xref: /openbmc/qemu/block/qed.c (revision 500eb6db)
1 /*
2  * QEMU Enhanced Disk Format
3  *
4  * Copyright IBM, Corp. 2010
5  *
6  * Authors:
7  *  Stefan Hajnoczi   <stefanha@linux.vnet.ibm.com>
8  *  Anthony Liguori   <aliguori@us.ibm.com>
9  *
10  * This work is licensed under the terms of the GNU LGPL, version 2 or later.
11  * See the COPYING.LIB file in the top-level directory.
12  *
13  */
14 
15 #include "qemu/osdep.h"
16 #include "block/qdict.h"
17 #include "qapi/error.h"
18 #include "qemu/timer.h"
19 #include "qemu/bswap.h"
20 #include "qemu/module.h"
21 #include "qemu/option.h"
22 #include "trace.h"
23 #include "qed.h"
24 #include "sysemu/block-backend.h"
25 #include "qapi/qmp/qdict.h"
26 #include "qapi/qobject-input-visitor.h"
27 #include "qapi/qapi-visit-block-core.h"
28 
29 static QemuOptsList qed_create_opts;
30 
31 static int bdrv_qed_probe(const uint8_t *buf, int buf_size,
32                           const char *filename)
33 {
34     const QEDHeader *header = (const QEDHeader *)buf;
35 
36     if (buf_size < sizeof(*header)) {
37         return 0;
38     }
39     if (le32_to_cpu(header->magic) != QED_MAGIC) {
40         return 0;
41     }
42     return 100;
43 }
44 
45 /**
46  * Check whether an image format is raw
47  *
48  * @fmt:    Backing file format, may be NULL
49  */
50 static bool qed_fmt_is_raw(const char *fmt)
51 {
52     return fmt && strcmp(fmt, "raw") == 0;
53 }
54 
55 static void qed_header_le_to_cpu(const QEDHeader *le, QEDHeader *cpu)
56 {
57     cpu->magic = le32_to_cpu(le->magic);
58     cpu->cluster_size = le32_to_cpu(le->cluster_size);
59     cpu->table_size = le32_to_cpu(le->table_size);
60     cpu->header_size = le32_to_cpu(le->header_size);
61     cpu->features = le64_to_cpu(le->features);
62     cpu->compat_features = le64_to_cpu(le->compat_features);
63     cpu->autoclear_features = le64_to_cpu(le->autoclear_features);
64     cpu->l1_table_offset = le64_to_cpu(le->l1_table_offset);
65     cpu->image_size = le64_to_cpu(le->image_size);
66     cpu->backing_filename_offset = le32_to_cpu(le->backing_filename_offset);
67     cpu->backing_filename_size = le32_to_cpu(le->backing_filename_size);
68 }
69 
70 static void qed_header_cpu_to_le(const QEDHeader *cpu, QEDHeader *le)
71 {
72     le->magic = cpu_to_le32(cpu->magic);
73     le->cluster_size = cpu_to_le32(cpu->cluster_size);
74     le->table_size = cpu_to_le32(cpu->table_size);
75     le->header_size = cpu_to_le32(cpu->header_size);
76     le->features = cpu_to_le64(cpu->features);
77     le->compat_features = cpu_to_le64(cpu->compat_features);
78     le->autoclear_features = cpu_to_le64(cpu->autoclear_features);
79     le->l1_table_offset = cpu_to_le64(cpu->l1_table_offset);
80     le->image_size = cpu_to_le64(cpu->image_size);
81     le->backing_filename_offset = cpu_to_le32(cpu->backing_filename_offset);
82     le->backing_filename_size = cpu_to_le32(cpu->backing_filename_size);
83 }
84 
85 int qed_write_header_sync(BDRVQEDState *s)
86 {
87     QEDHeader le;
88     int ret;
89 
90     qed_header_cpu_to_le(&s->header, &le);
91     ret = bdrv_pwrite(s->bs->file, 0, &le, sizeof(le));
92     if (ret != sizeof(le)) {
93         return ret;
94     }
95     return 0;
96 }
97 
98 /**
99  * Update header in-place (does not rewrite backing filename or other strings)
100  *
101  * This function only updates known header fields in-place and does not affect
102  * extra data after the QED header.
103  *
104  * No new allocating reqs can start while this function runs.
105  */
106 static int coroutine_fn qed_write_header(BDRVQEDState *s)
107 {
108     /* We must write full sectors for O_DIRECT but cannot necessarily generate
109      * the data following the header if an unrecognized compat feature is
110      * active.  Therefore, first read the sectors containing the header, update
111      * them, and write back.
112      */
113 
114     int nsectors = DIV_ROUND_UP(sizeof(QEDHeader), BDRV_SECTOR_SIZE);
115     size_t len = nsectors * BDRV_SECTOR_SIZE;
116     uint8_t *buf;
117     int ret;
118 
119     assert(s->allocating_acb || s->allocating_write_reqs_plugged);
120 
121     buf = qemu_blockalign(s->bs, len);
122 
123     ret = bdrv_co_pread(s->bs->file, 0, len, buf, 0);
124     if (ret < 0) {
125         goto out;
126     }
127 
128     /* Update header */
129     qed_header_cpu_to_le(&s->header, (QEDHeader *) buf);
130 
131     ret = bdrv_co_pwrite(s->bs->file, 0, len,  buf, 0);
132     if (ret < 0) {
133         goto out;
134     }
135 
136     ret = 0;
137 out:
138     qemu_vfree(buf);
139     return ret;
140 }
141 
142 static uint64_t qed_max_image_size(uint32_t cluster_size, uint32_t table_size)
143 {
144     uint64_t table_entries;
145     uint64_t l2_size;
146 
147     table_entries = (table_size * cluster_size) / sizeof(uint64_t);
148     l2_size = table_entries * cluster_size;
149 
150     return l2_size * table_entries;
151 }
152 
153 static bool qed_is_cluster_size_valid(uint32_t cluster_size)
154 {
155     if (cluster_size < QED_MIN_CLUSTER_SIZE ||
156         cluster_size > QED_MAX_CLUSTER_SIZE) {
157         return false;
158     }
159     if (cluster_size & (cluster_size - 1)) {
160         return false; /* not power of 2 */
161     }
162     return true;
163 }
164 
165 static bool qed_is_table_size_valid(uint32_t table_size)
166 {
167     if (table_size < QED_MIN_TABLE_SIZE ||
168         table_size > QED_MAX_TABLE_SIZE) {
169         return false;
170     }
171     if (table_size & (table_size - 1)) {
172         return false; /* not power of 2 */
173     }
174     return true;
175 }
176 
177 static bool qed_is_image_size_valid(uint64_t image_size, uint32_t cluster_size,
178                                     uint32_t table_size)
179 {
180     if (image_size % BDRV_SECTOR_SIZE != 0) {
181         return false; /* not multiple of sector size */
182     }
183     if (image_size > qed_max_image_size(cluster_size, table_size)) {
184         return false; /* image is too large */
185     }
186     return true;
187 }
188 
189 /**
190  * Read a string of known length from the image file
191  *
192  * @file:       Image file
193  * @offset:     File offset to start of string, in bytes
194  * @n:          String length in bytes
195  * @buf:        Destination buffer
196  * @buflen:     Destination buffer length in bytes
197  * @ret:        0 on success, -errno on failure
198  *
199  * The string is NUL-terminated.
200  */
201 static int qed_read_string(BdrvChild *file, uint64_t offset, size_t n,
202                            char *buf, size_t buflen)
203 {
204     int ret;
205     if (n >= buflen) {
206         return -EINVAL;
207     }
208     ret = bdrv_pread(file, offset, buf, n);
209     if (ret < 0) {
210         return ret;
211     }
212     buf[n] = '\0';
213     return 0;
214 }
215 
216 /**
217  * Allocate new clusters
218  *
219  * @s:          QED state
220  * @n:          Number of contiguous clusters to allocate
221  * @ret:        Offset of first allocated cluster
222  *
223  * This function only produces the offset where the new clusters should be
224  * written.  It updates BDRVQEDState but does not make any changes to the image
225  * file.
226  *
227  * Called with table_lock held.
228  */
229 static uint64_t qed_alloc_clusters(BDRVQEDState *s, unsigned int n)
230 {
231     uint64_t offset = s->file_size;
232     s->file_size += n * s->header.cluster_size;
233     return offset;
234 }
235 
236 QEDTable *qed_alloc_table(BDRVQEDState *s)
237 {
238     /* Honor O_DIRECT memory alignment requirements */
239     return qemu_blockalign(s->bs,
240                            s->header.cluster_size * s->header.table_size);
241 }
242 
243 /**
244  * Allocate a new zeroed L2 table
245  *
246  * Called with table_lock held.
247  */
248 static CachedL2Table *qed_new_l2_table(BDRVQEDState *s)
249 {
250     CachedL2Table *l2_table = qed_alloc_l2_cache_entry(&s->l2_cache);
251 
252     l2_table->table = qed_alloc_table(s);
253     l2_table->offset = qed_alloc_clusters(s, s->header.table_size);
254 
255     memset(l2_table->table->offsets, 0,
256            s->header.cluster_size * s->header.table_size);
257     return l2_table;
258 }
259 
260 static bool qed_plug_allocating_write_reqs(BDRVQEDState *s)
261 {
262     qemu_co_mutex_lock(&s->table_lock);
263 
264     /* No reentrancy is allowed.  */
265     assert(!s->allocating_write_reqs_plugged);
266     if (s->allocating_acb != NULL) {
267         /* Another allocating write came concurrently.  This cannot happen
268          * from bdrv_qed_co_drain_begin, but it can happen when the timer runs.
269          */
270         qemu_co_mutex_unlock(&s->table_lock);
271         return false;
272     }
273 
274     s->allocating_write_reqs_plugged = true;
275     qemu_co_mutex_unlock(&s->table_lock);
276     return true;
277 }
278 
279 static void qed_unplug_allocating_write_reqs(BDRVQEDState *s)
280 {
281     qemu_co_mutex_lock(&s->table_lock);
282     assert(s->allocating_write_reqs_plugged);
283     s->allocating_write_reqs_plugged = false;
284     qemu_co_queue_next(&s->allocating_write_reqs);
285     qemu_co_mutex_unlock(&s->table_lock);
286 }
287 
288 static void coroutine_fn qed_need_check_timer_entry(void *opaque)
289 {
290     BDRVQEDState *s = opaque;
291     int ret;
292 
293     trace_qed_need_check_timer_cb(s);
294 
295     if (!qed_plug_allocating_write_reqs(s)) {
296         return;
297     }
298 
299     /* Ensure writes are on disk before clearing flag */
300     ret = bdrv_co_flush(s->bs->file->bs);
301     if (ret < 0) {
302         qed_unplug_allocating_write_reqs(s);
303         return;
304     }
305 
306     s->header.features &= ~QED_F_NEED_CHECK;
307     ret = qed_write_header(s);
308     (void) ret;
309 
310     qed_unplug_allocating_write_reqs(s);
311 
312     ret = bdrv_co_flush(s->bs);
313     (void) ret;
314 }
315 
316 static void qed_need_check_timer_cb(void *opaque)
317 {
318     Coroutine *co = qemu_coroutine_create(qed_need_check_timer_entry, opaque);
319     qemu_coroutine_enter(co);
320 }
321 
322 static void qed_start_need_check_timer(BDRVQEDState *s)
323 {
324     trace_qed_start_need_check_timer(s);
325 
326     /* Use QEMU_CLOCK_VIRTUAL so we don't alter the image file while suspended for
327      * migration.
328      */
329     timer_mod(s->need_check_timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
330                    NANOSECONDS_PER_SECOND * QED_NEED_CHECK_TIMEOUT);
331 }
332 
333 /* It's okay to call this multiple times or when no timer is started */
334 static void qed_cancel_need_check_timer(BDRVQEDState *s)
335 {
336     trace_qed_cancel_need_check_timer(s);
337     timer_del(s->need_check_timer);
338 }
339 
340 static void bdrv_qed_detach_aio_context(BlockDriverState *bs)
341 {
342     BDRVQEDState *s = bs->opaque;
343 
344     qed_cancel_need_check_timer(s);
345     timer_free(s->need_check_timer);
346 }
347 
348 static void bdrv_qed_attach_aio_context(BlockDriverState *bs,
349                                         AioContext *new_context)
350 {
351     BDRVQEDState *s = bs->opaque;
352 
353     s->need_check_timer = aio_timer_new(new_context,
354                                         QEMU_CLOCK_VIRTUAL, SCALE_NS,
355                                         qed_need_check_timer_cb, s);
356     if (s->header.features & QED_F_NEED_CHECK) {
357         qed_start_need_check_timer(s);
358     }
359 }
360 
361 static void coroutine_fn bdrv_qed_co_drain_begin(BlockDriverState *bs)
362 {
363     BDRVQEDState *s = bs->opaque;
364 
365     /* Fire the timer immediately in order to start doing I/O as soon as the
366      * header is flushed.
367      */
368     if (s->need_check_timer && timer_pending(s->need_check_timer)) {
369         qed_cancel_need_check_timer(s);
370         qed_need_check_timer_entry(s);
371     }
372 }
373 
374 static void bdrv_qed_init_state(BlockDriverState *bs)
375 {
376     BDRVQEDState *s = bs->opaque;
377 
378     memset(s, 0, sizeof(BDRVQEDState));
379     s->bs = bs;
380     qemu_co_mutex_init(&s->table_lock);
381     qemu_co_queue_init(&s->allocating_write_reqs);
382 }
383 
384 /* Called with table_lock held.  */
385 static int coroutine_fn bdrv_qed_do_open(BlockDriverState *bs, QDict *options,
386                                          int flags, Error **errp)
387 {
388     BDRVQEDState *s = bs->opaque;
389     QEDHeader le_header;
390     int64_t file_size;
391     int ret;
392 
393     ret = bdrv_pread(bs->file, 0, &le_header, sizeof(le_header));
394     if (ret < 0) {
395         return ret;
396     }
397     qed_header_le_to_cpu(&le_header, &s->header);
398 
399     if (s->header.magic != QED_MAGIC) {
400         error_setg(errp, "Image not in QED format");
401         return -EINVAL;
402     }
403     if (s->header.features & ~QED_FEATURE_MASK) {
404         /* image uses unsupported feature bits */
405         error_setg(errp, "Unsupported QED features: %" PRIx64,
406                    s->header.features & ~QED_FEATURE_MASK);
407         return -ENOTSUP;
408     }
409     if (!qed_is_cluster_size_valid(s->header.cluster_size)) {
410         return -EINVAL;
411     }
412 
413     /* Round down file size to the last cluster */
414     file_size = bdrv_getlength(bs->file->bs);
415     if (file_size < 0) {
416         return file_size;
417     }
418     s->file_size = qed_start_of_cluster(s, file_size);
419 
420     if (!qed_is_table_size_valid(s->header.table_size)) {
421         return -EINVAL;
422     }
423     if (!qed_is_image_size_valid(s->header.image_size,
424                                  s->header.cluster_size,
425                                  s->header.table_size)) {
426         return -EINVAL;
427     }
428     if (!qed_check_table_offset(s, s->header.l1_table_offset)) {
429         return -EINVAL;
430     }
431 
432     s->table_nelems = (s->header.cluster_size * s->header.table_size) /
433                       sizeof(uint64_t);
434     s->l2_shift = ctz32(s->header.cluster_size);
435     s->l2_mask = s->table_nelems - 1;
436     s->l1_shift = s->l2_shift + ctz32(s->table_nelems);
437 
438     /* Header size calculation must not overflow uint32_t */
439     if (s->header.header_size > UINT32_MAX / s->header.cluster_size) {
440         return -EINVAL;
441     }
442 
443     if ((s->header.features & QED_F_BACKING_FILE)) {
444         if ((uint64_t)s->header.backing_filename_offset +
445             s->header.backing_filename_size >
446             s->header.cluster_size * s->header.header_size) {
447             return -EINVAL;
448         }
449 
450         ret = qed_read_string(bs->file, s->header.backing_filename_offset,
451                               s->header.backing_filename_size,
452                               bs->auto_backing_file,
453                               sizeof(bs->auto_backing_file));
454         if (ret < 0) {
455             return ret;
456         }
457         pstrcpy(bs->backing_file, sizeof(bs->backing_file),
458                 bs->auto_backing_file);
459 
460         if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) {
461             pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw");
462         }
463     }
464 
465     /* Reset unknown autoclear feature bits.  This is a backwards
466      * compatibility mechanism that allows images to be opened by older
467      * programs, which "knock out" unknown feature bits.  When an image is
468      * opened by a newer program again it can detect that the autoclear
469      * feature is no longer valid.
470      */
471     if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 &&
472         !bdrv_is_read_only(bs->file->bs) && !(flags & BDRV_O_INACTIVE)) {
473         s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK;
474 
475         ret = qed_write_header_sync(s);
476         if (ret) {
477             return ret;
478         }
479 
480         /* From here on only known autoclear feature bits are valid */
481         bdrv_flush(bs->file->bs);
482     }
483 
484     s->l1_table = qed_alloc_table(s);
485     qed_init_l2_cache(&s->l2_cache);
486 
487     ret = qed_read_l1_table_sync(s);
488     if (ret) {
489         goto out;
490     }
491 
492     /* If image was not closed cleanly, check consistency */
493     if (!(flags & BDRV_O_CHECK) && (s->header.features & QED_F_NEED_CHECK)) {
494         /* Read-only images cannot be fixed.  There is no risk of corruption
495          * since write operations are not possible.  Therefore, allow
496          * potentially inconsistent images to be opened read-only.  This can
497          * aid data recovery from an otherwise inconsistent image.
498          */
499         if (!bdrv_is_read_only(bs->file->bs) &&
500             !(flags & BDRV_O_INACTIVE)) {
501             BdrvCheckResult result = {0};
502 
503             ret = qed_check(s, &result, true);
504             if (ret) {
505                 goto out;
506             }
507         }
508     }
509 
510     bdrv_qed_attach_aio_context(bs, bdrv_get_aio_context(bs));
511 
512 out:
513     if (ret) {
514         qed_free_l2_cache(&s->l2_cache);
515         qemu_vfree(s->l1_table);
516     }
517     return ret;
518 }
519 
520 typedef struct QEDOpenCo {
521     BlockDriverState *bs;
522     QDict *options;
523     int flags;
524     Error **errp;
525     int ret;
526 } QEDOpenCo;
527 
528 static void coroutine_fn bdrv_qed_open_entry(void *opaque)
529 {
530     QEDOpenCo *qoc = opaque;
531     BDRVQEDState *s = qoc->bs->opaque;
532 
533     qemu_co_mutex_lock(&s->table_lock);
534     qoc->ret = bdrv_qed_do_open(qoc->bs, qoc->options, qoc->flags, qoc->errp);
535     qemu_co_mutex_unlock(&s->table_lock);
536 }
537 
538 static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags,
539                          Error **errp)
540 {
541     QEDOpenCo qoc = {
542         .bs = bs,
543         .options = options,
544         .flags = flags,
545         .errp = errp,
546         .ret = -EINPROGRESS
547     };
548 
549     bs->file = bdrv_open_child(NULL, options, "file", bs, &child_file,
550                                false, errp);
551     if (!bs->file) {
552         return -EINVAL;
553     }
554 
555     bdrv_qed_init_state(bs);
556     if (qemu_in_coroutine()) {
557         bdrv_qed_open_entry(&qoc);
558     } else {
559         assert(qemu_get_current_aio_context() == qemu_get_aio_context());
560         qemu_coroutine_enter(qemu_coroutine_create(bdrv_qed_open_entry, &qoc));
561         BDRV_POLL_WHILE(bs, qoc.ret == -EINPROGRESS);
562     }
563     BDRV_POLL_WHILE(bs, qoc.ret == -EINPROGRESS);
564     return qoc.ret;
565 }
566 
567 static void bdrv_qed_refresh_limits(BlockDriverState *bs, Error **errp)
568 {
569     BDRVQEDState *s = bs->opaque;
570 
571     bs->bl.pwrite_zeroes_alignment = s->header.cluster_size;
572 }
573 
574 /* We have nothing to do for QED reopen, stubs just return
575  * success */
576 static int bdrv_qed_reopen_prepare(BDRVReopenState *state,
577                                    BlockReopenQueue *queue, Error **errp)
578 {
579     return 0;
580 }
581 
582 static void bdrv_qed_close(BlockDriverState *bs)
583 {
584     BDRVQEDState *s = bs->opaque;
585 
586     bdrv_qed_detach_aio_context(bs);
587 
588     /* Ensure writes reach stable storage */
589     bdrv_flush(bs->file->bs);
590 
591     /* Clean shutdown, no check required on next open */
592     if (s->header.features & QED_F_NEED_CHECK) {
593         s->header.features &= ~QED_F_NEED_CHECK;
594         qed_write_header_sync(s);
595     }
596 
597     qed_free_l2_cache(&s->l2_cache);
598     qemu_vfree(s->l1_table);
599 }
600 
601 static int coroutine_fn bdrv_qed_co_create(BlockdevCreateOptions *opts,
602                                            Error **errp)
603 {
604     BlockdevCreateOptionsQed *qed_opts;
605     BlockBackend *blk = NULL;
606     BlockDriverState *bs = NULL;
607 
608     QEDHeader header;
609     QEDHeader le_header;
610     uint8_t *l1_table = NULL;
611     size_t l1_size;
612     int ret = 0;
613 
614     assert(opts->driver == BLOCKDEV_DRIVER_QED);
615     qed_opts = &opts->u.qed;
616 
617     /* Validate options and set default values */
618     if (!qed_opts->has_cluster_size) {
619         qed_opts->cluster_size = QED_DEFAULT_CLUSTER_SIZE;
620     }
621     if (!qed_opts->has_table_size) {
622         qed_opts->table_size = QED_DEFAULT_TABLE_SIZE;
623     }
624 
625     if (!qed_is_cluster_size_valid(qed_opts->cluster_size)) {
626         error_setg(errp, "QED cluster size must be within range [%u, %u] "
627                          "and power of 2",
628                    QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE);
629         return -EINVAL;
630     }
631     if (!qed_is_table_size_valid(qed_opts->table_size)) {
632         error_setg(errp, "QED table size must be within range [%u, %u] "
633                          "and power of 2",
634                    QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE);
635         return -EINVAL;
636     }
637     if (!qed_is_image_size_valid(qed_opts->size, qed_opts->cluster_size,
638                                  qed_opts->table_size))
639     {
640         error_setg(errp, "QED image size must be a non-zero multiple of "
641                          "cluster size and less than %" PRIu64 " bytes",
642                    qed_max_image_size(qed_opts->cluster_size,
643                                       qed_opts->table_size));
644         return -EINVAL;
645     }
646 
647     /* Create BlockBackend to write to the image */
648     bs = bdrv_open_blockdev_ref(qed_opts->file, errp);
649     if (bs == NULL) {
650         return -EIO;
651     }
652 
653     blk = blk_new(bdrv_get_aio_context(bs),
654                   BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
655     ret = blk_insert_bs(blk, bs, errp);
656     if (ret < 0) {
657         goto out;
658     }
659     blk_set_allow_write_beyond_eof(blk, true);
660 
661     /* Prepare image format */
662     header = (QEDHeader) {
663         .magic = QED_MAGIC,
664         .cluster_size = qed_opts->cluster_size,
665         .table_size = qed_opts->table_size,
666         .header_size = 1,
667         .features = 0,
668         .compat_features = 0,
669         .l1_table_offset = qed_opts->cluster_size,
670         .image_size = qed_opts->size,
671     };
672 
673     l1_size = header.cluster_size * header.table_size;
674 
675     /* File must start empty and grow, check truncate is supported */
676     ret = blk_truncate(blk, 0, PREALLOC_MODE_OFF, errp);
677     if (ret < 0) {
678         goto out;
679     }
680 
681     if (qed_opts->has_backing_file) {
682         header.features |= QED_F_BACKING_FILE;
683         header.backing_filename_offset = sizeof(le_header);
684         header.backing_filename_size = strlen(qed_opts->backing_file);
685 
686         if (qed_opts->has_backing_fmt) {
687             const char *backing_fmt = BlockdevDriver_str(qed_opts->backing_fmt);
688             if (qed_fmt_is_raw(backing_fmt)) {
689                 header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
690             }
691         }
692     }
693 
694     qed_header_cpu_to_le(&header, &le_header);
695     ret = blk_pwrite(blk, 0, &le_header, sizeof(le_header), 0);
696     if (ret < 0) {
697         goto out;
698     }
699     ret = blk_pwrite(blk, sizeof(le_header), qed_opts->backing_file,
700                      header.backing_filename_size, 0);
701     if (ret < 0) {
702         goto out;
703     }
704 
705     l1_table = g_malloc0(l1_size);
706     ret = blk_pwrite(blk, header.l1_table_offset, l1_table, l1_size, 0);
707     if (ret < 0) {
708         goto out;
709     }
710 
711     ret = 0; /* success */
712 out:
713     g_free(l1_table);
714     blk_unref(blk);
715     bdrv_unref(bs);
716     return ret;
717 }
718 
719 static int coroutine_fn bdrv_qed_co_create_opts(const char *filename,
720                                                 QemuOpts *opts,
721                                                 Error **errp)
722 {
723     BlockdevCreateOptions *create_options = NULL;
724     QDict *qdict;
725     Visitor *v;
726     BlockDriverState *bs = NULL;
727     Error *local_err = NULL;
728     int ret;
729 
730     static const QDictRenames opt_renames[] = {
731         { BLOCK_OPT_BACKING_FILE,       "backing-file" },
732         { BLOCK_OPT_BACKING_FMT,        "backing-fmt" },
733         { BLOCK_OPT_CLUSTER_SIZE,       "cluster-size" },
734         { BLOCK_OPT_TABLE_SIZE,         "table-size" },
735         { NULL, NULL },
736     };
737 
738     /* Parse options and convert legacy syntax */
739     qdict = qemu_opts_to_qdict_filtered(opts, NULL, &qed_create_opts, true);
740 
741     if (!qdict_rename_keys(qdict, opt_renames, errp)) {
742         ret = -EINVAL;
743         goto fail;
744     }
745 
746     /* Create and open the file (protocol layer) */
747     ret = bdrv_create_file(filename, opts, &local_err);
748     if (ret < 0) {
749         error_propagate(errp, local_err);
750         goto fail;
751     }
752 
753     bs = bdrv_open(filename, NULL, NULL,
754                    BDRV_O_RDWR | BDRV_O_RESIZE | BDRV_O_PROTOCOL, errp);
755     if (bs == NULL) {
756         ret = -EIO;
757         goto fail;
758     }
759 
760     /* Now get the QAPI type BlockdevCreateOptions */
761     qdict_put_str(qdict, "driver", "qed");
762     qdict_put_str(qdict, "file", bs->node_name);
763 
764     v = qobject_input_visitor_new_flat_confused(qdict, errp);
765     if (!v) {
766         ret = -EINVAL;
767         goto fail;
768     }
769 
770     visit_type_BlockdevCreateOptions(v, NULL, &create_options, &local_err);
771     visit_free(v);
772 
773     if (local_err) {
774         error_propagate(errp, local_err);
775         ret = -EINVAL;
776         goto fail;
777     }
778 
779     /* Silently round up size */
780     assert(create_options->driver == BLOCKDEV_DRIVER_QED);
781     create_options->u.qed.size =
782         ROUND_UP(create_options->u.qed.size, BDRV_SECTOR_SIZE);
783 
784     /* Create the qed image (format layer) */
785     ret = bdrv_qed_co_create(create_options, errp);
786 
787 fail:
788     qobject_unref(qdict);
789     bdrv_unref(bs);
790     qapi_free_BlockdevCreateOptions(create_options);
791     return ret;
792 }
793 
794 static int coroutine_fn bdrv_qed_co_block_status(BlockDriverState *bs,
795                                                  bool want_zero,
796                                                  int64_t pos, int64_t bytes,
797                                                  int64_t *pnum, int64_t *map,
798                                                  BlockDriverState **file)
799 {
800     BDRVQEDState *s = bs->opaque;
801     size_t len = MIN(bytes, SIZE_MAX);
802     int status;
803     QEDRequest request = { .l2_table = NULL };
804     uint64_t offset;
805     int ret;
806 
807     qemu_co_mutex_lock(&s->table_lock);
808     ret = qed_find_cluster(s, &request, pos, &len, &offset);
809 
810     *pnum = len;
811     switch (ret) {
812     case QED_CLUSTER_FOUND:
813         *map = offset | qed_offset_into_cluster(s, pos);
814         status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
815         *file = bs->file->bs;
816         break;
817     case QED_CLUSTER_ZERO:
818         status = BDRV_BLOCK_ZERO;
819         break;
820     case QED_CLUSTER_L2:
821     case QED_CLUSTER_L1:
822         status = 0;
823         break;
824     default:
825         assert(ret < 0);
826         status = ret;
827         break;
828     }
829 
830     qed_unref_l2_cache_entry(request.l2_table);
831     qemu_co_mutex_unlock(&s->table_lock);
832 
833     return status;
834 }
835 
836 static BDRVQEDState *acb_to_s(QEDAIOCB *acb)
837 {
838     return acb->bs->opaque;
839 }
840 
841 /**
842  * Read from the backing file or zero-fill if no backing file
843  *
844  * @s:              QED state
845  * @pos:            Byte position in device
846  * @qiov:           Destination I/O vector
847  * @backing_qiov:   Possibly shortened copy of qiov, to be allocated here
848  * @cb:             Completion function
849  * @opaque:         User data for completion function
850  *
851  * This function reads qiov->size bytes starting at pos from the backing file.
852  * If there is no backing file then zeroes are read.
853  */
854 static int coroutine_fn qed_read_backing_file(BDRVQEDState *s, uint64_t pos,
855                                               QEMUIOVector *qiov,
856                                               QEMUIOVector **backing_qiov)
857 {
858     uint64_t backing_length = 0;
859     size_t size;
860     int ret;
861 
862     /* If there is a backing file, get its length.  Treat the absence of a
863      * backing file like a zero length backing file.
864      */
865     if (s->bs->backing) {
866         int64_t l = bdrv_getlength(s->bs->backing->bs);
867         if (l < 0) {
868             return l;
869         }
870         backing_length = l;
871     }
872 
873     /* Zero all sectors if reading beyond the end of the backing file */
874     if (pos >= backing_length ||
875         pos + qiov->size > backing_length) {
876         qemu_iovec_memset(qiov, 0, 0, qiov->size);
877     }
878 
879     /* Complete now if there are no backing file sectors to read */
880     if (pos >= backing_length) {
881         return 0;
882     }
883 
884     /* If the read straddles the end of the backing file, shorten it */
885     size = MIN((uint64_t)backing_length - pos, qiov->size);
886 
887     assert(*backing_qiov == NULL);
888     *backing_qiov = g_new(QEMUIOVector, 1);
889     qemu_iovec_init(*backing_qiov, qiov->niov);
890     qemu_iovec_concat(*backing_qiov, qiov, 0, size);
891 
892     BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO);
893     ret = bdrv_co_preadv(s->bs->backing, pos, size, *backing_qiov, 0);
894     if (ret < 0) {
895         return ret;
896     }
897     return 0;
898 }
899 
900 /**
901  * Copy data from backing file into the image
902  *
903  * @s:          QED state
904  * @pos:        Byte position in device
905  * @len:        Number of bytes
906  * @offset:     Byte offset in image file
907  */
908 static int coroutine_fn qed_copy_from_backing_file(BDRVQEDState *s,
909                                                    uint64_t pos, uint64_t len,
910                                                    uint64_t offset)
911 {
912     QEMUIOVector qiov;
913     QEMUIOVector *backing_qiov = NULL;
914     int ret;
915 
916     /* Skip copy entirely if there is no work to do */
917     if (len == 0) {
918         return 0;
919     }
920 
921     qemu_iovec_init_buf(&qiov, qemu_blockalign(s->bs, len), len);
922 
923     ret = qed_read_backing_file(s, pos, &qiov, &backing_qiov);
924 
925     if (backing_qiov) {
926         qemu_iovec_destroy(backing_qiov);
927         g_free(backing_qiov);
928         backing_qiov = NULL;
929     }
930 
931     if (ret) {
932         goto out;
933     }
934 
935     BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE);
936     ret = bdrv_co_pwritev(s->bs->file, offset, qiov.size, &qiov, 0);
937     if (ret < 0) {
938         goto out;
939     }
940     ret = 0;
941 out:
942     qemu_vfree(qemu_iovec_buf(&qiov));
943     return ret;
944 }
945 
946 /**
947  * Link one or more contiguous clusters into a table
948  *
949  * @s:              QED state
950  * @table:          L2 table
951  * @index:          First cluster index
952  * @n:              Number of contiguous clusters
953  * @cluster:        First cluster offset
954  *
955  * The cluster offset may be an allocated byte offset in the image file, the
956  * zero cluster marker, or the unallocated cluster marker.
957  *
958  * Called with table_lock held.
959  */
960 static void coroutine_fn qed_update_l2_table(BDRVQEDState *s, QEDTable *table,
961                                              int index, unsigned int n,
962                                              uint64_t cluster)
963 {
964     int i;
965     for (i = index; i < index + n; i++) {
966         table->offsets[i] = cluster;
967         if (!qed_offset_is_unalloc_cluster(cluster) &&
968             !qed_offset_is_zero_cluster(cluster)) {
969             cluster += s->header.cluster_size;
970         }
971     }
972 }
973 
974 /* Called with table_lock held.  */
975 static void coroutine_fn qed_aio_complete(QEDAIOCB *acb)
976 {
977     BDRVQEDState *s = acb_to_s(acb);
978 
979     /* Free resources */
980     qemu_iovec_destroy(&acb->cur_qiov);
981     qed_unref_l2_cache_entry(acb->request.l2_table);
982 
983     /* Free the buffer we may have allocated for zero writes */
984     if (acb->flags & QED_AIOCB_ZERO) {
985         qemu_vfree(acb->qiov->iov[0].iov_base);
986         acb->qiov->iov[0].iov_base = NULL;
987     }
988 
989     /* Start next allocating write request waiting behind this one.  Note that
990      * requests enqueue themselves when they first hit an unallocated cluster
991      * but they wait until the entire request is finished before waking up the
992      * next request in the queue.  This ensures that we don't cycle through
993      * requests multiple times but rather finish one at a time completely.
994      */
995     if (acb == s->allocating_acb) {
996         s->allocating_acb = NULL;
997         if (!qemu_co_queue_empty(&s->allocating_write_reqs)) {
998             qemu_co_queue_next(&s->allocating_write_reqs);
999         } else if (s->header.features & QED_F_NEED_CHECK) {
1000             qed_start_need_check_timer(s);
1001         }
1002     }
1003 }
1004 
1005 /**
1006  * Update L1 table with new L2 table offset and write it out
1007  *
1008  * Called with table_lock held.
1009  */
1010 static int coroutine_fn qed_aio_write_l1_update(QEDAIOCB *acb)
1011 {
1012     BDRVQEDState *s = acb_to_s(acb);
1013     CachedL2Table *l2_table = acb->request.l2_table;
1014     uint64_t l2_offset = l2_table->offset;
1015     int index, ret;
1016 
1017     index = qed_l1_index(s, acb->cur_pos);
1018     s->l1_table->offsets[index] = l2_table->offset;
1019 
1020     ret = qed_write_l1_table(s, index, 1);
1021 
1022     /* Commit the current L2 table to the cache */
1023     qed_commit_l2_cache_entry(&s->l2_cache, l2_table);
1024 
1025     /* This is guaranteed to succeed because we just committed the entry to the
1026      * cache.
1027      */
1028     acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset);
1029     assert(acb->request.l2_table != NULL);
1030 
1031     return ret;
1032 }
1033 
1034 
1035 /**
1036  * Update L2 table with new cluster offsets and write them out
1037  *
1038  * Called with table_lock held.
1039  */
1040 static int coroutine_fn qed_aio_write_l2_update(QEDAIOCB *acb, uint64_t offset)
1041 {
1042     BDRVQEDState *s = acb_to_s(acb);
1043     bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1;
1044     int index, ret;
1045 
1046     if (need_alloc) {
1047         qed_unref_l2_cache_entry(acb->request.l2_table);
1048         acb->request.l2_table = qed_new_l2_table(s);
1049     }
1050 
1051     index = qed_l2_index(s, acb->cur_pos);
1052     qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters,
1053                          offset);
1054 
1055     if (need_alloc) {
1056         /* Write out the whole new L2 table */
1057         ret = qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true);
1058         if (ret) {
1059             return ret;
1060         }
1061         return qed_aio_write_l1_update(acb);
1062     } else {
1063         /* Write out only the updated part of the L2 table */
1064         ret = qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters,
1065                                  false);
1066         if (ret) {
1067             return ret;
1068         }
1069     }
1070     return 0;
1071 }
1072 
1073 /**
1074  * Write data to the image file
1075  *
1076  * Called with table_lock *not* held.
1077  */
1078 static int coroutine_fn qed_aio_write_main(QEDAIOCB *acb)
1079 {
1080     BDRVQEDState *s = acb_to_s(acb);
1081     uint64_t offset = acb->cur_cluster +
1082                       qed_offset_into_cluster(s, acb->cur_pos);
1083 
1084     trace_qed_aio_write_main(s, acb, 0, offset, acb->cur_qiov.size);
1085 
1086     BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO);
1087     return bdrv_co_pwritev(s->bs->file, offset, acb->cur_qiov.size,
1088                            &acb->cur_qiov, 0);
1089 }
1090 
1091 /**
1092  * Populate untouched regions of new data cluster
1093  *
1094  * Called with table_lock held.
1095  */
1096 static int coroutine_fn qed_aio_write_cow(QEDAIOCB *acb)
1097 {
1098     BDRVQEDState *s = acb_to_s(acb);
1099     uint64_t start, len, offset;
1100     int ret;
1101 
1102     qemu_co_mutex_unlock(&s->table_lock);
1103 
1104     /* Populate front untouched region of new data cluster */
1105     start = qed_start_of_cluster(s, acb->cur_pos);
1106     len = qed_offset_into_cluster(s, acb->cur_pos);
1107 
1108     trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster);
1109     ret = qed_copy_from_backing_file(s, start, len, acb->cur_cluster);
1110     if (ret < 0) {
1111         goto out;
1112     }
1113 
1114     /* Populate back untouched region of new data cluster */
1115     start = acb->cur_pos + acb->cur_qiov.size;
1116     len = qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start;
1117     offset = acb->cur_cluster +
1118              qed_offset_into_cluster(s, acb->cur_pos) +
1119              acb->cur_qiov.size;
1120 
1121     trace_qed_aio_write_postfill(s, acb, start, len, offset);
1122     ret = qed_copy_from_backing_file(s, start, len, offset);
1123     if (ret < 0) {
1124         goto out;
1125     }
1126 
1127     ret = qed_aio_write_main(acb);
1128     if (ret < 0) {
1129         goto out;
1130     }
1131 
1132     if (s->bs->backing) {
1133         /*
1134          * Flush new data clusters before updating the L2 table
1135          *
1136          * This flush is necessary when a backing file is in use.  A crash
1137          * during an allocating write could result in empty clusters in the
1138          * image.  If the write only touched a subregion of the cluster,
1139          * then backing image sectors have been lost in the untouched
1140          * region.  The solution is to flush after writing a new data
1141          * cluster and before updating the L2 table.
1142          */
1143         ret = bdrv_co_flush(s->bs->file->bs);
1144     }
1145 
1146 out:
1147     qemu_co_mutex_lock(&s->table_lock);
1148     return ret;
1149 }
1150 
1151 /**
1152  * Check if the QED_F_NEED_CHECK bit should be set during allocating write
1153  */
1154 static bool qed_should_set_need_check(BDRVQEDState *s)
1155 {
1156     /* The flush before L2 update path ensures consistency */
1157     if (s->bs->backing) {
1158         return false;
1159     }
1160 
1161     return !(s->header.features & QED_F_NEED_CHECK);
1162 }
1163 
1164 /**
1165  * Write new data cluster
1166  *
1167  * @acb:        Write request
1168  * @len:        Length in bytes
1169  *
1170  * This path is taken when writing to previously unallocated clusters.
1171  *
1172  * Called with table_lock held.
1173  */
1174 static int coroutine_fn qed_aio_write_alloc(QEDAIOCB *acb, size_t len)
1175 {
1176     BDRVQEDState *s = acb_to_s(acb);
1177     int ret;
1178 
1179     /* Cancel timer when the first allocating request comes in */
1180     if (s->allocating_acb == NULL) {
1181         qed_cancel_need_check_timer(s);
1182     }
1183 
1184     /* Freeze this request if another allocating write is in progress */
1185     if (s->allocating_acb != acb || s->allocating_write_reqs_plugged) {
1186         if (s->allocating_acb != NULL) {
1187             qemu_co_queue_wait(&s->allocating_write_reqs, &s->table_lock);
1188             assert(s->allocating_acb == NULL);
1189         }
1190         s->allocating_acb = acb;
1191         return -EAGAIN; /* start over with looking up table entries */
1192     }
1193 
1194     acb->cur_nclusters = qed_bytes_to_clusters(s,
1195             qed_offset_into_cluster(s, acb->cur_pos) + len);
1196     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1197 
1198     if (acb->flags & QED_AIOCB_ZERO) {
1199         /* Skip ahead if the clusters are already zero */
1200         if (acb->find_cluster_ret == QED_CLUSTER_ZERO) {
1201             return 0;
1202         }
1203         acb->cur_cluster = 1;
1204     } else {
1205         acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters);
1206     }
1207 
1208     if (qed_should_set_need_check(s)) {
1209         s->header.features |= QED_F_NEED_CHECK;
1210         ret = qed_write_header(s);
1211         if (ret < 0) {
1212             return ret;
1213         }
1214     }
1215 
1216     if (!(acb->flags & QED_AIOCB_ZERO)) {
1217         ret = qed_aio_write_cow(acb);
1218         if (ret < 0) {
1219             return ret;
1220         }
1221     }
1222 
1223     return qed_aio_write_l2_update(acb, acb->cur_cluster);
1224 }
1225 
1226 /**
1227  * Write data cluster in place
1228  *
1229  * @acb:        Write request
1230  * @offset:     Cluster offset in bytes
1231  * @len:        Length in bytes
1232  *
1233  * This path is taken when writing to already allocated clusters.
1234  *
1235  * Called with table_lock held.
1236  */
1237 static int coroutine_fn qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset,
1238                                               size_t len)
1239 {
1240     BDRVQEDState *s = acb_to_s(acb);
1241     int r;
1242 
1243     qemu_co_mutex_unlock(&s->table_lock);
1244 
1245     /* Allocate buffer for zero writes */
1246     if (acb->flags & QED_AIOCB_ZERO) {
1247         struct iovec *iov = acb->qiov->iov;
1248 
1249         if (!iov->iov_base) {
1250             iov->iov_base = qemu_try_blockalign(acb->bs, iov->iov_len);
1251             if (iov->iov_base == NULL) {
1252                 r = -ENOMEM;
1253                 goto out;
1254             }
1255             memset(iov->iov_base, 0, iov->iov_len);
1256         }
1257     }
1258 
1259     /* Calculate the I/O vector */
1260     acb->cur_cluster = offset;
1261     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1262 
1263     /* Do the actual write.  */
1264     r = qed_aio_write_main(acb);
1265 out:
1266     qemu_co_mutex_lock(&s->table_lock);
1267     return r;
1268 }
1269 
1270 /**
1271  * Write data cluster
1272  *
1273  * @opaque:     Write request
1274  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1275  * @offset:     Cluster offset in bytes
1276  * @len:        Length in bytes
1277  *
1278  * Called with table_lock held.
1279  */
1280 static int coroutine_fn qed_aio_write_data(void *opaque, int ret,
1281                                            uint64_t offset, size_t len)
1282 {
1283     QEDAIOCB *acb = opaque;
1284 
1285     trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len);
1286 
1287     acb->find_cluster_ret = ret;
1288 
1289     switch (ret) {
1290     case QED_CLUSTER_FOUND:
1291         return qed_aio_write_inplace(acb, offset, len);
1292 
1293     case QED_CLUSTER_L2:
1294     case QED_CLUSTER_L1:
1295     case QED_CLUSTER_ZERO:
1296         return qed_aio_write_alloc(acb, len);
1297 
1298     default:
1299         g_assert_not_reached();
1300     }
1301 }
1302 
1303 /**
1304  * Read data cluster
1305  *
1306  * @opaque:     Read request
1307  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1308  * @offset:     Cluster offset in bytes
1309  * @len:        Length in bytes
1310  *
1311  * Called with table_lock held.
1312  */
1313 static int coroutine_fn qed_aio_read_data(void *opaque, int ret,
1314                                           uint64_t offset, size_t len)
1315 {
1316     QEDAIOCB *acb = opaque;
1317     BDRVQEDState *s = acb_to_s(acb);
1318     BlockDriverState *bs = acb->bs;
1319     int r;
1320 
1321     qemu_co_mutex_unlock(&s->table_lock);
1322 
1323     /* Adjust offset into cluster */
1324     offset += qed_offset_into_cluster(s, acb->cur_pos);
1325 
1326     trace_qed_aio_read_data(s, acb, ret, offset, len);
1327 
1328     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1329 
1330     /* Handle zero cluster and backing file reads, otherwise read
1331      * data cluster directly.
1332      */
1333     if (ret == QED_CLUSTER_ZERO) {
1334         qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size);
1335         r = 0;
1336     } else if (ret != QED_CLUSTER_FOUND) {
1337         r = qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov,
1338                                   &acb->backing_qiov);
1339     } else {
1340         BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
1341         r = bdrv_co_preadv(bs->file, offset, acb->cur_qiov.size,
1342                            &acb->cur_qiov, 0);
1343     }
1344 
1345     qemu_co_mutex_lock(&s->table_lock);
1346     return r;
1347 }
1348 
1349 /**
1350  * Begin next I/O or complete the request
1351  */
1352 static int coroutine_fn qed_aio_next_io(QEDAIOCB *acb)
1353 {
1354     BDRVQEDState *s = acb_to_s(acb);
1355     uint64_t offset;
1356     size_t len;
1357     int ret;
1358 
1359     qemu_co_mutex_lock(&s->table_lock);
1360     while (1) {
1361         trace_qed_aio_next_io(s, acb, 0, acb->cur_pos + acb->cur_qiov.size);
1362 
1363         if (acb->backing_qiov) {
1364             qemu_iovec_destroy(acb->backing_qiov);
1365             g_free(acb->backing_qiov);
1366             acb->backing_qiov = NULL;
1367         }
1368 
1369         acb->qiov_offset += acb->cur_qiov.size;
1370         acb->cur_pos += acb->cur_qiov.size;
1371         qemu_iovec_reset(&acb->cur_qiov);
1372 
1373         /* Complete request */
1374         if (acb->cur_pos >= acb->end_pos) {
1375             ret = 0;
1376             break;
1377         }
1378 
1379         /* Find next cluster and start I/O */
1380         len = acb->end_pos - acb->cur_pos;
1381         ret = qed_find_cluster(s, &acb->request, acb->cur_pos, &len, &offset);
1382         if (ret < 0) {
1383             break;
1384         }
1385 
1386         if (acb->flags & QED_AIOCB_WRITE) {
1387             ret = qed_aio_write_data(acb, ret, offset, len);
1388         } else {
1389             ret = qed_aio_read_data(acb, ret, offset, len);
1390         }
1391 
1392         if (ret < 0 && ret != -EAGAIN) {
1393             break;
1394         }
1395     }
1396 
1397     trace_qed_aio_complete(s, acb, ret);
1398     qed_aio_complete(acb);
1399     qemu_co_mutex_unlock(&s->table_lock);
1400     return ret;
1401 }
1402 
1403 static int coroutine_fn qed_co_request(BlockDriverState *bs, int64_t sector_num,
1404                                        QEMUIOVector *qiov, int nb_sectors,
1405                                        int flags)
1406 {
1407     QEDAIOCB acb = {
1408         .bs         = bs,
1409         .cur_pos    = (uint64_t) sector_num * BDRV_SECTOR_SIZE,
1410         .end_pos    = (sector_num + nb_sectors) * BDRV_SECTOR_SIZE,
1411         .qiov       = qiov,
1412         .flags      = flags,
1413     };
1414     qemu_iovec_init(&acb.cur_qiov, qiov->niov);
1415 
1416     trace_qed_aio_setup(bs->opaque, &acb, sector_num, nb_sectors, NULL, flags);
1417 
1418     /* Start request */
1419     return qed_aio_next_io(&acb);
1420 }
1421 
1422 static int coroutine_fn bdrv_qed_co_readv(BlockDriverState *bs,
1423                                           int64_t sector_num, int nb_sectors,
1424                                           QEMUIOVector *qiov)
1425 {
1426     return qed_co_request(bs, sector_num, qiov, nb_sectors, 0);
1427 }
1428 
1429 static int coroutine_fn bdrv_qed_co_writev(BlockDriverState *bs,
1430                                            int64_t sector_num, int nb_sectors,
1431                                            QEMUIOVector *qiov, int flags)
1432 {
1433     assert(!flags);
1434     return qed_co_request(bs, sector_num, qiov, nb_sectors, QED_AIOCB_WRITE);
1435 }
1436 
1437 static int coroutine_fn bdrv_qed_co_pwrite_zeroes(BlockDriverState *bs,
1438                                                   int64_t offset,
1439                                                   int bytes,
1440                                                   BdrvRequestFlags flags)
1441 {
1442     BDRVQEDState *s = bs->opaque;
1443 
1444     /*
1445      * Zero writes start without an I/O buffer.  If a buffer becomes necessary
1446      * then it will be allocated during request processing.
1447      */
1448     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, bytes);
1449 
1450     /* Fall back if the request is not aligned */
1451     if (qed_offset_into_cluster(s, offset) ||
1452         qed_offset_into_cluster(s, bytes)) {
1453         return -ENOTSUP;
1454     }
1455 
1456     return qed_co_request(bs, offset >> BDRV_SECTOR_BITS, &qiov,
1457                           bytes >> BDRV_SECTOR_BITS,
1458                           QED_AIOCB_WRITE | QED_AIOCB_ZERO);
1459 }
1460 
1461 static int coroutine_fn bdrv_qed_co_truncate(BlockDriverState *bs,
1462                                              int64_t offset,
1463                                              PreallocMode prealloc,
1464                                              Error **errp)
1465 {
1466     BDRVQEDState *s = bs->opaque;
1467     uint64_t old_image_size;
1468     int ret;
1469 
1470     if (prealloc != PREALLOC_MODE_OFF) {
1471         error_setg(errp, "Unsupported preallocation mode '%s'",
1472                    PreallocMode_str(prealloc));
1473         return -ENOTSUP;
1474     }
1475 
1476     if (!qed_is_image_size_valid(offset, s->header.cluster_size,
1477                                  s->header.table_size)) {
1478         error_setg(errp, "Invalid image size specified");
1479         return -EINVAL;
1480     }
1481 
1482     if ((uint64_t)offset < s->header.image_size) {
1483         error_setg(errp, "Shrinking images is currently not supported");
1484         return -ENOTSUP;
1485     }
1486 
1487     old_image_size = s->header.image_size;
1488     s->header.image_size = offset;
1489     ret = qed_write_header_sync(s);
1490     if (ret < 0) {
1491         s->header.image_size = old_image_size;
1492         error_setg_errno(errp, -ret, "Failed to update the image size");
1493     }
1494     return ret;
1495 }
1496 
1497 static int64_t bdrv_qed_getlength(BlockDriverState *bs)
1498 {
1499     BDRVQEDState *s = bs->opaque;
1500     return s->header.image_size;
1501 }
1502 
1503 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1504 {
1505     BDRVQEDState *s = bs->opaque;
1506 
1507     memset(bdi, 0, sizeof(*bdi));
1508     bdi->cluster_size = s->header.cluster_size;
1509     bdi->is_dirty = s->header.features & QED_F_NEED_CHECK;
1510     bdi->unallocated_blocks_are_zero = true;
1511     return 0;
1512 }
1513 
1514 static int bdrv_qed_change_backing_file(BlockDriverState *bs,
1515                                         const char *backing_file,
1516                                         const char *backing_fmt)
1517 {
1518     BDRVQEDState *s = bs->opaque;
1519     QEDHeader new_header, le_header;
1520     void *buffer;
1521     size_t buffer_len, backing_file_len;
1522     int ret;
1523 
1524     /* Refuse to set backing filename if unknown compat feature bits are
1525      * active.  If the image uses an unknown compat feature then we may not
1526      * know the layout of data following the header structure and cannot safely
1527      * add a new string.
1528      */
1529     if (backing_file && (s->header.compat_features &
1530                          ~QED_COMPAT_FEATURE_MASK)) {
1531         return -ENOTSUP;
1532     }
1533 
1534     memcpy(&new_header, &s->header, sizeof(new_header));
1535 
1536     new_header.features &= ~(QED_F_BACKING_FILE |
1537                              QED_F_BACKING_FORMAT_NO_PROBE);
1538 
1539     /* Adjust feature flags */
1540     if (backing_file) {
1541         new_header.features |= QED_F_BACKING_FILE;
1542 
1543         if (qed_fmt_is_raw(backing_fmt)) {
1544             new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
1545         }
1546     }
1547 
1548     /* Calculate new header size */
1549     backing_file_len = 0;
1550 
1551     if (backing_file) {
1552         backing_file_len = strlen(backing_file);
1553     }
1554 
1555     buffer_len = sizeof(new_header);
1556     new_header.backing_filename_offset = buffer_len;
1557     new_header.backing_filename_size = backing_file_len;
1558     buffer_len += backing_file_len;
1559 
1560     /* Make sure we can rewrite header without failing */
1561     if (buffer_len > new_header.header_size * new_header.cluster_size) {
1562         return -ENOSPC;
1563     }
1564 
1565     /* Prepare new header */
1566     buffer = g_malloc(buffer_len);
1567 
1568     qed_header_cpu_to_le(&new_header, &le_header);
1569     memcpy(buffer, &le_header, sizeof(le_header));
1570     buffer_len = sizeof(le_header);
1571 
1572     if (backing_file) {
1573         memcpy(buffer + buffer_len, backing_file, backing_file_len);
1574         buffer_len += backing_file_len;
1575     }
1576 
1577     /* Write new header */
1578     ret = bdrv_pwrite_sync(bs->file, 0, buffer, buffer_len);
1579     g_free(buffer);
1580     if (ret == 0) {
1581         memcpy(&s->header, &new_header, sizeof(new_header));
1582     }
1583     return ret;
1584 }
1585 
1586 static void coroutine_fn bdrv_qed_co_invalidate_cache(BlockDriverState *bs,
1587                                                       Error **errp)
1588 {
1589     BDRVQEDState *s = bs->opaque;
1590     Error *local_err = NULL;
1591     int ret;
1592 
1593     bdrv_qed_close(bs);
1594 
1595     bdrv_qed_init_state(bs);
1596     qemu_co_mutex_lock(&s->table_lock);
1597     ret = bdrv_qed_do_open(bs, NULL, bs->open_flags, &local_err);
1598     qemu_co_mutex_unlock(&s->table_lock);
1599     if (local_err) {
1600         error_propagate_prepend(errp, local_err,
1601                                 "Could not reopen qed layer: ");
1602         return;
1603     } else if (ret < 0) {
1604         error_setg_errno(errp, -ret, "Could not reopen qed layer");
1605         return;
1606     }
1607 }
1608 
1609 static int coroutine_fn bdrv_qed_co_check(BlockDriverState *bs,
1610                                           BdrvCheckResult *result,
1611                                           BdrvCheckMode fix)
1612 {
1613     BDRVQEDState *s = bs->opaque;
1614     int ret;
1615 
1616     qemu_co_mutex_lock(&s->table_lock);
1617     ret = qed_check(s, result, !!fix);
1618     qemu_co_mutex_unlock(&s->table_lock);
1619 
1620     return ret;
1621 }
1622 
1623 static QemuOptsList qed_create_opts = {
1624     .name = "qed-create-opts",
1625     .head = QTAILQ_HEAD_INITIALIZER(qed_create_opts.head),
1626     .desc = {
1627         {
1628             .name = BLOCK_OPT_SIZE,
1629             .type = QEMU_OPT_SIZE,
1630             .help = "Virtual disk size"
1631         },
1632         {
1633             .name = BLOCK_OPT_BACKING_FILE,
1634             .type = QEMU_OPT_STRING,
1635             .help = "File name of a base image"
1636         },
1637         {
1638             .name = BLOCK_OPT_BACKING_FMT,
1639             .type = QEMU_OPT_STRING,
1640             .help = "Image format of the base image"
1641         },
1642         {
1643             .name = BLOCK_OPT_CLUSTER_SIZE,
1644             .type = QEMU_OPT_SIZE,
1645             .help = "Cluster size (in bytes)",
1646             .def_value_str = stringify(QED_DEFAULT_CLUSTER_SIZE)
1647         },
1648         {
1649             .name = BLOCK_OPT_TABLE_SIZE,
1650             .type = QEMU_OPT_SIZE,
1651             .help = "L1/L2 table size (in clusters)"
1652         },
1653         { /* end of list */ }
1654     }
1655 };
1656 
1657 static BlockDriver bdrv_qed = {
1658     .format_name              = "qed",
1659     .instance_size            = sizeof(BDRVQEDState),
1660     .create_opts              = &qed_create_opts,
1661     .supports_backing         = true,
1662 
1663     .bdrv_probe               = bdrv_qed_probe,
1664     .bdrv_open                = bdrv_qed_open,
1665     .bdrv_close               = bdrv_qed_close,
1666     .bdrv_reopen_prepare      = bdrv_qed_reopen_prepare,
1667     .bdrv_child_perm          = bdrv_format_default_perms,
1668     .bdrv_co_create           = bdrv_qed_co_create,
1669     .bdrv_co_create_opts      = bdrv_qed_co_create_opts,
1670     .bdrv_has_zero_init       = bdrv_has_zero_init_1,
1671     .bdrv_co_block_status     = bdrv_qed_co_block_status,
1672     .bdrv_co_readv            = bdrv_qed_co_readv,
1673     .bdrv_co_writev           = bdrv_qed_co_writev,
1674     .bdrv_co_pwrite_zeroes    = bdrv_qed_co_pwrite_zeroes,
1675     .bdrv_co_truncate         = bdrv_qed_co_truncate,
1676     .bdrv_getlength           = bdrv_qed_getlength,
1677     .bdrv_get_info            = bdrv_qed_get_info,
1678     .bdrv_refresh_limits      = bdrv_qed_refresh_limits,
1679     .bdrv_change_backing_file = bdrv_qed_change_backing_file,
1680     .bdrv_co_invalidate_cache = bdrv_qed_co_invalidate_cache,
1681     .bdrv_co_check            = bdrv_qed_co_check,
1682     .bdrv_detach_aio_context  = bdrv_qed_detach_aio_context,
1683     .bdrv_attach_aio_context  = bdrv_qed_attach_aio_context,
1684     .bdrv_co_drain_begin      = bdrv_qed_co_drain_begin,
1685 };
1686 
1687 static void bdrv_qed_init(void)
1688 {
1689     bdrv_register(&bdrv_qed);
1690 }
1691 
1692 block_init(bdrv_qed_init);
1693