xref: /openbmc/qemu/block/qed.c (revision 228aa992)
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 const AIOCBInfo qed_aiocb_info = {
22     .aiocb_size         = sizeof(QEDAIOCB),
23 };
24 
25 static int bdrv_qed_probe(const uint8_t *buf, int buf_size,
26                           const char *filename)
27 {
28     const QEDHeader *header = (const QEDHeader *)buf;
29 
30     if (buf_size < sizeof(*header)) {
31         return 0;
32     }
33     if (le32_to_cpu(header->magic) != QED_MAGIC) {
34         return 0;
35     }
36     return 100;
37 }
38 
39 /**
40  * Check whether an image format is raw
41  *
42  * @fmt:    Backing file format, may be NULL
43  */
44 static bool qed_fmt_is_raw(const char *fmt)
45 {
46     return fmt && strcmp(fmt, "raw") == 0;
47 }
48 
49 static void qed_header_le_to_cpu(const QEDHeader *le, QEDHeader *cpu)
50 {
51     cpu->magic = le32_to_cpu(le->magic);
52     cpu->cluster_size = le32_to_cpu(le->cluster_size);
53     cpu->table_size = le32_to_cpu(le->table_size);
54     cpu->header_size = le32_to_cpu(le->header_size);
55     cpu->features = le64_to_cpu(le->features);
56     cpu->compat_features = le64_to_cpu(le->compat_features);
57     cpu->autoclear_features = le64_to_cpu(le->autoclear_features);
58     cpu->l1_table_offset = le64_to_cpu(le->l1_table_offset);
59     cpu->image_size = le64_to_cpu(le->image_size);
60     cpu->backing_filename_offset = le32_to_cpu(le->backing_filename_offset);
61     cpu->backing_filename_size = le32_to_cpu(le->backing_filename_size);
62 }
63 
64 static void qed_header_cpu_to_le(const QEDHeader *cpu, QEDHeader *le)
65 {
66     le->magic = cpu_to_le32(cpu->magic);
67     le->cluster_size = cpu_to_le32(cpu->cluster_size);
68     le->table_size = cpu_to_le32(cpu->table_size);
69     le->header_size = cpu_to_le32(cpu->header_size);
70     le->features = cpu_to_le64(cpu->features);
71     le->compat_features = cpu_to_le64(cpu->compat_features);
72     le->autoclear_features = cpu_to_le64(cpu->autoclear_features);
73     le->l1_table_offset = cpu_to_le64(cpu->l1_table_offset);
74     le->image_size = cpu_to_le64(cpu->image_size);
75     le->backing_filename_offset = cpu_to_le32(cpu->backing_filename_offset);
76     le->backing_filename_size = cpu_to_le32(cpu->backing_filename_size);
77 }
78 
79 int qed_write_header_sync(BDRVQEDState *s)
80 {
81     QEDHeader le;
82     int ret;
83 
84     qed_header_cpu_to_le(&s->header, &le);
85     ret = bdrv_pwrite(s->bs->file, 0, &le, sizeof(le));
86     if (ret != sizeof(le)) {
87         return ret;
88     }
89     return 0;
90 }
91 
92 typedef struct {
93     GenericCB gencb;
94     BDRVQEDState *s;
95     struct iovec iov;
96     QEMUIOVector qiov;
97     int nsectors;
98     uint8_t *buf;
99 } QEDWriteHeaderCB;
100 
101 static void qed_write_header_cb(void *opaque, int ret)
102 {
103     QEDWriteHeaderCB *write_header_cb = opaque;
104 
105     qemu_vfree(write_header_cb->buf);
106     gencb_complete(write_header_cb, ret);
107 }
108 
109 static void qed_write_header_read_cb(void *opaque, int ret)
110 {
111     QEDWriteHeaderCB *write_header_cb = opaque;
112     BDRVQEDState *s = write_header_cb->s;
113 
114     if (ret) {
115         qed_write_header_cb(write_header_cb, ret);
116         return;
117     }
118 
119     /* Update header */
120     qed_header_cpu_to_le(&s->header, (QEDHeader *)write_header_cb->buf);
121 
122     bdrv_aio_writev(s->bs->file, 0, &write_header_cb->qiov,
123                     write_header_cb->nsectors, qed_write_header_cb,
124                     write_header_cb);
125 }
126 
127 /**
128  * Update header in-place (does not rewrite backing filename or other strings)
129  *
130  * This function only updates known header fields in-place and does not affect
131  * extra data after the QED header.
132  */
133 static void qed_write_header(BDRVQEDState *s, BlockCompletionFunc cb,
134                              void *opaque)
135 {
136     /* We must write full sectors for O_DIRECT but cannot necessarily generate
137      * the data following the header if an unrecognized compat feature is
138      * active.  Therefore, first read the sectors containing the header, update
139      * them, and write back.
140      */
141 
142     int nsectors = (sizeof(QEDHeader) + BDRV_SECTOR_SIZE - 1) /
143                    BDRV_SECTOR_SIZE;
144     size_t len = nsectors * BDRV_SECTOR_SIZE;
145     QEDWriteHeaderCB *write_header_cb = gencb_alloc(sizeof(*write_header_cb),
146                                                     cb, opaque);
147 
148     write_header_cb->s = s;
149     write_header_cb->nsectors = nsectors;
150     write_header_cb->buf = qemu_blockalign(s->bs, len);
151     write_header_cb->iov.iov_base = write_header_cb->buf;
152     write_header_cb->iov.iov_len = len;
153     qemu_iovec_init_external(&write_header_cb->qiov, &write_header_cb->iov, 1);
154 
155     bdrv_aio_readv(s->bs->file, 0, &write_header_cb->qiov, nsectors,
156                    qed_write_header_read_cb, write_header_cb);
157 }
158 
159 static uint64_t qed_max_image_size(uint32_t cluster_size, uint32_t table_size)
160 {
161     uint64_t table_entries;
162     uint64_t l2_size;
163 
164     table_entries = (table_size * cluster_size) / sizeof(uint64_t);
165     l2_size = table_entries * cluster_size;
166 
167     return l2_size * table_entries;
168 }
169 
170 static bool qed_is_cluster_size_valid(uint32_t cluster_size)
171 {
172     if (cluster_size < QED_MIN_CLUSTER_SIZE ||
173         cluster_size > QED_MAX_CLUSTER_SIZE) {
174         return false;
175     }
176     if (cluster_size & (cluster_size - 1)) {
177         return false; /* not power of 2 */
178     }
179     return true;
180 }
181 
182 static bool qed_is_table_size_valid(uint32_t table_size)
183 {
184     if (table_size < QED_MIN_TABLE_SIZE ||
185         table_size > QED_MAX_TABLE_SIZE) {
186         return false;
187     }
188     if (table_size & (table_size - 1)) {
189         return false; /* not power of 2 */
190     }
191     return true;
192 }
193 
194 static bool qed_is_image_size_valid(uint64_t image_size, uint32_t cluster_size,
195                                     uint32_t table_size)
196 {
197     if (image_size % BDRV_SECTOR_SIZE != 0) {
198         return false; /* not multiple of sector size */
199     }
200     if (image_size > qed_max_image_size(cluster_size, table_size)) {
201         return false; /* image is too large */
202     }
203     return true;
204 }
205 
206 /**
207  * Read a string of known length from the image file
208  *
209  * @file:       Image file
210  * @offset:     File offset to start of string, in bytes
211  * @n:          String length in bytes
212  * @buf:        Destination buffer
213  * @buflen:     Destination buffer length in bytes
214  * @ret:        0 on success, -errno on failure
215  *
216  * The string is NUL-terminated.
217  */
218 static int qed_read_string(BlockDriverState *file, uint64_t offset, size_t n,
219                            char *buf, size_t buflen)
220 {
221     int ret;
222     if (n >= buflen) {
223         return -EINVAL;
224     }
225     ret = bdrv_pread(file, offset, buf, n);
226     if (ret < 0) {
227         return ret;
228     }
229     buf[n] = '\0';
230     return 0;
231 }
232 
233 /**
234  * Allocate new clusters
235  *
236  * @s:          QED state
237  * @n:          Number of contiguous clusters to allocate
238  * @ret:        Offset of first allocated cluster
239  *
240  * This function only produces the offset where the new clusters should be
241  * written.  It updates BDRVQEDState but does not make any changes to the image
242  * file.
243  */
244 static uint64_t qed_alloc_clusters(BDRVQEDState *s, unsigned int n)
245 {
246     uint64_t offset = s->file_size;
247     s->file_size += n * s->header.cluster_size;
248     return offset;
249 }
250 
251 QEDTable *qed_alloc_table(BDRVQEDState *s)
252 {
253     /* Honor O_DIRECT memory alignment requirements */
254     return qemu_blockalign(s->bs,
255                            s->header.cluster_size * s->header.table_size);
256 }
257 
258 /**
259  * Allocate a new zeroed L2 table
260  */
261 static CachedL2Table *qed_new_l2_table(BDRVQEDState *s)
262 {
263     CachedL2Table *l2_table = qed_alloc_l2_cache_entry(&s->l2_cache);
264 
265     l2_table->table = qed_alloc_table(s);
266     l2_table->offset = qed_alloc_clusters(s, s->header.table_size);
267 
268     memset(l2_table->table->offsets, 0,
269            s->header.cluster_size * s->header.table_size);
270     return l2_table;
271 }
272 
273 static void qed_aio_next_io(void *opaque, int ret);
274 
275 static void qed_plug_allocating_write_reqs(BDRVQEDState *s)
276 {
277     assert(!s->allocating_write_reqs_plugged);
278 
279     s->allocating_write_reqs_plugged = true;
280 }
281 
282 static void qed_unplug_allocating_write_reqs(BDRVQEDState *s)
283 {
284     QEDAIOCB *acb;
285 
286     assert(s->allocating_write_reqs_plugged);
287 
288     s->allocating_write_reqs_plugged = false;
289 
290     acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs);
291     if (acb) {
292         qed_aio_next_io(acb, 0);
293     }
294 }
295 
296 static void qed_finish_clear_need_check(void *opaque, int ret)
297 {
298     /* Do nothing */
299 }
300 
301 static void qed_flush_after_clear_need_check(void *opaque, int ret)
302 {
303     BDRVQEDState *s = opaque;
304 
305     bdrv_aio_flush(s->bs, qed_finish_clear_need_check, s);
306 
307     /* No need to wait until flush completes */
308     qed_unplug_allocating_write_reqs(s);
309 }
310 
311 static void qed_clear_need_check(void *opaque, int ret)
312 {
313     BDRVQEDState *s = opaque;
314 
315     if (ret) {
316         qed_unplug_allocating_write_reqs(s);
317         return;
318     }
319 
320     s->header.features &= ~QED_F_NEED_CHECK;
321     qed_write_header(s, qed_flush_after_clear_need_check, s);
322 }
323 
324 static void qed_need_check_timer_cb(void *opaque)
325 {
326     BDRVQEDState *s = opaque;
327 
328     /* The timer should only fire when allocating writes have drained */
329     assert(!QSIMPLEQ_FIRST(&s->allocating_write_reqs));
330 
331     trace_qed_need_check_timer_cb(s);
332 
333     qed_plug_allocating_write_reqs(s);
334 
335     /* Ensure writes are on disk before clearing flag */
336     bdrv_aio_flush(s->bs, qed_clear_need_check, s);
337 }
338 
339 static void qed_start_need_check_timer(BDRVQEDState *s)
340 {
341     trace_qed_start_need_check_timer(s);
342 
343     /* Use QEMU_CLOCK_VIRTUAL so we don't alter the image file while suspended for
344      * migration.
345      */
346     timer_mod(s->need_check_timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
347                    get_ticks_per_sec() * QED_NEED_CHECK_TIMEOUT);
348 }
349 
350 /* It's okay to call this multiple times or when no timer is started */
351 static void qed_cancel_need_check_timer(BDRVQEDState *s)
352 {
353     trace_qed_cancel_need_check_timer(s);
354     timer_del(s->need_check_timer);
355 }
356 
357 static void bdrv_qed_rebind(BlockDriverState *bs)
358 {
359     BDRVQEDState *s = bs->opaque;
360     s->bs = bs;
361 }
362 
363 static void bdrv_qed_detach_aio_context(BlockDriverState *bs)
364 {
365     BDRVQEDState *s = bs->opaque;
366 
367     qed_cancel_need_check_timer(s);
368     timer_free(s->need_check_timer);
369 }
370 
371 static void bdrv_qed_attach_aio_context(BlockDriverState *bs,
372                                         AioContext *new_context)
373 {
374     BDRVQEDState *s = bs->opaque;
375 
376     s->need_check_timer = aio_timer_new(new_context,
377                                         QEMU_CLOCK_VIRTUAL, SCALE_NS,
378                                         qed_need_check_timer_cb, s);
379     if (s->header.features & QED_F_NEED_CHECK) {
380         qed_start_need_check_timer(s);
381     }
382 }
383 
384 static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags,
385                          Error **errp)
386 {
387     BDRVQEDState *s = bs->opaque;
388     QEDHeader le_header;
389     int64_t file_size;
390     int ret;
391 
392     s->bs = bs;
393     QSIMPLEQ_INIT(&s->allocating_write_reqs);
394 
395     ret = bdrv_pread(bs->file, 0, &le_header, sizeof(le_header));
396     if (ret < 0) {
397         return ret;
398     }
399     qed_header_le_to_cpu(&le_header, &s->header);
400 
401     if (s->header.magic != QED_MAGIC) {
402         error_setg(errp, "Image not in QED format");
403         return -EINVAL;
404     }
405     if (s->header.features & ~QED_FEATURE_MASK) {
406         /* image uses unsupported feature bits */
407         char buf[64];
408         snprintf(buf, sizeof(buf), "%" PRIx64,
409             s->header.features & ~QED_FEATURE_MASK);
410         error_set(errp, QERR_UNKNOWN_BLOCK_FORMAT_FEATURE,
411             bdrv_get_device_name(bs), "QED", buf);
412         return -ENOTSUP;
413     }
414     if (!qed_is_cluster_size_valid(s->header.cluster_size)) {
415         return -EINVAL;
416     }
417 
418     /* Round down file size to the last cluster */
419     file_size = bdrv_getlength(bs->file);
420     if (file_size < 0) {
421         return file_size;
422     }
423     s->file_size = qed_start_of_cluster(s, file_size);
424 
425     if (!qed_is_table_size_valid(s->header.table_size)) {
426         return -EINVAL;
427     }
428     if (!qed_is_image_size_valid(s->header.image_size,
429                                  s->header.cluster_size,
430                                  s->header.table_size)) {
431         return -EINVAL;
432     }
433     if (!qed_check_table_offset(s, s->header.l1_table_offset)) {
434         return -EINVAL;
435     }
436 
437     s->table_nelems = (s->header.cluster_size * s->header.table_size) /
438                       sizeof(uint64_t);
439     s->l2_shift = ffs(s->header.cluster_size) - 1;
440     s->l2_mask = s->table_nelems - 1;
441     s->l1_shift = s->l2_shift + ffs(s->table_nelems) - 1;
442 
443     if ((s->header.features & QED_F_BACKING_FILE)) {
444         if ((uint64_t)s->header.backing_filename_offset +
445             s->header.backing_filename_size >
446             s->header.cluster_size * s->header.header_size) {
447             return -EINVAL;
448         }
449 
450         ret = qed_read_string(bs->file, s->header.backing_filename_offset,
451                               s->header.backing_filename_size, bs->backing_file,
452                               sizeof(bs->backing_file));
453         if (ret < 0) {
454             return ret;
455         }
456 
457         if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) {
458             pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw");
459         }
460     }
461 
462     /* Reset unknown autoclear feature bits.  This is a backwards
463      * compatibility mechanism that allows images to be opened by older
464      * programs, which "knock out" unknown feature bits.  When an image is
465      * opened by a newer program again it can detect that the autoclear
466      * feature is no longer valid.
467      */
468     if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 &&
469         !bdrv_is_read_only(bs->file) && !(flags & BDRV_O_INCOMING)) {
470         s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK;
471 
472         ret = qed_write_header_sync(s);
473         if (ret) {
474             return ret;
475         }
476 
477         /* From here on only known autoclear feature bits are valid */
478         bdrv_flush(bs->file);
479     }
480 
481     s->l1_table = qed_alloc_table(s);
482     qed_init_l2_cache(&s->l2_cache);
483 
484     ret = qed_read_l1_table_sync(s);
485     if (ret) {
486         goto out;
487     }
488 
489     /* If image was not closed cleanly, check consistency */
490     if (!(flags & BDRV_O_CHECK) && (s->header.features & QED_F_NEED_CHECK)) {
491         /* Read-only images cannot be fixed.  There is no risk of corruption
492          * since write operations are not possible.  Therefore, allow
493          * potentially inconsistent images to be opened read-only.  This can
494          * aid data recovery from an otherwise inconsistent image.
495          */
496         if (!bdrv_is_read_only(bs->file) &&
497             !(flags & BDRV_O_INCOMING)) {
498             BdrvCheckResult result = {0};
499 
500             ret = qed_check(s, &result, true);
501             if (ret) {
502                 goto out;
503             }
504         }
505     }
506 
507     bdrv_qed_attach_aio_context(bs, bdrv_get_aio_context(bs));
508 
509 out:
510     if (ret) {
511         qed_free_l2_cache(&s->l2_cache);
512         qemu_vfree(s->l1_table);
513     }
514     return ret;
515 }
516 
517 static void bdrv_qed_refresh_limits(BlockDriverState *bs, Error **errp)
518 {
519     BDRVQEDState *s = bs->opaque;
520 
521     bs->bl.write_zeroes_alignment = s->header.cluster_size >> BDRV_SECTOR_BITS;
522 }
523 
524 /* We have nothing to do for QED reopen, stubs just return
525  * success */
526 static int bdrv_qed_reopen_prepare(BDRVReopenState *state,
527                                    BlockReopenQueue *queue, Error **errp)
528 {
529     return 0;
530 }
531 
532 static void bdrv_qed_close(BlockDriverState *bs)
533 {
534     BDRVQEDState *s = bs->opaque;
535 
536     bdrv_qed_detach_aio_context(bs);
537 
538     /* Ensure writes reach stable storage */
539     bdrv_flush(bs->file);
540 
541     /* Clean shutdown, no check required on next open */
542     if (s->header.features & QED_F_NEED_CHECK) {
543         s->header.features &= ~QED_F_NEED_CHECK;
544         qed_write_header_sync(s);
545     }
546 
547     qed_free_l2_cache(&s->l2_cache);
548     qemu_vfree(s->l1_table);
549 }
550 
551 static int qed_create(const char *filename, uint32_t cluster_size,
552                       uint64_t image_size, uint32_t table_size,
553                       const char *backing_file, const char *backing_fmt,
554                       QemuOpts *opts, Error **errp)
555 {
556     QEDHeader header = {
557         .magic = QED_MAGIC,
558         .cluster_size = cluster_size,
559         .table_size = table_size,
560         .header_size = 1,
561         .features = 0,
562         .compat_features = 0,
563         .l1_table_offset = cluster_size,
564         .image_size = image_size,
565     };
566     QEDHeader le_header;
567     uint8_t *l1_table = NULL;
568     size_t l1_size = header.cluster_size * header.table_size;
569     Error *local_err = NULL;
570     int ret = 0;
571     BlockDriverState *bs;
572 
573     ret = bdrv_create_file(filename, opts, &local_err);
574     if (ret < 0) {
575         error_propagate(errp, local_err);
576         return ret;
577     }
578 
579     bs = NULL;
580     ret = bdrv_open(&bs, filename, NULL, NULL,
581                     BDRV_O_RDWR | BDRV_O_CACHE_WB | BDRV_O_PROTOCOL, NULL,
582                     &local_err);
583     if (ret < 0) {
584         error_propagate(errp, local_err);
585         return ret;
586     }
587 
588     /* File must start empty and grow, check truncate is supported */
589     ret = bdrv_truncate(bs, 0);
590     if (ret < 0) {
591         goto out;
592     }
593 
594     if (backing_file) {
595         header.features |= QED_F_BACKING_FILE;
596         header.backing_filename_offset = sizeof(le_header);
597         header.backing_filename_size = strlen(backing_file);
598 
599         if (qed_fmt_is_raw(backing_fmt)) {
600             header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
601         }
602     }
603 
604     qed_header_cpu_to_le(&header, &le_header);
605     ret = bdrv_pwrite(bs, 0, &le_header, sizeof(le_header));
606     if (ret < 0) {
607         goto out;
608     }
609     ret = bdrv_pwrite(bs, sizeof(le_header), backing_file,
610                       header.backing_filename_size);
611     if (ret < 0) {
612         goto out;
613     }
614 
615     l1_table = g_malloc0(l1_size);
616     ret = bdrv_pwrite(bs, header.l1_table_offset, l1_table, l1_size);
617     if (ret < 0) {
618         goto out;
619     }
620 
621     ret = 0; /* success */
622 out:
623     g_free(l1_table);
624     bdrv_unref(bs);
625     return ret;
626 }
627 
628 static int bdrv_qed_create(const char *filename, QemuOpts *opts, Error **errp)
629 {
630     uint64_t image_size = 0;
631     uint32_t cluster_size = QED_DEFAULT_CLUSTER_SIZE;
632     uint32_t table_size = QED_DEFAULT_TABLE_SIZE;
633     char *backing_file = NULL;
634     char *backing_fmt = NULL;
635     int ret;
636 
637     image_size = ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
638                           BDRV_SECTOR_SIZE);
639     backing_file = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FILE);
640     backing_fmt = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FMT);
641     cluster_size = qemu_opt_get_size_del(opts,
642                                          BLOCK_OPT_CLUSTER_SIZE,
643                                          QED_DEFAULT_CLUSTER_SIZE);
644     table_size = qemu_opt_get_size_del(opts, BLOCK_OPT_TABLE_SIZE,
645                                        QED_DEFAULT_TABLE_SIZE);
646 
647     if (!qed_is_cluster_size_valid(cluster_size)) {
648         error_setg(errp, "QED cluster size must be within range [%u, %u] "
649                          "and power of 2",
650                    QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE);
651         ret = -EINVAL;
652         goto finish;
653     }
654     if (!qed_is_table_size_valid(table_size)) {
655         error_setg(errp, "QED table size must be within range [%u, %u] "
656                          "and power of 2",
657                    QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE);
658         ret = -EINVAL;
659         goto finish;
660     }
661     if (!qed_is_image_size_valid(image_size, cluster_size, table_size)) {
662         error_setg(errp, "QED image size must be a non-zero multiple of "
663                          "cluster size and less than %" PRIu64 " bytes",
664                    qed_max_image_size(cluster_size, table_size));
665         ret = -EINVAL;
666         goto finish;
667     }
668 
669     ret = qed_create(filename, cluster_size, image_size, table_size,
670                      backing_file, backing_fmt, opts, errp);
671 
672 finish:
673     g_free(backing_file);
674     g_free(backing_fmt);
675     return ret;
676 }
677 
678 typedef struct {
679     BlockDriverState *bs;
680     Coroutine *co;
681     uint64_t pos;
682     int64_t status;
683     int *pnum;
684 } QEDIsAllocatedCB;
685 
686 static void qed_is_allocated_cb(void *opaque, int ret, uint64_t offset, size_t len)
687 {
688     QEDIsAllocatedCB *cb = opaque;
689     BDRVQEDState *s = cb->bs->opaque;
690     *cb->pnum = len / BDRV_SECTOR_SIZE;
691     switch (ret) {
692     case QED_CLUSTER_FOUND:
693         offset |= qed_offset_into_cluster(s, cb->pos);
694         cb->status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID | offset;
695         break;
696     case QED_CLUSTER_ZERO:
697         cb->status = BDRV_BLOCK_ZERO;
698         break;
699     case QED_CLUSTER_L2:
700     case QED_CLUSTER_L1:
701         cb->status = 0;
702         break;
703     default:
704         assert(ret < 0);
705         cb->status = ret;
706         break;
707     }
708 
709     if (cb->co) {
710         qemu_coroutine_enter(cb->co, NULL);
711     }
712 }
713 
714 static int64_t coroutine_fn bdrv_qed_co_get_block_status(BlockDriverState *bs,
715                                                  int64_t sector_num,
716                                                  int nb_sectors, int *pnum)
717 {
718     BDRVQEDState *s = bs->opaque;
719     size_t len = (size_t)nb_sectors * BDRV_SECTOR_SIZE;
720     QEDIsAllocatedCB cb = {
721         .bs = bs,
722         .pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE,
723         .status = BDRV_BLOCK_OFFSET_MASK,
724         .pnum = pnum,
725     };
726     QEDRequest request = { .l2_table = NULL };
727 
728     qed_find_cluster(s, &request, cb.pos, len, qed_is_allocated_cb, &cb);
729 
730     /* Now sleep if the callback wasn't invoked immediately */
731     while (cb.status == BDRV_BLOCK_OFFSET_MASK) {
732         cb.co = qemu_coroutine_self();
733         qemu_coroutine_yield();
734     }
735 
736     qed_unref_l2_cache_entry(request.l2_table);
737 
738     return cb.status;
739 }
740 
741 static BDRVQEDState *acb_to_s(QEDAIOCB *acb)
742 {
743     return acb->common.bs->opaque;
744 }
745 
746 /**
747  * Read from the backing file or zero-fill if no backing file
748  *
749  * @s:              QED state
750  * @pos:            Byte position in device
751  * @qiov:           Destination I/O vector
752  * @backing_qiov:   Possibly shortened copy of qiov, to be allocated here
753  * @cb:             Completion function
754  * @opaque:         User data for completion function
755  *
756  * This function reads qiov->size bytes starting at pos from the backing file.
757  * If there is no backing file then zeroes are read.
758  */
759 static void qed_read_backing_file(BDRVQEDState *s, uint64_t pos,
760                                   QEMUIOVector *qiov,
761                                   QEMUIOVector **backing_qiov,
762                                   BlockCompletionFunc *cb, void *opaque)
763 {
764     uint64_t backing_length = 0;
765     size_t size;
766 
767     /* If there is a backing file, get its length.  Treat the absence of a
768      * backing file like a zero length backing file.
769      */
770     if (s->bs->backing_hd) {
771         int64_t l = bdrv_getlength(s->bs->backing_hd);
772         if (l < 0) {
773             cb(opaque, l);
774             return;
775         }
776         backing_length = l;
777     }
778 
779     /* Zero all sectors if reading beyond the end of the backing file */
780     if (pos >= backing_length ||
781         pos + qiov->size > backing_length) {
782         qemu_iovec_memset(qiov, 0, 0, qiov->size);
783     }
784 
785     /* Complete now if there are no backing file sectors to read */
786     if (pos >= backing_length) {
787         cb(opaque, 0);
788         return;
789     }
790 
791     /* If the read straddles the end of the backing file, shorten it */
792     size = MIN((uint64_t)backing_length - pos, qiov->size);
793 
794     assert(*backing_qiov == NULL);
795     *backing_qiov = g_new(QEMUIOVector, 1);
796     qemu_iovec_init(*backing_qiov, qiov->niov);
797     qemu_iovec_concat(*backing_qiov, qiov, 0, size);
798 
799     BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO);
800     bdrv_aio_readv(s->bs->backing_hd, pos / BDRV_SECTOR_SIZE,
801                    *backing_qiov, size / BDRV_SECTOR_SIZE, cb, opaque);
802 }
803 
804 typedef struct {
805     GenericCB gencb;
806     BDRVQEDState *s;
807     QEMUIOVector qiov;
808     QEMUIOVector *backing_qiov;
809     struct iovec iov;
810     uint64_t offset;
811 } CopyFromBackingFileCB;
812 
813 static void qed_copy_from_backing_file_cb(void *opaque, int ret)
814 {
815     CopyFromBackingFileCB *copy_cb = opaque;
816     qemu_vfree(copy_cb->iov.iov_base);
817     gencb_complete(&copy_cb->gencb, ret);
818 }
819 
820 static void qed_copy_from_backing_file_write(void *opaque, int ret)
821 {
822     CopyFromBackingFileCB *copy_cb = opaque;
823     BDRVQEDState *s = copy_cb->s;
824 
825     if (copy_cb->backing_qiov) {
826         qemu_iovec_destroy(copy_cb->backing_qiov);
827         g_free(copy_cb->backing_qiov);
828         copy_cb->backing_qiov = NULL;
829     }
830 
831     if (ret) {
832         qed_copy_from_backing_file_cb(copy_cb, ret);
833         return;
834     }
835 
836     BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE);
837     bdrv_aio_writev(s->bs->file, copy_cb->offset / BDRV_SECTOR_SIZE,
838                     &copy_cb->qiov, copy_cb->qiov.size / BDRV_SECTOR_SIZE,
839                     qed_copy_from_backing_file_cb, copy_cb);
840 }
841 
842 /**
843  * Copy data from backing file into the image
844  *
845  * @s:          QED state
846  * @pos:        Byte position in device
847  * @len:        Number of bytes
848  * @offset:     Byte offset in image file
849  * @cb:         Completion function
850  * @opaque:     User data for completion function
851  */
852 static void qed_copy_from_backing_file(BDRVQEDState *s, uint64_t pos,
853                                        uint64_t len, uint64_t offset,
854                                        BlockCompletionFunc *cb,
855                                        void *opaque)
856 {
857     CopyFromBackingFileCB *copy_cb;
858 
859     /* Skip copy entirely if there is no work to do */
860     if (len == 0) {
861         cb(opaque, 0);
862         return;
863     }
864 
865     copy_cb = gencb_alloc(sizeof(*copy_cb), cb, opaque);
866     copy_cb->s = s;
867     copy_cb->offset = offset;
868     copy_cb->backing_qiov = NULL;
869     copy_cb->iov.iov_base = qemu_blockalign(s->bs, len);
870     copy_cb->iov.iov_len = len;
871     qemu_iovec_init_external(&copy_cb->qiov, &copy_cb->iov, 1);
872 
873     qed_read_backing_file(s, pos, &copy_cb->qiov, &copy_cb->backing_qiov,
874                           qed_copy_from_backing_file_write, copy_cb);
875 }
876 
877 /**
878  * Link one or more contiguous clusters into a table
879  *
880  * @s:              QED state
881  * @table:          L2 table
882  * @index:          First cluster index
883  * @n:              Number of contiguous clusters
884  * @cluster:        First cluster offset
885  *
886  * The cluster offset may be an allocated byte offset in the image file, the
887  * zero cluster marker, or the unallocated cluster marker.
888  */
889 static void qed_update_l2_table(BDRVQEDState *s, QEDTable *table, int index,
890                                 unsigned int n, uint64_t cluster)
891 {
892     int i;
893     for (i = index; i < index + n; i++) {
894         table->offsets[i] = cluster;
895         if (!qed_offset_is_unalloc_cluster(cluster) &&
896             !qed_offset_is_zero_cluster(cluster)) {
897             cluster += s->header.cluster_size;
898         }
899     }
900 }
901 
902 static void qed_aio_complete_bh(void *opaque)
903 {
904     QEDAIOCB *acb = opaque;
905     BlockCompletionFunc *cb = acb->common.cb;
906     void *user_opaque = acb->common.opaque;
907     int ret = acb->bh_ret;
908 
909     qemu_bh_delete(acb->bh);
910     qemu_aio_unref(acb);
911 
912     /* Invoke callback */
913     cb(user_opaque, ret);
914 }
915 
916 static void qed_aio_complete(QEDAIOCB *acb, int ret)
917 {
918     BDRVQEDState *s = acb_to_s(acb);
919 
920     trace_qed_aio_complete(s, acb, ret);
921 
922     /* Free resources */
923     qemu_iovec_destroy(&acb->cur_qiov);
924     qed_unref_l2_cache_entry(acb->request.l2_table);
925 
926     /* Free the buffer we may have allocated for zero writes */
927     if (acb->flags & QED_AIOCB_ZERO) {
928         qemu_vfree(acb->qiov->iov[0].iov_base);
929         acb->qiov->iov[0].iov_base = NULL;
930     }
931 
932     /* Arrange for a bh to invoke the completion function */
933     acb->bh_ret = ret;
934     acb->bh = aio_bh_new(bdrv_get_aio_context(acb->common.bs),
935                          qed_aio_complete_bh, acb);
936     qemu_bh_schedule(acb->bh);
937 
938     /* Start next allocating write request waiting behind this one.  Note that
939      * requests enqueue themselves when they first hit an unallocated cluster
940      * but they wait until the entire request is finished before waking up the
941      * next request in the queue.  This ensures that we don't cycle through
942      * requests multiple times but rather finish one at a time completely.
943      */
944     if (acb == QSIMPLEQ_FIRST(&s->allocating_write_reqs)) {
945         QSIMPLEQ_REMOVE_HEAD(&s->allocating_write_reqs, next);
946         acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs);
947         if (acb) {
948             qed_aio_next_io(acb, 0);
949         } else if (s->header.features & QED_F_NEED_CHECK) {
950             qed_start_need_check_timer(s);
951         }
952     }
953 }
954 
955 /**
956  * Commit the current L2 table to the cache
957  */
958 static void qed_commit_l2_update(void *opaque, int ret)
959 {
960     QEDAIOCB *acb = opaque;
961     BDRVQEDState *s = acb_to_s(acb);
962     CachedL2Table *l2_table = acb->request.l2_table;
963     uint64_t l2_offset = l2_table->offset;
964 
965     qed_commit_l2_cache_entry(&s->l2_cache, l2_table);
966 
967     /* This is guaranteed to succeed because we just committed the entry to the
968      * cache.
969      */
970     acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset);
971     assert(acb->request.l2_table != NULL);
972 
973     qed_aio_next_io(opaque, ret);
974 }
975 
976 /**
977  * Update L1 table with new L2 table offset and write it out
978  */
979 static void qed_aio_write_l1_update(void *opaque, int ret)
980 {
981     QEDAIOCB *acb = opaque;
982     BDRVQEDState *s = acb_to_s(acb);
983     int index;
984 
985     if (ret) {
986         qed_aio_complete(acb, ret);
987         return;
988     }
989 
990     index = qed_l1_index(s, acb->cur_pos);
991     s->l1_table->offsets[index] = acb->request.l2_table->offset;
992 
993     qed_write_l1_table(s, index, 1, qed_commit_l2_update, acb);
994 }
995 
996 /**
997  * Update L2 table with new cluster offsets and write them out
998  */
999 static void qed_aio_write_l2_update(QEDAIOCB *acb, int ret, uint64_t offset)
1000 {
1001     BDRVQEDState *s = acb_to_s(acb);
1002     bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1;
1003     int index;
1004 
1005     if (ret) {
1006         goto err;
1007     }
1008 
1009     if (need_alloc) {
1010         qed_unref_l2_cache_entry(acb->request.l2_table);
1011         acb->request.l2_table = qed_new_l2_table(s);
1012     }
1013 
1014     index = qed_l2_index(s, acb->cur_pos);
1015     qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters,
1016                          offset);
1017 
1018     if (need_alloc) {
1019         /* Write out the whole new L2 table */
1020         qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true,
1021                             qed_aio_write_l1_update, acb);
1022     } else {
1023         /* Write out only the updated part of the L2 table */
1024         qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters, false,
1025                             qed_aio_next_io, acb);
1026     }
1027     return;
1028 
1029 err:
1030     qed_aio_complete(acb, ret);
1031 }
1032 
1033 static void qed_aio_write_l2_update_cb(void *opaque, int ret)
1034 {
1035     QEDAIOCB *acb = opaque;
1036     qed_aio_write_l2_update(acb, ret, acb->cur_cluster);
1037 }
1038 
1039 /**
1040  * Flush new data clusters before updating the L2 table
1041  *
1042  * This flush is necessary when a backing file is in use.  A crash during an
1043  * allocating write could result in empty clusters in the image.  If the write
1044  * only touched a subregion of the cluster, then backing image sectors have
1045  * been lost in the untouched region.  The solution is to flush after writing a
1046  * new data cluster and before updating the L2 table.
1047  */
1048 static void qed_aio_write_flush_before_l2_update(void *opaque, int ret)
1049 {
1050     QEDAIOCB *acb = opaque;
1051     BDRVQEDState *s = acb_to_s(acb);
1052 
1053     if (!bdrv_aio_flush(s->bs->file, qed_aio_write_l2_update_cb, opaque)) {
1054         qed_aio_complete(acb, -EIO);
1055     }
1056 }
1057 
1058 /**
1059  * Write data to the image file
1060  */
1061 static void qed_aio_write_main(void *opaque, int ret)
1062 {
1063     QEDAIOCB *acb = opaque;
1064     BDRVQEDState *s = acb_to_s(acb);
1065     uint64_t offset = acb->cur_cluster +
1066                       qed_offset_into_cluster(s, acb->cur_pos);
1067     BlockCompletionFunc *next_fn;
1068 
1069     trace_qed_aio_write_main(s, acb, ret, offset, acb->cur_qiov.size);
1070 
1071     if (ret) {
1072         qed_aio_complete(acb, ret);
1073         return;
1074     }
1075 
1076     if (acb->find_cluster_ret == QED_CLUSTER_FOUND) {
1077         next_fn = qed_aio_next_io;
1078     } else {
1079         if (s->bs->backing_hd) {
1080             next_fn = qed_aio_write_flush_before_l2_update;
1081         } else {
1082             next_fn = qed_aio_write_l2_update_cb;
1083         }
1084     }
1085 
1086     BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO);
1087     bdrv_aio_writev(s->bs->file, offset / BDRV_SECTOR_SIZE,
1088                     &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE,
1089                     next_fn, acb);
1090 }
1091 
1092 /**
1093  * Populate back untouched region of new data cluster
1094  */
1095 static void qed_aio_write_postfill(void *opaque, int ret)
1096 {
1097     QEDAIOCB *acb = opaque;
1098     BDRVQEDState *s = acb_to_s(acb);
1099     uint64_t start = acb->cur_pos + acb->cur_qiov.size;
1100     uint64_t len =
1101         qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start;
1102     uint64_t offset = acb->cur_cluster +
1103                       qed_offset_into_cluster(s, acb->cur_pos) +
1104                       acb->cur_qiov.size;
1105 
1106     if (ret) {
1107         qed_aio_complete(acb, ret);
1108         return;
1109     }
1110 
1111     trace_qed_aio_write_postfill(s, acb, start, len, offset);
1112     qed_copy_from_backing_file(s, start, len, offset,
1113                                 qed_aio_write_main, acb);
1114 }
1115 
1116 /**
1117  * Populate front untouched region of new data cluster
1118  */
1119 static void qed_aio_write_prefill(void *opaque, int ret)
1120 {
1121     QEDAIOCB *acb = opaque;
1122     BDRVQEDState *s = acb_to_s(acb);
1123     uint64_t start = qed_start_of_cluster(s, acb->cur_pos);
1124     uint64_t len = qed_offset_into_cluster(s, acb->cur_pos);
1125 
1126     trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster);
1127     qed_copy_from_backing_file(s, start, len, acb->cur_cluster,
1128                                 qed_aio_write_postfill, acb);
1129 }
1130 
1131 /**
1132  * Check if the QED_F_NEED_CHECK bit should be set during allocating write
1133  */
1134 static bool qed_should_set_need_check(BDRVQEDState *s)
1135 {
1136     /* The flush before L2 update path ensures consistency */
1137     if (s->bs->backing_hd) {
1138         return false;
1139     }
1140 
1141     return !(s->header.features & QED_F_NEED_CHECK);
1142 }
1143 
1144 static void qed_aio_write_zero_cluster(void *opaque, int ret)
1145 {
1146     QEDAIOCB *acb = opaque;
1147 
1148     if (ret) {
1149         qed_aio_complete(acb, ret);
1150         return;
1151     }
1152 
1153     qed_aio_write_l2_update(acb, 0, 1);
1154 }
1155 
1156 /**
1157  * Write new data cluster
1158  *
1159  * @acb:        Write request
1160  * @len:        Length in bytes
1161  *
1162  * This path is taken when writing to previously unallocated clusters.
1163  */
1164 static void qed_aio_write_alloc(QEDAIOCB *acb, size_t len)
1165 {
1166     BDRVQEDState *s = acb_to_s(acb);
1167     BlockCompletionFunc *cb;
1168 
1169     /* Cancel timer when the first allocating request comes in */
1170     if (QSIMPLEQ_EMPTY(&s->allocating_write_reqs)) {
1171         qed_cancel_need_check_timer(s);
1172     }
1173 
1174     /* Freeze this request if another allocating write is in progress */
1175     if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs)) {
1176         QSIMPLEQ_INSERT_TAIL(&s->allocating_write_reqs, acb, next);
1177     }
1178     if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs) ||
1179         s->allocating_write_reqs_plugged) {
1180         return; /* wait for existing request to finish */
1181     }
1182 
1183     acb->cur_nclusters = qed_bytes_to_clusters(s,
1184             qed_offset_into_cluster(s, acb->cur_pos) + len);
1185     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1186 
1187     if (acb->flags & QED_AIOCB_ZERO) {
1188         /* Skip ahead if the clusters are already zero */
1189         if (acb->find_cluster_ret == QED_CLUSTER_ZERO) {
1190             qed_aio_next_io(acb, 0);
1191             return;
1192         }
1193 
1194         cb = qed_aio_write_zero_cluster;
1195     } else {
1196         cb = qed_aio_write_prefill;
1197         acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters);
1198     }
1199 
1200     if (qed_should_set_need_check(s)) {
1201         s->header.features |= QED_F_NEED_CHECK;
1202         qed_write_header(s, cb, acb);
1203     } else {
1204         cb(acb, 0);
1205     }
1206 }
1207 
1208 /**
1209  * Write data cluster in place
1210  *
1211  * @acb:        Write request
1212  * @offset:     Cluster offset in bytes
1213  * @len:        Length in bytes
1214  *
1215  * This path is taken when writing to already allocated clusters.
1216  */
1217 static void qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset, size_t len)
1218 {
1219     /* Allocate buffer for zero writes */
1220     if (acb->flags & QED_AIOCB_ZERO) {
1221         struct iovec *iov = acb->qiov->iov;
1222 
1223         if (!iov->iov_base) {
1224             iov->iov_base = qemu_try_blockalign(acb->common.bs, iov->iov_len);
1225             if (iov->iov_base == NULL) {
1226                 qed_aio_complete(acb, -ENOMEM);
1227                 return;
1228             }
1229             memset(iov->iov_base, 0, iov->iov_len);
1230         }
1231     }
1232 
1233     /* Calculate the I/O vector */
1234     acb->cur_cluster = offset;
1235     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1236 
1237     /* Do the actual write */
1238     qed_aio_write_main(acb, 0);
1239 }
1240 
1241 /**
1242  * Write data cluster
1243  *
1244  * @opaque:     Write request
1245  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1,
1246  *              or -errno
1247  * @offset:     Cluster offset in bytes
1248  * @len:        Length in bytes
1249  *
1250  * Callback from qed_find_cluster().
1251  */
1252 static void qed_aio_write_data(void *opaque, int ret,
1253                                uint64_t offset, size_t len)
1254 {
1255     QEDAIOCB *acb = opaque;
1256 
1257     trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len);
1258 
1259     acb->find_cluster_ret = ret;
1260 
1261     switch (ret) {
1262     case QED_CLUSTER_FOUND:
1263         qed_aio_write_inplace(acb, offset, len);
1264         break;
1265 
1266     case QED_CLUSTER_L2:
1267     case QED_CLUSTER_L1:
1268     case QED_CLUSTER_ZERO:
1269         qed_aio_write_alloc(acb, len);
1270         break;
1271 
1272     default:
1273         qed_aio_complete(acb, ret);
1274         break;
1275     }
1276 }
1277 
1278 /**
1279  * Read data cluster
1280  *
1281  * @opaque:     Read request
1282  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1,
1283  *              or -errno
1284  * @offset:     Cluster offset in bytes
1285  * @len:        Length in bytes
1286  *
1287  * Callback from qed_find_cluster().
1288  */
1289 static void qed_aio_read_data(void *opaque, int ret,
1290                               uint64_t offset, size_t len)
1291 {
1292     QEDAIOCB *acb = opaque;
1293     BDRVQEDState *s = acb_to_s(acb);
1294     BlockDriverState *bs = acb->common.bs;
1295 
1296     /* Adjust offset into cluster */
1297     offset += qed_offset_into_cluster(s, acb->cur_pos);
1298 
1299     trace_qed_aio_read_data(s, acb, ret, offset, len);
1300 
1301     if (ret < 0) {
1302         goto err;
1303     }
1304 
1305     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1306 
1307     /* Handle zero cluster and backing file reads */
1308     if (ret == QED_CLUSTER_ZERO) {
1309         qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size);
1310         qed_aio_next_io(acb, 0);
1311         return;
1312     } else if (ret != QED_CLUSTER_FOUND) {
1313         qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov,
1314                               &acb->backing_qiov, qed_aio_next_io, acb);
1315         return;
1316     }
1317 
1318     BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
1319     bdrv_aio_readv(bs->file, offset / BDRV_SECTOR_SIZE,
1320                    &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE,
1321                    qed_aio_next_io, acb);
1322     return;
1323 
1324 err:
1325     qed_aio_complete(acb, ret);
1326 }
1327 
1328 /**
1329  * Begin next I/O or complete the request
1330  */
1331 static void qed_aio_next_io(void *opaque, int ret)
1332 {
1333     QEDAIOCB *acb = opaque;
1334     BDRVQEDState *s = acb_to_s(acb);
1335     QEDFindClusterFunc *io_fn = (acb->flags & QED_AIOCB_WRITE) ?
1336                                 qed_aio_write_data : qed_aio_read_data;
1337 
1338     trace_qed_aio_next_io(s, acb, ret, acb->cur_pos + acb->cur_qiov.size);
1339 
1340     if (acb->backing_qiov) {
1341         qemu_iovec_destroy(acb->backing_qiov);
1342         g_free(acb->backing_qiov);
1343         acb->backing_qiov = NULL;
1344     }
1345 
1346     /* Handle I/O error */
1347     if (ret) {
1348         qed_aio_complete(acb, ret);
1349         return;
1350     }
1351 
1352     acb->qiov_offset += acb->cur_qiov.size;
1353     acb->cur_pos += acb->cur_qiov.size;
1354     qemu_iovec_reset(&acb->cur_qiov);
1355 
1356     /* Complete request */
1357     if (acb->cur_pos >= acb->end_pos) {
1358         qed_aio_complete(acb, 0);
1359         return;
1360     }
1361 
1362     /* Find next cluster and start I/O */
1363     qed_find_cluster(s, &acb->request,
1364                       acb->cur_pos, acb->end_pos - acb->cur_pos,
1365                       io_fn, acb);
1366 }
1367 
1368 static BlockAIOCB *qed_aio_setup(BlockDriverState *bs,
1369                                  int64_t sector_num,
1370                                  QEMUIOVector *qiov, int nb_sectors,
1371                                  BlockCompletionFunc *cb,
1372                                  void *opaque, int flags)
1373 {
1374     QEDAIOCB *acb = qemu_aio_get(&qed_aiocb_info, bs, cb, opaque);
1375 
1376     trace_qed_aio_setup(bs->opaque, acb, sector_num, nb_sectors,
1377                         opaque, flags);
1378 
1379     acb->flags = flags;
1380     acb->qiov = qiov;
1381     acb->qiov_offset = 0;
1382     acb->cur_pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE;
1383     acb->end_pos = acb->cur_pos + nb_sectors * BDRV_SECTOR_SIZE;
1384     acb->backing_qiov = NULL;
1385     acb->request.l2_table = NULL;
1386     qemu_iovec_init(&acb->cur_qiov, qiov->niov);
1387 
1388     /* Start request */
1389     qed_aio_next_io(acb, 0);
1390     return &acb->common;
1391 }
1392 
1393 static BlockAIOCB *bdrv_qed_aio_readv(BlockDriverState *bs,
1394                                       int64_t sector_num,
1395                                       QEMUIOVector *qiov, int nb_sectors,
1396                                       BlockCompletionFunc *cb,
1397                                       void *opaque)
1398 {
1399     return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb, opaque, 0);
1400 }
1401 
1402 static BlockAIOCB *bdrv_qed_aio_writev(BlockDriverState *bs,
1403                                        int64_t sector_num,
1404                                        QEMUIOVector *qiov, int nb_sectors,
1405                                        BlockCompletionFunc *cb,
1406                                        void *opaque)
1407 {
1408     return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb,
1409                          opaque, QED_AIOCB_WRITE);
1410 }
1411 
1412 typedef struct {
1413     Coroutine *co;
1414     int ret;
1415     bool done;
1416 } QEDWriteZeroesCB;
1417 
1418 static void coroutine_fn qed_co_write_zeroes_cb(void *opaque, int ret)
1419 {
1420     QEDWriteZeroesCB *cb = opaque;
1421 
1422     cb->done = true;
1423     cb->ret = ret;
1424     if (cb->co) {
1425         qemu_coroutine_enter(cb->co, NULL);
1426     }
1427 }
1428 
1429 static int coroutine_fn bdrv_qed_co_write_zeroes(BlockDriverState *bs,
1430                                                  int64_t sector_num,
1431                                                  int nb_sectors,
1432                                                  BdrvRequestFlags flags)
1433 {
1434     BlockAIOCB *blockacb;
1435     BDRVQEDState *s = bs->opaque;
1436     QEDWriteZeroesCB cb = { .done = false };
1437     QEMUIOVector qiov;
1438     struct iovec iov;
1439 
1440     /* Refuse if there are untouched backing file sectors */
1441     if (bs->backing_hd) {
1442         if (qed_offset_into_cluster(s, sector_num * BDRV_SECTOR_SIZE) != 0) {
1443             return -ENOTSUP;
1444         }
1445         if (qed_offset_into_cluster(s, nb_sectors * BDRV_SECTOR_SIZE) != 0) {
1446             return -ENOTSUP;
1447         }
1448     }
1449 
1450     /* Zero writes start without an I/O buffer.  If a buffer becomes necessary
1451      * then it will be allocated during request processing.
1452      */
1453     iov.iov_base = NULL,
1454     iov.iov_len  = nb_sectors * BDRV_SECTOR_SIZE,
1455 
1456     qemu_iovec_init_external(&qiov, &iov, 1);
1457     blockacb = qed_aio_setup(bs, sector_num, &qiov, nb_sectors,
1458                              qed_co_write_zeroes_cb, &cb,
1459                              QED_AIOCB_WRITE | QED_AIOCB_ZERO);
1460     if (!blockacb) {
1461         return -EIO;
1462     }
1463     if (!cb.done) {
1464         cb.co = qemu_coroutine_self();
1465         qemu_coroutine_yield();
1466     }
1467     assert(cb.done);
1468     return cb.ret;
1469 }
1470 
1471 static int bdrv_qed_truncate(BlockDriverState *bs, int64_t offset)
1472 {
1473     BDRVQEDState *s = bs->opaque;
1474     uint64_t old_image_size;
1475     int ret;
1476 
1477     if (!qed_is_image_size_valid(offset, s->header.cluster_size,
1478                                  s->header.table_size)) {
1479         return -EINVAL;
1480     }
1481 
1482     /* Shrinking is currently not supported */
1483     if ((uint64_t)offset < s->header.image_size) {
1484         return -ENOTSUP;
1485     }
1486 
1487     old_image_size = s->header.image_size;
1488     s->header.image_size = offset;
1489     ret = qed_write_header_sync(s);
1490     if (ret < 0) {
1491         s->header.image_size = old_image_size;
1492     }
1493     return ret;
1494 }
1495 
1496 static int64_t bdrv_qed_getlength(BlockDriverState *bs)
1497 {
1498     BDRVQEDState *s = bs->opaque;
1499     return s->header.image_size;
1500 }
1501 
1502 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1503 {
1504     BDRVQEDState *s = bs->opaque;
1505 
1506     memset(bdi, 0, sizeof(*bdi));
1507     bdi->cluster_size = s->header.cluster_size;
1508     bdi->is_dirty = s->header.features & QED_F_NEED_CHECK;
1509     bdi->unallocated_blocks_are_zero = true;
1510     bdi->can_write_zeroes_with_unmap = true;
1511     return 0;
1512 }
1513 
1514 static int bdrv_qed_change_backing_file(BlockDriverState *bs,
1515                                         const char *backing_file,
1516                                         const char *backing_fmt)
1517 {
1518     BDRVQEDState *s = bs->opaque;
1519     QEDHeader new_header, le_header;
1520     void *buffer;
1521     size_t buffer_len, backing_file_len;
1522     int ret;
1523 
1524     /* Refuse to set backing filename if unknown compat feature bits are
1525      * active.  If the image uses an unknown compat feature then we may not
1526      * know the layout of data following the header structure and cannot safely
1527      * add a new string.
1528      */
1529     if (backing_file && (s->header.compat_features &
1530                          ~QED_COMPAT_FEATURE_MASK)) {
1531         return -ENOTSUP;
1532     }
1533 
1534     memcpy(&new_header, &s->header, sizeof(new_header));
1535 
1536     new_header.features &= ~(QED_F_BACKING_FILE |
1537                              QED_F_BACKING_FORMAT_NO_PROBE);
1538 
1539     /* Adjust feature flags */
1540     if (backing_file) {
1541         new_header.features |= QED_F_BACKING_FILE;
1542 
1543         if (qed_fmt_is_raw(backing_fmt)) {
1544             new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
1545         }
1546     }
1547 
1548     /* Calculate new header size */
1549     backing_file_len = 0;
1550 
1551     if (backing_file) {
1552         backing_file_len = strlen(backing_file);
1553     }
1554 
1555     buffer_len = sizeof(new_header);
1556     new_header.backing_filename_offset = buffer_len;
1557     new_header.backing_filename_size = backing_file_len;
1558     buffer_len += backing_file_len;
1559 
1560     /* Make sure we can rewrite header without failing */
1561     if (buffer_len > new_header.header_size * new_header.cluster_size) {
1562         return -ENOSPC;
1563     }
1564 
1565     /* Prepare new header */
1566     buffer = g_malloc(buffer_len);
1567 
1568     qed_header_cpu_to_le(&new_header, &le_header);
1569     memcpy(buffer, &le_header, sizeof(le_header));
1570     buffer_len = sizeof(le_header);
1571 
1572     if (backing_file) {
1573         memcpy(buffer + buffer_len, backing_file, backing_file_len);
1574         buffer_len += backing_file_len;
1575     }
1576 
1577     /* Write new header */
1578     ret = bdrv_pwrite_sync(bs->file, 0, buffer, buffer_len);
1579     g_free(buffer);
1580     if (ret == 0) {
1581         memcpy(&s->header, &new_header, sizeof(new_header));
1582     }
1583     return ret;
1584 }
1585 
1586 static void bdrv_qed_invalidate_cache(BlockDriverState *bs, Error **errp)
1587 {
1588     BDRVQEDState *s = bs->opaque;
1589     Error *local_err = NULL;
1590     int ret;
1591 
1592     bdrv_qed_close(bs);
1593 
1594     bdrv_invalidate_cache(bs->file, &local_err);
1595     if (local_err) {
1596         error_propagate(errp, local_err);
1597         return;
1598     }
1599 
1600     memset(s, 0, sizeof(BDRVQEDState));
1601     ret = bdrv_qed_open(bs, NULL, bs->open_flags, &local_err);
1602     if (local_err) {
1603         error_setg(errp, "Could not reopen qed layer: %s",
1604                    error_get_pretty(local_err));
1605         error_free(local_err);
1606         return;
1607     } else if (ret < 0) {
1608         error_setg_errno(errp, -ret, "Could not reopen qed layer");
1609         return;
1610     }
1611 }
1612 
1613 static int bdrv_qed_check(BlockDriverState *bs, BdrvCheckResult *result,
1614                           BdrvCheckMode fix)
1615 {
1616     BDRVQEDState *s = bs->opaque;
1617 
1618     return qed_check(s, result, !!fix);
1619 }
1620 
1621 static QemuOptsList qed_create_opts = {
1622     .name = "qed-create-opts",
1623     .head = QTAILQ_HEAD_INITIALIZER(qed_create_opts.head),
1624     .desc = {
1625         {
1626             .name = BLOCK_OPT_SIZE,
1627             .type = QEMU_OPT_SIZE,
1628             .help = "Virtual disk size"
1629         },
1630         {
1631             .name = BLOCK_OPT_BACKING_FILE,
1632             .type = QEMU_OPT_STRING,
1633             .help = "File name of a base image"
1634         },
1635         {
1636             .name = BLOCK_OPT_BACKING_FMT,
1637             .type = QEMU_OPT_STRING,
1638             .help = "Image format of the base image"
1639         },
1640         {
1641             .name = BLOCK_OPT_CLUSTER_SIZE,
1642             .type = QEMU_OPT_SIZE,
1643             .help = "Cluster size (in bytes)",
1644             .def_value_str = stringify(QED_DEFAULT_CLUSTER_SIZE)
1645         },
1646         {
1647             .name = BLOCK_OPT_TABLE_SIZE,
1648             .type = QEMU_OPT_SIZE,
1649             .help = "L1/L2 table size (in clusters)"
1650         },
1651         { /* end of list */ }
1652     }
1653 };
1654 
1655 static BlockDriver bdrv_qed = {
1656     .format_name              = "qed",
1657     .instance_size            = sizeof(BDRVQEDState),
1658     .create_opts              = &qed_create_opts,
1659     .supports_backing         = true,
1660 
1661     .bdrv_probe               = bdrv_qed_probe,
1662     .bdrv_rebind              = bdrv_qed_rebind,
1663     .bdrv_open                = bdrv_qed_open,
1664     .bdrv_close               = bdrv_qed_close,
1665     .bdrv_reopen_prepare      = bdrv_qed_reopen_prepare,
1666     .bdrv_create              = bdrv_qed_create,
1667     .bdrv_has_zero_init       = bdrv_has_zero_init_1,
1668     .bdrv_co_get_block_status = bdrv_qed_co_get_block_status,
1669     .bdrv_aio_readv           = bdrv_qed_aio_readv,
1670     .bdrv_aio_writev          = bdrv_qed_aio_writev,
1671     .bdrv_co_write_zeroes     = bdrv_qed_co_write_zeroes,
1672     .bdrv_truncate            = bdrv_qed_truncate,
1673     .bdrv_getlength           = bdrv_qed_getlength,
1674     .bdrv_get_info            = bdrv_qed_get_info,
1675     .bdrv_refresh_limits      = bdrv_qed_refresh_limits,
1676     .bdrv_change_backing_file = bdrv_qed_change_backing_file,
1677     .bdrv_invalidate_cache    = bdrv_qed_invalidate_cache,
1678     .bdrv_check               = bdrv_qed_check,
1679     .bdrv_detach_aio_context  = bdrv_qed_detach_aio_context,
1680     .bdrv_attach_aio_context  = bdrv_qed_attach_aio_context,
1681 };
1682 
1683 static void bdrv_qed_init(void)
1684 {
1685     bdrv_register(&bdrv_qed);
1686 }
1687 
1688 block_init(bdrv_qed_init);
1689