xref: /openbmc/qemu/block/qed.c (revision 4b4629d9)
1 /*
2  * QEMU Enhanced Disk Format
3  *
4  * Copyright IBM, Corp. 2010
5  *
6  * Authors:
7  *  Stefan Hajnoczi   <stefanha@linux.vnet.ibm.com>
8  *  Anthony Liguori   <aliguori@us.ibm.com>
9  *
10  * This work is licensed under the terms of the GNU LGPL, version 2 or later.
11  * See the COPYING.LIB file in the top-level directory.
12  *
13  */
14 
15 #include "qemu/osdep.h"
16 #include "qapi/error.h"
17 #include "qemu/timer.h"
18 #include "trace.h"
19 #include "qed.h"
20 #include "qapi/qmp/qerror.h"
21 #include "migration/migration.h"
22 #include "sysemu/block-backend.h"
23 
24 static const AIOCBInfo qed_aiocb_info = {
25     .aiocb_size         = sizeof(QEDAIOCB),
26 };
27 
28 static int bdrv_qed_probe(const uint8_t *buf, int buf_size,
29                           const char *filename)
30 {
31     const QEDHeader *header = (const QEDHeader *)buf;
32 
33     if (buf_size < sizeof(*header)) {
34         return 0;
35     }
36     if (le32_to_cpu(header->magic) != QED_MAGIC) {
37         return 0;
38     }
39     return 100;
40 }
41 
42 /**
43  * Check whether an image format is raw
44  *
45  * @fmt:    Backing file format, may be NULL
46  */
47 static bool qed_fmt_is_raw(const char *fmt)
48 {
49     return fmt && strcmp(fmt, "raw") == 0;
50 }
51 
52 static void qed_header_le_to_cpu(const QEDHeader *le, QEDHeader *cpu)
53 {
54     cpu->magic = le32_to_cpu(le->magic);
55     cpu->cluster_size = le32_to_cpu(le->cluster_size);
56     cpu->table_size = le32_to_cpu(le->table_size);
57     cpu->header_size = le32_to_cpu(le->header_size);
58     cpu->features = le64_to_cpu(le->features);
59     cpu->compat_features = le64_to_cpu(le->compat_features);
60     cpu->autoclear_features = le64_to_cpu(le->autoclear_features);
61     cpu->l1_table_offset = le64_to_cpu(le->l1_table_offset);
62     cpu->image_size = le64_to_cpu(le->image_size);
63     cpu->backing_filename_offset = le32_to_cpu(le->backing_filename_offset);
64     cpu->backing_filename_size = le32_to_cpu(le->backing_filename_size);
65 }
66 
67 static void qed_header_cpu_to_le(const QEDHeader *cpu, QEDHeader *le)
68 {
69     le->magic = cpu_to_le32(cpu->magic);
70     le->cluster_size = cpu_to_le32(cpu->cluster_size);
71     le->table_size = cpu_to_le32(cpu->table_size);
72     le->header_size = cpu_to_le32(cpu->header_size);
73     le->features = cpu_to_le64(cpu->features);
74     le->compat_features = cpu_to_le64(cpu->compat_features);
75     le->autoclear_features = cpu_to_le64(cpu->autoclear_features);
76     le->l1_table_offset = cpu_to_le64(cpu->l1_table_offset);
77     le->image_size = cpu_to_le64(cpu->image_size);
78     le->backing_filename_offset = cpu_to_le32(cpu->backing_filename_offset);
79     le->backing_filename_size = cpu_to_le32(cpu->backing_filename_size);
80 }
81 
82 int qed_write_header_sync(BDRVQEDState *s)
83 {
84     QEDHeader le;
85     int ret;
86 
87     qed_header_cpu_to_le(&s->header, &le);
88     ret = bdrv_pwrite(s->bs->file->bs, 0, &le, sizeof(le));
89     if (ret != sizeof(le)) {
90         return ret;
91     }
92     return 0;
93 }
94 
95 typedef struct {
96     GenericCB gencb;
97     BDRVQEDState *s;
98     struct iovec iov;
99     QEMUIOVector qiov;
100     int nsectors;
101     uint8_t *buf;
102 } QEDWriteHeaderCB;
103 
104 static void qed_write_header_cb(void *opaque, int ret)
105 {
106     QEDWriteHeaderCB *write_header_cb = opaque;
107 
108     qemu_vfree(write_header_cb->buf);
109     gencb_complete(write_header_cb, ret);
110 }
111 
112 static void qed_write_header_read_cb(void *opaque, int ret)
113 {
114     QEDWriteHeaderCB *write_header_cb = opaque;
115     BDRVQEDState *s = write_header_cb->s;
116 
117     if (ret) {
118         qed_write_header_cb(write_header_cb, ret);
119         return;
120     }
121 
122     /* Update header */
123     qed_header_cpu_to_le(&s->header, (QEDHeader *)write_header_cb->buf);
124 
125     bdrv_aio_writev(s->bs->file->bs, 0, &write_header_cb->qiov,
126                     write_header_cb->nsectors, qed_write_header_cb,
127                     write_header_cb);
128 }
129 
130 /**
131  * Update header in-place (does not rewrite backing filename or other strings)
132  *
133  * This function only updates known header fields in-place and does not affect
134  * extra data after the QED header.
135  */
136 static void qed_write_header(BDRVQEDState *s, BlockCompletionFunc cb,
137                              void *opaque)
138 {
139     /* We must write full sectors for O_DIRECT but cannot necessarily generate
140      * the data following the header if an unrecognized compat feature is
141      * active.  Therefore, first read the sectors containing the header, update
142      * them, and write back.
143      */
144 
145     int nsectors = (sizeof(QEDHeader) + BDRV_SECTOR_SIZE - 1) /
146                    BDRV_SECTOR_SIZE;
147     size_t len = nsectors * BDRV_SECTOR_SIZE;
148     QEDWriteHeaderCB *write_header_cb = gencb_alloc(sizeof(*write_header_cb),
149                                                     cb, opaque);
150 
151     write_header_cb->s = s;
152     write_header_cb->nsectors = nsectors;
153     write_header_cb->buf = qemu_blockalign(s->bs, len);
154     write_header_cb->iov.iov_base = write_header_cb->buf;
155     write_header_cb->iov.iov_len = len;
156     qemu_iovec_init_external(&write_header_cb->qiov, &write_header_cb->iov, 1);
157 
158     bdrv_aio_readv(s->bs->file->bs, 0, &write_header_cb->qiov, nsectors,
159                    qed_write_header_read_cb, write_header_cb);
160 }
161 
162 static uint64_t qed_max_image_size(uint32_t cluster_size, uint32_t table_size)
163 {
164     uint64_t table_entries;
165     uint64_t l2_size;
166 
167     table_entries = (table_size * cluster_size) / sizeof(uint64_t);
168     l2_size = table_entries * cluster_size;
169 
170     return l2_size * table_entries;
171 }
172 
173 static bool qed_is_cluster_size_valid(uint32_t cluster_size)
174 {
175     if (cluster_size < QED_MIN_CLUSTER_SIZE ||
176         cluster_size > QED_MAX_CLUSTER_SIZE) {
177         return false;
178     }
179     if (cluster_size & (cluster_size - 1)) {
180         return false; /* not power of 2 */
181     }
182     return true;
183 }
184 
185 static bool qed_is_table_size_valid(uint32_t table_size)
186 {
187     if (table_size < QED_MIN_TABLE_SIZE ||
188         table_size > QED_MAX_TABLE_SIZE) {
189         return false;
190     }
191     if (table_size & (table_size - 1)) {
192         return false; /* not power of 2 */
193     }
194     return true;
195 }
196 
197 static bool qed_is_image_size_valid(uint64_t image_size, uint32_t cluster_size,
198                                     uint32_t table_size)
199 {
200     if (image_size % BDRV_SECTOR_SIZE != 0) {
201         return false; /* not multiple of sector size */
202     }
203     if (image_size > qed_max_image_size(cluster_size, table_size)) {
204         return false; /* image is too large */
205     }
206     return true;
207 }
208 
209 /**
210  * Read a string of known length from the image file
211  *
212  * @file:       Image file
213  * @offset:     File offset to start of string, in bytes
214  * @n:          String length in bytes
215  * @buf:        Destination buffer
216  * @buflen:     Destination buffer length in bytes
217  * @ret:        0 on success, -errno on failure
218  *
219  * The string is NUL-terminated.
220  */
221 static int qed_read_string(BlockDriverState *file, uint64_t offset, size_t n,
222                            char *buf, size_t buflen)
223 {
224     int ret;
225     if (n >= buflen) {
226         return -EINVAL;
227     }
228     ret = bdrv_pread(file, offset, buf, n);
229     if (ret < 0) {
230         return ret;
231     }
232     buf[n] = '\0';
233     return 0;
234 }
235 
236 /**
237  * Allocate new clusters
238  *
239  * @s:          QED state
240  * @n:          Number of contiguous clusters to allocate
241  * @ret:        Offset of first allocated cluster
242  *
243  * This function only produces the offset where the new clusters should be
244  * written.  It updates BDRVQEDState but does not make any changes to the image
245  * file.
246  */
247 static uint64_t qed_alloc_clusters(BDRVQEDState *s, unsigned int n)
248 {
249     uint64_t offset = s->file_size;
250     s->file_size += n * s->header.cluster_size;
251     return offset;
252 }
253 
254 QEDTable *qed_alloc_table(BDRVQEDState *s)
255 {
256     /* Honor O_DIRECT memory alignment requirements */
257     return qemu_blockalign(s->bs,
258                            s->header.cluster_size * s->header.table_size);
259 }
260 
261 /**
262  * Allocate a new zeroed L2 table
263  */
264 static CachedL2Table *qed_new_l2_table(BDRVQEDState *s)
265 {
266     CachedL2Table *l2_table = qed_alloc_l2_cache_entry(&s->l2_cache);
267 
268     l2_table->table = qed_alloc_table(s);
269     l2_table->offset = qed_alloc_clusters(s, s->header.table_size);
270 
271     memset(l2_table->table->offsets, 0,
272            s->header.cluster_size * s->header.table_size);
273     return l2_table;
274 }
275 
276 static void qed_aio_next_io(void *opaque, int ret);
277 
278 static void qed_plug_allocating_write_reqs(BDRVQEDState *s)
279 {
280     assert(!s->allocating_write_reqs_plugged);
281 
282     s->allocating_write_reqs_plugged = true;
283 }
284 
285 static void qed_unplug_allocating_write_reqs(BDRVQEDState *s)
286 {
287     QEDAIOCB *acb;
288 
289     assert(s->allocating_write_reqs_plugged);
290 
291     s->allocating_write_reqs_plugged = false;
292 
293     acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs);
294     if (acb) {
295         qed_aio_next_io(acb, 0);
296     }
297 }
298 
299 static void qed_finish_clear_need_check(void *opaque, int ret)
300 {
301     /* Do nothing */
302 }
303 
304 static void qed_flush_after_clear_need_check(void *opaque, int ret)
305 {
306     BDRVQEDState *s = opaque;
307 
308     bdrv_aio_flush(s->bs, qed_finish_clear_need_check, s);
309 
310     /* No need to wait until flush completes */
311     qed_unplug_allocating_write_reqs(s);
312 }
313 
314 static void qed_clear_need_check(void *opaque, int ret)
315 {
316     BDRVQEDState *s = opaque;
317 
318     if (ret) {
319         qed_unplug_allocating_write_reqs(s);
320         return;
321     }
322 
323     s->header.features &= ~QED_F_NEED_CHECK;
324     qed_write_header(s, qed_flush_after_clear_need_check, s);
325 }
326 
327 static void qed_need_check_timer_cb(void *opaque)
328 {
329     BDRVQEDState *s = opaque;
330 
331     /* The timer should only fire when allocating writes have drained */
332     assert(!QSIMPLEQ_FIRST(&s->allocating_write_reqs));
333 
334     trace_qed_need_check_timer_cb(s);
335 
336     qed_plug_allocating_write_reqs(s);
337 
338     /* Ensure writes are on disk before clearing flag */
339     bdrv_aio_flush(s->bs, qed_clear_need_check, s);
340 }
341 
342 static void qed_start_need_check_timer(BDRVQEDState *s)
343 {
344     trace_qed_start_need_check_timer(s);
345 
346     /* Use QEMU_CLOCK_VIRTUAL so we don't alter the image file while suspended for
347      * migration.
348      */
349     timer_mod(s->need_check_timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
350                    NANOSECONDS_PER_SECOND * QED_NEED_CHECK_TIMEOUT);
351 }
352 
353 /* It's okay to call this multiple times or when no timer is started */
354 static void qed_cancel_need_check_timer(BDRVQEDState *s)
355 {
356     trace_qed_cancel_need_check_timer(s);
357     timer_del(s->need_check_timer);
358 }
359 
360 static void bdrv_qed_detach_aio_context(BlockDriverState *bs)
361 {
362     BDRVQEDState *s = bs->opaque;
363 
364     qed_cancel_need_check_timer(s);
365     timer_free(s->need_check_timer);
366 }
367 
368 static void bdrv_qed_attach_aio_context(BlockDriverState *bs,
369                                         AioContext *new_context)
370 {
371     BDRVQEDState *s = bs->opaque;
372 
373     s->need_check_timer = aio_timer_new(new_context,
374                                         QEMU_CLOCK_VIRTUAL, SCALE_NS,
375                                         qed_need_check_timer_cb, s);
376     if (s->header.features & QED_F_NEED_CHECK) {
377         qed_start_need_check_timer(s);
378     }
379 }
380 
381 static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags,
382                          Error **errp)
383 {
384     BDRVQEDState *s = bs->opaque;
385     QEDHeader le_header;
386     int64_t file_size;
387     int ret;
388 
389     s->bs = bs;
390     QSIMPLEQ_INIT(&s->allocating_write_reqs);
391 
392     ret = bdrv_pread(bs->file->bs, 0, &le_header, sizeof(le_header));
393     if (ret < 0) {
394         return ret;
395     }
396     qed_header_le_to_cpu(&le_header, &s->header);
397 
398     if (s->header.magic != QED_MAGIC) {
399         error_setg(errp, "Image not in QED format");
400         return -EINVAL;
401     }
402     if (s->header.features & ~QED_FEATURE_MASK) {
403         /* image uses unsupported feature bits */
404         error_setg(errp, "Unsupported QED features: %" PRIx64,
405                    s->header.features & ~QED_FEATURE_MASK);
406         return -ENOTSUP;
407     }
408     if (!qed_is_cluster_size_valid(s->header.cluster_size)) {
409         return -EINVAL;
410     }
411 
412     /* Round down file size to the last cluster */
413     file_size = bdrv_getlength(bs->file->bs);
414     if (file_size < 0) {
415         return file_size;
416     }
417     s->file_size = qed_start_of_cluster(s, file_size);
418 
419     if (!qed_is_table_size_valid(s->header.table_size)) {
420         return -EINVAL;
421     }
422     if (!qed_is_image_size_valid(s->header.image_size,
423                                  s->header.cluster_size,
424                                  s->header.table_size)) {
425         return -EINVAL;
426     }
427     if (!qed_check_table_offset(s, s->header.l1_table_offset)) {
428         return -EINVAL;
429     }
430 
431     s->table_nelems = (s->header.cluster_size * s->header.table_size) /
432                       sizeof(uint64_t);
433     s->l2_shift = ctz32(s->header.cluster_size);
434     s->l2_mask = s->table_nelems - 1;
435     s->l1_shift = s->l2_shift + ctz32(s->table_nelems);
436 
437     /* Header size calculation must not overflow uint32_t */
438     if (s->header.header_size > UINT32_MAX / s->header.cluster_size) {
439         return -EINVAL;
440     }
441 
442     if ((s->header.features & QED_F_BACKING_FILE)) {
443         if ((uint64_t)s->header.backing_filename_offset +
444             s->header.backing_filename_size >
445             s->header.cluster_size * s->header.header_size) {
446             return -EINVAL;
447         }
448 
449         ret = qed_read_string(bs->file->bs, s->header.backing_filename_offset,
450                               s->header.backing_filename_size, bs->backing_file,
451                               sizeof(bs->backing_file));
452         if (ret < 0) {
453             return ret;
454         }
455 
456         if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) {
457             pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw");
458         }
459     }
460 
461     /* Reset unknown autoclear feature bits.  This is a backwards
462      * compatibility mechanism that allows images to be opened by older
463      * programs, which "knock out" unknown feature bits.  When an image is
464      * opened by a newer program again it can detect that the autoclear
465      * feature is no longer valid.
466      */
467     if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 &&
468         !bdrv_is_read_only(bs->file->bs) && !(flags & BDRV_O_INACTIVE)) {
469         s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK;
470 
471         ret = qed_write_header_sync(s);
472         if (ret) {
473             return ret;
474         }
475 
476         /* From here on only known autoclear feature bits are valid */
477         bdrv_flush(bs->file->bs);
478     }
479 
480     s->l1_table = qed_alloc_table(s);
481     qed_init_l2_cache(&s->l2_cache);
482 
483     ret = qed_read_l1_table_sync(s);
484     if (ret) {
485         goto out;
486     }
487 
488     /* If image was not closed cleanly, check consistency */
489     if (!(flags & BDRV_O_CHECK) && (s->header.features & QED_F_NEED_CHECK)) {
490         /* Read-only images cannot be fixed.  There is no risk of corruption
491          * since write operations are not possible.  Therefore, allow
492          * potentially inconsistent images to be opened read-only.  This can
493          * aid data recovery from an otherwise inconsistent image.
494          */
495         if (!bdrv_is_read_only(bs->file->bs) &&
496             !(flags & BDRV_O_INACTIVE)) {
497             BdrvCheckResult result = {0};
498 
499             ret = qed_check(s, &result, true);
500             if (ret) {
501                 goto out;
502             }
503         }
504     }
505 
506     bdrv_qed_attach_aio_context(bs, bdrv_get_aio_context(bs));
507 
508 out:
509     if (ret) {
510         qed_free_l2_cache(&s->l2_cache);
511         qemu_vfree(s->l1_table);
512     }
513     return ret;
514 }
515 
516 static void bdrv_qed_refresh_limits(BlockDriverState *bs, Error **errp)
517 {
518     BDRVQEDState *s = bs->opaque;
519 
520     bs->bl.write_zeroes_alignment = s->header.cluster_size >> BDRV_SECTOR_BITS;
521 }
522 
523 /* We have nothing to do for QED reopen, stubs just return
524  * success */
525 static int bdrv_qed_reopen_prepare(BDRVReopenState *state,
526                                    BlockReopenQueue *queue, Error **errp)
527 {
528     return 0;
529 }
530 
531 static void bdrv_qed_close(BlockDriverState *bs)
532 {
533     BDRVQEDState *s = bs->opaque;
534 
535     bdrv_qed_detach_aio_context(bs);
536 
537     /* Ensure writes reach stable storage */
538     bdrv_flush(bs->file->bs);
539 
540     /* Clean shutdown, no check required on next open */
541     if (s->header.features & QED_F_NEED_CHECK) {
542         s->header.features &= ~QED_F_NEED_CHECK;
543         qed_write_header_sync(s);
544     }
545 
546     qed_free_l2_cache(&s->l2_cache);
547     qemu_vfree(s->l1_table);
548 }
549 
550 static int qed_create(const char *filename, uint32_t cluster_size,
551                       uint64_t image_size, uint32_t table_size,
552                       const char *backing_file, const char *backing_fmt,
553                       QemuOpts *opts, Error **errp)
554 {
555     QEDHeader header = {
556         .magic = QED_MAGIC,
557         .cluster_size = cluster_size,
558         .table_size = table_size,
559         .header_size = 1,
560         .features = 0,
561         .compat_features = 0,
562         .l1_table_offset = cluster_size,
563         .image_size = image_size,
564     };
565     QEDHeader le_header;
566     uint8_t *l1_table = NULL;
567     size_t l1_size = header.cluster_size * header.table_size;
568     Error *local_err = NULL;
569     int ret = 0;
570     BlockBackend *blk;
571 
572     ret = bdrv_create_file(filename, opts, &local_err);
573     if (ret < 0) {
574         error_propagate(errp, local_err);
575         return ret;
576     }
577 
578     blk = blk_new_open(filename, NULL, NULL,
579                        BDRV_O_RDWR | BDRV_O_PROTOCOL, &local_err);
580     if (blk == NULL) {
581         error_propagate(errp, local_err);
582         return -EIO;
583     }
584 
585     blk_set_allow_write_beyond_eof(blk, true);
586 
587     /* File must start empty and grow, check truncate is supported */
588     ret = blk_truncate(blk, 0);
589     if (ret < 0) {
590         goto out;
591     }
592 
593     if (backing_file) {
594         header.features |= QED_F_BACKING_FILE;
595         header.backing_filename_offset = sizeof(le_header);
596         header.backing_filename_size = strlen(backing_file);
597 
598         if (qed_fmt_is_raw(backing_fmt)) {
599             header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
600         }
601     }
602 
603     qed_header_cpu_to_le(&header, &le_header);
604     ret = blk_pwrite(blk, 0, &le_header, sizeof(le_header), 0);
605     if (ret < 0) {
606         goto out;
607     }
608     ret = blk_pwrite(blk, sizeof(le_header), backing_file,
609                      header.backing_filename_size, 0);
610     if (ret < 0) {
611         goto out;
612     }
613 
614     l1_table = g_malloc0(l1_size);
615     ret = blk_pwrite(blk, header.l1_table_offset, l1_table, l1_size, 0);
616     if (ret < 0) {
617         goto out;
618     }
619 
620     ret = 0; /* success */
621 out:
622     g_free(l1_table);
623     blk_unref(blk);
624     return ret;
625 }
626 
627 static int bdrv_qed_create(const char *filename, QemuOpts *opts, Error **errp)
628 {
629     uint64_t image_size = 0;
630     uint32_t cluster_size = QED_DEFAULT_CLUSTER_SIZE;
631     uint32_t table_size = QED_DEFAULT_TABLE_SIZE;
632     char *backing_file = NULL;
633     char *backing_fmt = NULL;
634     int ret;
635 
636     image_size = ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
637                           BDRV_SECTOR_SIZE);
638     backing_file = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FILE);
639     backing_fmt = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FMT);
640     cluster_size = qemu_opt_get_size_del(opts,
641                                          BLOCK_OPT_CLUSTER_SIZE,
642                                          QED_DEFAULT_CLUSTER_SIZE);
643     table_size = qemu_opt_get_size_del(opts, BLOCK_OPT_TABLE_SIZE,
644                                        QED_DEFAULT_TABLE_SIZE);
645 
646     if (!qed_is_cluster_size_valid(cluster_size)) {
647         error_setg(errp, "QED cluster size must be within range [%u, %u] "
648                          "and power of 2",
649                    QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE);
650         ret = -EINVAL;
651         goto finish;
652     }
653     if (!qed_is_table_size_valid(table_size)) {
654         error_setg(errp, "QED table size must be within range [%u, %u] "
655                          "and power of 2",
656                    QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE);
657         ret = -EINVAL;
658         goto finish;
659     }
660     if (!qed_is_image_size_valid(image_size, cluster_size, table_size)) {
661         error_setg(errp, "QED image size must be a non-zero multiple of "
662                          "cluster size and less than %" PRIu64 " bytes",
663                    qed_max_image_size(cluster_size, table_size));
664         ret = -EINVAL;
665         goto finish;
666     }
667 
668     ret = qed_create(filename, cluster_size, image_size, table_size,
669                      backing_file, backing_fmt, opts, errp);
670 
671 finish:
672     g_free(backing_file);
673     g_free(backing_fmt);
674     return ret;
675 }
676 
677 typedef struct {
678     BlockDriverState *bs;
679     Coroutine *co;
680     uint64_t pos;
681     int64_t status;
682     int *pnum;
683     BlockDriverState **file;
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         *cb->file = cb->bs->file->bs;
696         break;
697     case QED_CLUSTER_ZERO:
698         cb->status = BDRV_BLOCK_ZERO;
699         break;
700     case QED_CLUSTER_L2:
701     case QED_CLUSTER_L1:
702         cb->status = 0;
703         break;
704     default:
705         assert(ret < 0);
706         cb->status = ret;
707         break;
708     }
709 
710     if (cb->co) {
711         qemu_coroutine_enter(cb->co, NULL);
712     }
713 }
714 
715 static int64_t coroutine_fn bdrv_qed_co_get_block_status(BlockDriverState *bs,
716                                                  int64_t sector_num,
717                                                  int nb_sectors, int *pnum,
718                                                  BlockDriverState **file)
719 {
720     BDRVQEDState *s = bs->opaque;
721     size_t len = (size_t)nb_sectors * BDRV_SECTOR_SIZE;
722     QEDIsAllocatedCB cb = {
723         .bs = bs,
724         .pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE,
725         .status = BDRV_BLOCK_OFFSET_MASK,
726         .pnum = pnum,
727         .file = file,
728     };
729     QEDRequest request = { .l2_table = NULL };
730 
731     qed_find_cluster(s, &request, cb.pos, len, qed_is_allocated_cb, &cb);
732 
733     /* Now sleep if the callback wasn't invoked immediately */
734     while (cb.status == BDRV_BLOCK_OFFSET_MASK) {
735         cb.co = qemu_coroutine_self();
736         qemu_coroutine_yield();
737     }
738 
739     qed_unref_l2_cache_entry(request.l2_table);
740 
741     return cb.status;
742 }
743 
744 static BDRVQEDState *acb_to_s(QEDAIOCB *acb)
745 {
746     return acb->common.bs->opaque;
747 }
748 
749 /**
750  * Read from the backing file or zero-fill if no backing file
751  *
752  * @s:              QED state
753  * @pos:            Byte position in device
754  * @qiov:           Destination I/O vector
755  * @backing_qiov:   Possibly shortened copy of qiov, to be allocated here
756  * @cb:             Completion function
757  * @opaque:         User data for completion function
758  *
759  * This function reads qiov->size bytes starting at pos from the backing file.
760  * If there is no backing file then zeroes are read.
761  */
762 static void qed_read_backing_file(BDRVQEDState *s, uint64_t pos,
763                                   QEMUIOVector *qiov,
764                                   QEMUIOVector **backing_qiov,
765                                   BlockCompletionFunc *cb, void *opaque)
766 {
767     uint64_t backing_length = 0;
768     size_t size;
769 
770     /* If there is a backing file, get its length.  Treat the absence of a
771      * backing file like a zero length backing file.
772      */
773     if (s->bs->backing) {
774         int64_t l = bdrv_getlength(s->bs->backing->bs);
775         if (l < 0) {
776             cb(opaque, l);
777             return;
778         }
779         backing_length = l;
780     }
781 
782     /* Zero all sectors if reading beyond the end of the backing file */
783     if (pos >= backing_length ||
784         pos + qiov->size > backing_length) {
785         qemu_iovec_memset(qiov, 0, 0, qiov->size);
786     }
787 
788     /* Complete now if there are no backing file sectors to read */
789     if (pos >= backing_length) {
790         cb(opaque, 0);
791         return;
792     }
793 
794     /* If the read straddles the end of the backing file, shorten it */
795     size = MIN((uint64_t)backing_length - pos, qiov->size);
796 
797     assert(*backing_qiov == NULL);
798     *backing_qiov = g_new(QEMUIOVector, 1);
799     qemu_iovec_init(*backing_qiov, qiov->niov);
800     qemu_iovec_concat(*backing_qiov, qiov, 0, size);
801 
802     BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO);
803     bdrv_aio_readv(s->bs->backing->bs, pos / BDRV_SECTOR_SIZE,
804                    *backing_qiov, size / BDRV_SECTOR_SIZE, cb, opaque);
805 }
806 
807 typedef struct {
808     GenericCB gencb;
809     BDRVQEDState *s;
810     QEMUIOVector qiov;
811     QEMUIOVector *backing_qiov;
812     struct iovec iov;
813     uint64_t offset;
814 } CopyFromBackingFileCB;
815 
816 static void qed_copy_from_backing_file_cb(void *opaque, int ret)
817 {
818     CopyFromBackingFileCB *copy_cb = opaque;
819     qemu_vfree(copy_cb->iov.iov_base);
820     gencb_complete(&copy_cb->gencb, ret);
821 }
822 
823 static void qed_copy_from_backing_file_write(void *opaque, int ret)
824 {
825     CopyFromBackingFileCB *copy_cb = opaque;
826     BDRVQEDState *s = copy_cb->s;
827 
828     if (copy_cb->backing_qiov) {
829         qemu_iovec_destroy(copy_cb->backing_qiov);
830         g_free(copy_cb->backing_qiov);
831         copy_cb->backing_qiov = NULL;
832     }
833 
834     if (ret) {
835         qed_copy_from_backing_file_cb(copy_cb, ret);
836         return;
837     }
838 
839     BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE);
840     bdrv_aio_writev(s->bs->file->bs, copy_cb->offset / BDRV_SECTOR_SIZE,
841                     &copy_cb->qiov, copy_cb->qiov.size / BDRV_SECTOR_SIZE,
842                     qed_copy_from_backing_file_cb, copy_cb);
843 }
844 
845 /**
846  * Copy data from backing file into the image
847  *
848  * @s:          QED state
849  * @pos:        Byte position in device
850  * @len:        Number of bytes
851  * @offset:     Byte offset in image file
852  * @cb:         Completion function
853  * @opaque:     User data for completion function
854  */
855 static void qed_copy_from_backing_file(BDRVQEDState *s, uint64_t pos,
856                                        uint64_t len, uint64_t offset,
857                                        BlockCompletionFunc *cb,
858                                        void *opaque)
859 {
860     CopyFromBackingFileCB *copy_cb;
861 
862     /* Skip copy entirely if there is no work to do */
863     if (len == 0) {
864         cb(opaque, 0);
865         return;
866     }
867 
868     copy_cb = gencb_alloc(sizeof(*copy_cb), cb, opaque);
869     copy_cb->s = s;
870     copy_cb->offset = offset;
871     copy_cb->backing_qiov = NULL;
872     copy_cb->iov.iov_base = qemu_blockalign(s->bs, len);
873     copy_cb->iov.iov_len = len;
874     qemu_iovec_init_external(&copy_cb->qiov, &copy_cb->iov, 1);
875 
876     qed_read_backing_file(s, pos, &copy_cb->qiov, &copy_cb->backing_qiov,
877                           qed_copy_from_backing_file_write, copy_cb);
878 }
879 
880 /**
881  * Link one or more contiguous clusters into a table
882  *
883  * @s:              QED state
884  * @table:          L2 table
885  * @index:          First cluster index
886  * @n:              Number of contiguous clusters
887  * @cluster:        First cluster offset
888  *
889  * The cluster offset may be an allocated byte offset in the image file, the
890  * zero cluster marker, or the unallocated cluster marker.
891  */
892 static void qed_update_l2_table(BDRVQEDState *s, QEDTable *table, int index,
893                                 unsigned int n, uint64_t cluster)
894 {
895     int i;
896     for (i = index; i < index + n; i++) {
897         table->offsets[i] = cluster;
898         if (!qed_offset_is_unalloc_cluster(cluster) &&
899             !qed_offset_is_zero_cluster(cluster)) {
900             cluster += s->header.cluster_size;
901         }
902     }
903 }
904 
905 static void qed_aio_complete_bh(void *opaque)
906 {
907     QEDAIOCB *acb = opaque;
908     BlockCompletionFunc *cb = acb->common.cb;
909     void *user_opaque = acb->common.opaque;
910     int ret = acb->bh_ret;
911 
912     qemu_bh_delete(acb->bh);
913     qemu_aio_unref(acb);
914 
915     /* Invoke callback */
916     cb(user_opaque, ret);
917 }
918 
919 static void qed_aio_complete(QEDAIOCB *acb, int ret)
920 {
921     BDRVQEDState *s = acb_to_s(acb);
922 
923     trace_qed_aio_complete(s, acb, ret);
924 
925     /* Free resources */
926     qemu_iovec_destroy(&acb->cur_qiov);
927     qed_unref_l2_cache_entry(acb->request.l2_table);
928 
929     /* Free the buffer we may have allocated for zero writes */
930     if (acb->flags & QED_AIOCB_ZERO) {
931         qemu_vfree(acb->qiov->iov[0].iov_base);
932         acb->qiov->iov[0].iov_base = NULL;
933     }
934 
935     /* Arrange for a bh to invoke the completion function */
936     acb->bh_ret = ret;
937     acb->bh = aio_bh_new(bdrv_get_aio_context(acb->common.bs),
938                          qed_aio_complete_bh, acb);
939     qemu_bh_schedule(acb->bh);
940 
941     /* Start next allocating write request waiting behind this one.  Note that
942      * requests enqueue themselves when they first hit an unallocated cluster
943      * but they wait until the entire request is finished before waking up the
944      * next request in the queue.  This ensures that we don't cycle through
945      * requests multiple times but rather finish one at a time completely.
946      */
947     if (acb == QSIMPLEQ_FIRST(&s->allocating_write_reqs)) {
948         QSIMPLEQ_REMOVE_HEAD(&s->allocating_write_reqs, next);
949         acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs);
950         if (acb) {
951             qed_aio_next_io(acb, 0);
952         } else if (s->header.features & QED_F_NEED_CHECK) {
953             qed_start_need_check_timer(s);
954         }
955     }
956 }
957 
958 /**
959  * Commit the current L2 table to the cache
960  */
961 static void qed_commit_l2_update(void *opaque, int ret)
962 {
963     QEDAIOCB *acb = opaque;
964     BDRVQEDState *s = acb_to_s(acb);
965     CachedL2Table *l2_table = acb->request.l2_table;
966     uint64_t l2_offset = l2_table->offset;
967 
968     qed_commit_l2_cache_entry(&s->l2_cache, l2_table);
969 
970     /* This is guaranteed to succeed because we just committed the entry to the
971      * cache.
972      */
973     acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset);
974     assert(acb->request.l2_table != NULL);
975 
976     qed_aio_next_io(opaque, ret);
977 }
978 
979 /**
980  * Update L1 table with new L2 table offset and write it out
981  */
982 static void qed_aio_write_l1_update(void *opaque, int ret)
983 {
984     QEDAIOCB *acb = opaque;
985     BDRVQEDState *s = acb_to_s(acb);
986     int index;
987 
988     if (ret) {
989         qed_aio_complete(acb, ret);
990         return;
991     }
992 
993     index = qed_l1_index(s, acb->cur_pos);
994     s->l1_table->offsets[index] = acb->request.l2_table->offset;
995 
996     qed_write_l1_table(s, index, 1, qed_commit_l2_update, acb);
997 }
998 
999 /**
1000  * Update L2 table with new cluster offsets and write them out
1001  */
1002 static void qed_aio_write_l2_update(QEDAIOCB *acb, int ret, uint64_t offset)
1003 {
1004     BDRVQEDState *s = acb_to_s(acb);
1005     bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1;
1006     int index;
1007 
1008     if (ret) {
1009         goto err;
1010     }
1011 
1012     if (need_alloc) {
1013         qed_unref_l2_cache_entry(acb->request.l2_table);
1014         acb->request.l2_table = qed_new_l2_table(s);
1015     }
1016 
1017     index = qed_l2_index(s, acb->cur_pos);
1018     qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters,
1019                          offset);
1020 
1021     if (need_alloc) {
1022         /* Write out the whole new L2 table */
1023         qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true,
1024                             qed_aio_write_l1_update, acb);
1025     } else {
1026         /* Write out only the updated part of the L2 table */
1027         qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters, false,
1028                             qed_aio_next_io, acb);
1029     }
1030     return;
1031 
1032 err:
1033     qed_aio_complete(acb, ret);
1034 }
1035 
1036 static void qed_aio_write_l2_update_cb(void *opaque, int ret)
1037 {
1038     QEDAIOCB *acb = opaque;
1039     qed_aio_write_l2_update(acb, ret, acb->cur_cluster);
1040 }
1041 
1042 /**
1043  * Flush new data clusters before updating the L2 table
1044  *
1045  * This flush is necessary when a backing file is in use.  A crash during an
1046  * allocating write could result in empty clusters in the image.  If the write
1047  * only touched a subregion of the cluster, then backing image sectors have
1048  * been lost in the untouched region.  The solution is to flush after writing a
1049  * new data cluster and before updating the L2 table.
1050  */
1051 static void qed_aio_write_flush_before_l2_update(void *opaque, int ret)
1052 {
1053     QEDAIOCB *acb = opaque;
1054     BDRVQEDState *s = acb_to_s(acb);
1055 
1056     if (!bdrv_aio_flush(s->bs->file->bs, qed_aio_write_l2_update_cb, opaque)) {
1057         qed_aio_complete(acb, -EIO);
1058     }
1059 }
1060 
1061 /**
1062  * Write data to the image file
1063  */
1064 static void qed_aio_write_main(void *opaque, int ret)
1065 {
1066     QEDAIOCB *acb = opaque;
1067     BDRVQEDState *s = acb_to_s(acb);
1068     uint64_t offset = acb->cur_cluster +
1069                       qed_offset_into_cluster(s, acb->cur_pos);
1070     BlockCompletionFunc *next_fn;
1071 
1072     trace_qed_aio_write_main(s, acb, ret, offset, acb->cur_qiov.size);
1073 
1074     if (ret) {
1075         qed_aio_complete(acb, ret);
1076         return;
1077     }
1078 
1079     if (acb->find_cluster_ret == QED_CLUSTER_FOUND) {
1080         next_fn = qed_aio_next_io;
1081     } else {
1082         if (s->bs->backing) {
1083             next_fn = qed_aio_write_flush_before_l2_update;
1084         } else {
1085             next_fn = qed_aio_write_l2_update_cb;
1086         }
1087     }
1088 
1089     BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO);
1090     bdrv_aio_writev(s->bs->file->bs, offset / BDRV_SECTOR_SIZE,
1091                     &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE,
1092                     next_fn, acb);
1093 }
1094 
1095 /**
1096  * Populate back untouched region of new data cluster
1097  */
1098 static void qed_aio_write_postfill(void *opaque, int ret)
1099 {
1100     QEDAIOCB *acb = opaque;
1101     BDRVQEDState *s = acb_to_s(acb);
1102     uint64_t start = acb->cur_pos + acb->cur_qiov.size;
1103     uint64_t len =
1104         qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start;
1105     uint64_t offset = acb->cur_cluster +
1106                       qed_offset_into_cluster(s, acb->cur_pos) +
1107                       acb->cur_qiov.size;
1108 
1109     if (ret) {
1110         qed_aio_complete(acb, ret);
1111         return;
1112     }
1113 
1114     trace_qed_aio_write_postfill(s, acb, start, len, offset);
1115     qed_copy_from_backing_file(s, start, len, offset,
1116                                 qed_aio_write_main, acb);
1117 }
1118 
1119 /**
1120  * Populate front untouched region of new data cluster
1121  */
1122 static void qed_aio_write_prefill(void *opaque, int ret)
1123 {
1124     QEDAIOCB *acb = opaque;
1125     BDRVQEDState *s = acb_to_s(acb);
1126     uint64_t start = qed_start_of_cluster(s, acb->cur_pos);
1127     uint64_t len = qed_offset_into_cluster(s, acb->cur_pos);
1128 
1129     trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster);
1130     qed_copy_from_backing_file(s, start, len, acb->cur_cluster,
1131                                 qed_aio_write_postfill, acb);
1132 }
1133 
1134 /**
1135  * Check if the QED_F_NEED_CHECK bit should be set during allocating write
1136  */
1137 static bool qed_should_set_need_check(BDRVQEDState *s)
1138 {
1139     /* The flush before L2 update path ensures consistency */
1140     if (s->bs->backing) {
1141         return false;
1142     }
1143 
1144     return !(s->header.features & QED_F_NEED_CHECK);
1145 }
1146 
1147 static void qed_aio_write_zero_cluster(void *opaque, int ret)
1148 {
1149     QEDAIOCB *acb = opaque;
1150 
1151     if (ret) {
1152         qed_aio_complete(acb, ret);
1153         return;
1154     }
1155 
1156     qed_aio_write_l2_update(acb, 0, 1);
1157 }
1158 
1159 /**
1160  * Write new data cluster
1161  *
1162  * @acb:        Write request
1163  * @len:        Length in bytes
1164  *
1165  * This path is taken when writing to previously unallocated clusters.
1166  */
1167 static void qed_aio_write_alloc(QEDAIOCB *acb, size_t len)
1168 {
1169     BDRVQEDState *s = acb_to_s(acb);
1170     BlockCompletionFunc *cb;
1171 
1172     /* Cancel timer when the first allocating request comes in */
1173     if (QSIMPLEQ_EMPTY(&s->allocating_write_reqs)) {
1174         qed_cancel_need_check_timer(s);
1175     }
1176 
1177     /* Freeze this request if another allocating write is in progress */
1178     if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs)) {
1179         QSIMPLEQ_INSERT_TAIL(&s->allocating_write_reqs, acb, next);
1180     }
1181     if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs) ||
1182         s->allocating_write_reqs_plugged) {
1183         return; /* wait for existing request to finish */
1184     }
1185 
1186     acb->cur_nclusters = qed_bytes_to_clusters(s,
1187             qed_offset_into_cluster(s, acb->cur_pos) + len);
1188     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1189 
1190     if (acb->flags & QED_AIOCB_ZERO) {
1191         /* Skip ahead if the clusters are already zero */
1192         if (acb->find_cluster_ret == QED_CLUSTER_ZERO) {
1193             qed_aio_next_io(acb, 0);
1194             return;
1195         }
1196 
1197         cb = qed_aio_write_zero_cluster;
1198     } else {
1199         cb = qed_aio_write_prefill;
1200         acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters);
1201     }
1202 
1203     if (qed_should_set_need_check(s)) {
1204         s->header.features |= QED_F_NEED_CHECK;
1205         qed_write_header(s, cb, acb);
1206     } else {
1207         cb(acb, 0);
1208     }
1209 }
1210 
1211 /**
1212  * Write data cluster in place
1213  *
1214  * @acb:        Write request
1215  * @offset:     Cluster offset in bytes
1216  * @len:        Length in bytes
1217  *
1218  * This path is taken when writing to already allocated clusters.
1219  */
1220 static void qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset, size_t len)
1221 {
1222     /* Allocate buffer for zero writes */
1223     if (acb->flags & QED_AIOCB_ZERO) {
1224         struct iovec *iov = acb->qiov->iov;
1225 
1226         if (!iov->iov_base) {
1227             iov->iov_base = qemu_try_blockalign(acb->common.bs, iov->iov_len);
1228             if (iov->iov_base == NULL) {
1229                 qed_aio_complete(acb, -ENOMEM);
1230                 return;
1231             }
1232             memset(iov->iov_base, 0, iov->iov_len);
1233         }
1234     }
1235 
1236     /* Calculate the I/O vector */
1237     acb->cur_cluster = offset;
1238     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1239 
1240     /* Do the actual write */
1241     qed_aio_write_main(acb, 0);
1242 }
1243 
1244 /**
1245  * Write data cluster
1246  *
1247  * @opaque:     Write request
1248  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1,
1249  *              or -errno
1250  * @offset:     Cluster offset in bytes
1251  * @len:        Length in bytes
1252  *
1253  * Callback from qed_find_cluster().
1254  */
1255 static void qed_aio_write_data(void *opaque, int ret,
1256                                uint64_t offset, size_t len)
1257 {
1258     QEDAIOCB *acb = opaque;
1259 
1260     trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len);
1261 
1262     acb->find_cluster_ret = ret;
1263 
1264     switch (ret) {
1265     case QED_CLUSTER_FOUND:
1266         qed_aio_write_inplace(acb, offset, len);
1267         break;
1268 
1269     case QED_CLUSTER_L2:
1270     case QED_CLUSTER_L1:
1271     case QED_CLUSTER_ZERO:
1272         qed_aio_write_alloc(acb, len);
1273         break;
1274 
1275     default:
1276         qed_aio_complete(acb, ret);
1277         break;
1278     }
1279 }
1280 
1281 /**
1282  * Read data cluster
1283  *
1284  * @opaque:     Read request
1285  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1,
1286  *              or -errno
1287  * @offset:     Cluster offset in bytes
1288  * @len:        Length in bytes
1289  *
1290  * Callback from qed_find_cluster().
1291  */
1292 static void qed_aio_read_data(void *opaque, int ret,
1293                               uint64_t offset, size_t len)
1294 {
1295     QEDAIOCB *acb = opaque;
1296     BDRVQEDState *s = acb_to_s(acb);
1297     BlockDriverState *bs = acb->common.bs;
1298 
1299     /* Adjust offset into cluster */
1300     offset += qed_offset_into_cluster(s, acb->cur_pos);
1301 
1302     trace_qed_aio_read_data(s, acb, ret, offset, len);
1303 
1304     if (ret < 0) {
1305         goto err;
1306     }
1307 
1308     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1309 
1310     /* Handle zero cluster and backing file reads */
1311     if (ret == QED_CLUSTER_ZERO) {
1312         qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size);
1313         qed_aio_next_io(acb, 0);
1314         return;
1315     } else if (ret != QED_CLUSTER_FOUND) {
1316         qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov,
1317                               &acb->backing_qiov, qed_aio_next_io, acb);
1318         return;
1319     }
1320 
1321     BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
1322     bdrv_aio_readv(bs->file->bs, offset / BDRV_SECTOR_SIZE,
1323                    &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE,
1324                    qed_aio_next_io, acb);
1325     return;
1326 
1327 err:
1328     qed_aio_complete(acb, ret);
1329 }
1330 
1331 /**
1332  * Begin next I/O or complete the request
1333  */
1334 static void qed_aio_next_io(void *opaque, int ret)
1335 {
1336     QEDAIOCB *acb = opaque;
1337     BDRVQEDState *s = acb_to_s(acb);
1338     QEDFindClusterFunc *io_fn = (acb->flags & QED_AIOCB_WRITE) ?
1339                                 qed_aio_write_data : qed_aio_read_data;
1340 
1341     trace_qed_aio_next_io(s, acb, ret, acb->cur_pos + acb->cur_qiov.size);
1342 
1343     if (acb->backing_qiov) {
1344         qemu_iovec_destroy(acb->backing_qiov);
1345         g_free(acb->backing_qiov);
1346         acb->backing_qiov = NULL;
1347     }
1348 
1349     /* Handle I/O error */
1350     if (ret) {
1351         qed_aio_complete(acb, ret);
1352         return;
1353     }
1354 
1355     acb->qiov_offset += acb->cur_qiov.size;
1356     acb->cur_pos += acb->cur_qiov.size;
1357     qemu_iovec_reset(&acb->cur_qiov);
1358 
1359     /* Complete request */
1360     if (acb->cur_pos >= acb->end_pos) {
1361         qed_aio_complete(acb, 0);
1362         return;
1363     }
1364 
1365     /* Find next cluster and start I/O */
1366     qed_find_cluster(s, &acb->request,
1367                       acb->cur_pos, acb->end_pos - acb->cur_pos,
1368                       io_fn, acb);
1369 }
1370 
1371 static BlockAIOCB *qed_aio_setup(BlockDriverState *bs,
1372                                  int64_t sector_num,
1373                                  QEMUIOVector *qiov, int nb_sectors,
1374                                  BlockCompletionFunc *cb,
1375                                  void *opaque, int flags)
1376 {
1377     QEDAIOCB *acb = qemu_aio_get(&qed_aiocb_info, bs, cb, opaque);
1378 
1379     trace_qed_aio_setup(bs->opaque, acb, sector_num, nb_sectors,
1380                         opaque, flags);
1381 
1382     acb->flags = flags;
1383     acb->qiov = qiov;
1384     acb->qiov_offset = 0;
1385     acb->cur_pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE;
1386     acb->end_pos = acb->cur_pos + nb_sectors * BDRV_SECTOR_SIZE;
1387     acb->backing_qiov = NULL;
1388     acb->request.l2_table = NULL;
1389     qemu_iovec_init(&acb->cur_qiov, qiov->niov);
1390 
1391     /* Start request */
1392     qed_aio_next_io(acb, 0);
1393     return &acb->common;
1394 }
1395 
1396 static BlockAIOCB *bdrv_qed_aio_readv(BlockDriverState *bs,
1397                                       int64_t sector_num,
1398                                       QEMUIOVector *qiov, int nb_sectors,
1399                                       BlockCompletionFunc *cb,
1400                                       void *opaque)
1401 {
1402     return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb, opaque, 0);
1403 }
1404 
1405 static BlockAIOCB *bdrv_qed_aio_writev(BlockDriverState *bs,
1406                                        int64_t sector_num,
1407                                        QEMUIOVector *qiov, int nb_sectors,
1408                                        BlockCompletionFunc *cb,
1409                                        void *opaque)
1410 {
1411     return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb,
1412                          opaque, QED_AIOCB_WRITE);
1413 }
1414 
1415 typedef struct {
1416     Coroutine *co;
1417     int ret;
1418     bool done;
1419 } QEDWriteZeroesCB;
1420 
1421 static void coroutine_fn qed_co_write_zeroes_cb(void *opaque, int ret)
1422 {
1423     QEDWriteZeroesCB *cb = opaque;
1424 
1425     cb->done = true;
1426     cb->ret = ret;
1427     if (cb->co) {
1428         qemu_coroutine_enter(cb->co, NULL);
1429     }
1430 }
1431 
1432 static int coroutine_fn bdrv_qed_co_write_zeroes(BlockDriverState *bs,
1433                                                  int64_t sector_num,
1434                                                  int nb_sectors,
1435                                                  BdrvRequestFlags flags)
1436 {
1437     BlockAIOCB *blockacb;
1438     BDRVQEDState *s = bs->opaque;
1439     QEDWriteZeroesCB cb = { .done = false };
1440     QEMUIOVector qiov;
1441     struct iovec iov;
1442 
1443     /* Refuse if there are untouched backing file sectors */
1444     if (bs->backing) {
1445         if (qed_offset_into_cluster(s, sector_num * BDRV_SECTOR_SIZE) != 0) {
1446             return -ENOTSUP;
1447         }
1448         if (qed_offset_into_cluster(s, nb_sectors * BDRV_SECTOR_SIZE) != 0) {
1449             return -ENOTSUP;
1450         }
1451     }
1452 
1453     /* Zero writes start without an I/O buffer.  If a buffer becomes necessary
1454      * then it will be allocated during request processing.
1455      */
1456     iov.iov_base = NULL,
1457     iov.iov_len  = nb_sectors * BDRV_SECTOR_SIZE,
1458 
1459     qemu_iovec_init_external(&qiov, &iov, 1);
1460     blockacb = qed_aio_setup(bs, sector_num, &qiov, nb_sectors,
1461                              qed_co_write_zeroes_cb, &cb,
1462                              QED_AIOCB_WRITE | QED_AIOCB_ZERO);
1463     if (!blockacb) {
1464         return -EIO;
1465     }
1466     if (!cb.done) {
1467         cb.co = qemu_coroutine_self();
1468         qemu_coroutine_yield();
1469     }
1470     assert(cb.done);
1471     return cb.ret;
1472 }
1473 
1474 static int bdrv_qed_truncate(BlockDriverState *bs, int64_t offset)
1475 {
1476     BDRVQEDState *s = bs->opaque;
1477     uint64_t old_image_size;
1478     int ret;
1479 
1480     if (!qed_is_image_size_valid(offset, s->header.cluster_size,
1481                                  s->header.table_size)) {
1482         return -EINVAL;
1483     }
1484 
1485     /* Shrinking is currently not supported */
1486     if ((uint64_t)offset < s->header.image_size) {
1487         return -ENOTSUP;
1488     }
1489 
1490     old_image_size = s->header.image_size;
1491     s->header.image_size = offset;
1492     ret = qed_write_header_sync(s);
1493     if (ret < 0) {
1494         s->header.image_size = old_image_size;
1495     }
1496     return ret;
1497 }
1498 
1499 static int64_t bdrv_qed_getlength(BlockDriverState *bs)
1500 {
1501     BDRVQEDState *s = bs->opaque;
1502     return s->header.image_size;
1503 }
1504 
1505 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1506 {
1507     BDRVQEDState *s = bs->opaque;
1508 
1509     memset(bdi, 0, sizeof(*bdi));
1510     bdi->cluster_size = s->header.cluster_size;
1511     bdi->is_dirty = s->header.features & QED_F_NEED_CHECK;
1512     bdi->unallocated_blocks_are_zero = true;
1513     bdi->can_write_zeroes_with_unmap = true;
1514     return 0;
1515 }
1516 
1517 static int bdrv_qed_change_backing_file(BlockDriverState *bs,
1518                                         const char *backing_file,
1519                                         const char *backing_fmt)
1520 {
1521     BDRVQEDState *s = bs->opaque;
1522     QEDHeader new_header, le_header;
1523     void *buffer;
1524     size_t buffer_len, backing_file_len;
1525     int ret;
1526 
1527     /* Refuse to set backing filename if unknown compat feature bits are
1528      * active.  If the image uses an unknown compat feature then we may not
1529      * know the layout of data following the header structure and cannot safely
1530      * add a new string.
1531      */
1532     if (backing_file && (s->header.compat_features &
1533                          ~QED_COMPAT_FEATURE_MASK)) {
1534         return -ENOTSUP;
1535     }
1536 
1537     memcpy(&new_header, &s->header, sizeof(new_header));
1538 
1539     new_header.features &= ~(QED_F_BACKING_FILE |
1540                              QED_F_BACKING_FORMAT_NO_PROBE);
1541 
1542     /* Adjust feature flags */
1543     if (backing_file) {
1544         new_header.features |= QED_F_BACKING_FILE;
1545 
1546         if (qed_fmt_is_raw(backing_fmt)) {
1547             new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
1548         }
1549     }
1550 
1551     /* Calculate new header size */
1552     backing_file_len = 0;
1553 
1554     if (backing_file) {
1555         backing_file_len = strlen(backing_file);
1556     }
1557 
1558     buffer_len = sizeof(new_header);
1559     new_header.backing_filename_offset = buffer_len;
1560     new_header.backing_filename_size = backing_file_len;
1561     buffer_len += backing_file_len;
1562 
1563     /* Make sure we can rewrite header without failing */
1564     if (buffer_len > new_header.header_size * new_header.cluster_size) {
1565         return -ENOSPC;
1566     }
1567 
1568     /* Prepare new header */
1569     buffer = g_malloc(buffer_len);
1570 
1571     qed_header_cpu_to_le(&new_header, &le_header);
1572     memcpy(buffer, &le_header, sizeof(le_header));
1573     buffer_len = sizeof(le_header);
1574 
1575     if (backing_file) {
1576         memcpy(buffer + buffer_len, backing_file, backing_file_len);
1577         buffer_len += backing_file_len;
1578     }
1579 
1580     /* Write new header */
1581     ret = bdrv_pwrite_sync(bs->file->bs, 0, buffer, buffer_len);
1582     g_free(buffer);
1583     if (ret == 0) {
1584         memcpy(&s->header, &new_header, sizeof(new_header));
1585     }
1586     return ret;
1587 }
1588 
1589 static void bdrv_qed_invalidate_cache(BlockDriverState *bs, Error **errp)
1590 {
1591     BDRVQEDState *s = bs->opaque;
1592     Error *local_err = NULL;
1593     int ret;
1594 
1595     bdrv_qed_close(bs);
1596 
1597     memset(s, 0, sizeof(BDRVQEDState));
1598     ret = bdrv_qed_open(bs, NULL, bs->open_flags, &local_err);
1599     if (local_err) {
1600         error_propagate(errp, local_err);
1601         error_prepend(errp, "Could not reopen qed layer: ");
1602         return;
1603     } else if (ret < 0) {
1604         error_setg_errno(errp, -ret, "Could not reopen qed layer");
1605         return;
1606     }
1607 }
1608 
1609 static int bdrv_qed_check(BlockDriverState *bs, BdrvCheckResult *result,
1610                           BdrvCheckMode fix)
1611 {
1612     BDRVQEDState *s = bs->opaque;
1613 
1614     return qed_check(s, result, !!fix);
1615 }
1616 
1617 static QemuOptsList qed_create_opts = {
1618     .name = "qed-create-opts",
1619     .head = QTAILQ_HEAD_INITIALIZER(qed_create_opts.head),
1620     .desc = {
1621         {
1622             .name = BLOCK_OPT_SIZE,
1623             .type = QEMU_OPT_SIZE,
1624             .help = "Virtual disk size"
1625         },
1626         {
1627             .name = BLOCK_OPT_BACKING_FILE,
1628             .type = QEMU_OPT_STRING,
1629             .help = "File name of a base image"
1630         },
1631         {
1632             .name = BLOCK_OPT_BACKING_FMT,
1633             .type = QEMU_OPT_STRING,
1634             .help = "Image format of the base image"
1635         },
1636         {
1637             .name = BLOCK_OPT_CLUSTER_SIZE,
1638             .type = QEMU_OPT_SIZE,
1639             .help = "Cluster size (in bytes)",
1640             .def_value_str = stringify(QED_DEFAULT_CLUSTER_SIZE)
1641         },
1642         {
1643             .name = BLOCK_OPT_TABLE_SIZE,
1644             .type = QEMU_OPT_SIZE,
1645             .help = "L1/L2 table size (in clusters)"
1646         },
1647         { /* end of list */ }
1648     }
1649 };
1650 
1651 static BlockDriver bdrv_qed = {
1652     .format_name              = "qed",
1653     .instance_size            = sizeof(BDRVQEDState),
1654     .create_opts              = &qed_create_opts,
1655     .supports_backing         = true,
1656 
1657     .bdrv_probe               = bdrv_qed_probe,
1658     .bdrv_open                = bdrv_qed_open,
1659     .bdrv_close               = bdrv_qed_close,
1660     .bdrv_reopen_prepare      = bdrv_qed_reopen_prepare,
1661     .bdrv_create              = bdrv_qed_create,
1662     .bdrv_has_zero_init       = bdrv_has_zero_init_1,
1663     .bdrv_co_get_block_status = bdrv_qed_co_get_block_status,
1664     .bdrv_aio_readv           = bdrv_qed_aio_readv,
1665     .bdrv_aio_writev          = bdrv_qed_aio_writev,
1666     .bdrv_co_write_zeroes     = bdrv_qed_co_write_zeroes,
1667     .bdrv_truncate            = bdrv_qed_truncate,
1668     .bdrv_getlength           = bdrv_qed_getlength,
1669     .bdrv_get_info            = bdrv_qed_get_info,
1670     .bdrv_refresh_limits      = bdrv_qed_refresh_limits,
1671     .bdrv_change_backing_file = bdrv_qed_change_backing_file,
1672     .bdrv_invalidate_cache    = bdrv_qed_invalidate_cache,
1673     .bdrv_check               = bdrv_qed_check,
1674     .bdrv_detach_aio_context  = bdrv_qed_detach_aio_context,
1675     .bdrv_attach_aio_context  = bdrv_qed_attach_aio_context,
1676 };
1677 
1678 static void bdrv_qed_init(void)
1679 {
1680     bdrv_register(&bdrv_qed);
1681 }
1682 
1683 block_init(bdrv_qed_init);
1684