xref: /openbmc/qemu/block/qed.c (revision 75c5bb0b)
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/main-loop.h"
21 #include "qemu/module.h"
22 #include "qemu/option.h"
23 #include "trace.h"
24 #include "qed.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"
29 
30 static QemuOptsList qed_create_opts;
31 
32 static int bdrv_qed_probe(const uint8_t *buf, int buf_size,
33                           const char *filename)
34 {
35     const QEDHeader *header = (const QEDHeader *)buf;
36 
37     if (buf_size < sizeof(*header)) {
38         return 0;
39     }
40     if (le32_to_cpu(header->magic) != QED_MAGIC) {
41         return 0;
42     }
43     return 100;
44 }
45 
46 /**
47  * Check whether an image format is raw
48  *
49  * @fmt:    Backing file format, may be NULL
50  */
51 static bool qed_fmt_is_raw(const char *fmt)
52 {
53     return fmt && strcmp(fmt, "raw") == 0;
54 }
55 
56 static void qed_header_le_to_cpu(const QEDHeader *le, QEDHeader *cpu)
57 {
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);
69 }
70 
71 static void qed_header_cpu_to_le(const QEDHeader *cpu, QEDHeader *le)
72 {
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);
84 }
85 
86 int qed_write_header_sync(BDRVQEDState *s)
87 {
88     QEDHeader le;
89     int ret;
90 
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)) {
94         return ret;
95     }
96     return 0;
97 }
98 
99 /**
100  * Update header in-place (does not rewrite backing filename or other strings)
101  *
102  * This function only updates known header fields in-place and does not affect
103  * extra data after the QED header.
104  *
105  * No new allocating reqs can start while this function runs.
106  */
107 static int coroutine_fn qed_write_header(BDRVQEDState *s)
108 {
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.
113      */
114 
115     int nsectors = DIV_ROUND_UP(sizeof(QEDHeader), BDRV_SECTOR_SIZE);
116     size_t len = nsectors * BDRV_SECTOR_SIZE;
117     uint8_t *buf;
118     int ret;
119 
120     assert(s->allocating_acb || s->allocating_write_reqs_plugged);
121 
122     buf = qemu_blockalign(s->bs, len);
123 
124     ret = bdrv_co_pread(s->bs->file, 0, len, buf, 0);
125     if (ret < 0) {
126         goto out;
127     }
128 
129     /* Update header */
130     qed_header_cpu_to_le(&s->header, (QEDHeader *) buf);
131 
132     ret = bdrv_co_pwrite(s->bs->file, 0, len,  buf, 0);
133     if (ret < 0) {
134         goto out;
135     }
136 
137     ret = 0;
138 out:
139     qemu_vfree(buf);
140     return ret;
141 }
142 
143 static uint64_t qed_max_image_size(uint32_t cluster_size, uint32_t table_size)
144 {
145     uint64_t table_entries;
146     uint64_t l2_size;
147 
148     table_entries = (table_size * cluster_size) / sizeof(uint64_t);
149     l2_size = table_entries * cluster_size;
150 
151     return l2_size * table_entries;
152 }
153 
154 static bool qed_is_cluster_size_valid(uint32_t cluster_size)
155 {
156     if (cluster_size < QED_MIN_CLUSTER_SIZE ||
157         cluster_size > QED_MAX_CLUSTER_SIZE) {
158         return false;
159     }
160     if (cluster_size & (cluster_size - 1)) {
161         return false; /* not power of 2 */
162     }
163     return true;
164 }
165 
166 static bool qed_is_table_size_valid(uint32_t table_size)
167 {
168     if (table_size < QED_MIN_TABLE_SIZE ||
169         table_size > QED_MAX_TABLE_SIZE) {
170         return false;
171     }
172     if (table_size & (table_size - 1)) {
173         return false; /* not power of 2 */
174     }
175     return true;
176 }
177 
178 static bool qed_is_image_size_valid(uint64_t image_size, uint32_t cluster_size,
179                                     uint32_t table_size)
180 {
181     if (image_size % BDRV_SECTOR_SIZE != 0) {
182         return false; /* not multiple of sector size */
183     }
184     if (image_size > qed_max_image_size(cluster_size, table_size)) {
185         return false; /* image is too large */
186     }
187     return true;
188 }
189 
190 /**
191  * Read a string of known length from the image file
192  *
193  * @file:       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
199  *
200  * The string is NUL-terminated.
201  */
202 static int qed_read_string(BdrvChild *file, uint64_t offset, size_t n,
203                            char *buf, size_t buflen)
204 {
205     int ret;
206     if (n >= buflen) {
207         return -EINVAL;
208     }
209     ret = bdrv_pread(file, offset, buf, n);
210     if (ret < 0) {
211         return ret;
212     }
213     buf[n] = '\0';
214     return 0;
215 }
216 
217 /**
218  * Allocate new clusters
219  *
220  * @s:          QED state
221  * @n:          Number of contiguous clusters to allocate
222  * @ret:        Offset of first allocated cluster
223  *
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
226  * file.
227  *
228  * Called with table_lock held.
229  */
230 static uint64_t qed_alloc_clusters(BDRVQEDState *s, unsigned int n)
231 {
232     uint64_t offset = s->file_size;
233     s->file_size += n * s->header.cluster_size;
234     return offset;
235 }
236 
237 QEDTable *qed_alloc_table(BDRVQEDState *s)
238 {
239     /* Honor O_DIRECT memory alignment requirements */
240     return qemu_blockalign(s->bs,
241                            s->header.cluster_size * s->header.table_size);
242 }
243 
244 /**
245  * Allocate a new zeroed L2 table
246  *
247  * Called with table_lock held.
248  */
249 static CachedL2Table *qed_new_l2_table(BDRVQEDState *s)
250 {
251     CachedL2Table *l2_table = qed_alloc_l2_cache_entry(&s->l2_cache);
252 
253     l2_table->table = qed_alloc_table(s);
254     l2_table->offset = qed_alloc_clusters(s, s->header.table_size);
255 
256     memset(l2_table->table->offsets, 0,
257            s->header.cluster_size * s->header.table_size);
258     return l2_table;
259 }
260 
261 static bool qed_plug_allocating_write_reqs(BDRVQEDState *s)
262 {
263     qemu_co_mutex_lock(&s->table_lock);
264 
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.
270          */
271         qemu_co_mutex_unlock(&s->table_lock);
272         return false;
273     }
274 
275     s->allocating_write_reqs_plugged = true;
276     qemu_co_mutex_unlock(&s->table_lock);
277     return true;
278 }
279 
280 static void qed_unplug_allocating_write_reqs(BDRVQEDState *s)
281 {
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);
287 }
288 
289 static void coroutine_fn qed_need_check_timer_entry(void *opaque)
290 {
291     BDRVQEDState *s = opaque;
292     int ret;
293 
294     trace_qed_need_check_timer_cb(s);
295 
296     if (!qed_plug_allocating_write_reqs(s)) {
297         return;
298     }
299 
300     /* Ensure writes are on disk before clearing flag */
301     ret = bdrv_co_flush(s->bs->file->bs);
302     if (ret < 0) {
303         qed_unplug_allocating_write_reqs(s);
304         return;
305     }
306 
307     s->header.features &= ~QED_F_NEED_CHECK;
308     ret = qed_write_header(s);
309     (void) ret;
310 
311     qed_unplug_allocating_write_reqs(s);
312 
313     ret = bdrv_co_flush(s->bs);
314     (void) ret;
315 }
316 
317 static void qed_need_check_timer_cb(void *opaque)
318 {
319     Coroutine *co = qemu_coroutine_create(qed_need_check_timer_entry, opaque);
320     qemu_coroutine_enter(co);
321 }
322 
323 static void qed_start_need_check_timer(BDRVQEDState *s)
324 {
325     trace_qed_start_need_check_timer(s);
326 
327     /* Use QEMU_CLOCK_VIRTUAL so we don't alter the image file while suspended for
328      * migration.
329      */
330     timer_mod(s->need_check_timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
331                    NANOSECONDS_PER_SECOND * QED_NEED_CHECK_TIMEOUT);
332 }
333 
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)
336 {
337     trace_qed_cancel_need_check_timer(s);
338     timer_del(s->need_check_timer);
339 }
340 
341 static void bdrv_qed_detach_aio_context(BlockDriverState *bs)
342 {
343     BDRVQEDState *s = bs->opaque;
344 
345     qed_cancel_need_check_timer(s);
346     timer_free(s->need_check_timer);
347 }
348 
349 static void bdrv_qed_attach_aio_context(BlockDriverState *bs,
350                                         AioContext *new_context)
351 {
352     BDRVQEDState *s = bs->opaque;
353 
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);
359     }
360 }
361 
362 static void coroutine_fn bdrv_qed_co_drain_begin(BlockDriverState *bs)
363 {
364     BDRVQEDState *s = bs->opaque;
365 
366     /* Fire the timer immediately in order to start doing I/O as soon as the
367      * header is flushed.
368      */
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);
372     }
373 }
374 
375 static void bdrv_qed_init_state(BlockDriverState *bs)
376 {
377     BDRVQEDState *s = bs->opaque;
378 
379     memset(s, 0, sizeof(BDRVQEDState));
380     s->bs = bs;
381     qemu_co_mutex_init(&s->table_lock);
382     qemu_co_queue_init(&s->allocating_write_reqs);
383 }
384 
385 /* Called with table_lock held.  */
386 static int coroutine_fn bdrv_qed_do_open(BlockDriverState *bs, QDict *options,
387                                          int flags, Error **errp)
388 {
389     BDRVQEDState *s = bs->opaque;
390     QEDHeader le_header;
391     int64_t file_size;
392     int ret;
393 
394     ret = bdrv_pread(bs->file, 0, &le_header, sizeof(le_header));
395     if (ret < 0) {
396         return ret;
397     }
398     qed_header_le_to_cpu(&le_header, &s->header);
399 
400     if (s->header.magic != QED_MAGIC) {
401         error_setg(errp, "Image not in QED format");
402         return -EINVAL;
403     }
404     if (s->header.features & ~QED_FEATURE_MASK) {
405         /* image uses unsupported feature bits */
406         error_setg(errp, "Unsupported QED features: %" PRIx64,
407                    s->header.features & ~QED_FEATURE_MASK);
408         return -ENOTSUP;
409     }
410     if (!qed_is_cluster_size_valid(s->header.cluster_size)) {
411         return -EINVAL;
412     }
413 
414     /* Round down file size to the last cluster */
415     file_size = bdrv_getlength(bs->file->bs);
416     if (file_size < 0) {
417         return file_size;
418     }
419     s->file_size = qed_start_of_cluster(s, file_size);
420 
421     if (!qed_is_table_size_valid(s->header.table_size)) {
422         return -EINVAL;
423     }
424     if (!qed_is_image_size_valid(s->header.image_size,
425                                  s->header.cluster_size,
426                                  s->header.table_size)) {
427         return -EINVAL;
428     }
429     if (!qed_check_table_offset(s, s->header.l1_table_offset)) {
430         return -EINVAL;
431     }
432 
433     s->table_nelems = (s->header.cluster_size * s->header.table_size) /
434                       sizeof(uint64_t);
435     s->l2_shift = ctz32(s->header.cluster_size);
436     s->l2_mask = s->table_nelems - 1;
437     s->l1_shift = s->l2_shift + ctz32(s->table_nelems);
438 
439     /* Header size calculation must not overflow uint32_t */
440     if (s->header.header_size > UINT32_MAX / s->header.cluster_size) {
441         return -EINVAL;
442     }
443 
444     if ((s->header.features & QED_F_BACKING_FILE)) {
445         if ((uint64_t)s->header.backing_filename_offset +
446             s->header.backing_filename_size >
447             s->header.cluster_size * s->header.header_size) {
448             return -EINVAL;
449         }
450 
451         ret = qed_read_string(bs->file, s->header.backing_filename_offset,
452                               s->header.backing_filename_size,
453                               bs->auto_backing_file,
454                               sizeof(bs->auto_backing_file));
455         if (ret < 0) {
456             return ret;
457         }
458         pstrcpy(bs->backing_file, sizeof(bs->backing_file),
459                 bs->auto_backing_file);
460 
461         if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) {
462             pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw");
463         }
464     }
465 
466     /* Reset unknown autoclear feature bits.  This is a backwards
467      * compatibility mechanism that allows images to be opened by older
468      * programs, which "knock out" unknown feature bits.  When an image is
469      * opened by a newer program again it can detect that the autoclear
470      * feature is no longer valid.
471      */
472     if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 &&
473         !bdrv_is_read_only(bs->file->bs) && !(flags & BDRV_O_INACTIVE)) {
474         s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK;
475 
476         ret = qed_write_header_sync(s);
477         if (ret) {
478             return ret;
479         }
480 
481         /* From here on only known autoclear feature bits are valid */
482         bdrv_flush(bs->file->bs);
483     }
484 
485     s->l1_table = qed_alloc_table(s);
486     qed_init_l2_cache(&s->l2_cache);
487 
488     ret = qed_read_l1_table_sync(s);
489     if (ret) {
490         goto out;
491     }
492 
493     /* If image was not closed cleanly, check consistency */
494     if (!(flags & BDRV_O_CHECK) && (s->header.features & QED_F_NEED_CHECK)) {
495         /* Read-only images cannot be fixed.  There is no risk of corruption
496          * since write operations are not possible.  Therefore, allow
497          * potentially inconsistent images to be opened read-only.  This can
498          * aid data recovery from an otherwise inconsistent image.
499          */
500         if (!bdrv_is_read_only(bs->file->bs) &&
501             !(flags & BDRV_O_INACTIVE)) {
502             BdrvCheckResult result = {0};
503 
504             ret = qed_check(s, &result, true);
505             if (ret) {
506                 goto out;
507             }
508         }
509     }
510 
511     bdrv_qed_attach_aio_context(bs, bdrv_get_aio_context(bs));
512 
513 out:
514     if (ret) {
515         qed_free_l2_cache(&s->l2_cache);
516         qemu_vfree(s->l1_table);
517     }
518     return ret;
519 }
520 
521 typedef struct QEDOpenCo {
522     BlockDriverState *bs;
523     QDict *options;
524     int flags;
525     Error **errp;
526     int ret;
527 } QEDOpenCo;
528 
529 static void coroutine_fn bdrv_qed_open_entry(void *opaque)
530 {
531     QEDOpenCo *qoc = opaque;
532     BDRVQEDState *s = qoc->bs->opaque;
533 
534     qemu_co_mutex_lock(&s->table_lock);
535     qoc->ret = bdrv_qed_do_open(qoc->bs, qoc->options, qoc->flags, qoc->errp);
536     qemu_co_mutex_unlock(&s->table_lock);
537 }
538 
539 static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags,
540                          Error **errp)
541 {
542     QEDOpenCo qoc = {
543         .bs = bs,
544         .options = options,
545         .flags = flags,
546         .errp = errp,
547         .ret = -EINPROGRESS
548     };
549 
550     bs->file = bdrv_open_child(NULL, options, "file", bs, &child_file,
551                                false, errp);
552     if (!bs->file) {
553         return -EINVAL;
554     }
555 
556     bdrv_qed_init_state(bs);
557     if (qemu_in_coroutine()) {
558         bdrv_qed_open_entry(&qoc);
559     } else {
560         assert(qemu_get_current_aio_context() == qemu_get_aio_context());
561         qemu_coroutine_enter(qemu_coroutine_create(bdrv_qed_open_entry, &qoc));
562         BDRV_POLL_WHILE(bs, qoc.ret == -EINPROGRESS);
563     }
564     BDRV_POLL_WHILE(bs, qoc.ret == -EINPROGRESS);
565     return qoc.ret;
566 }
567 
568 static void bdrv_qed_refresh_limits(BlockDriverState *bs, Error **errp)
569 {
570     BDRVQEDState *s = bs->opaque;
571 
572     bs->bl.pwrite_zeroes_alignment = s->header.cluster_size;
573 }
574 
575 /* We have nothing to do for QED reopen, stubs just return
576  * success */
577 static int bdrv_qed_reopen_prepare(BDRVReopenState *state,
578                                    BlockReopenQueue *queue, Error **errp)
579 {
580     return 0;
581 }
582 
583 static void bdrv_qed_close(BlockDriverState *bs)
584 {
585     BDRVQEDState *s = bs->opaque;
586 
587     bdrv_qed_detach_aio_context(bs);
588 
589     /* Ensure writes reach stable storage */
590     bdrv_flush(bs->file->bs);
591 
592     /* Clean shutdown, no check required on next open */
593     if (s->header.features & QED_F_NEED_CHECK) {
594         s->header.features &= ~QED_F_NEED_CHECK;
595         qed_write_header_sync(s);
596     }
597 
598     qed_free_l2_cache(&s->l2_cache);
599     qemu_vfree(s->l1_table);
600 }
601 
602 static int coroutine_fn bdrv_qed_co_create(BlockdevCreateOptions *opts,
603                                            Error **errp)
604 {
605     BlockdevCreateOptionsQed *qed_opts;
606     BlockBackend *blk = NULL;
607     BlockDriverState *bs = NULL;
608 
609     QEDHeader header;
610     QEDHeader le_header;
611     uint8_t *l1_table = NULL;
612     size_t l1_size;
613     int ret = 0;
614 
615     assert(opts->driver == BLOCKDEV_DRIVER_QED);
616     qed_opts = &opts->u.qed;
617 
618     /* Validate options and set default values */
619     if (!qed_opts->has_cluster_size) {
620         qed_opts->cluster_size = QED_DEFAULT_CLUSTER_SIZE;
621     }
622     if (!qed_opts->has_table_size) {
623         qed_opts->table_size = QED_DEFAULT_TABLE_SIZE;
624     }
625 
626     if (!qed_is_cluster_size_valid(qed_opts->cluster_size)) {
627         error_setg(errp, "QED cluster size must be within range [%u, %u] "
628                          "and power of 2",
629                    QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE);
630         return -EINVAL;
631     }
632     if (!qed_is_table_size_valid(qed_opts->table_size)) {
633         error_setg(errp, "QED table size must be within range [%u, %u] "
634                          "and power of 2",
635                    QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE);
636         return -EINVAL;
637     }
638     if (!qed_is_image_size_valid(qed_opts->size, qed_opts->cluster_size,
639                                  qed_opts->table_size))
640     {
641         error_setg(errp, "QED image size must be a non-zero multiple of "
642                          "cluster size and less than %" PRIu64 " bytes",
643                    qed_max_image_size(qed_opts->cluster_size,
644                                       qed_opts->table_size));
645         return -EINVAL;
646     }
647 
648     /* Create BlockBackend to write to the image */
649     bs = bdrv_open_blockdev_ref(qed_opts->file, errp);
650     if (bs == NULL) {
651         return -EIO;
652     }
653 
654     blk = blk_new(bdrv_get_aio_context(bs),
655                   BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
656     ret = blk_insert_bs(blk, bs, errp);
657     if (ret < 0) {
658         goto out;
659     }
660     blk_set_allow_write_beyond_eof(blk, true);
661 
662     /* Prepare image format */
663     header = (QEDHeader) {
664         .magic = QED_MAGIC,
665         .cluster_size = qed_opts->cluster_size,
666         .table_size = qed_opts->table_size,
667         .header_size = 1,
668         .features = 0,
669         .compat_features = 0,
670         .l1_table_offset = qed_opts->cluster_size,
671         .image_size = qed_opts->size,
672     };
673 
674     l1_size = header.cluster_size * header.table_size;
675 
676     /*
677      * The QED format associates file length with allocation status,
678      * so a new file (which is empty) must have a length of 0.
679      */
680     ret = blk_truncate(blk, 0, true, PREALLOC_MODE_OFF, errp);
681     if (ret < 0) {
682         goto out;
683     }
684 
685     if (qed_opts->has_backing_file) {
686         header.features |= QED_F_BACKING_FILE;
687         header.backing_filename_offset = sizeof(le_header);
688         header.backing_filename_size = strlen(qed_opts->backing_file);
689 
690         if (qed_opts->has_backing_fmt) {
691             const char *backing_fmt = BlockdevDriver_str(qed_opts->backing_fmt);
692             if (qed_fmt_is_raw(backing_fmt)) {
693                 header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
694             }
695         }
696     }
697 
698     qed_header_cpu_to_le(&header, &le_header);
699     ret = blk_pwrite(blk, 0, &le_header, sizeof(le_header), 0);
700     if (ret < 0) {
701         goto out;
702     }
703     ret = blk_pwrite(blk, sizeof(le_header), qed_opts->backing_file,
704                      header.backing_filename_size, 0);
705     if (ret < 0) {
706         goto out;
707     }
708 
709     l1_table = g_malloc0(l1_size);
710     ret = blk_pwrite(blk, header.l1_table_offset, l1_table, l1_size, 0);
711     if (ret < 0) {
712         goto out;
713     }
714 
715     ret = 0; /* success */
716 out:
717     g_free(l1_table);
718     blk_unref(blk);
719     bdrv_unref(bs);
720     return ret;
721 }
722 
723 static int coroutine_fn bdrv_qed_co_create_opts(const char *filename,
724                                                 QemuOpts *opts,
725                                                 Error **errp)
726 {
727     BlockdevCreateOptions *create_options = NULL;
728     QDict *qdict;
729     Visitor *v;
730     BlockDriverState *bs = NULL;
731     Error *local_err = NULL;
732     int ret;
733 
734     static const QDictRenames opt_renames[] = {
735         { BLOCK_OPT_BACKING_FILE,       "backing-file" },
736         { BLOCK_OPT_BACKING_FMT,        "backing-fmt" },
737         { BLOCK_OPT_CLUSTER_SIZE,       "cluster-size" },
738         { BLOCK_OPT_TABLE_SIZE,         "table-size" },
739         { NULL, NULL },
740     };
741 
742     /* Parse options and convert legacy syntax */
743     qdict = qemu_opts_to_qdict_filtered(opts, NULL, &qed_create_opts, true);
744 
745     if (!qdict_rename_keys(qdict, opt_renames, errp)) {
746         ret = -EINVAL;
747         goto fail;
748     }
749 
750     /* Create and open the file (protocol layer) */
751     ret = bdrv_create_file(filename, opts, &local_err);
752     if (ret < 0) {
753         error_propagate(errp, local_err);
754         goto fail;
755     }
756 
757     bs = bdrv_open(filename, NULL, NULL,
758                    BDRV_O_RDWR | BDRV_O_RESIZE | BDRV_O_PROTOCOL, errp);
759     if (bs == NULL) {
760         ret = -EIO;
761         goto fail;
762     }
763 
764     /* Now get the QAPI type BlockdevCreateOptions */
765     qdict_put_str(qdict, "driver", "qed");
766     qdict_put_str(qdict, "file", bs->node_name);
767 
768     v = qobject_input_visitor_new_flat_confused(qdict, errp);
769     if (!v) {
770         ret = -EINVAL;
771         goto fail;
772     }
773 
774     visit_type_BlockdevCreateOptions(v, NULL, &create_options, &local_err);
775     visit_free(v);
776 
777     if (local_err) {
778         error_propagate(errp, local_err);
779         ret = -EINVAL;
780         goto fail;
781     }
782 
783     /* Silently round up size */
784     assert(create_options->driver == BLOCKDEV_DRIVER_QED);
785     create_options->u.qed.size =
786         ROUND_UP(create_options->u.qed.size, BDRV_SECTOR_SIZE);
787 
788     /* Create the qed image (format layer) */
789     ret = bdrv_qed_co_create(create_options, errp);
790 
791 fail:
792     qobject_unref(qdict);
793     bdrv_unref(bs);
794     qapi_free_BlockdevCreateOptions(create_options);
795     return ret;
796 }
797 
798 static int coroutine_fn bdrv_qed_co_block_status(BlockDriverState *bs,
799                                                  bool want_zero,
800                                                  int64_t pos, int64_t bytes,
801                                                  int64_t *pnum, int64_t *map,
802                                                  BlockDriverState **file)
803 {
804     BDRVQEDState *s = bs->opaque;
805     size_t len = MIN(bytes, SIZE_MAX);
806     int status;
807     QEDRequest request = { .l2_table = NULL };
808     uint64_t offset;
809     int ret;
810 
811     qemu_co_mutex_lock(&s->table_lock);
812     ret = qed_find_cluster(s, &request, pos, &len, &offset);
813 
814     *pnum = len;
815     switch (ret) {
816     case QED_CLUSTER_FOUND:
817         *map = offset | qed_offset_into_cluster(s, pos);
818         status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
819         *file = bs->file->bs;
820         break;
821     case QED_CLUSTER_ZERO:
822         status = BDRV_BLOCK_ZERO;
823         break;
824     case QED_CLUSTER_L2:
825     case QED_CLUSTER_L1:
826         status = 0;
827         break;
828     default:
829         assert(ret < 0);
830         status = ret;
831         break;
832     }
833 
834     qed_unref_l2_cache_entry(request.l2_table);
835     qemu_co_mutex_unlock(&s->table_lock);
836 
837     return status;
838 }
839 
840 static BDRVQEDState *acb_to_s(QEDAIOCB *acb)
841 {
842     return acb->bs->opaque;
843 }
844 
845 /**
846  * Read from the backing file or zero-fill if no backing file
847  *
848  * @s:              QED state
849  * @pos:            Byte position in device
850  * @qiov:           Destination I/O vector
851  * @backing_qiov:   Possibly shortened copy of qiov, to be allocated here
852  * @cb:             Completion function
853  * @opaque:         User data for completion function
854  *
855  * This function reads qiov->size bytes starting at pos from the backing file.
856  * If there is no backing file then zeroes are read.
857  */
858 static int coroutine_fn qed_read_backing_file(BDRVQEDState *s, uint64_t pos,
859                                               QEMUIOVector *qiov,
860                                               QEMUIOVector **backing_qiov)
861 {
862     uint64_t backing_length = 0;
863     size_t size;
864     int ret;
865 
866     /* If there is a backing file, get its length.  Treat the absence of a
867      * backing file like a zero length backing file.
868      */
869     if (s->bs->backing) {
870         int64_t l = bdrv_getlength(s->bs->backing->bs);
871         if (l < 0) {
872             return l;
873         }
874         backing_length = l;
875     }
876 
877     /* Zero all sectors if reading beyond the end of the backing file */
878     if (pos >= backing_length ||
879         pos + qiov->size > backing_length) {
880         qemu_iovec_memset(qiov, 0, 0, qiov->size);
881     }
882 
883     /* Complete now if there are no backing file sectors to read */
884     if (pos >= backing_length) {
885         return 0;
886     }
887 
888     /* If the read straddles the end of the backing file, shorten it */
889     size = MIN((uint64_t)backing_length - pos, qiov->size);
890 
891     assert(*backing_qiov == NULL);
892     *backing_qiov = g_new(QEMUIOVector, 1);
893     qemu_iovec_init(*backing_qiov, qiov->niov);
894     qemu_iovec_concat(*backing_qiov, qiov, 0, size);
895 
896     BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO);
897     ret = bdrv_co_preadv(s->bs->backing, pos, size, *backing_qiov, 0);
898     if (ret < 0) {
899         return ret;
900     }
901     return 0;
902 }
903 
904 /**
905  * Copy data from backing file into the image
906  *
907  * @s:          QED state
908  * @pos:        Byte position in device
909  * @len:        Number of bytes
910  * @offset:     Byte offset in image file
911  */
912 static int coroutine_fn qed_copy_from_backing_file(BDRVQEDState *s,
913                                                    uint64_t pos, uint64_t len,
914                                                    uint64_t offset)
915 {
916     QEMUIOVector qiov;
917     QEMUIOVector *backing_qiov = NULL;
918     int ret;
919 
920     /* Skip copy entirely if there is no work to do */
921     if (len == 0) {
922         return 0;
923     }
924 
925     qemu_iovec_init_buf(&qiov, qemu_blockalign(s->bs, len), len);
926 
927     ret = qed_read_backing_file(s, pos, &qiov, &backing_qiov);
928 
929     if (backing_qiov) {
930         qemu_iovec_destroy(backing_qiov);
931         g_free(backing_qiov);
932         backing_qiov = NULL;
933     }
934 
935     if (ret) {
936         goto out;
937     }
938 
939     BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE);
940     ret = bdrv_co_pwritev(s->bs->file, offset, qiov.size, &qiov, 0);
941     if (ret < 0) {
942         goto out;
943     }
944     ret = 0;
945 out:
946     qemu_vfree(qemu_iovec_buf(&qiov));
947     return ret;
948 }
949 
950 /**
951  * Link one or more contiguous clusters into a table
952  *
953  * @s:              QED state
954  * @table:          L2 table
955  * @index:          First cluster index
956  * @n:              Number of contiguous clusters
957  * @cluster:        First cluster offset
958  *
959  * The cluster offset may be an allocated byte offset in the image file, the
960  * zero cluster marker, or the unallocated cluster marker.
961  *
962  * Called with table_lock held.
963  */
964 static void coroutine_fn qed_update_l2_table(BDRVQEDState *s, QEDTable *table,
965                                              int index, unsigned int n,
966                                              uint64_t cluster)
967 {
968     int i;
969     for (i = index; i < index + n; i++) {
970         table->offsets[i] = cluster;
971         if (!qed_offset_is_unalloc_cluster(cluster) &&
972             !qed_offset_is_zero_cluster(cluster)) {
973             cluster += s->header.cluster_size;
974         }
975     }
976 }
977 
978 /* Called with table_lock held.  */
979 static void coroutine_fn qed_aio_complete(QEDAIOCB *acb)
980 {
981     BDRVQEDState *s = acb_to_s(acb);
982 
983     /* Free resources */
984     qemu_iovec_destroy(&acb->cur_qiov);
985     qed_unref_l2_cache_entry(acb->request.l2_table);
986 
987     /* Free the buffer we may have allocated for zero writes */
988     if (acb->flags & QED_AIOCB_ZERO) {
989         qemu_vfree(acb->qiov->iov[0].iov_base);
990         acb->qiov->iov[0].iov_base = NULL;
991     }
992 
993     /* Start next allocating write request waiting behind this one.  Note that
994      * requests enqueue themselves when they first hit an unallocated cluster
995      * but they wait until the entire request is finished before waking up the
996      * next request in the queue.  This ensures that we don't cycle through
997      * requests multiple times but rather finish one at a time completely.
998      */
999     if (acb == s->allocating_acb) {
1000         s->allocating_acb = NULL;
1001         if (!qemu_co_queue_empty(&s->allocating_write_reqs)) {
1002             qemu_co_queue_next(&s->allocating_write_reqs);
1003         } else if (s->header.features & QED_F_NEED_CHECK) {
1004             qed_start_need_check_timer(s);
1005         }
1006     }
1007 }
1008 
1009 /**
1010  * Update L1 table with new L2 table offset and write it out
1011  *
1012  * Called with table_lock held.
1013  */
1014 static int coroutine_fn qed_aio_write_l1_update(QEDAIOCB *acb)
1015 {
1016     BDRVQEDState *s = acb_to_s(acb);
1017     CachedL2Table *l2_table = acb->request.l2_table;
1018     uint64_t l2_offset = l2_table->offset;
1019     int index, ret;
1020 
1021     index = qed_l1_index(s, acb->cur_pos);
1022     s->l1_table->offsets[index] = l2_table->offset;
1023 
1024     ret = qed_write_l1_table(s, index, 1);
1025 
1026     /* Commit the current L2 table to the cache */
1027     qed_commit_l2_cache_entry(&s->l2_cache, l2_table);
1028 
1029     /* This is guaranteed to succeed because we just committed the entry to the
1030      * cache.
1031      */
1032     acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset);
1033     assert(acb->request.l2_table != NULL);
1034 
1035     return ret;
1036 }
1037 
1038 
1039 /**
1040  * Update L2 table with new cluster offsets and write them out
1041  *
1042  * Called with table_lock held.
1043  */
1044 static int coroutine_fn qed_aio_write_l2_update(QEDAIOCB *acb, uint64_t offset)
1045 {
1046     BDRVQEDState *s = acb_to_s(acb);
1047     bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1;
1048     int index, ret;
1049 
1050     if (need_alloc) {
1051         qed_unref_l2_cache_entry(acb->request.l2_table);
1052         acb->request.l2_table = qed_new_l2_table(s);
1053     }
1054 
1055     index = qed_l2_index(s, acb->cur_pos);
1056     qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters,
1057                          offset);
1058 
1059     if (need_alloc) {
1060         /* Write out the whole new L2 table */
1061         ret = qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true);
1062         if (ret) {
1063             return ret;
1064         }
1065         return qed_aio_write_l1_update(acb);
1066     } else {
1067         /* Write out only the updated part of the L2 table */
1068         ret = qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters,
1069                                  false);
1070         if (ret) {
1071             return ret;
1072         }
1073     }
1074     return 0;
1075 }
1076 
1077 /**
1078  * Write data to the image file
1079  *
1080  * Called with table_lock *not* held.
1081  */
1082 static int coroutine_fn qed_aio_write_main(QEDAIOCB *acb)
1083 {
1084     BDRVQEDState *s = acb_to_s(acb);
1085     uint64_t offset = acb->cur_cluster +
1086                       qed_offset_into_cluster(s, acb->cur_pos);
1087 
1088     trace_qed_aio_write_main(s, acb, 0, offset, acb->cur_qiov.size);
1089 
1090     BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO);
1091     return bdrv_co_pwritev(s->bs->file, offset, acb->cur_qiov.size,
1092                            &acb->cur_qiov, 0);
1093 }
1094 
1095 /**
1096  * Populate untouched regions of new data cluster
1097  *
1098  * Called with table_lock held.
1099  */
1100 static int coroutine_fn qed_aio_write_cow(QEDAIOCB *acb)
1101 {
1102     BDRVQEDState *s = acb_to_s(acb);
1103     uint64_t start, len, offset;
1104     int ret;
1105 
1106     qemu_co_mutex_unlock(&s->table_lock);
1107 
1108     /* Populate front untouched region of new data cluster */
1109     start = qed_start_of_cluster(s, acb->cur_pos);
1110     len = qed_offset_into_cluster(s, acb->cur_pos);
1111 
1112     trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster);
1113     ret = qed_copy_from_backing_file(s, start, len, acb->cur_cluster);
1114     if (ret < 0) {
1115         goto out;
1116     }
1117 
1118     /* Populate back untouched region of new data cluster */
1119     start = acb->cur_pos + acb->cur_qiov.size;
1120     len = qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start;
1121     offset = acb->cur_cluster +
1122              qed_offset_into_cluster(s, acb->cur_pos) +
1123              acb->cur_qiov.size;
1124 
1125     trace_qed_aio_write_postfill(s, acb, start, len, offset);
1126     ret = qed_copy_from_backing_file(s, start, len, offset);
1127     if (ret < 0) {
1128         goto out;
1129     }
1130 
1131     ret = qed_aio_write_main(acb);
1132     if (ret < 0) {
1133         goto out;
1134     }
1135 
1136     if (s->bs->backing) {
1137         /*
1138          * Flush new data clusters before updating the L2 table
1139          *
1140          * This flush is necessary when a backing file is in use.  A crash
1141          * during an allocating write could result in empty clusters in the
1142          * image.  If the write only touched a subregion of the cluster,
1143          * then backing image sectors have been lost in the untouched
1144          * region.  The solution is to flush after writing a new data
1145          * cluster and before updating the L2 table.
1146          */
1147         ret = bdrv_co_flush(s->bs->file->bs);
1148     }
1149 
1150 out:
1151     qemu_co_mutex_lock(&s->table_lock);
1152     return ret;
1153 }
1154 
1155 /**
1156  * Check if the QED_F_NEED_CHECK bit should be set during allocating write
1157  */
1158 static bool qed_should_set_need_check(BDRVQEDState *s)
1159 {
1160     /* The flush before L2 update path ensures consistency */
1161     if (s->bs->backing) {
1162         return false;
1163     }
1164 
1165     return !(s->header.features & QED_F_NEED_CHECK);
1166 }
1167 
1168 /**
1169  * Write new data cluster
1170  *
1171  * @acb:        Write request
1172  * @len:        Length in bytes
1173  *
1174  * This path is taken when writing to previously unallocated clusters.
1175  *
1176  * Called with table_lock held.
1177  */
1178 static int coroutine_fn qed_aio_write_alloc(QEDAIOCB *acb, size_t len)
1179 {
1180     BDRVQEDState *s = acb_to_s(acb);
1181     int ret;
1182 
1183     /* Cancel timer when the first allocating request comes in */
1184     if (s->allocating_acb == NULL) {
1185         qed_cancel_need_check_timer(s);
1186     }
1187 
1188     /* Freeze this request if another allocating write is in progress */
1189     if (s->allocating_acb != acb || s->allocating_write_reqs_plugged) {
1190         if (s->allocating_acb != NULL) {
1191             qemu_co_queue_wait(&s->allocating_write_reqs, &s->table_lock);
1192             assert(s->allocating_acb == NULL);
1193         }
1194         s->allocating_acb = acb;
1195         return -EAGAIN; /* start over with looking up table entries */
1196     }
1197 
1198     acb->cur_nclusters = qed_bytes_to_clusters(s,
1199             qed_offset_into_cluster(s, acb->cur_pos) + len);
1200     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1201 
1202     if (acb->flags & QED_AIOCB_ZERO) {
1203         /* Skip ahead if the clusters are already zero */
1204         if (acb->find_cluster_ret == QED_CLUSTER_ZERO) {
1205             return 0;
1206         }
1207         acb->cur_cluster = 1;
1208     } else {
1209         acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters);
1210     }
1211 
1212     if (qed_should_set_need_check(s)) {
1213         s->header.features |= QED_F_NEED_CHECK;
1214         ret = qed_write_header(s);
1215         if (ret < 0) {
1216             return ret;
1217         }
1218     }
1219 
1220     if (!(acb->flags & QED_AIOCB_ZERO)) {
1221         ret = qed_aio_write_cow(acb);
1222         if (ret < 0) {
1223             return ret;
1224         }
1225     }
1226 
1227     return qed_aio_write_l2_update(acb, acb->cur_cluster);
1228 }
1229 
1230 /**
1231  * Write data cluster in place
1232  *
1233  * @acb:        Write request
1234  * @offset:     Cluster offset in bytes
1235  * @len:        Length in bytes
1236  *
1237  * This path is taken when writing to already allocated clusters.
1238  *
1239  * Called with table_lock held.
1240  */
1241 static int coroutine_fn qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset,
1242                                               size_t len)
1243 {
1244     BDRVQEDState *s = acb_to_s(acb);
1245     int r;
1246 
1247     qemu_co_mutex_unlock(&s->table_lock);
1248 
1249     /* Allocate buffer for zero writes */
1250     if (acb->flags & QED_AIOCB_ZERO) {
1251         struct iovec *iov = acb->qiov->iov;
1252 
1253         if (!iov->iov_base) {
1254             iov->iov_base = qemu_try_blockalign(acb->bs, iov->iov_len);
1255             if (iov->iov_base == NULL) {
1256                 r = -ENOMEM;
1257                 goto out;
1258             }
1259             memset(iov->iov_base, 0, iov->iov_len);
1260         }
1261     }
1262 
1263     /* Calculate the I/O vector */
1264     acb->cur_cluster = offset;
1265     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1266 
1267     /* Do the actual write.  */
1268     r = qed_aio_write_main(acb);
1269 out:
1270     qemu_co_mutex_lock(&s->table_lock);
1271     return r;
1272 }
1273 
1274 /**
1275  * Write data cluster
1276  *
1277  * @opaque:     Write request
1278  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1279  * @offset:     Cluster offset in bytes
1280  * @len:        Length in bytes
1281  *
1282  * Called with table_lock held.
1283  */
1284 static int coroutine_fn qed_aio_write_data(void *opaque, int ret,
1285                                            uint64_t offset, size_t len)
1286 {
1287     QEDAIOCB *acb = opaque;
1288 
1289     trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len);
1290 
1291     acb->find_cluster_ret = ret;
1292 
1293     switch (ret) {
1294     case QED_CLUSTER_FOUND:
1295         return qed_aio_write_inplace(acb, offset, len);
1296 
1297     case QED_CLUSTER_L2:
1298     case QED_CLUSTER_L1:
1299     case QED_CLUSTER_ZERO:
1300         return qed_aio_write_alloc(acb, len);
1301 
1302     default:
1303         g_assert_not_reached();
1304     }
1305 }
1306 
1307 /**
1308  * Read data cluster
1309  *
1310  * @opaque:     Read request
1311  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1312  * @offset:     Cluster offset in bytes
1313  * @len:        Length in bytes
1314  *
1315  * Called with table_lock held.
1316  */
1317 static int coroutine_fn qed_aio_read_data(void *opaque, int ret,
1318                                           uint64_t offset, size_t len)
1319 {
1320     QEDAIOCB *acb = opaque;
1321     BDRVQEDState *s = acb_to_s(acb);
1322     BlockDriverState *bs = acb->bs;
1323     int r;
1324 
1325     qemu_co_mutex_unlock(&s->table_lock);
1326 
1327     /* Adjust offset into cluster */
1328     offset += qed_offset_into_cluster(s, acb->cur_pos);
1329 
1330     trace_qed_aio_read_data(s, acb, ret, offset, len);
1331 
1332     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1333 
1334     /* Handle zero cluster and backing file reads, otherwise read
1335      * data cluster directly.
1336      */
1337     if (ret == QED_CLUSTER_ZERO) {
1338         qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size);
1339         r = 0;
1340     } else if (ret != QED_CLUSTER_FOUND) {
1341         r = qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov,
1342                                   &acb->backing_qiov);
1343     } else {
1344         BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
1345         r = bdrv_co_preadv(bs->file, offset, acb->cur_qiov.size,
1346                            &acb->cur_qiov, 0);
1347     }
1348 
1349     qemu_co_mutex_lock(&s->table_lock);
1350     return r;
1351 }
1352 
1353 /**
1354  * Begin next I/O or complete the request
1355  */
1356 static int coroutine_fn qed_aio_next_io(QEDAIOCB *acb)
1357 {
1358     BDRVQEDState *s = acb_to_s(acb);
1359     uint64_t offset;
1360     size_t len;
1361     int ret;
1362 
1363     qemu_co_mutex_lock(&s->table_lock);
1364     while (1) {
1365         trace_qed_aio_next_io(s, acb, 0, acb->cur_pos + acb->cur_qiov.size);
1366 
1367         if (acb->backing_qiov) {
1368             qemu_iovec_destroy(acb->backing_qiov);
1369             g_free(acb->backing_qiov);
1370             acb->backing_qiov = NULL;
1371         }
1372 
1373         acb->qiov_offset += acb->cur_qiov.size;
1374         acb->cur_pos += acb->cur_qiov.size;
1375         qemu_iovec_reset(&acb->cur_qiov);
1376 
1377         /* Complete request */
1378         if (acb->cur_pos >= acb->end_pos) {
1379             ret = 0;
1380             break;
1381         }
1382 
1383         /* Find next cluster and start I/O */
1384         len = acb->end_pos - acb->cur_pos;
1385         ret = qed_find_cluster(s, &acb->request, acb->cur_pos, &len, &offset);
1386         if (ret < 0) {
1387             break;
1388         }
1389 
1390         if (acb->flags & QED_AIOCB_WRITE) {
1391             ret = qed_aio_write_data(acb, ret, offset, len);
1392         } else {
1393             ret = qed_aio_read_data(acb, ret, offset, len);
1394         }
1395 
1396         if (ret < 0 && ret != -EAGAIN) {
1397             break;
1398         }
1399     }
1400 
1401     trace_qed_aio_complete(s, acb, ret);
1402     qed_aio_complete(acb);
1403     qemu_co_mutex_unlock(&s->table_lock);
1404     return ret;
1405 }
1406 
1407 static int coroutine_fn qed_co_request(BlockDriverState *bs, int64_t sector_num,
1408                                        QEMUIOVector *qiov, int nb_sectors,
1409                                        int flags)
1410 {
1411     QEDAIOCB acb = {
1412         .bs         = bs,
1413         .cur_pos    = (uint64_t) sector_num * BDRV_SECTOR_SIZE,
1414         .end_pos    = (sector_num + nb_sectors) * BDRV_SECTOR_SIZE,
1415         .qiov       = qiov,
1416         .flags      = flags,
1417     };
1418     qemu_iovec_init(&acb.cur_qiov, qiov->niov);
1419 
1420     trace_qed_aio_setup(bs->opaque, &acb, sector_num, nb_sectors, NULL, flags);
1421 
1422     /* Start request */
1423     return qed_aio_next_io(&acb);
1424 }
1425 
1426 static int coroutine_fn bdrv_qed_co_readv(BlockDriverState *bs,
1427                                           int64_t sector_num, int nb_sectors,
1428                                           QEMUIOVector *qiov)
1429 {
1430     return qed_co_request(bs, sector_num, qiov, nb_sectors, 0);
1431 }
1432 
1433 static int coroutine_fn bdrv_qed_co_writev(BlockDriverState *bs,
1434                                            int64_t sector_num, int nb_sectors,
1435                                            QEMUIOVector *qiov, int flags)
1436 {
1437     assert(!flags);
1438     return qed_co_request(bs, sector_num, qiov, nb_sectors, QED_AIOCB_WRITE);
1439 }
1440 
1441 static int coroutine_fn bdrv_qed_co_pwrite_zeroes(BlockDriverState *bs,
1442                                                   int64_t offset,
1443                                                   int bytes,
1444                                                   BdrvRequestFlags flags)
1445 {
1446     BDRVQEDState *s = bs->opaque;
1447 
1448     /*
1449      * Zero writes start without an I/O buffer.  If a buffer becomes necessary
1450      * then it will be allocated during request processing.
1451      */
1452     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, bytes);
1453 
1454     /* Fall back if the request is not aligned */
1455     if (qed_offset_into_cluster(s, offset) ||
1456         qed_offset_into_cluster(s, bytes)) {
1457         return -ENOTSUP;
1458     }
1459 
1460     return qed_co_request(bs, offset >> BDRV_SECTOR_BITS, &qiov,
1461                           bytes >> BDRV_SECTOR_BITS,
1462                           QED_AIOCB_WRITE | QED_AIOCB_ZERO);
1463 }
1464 
1465 static int coroutine_fn bdrv_qed_co_truncate(BlockDriverState *bs,
1466                                              int64_t offset,
1467                                              bool exact,
1468                                              PreallocMode prealloc,
1469                                              Error **errp)
1470 {
1471     BDRVQEDState *s = bs->opaque;
1472     uint64_t old_image_size;
1473     int ret;
1474 
1475     if (prealloc != PREALLOC_MODE_OFF) {
1476         error_setg(errp, "Unsupported preallocation mode '%s'",
1477                    PreallocMode_str(prealloc));
1478         return -ENOTSUP;
1479     }
1480 
1481     if (!qed_is_image_size_valid(offset, s->header.cluster_size,
1482                                  s->header.table_size)) {
1483         error_setg(errp, "Invalid image size specified");
1484         return -EINVAL;
1485     }
1486 
1487     if ((uint64_t)offset < s->header.image_size) {
1488         error_setg(errp, "Shrinking images is currently not supported");
1489         return -ENOTSUP;
1490     }
1491 
1492     old_image_size = s->header.image_size;
1493     s->header.image_size = offset;
1494     ret = qed_write_header_sync(s);
1495     if (ret < 0) {
1496         s->header.image_size = old_image_size;
1497         error_setg_errno(errp, -ret, "Failed to update the image size");
1498     }
1499     return ret;
1500 }
1501 
1502 static int64_t bdrv_qed_getlength(BlockDriverState *bs)
1503 {
1504     BDRVQEDState *s = bs->opaque;
1505     return s->header.image_size;
1506 }
1507 
1508 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1509 {
1510     BDRVQEDState *s = bs->opaque;
1511 
1512     memset(bdi, 0, sizeof(*bdi));
1513     bdi->cluster_size = s->header.cluster_size;
1514     bdi->is_dirty = s->header.features & QED_F_NEED_CHECK;
1515     bdi->unallocated_blocks_are_zero = true;
1516     return 0;
1517 }
1518 
1519 static int bdrv_qed_change_backing_file(BlockDriverState *bs,
1520                                         const char *backing_file,
1521                                         const char *backing_fmt)
1522 {
1523     BDRVQEDState *s = bs->opaque;
1524     QEDHeader new_header, le_header;
1525     void *buffer;
1526     size_t buffer_len, backing_file_len;
1527     int ret;
1528 
1529     /* Refuse to set backing filename if unknown compat feature bits are
1530      * active.  If the image uses an unknown compat feature then we may not
1531      * know the layout of data following the header structure and cannot safely
1532      * add a new string.
1533      */
1534     if (backing_file && (s->header.compat_features &
1535                          ~QED_COMPAT_FEATURE_MASK)) {
1536         return -ENOTSUP;
1537     }
1538 
1539     memcpy(&new_header, &s->header, sizeof(new_header));
1540 
1541     new_header.features &= ~(QED_F_BACKING_FILE |
1542                              QED_F_BACKING_FORMAT_NO_PROBE);
1543 
1544     /* Adjust feature flags */
1545     if (backing_file) {
1546         new_header.features |= QED_F_BACKING_FILE;
1547 
1548         if (qed_fmt_is_raw(backing_fmt)) {
1549             new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
1550         }
1551     }
1552 
1553     /* Calculate new header size */
1554     backing_file_len = 0;
1555 
1556     if (backing_file) {
1557         backing_file_len = strlen(backing_file);
1558     }
1559 
1560     buffer_len = sizeof(new_header);
1561     new_header.backing_filename_offset = buffer_len;
1562     new_header.backing_filename_size = backing_file_len;
1563     buffer_len += backing_file_len;
1564 
1565     /* Make sure we can rewrite header without failing */
1566     if (buffer_len > new_header.header_size * new_header.cluster_size) {
1567         return -ENOSPC;
1568     }
1569 
1570     /* Prepare new header */
1571     buffer = g_malloc(buffer_len);
1572 
1573     qed_header_cpu_to_le(&new_header, &le_header);
1574     memcpy(buffer, &le_header, sizeof(le_header));
1575     buffer_len = sizeof(le_header);
1576 
1577     if (backing_file) {
1578         memcpy(buffer + buffer_len, backing_file, backing_file_len);
1579         buffer_len += backing_file_len;
1580     }
1581 
1582     /* Write new header */
1583     ret = bdrv_pwrite_sync(bs->file, 0, buffer, buffer_len);
1584     g_free(buffer);
1585     if (ret == 0) {
1586         memcpy(&s->header, &new_header, sizeof(new_header));
1587     }
1588     return ret;
1589 }
1590 
1591 static void coroutine_fn bdrv_qed_co_invalidate_cache(BlockDriverState *bs,
1592                                                       Error **errp)
1593 {
1594     BDRVQEDState *s = bs->opaque;
1595     Error *local_err = NULL;
1596     int ret;
1597 
1598     bdrv_qed_close(bs);
1599 
1600     bdrv_qed_init_state(bs);
1601     qemu_co_mutex_lock(&s->table_lock);
1602     ret = bdrv_qed_do_open(bs, NULL, bs->open_flags, &local_err);
1603     qemu_co_mutex_unlock(&s->table_lock);
1604     if (local_err) {
1605         error_propagate_prepend(errp, local_err,
1606                                 "Could not reopen qed layer: ");
1607         return;
1608     } else if (ret < 0) {
1609         error_setg_errno(errp, -ret, "Could not reopen qed layer");
1610         return;
1611     }
1612 }
1613 
1614 static int coroutine_fn bdrv_qed_co_check(BlockDriverState *bs,
1615                                           BdrvCheckResult *result,
1616                                           BdrvCheckMode fix)
1617 {
1618     BDRVQEDState *s = bs->opaque;
1619     int ret;
1620 
1621     qemu_co_mutex_lock(&s->table_lock);
1622     ret = qed_check(s, result, !!fix);
1623     qemu_co_mutex_unlock(&s->table_lock);
1624 
1625     return ret;
1626 }
1627 
1628 static QemuOptsList qed_create_opts = {
1629     .name = "qed-create-opts",
1630     .head = QTAILQ_HEAD_INITIALIZER(qed_create_opts.head),
1631     .desc = {
1632         {
1633             .name = BLOCK_OPT_SIZE,
1634             .type = QEMU_OPT_SIZE,
1635             .help = "Virtual disk size"
1636         },
1637         {
1638             .name = BLOCK_OPT_BACKING_FILE,
1639             .type = QEMU_OPT_STRING,
1640             .help = "File name of a base image"
1641         },
1642         {
1643             .name = BLOCK_OPT_BACKING_FMT,
1644             .type = QEMU_OPT_STRING,
1645             .help = "Image format of the base image"
1646         },
1647         {
1648             .name = BLOCK_OPT_CLUSTER_SIZE,
1649             .type = QEMU_OPT_SIZE,
1650             .help = "Cluster size (in bytes)",
1651             .def_value_str = stringify(QED_DEFAULT_CLUSTER_SIZE)
1652         },
1653         {
1654             .name = BLOCK_OPT_TABLE_SIZE,
1655             .type = QEMU_OPT_SIZE,
1656             .help = "L1/L2 table size (in clusters)"
1657         },
1658         { /* end of list */ }
1659     }
1660 };
1661 
1662 static BlockDriver bdrv_qed = {
1663     .format_name              = "qed",
1664     .instance_size            = sizeof(BDRVQEDState),
1665     .create_opts              = &qed_create_opts,
1666     .supports_backing         = true,
1667 
1668     .bdrv_probe               = bdrv_qed_probe,
1669     .bdrv_open                = bdrv_qed_open,
1670     .bdrv_close               = bdrv_qed_close,
1671     .bdrv_reopen_prepare      = bdrv_qed_reopen_prepare,
1672     .bdrv_child_perm          = bdrv_format_default_perms,
1673     .bdrv_co_create           = bdrv_qed_co_create,
1674     .bdrv_co_create_opts      = bdrv_qed_co_create_opts,
1675     .bdrv_has_zero_init       = bdrv_has_zero_init_1,
1676     .bdrv_has_zero_init_truncate = bdrv_has_zero_init_1,
1677     .bdrv_co_block_status     = bdrv_qed_co_block_status,
1678     .bdrv_co_readv            = bdrv_qed_co_readv,
1679     .bdrv_co_writev           = bdrv_qed_co_writev,
1680     .bdrv_co_pwrite_zeroes    = bdrv_qed_co_pwrite_zeroes,
1681     .bdrv_co_truncate         = bdrv_qed_co_truncate,
1682     .bdrv_getlength           = bdrv_qed_getlength,
1683     .bdrv_get_info            = bdrv_qed_get_info,
1684     .bdrv_refresh_limits      = bdrv_qed_refresh_limits,
1685     .bdrv_change_backing_file = bdrv_qed_change_backing_file,
1686     .bdrv_co_invalidate_cache = bdrv_qed_co_invalidate_cache,
1687     .bdrv_co_check            = bdrv_qed_co_check,
1688     .bdrv_detach_aio_context  = bdrv_qed_detach_aio_context,
1689     .bdrv_attach_aio_context  = bdrv_qed_attach_aio_context,
1690     .bdrv_co_drain_begin      = bdrv_qed_co_drain_begin,
1691 };
1692 
1693 static void bdrv_qed_init(void)
1694 {
1695     bdrv_register(&bdrv_qed);
1696 }
1697 
1698 block_init(bdrv_qed_init);
1699