *s = bs->opaque;
345 qed_cancel_need_check_timer(s);
346 timer_free(s->need_check_timer);
349 static void bdrv_qed_attach_aio_context(BlockDriverState *bs,
350 AioContext *new_context)
352 BDRVQEDState *s = bs->opaque;
354 s->need_check_timer = aio_timer_new(new_context,
355 QEMU_CLOCK_VIRTUAL, SCALE_NS,
356 qed_need_check_timer_cb, s);
357 if (s->header.features & QED_F_NEED_CHECK) {
358 qed_start_need_check_timer(s);
362 static void coroutine_fn bdrv_qed_co_drain_begin(BlockDriverState *bs)
364 BDRVQEDState *s = bs->opaque;
366 /* Fire the timer immediately in order to start doing I/O as soon as the
369 if (s->need_check_timer && timer_pending(s->need_check_timer)) {
370 qed_cancel_need_check_timer(s);
371 qed_need_check_timer_entry(s);
375 static void bdrv_qed_init_state(BlockDriverState *bs)
377 BDRVQEDState *s = bs->opaque;
379 memset(s, 0, sizeof(BDRVQEDState));
381 qemu_co_mutex_init(&s->table_lock);
382 qemu_co_queue_init(&s->allocating_write_reqs);
385 /* Called with table_lock held. */
386 static int coroutine_fn bdrv_qed_do_open(BlockDriverState *bs, QDict *options,
387 int flags, Error **errp)
389 BDRVQEDState *s = bs->opaque;
394 ret = bdrv_pread(bs->file, 0, &le_header, sizeof(le_header));
396 error_setg(errp, "Failed to read QED header");
399 qed_header_le_to_cpu(&le_header, &s->header);
401 if (s->header.magic != QED_MAGIC) {
402 error_setg(errp, "Image not in QED format");
405 if (s->header.features & ~QED_FEATURE_MASK) {
406 /* image uses unsupported feature bits */
407 error_setg(errp, "Unsupported QED features: %" PRIx64,
408 s->header.features & ~QED_FEATURE_MASK);
411 if (!qed_is_cluster_size_valid(s->header.cluster_size)) {
412 error_setg(errp, "QED cluster size is invalid");
416 /* Round down file size to the last cluster */
417 file_size = bdrv_getlength(bs->file->bs);
419 error_setg(errp, "Failed to get file length");
422 s->file_size = qed_start_of_cluster(s, file_size);
424 if (!qed_is_table_size_valid(s->header.table_size)) {
425 error_setg(errp, "QED table size is invalid");
428 if (!qed_is_image_size_valid(s->header.image_size,
429 s->header.cluster_size,
430 s->header.table_size)) {
431 error_setg(errp, "QED image size is invalid");
434 if (!qed_check_table_offset(s, s->header.l1_table_offset)) {
435 error_setg(errp, "QED table offset is invalid");
439 s->table_nelems = (s->header.cluster_size * s->header.table_size) /
441 s->l2_shift = ctz32(s->header.cluster_size);
442 s->l2_mask = s->table_nelems - 1;
443 s->l1_shift = s->l2_shift + ctz32(s->table_nelems);
445 /* Header size calculation must not overflow uint32_t */
446 if (s->header.header_size > UINT32_MAX / s->header.cluster_size) {
447 error_setg(errp, "QED header size is too large");
451 if ((s->header.features & QED_F_BACKING_FILE)) {
452 if ((uint64_t)s->header.backing_filename_offset +
453 s->header.backing_filename_size >
454 s->header.cluster_size * s->header.header_size) {
455 error_setg(errp, "QED backing filename offset is invalid");
459 ret = qed_read_string(bs->file, s->header.backing_filename_offset,
460 s->header.backing_filename_size,
461 bs->auto_backing_file,
462 sizeof(bs->auto_backing_file));
464 error_setg(errp, "Failed to read backing filename");
467 pstrcpy(bs->backing_file, sizeof(bs->backing_file),
468 bs->auto_backing_file);
470 if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) {
471 pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw");
475 /* Reset unknown autoclear feature bits. This is a backwards
476 * compatibility mechanism that allows images to be opened by older
477 * programs, which "knock out" unknown feature bits. When an image is
478 * opened by a newer program again it can detect that the autoclear
479 * feature is no longer valid.
481 if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 &&
482 !bdrv_is_read_only(bs->file->bs) && !(flags & BDRV_O_INACTIVE)) {
483 s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK;
485 ret = qed_write_header_sync(s);
487 error_setg(errp, "Failed to update header");
491 /* From here on only known autoclear feature bits are valid */
492 bdrv_flush(bs->file->bs);
495 s->l1_table = qed_alloc_table(s);
496 qed_init_l2_cache(&s->l2_cache);
498 ret = qed_read_l1_table_sync(s);
500 error_setg(errp, "Failed to read L1 table");
504 /* If image was not closed cleanly, check consistency */
505 if (!(flags & BDRV_O_CHECK) && (s->header.features & QED_F_NEED_CHECK)) {
506 /* Read-only images cannot be fixed. There is no risk of corruption
507 * since write operations are not possible. Therefore, allow
508 * potentially inconsistent images to be opened read-only. This can
509 * aid data recovery from an otherwise inconsistent image.
511 if (!bdrv_is_read_only(bs->file->bs) &&
512 !(flags & BDRV_O_INACTIVE)) {
513 BdrvCheckResult result = {0};
515 ret = qed_check(s, &result, true);
517 error_setg(errp, "Image corrupted");
523 bdrv_qed_attach_aio_context(bs, bdrv_get_aio_context(bs));
527 qed_free_l2_cache(&s->l2_cache);
528 qemu_vfree(s->l1_table);
533 typedef struct QEDOpenCo {
534 BlockDriverState *bs;
541 static void coroutine_fn bdrv_qed_open_entry(void *opaque)
543 QEDOpenCo *qoc = opaque;
544 BDRVQEDState *s = qoc->bs->opaque;
546 qemu_co_mutex_lock(&s->table_lock);
547 qoc->ret = bdrv_qed_do_open(qoc->bs, qoc->options, qoc->flags, qoc->errp);
548 qemu_co_mutex_unlock(&s->table_lock);
551 static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags,
562 bs->file = bdrv_open_child(NULL, options, "file", bs, &child_of_bds,
563 BDRV_CHILD_IMAGE, false, errp);
568 bdrv_qed_init_state(bs);
569 if (qemu_in_coroutine()) {
570 bdrv_qed_open_entry(&qoc);
572 assert(qemu_get_current_aio_context() == qemu_get_aio_context());
573 qemu_coroutine_enter(qemu_coroutine_create(bdrv_qed_open_entry, &qoc));
574 BDRV_POLL_WHILE(bs, qoc.ret == -EINPROGRESS);
576 BDRV_POLL_WHILE(bs, qoc.ret == -EINPROGRESS);
580 static void bdrv_qed_refresh_limits(BlockDriverState *bs, Error **errp)
582 BDRVQEDState *s = bs->opaque;
584 bs->bl.pwrite_zeroes_alignment = s->header.cluster_size;
587 /* We have nothing to do for QED reopen, stubs just return
589 static int bdrv_qed_reopen_prepare(BDRVReopenState *state,
590 BlockReopenQueue *queue, Error **errp)
595 static void bdrv_qed_close(BlockDriverState *bs)
597 BDRVQEDState *s = bs->opaque;
599 bdrv_qed_detach_aio_context(bs);
601 /* Ensure writes reach stable storage */
602 bdrv_flush(bs->file->bs);
604 /* Clean shutdown, no check required on next open */
605 if (s->header.features & QED_F_NEED_CHECK) {
606 s->header.features &= ~QED_F_NEED_CHECK;
607 qed_write_header_sync(s);
610 qed_free_l2_cache(&s->l2_cache);
611 qemu_vfree(s->l1_table);
614 static int coroutine_fn bdrv_qed_co_create(BlockdevCreateOptions *opts,
617 BlockdevCreateOptionsQed *qed_opts;
618 BlockBackend *blk = NULL;
619 BlockDriverState *bs = NULL;
623 uint8_t *l1_table = NULL;
627 assert(opts->driver == BLOCKDEV_DRIVER_QED);
628 qed_opts = &opts->u.qed;
630 /* Validate options and set default values */
631 if (!qed_opts->has_cluster_size) {
632 qed_opts->cluster_size = QED_DEFAULT_CLUSTER_SIZE;
634 if (!qed_opts->has_table_size) {
635 qed_opts->table_size = QED_DEFAULT_TABLE_SIZE;
638 if (!qed_is_cluster_size_valid(qed_opts->cluster_size)) {
639 error_setg(errp, "QED cluster size must be within range [%u, %u] "
641 QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE);
644 if (!qed_is_table_size_valid(qed_opts->table_size)) {
645 error_setg(errp, "QED table size must be within range [%u, %u] "
647 QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE);
650 if (!qed_is_image_size_valid(qed_opts->size, qed_opts->cluster_size,
651 qed_opts->table_size))
653 error_setg(errp, "QED image size must be a non-zero multiple of "
654 "cluster size and less than %" PRIu64 " bytes",
655 qed_max_image_size(qed_opts->cluster_size,
656 qed_opts->table_size));
660 /* Create BlockBackend to write to the image */
661 bs = bdrv_open_blockdev_ref(qed_opts->file, errp);
666 blk = blk_new_with_bs(bs, BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL,
672 blk_set_allow_write_beyond_eof(blk, true);
674 /* Prepare image format */
675 header = (QEDHeader) {
677 .cluster_size = qed_opts->cluster_size,
678 .table_size = qed_opts->table_size,
681 .compat_features = 0,
682 .l1_table_offset = qed_opts->cluster_size,
683 .image_size = qed_opts->size,
686 l1_size = header.cluster_size * header.table_size;
689 * The QED format associates file length with allocation status,
690 * so a new file (which is empty) must have a length of 0.
692 ret = blk_truncate(blk, 0, true, PREALLOC_MODE_OFF, 0, errp);
697 if (qed_opts->has_backing_file) {
698 header.features |= QED_F_BACKING_FILE;
699 header.backing_filename_offset = sizeof(le_header);
700 header.backing_filename_size = strlen(qed_opts->backing_file);
702 if (qed_opts->has_backing_fmt) {
703 const char *backing_fmt = BlockdevDriver_str(qed_opts->backing_fmt);
704 if (qed_fmt_is_raw(backing_fmt)) {
705 header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
710 qed_header_cpu_to_le(&header, &le_header);
711 ret = blk_pwrite(blk, 0, &le_header, sizeof(le_header), 0);
715 ret = blk_pwrite(blk, sizeof(le_header), qed_opts->backing_file,
716 header.backing_filename_size, 0);
721 l1_table = g_malloc0(l1_size);
722 ret = blk_pwrite(blk, header.l1_table_offset, l1_table, l1_size, 0);
727 ret = 0; /* success */
735 static int coroutine_fn bdrv_qed_co_create_opts(BlockDriver *drv,
736 const char *filename,
740 BlockdevCreateOptions *create_options = NULL;
743 BlockDriverState *bs = NULL;
746 static const QDictRenames opt_renames[] = {
747 { BLOCK_OPT_BACKING_FILE, "backing-file" },
748 { BLOCK_OPT_BACKING_FMT, "backing-fmt" },
749 { BLOCK_OPT_CLUSTER_SIZE, "cluster-size" },
750 { BLOCK_OPT_TABLE_SIZE, "table-size" },
754 /* Parse options and convert legacy syntax */
755 qdict = qemu_opts_to_qdict_filtered(opts, NULL, &qed_create_opts, true);
757 if (!qdict_rename_keys(qdict, opt_renames, errp)) {
762 /* Create and open the file (protocol layer) */
763 ret = bdrv_create_file(filename, opts, errp);
768 bs = bdrv_open(filename, NULL, NULL,
769 BDRV_O_RDWR | BDRV_O_RESIZE | BDRV_O_PROTOCOL, errp);
775 /* Now get the QAPI type BlockdevCreateOptions */
776 qdict_put_str(qdict, "driver", "qed");
777 qdict_put_str(qdict, "file", bs->node_name);
779 v = qobject_input_visitor_new_flat_confused(qdict, errp);
785 visit_type_BlockdevCreateOptions(v, NULL, &create_options, errp);
787 if (!create_options) {
792 /* Silently round up size */
793 assert(create_options->driver == BLOCKDEV_DRIVER_QED);
794 create_options->u.qed.size =
795 ROUND_UP(create_options->u.qed.size, BDRV_SECTOR_SIZE);
797 /* Create the qed image (format layer) */
798 ret = bdrv_qed_co_create(create_options, errp);
801 qobject_unref(qdict);
803 qapi_free_BlockdevCreateOptions(create_options);
807 static int coroutine_fn bdrv_qed_co_block_status(BlockDriverState *bs,
809 int64_t pos, int64_t bytes,
810 int64_t *pnum, int64_t *map,
811 BlockDriverState **file)
813 BDRVQEDState *s = bs->opaque;
814 size_t len = MIN(bytes, SIZE_MAX);
816 QEDRequest request = { .l2_table = NULL };
820 qemu_co_mutex_lock(&s->table_lock);
821 ret = qed_find_cluster(s, &request, pos, &len, &offset);
825 case QED_CLUSTER_FOUND:
826 *map = offset | qed_offset_into_cluster(s, pos);
827 status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
828 *file = bs->file->bs;
830 case QED_CLUSTER_ZERO:
831 status = BDRV_BLOCK_ZERO;
843 qed_unref_l2_cache_entry(request.l2_table);
844 qemu_co_mutex_unlock(&s->table_lock);
849 static BDRVQEDState *acb_to_s(QEDAIOCB *acb)
851 return acb->bs->opaque;
855 * Read from the backing file or zero-fill if no backing file
858 * @pos: Byte position in device
859 * @qiov: Destination I/O vector
861 * This function reads qiov->size bytes starting at pos from the backing file.
862 * If there is no backing file then zeroes are read.
864 static int coroutine_fn qed_read_backing_file(BDRVQEDState *s, uint64_t pos,
867 if (s->bs->backing) {
868 BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO);
869 return bdrv_co_preadv(s->bs->backing, pos, qiov->size, qiov, 0);
871 qemu_iovec_memset(qiov, 0, 0, qiov->size);
876 * Copy data from backing file into the image
879 * @pos: Byte position in device
880 * @len: Number of bytes
881 * @offset: Byte offset in image file
883 static int coroutine_fn qed_copy_from_backing_file(BDRVQEDState *s,
884 uint64_t pos, uint64_t len,
890 /* Skip copy entirely if there is no work to do */
895 qemu_iovec_init_buf(&qiov, qemu_blockalign(s->bs, len), len);
897 ret = qed_read_backing_file(s, pos, &qiov);
903 BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE);
904 ret = bdrv_co_pwritev(s->bs->file, offset, qiov.size, &qiov, 0);
910 qemu_vfree(qemu_iovec_buf(&qiov));
915 * Link one or more contiguous clusters into a table
919 * @index: First cluster index
920 * @n: Number of contiguous clusters
921 * @cluster: First cluster offset
923 * The cluster offset may be an allocated byte offset in the image file, the
924 * zero cluster marker, or the unallocated cluster marker.
926 * Called with table_lock held.
928 static void coroutine_fn qed_update_l2_table(BDRVQEDState *s, QEDTable *table,
929 int index, unsigned int n,
933 for (i = index; i < index + n; i++) {
934 table->offsets[i] = cluster;
935 if (!qed_offset_is_unalloc_cluster(cluster) &&
936 !qed_offset_is_zero_cluster(cluster)) {
937 cluster += s->header.cluster_size;
942 /* Called with table_lock held. */
943 static void coroutine_fn qed_aio_complete(QEDAIOCB *acb)
945 BDRVQEDState *s = acb_to_s(acb);
948 qemu_iovec_destroy(&acb->cur_qiov);
949 qed_unref_l2_cache_entry(acb->request.l2_table);
951 /* Free the buffer we may have allocated for zero writes */
952 if (acb->flags & QED_AIOCB_ZERO) {
953 qemu_vfree(acb->qiov->iov[0].iov_base);
954 acb->qiov->iov[0].iov_base = NULL;
957 /* Start next allocating write request waiting behind this one. Note that
958 * requests enqueue themselves when they first hit an unallocated cluster
959 * but they wait until the entire request is finished before waking up the
960 * next request in the queue. This ensures that we don't cycle through
961 * requests multiple times but rather finish one at a time completely.
963 if (acb == s->allocating_acb) {
964 s->allocating_acb = NULL;
965 if (!qemu_co_queue_empty(&s->allocating_write_reqs)) {
966 qemu_co_queue_next(&s->allocating_write_reqs);
967 } else if (s->header.features & QED_F_NEED_CHECK) {
968 qed_start_need_check_timer(s);
974 * Update L1 table with new L2 table offset and write it out
976 * Called with table_lock held.
978 static int coroutine_fn qed_aio_write_l1_update(QEDAIOCB *acb)
980 BDRVQEDState *s = acb_to_s(acb);
981 CachedL2Table *l2_table = acb->request.l2_table;
982 uint64_t l2_offset = l2_table->offset;
985 index = qed_l1_index(s, acb->cur_pos);
986 s->l1_table->offsets[index] = l2_table->offset;
988 ret = qed_write_l1_table(s, index, 1);
990 /* Commit the current L2 table to the cache */
991 qed_commit_l2_cache_entry(&s->l2_cache, l2_table);
993 /* This is guaranteed to succeed because we just committed the entry to the
996 acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset);
997 assert(acb->request.l2_table != NULL);
1004 * Update L2 table with new cluster offsets and write them out
1006 * Called with table_lock held.
1008 static int coroutine_fn qed_aio_write_l2_update(QEDAIOCB *acb, uint64_t offset)
1010 BDRVQEDState *s = acb_to_s(acb);
1011 bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1;
1015 qed_unref_l2_cache_entry(acb->request.l2_table);
1016 acb->request.l2_table = qed_new_l2_table(s);
1019 index = qed_l2_index(s, acb->cur_pos);
1020 qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters,
1024 /* Write out the whole new L2 table */
1025 ret = qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true);
1029 return qed_aio_write_l1_update(acb);
1031 /* Write out only the updated part of the L2 table */
1032 ret = qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters,
1042 * Write data to the image file
1044 * Called with table_lock *not* held.
1046 static int coroutine_fn qed_aio_write_main(QEDAIOCB *acb)
1048 BDRVQEDState *s = acb_to_s(acb);
1049 uint64_t offset = acb->cur_cluster +
1050 qed_offset_into_cluster(s, acb->cur_pos);
1052 trace_qed_aio_write_main(s, acb, 0, offset, acb->cur_qiov.size);
1054 BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO);
1055 return bdrv_co_pwritev(s->bs->file, offset, acb->cur_qiov.size,
1060 * Populate untouched regions of new data cluster
1062 * Called with table_lock held.
1064 static int coroutine_fn qed_aio_write_cow(QEDAIOCB *acb)
1066 BDRVQEDState *s = acb_to_s(acb);
1067 uint64_t start, len, offset;
1070 qemu_co_mutex_unlock(&s->table_lock);
1072 /* Populate front untouched region of new data cluster */
1073 start = qed_start_of_cluster(s, acb->cur_pos);
1074 len = qed_offset_into_cluster(s, acb->cur_pos);
1076 trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster);
1077 ret = qed_copy_from_backing_file(s, start, len, acb->cur_cluster);
1082 /* Populate back untouched region of new data cluster */
1083 start = acb->cur_pos + acb->cur_qiov.size;
1084 len = qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start;
1085 offset = acb->cur_cluster +
1086 qed_offset_into_cluster(s, acb->cur_pos) +
1089 trace_qed_aio_write_postfill(s, acb, start, len, offset);
1090 ret = qed_copy_from_backing_file(s, start, len, offset);
1095 ret = qed_aio_write_main(acb);
1100 if (s->bs->backing) {
1102 * Flush new data clusters before updating the L2 table
1104 * This flush is necessary when a backing file is in use. A crash
1105 * during an allocating write could result in empty clusters in the
1106 * image. If the write only touched a subregion of the cluster,
1107 * then backing image sectors have been lost in the untouched
1108 * region. The solution is to flush after writing a new data
1109 * cluster and before updating the L2 table.
1111 ret = bdrv_co_flush(s->bs->file->bs);
1115 qemu_co_mutex_lock(&s->table_lock);
1120 * Check if the QED_F_NEED_CHECK bit should be set during allocating write
1122 static bool qed_should_set_need_check(BDRVQEDState *s)
1124 /* The flush before L2 update path ensures consistency */
1125 if (s->bs->backing) {
1129 return !(s->header.features & QED_F_NEED_CHECK);
1133 * Write new data cluster
1135 * @acb: Write request
1136 * @len: Length in bytes
1138 * This path is taken when writing to previously unallocated clusters.
1140 * Called with table_lock held.
1142 static int coroutine_fn qed_aio_write_alloc(QEDAIOCB *acb, size_t len)
1144 BDRVQEDState *s = acb_to_s(acb);
1147 /* Cancel timer when the first allocating request comes in */
1148 if (s->allocating_acb == NULL) {
1149 qed_cancel_need_check_timer(s);
1152 /* Freeze this request if another allocating write is in progress */
1153 if (s->allocating_acb != acb || s->allocating_write_reqs_plugged) {
1154 if (s->allocating_acb != NULL) {
1155 qemu_co_queue_wait(&s->allocating_write_reqs, &s->table_lock);
1156 assert(s->allocating_acb == NULL);
1158 s->allocating_acb = acb;
1159 return -EAGAIN; /* start over with looking up table entries */
1162 acb->cur_nclusters = qed_bytes_to_clusters(s,
1163 qed_offset_into_cluster(s, acb->cur_pos) + len);
1164 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1166 if (acb->flags & QED_AIOCB_ZERO) {
1167 /* Skip ahead if the clusters are already zero */
1168 if (acb->find_cluster_ret == QED_CLUSTER_ZERO) {
1171 acb->cur_cluster = 1;
1173 acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters);
1176 if (qed_should_set_need_check(s)) {
1177 s->header.features |= QED_F_NEED_CHECK;
1178 ret = qed_write_header(s);
1184 if (!(acb->flags & QED_AIOCB_ZERO)) {
1185 ret = qed_aio_write_cow(acb);
1191 return qed_aio_write_l2_update(acb, acb->cur_cluster);
1195 * Write data cluster in place
1197 * @acb: Write request
1198 * @offset: Cluster offset in bytes
1199 * @len: Length in bytes
1201 * This path is taken when writing to already allocated clusters.
1203 * Called with table_lock held.
1205 static int coroutine_fn qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset,
1208 BDRVQEDState *s = acb_to_s(acb);
1211 qemu_co_mutex_unlock(&s->table_lock);
1213 /* Allocate buffer for zero writes */
1214 if (acb->flags & QED_AIOCB_ZERO) {
1215 struct iovec *iov = acb->qiov->iov;
1217 if (!iov->iov_base) {
1218 iov->iov_base = qemu_try_blockalign(acb->bs, iov->iov_len);
1219 if (iov->iov_base == NULL) {
1223 memset(iov->iov_base, 0, iov->iov_len);
1227 /* Calculate the I/O vector */
1228 acb->cur_cluster = offset;
1229 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1231 /* Do the actual write. */
1232 r = qed_aio_write_main(acb);
1234 qemu_co_mutex_lock(&s->table_lock);
1239 * Write data cluster
1241 * @opaque: Write request
1242 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1243 * @offset: Cluster offset in bytes
1244 * @len: Length in bytes
1246 * Called with table_lock held.
1248 static int coroutine_fn qed_aio_write_data(void *opaque, int ret,
1249 uint64_t offset, size_t len)
1251 QEDAIOCB *acb = opaque;
1253 trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len);
1255 acb->find_cluster_ret = ret;
1258 case QED_CLUSTER_FOUND:
1259 return qed_aio_write_inplace(acb, offset, len);
1261 case QED_CLUSTER_L2:
1262 case QED_CLUSTER_L1:
1263 case QED_CLUSTER_ZERO:
1264 return qed_aio_write_alloc(acb, len);
1267 g_assert_not_reached();
1274 * @opaque: Read request
1275 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1276 * @offset: Cluster offset in bytes
1277 * @len: Length in bytes
1279 * Called with table_lock held.
1281 static int coroutine_fn qed_aio_read_data(void *opaque, int ret,
1282 uint64_t offset, size_t len)
1284 QEDAIOCB *acb = opaque;
1285 BDRVQEDState *s = acb_to_s(acb);
1286 BlockDriverState *bs = acb->bs;
1289 qemu_co_mutex_unlock(&s->table_lock);
1291 /* Adjust offset into cluster */
1292 offset += qed_offset_into_cluster(s, acb->cur_pos);
1294 trace_qed_aio_read_data(s, acb, ret, offset, len);
1296 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1298 /* Handle zero cluster and backing file reads, otherwise read
1299 * data cluster directly.
1301 if (ret == QED_CLUSTER_ZERO) {
1302 qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size);
1304 } else if (ret != QED_CLUSTER_FOUND) {
1305 r = qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov);
1307 BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
1308 r = bdrv_co_preadv(bs->file, offset, acb->cur_qiov.size,
1312 qemu_co_mutex_lock(&s->table_lock);
1317 * Begin next I/O or complete the request
1319 static int coroutine_fn qed_aio_next_io(QEDAIOCB *acb)
1321 BDRVQEDState *s = acb_to_s(acb);
1326 qemu_co_mutex_lock(&s->table_lock);
1328 trace_qed_aio_next_io(s, acb, 0, acb->cur_pos + acb->cur_qiov.size);
1330 acb->qiov_offset += acb->cur_qiov.size;
1331 acb->cur_pos += acb->cur_qiov.size;
1332 qemu_iovec_reset(&acb->cur_qiov);
1334 /* Complete request */
1335 if (acb->cur_pos >= acb->end_pos) {
1340 /* Find next cluster and start I/O */
1341 len = acb->end_pos - acb->cur_pos;
1342 ret = qed_find_cluster(s, &acb->request, acb->cur_pos, &len, &offset);
1347 if (acb->flags & QED_AIOCB_WRITE) {
1348 ret = qed_aio_write_data(acb, ret, offset, len);
1350 ret = qed_aio_read_data(acb, ret, offset, len);
1353 if (ret < 0 && ret != -EAGAIN) {
1358 trace_qed_aio_complete(s, acb, ret);
1359 qed_aio_complete(acb);
1360 qemu_co_mutex_unlock(&s->table_lock);
1364 static int coroutine_fn qed_co_request(BlockDriverState *bs, int64_t sector_num,
1365 QEMUIOVector *qiov, int nb_sectors,
1370 .cur_pos = (uint64_t) sector_num * BDRV_SECTOR_SIZE,
1371 .end_pos = (sector_num + nb_sectors) * BDRV_SECTOR_SIZE,
1375 qemu_iovec_init(&acb.cur_qiov, qiov->niov);
1377 trace_qed_aio_setup(bs->opaque, &acb, sector_num, nb_sectors, NULL, flags);
1380 return qed_aio_next_io(&acb);
1383 static int coroutine_fn bdrv_qed_co_readv(BlockDriverState *bs,
1384 int64_t sector_num, int nb_sectors,
1387 return qed_co_request(bs, sector_num, qiov, nb_sectors, 0);
1390 static int coroutine_fn bdrv_qed_co_writev(BlockDriverState *bs,
1391 int64_t sector_num, int nb_sectors,
1392 QEMUIOVector *qiov, int flags)
1395 return qed_co_request(bs, sector_num, qiov, nb_sectors, QED_AIOCB_WRITE);
1398 static int coroutine_fn bdrv_qed_co_pwrite_zeroes(BlockDriverState *bs,
1401 BdrvRequestFlags flags)
1403 BDRVQEDState *s = bs->opaque;
1406 * Zero writes start without an I/O buffer. If a buffer becomes necessary
1407 * then it will be allocated during request processing.
1409 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, bytes);
1411 /* Fall back if the request is not aligned */
1412 if (qed_offset_into_cluster(s, offset) ||
1413 qed_offset_into_cluster(s, bytes)) {
1417 return qed_co_request(bs, offset >> BDRV_SECTOR_BITS, &qiov,
1418 bytes >> BDRV_SECTOR_BITS,
1419 QED_AIOCB_WRITE | QED_AIOCB_ZERO);
1422 static int coroutine_fn bdrv_qed_co_truncate(BlockDriverState *bs,
1425 PreallocMode prealloc,
1426 BdrvRequestFlags flags,
1429 BDRVQEDState *s = bs->opaque;
1430 uint64_t old_image_size;
1433 if (prealloc != PREALLOC_MODE_OFF) {
1434 error_setg(errp, "Unsupported preallocation mode '%s'",
1435 PreallocMode_str(prealloc));
1439 if (!qed_is_image_size_valid(offset, s->header.cluster_size,
1440 s->header.table_size)) {
1441 error_setg(errp, "Invalid image size specified");
1445 if ((uint64_t)offset < s->header.image_size) {
1446 error_setg(errp, "Shrinking images is currently not supported");
1450 old_image_size = s->header.image_size;
1451 s->header.image_size = offset;
1452 ret = qed_write_header_sync(s);
1454 s->header.image_size = old_image_size;
1455 error_setg_errno(errp, -ret, "Failed to update the image size");
1460 static int64_t bdrv_qed_getlength(BlockDriverState *bs)
1462 BDRVQEDState *s = bs->opaque;
1463 return s->header.image_size;
1466 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1468 BDRVQEDState *s = bs->opaque;
1470 memset(bdi, 0, sizeof(*bdi));
1471 bdi->cluster_size = s->header.cluster_size;
1472 bdi->is_dirty = s->header.features & QED_F_NEED_CHECK;
1476 static int bdrv_qed_change_backing_file(BlockDriverState *bs,
1477 const char *backing_file,
1478 const char *backing_fmt)
1480 BDRVQEDState *s = bs->opaque;
1481 QEDHeader new_header, le_header;
1483 size_t buffer_len, backing_file_len;
1486 /* Refuse to set backing filename if unknown compat feature bits are
1487 * active. If the image uses an unknown compat feature then we may not
1488 * know the layout of data following the header structure and cannot safely
1491 if (backing_file && (s->header.compat_features &
1492 ~QED_COMPAT_FEATURE_MASK)) {
1496 memcpy(&new_header, &s->header, sizeof(new_header));
1498 new_header.features &= ~(QED_F_BACKING_FILE |
1499 QED_F_BACKING_FORMAT_NO_PROBE);
1501 /* Adjust feature flags */
1503 new_header.features |= QED_F_BACKING_FILE;
1505 if (qed_fmt_is_raw(backing_fmt)) {
1506 new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
1510 /* Calculate new header size */
1511 backing_file_len = 0;
1514 backing_file_len = strlen(backing_file);
1517 buffer_len = sizeof(new_header);
1518 new_header.backing_filename_offset = buffer_len;
1519 new_header.backing_filename_size = backing_file_len;
1520 buffer_len += backing_file_len;
1522 /* Make sure we can rewrite header without failing */
1523 if (buffer_len > new_header.header_size * new_header.cluster_size) {
1527 /* Prepare new header */
1528 buffer = g_malloc(buffer_len);
1530 qed_header_cpu_to_le(&new_header, &le_header);
1531 memcpy(buffer, &le_header, sizeof(le_header));
1532 buffer_len = sizeof(le_header);
1535 memcpy(buffer + buffer_len, backing_file, backing_file_len);
1536 buffer_len += backing_file_len;
1539 /* Write new header */
1540 ret = bdrv_pwrite_sync(bs->file, 0, buffer, buffer_len);
1543 memcpy(&s->header, &new_header, sizeof(new_header));
1548 static void coroutine_fn bdrv_qed_co_invalidate_cache(BlockDriverState *bs,
1551 BDRVQEDState *s = bs->opaque;
1556 bdrv_qed_init_state(bs);
1557 qemu_co_mutex_lock(&s->table_lock);
1558 ret = bdrv_qed_do_open(bs, NULL, bs->open_flags, errp);
1559 qemu_co_mutex_unlock(&s->table_lock);
1561 error_prepend(errp, "Could not reopen qed layer: ");
1565 static int coroutine_fn bdrv_qed_co_check(BlockDriverState *bs,
1566 BdrvCheckResult *result,
1569 BDRVQEDState *s = bs->opaque;
1572 qemu_co_mutex_lock(&s->table_lock);
1573 ret = qed_check(s, result, !!fix);
1574 qemu_co_mutex_unlock(&s->table_lock);
1579 static QemuOptsList qed_create_opts = {
1580 .name = "qed-create-opts",
1581 .head = QTAILQ_HEAD_INITIALIZER(qed_create_opts.head),
1584 .name = BLOCK_OPT_SIZE,
1585 .type = QEMU_OPT_SIZE,
1586 .help = "Virtual disk size"
1589 .name = BLOCK_OPT_BACKING_FILE,
1590 .type = QEMU_OPT_STRING,
1591 .help = "File name of a base image"
1594 .name = BLOCK_OPT_BACKING_FMT,
1595 .type = QEMU_OPT_STRING,
1596 .help = "Image format of the base image"
1599 .name = BLOCK_OPT_CLUSTER_SIZE,
1600 .type = QEMU_OPT_SIZE,
1601 .help = "Cluster size (in bytes)",
1602 .def_value_str = stringify(QED_DEFAULT_CLUSTER_SIZE)
1605 .name = BLOCK_OPT_TABLE_SIZE,
1606 .type = QEMU_OPT_SIZE,
1607 .help = "L1/L2 table size (in clusters)"
1609 { /* end of list */ }
1613 static BlockDriver bdrv_qed = {
1614 .format_name = "qed",
1615 .instance_size = sizeof(BDRVQEDState),
1616 .create_opts = &qed_create_opts,
1618 .supports_backing = true,
1620 .bdrv_probe = bdrv_qed_probe,
1621 .bdrv_open = bdrv_qed_open,
1622 .bdrv_close = bdrv_qed_close,
1623 .bdrv_reopen_prepare = bdrv_qed_reopen_prepare,
1624 .bdrv_child_perm = bdrv_default_perms,
1625 .bdrv_co_create = bdrv_qed_co_create,
1626 .bdrv_co_create_opts = bdrv_qed_co_create_opts,
1627 .bdrv_has_zero_init = bdrv_has_zero_init_1,
1628 .bdrv_co_block_status = bdrv_qed_co_block_status,
1629 .bdrv_co_readv = bdrv_qed_co_readv,
1630 .bdrv_co_writev = bdrv_qed_co_writev,
1631 .bdrv_co_pwrite_zeroes = bdrv_qed_co_pwrite_zeroes,
1632 .bdrv_co_truncate = bdrv_qed_co_truncate,
1633 .bdrv_getlength = bdrv_qed_getlength,
1634 .bdrv_get_info = bdrv_qed_get_info,
1635 .bdrv_refresh_limits = bdrv_qed_refresh_limits,
1636 .bdrv_change_backing_file = bdrv_qed_change_backing_file,
1637 .bdrv_co_invalidate_cache = bdrv_qed_co_invalidate_cache,
1638 .bdrv_co_check = bdrv_qed_co_check,
1639 .bdrv_detach_aio_context = bdrv_qed_detach_aio_context,
1640 .bdrv_attach_aio_context = bdrv_qed_attach_aio_context,
1641 .bdrv_co_drain_begin = bdrv_qed_co_drain_begin,
1644 static void bdrv_qed_init(void)
1646 bdrv_register(&bdrv_qed);
1649 block_init(bdrv_qed_init);