Merge remote-tracking branch 'remotes/kraxel/tags/pull-vga-20170103-1' into staging
[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 "qapi/error.h"
17 #include "qemu/timer.h"
18 #include "qemu/bswap.h"
19 #include "trace.h"
20 #include "qed.h"
21 #include "qapi/qmp/qerror.h"
22 #include "migration/migration.h"
23 #include "sysemu/block-backend.h"
24
25 static const AIOCBInfo qed_aiocb_info = {
26 .aiocb_size = sizeof(QEDAIOCB),
27 };
28
29 static int bdrv_qed_probe(const uint8_t *buf, int buf_size,
30 const char *filename)
31 {
32 const QEDHeader *header = (const QEDHeader *)buf;
33
34 if (buf_size < sizeof(*header)) {
35 return 0;
36 }
37 if (le32_to_cpu(header->magic) != QED_MAGIC) {
38 return 0;
39 }
40 return 100;
41 }
42
43 /**
44 * Check whether an image format is raw
45 *
46 * @fmt: Backing file format, may be NULL
47 */
48 static bool qed_fmt_is_raw(const char *fmt)
49 {
50 return fmt && strcmp(fmt, "raw") == 0;
51 }
52
53 static void qed_header_le_to_cpu(const QEDHeader *le, QEDHeader *cpu)
54 {
55 cpu->magic = le32_to_cpu(le->magic);
56 cpu->cluster_size = le32_to_cpu(le->cluster_size);
57 cpu->table_size = le32_to_cpu(le->table_size);
58 cpu->header_size = le32_to_cpu(le->header_size);
59 cpu->features = le64_to_cpu(le->features);
60 cpu->compat_features = le64_to_cpu(le->compat_features);
61 cpu->autoclear_features = le64_to_cpu(le->autoclear_features);
62 cpu->l1_table_offset = le64_to_cpu(le->l1_table_offset);
63 cpu->image_size = le64_to_cpu(le->image_size);
64 cpu->backing_filename_offset = le32_to_cpu(le->backing_filename_offset);
65 cpu->backing_filename_size = le32_to_cpu(le->backing_filename_size);
66 }
67
68 static void qed_header_cpu_to_le(const QEDHeader *cpu, QEDHeader *le)
69 {
70 le->magic = cpu_to_le32(cpu->magic);
71 le->cluster_size = cpu_to_le32(cpu->cluster_size);
72 le->table_size = cpu_to_le32(cpu->table_size);
73 le->header_size = cpu_to_le32(cpu->header_size);
74 le->features = cpu_to_le64(cpu->features);
75 le->compat_features = cpu_to_le64(cpu->compat_features);
76 le->autoclear_features = cpu_to_le64(cpu->autoclear_features);
77 le->l1_table_offset = cpu_to_le64(cpu->l1_table_offset);
78 le->image_size = cpu_to_le64(cpu->image_size);
79 le->backing_filename_offset = cpu_to_le32(cpu->backing_filename_offset);
80 le->backing_filename_size = cpu_to_le32(cpu->backing_filename_size);
81 }
82
83 int qed_write_header_sync(BDRVQEDState *s)
84 {
85 QEDHeader le;
86 int ret;
87
88 qed_header_cpu_to_le(&s->header, &le);
89 ret = bdrv_pwrite(s->bs->file, 0, &le, sizeof(le));
90 if (ret != sizeof(le)) {
91 return ret;
92 }
93 return 0;
94 }
95
96 typedef struct {
97 GenericCB gencb;
98 BDRVQEDState *s;
99 struct iovec iov;
100 QEMUIOVector qiov;
101 int nsectors;
102 uint8_t *buf;
103 } QEDWriteHeaderCB;
104
105 static void qed_write_header_cb(void *opaque, int ret)
106 {
107 QEDWriteHeaderCB *write_header_cb = opaque;
108
109 qemu_vfree(write_header_cb->buf);
110 gencb_complete(write_header_cb, ret);
111 }
112
113 static void qed_write_header_read_cb(void *opaque, int ret)
114 {
115 QEDWriteHeaderCB *write_header_cb = opaque;
116 BDRVQEDState *s = write_header_cb->s;
117
118 if (ret) {
119 qed_write_header_cb(write_header_cb, ret);
120 return;
121 }
122
123 /* Update header */
124 qed_header_cpu_to_le(&s->header, (QEDHeader *)write_header_cb->buf);
125
126 bdrv_aio_writev(s->bs->file, 0, &write_header_cb->qiov,
127 write_header_cb->nsectors, qed_write_header_cb,
128 write_header_cb);
129 }
130
131 /**
132 * Update header in-place (does not rewrite backing filename or other strings)
133 *
134 * This function only updates known header fields in-place and does not affect
135 * extra data after the QED header.
136 */
137 static void qed_write_header(BDRVQEDState *s, BlockCompletionFunc cb,
138 void *opaque)
139 {
140 /* We must write full sectors for O_DIRECT but cannot necessarily generate
141 * the data following the header if an unrecognized compat feature is
142 * active. Therefore, first read the sectors containing the header, update
143 * them, and write back.
144 */
145
146 int nsectors = DIV_ROUND_UP(sizeof(QEDHeader), BDRV_SECTOR_SIZE);
147 size_t len = nsectors * BDRV_SECTOR_SIZE;
148 QEDWriteHeaderCB *write_header_cb = gencb_alloc(sizeof(*write_header_cb),
149 cb, opaque);
150
151 write_header_cb->s = s;
152 write_header_cb->nsectors = nsectors;
153 write_header_cb->buf = qemu_blockalign(s->bs, len);
154 write_header_cb->iov.iov_base = write_header_cb->buf;
155 write_header_cb->iov.iov_len = len;
156 qemu_iovec_init_external(&write_header_cb->qiov, &write_header_cb->iov, 1);
157
158 bdrv_aio_readv(s->bs->file, 0, &write_header_cb->qiov, nsectors,
159 qed_write_header_read_cb, write_header_cb);
160 }
161
162 static uint64_t qed_max_image_size(uint32_t cluster_size, uint32_t table_size)
163 {
164 uint64_t table_entries;
165 uint64_t l2_size;
166
167 table_entries = (table_size * cluster_size) / sizeof(uint64_t);
168 l2_size = table_entries * cluster_size;
169
170 return l2_size * table_entries;
171 }
172
173 static bool qed_is_cluster_size_valid(uint32_t cluster_size)
174 {
175 if (cluster_size < QED_MIN_CLUSTER_SIZE ||
176 cluster_size > QED_MAX_CLUSTER_SIZE) {
177 return false;
178 }
179 if (cluster_size & (cluster_size - 1)) {
180 return false; /* not power of 2 */
181 }
182 return true;
183 }
184
185 static bool qed_is_table_size_valid(uint32_t table_size)
186 {
187 if (table_size < QED_MIN_TABLE_SIZE ||
188 table_size > QED_MAX_TABLE_SIZE) {
189 return false;
190 }
191 if (table_size & (table_size - 1)) {
192 return false; /* not power of 2 */
193 }
194 return true;
195 }
196
197 static bool qed_is_image_size_valid(uint64_t image_size, uint32_t cluster_size,
198 uint32_t table_size)
199 {
200 if (image_size % BDRV_SECTOR_SIZE != 0) {
201 return false; /* not multiple of sector size */
202 }
203 if (image_size > qed_max_image_size(cluster_size, table_size)) {
204 return false; /* image is too large */
205 }
206 return true;
207 }
208
209 /**
210 * Read a string of known length from the image file
211 *
212 * @file: Image file
213 * @offset: File offset to start of string, in bytes
214 * @n: String length in bytes
215 * @buf: Destination buffer
216 * @buflen: Destination buffer length in bytes
217 * @ret: 0 on success, -errno on failure
218 *
219 * The string is NUL-terminated.
220 */
221 static int qed_read_string(BdrvChild *file, uint64_t offset, size_t n,
222 char *buf, size_t buflen)
223 {
224 int ret;
225 if (n >= buflen) {
226 return -EINVAL;
227 }
228 ret = bdrv_pread(file, offset, buf, n);
229 if (ret < 0) {
230 return ret;
231 }
232 buf[n] = '\0';
233 return 0;
234 }
235
236 /**
237 * Allocate new clusters
238 *
239 * @s: QED state
240 * @n: Number of contiguous clusters to allocate
241 * @ret: Offset of first allocated cluster
242 *
243 * This function only produces the offset where the new clusters should be
244 * written. It updates BDRVQEDState but does not make any changes to the image
245 * file.
246 */
247 static uint64_t qed_alloc_clusters(BDRVQEDState *s, unsigned int n)
248 {
249 uint64_t offset = s->file_size;
250 s->file_size += n * s->header.cluster_size;
251 return offset;
252 }
253
254 QEDTable *qed_alloc_table(BDRVQEDState *s)
255 {
256 /* Honor O_DIRECT memory alignment requirements */
257 return qemu_blockalign(s->bs,
258 s->header.cluster_size * s->header.table_size);
259 }
260
261 /**
262 * Allocate a new zeroed L2 table
263 */
264 static CachedL2Table *qed_new_l2_table(BDRVQEDState *s)
265 {
266 CachedL2Table *l2_table = qed_alloc_l2_cache_entry(&s->l2_cache);
267
268 l2_table->table = qed_alloc_table(s);
269 l2_table->offset = qed_alloc_clusters(s, s->header.table_size);
270
271 memset(l2_table->table->offsets, 0,
272 s->header.cluster_size * s->header.table_size);
273 return l2_table;
274 }
275
276 static void qed_aio_next_io(void *opaque, int ret);
277
278 static void qed_plug_allocating_write_reqs(BDRVQEDState *s)
279 {
280 assert(!s->allocating_write_reqs_plugged);
281
282 s->allocating_write_reqs_plugged = true;
283 }
284
285 static void qed_unplug_allocating_write_reqs(BDRVQEDState *s)
286 {
287 QEDAIOCB *acb;
288
289 assert(s->allocating_write_reqs_plugged);
290
291 s->allocating_write_reqs_plugged = false;
292
293 acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs);
294 if (acb) {
295 qed_aio_next_io(acb, 0);
296 }
297 }
298
299 static void qed_finish_clear_need_check(void *opaque, int ret)
300 {
301 /* Do nothing */
302 }
303
304 static void qed_flush_after_clear_need_check(void *opaque, int ret)
305 {
306 BDRVQEDState *s = opaque;
307
308 bdrv_aio_flush(s->bs, qed_finish_clear_need_check, s);
309
310 /* No need to wait until flush completes */
311 qed_unplug_allocating_write_reqs(s);
312 }
313
314 static void qed_clear_need_check(void *opaque, int ret)
315 {
316 BDRVQEDState *s = opaque;
317
318 if (ret) {
319 qed_unplug_allocating_write_reqs(s);
320 return;
321 }
322
323 s->header.features &= ~QED_F_NEED_CHECK;
324 qed_write_header(s, qed_flush_after_clear_need_check, s);
325 }
326
327 static void qed_need_check_timer_cb(void *opaque)
328 {
329 BDRVQEDState *s = opaque;
330
331 /* The timer should only fire when allocating writes have drained */
332 assert(!QSIMPLEQ_FIRST(&s->allocating_write_reqs));
333
334 trace_qed_need_check_timer_cb(s);
335
336 qed_plug_allocating_write_reqs(s);
337
338 /* Ensure writes are on disk before clearing flag */
339 bdrv_aio_flush(s->bs->file->bs, qed_clear_need_check, s);
340 }
341
342 static void qed_start_need_check_timer(BDRVQEDState *s)
343 {
344 trace_qed_start_need_check_timer(s);
345
346 /* Use QEMU_CLOCK_VIRTUAL so we don't alter the image file while suspended for
347 * migration.
348 */
349 timer_mod(s->need_check_timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
350 NANOSECONDS_PER_SECOND * QED_NEED_CHECK_TIMEOUT);
351 }
352
353 /* It's okay to call this multiple times or when no timer is started */
354 static void qed_cancel_need_check_timer(BDRVQEDState *s)
355 {
356 trace_qed_cancel_need_check_timer(s);
357 timer_del(s->need_check_timer);
358 }
359
360 static void bdrv_qed_detach_aio_context(BlockDriverState *bs)
361 {
362 BDRVQEDState *s = bs->opaque;
363
364 qed_cancel_need_check_timer(s);
365 timer_free(s->need_check_timer);
366 }
367
368 static void bdrv_qed_attach_aio_context(BlockDriverState *bs,
369 AioContext *new_context)
370 {
371 BDRVQEDState *s = bs->opaque;
372
373 s->need_check_timer = aio_timer_new(new_context,
374 QEMU_CLOCK_VIRTUAL, SCALE_NS,
375 qed_need_check_timer_cb, s);
376 if (s->header.features & QED_F_NEED_CHECK) {
377 qed_start_need_check_timer(s);
378 }
379 }
380
381 static void bdrv_qed_drain(BlockDriverState *bs)
382 {
383 BDRVQEDState *s = bs->opaque;
384
385 /* Fire the timer immediately in order to start doing I/O as soon as the
386 * header is flushed.
387 */
388 if (s->need_check_timer && timer_pending(s->need_check_timer)) {
389 qed_cancel_need_check_timer(s);
390 qed_need_check_timer_cb(s);
391 }
392 }
393
394 static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags,
395 Error **errp)
396 {
397 BDRVQEDState *s = bs->opaque;
398 QEDHeader le_header;
399 int64_t file_size;
400 int ret;
401
402 s->bs = bs;
403 QSIMPLEQ_INIT(&s->allocating_write_reqs);
404
405 ret = bdrv_pread(bs->file, 0, &le_header, sizeof(le_header));
406 if (ret < 0) {
407 return ret;
408 }
409 qed_header_le_to_cpu(&le_header, &s->header);
410
411 if (s->header.magic != QED_MAGIC) {
412 error_setg(errp, "Image not in QED format");
413 return -EINVAL;
414 }
415 if (s->header.features & ~QED_FEATURE_MASK) {
416 /* image uses unsupported feature bits */
417 error_setg(errp, "Unsupported QED features: %" PRIx64,
418 s->header.features & ~QED_FEATURE_MASK);
419 return -ENOTSUP;
420 }
421 if (!qed_is_cluster_size_valid(s->header.cluster_size)) {
422 return -EINVAL;
423 }
424
425 /* Round down file size to the last cluster */
426 file_size = bdrv_getlength(bs->file->bs);
427 if (file_size < 0) {
428 return file_size;
429 }
430 s->file_size = qed_start_of_cluster(s, file_size);
431
432 if (!qed_is_table_size_valid(s->header.table_size)) {
433 return -EINVAL;
434 }
435 if (!qed_is_image_size_valid(s->header.image_size,
436 s->header.cluster_size,
437 s->header.table_size)) {
438 return -EINVAL;
439 }
440 if (!qed_check_table_offset(s, s->header.l1_table_offset)) {
441 return -EINVAL;
442 }
443
444 s->table_nelems = (s->header.cluster_size * s->header.table_size) /
445 sizeof(uint64_t);
446 s->l2_shift = ctz32(s->header.cluster_size);
447 s->l2_mask = s->table_nelems - 1;
448 s->l1_shift = s->l2_shift + ctz32(s->table_nelems);
449
450 /* Header size calculation must not overflow uint32_t */
451 if (s->header.header_size > UINT32_MAX / s->header.cluster_size) {
452 return -EINVAL;
453 }
454
455 if ((s->header.features & QED_F_BACKING_FILE)) {
456 if ((uint64_t)s->header.backing_filename_offset +
457 s->header.backing_filename_size >
458 s->header.cluster_size * s->header.header_size) {
459 return -EINVAL;
460 }
461
462 ret = qed_read_string(bs->file, s->header.backing_filename_offset,
463 s->header.backing_filename_size, bs->backing_file,
464 sizeof(bs->backing_file));
465 if (ret < 0) {
466 return ret;
467 }
468
469 if (s->header.features & QED_F_BACKING_FORMAT_NO_PROBE) {
470 pstrcpy(bs->backing_format, sizeof(bs->backing_format), "raw");
471 }
472 }
473
474 /* Reset unknown autoclear feature bits. This is a backwards
475 * compatibility mechanism that allows images to be opened by older
476 * programs, which "knock out" unknown feature bits. When an image is
477 * opened by a newer program again it can detect that the autoclear
478 * feature is no longer valid.
479 */
480 if ((s->header.autoclear_features & ~QED_AUTOCLEAR_FEATURE_MASK) != 0 &&
481 !bdrv_is_read_only(bs->file->bs) && !(flags & BDRV_O_INACTIVE)) {
482 s->header.autoclear_features &= QED_AUTOCLEAR_FEATURE_MASK;
483
484 ret = qed_write_header_sync(s);
485 if (ret) {
486 return ret;
487 }
488
489 /* From here on only known autoclear feature bits are valid */
490 bdrv_flush(bs->file->bs);
491 }
492
493 s->l1_table = qed_alloc_table(s);
494 qed_init_l2_cache(&s->l2_cache);
495
496 ret = qed_read_l1_table_sync(s);
497 if (ret) {
498 goto out;
499 }
500
501 /* If image was not closed cleanly, check consistency */
502 if (!(flags & BDRV_O_CHECK) && (s->header.features & QED_F_NEED_CHECK)) {
503 /* Read-only images cannot be fixed. There is no risk of corruption
504 * since write operations are not possible. Therefore, allow
505 * potentially inconsistent images to be opened read-only. This can
506 * aid data recovery from an otherwise inconsistent image.
507 */
508 if (!bdrv_is_read_only(bs->file->bs) &&
509 !(flags & BDRV_O_INACTIVE)) {
510 BdrvCheckResult result = {0};
511
512 ret = qed_check(s, &result, true);
513 if (ret) {
514 goto out;
515 }
516 }
517 }
518
519 bdrv_qed_attach_aio_context(bs, bdrv_get_aio_context(bs));
520
521 out:
522 if (ret) {
523 qed_free_l2_cache(&s->l2_cache);
524 qemu_vfree(s->l1_table);
525 }
526 return ret;
527 }
528
529 static void bdrv_qed_refresh_limits(BlockDriverState *bs, Error **errp)
530 {
531 BDRVQEDState *s = bs->opaque;
532
533 bs->bl.pwrite_zeroes_alignment = s->header.cluster_size;
534 }
535
536 /* We have nothing to do for QED reopen, stubs just return
537 * success */
538 static int bdrv_qed_reopen_prepare(BDRVReopenState *state,
539 BlockReopenQueue *queue, Error **errp)
540 {
541 return 0;
542 }
543
544 static void bdrv_qed_close(BlockDriverState *bs)
545 {
546 BDRVQEDState *s = bs->opaque;
547
548 bdrv_qed_detach_aio_context(bs);
549
550 /* Ensure writes reach stable storage */
551 bdrv_flush(bs->file->bs);
552
553 /* Clean shutdown, no check required on next open */
554 if (s->header.features & QED_F_NEED_CHECK) {
555 s->header.features &= ~QED_F_NEED_CHECK;
556 qed_write_header_sync(s);
557 }
558
559 qed_free_l2_cache(&s->l2_cache);
560 qemu_vfree(s->l1_table);
561 }
562
563 static int qed_create(const char *filename, uint32_t cluster_size,
564 uint64_t image_size, uint32_t table_size,
565 const char *backing_file, const char *backing_fmt,
566 QemuOpts *opts, Error **errp)
567 {
568 QEDHeader header = {
569 .magic = QED_MAGIC,
570 .cluster_size = cluster_size,
571 .table_size = table_size,
572 .header_size = 1,
573 .features = 0,
574 .compat_features = 0,
575 .l1_table_offset = cluster_size,
576 .image_size = image_size,
577 };
578 QEDHeader le_header;
579 uint8_t *l1_table = NULL;
580 size_t l1_size = header.cluster_size * header.table_size;
581 Error *local_err = NULL;
582 int ret = 0;
583 BlockBackend *blk;
584
585 ret = bdrv_create_file(filename, opts, &local_err);
586 if (ret < 0) {
587 error_propagate(errp, local_err);
588 return ret;
589 }
590
591 blk = blk_new_open(filename, NULL, NULL,
592 BDRV_O_RDWR | BDRV_O_PROTOCOL, &local_err);
593 if (blk == NULL) {
594 error_propagate(errp, local_err);
595 return -EIO;
596 }
597
598 blk_set_allow_write_beyond_eof(blk, true);
599
600 /* File must start empty and grow, check truncate is supported */
601 ret = blk_truncate(blk, 0);
602 if (ret < 0) {
603 goto out;
604 }
605
606 if (backing_file) {
607 header.features |= QED_F_BACKING_FILE;
608 header.backing_filename_offset = sizeof(le_header);
609 header.backing_filename_size = strlen(backing_file);
610
611 if (qed_fmt_is_raw(backing_fmt)) {
612 header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
613 }
614 }
615
616 qed_header_cpu_to_le(&header, &le_header);
617 ret = blk_pwrite(blk, 0, &le_header, sizeof(le_header), 0);
618 if (ret < 0) {
619 goto out;
620 }
621 ret = blk_pwrite(blk, sizeof(le_header), backing_file,
622 header.backing_filename_size, 0);
623 if (ret < 0) {
624 goto out;
625 }
626
627 l1_table = g_malloc0(l1_size);
628 ret = blk_pwrite(blk, header.l1_table_offset, l1_table, l1_size, 0);
629 if (ret < 0) {
630 goto out;
631 }
632
633 ret = 0; /* success */
634 out:
635 g_free(l1_table);
636 blk_unref(blk);
637 return ret;
638 }
639
640 static int bdrv_qed_create(const char *filename, QemuOpts *opts, Error **errp)
641 {
642 uint64_t image_size = 0;
643 uint32_t cluster_size = QED_DEFAULT_CLUSTER_SIZE;
644 uint32_t table_size = QED_DEFAULT_TABLE_SIZE;
645 char *backing_file = NULL;
646 char *backing_fmt = NULL;
647 int ret;
648
649 image_size = ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
650 BDRV_SECTOR_SIZE);
651 backing_file = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FILE);
652 backing_fmt = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FMT);
653 cluster_size = qemu_opt_get_size_del(opts,
654 BLOCK_OPT_CLUSTER_SIZE,
655 QED_DEFAULT_CLUSTER_SIZE);
656 table_size = qemu_opt_get_size_del(opts, BLOCK_OPT_TABLE_SIZE,
657 QED_DEFAULT_TABLE_SIZE);
658
659 if (!qed_is_cluster_size_valid(cluster_size)) {
660 error_setg(errp, "QED cluster size must be within range [%u, %u] "
661 "and power of 2",
662 QED_MIN_CLUSTER_SIZE, QED_MAX_CLUSTER_SIZE);
663 ret = -EINVAL;
664 goto finish;
665 }
666 if (!qed_is_table_size_valid(table_size)) {
667 error_setg(errp, "QED table size must be within range [%u, %u] "
668 "and power of 2",
669 QED_MIN_TABLE_SIZE, QED_MAX_TABLE_SIZE);
670 ret = -EINVAL;
671 goto finish;
672 }
673 if (!qed_is_image_size_valid(image_size, cluster_size, table_size)) {
674 error_setg(errp, "QED image size must be a non-zero multiple of "
675 "cluster size and less than %" PRIu64 " bytes",
676 qed_max_image_size(cluster_size, table_size));
677 ret = -EINVAL;
678 goto finish;
679 }
680
681 ret = qed_create(filename, cluster_size, image_size, table_size,
682 backing_file, backing_fmt, opts, errp);
683
684 finish:
685 g_free(backing_file);
686 g_free(backing_fmt);
687 return ret;
688 }
689
690 typedef struct {
691 BlockDriverState *bs;
692 Coroutine *co;
693 uint64_t pos;
694 int64_t status;
695 int *pnum;
696 BlockDriverState **file;
697 } QEDIsAllocatedCB;
698
699 static void qed_is_allocated_cb(void *opaque, int ret, uint64_t offset, size_t len)
700 {
701 QEDIsAllocatedCB *cb = opaque;
702 BDRVQEDState *s = cb->bs->opaque;
703 *cb->pnum = len / BDRV_SECTOR_SIZE;
704 switch (ret) {
705 case QED_CLUSTER_FOUND:
706 offset |= qed_offset_into_cluster(s, cb->pos);
707 cb->status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID | offset;
708 *cb->file = cb->bs->file->bs;
709 break;
710 case QED_CLUSTER_ZERO:
711 cb->status = BDRV_BLOCK_ZERO;
712 break;
713 case QED_CLUSTER_L2:
714 case QED_CLUSTER_L1:
715 cb->status = 0;
716 break;
717 default:
718 assert(ret < 0);
719 cb->status = ret;
720 break;
721 }
722
723 if (cb->co) {
724 qemu_coroutine_enter(cb->co);
725 }
726 }
727
728 static int64_t coroutine_fn bdrv_qed_co_get_block_status(BlockDriverState *bs,
729 int64_t sector_num,
730 int nb_sectors, int *pnum,
731 BlockDriverState **file)
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 .file = file,
741 };
742 QEDRequest request = { .l2_table = NULL };
743
744 qed_find_cluster(s, &request, cb.pos, len, qed_is_allocated_cb, &cb);
745
746 /* Now sleep if the callback wasn't invoked immediately */
747 while (cb.status == BDRV_BLOCK_OFFSET_MASK) {
748 cb.co = qemu_coroutine_self();
749 qemu_coroutine_yield();
750 }
751
752 qed_unref_l2_cache_entry(request.l2_table);
753
754 return cb.status;
755 }
756
757 static BDRVQEDState *acb_to_s(QEDAIOCB *acb)
758 {
759 return acb->common.bs->opaque;
760 }
761
762 /**
763 * Read from the backing file or zero-fill if no backing file
764 *
765 * @s: QED state
766 * @pos: Byte position in device
767 * @qiov: Destination I/O vector
768 * @backing_qiov: Possibly shortened copy of qiov, to be allocated here
769 * @cb: Completion function
770 * @opaque: User data for completion function
771 *
772 * This function reads qiov->size bytes starting at pos from the backing file.
773 * If there is no backing file then zeroes are read.
774 */
775 static void qed_read_backing_file(BDRVQEDState *s, uint64_t pos,
776 QEMUIOVector *qiov,
777 QEMUIOVector **backing_qiov,
778 BlockCompletionFunc *cb, void *opaque)
779 {
780 uint64_t backing_length = 0;
781 size_t size;
782
783 /* If there is a backing file, get its length. Treat the absence of a
784 * backing file like a zero length backing file.
785 */
786 if (s->bs->backing) {
787 int64_t l = bdrv_getlength(s->bs->backing->bs);
788 if (l < 0) {
789 cb(opaque, l);
790 return;
791 }
792 backing_length = l;
793 }
794
795 /* Zero all sectors if reading beyond the end of the backing file */
796 if (pos >= backing_length ||
797 pos + qiov->size > backing_length) {
798 qemu_iovec_memset(qiov, 0, 0, qiov->size);
799 }
800
801 /* Complete now if there are no backing file sectors to read */
802 if (pos >= backing_length) {
803 cb(opaque, 0);
804 return;
805 }
806
807 /* If the read straddles the end of the backing file, shorten it */
808 size = MIN((uint64_t)backing_length - pos, qiov->size);
809
810 assert(*backing_qiov == NULL);
811 *backing_qiov = g_new(QEMUIOVector, 1);
812 qemu_iovec_init(*backing_qiov, qiov->niov);
813 qemu_iovec_concat(*backing_qiov, qiov, 0, size);
814
815 BLKDBG_EVENT(s->bs->file, BLKDBG_READ_BACKING_AIO);
816 bdrv_aio_readv(s->bs->backing, pos / BDRV_SECTOR_SIZE,
817 *backing_qiov, size / BDRV_SECTOR_SIZE, cb, opaque);
818 }
819
820 typedef struct {
821 GenericCB gencb;
822 BDRVQEDState *s;
823 QEMUIOVector qiov;
824 QEMUIOVector *backing_qiov;
825 struct iovec iov;
826 uint64_t offset;
827 } CopyFromBackingFileCB;
828
829 static void qed_copy_from_backing_file_cb(void *opaque, int ret)
830 {
831 CopyFromBackingFileCB *copy_cb = opaque;
832 qemu_vfree(copy_cb->iov.iov_base);
833 gencb_complete(&copy_cb->gencb, ret);
834 }
835
836 static void qed_copy_from_backing_file_write(void *opaque, int ret)
837 {
838 CopyFromBackingFileCB *copy_cb = opaque;
839 BDRVQEDState *s = copy_cb->s;
840
841 if (copy_cb->backing_qiov) {
842 qemu_iovec_destroy(copy_cb->backing_qiov);
843 g_free(copy_cb->backing_qiov);
844 copy_cb->backing_qiov = NULL;
845 }
846
847 if (ret) {
848 qed_copy_from_backing_file_cb(copy_cb, ret);
849 return;
850 }
851
852 BLKDBG_EVENT(s->bs->file, BLKDBG_COW_WRITE);
853 bdrv_aio_writev(s->bs->file, copy_cb->offset / BDRV_SECTOR_SIZE,
854 &copy_cb->qiov, copy_cb->qiov.size / BDRV_SECTOR_SIZE,
855 qed_copy_from_backing_file_cb, copy_cb);
856 }
857
858 /**
859 * Copy data from backing file into the image
860 *
861 * @s: QED state
862 * @pos: Byte position in device
863 * @len: Number of bytes
864 * @offset: Byte offset in image file
865 * @cb: Completion function
866 * @opaque: User data for completion function
867 */
868 static void qed_copy_from_backing_file(BDRVQEDState *s, uint64_t pos,
869 uint64_t len, uint64_t offset,
870 BlockCompletionFunc *cb,
871 void *opaque)
872 {
873 CopyFromBackingFileCB *copy_cb;
874
875 /* Skip copy entirely if there is no work to do */
876 if (len == 0) {
877 cb(opaque, 0);
878 return;
879 }
880
881 copy_cb = gencb_alloc(sizeof(*copy_cb), cb, opaque);
882 copy_cb->s = s;
883 copy_cb->offset = offset;
884 copy_cb->backing_qiov = NULL;
885 copy_cb->iov.iov_base = qemu_blockalign(s->bs, len);
886 copy_cb->iov.iov_len = len;
887 qemu_iovec_init_external(&copy_cb->qiov, &copy_cb->iov, 1);
888
889 qed_read_backing_file(s, pos, &copy_cb->qiov, &copy_cb->backing_qiov,
890 qed_copy_from_backing_file_write, copy_cb);
891 }
892
893 /**
894 * Link one or more contiguous clusters into a table
895 *
896 * @s: QED state
897 * @table: L2 table
898 * @index: First cluster index
899 * @n: Number of contiguous clusters
900 * @cluster: First cluster offset
901 *
902 * The cluster offset may be an allocated byte offset in the image file, the
903 * zero cluster marker, or the unallocated cluster marker.
904 */
905 static void qed_update_l2_table(BDRVQEDState *s, QEDTable *table, int index,
906 unsigned int n, uint64_t cluster)
907 {
908 int i;
909 for (i = index; i < index + n; i++) {
910 table->offsets[i] = cluster;
911 if (!qed_offset_is_unalloc_cluster(cluster) &&
912 !qed_offset_is_zero_cluster(cluster)) {
913 cluster += s->header.cluster_size;
914 }
915 }
916 }
917
918 static void qed_aio_complete_bh(void *opaque)
919 {
920 QEDAIOCB *acb = opaque;
921 BlockCompletionFunc *cb = acb->common.cb;
922 void *user_opaque = acb->common.opaque;
923 int ret = acb->bh_ret;
924
925 qemu_aio_unref(acb);
926
927 /* Invoke callback */
928 cb(user_opaque, ret);
929 }
930
931 static void qed_aio_complete(QEDAIOCB *acb, int ret)
932 {
933 BDRVQEDState *s = acb_to_s(acb);
934
935 trace_qed_aio_complete(s, acb, ret);
936
937 /* Free resources */
938 qemu_iovec_destroy(&acb->cur_qiov);
939 qed_unref_l2_cache_entry(acb->request.l2_table);
940
941 /* Free the buffer we may have allocated for zero writes */
942 if (acb->flags & QED_AIOCB_ZERO) {
943 qemu_vfree(acb->qiov->iov[0].iov_base);
944 acb->qiov->iov[0].iov_base = NULL;
945 }
946
947 /* Arrange for a bh to invoke the completion function */
948 acb->bh_ret = ret;
949 aio_bh_schedule_oneshot(bdrv_get_aio_context(acb->common.bs),
950 qed_aio_complete_bh, acb);
951
952 /* Start next allocating write request waiting behind this one. Note that
953 * requests enqueue themselves when they first hit an unallocated cluster
954 * but they wait until the entire request is finished before waking up the
955 * next request in the queue. This ensures that we don't cycle through
956 * requests multiple times but rather finish one at a time completely.
957 */
958 if (acb == QSIMPLEQ_FIRST(&s->allocating_write_reqs)) {
959 QSIMPLEQ_REMOVE_HEAD(&s->allocating_write_reqs, next);
960 acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs);
961 if (acb) {
962 qed_aio_next_io(acb, 0);
963 } else if (s->header.features & QED_F_NEED_CHECK) {
964 qed_start_need_check_timer(s);
965 }
966 }
967 }
968
969 /**
970 * Commit the current L2 table to the cache
971 */
972 static void qed_commit_l2_update(void *opaque, int ret)
973 {
974 QEDAIOCB *acb = opaque;
975 BDRVQEDState *s = acb_to_s(acb);
976 CachedL2Table *l2_table = acb->request.l2_table;
977 uint64_t l2_offset = l2_table->offset;
978
979 qed_commit_l2_cache_entry(&s->l2_cache, l2_table);
980
981 /* This is guaranteed to succeed because we just committed the entry to the
982 * cache.
983 */
984 acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset);
985 assert(acb->request.l2_table != NULL);
986
987 qed_aio_next_io(opaque, ret);
988 }
989
990 /**
991 * Update L1 table with new L2 table offset and write it out
992 */
993 static void qed_aio_write_l1_update(void *opaque, int ret)
994 {
995 QEDAIOCB *acb = opaque;
996 BDRVQEDState *s = acb_to_s(acb);
997 int index;
998
999 if (ret) {
1000 qed_aio_complete(acb, ret);
1001 return;
1002 }
1003
1004 index = qed_l1_index(s, acb->cur_pos);
1005 s->l1_table->offsets[index] = acb->request.l2_table->offset;
1006
1007 qed_write_l1_table(s, index, 1, qed_commit_l2_update, acb);
1008 }
1009
1010 /**
1011 * Update L2 table with new cluster offsets and write them out
1012 */
1013 static void qed_aio_write_l2_update(QEDAIOCB *acb, int ret, uint64_t offset)
1014 {
1015 BDRVQEDState *s = acb_to_s(acb);
1016 bool need_alloc = acb->find_cluster_ret == QED_CLUSTER_L1;
1017 int index;
1018
1019 if (ret) {
1020 goto err;
1021 }
1022
1023 if (need_alloc) {
1024 qed_unref_l2_cache_entry(acb->request.l2_table);
1025 acb->request.l2_table = qed_new_l2_table(s);
1026 }
1027
1028 index = qed_l2_index(s, acb->cur_pos);
1029 qed_update_l2_table(s, acb->request.l2_table->table, index, acb->cur_nclusters,
1030 offset);
1031
1032 if (need_alloc) {
1033 /* Write out the whole new L2 table */
1034 qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true,
1035 qed_aio_write_l1_update, acb);
1036 } else {
1037 /* Write out only the updated part of the L2 table */
1038 qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters, false,
1039 qed_aio_next_io, acb);
1040 }
1041 return;
1042
1043 err:
1044 qed_aio_complete(acb, ret);
1045 }
1046
1047 static void qed_aio_write_l2_update_cb(void *opaque, int ret)
1048 {
1049 QEDAIOCB *acb = opaque;
1050 qed_aio_write_l2_update(acb, ret, acb->cur_cluster);
1051 }
1052
1053 /**
1054 * Flush new data clusters before updating the L2 table
1055 *
1056 * This flush is necessary when a backing file is in use. A crash during an
1057 * allocating write could result in empty clusters in the image. If the write
1058 * only touched a subregion of the cluster, then backing image sectors have
1059 * been lost in the untouched region. The solution is to flush after writing a
1060 * new data cluster and before updating the L2 table.
1061 */
1062 static void qed_aio_write_flush_before_l2_update(void *opaque, int ret)
1063 {
1064 QEDAIOCB *acb = opaque;
1065 BDRVQEDState *s = acb_to_s(acb);
1066
1067 if (!bdrv_aio_flush(s->bs->file->bs, qed_aio_write_l2_update_cb, opaque)) {
1068 qed_aio_complete(acb, -EIO);
1069 }
1070 }
1071
1072 /**
1073 * Write data to the image file
1074 */
1075 static void qed_aio_write_main(void *opaque, int ret)
1076 {
1077 QEDAIOCB *acb = opaque;
1078 BDRVQEDState *s = acb_to_s(acb);
1079 uint64_t offset = acb->cur_cluster +
1080 qed_offset_into_cluster(s, acb->cur_pos);
1081 BlockCompletionFunc *next_fn;
1082
1083 trace_qed_aio_write_main(s, acb, ret, offset, acb->cur_qiov.size);
1084
1085 if (ret) {
1086 qed_aio_complete(acb, ret);
1087 return;
1088 }
1089
1090 if (acb->find_cluster_ret == QED_CLUSTER_FOUND) {
1091 next_fn = qed_aio_next_io;
1092 } else {
1093 if (s->bs->backing) {
1094 next_fn = qed_aio_write_flush_before_l2_update;
1095 } else {
1096 next_fn = qed_aio_write_l2_update_cb;
1097 }
1098 }
1099
1100 BLKDBG_EVENT(s->bs->file, BLKDBG_WRITE_AIO);
1101 bdrv_aio_writev(s->bs->file, offset / BDRV_SECTOR_SIZE,
1102 &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE,
1103 next_fn, acb);
1104 }
1105
1106 /**
1107 * Populate back untouched region of new data cluster
1108 */
1109 static void qed_aio_write_postfill(void *opaque, int ret)
1110 {
1111 QEDAIOCB *acb = opaque;
1112 BDRVQEDState *s = acb_to_s(acb);
1113 uint64_t start = acb->cur_pos + acb->cur_qiov.size;
1114 uint64_t len =
1115 qed_start_of_cluster(s, start + s->header.cluster_size - 1) - start;
1116 uint64_t offset = acb->cur_cluster +
1117 qed_offset_into_cluster(s, acb->cur_pos) +
1118 acb->cur_qiov.size;
1119
1120 if (ret) {
1121 qed_aio_complete(acb, ret);
1122 return;
1123 }
1124
1125 trace_qed_aio_write_postfill(s, acb, start, len, offset);
1126 qed_copy_from_backing_file(s, start, len, offset,
1127 qed_aio_write_main, acb);
1128 }
1129
1130 /**
1131 * Populate front untouched region of new data cluster
1132 */
1133 static void qed_aio_write_prefill(void *opaque, int ret)
1134 {
1135 QEDAIOCB *acb = opaque;
1136 BDRVQEDState *s = acb_to_s(acb);
1137 uint64_t start = qed_start_of_cluster(s, acb->cur_pos);
1138 uint64_t len = qed_offset_into_cluster(s, acb->cur_pos);
1139
1140 trace_qed_aio_write_prefill(s, acb, start, len, acb->cur_cluster);
1141 qed_copy_from_backing_file(s, start, len, acb->cur_cluster,
1142 qed_aio_write_postfill, acb);
1143 }
1144
1145 /**
1146 * Check if the QED_F_NEED_CHECK bit should be set during allocating write
1147 */
1148 static bool qed_should_set_need_check(BDRVQEDState *s)
1149 {
1150 /* The flush before L2 update path ensures consistency */
1151 if (s->bs->backing) {
1152 return false;
1153 }
1154
1155 return !(s->header.features & QED_F_NEED_CHECK);
1156 }
1157
1158 static void qed_aio_write_zero_cluster(void *opaque, int ret)
1159 {
1160 QEDAIOCB *acb = opaque;
1161
1162 if (ret) {
1163 qed_aio_complete(acb, ret);
1164 return;
1165 }
1166
1167 qed_aio_write_l2_update(acb, 0, 1);
1168 }
1169
1170 /**
1171 * Write new data cluster
1172 *
1173 * @acb: Write request
1174 * @len: Length in bytes
1175 *
1176 * This path is taken when writing to previously unallocated clusters.
1177 */
1178 static void qed_aio_write_alloc(QEDAIOCB *acb, size_t len)
1179 {
1180 BDRVQEDState *s = acb_to_s(acb);
1181 BlockCompletionFunc *cb;
1182
1183 /* Cancel timer when the first allocating request comes in */
1184 if (QSIMPLEQ_EMPTY(&s->allocating_write_reqs)) {
1185 qed_cancel_need_check_timer(s);
1186 }
1187
1188 /* Freeze this request if another allocating write is in progress */
1189 if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs)) {
1190 QSIMPLEQ_INSERT_TAIL(&s->allocating_write_reqs, acb, next);
1191 }
1192 if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs) ||
1193 s->allocating_write_reqs_plugged) {
1194 return; /* wait for existing request to finish */
1195 }
1196
1197 acb->cur_nclusters = qed_bytes_to_clusters(s,
1198 qed_offset_into_cluster(s, acb->cur_pos) + len);
1199 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1200
1201 if (acb->flags & QED_AIOCB_ZERO) {
1202 /* Skip ahead if the clusters are already zero */
1203 if (acb->find_cluster_ret == QED_CLUSTER_ZERO) {
1204 qed_aio_next_io(acb, 0);
1205 return;
1206 }
1207
1208 cb = qed_aio_write_zero_cluster;
1209 } else {
1210 cb = qed_aio_write_prefill;
1211 acb->cur_cluster = qed_alloc_clusters(s, acb->cur_nclusters);
1212 }
1213
1214 if (qed_should_set_need_check(s)) {
1215 s->header.features |= QED_F_NEED_CHECK;
1216 qed_write_header(s, cb, acb);
1217 } else {
1218 cb(acb, 0);
1219 }
1220 }
1221
1222 /**
1223 * Write data cluster in place
1224 *
1225 * @acb: Write request
1226 * @offset: Cluster offset in bytes
1227 * @len: Length in bytes
1228 *
1229 * This path is taken when writing to already allocated clusters.
1230 */
1231 static void qed_aio_write_inplace(QEDAIOCB *acb, uint64_t offset, size_t len)
1232 {
1233 /* Allocate buffer for zero writes */
1234 if (acb->flags & QED_AIOCB_ZERO) {
1235 struct iovec *iov = acb->qiov->iov;
1236
1237 if (!iov->iov_base) {
1238 iov->iov_base = qemu_try_blockalign(acb->common.bs, iov->iov_len);
1239 if (iov->iov_base == NULL) {
1240 qed_aio_complete(acb, -ENOMEM);
1241 return;
1242 }
1243 memset(iov->iov_base, 0, iov->iov_len);
1244 }
1245 }
1246
1247 /* Calculate the I/O vector */
1248 acb->cur_cluster = offset;
1249 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1250
1251 /* Do the actual write */
1252 qed_aio_write_main(acb, 0);
1253 }
1254
1255 /**
1256 * Write data cluster
1257 *
1258 * @opaque: Write request
1259 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1,
1260 * or -errno
1261 * @offset: Cluster offset in bytes
1262 * @len: Length in bytes
1263 *
1264 * Callback from qed_find_cluster().
1265 */
1266 static void qed_aio_write_data(void *opaque, int ret,
1267 uint64_t offset, size_t len)
1268 {
1269 QEDAIOCB *acb = opaque;
1270
1271 trace_qed_aio_write_data(acb_to_s(acb), acb, ret, offset, len);
1272
1273 acb->find_cluster_ret = ret;
1274
1275 switch (ret) {
1276 case QED_CLUSTER_FOUND:
1277 qed_aio_write_inplace(acb, offset, len);
1278 break;
1279
1280 case QED_CLUSTER_L2:
1281 case QED_CLUSTER_L1:
1282 case QED_CLUSTER_ZERO:
1283 qed_aio_write_alloc(acb, len);
1284 break;
1285
1286 default:
1287 qed_aio_complete(acb, ret);
1288 break;
1289 }
1290 }
1291
1292 /**
1293 * Read data cluster
1294 *
1295 * @opaque: Read request
1296 * @ret: QED_CLUSTER_FOUND, QED_CLUSTER_L2, QED_CLUSTER_L1,
1297 * or -errno
1298 * @offset: Cluster offset in bytes
1299 * @len: Length in bytes
1300 *
1301 * Callback from qed_find_cluster().
1302 */
1303 static void qed_aio_read_data(void *opaque, int ret,
1304 uint64_t offset, size_t len)
1305 {
1306 QEDAIOCB *acb = opaque;
1307 BDRVQEDState *s = acb_to_s(acb);
1308 BlockDriverState *bs = acb->common.bs;
1309
1310 /* Adjust offset into cluster */
1311 offset += qed_offset_into_cluster(s, acb->cur_pos);
1312
1313 trace_qed_aio_read_data(s, acb, ret, offset, len);
1314
1315 if (ret < 0) {
1316 goto err;
1317 }
1318
1319 qemu_iovec_concat(&acb->cur_qiov, acb->qiov, acb->qiov_offset, len);
1320
1321 /* Handle zero cluster and backing file reads */
1322 if (ret == QED_CLUSTER_ZERO) {
1323 qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size);
1324 qed_aio_next_io(acb, 0);
1325 return;
1326 } else if (ret != QED_CLUSTER_FOUND) {
1327 qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov,
1328 &acb->backing_qiov, qed_aio_next_io, acb);
1329 return;
1330 }
1331
1332 BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
1333 bdrv_aio_readv(bs->file, offset / BDRV_SECTOR_SIZE,
1334 &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE,
1335 qed_aio_next_io, acb);
1336 return;
1337
1338 err:
1339 qed_aio_complete(acb, ret);
1340 }
1341
1342 /**
1343 * Begin next I/O or complete the request
1344 */
1345 static void qed_aio_next_io(void *opaque, int ret)
1346 {
1347 QEDAIOCB *acb = opaque;
1348 BDRVQEDState *s = acb_to_s(acb);
1349 QEDFindClusterFunc *io_fn = (acb->flags & QED_AIOCB_WRITE) ?
1350 qed_aio_write_data : qed_aio_read_data;
1351
1352 trace_qed_aio_next_io(s, acb, ret, acb->cur_pos + acb->cur_qiov.size);
1353
1354 if (acb->backing_qiov) {
1355 qemu_iovec_destroy(acb->backing_qiov);
1356 g_free(acb->backing_qiov);
1357 acb->backing_qiov = NULL;
1358 }
1359
1360 /* Handle I/O error */
1361 if (ret) {
1362 qed_aio_complete(acb, ret);
1363 return;
1364 }
1365
1366 acb->qiov_offset += acb->cur_qiov.size;
1367 acb->cur_pos += acb->cur_qiov.size;
1368 qemu_iovec_reset(&acb->cur_qiov);
1369
1370 /* Complete request */
1371 if (acb->cur_pos >= acb->end_pos) {
1372 qed_aio_complete(acb, 0);
1373 return;
1374 }
1375
1376 /* Find next cluster and start I/O */
1377 qed_find_cluster(s, &acb->request,
1378 acb->cur_pos, acb->end_pos - acb->cur_pos,
1379 io_fn, acb);
1380 }
1381
1382 static BlockAIOCB *qed_aio_setup(BlockDriverState *bs,
1383 int64_t sector_num,
1384 QEMUIOVector *qiov, int nb_sectors,
1385 BlockCompletionFunc *cb,
1386 void *opaque, int flags)
1387 {
1388 QEDAIOCB *acb = qemu_aio_get(&qed_aiocb_info, bs, cb, opaque);
1389
1390 trace_qed_aio_setup(bs->opaque, acb, sector_num, nb_sectors,
1391 opaque, flags);
1392
1393 acb->flags = flags;
1394 acb->qiov = qiov;
1395 acb->qiov_offset = 0;
1396 acb->cur_pos = (uint64_t)sector_num * BDRV_SECTOR_SIZE;
1397 acb->end_pos = acb->cur_pos + nb_sectors * BDRV_SECTOR_SIZE;
1398 acb->backing_qiov = NULL;
1399 acb->request.l2_table = NULL;
1400 qemu_iovec_init(&acb->cur_qiov, qiov->niov);
1401
1402 /* Start request */
1403 qed_aio_next_io(acb, 0);
1404 return &acb->common;
1405 }
1406
1407 static BlockAIOCB *bdrv_qed_aio_readv(BlockDriverState *bs,
1408 int64_t sector_num,
1409 QEMUIOVector *qiov, int nb_sectors,
1410 BlockCompletionFunc *cb,
1411 void *opaque)
1412 {
1413 return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb, opaque, 0);
1414 }
1415
1416 static BlockAIOCB *bdrv_qed_aio_writev(BlockDriverState *bs,
1417 int64_t sector_num,
1418 QEMUIOVector *qiov, int nb_sectors,
1419 BlockCompletionFunc *cb,
1420 void *opaque)
1421 {
1422 return qed_aio_setup(bs, sector_num, qiov, nb_sectors, cb,
1423 opaque, QED_AIOCB_WRITE);
1424 }
1425
1426 typedef struct {
1427 Coroutine *co;
1428 int ret;
1429 bool done;
1430 } QEDWriteZeroesCB;
1431
1432 static void coroutine_fn qed_co_pwrite_zeroes_cb(void *opaque, int ret)
1433 {
1434 QEDWriteZeroesCB *cb = opaque;
1435
1436 cb->done = true;
1437 cb->ret = ret;
1438 if (cb->co) {
1439 qemu_coroutine_enter(cb->co);
1440 }
1441 }
1442
1443 static int coroutine_fn bdrv_qed_co_pwrite_zeroes(BlockDriverState *bs,
1444 int64_t offset,
1445 int count,
1446 BdrvRequestFlags flags)
1447 {
1448 BlockAIOCB *blockacb;
1449 BDRVQEDState *s = bs->opaque;
1450 QEDWriteZeroesCB cb = { .done = false };
1451 QEMUIOVector qiov;
1452 struct iovec iov;
1453
1454 /* Fall back if the request is not aligned */
1455 if (qed_offset_into_cluster(s, offset) ||
1456 qed_offset_into_cluster(s, count)) {
1457 return -ENOTSUP;
1458 }
1459
1460 /* Zero writes start without an I/O buffer. If a buffer becomes necessary
1461 * then it will be allocated during request processing.
1462 */
1463 iov.iov_base = NULL;
1464 iov.iov_len = count;
1465
1466 qemu_iovec_init_external(&qiov, &iov, 1);
1467 blockacb = qed_aio_setup(bs, offset >> BDRV_SECTOR_BITS, &qiov,
1468 count >> BDRV_SECTOR_BITS,
1469 qed_co_pwrite_zeroes_cb, &cb,
1470 QED_AIOCB_WRITE | QED_AIOCB_ZERO);
1471 if (!blockacb) {
1472 return -EIO;
1473 }
1474 if (!cb.done) {
1475 cb.co = qemu_coroutine_self();
1476 qemu_coroutine_yield();
1477 }
1478 assert(cb.done);
1479 return cb.ret;
1480 }
1481
1482 static int bdrv_qed_truncate(BlockDriverState *bs, int64_t offset)
1483 {
1484 BDRVQEDState *s = bs->opaque;
1485 uint64_t old_image_size;
1486 int ret;
1487
1488 if (!qed_is_image_size_valid(offset, s->header.cluster_size,
1489 s->header.table_size)) {
1490 return -EINVAL;
1491 }
1492
1493 /* Shrinking is currently not supported */
1494 if ((uint64_t)offset < s->header.image_size) {
1495 return -ENOTSUP;
1496 }
1497
1498 old_image_size = s->header.image_size;
1499 s->header.image_size = offset;
1500 ret = qed_write_header_sync(s);
1501 if (ret < 0) {
1502 s->header.image_size = old_image_size;
1503 }
1504 return ret;
1505 }
1506
1507 static int64_t bdrv_qed_getlength(BlockDriverState *bs)
1508 {
1509 BDRVQEDState *s = bs->opaque;
1510 return s->header.image_size;
1511 }
1512
1513 static int bdrv_qed_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1514 {
1515 BDRVQEDState *s = bs->opaque;
1516
1517 memset(bdi, 0, sizeof(*bdi));
1518 bdi->cluster_size = s->header.cluster_size;
1519 bdi->is_dirty = s->header.features & QED_F_NEED_CHECK;
1520 bdi->unallocated_blocks_are_zero = true;
1521 bdi->can_write_zeroes_with_unmap = true;
1522 return 0;
1523 }
1524
1525 static int bdrv_qed_change_backing_file(BlockDriverState *bs,
1526 const char *backing_file,
1527 const char *backing_fmt)
1528 {
1529 BDRVQEDState *s = bs->opaque;
1530 QEDHeader new_header, le_header;
1531 void *buffer;
1532 size_t buffer_len, backing_file_len;
1533 int ret;
1534
1535 /* Refuse to set backing filename if unknown compat feature bits are
1536 * active. If the image uses an unknown compat feature then we may not
1537 * know the layout of data following the header structure and cannot safely
1538 * add a new string.
1539 */
1540 if (backing_file && (s->header.compat_features &
1541 ~QED_COMPAT_FEATURE_MASK)) {
1542 return -ENOTSUP;
1543 }
1544
1545 memcpy(&new_header, &s->header, sizeof(new_header));
1546
1547 new_header.features &= ~(QED_F_BACKING_FILE |
1548 QED_F_BACKING_FORMAT_NO_PROBE);
1549
1550 /* Adjust feature flags */
1551 if (backing_file) {
1552 new_header.features |= QED_F_BACKING_FILE;
1553
1554 if (qed_fmt_is_raw(backing_fmt)) {
1555 new_header.features |= QED_F_BACKING_FORMAT_NO_PROBE;
1556 }
1557 }
1558
1559 /* Calculate new header size */
1560 backing_file_len = 0;
1561
1562 if (backing_file) {
1563 backing_file_len = strlen(backing_file);
1564 }
1565
1566 buffer_len = sizeof(new_header);
1567 new_header.backing_filename_offset = buffer_len;
1568 new_header.backing_filename_size = backing_file_len;
1569 buffer_len += backing_file_len;
1570
1571 /* Make sure we can rewrite header without failing */
1572 if (buffer_len > new_header.header_size * new_header.cluster_size) {
1573 return -ENOSPC;
1574 }
1575
1576 /* Prepare new header */
1577 buffer = g_malloc(buffer_len);
1578
1579 qed_header_cpu_to_le(&new_header, &le_header);
1580 memcpy(buffer, &le_header, sizeof(le_header));
1581 buffer_len = sizeof(le_header);
1582
1583 if (backing_file) {
1584 memcpy(buffer + buffer_len, backing_file, backing_file_len);
1585 buffer_len += backing_file_len;
1586 }
1587
1588 /* Write new header */
1589 ret = bdrv_pwrite_sync(bs->file, 0, buffer, buffer_len);
1590 g_free(buffer);
1591 if (ret == 0) {
1592 memcpy(&s->header, &new_header, sizeof(new_header));
1593 }
1594 return ret;
1595 }
1596
1597 static void bdrv_qed_invalidate_cache(BlockDriverState *bs, Error **errp)
1598 {
1599 BDRVQEDState *s = bs->opaque;
1600 Error *local_err = NULL;
1601 int ret;
1602
1603 bdrv_qed_close(bs);
1604
1605 memset(s, 0, sizeof(BDRVQEDState));
1606 ret = bdrv_qed_open(bs, NULL, bs->open_flags, &local_err);
1607 if (local_err) {
1608 error_propagate(errp, local_err);
1609 error_prepend(errp, "Could not reopen qed layer: ");
1610 return;
1611 } else if (ret < 0) {
1612 error_setg_errno(errp, -ret, "Could not reopen qed layer");
1613 return;
1614 }
1615 }
1616
1617 static int bdrv_qed_check(BlockDriverState *bs, BdrvCheckResult *result,
1618 BdrvCheckMode fix)
1619 {
1620 BDRVQEDState *s = bs->opaque;
1621
1622 return qed_check(s, result, !!fix);
1623 }
1624
1625 static QemuOptsList qed_create_opts = {
1626 .name = "qed-create-opts",
1627 .head = QTAILQ_HEAD_INITIALIZER(qed_create_opts.head),
1628 .desc = {
1629 {
1630 .name = BLOCK_OPT_SIZE,
1631 .type = QEMU_OPT_SIZE,
1632 .help = "Virtual disk size"
1633 },
1634 {
1635 .name = BLOCK_OPT_BACKING_FILE,
1636 .type = QEMU_OPT_STRING,
1637 .help = "File name of a base image"
1638 },
1639 {
1640 .name = BLOCK_OPT_BACKING_FMT,
1641 .type = QEMU_OPT_STRING,
1642 .help = "Image format of the base image"
1643 },
1644 {
1645 .name = BLOCK_OPT_CLUSTER_SIZE,
1646 .type = QEMU_OPT_SIZE,
1647 .help = "Cluster size (in bytes)",
1648 .def_value_str = stringify(QED_DEFAULT_CLUSTER_SIZE)
1649 },
1650 {
1651 .name = BLOCK_OPT_TABLE_SIZE,
1652 .type = QEMU_OPT_SIZE,
1653 .help = "L1/L2 table size (in clusters)"
1654 },
1655 { /* end of list */ }
1656 }
1657 };
1658
1659 static BlockDriver bdrv_qed = {
1660 .format_name = "qed",
1661 .instance_size = sizeof(BDRVQEDState),
1662 .create_opts = &qed_create_opts,
1663 .supports_backing = true,
1664
1665 .bdrv_probe = bdrv_qed_probe,
1666 .bdrv_open = bdrv_qed_open,
1667 .bdrv_close = bdrv_qed_close,
1668 .bdrv_reopen_prepare = bdrv_qed_reopen_prepare,
1669 .bdrv_create = bdrv_qed_create,
1670 .bdrv_has_zero_init = bdrv_has_zero_init_1,
1671 .bdrv_co_get_block_status = bdrv_qed_co_get_block_status,
1672 .bdrv_aio_readv = bdrv_qed_aio_readv,
1673 .bdrv_aio_writev = bdrv_qed_aio_writev,
1674 .bdrv_co_pwrite_zeroes = bdrv_qed_co_pwrite_zeroes,
1675 .bdrv_truncate = bdrv_qed_truncate,
1676 .bdrv_getlength = bdrv_qed_getlength,
1677 .bdrv_get_info = bdrv_qed_get_info,
1678 .bdrv_refresh_limits = bdrv_qed_refresh_limits,
1679 .bdrv_change_backing_file = bdrv_qed_change_backing_file,
1680 .bdrv_invalidate_cache = bdrv_qed_invalidate_cache,
1681 .bdrv_check = bdrv_qed_check,
1682 .bdrv_detach_aio_context = bdrv_qed_detach_aio_context,
1683 .bdrv_attach_aio_context = bdrv_qed_attach_aio_context,
1684 .bdrv_drain = bdrv_qed_drain,
1685 };
1686
1687 static void bdrv_qed_init(void)
1688 {
1689 bdrv_register(&bdrv_qed);
1690 }
1691
1692 block_init(bdrv_qed_init);