xref: /openbmc/qemu/block/qed.c (revision b2c623a3)
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/timer.h"
16 #include "trace.h"
17 #include "qed.h"
18 #include "qapi/qmp/qerror.h"
19 #include "migration/migration.h"
20 
21 static void qed_aio_cancel(BlockDriverAIOCB *blockacb)
22 {
23     QEDAIOCB *acb = (QEDAIOCB *)blockacb;
24     bool finished = false;
25 
26     /* Wait for the request to finish */
27     acb->finished = &finished;
28     while (!finished) {
29         qemu_aio_wait();
30     }
31 }
32 
33 static const AIOCBInfo qed_aiocb_info = {
34     .aiocb_size         = sizeof(QEDAIOCB),
35     .cancel             = qed_aio_cancel,
36 };
37 
38 static int bdrv_qed_probe(const uint8_t *buf, int buf_size,
39                           const char *filename)
40 {
41     const QEDHeader *header = (const QEDHeader *)buf;
42 
43     if (buf_size < sizeof(*header)) {
44         return 0;
45     }
46     if (le32_to_cpu(header->magic) != QED_MAGIC) {
47         return 0;
48     }
49     return 100;
50 }
51 
52 /**
53  * Check whether an image format is raw
54  *
55  * @fmt:    Backing file format, may be NULL
56  */
57 static bool qed_fmt_is_raw(const char *fmt)
58 {
59     return fmt && strcmp(fmt, "raw") == 0;
60 }
61 
62 static void qed_header_le_to_cpu(const QEDHeader *le, QEDHeader *cpu)
63 {
64     cpu->magic = le32_to_cpu(le->magic);
65     cpu->cluster_size = le32_to_cpu(le->cluster_size);
66     cpu->table_size = le32_to_cpu(le->table_size);
67     cpu->header_size = le32_to_cpu(le->header_size);
68     cpu->features = le64_to_cpu(le->features);
69     cpu->compat_features = le64_to_cpu(le->compat_features);
70     cpu->autoclear_features = le64_to_cpu(le->autoclear_features);
71     cpu->l1_table_offset = le64_to_cpu(le->l1_table_offset);
72     cpu->image_size = le64_to_cpu(le->image_size);
73     cpu->backing_filename_offset = le32_to_cpu(le->backing_filename_offset);
74     cpu->backing_filename_size = le32_to_cpu(le->backing_filename_size);
75 }
76 
77 static void qed_header_cpu_to_le(const QEDHeader *cpu, QEDHeader *le)
78 {
79     le->magic = cpu_to_le32(cpu->magic);
80     le->cluster_size = cpu_to_le32(cpu->cluster_size);
81     le->table_size = cpu_to_le32(cpu->table_size);
82     le->header_size = cpu_to_le32(cpu->header_size);
83     le->features = cpu_to_le64(cpu->features);
84     le->compat_features = cpu_to_le64(cpu->compat_features);
85     le->autoclear_features = cpu_to_le64(cpu->autoclear_features);
86     le->l1_table_offset = cpu_to_le64(cpu->l1_table_offset);
87     le->image_size = cpu_to_le64(cpu->image_size);
88     le->backing_filename_offset = cpu_to_le32(cpu->backing_filename_offset);
89     le->backing_filename_size = cpu_to_le32(cpu->backing_filename_size);
90 }
91 
92 int qed_write_header_sync(BDRVQEDState *s)
93 {
94     QEDHeader le;
95     int ret;
96 
97     qed_header_cpu_to_le(&s->header, &le);
98     ret = bdrv_pwrite(s->bs->file, 0, &le, sizeof(le));
99     if (ret != sizeof(le)) {
100         return ret;
101     }
102     return 0;
103 }
104 
105 typedef struct {
106     GenericCB gencb;
107     BDRVQEDState *s;
108     struct iovec iov;
109     QEMUIOVector qiov;
110     int nsectors;
111     uint8_t *buf;
112 } QEDWriteHeaderCB;
113 
114 static void qed_write_header_cb(void *opaque, int ret)
115 {
116     QEDWriteHeaderCB *write_header_cb = opaque;
117 
118     qemu_vfree(write_header_cb->buf);
119     gencb_complete(write_header_cb, ret);
120 }
121 
122 static void qed_write_header_read_cb(void *opaque, int ret)
123 {
124     QEDWriteHeaderCB *write_header_cb = opaque;
125     BDRVQEDState *s = write_header_cb->s;
126 
127     if (ret) {
128         qed_write_header_cb(write_header_cb, ret);
129         return;
130     }
131 
132     /* Update header */
133     qed_header_cpu_to_le(&s->header, (QEDHeader *)write_header_cb->buf);
134 
135     bdrv_aio_writev(s->bs->file, 0, &write_header_cb->qiov,
136                     write_header_cb->nsectors, qed_write_header_cb,
137                     write_header_cb);
138 }
139 
140 /**
141  * Update header in-place (does not rewrite backing filename or other strings)
142  *
143  * This function only updates known header fields in-place and does not affect
144  * extra data after the QED header.
145  */
146 static void qed_write_header(BDRVQEDState *s, BlockDriverCompletionFunc cb,
147                              void *opaque)
148 {
149     /* We must write full sectors for O_DIRECT but cannot necessarily generate
150      * the data following the header if an unrecognized compat feature is
151      * active.  Therefore, first read the sectors containing the header, update
152      * them, and write back.
153      */
154 
155     int nsectors = (sizeof(QEDHeader) + BDRV_SECTOR_SIZE - 1) /
156                    BDRV_SECTOR_SIZE;
157     size_t len = nsectors * BDRV_SECTOR_SIZE;
158     QEDWriteHeaderCB *write_header_cb = gencb_alloc(sizeof(*write_header_cb),
159                                                     cb, opaque);
160 
161     write_header_cb->s = s;
162     write_header_cb->nsectors = nsectors;
163     write_header_cb->buf = qemu_blockalign(s->bs, len);
164     write_header_cb->iov.iov_base = write_header_cb->buf;
165     write_header_cb->iov.iov_len = len;
166     qemu_iovec_init_external(&write_header_cb->qiov, &write_header_cb->iov, 1);
167 
168     bdrv_aio_readv(s->bs->file, 0, &write_header_cb->qiov, nsectors,
169                    qed_write_header_read_cb, write_header_cb);
170 }
171 
172 static uint64_t qed_max_image_size(uint32_t cluster_size, uint32_t table_size)
173 {
174     uint64_t table_entries;
175     uint64_t l2_size;
176 
177     table_entries = (table_size * cluster_size) / sizeof(uint64_t);
178     l2_size = table_entries * cluster_size;
179 
180     return l2_size * table_entries;
181 }
182 
183 static bool qed_is_cluster_size_valid(uint32_t cluster_size)
184 {
185     if (cluster_size < QED_MIN_CLUSTER_SIZE ||
186         cluster_size > QED_MAX_CLUSTER_SIZE) {
187         return false;
188     }
189     if (cluster_size & (cluster_size - 1)) {
190         return false; /* not power of 2 */
191     }
192     return true;
193 }
194 
195 static bool qed_is_table_size_valid(uint32_t table_size)
196 {
197     if (table_size < QED_MIN_TABLE_SIZE ||
198         table_size > QED_MAX_TABLE_SIZE) {
199         return false;
200     }
201     if (table_size & (table_size - 1)) {
202         return false; /* not power of 2 */
203     }
204     return true;
205 }
206 
207 static bool qed_is_image_size_valid(uint64_t image_size, uint32_t cluster_size,
208                                     uint32_t table_size)
209 {
210     if (image_size % BDRV_SECTOR_SIZE != 0) {
211         return false; /* not multiple of sector size */
212     }
213     if (image_size > qed_max_image_size(cluster_size, table_size)) {
214         return false; /* image is too large */
215     }
216     return true;
217 }
218 
219 /**
220  * Read a string of known length from the image file
221  *
222  * @file:       Image file
223  * @offset:     File offset to start of string, in bytes
224  * @n:          String length in bytes
225  * @buf:        Destination buffer
226  * @buflen:     Destination buffer length in bytes
227  * @ret:        0 on success, -errno on failure
228  *
229  * The string is NUL-terminated.
230  */
231 static int qed_read_string(BlockDriverState *file, uint64_t offset, size_t n,
232                            char *buf, size_t buflen)
233 {
234     int ret;
235     if (n >= buflen) {
236         return -EINVAL;
237     }
238     ret = bdrv_pread(file, offset, buf, n);
239     if (ret < 0) {
240         return ret;
241     }
242     buf[n] = '\0';
243     return 0;
244 }
245 
246 /**
247  * Allocate new clusters
248  *
249  * @s:          QED state
250  * @n:          Number of contiguous clusters to allocate
251  * @ret:        Offset of first allocated cluster
252  *
253  * This function only produces the offset where the new clusters should be
254  * written.  It updates BDRVQEDState but does not make any changes to the image
255  * file.
256  */
257 static uint64_t qed_alloc_clusters(BDRVQEDState *s, unsigned int n)
258 {
259     uint64_t offset = s->file_size;
260     s->file_size += n * s->header.cluster_size;
261     return offset;
262 }
263 
264 QEDTable *qed_alloc_table(BDRVQEDState *s)
265 {
266     /* Honor O_DIRECT memory alignment requirements */
267     return qemu_blockalign(s->bs,
268                            s->header.cluster_size * s->header.table_size);
269 }
270 
271 /**
272  * Allocate a new zeroed L2 table
273  */
274 static CachedL2Table *qed_new_l2_table(BDRVQEDState *s)
275 {
276     CachedL2Table *l2_table = qed_alloc_l2_cache_entry(&s->l2_cache);
277 
278     l2_table->table = qed_alloc_table(s);
279     l2_table->offset = qed_alloc_clusters(s, s->header.table_size);
280 
281     memset(l2_table->table->offsets, 0,
282            s->header.cluster_size * s->header.table_size);
283     return l2_table;
284 }
285 
286 static void qed_aio_next_io(void *opaque, int ret);
287 
288 static void qed_plug_allocating_write_reqs(BDRVQEDState *s)
289 {
290     assert(!s->allocating_write_reqs_plugged);
291 
292     s->allocating_write_reqs_plugged = true;
293 }
294 
295 static void qed_unplug_allocating_write_reqs(BDRVQEDState *s)
296 {
297     QEDAIOCB *acb;
298 
299     assert(s->allocating_write_reqs_plugged);
300 
301     s->allocating_write_reqs_plugged = false;
302 
303     acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs);
304     if (acb) {
305         qed_aio_next_io(acb, 0);
306     }
307 }
308 
309 static void qed_finish_clear_need_check(void *opaque, int ret)
310 {
311     /* Do nothing */
312 }
313 
314 static void qed_flush_after_clear_need_check(void *opaque, int ret)
315 {
316     BDRVQEDState *s = opaque;
317 
318     bdrv_aio_flush(s->bs, qed_finish_clear_need_check, s);
319 
320     /* No need to wait until flush completes */
321     qed_unplug_allocating_write_reqs(s);
322 }
323 
324 static void qed_clear_need_check(void *opaque, int ret)
325 {
326     BDRVQEDState *s = opaque;
327 
328     if (ret) {
329         qed_unplug_allocating_write_reqs(s);
330         return;
331     }
332 
333     s->header.features &= ~QED_F_NEED_CHECK;
334     qed_write_header(s, qed_flush_after_clear_need_check, s);
335 }
336 
337 static void qed_need_check_timer_cb(void *opaque)
338 {
339     BDRVQEDState *s = opaque;
340 
341     /* The timer should only fire when allocating writes have drained */
342     assert(!QSIMPLEQ_FIRST(&s->allocating_write_reqs));
343 
344     trace_qed_need_check_timer_cb(s);
345 
346     qed_plug_allocating_write_reqs(s);
347 
348     /* Ensure writes are on disk before clearing flag */
349     bdrv_aio_flush(s->bs, qed_clear_need_check, s);
350 }
351 
352 static void qed_start_need_check_timer(BDRVQEDState *s)
353 {
354     trace_qed_start_need_check_timer(s);
355 
356     /* Use QEMU_CLOCK_VIRTUAL so we don't alter the image file while suspended for
357      * migration.
358      */
359     timer_mod(s->need_check_timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
360                    get_ticks_per_sec() * QED_NEED_CHECK_TIMEOUT);
361 }
362 
363 /* It's okay to call this multiple times or when no timer is started */
364 static void qed_cancel_need_check_timer(BDRVQEDState *s)
365 {
366     trace_qed_cancel_need_check_timer(s);
367     timer_del(s->need_check_timer);
368 }
369 
370 static void bdrv_qed_rebind(BlockDriverState *bs)
371 {
372     BDRVQEDState *s = bs->opaque;
373     s->bs = bs;
374 }
375 
376 static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags,
377                          Error **errp)
378 {
379     BDRVQEDState *s = bs->opaque;
380     QEDHeader le_header;
381     int64_t file_size;
382     int ret;
383 
384     s->bs = bs;
385     QSIMPLEQ_INIT(&s->allocating_write_reqs);
386 
387     ret = bdrv_pread(bs->file, 0, &le_header, sizeof(le_header));
388     if (ret < 0) {
389         return ret;
390     }
391     qed_header_le_to_cpu(&le_header, &s->header);
392 
393     if (s->header.magic != QED_MAGIC) {
394         return -EMEDIUMTYPE;
395     }
396     if (s->header.features & ~QED_FEATURE_MASK) {
397         /* image uses unsupported feature bits */
398         char buf[64];
399         snprintf(buf, sizeof(buf), "%" PRIx64,
400             s->header.features & ~QED_FEATURE_MASK);
401         qerror_report(QERR_UNKNOWN_BLOCK_FORMAT_FEATURE,
402             bs->device_name, "QED", buf);
403         return -ENOTSUP;
404     }
405     if (!qed_is_cluster_size_valid(s->header.cluster_size)) {
406         return -EINVAL;
407     }
408 
409     /* Round down file size to the last cluster */
410     file_size = bdrv_getlength(bs->file);
411     if (file_size < 0) {
412         return file_size;
413     }
414     s->file_size = qed_start_of_cluster(s, file_size);
415 
416     if (!qed_is_table_size_valid(s->header.table_size)) {
417         return -EINVAL;
418     }
419     if (!qed_is_image_size_valid(s->header.image_size,
420                                  s->header.cluster_size,
421                                  s->header.table_size)) {
422         return -EINVAL;
423     }
424     if (!qed_check_table_offset(s, s->header.l1_table_offset)) {
425         return -EINVAL;
426     }
427 
428     s->table_nelems = (s->header.cluster_size * s->header.table_size) /
429                       sizeof(uint64_t);
430     s->l2_shift = ffs(s->header.cluster_size) - 1;
431     s->l2_mask = s->table_nelems - 1;
432     s->l1_shift = s->l2_shift + ffs(s->table_nelems) - 1;
433 
434     if ((s->header.features & QED_F_BACKING_FILE)) {
435         if ((uint64_t)s->header.backing_filename_offset +
436             s->header.backing_filename_size >
437             s->header.cluster_size * s->header.header_size) {
438             return -EINVAL;
439         }
440 
441         ret = qed_read_string(bs->file, s->header.backing_filename_offset,
442                               s->header.backing_filename_size, bs->backing_file,
443                               sizeof(bs->backing_file));
444         if (ret < 0) {
445             return ret;
446         }
447 
448         if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) {
449             pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw");
450         }
451     }
452 
453     /* Reset unknown autoclear feature bits.  This is a backwards
454      * compatibility mechanism that allows images to be opened by older
455      * programs, which "knock out" unknown feature bits.  When an image is
456      * opened by a newer program again it can detect that the autoclear
457      * feature is no longer valid.
458      */
459     if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 &&
460         !bdrv_is_read_only(bs->file) && !(flags & BDRV_O_INCOMING)) {
461         s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK;
462 
463         ret = qed_write_header_sync(s);
464         if (ret) {
465             return ret;
466         }
467 
468         /* From here on only known autoclear feature bits are valid */
469         bdrv_flush(bs->file);
470     }
471 
472     s->l1_table = qed_alloc_table(s);
473     qed_init_l2_cache(&s->l2_cache);
474 
475     ret = qed_read_l1_table_sync(s);
476     if (ret) {
477         goto out;
478     }
479 
480     /* If image was not closed cleanly, check consistency */
481     if (!(flags & BDRV_O_CHECK) && (s->header.features & QED_F_NEED_CHECK)) {
482         /* Read-only images cannot be fixed.  There is no risk of corruption
483          * since write operations are not possible.  Therefore, allow
484          * potentially inconsistent images to be opened read-only.  This can
485          * aid data recovery from an otherwise inconsistent image.
486          */
487         if (!bdrv_is_read_only(bs->file) &&
488             !(flags & BDRV_O_INCOMING)) {
489             BdrvCheckResult result = {0};
490 
491             ret = qed_check(s, &result, true);
492             if (ret) {
493                 goto out;
494             }
495         }
496     }
497 
498     s->need_check_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL,
499                                             qed_need_check_timer_cb, s);
500 
501 out:
502     if (ret) {
503         qed_free_l2_cache(&s->l2_cache);
504         qemu_vfree(s->l1_table);
505     }
506     return ret;
507 }
508 
509 static int bdrv_qed_refresh_limits(BlockDriverState *bs)
510 {
511     BDRVQEDState *s = bs->opaque;
512 
513     bs->bl.write_zeroes_alignment = s->header.cluster_size >> BDRV_SECTOR_BITS;
514 
515     return 0;
516 }
517 
518 /* We have nothing to do for QED reopen, stubs just return
519  * success */
520 static int bdrv_qed_reopen_prepare(BDRVReopenState *state,
521                                    BlockReopenQueue *queue, Error **errp)
522 {
523     return 0;
524 }
525 
526 static void bdrv_qed_close(BlockDriverState *bs)
527 {
528     BDRVQEDState *s = bs->opaque;
529 
530     qed_cancel_need_check_timer(s);
531     timer_free(s->need_check_timer);
532 
533     /* Ensure writes reach stable storage */
534     bdrv_flush(bs->file);
535 
536     /* Clean shutdown, no check required on next open */
537     if (s->header.features & QED_F_NEED_CHECK) {
538         s->header.features &= ~QED_F_NEED_CHECK;
539         qed_write_header_sync(s);
540     }
541 
542     qed_free_l2_cache(&s->l2_cache);
543     qemu_vfree(s->l1_table);
544 }
545 
546 static int qed_create(const char *filename, uint32_t cluster_size,
547                       uint64_t image_size, uint32_t table_size,
548                       const char *backing_file, const char *backing_fmt)
549 {
550     QEDHeader header = {
551         .magic = QED_MAGIC,
552         .cluster_size = cluster_size,
553         .table_size = table_size,
554         .header_size = 1,
555         .features = 0,
556         .compat_features = 0,
557         .l1_table_offset = cluster_size,
558         .image_size = image_size,
559     };
560     QEDHeader le_header;
561     uint8_t *l1_table = NULL;
562     size_t l1_size = header.cluster_size * header.table_size;
563     Error *local_err = NULL;
564     int ret = 0;
565     BlockDriverState *bs = NULL;
566 
567     ret = bdrv_create_file(filename, NULL, &local_err);
568     if (ret < 0) {
569         qerror_report_err(local_err);
570         error_free(local_err);
571         return ret;
572     }
573 
574     ret = bdrv_file_open(&bs, filename, NULL, NULL,
575                          BDRV_O_RDWR | BDRV_O_CACHE_WB, &local_err);
576     if (ret < 0) {
577         qerror_report_err(local_err);
578         error_free(local_err);
579         return ret;
580     }
581 
582     /* File must start empty and grow, check truncate is supported */
583     ret = bdrv_truncate(bs, 0);
584     if (ret < 0) {
585         goto out;
586     }
587 
588     if (backing_file) {
589         header.features |= QED_F_BACKING_FILE;
590         header.backing_filename_offset = sizeof(le_header);
591         header.backing_filename_size = strlen(backing_file);
592 
593         if (qed_fmt_is_raw(backing_fmt)) {
594             header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
595         }
596     }
597 
598     qed_header_cpu_to_le(&header, &le_header);
599     ret = bdrv_pwrite(bs, 0, &le_header, sizeof(le_header));
600     if (ret < 0) {
601         goto out;
602     }
603     ret = bdrv_pwrite(bs, sizeof(le_header), backing_file,
604                       header.backing_filename_size);
605     if (ret < 0) {
606         goto out;
607     }
608 
609     l1_table = g_malloc0(l1_size);
610     ret = bdrv_pwrite(bs, header.l1_table_offset, l1_table, l1_size);
611     if (ret < 0) {
612         goto out;
613     }
614 
615     ret = 0; /* success */
616 out:
617     g_free(l1_table);
618     bdrv_unref(bs);
619     return ret;
620 }
621 
622 static int bdrv_qed_create(const char *filename, QEMUOptionParameter *options,
623                            Error **errp)
624 {
625     uint64_t image_size = 0;
626     uint32_t cluster_size = QED_DEFAULT_CLUSTER_SIZE;
627     uint32_t table_size = QED_DEFAULT_TABLE_SIZE;
628     const char *backing_file = NULL;
629     const char *backing_fmt = NULL;
630 
631     while (options && options->name) {
632         if (!strcmp(options->name, BLOCK_OPT_SIZE)) {
633             image_size = options->value.n;
634         } else if (!strcmp(options->name, BLOCK_OPT_BACKING_FILE)) {
635             backing_file = options->value.s;
636         } else if (!strcmp(options->name, BLOCK_OPT_BACKING_FMT)) {
637             backing_fmt = options->value.s;
638         } else if (!strcmp(options->name, BLOCK_OPT_CLUSTER_SIZE)) {
639             if (options->value.n) {
640                 cluster_size = options->value.n;
641             }
642         } else if (!strcmp(options->name, BLOCK_OPT_TABLE_SIZE)) {
643             if (options->value.n) {
644                 table_size = options->value.n;
645             }
646         }
647         options++;
648     }
649 
650     if (!qed_is_cluster_size_valid(cluster_size)) {
651         fprintf(stderr, "QED cluster size must be within range [%u, %u] and power of 2\n",
652                 QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE);
653         return -EINVAL;
654     }
655     if (!qed_is_table_size_valid(table_size)) {
656         fprintf(stderr, "QED table size must be within range [%u, %u] and power of 2\n",
657                 QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE);
658         return -EINVAL;
659     }
660     if (!qed_is_image_size_valid(image_size, cluster_size, table_size)) {
661         fprintf(stderr, "QED image size must be a non-zero multiple of "
662                         "cluster size and less than %" PRIu64 " bytes\n",
663                 qed_max_image_size(cluster_size, table_size));
664         return -EINVAL;
665     }
666 
667     return qed_create(filename, cluster_size, image_size, table_size,
668                       backing_file, backing_fmt);
669 }
670 
671 typedef struct {
672     BlockDriverState *bs;
673     Coroutine *co;
674     uint64_t pos;
675     int64_t status;
676     int *pnum;
677 } QEDIsAllocatedCB;
678 
679 static void qed_is_allocated_cb(void *opaque, int ret, uint64_t offset, size_t len)
680 {
681     QEDIsAllocatedCB *cb = opaque;
682     BDRVQEDState *s = cb->bs->opaque;
683     *cb->pnum = len / BDRV_SECTOR_SIZE;
684     switch (ret) {
685     case QED_CLUSTER_FOUND:
686         offset |= qed_offset_into_cluster(s, cb->pos);
687         cb->status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID | offset;
688         break;
689     case QED_CLUSTER_ZERO:
690         cb->status = BDRV_BLOCK_ZERO;
691         break;
692     case QED_CLUSTER_L2:
693     case QED_CLUSTER_L1:
694         cb->status = 0;
695         break;
696     default:
697         assert(ret < 0);
698         cb->status = ret;
699         break;
700     }
701 
702     if (cb->co) {
703         qemu_coroutine_enter(cb->co, NULL);
704     }
705 }
706 
707 static int64_t coroutine_fn bdrv_qed_co_get_block_status(BlockDriverState *bs,
708                                                  int64_t sector_num,
709                                                  int nb_sectors, int *pnum)
710 {
711     BDRVQEDState *s = bs->opaque;
712     size_t len = (size_t)nb_sectors * BDRV_SECTOR_SIZE;
713     QEDIsAllocatedCB cb = {
714         .bs = bs,
715         .pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE,
716         .status = BDRV_BLOCK_OFFSET_MASK,
717         .pnum = pnum,
718     };
719     QEDRequest request = { .l2_table = NULL };
720 
721     qed_find_cluster(s, &request, cb.pos, len, qed_is_allocated_cb, &cb);
722 
723     /* Now sleep if the callback wasn't invoked immediately */
724     while (cb.status == BDRV_BLOCK_OFFSET_MASK) {
725         cb.co = qemu_coroutine_self();
726         qemu_coroutine_yield();
727     }
728 
729     qed_unref_l2_cache_entry(request.l2_table);
730 
731     return cb.status;
732 }
733 
734 static int bdrv_qed_make_empty(BlockDriverState *bs)
735 {
736     return -ENOTSUP;
737 }
738 
739 static BDRVQEDState *acb_to_s(QEDAIOCB *acb)
740 {
741     return acb->common.bs->opaque;
742 }
743 
744 /**
745  * Read from the backing file or zero-fill if no backing file
746  *
747  * @s:          QED state
748  * @pos:        Byte position in device
749  * @qiov:       Destination I/O vector
750  * @cb:         Completion function
751  * @opaque:     User data for completion function
752  *
753  * This function reads qiov->size bytes starting at pos from the backing file.
754  * If there is no backing file then zeroes are read.
755  */
756 static void qed_read_backing_file(BDRVQEDState *s, uint64_t pos,
757                                   QEMUIOVector *qiov,
758                                   BlockDriverCompletionFunc *cb, void *opaque)
759 {
760     uint64_t backing_length = 0;
761     size_t size;
762 
763     /* If there is a backing file, get its length.  Treat the absence of a
764      * backing file like a zero length backing file.
765      */
766     if (s->bs->backing_hd) {
767         int64_t l = bdrv_getlength(s->bs->backing_hd);
768         if (l < 0) {
769             cb(opaque, l);
770             return;
771         }
772         backing_length = l;
773     }
774 
775     /* Zero all sectors if reading beyond the end of the backing file */
776     if (pos >= backing_length ||
777         pos + qiov->size > backing_length) {
778         qemu_iovec_memset(qiov, 0, 0, qiov->size);
779     }
780 
781     /* Complete now if there are no backing file sectors to read */
782     if (pos >= backing_length) {
783         cb(opaque, 0);
784         return;
785     }
786 
787     /* If the read straddles the end of the backing file, shorten it */
788     size = MIN((uint64_t)backing_length - pos, qiov->size);
789 
790     BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO);
791     bdrv_aio_readv(s->bs->backing_hd, pos / BDRV_SECTOR_SIZE,
792                    qiov, size / BDRV_SECTOR_SIZE, cb, opaque);
793 }
794 
795 typedef struct {
796     GenericCB gencb;
797     BDRVQEDState *s;
798     QEMUIOVector qiov;
799     struct iovec iov;
800     uint64_t offset;
801 } CopyFromBackingFileCB;
802 
803 static void qed_copy_from_backing_file_cb(void *opaque, int ret)
804 {
805     CopyFromBackingFileCB *copy_cb = opaque;
806     qemu_vfree(copy_cb->iov.iov_base);
807     gencb_complete(&copy_cb->gencb, ret);
808 }
809 
810 static void qed_copy_from_backing_file_write(void *opaque, int ret)
811 {
812     CopyFromBackingFileCB *copy_cb = opaque;
813     BDRVQEDState *s = copy_cb->s;
814 
815     if (ret) {
816         qed_copy_from_backing_file_cb(copy_cb, ret);
817         return;
818     }
819 
820     BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE);
821     bdrv_aio_writev(s->bs->file, copy_cb->offset / BDRV_SECTOR_SIZE,
822                     &copy_cb->qiov, copy_cb->qiov.size / BDRV_SECTOR_SIZE,
823                     qed_copy_from_backing_file_cb, copy_cb);
824 }
825 
826 /**
827  * Copy data from backing file into the image
828  *
829  * @s:          QED state
830  * @pos:        Byte position in device
831  * @len:        Number of bytes
832  * @offset:     Byte offset in image file
833  * @cb:         Completion function
834  * @opaque:     User data for completion function
835  */
836 static void qed_copy_from_backing_file(BDRVQEDState *s, uint64_t pos,
837                                        uint64_t len, uint64_t offset,
838                                        BlockDriverCompletionFunc *cb,
839                                        void *opaque)
840 {
841     CopyFromBackingFileCB *copy_cb;
842 
843     /* Skip copy entirely if there is no work to do */
844     if (len == 0) {
845         cb(opaque, 0);
846         return;
847     }
848 
849     copy_cb = gencb_alloc(sizeof(*copy_cb), cb, opaque);
850     copy_cb->s = s;
851     copy_cb->offset = offset;
852     copy_cb->iov.iov_base = qemu_blockalign(s->bs, len);
853     copy_cb->iov.iov_len = len;
854     qemu_iovec_init_external(&copy_cb->qiov, &copy_cb->iov, 1);
855 
856     qed_read_backing_file(s, pos, &copy_cb->qiov,
857                           qed_copy_from_backing_file_write, copy_cb);
858 }
859 
860 /**
861  * Link one or more contiguous clusters into a table
862  *
863  * @s:              QED state
864  * @table:          L2 table
865  * @index:          First cluster index
866  * @n:              Number of contiguous clusters
867  * @cluster:        First cluster offset
868  *
869  * The cluster offset may be an allocated byte offset in the image file, the
870  * zero cluster marker, or the unallocated cluster marker.
871  */
872 static void qed_update_l2_table(BDRVQEDState *s, QEDTable *table, int index,
873                                 unsigned int n, uint64_t cluster)
874 {
875     int i;
876     for (i = index; i < index + n; i++) {
877         table->offsets[i] = cluster;
878         if (!qed_offset_is_unalloc_cluster(cluster) &&
879             !qed_offset_is_zero_cluster(cluster)) {
880             cluster += s->header.cluster_size;
881         }
882     }
883 }
884 
885 static void qed_aio_complete_bh(void *opaque)
886 {
887     QEDAIOCB *acb = opaque;
888     BlockDriverCompletionFunc *cb = acb->common.cb;
889     void *user_opaque = acb->common.opaque;
890     int ret = acb->bh_ret;
891     bool *finished = acb->finished;
892 
893     qemu_bh_delete(acb->bh);
894     qemu_aio_release(acb);
895 
896     /* Invoke callback */
897     cb(user_opaque, ret);
898 
899     /* Signal cancel completion */
900     if (finished) {
901         *finished = true;
902     }
903 }
904 
905 static void qed_aio_complete(QEDAIOCB *acb, int ret)
906 {
907     BDRVQEDState *s = acb_to_s(acb);
908 
909     trace_qed_aio_complete(s, acb, ret);
910 
911     /* Free resources */
912     qemu_iovec_destroy(&acb->cur_qiov);
913     qed_unref_l2_cache_entry(acb->request.l2_table);
914 
915     /* Free the buffer we may have allocated for zero writes */
916     if (acb->flags & QED_AIOCB_ZERO) {
917         qemu_vfree(acb->qiov->iov[0].iov_base);
918         acb->qiov->iov[0].iov_base = NULL;
919     }
920 
921     /* Arrange for a bh to invoke the completion function */
922     acb->bh_ret = ret;
923     acb->bh = qemu_bh_new(qed_aio_complete_bh, acb);
924     qemu_bh_schedule(acb->bh);
925 
926     /* Start next allocating write request waiting behind this one.  Note that
927      * requests enqueue themselves when they first hit an unallocated cluster
928      * but they wait until the entire request is finished before waking up the
929      * next request in the queue.  This ensures that we don't cycle through
930      * requests multiple times but rather finish one at a time completely.
931      */
932     if (acb == QSIMPLEQ_FIRST(&s->allocating_write_reqs)) {
933         QSIMPLEQ_REMOVE_HEAD(&s->allocating_write_reqs, next);
934         acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs);
935         if (acb) {
936             qed_aio_next_io(acb, 0);
937         } else if (s->header.features & QED_F_NEED_CHECK) {
938             qed_start_need_check_timer(s);
939         }
940     }
941 }
942 
943 /**
944  * Commit the current L2 table to the cache
945  */
946 static void qed_commit_l2_update(void *opaque, int ret)
947 {
948     QEDAIOCB *acb = opaque;
949     BDRVQEDState *s = acb_to_s(acb);
950     CachedL2Table *l2_table = acb->request.l2_table;
951     uint64_t l2_offset = l2_table->offset;
952 
953     qed_commit_l2_cache_entry(&s->l2_cache, l2_table);
954 
955     /* This is guaranteed to succeed because we just committed the entry to the
956      * cache.
957      */
958     acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset);
959     assert(acb->request.l2_table != NULL);
960 
961     qed_aio_next_io(opaque, ret);
962 }
963 
964 /**
965  * Update L1 table with new L2 table offset and write it out
966  */
967 static void qed_aio_write_l1_update(void *opaque, int ret)
968 {
969     QEDAIOCB *acb = opaque;
970     BDRVQEDState *s = acb_to_s(acb);
971     int index;
972 
973     if (ret) {
974         qed_aio_complete(acb, ret);
975         return;
976     }
977 
978     index = qed_l1_index(s, acb->cur_pos);
979     s->l1_table->offsets[index] = acb->request.l2_table->offset;
980 
981     qed_write_l1_table(s, index, 1, qed_commit_l2_update, acb);
982 }
983 
984 /**
985  * Update L2 table with new cluster offsets and write them out
986  */
987 static void qed_aio_write_l2_update(QEDAIOCB *acb, int ret, uint64_t offset)
988 {
989     BDRVQEDState *s = acb_to_s(acb);
990     bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1;
991     int index;
992 
993     if (ret) {
994         goto err;
995     }
996 
997     if (need_alloc) {
998         qed_unref_l2_cache_entry(acb->request.l2_table);
999         acb->request.l2_table = qed_new_l2_table(s);
1000     }
1001 
1002     index = qed_l2_index(s, acb->cur_pos);
1003     qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters,
1004                          offset);
1005 
1006     if (need_alloc) {
1007         /* Write out the whole new L2 table */
1008         qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true,
1009                             qed_aio_write_l1_update, acb);
1010     } else {
1011         /* Write out only the updated part of the L2 table */
1012         qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters, false,
1013                             qed_aio_next_io, acb);
1014     }
1015     return;
1016 
1017 err:
1018     qed_aio_complete(acb, ret);
1019 }
1020 
1021 static void qed_aio_write_l2_update_cb(void *opaque, int ret)
1022 {
1023     QEDAIOCB *acb = opaque;
1024     qed_aio_write_l2_update(acb, ret, acb->cur_cluster);
1025 }
1026 
1027 /**
1028  * Flush new data clusters before updating the L2 table
1029  *
1030  * This flush is necessary when a backing file is in use.  A crash during an
1031  * allocating write could result in empty clusters in the image.  If the write
1032  * only touched a subregion of the cluster, then backing image sectors have
1033  * been lost in the untouched region.  The solution is to flush after writing a
1034  * new data cluster and before updating the L2 table.
1035  */
1036 static void qed_aio_write_flush_before_l2_update(void *opaque, int ret)
1037 {
1038     QEDAIOCB *acb = opaque;
1039     BDRVQEDState *s = acb_to_s(acb);
1040 
1041     if (!bdrv_aio_flush(s->bs->file, qed_aio_write_l2_update_cb, opaque)) {
1042         qed_aio_complete(acb, -EIO);
1043     }
1044 }
1045 
1046 /**
1047  * Write data to the image file
1048  */
1049 static void qed_aio_write_main(void *opaque, int ret)
1050 {
1051     QEDAIOCB *acb = opaque;
1052     BDRVQEDState *s = acb_to_s(acb);
1053     uint64_t offset = acb->cur_cluster +
1054                       qed_offset_into_cluster(s, acb->cur_pos);
1055     BlockDriverCompletionFunc *next_fn;
1056 
1057     trace_qed_aio_write_main(s, acb, ret, offset, acb->cur_qiov.size);
1058 
1059     if (ret) {
1060         qed_aio_complete(acb, ret);
1061         return;
1062     }
1063 
1064     if (acb->find_cluster_ret == QED_CLUSTER_FOUND) {
1065         next_fn = qed_aio_next_io;
1066     } else {
1067         if (s->bs->backing_hd) {
1068             next_fn = qed_aio_write_flush_before_l2_update;
1069         } else {
1070             next_fn = qed_aio_write_l2_update_cb;
1071         }
1072     }
1073 
1074     BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO);
1075     bdrv_aio_writev(s->bs->file, offset / BDRV_SECTOR_SIZE,
1076                     &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE,
1077                     next_fn, acb);
1078 }
1079 
1080 /**
1081  * Populate back untouched region of new data cluster
1082  */
1083 static void qed_aio_write_postfill(void *opaque, int ret)
1084 {
1085     QEDAIOCB *acb = opaque;
1086     BDRVQEDState *s = acb_to_s(acb);
1087     uint64_t start = acb->cur_pos + acb->cur_qiov.size;
1088     uint64_t len =
1089         qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start;
1090     uint64_t offset = acb->cur_cluster +
1091                       qed_offset_into_cluster(s, acb->cur_pos) +
1092                       acb->cur_qiov.size;
1093 
1094     if (ret) {
1095         qed_aio_complete(acb, ret);
1096         return;
1097     }
1098 
1099     trace_qed_aio_write_postfill(s, acb, start, len, offset);
1100     qed_copy_from_backing_file(s, start, len, offset,
1101                                 qed_aio_write_main, acb);
1102 }
1103 
1104 /**
1105  * Populate front untouched region of new data cluster
1106  */
1107 static void qed_aio_write_prefill(void *opaque, int ret)
1108 {
1109     QEDAIOCB *acb = opaque;
1110     BDRVQEDState *s = acb_to_s(acb);
1111     uint64_t start = qed_start_of_cluster(s, acb->cur_pos);
1112     uint64_t len = qed_offset_into_cluster(s, acb->cur_pos);
1113 
1114     trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster);
1115     qed_copy_from_backing_file(s, start, len, acb->cur_cluster,
1116                                 qed_aio_write_postfill, acb);
1117 }
1118 
1119 /**
1120  * Check if the QED_F_NEED_CHECK bit should be set during allocating write
1121  */
1122 static bool qed_should_set_need_check(BDRVQEDState *s)
1123 {
1124     /* The flush before L2 update path ensures consistency */
1125     if (s->bs->backing_hd) {
1126         return false;
1127     }
1128 
1129     return !(s->header.features & QED_F_NEED_CHECK);
1130 }
1131 
1132 static void qed_aio_write_zero_cluster(void *opaque, int ret)
1133 {
1134     QEDAIOCB *acb = opaque;
1135 
1136     if (ret) {
1137         qed_aio_complete(acb, ret);
1138         return;
1139     }
1140 
1141     qed_aio_write_l2_update(acb, 0, 1);
1142 }
1143 
1144 /**
1145  * Write new data cluster
1146  *
1147  * @acb:        Write request
1148  * @len:        Length in bytes
1149  *
1150  * This path is taken when writing to previously unallocated clusters.
1151  */
1152 static void qed_aio_write_alloc(QEDAIOCB *acb, size_t len)
1153 {
1154     BDRVQEDState *s = acb_to_s(acb);
1155     BlockDriverCompletionFunc *cb;
1156 
1157     /* Cancel timer when the first allocating request comes in */
1158     if (QSIMPLEQ_EMPTY(&s->allocating_write_reqs)) {
1159         qed_cancel_need_check_timer(s);
1160     }
1161 
1162     /* Freeze this request if another allocating write is in progress */
1163     if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs)) {
1164         QSIMPLEQ_INSERT_TAIL(&s->allocating_write_reqs, acb, next);
1165     }
1166     if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs) ||
1167         s->allocating_write_reqs_plugged) {
1168         return; /* wait for existing request to finish */
1169     }
1170 
1171     acb->cur_nclusters = qed_bytes_to_clusters(s,
1172             qed_offset_into_cluster(s, acb->cur_pos) + len);
1173     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1174 
1175     if (acb->flags & QED_AIOCB_ZERO) {
1176         /* Skip ahead if the clusters are already zero */
1177         if (acb->find_cluster_ret == QED_CLUSTER_ZERO) {
1178             qed_aio_next_io(acb, 0);
1179             return;
1180         }
1181 
1182         cb = qed_aio_write_zero_cluster;
1183     } else {
1184         cb = qed_aio_write_prefill;
1185         acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters);
1186     }
1187 
1188     if (qed_should_set_need_check(s)) {
1189         s->header.features |= QED_F_NEED_CHECK;
1190         qed_write_header(s, cb, acb);
1191     } else {
1192         cb(acb, 0);
1193     }
1194 }
1195 
1196 /**
1197  * Write data cluster in place
1198  *
1199  * @acb:        Write request
1200  * @offset:     Cluster offset in bytes
1201  * @len:        Length in bytes
1202  *
1203  * This path is taken when writing to already allocated clusters.
1204  */
1205 static void qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset, size_t len)
1206 {
1207     /* Allocate buffer for zero writes */
1208     if (acb->flags & QED_AIOCB_ZERO) {
1209         struct iovec *iov = acb->qiov->iov;
1210 
1211         if (!iov->iov_base) {
1212             iov->iov_base = qemu_blockalign(acb->common.bs, iov->iov_len);
1213             memset(iov->iov_base, 0, iov->iov_len);
1214         }
1215     }
1216 
1217     /* Calculate the I/O vector */
1218     acb->cur_cluster = offset;
1219     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1220 
1221     /* Do the actual write */
1222     qed_aio_write_main(acb, 0);
1223 }
1224 
1225 /**
1226  * Write data cluster
1227  *
1228  * @opaque:     Write request
1229  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1,
1230  *              or -errno
1231  * @offset:     Cluster offset in bytes
1232  * @len:        Length in bytes
1233  *
1234  * Callback from qed_find_cluster().
1235  */
1236 static void qed_aio_write_data(void *opaque, int ret,
1237                                uint64_t offset, size_t len)
1238 {
1239     QEDAIOCB *acb = opaque;
1240 
1241     trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len);
1242 
1243     acb->find_cluster_ret = ret;
1244 
1245     switch (ret) {
1246     case QED_CLUSTER_FOUND:
1247         qed_aio_write_inplace(acb, offset, len);
1248         break;
1249 
1250     case QED_CLUSTER_L2:
1251     case QED_CLUSTER_L1:
1252     case QED_CLUSTER_ZERO:
1253         qed_aio_write_alloc(acb, len);
1254         break;
1255 
1256     default:
1257         qed_aio_complete(acb, ret);
1258         break;
1259     }
1260 }
1261 
1262 /**
1263  * Read data cluster
1264  *
1265  * @opaque:     Read request
1266  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1,
1267  *              or -errno
1268  * @offset:     Cluster offset in bytes
1269  * @len:        Length in bytes
1270  *
1271  * Callback from qed_find_cluster().
1272  */
1273 static void qed_aio_read_data(void *opaque, int ret,
1274                               uint64_t offset, size_t len)
1275 {
1276     QEDAIOCB *acb = opaque;
1277     BDRVQEDState *s = acb_to_s(acb);
1278     BlockDriverState *bs = acb->common.bs;
1279 
1280     /* Adjust offset into cluster */
1281     offset += qed_offset_into_cluster(s, acb->cur_pos);
1282 
1283     trace_qed_aio_read_data(s, acb, ret, offset, len);
1284 
1285     if (ret < 0) {
1286         goto err;
1287     }
1288 
1289     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1290 
1291     /* Handle zero cluster and backing file reads */
1292     if (ret == QED_CLUSTER_ZERO) {
1293         qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size);
1294         qed_aio_next_io(acb, 0);
1295         return;
1296     } else if (ret != QED_CLUSTER_FOUND) {
1297         qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov,
1298                               qed_aio_next_io, acb);
1299         return;
1300     }
1301 
1302     BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
1303     bdrv_aio_readv(bs->file, offset / BDRV_SECTOR_SIZE,
1304                    &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE,
1305                    qed_aio_next_io, acb);
1306     return;
1307 
1308 err:
1309     qed_aio_complete(acb, ret);
1310 }
1311 
1312 /**
1313  * Begin next I/O or complete the request
1314  */
1315 static void qed_aio_next_io(void *opaque, int ret)
1316 {
1317     QEDAIOCB *acb = opaque;
1318     BDRVQEDState *s = acb_to_s(acb);
1319     QEDFindClusterFunc *io_fn = (acb->flags & QED_AIOCB_WRITE) ?
1320                                 qed_aio_write_data : qed_aio_read_data;
1321 
1322     trace_qed_aio_next_io(s, acb, ret, acb->cur_pos + acb->cur_qiov.size);
1323 
1324     /* Handle I/O error */
1325     if (ret) {
1326         qed_aio_complete(acb, ret);
1327         return;
1328     }
1329 
1330     acb->qiov_offset += acb->cur_qiov.size;
1331     acb->cur_pos += acb->cur_qiov.size;
1332     qemu_iovec_reset(&acb->cur_qiov);
1333 
1334     /* Complete request */
1335     if (acb->cur_pos >= acb->end_pos) {
1336         qed_aio_complete(acb, 0);
1337         return;
1338     }
1339 
1340     /* Find next cluster and start I/O */
1341     qed_find_cluster(s, &acb->request,
1342                       acb->cur_pos, acb->end_pos - acb->cur_pos,
1343                       io_fn, acb);
1344 }
1345 
1346 static BlockDriverAIOCB *qed_aio_setup(BlockDriverState *bs,
1347                                        int64_t sector_num,
1348                                        QEMUIOVector *qiov, int nb_sectors,
1349                                        BlockDriverCompletionFunc *cb,
1350                                        void *opaque, int flags)
1351 {
1352     QEDAIOCB *acb = qemu_aio_get(&qed_aiocb_info, bs, cb, opaque);
1353 
1354     trace_qed_aio_setup(bs->opaque, acb, sector_num, nb_sectors,
1355                         opaque, flags);
1356 
1357     acb->flags = flags;
1358     acb->finished = NULL;
1359     acb->qiov = qiov;
1360     acb->qiov_offset = 0;
1361     acb->cur_pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE;
1362     acb->end_pos = acb->cur_pos + nb_sectors * BDRV_SECTOR_SIZE;
1363     acb->request.l2_table = NULL;
1364     qemu_iovec_init(&acb->cur_qiov, qiov->niov);
1365 
1366     /* Start request */
1367     qed_aio_next_io(acb, 0);
1368     return &acb->common;
1369 }
1370 
1371 static BlockDriverAIOCB *bdrv_qed_aio_readv(BlockDriverState *bs,
1372                                             int64_t sector_num,
1373                                             QEMUIOVector *qiov, int nb_sectors,
1374                                             BlockDriverCompletionFunc *cb,
1375                                             void *opaque)
1376 {
1377     return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb, opaque, 0);
1378 }
1379 
1380 static BlockDriverAIOCB *bdrv_qed_aio_writev(BlockDriverState *bs,
1381                                              int64_t sector_num,
1382                                              QEMUIOVector *qiov, int nb_sectors,
1383                                              BlockDriverCompletionFunc *cb,
1384                                              void *opaque)
1385 {
1386     return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb,
1387                          opaque, QED_AIOCB_WRITE);
1388 }
1389 
1390 typedef struct {
1391     Coroutine *co;
1392     int ret;
1393     bool done;
1394 } QEDWriteZeroesCB;
1395 
1396 static void coroutine_fn qed_co_write_zeroes_cb(void *opaque, int ret)
1397 {
1398     QEDWriteZeroesCB *cb = opaque;
1399 
1400     cb->done = true;
1401     cb->ret = ret;
1402     if (cb->co) {
1403         qemu_coroutine_enter(cb->co, NULL);
1404     }
1405 }
1406 
1407 static int coroutine_fn bdrv_qed_co_write_zeroes(BlockDriverState *bs,
1408                                                  int64_t sector_num,
1409                                                  int nb_sectors,
1410                                                  BdrvRequestFlags flags)
1411 {
1412     BlockDriverAIOCB *blockacb;
1413     BDRVQEDState *s = bs->opaque;
1414     QEDWriteZeroesCB cb = { .done = false };
1415     QEMUIOVector qiov;
1416     struct iovec iov;
1417 
1418     /* Refuse if there are untouched backing file sectors */
1419     if (bs->backing_hd) {
1420         if (qed_offset_into_cluster(s, sector_num * BDRV_SECTOR_SIZE) != 0) {
1421             return -ENOTSUP;
1422         }
1423         if (qed_offset_into_cluster(s, nb_sectors * BDRV_SECTOR_SIZE) != 0) {
1424             return -ENOTSUP;
1425         }
1426     }
1427 
1428     /* Zero writes start without an I/O buffer.  If a buffer becomes necessary
1429      * then it will be allocated during request processing.
1430      */
1431     iov.iov_base = NULL,
1432     iov.iov_len  = nb_sectors * BDRV_SECTOR_SIZE,
1433 
1434     qemu_iovec_init_external(&qiov, &iov, 1);
1435     blockacb = qed_aio_setup(bs, sector_num, &qiov, nb_sectors,
1436                              qed_co_write_zeroes_cb, &cb,
1437                              QED_AIOCB_WRITE | QED_AIOCB_ZERO);
1438     if (!blockacb) {
1439         return -EIO;
1440     }
1441     if (!cb.done) {
1442         cb.co = qemu_coroutine_self();
1443         qemu_coroutine_yield();
1444     }
1445     assert(cb.done);
1446     return cb.ret;
1447 }
1448 
1449 static int bdrv_qed_truncate(BlockDriverState *bs, int64_t offset)
1450 {
1451     BDRVQEDState *s = bs->opaque;
1452     uint64_t old_image_size;
1453     int ret;
1454 
1455     if (!qed_is_image_size_valid(offset, s->header.cluster_size,
1456                                  s->header.table_size)) {
1457         return -EINVAL;
1458     }
1459 
1460     /* Shrinking is currently not supported */
1461     if ((uint64_t)offset < s->header.image_size) {
1462         return -ENOTSUP;
1463     }
1464 
1465     old_image_size = s->header.image_size;
1466     s->header.image_size = offset;
1467     ret = qed_write_header_sync(s);
1468     if (ret < 0) {
1469         s->header.image_size = old_image_size;
1470     }
1471     return ret;
1472 }
1473 
1474 static int64_t bdrv_qed_getlength(BlockDriverState *bs)
1475 {
1476     BDRVQEDState *s = bs->opaque;
1477     return s->header.image_size;
1478 }
1479 
1480 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1481 {
1482     BDRVQEDState *s = bs->opaque;
1483 
1484     memset(bdi, 0, sizeof(*bdi));
1485     bdi->cluster_size = s->header.cluster_size;
1486     bdi->is_dirty = s->header.features & QED_F_NEED_CHECK;
1487     bdi->unallocated_blocks_are_zero = true;
1488     bdi->can_write_zeroes_with_unmap = true;
1489     return 0;
1490 }
1491 
1492 static int bdrv_qed_change_backing_file(BlockDriverState *bs,
1493                                         const char *backing_file,
1494                                         const char *backing_fmt)
1495 {
1496     BDRVQEDState *s = bs->opaque;
1497     QEDHeader new_header, le_header;
1498     void *buffer;
1499     size_t buffer_len, backing_file_len;
1500     int ret;
1501 
1502     /* Refuse to set backing filename if unknown compat feature bits are
1503      * active.  If the image uses an unknown compat feature then we may not
1504      * know the layout of data following the header structure and cannot safely
1505      * add a new string.
1506      */
1507     if (backing_file && (s->header.compat_features &
1508                          ~QED_COMPAT_FEATURE_MASK)) {
1509         return -ENOTSUP;
1510     }
1511 
1512     memcpy(&new_header, &s->header, sizeof(new_header));
1513 
1514     new_header.features &= ~(QED_F_BACKING_FILE |
1515                              QED_F_BACKING_FORMAT_NO_PROBE);
1516 
1517     /* Adjust feature flags */
1518     if (backing_file) {
1519         new_header.features |= QED_F_BACKING_FILE;
1520 
1521         if (qed_fmt_is_raw(backing_fmt)) {
1522             new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
1523         }
1524     }
1525 
1526     /* Calculate new header size */
1527     backing_file_len = 0;
1528 
1529     if (backing_file) {
1530         backing_file_len = strlen(backing_file);
1531     }
1532 
1533     buffer_len = sizeof(new_header);
1534     new_header.backing_filename_offset = buffer_len;
1535     new_header.backing_filename_size = backing_file_len;
1536     buffer_len += backing_file_len;
1537 
1538     /* Make sure we can rewrite header without failing */
1539     if (buffer_len > new_header.header_size * new_header.cluster_size) {
1540         return -ENOSPC;
1541     }
1542 
1543     /* Prepare new header */
1544     buffer = g_malloc(buffer_len);
1545 
1546     qed_header_cpu_to_le(&new_header, &le_header);
1547     memcpy(buffer, &le_header, sizeof(le_header));
1548     buffer_len = sizeof(le_header);
1549 
1550     if (backing_file) {
1551         memcpy(buffer + buffer_len, backing_file, backing_file_len);
1552         buffer_len += backing_file_len;
1553     }
1554 
1555     /* Write new header */
1556     ret = bdrv_pwrite_sync(bs->file, 0, buffer, buffer_len);
1557     g_free(buffer);
1558     if (ret == 0) {
1559         memcpy(&s->header, &new_header, sizeof(new_header));
1560     }
1561     return ret;
1562 }
1563 
1564 static void bdrv_qed_invalidate_cache(BlockDriverState *bs)
1565 {
1566     BDRVQEDState *s = bs->opaque;
1567 
1568     bdrv_qed_close(bs);
1569     memset(s, 0, sizeof(BDRVQEDState));
1570     bdrv_qed_open(bs, NULL, bs->open_flags, NULL);
1571 }
1572 
1573 static int bdrv_qed_check(BlockDriverState *bs, BdrvCheckResult *result,
1574                           BdrvCheckMode fix)
1575 {
1576     BDRVQEDState *s = bs->opaque;
1577 
1578     return qed_check(s, result, !!fix);
1579 }
1580 
1581 static QEMUOptionParameter qed_create_options[] = {
1582     {
1583         .name = BLOCK_OPT_SIZE,
1584         .type = OPT_SIZE,
1585         .help = "Virtual disk size (in bytes)"
1586     }, {
1587         .name = BLOCK_OPT_BACKING_FILE,
1588         .type = OPT_STRING,
1589         .help = "File name of a base image"
1590     }, {
1591         .name = BLOCK_OPT_BACKING_FMT,
1592         .type = OPT_STRING,
1593         .help = "Image format of the base image"
1594     }, {
1595         .name = BLOCK_OPT_CLUSTER_SIZE,
1596         .type = OPT_SIZE,
1597         .help = "Cluster size (in bytes)",
1598         .value = { .n = QED_DEFAULT_CLUSTER_SIZE },
1599     }, {
1600         .name = BLOCK_OPT_TABLE_SIZE,
1601         .type = OPT_SIZE,
1602         .help = "L1/L2 table size (in clusters)"
1603     },
1604     { /* end of list */ }
1605 };
1606 
1607 static BlockDriver bdrv_qed = {
1608     .format_name              = "qed",
1609     .instance_size            = sizeof(BDRVQEDState),
1610     .create_options           = qed_create_options,
1611 
1612     .bdrv_probe               = bdrv_qed_probe,
1613     .bdrv_rebind              = bdrv_qed_rebind,
1614     .bdrv_open                = bdrv_qed_open,
1615     .bdrv_close               = bdrv_qed_close,
1616     .bdrv_reopen_prepare      = bdrv_qed_reopen_prepare,
1617     .bdrv_create              = bdrv_qed_create,
1618     .bdrv_has_zero_init       = bdrv_has_zero_init_1,
1619     .bdrv_co_get_block_status = bdrv_qed_co_get_block_status,
1620     .bdrv_make_empty          = bdrv_qed_make_empty,
1621     .bdrv_aio_readv           = bdrv_qed_aio_readv,
1622     .bdrv_aio_writev          = bdrv_qed_aio_writev,
1623     .bdrv_co_write_zeroes     = bdrv_qed_co_write_zeroes,
1624     .bdrv_truncate            = bdrv_qed_truncate,
1625     .bdrv_getlength           = bdrv_qed_getlength,
1626     .bdrv_get_info            = bdrv_qed_get_info,
1627     .bdrv_refresh_limits      = bdrv_qed_refresh_limits,
1628     .bdrv_change_backing_file = bdrv_qed_change_backing_file,
1629     .bdrv_invalidate_cache    = bdrv_qed_invalidate_cache,
1630     .bdrv_check               = bdrv_qed_check,
1631 };
1632 
1633 static void bdrv_qed_init(void)
1634 {
1635     bdrv_register(&bdrv_qed);
1636 }
1637 
1638 block_init(bdrv_qed_init);
1639