xref: /openbmc/qemu/block/qed.c (revision 469b046e)
1 /*
2  * QEMU Enhanced Disk Format
3  *
4  * Copyright IBM, Corp. 2010
5  *
6  * Authors:
7  *  Stefan Hajnoczi   <stefanha@linux.vnet.ibm.com>
8  *  Anthony Liguori   <aliguori@us.ibm.com>
9  *
10  * This work is licensed under the terms of the GNU LGPL, version 2 or later.
11  * See the COPYING.LIB file in the top-level directory.
12  *
13  */
14 
15 #include "qemu/timer.h"
16 #include "trace.h"
17 #include "qed.h"
18 #include "qapi/qmp/qerror.h"
19 #include "migration/migration.h"
20 
21 static void qed_aio_cancel(BlockDriverAIOCB *blockacb)
22 {
23     QEDAIOCB *acb = (QEDAIOCB *)blockacb;
24     AioContext *aio_context = bdrv_get_aio_context(blockacb->bs);
25     bool finished = false;
26 
27     /* Wait for the request to finish */
28     acb->finished = &finished;
29     while (!finished) {
30         aio_poll(aio_context, true);
31     }
32 }
33 
34 static const AIOCBInfo qed_aiocb_info = {
35     .aiocb_size         = sizeof(QEDAIOCB),
36     .cancel             = qed_aio_cancel,
37 };
38 
39 static int bdrv_qed_probe(const uint8_t *buf, int buf_size,
40                           const char *filename)
41 {
42     const QEDHeader *header = (const QEDHeader *)buf;
43 
44     if (buf_size < sizeof(*header)) {
45         return 0;
46     }
47     if (le32_to_cpu(header->magic) != QED_MAGIC) {
48         return 0;
49     }
50     return 100;
51 }
52 
53 /**
54  * Check whether an image format is raw
55  *
56  * @fmt:    Backing file format, may be NULL
57  */
58 static bool qed_fmt_is_raw(const char *fmt)
59 {
60     return fmt && strcmp(fmt, "raw") == 0;
61 }
62 
63 static void qed_header_le_to_cpu(const QEDHeader *le, QEDHeader *cpu)
64 {
65     cpu->magic = le32_to_cpu(le->magic);
66     cpu->cluster_size = le32_to_cpu(le->cluster_size);
67     cpu->table_size = le32_to_cpu(le->table_size);
68     cpu->header_size = le32_to_cpu(le->header_size);
69     cpu->features = le64_to_cpu(le->features);
70     cpu->compat_features = le64_to_cpu(le->compat_features);
71     cpu->autoclear_features = le64_to_cpu(le->autoclear_features);
72     cpu->l1_table_offset = le64_to_cpu(le->l1_table_offset);
73     cpu->image_size = le64_to_cpu(le->image_size);
74     cpu->backing_filename_offset = le32_to_cpu(le->backing_filename_offset);
75     cpu->backing_filename_size = le32_to_cpu(le->backing_filename_size);
76 }
77 
78 static void qed_header_cpu_to_le(const QEDHeader *cpu, QEDHeader *le)
79 {
80     le->magic = cpu_to_le32(cpu->magic);
81     le->cluster_size = cpu_to_le32(cpu->cluster_size);
82     le->table_size = cpu_to_le32(cpu->table_size);
83     le->header_size = cpu_to_le32(cpu->header_size);
84     le->features = cpu_to_le64(cpu->features);
85     le->compat_features = cpu_to_le64(cpu->compat_features);
86     le->autoclear_features = cpu_to_le64(cpu->autoclear_features);
87     le->l1_table_offset = cpu_to_le64(cpu->l1_table_offset);
88     le->image_size = cpu_to_le64(cpu->image_size);
89     le->backing_filename_offset = cpu_to_le32(cpu->backing_filename_offset);
90     le->backing_filename_size = cpu_to_le32(cpu->backing_filename_size);
91 }
92 
93 int qed_write_header_sync(BDRVQEDState *s)
94 {
95     QEDHeader le;
96     int ret;
97 
98     qed_header_cpu_to_le(&s->header, &le);
99     ret = bdrv_pwrite(s->bs->file, 0, &le, sizeof(le));
100     if (ret != sizeof(le)) {
101         return ret;
102     }
103     return 0;
104 }
105 
106 typedef struct {
107     GenericCB gencb;
108     BDRVQEDState *s;
109     struct iovec iov;
110     QEMUIOVector qiov;
111     int nsectors;
112     uint8_t *buf;
113 } QEDWriteHeaderCB;
114 
115 static void qed_write_header_cb(void *opaque, int ret)
116 {
117     QEDWriteHeaderCB *write_header_cb = opaque;
118 
119     qemu_vfree(write_header_cb->buf);
120     gencb_complete(write_header_cb, ret);
121 }
122 
123 static void qed_write_header_read_cb(void *opaque, int ret)
124 {
125     QEDWriteHeaderCB *write_header_cb = opaque;
126     BDRVQEDState *s = write_header_cb->s;
127 
128     if (ret) {
129         qed_write_header_cb(write_header_cb, ret);
130         return;
131     }
132 
133     /* Update header */
134     qed_header_cpu_to_le(&s->header, (QEDHeader *)write_header_cb->buf);
135 
136     bdrv_aio_writev(s->bs->file, 0, &write_header_cb->qiov,
137                     write_header_cb->nsectors, qed_write_header_cb,
138                     write_header_cb);
139 }
140 
141 /**
142  * Update header in-place (does not rewrite backing filename or other strings)
143  *
144  * This function only updates known header fields in-place and does not affect
145  * extra data after the QED header.
146  */
147 static void qed_write_header(BDRVQEDState *s, BlockDriverCompletionFunc cb,
148                              void *opaque)
149 {
150     /* We must write full sectors for O_DIRECT but cannot necessarily generate
151      * the data following the header if an unrecognized compat feature is
152      * active.  Therefore, first read the sectors containing the header, update
153      * them, and write back.
154      */
155 
156     int nsectors = (sizeof(QEDHeader) + BDRV_SECTOR_SIZE - 1) /
157                    BDRV_SECTOR_SIZE;
158     size_t len = nsectors * BDRV_SECTOR_SIZE;
159     QEDWriteHeaderCB *write_header_cb = gencb_alloc(sizeof(*write_header_cb),
160                                                     cb, opaque);
161 
162     write_header_cb->s = s;
163     write_header_cb->nsectors = nsectors;
164     write_header_cb->buf = qemu_blockalign(s->bs, len);
165     write_header_cb->iov.iov_base = write_header_cb->buf;
166     write_header_cb->iov.iov_len = len;
167     qemu_iovec_init_external(&write_header_cb->qiov, &write_header_cb->iov, 1);
168 
169     bdrv_aio_readv(s->bs->file, 0, &write_header_cb->qiov, nsectors,
170                    qed_write_header_read_cb, write_header_cb);
171 }
172 
173 static uint64_t qed_max_image_size(uint32_t cluster_size, uint32_t table_size)
174 {
175     uint64_t table_entries;
176     uint64_t l2_size;
177 
178     table_entries = (table_size * cluster_size) / sizeof(uint64_t);
179     l2_size = table_entries * cluster_size;
180 
181     return l2_size * table_entries;
182 }
183 
184 static bool qed_is_cluster_size_valid(uint32_t cluster_size)
185 {
186     if (cluster_size < QED_MIN_CLUSTER_SIZE ||
187         cluster_size > QED_MAX_CLUSTER_SIZE) {
188         return false;
189     }
190     if (cluster_size & (cluster_size - 1)) {
191         return false; /* not power of 2 */
192     }
193     return true;
194 }
195 
196 static bool qed_is_table_size_valid(uint32_t table_size)
197 {
198     if (table_size < QED_MIN_TABLE_SIZE ||
199         table_size > QED_MAX_TABLE_SIZE) {
200         return false;
201     }
202     if (table_size & (table_size - 1)) {
203         return false; /* not power of 2 */
204     }
205     return true;
206 }
207 
208 static bool qed_is_image_size_valid(uint64_t image_size, uint32_t cluster_size,
209                                     uint32_t table_size)
210 {
211     if (image_size % BDRV_SECTOR_SIZE != 0) {
212         return false; /* not multiple of sector size */
213     }
214     if (image_size > qed_max_image_size(cluster_size, table_size)) {
215         return false; /* image is too large */
216     }
217     return true;
218 }
219 
220 /**
221  * Read a string of known length from the image file
222  *
223  * @file:       Image file
224  * @offset:     File offset to start of string, in bytes
225  * @n:          String length in bytes
226  * @buf:        Destination buffer
227  * @buflen:     Destination buffer length in bytes
228  * @ret:        0 on success, -errno on failure
229  *
230  * The string is NUL-terminated.
231  */
232 static int qed_read_string(BlockDriverState *file, uint64_t offset, size_t n,
233                            char *buf, size_t buflen)
234 {
235     int ret;
236     if (n >= buflen) {
237         return -EINVAL;
238     }
239     ret = bdrv_pread(file, offset, buf, n);
240     if (ret < 0) {
241         return ret;
242     }
243     buf[n] = '\0';
244     return 0;
245 }
246 
247 /**
248  * Allocate new clusters
249  *
250  * @s:          QED state
251  * @n:          Number of contiguous clusters to allocate
252  * @ret:        Offset of first allocated cluster
253  *
254  * This function only produces the offset where the new clusters should be
255  * written.  It updates BDRVQEDState but does not make any changes to the image
256  * file.
257  */
258 static uint64_t qed_alloc_clusters(BDRVQEDState *s, unsigned int n)
259 {
260     uint64_t offset = s->file_size;
261     s->file_size += n * s->header.cluster_size;
262     return offset;
263 }
264 
265 QEDTable *qed_alloc_table(BDRVQEDState *s)
266 {
267     /* Honor O_DIRECT memory alignment requirements */
268     return qemu_blockalign(s->bs,
269                            s->header.cluster_size * s->header.table_size);
270 }
271 
272 /**
273  * Allocate a new zeroed L2 table
274  */
275 static CachedL2Table *qed_new_l2_table(BDRVQEDState *s)
276 {
277     CachedL2Table *l2_table = qed_alloc_l2_cache_entry(&s->l2_cache);
278 
279     l2_table->table = qed_alloc_table(s);
280     l2_table->offset = qed_alloc_clusters(s, s->header.table_size);
281 
282     memset(l2_table->table->offsets, 0,
283            s->header.cluster_size * s->header.table_size);
284     return l2_table;
285 }
286 
287 static void qed_aio_next_io(void *opaque, int ret);
288 
289 static void qed_plug_allocating_write_reqs(BDRVQEDState *s)
290 {
291     assert(!s->allocating_write_reqs_plugged);
292 
293     s->allocating_write_reqs_plugged = true;
294 }
295 
296 static void qed_unplug_allocating_write_reqs(BDRVQEDState *s)
297 {
298     QEDAIOCB *acb;
299 
300     assert(s->allocating_write_reqs_plugged);
301 
302     s->allocating_write_reqs_plugged = false;
303 
304     acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs);
305     if (acb) {
306         qed_aio_next_io(acb, 0);
307     }
308 }
309 
310 static void qed_finish_clear_need_check(void *opaque, int ret)
311 {
312     /* Do nothing */
313 }
314 
315 static void qed_flush_after_clear_need_check(void *opaque, int ret)
316 {
317     BDRVQEDState *s = opaque;
318 
319     bdrv_aio_flush(s->bs, qed_finish_clear_need_check, s);
320 
321     /* No need to wait until flush completes */
322     qed_unplug_allocating_write_reqs(s);
323 }
324 
325 static void qed_clear_need_check(void *opaque, int ret)
326 {
327     BDRVQEDState *s = opaque;
328 
329     if (ret) {
330         qed_unplug_allocating_write_reqs(s);
331         return;
332     }
333 
334     s->header.features &= ~QED_F_NEED_CHECK;
335     qed_write_header(s, qed_flush_after_clear_need_check, s);
336 }
337 
338 static void qed_need_check_timer_cb(void *opaque)
339 {
340     BDRVQEDState *s = opaque;
341 
342     /* The timer should only fire when allocating writes have drained */
343     assert(!QSIMPLEQ_FIRST(&s->allocating_write_reqs));
344 
345     trace_qed_need_check_timer_cb(s);
346 
347     qed_plug_allocating_write_reqs(s);
348 
349     /* Ensure writes are on disk before clearing flag */
350     bdrv_aio_flush(s->bs, qed_clear_need_check, s);
351 }
352 
353 static void qed_start_need_check_timer(BDRVQEDState *s)
354 {
355     trace_qed_start_need_check_timer(s);
356 
357     /* Use QEMU_CLOCK_VIRTUAL so we don't alter the image file while suspended for
358      * migration.
359      */
360     timer_mod(s->need_check_timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
361                    get_ticks_per_sec() * QED_NEED_CHECK_TIMEOUT);
362 }
363 
364 /* It's okay to call this multiple times or when no timer is started */
365 static void qed_cancel_need_check_timer(BDRVQEDState *s)
366 {
367     trace_qed_cancel_need_check_timer(s);
368     timer_del(s->need_check_timer);
369 }
370 
371 static void bdrv_qed_rebind(BlockDriverState *bs)
372 {
373     BDRVQEDState *s = bs->opaque;
374     s->bs = bs;
375 }
376 
377 static void bdrv_qed_detach_aio_context(BlockDriverState *bs)
378 {
379     BDRVQEDState *s = bs->opaque;
380 
381     qed_cancel_need_check_timer(s);
382     timer_free(s->need_check_timer);
383 }
384 
385 static void bdrv_qed_attach_aio_context(BlockDriverState *bs,
386                                         AioContext *new_context)
387 {
388     BDRVQEDState *s = bs->opaque;
389 
390     s->need_check_timer = aio_timer_new(new_context,
391                                         QEMU_CLOCK_VIRTUAL, SCALE_NS,
392                                         qed_need_check_timer_cb, s);
393     if (s->header.features & QED_F_NEED_CHECK) {
394         qed_start_need_check_timer(s);
395     }
396 }
397 
398 static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags,
399                          Error **errp)
400 {
401     BDRVQEDState *s = bs->opaque;
402     QEDHeader le_header;
403     int64_t file_size;
404     int ret;
405 
406     s->bs = bs;
407     QSIMPLEQ_INIT(&s->allocating_write_reqs);
408 
409     ret = bdrv_pread(bs->file, 0, &le_header, sizeof(le_header));
410     if (ret < 0) {
411         return ret;
412     }
413     qed_header_le_to_cpu(&le_header, &s->header);
414 
415     if (s->header.magic != QED_MAGIC) {
416         error_setg(errp, "Image not in QED format");
417         return -EINVAL;
418     }
419     if (s->header.features & ~QED_FEATURE_MASK) {
420         /* image uses unsupported feature bits */
421         char buf[64];
422         snprintf(buf, sizeof(buf), "%" PRIx64,
423             s->header.features & ~QED_FEATURE_MASK);
424         error_set(errp, QERR_UNKNOWN_BLOCK_FORMAT_FEATURE,
425             bs->device_name, "QED", buf);
426         return -ENOTSUP;
427     }
428     if (!qed_is_cluster_size_valid(s->header.cluster_size)) {
429         return -EINVAL;
430     }
431 
432     /* Round down file size to the last cluster */
433     file_size = bdrv_getlength(bs->file);
434     if (file_size < 0) {
435         return file_size;
436     }
437     s->file_size = qed_start_of_cluster(s, file_size);
438 
439     if (!qed_is_table_size_valid(s->header.table_size)) {
440         return -EINVAL;
441     }
442     if (!qed_is_image_size_valid(s->header.image_size,
443                                  s->header.cluster_size,
444                                  s->header.table_size)) {
445         return -EINVAL;
446     }
447     if (!qed_check_table_offset(s, s->header.l1_table_offset)) {
448         return -EINVAL;
449     }
450 
451     s->table_nelems = (s->header.cluster_size * s->header.table_size) /
452                       sizeof(uint64_t);
453     s->l2_shift = ffs(s->header.cluster_size) - 1;
454     s->l2_mask = s->table_nelems - 1;
455     s->l1_shift = s->l2_shift + ffs(s->table_nelems) - 1;
456 
457     if ((s->header.features & QED_F_BACKING_FILE)) {
458         if ((uint64_t)s->header.backing_filename_offset +
459             s->header.backing_filename_size >
460             s->header.cluster_size * s->header.header_size) {
461             return -EINVAL;
462         }
463 
464         ret = qed_read_string(bs->file, s->header.backing_filename_offset,
465                               s->header.backing_filename_size, bs->backing_file,
466                               sizeof(bs->backing_file));
467         if (ret < 0) {
468             return ret;
469         }
470 
471         if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) {
472             pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw");
473         }
474     }
475 
476     /* Reset unknown autoclear feature bits.  This is a backwards
477      * compatibility mechanism that allows images to be opened by older
478      * programs, which "knock out" unknown feature bits.  When an image is
479      * opened by a newer program again it can detect that the autoclear
480      * feature is no longer valid.
481      */
482     if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 &&
483         !bdrv_is_read_only(bs->file) && !(flags & BDRV_O_INCOMING)) {
484         s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK;
485 
486         ret = qed_write_header_sync(s);
487         if (ret) {
488             return ret;
489         }
490 
491         /* From here on only known autoclear feature bits are valid */
492         bdrv_flush(bs->file);
493     }
494 
495     s->l1_table = qed_alloc_table(s);
496     qed_init_l2_cache(&s->l2_cache);
497 
498     ret = qed_read_l1_table_sync(s);
499     if (ret) {
500         goto out;
501     }
502 
503     /* If image was not closed cleanly, check consistency */
504     if (!(flags & BDRV_O_CHECK) && (s->header.features & QED_F_NEED_CHECK)) {
505         /* Read-only images cannot be fixed.  There is no risk of corruption
506          * since write operations are not possible.  Therefore, allow
507          * potentially inconsistent images to be opened read-only.  This can
508          * aid data recovery from an otherwise inconsistent image.
509          */
510         if (!bdrv_is_read_only(bs->file) &&
511             !(flags & BDRV_O_INCOMING)) {
512             BdrvCheckResult result = {0};
513 
514             ret = qed_check(s, &result, true);
515             if (ret) {
516                 goto out;
517             }
518         }
519     }
520 
521     bdrv_qed_attach_aio_context(bs, bdrv_get_aio_context(bs));
522 
523 out:
524     if (ret) {
525         qed_free_l2_cache(&s->l2_cache);
526         qemu_vfree(s->l1_table);
527     }
528     return ret;
529 }
530 
531 static int bdrv_qed_refresh_limits(BlockDriverState *bs)
532 {
533     BDRVQEDState *s = bs->opaque;
534 
535     bs->bl.write_zeroes_alignment = s->header.cluster_size >> BDRV_SECTOR_BITS;
536 
537     return 0;
538 }
539 
540 /* We have nothing to do for QED reopen, stubs just return
541  * success */
542 static int bdrv_qed_reopen_prepare(BDRVReopenState *state,
543                                    BlockReopenQueue *queue, Error **errp)
544 {
545     return 0;
546 }
547 
548 static void bdrv_qed_close(BlockDriverState *bs)
549 {
550     BDRVQEDState *s = bs->opaque;
551 
552     bdrv_qed_detach_aio_context(bs);
553 
554     /* Ensure writes reach stable storage */
555     bdrv_flush(bs->file);
556 
557     /* Clean shutdown, no check required on next open */
558     if (s->header.features & QED_F_NEED_CHECK) {
559         s->header.features &= ~QED_F_NEED_CHECK;
560         qed_write_header_sync(s);
561     }
562 
563     qed_free_l2_cache(&s->l2_cache);
564     qemu_vfree(s->l1_table);
565 }
566 
567 static int qed_create(const char *filename, uint32_t cluster_size,
568                       uint64_t image_size, uint32_t table_size,
569                       const char *backing_file, const char *backing_fmt,
570                       QemuOpts *opts, Error **errp)
571 {
572     QEDHeader header = {
573         .magic = QED_MAGIC,
574         .cluster_size = cluster_size,
575         .table_size = table_size,
576         .header_size = 1,
577         .features = 0,
578         .compat_features = 0,
579         .l1_table_offset = cluster_size,
580         .image_size = image_size,
581     };
582     QEDHeader le_header;
583     uint8_t *l1_table = NULL;
584     size_t l1_size = header.cluster_size * header.table_size;
585     Error *local_err = NULL;
586     int ret = 0;
587     BlockDriverState *bs;
588 
589     ret = bdrv_create_file(filename, opts, &local_err);
590     if (ret < 0) {
591         error_propagate(errp, local_err);
592         return ret;
593     }
594 
595     bs = NULL;
596     ret = bdrv_open(&bs, filename, NULL, NULL,
597                     BDRV_O_RDWR | BDRV_O_CACHE_WB | BDRV_O_PROTOCOL, NULL,
598                     &local_err);
599     if (ret < 0) {
600         error_propagate(errp, local_err);
601         return ret;
602     }
603 
604     /* File must start empty and grow, check truncate is supported */
605     ret = bdrv_truncate(bs, 0);
606     if (ret < 0) {
607         goto out;
608     }
609 
610     if (backing_file) {
611         header.features |= QED_F_BACKING_FILE;
612         header.backing_filename_offset = sizeof(le_header);
613         header.backing_filename_size = strlen(backing_file);
614 
615         if (qed_fmt_is_raw(backing_fmt)) {
616             header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
617         }
618     }
619 
620     qed_header_cpu_to_le(&header, &le_header);
621     ret = bdrv_pwrite(bs, 0, &le_header, sizeof(le_header));
622     if (ret < 0) {
623         goto out;
624     }
625     ret = bdrv_pwrite(bs, sizeof(le_header), backing_file,
626                       header.backing_filename_size);
627     if (ret < 0) {
628         goto out;
629     }
630 
631     l1_table = g_malloc0(l1_size);
632     ret = bdrv_pwrite(bs, header.l1_table_offset, l1_table, l1_size);
633     if (ret < 0) {
634         goto out;
635     }
636 
637     ret = 0; /* success */
638 out:
639     g_free(l1_table);
640     bdrv_unref(bs);
641     return ret;
642 }
643 
644 static int bdrv_qed_create(const char *filename, QemuOpts *opts, Error **errp)
645 {
646     uint64_t image_size = 0;
647     uint32_t cluster_size = QED_DEFAULT_CLUSTER_SIZE;
648     uint32_t table_size = QED_DEFAULT_TABLE_SIZE;
649     char *backing_file = NULL;
650     char *backing_fmt = NULL;
651     int ret;
652 
653     image_size = qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0);
654     backing_file = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FILE);
655     backing_fmt = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FMT);
656     cluster_size = qemu_opt_get_size_del(opts,
657                                          BLOCK_OPT_CLUSTER_SIZE,
658                                          QED_DEFAULT_CLUSTER_SIZE);
659     table_size = qemu_opt_get_size_del(opts, BLOCK_OPT_TABLE_SIZE,
660                                        QED_DEFAULT_TABLE_SIZE);
661 
662     if (!qed_is_cluster_size_valid(cluster_size)) {
663         error_setg(errp, "QED cluster size must be within range [%u, %u] "
664                          "and power of 2",
665                    QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE);
666         ret = -EINVAL;
667         goto finish;
668     }
669     if (!qed_is_table_size_valid(table_size)) {
670         error_setg(errp, "QED table size must be within range [%u, %u] "
671                          "and power of 2",
672                    QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE);
673         ret = -EINVAL;
674         goto finish;
675     }
676     if (!qed_is_image_size_valid(image_size, cluster_size, table_size)) {
677         error_setg(errp, "QED image size must be a non-zero multiple of "
678                          "cluster size and less than %" PRIu64 " bytes",
679                    qed_max_image_size(cluster_size, table_size));
680         ret = -EINVAL;
681         goto finish;
682     }
683 
684     ret = qed_create(filename, cluster_size, image_size, table_size,
685                      backing_file, backing_fmt, opts, errp);
686 
687 finish:
688     g_free(backing_file);
689     g_free(backing_fmt);
690     return ret;
691 }
692 
693 typedef struct {
694     BlockDriverState *bs;
695     Coroutine *co;
696     uint64_t pos;
697     int64_t status;
698     int *pnum;
699 } QEDIsAllocatedCB;
700 
701 static void qed_is_allocated_cb(void *opaque, int ret, uint64_t offset, size_t len)
702 {
703     QEDIsAllocatedCB *cb = opaque;
704     BDRVQEDState *s = cb->bs->opaque;
705     *cb->pnum = len / BDRV_SECTOR_SIZE;
706     switch (ret) {
707     case QED_CLUSTER_FOUND:
708         offset |= qed_offset_into_cluster(s, cb->pos);
709         cb->status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID | offset;
710         break;
711     case QED_CLUSTER_ZERO:
712         cb->status = BDRV_BLOCK_ZERO;
713         break;
714     case QED_CLUSTER_L2:
715     case QED_CLUSTER_L1:
716         cb->status = 0;
717         break;
718     default:
719         assert(ret < 0);
720         cb->status = ret;
721         break;
722     }
723 
724     if (cb->co) {
725         qemu_coroutine_enter(cb->co, NULL);
726     }
727 }
728 
729 static int64_t coroutine_fn bdrv_qed_co_get_block_status(BlockDriverState *bs,
730                                                  int64_t sector_num,
731                                                  int nb_sectors, int *pnum)
732 {
733     BDRVQEDState *s = bs->opaque;
734     size_t len = (size_t)nb_sectors * BDRV_SECTOR_SIZE;
735     QEDIsAllocatedCB cb = {
736         .bs = bs,
737         .pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE,
738         .status = BDRV_BLOCK_OFFSET_MASK,
739         .pnum = pnum,
740     };
741     QEDRequest request = { .l2_table = NULL };
742 
743     qed_find_cluster(s, &request, cb.pos, len, qed_is_allocated_cb, &cb);
744 
745     /* Now sleep if the callback wasn't invoked immediately */
746     while (cb.status == BDRV_BLOCK_OFFSET_MASK) {
747         cb.co = qemu_coroutine_self();
748         qemu_coroutine_yield();
749     }
750 
751     qed_unref_l2_cache_entry(request.l2_table);
752 
753     return cb.status;
754 }
755 
756 static BDRVQEDState *acb_to_s(QEDAIOCB *acb)
757 {
758     return acb->common.bs->opaque;
759 }
760 
761 /**
762  * Read from the backing file or zero-fill if no backing file
763  *
764  * @s:              QED state
765  * @pos:            Byte position in device
766  * @qiov:           Destination I/O vector
767  * @backing_qiov:   Possibly shortened copy of qiov, to be allocated here
768  * @cb:             Completion function
769  * @opaque:         User data for completion function
770  *
771  * This function reads qiov->size bytes starting at pos from the backing file.
772  * If there is no backing file then zeroes are read.
773  */
774 static void qed_read_backing_file(BDRVQEDState *s, uint64_t pos,
775                                   QEMUIOVector *qiov,
776                                   QEMUIOVector **backing_qiov,
777                                   BlockDriverCompletionFunc *cb, void *opaque)
778 {
779     uint64_t backing_length = 0;
780     size_t size;
781 
782     /* If there is a backing file, get its length.  Treat the absence of a
783      * backing file like a zero length backing file.
784      */
785     if (s->bs->backing_hd) {
786         int64_t l = bdrv_getlength(s->bs->backing_hd);
787         if (l < 0) {
788             cb(opaque, l);
789             return;
790         }
791         backing_length = l;
792     }
793 
794     /* Zero all sectors if reading beyond the end of the backing file */
795     if (pos >= backing_length ||
796         pos + qiov->size > backing_length) {
797         qemu_iovec_memset(qiov, 0, 0, qiov->size);
798     }
799 
800     /* Complete now if there are no backing file sectors to read */
801     if (pos >= backing_length) {
802         cb(opaque, 0);
803         return;
804     }
805 
806     /* If the read straddles the end of the backing file, shorten it */
807     size = MIN((uint64_t)backing_length - pos, qiov->size);
808 
809     assert(*backing_qiov == NULL);
810     *backing_qiov = g_new(QEMUIOVector, 1);
811     qemu_iovec_init(*backing_qiov, qiov->niov);
812     qemu_iovec_concat(*backing_qiov, qiov, 0, size);
813 
814     BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO);
815     bdrv_aio_readv(s->bs->backing_hd, pos / BDRV_SECTOR_SIZE,
816                    *backing_qiov, size / BDRV_SECTOR_SIZE, cb, opaque);
817 }
818 
819 typedef struct {
820     GenericCB gencb;
821     BDRVQEDState *s;
822     QEMUIOVector qiov;
823     QEMUIOVector *backing_qiov;
824     struct iovec iov;
825     uint64_t offset;
826 } CopyFromBackingFileCB;
827 
828 static void qed_copy_from_backing_file_cb(void *opaque, int ret)
829 {
830     CopyFromBackingFileCB *copy_cb = opaque;
831     qemu_vfree(copy_cb->iov.iov_base);
832     gencb_complete(&copy_cb->gencb, ret);
833 }
834 
835 static void qed_copy_from_backing_file_write(void *opaque, int ret)
836 {
837     CopyFromBackingFileCB *copy_cb = opaque;
838     BDRVQEDState *s = copy_cb->s;
839 
840     if (copy_cb->backing_qiov) {
841         qemu_iovec_destroy(copy_cb->backing_qiov);
842         g_free(copy_cb->backing_qiov);
843         copy_cb->backing_qiov = NULL;
844     }
845 
846     if (ret) {
847         qed_copy_from_backing_file_cb(copy_cb, ret);
848         return;
849     }
850 
851     BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE);
852     bdrv_aio_writev(s->bs->file, copy_cb->offset / BDRV_SECTOR_SIZE,
853                     &copy_cb->qiov, copy_cb->qiov.size / BDRV_SECTOR_SIZE,
854                     qed_copy_from_backing_file_cb, copy_cb);
855 }
856 
857 /**
858  * Copy data from backing file into the image
859  *
860  * @s:          QED state
861  * @pos:        Byte position in device
862  * @len:        Number of bytes
863  * @offset:     Byte offset in image file
864  * @cb:         Completion function
865  * @opaque:     User data for completion function
866  */
867 static void qed_copy_from_backing_file(BDRVQEDState *s, uint64_t pos,
868                                        uint64_t len, uint64_t offset,
869                                        BlockDriverCompletionFunc *cb,
870                                        void *opaque)
871 {
872     CopyFromBackingFileCB *copy_cb;
873 
874     /* Skip copy entirely if there is no work to do */
875     if (len == 0) {
876         cb(opaque, 0);
877         return;
878     }
879 
880     copy_cb = gencb_alloc(sizeof(*copy_cb), cb, opaque);
881     copy_cb->s = s;
882     copy_cb->offset = offset;
883     copy_cb->backing_qiov = NULL;
884     copy_cb->iov.iov_base = qemu_blockalign(s->bs, len);
885     copy_cb->iov.iov_len = len;
886     qemu_iovec_init_external(&copy_cb->qiov, &copy_cb->iov, 1);
887 
888     qed_read_backing_file(s, pos, &copy_cb->qiov, &copy_cb->backing_qiov,
889                           qed_copy_from_backing_file_write, copy_cb);
890 }
891 
892 /**
893  * Link one or more contiguous clusters into a table
894  *
895  * @s:              QED state
896  * @table:          L2 table
897  * @index:          First cluster index
898  * @n:              Number of contiguous clusters
899  * @cluster:        First cluster offset
900  *
901  * The cluster offset may be an allocated byte offset in the image file, the
902  * zero cluster marker, or the unallocated cluster marker.
903  */
904 static void qed_update_l2_table(BDRVQEDState *s, QEDTable *table, int index,
905                                 unsigned int n, uint64_t cluster)
906 {
907     int i;
908     for (i = index; i < index + n; i++) {
909         table->offsets[i] = cluster;
910         if (!qed_offset_is_unalloc_cluster(cluster) &&
911             !qed_offset_is_zero_cluster(cluster)) {
912             cluster += s->header.cluster_size;
913         }
914     }
915 }
916 
917 static void qed_aio_complete_bh(void *opaque)
918 {
919     QEDAIOCB *acb = opaque;
920     BlockDriverCompletionFunc *cb = acb->common.cb;
921     void *user_opaque = acb->common.opaque;
922     int ret = acb->bh_ret;
923     bool *finished = acb->finished;
924 
925     qemu_bh_delete(acb->bh);
926     qemu_aio_release(acb);
927 
928     /* Invoke callback */
929     cb(user_opaque, ret);
930 
931     /* Signal cancel completion */
932     if (finished) {
933         *finished = true;
934     }
935 }
936 
937 static void qed_aio_complete(QEDAIOCB *acb, int ret)
938 {
939     BDRVQEDState *s = acb_to_s(acb);
940 
941     trace_qed_aio_complete(s, acb, ret);
942 
943     /* Free resources */
944     qemu_iovec_destroy(&acb->cur_qiov);
945     qed_unref_l2_cache_entry(acb->request.l2_table);
946 
947     /* Free the buffer we may have allocated for zero writes */
948     if (acb->flags & QED_AIOCB_ZERO) {
949         qemu_vfree(acb->qiov->iov[0].iov_base);
950         acb->qiov->iov[0].iov_base = NULL;
951     }
952 
953     /* Arrange for a bh to invoke the completion function */
954     acb->bh_ret = ret;
955     acb->bh = aio_bh_new(bdrv_get_aio_context(acb->common.bs),
956                          qed_aio_complete_bh, acb);
957     qemu_bh_schedule(acb->bh);
958 
959     /* Start next allocating write request waiting behind this one.  Note that
960      * requests enqueue themselves when they first hit an unallocated cluster
961      * but they wait until the entire request is finished before waking up the
962      * next request in the queue.  This ensures that we don't cycle through
963      * requests multiple times but rather finish one at a time completely.
964      */
965     if (acb == QSIMPLEQ_FIRST(&s->allocating_write_reqs)) {
966         QSIMPLEQ_REMOVE_HEAD(&s->allocating_write_reqs, next);
967         acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs);
968         if (acb) {
969             qed_aio_next_io(acb, 0);
970         } else if (s->header.features & QED_F_NEED_CHECK) {
971             qed_start_need_check_timer(s);
972         }
973     }
974 }
975 
976 /**
977  * Commit the current L2 table to the cache
978  */
979 static void qed_commit_l2_update(void *opaque, int ret)
980 {
981     QEDAIOCB *acb = opaque;
982     BDRVQEDState *s = acb_to_s(acb);
983     CachedL2Table *l2_table = acb->request.l2_table;
984     uint64_t l2_offset = l2_table->offset;
985 
986     qed_commit_l2_cache_entry(&s->l2_cache, l2_table);
987 
988     /* This is guaranteed to succeed because we just committed the entry to the
989      * cache.
990      */
991     acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset);
992     assert(acb->request.l2_table != NULL);
993 
994     qed_aio_next_io(opaque, ret);
995 }
996 
997 /**
998  * Update L1 table with new L2 table offset and write it out
999  */
1000 static void qed_aio_write_l1_update(void *opaque, int ret)
1001 {
1002     QEDAIOCB *acb = opaque;
1003     BDRVQEDState *s = acb_to_s(acb);
1004     int index;
1005 
1006     if (ret) {
1007         qed_aio_complete(acb, ret);
1008         return;
1009     }
1010 
1011     index = qed_l1_index(s, acb->cur_pos);
1012     s->l1_table->offsets[index] = acb->request.l2_table->offset;
1013 
1014     qed_write_l1_table(s, index, 1, qed_commit_l2_update, acb);
1015 }
1016 
1017 /**
1018  * Update L2 table with new cluster offsets and write them out
1019  */
1020 static void qed_aio_write_l2_update(QEDAIOCB *acb, int ret, uint64_t offset)
1021 {
1022     BDRVQEDState *s = acb_to_s(acb);
1023     bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1;
1024     int index;
1025 
1026     if (ret) {
1027         goto err;
1028     }
1029 
1030     if (need_alloc) {
1031         qed_unref_l2_cache_entry(acb->request.l2_table);
1032         acb->request.l2_table = qed_new_l2_table(s);
1033     }
1034 
1035     index = qed_l2_index(s, acb->cur_pos);
1036     qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters,
1037                          offset);
1038 
1039     if (need_alloc) {
1040         /* Write out the whole new L2 table */
1041         qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true,
1042                             qed_aio_write_l1_update, acb);
1043     } else {
1044         /* Write out only the updated part of the L2 table */
1045         qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters, false,
1046                             qed_aio_next_io, acb);
1047     }
1048     return;
1049 
1050 err:
1051     qed_aio_complete(acb, ret);
1052 }
1053 
1054 static void qed_aio_write_l2_update_cb(void *opaque, int ret)
1055 {
1056     QEDAIOCB *acb = opaque;
1057     qed_aio_write_l2_update(acb, ret, acb->cur_cluster);
1058 }
1059 
1060 /**
1061  * Flush new data clusters before updating the L2 table
1062  *
1063  * This flush is necessary when a backing file is in use.  A crash during an
1064  * allocating write could result in empty clusters in the image.  If the write
1065  * only touched a subregion of the cluster, then backing image sectors have
1066  * been lost in the untouched region.  The solution is to flush after writing a
1067  * new data cluster and before updating the L2 table.
1068  */
1069 static void qed_aio_write_flush_before_l2_update(void *opaque, int ret)
1070 {
1071     QEDAIOCB *acb = opaque;
1072     BDRVQEDState *s = acb_to_s(acb);
1073 
1074     if (!bdrv_aio_flush(s->bs->file, qed_aio_write_l2_update_cb, opaque)) {
1075         qed_aio_complete(acb, -EIO);
1076     }
1077 }
1078 
1079 /**
1080  * Write data to the image file
1081  */
1082 static void qed_aio_write_main(void *opaque, int ret)
1083 {
1084     QEDAIOCB *acb = opaque;
1085     BDRVQEDState *s = acb_to_s(acb);
1086     uint64_t offset = acb->cur_cluster +
1087                       qed_offset_into_cluster(s, acb->cur_pos);
1088     BlockDriverCompletionFunc *next_fn;
1089 
1090     trace_qed_aio_write_main(s, acb, ret, offset, acb->cur_qiov.size);
1091 
1092     if (ret) {
1093         qed_aio_complete(acb, ret);
1094         return;
1095     }
1096 
1097     if (acb->find_cluster_ret == QED_CLUSTER_FOUND) {
1098         next_fn = qed_aio_next_io;
1099     } else {
1100         if (s->bs->backing_hd) {
1101             next_fn = qed_aio_write_flush_before_l2_update;
1102         } else {
1103             next_fn = qed_aio_write_l2_update_cb;
1104         }
1105     }
1106 
1107     BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO);
1108     bdrv_aio_writev(s->bs->file, offset / BDRV_SECTOR_SIZE,
1109                     &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE,
1110                     next_fn, acb);
1111 }
1112 
1113 /**
1114  * Populate back untouched region of new data cluster
1115  */
1116 static void qed_aio_write_postfill(void *opaque, int ret)
1117 {
1118     QEDAIOCB *acb = opaque;
1119     BDRVQEDState *s = acb_to_s(acb);
1120     uint64_t start = acb->cur_pos + acb->cur_qiov.size;
1121     uint64_t len =
1122         qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start;
1123     uint64_t offset = acb->cur_cluster +
1124                       qed_offset_into_cluster(s, acb->cur_pos) +
1125                       acb->cur_qiov.size;
1126 
1127     if (ret) {
1128         qed_aio_complete(acb, ret);
1129         return;
1130     }
1131 
1132     trace_qed_aio_write_postfill(s, acb, start, len, offset);
1133     qed_copy_from_backing_file(s, start, len, offset,
1134                                 qed_aio_write_main, acb);
1135 }
1136 
1137 /**
1138  * Populate front untouched region of new data cluster
1139  */
1140 static void qed_aio_write_prefill(void *opaque, int ret)
1141 {
1142     QEDAIOCB *acb = opaque;
1143     BDRVQEDState *s = acb_to_s(acb);
1144     uint64_t start = qed_start_of_cluster(s, acb->cur_pos);
1145     uint64_t len = qed_offset_into_cluster(s, acb->cur_pos);
1146 
1147     trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster);
1148     qed_copy_from_backing_file(s, start, len, acb->cur_cluster,
1149                                 qed_aio_write_postfill, acb);
1150 }
1151 
1152 /**
1153  * Check if the QED_F_NEED_CHECK bit should be set during allocating write
1154  */
1155 static bool qed_should_set_need_check(BDRVQEDState *s)
1156 {
1157     /* The flush before L2 update path ensures consistency */
1158     if (s->bs->backing_hd) {
1159         return false;
1160     }
1161 
1162     return !(s->header.features & QED_F_NEED_CHECK);
1163 }
1164 
1165 static void qed_aio_write_zero_cluster(void *opaque, int ret)
1166 {
1167     QEDAIOCB *acb = opaque;
1168 
1169     if (ret) {
1170         qed_aio_complete(acb, ret);
1171         return;
1172     }
1173 
1174     qed_aio_write_l2_update(acb, 0, 1);
1175 }
1176 
1177 /**
1178  * Write new data cluster
1179  *
1180  * @acb:        Write request
1181  * @len:        Length in bytes
1182  *
1183  * This path is taken when writing to previously unallocated clusters.
1184  */
1185 static void qed_aio_write_alloc(QEDAIOCB *acb, size_t len)
1186 {
1187     BDRVQEDState *s = acb_to_s(acb);
1188     BlockDriverCompletionFunc *cb;
1189 
1190     /* Cancel timer when the first allocating request comes in */
1191     if (QSIMPLEQ_EMPTY(&s->allocating_write_reqs)) {
1192         qed_cancel_need_check_timer(s);
1193     }
1194 
1195     /* Freeze this request if another allocating write is in progress */
1196     if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs)) {
1197         QSIMPLEQ_INSERT_TAIL(&s->allocating_write_reqs, acb, next);
1198     }
1199     if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs) ||
1200         s->allocating_write_reqs_plugged) {
1201         return; /* wait for existing request to finish */
1202     }
1203 
1204     acb->cur_nclusters = qed_bytes_to_clusters(s,
1205             qed_offset_into_cluster(s, acb->cur_pos) + len);
1206     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1207 
1208     if (acb->flags & QED_AIOCB_ZERO) {
1209         /* Skip ahead if the clusters are already zero */
1210         if (acb->find_cluster_ret == QED_CLUSTER_ZERO) {
1211             qed_aio_next_io(acb, 0);
1212             return;
1213         }
1214 
1215         cb = qed_aio_write_zero_cluster;
1216     } else {
1217         cb = qed_aio_write_prefill;
1218         acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters);
1219     }
1220 
1221     if (qed_should_set_need_check(s)) {
1222         s->header.features |= QED_F_NEED_CHECK;
1223         qed_write_header(s, cb, acb);
1224     } else {
1225         cb(acb, 0);
1226     }
1227 }
1228 
1229 /**
1230  * Write data cluster in place
1231  *
1232  * @acb:        Write request
1233  * @offset:     Cluster offset in bytes
1234  * @len:        Length in bytes
1235  *
1236  * This path is taken when writing to already allocated clusters.
1237  */
1238 static void qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset, size_t len)
1239 {
1240     /* Allocate buffer for zero writes */
1241     if (acb->flags & QED_AIOCB_ZERO) {
1242         struct iovec *iov = acb->qiov->iov;
1243 
1244         if (!iov->iov_base) {
1245             iov->iov_base = qemu_blockalign(acb->common.bs, iov->iov_len);
1246             memset(iov->iov_base, 0, iov->iov_len);
1247         }
1248     }
1249 
1250     /* Calculate the I/O vector */
1251     acb->cur_cluster = offset;
1252     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1253 
1254     /* Do the actual write */
1255     qed_aio_write_main(acb, 0);
1256 }
1257 
1258 /**
1259  * Write data cluster
1260  *
1261  * @opaque:     Write request
1262  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1,
1263  *              or -errno
1264  * @offset:     Cluster offset in bytes
1265  * @len:        Length in bytes
1266  *
1267  * Callback from qed_find_cluster().
1268  */
1269 static void qed_aio_write_data(void *opaque, int ret,
1270                                uint64_t offset, size_t len)
1271 {
1272     QEDAIOCB *acb = opaque;
1273 
1274     trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len);
1275 
1276     acb->find_cluster_ret = ret;
1277 
1278     switch (ret) {
1279     case QED_CLUSTER_FOUND:
1280         qed_aio_write_inplace(acb, offset, len);
1281         break;
1282 
1283     case QED_CLUSTER_L2:
1284     case QED_CLUSTER_L1:
1285     case QED_CLUSTER_ZERO:
1286         qed_aio_write_alloc(acb, len);
1287         break;
1288 
1289     default:
1290         qed_aio_complete(acb, ret);
1291         break;
1292     }
1293 }
1294 
1295 /**
1296  * Read data cluster
1297  *
1298  * @opaque:     Read request
1299  * @ret:        QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1,
1300  *              or -errno
1301  * @offset:     Cluster offset in bytes
1302  * @len:        Length in bytes
1303  *
1304  * Callback from qed_find_cluster().
1305  */
1306 static void qed_aio_read_data(void *opaque, int ret,
1307                               uint64_t offset, size_t len)
1308 {
1309     QEDAIOCB *acb = opaque;
1310     BDRVQEDState *s = acb_to_s(acb);
1311     BlockDriverState *bs = acb->common.bs;
1312 
1313     /* Adjust offset into cluster */
1314     offset += qed_offset_into_cluster(s, acb->cur_pos);
1315 
1316     trace_qed_aio_read_data(s, acb, ret, offset, len);
1317 
1318     if (ret < 0) {
1319         goto err;
1320     }
1321 
1322     qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1323 
1324     /* Handle zero cluster and backing file reads */
1325     if (ret == QED_CLUSTER_ZERO) {
1326         qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size);
1327         qed_aio_next_io(acb, 0);
1328         return;
1329     } else if (ret != QED_CLUSTER_FOUND) {
1330         qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov,
1331                               &acb->backing_qiov, qed_aio_next_io, acb);
1332         return;
1333     }
1334 
1335     BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
1336     bdrv_aio_readv(bs->file, offset / BDRV_SECTOR_SIZE,
1337                    &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE,
1338                    qed_aio_next_io, acb);
1339     return;
1340 
1341 err:
1342     qed_aio_complete(acb, ret);
1343 }
1344 
1345 /**
1346  * Begin next I/O or complete the request
1347  */
1348 static void qed_aio_next_io(void *opaque, int ret)
1349 {
1350     QEDAIOCB *acb = opaque;
1351     BDRVQEDState *s = acb_to_s(acb);
1352     QEDFindClusterFunc *io_fn = (acb->flags & QED_AIOCB_WRITE) ?
1353                                 qed_aio_write_data : qed_aio_read_data;
1354 
1355     trace_qed_aio_next_io(s, acb, ret, acb->cur_pos + acb->cur_qiov.size);
1356 
1357     if (acb->backing_qiov) {
1358         qemu_iovec_destroy(acb->backing_qiov);
1359         g_free(acb->backing_qiov);
1360         acb->backing_qiov = NULL;
1361     }
1362 
1363     /* Handle I/O error */
1364     if (ret) {
1365         qed_aio_complete(acb, ret);
1366         return;
1367     }
1368 
1369     acb->qiov_offset += acb->cur_qiov.size;
1370     acb->cur_pos += acb->cur_qiov.size;
1371     qemu_iovec_reset(&acb->cur_qiov);
1372 
1373     /* Complete request */
1374     if (acb->cur_pos >= acb->end_pos) {
1375         qed_aio_complete(acb, 0);
1376         return;
1377     }
1378 
1379     /* Find next cluster and start I/O */
1380     qed_find_cluster(s, &acb->request,
1381                       acb->cur_pos, acb->end_pos - acb->cur_pos,
1382                       io_fn, acb);
1383 }
1384 
1385 static BlockDriverAIOCB *qed_aio_setup(BlockDriverState *bs,
1386                                        int64_t sector_num,
1387                                        QEMUIOVector *qiov, int nb_sectors,
1388                                        BlockDriverCompletionFunc *cb,
1389                                        void *opaque, int flags)
1390 {
1391     QEDAIOCB *acb = qemu_aio_get(&qed_aiocb_info, bs, cb, opaque);
1392 
1393     trace_qed_aio_setup(bs->opaque, acb, sector_num, nb_sectors,
1394                         opaque, flags);
1395 
1396     acb->flags = flags;
1397     acb->finished = NULL;
1398     acb->qiov = qiov;
1399     acb->qiov_offset = 0;
1400     acb->cur_pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE;
1401     acb->end_pos = acb->cur_pos + nb_sectors * BDRV_SECTOR_SIZE;
1402     acb->backing_qiov = NULL;
1403     acb->request.l2_table = NULL;
1404     qemu_iovec_init(&acb->cur_qiov, qiov->niov);
1405 
1406     /* Start request */
1407     qed_aio_next_io(acb, 0);
1408     return &acb->common;
1409 }
1410 
1411 static BlockDriverAIOCB *bdrv_qed_aio_readv(BlockDriverState *bs,
1412                                             int64_t sector_num,
1413                                             QEMUIOVector *qiov, int nb_sectors,
1414                                             BlockDriverCompletionFunc *cb,
1415                                             void *opaque)
1416 {
1417     return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb, opaque, 0);
1418 }
1419 
1420 static BlockDriverAIOCB *bdrv_qed_aio_writev(BlockDriverState *bs,
1421                                              int64_t sector_num,
1422                                              QEMUIOVector *qiov, int nb_sectors,
1423                                              BlockDriverCompletionFunc *cb,
1424                                              void *opaque)
1425 {
1426     return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb,
1427                          opaque, QED_AIOCB_WRITE);
1428 }
1429 
1430 typedef struct {
1431     Coroutine *co;
1432     int ret;
1433     bool done;
1434 } QEDWriteZeroesCB;
1435 
1436 static void coroutine_fn qed_co_write_zeroes_cb(void *opaque, int ret)
1437 {
1438     QEDWriteZeroesCB *cb = opaque;
1439 
1440     cb->done = true;
1441     cb->ret = ret;
1442     if (cb->co) {
1443         qemu_coroutine_enter(cb->co, NULL);
1444     }
1445 }
1446 
1447 static int coroutine_fn bdrv_qed_co_write_zeroes(BlockDriverState *bs,
1448                                                  int64_t sector_num,
1449                                                  int nb_sectors,
1450                                                  BdrvRequestFlags flags)
1451 {
1452     BlockDriverAIOCB *blockacb;
1453     BDRVQEDState *s = bs->opaque;
1454     QEDWriteZeroesCB cb = { .done = false };
1455     QEMUIOVector qiov;
1456     struct iovec iov;
1457 
1458     /* Refuse if there are untouched backing file sectors */
1459     if (bs->backing_hd) {
1460         if (qed_offset_into_cluster(s, sector_num * BDRV_SECTOR_SIZE) != 0) {
1461             return -ENOTSUP;
1462         }
1463         if (qed_offset_into_cluster(s, nb_sectors * BDRV_SECTOR_SIZE) != 0) {
1464             return -ENOTSUP;
1465         }
1466     }
1467 
1468     /* Zero writes start without an I/O buffer.  If a buffer becomes necessary
1469      * then it will be allocated during request processing.
1470      */
1471     iov.iov_base = NULL,
1472     iov.iov_len  = nb_sectors * BDRV_SECTOR_SIZE,
1473 
1474     qemu_iovec_init_external(&qiov, &iov, 1);
1475     blockacb = qed_aio_setup(bs, sector_num, &qiov, nb_sectors,
1476                              qed_co_write_zeroes_cb, &cb,
1477                              QED_AIOCB_WRITE | QED_AIOCB_ZERO);
1478     if (!blockacb) {
1479         return -EIO;
1480     }
1481     if (!cb.done) {
1482         cb.co = qemu_coroutine_self();
1483         qemu_coroutine_yield();
1484     }
1485     assert(cb.done);
1486     return cb.ret;
1487 }
1488 
1489 static int bdrv_qed_truncate(BlockDriverState *bs, int64_t offset)
1490 {
1491     BDRVQEDState *s = bs->opaque;
1492     uint64_t old_image_size;
1493     int ret;
1494 
1495     if (!qed_is_image_size_valid(offset, s->header.cluster_size,
1496                                  s->header.table_size)) {
1497         return -EINVAL;
1498     }
1499 
1500     /* Shrinking is currently not supported */
1501     if ((uint64_t)offset < s->header.image_size) {
1502         return -ENOTSUP;
1503     }
1504 
1505     old_image_size = s->header.image_size;
1506     s->header.image_size = offset;
1507     ret = qed_write_header_sync(s);
1508     if (ret < 0) {
1509         s->header.image_size = old_image_size;
1510     }
1511     return ret;
1512 }
1513 
1514 static int64_t bdrv_qed_getlength(BlockDriverState *bs)
1515 {
1516     BDRVQEDState *s = bs->opaque;
1517     return s->header.image_size;
1518 }
1519 
1520 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1521 {
1522     BDRVQEDState *s = bs->opaque;
1523 
1524     memset(bdi, 0, sizeof(*bdi));
1525     bdi->cluster_size = s->header.cluster_size;
1526     bdi->is_dirty = s->header.features & QED_F_NEED_CHECK;
1527     bdi->unallocated_blocks_are_zero = true;
1528     bdi->can_write_zeroes_with_unmap = true;
1529     return 0;
1530 }
1531 
1532 static int bdrv_qed_change_backing_file(BlockDriverState *bs,
1533                                         const char *backing_file,
1534                                         const char *backing_fmt)
1535 {
1536     BDRVQEDState *s = bs->opaque;
1537     QEDHeader new_header, le_header;
1538     void *buffer;
1539     size_t buffer_len, backing_file_len;
1540     int ret;
1541 
1542     /* Refuse to set backing filename if unknown compat feature bits are
1543      * active.  If the image uses an unknown compat feature then we may not
1544      * know the layout of data following the header structure and cannot safely
1545      * add a new string.
1546      */
1547     if (backing_file && (s->header.compat_features &
1548                          ~QED_COMPAT_FEATURE_MASK)) {
1549         return -ENOTSUP;
1550     }
1551 
1552     memcpy(&new_header, &s->header, sizeof(new_header));
1553 
1554     new_header.features &= ~(QED_F_BACKING_FILE |
1555                              QED_F_BACKING_FORMAT_NO_PROBE);
1556 
1557     /* Adjust feature flags */
1558     if (backing_file) {
1559         new_header.features |= QED_F_BACKING_FILE;
1560 
1561         if (qed_fmt_is_raw(backing_fmt)) {
1562             new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
1563         }
1564     }
1565 
1566     /* Calculate new header size */
1567     backing_file_len = 0;
1568 
1569     if (backing_file) {
1570         backing_file_len = strlen(backing_file);
1571     }
1572 
1573     buffer_len = sizeof(new_header);
1574     new_header.backing_filename_offset = buffer_len;
1575     new_header.backing_filename_size = backing_file_len;
1576     buffer_len += backing_file_len;
1577 
1578     /* Make sure we can rewrite header without failing */
1579     if (buffer_len > new_header.header_size * new_header.cluster_size) {
1580         return -ENOSPC;
1581     }
1582 
1583     /* Prepare new header */
1584     buffer = g_malloc(buffer_len);
1585 
1586     qed_header_cpu_to_le(&new_header, &le_header);
1587     memcpy(buffer, &le_header, sizeof(le_header));
1588     buffer_len = sizeof(le_header);
1589 
1590     if (backing_file) {
1591         memcpy(buffer + buffer_len, backing_file, backing_file_len);
1592         buffer_len += backing_file_len;
1593     }
1594 
1595     /* Write new header */
1596     ret = bdrv_pwrite_sync(bs->file, 0, buffer, buffer_len);
1597     g_free(buffer);
1598     if (ret == 0) {
1599         memcpy(&s->header, &new_header, sizeof(new_header));
1600     }
1601     return ret;
1602 }
1603 
1604 static void bdrv_qed_invalidate_cache(BlockDriverState *bs, Error **errp)
1605 {
1606     BDRVQEDState *s = bs->opaque;
1607     Error *local_err = NULL;
1608     int ret;
1609 
1610     bdrv_qed_close(bs);
1611 
1612     bdrv_invalidate_cache(bs->file, &local_err);
1613     if (local_err) {
1614         error_propagate(errp, local_err);
1615         return;
1616     }
1617 
1618     memset(s, 0, sizeof(BDRVQEDState));
1619     ret = bdrv_qed_open(bs, NULL, bs->open_flags, &local_err);
1620     if (local_err) {
1621         error_setg(errp, "Could not reopen qed layer: %s",
1622                    error_get_pretty(local_err));
1623         error_free(local_err);
1624         return;
1625     } else if (ret < 0) {
1626         error_setg_errno(errp, -ret, "Could not reopen qed layer");
1627         return;
1628     }
1629 }
1630 
1631 static int bdrv_qed_check(BlockDriverState *bs, BdrvCheckResult *result,
1632                           BdrvCheckMode fix)
1633 {
1634     BDRVQEDState *s = bs->opaque;
1635 
1636     return qed_check(s, result, !!fix);
1637 }
1638 
1639 static QemuOptsList qed_create_opts = {
1640     .name = "qed-create-opts",
1641     .head = QTAILQ_HEAD_INITIALIZER(qed_create_opts.head),
1642     .desc = {
1643         {
1644             .name = BLOCK_OPT_SIZE,
1645             .type = QEMU_OPT_SIZE,
1646             .help = "Virtual disk size"
1647         },
1648         {
1649             .name = BLOCK_OPT_BACKING_FILE,
1650             .type = QEMU_OPT_STRING,
1651             .help = "File name of a base image"
1652         },
1653         {
1654             .name = BLOCK_OPT_BACKING_FMT,
1655             .type = QEMU_OPT_STRING,
1656             .help = "Image format of the base image"
1657         },
1658         {
1659             .name = BLOCK_OPT_CLUSTER_SIZE,
1660             .type = QEMU_OPT_SIZE,
1661             .help = "Cluster size (in bytes)",
1662             .def_value_str = stringify(QED_DEFAULT_CLUSTER_SIZE)
1663         },
1664         {
1665             .name = BLOCK_OPT_TABLE_SIZE,
1666             .type = QEMU_OPT_SIZE,
1667             .help = "L1/L2 table size (in clusters)"
1668         },
1669         { /* end of list */ }
1670     }
1671 };
1672 
1673 static BlockDriver bdrv_qed = {
1674     .format_name              = "qed",
1675     .instance_size            = sizeof(BDRVQEDState),
1676     .create_opts              = &qed_create_opts,
1677     .supports_backing         = true,
1678 
1679     .bdrv_probe               = bdrv_qed_probe,
1680     .bdrv_rebind              = bdrv_qed_rebind,
1681     .bdrv_open                = bdrv_qed_open,
1682     .bdrv_close               = bdrv_qed_close,
1683     .bdrv_reopen_prepare      = bdrv_qed_reopen_prepare,
1684     .bdrv_create              = bdrv_qed_create,
1685     .bdrv_has_zero_init       = bdrv_has_zero_init_1,
1686     .bdrv_co_get_block_status = bdrv_qed_co_get_block_status,
1687     .bdrv_aio_readv           = bdrv_qed_aio_readv,
1688     .bdrv_aio_writev          = bdrv_qed_aio_writev,
1689     .bdrv_co_write_zeroes     = bdrv_qed_co_write_zeroes,
1690     .bdrv_truncate            = bdrv_qed_truncate,
1691     .bdrv_getlength           = bdrv_qed_getlength,
1692     .bdrv_get_info            = bdrv_qed_get_info,
1693     .bdrv_refresh_limits      = bdrv_qed_refresh_limits,
1694     .bdrv_change_backing_file = bdrv_qed_change_backing_file,
1695     .bdrv_invalidate_cache    = bdrv_qed_invalidate_cache,
1696     .bdrv_check               = bdrv_qed_check,
1697     .bdrv_detach_aio_context  = bdrv_qed_detach_aio_context,
1698     .bdrv_attach_aio_context  = bdrv_qed_attach_aio_context,
1699 };
1700 
1701 static void bdrv_qed_init(void)
1702 {
1703     bdrv_register(&bdrv_qed);
1704 }
1705 
1706 block_init(bdrv_qed_init);
1707