file-posix: add drop-cache=on|off option
[qemu/kevin.git] / block / qcow2-refcount.c
blobe0fe322500bf61a44e2b1dddb64475e478cca19f
1 /*
2 * Block driver for the QCOW version 2 format
4 * Copyright (c) 2004-2006 Fabrice Bellard
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22 * THE SOFTWARE.
25 #include "qemu/osdep.h"
26 #include "qapi/error.h"
27 #include "qemu-common.h"
28 #include "block/block_int.h"
29 #include "qcow2.h"
30 #include "qemu/range.h"
31 #include "qemu/bswap.h"
32 #include "qemu/cutils.h"
34 static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size,
35 uint64_t max);
36 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
37 int64_t offset, int64_t length, uint64_t addend,
38 bool decrease, enum qcow2_discard_type type);
40 static uint64_t get_refcount_ro0(const void *refcount_array, uint64_t index);
41 static uint64_t get_refcount_ro1(const void *refcount_array, uint64_t index);
42 static uint64_t get_refcount_ro2(const void *refcount_array, uint64_t index);
43 static uint64_t get_refcount_ro3(const void *refcount_array, uint64_t index);
44 static uint64_t get_refcount_ro4(const void *refcount_array, uint64_t index);
45 static uint64_t get_refcount_ro5(const void *refcount_array, uint64_t index);
46 static uint64_t get_refcount_ro6(const void *refcount_array, uint64_t index);
48 static void set_refcount_ro0(void *refcount_array, uint64_t index,
49 uint64_t value);
50 static void set_refcount_ro1(void *refcount_array, uint64_t index,
51 uint64_t value);
52 static void set_refcount_ro2(void *refcount_array, uint64_t index,
53 uint64_t value);
54 static void set_refcount_ro3(void *refcount_array, uint64_t index,
55 uint64_t value);
56 static void set_refcount_ro4(void *refcount_array, uint64_t index,
57 uint64_t value);
58 static void set_refcount_ro5(void *refcount_array, uint64_t index,
59 uint64_t value);
60 static void set_refcount_ro6(void *refcount_array, uint64_t index,
61 uint64_t value);
64 static Qcow2GetRefcountFunc *const get_refcount_funcs[] = {
65 &get_refcount_ro0,
66 &get_refcount_ro1,
67 &get_refcount_ro2,
68 &get_refcount_ro3,
69 &get_refcount_ro4,
70 &get_refcount_ro5,
71 &get_refcount_ro6
74 static Qcow2SetRefcountFunc *const set_refcount_funcs[] = {
75 &set_refcount_ro0,
76 &set_refcount_ro1,
77 &set_refcount_ro2,
78 &set_refcount_ro3,
79 &set_refcount_ro4,
80 &set_refcount_ro5,
81 &set_refcount_ro6
85 /*********************************************************/
86 /* refcount handling */
88 static void update_max_refcount_table_index(BDRVQcow2State *s)
90 unsigned i = s->refcount_table_size - 1;
91 while (i > 0 && (s->refcount_table[i] & REFT_OFFSET_MASK) == 0) {
92 i--;
94 /* Set s->max_refcount_table_index to the index of the last used entry */
95 s->max_refcount_table_index = i;
98 int qcow2_refcount_init(BlockDriverState *bs)
100 BDRVQcow2State *s = bs->opaque;
101 unsigned int refcount_table_size2, i;
102 int ret;
104 assert(s->refcount_order >= 0 && s->refcount_order <= 6);
106 s->get_refcount = get_refcount_funcs[s->refcount_order];
107 s->set_refcount = set_refcount_funcs[s->refcount_order];
109 assert(s->refcount_table_size <= INT_MAX / sizeof(uint64_t));
110 refcount_table_size2 = s->refcount_table_size * sizeof(uint64_t);
111 s->refcount_table = g_try_malloc(refcount_table_size2);
113 if (s->refcount_table_size > 0) {
114 if (s->refcount_table == NULL) {
115 ret = -ENOMEM;
116 goto fail;
118 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_LOAD);
119 ret = bdrv_pread(bs->file, s->refcount_table_offset,
120 s->refcount_table, refcount_table_size2);
121 if (ret < 0) {
122 goto fail;
124 for(i = 0; i < s->refcount_table_size; i++)
125 be64_to_cpus(&s->refcount_table[i]);
126 update_max_refcount_table_index(s);
128 return 0;
129 fail:
130 return ret;
133 void qcow2_refcount_close(BlockDriverState *bs)
135 BDRVQcow2State *s = bs->opaque;
136 g_free(s->refcount_table);
140 static uint64_t get_refcount_ro0(const void *refcount_array, uint64_t index)
142 return (((const uint8_t *)refcount_array)[index / 8] >> (index % 8)) & 0x1;
145 static void set_refcount_ro0(void *refcount_array, uint64_t index,
146 uint64_t value)
148 assert(!(value >> 1));
149 ((uint8_t *)refcount_array)[index / 8] &= ~(0x1 << (index % 8));
150 ((uint8_t *)refcount_array)[index / 8] |= value << (index % 8);
153 static uint64_t get_refcount_ro1(const void *refcount_array, uint64_t index)
155 return (((const uint8_t *)refcount_array)[index / 4] >> (2 * (index % 4)))
156 & 0x3;
159 static void set_refcount_ro1(void *refcount_array, uint64_t index,
160 uint64_t value)
162 assert(!(value >> 2));
163 ((uint8_t *)refcount_array)[index / 4] &= ~(0x3 << (2 * (index % 4)));
164 ((uint8_t *)refcount_array)[index / 4] |= value << (2 * (index % 4));
167 static uint64_t get_refcount_ro2(const void *refcount_array, uint64_t index)
169 return (((const uint8_t *)refcount_array)[index / 2] >> (4 * (index % 2)))
170 & 0xf;
173 static void set_refcount_ro2(void *refcount_array, uint64_t index,
174 uint64_t value)
176 assert(!(value >> 4));
177 ((uint8_t *)refcount_array)[index / 2] &= ~(0xf << (4 * (index % 2)));
178 ((uint8_t *)refcount_array)[index / 2] |= value << (4 * (index % 2));
181 static uint64_t get_refcount_ro3(const void *refcount_array, uint64_t index)
183 return ((const uint8_t *)refcount_array)[index];
186 static void set_refcount_ro3(void *refcount_array, uint64_t index,
187 uint64_t value)
189 assert(!(value >> 8));
190 ((uint8_t *)refcount_array)[index] = value;
193 static uint64_t get_refcount_ro4(const void *refcount_array, uint64_t index)
195 return be16_to_cpu(((const uint16_t *)refcount_array)[index]);
198 static void set_refcount_ro4(void *refcount_array, uint64_t index,
199 uint64_t value)
201 assert(!(value >> 16));
202 ((uint16_t *)refcount_array)[index] = cpu_to_be16(value);
205 static uint64_t get_refcount_ro5(const void *refcount_array, uint64_t index)
207 return be32_to_cpu(((const uint32_t *)refcount_array)[index]);
210 static void set_refcount_ro5(void *refcount_array, uint64_t index,
211 uint64_t value)
213 assert(!(value >> 32));
214 ((uint32_t *)refcount_array)[index] = cpu_to_be32(value);
217 static uint64_t get_refcount_ro6(const void *refcount_array, uint64_t index)
219 return be64_to_cpu(((const uint64_t *)refcount_array)[index]);
222 static void set_refcount_ro6(void *refcount_array, uint64_t index,
223 uint64_t value)
225 ((uint64_t *)refcount_array)[index] = cpu_to_be64(value);
229 static int load_refcount_block(BlockDriverState *bs,
230 int64_t refcount_block_offset,
231 void **refcount_block)
233 BDRVQcow2State *s = bs->opaque;
235 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_LOAD);
236 return qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
237 refcount_block);
241 * Retrieves the refcount of the cluster given by its index and stores it in
242 * *refcount. Returns 0 on success and -errno on failure.
244 int qcow2_get_refcount(BlockDriverState *bs, int64_t cluster_index,
245 uint64_t *refcount)
247 BDRVQcow2State *s = bs->opaque;
248 uint64_t refcount_table_index, block_index;
249 int64_t refcount_block_offset;
250 int ret;
251 void *refcount_block;
253 refcount_table_index = cluster_index >> s->refcount_block_bits;
254 if (refcount_table_index >= s->refcount_table_size) {
255 *refcount = 0;
256 return 0;
258 refcount_block_offset =
259 s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
260 if (!refcount_block_offset) {
261 *refcount = 0;
262 return 0;
265 if (offset_into_cluster(s, refcount_block_offset)) {
266 qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#" PRIx64
267 " unaligned (reftable index: %#" PRIx64 ")",
268 refcount_block_offset, refcount_table_index);
269 return -EIO;
272 ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
273 &refcount_block);
274 if (ret < 0) {
275 return ret;
278 block_index = cluster_index & (s->refcount_block_size - 1);
279 *refcount = s->get_refcount(refcount_block, block_index);
281 qcow2_cache_put(s->refcount_block_cache, &refcount_block);
283 return 0;
286 /* Checks if two offsets are described by the same refcount block */
287 static int in_same_refcount_block(BDRVQcow2State *s, uint64_t offset_a,
288 uint64_t offset_b)
290 uint64_t block_a = offset_a >> (s->cluster_bits + s->refcount_block_bits);
291 uint64_t block_b = offset_b >> (s->cluster_bits + s->refcount_block_bits);
293 return (block_a == block_b);
297 * Loads a refcount block. If it doesn't exist yet, it is allocated first
298 * (including growing the refcount table if needed).
300 * Returns 0 on success or -errno in error case
302 static int alloc_refcount_block(BlockDriverState *bs,
303 int64_t cluster_index, void **refcount_block)
305 BDRVQcow2State *s = bs->opaque;
306 unsigned int refcount_table_index;
307 int64_t ret;
309 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC);
311 /* Find the refcount block for the given cluster */
312 refcount_table_index = cluster_index >> s->refcount_block_bits;
314 if (refcount_table_index < s->refcount_table_size) {
316 uint64_t refcount_block_offset =
317 s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
319 /* If it's already there, we're done */
320 if (refcount_block_offset) {
321 if (offset_into_cluster(s, refcount_block_offset)) {
322 qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#"
323 PRIx64 " unaligned (reftable index: "
324 "%#x)", refcount_block_offset,
325 refcount_table_index);
326 return -EIO;
329 return load_refcount_block(bs, refcount_block_offset,
330 refcount_block);
335 * If we came here, we need to allocate something. Something is at least
336 * a cluster for the new refcount block. It may also include a new refcount
337 * table if the old refcount table is too small.
339 * Note that allocating clusters here needs some special care:
341 * - We can't use the normal qcow2_alloc_clusters(), it would try to
342 * increase the refcount and very likely we would end up with an endless
343 * recursion. Instead we must place the refcount blocks in a way that
344 * they can describe them themselves.
346 * - We need to consider that at this point we are inside update_refcounts
347 * and potentially doing an initial refcount increase. This means that
348 * some clusters have already been allocated by the caller, but their
349 * refcount isn't accurate yet. If we allocate clusters for metadata, we
350 * need to return -EAGAIN to signal the caller that it needs to restart
351 * the search for free clusters.
353 * - alloc_clusters_noref and qcow2_free_clusters may load a different
354 * refcount block into the cache
357 *refcount_block = NULL;
359 /* We write to the refcount table, so we might depend on L2 tables */
360 ret = qcow2_cache_flush(bs, s->l2_table_cache);
361 if (ret < 0) {
362 return ret;
365 /* Allocate the refcount block itself and mark it as used */
366 int64_t new_block = alloc_clusters_noref(bs, s->cluster_size, INT64_MAX);
367 if (new_block < 0) {
368 return new_block;
371 /* The offset must fit in the offset field of the refcount table entry */
372 assert((new_block & REFT_OFFSET_MASK) == new_block);
374 /* If we're allocating the block at offset 0 then something is wrong */
375 if (new_block == 0) {
376 qcow2_signal_corruption(bs, true, -1, -1, "Preventing invalid "
377 "allocation of refcount block at offset 0");
378 return -EIO;
381 #ifdef DEBUG_ALLOC2
382 fprintf(stderr, "qcow2: Allocate refcount block %d for %" PRIx64
383 " at %" PRIx64 "\n",
384 refcount_table_index, cluster_index << s->cluster_bits, new_block);
385 #endif
387 if (in_same_refcount_block(s, new_block, cluster_index << s->cluster_bits)) {
388 /* Zero the new refcount block before updating it */
389 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
390 refcount_block);
391 if (ret < 0) {
392 goto fail;
395 memset(*refcount_block, 0, s->cluster_size);
397 /* The block describes itself, need to update the cache */
398 int block_index = (new_block >> s->cluster_bits) &
399 (s->refcount_block_size - 1);
400 s->set_refcount(*refcount_block, block_index, 1);
401 } else {
402 /* Described somewhere else. This can recurse at most twice before we
403 * arrive at a block that describes itself. */
404 ret = update_refcount(bs, new_block, s->cluster_size, 1, false,
405 QCOW2_DISCARD_NEVER);
406 if (ret < 0) {
407 goto fail;
410 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
411 if (ret < 0) {
412 goto fail;
415 /* Initialize the new refcount block only after updating its refcount,
416 * update_refcount uses the refcount cache itself */
417 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
418 refcount_block);
419 if (ret < 0) {
420 goto fail;
423 memset(*refcount_block, 0, s->cluster_size);
426 /* Now the new refcount block needs to be written to disk */
427 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE);
428 qcow2_cache_entry_mark_dirty(s->refcount_block_cache, *refcount_block);
429 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
430 if (ret < 0) {
431 goto fail;
434 /* If the refcount table is big enough, just hook the block up there */
435 if (refcount_table_index < s->refcount_table_size) {
436 uint64_t data64 = cpu_to_be64(new_block);
437 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_HOOKUP);
438 ret = bdrv_pwrite_sync(bs->file,
439 s->refcount_table_offset + refcount_table_index * sizeof(uint64_t),
440 &data64, sizeof(data64));
441 if (ret < 0) {
442 goto fail;
445 s->refcount_table[refcount_table_index] = new_block;
446 /* If there's a hole in s->refcount_table then it can happen
447 * that refcount_table_index < s->max_refcount_table_index */
448 s->max_refcount_table_index =
449 MAX(s->max_refcount_table_index, refcount_table_index);
451 /* The new refcount block may be where the caller intended to put its
452 * data, so let it restart the search. */
453 return -EAGAIN;
456 qcow2_cache_put(s->refcount_block_cache, refcount_block);
459 * If we come here, we need to grow the refcount table. Again, a new
460 * refcount table needs some space and we can't simply allocate to avoid
461 * endless recursion.
463 * Therefore let's grab new refcount blocks at the end of the image, which
464 * will describe themselves and the new refcount table. This way we can
465 * reference them only in the new table and do the switch to the new
466 * refcount table at once without producing an inconsistent state in
467 * between.
469 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_GROW);
471 /* Calculate the number of refcount blocks needed so far; this will be the
472 * basis for calculating the index of the first cluster used for the
473 * self-describing refcount structures which we are about to create.
475 * Because we reached this point, there cannot be any refcount entries for
476 * cluster_index or higher indices yet. However, because new_block has been
477 * allocated to describe that cluster (and it will assume this role later
478 * on), we cannot use that index; also, new_block may actually have a higher
479 * cluster index than cluster_index, so it needs to be taken into account
480 * here (and 1 needs to be added to its value because that cluster is used).
482 uint64_t blocks_used = DIV_ROUND_UP(MAX(cluster_index + 1,
483 (new_block >> s->cluster_bits) + 1),
484 s->refcount_block_size);
486 /* Create the new refcount table and blocks */
487 uint64_t meta_offset = (blocks_used * s->refcount_block_size) *
488 s->cluster_size;
490 ret = qcow2_refcount_area(bs, meta_offset, 0, false,
491 refcount_table_index, new_block);
492 if (ret < 0) {
493 return ret;
496 ret = load_refcount_block(bs, new_block, refcount_block);
497 if (ret < 0) {
498 return ret;
501 /* If we were trying to do the initial refcount update for some cluster
502 * allocation, we might have used the same clusters to store newly
503 * allocated metadata. Make the caller search some new space. */
504 return -EAGAIN;
506 fail:
507 if (*refcount_block != NULL) {
508 qcow2_cache_put(s->refcount_block_cache, refcount_block);
510 return ret;
514 * Starting at @start_offset, this function creates new self-covering refcount
515 * structures: A new refcount table and refcount blocks which cover all of
516 * themselves, and a number of @additional_clusters beyond their end.
517 * @start_offset must be at the end of the image file, that is, there must be
518 * only empty space beyond it.
519 * If @exact_size is false, the refcount table will have 50 % more entries than
520 * necessary so it will not need to grow again soon.
521 * If @new_refblock_offset is not zero, it contains the offset of a refcount
522 * block that should be entered into the new refcount table at index
523 * @new_refblock_index.
525 * Returns: The offset after the new refcount structures (i.e. where the
526 * @additional_clusters may be placed) on success, -errno on error.
528 int64_t qcow2_refcount_area(BlockDriverState *bs, uint64_t start_offset,
529 uint64_t additional_clusters, bool exact_size,
530 int new_refblock_index,
531 uint64_t new_refblock_offset)
533 BDRVQcow2State *s = bs->opaque;
534 uint64_t total_refblock_count_u64, additional_refblock_count;
535 int total_refblock_count, table_size, area_reftable_index, table_clusters;
536 int i;
537 uint64_t table_offset, block_offset, end_offset;
538 int ret;
539 uint64_t *new_table;
541 assert(!(start_offset % s->cluster_size));
543 qcow2_refcount_metadata_size(start_offset / s->cluster_size +
544 additional_clusters,
545 s->cluster_size, s->refcount_order,
546 !exact_size, &total_refblock_count_u64);
547 if (total_refblock_count_u64 > QCOW_MAX_REFTABLE_SIZE) {
548 return -EFBIG;
550 total_refblock_count = total_refblock_count_u64;
552 /* Index in the refcount table of the first refcount block to cover the area
553 * of refcount structures we are about to create; we know that
554 * @total_refblock_count can cover @start_offset, so this will definitely
555 * fit into an int. */
556 area_reftable_index = (start_offset / s->cluster_size) /
557 s->refcount_block_size;
559 if (exact_size) {
560 table_size = total_refblock_count;
561 } else {
562 table_size = total_refblock_count +
563 DIV_ROUND_UP(total_refblock_count, 2);
565 /* The qcow2 file can only store the reftable size in number of clusters */
566 table_size = ROUND_UP(table_size, s->cluster_size / sizeof(uint64_t));
567 table_clusters = (table_size * sizeof(uint64_t)) / s->cluster_size;
569 if (table_size > QCOW_MAX_REFTABLE_SIZE) {
570 return -EFBIG;
573 new_table = g_try_new0(uint64_t, table_size);
575 assert(table_size > 0);
576 if (new_table == NULL) {
577 ret = -ENOMEM;
578 goto fail;
581 /* Fill the new refcount table */
582 if (table_size > s->max_refcount_table_index) {
583 /* We're actually growing the reftable */
584 memcpy(new_table, s->refcount_table,
585 (s->max_refcount_table_index + 1) * sizeof(uint64_t));
586 } else {
587 /* Improbable case: We're shrinking the reftable. However, the caller
588 * has assured us that there is only empty space beyond @start_offset,
589 * so we can simply drop all of the refblocks that won't fit into the
590 * new reftable. */
591 memcpy(new_table, s->refcount_table, table_size * sizeof(uint64_t));
594 if (new_refblock_offset) {
595 assert(new_refblock_index < total_refblock_count);
596 new_table[new_refblock_index] = new_refblock_offset;
599 /* Count how many new refblocks we have to create */
600 additional_refblock_count = 0;
601 for (i = area_reftable_index; i < total_refblock_count; i++) {
602 if (!new_table[i]) {
603 additional_refblock_count++;
607 table_offset = start_offset + additional_refblock_count * s->cluster_size;
608 end_offset = table_offset + table_clusters * s->cluster_size;
610 /* Fill the refcount blocks, and create new ones, if necessary */
611 block_offset = start_offset;
612 for (i = area_reftable_index; i < total_refblock_count; i++) {
613 void *refblock_data;
614 uint64_t first_offset_covered;
616 /* Reuse an existing refblock if possible, create a new one otherwise */
617 if (new_table[i]) {
618 ret = qcow2_cache_get(bs, s->refcount_block_cache, new_table[i],
619 &refblock_data);
620 if (ret < 0) {
621 goto fail;
623 } else {
624 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache,
625 block_offset, &refblock_data);
626 if (ret < 0) {
627 goto fail;
629 memset(refblock_data, 0, s->cluster_size);
630 qcow2_cache_entry_mark_dirty(s->refcount_block_cache,
631 refblock_data);
633 new_table[i] = block_offset;
634 block_offset += s->cluster_size;
637 /* First host offset covered by this refblock */
638 first_offset_covered = (uint64_t)i * s->refcount_block_size *
639 s->cluster_size;
640 if (first_offset_covered < end_offset) {
641 int j, end_index;
643 /* Set the refcount of all of the new refcount structures to 1 */
645 if (first_offset_covered < start_offset) {
646 assert(i == area_reftable_index);
647 j = (start_offset - first_offset_covered) / s->cluster_size;
648 assert(j < s->refcount_block_size);
649 } else {
650 j = 0;
653 end_index = MIN((end_offset - first_offset_covered) /
654 s->cluster_size,
655 s->refcount_block_size);
657 for (; j < end_index; j++) {
658 /* The caller guaranteed us this space would be empty */
659 assert(s->get_refcount(refblock_data, j) == 0);
660 s->set_refcount(refblock_data, j, 1);
663 qcow2_cache_entry_mark_dirty(s->refcount_block_cache,
664 refblock_data);
667 qcow2_cache_put(s->refcount_block_cache, &refblock_data);
670 assert(block_offset == table_offset);
672 /* Write refcount blocks to disk */
673 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_BLOCKS);
674 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
675 if (ret < 0) {
676 goto fail;
679 /* Write refcount table to disk */
680 for (i = 0; i < total_refblock_count; i++) {
681 cpu_to_be64s(&new_table[i]);
684 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_TABLE);
685 ret = bdrv_pwrite_sync(bs->file, table_offset, new_table,
686 table_size * sizeof(uint64_t));
687 if (ret < 0) {
688 goto fail;
691 for (i = 0; i < total_refblock_count; i++) {
692 be64_to_cpus(&new_table[i]);
695 /* Hook up the new refcount table in the qcow2 header */
696 struct QEMU_PACKED {
697 uint64_t d64;
698 uint32_t d32;
699 } data;
700 data.d64 = cpu_to_be64(table_offset);
701 data.d32 = cpu_to_be32(table_clusters);
702 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_SWITCH_TABLE);
703 ret = bdrv_pwrite_sync(bs->file,
704 offsetof(QCowHeader, refcount_table_offset),
705 &data, sizeof(data));
706 if (ret < 0) {
707 goto fail;
710 /* And switch it in memory */
711 uint64_t old_table_offset = s->refcount_table_offset;
712 uint64_t old_table_size = s->refcount_table_size;
714 g_free(s->refcount_table);
715 s->refcount_table = new_table;
716 s->refcount_table_size = table_size;
717 s->refcount_table_offset = table_offset;
718 update_max_refcount_table_index(s);
720 /* Free old table. */
721 qcow2_free_clusters(bs, old_table_offset, old_table_size * sizeof(uint64_t),
722 QCOW2_DISCARD_OTHER);
724 return end_offset;
726 fail:
727 g_free(new_table);
728 return ret;
731 void qcow2_process_discards(BlockDriverState *bs, int ret)
733 BDRVQcow2State *s = bs->opaque;
734 Qcow2DiscardRegion *d, *next;
736 QTAILQ_FOREACH_SAFE(d, &s->discards, next, next) {
737 QTAILQ_REMOVE(&s->discards, d, next);
739 /* Discard is optional, ignore the return value */
740 if (ret >= 0) {
741 bdrv_pdiscard(bs->file, d->offset, d->bytes);
744 g_free(d);
748 static void update_refcount_discard(BlockDriverState *bs,
749 uint64_t offset, uint64_t length)
751 BDRVQcow2State *s = bs->opaque;
752 Qcow2DiscardRegion *d, *p, *next;
754 QTAILQ_FOREACH(d, &s->discards, next) {
755 uint64_t new_start = MIN(offset, d->offset);
756 uint64_t new_end = MAX(offset + length, d->offset + d->bytes);
758 if (new_end - new_start <= length + d->bytes) {
759 /* There can't be any overlap, areas ending up here have no
760 * references any more and therefore shouldn't get freed another
761 * time. */
762 assert(d->bytes + length == new_end - new_start);
763 d->offset = new_start;
764 d->bytes = new_end - new_start;
765 goto found;
769 d = g_malloc(sizeof(*d));
770 *d = (Qcow2DiscardRegion) {
771 .bs = bs,
772 .offset = offset,
773 .bytes = length,
775 QTAILQ_INSERT_TAIL(&s->discards, d, next);
777 found:
778 /* Merge discard requests if they are adjacent now */
779 QTAILQ_FOREACH_SAFE(p, &s->discards, next, next) {
780 if (p == d
781 || p->offset > d->offset + d->bytes
782 || d->offset > p->offset + p->bytes)
784 continue;
787 /* Still no overlap possible */
788 assert(p->offset == d->offset + d->bytes
789 || d->offset == p->offset + p->bytes);
791 QTAILQ_REMOVE(&s->discards, p, next);
792 d->offset = MIN(d->offset, p->offset);
793 d->bytes += p->bytes;
794 g_free(p);
798 /* XXX: cache several refcount block clusters ? */
799 /* @addend is the absolute value of the addend; if @decrease is set, @addend
800 * will be subtracted from the current refcount, otherwise it will be added */
801 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
802 int64_t offset,
803 int64_t length,
804 uint64_t addend,
805 bool decrease,
806 enum qcow2_discard_type type)
808 BDRVQcow2State *s = bs->opaque;
809 int64_t start, last, cluster_offset;
810 void *refcount_block = NULL;
811 int64_t old_table_index = -1;
812 int ret;
814 #ifdef DEBUG_ALLOC2
815 fprintf(stderr, "update_refcount: offset=%" PRId64 " size=%" PRId64
816 " addend=%s%" PRIu64 "\n", offset, length, decrease ? "-" : "",
817 addend);
818 #endif
819 if (length < 0) {
820 return -EINVAL;
821 } else if (length == 0) {
822 return 0;
825 if (decrease) {
826 qcow2_cache_set_dependency(bs, s->refcount_block_cache,
827 s->l2_table_cache);
830 start = start_of_cluster(s, offset);
831 last = start_of_cluster(s, offset + length - 1);
832 for(cluster_offset = start; cluster_offset <= last;
833 cluster_offset += s->cluster_size)
835 int block_index;
836 uint64_t refcount;
837 int64_t cluster_index = cluster_offset >> s->cluster_bits;
838 int64_t table_index = cluster_index >> s->refcount_block_bits;
840 /* Load the refcount block and allocate it if needed */
841 if (table_index != old_table_index) {
842 if (refcount_block) {
843 qcow2_cache_put(s->refcount_block_cache, &refcount_block);
845 ret = alloc_refcount_block(bs, cluster_index, &refcount_block);
846 /* If the caller needs to restart the search for free clusters,
847 * try the same ones first to see if they're still free. */
848 if (ret == -EAGAIN) {
849 if (s->free_cluster_index > (start >> s->cluster_bits)) {
850 s->free_cluster_index = (start >> s->cluster_bits);
853 if (ret < 0) {
854 goto fail;
857 old_table_index = table_index;
859 qcow2_cache_entry_mark_dirty(s->refcount_block_cache, refcount_block);
861 /* we can update the count and save it */
862 block_index = cluster_index & (s->refcount_block_size - 1);
864 refcount = s->get_refcount(refcount_block, block_index);
865 if (decrease ? (refcount - addend > refcount)
866 : (refcount + addend < refcount ||
867 refcount + addend > s->refcount_max))
869 ret = -EINVAL;
870 goto fail;
872 if (decrease) {
873 refcount -= addend;
874 } else {
875 refcount += addend;
877 if (refcount == 0 && cluster_index < s->free_cluster_index) {
878 s->free_cluster_index = cluster_index;
880 s->set_refcount(refcount_block, block_index, refcount);
882 if (refcount == 0) {
883 void *table;
885 table = qcow2_cache_is_table_offset(s->refcount_block_cache,
886 offset);
887 if (table != NULL) {
888 qcow2_cache_put(s->refcount_block_cache, &refcount_block);
889 qcow2_cache_discard(s->refcount_block_cache, table);
892 table = qcow2_cache_is_table_offset(s->l2_table_cache, offset);
893 if (table != NULL) {
894 qcow2_cache_discard(s->l2_table_cache, table);
897 if (s->discard_passthrough[type]) {
898 update_refcount_discard(bs, cluster_offset, s->cluster_size);
903 ret = 0;
904 fail:
905 if (!s->cache_discards) {
906 qcow2_process_discards(bs, ret);
909 /* Write last changed block to disk */
910 if (refcount_block) {
911 qcow2_cache_put(s->refcount_block_cache, &refcount_block);
915 * Try do undo any updates if an error is returned (This may succeed in
916 * some cases like ENOSPC for allocating a new refcount block)
918 if (ret < 0) {
919 int dummy;
920 dummy = update_refcount(bs, offset, cluster_offset - offset, addend,
921 !decrease, QCOW2_DISCARD_NEVER);
922 (void)dummy;
925 return ret;
929 * Increases or decreases the refcount of a given cluster.
931 * @addend is the absolute value of the addend; if @decrease is set, @addend
932 * will be subtracted from the current refcount, otherwise it will be added.
934 * On success 0 is returned; on failure -errno is returned.
936 int qcow2_update_cluster_refcount(BlockDriverState *bs,
937 int64_t cluster_index,
938 uint64_t addend, bool decrease,
939 enum qcow2_discard_type type)
941 BDRVQcow2State *s = bs->opaque;
942 int ret;
944 ret = update_refcount(bs, cluster_index << s->cluster_bits, 1, addend,
945 decrease, type);
946 if (ret < 0) {
947 return ret;
950 return 0;
955 /*********************************************************/
956 /* cluster allocation functions */
960 /* return < 0 if error */
961 static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size,
962 uint64_t max)
964 BDRVQcow2State *s = bs->opaque;
965 uint64_t i, nb_clusters, refcount;
966 int ret;
968 /* We can't allocate clusters if they may still be queued for discard. */
969 if (s->cache_discards) {
970 qcow2_process_discards(bs, 0);
973 nb_clusters = size_to_clusters(s, size);
974 retry:
975 for(i = 0; i < nb_clusters; i++) {
976 uint64_t next_cluster_index = s->free_cluster_index++;
977 ret = qcow2_get_refcount(bs, next_cluster_index, &refcount);
979 if (ret < 0) {
980 return ret;
981 } else if (refcount != 0) {
982 goto retry;
986 /* Make sure that all offsets in the "allocated" range are representable
987 * in the requested max */
988 if (s->free_cluster_index > 0 &&
989 s->free_cluster_index - 1 > (max >> s->cluster_bits))
991 return -EFBIG;
994 #ifdef DEBUG_ALLOC2
995 fprintf(stderr, "alloc_clusters: size=%" PRId64 " -> %" PRId64 "\n",
996 size,
997 (s->free_cluster_index - nb_clusters) << s->cluster_bits);
998 #endif
999 return (s->free_cluster_index - nb_clusters) << s->cluster_bits;
1002 int64_t qcow2_alloc_clusters(BlockDriverState *bs, uint64_t size)
1004 int64_t offset;
1005 int ret;
1007 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC);
1008 do {
1009 offset = alloc_clusters_noref(bs, size, QCOW_MAX_CLUSTER_OFFSET);
1010 if (offset < 0) {
1011 return offset;
1014 ret = update_refcount(bs, offset, size, 1, false, QCOW2_DISCARD_NEVER);
1015 } while (ret == -EAGAIN);
1017 if (ret < 0) {
1018 return ret;
1021 return offset;
1024 int64_t qcow2_alloc_clusters_at(BlockDriverState *bs, uint64_t offset,
1025 int64_t nb_clusters)
1027 BDRVQcow2State *s = bs->opaque;
1028 uint64_t cluster_index, refcount;
1029 uint64_t i;
1030 int ret;
1032 assert(nb_clusters >= 0);
1033 if (nb_clusters == 0) {
1034 return 0;
1037 do {
1038 /* Check how many clusters there are free */
1039 cluster_index = offset >> s->cluster_bits;
1040 for(i = 0; i < nb_clusters; i++) {
1041 ret = qcow2_get_refcount(bs, cluster_index++, &refcount);
1042 if (ret < 0) {
1043 return ret;
1044 } else if (refcount != 0) {
1045 break;
1049 /* And then allocate them */
1050 ret = update_refcount(bs, offset, i << s->cluster_bits, 1, false,
1051 QCOW2_DISCARD_NEVER);
1052 } while (ret == -EAGAIN);
1054 if (ret < 0) {
1055 return ret;
1058 return i;
1061 /* only used to allocate compressed sectors. We try to allocate
1062 contiguous sectors. size must be <= cluster_size */
1063 int64_t qcow2_alloc_bytes(BlockDriverState *bs, int size)
1065 BDRVQcow2State *s = bs->opaque;
1066 int64_t offset;
1067 size_t free_in_cluster;
1068 int ret;
1070 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC_BYTES);
1071 assert(size > 0 && size <= s->cluster_size);
1072 assert(!s->free_byte_offset || offset_into_cluster(s, s->free_byte_offset));
1074 offset = s->free_byte_offset;
1076 if (offset) {
1077 uint64_t refcount;
1078 ret = qcow2_get_refcount(bs, offset >> s->cluster_bits, &refcount);
1079 if (ret < 0) {
1080 return ret;
1083 if (refcount == s->refcount_max) {
1084 offset = 0;
1088 free_in_cluster = s->cluster_size - offset_into_cluster(s, offset);
1089 do {
1090 if (!offset || free_in_cluster < size) {
1091 int64_t new_cluster;
1093 new_cluster = alloc_clusters_noref(bs, s->cluster_size,
1094 MIN(s->cluster_offset_mask,
1095 QCOW_MAX_CLUSTER_OFFSET));
1096 if (new_cluster < 0) {
1097 return new_cluster;
1100 if (new_cluster == 0) {
1101 qcow2_signal_corruption(bs, true, -1, -1, "Preventing invalid "
1102 "allocation of compressed cluster "
1103 "at offset 0");
1104 return -EIO;
1107 if (!offset || ROUND_UP(offset, s->cluster_size) != new_cluster) {
1108 offset = new_cluster;
1109 free_in_cluster = s->cluster_size;
1110 } else {
1111 free_in_cluster += s->cluster_size;
1115 assert(offset);
1116 ret = update_refcount(bs, offset, size, 1, false, QCOW2_DISCARD_NEVER);
1117 if (ret < 0) {
1118 offset = 0;
1120 } while (ret == -EAGAIN);
1121 if (ret < 0) {
1122 return ret;
1125 /* The cluster refcount was incremented; refcount blocks must be flushed
1126 * before the caller's L2 table updates. */
1127 qcow2_cache_set_dependency(bs, s->l2_table_cache, s->refcount_block_cache);
1129 s->free_byte_offset = offset + size;
1130 if (!offset_into_cluster(s, s->free_byte_offset)) {
1131 s->free_byte_offset = 0;
1134 return offset;
1137 void qcow2_free_clusters(BlockDriverState *bs,
1138 int64_t offset, int64_t size,
1139 enum qcow2_discard_type type)
1141 int ret;
1143 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_FREE);
1144 ret = update_refcount(bs, offset, size, 1, true, type);
1145 if (ret < 0) {
1146 fprintf(stderr, "qcow2_free_clusters failed: %s\n", strerror(-ret));
1147 /* TODO Remember the clusters to free them later and avoid leaking */
1152 * Free a cluster using its L2 entry (handles clusters of all types, e.g.
1153 * normal cluster, compressed cluster, etc.)
1155 void qcow2_free_any_clusters(BlockDriverState *bs, uint64_t l2_entry,
1156 int nb_clusters, enum qcow2_discard_type type)
1158 BDRVQcow2State *s = bs->opaque;
1159 QCow2ClusterType ctype = qcow2_get_cluster_type(bs, l2_entry);
1161 if (has_data_file(bs)) {
1162 if (s->discard_passthrough[type] &&
1163 (ctype == QCOW2_CLUSTER_NORMAL ||
1164 ctype == QCOW2_CLUSTER_ZERO_ALLOC))
1166 bdrv_pdiscard(s->data_file, l2_entry & L2E_OFFSET_MASK,
1167 nb_clusters << s->cluster_bits);
1169 return;
1172 switch (ctype) {
1173 case QCOW2_CLUSTER_COMPRESSED:
1175 int nb_csectors;
1176 nb_csectors = ((l2_entry >> s->csize_shift) &
1177 s->csize_mask) + 1;
1178 qcow2_free_clusters(bs,
1179 (l2_entry & s->cluster_offset_mask) & ~511,
1180 nb_csectors * 512, type);
1182 break;
1183 case QCOW2_CLUSTER_NORMAL:
1184 case QCOW2_CLUSTER_ZERO_ALLOC:
1185 if (offset_into_cluster(s, l2_entry & L2E_OFFSET_MASK)) {
1186 qcow2_signal_corruption(bs, false, -1, -1,
1187 "Cannot free unaligned cluster %#llx",
1188 l2_entry & L2E_OFFSET_MASK);
1189 } else {
1190 qcow2_free_clusters(bs, l2_entry & L2E_OFFSET_MASK,
1191 nb_clusters << s->cluster_bits, type);
1193 break;
1194 case QCOW2_CLUSTER_ZERO_PLAIN:
1195 case QCOW2_CLUSTER_UNALLOCATED:
1196 break;
1197 default:
1198 abort();
1202 int coroutine_fn qcow2_write_caches(BlockDriverState *bs)
1204 BDRVQcow2State *s = bs->opaque;
1205 int ret;
1207 ret = qcow2_cache_write(bs, s->l2_table_cache);
1208 if (ret < 0) {
1209 return ret;
1212 if (qcow2_need_accurate_refcounts(s)) {
1213 ret = qcow2_cache_write(bs, s->refcount_block_cache);
1214 if (ret < 0) {
1215 return ret;
1219 return 0;
1222 int coroutine_fn qcow2_flush_caches(BlockDriverState *bs)
1224 int ret = qcow2_write_caches(bs);
1225 if (ret < 0) {
1226 return ret;
1229 return bdrv_flush(bs->file->bs);
1232 /*********************************************************/
1233 /* snapshots and image creation */
1237 /* update the refcounts of snapshots and the copied flag */
1238 int qcow2_update_snapshot_refcount(BlockDriverState *bs,
1239 int64_t l1_table_offset, int l1_size, int addend)
1241 BDRVQcow2State *s = bs->opaque;
1242 uint64_t *l1_table, *l2_slice, l2_offset, entry, l1_size2, refcount;
1243 bool l1_allocated = false;
1244 int64_t old_entry, old_l2_offset;
1245 unsigned slice, slice_size2, n_slices;
1246 int i, j, l1_modified = 0, nb_csectors;
1247 int ret;
1249 assert(addend >= -1 && addend <= 1);
1251 l2_slice = NULL;
1252 l1_table = NULL;
1253 l1_size2 = l1_size * sizeof(uint64_t);
1254 slice_size2 = s->l2_slice_size * sizeof(uint64_t);
1255 n_slices = s->cluster_size / slice_size2;
1257 s->cache_discards = true;
1259 /* WARNING: qcow2_snapshot_goto relies on this function not using the
1260 * l1_table_offset when it is the current s->l1_table_offset! Be careful
1261 * when changing this! */
1262 if (l1_table_offset != s->l1_table_offset) {
1263 l1_table = g_try_malloc0(ROUND_UP(l1_size2, 512));
1264 if (l1_size2 && l1_table == NULL) {
1265 ret = -ENOMEM;
1266 goto fail;
1268 l1_allocated = true;
1270 ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2);
1271 if (ret < 0) {
1272 goto fail;
1275 for (i = 0; i < l1_size; i++) {
1276 be64_to_cpus(&l1_table[i]);
1278 } else {
1279 assert(l1_size == s->l1_size);
1280 l1_table = s->l1_table;
1281 l1_allocated = false;
1284 for (i = 0; i < l1_size; i++) {
1285 l2_offset = l1_table[i];
1286 if (l2_offset) {
1287 old_l2_offset = l2_offset;
1288 l2_offset &= L1E_OFFSET_MASK;
1290 if (offset_into_cluster(s, l2_offset)) {
1291 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#"
1292 PRIx64 " unaligned (L1 index: %#x)",
1293 l2_offset, i);
1294 ret = -EIO;
1295 goto fail;
1298 for (slice = 0; slice < n_slices; slice++) {
1299 ret = qcow2_cache_get(bs, s->l2_table_cache,
1300 l2_offset + slice * slice_size2,
1301 (void **) &l2_slice);
1302 if (ret < 0) {
1303 goto fail;
1306 for (j = 0; j < s->l2_slice_size; j++) {
1307 uint64_t cluster_index;
1308 uint64_t offset;
1310 entry = be64_to_cpu(l2_slice[j]);
1311 old_entry = entry;
1312 entry &= ~QCOW_OFLAG_COPIED;
1313 offset = entry & L2E_OFFSET_MASK;
1315 switch (qcow2_get_cluster_type(bs, entry)) {
1316 case QCOW2_CLUSTER_COMPRESSED:
1317 nb_csectors = ((entry >> s->csize_shift) &
1318 s->csize_mask) + 1;
1319 if (addend != 0) {
1320 ret = update_refcount(
1321 bs, (entry & s->cluster_offset_mask) & ~511,
1322 nb_csectors * 512, abs(addend), addend < 0,
1323 QCOW2_DISCARD_SNAPSHOT);
1324 if (ret < 0) {
1325 goto fail;
1328 /* compressed clusters are never modified */
1329 refcount = 2;
1330 break;
1332 case QCOW2_CLUSTER_NORMAL:
1333 case QCOW2_CLUSTER_ZERO_ALLOC:
1334 if (offset_into_cluster(s, offset)) {
1335 /* Here l2_index means table (not slice) index */
1336 int l2_index = slice * s->l2_slice_size + j;
1337 qcow2_signal_corruption(
1338 bs, true, -1, -1, "Cluster "
1339 "allocation offset %#" PRIx64
1340 " unaligned (L2 offset: %#"
1341 PRIx64 ", L2 index: %#x)",
1342 offset, l2_offset, l2_index);
1343 ret = -EIO;
1344 goto fail;
1347 cluster_index = offset >> s->cluster_bits;
1348 assert(cluster_index);
1349 if (addend != 0) {
1350 ret = qcow2_update_cluster_refcount(
1351 bs, cluster_index, abs(addend), addend < 0,
1352 QCOW2_DISCARD_SNAPSHOT);
1353 if (ret < 0) {
1354 goto fail;
1358 ret = qcow2_get_refcount(bs, cluster_index, &refcount);
1359 if (ret < 0) {
1360 goto fail;
1362 break;
1364 case QCOW2_CLUSTER_ZERO_PLAIN:
1365 case QCOW2_CLUSTER_UNALLOCATED:
1366 refcount = 0;
1367 break;
1369 default:
1370 abort();
1373 if (refcount == 1) {
1374 entry |= QCOW_OFLAG_COPIED;
1376 if (entry != old_entry) {
1377 if (addend > 0) {
1378 qcow2_cache_set_dependency(bs, s->l2_table_cache,
1379 s->refcount_block_cache);
1381 l2_slice[j] = cpu_to_be64(entry);
1382 qcow2_cache_entry_mark_dirty(s->l2_table_cache,
1383 l2_slice);
1387 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1390 if (addend != 0) {
1391 ret = qcow2_update_cluster_refcount(bs, l2_offset >>
1392 s->cluster_bits,
1393 abs(addend), addend < 0,
1394 QCOW2_DISCARD_SNAPSHOT);
1395 if (ret < 0) {
1396 goto fail;
1399 ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits,
1400 &refcount);
1401 if (ret < 0) {
1402 goto fail;
1403 } else if (refcount == 1) {
1404 l2_offset |= QCOW_OFLAG_COPIED;
1406 if (l2_offset != old_l2_offset) {
1407 l1_table[i] = l2_offset;
1408 l1_modified = 1;
1413 ret = bdrv_flush(bs);
1414 fail:
1415 if (l2_slice) {
1416 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1419 s->cache_discards = false;
1420 qcow2_process_discards(bs, ret);
1422 /* Update L1 only if it isn't deleted anyway (addend = -1) */
1423 if (ret == 0 && addend >= 0 && l1_modified) {
1424 for (i = 0; i < l1_size; i++) {
1425 cpu_to_be64s(&l1_table[i]);
1428 ret = bdrv_pwrite_sync(bs->file, l1_table_offset,
1429 l1_table, l1_size2);
1431 for (i = 0; i < l1_size; i++) {
1432 be64_to_cpus(&l1_table[i]);
1435 if (l1_allocated)
1436 g_free(l1_table);
1437 return ret;
1443 /*********************************************************/
1444 /* refcount checking functions */
1447 static uint64_t refcount_array_byte_size(BDRVQcow2State *s, uint64_t entries)
1449 /* This assertion holds because there is no way we can address more than
1450 * 2^(64 - 9) clusters at once (with cluster size 512 = 2^9, and because
1451 * offsets have to be representable in bytes); due to every cluster
1452 * corresponding to one refcount entry, we are well below that limit */
1453 assert(entries < (UINT64_C(1) << (64 - 9)));
1455 /* Thanks to the assertion this will not overflow, because
1456 * s->refcount_order < 7.
1457 * (note: x << s->refcount_order == x * s->refcount_bits) */
1458 return DIV_ROUND_UP(entries << s->refcount_order, 8);
1462 * Reallocates *array so that it can hold new_size entries. *size must contain
1463 * the current number of entries in *array. If the reallocation fails, *array
1464 * and *size will not be modified and -errno will be returned. If the
1465 * reallocation is successful, *array will be set to the new buffer, *size
1466 * will be set to new_size and 0 will be returned. The size of the reallocated
1467 * refcount array buffer will be aligned to a cluster boundary, and the newly
1468 * allocated area will be zeroed.
1470 static int realloc_refcount_array(BDRVQcow2State *s, void **array,
1471 int64_t *size, int64_t new_size)
1473 int64_t old_byte_size, new_byte_size;
1474 void *new_ptr;
1476 /* Round to clusters so the array can be directly written to disk */
1477 old_byte_size = size_to_clusters(s, refcount_array_byte_size(s, *size))
1478 * s->cluster_size;
1479 new_byte_size = size_to_clusters(s, refcount_array_byte_size(s, new_size))
1480 * s->cluster_size;
1482 if (new_byte_size == old_byte_size) {
1483 *size = new_size;
1484 return 0;
1487 assert(new_byte_size > 0);
1489 if (new_byte_size > SIZE_MAX) {
1490 return -ENOMEM;
1493 new_ptr = g_try_realloc(*array, new_byte_size);
1494 if (!new_ptr) {
1495 return -ENOMEM;
1498 if (new_byte_size > old_byte_size) {
1499 memset((char *)new_ptr + old_byte_size, 0,
1500 new_byte_size - old_byte_size);
1503 *array = new_ptr;
1504 *size = new_size;
1506 return 0;
1510 * Increases the refcount for a range of clusters in a given refcount table.
1511 * This is used to construct a temporary refcount table out of L1 and L2 tables
1512 * which can be compared to the refcount table saved in the image.
1514 * Modifies the number of errors in res.
1516 int qcow2_inc_refcounts_imrt(BlockDriverState *bs, BdrvCheckResult *res,
1517 void **refcount_table,
1518 int64_t *refcount_table_size,
1519 int64_t offset, int64_t size)
1521 BDRVQcow2State *s = bs->opaque;
1522 uint64_t start, last, cluster_offset, k, refcount;
1523 int ret;
1525 if (size <= 0) {
1526 return 0;
1529 start = start_of_cluster(s, offset);
1530 last = start_of_cluster(s, offset + size - 1);
1531 for(cluster_offset = start; cluster_offset <= last;
1532 cluster_offset += s->cluster_size) {
1533 k = cluster_offset >> s->cluster_bits;
1534 if (k >= *refcount_table_size) {
1535 ret = realloc_refcount_array(s, refcount_table,
1536 refcount_table_size, k + 1);
1537 if (ret < 0) {
1538 res->check_errors++;
1539 return ret;
1543 refcount = s->get_refcount(*refcount_table, k);
1544 if (refcount == s->refcount_max) {
1545 fprintf(stderr, "ERROR: overflow cluster offset=0x%" PRIx64
1546 "\n", cluster_offset);
1547 fprintf(stderr, "Use qemu-img amend to increase the refcount entry "
1548 "width or qemu-img convert to create a clean copy if the "
1549 "image cannot be opened for writing\n");
1550 res->corruptions++;
1551 continue;
1553 s->set_refcount(*refcount_table, k, refcount + 1);
1556 return 0;
1559 /* Flags for check_refcounts_l1() and check_refcounts_l2() */
1560 enum {
1561 CHECK_FRAG_INFO = 0x2, /* update BlockFragInfo counters */
1565 * Increases the refcount in the given refcount table for the all clusters
1566 * referenced in the L2 table. While doing so, performs some checks on L2
1567 * entries.
1569 * Returns the number of errors found by the checks or -errno if an internal
1570 * error occurred.
1572 static int check_refcounts_l2(BlockDriverState *bs, BdrvCheckResult *res,
1573 void **refcount_table,
1574 int64_t *refcount_table_size, int64_t l2_offset,
1575 int flags, BdrvCheckMode fix)
1577 BDRVQcow2State *s = bs->opaque;
1578 uint64_t *l2_table, l2_entry;
1579 uint64_t next_contiguous_offset = 0;
1580 int i, l2_size, nb_csectors, ret;
1582 /* Read L2 table from disk */
1583 l2_size = s->l2_size * sizeof(uint64_t);
1584 l2_table = g_malloc(l2_size);
1586 ret = bdrv_pread(bs->file, l2_offset, l2_table, l2_size);
1587 if (ret < 0) {
1588 fprintf(stderr, "ERROR: I/O error in check_refcounts_l2\n");
1589 res->check_errors++;
1590 goto fail;
1593 /* Do the actual checks */
1594 for(i = 0; i < s->l2_size; i++) {
1595 l2_entry = be64_to_cpu(l2_table[i]);
1597 switch (qcow2_get_cluster_type(bs, l2_entry)) {
1598 case QCOW2_CLUSTER_COMPRESSED:
1599 /* Compressed clusters don't have QCOW_OFLAG_COPIED */
1600 if (l2_entry & QCOW_OFLAG_COPIED) {
1601 fprintf(stderr, "ERROR: coffset=0x%" PRIx64 ": "
1602 "copied flag must never be set for compressed "
1603 "clusters\n", l2_entry & s->cluster_offset_mask);
1604 l2_entry &= ~QCOW_OFLAG_COPIED;
1605 res->corruptions++;
1608 if (has_data_file(bs)) {
1609 fprintf(stderr, "ERROR compressed cluster %d with data file, "
1610 "entry=0x%" PRIx64 "\n", i, l2_entry);
1611 res->corruptions++;
1612 break;
1615 /* Mark cluster as used */
1616 nb_csectors = ((l2_entry >> s->csize_shift) &
1617 s->csize_mask) + 1;
1618 l2_entry &= s->cluster_offset_mask;
1619 ret = qcow2_inc_refcounts_imrt(bs, res,
1620 refcount_table, refcount_table_size,
1621 l2_entry & ~511, nb_csectors * 512);
1622 if (ret < 0) {
1623 goto fail;
1626 if (flags & CHECK_FRAG_INFO) {
1627 res->bfi.allocated_clusters++;
1628 res->bfi.compressed_clusters++;
1630 /* Compressed clusters are fragmented by nature. Since they
1631 * take up sub-sector space but we only have sector granularity
1632 * I/O we need to re-read the same sectors even for adjacent
1633 * compressed clusters.
1635 res->bfi.fragmented_clusters++;
1637 break;
1639 case QCOW2_CLUSTER_ZERO_ALLOC:
1640 case QCOW2_CLUSTER_NORMAL:
1642 uint64_t offset = l2_entry & L2E_OFFSET_MASK;
1644 if (flags & CHECK_FRAG_INFO) {
1645 res->bfi.allocated_clusters++;
1646 if (next_contiguous_offset &&
1647 offset != next_contiguous_offset) {
1648 res->bfi.fragmented_clusters++;
1650 next_contiguous_offset = offset + s->cluster_size;
1653 /* Correct offsets are cluster aligned */
1654 if (offset_into_cluster(s, offset)) {
1655 if (qcow2_get_cluster_type(bs, l2_entry) ==
1656 QCOW2_CLUSTER_ZERO_ALLOC)
1658 fprintf(stderr, "%s offset=%" PRIx64 ": Preallocated zero "
1659 "cluster is not properly aligned; L2 entry "
1660 "corrupted.\n",
1661 fix & BDRV_FIX_ERRORS ? "Repairing" : "ERROR",
1662 offset);
1663 if (fix & BDRV_FIX_ERRORS) {
1664 uint64_t l2e_offset =
1665 l2_offset + (uint64_t)i * sizeof(uint64_t);
1667 l2_entry = QCOW_OFLAG_ZERO;
1668 l2_table[i] = cpu_to_be64(l2_entry);
1669 ret = qcow2_pre_write_overlap_check(bs,
1670 QCOW2_OL_ACTIVE_L2 | QCOW2_OL_INACTIVE_L2,
1671 l2e_offset, sizeof(uint64_t), false);
1672 if (ret < 0) {
1673 fprintf(stderr, "ERROR: Overlap check failed\n");
1674 res->check_errors++;
1675 /* Something is seriously wrong, so abort checking
1676 * this L2 table */
1677 goto fail;
1680 ret = bdrv_pwrite_sync(bs->file, l2e_offset,
1681 &l2_table[i], sizeof(uint64_t));
1682 if (ret < 0) {
1683 fprintf(stderr, "ERROR: Failed to overwrite L2 "
1684 "table entry: %s\n", strerror(-ret));
1685 res->check_errors++;
1686 /* Do not abort, continue checking the rest of this
1687 * L2 table's entries */
1688 } else {
1689 res->corruptions_fixed++;
1690 /* Skip marking the cluster as used
1691 * (it is unused now) */
1692 continue;
1694 } else {
1695 res->corruptions++;
1697 } else {
1698 fprintf(stderr, "ERROR offset=%" PRIx64 ": Data cluster is "
1699 "not properly aligned; L2 entry corrupted.\n", offset);
1700 res->corruptions++;
1704 /* Mark cluster as used */
1705 if (!has_data_file(bs)) {
1706 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table,
1707 refcount_table_size,
1708 offset, s->cluster_size);
1709 if (ret < 0) {
1710 goto fail;
1713 break;
1716 case QCOW2_CLUSTER_ZERO_PLAIN:
1717 case QCOW2_CLUSTER_UNALLOCATED:
1718 break;
1720 default:
1721 abort();
1725 g_free(l2_table);
1726 return 0;
1728 fail:
1729 g_free(l2_table);
1730 return ret;
1734 * Increases the refcount for the L1 table, its L2 tables and all referenced
1735 * clusters in the given refcount table. While doing so, performs some checks
1736 * on L1 and L2 entries.
1738 * Returns the number of errors found by the checks or -errno if an internal
1739 * error occurred.
1741 static int check_refcounts_l1(BlockDriverState *bs,
1742 BdrvCheckResult *res,
1743 void **refcount_table,
1744 int64_t *refcount_table_size,
1745 int64_t l1_table_offset, int l1_size,
1746 int flags, BdrvCheckMode fix)
1748 BDRVQcow2State *s = bs->opaque;
1749 uint64_t *l1_table = NULL, l2_offset, l1_size2;
1750 int i, ret;
1752 l1_size2 = l1_size * sizeof(uint64_t);
1754 /* Mark L1 table as used */
1755 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, refcount_table_size,
1756 l1_table_offset, l1_size2);
1757 if (ret < 0) {
1758 goto fail;
1761 /* Read L1 table entries from disk */
1762 if (l1_size2 > 0) {
1763 l1_table = g_try_malloc(l1_size2);
1764 if (l1_table == NULL) {
1765 ret = -ENOMEM;
1766 res->check_errors++;
1767 goto fail;
1769 ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2);
1770 if (ret < 0) {
1771 fprintf(stderr, "ERROR: I/O error in check_refcounts_l1\n");
1772 res->check_errors++;
1773 goto fail;
1775 for(i = 0;i < l1_size; i++)
1776 be64_to_cpus(&l1_table[i]);
1779 /* Do the actual checks */
1780 for(i = 0; i < l1_size; i++) {
1781 l2_offset = l1_table[i];
1782 if (l2_offset) {
1783 /* Mark L2 table as used */
1784 l2_offset &= L1E_OFFSET_MASK;
1785 ret = qcow2_inc_refcounts_imrt(bs, res,
1786 refcount_table, refcount_table_size,
1787 l2_offset, s->cluster_size);
1788 if (ret < 0) {
1789 goto fail;
1792 /* L2 tables are cluster aligned */
1793 if (offset_into_cluster(s, l2_offset)) {
1794 fprintf(stderr, "ERROR l2_offset=%" PRIx64 ": Table is not "
1795 "cluster aligned; L1 entry corrupted\n", l2_offset);
1796 res->corruptions++;
1799 /* Process and check L2 entries */
1800 ret = check_refcounts_l2(bs, res, refcount_table,
1801 refcount_table_size, l2_offset, flags,
1802 fix);
1803 if (ret < 0) {
1804 goto fail;
1808 g_free(l1_table);
1809 return 0;
1811 fail:
1812 g_free(l1_table);
1813 return ret;
1817 * Checks the OFLAG_COPIED flag for all L1 and L2 entries.
1819 * This function does not print an error message nor does it increment
1820 * check_errors if qcow2_get_refcount fails (this is because such an error will
1821 * have been already detected and sufficiently signaled by the calling function
1822 * (qcow2_check_refcounts) by the time this function is called).
1824 static int check_oflag_copied(BlockDriverState *bs, BdrvCheckResult *res,
1825 BdrvCheckMode fix)
1827 BDRVQcow2State *s = bs->opaque;
1828 uint64_t *l2_table = qemu_blockalign(bs, s->cluster_size);
1829 int ret;
1830 uint64_t refcount;
1831 int i, j;
1832 bool repair;
1834 if (fix & BDRV_FIX_ERRORS) {
1835 /* Always repair */
1836 repair = true;
1837 } else if (fix & BDRV_FIX_LEAKS) {
1838 /* Repair only if that seems safe: This function is always
1839 * called after the refcounts have been fixed, so the refcount
1840 * is accurate if that repair was successful */
1841 repair = !res->check_errors && !res->corruptions && !res->leaks;
1842 } else {
1843 repair = false;
1846 for (i = 0; i < s->l1_size; i++) {
1847 uint64_t l1_entry = s->l1_table[i];
1848 uint64_t l2_offset = l1_entry & L1E_OFFSET_MASK;
1849 bool l2_dirty = false;
1851 if (!l2_offset) {
1852 continue;
1855 ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits,
1856 &refcount);
1857 if (ret < 0) {
1858 /* don't print message nor increment check_errors */
1859 continue;
1861 if ((refcount == 1) != ((l1_entry & QCOW_OFLAG_COPIED) != 0)) {
1862 fprintf(stderr, "%s OFLAG_COPIED L2 cluster: l1_index=%d "
1863 "l1_entry=%" PRIx64 " refcount=%" PRIu64 "\n",
1864 repair ? "Repairing" : "ERROR", i, l1_entry, refcount);
1865 if (repair) {
1866 s->l1_table[i] = refcount == 1
1867 ? l1_entry | QCOW_OFLAG_COPIED
1868 : l1_entry & ~QCOW_OFLAG_COPIED;
1869 ret = qcow2_write_l1_entry(bs, i);
1870 if (ret < 0) {
1871 res->check_errors++;
1872 goto fail;
1874 res->corruptions_fixed++;
1875 } else {
1876 res->corruptions++;
1880 ret = bdrv_pread(bs->file, l2_offset, l2_table,
1881 s->l2_size * sizeof(uint64_t));
1882 if (ret < 0) {
1883 fprintf(stderr, "ERROR: Could not read L2 table: %s\n",
1884 strerror(-ret));
1885 res->check_errors++;
1886 goto fail;
1889 for (j = 0; j < s->l2_size; j++) {
1890 uint64_t l2_entry = be64_to_cpu(l2_table[j]);
1891 uint64_t data_offset = l2_entry & L2E_OFFSET_MASK;
1892 QCow2ClusterType cluster_type = qcow2_get_cluster_type(bs, l2_entry);
1894 if (cluster_type == QCOW2_CLUSTER_NORMAL ||
1895 cluster_type == QCOW2_CLUSTER_ZERO_ALLOC) {
1896 if (has_data_file(bs)) {
1897 refcount = 1;
1898 } else {
1899 ret = qcow2_get_refcount(bs,
1900 data_offset >> s->cluster_bits,
1901 &refcount);
1902 if (ret < 0) {
1903 /* don't print message nor increment check_errors */
1904 continue;
1907 if ((refcount == 1) != ((l2_entry & QCOW_OFLAG_COPIED) != 0)) {
1908 fprintf(stderr, "%s OFLAG_COPIED data cluster: "
1909 "l2_entry=%" PRIx64 " refcount=%" PRIu64 "\n",
1910 repair ? "Repairing" : "ERROR", l2_entry, refcount);
1911 if (repair) {
1912 l2_table[j] = cpu_to_be64(refcount == 1
1913 ? l2_entry | QCOW_OFLAG_COPIED
1914 : l2_entry & ~QCOW_OFLAG_COPIED);
1915 l2_dirty = true;
1916 res->corruptions_fixed++;
1917 } else {
1918 res->corruptions++;
1924 if (l2_dirty) {
1925 ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L2,
1926 l2_offset, s->cluster_size,
1927 false);
1928 if (ret < 0) {
1929 fprintf(stderr, "ERROR: Could not write L2 table; metadata "
1930 "overlap check failed: %s\n", strerror(-ret));
1931 res->check_errors++;
1932 goto fail;
1935 ret = bdrv_pwrite(bs->file, l2_offset, l2_table,
1936 s->cluster_size);
1937 if (ret < 0) {
1938 fprintf(stderr, "ERROR: Could not write L2 table: %s\n",
1939 strerror(-ret));
1940 res->check_errors++;
1941 goto fail;
1946 ret = 0;
1948 fail:
1949 qemu_vfree(l2_table);
1950 return ret;
1954 * Checks consistency of refblocks and accounts for each refblock in
1955 * *refcount_table.
1957 static int check_refblocks(BlockDriverState *bs, BdrvCheckResult *res,
1958 BdrvCheckMode fix, bool *rebuild,
1959 void **refcount_table, int64_t *nb_clusters)
1961 BDRVQcow2State *s = bs->opaque;
1962 int64_t i, size;
1963 int ret;
1965 for(i = 0; i < s->refcount_table_size; i++) {
1966 uint64_t offset, cluster;
1967 offset = s->refcount_table[i];
1968 cluster = offset >> s->cluster_bits;
1970 /* Refcount blocks are cluster aligned */
1971 if (offset_into_cluster(s, offset)) {
1972 fprintf(stderr, "ERROR refcount block %" PRId64 " is not "
1973 "cluster aligned; refcount table entry corrupted\n", i);
1974 res->corruptions++;
1975 *rebuild = true;
1976 continue;
1979 if (cluster >= *nb_clusters) {
1980 fprintf(stderr, "%s refcount block %" PRId64 " is outside image\n",
1981 fix & BDRV_FIX_ERRORS ? "Repairing" : "ERROR", i);
1983 if (fix & BDRV_FIX_ERRORS) {
1984 int64_t new_nb_clusters;
1985 Error *local_err = NULL;
1987 if (offset > INT64_MAX - s->cluster_size) {
1988 ret = -EINVAL;
1989 goto resize_fail;
1992 ret = bdrv_truncate(bs->file, offset + s->cluster_size,
1993 PREALLOC_MODE_OFF, &local_err);
1994 if (ret < 0) {
1995 error_report_err(local_err);
1996 goto resize_fail;
1998 size = bdrv_getlength(bs->file->bs);
1999 if (size < 0) {
2000 ret = size;
2001 goto resize_fail;
2004 new_nb_clusters = size_to_clusters(s, size);
2005 assert(new_nb_clusters >= *nb_clusters);
2007 ret = realloc_refcount_array(s, refcount_table,
2008 nb_clusters, new_nb_clusters);
2009 if (ret < 0) {
2010 res->check_errors++;
2011 return ret;
2014 if (cluster >= *nb_clusters) {
2015 ret = -EINVAL;
2016 goto resize_fail;
2019 res->corruptions_fixed++;
2020 ret = qcow2_inc_refcounts_imrt(bs, res,
2021 refcount_table, nb_clusters,
2022 offset, s->cluster_size);
2023 if (ret < 0) {
2024 return ret;
2026 /* No need to check whether the refcount is now greater than 1:
2027 * This area was just allocated and zeroed, so it can only be
2028 * exactly 1 after qcow2_inc_refcounts_imrt() */
2029 continue;
2031 resize_fail:
2032 res->corruptions++;
2033 *rebuild = true;
2034 fprintf(stderr, "ERROR could not resize image: %s\n",
2035 strerror(-ret));
2036 } else {
2037 res->corruptions++;
2039 continue;
2042 if (offset != 0) {
2043 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters,
2044 offset, s->cluster_size);
2045 if (ret < 0) {
2046 return ret;
2048 if (s->get_refcount(*refcount_table, cluster) != 1) {
2049 fprintf(stderr, "ERROR refcount block %" PRId64
2050 " refcount=%" PRIu64 "\n", i,
2051 s->get_refcount(*refcount_table, cluster));
2052 res->corruptions++;
2053 *rebuild = true;
2058 return 0;
2062 * Calculates an in-memory refcount table.
2064 static int calculate_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
2065 BdrvCheckMode fix, bool *rebuild,
2066 void **refcount_table, int64_t *nb_clusters)
2068 BDRVQcow2State *s = bs->opaque;
2069 int64_t i;
2070 QCowSnapshot *sn;
2071 int ret;
2073 if (!*refcount_table) {
2074 int64_t old_size = 0;
2075 ret = realloc_refcount_array(s, refcount_table,
2076 &old_size, *nb_clusters);
2077 if (ret < 0) {
2078 res->check_errors++;
2079 return ret;
2083 /* header */
2084 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters,
2085 0, s->cluster_size);
2086 if (ret < 0) {
2087 return ret;
2090 /* current L1 table */
2091 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
2092 s->l1_table_offset, s->l1_size, CHECK_FRAG_INFO,
2093 fix);
2094 if (ret < 0) {
2095 return ret;
2098 /* snapshots */
2099 if (has_data_file(bs) && s->nb_snapshots) {
2100 fprintf(stderr, "ERROR %d snapshots in image with data file\n",
2101 s->nb_snapshots);
2102 res->corruptions++;
2105 for (i = 0; i < s->nb_snapshots; i++) {
2106 sn = s->snapshots + i;
2107 if (offset_into_cluster(s, sn->l1_table_offset)) {
2108 fprintf(stderr, "ERROR snapshot %s (%s) l1_offset=%#" PRIx64 ": "
2109 "L1 table is not cluster aligned; snapshot table entry "
2110 "corrupted\n", sn->id_str, sn->name, sn->l1_table_offset);
2111 res->corruptions++;
2112 continue;
2114 if (sn->l1_size > QCOW_MAX_L1_SIZE / sizeof(uint64_t)) {
2115 fprintf(stderr, "ERROR snapshot %s (%s) l1_size=%#" PRIx32 ": "
2116 "L1 table is too large; snapshot table entry corrupted\n",
2117 sn->id_str, sn->name, sn->l1_size);
2118 res->corruptions++;
2119 continue;
2121 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
2122 sn->l1_table_offset, sn->l1_size, 0, fix);
2123 if (ret < 0) {
2124 return ret;
2127 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters,
2128 s->snapshots_offset, s->snapshots_size);
2129 if (ret < 0) {
2130 return ret;
2133 /* refcount data */
2134 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters,
2135 s->refcount_table_offset,
2136 s->refcount_table_size * sizeof(uint64_t));
2137 if (ret < 0) {
2138 return ret;
2141 /* encryption */
2142 if (s->crypto_header.length) {
2143 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters,
2144 s->crypto_header.offset,
2145 s->crypto_header.length);
2146 if (ret < 0) {
2147 return ret;
2151 /* bitmaps */
2152 ret = qcow2_check_bitmaps_refcounts(bs, res, refcount_table, nb_clusters);
2153 if (ret < 0) {
2154 return ret;
2157 return check_refblocks(bs, res, fix, rebuild, refcount_table, nb_clusters);
2161 * Compares the actual reference count for each cluster in the image against the
2162 * refcount as reported by the refcount structures on-disk.
2164 static void compare_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
2165 BdrvCheckMode fix, bool *rebuild,
2166 int64_t *highest_cluster,
2167 void *refcount_table, int64_t nb_clusters)
2169 BDRVQcow2State *s = bs->opaque;
2170 int64_t i;
2171 uint64_t refcount1, refcount2;
2172 int ret;
2174 for (i = 0, *highest_cluster = 0; i < nb_clusters; i++) {
2175 ret = qcow2_get_refcount(bs, i, &refcount1);
2176 if (ret < 0) {
2177 fprintf(stderr, "Can't get refcount for cluster %" PRId64 ": %s\n",
2178 i, strerror(-ret));
2179 res->check_errors++;
2180 continue;
2183 refcount2 = s->get_refcount(refcount_table, i);
2185 if (refcount1 > 0 || refcount2 > 0) {
2186 *highest_cluster = i;
2189 if (refcount1 != refcount2) {
2190 /* Check if we're allowed to fix the mismatch */
2191 int *num_fixed = NULL;
2192 if (refcount1 == 0) {
2193 *rebuild = true;
2194 } else if (refcount1 > refcount2 && (fix & BDRV_FIX_LEAKS)) {
2195 num_fixed = &res->leaks_fixed;
2196 } else if (refcount1 < refcount2 && (fix & BDRV_FIX_ERRORS)) {
2197 num_fixed = &res->corruptions_fixed;
2200 fprintf(stderr, "%s cluster %" PRId64 " refcount=%" PRIu64
2201 " reference=%" PRIu64 "\n",
2202 num_fixed != NULL ? "Repairing" :
2203 refcount1 < refcount2 ? "ERROR" :
2204 "Leaked",
2205 i, refcount1, refcount2);
2207 if (num_fixed) {
2208 ret = update_refcount(bs, i << s->cluster_bits, 1,
2209 refcount_diff(refcount1, refcount2),
2210 refcount1 > refcount2,
2211 QCOW2_DISCARD_ALWAYS);
2212 if (ret >= 0) {
2213 (*num_fixed)++;
2214 continue;
2218 /* And if we couldn't, print an error */
2219 if (refcount1 < refcount2) {
2220 res->corruptions++;
2221 } else {
2222 res->leaks++;
2229 * Allocates clusters using an in-memory refcount table (IMRT) in contrast to
2230 * the on-disk refcount structures.
2232 * On input, *first_free_cluster tells where to start looking, and need not
2233 * actually be a free cluster; the returned offset will not be before that
2234 * cluster. On output, *first_free_cluster points to the first gap found, even
2235 * if that gap was too small to be used as the returned offset.
2237 * Note that *first_free_cluster is a cluster index whereas the return value is
2238 * an offset.
2240 static int64_t alloc_clusters_imrt(BlockDriverState *bs,
2241 int cluster_count,
2242 void **refcount_table,
2243 int64_t *imrt_nb_clusters,
2244 int64_t *first_free_cluster)
2246 BDRVQcow2State *s = bs->opaque;
2247 int64_t cluster = *first_free_cluster, i;
2248 bool first_gap = true;
2249 int contiguous_free_clusters;
2250 int ret;
2252 /* Starting at *first_free_cluster, find a range of at least cluster_count
2253 * continuously free clusters */
2254 for (contiguous_free_clusters = 0;
2255 cluster < *imrt_nb_clusters &&
2256 contiguous_free_clusters < cluster_count;
2257 cluster++)
2259 if (!s->get_refcount(*refcount_table, cluster)) {
2260 contiguous_free_clusters++;
2261 if (first_gap) {
2262 /* If this is the first free cluster found, update
2263 * *first_free_cluster accordingly */
2264 *first_free_cluster = cluster;
2265 first_gap = false;
2267 } else if (contiguous_free_clusters) {
2268 contiguous_free_clusters = 0;
2272 /* If contiguous_free_clusters is greater than zero, it contains the number
2273 * of continuously free clusters until the current cluster; the first free
2274 * cluster in the current "gap" is therefore
2275 * cluster - contiguous_free_clusters */
2277 /* If no such range could be found, grow the in-memory refcount table
2278 * accordingly to append free clusters at the end of the image */
2279 if (contiguous_free_clusters < cluster_count) {
2280 /* contiguous_free_clusters clusters are already empty at the image end;
2281 * we need cluster_count clusters; therefore, we have to allocate
2282 * cluster_count - contiguous_free_clusters new clusters at the end of
2283 * the image (which is the current value of cluster; note that cluster
2284 * may exceed old_imrt_nb_clusters if *first_free_cluster pointed beyond
2285 * the image end) */
2286 ret = realloc_refcount_array(s, refcount_table, imrt_nb_clusters,
2287 cluster + cluster_count
2288 - contiguous_free_clusters);
2289 if (ret < 0) {
2290 return ret;
2294 /* Go back to the first free cluster */
2295 cluster -= contiguous_free_clusters;
2296 for (i = 0; i < cluster_count; i++) {
2297 s->set_refcount(*refcount_table, cluster + i, 1);
2300 return cluster << s->cluster_bits;
2304 * Creates a new refcount structure based solely on the in-memory information
2305 * given through *refcount_table. All necessary allocations will be reflected
2306 * in that array.
2308 * On success, the old refcount structure is leaked (it will be covered by the
2309 * new refcount structure).
2311 static int rebuild_refcount_structure(BlockDriverState *bs,
2312 BdrvCheckResult *res,
2313 void **refcount_table,
2314 int64_t *nb_clusters)
2316 BDRVQcow2State *s = bs->opaque;
2317 int64_t first_free_cluster = 0, reftable_offset = -1, cluster = 0;
2318 int64_t refblock_offset, refblock_start, refblock_index;
2319 uint32_t reftable_size = 0;
2320 uint64_t *on_disk_reftable = NULL;
2321 void *on_disk_refblock;
2322 int ret = 0;
2323 struct {
2324 uint64_t reftable_offset;
2325 uint32_t reftable_clusters;
2326 } QEMU_PACKED reftable_offset_and_clusters;
2328 qcow2_cache_empty(bs, s->refcount_block_cache);
2330 write_refblocks:
2331 for (; cluster < *nb_clusters; cluster++) {
2332 if (!s->get_refcount(*refcount_table, cluster)) {
2333 continue;
2336 refblock_index = cluster >> s->refcount_block_bits;
2337 refblock_start = refblock_index << s->refcount_block_bits;
2339 /* Don't allocate a cluster in a refblock already written to disk */
2340 if (first_free_cluster < refblock_start) {
2341 first_free_cluster = refblock_start;
2343 refblock_offset = alloc_clusters_imrt(bs, 1, refcount_table,
2344 nb_clusters, &first_free_cluster);
2345 if (refblock_offset < 0) {
2346 fprintf(stderr, "ERROR allocating refblock: %s\n",
2347 strerror(-refblock_offset));
2348 res->check_errors++;
2349 ret = refblock_offset;
2350 goto fail;
2353 if (reftable_size <= refblock_index) {
2354 uint32_t old_reftable_size = reftable_size;
2355 uint64_t *new_on_disk_reftable;
2357 reftable_size = ROUND_UP((refblock_index + 1) * sizeof(uint64_t),
2358 s->cluster_size) / sizeof(uint64_t);
2359 new_on_disk_reftable = g_try_realloc(on_disk_reftable,
2360 reftable_size *
2361 sizeof(uint64_t));
2362 if (!new_on_disk_reftable) {
2363 res->check_errors++;
2364 ret = -ENOMEM;
2365 goto fail;
2367 on_disk_reftable = new_on_disk_reftable;
2369 memset(on_disk_reftable + old_reftable_size, 0,
2370 (reftable_size - old_reftable_size) * sizeof(uint64_t));
2372 /* The offset we have for the reftable is now no longer valid;
2373 * this will leak that range, but we can easily fix that by running
2374 * a leak-fixing check after this rebuild operation */
2375 reftable_offset = -1;
2376 } else {
2377 assert(on_disk_reftable);
2379 on_disk_reftable[refblock_index] = refblock_offset;
2381 /* If this is apparently the last refblock (for now), try to squeeze the
2382 * reftable in */
2383 if (refblock_index == (*nb_clusters - 1) >> s->refcount_block_bits &&
2384 reftable_offset < 0)
2386 uint64_t reftable_clusters = size_to_clusters(s, reftable_size *
2387 sizeof(uint64_t));
2388 reftable_offset = alloc_clusters_imrt(bs, reftable_clusters,
2389 refcount_table, nb_clusters,
2390 &first_free_cluster);
2391 if (reftable_offset < 0) {
2392 fprintf(stderr, "ERROR allocating reftable: %s\n",
2393 strerror(-reftable_offset));
2394 res->check_errors++;
2395 ret = reftable_offset;
2396 goto fail;
2400 ret = qcow2_pre_write_overlap_check(bs, 0, refblock_offset,
2401 s->cluster_size, false);
2402 if (ret < 0) {
2403 fprintf(stderr, "ERROR writing refblock: %s\n", strerror(-ret));
2404 goto fail;
2407 /* The size of *refcount_table is always cluster-aligned, therefore the
2408 * write operation will not overflow */
2409 on_disk_refblock = (void *)((char *) *refcount_table +
2410 refblock_index * s->cluster_size);
2412 ret = bdrv_write(bs->file, refblock_offset / BDRV_SECTOR_SIZE,
2413 on_disk_refblock, s->cluster_sectors);
2414 if (ret < 0) {
2415 fprintf(stderr, "ERROR writing refblock: %s\n", strerror(-ret));
2416 goto fail;
2419 /* Go to the end of this refblock */
2420 cluster = refblock_start + s->refcount_block_size - 1;
2423 if (reftable_offset < 0) {
2424 uint64_t post_refblock_start, reftable_clusters;
2426 post_refblock_start = ROUND_UP(*nb_clusters, s->refcount_block_size);
2427 reftable_clusters = size_to_clusters(s,
2428 reftable_size * sizeof(uint64_t));
2429 /* Not pretty but simple */
2430 if (first_free_cluster < post_refblock_start) {
2431 first_free_cluster = post_refblock_start;
2433 reftable_offset = alloc_clusters_imrt(bs, reftable_clusters,
2434 refcount_table, nb_clusters,
2435 &first_free_cluster);
2436 if (reftable_offset < 0) {
2437 fprintf(stderr, "ERROR allocating reftable: %s\n",
2438 strerror(-reftable_offset));
2439 res->check_errors++;
2440 ret = reftable_offset;
2441 goto fail;
2444 goto write_refblocks;
2447 for (refblock_index = 0; refblock_index < reftable_size; refblock_index++) {
2448 cpu_to_be64s(&on_disk_reftable[refblock_index]);
2451 ret = qcow2_pre_write_overlap_check(bs, 0, reftable_offset,
2452 reftable_size * sizeof(uint64_t),
2453 false);
2454 if (ret < 0) {
2455 fprintf(stderr, "ERROR writing reftable: %s\n", strerror(-ret));
2456 goto fail;
2459 assert(reftable_size < INT_MAX / sizeof(uint64_t));
2460 ret = bdrv_pwrite(bs->file, reftable_offset, on_disk_reftable,
2461 reftable_size * sizeof(uint64_t));
2462 if (ret < 0) {
2463 fprintf(stderr, "ERROR writing reftable: %s\n", strerror(-ret));
2464 goto fail;
2467 /* Enter new reftable into the image header */
2468 reftable_offset_and_clusters.reftable_offset = cpu_to_be64(reftable_offset);
2469 reftable_offset_and_clusters.reftable_clusters =
2470 cpu_to_be32(size_to_clusters(s, reftable_size * sizeof(uint64_t)));
2471 ret = bdrv_pwrite_sync(bs->file,
2472 offsetof(QCowHeader, refcount_table_offset),
2473 &reftable_offset_and_clusters,
2474 sizeof(reftable_offset_and_clusters));
2475 if (ret < 0) {
2476 fprintf(stderr, "ERROR setting reftable: %s\n", strerror(-ret));
2477 goto fail;
2480 for (refblock_index = 0; refblock_index < reftable_size; refblock_index++) {
2481 be64_to_cpus(&on_disk_reftable[refblock_index]);
2483 s->refcount_table = on_disk_reftable;
2484 s->refcount_table_offset = reftable_offset;
2485 s->refcount_table_size = reftable_size;
2486 update_max_refcount_table_index(s);
2488 return 0;
2490 fail:
2491 g_free(on_disk_reftable);
2492 return ret;
2496 * Checks an image for refcount consistency.
2498 * Returns 0 if no errors are found, the number of errors in case the image is
2499 * detected as corrupted, and -errno when an internal error occurred.
2501 int qcow2_check_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
2502 BdrvCheckMode fix)
2504 BDRVQcow2State *s = bs->opaque;
2505 BdrvCheckResult pre_compare_res;
2506 int64_t size, highest_cluster, nb_clusters;
2507 void *refcount_table = NULL;
2508 bool rebuild = false;
2509 int ret;
2511 size = bdrv_getlength(bs->file->bs);
2512 if (size < 0) {
2513 res->check_errors++;
2514 return size;
2517 nb_clusters = size_to_clusters(s, size);
2518 if (nb_clusters > INT_MAX) {
2519 res->check_errors++;
2520 return -EFBIG;
2523 res->bfi.total_clusters =
2524 size_to_clusters(s, bs->total_sectors * BDRV_SECTOR_SIZE);
2526 ret = calculate_refcounts(bs, res, fix, &rebuild, &refcount_table,
2527 &nb_clusters);
2528 if (ret < 0) {
2529 goto fail;
2532 /* In case we don't need to rebuild the refcount structure (but want to fix
2533 * something), this function is immediately called again, in which case the
2534 * result should be ignored */
2535 pre_compare_res = *res;
2536 compare_refcounts(bs, res, 0, &rebuild, &highest_cluster, refcount_table,
2537 nb_clusters);
2539 if (rebuild && (fix & BDRV_FIX_ERRORS)) {
2540 BdrvCheckResult old_res = *res;
2541 int fresh_leaks = 0;
2543 fprintf(stderr, "Rebuilding refcount structure\n");
2544 ret = rebuild_refcount_structure(bs, res, &refcount_table,
2545 &nb_clusters);
2546 if (ret < 0) {
2547 goto fail;
2550 res->corruptions = 0;
2551 res->leaks = 0;
2553 /* Because the old reftable has been exchanged for a new one the
2554 * references have to be recalculated */
2555 rebuild = false;
2556 memset(refcount_table, 0, refcount_array_byte_size(s, nb_clusters));
2557 ret = calculate_refcounts(bs, res, 0, &rebuild, &refcount_table,
2558 &nb_clusters);
2559 if (ret < 0) {
2560 goto fail;
2563 if (fix & BDRV_FIX_LEAKS) {
2564 /* The old refcount structures are now leaked, fix it; the result
2565 * can be ignored, aside from leaks which were introduced by
2566 * rebuild_refcount_structure() that could not be fixed */
2567 BdrvCheckResult saved_res = *res;
2568 *res = (BdrvCheckResult){ 0 };
2570 compare_refcounts(bs, res, BDRV_FIX_LEAKS, &rebuild,
2571 &highest_cluster, refcount_table, nb_clusters);
2572 if (rebuild) {
2573 fprintf(stderr, "ERROR rebuilt refcount structure is still "
2574 "broken\n");
2577 /* Any leaks accounted for here were introduced by
2578 * rebuild_refcount_structure() because that function has created a
2579 * new refcount structure from scratch */
2580 fresh_leaks = res->leaks;
2581 *res = saved_res;
2584 if (res->corruptions < old_res.corruptions) {
2585 res->corruptions_fixed += old_res.corruptions - res->corruptions;
2587 if (res->leaks < old_res.leaks) {
2588 res->leaks_fixed += old_res.leaks - res->leaks;
2590 res->leaks += fresh_leaks;
2591 } else if (fix) {
2592 if (rebuild) {
2593 fprintf(stderr, "ERROR need to rebuild refcount structures\n");
2594 res->check_errors++;
2595 ret = -EIO;
2596 goto fail;
2599 if (res->leaks || res->corruptions) {
2600 *res = pre_compare_res;
2601 compare_refcounts(bs, res, fix, &rebuild, &highest_cluster,
2602 refcount_table, nb_clusters);
2606 /* check OFLAG_COPIED */
2607 ret = check_oflag_copied(bs, res, fix);
2608 if (ret < 0) {
2609 goto fail;
2612 res->image_end_offset = (highest_cluster + 1) * s->cluster_size;
2613 ret = 0;
2615 fail:
2616 g_free(refcount_table);
2618 return ret;
2621 #define overlaps_with(ofs, sz) \
2622 ranges_overlap(offset, size, ofs, sz)
2625 * Checks if the given offset into the image file is actually free to use by
2626 * looking for overlaps with important metadata sections (L1/L2 tables etc.),
2627 * i.e. a sanity check without relying on the refcount tables.
2629 * The ign parameter specifies what checks not to perform (being a bitmask of
2630 * QCow2MetadataOverlap values), i.e., what sections to ignore.
2632 * Returns:
2633 * - 0 if writing to this offset will not affect the mentioned metadata
2634 * - a positive QCow2MetadataOverlap value indicating one overlapping section
2635 * - a negative value (-errno) indicating an error while performing a check,
2636 * e.g. when bdrv_read failed on QCOW2_OL_INACTIVE_L2
2638 int qcow2_check_metadata_overlap(BlockDriverState *bs, int ign, int64_t offset,
2639 int64_t size)
2641 BDRVQcow2State *s = bs->opaque;
2642 int chk = s->overlap_check & ~ign;
2643 int i, j;
2645 if (!size) {
2646 return 0;
2649 if (chk & QCOW2_OL_MAIN_HEADER) {
2650 if (offset < s->cluster_size) {
2651 return QCOW2_OL_MAIN_HEADER;
2655 /* align range to test to cluster boundaries */
2656 size = ROUND_UP(offset_into_cluster(s, offset) + size, s->cluster_size);
2657 offset = start_of_cluster(s, offset);
2659 if ((chk & QCOW2_OL_ACTIVE_L1) && s->l1_size) {
2660 if (overlaps_with(s->l1_table_offset, s->l1_size * sizeof(uint64_t))) {
2661 return QCOW2_OL_ACTIVE_L1;
2665 if ((chk & QCOW2_OL_REFCOUNT_TABLE) && s->refcount_table_size) {
2666 if (overlaps_with(s->refcount_table_offset,
2667 s->refcount_table_size * sizeof(uint64_t))) {
2668 return QCOW2_OL_REFCOUNT_TABLE;
2672 if ((chk & QCOW2_OL_SNAPSHOT_TABLE) && s->snapshots_size) {
2673 if (overlaps_with(s->snapshots_offset, s->snapshots_size)) {
2674 return QCOW2_OL_SNAPSHOT_TABLE;
2678 if ((chk & QCOW2_OL_INACTIVE_L1) && s->snapshots) {
2679 for (i = 0; i < s->nb_snapshots; i++) {
2680 if (s->snapshots[i].l1_size &&
2681 overlaps_with(s->snapshots[i].l1_table_offset,
2682 s->snapshots[i].l1_size * sizeof(uint64_t))) {
2683 return QCOW2_OL_INACTIVE_L1;
2688 if ((chk & QCOW2_OL_ACTIVE_L2) && s->l1_table) {
2689 for (i = 0; i < s->l1_size; i++) {
2690 if ((s->l1_table[i] & L1E_OFFSET_MASK) &&
2691 overlaps_with(s->l1_table[i] & L1E_OFFSET_MASK,
2692 s->cluster_size)) {
2693 return QCOW2_OL_ACTIVE_L2;
2698 if ((chk & QCOW2_OL_REFCOUNT_BLOCK) && s->refcount_table) {
2699 unsigned last_entry = s->max_refcount_table_index;
2700 assert(last_entry < s->refcount_table_size);
2701 assert(last_entry + 1 == s->refcount_table_size ||
2702 (s->refcount_table[last_entry + 1] & REFT_OFFSET_MASK) == 0);
2703 for (i = 0; i <= last_entry; i++) {
2704 if ((s->refcount_table[i] & REFT_OFFSET_MASK) &&
2705 overlaps_with(s->refcount_table[i] & REFT_OFFSET_MASK,
2706 s->cluster_size)) {
2707 return QCOW2_OL_REFCOUNT_BLOCK;
2712 if ((chk & QCOW2_OL_INACTIVE_L2) && s->snapshots) {
2713 for (i = 0; i < s->nb_snapshots; i++) {
2714 uint64_t l1_ofs = s->snapshots[i].l1_table_offset;
2715 uint32_t l1_sz = s->snapshots[i].l1_size;
2716 uint64_t l1_sz2 = l1_sz * sizeof(uint64_t);
2717 uint64_t *l1;
2718 int ret;
2720 ret = qcow2_validate_table(bs, l1_ofs, l1_sz, sizeof(uint64_t),
2721 QCOW_MAX_L1_SIZE, "", NULL);
2722 if (ret < 0) {
2723 return ret;
2726 l1 = g_try_malloc(l1_sz2);
2728 if (l1_sz2 && l1 == NULL) {
2729 return -ENOMEM;
2732 ret = bdrv_pread(bs->file, l1_ofs, l1, l1_sz2);
2733 if (ret < 0) {
2734 g_free(l1);
2735 return ret;
2738 for (j = 0; j < l1_sz; j++) {
2739 uint64_t l2_ofs = be64_to_cpu(l1[j]) & L1E_OFFSET_MASK;
2740 if (l2_ofs && overlaps_with(l2_ofs, s->cluster_size)) {
2741 g_free(l1);
2742 return QCOW2_OL_INACTIVE_L2;
2746 g_free(l1);
2750 if ((chk & QCOW2_OL_BITMAP_DIRECTORY) &&
2751 (s->autoclear_features & QCOW2_AUTOCLEAR_BITMAPS))
2753 if (overlaps_with(s->bitmap_directory_offset,
2754 s->bitmap_directory_size))
2756 return QCOW2_OL_BITMAP_DIRECTORY;
2760 return 0;
2763 static const char *metadata_ol_names[] = {
2764 [QCOW2_OL_MAIN_HEADER_BITNR] = "qcow2_header",
2765 [QCOW2_OL_ACTIVE_L1_BITNR] = "active L1 table",
2766 [QCOW2_OL_ACTIVE_L2_BITNR] = "active L2 table",
2767 [QCOW2_OL_REFCOUNT_TABLE_BITNR] = "refcount table",
2768 [QCOW2_OL_REFCOUNT_BLOCK_BITNR] = "refcount block",
2769 [QCOW2_OL_SNAPSHOT_TABLE_BITNR] = "snapshot table",
2770 [QCOW2_OL_INACTIVE_L1_BITNR] = "inactive L1 table",
2771 [QCOW2_OL_INACTIVE_L2_BITNR] = "inactive L2 table",
2772 [QCOW2_OL_BITMAP_DIRECTORY_BITNR] = "bitmap directory",
2774 QEMU_BUILD_BUG_ON(QCOW2_OL_MAX_BITNR != ARRAY_SIZE(metadata_ol_names));
2777 * First performs a check for metadata overlaps (through
2778 * qcow2_check_metadata_overlap); if that fails with a negative value (error
2779 * while performing a check), that value is returned. If an impending overlap
2780 * is detected, the BDS will be made unusable, the qcow2 file marked corrupt
2781 * and -EIO returned.
2783 * Returns 0 if there were neither overlaps nor errors while checking for
2784 * overlaps; or a negative value (-errno) on error.
2786 int qcow2_pre_write_overlap_check(BlockDriverState *bs, int ign, int64_t offset,
2787 int64_t size, bool data_file)
2789 int ret;
2791 if (data_file && has_data_file(bs)) {
2792 return 0;
2795 ret = qcow2_check_metadata_overlap(bs, ign, offset, size);
2796 if (ret < 0) {
2797 return ret;
2798 } else if (ret > 0) {
2799 int metadata_ol_bitnr = ctz32(ret);
2800 assert(metadata_ol_bitnr < QCOW2_OL_MAX_BITNR);
2802 qcow2_signal_corruption(bs, true, offset, size, "Preventing invalid "
2803 "write on metadata (overlaps with %s)",
2804 metadata_ol_names[metadata_ol_bitnr]);
2805 return -EIO;
2808 return 0;
2811 /* A pointer to a function of this type is given to walk_over_reftable(). That
2812 * function will create refblocks and pass them to a RefblockFinishOp once they
2813 * are completed (@refblock). @refblock_empty is set if the refblock is
2814 * completely empty.
2816 * Along with the refblock, a corresponding reftable entry is passed, in the
2817 * reftable @reftable (which may be reallocated) at @reftable_index.
2819 * @allocated should be set to true if a new cluster has been allocated.
2821 typedef int (RefblockFinishOp)(BlockDriverState *bs, uint64_t **reftable,
2822 uint64_t reftable_index, uint64_t *reftable_size,
2823 void *refblock, bool refblock_empty,
2824 bool *allocated, Error **errp);
2827 * This "operation" for walk_over_reftable() allocates the refblock on disk (if
2828 * it is not empty) and inserts its offset into the new reftable. The size of
2829 * this new reftable is increased as required.
2831 static int alloc_refblock(BlockDriverState *bs, uint64_t **reftable,
2832 uint64_t reftable_index, uint64_t *reftable_size,
2833 void *refblock, bool refblock_empty, bool *allocated,
2834 Error **errp)
2836 BDRVQcow2State *s = bs->opaque;
2837 int64_t offset;
2839 if (!refblock_empty && reftable_index >= *reftable_size) {
2840 uint64_t *new_reftable;
2841 uint64_t new_reftable_size;
2843 new_reftable_size = ROUND_UP(reftable_index + 1,
2844 s->cluster_size / sizeof(uint64_t));
2845 if (new_reftable_size > QCOW_MAX_REFTABLE_SIZE / sizeof(uint64_t)) {
2846 error_setg(errp,
2847 "This operation would make the refcount table grow "
2848 "beyond the maximum size supported by QEMU, aborting");
2849 return -ENOTSUP;
2852 new_reftable = g_try_realloc(*reftable, new_reftable_size *
2853 sizeof(uint64_t));
2854 if (!new_reftable) {
2855 error_setg(errp, "Failed to increase reftable buffer size");
2856 return -ENOMEM;
2859 memset(new_reftable + *reftable_size, 0,
2860 (new_reftable_size - *reftable_size) * sizeof(uint64_t));
2862 *reftable = new_reftable;
2863 *reftable_size = new_reftable_size;
2866 if (!refblock_empty && !(*reftable)[reftable_index]) {
2867 offset = qcow2_alloc_clusters(bs, s->cluster_size);
2868 if (offset < 0) {
2869 error_setg_errno(errp, -offset, "Failed to allocate refblock");
2870 return offset;
2872 (*reftable)[reftable_index] = offset;
2873 *allocated = true;
2876 return 0;
2880 * This "operation" for walk_over_reftable() writes the refblock to disk at the
2881 * offset specified by the new reftable's entry. It does not modify the new
2882 * reftable or change any refcounts.
2884 static int flush_refblock(BlockDriverState *bs, uint64_t **reftable,
2885 uint64_t reftable_index, uint64_t *reftable_size,
2886 void *refblock, bool refblock_empty, bool *allocated,
2887 Error **errp)
2889 BDRVQcow2State *s = bs->opaque;
2890 int64_t offset;
2891 int ret;
2893 if (reftable_index < *reftable_size && (*reftable)[reftable_index]) {
2894 offset = (*reftable)[reftable_index];
2896 ret = qcow2_pre_write_overlap_check(bs, 0, offset, s->cluster_size,
2897 false);
2898 if (ret < 0) {
2899 error_setg_errno(errp, -ret, "Overlap check failed");
2900 return ret;
2903 ret = bdrv_pwrite(bs->file, offset, refblock, s->cluster_size);
2904 if (ret < 0) {
2905 error_setg_errno(errp, -ret, "Failed to write refblock");
2906 return ret;
2908 } else {
2909 assert(refblock_empty);
2912 return 0;
2916 * This function walks over the existing reftable and every referenced refblock;
2917 * if @new_set_refcount is non-NULL, it is called for every refcount entry to
2918 * create an equal new entry in the passed @new_refblock. Once that
2919 * @new_refblock is completely filled, @operation will be called.
2921 * @status_cb and @cb_opaque are used for the amend operation's status callback.
2922 * @index is the index of the walk_over_reftable() calls and @total is the total
2923 * number of walk_over_reftable() calls per amend operation. Both are used for
2924 * calculating the parameters for the status callback.
2926 * @allocated is set to true if a new cluster has been allocated.
2928 static int walk_over_reftable(BlockDriverState *bs, uint64_t **new_reftable,
2929 uint64_t *new_reftable_index,
2930 uint64_t *new_reftable_size,
2931 void *new_refblock, int new_refblock_size,
2932 int new_refcount_bits,
2933 RefblockFinishOp *operation, bool *allocated,
2934 Qcow2SetRefcountFunc *new_set_refcount,
2935 BlockDriverAmendStatusCB *status_cb,
2936 void *cb_opaque, int index, int total,
2937 Error **errp)
2939 BDRVQcow2State *s = bs->opaque;
2940 uint64_t reftable_index;
2941 bool new_refblock_empty = true;
2942 int refblock_index;
2943 int new_refblock_index = 0;
2944 int ret;
2946 for (reftable_index = 0; reftable_index < s->refcount_table_size;
2947 reftable_index++)
2949 uint64_t refblock_offset = s->refcount_table[reftable_index]
2950 & REFT_OFFSET_MASK;
2952 status_cb(bs, (uint64_t)index * s->refcount_table_size + reftable_index,
2953 (uint64_t)total * s->refcount_table_size, cb_opaque);
2955 if (refblock_offset) {
2956 void *refblock;
2958 if (offset_into_cluster(s, refblock_offset)) {
2959 qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#"
2960 PRIx64 " unaligned (reftable index: %#"
2961 PRIx64 ")", refblock_offset,
2962 reftable_index);
2963 error_setg(errp,
2964 "Image is corrupt (unaligned refblock offset)");
2965 return -EIO;
2968 ret = qcow2_cache_get(bs, s->refcount_block_cache, refblock_offset,
2969 &refblock);
2970 if (ret < 0) {
2971 error_setg_errno(errp, -ret, "Failed to retrieve refblock");
2972 return ret;
2975 for (refblock_index = 0; refblock_index < s->refcount_block_size;
2976 refblock_index++)
2978 uint64_t refcount;
2980 if (new_refblock_index >= new_refblock_size) {
2981 /* new_refblock is now complete */
2982 ret = operation(bs, new_reftable, *new_reftable_index,
2983 new_reftable_size, new_refblock,
2984 new_refblock_empty, allocated, errp);
2985 if (ret < 0) {
2986 qcow2_cache_put(s->refcount_block_cache, &refblock);
2987 return ret;
2990 (*new_reftable_index)++;
2991 new_refblock_index = 0;
2992 new_refblock_empty = true;
2995 refcount = s->get_refcount(refblock, refblock_index);
2996 if (new_refcount_bits < 64 && refcount >> new_refcount_bits) {
2997 uint64_t offset;
2999 qcow2_cache_put(s->refcount_block_cache, &refblock);
3001 offset = ((reftable_index << s->refcount_block_bits)
3002 + refblock_index) << s->cluster_bits;
3004 error_setg(errp, "Cannot decrease refcount entry width to "
3005 "%i bits: Cluster at offset %#" PRIx64 " has a "
3006 "refcount of %" PRIu64, new_refcount_bits,
3007 offset, refcount);
3008 return -EINVAL;
3011 if (new_set_refcount) {
3012 new_set_refcount(new_refblock, new_refblock_index++,
3013 refcount);
3014 } else {
3015 new_refblock_index++;
3017 new_refblock_empty = new_refblock_empty && refcount == 0;
3020 qcow2_cache_put(s->refcount_block_cache, &refblock);
3021 } else {
3022 /* No refblock means every refcount is 0 */
3023 for (refblock_index = 0; refblock_index < s->refcount_block_size;
3024 refblock_index++)
3026 if (new_refblock_index >= new_refblock_size) {
3027 /* new_refblock is now complete */
3028 ret = operation(bs, new_reftable, *new_reftable_index,
3029 new_reftable_size, new_refblock,
3030 new_refblock_empty, allocated, errp);
3031 if (ret < 0) {
3032 return ret;
3035 (*new_reftable_index)++;
3036 new_refblock_index = 0;
3037 new_refblock_empty = true;
3040 if (new_set_refcount) {
3041 new_set_refcount(new_refblock, new_refblock_index++, 0);
3042 } else {
3043 new_refblock_index++;
3049 if (new_refblock_index > 0) {
3050 /* Complete the potentially existing partially filled final refblock */
3051 if (new_set_refcount) {
3052 for (; new_refblock_index < new_refblock_size;
3053 new_refblock_index++)
3055 new_set_refcount(new_refblock, new_refblock_index, 0);
3059 ret = operation(bs, new_reftable, *new_reftable_index,
3060 new_reftable_size, new_refblock, new_refblock_empty,
3061 allocated, errp);
3062 if (ret < 0) {
3063 return ret;
3066 (*new_reftable_index)++;
3069 status_cb(bs, (uint64_t)(index + 1) * s->refcount_table_size,
3070 (uint64_t)total * s->refcount_table_size, cb_opaque);
3072 return 0;
3075 int qcow2_change_refcount_order(BlockDriverState *bs, int refcount_order,
3076 BlockDriverAmendStatusCB *status_cb,
3077 void *cb_opaque, Error **errp)
3079 BDRVQcow2State *s = bs->opaque;
3080 Qcow2GetRefcountFunc *new_get_refcount;
3081 Qcow2SetRefcountFunc *new_set_refcount;
3082 void *new_refblock = qemu_blockalign(bs->file->bs, s->cluster_size);
3083 uint64_t *new_reftable = NULL, new_reftable_size = 0;
3084 uint64_t *old_reftable, old_reftable_size, old_reftable_offset;
3085 uint64_t new_reftable_index = 0;
3086 uint64_t i;
3087 int64_t new_reftable_offset = 0, allocated_reftable_size = 0;
3088 int new_refblock_size, new_refcount_bits = 1 << refcount_order;
3089 int old_refcount_order;
3090 int walk_index = 0;
3091 int ret;
3092 bool new_allocation;
3094 assert(s->qcow_version >= 3);
3095 assert(refcount_order >= 0 && refcount_order <= 6);
3097 /* see qcow2_open() */
3098 new_refblock_size = 1 << (s->cluster_bits - (refcount_order - 3));
3100 new_get_refcount = get_refcount_funcs[refcount_order];
3101 new_set_refcount = set_refcount_funcs[refcount_order];
3104 do {
3105 int total_walks;
3107 new_allocation = false;
3109 /* At least we have to do this walk and the one which writes the
3110 * refblocks; also, at least we have to do this loop here at least
3111 * twice (normally), first to do the allocations, and second to
3112 * determine that everything is correctly allocated, this then makes
3113 * three walks in total */
3114 total_walks = MAX(walk_index + 2, 3);
3116 /* First, allocate the structures so they are present in the refcount
3117 * structures */
3118 ret = walk_over_reftable(bs, &new_reftable, &new_reftable_index,
3119 &new_reftable_size, NULL, new_refblock_size,
3120 new_refcount_bits, &alloc_refblock,
3121 &new_allocation, NULL, status_cb, cb_opaque,
3122 walk_index++, total_walks, errp);
3123 if (ret < 0) {
3124 goto done;
3127 new_reftable_index = 0;
3129 if (new_allocation) {
3130 if (new_reftable_offset) {
3131 qcow2_free_clusters(bs, new_reftable_offset,
3132 allocated_reftable_size * sizeof(uint64_t),
3133 QCOW2_DISCARD_NEVER);
3136 new_reftable_offset = qcow2_alloc_clusters(bs, new_reftable_size *
3137 sizeof(uint64_t));
3138 if (new_reftable_offset < 0) {
3139 error_setg_errno(errp, -new_reftable_offset,
3140 "Failed to allocate the new reftable");
3141 ret = new_reftable_offset;
3142 goto done;
3144 allocated_reftable_size = new_reftable_size;
3146 } while (new_allocation);
3148 /* Second, write the new refblocks */
3149 ret = walk_over_reftable(bs, &new_reftable, &new_reftable_index,
3150 &new_reftable_size, new_refblock,
3151 new_refblock_size, new_refcount_bits,
3152 &flush_refblock, &new_allocation, new_set_refcount,
3153 status_cb, cb_opaque, walk_index, walk_index + 1,
3154 errp);
3155 if (ret < 0) {
3156 goto done;
3158 assert(!new_allocation);
3161 /* Write the new reftable */
3162 ret = qcow2_pre_write_overlap_check(bs, 0, new_reftable_offset,
3163 new_reftable_size * sizeof(uint64_t),
3164 false);
3165 if (ret < 0) {
3166 error_setg_errno(errp, -ret, "Overlap check failed");
3167 goto done;
3170 for (i = 0; i < new_reftable_size; i++) {
3171 cpu_to_be64s(&new_reftable[i]);
3174 ret = bdrv_pwrite(bs->file, new_reftable_offset, new_reftable,
3175 new_reftable_size * sizeof(uint64_t));
3177 for (i = 0; i < new_reftable_size; i++) {
3178 be64_to_cpus(&new_reftable[i]);
3181 if (ret < 0) {
3182 error_setg_errno(errp, -ret, "Failed to write the new reftable");
3183 goto done;
3187 /* Empty the refcount cache */
3188 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
3189 if (ret < 0) {
3190 error_setg_errno(errp, -ret, "Failed to flush the refblock cache");
3191 goto done;
3194 /* Update the image header to point to the new reftable; this only updates
3195 * the fields which are relevant to qcow2_update_header(); other fields
3196 * such as s->refcount_table or s->refcount_bits stay stale for now
3197 * (because we have to restore everything if qcow2_update_header() fails) */
3198 old_refcount_order = s->refcount_order;
3199 old_reftable_size = s->refcount_table_size;
3200 old_reftable_offset = s->refcount_table_offset;
3202 s->refcount_order = refcount_order;
3203 s->refcount_table_size = new_reftable_size;
3204 s->refcount_table_offset = new_reftable_offset;
3206 ret = qcow2_update_header(bs);
3207 if (ret < 0) {
3208 s->refcount_order = old_refcount_order;
3209 s->refcount_table_size = old_reftable_size;
3210 s->refcount_table_offset = old_reftable_offset;
3211 error_setg_errno(errp, -ret, "Failed to update the qcow2 header");
3212 goto done;
3215 /* Now update the rest of the in-memory information */
3216 old_reftable = s->refcount_table;
3217 s->refcount_table = new_reftable;
3218 update_max_refcount_table_index(s);
3220 s->refcount_bits = 1 << refcount_order;
3221 s->refcount_max = UINT64_C(1) << (s->refcount_bits - 1);
3222 s->refcount_max += s->refcount_max - 1;
3224 s->refcount_block_bits = s->cluster_bits - (refcount_order - 3);
3225 s->refcount_block_size = 1 << s->refcount_block_bits;
3227 s->get_refcount = new_get_refcount;
3228 s->set_refcount = new_set_refcount;
3230 /* For cleaning up all old refblocks and the old reftable below the "done"
3231 * label */
3232 new_reftable = old_reftable;
3233 new_reftable_size = old_reftable_size;
3234 new_reftable_offset = old_reftable_offset;
3236 done:
3237 if (new_reftable) {
3238 /* On success, new_reftable actually points to the old reftable (and
3239 * new_reftable_size is the old reftable's size); but that is just
3240 * fine */
3241 for (i = 0; i < new_reftable_size; i++) {
3242 uint64_t offset = new_reftable[i] & REFT_OFFSET_MASK;
3243 if (offset) {
3244 qcow2_free_clusters(bs, offset, s->cluster_size,
3245 QCOW2_DISCARD_OTHER);
3248 g_free(new_reftable);
3250 if (new_reftable_offset > 0) {
3251 qcow2_free_clusters(bs, new_reftable_offset,
3252 new_reftable_size * sizeof(uint64_t),
3253 QCOW2_DISCARD_OTHER);
3257 qemu_vfree(new_refblock);
3258 return ret;
3261 static int64_t get_refblock_offset(BlockDriverState *bs, uint64_t offset)
3263 BDRVQcow2State *s = bs->opaque;
3264 uint32_t index = offset_to_reftable_index(s, offset);
3265 int64_t covering_refblock_offset = 0;
3267 if (index < s->refcount_table_size) {
3268 covering_refblock_offset = s->refcount_table[index] & REFT_OFFSET_MASK;
3270 if (!covering_refblock_offset) {
3271 qcow2_signal_corruption(bs, true, -1, -1, "Refblock at %#" PRIx64 " is "
3272 "not covered by the refcount structures",
3273 offset);
3274 return -EIO;
3277 return covering_refblock_offset;
3280 static int qcow2_discard_refcount_block(BlockDriverState *bs,
3281 uint64_t discard_block_offs)
3283 BDRVQcow2State *s = bs->opaque;
3284 int64_t refblock_offs;
3285 uint64_t cluster_index = discard_block_offs >> s->cluster_bits;
3286 uint32_t block_index = cluster_index & (s->refcount_block_size - 1);
3287 void *refblock;
3288 int ret;
3290 refblock_offs = get_refblock_offset(bs, discard_block_offs);
3291 if (refblock_offs < 0) {
3292 return refblock_offs;
3295 assert(discard_block_offs != 0);
3297 ret = qcow2_cache_get(bs, s->refcount_block_cache, refblock_offs,
3298 &refblock);
3299 if (ret < 0) {
3300 return ret;
3303 if (s->get_refcount(refblock, block_index) != 1) {
3304 qcow2_signal_corruption(bs, true, -1, -1, "Invalid refcount:"
3305 " refblock offset %#" PRIx64
3306 ", reftable index %u"
3307 ", block offset %#" PRIx64
3308 ", refcount %#" PRIx64,
3309 refblock_offs,
3310 offset_to_reftable_index(s, discard_block_offs),
3311 discard_block_offs,
3312 s->get_refcount(refblock, block_index));
3313 qcow2_cache_put(s->refcount_block_cache, &refblock);
3314 return -EINVAL;
3316 s->set_refcount(refblock, block_index, 0);
3318 qcow2_cache_entry_mark_dirty(s->refcount_block_cache, refblock);
3320 qcow2_cache_put(s->refcount_block_cache, &refblock);
3322 if (cluster_index < s->free_cluster_index) {
3323 s->free_cluster_index = cluster_index;
3326 refblock = qcow2_cache_is_table_offset(s->refcount_block_cache,
3327 discard_block_offs);
3328 if (refblock) {
3329 /* discard refblock from the cache if refblock is cached */
3330 qcow2_cache_discard(s->refcount_block_cache, refblock);
3332 update_refcount_discard(bs, discard_block_offs, s->cluster_size);
3334 return 0;
3337 int qcow2_shrink_reftable(BlockDriverState *bs)
3339 BDRVQcow2State *s = bs->opaque;
3340 uint64_t *reftable_tmp =
3341 g_malloc(s->refcount_table_size * sizeof(uint64_t));
3342 int i, ret;
3344 for (i = 0; i < s->refcount_table_size; i++) {
3345 int64_t refblock_offs = s->refcount_table[i] & REFT_OFFSET_MASK;
3346 void *refblock;
3347 bool unused_block;
3349 if (refblock_offs == 0) {
3350 reftable_tmp[i] = 0;
3351 continue;
3353 ret = qcow2_cache_get(bs, s->refcount_block_cache, refblock_offs,
3354 &refblock);
3355 if (ret < 0) {
3356 goto out;
3359 /* the refblock has own reference */
3360 if (i == offset_to_reftable_index(s, refblock_offs)) {
3361 uint64_t block_index = (refblock_offs >> s->cluster_bits) &
3362 (s->refcount_block_size - 1);
3363 uint64_t refcount = s->get_refcount(refblock, block_index);
3365 s->set_refcount(refblock, block_index, 0);
3367 unused_block = buffer_is_zero(refblock, s->cluster_size);
3369 s->set_refcount(refblock, block_index, refcount);
3370 } else {
3371 unused_block = buffer_is_zero(refblock, s->cluster_size);
3373 qcow2_cache_put(s->refcount_block_cache, &refblock);
3375 reftable_tmp[i] = unused_block ? 0 : cpu_to_be64(s->refcount_table[i]);
3378 ret = bdrv_pwrite_sync(bs->file, s->refcount_table_offset, reftable_tmp,
3379 s->refcount_table_size * sizeof(uint64_t));
3381 * If the write in the reftable failed the image may contain a partially
3382 * overwritten reftable. In this case it would be better to clear the
3383 * reftable in memory to avoid possible image corruption.
3385 for (i = 0; i < s->refcount_table_size; i++) {
3386 if (s->refcount_table[i] && !reftable_tmp[i]) {
3387 if (ret == 0) {
3388 ret = qcow2_discard_refcount_block(bs, s->refcount_table[i] &
3389 REFT_OFFSET_MASK);
3391 s->refcount_table[i] = 0;
3395 if (!s->cache_discards) {
3396 qcow2_process_discards(bs, ret);
3399 out:
3400 g_free(reftable_tmp);
3401 return ret;
3404 int64_t qcow2_get_last_cluster(BlockDriverState *bs, int64_t size)
3406 BDRVQcow2State *s = bs->opaque;
3407 int64_t i;
3409 for (i = size_to_clusters(s, size) - 1; i >= 0; i--) {
3410 uint64_t refcount;
3411 int ret = qcow2_get_refcount(bs, i, &refcount);
3412 if (ret < 0) {
3413 fprintf(stderr, "Can't get refcount for cluster %" PRId64 ": %s\n",
3414 i, strerror(-ret));
3415 return ret;
3417 if (refcount > 0) {
3418 return i;
3421 qcow2_signal_corruption(bs, true, -1, -1,
3422 "There are no references in the refcount table.");
3423 return -EIO;