xref: /openbmc/qemu/block/qed.c (revision 0c4e9931)
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_of_bds,
551                                BDRV_CHILD_IMAGE, 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_with_bs(bs, BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL,
655                           errp);
656     if (!blk) {
657         ret = -EPERM;
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, 0, 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(BlockDriver *drv,
724                                                 const char *filename,
725                                                 QemuOpts *opts,
726                                                 Error **errp)
727 {
728     BlockdevCreateOptions *create_options = NULL;
729     QDict *qdict;
730     Visitor *v;
731     BlockDriverState *bs = NULL;
732     Error *local_err = NULL;
733     int ret;
734 
735     static const QDictRenames opt_renames[] = {
736         { BLOCK_OPT_BACKING_FILE,       "backing-file" },
737         { BLOCK_OPT_BACKING_FMT,        "backing-fmt" },
738         { BLOCK_OPT_CLUSTER_SIZE,       "cluster-size" },
739         { BLOCK_OPT_TABLE_SIZE,         "table-size" },
740         { NULL, NULL },
741     };
742 
743     /* Parse options and convert legacy syntax */
744     qdict = qemu_opts_to_qdict_filtered(opts, NULL, &qed_create_opts, true);
745 
746     if (!qdict_rename_keys(qdict, opt_renames, errp)) {
747         ret = -EINVAL;
748         goto fail;
749     }
750 
751     /* Create and open the file (protocol layer) */
752     ret = bdrv_create_file(filename, opts, &local_err);
753     if (ret < 0) {
754         error_propagate(errp, local_err);
755         goto fail;
756     }
757 
758     bs = bdrv_open(filename, NULL, NULL,
759                    BDRV_O_RDWR | BDRV_O_RESIZE | BDRV_O_PROTOCOL, errp);
760     if (bs == NULL) {
761         ret = -EIO;
762         goto fail;
763     }
764 
765     /* Now get the QAPI type BlockdevCreateOptions */
766     qdict_put_str(qdict, "driver", "qed");
767     qdict_put_str(qdict, "file", bs->node_name);
768 
769     v = qobject_input_visitor_new_flat_confused(qdict, errp);
770     if (!v) {
771         ret = -EINVAL;
772         goto fail;
773     }
774 
775     visit_type_BlockdevCreateOptions(v, NULL, &create_options, &local_err);
776     visit_free(v);
777 
778     if (local_err) {
779         error_propagate(errp, local_err);
780         ret = -EINVAL;
781         goto fail;
782     }
783 
784     /* Silently round up size */
785     assert(create_options->driver == BLOCKDEV_DRIVER_QED);
786     create_options->u.qed.size =
787         ROUND_UP(create_options->u.qed.size, BDRV_SECTOR_SIZE);
788 
789     /* Create the qed image (format layer) */
790     ret = bdrv_qed_co_create(create_options, errp);
791 
792 fail:
793     qobject_unref(qdict);
794     bdrv_unref(bs);
795     qapi_free_BlockdevCreateOptions(create_options);
796     return ret;
797 }
798 
799 static int coroutine_fn bdrv_qed_co_block_status(BlockDriverState *bs,
800                                                  bool want_zero,
801                                                  int64_t pos, int64_t bytes,
802                                                  int64_t *pnum, int64_t *map,
803                                                  BlockDriverState **file)
804 {
805     BDRVQEDState *s = bs->opaque;
806     size_t len = MIN(bytes, SIZE_MAX);
807     int status;
808     QEDRequest request = { .l2_table = NULL };
809     uint64_t offset;
810     int ret;
811 
812     qemu_co_mutex_lock(&s->table_lock);
813     ret = qed_find_cluster(s, &request, pos, &len, &offset);
814 
815     *pnum = len;
816     switch (ret) {
817     case QED_CLUSTER_FOUND:
818         *map = offset | qed_offset_into_cluster(s, pos);
819         status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
820         *file = bs->file->bs;
821         break;
822     case QED_CLUSTER_ZERO:
823         status = BDRV_BLOCK_ZERO;
824         break;
825     case QED_CLUSTER_L2:
826     case QED_CLUSTER_L1:
827         status = 0;
828         break;
829     default:
830         assert(ret < 0);
831         status = ret;
832         break;
833     }
834 
835     qed_unref_l2_cache_entry(request.l2_table);
836     qemu_co_mutex_unlock(&s->table_lock);
837 
838     return status;
839 }
840 
841 static BDRVQEDState *acb_to_s(QEDAIOCB *acb)
842 {
843     return acb->bs->opaque;
844 }
845 
846 /**
847  * Read from the backing file or zero-fill if no backing file
848  *
849  * @s:              QED state
850  * @pos:            Byte position in device
851  * @qiov:           Destination I/O vector
852  * @backing_qiov:   Possibly shortened copy of qiov, to be allocated here
853  * @cb:             Completion function
854  * @opaque:         User data for completion function
855  *
856  * This function reads qiov->size bytes starting at pos from the backing file.
857  * If there is no backing file then zeroes are read.
858  */
859 static int coroutine_fn qed_read_backing_file(BDRVQEDState *s, uint64_t pos,
860                                               QEMUIOVector *qiov,
861                                               QEMUIOVector **backing_qiov)
862 {
863     uint64_t backing_length = 0;
864     size_t size;
865     int ret;
866 
867     /* If there is a backing file, get its length.  Treat the absence of a
868      * backing file like a zero length backing file.
869      */
870     if (s->bs->backing) {
871         int64_t l = bdrv_getlength(s->bs->backing->bs);
872         if (l < 0) {
873             return l;
874         }
875         backing_length = l;
876     }
877 
878     /* Zero all sectors if reading beyond the end of the backing file */
879     if (pos >= backing_length ||
880         pos + qiov->size > backing_length) {
881         qemu_iovec_memset(qiov, 0, 0, qiov->size);
882     }
883 
884     /* Complete now if there are no backing file sectors to read */
885     if (pos >= backing_length) {
886         return 0;
887     }
888 
889     /* If the read straddles the end of the backing file, shorten it */
890     size = MIN((uint64_t)backing_length - pos, qiov->size);
891 
892     assert(*backing_qiov == NULL);
893     *backing_qiov = g_new(QEMUIOVector, 1);
894     qemu_iovec_init(*backing_qiov, qiov->niov);
895     qemu_iovec_concat(*backing_qiov, qiov, 0, size);
896 
897     BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO);
898     ret = bdrv_co_preadv(s->bs->backing, pos, size, *backing_qiov, 0);
899     if (ret < 0) {
900         return ret;
901     }
902     return 0;
903 }
904 
905 /**
906  * Copy data from backing file into the image
907  *
908  * @s:          QED state
909  * @pos:        Byte position in device
910  * @len:        Number of bytes
911  * @offset:     Byte offset in image file
912  */
913 static int coroutine_fn qed_copy_from_backing_file(BDRVQEDState *s,
914                                                    uint64_t pos, uint64_t len,
915                                                    uint64_t offset)
916 {
917     QEMUIOVector qiov;
918     QEMUIOVector *backing_qiov = NULL;
919     int ret;
920 
921     /* Skip copy entirely if there is no work to do */
922     if (len == 0) {
923         return 0;
924     }
925 
926     qemu_iovec_init_buf(&qiov, qemu_blockalign(s->bs, len), len);
927 
928     ret = qed_read_backing_file(s, pos, &qiov, &backing_qiov);
929 
930     if (backing_qiov) {
931         qemu_iovec_destroy(backing_qiov);
932         g_free(backing_qiov);
933         backing_qiov = NULL;
934     }
935 
936     if (ret) {
937         goto out;
938     }
939 
940     BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE);
941     ret = bdrv_co_pwritev(s->bs->file, offset, qiov.size, &qiov, 0);
942     if (ret < 0) {
943         goto out;
944     }
945     ret = 0;
946 out:
947     qemu_vfree(qemu_iovec_buf(&qiov));
948     return ret;
949 }
950 
951 /**
952  * Link one or more contiguous clusters into a table
953  *
954  * @s:              QED state
955  * @table:          L2 table
956  * @index:          First cluster index
957  * @n:              Number of contiguous clusters
958  * @cluster:        First cluster offset
959  *
960  * The cluster offset may be an allocated byte offset in the image file, the
961  * zero cluster marker, or the unallocated cluster marker.
962  *
963  * Called with table_lock held.
964  */
965 static void coroutine_fn qed_update_l2_table(BDRVQEDState *s, QEDTable *table,
966                                              int index, unsigned int n,
967                                              uint64_t cluster)
968 {
969     int i;
970     for (i = index; i < index + n; i++) {
971         table->offsets[i] = cluster;
972         if (!qed_offset_is_unalloc_cluster(cluster) &&
973             !qed_offset_is_zero_cluster(cluster)) {
974             cluster += s->header.cluster_size;
975         }
976     }
977 }
978 
979 /* Called with table_lock held.  */
980 static void coroutine_fn qed_aio_complete(QEDAIOCB *acb)
981 {
982     BDRVQEDState *s = acb_to_s(acb);
983 
984     /* Free resources */
985     qemu_iovec_destroy(&acb->cur_qiov);
986     qed_unref_l2_cache_entry(acb->request.l2_table);
987 
988     /* Free the buffer we may have allocated for zero writes */
989     if (acb->flags & QED_AIOCB_ZERO) {
990         qemu_vfree(acb->qiov->iov[0].iov_base);
991         acb->qiov->iov[0].iov_base = NULL;
992     }
993 
994     /* Start next allocating write request waiting behind this one.  Note that
995      * requests enqueue themselves when they first hit an unallocated cluster
996      * but they wait until the entire request is finished before waking up the
997      * next request in the queue.  This ensures that we don't cycle through
998      * requests multiple times but rather finish one at a time completely.
999      */
1000     if (acb == s->allocating_acb) {
1001         s->allocating_acb = NULL;
1002         if (!qemu_co_queue_empty(&s->allocating_write_reqs)) {
1003             qemu_co_queue_next(&s->allocating_write_reqs);
1004         } else if (s->header.features & QED_F_NEED_CHECK) {
1005             qed_start_need_check_timer(s);
1006         }
1007     }
1008 }
1009 
1010 /**
1011  * Update L1 table with new L2 table offset and write it out
1012  *
1013  * Called with table_lock held.
1014  */
1015 static int coroutine_fn qed_aio_write_l1_update(QEDAIOCB *acb)
1016 {
1017     BDRVQEDState *s = acb_to_s(acb);
1018     CachedL2Table *l2_table = acb->request.l2_table;
1019     uint64_t l2_offset = l2_table->offset;
1020     int index, ret;
1021 
1022     index = qed_l1_index(s, acb->cur_pos);
1023     s->l1_table->offsets[index] = l2_table->offset;
1024 
1025     ret = qed_write_l1_table(s, index, 1);
1026 
1027     /* Commit the current L2 table to the cache */
1028     qed_commit_l2_cache_entry(&s->l2_cache, l2_table);
1029 
1030     /* This is guaranteed to succeed because we just committed the entry to the
1031      * cache.
1032      */
1033     acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset);
1034     assert(acb->request.l2_table != NULL);
1035 
1036     return ret;
1037 }
1038 
1039 
1040 /**
1041  * Update L2 table with new cluster offsets and write them out
1042  *
1043  * Called with table_lock held.
1044  */
1045 static int coroutine_fn qed_aio_write_l2_update(QEDAIOCB *acb, uint64_t offset)
1046 {
1047     BDRVQEDState *s = acb_to_s(acb);
1048     bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1;
1049     int index, ret;
1050 
1051     if (need_alloc) {
1052         qed_unref_l2_cache_entry(acb->request.l2_table);
1053         acb->request.l2_table = qed_new_l2_table(s);
1054     }
1055 
1056     index = qed_l2_index(s, acb->cur_pos);
1057     qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters,
1058                          offset);
1059 
1060     if (need_alloc) {
1061         /* Write out the whole new L2 table */
1062         ret = qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true);
1063         if (ret) {
1064             return ret;
1065         }
1066         return qed_aio_write_l1_update(acb);
1067     } else {
1068         /* Write out only the updated part of the L2 table */
1069         ret = qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters,
1070                                  false);
1071         if (ret) {
1072             return ret;
1073         }
1074     }
1075     return 0;
1076 }
1077 
1078 /**
1079  * Write data to the image file
1080  *
1081  * Called with table_lock *not* held.
1082  */
1083 static int coroutine_fn qed_aio_write_main(QEDAIOCB *acb)
1084 {
1085     BDRVQEDState *s = acb_to_s(acb);
1086     uint64_t offset = acb->cur_cluster +
1087                       qed_offset_into_cluster(s, acb->cur_pos);
1088 
1089     trace_qed_aio_write_main(s, acb, 0, offset, acb->cur_qiov.size);
1090 
1091     BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO);
1092     return bdrv_co_pwritev(s->bs->file, offset, acb->cur_qiov.size,
1093                            &acb->cur_qiov, 0);
1094 }
1095 
1096 /**
1097  * Populate untouched regions of new data cluster
1098  *
1099  * Called with table_lock held.
1100  */
1101 static int coroutine_fn qed_aio_write_cow(QEDAIOCB *acb)
1102 {
1103     BDRVQEDState *s = acb_to_s(acb);
1104     uint64_t start, len, offset;
1105     int ret;
1106 
1107     qemu_co_mutex_unlock(&s->table_lock);
1108 
1109     /* Populate front untouched region of new data cluster */
1110     start = qed_start_of_cluster(s, acb->cur_pos);
1111     len = qed_offset_into_cluster(s, acb->cur_pos);
1112 
1113     trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster);
1114     ret = qed_copy_from_backing_file(s, start, len, acb->cur_cluster);
1115     if (ret < 0) {
1116         goto out;
1117     }
1118 
1119     /* Populate back untouched region of new data cluster */
1120     start = acb->cur_pos + acb->cur_qiov.size;
1121     len = qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start;
1122     offset = acb->cur_cluster +
1123              qed_offset_into_cluster(s, acb->cur_pos) +
1124              acb->cur_qiov.size;
1125 
1126     trace_qed_aio_write_postfill(s, acb, start, len, offset);
1127     ret = qed_copy_from_backing_file(s, start, len, offset);
1128     if (ret < 0) {
1129         goto out;
1130     }
1131 
1132     ret = qed_aio_write_main(acb);
1133     if (ret < 0) {
1134         goto out;
1135     }
1136 
1137     if (s->bs->backing) {
1138         /*
1139          * Flush new data clusters before updating the L2 table
1140          *
1141          * This flush is necessary when a backing file is in use.  A crash
1142          * during an allocating write could result in empty clusters in the
1143          * image.  If the write only touched a subregion of the cluster,
1144          * then backing image sectors have been lost in the untouched
1145          * region.  The solution is to flush after writing a new data
1146          * cluster and before updating the L2 table.
1147          */
1148         ret = bdrv_co_flush(s->bs->file->bs);
1149     }
1150 
1151 out:
1152     qemu_co_mutex_lock(&s->table_lock);
1153     return ret;
1154 }
1155 
1156 /**
1157  * Check if the QED_F_NEED_CHECK bit should be set during allocating write
1158  */
1159 static bool qed_should_set_need_check(BDRVQEDState *s)
1160 {
1161     /* The flush before L2 update path ensures consistency */
1162     if (s->bs->backing) {
1163         return false;
1164     }
1165 
1166     return !(s->header.features & QED_F_NEED_CHECK);
1167 }
1168 
1169 /**
1170  * Write new data cluster
1171  *
1172  * @acb:        Write request
1173  * @len:        Length in bytes
1174  *
1175  * This path is taken when writing to previously unallocated clusters.
1176  *
1177  * Called with table_lock held.
1178  */
1179 static int coroutine_fn qed_aio_write_alloc(QEDAIOCB *acb, size_t len)
1180 {
1181     BDRVQEDState *s = acb_to_s(acb);
1182     int ret;
1183 
1184     /* Cancel timer when the first allocating request comes in */
1185     if (s->allocating_acb == NULL) {
1186         qed_cancel_need_check_timer(s);
1187     }
1188 
1189     /* Freeze this request if another allocating write is in progress */
1190     if (s->allocating_acb != acb || s->allocating_write_reqs_plugged) {
1191         if (s->allocating_acb != NULL) {
1192             qemu_co_queue_wait(&s->allocating_write_reqs, &s->table_lock);
1193             assert(s->allocating_acb == NULL);
1194         }
1195         s->allocating_acb = acb;
1196         return -EAGAIN; /* start over with looking up table entries */
1197     }
1198 
1199     acb->cur_nclusters = qed_bytes_to_clusters(s,
1200             qed_offset_into_cluster(s, acb->cur_pos) + len);
1201     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1202 
1203     if (acb->flags & QED_AIOCB_ZERO) {
1204         /* Skip ahead if the clusters are already zero */
1205         if (acb->find_cluster_ret == QED_CLUSTER_ZERO) {
1206             return 0;
1207         }
1208         acb->cur_cluster = 1;
1209     } else {
1210         acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters);
1211     }
1212 
1213     if (qed_should_set_need_check(s)) {
1214         s->header.features |= QED_F_NEED_CHECK;
1215         ret = qed_write_header(s);
1216         if (ret < 0) {
1217             return ret;
1218         }
1219     }
1220 
1221     if (!(acb->flags & QED_AIOCB_ZERO)) {
1222         ret = qed_aio_write_cow(acb);
1223         if (ret < 0) {
1224             return ret;
1225         }
1226     }
1227 
1228     return qed_aio_write_l2_update(acb, acb->cur_cluster);
1229 }
1230 
1231 /**
1232  * Write data cluster in place
1233  *
1234  * @acb:        Write request
1235  * @offset:     Cluster offset in bytes
1236  * @len:        Length in bytes
1237  *
1238  * This path is taken when writing to already allocated clusters.
1239  *
1240  * Called with table_lock held.
1241  */
1242 static int coroutine_fn qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset,
1243                                               size_t len)
1244 {
1245     BDRVQEDState *s = acb_to_s(acb);
1246     int r;
1247 
1248     qemu_co_mutex_unlock(&s->table_lock);
1249 
1250     /* Allocate buffer for zero writes */
1251     if (acb->flags & QED_AIOCB_ZERO) {
1252         struct iovec *iov = acb->qiov->iov;
1253 
1254         if (!iov->iov_base) {
1255             iov->iov_base = qemu_try_blockalign(acb->bs, iov->iov_len);
1256             if (iov->iov_base == NULL) {
1257                 r = -ENOMEM;
1258                 goto out;
1259             }
1260             memset(iov->iov_base, 0, iov->iov_len);
1261         }
1262     }
1263 
1264     /* Calculate the I/O vector */
1265     acb->cur_cluster = offset;
1266     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1267 
1268     /* Do the actual write.  */
1269     r = qed_aio_write_main(acb);
1270 out:
1271     qemu_co_mutex_lock(&s->table_lock);
1272     return r;
1273 }
1274 
1275 /**
1276  * Write data cluster
1277  *
1278  * @opaque:     Write request
1279  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1280  * @offset:     Cluster offset in bytes
1281  * @len:        Length in bytes
1282  *
1283  * Called with table_lock held.
1284  */
1285 static int coroutine_fn qed_aio_write_data(void *opaque, int ret,
1286                                            uint64_t offset, size_t len)
1287 {
1288     QEDAIOCB *acb = opaque;
1289 
1290     trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len);
1291 
1292     acb->find_cluster_ret = ret;
1293 
1294     switch (ret) {
1295     case QED_CLUSTER_FOUND:
1296         return qed_aio_write_inplace(acb, offset, len);
1297 
1298     case QED_CLUSTER_L2:
1299     case QED_CLUSTER_L1:
1300     case QED_CLUSTER_ZERO:
1301         return qed_aio_write_alloc(acb, len);
1302 
1303     default:
1304         g_assert_not_reached();
1305     }
1306 }
1307 
1308 /**
1309  * Read data cluster
1310  *
1311  * @opaque:     Read request
1312  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1313  * @offset:     Cluster offset in bytes
1314  * @len:        Length in bytes
1315  *
1316  * Called with table_lock held.
1317  */
1318 static int coroutine_fn qed_aio_read_data(void *opaque, int ret,
1319                                           uint64_t offset, size_t len)
1320 {
1321     QEDAIOCB *acb = opaque;
1322     BDRVQEDState *s = acb_to_s(acb);
1323     BlockDriverState *bs = acb->bs;
1324     int r;
1325 
1326     qemu_co_mutex_unlock(&s->table_lock);
1327 
1328     /* Adjust offset into cluster */
1329     offset += qed_offset_into_cluster(s, acb->cur_pos);
1330 
1331     trace_qed_aio_read_data(s, acb, ret, offset, len);
1332 
1333     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1334 
1335     /* Handle zero cluster and backing file reads, otherwise read
1336      * data cluster directly.
1337      */
1338     if (ret == QED_CLUSTER_ZERO) {
1339         qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size);
1340         r = 0;
1341     } else if (ret != QED_CLUSTER_FOUND) {
1342         r = qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov,
1343                                   &acb->backing_qiov);
1344     } else {
1345         BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
1346         r = bdrv_co_preadv(bs->file, offset, acb->cur_qiov.size,
1347                            &acb->cur_qiov, 0);
1348     }
1349 
1350     qemu_co_mutex_lock(&s->table_lock);
1351     return r;
1352 }
1353 
1354 /**
1355  * Begin next I/O or complete the request
1356  */
1357 static int coroutine_fn qed_aio_next_io(QEDAIOCB *acb)
1358 {
1359     BDRVQEDState *s = acb_to_s(acb);
1360     uint64_t offset;
1361     size_t len;
1362     int ret;
1363 
1364     qemu_co_mutex_lock(&s->table_lock);
1365     while (1) {
1366         trace_qed_aio_next_io(s, acb, 0, acb->cur_pos + acb->cur_qiov.size);
1367 
1368         if (acb->backing_qiov) {
1369             qemu_iovec_destroy(acb->backing_qiov);
1370             g_free(acb->backing_qiov);
1371             acb->backing_qiov = NULL;
1372         }
1373 
1374         acb->qiov_offset += acb->cur_qiov.size;
1375         acb->cur_pos += acb->cur_qiov.size;
1376         qemu_iovec_reset(&acb->cur_qiov);
1377 
1378         /* Complete request */
1379         if (acb->cur_pos >= acb->end_pos) {
1380             ret = 0;
1381             break;
1382         }
1383 
1384         /* Find next cluster and start I/O */
1385         len = acb->end_pos - acb->cur_pos;
1386         ret = qed_find_cluster(s, &acb->request, acb->cur_pos, &len, &offset);
1387         if (ret < 0) {
1388             break;
1389         }
1390 
1391         if (acb->flags & QED_AIOCB_WRITE) {
1392             ret = qed_aio_write_data(acb, ret, offset, len);
1393         } else {
1394             ret = qed_aio_read_data(acb, ret, offset, len);
1395         }
1396 
1397         if (ret < 0 && ret != -EAGAIN) {
1398             break;
1399         }
1400     }
1401 
1402     trace_qed_aio_complete(s, acb, ret);
1403     qed_aio_complete(acb);
1404     qemu_co_mutex_unlock(&s->table_lock);
1405     return ret;
1406 }
1407 
1408 static int coroutine_fn qed_co_request(BlockDriverState *bs, int64_t sector_num,
1409                                        QEMUIOVector *qiov, int nb_sectors,
1410                                        int flags)
1411 {
1412     QEDAIOCB acb = {
1413         .bs         = bs,
1414         .cur_pos    = (uint64_t) sector_num * BDRV_SECTOR_SIZE,
1415         .end_pos    = (sector_num + nb_sectors) * BDRV_SECTOR_SIZE,
1416         .qiov       = qiov,
1417         .flags      = flags,
1418     };
1419     qemu_iovec_init(&acb.cur_qiov, qiov->niov);
1420 
1421     trace_qed_aio_setup(bs->opaque, &acb, sector_num, nb_sectors, NULL, flags);
1422 
1423     /* Start request */
1424     return qed_aio_next_io(&acb);
1425 }
1426 
1427 static int coroutine_fn bdrv_qed_co_readv(BlockDriverState *bs,
1428                                           int64_t sector_num, int nb_sectors,
1429                                           QEMUIOVector *qiov)
1430 {
1431     return qed_co_request(bs, sector_num, qiov, nb_sectors, 0);
1432 }
1433 
1434 static int coroutine_fn bdrv_qed_co_writev(BlockDriverState *bs,
1435                                            int64_t sector_num, int nb_sectors,
1436                                            QEMUIOVector *qiov, int flags)
1437 {
1438     assert(!flags);
1439     return qed_co_request(bs, sector_num, qiov, nb_sectors, QED_AIOCB_WRITE);
1440 }
1441 
1442 static int coroutine_fn bdrv_qed_co_pwrite_zeroes(BlockDriverState *bs,
1443                                                   int64_t offset,
1444                                                   int bytes,
1445                                                   BdrvRequestFlags flags)
1446 {
1447     BDRVQEDState *s = bs->opaque;
1448 
1449     /*
1450      * Zero writes start without an I/O buffer.  If a buffer becomes necessary
1451      * then it will be allocated during request processing.
1452      */
1453     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, bytes);
1454 
1455     /* Fall back if the request is not aligned */
1456     if (qed_offset_into_cluster(s, offset) ||
1457         qed_offset_into_cluster(s, bytes)) {
1458         return -ENOTSUP;
1459     }
1460 
1461     return qed_co_request(bs, offset >> BDRV_SECTOR_BITS, &qiov,
1462                           bytes >> BDRV_SECTOR_BITS,
1463                           QED_AIOCB_WRITE | QED_AIOCB_ZERO);
1464 }
1465 
1466 static int coroutine_fn bdrv_qed_co_truncate(BlockDriverState *bs,
1467                                              int64_t offset,
1468                                              bool exact,
1469                                              PreallocMode prealloc,
1470                                              BdrvRequestFlags flags,
1471                                              Error **errp)
1472 {
1473     BDRVQEDState *s = bs->opaque;
1474     uint64_t old_image_size;
1475     int ret;
1476 
1477     if (prealloc != PREALLOC_MODE_OFF) {
1478         error_setg(errp, "Unsupported preallocation mode '%s'",
1479                    PreallocMode_str(prealloc));
1480         return -ENOTSUP;
1481     }
1482 
1483     if (!qed_is_image_size_valid(offset, s->header.cluster_size,
1484                                  s->header.table_size)) {
1485         error_setg(errp, "Invalid image size specified");
1486         return -EINVAL;
1487     }
1488 
1489     if ((uint64_t)offset < s->header.image_size) {
1490         error_setg(errp, "Shrinking images is currently not supported");
1491         return -ENOTSUP;
1492     }
1493 
1494     old_image_size = s->header.image_size;
1495     s->header.image_size = offset;
1496     ret = qed_write_header_sync(s);
1497     if (ret < 0) {
1498         s->header.image_size = old_image_size;
1499         error_setg_errno(errp, -ret, "Failed to update the image size");
1500     }
1501     return ret;
1502 }
1503 
1504 static int64_t bdrv_qed_getlength(BlockDriverState *bs)
1505 {
1506     BDRVQEDState *s = bs->opaque;
1507     return s->header.image_size;
1508 }
1509 
1510 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1511 {
1512     BDRVQEDState *s = bs->opaque;
1513 
1514     memset(bdi, 0, sizeof(*bdi));
1515     bdi->cluster_size = s->header.cluster_size;
1516     bdi->is_dirty = s->header.features & QED_F_NEED_CHECK;
1517     bdi->unallocated_blocks_are_zero = true;
1518     return 0;
1519 }
1520 
1521 static int bdrv_qed_change_backing_file(BlockDriverState *bs,
1522                                         const char *backing_file,
1523                                         const char *backing_fmt)
1524 {
1525     BDRVQEDState *s = bs->opaque;
1526     QEDHeader new_header, le_header;
1527     void *buffer;
1528     size_t buffer_len, backing_file_len;
1529     int ret;
1530 
1531     /* Refuse to set backing filename if unknown compat feature bits are
1532      * active.  If the image uses an unknown compat feature then we may not
1533      * know the layout of data following the header structure and cannot safely
1534      * add a new string.
1535      */
1536     if (backing_file && (s->header.compat_features &
1537                          ~QED_COMPAT_FEATURE_MASK)) {
1538         return -ENOTSUP;
1539     }
1540 
1541     memcpy(&new_header, &s->header, sizeof(new_header));
1542 
1543     new_header.features &= ~(QED_F_BACKING_FILE |
1544                              QED_F_BACKING_FORMAT_NO_PROBE);
1545 
1546     /* Adjust feature flags */
1547     if (backing_file) {
1548         new_header.features |= QED_F_BACKING_FILE;
1549 
1550         if (qed_fmt_is_raw(backing_fmt)) {
1551             new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
1552         }
1553     }
1554 
1555     /* Calculate new header size */
1556     backing_file_len = 0;
1557 
1558     if (backing_file) {
1559         backing_file_len = strlen(backing_file);
1560     }
1561 
1562     buffer_len = sizeof(new_header);
1563     new_header.backing_filename_offset = buffer_len;
1564     new_header.backing_filename_size = backing_file_len;
1565     buffer_len += backing_file_len;
1566 
1567     /* Make sure we can rewrite header without failing */
1568     if (buffer_len > new_header.header_size * new_header.cluster_size) {
1569         return -ENOSPC;
1570     }
1571 
1572     /* Prepare new header */
1573     buffer = g_malloc(buffer_len);
1574 
1575     qed_header_cpu_to_le(&new_header, &le_header);
1576     memcpy(buffer, &le_header, sizeof(le_header));
1577     buffer_len = sizeof(le_header);
1578 
1579     if (backing_file) {
1580         memcpy(buffer + buffer_len, backing_file, backing_file_len);
1581         buffer_len += backing_file_len;
1582     }
1583 
1584     /* Write new header */
1585     ret = bdrv_pwrite_sync(bs->file, 0, buffer, buffer_len);
1586     g_free(buffer);
1587     if (ret == 0) {
1588         memcpy(&s->header, &new_header, sizeof(new_header));
1589     }
1590     return ret;
1591 }
1592 
1593 static void coroutine_fn bdrv_qed_co_invalidate_cache(BlockDriverState *bs,
1594                                                       Error **errp)
1595 {
1596     BDRVQEDState *s = bs->opaque;
1597     Error *local_err = NULL;
1598     int ret;
1599 
1600     bdrv_qed_close(bs);
1601 
1602     bdrv_qed_init_state(bs);
1603     qemu_co_mutex_lock(&s->table_lock);
1604     ret = bdrv_qed_do_open(bs, NULL, bs->open_flags, &local_err);
1605     qemu_co_mutex_unlock(&s->table_lock);
1606     if (local_err) {
1607         error_propagate_prepend(errp, local_err,
1608                                 "Could not reopen qed layer: ");
1609         return;
1610     } else if (ret < 0) {
1611         error_setg_errno(errp, -ret, "Could not reopen qed layer");
1612         return;
1613     }
1614 }
1615 
1616 static int coroutine_fn bdrv_qed_co_check(BlockDriverState *bs,
1617                                           BdrvCheckResult *result,
1618                                           BdrvCheckMode fix)
1619 {
1620     BDRVQEDState *s = bs->opaque;
1621     int ret;
1622 
1623     qemu_co_mutex_lock(&s->table_lock);
1624     ret = qed_check(s, result, !!fix);
1625     qemu_co_mutex_unlock(&s->table_lock);
1626 
1627     return ret;
1628 }
1629 
1630 static QemuOptsList qed_create_opts = {
1631     .name = "qed-create-opts",
1632     .head = QTAILQ_HEAD_INITIALIZER(qed_create_opts.head),
1633     .desc = {
1634         {
1635             .name = BLOCK_OPT_SIZE,
1636             .type = QEMU_OPT_SIZE,
1637             .help = "Virtual disk size"
1638         },
1639         {
1640             .name = BLOCK_OPT_BACKING_FILE,
1641             .type = QEMU_OPT_STRING,
1642             .help = "File name of a base image"
1643         },
1644         {
1645             .name = BLOCK_OPT_BACKING_FMT,
1646             .type = QEMU_OPT_STRING,
1647             .help = "Image format of the base image"
1648         },
1649         {
1650             .name = BLOCK_OPT_CLUSTER_SIZE,
1651             .type = QEMU_OPT_SIZE,
1652             .help = "Cluster size (in bytes)",
1653             .def_value_str = stringify(QED_DEFAULT_CLUSTER_SIZE)
1654         },
1655         {
1656             .name = BLOCK_OPT_TABLE_SIZE,
1657             .type = QEMU_OPT_SIZE,
1658             .help = "L1/L2 table size (in clusters)"
1659         },
1660         { /* end of list */ }
1661     }
1662 };
1663 
1664 static BlockDriver bdrv_qed = {
1665     .format_name              = "qed",
1666     .instance_size            = sizeof(BDRVQEDState),
1667     .create_opts              = &qed_create_opts,
1668     .is_format                = true,
1669     .supports_backing         = true,
1670 
1671     .bdrv_probe               = bdrv_qed_probe,
1672     .bdrv_open                = bdrv_qed_open,
1673     .bdrv_close               = bdrv_qed_close,
1674     .bdrv_reopen_prepare      = bdrv_qed_reopen_prepare,
1675     .bdrv_child_perm          = bdrv_default_perms,
1676     .bdrv_co_create           = bdrv_qed_co_create,
1677     .bdrv_co_create_opts      = bdrv_qed_co_create_opts,
1678     .bdrv_has_zero_init       = bdrv_has_zero_init_1,
1679     .bdrv_co_block_status     = bdrv_qed_co_block_status,
1680     .bdrv_co_readv            = bdrv_qed_co_readv,
1681     .bdrv_co_writev           = bdrv_qed_co_writev,
1682     .bdrv_co_pwrite_zeroes    = bdrv_qed_co_pwrite_zeroes,
1683     .bdrv_co_truncate         = bdrv_qed_co_truncate,
1684     .bdrv_getlength           = bdrv_qed_getlength,
1685     .bdrv_get_info            = bdrv_qed_get_info,
1686     .bdrv_refresh_limits      = bdrv_qed_refresh_limits,
1687     .bdrv_change_backing_file = bdrv_qed_change_backing_file,
1688     .bdrv_co_invalidate_cache = bdrv_qed_co_invalidate_cache,
1689     .bdrv_co_check            = bdrv_qed_co_check,
1690     .bdrv_detach_aio_context  = bdrv_qed_detach_aio_context,
1691     .bdrv_attach_aio_context  = bdrv_qed_attach_aio_context,
1692     .bdrv_co_drain_begin      = bdrv_qed_co_drain_begin,
1693 };
1694 
1695 static void bdrv_qed_init(void)
1696 {
1697     bdrv_register(&bdrv_qed);
1698 }
1699 
1700 block_init(bdrv_qed_init);
1701