qed: Simplify backing reads
[qemu.git] / block / qed.c
1 /*
2 * QEMU Enhanced Disk Format
3 *
4 * Copyright IBM, Corp. 2010
5 *
6 * Authors:
7 * Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
8 * Anthony Liguori <aliguori@us.ibm.com>
9 *
10 * This work is licensed under the terms of the GNU LGPL, version 2 or later.
11 * See the COPYING.LIB file in the top-level directory.
12 *
13 */
14
15 #include "qemu/osdep.h"
16 #include "block/qdict.h"
17 #include "qapi/error.h"
18 #include "qemu/timer.h"
19 #include "qemu/bswap.h"
20 #include "qemu/main-loop.h"
21 #include "qemu/module.h"
22 #include "qemu/option.h"
23 #include "trace.h"
24 #include "qed.h"
25 #include "sysemu/block-backend.h"
26 #include "qapi/qmp/qdict.h"
27 #include "qapi/qobject-input-visitor.h"
28 #include "qapi/qapi-visit-block-core.h"
29
30 static QemuOptsList qed_create_opts;
31
32 static int bdrv_qed_probe(const uint8_t *buf, int buf_size,
33 const char *filename)
34 {
35 const QEDHeader *header = (const QEDHeader *)buf;
36
37 if (buf_size < sizeof(*header)) {
38 return 0;
39 }
40 if (le32_to_cpu(header->magic) != QED_MAGIC) {
41 return 0;
42 }
43 return 100;
44 }
45
46 /**
47 * Check whether an image format is raw
48 *
49 * @fmt: Backing file format, may be NULL
50 */
51 static bool qed_fmt_is_raw(const char *fmt)
52 {
53 return fmt && strcmp(fmt, "raw") == 0;
54 }
55
56 static void qed_header_le_to_cpu(const QEDHeader *le, QEDHeader *cpu)
57 {
58 cpu->magic = le32_to_cpu(le->magic);
59 cpu->cluster_size = le32_to_cpu(le->cluster_size);
60 cpu->table_size = le32_to_cpu(le->table_size);
61 cpu->header_size = le32_to_cpu(le->header_size);
62 cpu->features = le64_to_cpu(le->features);
63 cpu->compat_features = le64_to_cpu(le->compat_features);
64 cpu->autoclear_features = le64_to_cpu(le->autoclear_features);
65 cpu->l1_table_offset = le64_to_cpu(le->l1_table_offset);
66 cpu->image_size = le64_to_cpu(le->image_size);
67 cpu->backing_filename_offset = le32_to_cpu(le->backing_filename_offset);
68 cpu->backing_filename_size = le32_to_cpu(le->backing_filename_size);
69 }
70
71 static void qed_header_cpu_to_le(const QEDHeader *cpu, QEDHeader *le)
72 {
73 le->magic = cpu_to_le32(cpu->magic);
74 le->cluster_size = cpu_to_le32(cpu->cluster_size);
75 le->table_size = cpu_to_le32(cpu->table_size);
76 le->header_size = cpu_to_le32(cpu->header_size);
77 le->features = cpu_to_le64(cpu->features);
78 le->compat_features = cpu_to_le64(cpu->compat_features);
79 le->autoclear_features = cpu_to_le64(cpu->autoclear_features);
80 le->l1_table_offset = cpu_to_le64(cpu->l1_table_offset);
81 le->image_size = cpu_to_le64(cpu->image_size);
82 le->backing_filename_offset = cpu_to_le32(cpu->backing_filename_offset);
83 le->backing_filename_size = cpu_to_le32(cpu->backing_filename_size);
84 }
85
86 int qed_write_header_sync(BDRVQEDState *s)
87 {
88 QEDHeader le;
89 int ret;
90
91 qed_header_cpu_to_le(&s->header, &le);
92 ret = bdrv_pwrite(s->bs->file, 0, &le, sizeof(le));
93 if (ret != sizeof(le)) {
94 return ret;
95 }
96 return 0;
97 }
98
99 /**
100 * Update header in-place (does not rewrite backing filename or other strings)
101 *
102 * This function only updates known header fields in-place and does not affect
103 * extra data after the QED header.
104 *
105 * No new allocating reqs can start while this function runs.
106 */
107 static int coroutine_fn qed_write_header(BDRVQEDState *s)
108 {
109 /* We must write full sectors for O_DIRECT but cannot necessarily generate
110 * the data following the header if an unrecognized compat feature is
111 * active. Therefore, first read the sectors containing the header, update
112 * them, and write back.
113 */
114
115 int nsectors = DIV_ROUND_UP(sizeof(QEDHeader), BDRV_SECTOR_SIZE);
116 size_t len = nsectors * BDRV_SECTOR_SIZE;
117 uint8_t *buf;
118 int ret;
119
120 assert(s->allocating_acb || s->allocating_write_reqs_plugged);
121
122 buf = qemu_blockalign(s->bs, len);
123
124 ret = bdrv_co_pread(s->bs->file, 0, len, buf, 0);
125 if (ret < 0) {
126 goto out;
127 }
128
129 /* Update header */
130 qed_header_cpu_to_le(&s->header, (QEDHeader *) buf);
131
132 ret = bdrv_co_pwrite(s->bs->file, 0, len, buf, 0);
133 if (ret < 0) {
134 goto out;
135 }
136
137 ret = 0;
138 out:
139 qemu_vfree(buf);
140 return ret;
141 }
142
143 static uint64_t qed_max_image_size(uint32_t cluster_size, uint32_t table_size)
144 {
145 uint64_t table_entries;
146 uint64_t l2_size;
147
148 table_entries = (table_size * cluster_size) / sizeof(uint64_t);
149 l2_size = table_entries * cluster_size;
150
151 return l2_size * table_entries;
152 }
153
154 static bool qed_is_cluster_size_valid(uint32_t cluster_size)
155 {
156 if (cluster_size < QED_MIN_CLUSTER_SIZE ||
157 cluster_size > QED_MAX_CLUSTER_SIZE) {
158 return false;
159 }
160 if (cluster_size & (cluster_size - 1)) {
161 return false; /* not power of 2 */
162 }
163 return true;
164 }
165
166 static bool qed_is_table_size_valid(uint32_t table_size)
167 {
168 if (table_size < QED_MIN_TABLE_SIZE ||
169 table_size > QED_MAX_TABLE_SIZE) {
170 return false;
171 }
172 if (table_size & (table_size - 1)) {
173 return false; /* not power of 2 */
174 }
175 return true;
176 }
177
178 static bool qed_is_image_size_valid(uint64_t image_size, uint32_t cluster_size,
179 uint32_t table_size)
180 {
181 if (image_size % BDRV_SECTOR_SIZE != 0) {
182 return false; /* not multiple of sector size */
183 }
184 if (image_size > qed_max_image_size(cluster_size, table_size)) {
185 return false; /* image is too large */
186 }
187 return true;
188 }
189
190 /**
191 * Read a string of known length from the image file
192 *
193 * @file: Image file
194 * @offset: File offset to start of string, in bytes
195 * @n: String length in bytes
196 * @buf: Destination buffer
197 * @buflen: Destination buffer length in bytes
198 * @ret: 0 on success, -errno on failure
199 *
200 * The string is NUL-terminated.
201 */
202 static int qed_read_string(BdrvChild *file, uint64_t offset, size_t n,
203 char *buf, size_t buflen)
204 {
205 int ret;
206 if (n >= buflen) {
207 return -EINVAL;
208 }
209 ret = bdrv_pread(file, offset, buf, n);
210 if (ret < 0) {
211 return ret;
212 }
213 buf[n] = '\0';
214 return 0;
215 }
216
217 /**
218 * Allocate new clusters
219 *
220 * @s: QED state
221 * @n: Number of contiguous clusters to allocate
222 * @ret: Offset of first allocated cluster
223 *
224 * This function only produces the offset where the new clusters should be
225 * written. It updates BDRVQEDState but does not make any changes to the image
226 * file.
227 *
228 * Called with table_lock held.
229 */
230 static uint64_t qed_alloc_clusters(BDRVQEDState *s, unsigned int n)
231 {
232 uint64_t offset = s->file_size;
233 s->file_size += n * s->header.cluster_size;
234 return offset;
235 }
236
237 QEDTable *qed_alloc_table(BDRVQEDState *s)
238 {
239 /* Honor O_DIRECT memory alignment requirements */
240 return qemu_blockalign(s->bs,
241 s->header.cluster_size * s->header.table_size);
242 }
243
244 /**
245 * Allocate a new zeroed L2 table
246 *
247 * Called with table_lock held.
248 */
249 static CachedL2Table *qed_new_l2_table(BDRVQEDState *s)
250 {
251 CachedL2Table *l2_table = qed_alloc_l2_cache_entry(&s->l2_cache);
252
253 l2_table->table = qed_alloc_table(s);
254 l2_table->offset = qed_alloc_clusters(s, s->header.table_size);
255
256 memset(l2_table->table->offsets, 0,
257 s->header.cluster_size * s->header.table_size);
258 return l2_table;
259 }
260
261 static bool qed_plug_allocating_write_reqs(BDRVQEDState *s)
262 {
263 qemu_co_mutex_lock(&s->table_lock);
264
265 /* No reentrancy is allowed. */
266 assert(!s->allocating_write_reqs_plugged);
267 if (s->allocating_acb != NULL) {
268 /* Another allocating write came concurrently. This cannot happen
269 * from bdrv_qed_co_drain_begin, but it can happen when the timer runs.
270 */
271 qemu_co_mutex_unlock(&s->table_lock);
272 return false;
273 }
274
275 s->allocating_write_reqs_plugged = true;
276 qemu_co_mutex_unlock(&s->table_lock);
277 return true;
278 }
279
280 static void qed_unplug_allocating_write_reqs(BDRVQEDState *s)
281 {
282 qemu_co_mutex_lock(&s->table_lock);
283 assert(s->allocating_write_reqs_plugged);
284 s->allocating_write_reqs_plugged = false;
285 qemu_co_queue_next(&s->allocating_write_reqs);
286 qemu_co_mutex_unlock(&s->table_lock);
287 }
288
289 static void coroutine_fn qed_need_check_timer_entry(void *opaque)
290 {
291 BDRVQEDState *s = opaque;
292 int ret;
293
294 trace_qed_need_check_timer_cb(s);
295
296 if (!qed_plug_allocating_write_reqs(s)) {
297 return;
298 }
299
300 /* Ensure writes are on disk before clearing flag */
301 ret = bdrv_co_flush(s->bs->file->bs);
302 if (ret < 0) {
303 qed_unplug_allocating_write_reqs(s);
304 return;
305 }
306
307 s->header.features &= ~QED_F_NEED_CHECK;
308 ret = qed_write_header(s);
309 (void) ret;
310
311 qed_unplug_allocating_write_reqs(s);
312
313 ret = bdrv_co_flush(s->bs);
314 (void) ret;
315 }
316
317 static void qed_need_check_timer_cb(void *opaque)
318 {
319 Coroutine *co = qemu_coroutine_create(qed_need_check_timer_entry, opaque);
320 qemu_coroutine_enter(co);
321 }
322
323 static void qed_start_need_check_timer(BDRVQEDState *s)
324 {
325 trace_qed_start_need_check_timer(s);
326
327 /* Use QEMU_CLOCK_VIRTUAL so we don't alter the image file while suspended for
328 * migration.
329 */
330 timer_mod(s->need_check_timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
331 NANOSECONDS_PER_SECOND * QED_NEED_CHECK_TIMEOUT);
332 }
333
334 /* It's okay to call this multiple times or when no timer is started */
335 static void qed_cancel_need_check_timer(BDRVQEDState *s)
336 {
337 trace_qed_cancel_need_check_timer(s);
338 timer_del(s->need_check_timer);
339 }
340
341 static void bdrv_qed_detach_aio_context(BlockDriverState *bs)
342 {
343 BDRVQEDState *s = bs->opaque;
344
345 qed_cancel_need_check_timer(s);
346 timer_free(s->need_check_timer);
347 }
348
349 static void bdrv_qed_attach_aio_context(BlockDriverState *bs,
350 AioContext *new_context)
351 {
352 BDRVQEDState *s = bs->opaque;
353
354 s->need_check_timer = aio_timer_new(new_context,
355 QEMU_CLOCK_VIRTUAL, SCALE_NS,
356 qed_need_check_timer_cb, s);
357 if (s->header.features & QED_F_NEED_CHECK) {
358 qed_start_need_check_timer(s);
359 }
360 }
361
362 static void coroutine_fn bdrv_qed_co_drain_begin(BlockDriverState *bs)
363 {
364 BDRVQEDState *s = bs->opaque;
365
366 /* Fire the timer immediately in order to start doing I/O as soon as the
367 * header is flushed.
368 */
369 if (s->need_check_timer && timer_pending(s->need_check_timer)) {
370 qed_cancel_need_check_timer(s);
371 qed_need_check_timer_entry(s);
372 }
373 }
374
375 static void bdrv_qed_init_state(BlockDriverState *bs)
376 {
377 BDRVQEDState *s = bs->opaque;
378
379 memset(s, 0, sizeof(BDRVQEDState));
380 s->bs = bs;
381 qemu_co_mutex_init(&s->table_lock);
382 qemu_co_queue_init(&s->allocating_write_reqs);
383 }
384
385 /* Called with table_lock held. */
386 static int coroutine_fn bdrv_qed_do_open(BlockDriverState *bs, QDict *options,
387 int flags, Error **errp)
388 {
389 BDRVQEDState *s = bs->opaque;
390 QEDHeader le_header;
391 int64_t file_size;
392 int ret;
393
394 ret = bdrv_pread(bs->file, 0, &le_header, sizeof(le_header));
395 if (ret < 0) {
396 return ret;
397 }
398 qed_header_le_to_cpu(&le_header, &s->header);
399
400 if (s->header.magic != QED_MAGIC) {
401 error_setg(errp, "Image not in QED format");
402 return -EINVAL;
403 }
404 if (s->header.features & ~QED_FEATURE_MASK) {
405 /* image uses unsupported feature bits */
406 error_setg(errp, "Unsupported QED features: %" PRIx64,
407 s->header.features & ~QED_FEATURE_MASK);
408 return -ENOTSUP;
409 }
410 if (!qed_is_cluster_size_valid(s->header.cluster_size)) {
411 return -EINVAL;
412 }
413
414 /* Round down file size to the last cluster */
415 file_size = bdrv_getlength(bs->file->bs);
416 if (file_size < 0) {
417 return file_size;
418 }
419 s->file_size = qed_start_of_cluster(s, file_size);
420
421 if (!qed_is_table_size_valid(s->header.table_size)) {
422 return -EINVAL;
423 }
424 if (!qed_is_image_size_valid(s->header.image_size,
425 s->header.cluster_size,
426 s->header.table_size)) {
427 return -EINVAL;
428 }
429 if (!qed_check_table_offset(s, s->header.l1_table_offset)) {
430 return -EINVAL;
431 }
432
433 s->table_nelems = (s->header.cluster_size * s->header.table_size) /
434 sizeof(uint64_t);
435 s->l2_shift = ctz32(s->header.cluster_size);
436 s->l2_mask = s->table_nelems - 1;
437 s->l1_shift = s->l2_shift + ctz32(s->table_nelems);
438
439 /* Header size calculation must not overflow uint32_t */
440 if (s->header.header_size > UINT32_MAX / s->header.cluster_size) {
441 return -EINVAL;
442 }
443
444 if ((s->header.features & QED_F_BACKING_FILE)) {
445 if ((uint64_t)s->header.backing_filename_offset +
446 s->header.backing_filename_size >
447 s->header.cluster_size * s->header.header_size) {
448 return -EINVAL;
449 }
450
451 ret = qed_read_string(bs->file, s->header.backing_filename_offset,
452 s->header.backing_filename_size,
453 bs->auto_backing_file,
454 sizeof(bs->auto_backing_file));
455 if (ret < 0) {
456 return ret;
457 }
458 pstrcpy(bs->backing_file, sizeof(bs->backing_file),
459 bs->auto_backing_file);
460
461 if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) {
462 pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw");
463 }
464 }
465
466 /* Reset unknown autoclear feature bits. This is a backwards
467 * compatibility mechanism that allows images to be opened by older
468 * programs, which "knock out" unknown feature bits. When an image is
469 * opened by a newer program again it can detect that the autoclear
470 * feature is no longer valid.
471 */
472 if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 &&
473 !bdrv_is_read_only(bs->file->bs) && !(flags & BDRV_O_INACTIVE)) {
474 s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK;
475
476 ret = qed_write_header_sync(s);
477 if (ret) {
478 return ret;
479 }
480
481 /* From here on only known autoclear feature bits are valid */
482 bdrv_flush(bs->file->bs);
483 }
484
485 s->l1_table = qed_alloc_table(s);
486 qed_init_l2_cache(&s->l2_cache);
487
488 ret = qed_read_l1_table_sync(s);
489 if (ret) {
490 goto out;
491 }
492
493 /* If image was not closed cleanly, check consistency */
494 if (!(flags & BDRV_O_CHECK) && (s->header.features & QED_F_NEED_CHECK)) {
495 /* Read-only images cannot be fixed. There is no risk of corruption
496 * since write operations are not possible. Therefore, allow
497 * potentially inconsistent images to be opened read-only. This can
498 * aid data recovery from an otherwise inconsistent image.
499 */
500 if (!bdrv_is_read_only(bs->file->bs) &&
501 !(flags & BDRV_O_INACTIVE)) {
502 BdrvCheckResult result = {0};
503
504 ret = qed_check(s, &result, true);
505 if (ret) {
506 goto out;
507 }
508 }
509 }
510
511 bdrv_qed_attach_aio_context(bs, bdrv_get_aio_context(bs));
512
513 out:
514 if (ret) {
515 qed_free_l2_cache(&s->l2_cache);
516 qemu_vfree(s->l1_table);
517 }
518 return ret;
519 }
520
521 typedef struct QEDOpenCo {
522 BlockDriverState *bs;
523 QDict *options;
524 int flags;
525 Error **errp;
526 int ret;
527 } QEDOpenCo;
528
529 static void coroutine_fn bdrv_qed_open_entry(void *opaque)
530 {
531 QEDOpenCo *qoc = opaque;
532 BDRVQEDState *s = qoc->bs->opaque;
533
534 qemu_co_mutex_lock(&s->table_lock);
535 qoc->ret = bdrv_qed_do_open(qoc->bs, qoc->options, qoc->flags, qoc->errp);
536 qemu_co_mutex_unlock(&s->table_lock);
537 }
538
539 static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags,
540 Error **errp)
541 {
542 QEDOpenCo qoc = {
543 .bs = bs,
544 .options = options,
545 .flags = flags,
546 .errp = errp,
547 .ret = -EINPROGRESS
548 };
549
550 bs->file = bdrv_open_child(NULL, options, "file", bs, &child_of_bds,
551 BDRV_CHILD_IMAGE, false, errp);
552 if (!bs->file) {
553 return -EINVAL;
554 }
555
556 bdrv_qed_init_state(bs);
557 if (qemu_in_coroutine()) {
558 bdrv_qed_open_entry(&qoc);
559 } else {
560 assert(qemu_get_current_aio_context() == qemu_get_aio_context());
561 qemu_coroutine_enter(qemu_coroutine_create(bdrv_qed_open_entry, &qoc));
562 BDRV_POLL_WHILE(bs, qoc.ret == -EINPROGRESS);
563 }
564 BDRV_POLL_WHILE(bs, qoc.ret == -EINPROGRESS);
565 return qoc.ret;
566 }
567
568 static void bdrv_qed_refresh_limits(BlockDriverState *bs, Error **errp)
569 {
570 BDRVQEDState *s = bs->opaque;
571
572 bs->bl.pwrite_zeroes_alignment = s->header.cluster_size;
573 }
574
575 /* We have nothing to do for QED reopen, stubs just return
576 * success */
577 static int bdrv_qed_reopen_prepare(BDRVReopenState *state,
578 BlockReopenQueue *queue, Error **errp)
579 {
580 return 0;
581 }
582
583 static void bdrv_qed_close(BlockDriverState *bs)
584 {
585 BDRVQEDState *s = bs->opaque;
586
587 bdrv_qed_detach_aio_context(bs);
588
589 /* Ensure writes reach stable storage */
590 bdrv_flush(bs->file->bs);
591
592 /* Clean shutdown, no check required on next open */
593 if (s->header.features & QED_F_NEED_CHECK) {
594 s->header.features &= ~QED_F_NEED_CHECK;
595 qed_write_header_sync(s);
596 }
597
598 qed_free_l2_cache(&s->l2_cache);
599 qemu_vfree(s->l1_table);
600 }
601
602 static int coroutine_fn bdrv_qed_co_create(BlockdevCreateOptions *opts,
603 Error **errp)
604 {
605 BlockdevCreateOptionsQed *qed_opts;
606 BlockBackend *blk = NULL;
607 BlockDriverState *bs = NULL;
608
609 QEDHeader header;
610 QEDHeader le_header;
611 uint8_t *l1_table = NULL;
612 size_t l1_size;
613 int ret = 0;
614
615 assert(opts->driver == BLOCKDEV_DRIVER_QED);
616 qed_opts = &opts->u.qed;
617
618 /* Validate options and set default values */
619 if (!qed_opts->has_cluster_size) {
620 qed_opts->cluster_size = QED_DEFAULT_CLUSTER_SIZE;
621 }
622 if (!qed_opts->has_table_size) {
623 qed_opts->table_size = QED_DEFAULT_TABLE_SIZE;
624 }
625
626 if (!qed_is_cluster_size_valid(qed_opts->cluster_size)) {
627 error_setg(errp, "QED cluster size must be within range [%u, %u] "
628 "and power of 2",
629 QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE);
630 return -EINVAL;
631 }
632 if (!qed_is_table_size_valid(qed_opts->table_size)) {
633 error_setg(errp, "QED table size must be within range [%u, %u] "
634 "and power of 2",
635 QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE);
636 return -EINVAL;
637 }
638 if (!qed_is_image_size_valid(qed_opts->size, qed_opts->cluster_size,
639 qed_opts->table_size))
640 {
641 error_setg(errp, "QED image size must be a non-zero multiple of "
642 "cluster size and less than %" PRIu64 " bytes",
643 qed_max_image_size(qed_opts->cluster_size,
644 qed_opts->table_size));
645 return -EINVAL;
646 }
647
648 /* Create BlockBackend to write to the image */
649 bs = bdrv_open_blockdev_ref(qed_opts->file, errp);
650 if (bs == NULL) {
651 return -EIO;
652 }
653
654 blk = blk_new_with_bs(bs, BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL,
655 errp);
656 if (!blk) {
657 ret = -EPERM;
658 goto out;
659 }
660 blk_set_allow_write_beyond_eof(blk, true);
661
662 /* Prepare image format */
663 header = (QEDHeader) {
664 .magic = QED_MAGIC,
665 .cluster_size = qed_opts->cluster_size,
666 .table_size = qed_opts->table_size,
667 .header_size = 1,
668 .features = 0,
669 .compat_features = 0,
670 .l1_table_offset = qed_opts->cluster_size,
671 .image_size = qed_opts->size,
672 };
673
674 l1_size = header.cluster_size * header.table_size;
675
676 /*
677 * The QED format associates file length with allocation status,
678 * so a new file (which is empty) must have a length of 0.
679 */
680 ret = blk_truncate(blk, 0, true, PREALLOC_MODE_OFF, 0, errp);
681 if (ret < 0) {
682 goto out;
683 }
684
685 if (qed_opts->has_backing_file) {
686 header.features |= QED_F_BACKING_FILE;
687 header.backing_filename_offset = sizeof(le_header);
688 header.backing_filename_size = strlen(qed_opts->backing_file);
689
690 if (qed_opts->has_backing_fmt) {
691 const char *backing_fmt = BlockdevDriver_str(qed_opts->backing_fmt);
692 if (qed_fmt_is_raw(backing_fmt)) {
693 header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
694 }
695 }
696 }
697
698 qed_header_cpu_to_le(&header, &le_header);
699 ret = blk_pwrite(blk, 0, &le_header, sizeof(le_header), 0);
700 if (ret < 0) {
701 goto out;
702 }
703 ret = blk_pwrite(blk, sizeof(le_header), qed_opts->backing_file,
704 header.backing_filename_size, 0);
705 if (ret < 0) {
706 goto out;
707 }
708
709 l1_table = g_malloc0(l1_size);
710 ret = blk_pwrite(blk, header.l1_table_offset, l1_table, l1_size, 0);
711 if (ret < 0) {
712 goto out;
713 }
714
715 ret = 0; /* success */
716 out:
717 g_free(l1_table);
718 blk_unref(blk);
719 bdrv_unref(bs);
720 return ret;
721 }
722
723 static int coroutine_fn bdrv_qed_co_create_opts(BlockDriver *drv,
724 const char *filename,
725 QemuOpts *opts,
726 Error **errp)
727 {
728 BlockdevCreateOptions *create_options = NULL;
729 QDict *qdict;
730 Visitor *v;
731 BlockDriverState *bs = NULL;
732 Error *local_err = NULL;
733 int ret;
734
735 static const QDictRenames opt_renames[] = {
736 { BLOCK_OPT_BACKING_FILE, "backing-file" },
737 { BLOCK_OPT_BACKING_FMT, "backing-fmt" },
738 { BLOCK_OPT_CLUSTER_SIZE, "cluster-size" },
739 { BLOCK_OPT_TABLE_SIZE, "table-size" },
740 { NULL, NULL },
741 };
742
743 /* Parse options and convert legacy syntax */
744 qdict = qemu_opts_to_qdict_filtered(opts, NULL, &qed_create_opts, true);
745
746 if (!qdict_rename_keys(qdict, opt_renames, errp)) {
747 ret = -EINVAL;
748 goto fail;
749 }
750
751 /* Create and open the file (protocol layer) */
752 ret = bdrv_create_file(filename, opts, &local_err);
753 if (ret < 0) {
754 error_propagate(errp, local_err);
755 goto fail;
756 }
757
758 bs = bdrv_open(filename, NULL, NULL,
759 BDRV_O_RDWR | BDRV_O_RESIZE | BDRV_O_PROTOCOL, errp);
760 if (bs == NULL) {
761 ret = -EIO;
762 goto fail;
763 }
764
765 /* Now get the QAPI type BlockdevCreateOptions */
766 qdict_put_str(qdict, "driver", "qed");
767 qdict_put_str(qdict, "file", bs->node_name);
768
769 v = qobject_input_visitor_new_flat_confused(qdict, errp);
770 if (!v) {
771 ret = -EINVAL;
772 goto fail;
773 }
774
775 visit_type_BlockdevCreateOptions(v, NULL, &create_options, &local_err);
776 visit_free(v);
777
778 if (local_err) {
779 error_propagate(errp, local_err);
780 ret = -EINVAL;
781 goto fail;
782 }
783
784 /* Silently round up size */
785 assert(create_options->driver == BLOCKDEV_DRIVER_QED);
786 create_options->u.qed.size =
787 ROUND_UP(create_options->u.qed.size, BDRV_SECTOR_SIZE);
788
789 /* Create the qed image (format layer) */
790 ret = bdrv_qed_co_create(create_options, errp);
791
792 fail:
793 qobject_unref(qdict);
794 bdrv_unref(bs);
795 qapi_free_BlockdevCreateOptions(create_options);
796 return ret;
797 }
798
799 static int coroutine_fn bdrv_qed_co_block_status(BlockDriverState *bs,
800 bool want_zero,
801 int64_t pos, int64_t bytes,
802 int64_t *pnum, int64_t *map,
803 BlockDriverState **file)
804 {
805 BDRVQEDState *s = bs->opaque;
806 size_t len = MIN(bytes, SIZE_MAX);
807 int status;
808 QEDRequest request = { .l2_table = NULL };
809 uint64_t offset;
810 int ret;
811
812 qemu_co_mutex_lock(&s->table_lock);
813 ret = qed_find_cluster(s, &request, pos, &len, &offset);
814
815 *pnum = len;
816 switch (ret) {
817 case QED_CLUSTER_FOUND:
818 *map = offset | qed_offset_into_cluster(s, pos);
819 status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
820 *file = bs->file->bs;
821 break;
822 case QED_CLUSTER_ZERO:
823 status = BDRV_BLOCK_ZERO;
824 break;
825 case QED_CLUSTER_L2:
826 case QED_CLUSTER_L1:
827 status = 0;
828 break;
829 default:
830 assert(ret < 0);
831 status = ret;
832 break;
833 }
834
835 qed_unref_l2_cache_entry(request.l2_table);
836 qemu_co_mutex_unlock(&s->table_lock);
837
838 return status;
839 }
840
841 static BDRVQEDState *acb_to_s(QEDAIOCB *acb)
842 {
843 return acb->bs->opaque;
844 }
845
846 /**
847 * Read from the backing file or zero-fill if no backing file
848 *
849 * @s: QED state
850 * @pos: Byte position in device
851 * @qiov: Destination I/O vector
852 *
853 * This function reads qiov->size bytes starting at pos from the backing file.
854 * If there is no backing file then zeroes are read.
855 */
856 static int coroutine_fn qed_read_backing_file(BDRVQEDState *s, uint64_t pos,
857 QEMUIOVector *qiov)
858 {
859 if (s->bs->backing) {
860 BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO);
861 return bdrv_co_preadv(s->bs->backing, pos, qiov->size, qiov, 0);
862 }
863 qemu_iovec_memset(qiov, 0, 0, qiov->size);
864 return 0;
865 }
866
867 /**
868 * Copy data from backing file into the image
869 *
870 * @s: QED state
871 * @pos: Byte position in device
872 * @len: Number of bytes
873 * @offset: Byte offset in image file
874 */
875 static int coroutine_fn qed_copy_from_backing_file(BDRVQEDState *s,
876 uint64_t pos, uint64_t len,
877 uint64_t offset)
878 {
879 QEMUIOVector qiov;
880 int ret;
881
882 /* Skip copy entirely if there is no work to do */
883 if (len == 0) {
884 return 0;
885 }
886
887 qemu_iovec_init_buf(&qiov, qemu_blockalign(s->bs, len), len);
888
889 ret = qed_read_backing_file(s, pos, &qiov);
890
891 if (ret) {
892 goto out;
893 }
894
895 BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE);
896 ret = bdrv_co_pwritev(s->bs->file, offset, qiov.size, &qiov, 0);
897 if (ret < 0) {
898 goto out;
899 }
900 ret = 0;
901 out:
902 qemu_vfree(qemu_iovec_buf(&qiov));
903 return ret;
904 }
905
906 /**
907 * Link one or more contiguous clusters into a table
908 *
909 * @s: QED state
910 * @table: L2 table
911 * @index: First cluster index
912 * @n: Number of contiguous clusters
913 * @cluster: First cluster offset
914 *
915 * The cluster offset may be an allocated byte offset in the image file, the
916 * zero cluster marker, or the unallocated cluster marker.
917 *
918 * Called with table_lock held.
919 */
920 static void coroutine_fn qed_update_l2_table(BDRVQEDState *s, QEDTable *table,
921 int index, unsigned int n,
922 uint64_t cluster)
923 {
924 int i;
925 for (i = index; i < index + n; i++) {
926 table->offsets[i] = cluster;
927 if (!qed_offset_is_unalloc_cluster(cluster) &&
928 !qed_offset_is_zero_cluster(cluster)) {
929 cluster += s->header.cluster_size;
930 }
931 }
932 }
933
934 /* Called with table_lock held. */
935 static void coroutine_fn qed_aio_complete(QEDAIOCB *acb)
936 {
937 BDRVQEDState *s = acb_to_s(acb);
938
939 /* Free resources */
940 qemu_iovec_destroy(&acb->cur_qiov);
941 qed_unref_l2_cache_entry(acb->request.l2_table);
942
943 /* Free the buffer we may have allocated for zero writes */
944 if (acb->flags & QED_AIOCB_ZERO) {
945 qemu_vfree(acb->qiov->iov[0].iov_base);
946 acb->qiov->iov[0].iov_base = NULL;
947 }
948
949 /* Start next allocating write request waiting behind this one. Note that
950 * requests enqueue themselves when they first hit an unallocated cluster
951 * but they wait until the entire request is finished before waking up the
952 * next request in the queue. This ensures that we don't cycle through
953 * requests multiple times but rather finish one at a time completely.
954 */
955 if (acb == s->allocating_acb) {
956 s->allocating_acb = NULL;
957 if (!qemu_co_queue_empty(&s->allocating_write_reqs)) {
958 qemu_co_queue_next(&s->allocating_write_reqs);
959 } else if (s->header.features & QED_F_NEED_CHECK) {
960 qed_start_need_check_timer(s);
961 }
962 }
963 }
964
965 /**
966 * Update L1 table with new L2 table offset and write it out
967 *
968 * Called with table_lock held.
969 */
970 static int coroutine_fn qed_aio_write_l1_update(QEDAIOCB *acb)
971 {
972 BDRVQEDState *s = acb_to_s(acb);
973 CachedL2Table *l2_table = acb->request.l2_table;
974 uint64_t l2_offset = l2_table->offset;
975 int index, ret;
976
977 index = qed_l1_index(s, acb->cur_pos);
978 s->l1_table->offsets[index] = l2_table->offset;
979
980 ret = qed_write_l1_table(s, index, 1);
981
982 /* Commit the current L2 table to the cache */
983 qed_commit_l2_cache_entry(&s->l2_cache, l2_table);
984
985 /* This is guaranteed to succeed because we just committed the entry to the
986 * cache.
987 */
988 acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset);
989 assert(acb->request.l2_table != NULL);
990
991 return ret;
992 }
993
994
995 /**
996 * Update L2 table with new cluster offsets and write them out
997 *
998 * Called with table_lock held.
999 */
1000 static int coroutine_fn qed_aio_write_l2_update(QEDAIOCB *acb, uint64_t offset)
1001 {
1002 BDRVQEDState *s = acb_to_s(acb);
1003 bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1;
1004 int index, ret;
1005
1006 if (need_alloc) {
1007 qed_unref_l2_cache_entry(acb->request.l2_table);
1008 acb->request.l2_table = qed_new_l2_table(s);
1009 }
1010
1011 index = qed_l2_index(s, acb->cur_pos);
1012 qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters,
1013 offset);
1014
1015 if (need_alloc) {
1016 /* Write out the whole new L2 table */
1017 ret = qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true);
1018 if (ret) {
1019 return ret;
1020 }
1021 return qed_aio_write_l1_update(acb);
1022 } else {
1023 /* Write out only the updated part of the L2 table */
1024 ret = qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters,
1025 false);
1026 if (ret) {
1027 return ret;
1028 }
1029 }
1030 return 0;
1031 }
1032
1033 /**
1034 * Write data to the image file
1035 *
1036 * Called with table_lock *not* held.
1037 */
1038 static int coroutine_fn qed_aio_write_main(QEDAIOCB *acb)
1039 {
1040 BDRVQEDState *s = acb_to_s(acb);
1041 uint64_t offset = acb->cur_cluster +
1042 qed_offset_into_cluster(s, acb->cur_pos);
1043
1044 trace_qed_aio_write_main(s, acb, 0, offset, acb->cur_qiov.size);
1045
1046 BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO);
1047 return bdrv_co_pwritev(s->bs->file, offset, acb->cur_qiov.size,
1048 &acb->cur_qiov, 0);
1049 }
1050
1051 /**
1052 * Populate untouched regions of new data cluster
1053 *
1054 * Called with table_lock held.
1055 */
1056 static int coroutine_fn qed_aio_write_cow(QEDAIOCB *acb)
1057 {
1058 BDRVQEDState *s = acb_to_s(acb);
1059 uint64_t start, len, offset;
1060 int ret;
1061
1062 qemu_co_mutex_unlock(&s->table_lock);
1063
1064 /* Populate front untouched region of new data cluster */
1065 start = qed_start_of_cluster(s, acb->cur_pos);
1066 len = qed_offset_into_cluster(s, acb->cur_pos);
1067
1068 trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster);
1069 ret = qed_copy_from_backing_file(s, start, len, acb->cur_cluster);
1070 if (ret < 0) {
1071 goto out;
1072 }
1073
1074 /* Populate back untouched region of new data cluster */
1075 start = acb->cur_pos + acb->cur_qiov.size;
1076 len = qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start;
1077 offset = acb->cur_cluster +
1078 qed_offset_into_cluster(s, acb->cur_pos) +
1079 acb->cur_qiov.size;
1080
1081 trace_qed_aio_write_postfill(s, acb, start, len, offset);
1082 ret = qed_copy_from_backing_file(s, start, len, offset);
1083 if (ret < 0) {
1084 goto out;
1085 }
1086
1087 ret = qed_aio_write_main(acb);
1088 if (ret < 0) {
1089 goto out;
1090 }
1091
1092 if (s->bs->backing) {
1093 /*
1094 * Flush new data clusters before updating the L2 table
1095 *
1096 * This flush is necessary when a backing file is in use. A crash
1097 * during an allocating write could result in empty clusters in the
1098 * image. If the write only touched a subregion of the cluster,
1099 * then backing image sectors have been lost in the untouched
1100 * region. The solution is to flush after writing a new data
1101 * cluster and before updating the L2 table.
1102 */
1103 ret = bdrv_co_flush(s->bs->file->bs);
1104 }
1105
1106 out:
1107 qemu_co_mutex_lock(&s->table_lock);
1108 return ret;
1109 }
1110
1111 /**
1112 * Check if the QED_F_NEED_CHECK bit should be set during allocating write
1113 */
1114 static bool qed_should_set_need_check(BDRVQEDState *s)
1115 {
1116 /* The flush before L2 update path ensures consistency */
1117 if (s->bs->backing) {
1118 return false;
1119 }
1120
1121 return !(s->header.features & QED_F_NEED_CHECK);
1122 }
1123
1124 /**
1125 * Write new data cluster
1126 *
1127 * @acb: Write request
1128 * @len: Length in bytes
1129 *
1130 * This path is taken when writing to previously unallocated clusters.
1131 *
1132 * Called with table_lock held.
1133 */
1134 static int coroutine_fn qed_aio_write_alloc(QEDAIOCB *acb, size_t len)
1135 {
1136 BDRVQEDState *s = acb_to_s(acb);
1137 int ret;
1138
1139 /* Cancel timer when the first allocating request comes in */
1140 if (s->allocating_acb == NULL) {
1141 qed_cancel_need_check_timer(s);
1142 }
1143
1144 /* Freeze this request if another allocating write is in progress */
1145 if (s->allocating_acb != acb || s->allocating_write_reqs_plugged) {
1146 if (s->allocating_acb != NULL) {
1147 qemu_co_queue_wait(&s->allocating_write_reqs, &s->table_lock);
1148 assert(s->allocating_acb == NULL);
1149 }
1150 s->allocating_acb = acb;
1151 return -EAGAIN; /* start over with looking up table entries */
1152 }
1153
1154 acb->cur_nclusters = qed_bytes_to_clusters(s,
1155 qed_offset_into_cluster(s, acb->cur_pos) + len);
1156 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1157
1158 if (acb->flags & QED_AIOCB_ZERO) {
1159 /* Skip ahead if the clusters are already zero */
1160 if (acb->find_cluster_ret == QED_CLUSTER_ZERO) {
1161 return 0;
1162 }
1163 acb->cur_cluster = 1;
1164 } else {
1165 acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters);
1166 }
1167
1168 if (qed_should_set_need_check(s)) {
1169 s->header.features |= QED_F_NEED_CHECK;
1170 ret = qed_write_header(s);
1171 if (ret < 0) {
1172 return ret;
1173 }
1174 }
1175
1176 if (!(acb->flags & QED_AIOCB_ZERO)) {
1177 ret = qed_aio_write_cow(acb);
1178 if (ret < 0) {
1179 return ret;
1180 }
1181 }
1182
1183 return qed_aio_write_l2_update(acb, acb->cur_cluster);
1184 }
1185
1186 /**
1187 * Write data cluster in place
1188 *
1189 * @acb: Write request
1190 * @offset: Cluster offset in bytes
1191 * @len: Length in bytes
1192 *
1193 * This path is taken when writing to already allocated clusters.
1194 *
1195 * Called with table_lock held.
1196 */
1197 static int coroutine_fn qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset,
1198 size_t len)
1199 {
1200 BDRVQEDState *s = acb_to_s(acb);
1201 int r;
1202
1203 qemu_co_mutex_unlock(&s->table_lock);
1204
1205 /* Allocate buffer for zero writes */
1206 if (acb->flags & QED_AIOCB_ZERO) {
1207 struct iovec *iov = acb->qiov->iov;
1208
1209 if (!iov->iov_base) {
1210 iov->iov_base = qemu_try_blockalign(acb->bs, iov->iov_len);
1211 if (iov->iov_base == NULL) {
1212 r = -ENOMEM;
1213 goto out;
1214 }
1215 memset(iov->iov_base, 0, iov->iov_len);
1216 }
1217 }
1218
1219 /* Calculate the I/O vector */
1220 acb->cur_cluster = offset;
1221 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1222
1223 /* Do the actual write. */
1224 r = qed_aio_write_main(acb);
1225 out:
1226 qemu_co_mutex_lock(&s->table_lock);
1227 return r;
1228 }
1229
1230 /**
1231 * Write data cluster
1232 *
1233 * @opaque: Write request
1234 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1235 * @offset: Cluster offset in bytes
1236 * @len: Length in bytes
1237 *
1238 * Called with table_lock held.
1239 */
1240 static int coroutine_fn qed_aio_write_data(void *opaque, int ret,
1241 uint64_t offset, size_t len)
1242 {
1243 QEDAIOCB *acb = opaque;
1244
1245 trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len);
1246
1247 acb->find_cluster_ret = ret;
1248
1249 switch (ret) {
1250 case QED_CLUSTER_FOUND:
1251 return qed_aio_write_inplace(acb, offset, len);
1252
1253 case QED_CLUSTER_L2:
1254 case QED_CLUSTER_L1:
1255 case QED_CLUSTER_ZERO:
1256 return qed_aio_write_alloc(acb, len);
1257
1258 default:
1259 g_assert_not_reached();
1260 }
1261 }
1262
1263 /**
1264 * Read data cluster
1265 *
1266 * @opaque: Read request
1267 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2 or QED_CLUSTER_L1
1268 * @offset: Cluster offset in bytes
1269 * @len: Length in bytes
1270 *
1271 * Called with table_lock held.
1272 */
1273 static int coroutine_fn qed_aio_read_data(void *opaque, int ret,
1274 uint64_t offset, size_t len)
1275 {
1276 QEDAIOCB *acb = opaque;
1277 BDRVQEDState *s = acb_to_s(acb);
1278 BlockDriverState *bs = acb->bs;
1279 int r;
1280
1281 qemu_co_mutex_unlock(&s->table_lock);
1282
1283 /* Adjust offset into cluster */
1284 offset += qed_offset_into_cluster(s, acb->cur_pos);
1285
1286 trace_qed_aio_read_data(s, acb, ret, offset, len);
1287
1288 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1289
1290 /* Handle zero cluster and backing file reads, otherwise read
1291 * data cluster directly.
1292 */
1293 if (ret == QED_CLUSTER_ZERO) {
1294 qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size);
1295 r = 0;
1296 } else if (ret != QED_CLUSTER_FOUND) {
1297 r = qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov);
1298 } else {
1299 BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
1300 r = bdrv_co_preadv(bs->file, offset, acb->cur_qiov.size,
1301 &acb->cur_qiov, 0);
1302 }
1303
1304 qemu_co_mutex_lock(&s->table_lock);
1305 return r;
1306 }
1307
1308 /**
1309 * Begin next I/O or complete the request
1310 */
1311 static int coroutine_fn qed_aio_next_io(QEDAIOCB *acb)
1312 {
1313 BDRVQEDState *s = acb_to_s(acb);
1314 uint64_t offset;
1315 size_t len;
1316 int ret;
1317
1318 qemu_co_mutex_lock(&s->table_lock);
1319 while (1) {
1320 trace_qed_aio_next_io(s, acb, 0, acb->cur_pos + acb->cur_qiov.size);
1321
1322 acb->qiov_offset += acb->cur_qiov.size;
1323 acb->cur_pos += acb->cur_qiov.size;
1324 qemu_iovec_reset(&acb->cur_qiov);
1325
1326 /* Complete request */
1327 if (acb->cur_pos >= acb->end_pos) {
1328 ret = 0;
1329 break;
1330 }
1331
1332 /* Find next cluster and start I/O */
1333 len = acb->end_pos - acb->cur_pos;
1334 ret = qed_find_cluster(s, &acb->request, acb->cur_pos, &len, &offset);
1335 if (ret < 0) {
1336 break;
1337 }
1338
1339 if (acb->flags & QED_AIOCB_WRITE) {
1340 ret = qed_aio_write_data(acb, ret, offset, len);
1341 } else {
1342 ret = qed_aio_read_data(acb, ret, offset, len);
1343 }
1344
1345 if (ret < 0 && ret != -EAGAIN) {
1346 break;
1347 }
1348 }
1349
1350 trace_qed_aio_complete(s, acb, ret);
1351 qed_aio_complete(acb);
1352 qemu_co_mutex_unlock(&s->table_lock);
1353 return ret;
1354 }
1355
1356 static int coroutine_fn qed_co_request(BlockDriverState *bs, int64_t sector_num,
1357 QEMUIOVector *qiov, int nb_sectors,
1358 int flags)
1359 {
1360 QEDAIOCB acb = {
1361 .bs = bs,
1362 .cur_pos = (uint64_t) sector_num * BDRV_SECTOR_SIZE,
1363 .end_pos = (sector_num + nb_sectors) * BDRV_SECTOR_SIZE,
1364 .qiov = qiov,
1365 .flags = flags,
1366 };
1367 qemu_iovec_init(&acb.cur_qiov, qiov->niov);
1368
1369 trace_qed_aio_setup(bs->opaque, &acb, sector_num, nb_sectors, NULL, flags);
1370
1371 /* Start request */
1372 return qed_aio_next_io(&acb);
1373 }
1374
1375 static int coroutine_fn bdrv_qed_co_readv(BlockDriverState *bs,
1376 int64_t sector_num, int nb_sectors,
1377 QEMUIOVector *qiov)
1378 {
1379 return qed_co_request(bs, sector_num, qiov, nb_sectors, 0);
1380 }
1381
1382 static int coroutine_fn bdrv_qed_co_writev(BlockDriverState *bs,
1383 int64_t sector_num, int nb_sectors,
1384 QEMUIOVector *qiov, int flags)
1385 {
1386 assert(!flags);
1387 return qed_co_request(bs, sector_num, qiov, nb_sectors, QED_AIOCB_WRITE);
1388 }
1389
1390 static int coroutine_fn bdrv_qed_co_pwrite_zeroes(BlockDriverState *bs,
1391 int64_t offset,
1392 int bytes,
1393 BdrvRequestFlags flags)
1394 {
1395 BDRVQEDState *s = bs->opaque;
1396
1397 /*
1398 * Zero writes start without an I/O buffer. If a buffer becomes necessary
1399 * then it will be allocated during request processing.
1400 */
1401 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, bytes);
1402
1403 /* Fall back if the request is not aligned */
1404 if (qed_offset_into_cluster(s, offset) ||
1405 qed_offset_into_cluster(s, bytes)) {
1406 return -ENOTSUP;
1407 }
1408
1409 return qed_co_request(bs, offset >> BDRV_SECTOR_BITS, &qiov,
1410 bytes >> BDRV_SECTOR_BITS,
1411 QED_AIOCB_WRITE | QED_AIOCB_ZERO);
1412 }
1413
1414 static int coroutine_fn bdrv_qed_co_truncate(BlockDriverState *bs,
1415 int64_t offset,
1416 bool exact,
1417 PreallocMode prealloc,
1418 BdrvRequestFlags flags,
1419 Error **errp)
1420 {
1421 BDRVQEDState *s = bs->opaque;
1422 uint64_t old_image_size;
1423 int ret;
1424
1425 if (prealloc != PREALLOC_MODE_OFF) {
1426 error_setg(errp, "Unsupported preallocation mode '%s'",
1427 PreallocMode_str(prealloc));
1428 return -ENOTSUP;
1429 }
1430
1431 if (!qed_is_image_size_valid(offset, s->header.cluster_size,
1432 s->header.table_size)) {
1433 error_setg(errp, "Invalid image size specified");
1434 return -EINVAL;
1435 }
1436
1437 if ((uint64_t)offset < s->header.image_size) {
1438 error_setg(errp, "Shrinking images is currently not supported");
1439 return -ENOTSUP;
1440 }
1441
1442 old_image_size = s->header.image_size;
1443 s->header.image_size = offset;
1444 ret = qed_write_header_sync(s);
1445 if (ret < 0) {
1446 s->header.image_size = old_image_size;
1447 error_setg_errno(errp, -ret, "Failed to update the image size");
1448 }
1449 return ret;
1450 }
1451
1452 static int64_t bdrv_qed_getlength(BlockDriverState *bs)
1453 {
1454 BDRVQEDState *s = bs->opaque;
1455 return s->header.image_size;
1456 }
1457
1458 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1459 {
1460 BDRVQEDState *s = bs->opaque;
1461
1462 memset(bdi, 0, sizeof(*bdi));
1463 bdi->cluster_size = s->header.cluster_size;
1464 bdi->is_dirty = s->header.features & QED_F_NEED_CHECK;
1465 return 0;
1466 }
1467
1468 static int bdrv_qed_change_backing_file(BlockDriverState *bs,
1469 const char *backing_file,
1470 const char *backing_fmt)
1471 {
1472 BDRVQEDState *s = bs->opaque;
1473 QEDHeader new_header, le_header;
1474 void *buffer;
1475 size_t buffer_len, backing_file_len;
1476 int ret;
1477
1478 /* Refuse to set backing filename if unknown compat feature bits are
1479 * active. If the image uses an unknown compat feature then we may not
1480 * know the layout of data following the header structure and cannot safely
1481 * add a new string.
1482 */
1483 if (backing_file && (s->header.compat_features &
1484 ~QED_COMPAT_FEATURE_MASK)) {
1485 return -ENOTSUP;
1486 }
1487
1488 memcpy(&new_header, &s->header, sizeof(new_header));
1489
1490 new_header.features &= ~(QED_F_BACKING_FILE |
1491 QED_F_BACKING_FORMAT_NO_PROBE);
1492
1493 /* Adjust feature flags */
1494 if (backing_file) {
1495 new_header.features |= QED_F_BACKING_FILE;
1496
1497 if (qed_fmt_is_raw(backing_fmt)) {
1498 new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
1499 }
1500 }
1501
1502 /* Calculate new header size */
1503 backing_file_len = 0;
1504
1505 if (backing_file) {
1506 backing_file_len = strlen(backing_file);
1507 }
1508
1509 buffer_len = sizeof(new_header);
1510 new_header.backing_filename_offset = buffer_len;
1511 new_header.backing_filename_size = backing_file_len;
1512 buffer_len += backing_file_len;
1513
1514 /* Make sure we can rewrite header without failing */
1515 if (buffer_len > new_header.header_size * new_header.cluster_size) {
1516 return -ENOSPC;
1517 }
1518
1519 /* Prepare new header */
1520 buffer = g_malloc(buffer_len);
1521
1522 qed_header_cpu_to_le(&new_header, &le_header);
1523 memcpy(buffer, &le_header, sizeof(le_header));
1524 buffer_len = sizeof(le_header);
1525
1526 if (backing_file) {
1527 memcpy(buffer + buffer_len, backing_file, backing_file_len);
1528 buffer_len += backing_file_len;
1529 }
1530
1531 /* Write new header */
1532 ret = bdrv_pwrite_sync(bs->file, 0, buffer, buffer_len);
1533 g_free(buffer);
1534 if (ret == 0) {
1535 memcpy(&s->header, &new_header, sizeof(new_header));
1536 }
1537 return ret;
1538 }
1539
1540 static void coroutine_fn bdrv_qed_co_invalidate_cache(BlockDriverState *bs,
1541 Error **errp)
1542 {
1543 BDRVQEDState *s = bs->opaque;
1544 Error *local_err = NULL;
1545 int ret;
1546
1547 bdrv_qed_close(bs);
1548
1549 bdrv_qed_init_state(bs);
1550 qemu_co_mutex_lock(&s->table_lock);
1551 ret = bdrv_qed_do_open(bs, NULL, bs->open_flags, &local_err);
1552 qemu_co_mutex_unlock(&s->table_lock);
1553 if (local_err) {
1554 error_propagate_prepend(errp, local_err,
1555 "Could not reopen qed layer: ");
1556 return;
1557 } else if (ret < 0) {
1558 error_setg_errno(errp, -ret, "Could not reopen qed layer");
1559 return;
1560 }
1561 }
1562
1563 static int coroutine_fn bdrv_qed_co_check(BlockDriverState *bs,
1564 BdrvCheckResult *result,
1565 BdrvCheckMode fix)
1566 {
1567 BDRVQEDState *s = bs->opaque;
1568 int ret;
1569
1570 qemu_co_mutex_lock(&s->table_lock);
1571 ret = qed_check(s, result, !!fix);
1572 qemu_co_mutex_unlock(&s->table_lock);
1573
1574 return ret;
1575 }
1576
1577 static QemuOptsList qed_create_opts = {
1578 .name = "qed-create-opts",
1579 .head = QTAILQ_HEAD_INITIALIZER(qed_create_opts.head),
1580 .desc = {
1581 {
1582 .name = BLOCK_OPT_SIZE,
1583 .type = QEMU_OPT_SIZE,
1584 .help = "Virtual disk size"
1585 },
1586 {
1587 .name = BLOCK_OPT_BACKING_FILE,
1588 .type = QEMU_OPT_STRING,
1589 .help = "File name of a base image"
1590 },
1591 {
1592 .name = BLOCK_OPT_BACKING_FMT,
1593 .type = QEMU_OPT_STRING,
1594 .help = "Image format of the base image"
1595 },
1596 {
1597 .name = BLOCK_OPT_CLUSTER_SIZE,
1598 .type = QEMU_OPT_SIZE,
1599 .help = "Cluster size (in bytes)",
1600 .def_value_str = stringify(QED_DEFAULT_CLUSTER_SIZE)
1601 },
1602 {
1603 .name = BLOCK_OPT_TABLE_SIZE,
1604 .type = QEMU_OPT_SIZE,
1605 .help = "L1/L2 table size (in clusters)"
1606 },
1607 { /* end of list */ }
1608 }
1609 };
1610
1611 static BlockDriver bdrv_qed = {
1612 .format_name = "qed",
1613 .instance_size = sizeof(BDRVQEDState),
1614 .create_opts = &qed_create_opts,
1615 .is_format = true,
1616 .supports_backing = true,
1617
1618 .bdrv_probe = bdrv_qed_probe,
1619 .bdrv_open = bdrv_qed_open,
1620 .bdrv_close = bdrv_qed_close,
1621 .bdrv_reopen_prepare = bdrv_qed_reopen_prepare,
1622 .bdrv_child_perm = bdrv_default_perms,
1623 .bdrv_co_create = bdrv_qed_co_create,
1624 .bdrv_co_create_opts = bdrv_qed_co_create_opts,
1625 .bdrv_has_zero_init = bdrv_has_zero_init_1,
1626 .bdrv_co_block_status = bdrv_qed_co_block_status,
1627 .bdrv_co_readv = bdrv_qed_co_readv,
1628 .bdrv_co_writev = bdrv_qed_co_writev,
1629 .bdrv_co_pwrite_zeroes = bdrv_qed_co_pwrite_zeroes,
1630 .bdrv_co_truncate = bdrv_qed_co_truncate,
1631 .bdrv_getlength = bdrv_qed_getlength,
1632 .bdrv_get_info = bdrv_qed_get_info,
1633 .bdrv_refresh_limits = bdrv_qed_refresh_limits,
1634 .bdrv_change_backing_file = bdrv_qed_change_backing_file,
1635 .bdrv_co_invalidate_cache = bdrv_qed_co_invalidate_cache,
1636 .bdrv_co_check = bdrv_qed_co_check,
1637 .bdrv_detach_aio_context = bdrv_qed_detach_aio_context,
1638 .bdrv_attach_aio_context = bdrv_qed_attach_aio_context,
1639 .bdrv_co_drain_begin = bdrv_qed_co_drain_begin,
1640 };
1641
1642 static void bdrv_qed_init(void)
1643 {
1644 bdrv_register(&bdrv_qed);
1645 }
1646
1647 block_init(bdrv_qed_init);