xref: /openbmc/qemu/block/qed.c (revision f363d039)
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     /* File must start empty and grow, check truncate is supported */
677     ret = blk_truncate(blk, 0, PREALLOC_MODE_OFF, errp);
678     if (ret < 0) {
679         goto out;
680     }
681 
682     if (qed_opts->has_backing_file) {
683         header.features |= QED_F_BACKING_FILE;
684         header.backing_filename_offset = sizeof(le_header);
685         header.backing_filename_size = strlen(qed_opts->backing_file);
686 
687         if (qed_opts->has_backing_fmt) {
688             const char *backing_fmt = BlockdevDriver_str(qed_opts->backing_fmt);
689             if (qed_fmt_is_raw(backing_fmt)) {
690                 header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
691             }
692         }
693     }
694 
695     qed_header_cpu_to_le(&header, &le_header);
696     ret = blk_pwrite(blk, 0, &le_header, sizeof(le_header), 0);
697     if (ret < 0) {
698         goto out;
699     }
700     ret = blk_pwrite(blk, sizeof(le_header), qed_opts->backing_file,
701                      header.backing_filename_size, 0);
702     if (ret < 0) {
703         goto out;
704     }
705 
706     l1_table = g_malloc0(l1_size);
707     ret = blk_pwrite(blk, header.l1_table_offset, l1_table, l1_size, 0);
708     if (ret < 0) {
709         goto out;
710     }
711 
712     ret = 0; /* success */
713 out:
714     g_free(l1_table);
715     blk_unref(blk);
716     bdrv_unref(bs);
717     return ret;
718 }
719 
720 static int coroutine_fn bdrv_qed_co_create_opts(const char *filename,
721                                                 QemuOpts *opts,
722                                                 Error **errp)
723 {
724     BlockdevCreateOptions *create_options = NULL;
725     QDict *qdict;
726     Visitor *v;
727     BlockDriverState *bs = NULL;
728     Error *local_err = NULL;
729     int ret;
730 
731     static const QDictRenames opt_renames[] = {
732         { BLOCK_OPT_BACKING_FILE,       "backing-file" },
733         { BLOCK_OPT_BACKING_FMT,        "backing-fmt" },
734         { BLOCK_OPT_CLUSTER_SIZE,       "cluster-size" },
735         { BLOCK_OPT_TABLE_SIZE,         "table-size" },
736         { NULL, NULL },
737     };
738 
739     /* Parse options and convert legacy syntax */
740     qdict = qemu_opts_to_qdict_filtered(opts, NULL, &qed_create_opts, true);
741 
742     if (!qdict_rename_keys(qdict, opt_renames, errp)) {
743         ret = -EINVAL;
744         goto fail;
745     }
746 
747     /* Create and open the file (protocol layer) */
748     ret = bdrv_create_file(filename, opts, &local_err);
749     if (ret < 0) {
750         error_propagate(errp, local_err);
751         goto fail;
752     }
753 
754     bs = bdrv_open(filename, NULL, NULL,
755                    BDRV_O_RDWR | BDRV_O_RESIZE | BDRV_O_PROTOCOL, errp);
756     if (bs == NULL) {
757         ret = -EIO;
758         goto fail;
759     }
760 
761     /* Now get the QAPI type BlockdevCreateOptions */
762     qdict_put_str(qdict, "driver", "qed");
763     qdict_put_str(qdict, "file", bs->node_name);
764 
765     v = qobject_input_visitor_new_flat_confused(qdict, errp);
766     if (!v) {
767         ret = -EINVAL;
768         goto fail;
769     }
770 
771     visit_type_BlockdevCreateOptions(v, NULL, &create_options, &local_err);
772     visit_free(v);
773 
774     if (local_err) {
775         error_propagate(errp, local_err);
776         ret = -EINVAL;
777         goto fail;
778     }
779 
780     /* Silently round up size */
781     assert(create_options->driver == BLOCKDEV_DRIVER_QED);
782     create_options->u.qed.size =
783         ROUND_UP(create_options->u.qed.size, BDRV_SECTOR_SIZE);
784 
785     /* Create the qed image (format layer) */
786     ret = bdrv_qed_co_create(create_options, errp);
787 
788 fail:
789     qobject_unref(qdict);
790     bdrv_unref(bs);
791     qapi_free_BlockdevCreateOptions(create_options);
792     return ret;
793 }
794 
795 static int coroutine_fn bdrv_qed_co_block_status(BlockDriverState *bs,
796                                                  bool want_zero,
797                                                  int64_t pos, int64_t bytes,
798                                                  int64_t *pnum, int64_t *map,
799                                                  BlockDriverState **file)
800 {
801     BDRVQEDState *s = bs->opaque;
802     size_t len = MIN(bytes, SIZE_MAX);
803     int status;
804     QEDRequest request = { .l2_table = NULL };
805     uint64_t offset;
806     int ret;
807 
808     qemu_co_mutex_lock(&s->table_lock);
809     ret = qed_find_cluster(s, &request, pos, &len, &offset);
810 
811     *pnum = len;
812     switch (ret) {
813     case QED_CLUSTER_FOUND:
814         *map = offset | qed_offset_into_cluster(s, pos);
815         status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
816         *file = bs->file->bs;
817         break;
818     case QED_CLUSTER_ZERO:
819         status = BDRV_BLOCK_ZERO;
820         break;
821     case QED_CLUSTER_L2:
822     case QED_CLUSTER_L1:
823         status = 0;
824         break;
825     default:
826         assert(ret < 0);
827         status = ret;
828         break;
829     }
830 
831     qed_unref_l2_cache_entry(request.l2_table);
832     qemu_co_mutex_unlock(&s->table_lock);
833 
834     return status;
835 }
836 
837 static BDRVQEDState *acb_to_s(QEDAIOCB *acb)
838 {
839     return acb->bs->opaque;
840 }
841 
842 /**
843  * Read from the backing file or zero-fill if no backing file
844  *
845  * @s:              QED state
846  * @pos:            Byte position in device
847  * @qiov:           Destination I/O vector
848  * @backing_qiov:   Possibly shortened copy of qiov, to be allocated here
849  * @cb:             Completion function
850  * @opaque:         User data for completion function
851  *
852  * This function reads qiov->size bytes starting at pos from the backing file.
853  * If there is no backing file then zeroes are read.
854  */
855 static int coroutine_fn qed_read_backing_file(BDRVQEDState *s, uint64_t pos,
856                                               QEMUIOVector *qiov,
857                                               QEMUIOVector **backing_qiov)
858 {
859     uint64_t backing_length = 0;
860     size_t size;
861     int ret;
862 
863     /* If there is a backing file, get its length.  Treat the absence of a
864      * backing file like a zero length backing file.
865      */
866     if (s->bs->backing) {
867         int64_t l = bdrv_getlength(s->bs->backing->bs);
868         if (l < 0) {
869             return l;
870         }
871         backing_length = l;
872     }
873 
874     /* Zero all sectors if reading beyond the end of the backing file */
875     if (pos >= backing_length ||
876         pos + qiov->size > backing_length) {
877         qemu_iovec_memset(qiov, 0, 0, qiov->size);
878     }
879 
880     /* Complete now if there are no backing file sectors to read */
881     if (pos >= backing_length) {
882         return 0;
883     }
884 
885     /* If the read straddles the end of the backing file, shorten it */
886     size = MIN((uint64_t)backing_length - pos, qiov->size);
887 
888     assert(*backing_qiov == NULL);
889     *backing_qiov = g_new(QEMUIOVector, 1);
890     qemu_iovec_init(*backing_qiov, qiov->niov);
891     qemu_iovec_concat(*backing_qiov, qiov, 0, size);
892 
893     BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO);
894     ret = bdrv_co_preadv(s->bs->backing, pos, size, *backing_qiov, 0);
895     if (ret < 0) {
896         return ret;
897     }
898     return 0;
899 }
900 
901 /**
902  * Copy data from backing file into the image
903  *
904  * @s:          QED state
905  * @pos:        Byte position in device
906  * @len:        Number of bytes
907  * @offset:     Byte offset in image file
908  */
909 static int coroutine_fn qed_copy_from_backing_file(BDRVQEDState *s,
910                                                    uint64_t pos, uint64_t len,
911                                                    uint64_t offset)
912 {
913     QEMUIOVector qiov;
914     QEMUIOVector *backing_qiov = NULL;
915     int ret;
916 
917     /* Skip copy entirely if there is no work to do */
918     if (len == 0) {
919         return 0;
920     }
921 
922     qemu_iovec_init_buf(&qiov, qemu_blockalign(s->bs, len), len);
923 
924     ret = qed_read_backing_file(s, pos, &qiov, &backing_qiov);
925 
926     if (backing_qiov) {
927         qemu_iovec_destroy(backing_qiov);
928         g_free(backing_qiov);
929         backing_qiov = NULL;
930     }
931 
932     if (ret) {
933         goto out;
934     }
935 
936     BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE);
937     ret = bdrv_co_pwritev(s->bs->file, offset, qiov.size, &qiov, 0);
938     if (ret < 0) {
939         goto out;
940     }
941     ret = 0;
942 out:
943     qemu_vfree(qemu_iovec_buf(&qiov));
944     return ret;
945 }
946 
947 /**
948  * Link one or more contiguous clusters into a table
949  *
950  * @s:              QED state
951  * @table:          L2 table
952  * @index:          First cluster index
953  * @n:              Number of contiguous clusters
954  * @cluster:        First cluster offset
955  *
956  * The cluster offset may be an allocated byte offset in the image file, the
957  * zero cluster marker, or the unallocated cluster marker.
958  *
959  * Called with table_lock held.
960  */
961 static void coroutine_fn qed_update_l2_table(BDRVQEDState *s, QEDTable *table,
962                                              int index, unsigned int n,
963                                              uint64_t cluster)
964 {
965     int i;
966     for (i = index; i < index + n; i++) {
967         table->offsets[i] = cluster;
968         if (!qed_offset_is_unalloc_cluster(cluster) &&
969             !qed_offset_is_zero_cluster(cluster)) {
970             cluster += s->header.cluster_size;
971         }
972     }
973 }
974 
975 /* Called with table_lock held.  */
976 static void coroutine_fn qed_aio_complete(QEDAIOCB *acb)
977 {
978     BDRVQEDState *s = acb_to_s(acb);
979 
980     /* Free resources */
981     qemu_iovec_destroy(&acb->cur_qiov);
982     qed_unref_l2_cache_entry(acb->request.l2_table);
983 
984     /* Free the buffer we may have allocated for zero writes */
985     if (acb->flags & QED_AIOCB_ZERO) {
986         qemu_vfree(acb->qiov->iov[0].iov_base);
987         acb->qiov->iov[0].iov_base = NULL;
988     }
989 
990     /* Start next allocating write request waiting behind this one.  Note that
991      * requests enqueue themselves when they first hit an unallocated cluster
992      * but they wait until the entire request is finished before waking up the
993      * next request in the queue.  This ensures that we don't cycle through
994      * requests multiple times but rather finish one at a time completely.
995      */
996     if (acb == s->allocating_acb) {
997         s->allocating_acb = NULL;
998         if (!qemu_co_queue_empty(&s->allocating_write_reqs)) {
999             qemu_co_queue_next(&s->allocating_write_reqs);
1000         } else if (s->header.features & QED_F_NEED_CHECK) {
1001             qed_start_need_check_timer(s);
1002         }
1003     }
1004 }
1005 
1006 /**
1007  * Update L1 table with new L2 table offset and write it out
1008  *
1009  * Called with table_lock held.
1010  */
1011 static int coroutine_fn qed_aio_write_l1_update(QEDAIOCB *acb)
1012 {
1013     BDRVQEDState *s = acb_to_s(acb);
1014     CachedL2Table *l2_table = acb->request.l2_table;
1015     uint64_t l2_offset = l2_table->offset;
1016     int index, ret;
1017 
1018     index = qed_l1_index(s, acb->cur_pos);
1019     s->l1_table->offsets[index] = l2_table->offset;
1020 
1021     ret = qed_write_l1_table(s, index, 1);
1022 
1023     /* Commit the current L2 table to the cache */
1024     qed_commit_l2_cache_entry(&s->l2_cache, l2_table);
1025 
1026     /* This is guaranteed to succeed because we just committed the entry to the
1027      * cache.
1028      */
1029     acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset);
1030     assert(acb->request.l2_table != NULL);
1031 
1032     return ret;
1033 }
1034 
1035 
1036 /**
1037  * Update L2 table with new cluster offsets and write them out
1038  *
1039  * Called with table_lock held.
1040  */
1041 static int coroutine_fn qed_aio_write_l2_update(QEDAIOCB *acb, uint64_t offset)
1042 {
1043     BDRVQEDState *s = acb_to_s(acb);
1044     bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1;
1045     int index, ret;
1046 
1047     if (need_alloc) {
1048         qed_unref_l2_cache_entry(acb->request.l2_table);
1049         acb->request.l2_table = qed_new_l2_table(s);
1050     }
1051 
1052     index = qed_l2_index(s, acb->cur_pos);
1053     qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters,
1054                          offset);
1055 
1056     if (need_alloc) {
1057         /* Write out the whole new L2 table */
1058         ret = qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true);
1059         if (ret) {
1060             return ret;
1061         }
1062         return qed_aio_write_l1_update(acb);
1063     } else {
1064         /* Write out only the updated part of the L2 table */
1065         ret = qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters,
1066                                  false);
1067         if (ret) {
1068             return ret;
1069         }
1070     }
1071     return 0;
1072 }
1073 
1074 /**
1075  * Write data to the image file
1076  *
1077  * Called with table_lock *not* held.
1078  */
1079 static int coroutine_fn qed_aio_write_main(QEDAIOCB *acb)
1080 {
1081     BDRVQEDState *s = acb_to_s(acb);
1082     uint64_t offset = acb->cur_cluster +
1083                       qed_offset_into_cluster(s, acb->cur_pos);
1084 
1085     trace_qed_aio_write_main(s, acb, 0, offset, acb->cur_qiov.size);
1086 
1087     BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO);
1088     return bdrv_co_pwritev(s->bs->file, offset, acb->cur_qiov.size,
1089                            &acb->cur_qiov, 0);
1090 }
1091 
1092 /**
1093  * Populate untouched regions of new data cluster
1094  *
1095  * Called with table_lock held.
1096  */
1097 static int coroutine_fn qed_aio_write_cow(QEDAIOCB *acb)
1098 {
1099     BDRVQEDState *s = acb_to_s(acb);
1100     uint64_t start, len, offset;
1101     int ret;
1102 
1103     qemu_co_mutex_unlock(&s->table_lock);
1104 
1105     /* Populate front untouched region of new data cluster */
1106     start = qed_start_of_cluster(s, acb->cur_pos);
1107     len = qed_offset_into_cluster(s, acb->cur_pos);
1108 
1109     trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster);
1110     ret = qed_copy_from_backing_file(s, start, len, acb->cur_cluster);
1111     if (ret < 0) {
1112         goto out;
1113     }
1114 
1115     /* Populate back untouched region of new data cluster */
1116     start = acb->cur_pos + acb->cur_qiov.size;
1117     len = qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start;
1118     offset = acb->cur_cluster +
1119              qed_offset_into_cluster(s, acb->cur_pos) +
1120              acb->cur_qiov.size;
1121 
1122     trace_qed_aio_write_postfill(s, acb, start, len, offset);
1123     ret = qed_copy_from_backing_file(s, start, len, offset);
1124     if (ret < 0) {
1125         goto out;
1126     }
1127 
1128     ret = qed_aio_write_main(acb);
1129     if (ret < 0) {
1130         goto out;
1131     }
1132 
1133     if (s->bs->backing) {
1134         /*
1135          * Flush new data clusters before updating the L2 table
1136          *
1137          * This flush is necessary when a backing file is in use.  A crash
1138          * during an allocating write could result in empty clusters in the
1139          * image.  If the write only touched a subregion of the cluster,
1140          * then backing image sectors have been lost in the untouched
1141          * region.  The solution is to flush after writing a new data
1142          * cluster and before updating the L2 table.
1143          */
1144         ret = bdrv_co_flush(s->bs->file->bs);
1145     }
1146 
1147 out:
1148     qemu_co_mutex_lock(&s->table_lock);
1149     return ret;
1150 }
1151 
1152 /**
1153  * Check if the QED_F_NEED_CHECK bit should be set during allocating write
1154  */
1155 static bool qed_should_set_need_check(BDRVQEDState *s)
1156 {
1157     /* The flush before L2 update path ensures consistency */
1158     if (s->bs->backing) {
1159         return false;
1160     }
1161 
1162     return !(s->header.features & QED_F_NEED_CHECK);
1163 }
1164 
1165 /**
1166  * Write new data cluster
1167  *
1168  * @acb:        Write request
1169  * @len:        Length in bytes
1170  *
1171  * This path is taken when writing to previously unallocated clusters.
1172  *
1173  * Called with table_lock held.
1174  */
1175 static int coroutine_fn qed_aio_write_alloc(QEDAIOCB *acb, size_t len)
1176 {
1177     BDRVQEDState *s = acb_to_s(acb);
1178     int ret;
1179 
1180     /* Cancel timer when the first allocating request comes in */
1181     if (s->allocating_acb == NULL) {
1182         qed_cancel_need_check_timer(s);
1183     }
1184 
1185     /* Freeze this request if another allocating write is in progress */
1186     if (s->allocating_acb != acb || s->allocating_write_reqs_plugged) {
1187         if (s->allocating_acb != NULL) {
1188             qemu_co_queue_wait(&s->allocating_write_reqs, &s->table_lock);
1189             assert(s->allocating_acb == NULL);
1190         }
1191         s->allocating_acb = acb;
1192         return -EAGAIN; /* start over with looking up table entries */
1193     }
1194 
1195     acb->cur_nclusters = qed_bytes_to_clusters(s,
1196             qed_offset_into_cluster(s, acb->cur_pos) + len);
1197     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1198 
1199     if (acb->flags & QED_AIOCB_ZERO) {
1200         /* Skip ahead if the clusters are already zero */
1201         if (acb->find_cluster_ret == QED_CLUSTER_ZERO) {
1202             return 0;
1203         }
1204         acb->cur_cluster = 1;
1205     } else {
1206         acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters);
1207     }
1208 
1209     if (qed_should_set_need_check(s)) {
1210         s->header.features |= QED_F_NEED_CHECK;
1211         ret = qed_write_header(s);
1212         if (ret < 0) {
1213             return ret;
1214         }
1215     }
1216 
1217     if (!(acb->flags & QED_AIOCB_ZERO)) {
1218         ret = qed_aio_write_cow(acb);
1219         if (ret < 0) {
1220             return ret;
1221         }
1222     }
1223 
1224     return qed_aio_write_l2_update(acb, acb->cur_cluster);
1225 }
1226 
1227 /**
1228  * Write data cluster in place
1229  *
1230  * @acb:        Write request
1231  * @offset:     Cluster offset in bytes
1232  * @len:        Length in bytes
1233  *
1234  * This path is taken when writing to already allocated clusters.
1235  *
1236  * Called with table_lock held.
1237  */
1238 static int coroutine_fn qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset,
1239                                               size_t len)
1240 {
1241     BDRVQEDState *s = acb_to_s(acb);
1242     int r;
1243 
1244     qemu_co_mutex_unlock(&s->table_lock);
1245 
1246     /* Allocate buffer for zero writes */
1247     if (acb->flags & QED_AIOCB_ZERO) {
1248         struct iovec *iov = acb->qiov->iov;
1249 
1250         if (!iov->iov_base) {
1251             iov->iov_base = qemu_try_blockalign(acb->bs, iov->iov_len);
1252             if (iov->iov_base == NULL) {
1253                 r = -ENOMEM;
1254                 goto out;
1255             }
1256             memset(iov->iov_base, 0, iov->iov_len);
1257         }
1258     }
1259 
1260     /* Calculate the I/O vector */
1261     acb->cur_cluster = offset;
1262     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1263 
1264     /* Do the actual write.  */
1265     r = qed_aio_write_main(acb);
1266 out:
1267     qemu_co_mutex_lock(&s->table_lock);
1268     return r;
1269 }
1270 
1271 /**
1272  * Write data cluster
1273  *
1274  * @opaque:     Write request
1275  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1276  * @offset:     Cluster offset in bytes
1277  * @len:        Length in bytes
1278  *
1279  * Called with table_lock held.
1280  */
1281 static int coroutine_fn qed_aio_write_data(void *opaque, int ret,
1282                                            uint64_t offset, size_t len)
1283 {
1284     QEDAIOCB *acb = opaque;
1285 
1286     trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len);
1287 
1288     acb->find_cluster_ret = ret;
1289 
1290     switch (ret) {
1291     case QED_CLUSTER_FOUND:
1292         return qed_aio_write_inplace(acb, offset, len);
1293 
1294     case QED_CLUSTER_L2:
1295     case QED_CLUSTER_L1:
1296     case QED_CLUSTER_ZERO:
1297         return qed_aio_write_alloc(acb, len);
1298 
1299     default:
1300         g_assert_not_reached();
1301     }
1302 }
1303 
1304 /**
1305  * Read data cluster
1306  *
1307  * @opaque:     Read request
1308  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1309  * @offset:     Cluster offset in bytes
1310  * @len:        Length in bytes
1311  *
1312  * Called with table_lock held.
1313  */
1314 static int coroutine_fn qed_aio_read_data(void *opaque, int ret,
1315                                           uint64_t offset, size_t len)
1316 {
1317     QEDAIOCB *acb = opaque;
1318     BDRVQEDState *s = acb_to_s(acb);
1319     BlockDriverState *bs = acb->bs;
1320     int r;
1321 
1322     qemu_co_mutex_unlock(&s->table_lock);
1323 
1324     /* Adjust offset into cluster */
1325     offset += qed_offset_into_cluster(s, acb->cur_pos);
1326 
1327     trace_qed_aio_read_data(s, acb, ret, offset, len);
1328 
1329     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1330 
1331     /* Handle zero cluster and backing file reads, otherwise read
1332      * data cluster directly.
1333      */
1334     if (ret == QED_CLUSTER_ZERO) {
1335         qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size);
1336         r = 0;
1337     } else if (ret != QED_CLUSTER_FOUND) {
1338         r = qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov,
1339                                   &acb->backing_qiov);
1340     } else {
1341         BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
1342         r = bdrv_co_preadv(bs->file, offset, acb->cur_qiov.size,
1343                            &acb->cur_qiov, 0);
1344     }
1345 
1346     qemu_co_mutex_lock(&s->table_lock);
1347     return r;
1348 }
1349 
1350 /**
1351  * Begin next I/O or complete the request
1352  */
1353 static int coroutine_fn qed_aio_next_io(QEDAIOCB *acb)
1354 {
1355     BDRVQEDState *s = acb_to_s(acb);
1356     uint64_t offset;
1357     size_t len;
1358     int ret;
1359 
1360     qemu_co_mutex_lock(&s->table_lock);
1361     while (1) {
1362         trace_qed_aio_next_io(s, acb, 0, acb->cur_pos + acb->cur_qiov.size);
1363 
1364         if (acb->backing_qiov) {
1365             qemu_iovec_destroy(acb->backing_qiov);
1366             g_free(acb->backing_qiov);
1367             acb->backing_qiov = NULL;
1368         }
1369 
1370         acb->qiov_offset += acb->cur_qiov.size;
1371         acb->cur_pos += acb->cur_qiov.size;
1372         qemu_iovec_reset(&acb->cur_qiov);
1373 
1374         /* Complete request */
1375         if (acb->cur_pos >= acb->end_pos) {
1376             ret = 0;
1377             break;
1378         }
1379 
1380         /* Find next cluster and start I/O */
1381         len = acb->end_pos - acb->cur_pos;
1382         ret = qed_find_cluster(s, &acb->request, acb->cur_pos, &len, &offset);
1383         if (ret < 0) {
1384             break;
1385         }
1386 
1387         if (acb->flags & QED_AIOCB_WRITE) {
1388             ret = qed_aio_write_data(acb, ret, offset, len);
1389         } else {
1390             ret = qed_aio_read_data(acb, ret, offset, len);
1391         }
1392 
1393         if (ret < 0 && ret != -EAGAIN) {
1394             break;
1395         }
1396     }
1397 
1398     trace_qed_aio_complete(s, acb, ret);
1399     qed_aio_complete(acb);
1400     qemu_co_mutex_unlock(&s->table_lock);
1401     return ret;
1402 }
1403 
1404 static int coroutine_fn qed_co_request(BlockDriverState *bs, int64_t sector_num,
1405                                        QEMUIOVector *qiov, int nb_sectors,
1406                                        int flags)
1407 {
1408     QEDAIOCB acb = {
1409         .bs         = bs,
1410         .cur_pos    = (uint64_t) sector_num * BDRV_SECTOR_SIZE,
1411         .end_pos    = (sector_num + nb_sectors) * BDRV_SECTOR_SIZE,
1412         .qiov       = qiov,
1413         .flags      = flags,
1414     };
1415     qemu_iovec_init(&acb.cur_qiov, qiov->niov);
1416 
1417     trace_qed_aio_setup(bs->opaque, &acb, sector_num, nb_sectors, NULL, flags);
1418 
1419     /* Start request */
1420     return qed_aio_next_io(&acb);
1421 }
1422 
1423 static int coroutine_fn bdrv_qed_co_readv(BlockDriverState *bs,
1424                                           int64_t sector_num, int nb_sectors,
1425                                           QEMUIOVector *qiov)
1426 {
1427     return qed_co_request(bs, sector_num, qiov, nb_sectors, 0);
1428 }
1429 
1430 static int coroutine_fn bdrv_qed_co_writev(BlockDriverState *bs,
1431                                            int64_t sector_num, int nb_sectors,
1432                                            QEMUIOVector *qiov, int flags)
1433 {
1434     assert(!flags);
1435     return qed_co_request(bs, sector_num, qiov, nb_sectors, QED_AIOCB_WRITE);
1436 }
1437 
1438 static int coroutine_fn bdrv_qed_co_pwrite_zeroes(BlockDriverState *bs,
1439                                                   int64_t offset,
1440                                                   int bytes,
1441                                                   BdrvRequestFlags flags)
1442 {
1443     BDRVQEDState *s = bs->opaque;
1444 
1445     /*
1446      * Zero writes start without an I/O buffer.  If a buffer becomes necessary
1447      * then it will be allocated during request processing.
1448      */
1449     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, bytes);
1450 
1451     /* Fall back if the request is not aligned */
1452     if (qed_offset_into_cluster(s, offset) ||
1453         qed_offset_into_cluster(s, bytes)) {
1454         return -ENOTSUP;
1455     }
1456 
1457     return qed_co_request(bs, offset >> BDRV_SECTOR_BITS, &qiov,
1458                           bytes >> BDRV_SECTOR_BITS,
1459                           QED_AIOCB_WRITE | QED_AIOCB_ZERO);
1460 }
1461 
1462 static int coroutine_fn bdrv_qed_co_truncate(BlockDriverState *bs,
1463                                              int64_t offset,
1464                                              PreallocMode prealloc,
1465                                              Error **errp)
1466 {
1467     BDRVQEDState *s = bs->opaque;
1468     uint64_t old_image_size;
1469     int ret;
1470 
1471     if (prealloc != PREALLOC_MODE_OFF) {
1472         error_setg(errp, "Unsupported preallocation mode '%s'",
1473                    PreallocMode_str(prealloc));
1474         return -ENOTSUP;
1475     }
1476 
1477     if (!qed_is_image_size_valid(offset, s->header.cluster_size,
1478                                  s->header.table_size)) {
1479         error_setg(errp, "Invalid image size specified");
1480         return -EINVAL;
1481     }
1482 
1483     if ((uint64_t)offset < s->header.image_size) {
1484         error_setg(errp, "Shrinking images is currently not supported");
1485         return -ENOTSUP;
1486     }
1487 
1488     old_image_size = s->header.image_size;
1489     s->header.image_size = offset;
1490     ret = qed_write_header_sync(s);
1491     if (ret < 0) {
1492         s->header.image_size = old_image_size;
1493         error_setg_errno(errp, -ret, "Failed to update the image size");
1494     }
1495     return ret;
1496 }
1497 
1498 static int64_t bdrv_qed_getlength(BlockDriverState *bs)
1499 {
1500     BDRVQEDState *s = bs->opaque;
1501     return s->header.image_size;
1502 }
1503 
1504 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1505 {
1506     BDRVQEDState *s = bs->opaque;
1507 
1508     memset(bdi, 0, sizeof(*bdi));
1509     bdi->cluster_size = s->header.cluster_size;
1510     bdi->is_dirty = s->header.features & QED_F_NEED_CHECK;
1511     bdi->unallocated_blocks_are_zero = true;
1512     return 0;
1513 }
1514 
1515 static int bdrv_qed_change_backing_file(BlockDriverState *bs,
1516                                         const char *backing_file,
1517                                         const char *backing_fmt)
1518 {
1519     BDRVQEDState *s = bs->opaque;
1520     QEDHeader new_header, le_header;
1521     void *buffer;
1522     size_t buffer_len, backing_file_len;
1523     int ret;
1524 
1525     /* Refuse to set backing filename if unknown compat feature bits are
1526      * active.  If the image uses an unknown compat feature then we may not
1527      * know the layout of data following the header structure and cannot safely
1528      * add a new string.
1529      */
1530     if (backing_file && (s->header.compat_features &
1531                          ~QED_COMPAT_FEATURE_MASK)) {
1532         return -ENOTSUP;
1533     }
1534 
1535     memcpy(&new_header, &s->header, sizeof(new_header));
1536 
1537     new_header.features &= ~(QED_F_BACKING_FILE |
1538                              QED_F_BACKING_FORMAT_NO_PROBE);
1539 
1540     /* Adjust feature flags */
1541     if (backing_file) {
1542         new_header.features |= QED_F_BACKING_FILE;
1543 
1544         if (qed_fmt_is_raw(backing_fmt)) {
1545             new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
1546         }
1547     }
1548 
1549     /* Calculate new header size */
1550     backing_file_len = 0;
1551 
1552     if (backing_file) {
1553         backing_file_len = strlen(backing_file);
1554     }
1555 
1556     buffer_len = sizeof(new_header);
1557     new_header.backing_filename_offset = buffer_len;
1558     new_header.backing_filename_size = backing_file_len;
1559     buffer_len += backing_file_len;
1560 
1561     /* Make sure we can rewrite header without failing */
1562     if (buffer_len > new_header.header_size * new_header.cluster_size) {
1563         return -ENOSPC;
1564     }
1565 
1566     /* Prepare new header */
1567     buffer = g_malloc(buffer_len);
1568 
1569     qed_header_cpu_to_le(&new_header, &le_header);
1570     memcpy(buffer, &le_header, sizeof(le_header));
1571     buffer_len = sizeof(le_header);
1572 
1573     if (backing_file) {
1574         memcpy(buffer + buffer_len, backing_file, backing_file_len);
1575         buffer_len += backing_file_len;
1576     }
1577 
1578     /* Write new header */
1579     ret = bdrv_pwrite_sync(bs->file, 0, buffer, buffer_len);
1580     g_free(buffer);
1581     if (ret == 0) {
1582         memcpy(&s->header, &new_header, sizeof(new_header));
1583     }
1584     return ret;
1585 }
1586 
1587 static void coroutine_fn bdrv_qed_co_invalidate_cache(BlockDriverState *bs,
1588                                                       Error **errp)
1589 {
1590     BDRVQEDState *s = bs->opaque;
1591     Error *local_err = NULL;
1592     int ret;
1593 
1594     bdrv_qed_close(bs);
1595 
1596     bdrv_qed_init_state(bs);
1597     qemu_co_mutex_lock(&s->table_lock);
1598     ret = bdrv_qed_do_open(bs, NULL, bs->open_flags, &local_err);
1599     qemu_co_mutex_unlock(&s->table_lock);
1600     if (local_err) {
1601         error_propagate_prepend(errp, local_err,
1602                                 "Could not reopen qed layer: ");
1603         return;
1604     } else if (ret < 0) {
1605         error_setg_errno(errp, -ret, "Could not reopen qed layer");
1606         return;
1607     }
1608 }
1609 
1610 static int coroutine_fn bdrv_qed_co_check(BlockDriverState *bs,
1611                                           BdrvCheckResult *result,
1612                                           BdrvCheckMode fix)
1613 {
1614     BDRVQEDState *s = bs->opaque;
1615     int ret;
1616 
1617     qemu_co_mutex_lock(&s->table_lock);
1618     ret = qed_check(s, result, !!fix);
1619     qemu_co_mutex_unlock(&s->table_lock);
1620 
1621     return ret;
1622 }
1623 
1624 static QemuOptsList qed_create_opts = {
1625     .name = "qed-create-opts",
1626     .head = QTAILQ_HEAD_INITIALIZER(qed_create_opts.head),
1627     .desc = {
1628         {
1629             .name = BLOCK_OPT_SIZE,
1630             .type = QEMU_OPT_SIZE,
1631             .help = "Virtual disk size"
1632         },
1633         {
1634             .name = BLOCK_OPT_BACKING_FILE,
1635             .type = QEMU_OPT_STRING,
1636             .help = "File name of a base image"
1637         },
1638         {
1639             .name = BLOCK_OPT_BACKING_FMT,
1640             .type = QEMU_OPT_STRING,
1641             .help = "Image format of the base image"
1642         },
1643         {
1644             .name = BLOCK_OPT_CLUSTER_SIZE,
1645             .type = QEMU_OPT_SIZE,
1646             .help = "Cluster size (in bytes)",
1647             .def_value_str = stringify(QED_DEFAULT_CLUSTER_SIZE)
1648         },
1649         {
1650             .name = BLOCK_OPT_TABLE_SIZE,
1651             .type = QEMU_OPT_SIZE,
1652             .help = "L1/L2 table size (in clusters)"
1653         },
1654         { /* end of list */ }
1655     }
1656 };
1657 
1658 static BlockDriver bdrv_qed = {
1659     .format_name              = "qed",
1660     .instance_size            = sizeof(BDRVQEDState),
1661     .create_opts              = &qed_create_opts,
1662     .supports_backing         = true,
1663 
1664     .bdrv_probe               = bdrv_qed_probe,
1665     .bdrv_open                = bdrv_qed_open,
1666     .bdrv_close               = bdrv_qed_close,
1667     .bdrv_reopen_prepare      = bdrv_qed_reopen_prepare,
1668     .bdrv_child_perm          = bdrv_format_default_perms,
1669     .bdrv_co_create           = bdrv_qed_co_create,
1670     .bdrv_co_create_opts      = bdrv_qed_co_create_opts,
1671     .bdrv_has_zero_init       = bdrv_has_zero_init_1,
1672     .bdrv_has_zero_init_truncate = bdrv_has_zero_init_1,
1673     .bdrv_co_block_status     = bdrv_qed_co_block_status,
1674     .bdrv_co_readv            = bdrv_qed_co_readv,
1675     .bdrv_co_writev           = bdrv_qed_co_writev,
1676     .bdrv_co_pwrite_zeroes    = bdrv_qed_co_pwrite_zeroes,
1677     .bdrv_co_truncate         = bdrv_qed_co_truncate,
1678     .bdrv_getlength           = bdrv_qed_getlength,
1679     .bdrv_get_info            = bdrv_qed_get_info,
1680     .bdrv_refresh_limits      = bdrv_qed_refresh_limits,
1681     .bdrv_change_backing_file = bdrv_qed_change_backing_file,
1682     .bdrv_co_invalidate_cache = bdrv_qed_co_invalidate_cache,
1683     .bdrv_co_check            = bdrv_qed_co_check,
1684     .bdrv_detach_aio_context  = bdrv_qed_detach_aio_context,
1685     .bdrv_attach_aio_context  = bdrv_qed_attach_aio_context,
1686     .bdrv_co_drain_begin      = bdrv_qed_co_drain_begin,
1687 };
1688 
1689 static void bdrv_qed_init(void)
1690 {
1691     bdrv_register(&bdrv_qed);
1692 }
1693 
1694 block_init(bdrv_qed_init);
1695