xref: /openbmc/qemu/block/qcow2.h (revision 4c25f365) 
1 /*
2  * Block driver for the QCOW version 2 format
3  *
4  * Copyright (c) 2004-2006 Fabrice Bellard
5  *
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:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
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.
23  */
24 
25 #ifndef BLOCK_QCOW2_H
26 #define BLOCK_QCOW2_H
27 
28 #include "qemu/aes.h"
29 #include "block/coroutine.h"
30 
31 //#define DEBUG_ALLOC
32 //#define DEBUG_ALLOC2
33 //#define DEBUG_EXT
34 
35 #define QCOW_MAGIC (('Q' << 24) | ('F' << 16) | ('I' << 8) | 0xfb)
36 
37 #define QCOW_CRYPT_NONE 0
38 #define QCOW_CRYPT_AES  1
39 
40 #define QCOW_MAX_CRYPT_CLUSTERS 32
41 
42 /* indicate that the refcount of the referenced cluster is exactly one. */
43 #define QCOW_OFLAG_COPIED     (1ULL << 63)
44 /* indicate that the cluster is compressed (they never have the copied flag) */
45 #define QCOW_OFLAG_COMPRESSED (1ULL << 62)
46 /* The cluster reads as all zeros */
47 #define QCOW_OFLAG_ZERO (1ULL << 0)
48 
49 #define REFCOUNT_SHIFT 1 /* refcount size is 2 bytes */
50 
51 #define MIN_CLUSTER_BITS 9
52 #define MAX_CLUSTER_BITS 21
53 
54 #define L2_CACHE_SIZE 16
55 
56 /* Must be at least 4 to cover all cases of refcount table growth */
57 #define REFCOUNT_CACHE_SIZE 4
58 
59 #define DEFAULT_CLUSTER_SIZE 65536
60 
61 
62 #define QCOW2_OPT_LAZY_REFCOUNTS "lazy-refcounts"
63 #define QCOW2_OPT_DISCARD_REQUEST "pass-discard-request"
64 #define QCOW2_OPT_DISCARD_SNAPSHOT "pass-discard-snapshot"
65 #define QCOW2_OPT_DISCARD_OTHER "pass-discard-other"
66 #define QCOW2_OPT_OVERLAP "overlap-check"
67 #define QCOW2_OPT_OVERLAP_MAIN_HEADER "overlap-check.main-header"
68 #define QCOW2_OPT_OVERLAP_ACTIVE_L1 "overlap-check.active-l1"
69 #define QCOW2_OPT_OVERLAP_ACTIVE_L2 "overlap-check.active-l2"
70 #define QCOW2_OPT_OVERLAP_REFCOUNT_TABLE "overlap-check.refcount-table"
71 #define QCOW2_OPT_OVERLAP_REFCOUNT_BLOCK "overlap-check.refcount-block"
72 #define QCOW2_OPT_OVERLAP_SNAPSHOT_TABLE "overlap-check.snapshot-table"
73 #define QCOW2_OPT_OVERLAP_INACTIVE_L1 "overlap-check.inactive-l1"
74 #define QCOW2_OPT_OVERLAP_INACTIVE_L2 "overlap-check.inactive-l2"
75 
76 typedef struct QCowHeader {
77     uint32_t magic;
78     uint32_t version;
79     uint64_t backing_file_offset;
80     uint32_t backing_file_size;
81     uint32_t cluster_bits;
82     uint64_t size; /* in bytes */
83     uint32_t crypt_method;
84     uint32_t l1_size; /* XXX: save number of clusters instead ? */
85     uint64_t l1_table_offset;
86     uint64_t refcount_table_offset;
87     uint32_t refcount_table_clusters;
88     uint32_t nb_snapshots;
89     uint64_t snapshots_offset;
90 
91     /* The following fields are only valid for version >= 3 */
92     uint64_t incompatible_features;
93     uint64_t compatible_features;
94     uint64_t autoclear_features;
95 
96     uint32_t refcount_order;
97     uint32_t header_length;
98 } QEMU_PACKED QCowHeader;
99 
100 typedef struct QCowSnapshot {
101     uint64_t l1_table_offset;
102     uint32_t l1_size;
103     char *id_str;
104     char *name;
105     uint64_t disk_size;
106     uint64_t vm_state_size;
107     uint32_t date_sec;
108     uint32_t date_nsec;
109     uint64_t vm_clock_nsec;
110 } QCowSnapshot;
111 
112 struct Qcow2Cache;
113 typedef struct Qcow2Cache Qcow2Cache;
114 
115 typedef struct Qcow2UnknownHeaderExtension {
116     uint32_t magic;
117     uint32_t len;
118     QLIST_ENTRY(Qcow2UnknownHeaderExtension) next;
119     uint8_t data[];
120 } Qcow2UnknownHeaderExtension;
121 
122 enum {
123     QCOW2_FEAT_TYPE_INCOMPATIBLE    = 0,
124     QCOW2_FEAT_TYPE_COMPATIBLE      = 1,
125     QCOW2_FEAT_TYPE_AUTOCLEAR       = 2,
126 };
127 
128 /* Incompatible feature bits */
129 enum {
130     QCOW2_INCOMPAT_DIRTY_BITNR   = 0,
131     QCOW2_INCOMPAT_CORRUPT_BITNR = 1,
132     QCOW2_INCOMPAT_DIRTY         = 1 << QCOW2_INCOMPAT_DIRTY_BITNR,
133     QCOW2_INCOMPAT_CORRUPT       = 1 << QCOW2_INCOMPAT_CORRUPT_BITNR,
134 
135     QCOW2_INCOMPAT_MASK          = QCOW2_INCOMPAT_DIRTY
136                                  | QCOW2_INCOMPAT_CORRUPT,
137 };
138 
139 /* Compatible feature bits */
140 enum {
141     QCOW2_COMPAT_LAZY_REFCOUNTS_BITNR = 0,
142     QCOW2_COMPAT_LAZY_REFCOUNTS       = 1 << QCOW2_COMPAT_LAZY_REFCOUNTS_BITNR,
143 
144     QCOW2_COMPAT_FEAT_MASK            = QCOW2_COMPAT_LAZY_REFCOUNTS,
145 };
146 
147 enum qcow2_discard_type {
148     QCOW2_DISCARD_NEVER = 0,
149     QCOW2_DISCARD_ALWAYS,
150     QCOW2_DISCARD_REQUEST,
151     QCOW2_DISCARD_SNAPSHOT,
152     QCOW2_DISCARD_OTHER,
153     QCOW2_DISCARD_MAX
154 };
155 
156 typedef struct Qcow2Feature {
157     uint8_t type;
158     uint8_t bit;
159     char    name[46];
160 } QEMU_PACKED Qcow2Feature;
161 
162 typedef struct Qcow2DiscardRegion {
163     BlockDriverState *bs;
164     uint64_t offset;
165     uint64_t bytes;
166     QTAILQ_ENTRY(Qcow2DiscardRegion) next;
167 } Qcow2DiscardRegion;
168 
169 typedef struct BDRVQcowState {
170     int cluster_bits;
171     int cluster_size;
172     int cluster_sectors;
173     int l2_bits;
174     int l2_size;
175     int l1_size;
176     int l1_vm_state_index;
177     int csize_shift;
178     int csize_mask;
179     uint64_t cluster_offset_mask;
180     uint64_t l1_table_offset;
181     uint64_t *l1_table;
182 
183     Qcow2Cache* l2_table_cache;
184     Qcow2Cache* refcount_block_cache;
185 
186     uint8_t *cluster_cache;
187     uint8_t *cluster_data;
188     uint64_t cluster_cache_offset;
189     QLIST_HEAD(QCowClusterAlloc, QCowL2Meta) cluster_allocs;
190 
191     uint64_t *refcount_table;
192     uint64_t refcount_table_offset;
193     uint32_t refcount_table_size;
194     int64_t free_cluster_index;
195     int64_t free_byte_offset;
196 
197     CoMutex lock;
198 
199     uint32_t crypt_method; /* current crypt method, 0 if no key yet */
200     uint32_t crypt_method_header;
201     AES_KEY aes_encrypt_key;
202     AES_KEY aes_decrypt_key;
203     uint64_t snapshots_offset;
204     int snapshots_size;
205     int nb_snapshots;
206     QCowSnapshot *snapshots;
207 
208     int flags;
209     int qcow_version;
210     bool use_lazy_refcounts;
211     int refcount_order;
212 
213     bool discard_passthrough[QCOW2_DISCARD_MAX];
214 
215     int overlap_check; /* bitmask of Qcow2MetadataOverlap values */
216 
217     uint64_t incompatible_features;
218     uint64_t compatible_features;
219     uint64_t autoclear_features;
220 
221     size_t unknown_header_fields_size;
222     void* unknown_header_fields;
223     QLIST_HEAD(, Qcow2UnknownHeaderExtension) unknown_header_ext;
224     QTAILQ_HEAD (, Qcow2DiscardRegion) discards;
225     bool cache_discards;
226 } BDRVQcowState;
227 
228 /* XXX: use std qcow open function ? */
229 typedef struct QCowCreateState {
230     int cluster_size;
231     int cluster_bits;
232     uint16_t *refcount_block;
233     uint64_t *refcount_table;
234     int64_t l1_table_offset;
235     int64_t refcount_table_offset;
236     int64_t refcount_block_offset;
237 } QCowCreateState;
238 
239 struct QCowAIOCB;
240 
241 typedef struct Qcow2COWRegion {
242     /**
243      * Offset of the COW region in bytes from the start of the first cluster
244      * touched by the request.
245      */
246     uint64_t    offset;
247 
248     /** Number of sectors to copy */
249     int         nb_sectors;
250 } Qcow2COWRegion;
251 
252 /**
253  * Describes an in-flight (part of a) write request that writes to clusters
254  * that are not referenced in their L2 table yet.
255  */
256 typedef struct QCowL2Meta
257 {
258     /** Guest offset of the first newly allocated cluster */
259     uint64_t offset;
260 
261     /** Host offset of the first newly allocated cluster */
262     uint64_t alloc_offset;
263 
264     /**
265      * Number of sectors from the start of the first allocated cluster to
266      * the end of the (possibly shortened) request
267      */
268     int nb_available;
269 
270     /** Number of newly allocated clusters */
271     int nb_clusters;
272 
273     /**
274      * Requests that overlap with this allocation and wait to be restarted
275      * when the allocating request has completed.
276      */
277     CoQueue dependent_requests;
278 
279     /**
280      * The COW Region between the start of the first allocated cluster and the
281      * area the guest actually writes to.
282      */
283     Qcow2COWRegion cow_start;
284 
285     /**
286      * The COW Region between the area the guest actually writes to and the
287      * end of the last allocated cluster.
288      */
289     Qcow2COWRegion cow_end;
290 
291     /** Pointer to next L2Meta of the same write request */
292     struct QCowL2Meta *next;
293 
294     QLIST_ENTRY(QCowL2Meta) next_in_flight;
295 } QCowL2Meta;
296 
297 enum {
298     QCOW2_CLUSTER_UNALLOCATED,
299     QCOW2_CLUSTER_NORMAL,
300     QCOW2_CLUSTER_COMPRESSED,
301     QCOW2_CLUSTER_ZERO
302 };
303 
304 typedef enum QCow2MetadataOverlap {
305     QCOW2_OL_MAIN_HEADER_BITNR    = 0,
306     QCOW2_OL_ACTIVE_L1_BITNR      = 1,
307     QCOW2_OL_ACTIVE_L2_BITNR      = 2,
308     QCOW2_OL_REFCOUNT_TABLE_BITNR = 3,
309     QCOW2_OL_REFCOUNT_BLOCK_BITNR = 4,
310     QCOW2_OL_SNAPSHOT_TABLE_BITNR = 5,
311     QCOW2_OL_INACTIVE_L1_BITNR    = 6,
312     QCOW2_OL_INACTIVE_L2_BITNR    = 7,
313 
314     QCOW2_OL_MAX_BITNR            = 8,
315 
316     QCOW2_OL_NONE           = 0,
317     QCOW2_OL_MAIN_HEADER    = (1 << QCOW2_OL_MAIN_HEADER_BITNR),
318     QCOW2_OL_ACTIVE_L1      = (1 << QCOW2_OL_ACTIVE_L1_BITNR),
319     QCOW2_OL_ACTIVE_L2      = (1 << QCOW2_OL_ACTIVE_L2_BITNR),
320     QCOW2_OL_REFCOUNT_TABLE = (1 << QCOW2_OL_REFCOUNT_TABLE_BITNR),
321     QCOW2_OL_REFCOUNT_BLOCK = (1 << QCOW2_OL_REFCOUNT_BLOCK_BITNR),
322     QCOW2_OL_SNAPSHOT_TABLE = (1 << QCOW2_OL_SNAPSHOT_TABLE_BITNR),
323     QCOW2_OL_INACTIVE_L1    = (1 << QCOW2_OL_INACTIVE_L1_BITNR),
324     /* NOTE: Checking overlaps with inactive L2 tables will result in bdrv
325      * reads. */
326     QCOW2_OL_INACTIVE_L2    = (1 << QCOW2_OL_INACTIVE_L2_BITNR),
327 } QCow2MetadataOverlap;
328 
329 /* Perform all overlap checks which can be done in constant time */
330 #define QCOW2_OL_CONSTANT \
331     (QCOW2_OL_MAIN_HEADER | QCOW2_OL_ACTIVE_L1 | QCOW2_OL_REFCOUNT_TABLE | \
332      QCOW2_OL_SNAPSHOT_TABLE)
333 
334 /* Perform all overlap checks which don't require disk access */
335 #define QCOW2_OL_CACHED \
336     (QCOW2_OL_CONSTANT | QCOW2_OL_ACTIVE_L2 | QCOW2_OL_REFCOUNT_BLOCK | \
337      QCOW2_OL_INACTIVE_L1)
338 
339 /* Perform all overlap checks */
340 #define QCOW2_OL_ALL \
341     (QCOW2_OL_CACHED | QCOW2_OL_INACTIVE_L2)
342 
343 #define L1E_OFFSET_MASK 0x00ffffffffffff00ULL
344 #define L2E_OFFSET_MASK 0x00ffffffffffff00ULL
345 #define L2E_COMPRESSED_OFFSET_SIZE_MASK 0x3fffffffffffffffULL
346 
347 #define REFT_OFFSET_MASK 0xffffffffffffff00ULL
348 
349 static inline int64_t start_of_cluster(BDRVQcowState *s, int64_t offset)
350 {
351     return offset & ~(s->cluster_size - 1);
352 }
353 
354 static inline int64_t offset_into_cluster(BDRVQcowState *s, int64_t offset)
355 {
356     return offset & (s->cluster_size - 1);
357 }
358 
359 static inline int size_to_clusters(BDRVQcowState *s, int64_t size)
360 {
361     return (size + (s->cluster_size - 1)) >> s->cluster_bits;
362 }
363 
364 static inline int64_t size_to_l1(BDRVQcowState *s, int64_t size)
365 {
366     int shift = s->cluster_bits + s->l2_bits;
367     return (size + (1ULL << shift) - 1) >> shift;
368 }
369 
370 static inline int offset_to_l2_index(BDRVQcowState *s, int64_t offset)
371 {
372     return (offset >> s->cluster_bits) & (s->l2_size - 1);
373 }
374 
375 static inline int64_t align_offset(int64_t offset, int n)
376 {
377     offset = (offset + n - 1) & ~(n - 1);
378     return offset;
379 }
380 
381 static inline int64_t qcow2_vm_state_offset(BDRVQcowState *s)
382 {
383     return (int64_t)s->l1_vm_state_index << (s->cluster_bits + s->l2_bits);
384 }
385 
386 static inline int qcow2_get_cluster_type(uint64_t l2_entry)
387 {
388     if (l2_entry & QCOW_OFLAG_COMPRESSED) {
389         return QCOW2_CLUSTER_COMPRESSED;
390     } else if (l2_entry & QCOW_OFLAG_ZERO) {
391         return QCOW2_CLUSTER_ZERO;
392     } else if (!(l2_entry & L2E_OFFSET_MASK)) {
393         return QCOW2_CLUSTER_UNALLOCATED;
394     } else {
395         return QCOW2_CLUSTER_NORMAL;
396     }
397 }
398 
399 /* Check whether refcounts are eager or lazy */
400 static inline bool qcow2_need_accurate_refcounts(BDRVQcowState *s)
401 {
402     return !(s->incompatible_features & QCOW2_INCOMPAT_DIRTY);
403 }
404 
405 static inline uint64_t l2meta_cow_start(QCowL2Meta *m)
406 {
407     return m->offset + m->cow_start.offset;
408 }
409 
410 static inline uint64_t l2meta_cow_end(QCowL2Meta *m)
411 {
412     return m->offset + m->cow_end.offset
413         + (m->cow_end.nb_sectors << BDRV_SECTOR_BITS);
414 }
415 
416 // FIXME Need qcow2_ prefix to global functions
417 
418 /* qcow2.c functions */
419 int qcow2_backing_read1(BlockDriverState *bs, QEMUIOVector *qiov,
420                   int64_t sector_num, int nb_sectors);
421 
422 int qcow2_mark_dirty(BlockDriverState *bs);
423 int qcow2_mark_corrupt(BlockDriverState *bs);
424 int qcow2_mark_consistent(BlockDriverState *bs);
425 int qcow2_update_header(BlockDriverState *bs);
426 
427 /* qcow2-refcount.c functions */
428 int qcow2_refcount_init(BlockDriverState *bs);
429 void qcow2_refcount_close(BlockDriverState *bs);
430 
431 int qcow2_update_cluster_refcount(BlockDriverState *bs, int64_t cluster_index,
432                                   int addend, enum qcow2_discard_type type);
433 
434 int64_t qcow2_alloc_clusters(BlockDriverState *bs, int64_t size);
435 int qcow2_alloc_clusters_at(BlockDriverState *bs, uint64_t offset,
436     int nb_clusters);
437 int64_t qcow2_alloc_bytes(BlockDriverState *bs, int size);
438 void qcow2_free_clusters(BlockDriverState *bs,
439                           int64_t offset, int64_t size,
440                           enum qcow2_discard_type type);
441 void qcow2_free_any_clusters(BlockDriverState *bs, uint64_t l2_entry,
442                              int nb_clusters, enum qcow2_discard_type type);
443 
444 int qcow2_update_snapshot_refcount(BlockDriverState *bs,
445     int64_t l1_table_offset, int l1_size, int addend);
446 
447 int qcow2_check_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
448                           BdrvCheckMode fix);
449 
450 void qcow2_process_discards(BlockDriverState *bs, int ret);
451 
452 int qcow2_check_metadata_overlap(BlockDriverState *bs, int ign, int64_t offset,
453                                  int64_t size);
454 int qcow2_pre_write_overlap_check(BlockDriverState *bs, int ign, int64_t offset,
455                                   int64_t size);
456 
457 /* qcow2-cluster.c functions */
458 int qcow2_grow_l1_table(BlockDriverState *bs, uint64_t min_size,
459                         bool exact_size);
460 int qcow2_write_l1_entry(BlockDriverState *bs, int l1_index);
461 void qcow2_l2_cache_reset(BlockDriverState *bs);
462 int qcow2_decompress_cluster(BlockDriverState *bs, uint64_t cluster_offset);
463 void qcow2_encrypt_sectors(BDRVQcowState *s, int64_t sector_num,
464                      uint8_t *out_buf, const uint8_t *in_buf,
465                      int nb_sectors, int enc,
466                      const AES_KEY *key);
467 
468 int qcow2_get_cluster_offset(BlockDriverState *bs, uint64_t offset,
469     int *num, uint64_t *cluster_offset);
470 int qcow2_alloc_cluster_offset(BlockDriverState *bs, uint64_t offset,
471     int n_start, int n_end, int *num, uint64_t *host_offset, QCowL2Meta **m);
472 uint64_t qcow2_alloc_compressed_cluster_offset(BlockDriverState *bs,
473                                          uint64_t offset,
474                                          int compressed_size);
475 
476 int qcow2_alloc_cluster_link_l2(BlockDriverState *bs, QCowL2Meta *m);
477 int qcow2_discard_clusters(BlockDriverState *bs, uint64_t offset,
478     int nb_sectors, enum qcow2_discard_type type);
479 int qcow2_zero_clusters(BlockDriverState *bs, uint64_t offset, int nb_sectors);
480 
481 int qcow2_expand_zero_clusters(BlockDriverState *bs);
482 
483 /* qcow2-snapshot.c functions */
484 int qcow2_snapshot_create(BlockDriverState *bs, QEMUSnapshotInfo *sn_info);
485 int qcow2_snapshot_goto(BlockDriverState *bs, const char *snapshot_id);
486 int qcow2_snapshot_delete(BlockDriverState *bs,
487                           const char *snapshot_id,
488                           const char *name,
489                           Error **errp);
490 int qcow2_snapshot_list(BlockDriverState *bs, QEMUSnapshotInfo **psn_tab);
491 int qcow2_snapshot_load_tmp(BlockDriverState *bs,
492                             const char *snapshot_id,
493                             const char *name,
494                             Error **errp);
495 
496 void qcow2_free_snapshots(BlockDriverState *bs);
497 int qcow2_read_snapshots(BlockDriverState *bs);
498 
499 /* qcow2-cache.c functions */
500 Qcow2Cache *qcow2_cache_create(BlockDriverState *bs, int num_tables);
501 int qcow2_cache_destroy(BlockDriverState* bs, Qcow2Cache *c);
502 
503 void qcow2_cache_entry_mark_dirty(Qcow2Cache *c, void *table);
504 int qcow2_cache_flush(BlockDriverState *bs, Qcow2Cache *c);
505 int qcow2_cache_set_dependency(BlockDriverState *bs, Qcow2Cache *c,
506     Qcow2Cache *dependency);
507 void qcow2_cache_depends_on_flush(Qcow2Cache *c);
508 
509 int qcow2_cache_empty(BlockDriverState *bs, Qcow2Cache *c);
510 
511 int qcow2_cache_get(BlockDriverState *bs, Qcow2Cache *c, uint64_t offset,
512     void **table);
513 int qcow2_cache_get_empty(BlockDriverState *bs, Qcow2Cache *c, uint64_t offset,
514     void **table);
515 int qcow2_cache_put(BlockDriverState *bs, Qcow2Cache *c, void **table);
516 
517 #endif
518