xref: /openbmc/qemu/block/qcow2.h (revision 5accc840) 
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 
67 typedef struct QCowHeader {
68     uint32_t magic;
69     uint32_t version;
70     uint64_t backing_file_offset;
71     uint32_t backing_file_size;
72     uint32_t cluster_bits;
73     uint64_t size; /* in bytes */
74     uint32_t crypt_method;
75     uint32_t l1_size; /* XXX: save number of clusters instead ? */
76     uint64_t l1_table_offset;
77     uint64_t refcount_table_offset;
78     uint32_t refcount_table_clusters;
79     uint32_t nb_snapshots;
80     uint64_t snapshots_offset;
81 
82     /* The following fields are only valid for version >= 3 */
83     uint64_t incompatible_features;
84     uint64_t compatible_features;
85     uint64_t autoclear_features;
86 
87     uint32_t refcount_order;
88     uint32_t header_length;
89 } QCowHeader;
90 
91 typedef struct QCowSnapshot {
92     uint64_t l1_table_offset;
93     uint32_t l1_size;
94     char *id_str;
95     char *name;
96     uint64_t disk_size;
97     uint64_t vm_state_size;
98     uint32_t date_sec;
99     uint32_t date_nsec;
100     uint64_t vm_clock_nsec;
101 } QCowSnapshot;
102 
103 struct Qcow2Cache;
104 typedef struct Qcow2Cache Qcow2Cache;
105 
106 typedef struct Qcow2UnknownHeaderExtension {
107     uint32_t magic;
108     uint32_t len;
109     QLIST_ENTRY(Qcow2UnknownHeaderExtension) next;
110     uint8_t data[];
111 } Qcow2UnknownHeaderExtension;
112 
113 enum {
114     QCOW2_FEAT_TYPE_INCOMPATIBLE    = 0,
115     QCOW2_FEAT_TYPE_COMPATIBLE      = 1,
116     QCOW2_FEAT_TYPE_AUTOCLEAR       = 2,
117 };
118 
119 /* Incompatible feature bits */
120 enum {
121     QCOW2_INCOMPAT_DIRTY_BITNR   = 0,
122     QCOW2_INCOMPAT_CORRUPT_BITNR = 1,
123     QCOW2_INCOMPAT_DIRTY         = 1 << QCOW2_INCOMPAT_DIRTY_BITNR,
124     QCOW2_INCOMPAT_CORRUPT       = 1 << QCOW2_INCOMPAT_CORRUPT_BITNR,
125 
126     QCOW2_INCOMPAT_MASK          = QCOW2_INCOMPAT_DIRTY
127                                  | QCOW2_INCOMPAT_CORRUPT,
128 };
129 
130 /* Compatible feature bits */
131 enum {
132     QCOW2_COMPAT_LAZY_REFCOUNTS_BITNR = 0,
133     QCOW2_COMPAT_LAZY_REFCOUNTS       = 1 << QCOW2_COMPAT_LAZY_REFCOUNTS_BITNR,
134 
135     QCOW2_COMPAT_FEAT_MASK            = QCOW2_COMPAT_LAZY_REFCOUNTS,
136 };
137 
138 enum qcow2_discard_type {
139     QCOW2_DISCARD_NEVER = 0,
140     QCOW2_DISCARD_ALWAYS,
141     QCOW2_DISCARD_REQUEST,
142     QCOW2_DISCARD_SNAPSHOT,
143     QCOW2_DISCARD_OTHER,
144     QCOW2_DISCARD_MAX
145 };
146 
147 typedef struct Qcow2Feature {
148     uint8_t type;
149     uint8_t bit;
150     char    name[46];
151 } QEMU_PACKED Qcow2Feature;
152 
153 typedef struct Qcow2DiscardRegion {
154     BlockDriverState *bs;
155     uint64_t offset;
156     uint64_t bytes;
157     QTAILQ_ENTRY(Qcow2DiscardRegion) next;
158 } Qcow2DiscardRegion;
159 
160 typedef struct BDRVQcowState {
161     int cluster_bits;
162     int cluster_size;
163     int cluster_sectors;
164     int l2_bits;
165     int l2_size;
166     int l1_size;
167     int l1_vm_state_index;
168     int csize_shift;
169     int csize_mask;
170     uint64_t cluster_offset_mask;
171     uint64_t l1_table_offset;
172     uint64_t *l1_table;
173 
174     Qcow2Cache* l2_table_cache;
175     Qcow2Cache* refcount_block_cache;
176 
177     uint8_t *cluster_cache;
178     uint8_t *cluster_data;
179     uint64_t cluster_cache_offset;
180     QLIST_HEAD(QCowClusterAlloc, QCowL2Meta) cluster_allocs;
181 
182     uint64_t *refcount_table;
183     uint64_t refcount_table_offset;
184     uint32_t refcount_table_size;
185     int64_t free_cluster_index;
186     int64_t free_byte_offset;
187 
188     CoMutex lock;
189 
190     uint32_t crypt_method; /* current crypt method, 0 if no key yet */
191     uint32_t crypt_method_header;
192     AES_KEY aes_encrypt_key;
193     AES_KEY aes_decrypt_key;
194     uint64_t snapshots_offset;
195     int snapshots_size;
196     int nb_snapshots;
197     QCowSnapshot *snapshots;
198 
199     int flags;
200     int qcow_version;
201     bool use_lazy_refcounts;
202 
203     bool discard_passthrough[QCOW2_DISCARD_MAX];
204 
205     uint64_t incompatible_features;
206     uint64_t compatible_features;
207     uint64_t autoclear_features;
208 
209     size_t unknown_header_fields_size;
210     void* unknown_header_fields;
211     QLIST_HEAD(, Qcow2UnknownHeaderExtension) unknown_header_ext;
212     QTAILQ_HEAD (, Qcow2DiscardRegion) discards;
213     bool cache_discards;
214 } BDRVQcowState;
215 
216 /* XXX: use std qcow open function ? */
217 typedef struct QCowCreateState {
218     int cluster_size;
219     int cluster_bits;
220     uint16_t *refcount_block;
221     uint64_t *refcount_table;
222     int64_t l1_table_offset;
223     int64_t refcount_table_offset;
224     int64_t refcount_block_offset;
225 } QCowCreateState;
226 
227 struct QCowAIOCB;
228 
229 typedef struct Qcow2COWRegion {
230     /**
231      * Offset of the COW region in bytes from the start of the first cluster
232      * touched by the request.
233      */
234     uint64_t    offset;
235 
236     /** Number of sectors to copy */
237     int         nb_sectors;
238 } Qcow2COWRegion;
239 
240 /**
241  * Describes an in-flight (part of a) write request that writes to clusters
242  * that are not referenced in their L2 table yet.
243  */
244 typedef struct QCowL2Meta
245 {
246     /** Guest offset of the first newly allocated cluster */
247     uint64_t offset;
248 
249     /** Host offset of the first newly allocated cluster */
250     uint64_t alloc_offset;
251 
252     /**
253      * Number of sectors from the start of the first allocated cluster to
254      * the end of the (possibly shortened) request
255      */
256     int nb_available;
257 
258     /** Number of newly allocated clusters */
259     int nb_clusters;
260 
261     /**
262      * Requests that overlap with this allocation and wait to be restarted
263      * when the allocating request has completed.
264      */
265     CoQueue dependent_requests;
266 
267     /**
268      * The COW Region between the start of the first allocated cluster and the
269      * area the guest actually writes to.
270      */
271     Qcow2COWRegion cow_start;
272 
273     /**
274      * The COW Region between the area the guest actually writes to and the
275      * end of the last allocated cluster.
276      */
277     Qcow2COWRegion cow_end;
278 
279     /** Pointer to next L2Meta of the same write request */
280     struct QCowL2Meta *next;
281 
282     QLIST_ENTRY(QCowL2Meta) next_in_flight;
283 } QCowL2Meta;
284 
285 enum {
286     QCOW2_CLUSTER_UNALLOCATED,
287     QCOW2_CLUSTER_NORMAL,
288     QCOW2_CLUSTER_COMPRESSED,
289     QCOW2_CLUSTER_ZERO
290 };
291 
292 typedef enum QCow2MetadataOverlap {
293     QCOW2_OL_MAIN_HEADER_BITNR    = 0,
294     QCOW2_OL_ACTIVE_L1_BITNR      = 1,
295     QCOW2_OL_ACTIVE_L2_BITNR      = 2,
296     QCOW2_OL_REFCOUNT_TABLE_BITNR = 3,
297     QCOW2_OL_REFCOUNT_BLOCK_BITNR = 4,
298     QCOW2_OL_SNAPSHOT_TABLE_BITNR = 5,
299     QCOW2_OL_INACTIVE_L1_BITNR    = 6,
300     QCOW2_OL_INACTIVE_L2_BITNR    = 7,
301 
302     QCOW2_OL_MAX_BITNR            = 8,
303 
304     QCOW2_OL_NONE           = 0,
305     QCOW2_OL_MAIN_HEADER    = (1 << QCOW2_OL_MAIN_HEADER_BITNR),
306     QCOW2_OL_ACTIVE_L1      = (1 << QCOW2_OL_ACTIVE_L1_BITNR),
307     QCOW2_OL_ACTIVE_L2      = (1 << QCOW2_OL_ACTIVE_L2_BITNR),
308     QCOW2_OL_REFCOUNT_TABLE = (1 << QCOW2_OL_REFCOUNT_TABLE_BITNR),
309     QCOW2_OL_REFCOUNT_BLOCK = (1 << QCOW2_OL_REFCOUNT_BLOCK_BITNR),
310     QCOW2_OL_SNAPSHOT_TABLE = (1 << QCOW2_OL_SNAPSHOT_TABLE_BITNR),
311     QCOW2_OL_INACTIVE_L1    = (1 << QCOW2_OL_INACTIVE_L1_BITNR),
312     /* NOTE: Checking overlaps with inactive L2 tables will result in bdrv
313      * reads. */
314     QCOW2_OL_INACTIVE_L2    = (1 << QCOW2_OL_INACTIVE_L2_BITNR),
315 } QCow2MetadataOverlap;
316 
317 /* Perform all overlap checks which don't require disk access */
318 #define QCOW2_OL_CACHED \
319     (QCOW2_OL_MAIN_HEADER | QCOW2_OL_ACTIVE_L1 | QCOW2_OL_ACTIVE_L2 | \
320      QCOW2_OL_REFCOUNT_TABLE | QCOW2_OL_REFCOUNT_BLOCK | \
321      QCOW2_OL_SNAPSHOT_TABLE | QCOW2_OL_INACTIVE_L1)
322 
323 /* The default checks to perform */
324 #define QCOW2_OL_DEFAULT QCOW2_OL_CACHED
325 
326 #define L1E_OFFSET_MASK 0x00ffffffffffff00ULL
327 #define L2E_OFFSET_MASK 0x00ffffffffffff00ULL
328 #define L2E_COMPRESSED_OFFSET_SIZE_MASK 0x3fffffffffffffffULL
329 
330 #define REFT_OFFSET_MASK 0xffffffffffffff00ULL
331 
332 static inline int64_t start_of_cluster(BDRVQcowState *s, int64_t offset)
333 {
334     return offset & ~(s->cluster_size - 1);
335 }
336 
337 static inline int64_t offset_into_cluster(BDRVQcowState *s, int64_t offset)
338 {
339     return offset & (s->cluster_size - 1);
340 }
341 
342 static inline int size_to_clusters(BDRVQcowState *s, int64_t size)
343 {
344     return (size + (s->cluster_size - 1)) >> s->cluster_bits;
345 }
346 
347 static inline int64_t size_to_l1(BDRVQcowState *s, int64_t size)
348 {
349     int shift = s->cluster_bits + s->l2_bits;
350     return (size + (1ULL << shift) - 1) >> shift;
351 }
352 
353 static inline int offset_to_l2_index(BDRVQcowState *s, int64_t offset)
354 {
355     return (offset >> s->cluster_bits) & (s->l2_size - 1);
356 }
357 
358 static inline int64_t align_offset(int64_t offset, int n)
359 {
360     offset = (offset + n - 1) & ~(n - 1);
361     return offset;
362 }
363 
364 static inline int qcow2_get_cluster_type(uint64_t l2_entry)
365 {
366     if (l2_entry & QCOW_OFLAG_COMPRESSED) {
367         return QCOW2_CLUSTER_COMPRESSED;
368     } else if (l2_entry & QCOW_OFLAG_ZERO) {
369         return QCOW2_CLUSTER_ZERO;
370     } else if (!(l2_entry & L2E_OFFSET_MASK)) {
371         return QCOW2_CLUSTER_UNALLOCATED;
372     } else {
373         return QCOW2_CLUSTER_NORMAL;
374     }
375 }
376 
377 /* Check whether refcounts are eager or lazy */
378 static inline bool qcow2_need_accurate_refcounts(BDRVQcowState *s)
379 {
380     return !(s->incompatible_features & QCOW2_INCOMPAT_DIRTY);
381 }
382 
383 static inline uint64_t l2meta_cow_start(QCowL2Meta *m)
384 {
385     return m->offset + m->cow_start.offset;
386 }
387 
388 static inline uint64_t l2meta_cow_end(QCowL2Meta *m)
389 {
390     return m->offset + m->cow_end.offset
391         + (m->cow_end.nb_sectors << BDRV_SECTOR_BITS);
392 }
393 
394 // FIXME Need qcow2_ prefix to global functions
395 
396 /* qcow2.c functions */
397 int qcow2_backing_read1(BlockDriverState *bs, QEMUIOVector *qiov,
398                   int64_t sector_num, int nb_sectors);
399 
400 int qcow2_mark_dirty(BlockDriverState *bs);
401 int qcow2_mark_corrupt(BlockDriverState *bs);
402 int qcow2_mark_consistent(BlockDriverState *bs);
403 int qcow2_update_header(BlockDriverState *bs);
404 
405 /* qcow2-refcount.c functions */
406 int qcow2_refcount_init(BlockDriverState *bs);
407 void qcow2_refcount_close(BlockDriverState *bs);
408 
409 int64_t qcow2_alloc_clusters(BlockDriverState *bs, int64_t size);
410 int qcow2_alloc_clusters_at(BlockDriverState *bs, uint64_t offset,
411     int nb_clusters);
412 int64_t qcow2_alloc_bytes(BlockDriverState *bs, int size);
413 void qcow2_free_clusters(BlockDriverState *bs,
414                           int64_t offset, int64_t size,
415                           enum qcow2_discard_type type);
416 void qcow2_free_any_clusters(BlockDriverState *bs, uint64_t l2_entry,
417                              int nb_clusters, enum qcow2_discard_type type);
418 
419 int qcow2_update_snapshot_refcount(BlockDriverState *bs,
420     int64_t l1_table_offset, int l1_size, int addend);
421 
422 int qcow2_check_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
423                           BdrvCheckMode fix);
424 
425 void qcow2_process_discards(BlockDriverState *bs, int ret);
426 
427 int qcow2_check_metadata_overlap(BlockDriverState *bs, int chk, int64_t offset,
428                                  int64_t size);
429 int qcow2_pre_write_overlap_check(BlockDriverState *bs, int chk, int64_t offset,
430                                   int64_t size);
431 
432 /* qcow2-cluster.c functions */
433 int qcow2_grow_l1_table(BlockDriverState *bs, uint64_t min_size,
434                         bool exact_size);
435 int qcow2_write_l1_entry(BlockDriverState *bs, int l1_index);
436 void qcow2_l2_cache_reset(BlockDriverState *bs);
437 int qcow2_decompress_cluster(BlockDriverState *bs, uint64_t cluster_offset);
438 void qcow2_encrypt_sectors(BDRVQcowState *s, int64_t sector_num,
439                      uint8_t *out_buf, const uint8_t *in_buf,
440                      int nb_sectors, int enc,
441                      const AES_KEY *key);
442 
443 int qcow2_get_cluster_offset(BlockDriverState *bs, uint64_t offset,
444     int *num, uint64_t *cluster_offset);
445 int qcow2_alloc_cluster_offset(BlockDriverState *bs, uint64_t offset,
446     int n_start, int n_end, int *num, uint64_t *host_offset, QCowL2Meta **m);
447 uint64_t qcow2_alloc_compressed_cluster_offset(BlockDriverState *bs,
448                                          uint64_t offset,
449                                          int compressed_size);
450 
451 int qcow2_alloc_cluster_link_l2(BlockDriverState *bs, QCowL2Meta *m);
452 int qcow2_discard_clusters(BlockDriverState *bs, uint64_t offset,
453     int nb_sectors);
454 int qcow2_zero_clusters(BlockDriverState *bs, uint64_t offset, int nb_sectors);
455 
456 /* qcow2-snapshot.c functions */
457 int qcow2_snapshot_create(BlockDriverState *bs, QEMUSnapshotInfo *sn_info);
458 int qcow2_snapshot_goto(BlockDriverState *bs, const char *snapshot_id);
459 int qcow2_snapshot_delete(BlockDriverState *bs, const char *snapshot_id);
460 int qcow2_snapshot_list(BlockDriverState *bs, QEMUSnapshotInfo **psn_tab);
461 int qcow2_snapshot_load_tmp(BlockDriverState *bs, const char *snapshot_name);
462 
463 void qcow2_free_snapshots(BlockDriverState *bs);
464 int qcow2_read_snapshots(BlockDriverState *bs);
465 
466 /* qcow2-cache.c functions */
467 Qcow2Cache *qcow2_cache_create(BlockDriverState *bs, int num_tables);
468 int qcow2_cache_destroy(BlockDriverState* bs, Qcow2Cache *c);
469 
470 void qcow2_cache_entry_mark_dirty(Qcow2Cache *c, void *table);
471 int qcow2_cache_flush(BlockDriverState *bs, Qcow2Cache *c);
472 int qcow2_cache_set_dependency(BlockDriverState *bs, Qcow2Cache *c,
473     Qcow2Cache *dependency);
474 void qcow2_cache_depends_on_flush(Qcow2Cache *c);
475 
476 int qcow2_cache_get(BlockDriverState *bs, Qcow2Cache *c, uint64_t offset,
477     void **table);
478 int qcow2_cache_get_empty(BlockDriverState *bs, Qcow2Cache *c, uint64_t offset,
479     void **table);
480 int qcow2_cache_put(BlockDriverState *bs, Qcow2Cache *c, void **table);
481 
482 #endif
483