blob: 78f64336079956b6cf5f05ad213e5af69e59e722 [file] [log] [blame]
bellardea2384d2004-08-01 21:59:26 +00001/*
2 * Block driver for the VMDK format
ths5fafdf22007-09-16 21:08:06 +00003 *
bellardea2384d2004-08-01 21:59:26 +00004 * Copyright (c) 2004 Fabrice Bellard
bellardff1afc72005-04-26 21:08:00 +00005 * Copyright (c) 2005 Filip Navara
ths5fafdf22007-09-16 21:08:06 +00006 *
bellardea2384d2004-08-01 21:59:26 +00007 * Permission is hereby granted, free of charge, to any person obtaining a copy
8 * of this software and associated documentation files (the "Software"), to deal
9 * in the Software without restriction, including without limitation the rights
10 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
11 * copies of the Software, and to permit persons to whom the Software is
12 * furnished to do so, subject to the following conditions:
13 *
14 * The above copyright notice and this permission notice shall be included in
15 * all copies or substantial portions of the Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
23 * THE SOFTWARE.
24 */
ths5f4da8c2007-01-24 21:05:24 +000025
Peter Maydell80c71a22016-01-18 18:01:42 +000026#include "qemu/osdep.h"
Markus Armbrusterda34e652016-03-14 09:01:28 +010027#include "qapi/error.h"
Paolo Bonzini737e1502012-12-17 18:19:44 +010028#include "block/block_int.h"
Kevin Wolfc4bea162016-03-08 15:57:05 +010029#include "sysemu/block-backend.h"
Max Reitzabc521a2019-02-01 20:29:26 +010030#include "qapi/qmp/qdict.h"
Markus Armbrusterd49b6832015-03-17 18:29:20 +010031#include "qemu/error-report.h"
Paolo Bonzini1de7afc2012-12-17 18:20:00 +010032#include "qemu/module.h"
Markus Armbruster922a01a2018-02-01 12:18:46 +010033#include "qemu/option.h"
Paolo Bonzini58369e22016-03-15 17:22:36 +010034#include "qemu/bswap.h"
Peter Maydell5df022c2022-02-26 18:07:23 +000035#include "qemu/memalign.h"
Juan Quintela795c40b2017-04-06 12:00:28 +020036#include "migration/blocker.h"
Veronia Bahaaf348b6d2016-03-20 19:16:19 +020037#include "qemu/cutils.h"
Stefan Weil2923d342011-11-20 12:34:30 +010038#include <zlib.h>
bellardea2384d2004-08-01 21:59:26 +000039
bellardea2384d2004-08-01 21:59:26 +000040#define VMDK3_MAGIC (('C' << 24) | ('O' << 16) | ('W' << 8) | 'D')
41#define VMDK4_MAGIC (('K' << 24) | ('D' << 16) | ('M' << 8) | 'V')
Fam Zheng432bb172011-08-12 23:19:30 +080042#define VMDK4_COMPRESSION_DEFLATE 1
Fam Zheng95b0aa42013-05-02 10:25:25 +080043#define VMDK4_FLAG_NL_DETECT (1 << 0)
Fam Zhengbb45ded2011-08-12 23:19:34 +080044#define VMDK4_FLAG_RGD (1 << 1)
Fam Zheng14ead642013-05-02 10:25:23 +080045/* Zeroed-grain enable bit */
46#define VMDK4_FLAG_ZERO_GRAIN (1 << 2)
Fam Zheng432bb172011-08-12 23:19:30 +080047#define VMDK4_FLAG_COMPRESS (1 << 16)
48#define VMDK4_FLAG_MARKER (1 << 17)
Kevin Wolf65bd1552012-08-16 10:39:33 +020049#define VMDK4_GD_AT_END 0xffffffffffffffffULL
bellardea2384d2004-08-01 21:59:26 +000050
yuchenlina77672e2018-03-22 21:33:37 +080051#define VMDK_EXTENT_MAX_SECTORS (1ULL << 32)
52
Fam Zheng14ead642013-05-02 10:25:23 +080053#define VMDK_GTE_ZEROED 0x1
Fam Zheng65f74722013-05-02 10:25:22 +080054
55/* VMDK internal error codes */
56#define VMDK_OK 0
57#define VMDK_ERROR (-1)
58/* Cluster not allocated */
59#define VMDK_UNALLOC (-2)
60#define VMDK_ZEROED (-3)
61
Fam Zheng69e0b6d2013-05-02 10:25:24 +080062#define BLOCK_OPT_ZEROED_GRAIN "zeroed_grain"
Thomas Weißschuhf3d43df2021-09-13 15:04:19 +020063#define BLOCK_OPT_TOOLSVERSION "toolsversion"
Fam Zheng69e0b6d2013-05-02 10:25:24 +080064
bellardea2384d2004-08-01 21:59:26 +000065typedef struct {
66 uint32_t version;
67 uint32_t flags;
68 uint32_t disk_sectors;
69 uint32_t granularity;
70 uint32_t l1dir_offset;
71 uint32_t l1dir_size;
72 uint32_t file_sectors;
73 uint32_t cylinders;
74 uint32_t heads;
75 uint32_t sectors_per_track;
Fam Zheng5d8caa52013-08-06 15:44:47 +080076} QEMU_PACKED VMDK3Header;
bellardea2384d2004-08-01 21:59:26 +000077
78typedef struct {
79 uint32_t version;
80 uint32_t flags;
Fam Zhenge98768d2013-08-06 15:44:48 +080081 uint64_t capacity;
82 uint64_t granularity;
83 uint64_t desc_offset;
84 uint64_t desc_size;
Fam Zhengca8804c2013-08-06 15:44:55 +080085 /* Number of GrainTableEntries per GrainTable */
86 uint32_t num_gtes_per_gt;
Fam Zhenge98768d2013-08-06 15:44:48 +080087 uint64_t rgd_offset;
88 uint64_t gd_offset;
89 uint64_t grain_offset;
bellardea2384d2004-08-01 21:59:26 +000090 char filler[1];
91 char check_bytes[4];
Fam Zheng432bb172011-08-12 23:19:30 +080092 uint16_t compressAlgorithm;
Stefan Weil541dc0d2011-08-31 12:38:01 +020093} QEMU_PACKED VMDK4Header;
bellardea2384d2004-08-01 21:59:26 +000094
Sam Eiderman98eb9732019-06-20 12:10:57 +030095typedef struct VMDKSESparseConstHeader {
96 uint64_t magic;
97 uint64_t version;
98 uint64_t capacity;
99 uint64_t grain_size;
100 uint64_t grain_table_size;
101 uint64_t flags;
102 uint64_t reserved1;
103 uint64_t reserved2;
104 uint64_t reserved3;
105 uint64_t reserved4;
106 uint64_t volatile_header_offset;
107 uint64_t volatile_header_size;
108 uint64_t journal_header_offset;
109 uint64_t journal_header_size;
110 uint64_t journal_offset;
111 uint64_t journal_size;
112 uint64_t grain_dir_offset;
113 uint64_t grain_dir_size;
114 uint64_t grain_tables_offset;
115 uint64_t grain_tables_size;
116 uint64_t free_bitmap_offset;
117 uint64_t free_bitmap_size;
118 uint64_t backmap_offset;
119 uint64_t backmap_size;
120 uint64_t grains_offset;
121 uint64_t grains_size;
122 uint8_t pad[304];
123} QEMU_PACKED VMDKSESparseConstHeader;
124
125typedef struct VMDKSESparseVolatileHeader {
126 uint64_t magic;
127 uint64_t free_gt_number;
128 uint64_t next_txn_seq_number;
129 uint64_t replay_journal;
130 uint8_t pad[480];
131} QEMU_PACKED VMDKSESparseVolatileHeader;
132
bellardea2384d2004-08-01 21:59:26 +0000133#define L2_CACHE_SIZE 16
134
Fam Zhengb3976d32011-07-12 19:56:28 +0800135typedef struct VmdkExtent {
Kevin Wolf24bc15d2015-06-15 13:50:20 +0200136 BdrvChild *file;
Fam Zhengb3976d32011-07-12 19:56:28 +0800137 bool flat;
Fam Zheng432bb172011-08-12 23:19:30 +0800138 bool compressed;
139 bool has_marker;
Fam Zheng14ead642013-05-02 10:25:23 +0800140 bool has_zero_grain;
Sam Eiderman98eb9732019-06-20 12:10:57 +0300141 bool sesparse;
142 uint64_t sesparse_l2_tables_offset;
143 uint64_t sesparse_clusters_offset;
144 int32_t entry_size;
Fam Zheng14ead642013-05-02 10:25:23 +0800145 int version;
Fam Zhengb3976d32011-07-12 19:56:28 +0800146 int64_t sectors;
147 int64_t end_sector;
Fam Zheng7fa60fa2011-07-19 08:38:22 +0800148 int64_t flat_start_offset;
bellardea2384d2004-08-01 21:59:26 +0000149 int64_t l1_table_offset;
bellardff1afc72005-04-26 21:08:00 +0000150 int64_t l1_backup_table_offset;
Sam Eiderman98eb9732019-06-20 12:10:57 +0300151 void *l1_table;
bellardff1afc72005-04-26 21:08:00 +0000152 uint32_t *l1_backup_table;
bellardea2384d2004-08-01 21:59:26 +0000153 unsigned int l1_size;
154 uint32_t l1_entry_sectors;
155
156 unsigned int l2_size;
Sam Eiderman98eb9732019-06-20 12:10:57 +0300157 void *l2_cache;
bellardea2384d2004-08-01 21:59:26 +0000158 uint32_t l2_cache_offsets[L2_CACHE_SIZE];
159 uint32_t l2_cache_counts[L2_CACHE_SIZE];
160
Fam Zheng301c7d32013-09-23 17:18:29 +0800161 int64_t cluster_sectors;
Fam Zhengc6ac36e2014-07-30 14:39:10 +0800162 int64_t next_cluster_sector;
Fam Zhengf4c129a2013-10-31 10:06:23 +0800163 char *type;
Fam Zhengb3976d32011-07-12 19:56:28 +0800164} VmdkExtent;
165
166typedef struct BDRVVmdkState {
Paolo Bonzini848c66e2011-10-20 13:16:21 +0200167 CoMutex lock;
Fam Zhenge98768d2013-08-06 15:44:48 +0800168 uint64_t desc_offset;
Fam Zheng69b4d862011-07-12 19:56:34 +0800169 bool cid_updated;
Fam Zhengc338b6a2013-10-18 13:17:19 +0800170 bool cid_checked;
Fam Zhengf4c129a2013-10-31 10:06:23 +0800171 uint32_t cid;
ths5f4da8c2007-01-24 21:05:24 +0000172 uint32_t parent_cid;
Fam Zhengb3976d32011-07-12 19:56:28 +0800173 int num_extents;
174 /* Extent array with num_extents entries, ascend ordered by address */
175 VmdkExtent *extents;
Kevin Wolf2bc31662011-11-22 16:50:27 +0100176 Error *migration_blocker;
Fam Zhengf4c129a2013-10-31 10:06:23 +0800177 char *create_type;
bellardea2384d2004-08-01 21:59:26 +0000178} BDRVVmdkState;
179
Hanna Reitz6d17e282022-03-14 17:27:18 +0100180typedef struct BDRVVmdkReopenState {
181 bool *extents_using_bs_file;
182} BDRVVmdkReopenState;
183
ths630530a2007-06-18 15:01:30 +0000184typedef struct VmdkMetaData {
ths630530a2007-06-18 15:01:30 +0000185 unsigned int l1_index;
186 unsigned int l2_index;
187 unsigned int l2_offset;
Kevin Wolf4dc20e62020-04-30 15:30:02 +0200188 bool new_allocation;
Fam Zhengcdeaf1f2013-05-02 10:25:27 +0800189 uint32_t *l2_cache_entry;
ths630530a2007-06-18 15:01:30 +0000190} VmdkMetaData;
191
Fam Zheng432bb172011-08-12 23:19:30 +0800192typedef struct VmdkGrainMarker {
193 uint64_t lba;
194 uint32_t size;
Philippe Mathieu-Daudé880a7812020-03-04 16:38:16 +0100195 uint8_t data[];
Fam Zheng5d8caa52013-08-06 15:44:47 +0800196} QEMU_PACKED VmdkGrainMarker;
Fam Zheng432bb172011-08-12 23:19:30 +0800197
Kevin Wolf65bd1552012-08-16 10:39:33 +0200198enum {
199 MARKER_END_OF_STREAM = 0,
200 MARKER_GRAIN_TABLE = 1,
201 MARKER_GRAIN_DIRECTORY = 2,
202 MARKER_FOOTER = 3,
203};
204
bellardea2384d2004-08-01 21:59:26 +0000205static int vmdk_probe(const uint8_t *buf, int buf_size, const char *filename)
206{
207 uint32_t magic;
208
Fam Zhengae261c82011-07-12 19:56:38 +0800209 if (buf_size < 4) {
bellardea2384d2004-08-01 21:59:26 +0000210 return 0;
Fam Zhengae261c82011-07-12 19:56:38 +0800211 }
bellardea2384d2004-08-01 21:59:26 +0000212 magic = be32_to_cpu(*(uint32_t *)buf);
213 if (magic == VMDK3_MAGIC ||
Fam Zheng01fc99d2011-07-12 19:56:30 +0800214 magic == VMDK4_MAGIC) {
bellardea2384d2004-08-01 21:59:26 +0000215 return 100;
Fam Zheng01fc99d2011-07-12 19:56:30 +0800216 } else {
217 const char *p = (const char *)buf;
218 const char *end = p + buf_size;
219 while (p < end) {
220 if (*p == '#') {
221 /* skip comment line */
222 while (p < end && *p != '\n') {
223 p++;
224 }
225 p++;
226 continue;
227 }
228 if (*p == ' ') {
229 while (p < end && *p == ' ') {
230 p++;
231 }
232 /* skip '\r' if windows line endings used. */
233 if (p < end && *p == '\r') {
234 p++;
235 }
236 /* only accept blank lines before 'version=' line */
237 if (p == end || *p != '\n') {
238 return 0;
239 }
240 p++;
241 continue;
242 }
243 if (end - p >= strlen("version=X\n")) {
244 if (strncmp("version=1\n", p, strlen("version=1\n")) == 0 ||
Sam Eidermanb69864e2019-03-14 16:14:37 +0200245 strncmp("version=2\n", p, strlen("version=2\n")) == 0 ||
246 strncmp("version=3\n", p, strlen("version=3\n")) == 0) {
Fam Zheng01fc99d2011-07-12 19:56:30 +0800247 return 100;
248 }
249 }
250 if (end - p >= strlen("version=X\r\n")) {
251 if (strncmp("version=1\r\n", p, strlen("version=1\r\n")) == 0 ||
Sam Eidermanb69864e2019-03-14 16:14:37 +0200252 strncmp("version=2\r\n", p, strlen("version=2\r\n")) == 0 ||
253 strncmp("version=3\r\n", p, strlen("version=3\r\n")) == 0) {
Fam Zheng01fc99d2011-07-12 19:56:30 +0800254 return 100;
255 }
256 }
257 return 0;
258 }
bellardea2384d2004-08-01 21:59:26 +0000259 return 0;
Fam Zheng01fc99d2011-07-12 19:56:30 +0800260 }
bellardea2384d2004-08-01 21:59:26 +0000261}
262
ths3b46e622007-09-17 08:09:54 +0000263#define SECTOR_SIZE 512
Fam Zhengf66fd6c2011-07-19 08:45:23 +0800264#define DESC_SIZE (20 * SECTOR_SIZE) /* 20 sectors of 512 bytes each */
265#define BUF_SIZE 4096
266#define HEADER_SIZE 512 /* first sector of 512 bytes */
ths5f4da8c2007-01-24 21:05:24 +0000267
Fam Zhengb3976d32011-07-12 19:56:28 +0800268static void vmdk_free_extents(BlockDriverState *bs)
269{
270 int i;
271 BDRVVmdkState *s = bs->opaque;
Fam Zhengb3c0bfb2011-09-19 10:26:42 +0800272 VmdkExtent *e;
Fam Zhengb3976d32011-07-12 19:56:28 +0800273
Stefan Hajnoczi6bc30f12023-12-05 13:20:02 -0500274 bdrv_graph_wrlock();
Fam Zhengb3976d32011-07-12 19:56:28 +0800275 for (i = 0; i < s->num_extents; i++) {
Fam Zhengb3c0bfb2011-09-19 10:26:42 +0800276 e = &s->extents[i];
277 g_free(e->l1_table);
278 g_free(e->l2_cache);
279 g_free(e->l1_backup_table);
Fam Zhengf4c129a2013-10-31 10:06:23 +0800280 g_free(e->type);
Kevin Wolf9a4f4c32015-06-16 14:19:22 +0200281 if (e->file != bs->file) {
Kevin Wolf24bc15d2015-06-15 13:50:20 +0200282 bdrv_unref_child(bs, e->file);
Fam Zhengb3c0bfb2011-09-19 10:26:42 +0800283 }
Fam Zhengb3976d32011-07-12 19:56:28 +0800284 }
Stefan Hajnoczi6bc30f12023-12-05 13:20:02 -0500285 bdrv_graph_wrunlock();
Kevin Wolf32a8aba2023-09-11 11:46:19 +0200286
Anthony Liguori7267c092011-08-20 22:09:37 -0500287 g_free(s->extents);
Fam Zhengb3976d32011-07-12 19:56:28 +0800288}
289
Fam Zheng86c6b422011-08-12 23:19:28 +0800290static void vmdk_free_last_extent(BlockDriverState *bs)
291{
292 BDRVVmdkState *s = bs->opaque;
293
294 if (s->num_extents == 0) {
295 return;
296 }
297 s->num_extents--;
Markus Armbruster5839e532014-08-19 10:31:08 +0200298 s->extents = g_renew(VmdkExtent, s->extents, s->num_extents);
Fam Zheng86c6b422011-08-12 23:19:28 +0800299}
300
Peter Maydell98778602017-07-09 18:06:14 +0100301/* Return -ve errno, or 0 on success and write CID into *pcid. */
Kevin Wolf1f051dc2023-10-27 17:53:33 +0200302static int GRAPH_RDLOCK
303vmdk_read_cid(BlockDriverState *bs, int parent, uint32_t *pcid)
ths5f4da8c2007-01-24 21:05:24 +0000304{
Fam Zheng5997c212016-03-08 16:24:35 +0800305 char *desc;
Peter Maydell98778602017-07-09 18:06:14 +0100306 uint32_t cid;
blueswir17ccfb2e2008-09-14 06:45:34 +0000307 const char *p_name, *cid_str;
ths5f4da8c2007-01-24 21:05:24 +0000308 size_t cid_str_size;
Fam Zhenge1da9b22011-07-12 19:56:32 +0800309 BDRVVmdkState *s = bs->opaque;
Kevin Wolf99f18352011-10-26 12:25:25 +0200310 int ret;
ths5f4da8c2007-01-24 21:05:24 +0000311
Fam Zheng5997c212016-03-08 16:24:35 +0800312 desc = g_malloc0(DESC_SIZE);
Alberto Faria32cc71d2022-06-09 16:27:36 +0100313 ret = bdrv_pread(bs->file, s->desc_offset, DESC_SIZE, desc, 0);
Kevin Wolf99f18352011-10-26 12:25:25 +0200314 if (ret < 0) {
Peter Maydell98778602017-07-09 18:06:14 +0100315 goto out;
Fam Zhenge1da9b22011-07-12 19:56:32 +0800316 }
ths5f4da8c2007-01-24 21:05:24 +0000317
318 if (parent) {
319 cid_str = "parentCID";
320 cid_str_size = sizeof("parentCID");
321 } else {
322 cid_str = "CID";
323 cid_str_size = sizeof("CID");
324 }
325
Kevin Wolf93897b92011-10-26 12:25:52 +0200326 desc[DESC_SIZE - 1] = '\0';
Fam Zhengae261c82011-07-12 19:56:38 +0800327 p_name = strstr(desc, cid_str);
Peter Maydell98778602017-07-09 18:06:14 +0100328 if (p_name == NULL) {
329 ret = -EINVAL;
330 goto out;
ths5f4da8c2007-01-24 21:05:24 +0000331 }
Peter Maydell98778602017-07-09 18:06:14 +0100332 p_name += cid_str_size;
333 if (sscanf(p_name, "%" SCNx32, &cid) != 1) {
334 ret = -EINVAL;
335 goto out;
336 }
337 *pcid = cid;
338 ret = 0;
ths5f4da8c2007-01-24 21:05:24 +0000339
Peter Maydell98778602017-07-09 18:06:14 +0100340out:
Fam Zheng5997c212016-03-08 16:24:35 +0800341 g_free(desc);
Peter Maydell98778602017-07-09 18:06:14 +0100342 return ret;
ths5f4da8c2007-01-24 21:05:24 +0000343}
344
Paolo Bonzini28944f92023-06-01 13:51:41 +0200345static int coroutine_fn GRAPH_RDLOCK
346vmdk_write_cid(BlockDriverState *bs, uint32_t cid)
ths5f4da8c2007-01-24 21:05:24 +0000347{
Fam Zheng965415e2016-03-08 16:24:34 +0800348 char *desc, *tmp_desc;
ths5f4da8c2007-01-24 21:05:24 +0000349 char *p_name, *tmp_str;
Fam Zhenge1da9b22011-07-12 19:56:32 +0800350 BDRVVmdkState *s = bs->opaque;
Fam Zheng965415e2016-03-08 16:24:34 +0800351 int ret = 0;
ths5f4da8c2007-01-24 21:05:24 +0000352
Fam Zheng9fb7b352023-11-24 11:56:54 +0000353 size_t desc_buf_size;
354
355 if (s->desc_offset == 0) {
356 desc_buf_size = bdrv_getlength(bs->file->bs);
357 if (desc_buf_size > 16ULL << 20) {
358 error_report("VMDK description file too big");
359 return -EFBIG;
360 }
361 } else {
362 desc_buf_size = DESC_SIZE;
363 }
364
365 desc = g_malloc0(desc_buf_size);
366 tmp_desc = g_malloc0(desc_buf_size);
367 ret = bdrv_co_pread(bs->file, s->desc_offset, desc_buf_size, desc, 0);
Kevin Wolf99f18352011-10-26 12:25:25 +0200368 if (ret < 0) {
Fam Zheng965415e2016-03-08 16:24:34 +0800369 goto out;
Fam Zhenge1da9b22011-07-12 19:56:32 +0800370 }
ths5f4da8c2007-01-24 21:05:24 +0000371
Fam Zheng9fb7b352023-11-24 11:56:54 +0000372 desc[desc_buf_size - 1] = '\0';
Fam Zhengae261c82011-07-12 19:56:38 +0800373 tmp_str = strstr(desc, "parentCID");
Kevin Wolf93897b92011-10-26 12:25:52 +0200374 if (tmp_str == NULL) {
Fam Zheng965415e2016-03-08 16:24:34 +0800375 ret = -EINVAL;
376 goto out;
Kevin Wolf93897b92011-10-26 12:25:52 +0200377 }
378
Fam Zheng9fb7b352023-11-24 11:56:54 +0000379 pstrcpy(tmp_desc, desc_buf_size, tmp_str);
Fam Zhengae261c82011-07-12 19:56:38 +0800380 p_name = strstr(desc, "CID");
381 if (p_name != NULL) {
ths5f4da8c2007-01-24 21:05:24 +0000382 p_name += sizeof("CID");
Fam Zheng9fb7b352023-11-24 11:56:54 +0000383 snprintf(p_name, desc_buf_size - (p_name - desc), "%" PRIx32 "\n", cid);
384 pstrcat(desc, desc_buf_size, tmp_desc);
ths5f4da8c2007-01-24 21:05:24 +0000385 }
386
Fam Zheng9fb7b352023-11-24 11:56:54 +0000387 ret = bdrv_co_pwrite_sync(bs->file, s->desc_offset, desc_buf_size, desc, 0);
Kevin Wolf99f18352011-10-26 12:25:25 +0200388
Fam Zheng965415e2016-03-08 16:24:34 +0800389out:
390 g_free(desc);
391 g_free(tmp_desc);
392 return ret;
ths5f4da8c2007-01-24 21:05:24 +0000393}
394
Kevin Wolf004915a2023-10-27 17:53:26 +0200395static int coroutine_fn GRAPH_RDLOCK vmdk_is_cid_valid(BlockDriverState *bs)
ths5f4da8c2007-01-24 21:05:24 +0000396{
ths5f4da8c2007-01-24 21:05:24 +0000397 BDRVVmdkState *s = bs->opaque;
ths5f4da8c2007-01-24 21:05:24 +0000398 uint32_t cur_pcid;
399
Kevin Wolf760e0062015-06-17 14:55:21 +0200400 if (!s->cid_checked && bs->backing) {
401 BlockDriverState *p_bs = bs->backing->bs;
402
Max Reitz439e89f2018-07-02 23:07:20 +0200403 if (strcmp(p_bs->drv->format_name, "vmdk")) {
404 /* Backing file is not in vmdk format, so it does not have
405 * a CID, which makes the overlay's parent CID invalid */
406 return 0;
407 }
408
Peter Maydell98778602017-07-09 18:06:14 +0100409 if (vmdk_read_cid(p_bs, 0, &cur_pcid) != 0) {
410 /* read failure: report as not valid */
411 return 0;
412 }
Fam Zhengae261c82011-07-12 19:56:38 +0800413 if (s->parent_cid != cur_pcid) {
414 /* CID not valid */
ths5f4da8c2007-01-24 21:05:24 +0000415 return 0;
Fam Zhengae261c82011-07-12 19:56:38 +0800416 }
ths5f4da8c2007-01-24 21:05:24 +0000417 }
Fam Zhengc338b6a2013-10-18 13:17:19 +0800418 s->cid_checked = true;
Fam Zhengae261c82011-07-12 19:56:38 +0800419 /* CID valid */
ths5f4da8c2007-01-24 21:05:24 +0000420 return 1;
421}
422
Jeff Cody38975752012-09-20 15:13:30 -0400423static int vmdk_reopen_prepare(BDRVReopenState *state,
424 BlockReopenQueue *queue, Error **errp)
425{
Hanna Reitz6d17e282022-03-14 17:27:18 +0100426 BDRVVmdkState *s;
427 BDRVVmdkReopenState *rs;
428 int i;
429
Kevin Wolf1f051dc2023-10-27 17:53:33 +0200430 GLOBAL_STATE_CODE();
431 GRAPH_RDLOCK_GUARD_MAINLOOP();
432
Jeff Cody38975752012-09-20 15:13:30 -0400433 assert(state != NULL);
434 assert(state->bs != NULL);
Hanna Reitz6d17e282022-03-14 17:27:18 +0100435 assert(state->opaque == NULL);
436
437 s = state->bs->opaque;
438
439 rs = g_new0(BDRVVmdkReopenState, 1);
440 state->opaque = rs;
441
442 /*
443 * Check whether there are any extents stored in bs->file; if bs->file
444 * changes, we will need to update their .file pointers to follow suit
445 */
446 rs->extents_using_bs_file = g_new(bool, s->num_extents);
447 for (i = 0; i < s->num_extents; i++) {
448 rs->extents_using_bs_file[i] = s->extents[i].file == state->bs->file;
449 }
450
Kevin Wolf67251a32015-04-09 18:54:04 +0200451 return 0;
Jeff Cody38975752012-09-20 15:13:30 -0400452}
453
Hanna Reitz6d17e282022-03-14 17:27:18 +0100454static void vmdk_reopen_clean(BDRVReopenState *state)
455{
456 BDRVVmdkReopenState *rs = state->opaque;
457
458 g_free(rs->extents_using_bs_file);
459 g_free(rs);
460 state->opaque = NULL;
461}
462
463static void vmdk_reopen_commit(BDRVReopenState *state)
464{
465 BDRVVmdkState *s = state->bs->opaque;
466 BDRVVmdkReopenState *rs = state->opaque;
467 int i;
468
Kevin Wolf1f051dc2023-10-27 17:53:33 +0200469 GLOBAL_STATE_CODE();
470 GRAPH_RDLOCK_GUARD_MAINLOOP();
471
Hanna Reitz6d17e282022-03-14 17:27:18 +0100472 for (i = 0; i < s->num_extents; i++) {
473 if (rs->extents_using_bs_file[i]) {
474 s->extents[i].file = state->bs->file;
475 }
476 }
477
478 vmdk_reopen_clean(state);
479}
480
481static void vmdk_reopen_abort(BDRVReopenState *state)
482{
483 vmdk_reopen_clean(state);
484}
485
Kevin Wolf1f051dc2023-10-27 17:53:33 +0200486static int GRAPH_RDLOCK vmdk_parent_open(BlockDriverState *bs)
ths5f4da8c2007-01-24 21:05:24 +0000487{
ths5fafdf22007-09-16 21:08:06 +0000488 char *p_name;
Fam Zheng71968db2016-03-08 16:24:36 +0800489 char *desc;
Fam Zhenge1da9b22011-07-12 19:56:32 +0800490 BDRVVmdkState *s = bs->opaque;
Paolo Bonzini588b65a2011-10-20 13:16:19 +0200491 int ret;
ths5f4da8c2007-01-24 21:05:24 +0000492
Fam Zheng71968db2016-03-08 16:24:36 +0800493 desc = g_malloc0(DESC_SIZE + 1);
Alberto Faria32cc71d2022-06-09 16:27:36 +0100494 ret = bdrv_pread(bs->file, s->desc_offset, DESC_SIZE, desc, 0);
Paolo Bonzini588b65a2011-10-20 13:16:19 +0200495 if (ret < 0) {
Fam Zheng71968db2016-03-08 16:24:36 +0800496 goto out;
Fam Zhenge1da9b22011-07-12 19:56:32 +0800497 }
ths5f4da8c2007-01-24 21:05:24 +0000498
Fam Zhengae261c82011-07-12 19:56:38 +0800499 p_name = strstr(desc, "parentFileNameHint");
500 if (p_name != NULL) {
ths5f4da8c2007-01-24 21:05:24 +0000501 char *end_name;
ths5f4da8c2007-01-24 21:05:24 +0000502
503 p_name += sizeof("parentFileNameHint") + 1;
Fam Zhengae261c82011-07-12 19:56:38 +0800504 end_name = strchr(p_name, '\"');
505 if (end_name == NULL) {
Fam Zheng71968db2016-03-08 16:24:36 +0800506 ret = -EINVAL;
507 goto out;
Fam Zhengae261c82011-07-12 19:56:38 +0800508 }
Max Reitz998c2012019-02-01 20:29:08 +0100509 if ((end_name - p_name) > sizeof(bs->auto_backing_file) - 1) {
Fam Zheng71968db2016-03-08 16:24:36 +0800510 ret = -EINVAL;
511 goto out;
Fam Zhengae261c82011-07-12 19:56:38 +0800512 }
ths3b46e622007-09-17 08:09:54 +0000513
Max Reitz998c2012019-02-01 20:29:08 +0100514 pstrcpy(bs->auto_backing_file, end_name - p_name + 1, p_name);
515 pstrcpy(bs->backing_file, sizeof(bs->backing_file),
516 bs->auto_backing_file);
Sam Eiderman7502be82019-03-26 21:58:37 +0200517 pstrcpy(bs->backing_format, sizeof(bs->backing_format),
518 "vmdk");
ths5f4da8c2007-01-24 21:05:24 +0000519 }
520
Fam Zheng71968db2016-03-08 16:24:36 +0800521out:
522 g_free(desc);
523 return ret;
ths5f4da8c2007-01-24 21:05:24 +0000524}
525
Fam Zhengb3976d32011-07-12 19:56:28 +0800526/* Create and append extent to the extent array. Return the added VmdkExtent
527 * address. return NULL if allocation failed. */
Fam Zheng8aa13312013-08-06 15:44:51 +0800528static int vmdk_add_extent(BlockDriverState *bs,
Kevin Wolf24bc15d2015-06-15 13:50:20 +0200529 BdrvChild *file, bool flat, int64_t sectors,
Fam Zhengb3976d32011-07-12 19:56:28 +0800530 int64_t l1_offset, int64_t l1_backup_offset,
531 uint32_t l1_size,
Fam Zheng8aa13312013-08-06 15:44:51 +0800532 int l2_size, uint64_t cluster_sectors,
Fam Zheng48239702013-10-11 15:43:22 +0800533 VmdkExtent **new_extent,
534 Error **errp)
Fam Zhengb3976d32011-07-12 19:56:28 +0800535{
536 VmdkExtent *extent;
537 BDRVVmdkState *s = bs->opaque;
Markus Armbruster0a156f72014-08-21 14:36:19 +0200538 int64_t nb_sectors;
Fam Zhengb3976d32011-07-12 19:56:28 +0800539
Fam Zheng8aa13312013-08-06 15:44:51 +0800540 if (cluster_sectors > 0x200000) {
541 /* 0x200000 * 512Bytes = 1GB for one cluster is unrealistic */
Fam Zheng48239702013-10-11 15:43:22 +0800542 error_setg(errp, "Invalid granularity, image may be corrupt");
543 return -EFBIG;
Fam Zheng8aa13312013-08-06 15:44:51 +0800544 }
Sam Eiderman59d6ee42019-06-20 12:10:56 +0300545 if (l1_size > 32 * 1024 * 1024) {
Sam Eiderman940a2cd2019-06-20 12:10:55 +0300546 /*
547 * Although with big capacity and small l1_entry_sectors, we can get a
Fam Zhengb0651b82013-08-19 18:54:25 +0800548 * big l1_size, we don't want unbounded value to allocate the table.
Sam Eiderman59d6ee42019-06-20 12:10:56 +0300549 * Limit it to 32M, which is enough to store:
550 * 8TB - for both VMDK3 & VMDK4 with
551 * minimal cluster size: 512B
552 * minimal L2 table size: 512 entries
553 * 8 TB is still more than the maximal value supported for
554 * VMDK3 & VMDK4 which is 2TB.
Sam Eiderman98eb9732019-06-20 12:10:57 +0300555 * 64TB - for "ESXi seSparse Extent"
556 * minimal cluster size: 512B (default is 4KB)
557 * L2 table size: 4096 entries (const).
558 * 64TB is more than the maximal value supported for
559 * seSparse VMDKs (which is slightly less than 64TB)
Sam Eiderman940a2cd2019-06-20 12:10:55 +0300560 */
Fam Zheng48239702013-10-11 15:43:22 +0800561 error_setg(errp, "L1 size too big");
Fam Zhengb0651b82013-08-19 18:54:25 +0800562 return -EFBIG;
563 }
Fam Zheng8aa13312013-08-06 15:44:51 +0800564
Kevin Wolf24bc15d2015-06-15 13:50:20 +0200565 nb_sectors = bdrv_nb_sectors(file->bs);
Markus Armbruster0a156f72014-08-21 14:36:19 +0200566 if (nb_sectors < 0) {
567 return nb_sectors;
Fam Zhengc6ac36e2014-07-30 14:39:10 +0800568 }
569
Markus Armbruster5839e532014-08-19 10:31:08 +0200570 s->extents = g_renew(VmdkExtent, s->extents, s->num_extents + 1);
Fam Zhengb3976d32011-07-12 19:56:28 +0800571 extent = &s->extents[s->num_extents];
572 s->num_extents++;
573
574 memset(extent, 0, sizeof(VmdkExtent));
575 extent->file = file;
576 extent->flat = flat;
577 extent->sectors = sectors;
578 extent->l1_table_offset = l1_offset;
579 extent->l1_backup_table_offset = l1_backup_offset;
580 extent->l1_size = l1_size;
581 extent->l1_entry_sectors = l2_size * cluster_sectors;
582 extent->l2_size = l2_size;
Fam Zheng301c7d32013-09-23 17:18:29 +0800583 extent->cluster_sectors = flat ? sectors : cluster_sectors;
Markus Armbruster0a156f72014-08-21 14:36:19 +0200584 extent->next_cluster_sector = ROUND_UP(nb_sectors, cluster_sectors);
Sam Eiderman98eb9732019-06-20 12:10:57 +0300585 extent->entry_size = sizeof(uint32_t);
Fam Zhengb3976d32011-07-12 19:56:28 +0800586
587 if (s->num_extents > 1) {
588 extent->end_sector = (*(extent - 1)).end_sector + extent->sectors;
589 } else {
590 extent->end_sector = extent->sectors;
591 }
592 bs->total_sectors = extent->end_sector;
Fam Zheng8aa13312013-08-06 15:44:51 +0800593 if (new_extent) {
594 *new_extent = extent;
595 }
596 return 0;
Fam Zhengb3976d32011-07-12 19:56:28 +0800597}
598
Kevin Wolfb7cfc7d2023-09-29 16:51:45 +0200599static int GRAPH_RDLOCK
600vmdk_init_tables(BlockDriverState *bs, VmdkExtent *extent, Error **errp)
bellardea2384d2004-08-01 21:59:26 +0000601{
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800602 int ret;
Fam Zheng13c49412015-05-05 17:28:13 +0800603 size_t l1_size;
604 int i;
ths5f4da8c2007-01-24 21:05:24 +0000605
bellardea2384d2004-08-01 21:59:26 +0000606 /* read the L1 table */
Sam Eiderman98eb9732019-06-20 12:10:57 +0300607 l1_size = extent->l1_size * extent->entry_size;
Kevin Wolfd6e59932014-05-20 13:56:27 +0200608 extent->l1_table = g_try_malloc(l1_size);
609 if (l1_size && extent->l1_table == NULL) {
610 return -ENOMEM;
611 }
612
Alberto Faria32cc71d2022-06-09 16:27:36 +0100613 ret = bdrv_pread(extent->file, extent->l1_table_offset, l1_size,
614 extent->l1_table, 0);
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800615 if (ret < 0) {
Max Reitzf30c66b2019-02-01 20:29:05 +0100616 bdrv_refresh_filename(extent->file->bs);
Fam Zheng48239702013-10-11 15:43:22 +0800617 error_setg_errno(errp, -ret,
618 "Could not read l1 table from extent '%s'",
Kevin Wolf24bc15d2015-06-15 13:50:20 +0200619 extent->file->bs->filename);
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800620 goto fail_l1;
Fam Zhengb3976d32011-07-12 19:56:28 +0800621 }
622 for (i = 0; i < extent->l1_size; i++) {
Sam Eiderman98eb9732019-06-20 12:10:57 +0300623 if (extent->entry_size == sizeof(uint64_t)) {
624 le64_to_cpus((uint64_t *)extent->l1_table + i);
625 } else {
626 assert(extent->entry_size == sizeof(uint32_t));
627 le32_to_cpus((uint32_t *)extent->l1_table + i);
628 }
bellardea2384d2004-08-01 21:59:26 +0000629 }
630
Fam Zhengb3976d32011-07-12 19:56:28 +0800631 if (extent->l1_backup_table_offset) {
Sam Eiderman98eb9732019-06-20 12:10:57 +0300632 assert(!extent->sesparse);
Kevin Wolfd6e59932014-05-20 13:56:27 +0200633 extent->l1_backup_table = g_try_malloc(l1_size);
634 if (l1_size && extent->l1_backup_table == NULL) {
635 ret = -ENOMEM;
636 goto fail_l1;
637 }
Alberto Faria53fb7842022-06-09 16:27:35 +0100638 ret = bdrv_pread(extent->file, extent->l1_backup_table_offset,
Alberto Faria32cc71d2022-06-09 16:27:36 +0100639 l1_size, extent->l1_backup_table, 0);
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800640 if (ret < 0) {
Max Reitzf30c66b2019-02-01 20:29:05 +0100641 bdrv_refresh_filename(extent->file->bs);
Fam Zheng48239702013-10-11 15:43:22 +0800642 error_setg_errno(errp, -ret,
643 "Could not read l1 backup table from extent '%s'",
Kevin Wolf24bc15d2015-06-15 13:50:20 +0200644 extent->file->bs->filename);
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800645 goto fail_l1b;
Fam Zhengb3976d32011-07-12 19:56:28 +0800646 }
647 for (i = 0; i < extent->l1_size; i++) {
648 le32_to_cpus(&extent->l1_backup_table[i]);
bellardff1afc72005-04-26 21:08:00 +0000649 }
650 }
651
Fam Zhengb3976d32011-07-12 19:56:28 +0800652 extent->l2_cache =
Sam Eiderman98eb9732019-06-20 12:10:57 +0300653 g_malloc(extent->entry_size * extent->l2_size * L2_CACHE_SIZE);
bellardea2384d2004-08-01 21:59:26 +0000654 return 0;
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800655 fail_l1b:
Anthony Liguori7267c092011-08-20 22:09:37 -0500656 g_free(extent->l1_backup_table);
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800657 fail_l1:
Anthony Liguori7267c092011-08-20 22:09:37 -0500658 g_free(extent->l1_table);
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800659 return ret;
660}
661
Kevin Wolfb7cfc7d2023-09-29 16:51:45 +0200662static int GRAPH_RDLOCK
663vmdk_open_vmfs_sparse(BlockDriverState *bs, BdrvChild *file, int flags,
664 Error **errp)
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800665{
666 int ret;
667 uint32_t magic;
668 VMDK3Header header;
Christian Borntraegercd466702020-09-30 17:58:56 +0200669 VmdkExtent *extent = NULL;
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800670
Alberto Faria32cc71d2022-06-09 16:27:36 +0100671 ret = bdrv_pread(file, sizeof(magic), sizeof(header), &header, 0);
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800672 if (ret < 0) {
Max Reitzf30c66b2019-02-01 20:29:05 +0100673 bdrv_refresh_filename(file->bs);
Fam Zheng48239702013-10-11 15:43:22 +0800674 error_setg_errno(errp, -ret,
675 "Could not read header from file '%s'",
Kevin Wolf24bc15d2015-06-15 13:50:20 +0200676 file->bs->filename);
Fam Zheng86c6b422011-08-12 23:19:28 +0800677 return ret;
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800678 }
Fam Zhengf6b61e52013-08-19 18:54:26 +0800679 ret = vmdk_add_extent(bs, file, false,
680 le32_to_cpu(header.disk_sectors),
Fam Zheng7237aec2015-04-27 22:23:01 +0800681 (int64_t)le32_to_cpu(header.l1dir_offset) << 9,
Fam Zhengf6b61e52013-08-19 18:54:26 +0800682 0,
683 le32_to_cpu(header.l1dir_size),
684 4096,
685 le32_to_cpu(header.granularity),
Fam Zheng48239702013-10-11 15:43:22 +0800686 &extent,
687 errp);
Fam Zheng8aa13312013-08-06 15:44:51 +0800688 if (ret < 0) {
689 return ret;
690 }
Fam Zheng48239702013-10-11 15:43:22 +0800691 ret = vmdk_init_tables(bs, extent, errp);
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800692 if (ret) {
Fam Zheng86c6b422011-08-12 23:19:28 +0800693 /* free extent allocated by vmdk_add_extent */
694 vmdk_free_last_extent(bs);
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800695 }
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800696 return ret;
697}
698
Sam Eiderman98eb9732019-06-20 12:10:57 +0300699#define SESPARSE_CONST_HEADER_MAGIC UINT64_C(0x00000000cafebabe)
700#define SESPARSE_VOLATILE_HEADER_MAGIC UINT64_C(0x00000000cafecafe)
701
702/* Strict checks - format not officially documented */
703static int check_se_sparse_const_header(VMDKSESparseConstHeader *header,
704 Error **errp)
705{
706 header->magic = le64_to_cpu(header->magic);
707 header->version = le64_to_cpu(header->version);
708 header->grain_size = le64_to_cpu(header->grain_size);
709 header->grain_table_size = le64_to_cpu(header->grain_table_size);
710 header->flags = le64_to_cpu(header->flags);
711 header->reserved1 = le64_to_cpu(header->reserved1);
712 header->reserved2 = le64_to_cpu(header->reserved2);
713 header->reserved3 = le64_to_cpu(header->reserved3);
714 header->reserved4 = le64_to_cpu(header->reserved4);
715
716 header->volatile_header_offset =
717 le64_to_cpu(header->volatile_header_offset);
718 header->volatile_header_size = le64_to_cpu(header->volatile_header_size);
719
720 header->journal_header_offset = le64_to_cpu(header->journal_header_offset);
721 header->journal_header_size = le64_to_cpu(header->journal_header_size);
722
723 header->journal_offset = le64_to_cpu(header->journal_offset);
724 header->journal_size = le64_to_cpu(header->journal_size);
725
726 header->grain_dir_offset = le64_to_cpu(header->grain_dir_offset);
727 header->grain_dir_size = le64_to_cpu(header->grain_dir_size);
728
729 header->grain_tables_offset = le64_to_cpu(header->grain_tables_offset);
730 header->grain_tables_size = le64_to_cpu(header->grain_tables_size);
731
732 header->free_bitmap_offset = le64_to_cpu(header->free_bitmap_offset);
733 header->free_bitmap_size = le64_to_cpu(header->free_bitmap_size);
734
735 header->backmap_offset = le64_to_cpu(header->backmap_offset);
736 header->backmap_size = le64_to_cpu(header->backmap_size);
737
738 header->grains_offset = le64_to_cpu(header->grains_offset);
739 header->grains_size = le64_to_cpu(header->grains_size);
740
741 if (header->magic != SESPARSE_CONST_HEADER_MAGIC) {
742 error_setg(errp, "Bad const header magic: 0x%016" PRIx64,
743 header->magic);
744 return -EINVAL;
745 }
746
747 if (header->version != 0x0000000200000001) {
748 error_setg(errp, "Unsupported version: 0x%016" PRIx64,
749 header->version);
750 return -ENOTSUP;
751 }
752
753 if (header->grain_size != 8) {
754 error_setg(errp, "Unsupported grain size: %" PRIu64,
755 header->grain_size);
756 return -ENOTSUP;
757 }
758
759 if (header->grain_table_size != 64) {
760 error_setg(errp, "Unsupported grain table size: %" PRIu64,
761 header->grain_table_size);
762 return -ENOTSUP;
763 }
764
765 if (header->flags != 0) {
766 error_setg(errp, "Unsupported flags: 0x%016" PRIx64,
767 header->flags);
768 return -ENOTSUP;
769 }
770
771 if (header->reserved1 != 0 || header->reserved2 != 0 ||
772 header->reserved3 != 0 || header->reserved4 != 0) {
773 error_setg(errp, "Unsupported reserved bits:"
774 " 0x%016" PRIx64 " 0x%016" PRIx64
775 " 0x%016" PRIx64 " 0x%016" PRIx64,
776 header->reserved1, header->reserved2,
777 header->reserved3, header->reserved4);
778 return -ENOTSUP;
779 }
780
781 /* check that padding is 0 */
782 if (!buffer_is_zero(header->pad, sizeof(header->pad))) {
783 error_setg(errp, "Unsupported non-zero const header padding");
784 return -ENOTSUP;
785 }
786
787 return 0;
788}
789
790static int check_se_sparse_volatile_header(VMDKSESparseVolatileHeader *header,
791 Error **errp)
792{
793 header->magic = le64_to_cpu(header->magic);
794 header->free_gt_number = le64_to_cpu(header->free_gt_number);
795 header->next_txn_seq_number = le64_to_cpu(header->next_txn_seq_number);
796 header->replay_journal = le64_to_cpu(header->replay_journal);
797
798 if (header->magic != SESPARSE_VOLATILE_HEADER_MAGIC) {
799 error_setg(errp, "Bad volatile header magic: 0x%016" PRIx64,
800 header->magic);
801 return -EINVAL;
802 }
803
804 if (header->replay_journal) {
805 error_setg(errp, "Image is dirty, Replaying journal not supported");
806 return -ENOTSUP;
807 }
808
809 /* check that padding is 0 */
810 if (!buffer_is_zero(header->pad, sizeof(header->pad))) {
811 error_setg(errp, "Unsupported non-zero volatile header padding");
812 return -ENOTSUP;
813 }
814
815 return 0;
816}
817
Kevin Wolfb7cfc7d2023-09-29 16:51:45 +0200818static int GRAPH_RDLOCK
819vmdk_open_se_sparse(BlockDriverState *bs, BdrvChild *file, int flags,
820 Error **errp)
Sam Eiderman98eb9732019-06-20 12:10:57 +0300821{
822 int ret;
823 VMDKSESparseConstHeader const_header;
824 VMDKSESparseVolatileHeader volatile_header;
Christian Borntraegercd466702020-09-30 17:58:56 +0200825 VmdkExtent *extent = NULL;
Sam Eiderman98eb9732019-06-20 12:10:57 +0300826
827 ret = bdrv_apply_auto_read_only(bs,
828 "No write support for seSparse images available", errp);
829 if (ret < 0) {
830 return ret;
831 }
832
833 assert(sizeof(const_header) == SECTOR_SIZE);
834
Alberto Faria32cc71d2022-06-09 16:27:36 +0100835 ret = bdrv_pread(file, 0, sizeof(const_header), &const_header, 0);
Sam Eiderman98eb9732019-06-20 12:10:57 +0300836 if (ret < 0) {
837 bdrv_refresh_filename(file->bs);
838 error_setg_errno(errp, -ret,
839 "Could not read const header from file '%s'",
840 file->bs->filename);
841 return ret;
842 }
843
844 /* check const header */
845 ret = check_se_sparse_const_header(&const_header, errp);
846 if (ret < 0) {
847 return ret;
848 }
849
850 assert(sizeof(volatile_header) == SECTOR_SIZE);
851
Alberto Faria53fb7842022-06-09 16:27:35 +0100852 ret = bdrv_pread(file, const_header.volatile_header_offset * SECTOR_SIZE,
Alberto Faria32cc71d2022-06-09 16:27:36 +0100853 sizeof(volatile_header), &volatile_header, 0);
Sam Eiderman98eb9732019-06-20 12:10:57 +0300854 if (ret < 0) {
855 bdrv_refresh_filename(file->bs);
856 error_setg_errno(errp, -ret,
857 "Could not read volatile header from file '%s'",
858 file->bs->filename);
859 return ret;
860 }
861
862 /* check volatile header */
863 ret = check_se_sparse_volatile_header(&volatile_header, errp);
864 if (ret < 0) {
865 return ret;
866 }
867
868 ret = vmdk_add_extent(bs, file, false,
869 const_header.capacity,
870 const_header.grain_dir_offset * SECTOR_SIZE,
871 0,
872 const_header.grain_dir_size *
873 SECTOR_SIZE / sizeof(uint64_t),
874 const_header.grain_table_size *
875 SECTOR_SIZE / sizeof(uint64_t),
876 const_header.grain_size,
877 &extent,
878 errp);
879 if (ret < 0) {
880 return ret;
881 }
882
883 extent->sesparse = true;
884 extent->sesparse_l2_tables_offset = const_header.grain_tables_offset;
885 extent->sesparse_clusters_offset = const_header.grains_offset;
886 extent->entry_size = sizeof(uint64_t);
887
888 ret = vmdk_init_tables(bs, extent, errp);
889 if (ret) {
890 /* free extent allocated by vmdk_add_extent */
891 vmdk_free_last_extent(bs);
892 }
893
894 return ret;
895}
896
Paolo Bonzinid1833ef2014-02-17 14:44:03 +0100897static int vmdk_open_desc_file(BlockDriverState *bs, int flags, char *buf,
Kevin Wolfa6468362015-04-07 15:35:59 +0200898 QDict *options, Error **errp);
Fam Zhengf16f5092011-08-12 23:19:33 +0800899
Kevin Wolfcf2ab8f2016-06-20 18:24:02 +0200900static char *vmdk_read_desc(BdrvChild *file, uint64_t desc_offset, Error **errp)
Paolo Bonzinia8842e62014-02-17 14:44:02 +0100901{
902 int64_t size;
903 char *buf;
904 int ret;
905
Kevin Wolfcf2ab8f2016-06-20 18:24:02 +0200906 size = bdrv_getlength(file->bs);
Paolo Bonzinia8842e62014-02-17 14:44:02 +0100907 if (size < 0) {
908 error_setg_errno(errp, -size, "Could not access file");
909 return NULL;
910 }
911
Fam Zheng03c33592014-12-04 07:28:32 +0800912 if (size < 4) {
913 /* Both descriptor file and sparse image must be much larger than 4
914 * bytes, also callers of vmdk_read_desc want to compare the first 4
915 * bytes with VMDK4_MAGIC, let's error out if less is read. */
916 error_setg(errp, "File is too small, not a valid image");
917 return NULL;
918 }
919
Fam Zheng73b7bca2014-12-04 07:28:31 +0800920 size = MIN(size, (1 << 20) - 1); /* avoid unbounded allocation */
921 buf = g_malloc(size + 1);
Paolo Bonzinia8842e62014-02-17 14:44:02 +0100922
Alberto Faria32cc71d2022-06-09 16:27:36 +0100923 ret = bdrv_pread(file, desc_offset, size, buf, 0);
Paolo Bonzinia8842e62014-02-17 14:44:02 +0100924 if (ret < 0) {
925 error_setg_errno(errp, -ret, "Could not read from file");
926 g_free(buf);
927 return NULL;
928 }
Alberto Faria353a5d82022-06-09 16:27:37 +0100929 buf[size] = 0;
Paolo Bonzinia8842e62014-02-17 14:44:02 +0100930
931 return buf;
932}
933
Kevin Wolfb7cfc7d2023-09-29 16:51:45 +0200934static int GRAPH_RDLOCK
935vmdk_open_vmdk4(BlockDriverState *bs, BdrvChild *file, int flags,
936 QDict *options, Error **errp)
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800937{
938 int ret;
939 uint32_t magic;
940 uint32_t l1_size, l1_entry_sectors;
941 VMDK4Header header;
Christian Borntraegercd466702020-09-30 17:58:56 +0200942 VmdkExtent *extent = NULL;
Fam Zhengf4c129a2013-10-31 10:06:23 +0800943 BDRVVmdkState *s = bs->opaque;
Fam Zhengbb45ded2011-08-12 23:19:34 +0800944 int64_t l1_backup_offset = 0;
Fam Zheng3db1d982016-01-25 10:26:23 +0800945 bool compressed;
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800946
Alberto Faria32cc71d2022-06-09 16:27:36 +0100947 ret = bdrv_pread(file, sizeof(magic), sizeof(header), &header, 0);
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800948 if (ret < 0) {
Max Reitzf30c66b2019-02-01 20:29:05 +0100949 bdrv_refresh_filename(file->bs);
Fam Zheng48239702013-10-11 15:43:22 +0800950 error_setg_errno(errp, -ret,
951 "Could not read header from file '%s'",
Kevin Wolf24bc15d2015-06-15 13:50:20 +0200952 file->bs->filename);
Paolo Bonzini89ac8482014-02-17 14:44:05 +0100953 return -EINVAL;
Fam Zhengb4b3ab12011-07-12 19:56:31 +0800954 }
Stefan Hajnoczi5a394b92013-06-10 11:07:33 +0200955 if (header.capacity == 0) {
Fam Zhenge98768d2013-08-06 15:44:48 +0800956 uint64_t desc_offset = le64_to_cpu(header.desc_offset);
Stefan Hajnoczi5a394b92013-06-10 11:07:33 +0200957 if (desc_offset) {
Kevin Wolfcf2ab8f2016-06-20 18:24:02 +0200958 char *buf = vmdk_read_desc(file, desc_offset << 9, errp);
Paolo Bonzinid1833ef2014-02-17 14:44:03 +0100959 if (!buf) {
960 return -EINVAL;
961 }
Kevin Wolfa6468362015-04-07 15:35:59 +0200962 ret = vmdk_open_desc_file(bs, flags, buf, options, errp);
Paolo Bonzinid1833ef2014-02-17 14:44:03 +0100963 g_free(buf);
964 return ret;
Stefan Hajnoczi5a394b92013-06-10 11:07:33 +0200965 }
Fam Zhengf16f5092011-08-12 23:19:33 +0800966 }
Kevin Wolf65bd1552012-08-16 10:39:33 +0200967
Fam Zhengf4c129a2013-10-31 10:06:23 +0800968 if (!s->create_type) {
969 s->create_type = g_strdup("monolithicSparse");
970 }
971
Kevin Wolf65bd1552012-08-16 10:39:33 +0200972 if (le64_to_cpu(header.gd_offset) == VMDK4_GD_AT_END) {
973 /*
974 * The footer takes precedence over the header, so read it in. The
975 * footer starts at offset -1024 from the end: One sector for the
976 * footer, and another one for the end-of-stream marker.
977 */
978 struct {
979 struct {
980 uint64_t val;
981 uint32_t size;
982 uint32_t type;
983 uint8_t pad[512 - 16];
984 } QEMU_PACKED footer_marker;
985
986 uint32_t magic;
987 VMDK4Header header;
988 uint8_t pad[512 - 4 - sizeof(VMDK4Header)];
989
990 struct {
991 uint64_t val;
992 uint32_t size;
993 uint32_t type;
994 uint8_t pad[512 - 16];
995 } QEMU_PACKED eos_marker;
996 } QEMU_PACKED footer;
997
Alberto Faria53fb7842022-06-09 16:27:35 +0100998 ret = bdrv_pread(file, bs->file->bs->total_sectors * 512 - 1536,
Alberto Faria32cc71d2022-06-09 16:27:36 +0100999 sizeof(footer), &footer, 0);
Kevin Wolf65bd1552012-08-16 10:39:33 +02001000 if (ret < 0) {
Fam Zhengd899d2e2014-12-04 07:28:34 +08001001 error_setg_errno(errp, -ret, "Failed to read footer");
Kevin Wolf65bd1552012-08-16 10:39:33 +02001002 return ret;
1003 }
1004
1005 /* Some sanity checks for the footer */
1006 if (be32_to_cpu(footer.magic) != VMDK4_MAGIC ||
1007 le32_to_cpu(footer.footer_marker.size) != 0 ||
1008 le32_to_cpu(footer.footer_marker.type) != MARKER_FOOTER ||
1009 le64_to_cpu(footer.eos_marker.val) != 0 ||
1010 le32_to_cpu(footer.eos_marker.size) != 0 ||
1011 le32_to_cpu(footer.eos_marker.type) != MARKER_END_OF_STREAM)
1012 {
Fam Zhengd899d2e2014-12-04 07:28:34 +08001013 error_setg(errp, "Invalid footer");
Kevin Wolf65bd1552012-08-16 10:39:33 +02001014 return -EINVAL;
1015 }
1016
1017 header = footer.header;
1018 }
1019
Fam Zheng3db1d982016-01-25 10:26:23 +08001020 compressed =
1021 le16_to_cpu(header.compressAlgorithm) == VMDK4_COMPRESSION_DEFLATE;
Fam Zheng509d39a2013-11-28 09:48:03 +08001022 if (le32_to_cpu(header.version) > 3) {
Max Reitza55448b2016-03-16 19:54:33 +01001023 error_setg(errp, "Unsupported VMDK version %" PRIu32,
1024 le32_to_cpu(header.version));
Fam Zheng96c51eb2013-06-13 11:21:29 +08001025 return -ENOTSUP;
Fam Zheng3db1d982016-01-25 10:26:23 +08001026 } else if (le32_to_cpu(header.version) == 3 && (flags & BDRV_O_RDWR) &&
1027 !compressed) {
Fam Zheng509d39a2013-11-28 09:48:03 +08001028 /* VMware KB 2064959 explains that version 3 added support for
1029 * persistent changed block tracking (CBT), and backup software can
1030 * read it as version=1 if it doesn't care about the changed area
1031 * information. So we are safe to enable read only. */
1032 error_setg(errp, "VMDK version 3 must be read only");
1033 return -EINVAL;
Fam Zheng96c51eb2013-06-13 11:21:29 +08001034 }
1035
Fam Zhengca8804c2013-08-06 15:44:55 +08001036 if (le32_to_cpu(header.num_gtes_per_gt) > 512) {
Paolo Bonzini89ac8482014-02-17 14:44:05 +01001037 error_setg(errp, "L2 table size too big");
Fam Zhengf8ce0402013-08-06 15:44:52 +08001038 return -EINVAL;
1039 }
1040
Fam Zhengca8804c2013-08-06 15:44:55 +08001041 l1_entry_sectors = le32_to_cpu(header.num_gtes_per_gt)
Fam Zhengb4b3ab12011-07-12 19:56:31 +08001042 * le64_to_cpu(header.granularity);
Stefan Weil75d12342012-02-25 14:01:42 +01001043 if (l1_entry_sectors == 0) {
Fam Zhengd899d2e2014-12-04 07:28:34 +08001044 error_setg(errp, "L1 entry size is invalid");
Fam Zheng86c6b422011-08-12 23:19:28 +08001045 return -EINVAL;
1046 }
Fam Zhengb4b3ab12011-07-12 19:56:31 +08001047 l1_size = (le64_to_cpu(header.capacity) + l1_entry_sectors - 1)
1048 / l1_entry_sectors;
Fam Zhengbb45ded2011-08-12 23:19:34 +08001049 if (le32_to_cpu(header.flags) & VMDK4_FLAG_RGD) {
1050 l1_backup_offset = le64_to_cpu(header.rgd_offset) << 9;
1051 }
Kevin Wolf24bc15d2015-06-15 13:50:20 +02001052 if (bdrv_nb_sectors(file->bs) < le64_to_cpu(header.grain_offset)) {
Fam Zheng4ab9dab2014-04-17 11:34:37 +08001053 error_setg(errp, "File truncated, expecting at least %" PRId64 " bytes",
1054 (int64_t)(le64_to_cpu(header.grain_offset)
1055 * BDRV_SECTOR_SIZE));
Fam Zheng34ceed82014-01-21 15:07:43 +08001056 return -EINVAL;
1057 }
1058
Fam Zheng8aa13312013-08-06 15:44:51 +08001059 ret = vmdk_add_extent(bs, file, false,
Fam Zhengb4b3ab12011-07-12 19:56:31 +08001060 le64_to_cpu(header.capacity),
1061 le64_to_cpu(header.gd_offset) << 9,
Fam Zhengbb45ded2011-08-12 23:19:34 +08001062 l1_backup_offset,
Fam Zhengb4b3ab12011-07-12 19:56:31 +08001063 l1_size,
Fam Zhengca8804c2013-08-06 15:44:55 +08001064 le32_to_cpu(header.num_gtes_per_gt),
Fam Zheng8aa13312013-08-06 15:44:51 +08001065 le64_to_cpu(header.granularity),
Fam Zheng48239702013-10-11 15:43:22 +08001066 &extent,
1067 errp);
Fam Zheng8aa13312013-08-06 15:44:51 +08001068 if (ret < 0) {
1069 return ret;
1070 }
Fam Zheng432bb172011-08-12 23:19:30 +08001071 extent->compressed =
1072 le16_to_cpu(header.compressAlgorithm) == VMDK4_COMPRESSION_DEFLATE;
Fam Zhengd8a7b062014-01-23 15:10:52 +08001073 if (extent->compressed) {
1074 g_free(s->create_type);
1075 s->create_type = g_strdup("streamOptimized");
1076 }
Fam Zheng432bb172011-08-12 23:19:30 +08001077 extent->has_marker = le32_to_cpu(header.flags) & VMDK4_FLAG_MARKER;
Fam Zheng14ead642013-05-02 10:25:23 +08001078 extent->version = le32_to_cpu(header.version);
1079 extent->has_zero_grain = le32_to_cpu(header.flags) & VMDK4_FLAG_ZERO_GRAIN;
Fam Zheng48239702013-10-11 15:43:22 +08001080 ret = vmdk_init_tables(bs, extent, errp);
Fam Zhengb4b3ab12011-07-12 19:56:31 +08001081 if (ret) {
Fam Zheng86c6b422011-08-12 23:19:28 +08001082 /* free extent allocated by vmdk_add_extent */
1083 vmdk_free_last_extent(bs);
Fam Zhengb4b3ab12011-07-12 19:56:31 +08001084 }
Fam Zhengb4b3ab12011-07-12 19:56:31 +08001085 return ret;
1086}
1087
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001088/* find an option value out of descriptor file */
1089static int vmdk_parse_description(const char *desc, const char *opt_name,
1090 char *buf, int buf_size)
1091{
1092 char *opt_pos, *opt_end;
1093 const char *end = desc + strlen(desc);
1094
1095 opt_pos = strstr(desc, opt_name);
1096 if (!opt_pos) {
Fam Zheng65f74722013-05-02 10:25:22 +08001097 return VMDK_ERROR;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001098 }
1099 /* Skip "=\"" following opt_name */
1100 opt_pos += strlen(opt_name) + 2;
1101 if (opt_pos >= end) {
Fam Zheng65f74722013-05-02 10:25:22 +08001102 return VMDK_ERROR;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001103 }
1104 opt_end = opt_pos;
1105 while (opt_end < end && *opt_end != '"') {
1106 opt_end++;
1107 }
1108 if (opt_end == end || buf_size < opt_end - opt_pos + 1) {
Fam Zheng65f74722013-05-02 10:25:22 +08001109 return VMDK_ERROR;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001110 }
1111 pstrcpy(buf, opt_end - opt_pos + 1, opt_pos);
Fam Zheng65f74722013-05-02 10:25:22 +08001112 return VMDK_OK;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001113}
1114
Fam Zheng86c6b422011-08-12 23:19:28 +08001115/* Open an extent file and append to bs array */
Kevin Wolfb7cfc7d2023-09-29 16:51:45 +02001116static int GRAPH_RDLOCK
1117vmdk_open_sparse(BlockDriverState *bs, BdrvChild *file, int flags,
1118 char *buf, QDict *options, Error **errp)
Fam Zheng86c6b422011-08-12 23:19:28 +08001119{
1120 uint32_t magic;
1121
Paolo Bonzinid1833ef2014-02-17 14:44:03 +01001122 magic = ldl_be_p(buf);
Fam Zheng86c6b422011-08-12 23:19:28 +08001123 switch (magic) {
1124 case VMDK3_MAGIC:
Fam Zheng48239702013-10-11 15:43:22 +08001125 return vmdk_open_vmfs_sparse(bs, file, flags, errp);
Fam Zheng86c6b422011-08-12 23:19:28 +08001126 case VMDK4_MAGIC:
Kevin Wolfa6468362015-04-07 15:35:59 +02001127 return vmdk_open_vmdk4(bs, file, flags, options, errp);
Fam Zheng86c6b422011-08-12 23:19:28 +08001128 default:
Paolo Bonzini76abe402014-02-17 14:44:06 +01001129 error_setg(errp, "Image not in VMDK format");
1130 return -EINVAL;
Fam Zheng86c6b422011-08-12 23:19:28 +08001131 }
1132}
1133
Markus Armbrustere4937692015-12-18 16:35:20 +01001134static const char *next_line(const char *s)
1135{
1136 while (*s) {
1137 if (*s == '\n') {
1138 return s + 1;
1139 }
1140 s++;
1141 }
1142 return s;
1143}
1144
Kevin Wolfb7cfc7d2023-09-29 16:51:45 +02001145static int GRAPH_RDLOCK
1146vmdk_parse_extents(const char *desc, BlockDriverState *bs, QDict *options,
1147 Error **errp)
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001148{
Zhao Liu76db0ea2024-03-11 11:38:05 +08001149 ERRP_GUARD();
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001150 int ret;
Jeff Cody395a22f2015-01-22 08:03:25 -05001151 int matches;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001152 char access[11];
1153 char type[11];
1154 char fname[512];
Markus Armbrusterd28d7372015-12-18 16:35:21 +01001155 const char *p, *np;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001156 int64_t sectors = 0;
1157 int64_t flat_offset;
Max Reitzcdc0dd22019-08-15 17:36:33 +02001158 char *desc_file_dir = NULL;
Jeff Codyfe206562015-01-22 08:03:26 -05001159 char *extent_path;
Kevin Wolf24bc15d2015-06-15 13:50:20 +02001160 BdrvChild *extent_file;
Max Reitz8b1869d2020-05-13 13:05:35 +02001161 BdrvChildRole extent_role;
Fam Zhengf4c129a2013-10-31 10:06:23 +08001162 BDRVVmdkState *s = bs->opaque;
Christian Borntraegercd466702020-09-30 17:58:56 +02001163 VmdkExtent *extent = NULL;
Kevin Wolfa6468362015-04-07 15:35:59 +02001164 char extent_opt_prefix[32];
Kevin Wolf24bc15d2015-06-15 13:50:20 +02001165 Error *local_err = NULL;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001166
Kevin Wolfb7cfc7d2023-09-29 16:51:45 +02001167 GLOBAL_STATE_CODE();
1168
Markus Armbrustere4937692015-12-18 16:35:20 +01001169 for (p = desc; *p; p = next_line(p)) {
Fam Zheng8a3e0bc2014-12-04 07:28:30 +08001170 /* parse extent line in one of below formats:
1171 *
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001172 * RW [size in sectors] FLAT "file-name.vmdk" OFFSET
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001173 * RW [size in sectors] SPARSE "file-name.vmdk"
Fam Zheng8a3e0bc2014-12-04 07:28:30 +08001174 * RW [size in sectors] VMFS "file-name.vmdk"
1175 * RW [size in sectors] VMFSSPARSE "file-name.vmdk"
Sam Eiderman98eb9732019-06-20 12:10:57 +03001176 * RW [size in sectors] SESPARSE "file-name.vmdk"
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001177 */
1178 flat_offset = -1;
Jeff Cody395a22f2015-01-22 08:03:25 -05001179 matches = sscanf(p, "%10s %" SCNd64 " %10s \"%511[^\n\r\"]\" %" SCNd64,
1180 access, &sectors, type, fname, &flat_offset);
1181 if (matches < 4 || strcmp(access, "RW")) {
Markus Armbrustere4937692015-12-18 16:35:20 +01001182 continue;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001183 } else if (!strcmp(type, "FLAT")) {
Jeff Cody395a22f2015-01-22 08:03:25 -05001184 if (matches != 5 || flat_offset < 0) {
Markus Armbrusterd28d7372015-12-18 16:35:21 +01001185 goto invalid;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001186 }
Fam Zhengdbbcaa82013-10-18 15:07:33 +08001187 } else if (!strcmp(type, "VMFS")) {
Jeff Cody395a22f2015-01-22 08:03:25 -05001188 if (matches == 4) {
Fam Zhengb47053b2013-12-09 13:24:36 +08001189 flat_offset = 0;
1190 } else {
Markus Armbrusterd28d7372015-12-18 16:35:21 +01001191 goto invalid;
Fam Zhengb47053b2013-12-09 13:24:36 +08001192 }
Jeff Cody395a22f2015-01-22 08:03:25 -05001193 } else if (matches != 4) {
Markus Armbrusterd28d7372015-12-18 16:35:21 +01001194 goto invalid;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001195 }
1196
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001197 if (sectors <= 0 ||
Fam Zhengdaac8fd2013-08-19 18:54:27 +08001198 (strcmp(type, "FLAT") && strcmp(type, "SPARSE") &&
Sam Eiderman98eb9732019-06-20 12:10:57 +03001199 strcmp(type, "VMFS") && strcmp(type, "VMFSSPARSE") &&
1200 strcmp(type, "SESPARSE")) ||
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001201 (strcmp(access, "RW"))) {
Markus Armbrustere4937692015-12-18 16:35:20 +01001202 continue;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001203 }
1204
Max Reitzcdc0dd22019-08-15 17:36:33 +02001205 if (path_is_absolute(fname)) {
1206 extent_path = g_strdup(fname);
1207 } else {
1208 if (!desc_file_dir) {
1209 desc_file_dir = bdrv_dirname(bs->file->bs, errp);
1210 if (!desc_file_dir) {
1211 bdrv_refresh_filename(bs->file->bs);
1212 error_prepend(errp, "Cannot use relative paths with VMDK "
1213 "descriptor file '%s': ",
1214 bs->file->bs->filename);
1215 ret = -EINVAL;
1216 goto out;
1217 }
1218 }
Max Reitz5c984152014-12-03 14:57:22 +01001219
Max Reitzcdc0dd22019-08-15 17:36:33 +02001220 extent_path = g_strconcat(desc_file_dir, fname, NULL);
1221 }
Kevin Wolfa6468362015-04-07 15:35:59 +02001222
1223 ret = snprintf(extent_opt_prefix, 32, "extents.%d", s->num_extents);
1224 assert(ret < 32);
1225
Max Reitz8b1869d2020-05-13 13:05:35 +02001226 extent_role = BDRV_CHILD_DATA;
1227 if (strcmp(type, "FLAT") != 0 && strcmp(type, "VMFS") != 0) {
1228 /* non-flat extents have metadata */
1229 extent_role |= BDRV_CHILD_METADATA;
1230 }
1231
Kevin Wolf24bc15d2015-06-15 13:50:20 +02001232 extent_file = bdrv_open_child(extent_path, options, extent_opt_prefix,
Max Reitz8b1869d2020-05-13 13:05:35 +02001233 bs, &child_of_bds, extent_role, false,
1234 &local_err);
Jeff Codyfe206562015-01-22 08:03:26 -05001235 g_free(extent_path);
Dmitry Frolova8d99c02023-08-31 15:59:27 +03001236 if (!extent_file) {
Kevin Wolf24bc15d2015-06-15 13:50:20 +02001237 error_propagate(errp, local_err);
Max Reitzcdc0dd22019-08-15 17:36:33 +02001238 ret = -EINVAL;
1239 goto out;
Fam Zheng86c6b422011-08-12 23:19:28 +08001240 }
1241
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001242 /* save to extents array */
Paolo Bonzini04d542c2013-08-19 18:54:28 +08001243 if (!strcmp(type, "FLAT") || !strcmp(type, "VMFS")) {
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001244 /* FLAT extent */
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001245
Fam Zheng8aa13312013-08-06 15:44:51 +08001246 ret = vmdk_add_extent(bs, extent_file, true, sectors,
Fam Zheng48239702013-10-11 15:43:22 +08001247 0, 0, 0, 0, 0, &extent, errp);
Fam Zheng8aa13312013-08-06 15:44:51 +08001248 if (ret < 0) {
Kevin Wolfb7cfc7d2023-09-29 16:51:45 +02001249 bdrv_graph_rdunlock_main_loop();
Stefan Hajnoczi6bc30f12023-12-05 13:20:02 -05001250 bdrv_graph_wrlock();
Kevin Wolf24bc15d2015-06-15 13:50:20 +02001251 bdrv_unref_child(bs, extent_file);
Stefan Hajnoczi6bc30f12023-12-05 13:20:02 -05001252 bdrv_graph_wrunlock();
Kevin Wolfb7cfc7d2023-09-29 16:51:45 +02001253 bdrv_graph_rdlock_main_loop();
Max Reitzcdc0dd22019-08-15 17:36:33 +02001254 goto out;
Fam Zheng8aa13312013-08-06 15:44:51 +08001255 }
Fam Zhengf16f5092011-08-12 23:19:33 +08001256 extent->flat_start_offset = flat_offset << 9;
Fam Zhengdaac8fd2013-08-19 18:54:27 +08001257 } else if (!strcmp(type, "SPARSE") || !strcmp(type, "VMFSSPARSE")) {
1258 /* SPARSE extent and VMFSSPARSE extent are both "COWD" sparse file*/
Kevin Wolfcf2ab8f2016-06-20 18:24:02 +02001259 char *buf = vmdk_read_desc(extent_file, 0, errp);
Paolo Bonzinid1833ef2014-02-17 14:44:03 +01001260 if (!buf) {
1261 ret = -EINVAL;
1262 } else {
Kevin Wolfa6468362015-04-07 15:35:59 +02001263 ret = vmdk_open_sparse(bs, extent_file, bs->open_flags, buf,
1264 options, errp);
Paolo Bonzinid1833ef2014-02-17 14:44:03 +01001265 }
Stefan Hajnoczib6b1d312014-09-04 21:04:43 +01001266 g_free(buf);
Fam Zheng86c6b422011-08-12 23:19:28 +08001267 if (ret) {
Kevin Wolfb7cfc7d2023-09-29 16:51:45 +02001268 bdrv_graph_rdunlock_main_loop();
Stefan Hajnoczi6bc30f12023-12-05 13:20:02 -05001269 bdrv_graph_wrlock();
Kevin Wolf24bc15d2015-06-15 13:50:20 +02001270 bdrv_unref_child(bs, extent_file);
Stefan Hajnoczi6bc30f12023-12-05 13:20:02 -05001271 bdrv_graph_wrunlock();
Kevin Wolfb7cfc7d2023-09-29 16:51:45 +02001272 bdrv_graph_rdlock_main_loop();
Max Reitzcdc0dd22019-08-15 17:36:33 +02001273 goto out;
Fam Zheng86c6b422011-08-12 23:19:28 +08001274 }
Fam Zhengf4c129a2013-10-31 10:06:23 +08001275 extent = &s->extents[s->num_extents - 1];
Sam Eiderman98eb9732019-06-20 12:10:57 +03001276 } else if (!strcmp(type, "SESPARSE")) {
1277 ret = vmdk_open_se_sparse(bs, extent_file, bs->open_flags, errp);
1278 if (ret) {
Kevin Wolfb7cfc7d2023-09-29 16:51:45 +02001279 bdrv_graph_rdunlock_main_loop();
Stefan Hajnoczi6bc30f12023-12-05 13:20:02 -05001280 bdrv_graph_wrlock();
Sam Eiderman98eb9732019-06-20 12:10:57 +03001281 bdrv_unref_child(bs, extent_file);
Stefan Hajnoczi6bc30f12023-12-05 13:20:02 -05001282 bdrv_graph_wrunlock();
Kevin Wolfb7cfc7d2023-09-29 16:51:45 +02001283 bdrv_graph_rdlock_main_loop();
Max Reitzcdc0dd22019-08-15 17:36:33 +02001284 goto out;
Sam Eiderman98eb9732019-06-20 12:10:57 +03001285 }
1286 extent = &s->extents[s->num_extents - 1];
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001287 } else {
Fam Zheng48239702013-10-11 15:43:22 +08001288 error_setg(errp, "Unsupported extent type '%s'", type);
Kevin Wolfb7cfc7d2023-09-29 16:51:45 +02001289 bdrv_graph_rdunlock_main_loop();
Stefan Hajnoczi6bc30f12023-12-05 13:20:02 -05001290 bdrv_graph_wrlock();
Kevin Wolf24bc15d2015-06-15 13:50:20 +02001291 bdrv_unref_child(bs, extent_file);
Stefan Hajnoczi6bc30f12023-12-05 13:20:02 -05001292 bdrv_graph_wrunlock();
Kevin Wolfb7cfc7d2023-09-29 16:51:45 +02001293 bdrv_graph_rdlock_main_loop();
Max Reitzcdc0dd22019-08-15 17:36:33 +02001294 ret = -ENOTSUP;
1295 goto out;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001296 }
Fam Zhengf4c129a2013-10-31 10:06:23 +08001297 extent->type = g_strdup(type);
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001298 }
Max Reitzcdc0dd22019-08-15 17:36:33 +02001299
1300 ret = 0;
1301 goto out;
Markus Armbrusterd28d7372015-12-18 16:35:21 +01001302
1303invalid:
1304 np = next_line(p);
1305 assert(np != p);
1306 if (np[-1] == '\n') {
1307 np--;
1308 }
1309 error_setg(errp, "Invalid extent line: %.*s", (int)(np - p), p);
Max Reitzcdc0dd22019-08-15 17:36:33 +02001310 ret = -EINVAL;
1311
1312out:
1313 g_free(desc_file_dir);
1314 return ret;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001315}
1316
Kevin Wolfb7cfc7d2023-09-29 16:51:45 +02001317static int GRAPH_RDLOCK
1318vmdk_open_desc_file(BlockDriverState *bs, int flags, char *buf, QDict *options,
1319 Error **errp)
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001320{
1321 int ret;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001322 char ct[128];
1323 BDRVVmdkState *s = bs->opaque;
1324
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001325 if (vmdk_parse_description(buf, "createType", ct, sizeof(ct))) {
Paolo Bonzini76abe402014-02-17 14:44:06 +01001326 error_setg(errp, "invalid VMDK image descriptor");
1327 ret = -EINVAL;
Evgeny Budilovsky0bed0872013-06-12 14:06:30 +03001328 goto exit;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001329 }
Fam Zheng6398de52011-08-12 23:19:27 +08001330 if (strcmp(ct, "monolithicFlat") &&
Paolo Bonzini04d542c2013-08-19 18:54:28 +08001331 strcmp(ct, "vmfs") &&
Fam Zhengdaac8fd2013-08-19 18:54:27 +08001332 strcmp(ct, "vmfsSparse") &&
Sam Eiderman98eb9732019-06-20 12:10:57 +03001333 strcmp(ct, "seSparse") &&
Fam Zheng86c6b422011-08-12 23:19:28 +08001334 strcmp(ct, "twoGbMaxExtentSparse") &&
Fam Zheng6398de52011-08-12 23:19:27 +08001335 strcmp(ct, "twoGbMaxExtentFlat")) {
Fam Zheng48239702013-10-11 15:43:22 +08001336 error_setg(errp, "Unsupported image type '%s'", ct);
Evgeny Budilovsky0bed0872013-06-12 14:06:30 +03001337 ret = -ENOTSUP;
1338 goto exit;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001339 }
Fam Zhengf4c129a2013-10-31 10:06:23 +08001340 s->create_type = g_strdup(ct);
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001341 s->desc_offset = 0;
Max Reitzcdc0dd22019-08-15 17:36:33 +02001342 ret = vmdk_parse_extents(buf, bs, options, errp);
Evgeny Budilovsky0bed0872013-06-12 14:06:30 +03001343exit:
Evgeny Budilovsky0bed0872013-06-12 14:06:30 +03001344 return ret;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001345}
1346
Max Reitz015a1032013-09-05 14:22:29 +02001347static int vmdk_open(BlockDriverState *bs, QDict *options, int flags,
1348 Error **errp)
Fam Zhengb4b3ab12011-07-12 19:56:31 +08001349{
Fam Zheng9aeecbb2014-12-04 07:28:33 +08001350 char *buf;
Fam Zheng86c6b422011-08-12 23:19:28 +08001351 int ret;
1352 BDRVVmdkState *s = bs->opaque;
Paolo Bonzini37f09e52014-02-17 14:44:04 +01001353 uint32_t magic;
Fam Zhengb4b3ab12011-07-12 19:56:31 +08001354
Kevin Wolf3804e3c2023-09-11 11:46:12 +02001355 GRAPH_RDLOCK_GUARD_MAINLOOP();
1356
Vladimir Sementsov-Ogievskiy83930782022-07-26 23:11:21 +03001357 ret = bdrv_open_file_child(NULL, options, "file", bs, errp);
1358 if (ret < 0) {
1359 return ret;
Kevin Wolf4e4bf5c2016-12-16 18:52:37 +01001360 }
1361
Kevin Wolfcf2ab8f2016-06-20 18:24:02 +02001362 buf = vmdk_read_desc(bs->file, 0, errp);
Paolo Bonzinid1833ef2014-02-17 14:44:03 +01001363 if (!buf) {
1364 return -EINVAL;
1365 }
1366
Paolo Bonzini37f09e52014-02-17 14:44:04 +01001367 magic = ldl_be_p(buf);
1368 switch (magic) {
1369 case VMDK3_MAGIC:
1370 case VMDK4_MAGIC:
Kevin Wolf9a4f4c32015-06-16 14:19:22 +02001371 ret = vmdk_open_sparse(bs, bs->file, flags, buf, options,
Kevin Wolf24bc15d2015-06-15 13:50:20 +02001372 errp);
Paolo Bonzini37f09e52014-02-17 14:44:04 +01001373 s->desc_offset = 0x200;
1374 break;
1375 default:
Max Reitz8b1869d2020-05-13 13:05:35 +02001376 /* No data in the descriptor file */
1377 bs->file->role &= ~BDRV_CHILD_DATA;
1378
1379 /* Must succeed because we have given up permissions if anything */
1380 bdrv_child_refresh_perms(bs, bs->file, &error_abort);
1381
Kevin Wolfa6468362015-04-07 15:35:59 +02001382 ret = vmdk_open_desc_file(bs, flags, buf, options, errp);
Paolo Bonzini37f09e52014-02-17 14:44:04 +01001383 break;
Fam Zhengb4b3ab12011-07-12 19:56:31 +08001384 }
Paolo Bonzini37f09e52014-02-17 14:44:04 +01001385 if (ret) {
1386 goto fail;
1387 }
1388
Paolo Bonzinibae0a0c2011-10-20 13:16:20 +02001389 /* try to open parent images, if exist */
1390 ret = vmdk_parent_open(bs);
1391 if (ret) {
1392 goto fail;
1393 }
Peter Maydell98778602017-07-09 18:06:14 +01001394 ret = vmdk_read_cid(bs, 0, &s->cid);
1395 if (ret) {
1396 goto fail;
1397 }
1398 ret = vmdk_read_cid(bs, 1, &s->parent_cid);
1399 if (ret) {
1400 goto fail;
1401 }
Paolo Bonzini848c66e2011-10-20 13:16:21 +02001402 qemu_co_mutex_init(&s->lock);
Kevin Wolf2bc31662011-11-22 16:50:27 +01001403
1404 /* Disable migration when VMDK images are used */
Alberto Garcia81e5f782015-04-08 12:29:19 +03001405 error_setg(&s->migration_blocker, "The vmdk format used by node '%s' "
1406 "does not support live migration",
1407 bdrv_get_device_or_node_name(bs));
Steve Sistaree0ee3a82023-10-25 12:44:26 -07001408 ret = migrate_add_blocker_normal(&s->migration_blocker, errp);
Markus Armbruster386f6c02020-07-07 18:06:10 +02001409 if (ret < 0) {
Ashijeet Acharyafe44dc92017-01-16 17:01:53 +05301410 goto fail;
1411 }
1412
Paolo Bonzinid1833ef2014-02-17 14:44:03 +01001413 g_free(buf);
Kevin Wolf2bc31662011-11-22 16:50:27 +01001414 return 0;
Paolo Bonzinibae0a0c2011-10-20 13:16:20 +02001415
1416fail:
Paolo Bonzinid1833ef2014-02-17 14:44:03 +01001417 g_free(buf);
Fam Zhengf4c129a2013-10-31 10:06:23 +08001418 g_free(s->create_type);
1419 s->create_type = NULL;
Paolo Bonzinibae0a0c2011-10-20 13:16:20 +02001420 vmdk_free_extents(bs);
1421 return ret;
bellardea2384d2004-08-01 21:59:26 +00001422}
1423
Kevin Wolfd34682c2013-12-11 19:26:16 +01001424
Kevin Wolf3baca892014-07-16 17:48:16 +02001425static void vmdk_refresh_limits(BlockDriverState *bs, Error **errp)
Kevin Wolfd34682c2013-12-11 19:26:16 +01001426{
1427 BDRVVmdkState *s = bs->opaque;
1428 int i;
1429
1430 for (i = 0; i < s->num_extents; i++) {
1431 if (!s->extents[i].flat) {
Eric Blakecf081fc2016-06-01 15:10:02 -06001432 bs->bl.pwrite_zeroes_alignment =
1433 MAX(bs->bl.pwrite_zeroes_alignment,
1434 s->extents[i].cluster_sectors << BDRV_SECTOR_BITS);
Kevin Wolfd34682c2013-12-11 19:26:16 +01001435 }
1436 }
Kevin Wolfd34682c2013-12-11 19:26:16 +01001437}
1438
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001439/**
1440 * get_whole_cluster
1441 *
1442 * Copy backing file's cluster that covers @sector_num, otherwise write zero,
Kevin Wolf4823cde2020-04-30 15:30:04 +02001443 * to the cluster at @cluster_sector_num. If @zeroed is true, we're overwriting
1444 * a zeroed cluster in the current layer and must not copy data from the
1445 * backing file.
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001446 *
1447 * If @skip_start_sector < @skip_end_sector, the relative range
1448 * [@skip_start_sector, @skip_end_sector) is not copied or written, and leave
1449 * it for call to write user data in the request.
1450 */
Kevin Wolfb9b10c32023-02-03 16:21:50 +01001451static int coroutine_fn GRAPH_RDLOCK
1452get_whole_cluster(BlockDriverState *bs, VmdkExtent *extent,
1453 uint64_t cluster_offset, uint64_t offset,
1454 uint64_t skip_start_bytes, uint64_t skip_end_bytes,
1455 bool zeroed)
ths5f4da8c2007-01-24 21:05:24 +00001456{
Fam Zhengbf815072013-08-06 15:44:54 +08001457 int ret = VMDK_OK;
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001458 int64_t cluster_bytes;
1459 uint8_t *whole_grain;
Kevin Wolf4823cde2020-04-30 15:30:04 +02001460 bool copy_from_backing;
ths5f4da8c2007-01-24 21:05:24 +00001461
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001462 /* For COW, align request sector_num to cluster start */
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001463 cluster_bytes = extent->cluster_sectors << BDRV_SECTOR_BITS;
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02001464 offset = QEMU_ALIGN_DOWN(offset, cluster_bytes);
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001465 whole_grain = qemu_blockalign(bs, cluster_bytes);
Kevin Wolf4823cde2020-04-30 15:30:04 +02001466 copy_from_backing = bs->backing && !zeroed;
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001467
Kevin Wolf4823cde2020-04-30 15:30:04 +02001468 if (!copy_from_backing) {
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02001469 memset(whole_grain, 0, skip_start_bytes);
1470 memset(whole_grain + skip_end_bytes, 0, cluster_bytes - skip_end_bytes);
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001471 }
1472
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02001473 assert(skip_end_bytes <= cluster_bytes);
Fam Zheng0e69c542011-07-12 19:56:29 +08001474 /* we will be here if it's first write on non-exist grain(cluster).
1475 * try to read from parent image, if exist */
Kevin Wolf760e0062015-06-17 14:55:21 +02001476 if (bs->backing && !vmdk_is_cid_valid(bs)) {
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001477 ret = VMDK_ERROR;
1478 goto exit;
1479 }
ths5f4da8c2007-01-24 21:05:24 +00001480
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001481 /* Read backing data before skip range */
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02001482 if (skip_start_bytes > 0) {
Kevin Wolf4823cde2020-04-30 15:30:04 +02001483 if (copy_from_backing) {
Max Reitz23c4b2a2017-11-23 03:08:19 +01001484 /* qcow2 emits this on bs->file instead of bs->backing */
Paolo Bonzini17362392023-06-01 13:51:45 +02001485 BLKDBG_CO_EVENT(extent->file, BLKDBG_COW_READ);
Alberto Fariaa5c4e5b2022-10-13 14:37:10 +02001486 ret = bdrv_co_pread(bs->backing, offset, skip_start_bytes,
1487 whole_grain, 0);
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001488 if (ret < 0) {
1489 ret = VMDK_ERROR;
1490 goto exit;
1491 }
Kevin Wolfc3365002010-04-16 19:28:14 +02001492 }
Paolo Bonzini17362392023-06-01 13:51:45 +02001493 BLKDBG_CO_EVENT(extent->file, BLKDBG_COW_WRITE);
Alberto Fariaa5c4e5b2022-10-13 14:37:10 +02001494 ret = bdrv_co_pwrite(extent->file, cluster_offset, skip_start_bytes,
1495 whole_grain, 0);
Kevin Wolfc3365002010-04-16 19:28:14 +02001496 if (ret < 0) {
Fam Zhengbf815072013-08-06 15:44:54 +08001497 ret = VMDK_ERROR;
1498 goto exit;
ths630530a2007-06-18 15:01:30 +00001499 }
ths5f4da8c2007-01-24 21:05:24 +00001500 }
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001501 /* Read backing data after skip range */
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02001502 if (skip_end_bytes < cluster_bytes) {
Kevin Wolf4823cde2020-04-30 15:30:04 +02001503 if (copy_from_backing) {
Max Reitz23c4b2a2017-11-23 03:08:19 +01001504 /* qcow2 emits this on bs->file instead of bs->backing */
Paolo Bonzini17362392023-06-01 13:51:45 +02001505 BLKDBG_CO_EVENT(extent->file, BLKDBG_COW_READ);
Alberto Fariaa5c4e5b2022-10-13 14:37:10 +02001506 ret = bdrv_co_pread(bs->backing, offset + skip_end_bytes,
1507 cluster_bytes - skip_end_bytes,
1508 whole_grain + skip_end_bytes, 0);
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001509 if (ret < 0) {
1510 ret = VMDK_ERROR;
1511 goto exit;
1512 }
1513 }
Paolo Bonzini17362392023-06-01 13:51:45 +02001514 BLKDBG_CO_EVENT(extent->file, BLKDBG_COW_WRITE);
Alberto Fariaa5c4e5b2022-10-13 14:37:10 +02001515 ret = bdrv_co_pwrite(extent->file, cluster_offset + skip_end_bytes,
1516 cluster_bytes - skip_end_bytes,
1517 whole_grain + skip_end_bytes, 0);
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001518 if (ret < 0) {
1519 ret = VMDK_ERROR;
1520 goto exit;
1521 }
1522 }
1523
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02001524 ret = VMDK_OK;
Fam Zhengbf815072013-08-06 15:44:54 +08001525exit:
1526 qemu_vfree(whole_grain);
1527 return ret;
ths5f4da8c2007-01-24 21:05:24 +00001528}
1529
Emanuele Giuseppe Esposito88095342023-02-03 16:21:46 +01001530static int coroutine_fn GRAPH_RDLOCK
1531vmdk_L2update(VmdkExtent *extent, VmdkMetaData *m_data, uint32_t offset)
ths630530a2007-06-18 15:01:30 +00001532{
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001533 offset = cpu_to_le32(offset);
ths630530a2007-06-18 15:01:30 +00001534 /* update L2 table */
Paolo Bonzini17362392023-06-01 13:51:45 +02001535 BLKDBG_CO_EVENT(extent->file, BLKDBG_L2_UPDATE);
Alberto Fariaa5c4e5b2022-10-13 14:37:10 +02001536 if (bdrv_co_pwrite(extent->file,
1537 ((int64_t)m_data->l2_offset * 512)
1538 + (m_data->l2_index * sizeof(offset)),
1539 sizeof(offset), &offset, 0) < 0) {
Fam Zheng65f74722013-05-02 10:25:22 +08001540 return VMDK_ERROR;
Fam Zhengb3976d32011-07-12 19:56:28 +08001541 }
ths630530a2007-06-18 15:01:30 +00001542 /* update backup L2 table */
Fam Zhengb3976d32011-07-12 19:56:28 +08001543 if (extent->l1_backup_table_offset != 0) {
1544 m_data->l2_offset = extent->l1_backup_table[m_data->l1_index];
Alberto Fariaa5c4e5b2022-10-13 14:37:10 +02001545 if (bdrv_co_pwrite(extent->file,
1546 ((int64_t)m_data->l2_offset * 512)
1547 + (m_data->l2_index * sizeof(offset)),
1548 sizeof(offset), &offset, 0) < 0) {
Fam Zheng65f74722013-05-02 10:25:22 +08001549 return VMDK_ERROR;
Fam Zhengb3976d32011-07-12 19:56:28 +08001550 }
ths630530a2007-06-18 15:01:30 +00001551 }
Alberto Fariaa5c4e5b2022-10-13 14:37:10 +02001552 if (bdrv_co_flush(extent->file->bs) < 0) {
Kevin Wolf2758be02020-04-30 15:30:06 +02001553 return VMDK_ERROR;
1554 }
Fam Zhengcdeaf1f2013-05-02 10:25:27 +08001555 if (m_data->l2_cache_entry) {
1556 *m_data->l2_cache_entry = offset;
1557 }
ths630530a2007-06-18 15:01:30 +00001558
Fam Zheng65f74722013-05-02 10:25:22 +08001559 return VMDK_OK;
ths630530a2007-06-18 15:01:30 +00001560}
1561
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001562/**
1563 * get_cluster_offset
1564 *
1565 * Look up cluster offset in extent file by sector number, and store in
1566 * @cluster_offset.
1567 *
1568 * For flat extents, the start offset as parsed from the description file is
1569 * returned.
1570 *
1571 * For sparse extents, look up in L1, L2 table. If allocate is true, return an
1572 * offset for a new cluster and update L2 cache. If there is a backing file,
1573 * COW is done before returning; otherwise, zeroes are written to the allocated
1574 * cluster. Both COW and zero writing skips the sector range
1575 * [@skip_start_sector, @skip_end_sector) passed in by caller, because caller
1576 * has new data to write there.
1577 *
1578 * Returns: VMDK_OK if cluster exists and mapped in the image.
1579 * VMDK_UNALLOC if cluster is not mapped and @allocate is false.
1580 * VMDK_ERROR if failed.
1581 */
Kevin Wolfb9b10c32023-02-03 16:21:50 +01001582static int coroutine_fn GRAPH_RDLOCK
1583get_cluster_offset(BlockDriverState *bs, VmdkExtent *extent,
1584 VmdkMetaData *m_data, uint64_t offset, bool allocate,
1585 uint64_t *cluster_offset, uint64_t skip_start_bytes,
1586 uint64_t skip_end_bytes)
bellardea2384d2004-08-01 21:59:26 +00001587{
bellardea2384d2004-08-01 21:59:26 +00001588 unsigned int l1_index, l2_offset, l2_index;
1589 int min_index, i, j;
Sam Eiderman98eb9732019-06-20 12:10:57 +03001590 uint32_t min_count;
1591 void *l2_table;
Fam Zheng14ead642013-05-02 10:25:23 +08001592 bool zeroed = false;
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001593 int64_t ret;
Fam Zhengd1319b02014-09-23 09:56:21 +08001594 int64_t cluster_sector;
Sam Eiderman98eb9732019-06-20 12:10:57 +03001595 unsigned int l2_size_bytes = extent->l2_size * extent->entry_size;
ths630530a2007-06-18 15:01:30 +00001596
Fam Zhengae261c82011-07-12 19:56:38 +08001597 if (m_data) {
Kevin Wolf4dc20e62020-04-30 15:30:02 +02001598 m_data->new_allocation = false;
Fam Zhengae261c82011-07-12 19:56:38 +08001599 }
Fam Zheng91b85bd2011-07-12 19:56:35 +08001600 if (extent->flat) {
Fam Zheng7fa60fa2011-07-19 08:38:22 +08001601 *cluster_offset = extent->flat_start_offset;
Fam Zheng65f74722013-05-02 10:25:22 +08001602 return VMDK_OK;
Fam Zheng91b85bd2011-07-12 19:56:35 +08001603 }
ths630530a2007-06-18 15:01:30 +00001604
Fam Zheng6398de52011-08-12 23:19:27 +08001605 offset -= (extent->end_sector - extent->sectors) * SECTOR_SIZE;
Fam Zhengb3976d32011-07-12 19:56:28 +08001606 l1_index = (offset >> 9) / extent->l1_entry_sectors;
1607 if (l1_index >= extent->l1_size) {
Fam Zheng65f74722013-05-02 10:25:22 +08001608 return VMDK_ERROR;
Fam Zhengb3976d32011-07-12 19:56:28 +08001609 }
Sam Eiderman98eb9732019-06-20 12:10:57 +03001610 if (extent->sesparse) {
1611 uint64_t l2_offset_u64;
1612
1613 assert(extent->entry_size == sizeof(uint64_t));
1614
1615 l2_offset_u64 = ((uint64_t *)extent->l1_table)[l1_index];
1616 if (l2_offset_u64 == 0) {
1617 l2_offset = 0;
1618 } else if ((l2_offset_u64 & 0xffffffff00000000) != 0x1000000000000000) {
1619 /*
1620 * Top most nibble is 0x1 if grain table is allocated.
1621 * strict check - top most 4 bytes must be 0x10000000 since max
1622 * supported size is 64TB for disk - so no more than 64TB / 16MB
1623 * grain directories which is smaller than uint32,
1624 * where 16MB is the only supported default grain table coverage.
1625 */
1626 return VMDK_ERROR;
1627 } else {
1628 l2_offset_u64 = l2_offset_u64 & 0x00000000ffffffff;
1629 l2_offset_u64 = extent->sesparse_l2_tables_offset +
1630 l2_offset_u64 * l2_size_bytes / SECTOR_SIZE;
1631 if (l2_offset_u64 > 0x00000000ffffffff) {
1632 return VMDK_ERROR;
1633 }
1634 l2_offset = (unsigned int)(l2_offset_u64);
1635 }
1636 } else {
1637 assert(extent->entry_size == sizeof(uint32_t));
1638 l2_offset = ((uint32_t *)extent->l1_table)[l1_index];
1639 }
Fam Zhengb3976d32011-07-12 19:56:28 +08001640 if (!l2_offset) {
Fam Zheng65f74722013-05-02 10:25:22 +08001641 return VMDK_UNALLOC;
Fam Zhengb3976d32011-07-12 19:56:28 +08001642 }
Fam Zhengb4b3ab12011-07-12 19:56:31 +08001643 for (i = 0; i < L2_CACHE_SIZE; i++) {
Fam Zhengb3976d32011-07-12 19:56:28 +08001644 if (l2_offset == extent->l2_cache_offsets[i]) {
bellardea2384d2004-08-01 21:59:26 +00001645 /* increment the hit count */
Fam Zhengb3976d32011-07-12 19:56:28 +08001646 if (++extent->l2_cache_counts[i] == 0xffffffff) {
Fam Zhengb4b3ab12011-07-12 19:56:31 +08001647 for (j = 0; j < L2_CACHE_SIZE; j++) {
Fam Zhengb3976d32011-07-12 19:56:28 +08001648 extent->l2_cache_counts[j] >>= 1;
bellardea2384d2004-08-01 21:59:26 +00001649 }
1650 }
Sam Eiderman98eb9732019-06-20 12:10:57 +03001651 l2_table = (char *)extent->l2_cache + (i * l2_size_bytes);
bellardea2384d2004-08-01 21:59:26 +00001652 goto found;
1653 }
1654 }
1655 /* not found: load a new entry in the least used one */
1656 min_index = 0;
1657 min_count = 0xffffffff;
Fam Zhengb4b3ab12011-07-12 19:56:31 +08001658 for (i = 0; i < L2_CACHE_SIZE; i++) {
Fam Zhengb3976d32011-07-12 19:56:28 +08001659 if (extent->l2_cache_counts[i] < min_count) {
1660 min_count = extent->l2_cache_counts[i];
bellardea2384d2004-08-01 21:59:26 +00001661 min_index = i;
1662 }
1663 }
Sam Eiderman98eb9732019-06-20 12:10:57 +03001664 l2_table = (char *)extent->l2_cache + (min_index * l2_size_bytes);
Paolo Bonzini17362392023-06-01 13:51:45 +02001665 BLKDBG_CO_EVENT(extent->file, BLKDBG_L2_LOAD);
Alberto Fariaa5c4e5b2022-10-13 14:37:10 +02001666 if (bdrv_co_pread(extent->file,
Fam Zhengb3976d32011-07-12 19:56:28 +08001667 (int64_t)l2_offset * 512,
Alberto Faria53fb7842022-06-09 16:27:35 +01001668 l2_size_bytes,
Alberto Fariaa5c4e5b2022-10-13 14:37:10 +02001669 l2_table, 0
Alberto Faria353a5d82022-06-09 16:27:37 +01001670 ) < 0) {
Fam Zheng65f74722013-05-02 10:25:22 +08001671 return VMDK_ERROR;
Fam Zhengb3976d32011-07-12 19:56:28 +08001672 }
ths5f4da8c2007-01-24 21:05:24 +00001673
Fam Zhengb3976d32011-07-12 19:56:28 +08001674 extent->l2_cache_offsets[min_index] = l2_offset;
1675 extent->l2_cache_counts[min_index] = 1;
bellardea2384d2004-08-01 21:59:26 +00001676 found:
Fam Zhengb3976d32011-07-12 19:56:28 +08001677 l2_index = ((offset >> 9) / extent->cluster_sectors) % extent->l2_size;
Kevin Wolf2821c1c2020-04-30 15:30:03 +02001678 if (m_data) {
1679 m_data->l1_index = l1_index;
1680 m_data->l2_index = l2_index;
1681 m_data->l2_offset = l2_offset;
1682 m_data->l2_cache_entry = ((uint32_t *)l2_table) + l2_index;
1683 }
ths630530a2007-06-18 15:01:30 +00001684
Sam Eiderman98eb9732019-06-20 12:10:57 +03001685 if (extent->sesparse) {
1686 cluster_sector = le64_to_cpu(((uint64_t *)l2_table)[l2_index]);
1687 switch (cluster_sector & 0xf000000000000000) {
1688 case 0x0000000000000000:
1689 /* unallocated grain */
1690 if (cluster_sector != 0) {
1691 return VMDK_ERROR;
1692 }
1693 break;
1694 case 0x1000000000000000:
1695 /* scsi-unmapped grain - fallthrough */
1696 case 0x2000000000000000:
1697 /* zero grain */
1698 zeroed = true;
1699 break;
1700 case 0x3000000000000000:
1701 /* allocated grain */
1702 cluster_sector = (((cluster_sector & 0x0fff000000000000) >> 48) |
1703 ((cluster_sector & 0x0000ffffffffffff) << 12));
1704 cluster_sector = extent->sesparse_clusters_offset +
1705 cluster_sector * extent->cluster_sectors;
1706 break;
1707 default:
1708 return VMDK_ERROR;
1709 }
1710 } else {
1711 cluster_sector = le32_to_cpu(((uint32_t *)l2_table)[l2_index]);
1712
1713 if (extent->has_zero_grain && cluster_sector == VMDK_GTE_ZEROED) {
1714 zeroed = true;
1715 }
Fam Zheng14ead642013-05-02 10:25:23 +08001716 }
1717
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001718 if (!cluster_sector || zeroed) {
Fam Zheng91b85bd2011-07-12 19:56:35 +08001719 if (!allocate) {
Fam Zheng14ead642013-05-02 10:25:23 +08001720 return zeroed ? VMDK_ZEROED : VMDK_UNALLOC;
Fam Zheng91b85bd2011-07-12 19:56:35 +08001721 }
Sam Eiderman98eb9732019-06-20 12:10:57 +03001722 assert(!extent->sesparse);
ths5f4da8c2007-01-24 21:05:24 +00001723
yuchenlina77672e2018-03-22 21:33:37 +08001724 if (extent->next_cluster_sector >= VMDK_EXTENT_MAX_SECTORS) {
1725 return VMDK_ERROR;
1726 }
1727
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001728 cluster_sector = extent->next_cluster_sector;
1729 extent->next_cluster_sector += extent->cluster_sectors;
Kevin Wolf9949f972010-04-16 21:07:19 +02001730
ths630530a2007-06-18 15:01:30 +00001731 /* First of all we write grain itself, to avoid race condition
1732 * that may to corrupt the image.
1733 * This problem may occur because of insufficient space on host disk
1734 * or inappropriate VM shutdown.
1735 */
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02001736 ret = get_whole_cluster(bs, extent, cluster_sector * BDRV_SECTOR_SIZE,
Kevin Wolf4823cde2020-04-30 15:30:04 +02001737 offset, skip_start_bytes, skip_end_bytes,
1738 zeroed);
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001739 if (ret) {
1740 return ret;
ths630530a2007-06-18 15:01:30 +00001741 }
Reda Sallahi524089b2016-07-07 10:42:49 +02001742 if (m_data) {
Kevin Wolf4dc20e62020-04-30 15:30:02 +02001743 m_data->new_allocation = true;
Reda Sallahi524089b2016-07-07 10:42:49 +02001744 }
bellardff1afc72005-04-26 21:08:00 +00001745 }
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001746 *cluster_offset = cluster_sector << BDRV_SECTOR_BITS;
Fam Zheng65f74722013-05-02 10:25:22 +08001747 return VMDK_OK;
bellardea2384d2004-08-01 21:59:26 +00001748}
1749
Fam Zhengb3976d32011-07-12 19:56:28 +08001750static VmdkExtent *find_extent(BDRVVmdkState *s,
1751 int64_t sector_num, VmdkExtent *start_hint)
1752{
1753 VmdkExtent *extent = start_hint;
1754
1755 if (!extent) {
1756 extent = &s->extents[0];
1757 }
1758 while (extent < &s->extents[s->num_extents]) {
1759 if (sector_num < extent->end_sector) {
1760 return extent;
1761 }
1762 extent++;
1763 }
1764 return NULL;
1765}
1766
Kevin Wolfa844a2b2016-04-25 17:14:48 +02001767static inline uint64_t vmdk_find_offset_in_cluster(VmdkExtent *extent,
1768 int64_t offset)
1769{
Eduardo Habkost9be38592016-06-13 18:57:58 -03001770 uint64_t extent_begin_offset, extent_relative_offset;
Kevin Wolfa844a2b2016-04-25 17:14:48 +02001771 uint64_t cluster_size = extent->cluster_sectors * BDRV_SECTOR_SIZE;
1772
1773 extent_begin_offset =
1774 (extent->end_sector - extent->sectors) * BDRV_SECTOR_SIZE;
1775 extent_relative_offset = offset - extent_begin_offset;
Eduardo Habkost9be38592016-06-13 18:57:58 -03001776 return extent_relative_offset % cluster_size;
Kevin Wolfa844a2b2016-04-25 17:14:48 +02001777}
1778
Kevin Wolfb9b10c32023-02-03 16:21:50 +01001779static int coroutine_fn GRAPH_RDLOCK
1780vmdk_co_block_status(BlockDriverState *bs, bool want_zero,
1781 int64_t offset, int64_t bytes, int64_t *pnum,
1782 int64_t *map, BlockDriverState **file)
bellardea2384d2004-08-01 21:59:26 +00001783{
1784 BDRVVmdkState *s = bs->opaque;
Fam Zhengb3976d32011-07-12 19:56:28 +08001785 int64_t index_in_cluster, n, ret;
Eric Blakec72080b2018-02-13 14:26:58 -06001786 uint64_t cluster_offset;
Fam Zhengb3976d32011-07-12 19:56:28 +08001787 VmdkExtent *extent;
1788
Eric Blakec72080b2018-02-13 14:26:58 -06001789 extent = find_extent(s, offset >> BDRV_SECTOR_BITS, NULL);
Fam Zhengb3976d32011-07-12 19:56:28 +08001790 if (!extent) {
Eric Blakec72080b2018-02-13 14:26:58 -06001791 return -EIO;
Fam Zhengb3976d32011-07-12 19:56:28 +08001792 }
Stefan Hajnoczif8a2e5e2011-11-14 12:44:21 +00001793 qemu_co_mutex_lock(&s->lock);
Eric Blakec72080b2018-02-13 14:26:58 -06001794 ret = get_cluster_offset(bs, extent, NULL, offset, false, &cluster_offset,
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001795 0, 0);
Stefan Hajnoczif8a2e5e2011-11-14 12:44:21 +00001796 qemu_co_mutex_unlock(&s->lock);
Fam Zheng14ead642013-05-02 10:25:23 +08001797
Eric Blakec72080b2018-02-13 14:26:58 -06001798 index_in_cluster = vmdk_find_offset_in_cluster(extent, offset);
Paolo Bonzini4bc74be2013-09-04 19:00:30 +02001799 switch (ret) {
1800 case VMDK_ERROR:
1801 ret = -EIO;
1802 break;
1803 case VMDK_UNALLOC:
1804 ret = 0;
1805 break;
1806 case VMDK_ZEROED:
1807 ret = BDRV_BLOCK_ZERO;
1808 break;
1809 case VMDK_OK:
1810 ret = BDRV_BLOCK_DATA;
Fam Zhenge0f100f2016-01-26 11:58:59 +08001811 if (!extent->compressed) {
Fam Zhengd0a18f12016-01-26 11:58:58 +08001812 ret |= BDRV_BLOCK_OFFSET_VALID;
Eric Blakec72080b2018-02-13 14:26:58 -06001813 *map = cluster_offset + index_in_cluster;
Max Reitz4dd84ac2019-07-25 17:55:11 +02001814 if (extent->flat) {
1815 ret |= BDRV_BLOCK_RECURSE;
1816 }
Andrey Drobyshev via28482892023-09-08 00:02:25 +03001817 } else {
1818 ret |= BDRV_BLOCK_COMPRESSED;
Paolo Bonzini4bc74be2013-09-04 19:00:30 +02001819 }
Fam Zhenge0f100f2016-01-26 11:58:59 +08001820 *file = extent->file->bs;
Paolo Bonzini4bc74be2013-09-04 19:00:30 +02001821 break;
1822 }
Fam Zheng91b85bd2011-07-12 19:56:35 +08001823
Eric Blakec72080b2018-02-13 14:26:58 -06001824 n = extent->cluster_sectors * BDRV_SECTOR_SIZE - index_in_cluster;
1825 *pnum = MIN(n, bytes);
Fam Zhengb3976d32011-07-12 19:56:28 +08001826 return ret;
bellardea2384d2004-08-01 21:59:26 +00001827}
1828
Kevin Wolfb9b10c32023-02-03 16:21:50 +01001829static int coroutine_fn GRAPH_RDLOCK
Paolo Bonzinib4df9902022-09-22 10:49:18 +02001830vmdk_write_extent(VmdkExtent *extent, int64_t cluster_offset,
1831 int64_t offset_in_cluster, QEMUIOVector *qiov,
1832 uint64_t qiov_offset, uint64_t n_bytes,
1833 uint64_t offset)
Fam Zhengdd3f6ee2011-08-12 23:19:29 +08001834{
1835 int ret;
Fam Zheng2b2c8c52011-08-12 23:19:31 +08001836 VmdkGrainMarker *data = NULL;
1837 uLongf buf_len;
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02001838 QEMUIOVector local_qiov;
Fam Zheng5e82a312015-05-06 20:23:46 +08001839 int64_t write_offset;
1840 int64_t write_end_sector;
Fam Zhengdd3f6ee2011-08-12 23:19:29 +08001841
Fam Zheng2b2c8c52011-08-12 23:19:31 +08001842 if (extent->compressed) {
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02001843 void *compressed_data;
1844
Max Reitzbedb8bb2019-08-15 17:36:35 +02001845 /* Only whole clusters */
1846 if (offset_in_cluster ||
1847 n_bytes > (extent->cluster_sectors * SECTOR_SIZE) ||
1848 (n_bytes < (extent->cluster_sectors * SECTOR_SIZE) &&
1849 offset + n_bytes != extent->end_sector * SECTOR_SIZE))
1850 {
1851 ret = -EINVAL;
1852 goto out;
1853 }
1854
Fam Zheng2b2c8c52011-08-12 23:19:31 +08001855 if (!extent->has_marker) {
1856 ret = -EINVAL;
1857 goto out;
1858 }
1859 buf_len = (extent->cluster_sectors << 9) * 2;
1860 data = g_malloc(buf_len + sizeof(VmdkGrainMarker));
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02001861
1862 compressed_data = g_malloc(n_bytes);
1863 qemu_iovec_to_buf(qiov, qiov_offset, compressed_data, n_bytes);
1864 ret = compress(data->data, &buf_len, compressed_data, n_bytes);
1865 g_free(compressed_data);
1866
1867 if (ret != Z_OK || buf_len == 0) {
Fam Zheng2b2c8c52011-08-12 23:19:31 +08001868 ret = -EINVAL;
1869 goto out;
1870 }
Fam Zheng5e82a312015-05-06 20:23:46 +08001871
QingFeng Hao4545d4f2016-12-16 06:20:40 +01001872 data->lba = cpu_to_le64(offset >> BDRV_SECTOR_BITS);
1873 data->size = cpu_to_le32(buf_len);
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02001874
1875 n_bytes = buf_len + sizeof(VmdkGrainMarker);
Vladimir Sementsov-Ogievskiy199d95b2019-02-18 17:09:20 +03001876 qemu_iovec_init_buf(&local_qiov, data, n_bytes);
Max Reitz23c4b2a2017-11-23 03:08:19 +01001877
Paolo Bonzini17362392023-06-01 13:51:45 +02001878 BLKDBG_CO_EVENT(extent->file, BLKDBG_WRITE_COMPRESSED);
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02001879 } else {
1880 qemu_iovec_init(&local_qiov, qiov->niov);
1881 qemu_iovec_concat(&local_qiov, qiov, qiov_offset, n_bytes);
Max Reitz23c4b2a2017-11-23 03:08:19 +01001882
Paolo Bonzini17362392023-06-01 13:51:45 +02001883 BLKDBG_CO_EVENT(extent->file, BLKDBG_WRITE_AIO);
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02001884 }
1885
Max Reitz3c363572017-11-23 03:08:16 +01001886 write_offset = cluster_offset + offset_in_cluster;
Kevin Wolfa03ef882016-06-20 21:31:46 +02001887 ret = bdrv_co_pwritev(extent->file, write_offset, n_bytes,
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02001888 &local_qiov, 0);
1889
1890 write_end_sector = DIV_ROUND_UP(write_offset + n_bytes, BDRV_SECTOR_SIZE);
Fam Zheng5e82a312015-05-06 20:23:46 +08001891
Radoslav Gerganov3efffc32015-09-10 10:53:14 +03001892 if (extent->compressed) {
1893 extent->next_cluster_sector = write_end_sector;
1894 } else {
1895 extent->next_cluster_sector = MAX(extent->next_cluster_sector,
1896 write_end_sector);
1897 }
Fam Zheng5e82a312015-05-06 20:23:46 +08001898
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02001899 if (ret < 0) {
Fam Zhengdd3f6ee2011-08-12 23:19:29 +08001900 goto out;
1901 }
1902 ret = 0;
1903 out:
Fam Zheng2b2c8c52011-08-12 23:19:31 +08001904 g_free(data);
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02001905 if (!extent->compressed) {
1906 qemu_iovec_destroy(&local_qiov);
1907 }
Fam Zhengdd3f6ee2011-08-12 23:19:29 +08001908 return ret;
1909}
1910
Kevin Wolfb9b10c32023-02-03 16:21:50 +01001911static int coroutine_fn GRAPH_RDLOCK
Paolo Bonzinib4df9902022-09-22 10:49:18 +02001912vmdk_read_extent(VmdkExtent *extent, int64_t cluster_offset,
Kevin Wolfb9b10c32023-02-03 16:21:50 +01001913 int64_t offset_in_cluster, QEMUIOVector *qiov, int bytes)
Fam Zhengdd3f6ee2011-08-12 23:19:29 +08001914{
1915 int ret;
Fam Zheng2b2c8c52011-08-12 23:19:31 +08001916 int cluster_bytes, buf_bytes;
1917 uint8_t *cluster_buf, *compressed_data;
1918 uint8_t *uncomp_buf;
1919 uint32_t data_len;
1920 VmdkGrainMarker *marker;
1921 uLongf buf_len;
Fam Zhengdd3f6ee2011-08-12 23:19:29 +08001922
Fam Zheng2b2c8c52011-08-12 23:19:31 +08001923
1924 if (!extent->compressed) {
Paolo Bonzini17362392023-06-01 13:51:45 +02001925 BLKDBG_CO_EVENT(extent->file, BLKDBG_READ_AIO);
Kevin Wolfa03ef882016-06-20 21:31:46 +02001926 ret = bdrv_co_preadv(extent->file,
Kevin Wolff10cc242016-04-25 17:34:41 +02001927 cluster_offset + offset_in_cluster, bytes,
1928 qiov, 0);
1929 if (ret < 0) {
1930 return ret;
Fam Zheng2b2c8c52011-08-12 23:19:31 +08001931 }
Kevin Wolff10cc242016-04-25 17:34:41 +02001932 return 0;
Fam Zhengdd3f6ee2011-08-12 23:19:29 +08001933 }
Fam Zheng2b2c8c52011-08-12 23:19:31 +08001934 cluster_bytes = extent->cluster_sectors * 512;
1935 /* Read two clusters in case GrainMarker + compressed data > one cluster */
1936 buf_bytes = cluster_bytes * 2;
1937 cluster_buf = g_malloc(buf_bytes);
1938 uncomp_buf = g_malloc(cluster_bytes);
Paolo Bonzini17362392023-06-01 13:51:45 +02001939 BLKDBG_CO_EVENT(extent->file, BLKDBG_READ_COMPRESSED);
Alberto Fariaa5c4e5b2022-10-13 14:37:10 +02001940 ret = bdrv_co_pread(extent->file, cluster_offset, buf_bytes, cluster_buf,
1941 0);
Fam Zheng2b2c8c52011-08-12 23:19:31 +08001942 if (ret < 0) {
1943 goto out;
1944 }
1945 compressed_data = cluster_buf;
1946 buf_len = cluster_bytes;
1947 data_len = cluster_bytes;
1948 if (extent->has_marker) {
1949 marker = (VmdkGrainMarker *)cluster_buf;
1950 compressed_data = marker->data;
1951 data_len = le32_to_cpu(marker->size);
1952 }
1953 if (!data_len || data_len > buf_bytes) {
1954 ret = -EINVAL;
1955 goto out;
1956 }
1957 ret = uncompress(uncomp_buf, &buf_len, compressed_data, data_len);
1958 if (ret != Z_OK) {
1959 ret = -EINVAL;
1960 goto out;
1961
1962 }
1963 if (offset_in_cluster < 0 ||
Kevin Wolff10cc242016-04-25 17:34:41 +02001964 offset_in_cluster + bytes > buf_len) {
Fam Zheng2b2c8c52011-08-12 23:19:31 +08001965 ret = -EINVAL;
1966 goto out;
1967 }
Kevin Wolff10cc242016-04-25 17:34:41 +02001968 qemu_iovec_from_buf(qiov, 0, uncomp_buf + offset_in_cluster, bytes);
Fam Zheng2b2c8c52011-08-12 23:19:31 +08001969 ret = 0;
1970
1971 out:
1972 g_free(uncomp_buf);
1973 g_free(cluster_buf);
1974 return ret;
Fam Zhengdd3f6ee2011-08-12 23:19:29 +08001975}
1976
Kevin Wolfb9b10c32023-02-03 16:21:50 +01001977static int coroutine_fn GRAPH_RDLOCK
Vladimir Sementsov-Ogievskiyf7ef38d2021-09-03 13:27:59 +03001978vmdk_co_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
1979 QEMUIOVector *qiov, BdrvRequestFlags flags)
bellardea2384d2004-08-01 21:59:26 +00001980{
1981 BDRVVmdkState *s = bs->opaque;
Fam Zhengb3976d32011-07-12 19:56:28 +08001982 int ret;
Kevin Wolff10cc242016-04-25 17:34:41 +02001983 uint64_t n_bytes, offset_in_cluster;
Fam Zhengb3976d32011-07-12 19:56:28 +08001984 VmdkExtent *extent = NULL;
Kevin Wolff10cc242016-04-25 17:34:41 +02001985 QEMUIOVector local_qiov;
bellardea2384d2004-08-01 21:59:26 +00001986 uint64_t cluster_offset;
Kevin Wolff10cc242016-04-25 17:34:41 +02001987 uint64_t bytes_done = 0;
ths5f4da8c2007-01-24 21:05:24 +00001988
Kevin Wolff10cc242016-04-25 17:34:41 +02001989 qemu_iovec_init(&local_qiov, qiov->niov);
1990 qemu_co_mutex_lock(&s->lock);
1991
1992 while (bytes > 0) {
1993 extent = find_extent(s, offset >> BDRV_SECTOR_BITS, extent);
Fam Zhengb3976d32011-07-12 19:56:28 +08001994 if (!extent) {
Kevin Wolff10cc242016-04-25 17:34:41 +02001995 ret = -EIO;
1996 goto fail;
Fam Zhengb3976d32011-07-12 19:56:28 +08001997 }
Fam Zhengc6ac36e2014-07-30 14:39:10 +08001998 ret = get_cluster_offset(bs, extent, NULL,
Kevin Wolff10cc242016-04-25 17:34:41 +02001999 offset, false, &cluster_offset, 0, 0);
2000 offset_in_cluster = vmdk_find_offset_in_cluster(extent, offset);
2001
2002 n_bytes = MIN(bytes, extent->cluster_sectors * BDRV_SECTOR_SIZE
2003 - offset_in_cluster);
2004
Fam Zheng14ead642013-05-02 10:25:23 +08002005 if (ret != VMDK_OK) {
Fam Zheng91b85bd2011-07-12 19:56:35 +08002006 /* if not allocated, try to read from parent image, if exist */
Kevin Wolf760e0062015-06-17 14:55:21 +02002007 if (bs->backing && ret != VMDK_ZEROED) {
Fam Zhengae261c82011-07-12 19:56:38 +08002008 if (!vmdk_is_cid_valid(bs)) {
Kevin Wolff10cc242016-04-25 17:34:41 +02002009 ret = -EINVAL;
2010 goto fail;
Fam Zhengae261c82011-07-12 19:56:38 +08002011 }
Kevin Wolff10cc242016-04-25 17:34:41 +02002012
2013 qemu_iovec_reset(&local_qiov);
2014 qemu_iovec_concat(&local_qiov, qiov, bytes_done, n_bytes);
2015
Max Reitz23c4b2a2017-11-23 03:08:19 +01002016 /* qcow2 emits this on bs->file instead of bs->backing */
Paolo Bonzini17362392023-06-01 13:51:45 +02002017 BLKDBG_CO_EVENT(bs->file, BLKDBG_READ_BACKING_AIO);
Kevin Wolfa03ef882016-06-20 21:31:46 +02002018 ret = bdrv_co_preadv(bs->backing, offset, n_bytes,
Kevin Wolff10cc242016-04-25 17:34:41 +02002019 &local_qiov, 0);
Fam Zhengae261c82011-07-12 19:56:38 +08002020 if (ret < 0) {
Kevin Wolff10cc242016-04-25 17:34:41 +02002021 goto fail;
Fam Zhengae261c82011-07-12 19:56:38 +08002022 }
ths5f4da8c2007-01-24 21:05:24 +00002023 } else {
Kevin Wolff10cc242016-04-25 17:34:41 +02002024 qemu_iovec_memset(qiov, bytes_done, 0, n_bytes);
ths5f4da8c2007-01-24 21:05:24 +00002025 }
bellardea2384d2004-08-01 21:59:26 +00002026 } else {
Kevin Wolff10cc242016-04-25 17:34:41 +02002027 qemu_iovec_reset(&local_qiov);
2028 qemu_iovec_concat(&local_qiov, qiov, bytes_done, n_bytes);
2029
2030 ret = vmdk_read_extent(extent, cluster_offset, offset_in_cluster,
2031 &local_qiov, n_bytes);
Fam Zhengdd3f6ee2011-08-12 23:19:29 +08002032 if (ret) {
Kevin Wolff10cc242016-04-25 17:34:41 +02002033 goto fail;
Fam Zheng7fa60fa2011-07-19 08:38:22 +08002034 }
bellardea2384d2004-08-01 21:59:26 +00002035 }
Kevin Wolff10cc242016-04-25 17:34:41 +02002036 bytes -= n_bytes;
2037 offset += n_bytes;
2038 bytes_done += n_bytes;
bellardea2384d2004-08-01 21:59:26 +00002039 }
bellardea2384d2004-08-01 21:59:26 +00002040
Kevin Wolff10cc242016-04-25 17:34:41 +02002041 ret = 0;
2042fail:
Paolo Bonzini2914caa2011-10-20 13:16:22 +02002043 qemu_co_mutex_unlock(&s->lock);
Kevin Wolff10cc242016-04-25 17:34:41 +02002044 qemu_iovec_destroy(&local_qiov);
2045
Paolo Bonzini2914caa2011-10-20 13:16:22 +02002046 return ret;
2047}
2048
Fam Zhengcdeaf1f2013-05-02 10:25:27 +08002049/**
2050 * vmdk_write:
2051 * @zeroed: buf is ignored (data is zero), use zeroed_grain GTE feature
Fam Zheng8e507242013-08-01 18:12:17 +08002052 * if possible, otherwise return -ENOTSUP.
2053 * @zero_dry_run: used for zeroed == true only, don't update L2 table, just try
2054 * with each cluster. By dry run we can find if the zero write
2055 * is possible without modifying image data.
Fam Zhengcdeaf1f2013-05-02 10:25:27 +08002056 *
2057 * Returns: error code with 0 for success.
2058 */
Kevin Wolfb9b10c32023-02-03 16:21:50 +01002059static int coroutine_fn GRAPH_RDLOCK
2060vmdk_pwritev(BlockDriverState *bs, uint64_t offset, uint64_t bytes,
2061 QEMUIOVector *qiov, bool zeroed, bool zero_dry_run)
bellardea2384d2004-08-01 21:59:26 +00002062{
bellardff1afc72005-04-26 21:08:00 +00002063 BDRVVmdkState *s = bs->opaque;
Fam Zhengb3976d32011-07-12 19:56:28 +08002064 VmdkExtent *extent = NULL;
Fam Zheng585ea0c2014-01-08 09:42:07 +08002065 int ret;
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02002066 int64_t offset_in_cluster, n_bytes;
bellardff1afc72005-04-26 21:08:00 +00002067 uint64_t cluster_offset;
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02002068 uint64_t bytes_done = 0;
Fam Zhengb3976d32011-07-12 19:56:28 +08002069 VmdkMetaData m_data;
bellardff1afc72005-04-26 21:08:00 +00002070
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02002071 if (DIV_ROUND_UP(offset, BDRV_SECTOR_SIZE) > bs->total_sectors) {
2072 error_report("Wrong offset: offset=0x%" PRIx64
Markus Armbruster9af9e0f2015-12-18 16:35:19 +01002073 " total_sectors=0x%" PRIx64,
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02002074 offset, bs->total_sectors);
Fam Zheng7fa60fa2011-07-19 08:38:22 +08002075 return -EIO;
ths630530a2007-06-18 15:01:30 +00002076 }
2077
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02002078 while (bytes > 0) {
2079 extent = find_extent(s, offset >> BDRV_SECTOR_BITS, extent);
Fam Zhengb3976d32011-07-12 19:56:28 +08002080 if (!extent) {
2081 return -EIO;
2082 }
Sam Eiderman98eb9732019-06-20 12:10:57 +03002083 if (extent->sesparse) {
2084 return -ENOTSUP;
2085 }
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02002086 offset_in_cluster = vmdk_find_offset_in_cluster(extent, offset);
2087 n_bytes = MIN(bytes, extent->cluster_sectors * BDRV_SECTOR_SIZE
2088 - offset_in_cluster);
2089
2090 ret = get_cluster_offset(bs, extent, &m_data, offset,
Fam Zhengc6ac36e2014-07-30 14:39:10 +08002091 !(extent->compressed || zeroed),
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02002092 &cluster_offset, offset_in_cluster,
2093 offset_in_cluster + n_bytes);
Fam Zheng2b2c8c52011-08-12 23:19:31 +08002094 if (extent->compressed) {
Fam Zheng65f74722013-05-02 10:25:22 +08002095 if (ret == VMDK_OK) {
Fam Zheng2b2c8c52011-08-12 23:19:31 +08002096 /* Refuse write to allocated cluster for streamOptimized */
Fam Zheng48239702013-10-11 15:43:22 +08002097 error_report("Could not write to allocated cluster"
2098 " for streamOptimized");
Fam Zheng2b2c8c52011-08-12 23:19:31 +08002099 return -EIO;
Kevin Wolf2821c1c2020-04-30 15:30:03 +02002100 } else if (!zeroed) {
Fam Zheng2b2c8c52011-08-12 23:19:31 +08002101 /* allocate */
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02002102 ret = get_cluster_offset(bs, extent, &m_data, offset,
Fam Zhengc6ac36e2014-07-30 14:39:10 +08002103 true, &cluster_offset, 0, 0);
Fam Zheng2b2c8c52011-08-12 23:19:31 +08002104 }
2105 }
Fam Zhengcdeaf1f2013-05-02 10:25:27 +08002106 if (ret == VMDK_ERROR) {
Fam Zheng91b85bd2011-07-12 19:56:35 +08002107 return -EINVAL;
Fam Zhengb3976d32011-07-12 19:56:28 +08002108 }
Fam Zhengcdeaf1f2013-05-02 10:25:27 +08002109 if (zeroed) {
2110 /* Do zeroed write, buf is ignored */
2111 if (extent->has_zero_grain &&
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02002112 offset_in_cluster == 0 &&
2113 n_bytes >= extent->cluster_sectors * BDRV_SECTOR_SIZE) {
2114 n_bytes = extent->cluster_sectors * BDRV_SECTOR_SIZE;
Kevin Wolf78cae782020-04-30 15:30:05 +02002115 if (!zero_dry_run && ret != VMDK_ZEROED) {
Fam Zhengcdeaf1f2013-05-02 10:25:27 +08002116 /* update L2 tables */
Fam Zhengc6ac36e2014-07-30 14:39:10 +08002117 if (vmdk_L2update(extent, &m_data, VMDK_GTE_ZEROED)
2118 != VMDK_OK) {
Fam Zhengcdeaf1f2013-05-02 10:25:27 +08002119 return -EIO;
2120 }
2121 }
2122 } else {
2123 return -ENOTSUP;
2124 }
2125 } else {
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02002126 ret = vmdk_write_extent(extent, cluster_offset, offset_in_cluster,
2127 qiov, bytes_done, n_bytes, offset);
Fam Zhengcdeaf1f2013-05-02 10:25:27 +08002128 if (ret) {
2129 return ret;
2130 }
Kevin Wolf4dc20e62020-04-30 15:30:02 +02002131 if (m_data.new_allocation) {
Fam Zhengcdeaf1f2013-05-02 10:25:27 +08002132 /* update L2 tables */
Fam Zhengc6ac36e2014-07-30 14:39:10 +08002133 if (vmdk_L2update(extent, &m_data,
2134 cluster_offset >> BDRV_SECTOR_BITS)
2135 != VMDK_OK) {
Fam Zhengcdeaf1f2013-05-02 10:25:27 +08002136 return -EIO;
2137 }
Fam Zhengb3976d32011-07-12 19:56:28 +08002138 }
ths630530a2007-06-18 15:01:30 +00002139 }
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02002140 bytes -= n_bytes;
2141 offset += n_bytes;
2142 bytes_done += n_bytes;
ths5f4da8c2007-01-24 21:05:24 +00002143
Fam Zhengae261c82011-07-12 19:56:38 +08002144 /* update CID on the first write every time the virtual disk is
2145 * opened */
Fam Zheng69b4d862011-07-12 19:56:34 +08002146 if (!s->cid_updated) {
Fam Zhenge5dc64b2014-12-04 07:28:29 +08002147 ret = vmdk_write_cid(bs, g_random_int());
Kevin Wolf99f18352011-10-26 12:25:25 +02002148 if (ret < 0) {
2149 return ret;
2150 }
Fam Zheng69b4d862011-07-12 19:56:34 +08002151 s->cid_updated = true;
ths5f4da8c2007-01-24 21:05:24 +00002152 }
bellardff1afc72005-04-26 21:08:00 +00002153 }
2154 return 0;
bellardea2384d2004-08-01 21:59:26 +00002155}
2156
Kevin Wolfb9b10c32023-02-03 16:21:50 +01002157static int coroutine_fn GRAPH_RDLOCK
Vladimir Sementsov-Ogievskiye75abed2021-09-03 13:28:00 +03002158vmdk_co_pwritev(BlockDriverState *bs, int64_t offset, int64_t bytes,
2159 QEMUIOVector *qiov, BdrvRequestFlags flags)
Paolo Bonzinie183ef72011-10-20 13:16:23 +02002160{
2161 int ret;
2162 BDRVVmdkState *s = bs->opaque;
2163 qemu_co_mutex_lock(&s->lock);
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02002164 ret = vmdk_pwritev(bs, offset, bytes, qiov, false, false);
Fam Zhengcdeaf1f2013-05-02 10:25:27 +08002165 qemu_co_mutex_unlock(&s->lock);
2166 return ret;
2167}
2168
Kevin Wolf7b1fb722023-02-03 16:21:49 +01002169static int coroutine_fn GRAPH_RDLOCK
Vladimir Sementsov-Ogievskiye75abed2021-09-03 13:28:00 +03002170vmdk_co_pwritev_compressed(BlockDriverState *bs, int64_t offset, int64_t bytes,
2171 QEMUIOVector *qiov)
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02002172{
yuchenlin51b3c6b2018-09-13 16:29:52 +08002173 if (bytes == 0) {
2174 /* The caller will write bytes 0 to signal EOF.
2175 * When receive it, we align EOF to a sector boundary. */
2176 BDRVVmdkState *s = bs->opaque;
2177 int i, ret;
2178 int64_t length;
2179
2180 for (i = 0; i < s->num_extents; i++) {
Paolo Bonzini0af02bd2023-06-01 13:51:44 +02002181 length = bdrv_co_getlength(s->extents[i].file->bs);
yuchenlin51b3c6b2018-09-13 16:29:52 +08002182 if (length < 0) {
2183 return length;
2184 }
2185 length = QEMU_ALIGN_UP(length, BDRV_SECTOR_SIZE);
Alberto Fariaa5c4e5b2022-10-13 14:37:10 +02002186 ret = bdrv_co_truncate(s->extents[i].file, length, false,
2187 PREALLOC_MODE_OFF, 0, NULL);
yuchenlin51b3c6b2018-09-13 16:29:52 +08002188 if (ret < 0) {
2189 return ret;
2190 }
2191 }
2192 return 0;
2193 }
Pavel Butsykinb2c622d2016-07-22 11:17:45 +03002194 return vmdk_co_pwritev(bs, offset, bytes, qiov, 0);
Fam Zhengba0ad892014-05-06 21:08:44 +08002195}
2196
Kevin Wolfb9b10c32023-02-03 16:21:50 +01002197static int coroutine_fn GRAPH_RDLOCK
2198vmdk_co_pwrite_zeroes(BlockDriverState *bs, int64_t offset, int64_t bytes,
2199 BdrvRequestFlags flags)
Fam Zhengcdeaf1f2013-05-02 10:25:27 +08002200{
2201 int ret;
2202 BDRVVmdkState *s = bs->opaque;
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02002203
Fam Zhengcdeaf1f2013-05-02 10:25:27 +08002204 qemu_co_mutex_lock(&s->lock);
Fam Zheng8e507242013-08-01 18:12:17 +08002205 /* write zeroes could fail if sectors not aligned to cluster, test it with
2206 * dry_run == true before really updating image */
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02002207 ret = vmdk_pwritev(bs, offset, bytes, NULL, true, true);
Fam Zhengcdeaf1f2013-05-02 10:25:27 +08002208 if (!ret) {
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02002209 ret = vmdk_pwritev(bs, offset, bytes, NULL, true, false);
Fam Zhengcdeaf1f2013-05-02 10:25:27 +08002210 }
Paolo Bonzinie183ef72011-10-20 13:16:23 +02002211 qemu_co_mutex_unlock(&s->lock);
2212 return ret;
2213}
2214
Paolo Bonzini28944f92023-06-01 13:51:41 +02002215static int coroutine_fn GRAPH_UNLOCKED
Kevin Wolf4db7ba32023-05-10 22:35:54 +02002216vmdk_init_extent(BlockBackend *blk, int64_t filesize, bool flat, bool compress,
2217 bool zeroed_grain, Error **errp)
bellard8979b222005-07-02 14:02:54 +00002218{
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002219 int ret, i;
bellard8979b222005-07-02 14:02:54 +00002220 VMDK4Header header;
Fam Zheng917703c2013-12-20 09:48:48 +08002221 uint32_t tmp, magic, grains, gd_sectors, gt_size, gt_count;
2222 uint32_t *gd_buf = NULL;
2223 int gd_buf_size;
Kevin Wolf0e7e1982009-05-18 16:42:10 +02002224
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002225 if (flat) {
Paolo Bonzini28944f92023-06-01 13:51:41 +02002226 ret = blk_co_truncate(blk, filesize, false, PREALLOC_MODE_OFF, 0, errp);
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002227 goto exit;
2228 }
bellard8979b222005-07-02 14:02:54 +00002229 magic = cpu_to_be32(VMDK4_MAGIC);
2230 memset(&header, 0, sizeof(header));
Fam Zhengd62d9dc2015-09-17 13:04:10 +08002231 if (compress) {
2232 header.version = 3;
2233 } else if (zeroed_grain) {
2234 header.version = 2;
2235 } else {
2236 header.version = 1;
2237 }
Fam Zheng95b0aa42013-05-02 10:25:25 +08002238 header.flags = VMDK4_FLAG_RGD | VMDK4_FLAG_NL_DETECT
Fam Zheng69e0b6d2013-05-02 10:25:24 +08002239 | (compress ? VMDK4_FLAG_COMPRESS | VMDK4_FLAG_MARKER : 0)
2240 | (zeroed_grain ? VMDK4_FLAG_ZERO_GRAIN : 0);
Fam Zheng6c031aa2011-08-12 23:19:32 +08002241 header.compressAlgorithm = compress ? VMDK4_COMPRESSION_DEFLATE : 0;
Fam Zheng917703c2013-12-20 09:48:48 +08002242 header.capacity = filesize / BDRV_SECTOR_SIZE;
Alexander Graf16372ff2011-05-25 00:46:55 +02002243 header.granularity = 128;
Fam Zheng917703c2013-12-20 09:48:48 +08002244 header.num_gtes_per_gt = BDRV_SECTOR_SIZE;
bellard8979b222005-07-02 14:02:54 +00002245
Fam Zheng917703c2013-12-20 09:48:48 +08002246 grains = DIV_ROUND_UP(filesize / BDRV_SECTOR_SIZE, header.granularity);
2247 gt_size = DIV_ROUND_UP(header.num_gtes_per_gt * sizeof(uint32_t),
2248 BDRV_SECTOR_SIZE);
2249 gt_count = DIV_ROUND_UP(grains, header.num_gtes_per_gt);
2250 gd_sectors = DIV_ROUND_UP(gt_count * sizeof(uint32_t), BDRV_SECTOR_SIZE);
bellard8979b222005-07-02 14:02:54 +00002251
2252 header.desc_offset = 1;
2253 header.desc_size = 20;
2254 header.rgd_offset = header.desc_offset + header.desc_size;
Fam Zheng917703c2013-12-20 09:48:48 +08002255 header.gd_offset = header.rgd_offset + gd_sectors + (gt_size * gt_count);
bellard8979b222005-07-02 14:02:54 +00002256 header.grain_offset =
Fam Zheng917703c2013-12-20 09:48:48 +08002257 ROUND_UP(header.gd_offset + gd_sectors + (gt_size * gt_count),
2258 header.granularity);
Alexander Graf16372ff2011-05-25 00:46:55 +02002259 /* swap endianness for all header fields */
2260 header.version = cpu_to_le32(header.version);
2261 header.flags = cpu_to_le32(header.flags);
2262 header.capacity = cpu_to_le64(header.capacity);
2263 header.granularity = cpu_to_le64(header.granularity);
Fam Zhengca8804c2013-08-06 15:44:55 +08002264 header.num_gtes_per_gt = cpu_to_le32(header.num_gtes_per_gt);
bellard8979b222005-07-02 14:02:54 +00002265 header.desc_offset = cpu_to_le64(header.desc_offset);
2266 header.desc_size = cpu_to_le64(header.desc_size);
2267 header.rgd_offset = cpu_to_le64(header.rgd_offset);
2268 header.gd_offset = cpu_to_le64(header.gd_offset);
2269 header.grain_offset = cpu_to_le64(header.grain_offset);
Fam Zheng6c031aa2011-08-12 23:19:32 +08002270 header.compressAlgorithm = cpu_to_le16(header.compressAlgorithm);
bellard8979b222005-07-02 14:02:54 +00002271
2272 header.check_bytes[0] = 0xa;
2273 header.check_bytes[1] = 0x20;
2274 header.check_bytes[2] = 0xd;
2275 header.check_bytes[3] = 0xa;
ths3b46e622007-09-17 08:09:54 +00002276
2277 /* write all the data */
Paolo Bonzini28944f92023-06-01 13:51:41 +02002278 ret = blk_co_pwrite(blk, 0, sizeof(magic), &magic, 0);
Fam Zheng917703c2013-12-20 09:48:48 +08002279 if (ret < 0) {
Markus Armbruster29ad1872024-05-13 16:17:00 +02002280 error_setg_errno(errp, -ret, "failed to write VMDK magic");
Kirill A. Shutemov16403662010-01-20 00:56:13 +01002281 goto exit;
2282 }
Paolo Bonzini28944f92023-06-01 13:51:41 +02002283 ret = blk_co_pwrite(blk, sizeof(magic), sizeof(header), &header, 0);
Fam Zheng917703c2013-12-20 09:48:48 +08002284 if (ret < 0) {
Markus Armbruster29ad1872024-05-13 16:17:00 +02002285 error_setg_errno(errp, -ret, "failed to write VMDK header");
Kirill A. Shutemov16403662010-01-20 00:56:13 +01002286 goto exit;
2287 }
bellard8979b222005-07-02 14:02:54 +00002288
Paolo Bonzini28944f92023-06-01 13:51:41 +02002289 ret = blk_co_truncate(blk, le64_to_cpu(header.grain_offset) << 9, false,
2290 PREALLOC_MODE_OFF, 0, errp);
Kirill A. Shutemov16403662010-01-20 00:56:13 +01002291 if (ret < 0) {
Kirill A. Shutemov16403662010-01-20 00:56:13 +01002292 goto exit;
2293 }
bellard8979b222005-07-02 14:02:54 +00002294
2295 /* write grain directory */
Fam Zheng917703c2013-12-20 09:48:48 +08002296 gd_buf_size = gd_sectors * BDRV_SECTOR_SIZE;
2297 gd_buf = g_malloc0(gd_buf_size);
2298 for (i = 0, tmp = le64_to_cpu(header.rgd_offset) + gd_sectors;
Kirill A. Shutemov16403662010-01-20 00:56:13 +01002299 i < gt_count; i++, tmp += gt_size) {
Fam Zheng917703c2013-12-20 09:48:48 +08002300 gd_buf[i] = cpu_to_le32(tmp);
2301 }
Paolo Bonzini28944f92023-06-01 13:51:41 +02002302 ret = blk_co_pwrite(blk, le64_to_cpu(header.rgd_offset) * BDRV_SECTOR_SIZE,
2303 gd_buf_size, gd_buf, 0);
Fam Zheng917703c2013-12-20 09:48:48 +08002304 if (ret < 0) {
Markus Armbruster29ad1872024-05-13 16:17:00 +02002305 error_setg_errno(errp, -ret, "failed to write VMDK grain directory");
Fam Zheng917703c2013-12-20 09:48:48 +08002306 goto exit;
Kirill A. Shutemov16403662010-01-20 00:56:13 +01002307 }
ths3b46e622007-09-17 08:09:54 +00002308
bellard8979b222005-07-02 14:02:54 +00002309 /* write backup grain directory */
Fam Zheng917703c2013-12-20 09:48:48 +08002310 for (i = 0, tmp = le64_to_cpu(header.gd_offset) + gd_sectors;
Kirill A. Shutemov16403662010-01-20 00:56:13 +01002311 i < gt_count; i++, tmp += gt_size) {
Fam Zheng917703c2013-12-20 09:48:48 +08002312 gd_buf[i] = cpu_to_le32(tmp);
2313 }
Paolo Bonzini28944f92023-06-01 13:51:41 +02002314 ret = blk_co_pwrite(blk, le64_to_cpu(header.gd_offset) * BDRV_SECTOR_SIZE,
2315 gd_buf_size, gd_buf, 0);
Fam Zheng917703c2013-12-20 09:48:48 +08002316 if (ret < 0) {
Markus Armbruster29ad1872024-05-13 16:17:00 +02002317 error_setg_errno(errp, -ret,
2318 "failed to write VMDK backup grain directory");
Kirill A. Shutemov16403662010-01-20 00:56:13 +01002319 }
bellard8979b222005-07-02 14:02:54 +00002320
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002321 ret = 0;
Fam Zheng917703c2013-12-20 09:48:48 +08002322exit:
Fam Zheng917703c2013-12-20 09:48:48 +08002323 g_free(gd_buf);
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002324 return ret;
2325}
bellard8979b222005-07-02 14:02:54 +00002326
Kevin Wolf4db7ba32023-05-10 22:35:54 +02002327static int coroutine_fn GRAPH_UNLOCKED
Kevin Wolf4ec8df02023-02-03 16:21:55 +01002328vmdk_create_extent(const char *filename, int64_t filesize, bool flat,
2329 bool compress, bool zeroed_grain, BlockBackend **pbb,
2330 QemuOpts *opts, Error **errp)
Fam Zheng5be28492018-05-15 23:36:31 +08002331{
2332 int ret;
2333 BlockBackend *blk = NULL;
Fam Zheng5be28492018-05-15 23:36:31 +08002334
Emanuele Giuseppe Esposito2475a0d2022-11-28 09:23:31 -05002335 ret = bdrv_co_create_file(filename, opts, errp);
Fam Zheng5be28492018-05-15 23:36:31 +08002336 if (ret < 0) {
Fam Zheng5be28492018-05-15 23:36:31 +08002337 goto exit;
2338 }
2339
Kevin Wolf882f2022023-01-26 18:24:29 +01002340 blk = blk_co_new_open(filename, NULL, NULL,
2341 BDRV_O_RDWR | BDRV_O_RESIZE | BDRV_O_PROTOCOL,
2342 errp);
Fam Zheng5be28492018-05-15 23:36:31 +08002343 if (blk == NULL) {
Fam Zheng5be28492018-05-15 23:36:31 +08002344 ret = -EIO;
2345 goto exit;
2346 }
2347
2348 blk_set_allow_write_beyond_eof(blk, true);
2349
2350 ret = vmdk_init_extent(blk, filesize, flat, compress, zeroed_grain, errp);
2351exit:
2352 if (blk) {
2353 if (pbb) {
2354 *pbb = blk;
2355 } else {
Kevin Wolfb2ab5f52023-05-04 13:57:33 +02002356 blk_co_unref(blk);
Fam Zheng5be28492018-05-15 23:36:31 +08002357 blk = NULL;
2358 }
2359 }
2360 return ret;
2361}
2362
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002363static int filename_decompose(const char *filename, char *path, char *prefix,
Fam Zheng48239702013-10-11 15:43:22 +08002364 char *postfix, size_t buf_len, Error **errp)
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002365{
2366 const char *p, *q;
2367
2368 if (filename == NULL || !strlen(filename)) {
Fam Zheng48239702013-10-11 15:43:22 +08002369 error_setg(errp, "No filename provided");
Fam Zheng65f74722013-05-02 10:25:22 +08002370 return VMDK_ERROR;
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002371 }
2372 p = strrchr(filename, '/');
2373 if (p == NULL) {
2374 p = strrchr(filename, '\\');
2375 }
2376 if (p == NULL) {
2377 p = strrchr(filename, ':');
2378 }
2379 if (p != NULL) {
2380 p++;
2381 if (p - filename >= buf_len) {
Fam Zheng65f74722013-05-02 10:25:22 +08002382 return VMDK_ERROR;
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002383 }
2384 pstrcpy(path, p - filename + 1, filename);
2385 } else {
2386 p = filename;
2387 path[0] = '\0';
2388 }
2389 q = strrchr(p, '.');
2390 if (q == NULL) {
2391 pstrcpy(prefix, buf_len, p);
2392 postfix[0] = '\0';
2393 } else {
2394 if (q - p >= buf_len) {
Fam Zheng65f74722013-05-02 10:25:22 +08002395 return VMDK_ERROR;
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002396 }
2397 pstrcpy(prefix, q - p + 1, p);
2398 pstrcpy(postfix, buf_len, q);
2399 }
Fam Zheng65f74722013-05-02 10:25:22 +08002400 return VMDK_OK;
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002401}
2402
Fam Zheng30153722018-05-15 23:36:32 +08002403/*
2404 * idx == 0: get or create the descriptor file (also the image file if in a
2405 * non-split format.
2406 * idx >= 1: get the n-th extent if in a split subformat
2407 */
Kevin Wolf4db7ba32023-05-10 22:35:54 +02002408typedef BlockBackend * coroutine_fn GRAPH_UNLOCKED_PTR
Kevin Wolf4ec8df02023-02-03 16:21:55 +01002409 (*vmdk_create_extent_fn)(int64_t size, int idx, bool flat, bool split,
2410 bool compress, bool zeroed_grain, void *opaque,
2411 Error **errp);
Fam Zheng30153722018-05-15 23:36:32 +08002412
2413static void vmdk_desc_add_extent(GString *desc,
2414 const char *extent_line_fmt,
2415 int64_t size, const char *filename)
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002416{
Fam Zheng30153722018-05-15 23:36:32 +08002417 char *basename = g_path_get_basename(filename);
2418
2419 g_string_append_printf(desc, extent_line_fmt,
2420 DIV_ROUND_UP(size, BDRV_SECTOR_SIZE), basename);
2421 g_free(basename);
2422}
2423
Kevin Wolf4db7ba32023-05-10 22:35:54 +02002424static int coroutine_fn GRAPH_UNLOCKED
Kevin Wolf4ec8df02023-02-03 16:21:55 +01002425vmdk_co_do_create(int64_t size,
2426 BlockdevVmdkSubformat subformat,
2427 BlockdevVmdkAdapterType adapter_type,
2428 const char *backing_file,
2429 const char *hw_version,
2430 const char *toolsversion,
2431 bool compat6,
2432 bool zeroed_grain,
2433 vmdk_create_extent_fn extent_fn,
2434 void *opaque,
2435 Error **errp)
Fam Zheng30153722018-05-15 23:36:32 +08002436{
2437 int extent_idx;
2438 BlockBackend *blk = NULL;
Kevin Wolf4a960ec2018-12-07 12:42:19 +01002439 BlockBackend *extent_blk;
Fam Zhengc13959c2014-05-28 11:38:58 +08002440 Error *local_err = NULL;
Fam Zhengaf057fe2013-12-03 10:41:05 +08002441 char *desc = NULL;
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002442 int ret = 0;
Fam Zheng6c031aa2011-08-12 23:19:32 +08002443 bool flat, split, compress;
Fam Zhengaf057fe2013-12-03 10:41:05 +08002444 GString *ext_desc_lines;
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002445 const int64_t split_size = 0x80000000; /* VMDK has constant split size */
Fam Zheng30153722018-05-15 23:36:32 +08002446 int64_t extent_size;
2447 int64_t created_size = 0;
2448 const char *extent_line_fmt;
Jeff Codyfe206562015-01-22 08:03:26 -05002449 char *parent_desc_line = g_malloc0(BUF_SIZE);
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002450 uint32_t parent_cid = 0xffffffff;
Othmar Pasteka7f2039f2013-01-30 00:26:52 +01002451 uint32_t number_heads = 16;
Fam Zheng917703c2013-12-20 09:48:48 +08002452 uint32_t desc_offset = 0, desc_len;
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002453 const char desc_template[] =
2454 "# Disk DescriptorFile\n"
2455 "version=1\n"
Fam Zheng9b170312014-04-17 18:43:53 +08002456 "CID=%" PRIx32 "\n"
2457 "parentCID=%" PRIx32 "\n"
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002458 "createType=\"%s\"\n"
2459 "%s"
2460 "\n"
2461 "# Extent description\n"
2462 "%s"
2463 "\n"
2464 "# The Disk Data Base\n"
2465 "#DDB\n"
2466 "\n"
Janne Karhunenf2499242016-05-03 02:43:30 -07002467 "ddb.virtualHWVersion = \"%s\"\n"
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002468 "ddb.geometry.cylinders = \"%" PRId64 "\"\n"
Fam Zheng4ab9dab2014-04-17 11:34:37 +08002469 "ddb.geometry.heads = \"%" PRIu32 "\"\n"
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002470 "ddb.geometry.sectors = \"63\"\n"
Thomas Weißschuhf3d43df2021-09-13 15:04:19 +02002471 "ddb.adapterType = \"%s\"\n"
2472 "ddb.toolsVersion = \"%s\"\n";
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002473
Fam Zhengaf057fe2013-12-03 10:41:05 +08002474 ext_desc_lines = g_string_new(NULL);
2475
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002476 /* Read out options */
Fam Zheng30153722018-05-15 23:36:32 +08002477 if (compat6) {
2478 if (hw_version) {
Janne Karhunenf2499242016-05-03 02:43:30 -07002479 error_setg(errp,
2480 "compat6 cannot be enabled with hwversion set");
2481 ret = -EINVAL;
2482 goto exit;
2483 }
Fam Zheng30153722018-05-15 23:36:32 +08002484 hw_version = "6";
Janne Karhunenf2499242016-05-03 02:43:30 -07002485 }
Fam Zheng30153722018-05-15 23:36:32 +08002486 if (!hw_version) {
2487 hw_version = "4";
Chunyan Liu5820f1d2014-06-05 17:21:09 +08002488 }
Thomas Weißschuhf3d43df2021-09-13 15:04:19 +02002489 if (!toolsversion) {
2490 toolsversion = "2147483647";
2491 }
Chunyan Liu5820f1d2014-06-05 17:21:09 +08002492
Fam Zheng30153722018-05-15 23:36:32 +08002493 if (adapter_type != BLOCKDEV_VMDK_ADAPTER_TYPE_IDE) {
Othmar Pasteka7f2039f2013-01-30 00:26:52 +01002494 /* that's the number of heads with which vmware operates when
2495 creating, exporting, etc. vmdk files with a non-ide adapter type */
2496 number_heads = 255;
2497 }
Fam Zheng30153722018-05-15 23:36:32 +08002498 split = (subformat == BLOCKDEV_VMDK_SUBFORMAT_TWOGBMAXEXTENTFLAT) ||
2499 (subformat == BLOCKDEV_VMDK_SUBFORMAT_TWOGBMAXEXTENTSPARSE);
2500 flat = (subformat == BLOCKDEV_VMDK_SUBFORMAT_MONOLITHICFLAT) ||
2501 (subformat == BLOCKDEV_VMDK_SUBFORMAT_TWOGBMAXEXTENTFLAT);
2502 compress = subformat == BLOCKDEV_VMDK_SUBFORMAT_STREAMOPTIMIZED;
2503
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002504 if (flat) {
Fam Zheng30153722018-05-15 23:36:32 +08002505 extent_line_fmt = "RW %" PRId64 " FLAT \"%s\" 0\n";
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002506 } else {
Fam Zheng30153722018-05-15 23:36:32 +08002507 extent_line_fmt = "RW %" PRId64 " SPARSE \"%s\"\n";
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002508 }
2509 if (flat && backing_file) {
Fam Zheng48239702013-10-11 15:43:22 +08002510 error_setg(errp, "Flat image can't have backing file");
Fam Zhengaf057fe2013-12-03 10:41:05 +08002511 ret = -ENOTSUP;
2512 goto exit;
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002513 }
Fam Zheng52c8d622013-10-11 15:43:23 +08002514 if (flat && zeroed_grain) {
2515 error_setg(errp, "Flat image can't enable zeroed grain");
Fam Zhengaf057fe2013-12-03 10:41:05 +08002516 ret = -ENOTSUP;
2517 goto exit;
Fam Zheng52c8d622013-10-11 15:43:23 +08002518 }
Fam Zheng30153722018-05-15 23:36:32 +08002519
2520 /* Create extents */
2521 if (split) {
2522 extent_size = split_size;
2523 } else {
2524 extent_size = size;
2525 }
2526 if (!split && !flat) {
2527 created_size = extent_size;
2528 } else {
2529 created_size = 0;
2530 }
2531 /* Get the descriptor file BDS */
2532 blk = extent_fn(created_size, 0, flat, split, compress, zeroed_grain,
2533 opaque, errp);
2534 if (!blk) {
2535 ret = -EIO;
2536 goto exit;
2537 }
2538 if (!split && !flat) {
2539 vmdk_desc_add_extent(ext_desc_lines, extent_line_fmt, created_size,
2540 blk_bs(blk)->filename);
2541 }
2542
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002543 if (backing_file) {
Fam Zheng30153722018-05-15 23:36:32 +08002544 BlockBackend *backing;
Max Reitz645ae7d2019-02-01 20:29:14 +01002545 char *full_backing =
2546 bdrv_get_full_backing_filename_from_filename(blk_bs(blk)->filename,
2547 backing_file,
2548 &local_err);
Max Reitz1085daf2014-11-26 17:20:28 +01002549 if (local_err) {
Max Reitz1085daf2014-11-26 17:20:28 +01002550 error_propagate(errp, local_err);
2551 ret = -ENOENT;
2552 goto exit;
2553 }
Max Reitz645ae7d2019-02-01 20:29:14 +01002554 assert(full_backing);
Kevin Wolfc4bea162016-03-08 15:57:05 +01002555
Kevin Wolf882f2022023-01-26 18:24:29 +01002556 backing = blk_co_new_open(full_backing, NULL, NULL,
2557 BDRV_O_NO_BACKING, errp);
Max Reitz1085daf2014-11-26 17:20:28 +01002558 g_free(full_backing);
Fam Zheng30153722018-05-15 23:36:32 +08002559 if (backing == NULL) {
Kevin Wolfc4bea162016-03-08 15:57:05 +01002560 ret = -EIO;
Fam Zhengaf057fe2013-12-03 10:41:05 +08002561 goto exit;
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002562 }
Fam Zheng30153722018-05-15 23:36:32 +08002563 if (strcmp(blk_bs(backing)->drv->format_name, "vmdk")) {
2564 error_setg(errp, "Invalid backing file format: %s. Must be vmdk",
2565 blk_bs(backing)->drv->format_name);
Kevin Wolfb2ab5f52023-05-04 13:57:33 +02002566 blk_co_unref(backing);
Fam Zhengaf057fe2013-12-03 10:41:05 +08002567 ret = -EINVAL;
2568 goto exit;
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002569 }
Kevin Wolf1f051dc2023-10-27 17:53:33 +02002570
2571 bdrv_graph_co_rdlock();
Fam Zheng30153722018-05-15 23:36:32 +08002572 ret = vmdk_read_cid(blk_bs(backing), 0, &parent_cid);
Kevin Wolf1f051dc2023-10-27 17:53:33 +02002573 bdrv_graph_co_rdunlock();
Kevin Wolfb2ab5f52023-05-04 13:57:33 +02002574 blk_co_unref(backing);
Peter Maydell98778602017-07-09 18:06:14 +01002575 if (ret) {
Fam Zheng30153722018-05-15 23:36:32 +08002576 error_setg(errp, "Failed to read parent CID");
Peter Maydell98778602017-07-09 18:06:14 +01002577 goto exit;
2578 }
Jeff Codyfe206562015-01-22 08:03:26 -05002579 snprintf(parent_desc_line, BUF_SIZE,
Fam Zheng8ed610a2013-06-26 17:24:32 +08002580 "parentFileNameHint=\"%s\"", backing_file);
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002581 }
Fam Zheng30153722018-05-15 23:36:32 +08002582 extent_idx = 1;
2583 while (created_size < size) {
Fam Zheng30153722018-05-15 23:36:32 +08002584 int64_t cur_size = MIN(size - created_size, extent_size);
2585 extent_blk = extent_fn(cur_size, extent_idx, flat, split, compress,
2586 zeroed_grain, opaque, errp);
2587 if (!extent_blk) {
Fam Zhengaf057fe2013-12-03 10:41:05 +08002588 ret = -EINVAL;
2589 goto exit;
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002590 }
Fam Zheng30153722018-05-15 23:36:32 +08002591 vmdk_desc_add_extent(ext_desc_lines, extent_line_fmt, cur_size,
2592 blk_bs(extent_blk)->filename);
2593 created_size += cur_size;
2594 extent_idx++;
Kevin Wolfb2ab5f52023-05-04 13:57:33 +02002595 blk_co_unref(extent_blk);
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002596 }
Kevin Wolf4a960ec2018-12-07 12:42:19 +01002597
2598 /* Check whether we got excess extents */
2599 extent_blk = extent_fn(-1, extent_idx, flat, split, compress, zeroed_grain,
2600 opaque, NULL);
2601 if (extent_blk) {
Kevin Wolfb2ab5f52023-05-04 13:57:33 +02002602 blk_co_unref(extent_blk);
Kevin Wolf4a960ec2018-12-07 12:42:19 +01002603 error_setg(errp, "List of extents contains unused extents");
2604 ret = -EINVAL;
2605 goto exit;
2606 }
2607
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002608 /* generate descriptor file */
Fam Zhengaf057fe2013-12-03 10:41:05 +08002609 desc = g_strdup_printf(desc_template,
Fam Zhenge5dc64b2014-12-04 07:28:29 +08002610 g_random_int(),
Fam Zhengaf057fe2013-12-03 10:41:05 +08002611 parent_cid,
Fam Zheng30153722018-05-15 23:36:32 +08002612 BlockdevVmdkSubformat_str(subformat),
Fam Zhengaf057fe2013-12-03 10:41:05 +08002613 parent_desc_line,
2614 ext_desc_lines->str,
Janne Karhunenf2499242016-05-03 02:43:30 -07002615 hw_version,
Fam Zheng30153722018-05-15 23:36:32 +08002616 size /
Fam Zheng917703c2013-12-20 09:48:48 +08002617 (int64_t)(63 * number_heads * BDRV_SECTOR_SIZE),
Fam Zhengaf057fe2013-12-03 10:41:05 +08002618 number_heads,
Thomas Weißschuhf3d43df2021-09-13 15:04:19 +02002619 BlockdevVmdkAdapterType_str(adapter_type),
2620 toolsversion);
Fam Zheng917703c2013-12-20 09:48:48 +08002621 desc_len = strlen(desc);
2622 /* the descriptor offset = 0x200 */
2623 if (!split && !flat) {
2624 desc_offset = 0x200;
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002625 }
Kevin Wolfc4bea162016-03-08 15:57:05 +01002626
Alberto Fariaa5c4e5b2022-10-13 14:37:10 +02002627 ret = blk_co_pwrite(blk, desc_offset, desc_len, desc, 0);
Fam Zheng917703c2013-12-20 09:48:48 +08002628 if (ret < 0) {
2629 error_setg_errno(errp, -ret, "Could not write description");
2630 goto exit;
Fam Zhengf66fd6c2011-07-19 08:45:23 +08002631 }
Fam Zheng917703c2013-12-20 09:48:48 +08002632 /* bdrv_pwrite write padding zeros to align to sector, we don't need that
2633 * for description file */
2634 if (desc_offset == 0) {
Alberto Fariaa5c4e5b2022-10-13 14:37:10 +02002635 ret = blk_co_truncate(blk, desc_len, false, PREALLOC_MODE_OFF, 0, errp);
Fam Zheng30153722018-05-15 23:36:32 +08002636 if (ret < 0) {
2637 goto exit;
2638 }
Kirill A. Shutemov16403662010-01-20 00:56:13 +01002639 }
Fam Zheng30153722018-05-15 23:36:32 +08002640 ret = 0;
Fam Zhengaf057fe2013-12-03 10:41:05 +08002641exit:
Fam Zheng30153722018-05-15 23:36:32 +08002642 if (blk) {
Kevin Wolfb2ab5f52023-05-04 13:57:33 +02002643 blk_co_unref(blk);
Fam Zheng917703c2013-12-20 09:48:48 +08002644 }
Fam Zheng30153722018-05-15 23:36:32 +08002645 g_free(desc);
2646 g_free(parent_desc_line);
2647 g_string_free(ext_desc_lines, true);
2648 return ret;
2649}
2650
2651typedef struct {
2652 char *path;
2653 char *prefix;
2654 char *postfix;
2655 QemuOpts *opts;
2656} VMDKCreateOptsData;
2657
Kevin Wolf4db7ba32023-05-10 22:35:54 +02002658static BlockBackend * coroutine_fn GRAPH_UNLOCKED
Kevin Wolf4ec8df02023-02-03 16:21:55 +01002659vmdk_co_create_opts_cb(int64_t size, int idx, bool flat, bool split,
2660 bool compress, bool zeroed_grain, void *opaque,
2661 Error **errp)
Fam Zheng30153722018-05-15 23:36:32 +08002662{
2663 BlockBackend *blk = NULL;
2664 BlockDriverState *bs = NULL;
2665 VMDKCreateOptsData *data = opaque;
2666 char *ext_filename = NULL;
2667 char *rel_filename = NULL;
2668
Kevin Wolf4a960ec2018-12-07 12:42:19 +01002669 /* We're done, don't create excess extents. */
2670 if (size == -1) {
2671 assert(errp == NULL);
2672 return NULL;
2673 }
2674
Fam Zheng30153722018-05-15 23:36:32 +08002675 if (idx == 0) {
2676 rel_filename = g_strdup_printf("%s%s", data->prefix, data->postfix);
2677 } else if (split) {
2678 rel_filename = g_strdup_printf("%s-%c%03d%s",
2679 data->prefix,
2680 flat ? 'f' : 's', idx, data->postfix);
2681 } else {
2682 assert(idx == 1);
2683 rel_filename = g_strdup_printf("%s-flat%s", data->prefix, data->postfix);
2684 }
2685
2686 ext_filename = g_strdup_printf("%s%s", data->path, rel_filename);
2687 g_free(rel_filename);
2688
2689 if (vmdk_create_extent(ext_filename, size,
2690 flat, compress, zeroed_grain, &blk, data->opts,
2691 errp)) {
2692 goto exit;
2693 }
Kevin Wolfb2ab5f52023-05-04 13:57:33 +02002694 bdrv_co_unref(bs);
Fam Zheng30153722018-05-15 23:36:32 +08002695exit:
2696 g_free(ext_filename);
2697 return blk;
2698}
2699
Kevin Wolf4db7ba32023-05-10 22:35:54 +02002700static int coroutine_fn GRAPH_UNLOCKED
Kevin Wolf4ec8df02023-02-03 16:21:55 +01002701vmdk_co_create_opts(BlockDriver *drv, const char *filename,
2702 QemuOpts *opts, Error **errp)
Fam Zheng30153722018-05-15 23:36:32 +08002703{
2704 Error *local_err = NULL;
2705 char *desc = NULL;
2706 int64_t total_size = 0;
2707 char *adapter_type = NULL;
2708 BlockdevVmdkAdapterType adapter_type_enum;
2709 char *backing_file = NULL;
2710 char *hw_version = NULL;
Thomas Weißschuhf3d43df2021-09-13 15:04:19 +02002711 char *toolsversion = NULL;
Fam Zheng30153722018-05-15 23:36:32 +08002712 char *fmt = NULL;
2713 BlockdevVmdkSubformat subformat;
2714 int ret = 0;
2715 char *path = g_malloc0(PATH_MAX);
2716 char *prefix = g_malloc0(PATH_MAX);
2717 char *postfix = g_malloc0(PATH_MAX);
2718 char *desc_line = g_malloc0(BUF_SIZE);
2719 char *ext_filename = g_malloc0(PATH_MAX);
2720 char *desc_filename = g_malloc0(PATH_MAX);
2721 char *parent_desc_line = g_malloc0(BUF_SIZE);
2722 bool zeroed_grain;
2723 bool compat6;
2724 VMDKCreateOptsData data;
Eric Blaked51a8142020-07-06 15:39:48 -05002725 char *backing_fmt = NULL;
2726
2727 backing_fmt = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FMT);
2728 if (backing_fmt && strcmp(backing_fmt, "vmdk") != 0) {
2729 error_setg(errp, "backing_file must be a vmdk image");
2730 ret = -EINVAL;
2731 goto exit;
2732 }
Fam Zheng30153722018-05-15 23:36:32 +08002733
2734 if (filename_decompose(filename, path, prefix, postfix, PATH_MAX, errp)) {
2735 ret = -EINVAL;
2736 goto exit;
2737 }
2738 /* Read out options */
2739 total_size = ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
2740 BDRV_SECTOR_SIZE);
2741 adapter_type = qemu_opt_get_del(opts, BLOCK_OPT_ADAPTER_TYPE);
2742 backing_file = qemu_opt_get_del(opts, BLOCK_OPT_BACKING_FILE);
2743 hw_version = qemu_opt_get_del(opts, BLOCK_OPT_HWVERSION);
Thomas Weißschuhf3d43df2021-09-13 15:04:19 +02002744 toolsversion = qemu_opt_get_del(opts, BLOCK_OPT_TOOLSVERSION);
Fam Zheng30153722018-05-15 23:36:32 +08002745 compat6 = qemu_opt_get_bool_del(opts, BLOCK_OPT_COMPAT6, false);
2746 if (strcmp(hw_version, "undefined") == 0) {
2747 g_free(hw_version);
yuchenlin26c92962019-02-21 11:08:05 +00002748 hw_version = NULL;
Fam Zheng30153722018-05-15 23:36:32 +08002749 }
2750 fmt = qemu_opt_get_del(opts, BLOCK_OPT_SUBFMT);
2751 zeroed_grain = qemu_opt_get_bool_del(opts, BLOCK_OPT_ZEROED_GRAIN, false);
2752
2753 if (adapter_type) {
2754 adapter_type_enum = qapi_enum_parse(&BlockdevVmdkAdapterType_lookup,
2755 adapter_type,
2756 BLOCKDEV_VMDK_ADAPTER_TYPE_IDE,
2757 &local_err);
2758 if (local_err) {
2759 error_propagate(errp, local_err);
2760 ret = -EINVAL;
2761 goto exit;
2762 }
2763 } else {
2764 adapter_type_enum = BLOCKDEV_VMDK_ADAPTER_TYPE_IDE;
2765 }
2766
2767 if (!fmt) {
2768 /* Default format to monolithicSparse */
2769 subformat = BLOCKDEV_VMDK_SUBFORMAT_MONOLITHICSPARSE;
2770 } else {
2771 subformat = qapi_enum_parse(&BlockdevVmdkSubformat_lookup,
2772 fmt,
2773 BLOCKDEV_VMDK_SUBFORMAT_MONOLITHICSPARSE,
2774 &local_err);
2775 if (local_err) {
2776 error_propagate(errp, local_err);
2777 ret = -EINVAL;
2778 goto exit;
2779 }
2780 }
2781 data = (VMDKCreateOptsData){
2782 .prefix = prefix,
2783 .postfix = postfix,
2784 .path = path,
2785 .opts = opts,
2786 };
2787 ret = vmdk_co_do_create(total_size, subformat, adapter_type_enum,
Thomas Weißschuhf3d43df2021-09-13 15:04:19 +02002788 backing_file, hw_version, toolsversion, compat6,
2789 zeroed_grain, vmdk_co_create_opts_cb, &data, errp);
Fam Zheng30153722018-05-15 23:36:32 +08002790
2791exit:
Eric Blaked51a8142020-07-06 15:39:48 -05002792 g_free(backing_fmt);
Chunyan Liu5820f1d2014-06-05 17:21:09 +08002793 g_free(adapter_type);
2794 g_free(backing_file);
Janne Karhunenf2499242016-05-03 02:43:30 -07002795 g_free(hw_version);
Thomas Weißschuhf3d43df2021-09-13 15:04:19 +02002796 g_free(toolsversion);
Chunyan Liu5820f1d2014-06-05 17:21:09 +08002797 g_free(fmt);
Fam Zhengaf057fe2013-12-03 10:41:05 +08002798 g_free(desc);
Jeff Codyfe206562015-01-22 08:03:26 -05002799 g_free(path);
2800 g_free(prefix);
2801 g_free(postfix);
2802 g_free(desc_line);
2803 g_free(ext_filename);
2804 g_free(desc_filename);
2805 g_free(parent_desc_line);
Fam Zheng30153722018-05-15 23:36:32 +08002806 return ret;
2807}
2808
Kevin Wolf4db7ba32023-05-10 22:35:54 +02002809static BlockBackend * coroutine_fn GRAPH_UNLOCKED
2810vmdk_co_create_cb(int64_t size, int idx, bool flat, bool split, bool compress,
2811 bool zeroed_grain, void *opaque, Error **errp)
Fam Zheng30153722018-05-15 23:36:32 +08002812{
2813 int ret;
2814 BlockDriverState *bs;
2815 BlockBackend *blk;
2816 BlockdevCreateOptionsVmdk *opts = opaque;
2817
2818 if (idx == 0) {
Kevin Wolf882f2022023-01-26 18:24:29 +01002819 bs = bdrv_co_open_blockdev_ref(opts->file, errp);
Fam Zheng30153722018-05-15 23:36:32 +08002820 } else {
2821 int i;
2822 BlockdevRefList *list = opts->extents;
2823 for (i = 1; i < idx; i++) {
2824 if (!list || !list->next) {
2825 error_setg(errp, "Extent [%d] not specified", i);
2826 return NULL;
2827 }
2828 list = list->next;
2829 }
2830 if (!list) {
2831 error_setg(errp, "Extent [%d] not specified", idx - 1);
2832 return NULL;
2833 }
Kevin Wolf882f2022023-01-26 18:24:29 +01002834 bs = bdrv_co_open_blockdev_ref(list->value, errp);
Fam Zheng30153722018-05-15 23:36:32 +08002835 }
2836 if (!bs) {
2837 return NULL;
2838 }
Kevin Wolf882f2022023-01-26 18:24:29 +01002839 blk = blk_co_new_with_bs(bs,
2840 BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE |
2841 BLK_PERM_RESIZE,
2842 BLK_PERM_ALL,
2843 errp);
Eric Blakea3aeeab2020-04-28 14:26:46 -05002844 if (!blk) {
Fam Zheng30153722018-05-15 23:36:32 +08002845 return NULL;
2846 }
2847 blk_set_allow_write_beyond_eof(blk, true);
Kevin Wolfb2ab5f52023-05-04 13:57:33 +02002848 bdrv_co_unref(bs);
Fam Zheng30153722018-05-15 23:36:32 +08002849
Kevin Wolf4a960ec2018-12-07 12:42:19 +01002850 if (size != -1) {
2851 ret = vmdk_init_extent(blk, size, flat, compress, zeroed_grain, errp);
2852 if (ret) {
Kevin Wolfb2ab5f52023-05-04 13:57:33 +02002853 blk_co_unref(blk);
Kevin Wolf4a960ec2018-12-07 12:42:19 +01002854 blk = NULL;
2855 }
Fam Zheng30153722018-05-15 23:36:32 +08002856 }
2857 return blk;
2858}
2859
Kevin Wolf4db7ba32023-05-10 22:35:54 +02002860static int coroutine_fn GRAPH_UNLOCKED
Kevin Wolf4ec8df02023-02-03 16:21:55 +01002861vmdk_co_create(BlockdevCreateOptions *create_options, Error **errp)
Fam Zheng30153722018-05-15 23:36:32 +08002862{
Fam Zheng30153722018-05-15 23:36:32 +08002863 BlockdevCreateOptionsVmdk *opts;
2864
2865 opts = &create_options->u.vmdk;
2866
2867 /* Validate options */
2868 if (!QEMU_IS_ALIGNED(opts->size, BDRV_SECTOR_SIZE)) {
2869 error_setg(errp, "Image size must be a multiple of 512 bytes");
Markus Armbruster851fd4a2022-11-22 14:49:17 +01002870 return -EINVAL;
Fam Zheng30153722018-05-15 23:36:32 +08002871 }
2872
Markus Armbruster851fd4a2022-11-22 14:49:17 +01002873 return vmdk_co_do_create(opts->size,
2874 opts->subformat,
2875 opts->adapter_type,
2876 opts->backing_file,
2877 opts->hwversion,
2878 opts->toolsversion,
2879 false,
2880 opts->zeroed_grain,
2881 vmdk_co_create_cb,
2882 opts, errp);
bellard8979b222005-07-02 14:02:54 +00002883}
2884
bellarde2731ad2004-09-18 19:32:11 +00002885static void vmdk_close(BlockDriverState *bs)
bellardea2384d2004-08-01 21:59:26 +00002886{
Kevin Wolf2bc31662011-11-22 16:50:27 +01002887 BDRVVmdkState *s = bs->opaque;
2888
Fam Zhengb3976d32011-07-12 19:56:28 +08002889 vmdk_free_extents(bs);
Fam Zhengf4c129a2013-10-31 10:06:23 +08002890 g_free(s->create_type);
Kevin Wolf2bc31662011-11-22 16:50:27 +01002891
Steve Sistarec8a7fc52023-10-18 06:03:36 -07002892 migrate_del_blocker(&s->migration_blocker);
bellardea2384d2004-08-01 21:59:26 +00002893}
2894
Emanuele Giuseppe Espositode335632023-05-04 13:57:43 +02002895static int64_t coroutine_fn GRAPH_RDLOCK
Emanuele Giuseppe Esposito82618d72023-01-13 21:42:07 +01002896vmdk_co_get_allocated_file_size(BlockDriverState *bs)
Fam Zheng4a1d5e12011-07-12 19:56:39 +08002897{
2898 int i;
2899 int64_t ret = 0;
2900 int64_t r;
2901 BDRVVmdkState *s = bs->opaque;
2902
Emanuele Giuseppe Esposito82618d72023-01-13 21:42:07 +01002903 ret = bdrv_co_get_allocated_file_size(bs->file->bs);
Fam Zheng4a1d5e12011-07-12 19:56:39 +08002904 if (ret < 0) {
2905 return ret;
2906 }
2907 for (i = 0; i < s->num_extents; i++) {
Kevin Wolf9a4f4c32015-06-16 14:19:22 +02002908 if (s->extents[i].file == bs->file) {
Fam Zheng4a1d5e12011-07-12 19:56:39 +08002909 continue;
2910 }
Emanuele Giuseppe Esposito82618d72023-01-13 21:42:07 +01002911 r = bdrv_co_get_allocated_file_size(s->extents[i].file->bs);
Fam Zheng4a1d5e12011-07-12 19:56:39 +08002912 if (r < 0) {
2913 return r;
2914 }
2915 ret += r;
2916 }
2917 return ret;
2918}
Kevin Wolf0e7e1982009-05-18 16:42:10 +02002919
Kevin Wolf06717982023-10-27 17:53:11 +02002920static int GRAPH_RDLOCK vmdk_has_zero_init(BlockDriverState *bs)
Fam Zhengda7a50f2013-07-01 11:33:17 +08002921{
2922 int i;
2923 BDRVVmdkState *s = bs->opaque;
2924
2925 /* If has a flat extent and its underlying storage doesn't have zero init,
2926 * return 0. */
2927 for (i = 0; i < s->num_extents; i++) {
2928 if (s->extents[i].flat) {
Kevin Wolf24bc15d2015-06-15 13:50:20 +02002929 if (!bdrv_has_zero_init(s->extents[i].file->bs)) {
Fam Zhengda7a50f2013-07-01 11:33:17 +08002930 return 0;
2931 }
2932 }
2933 }
2934 return 1;
2935}
2936
Kevin Wolfb7cfc7d2023-09-29 16:51:45 +02002937static VmdkExtentInfo * GRAPH_RDLOCK vmdk_get_extent_info(VmdkExtent *extent)
Fam Zhengf4c129a2013-10-31 10:06:23 +08002938{
Hanna Reitz456e7512022-06-20 18:26:55 +02002939 VmdkExtentInfo *info = g_new0(VmdkExtentInfo, 1);
Fam Zhengf4c129a2013-10-31 10:06:23 +08002940
Max Reitzf30c66b2019-02-01 20:29:05 +01002941 bdrv_refresh_filename(extent->file->bs);
Hanna Reitz456e7512022-06-20 18:26:55 +02002942 *info = (VmdkExtentInfo){
Kevin Wolf24bc15d2015-06-15 13:50:20 +02002943 .filename = g_strdup(extent->file->bs->filename),
Fam Zhengf4c129a2013-10-31 10:06:23 +08002944 .format = g_strdup(extent->type),
2945 .virtual_size = extent->sectors * BDRV_SECTOR_SIZE,
2946 .compressed = extent->compressed,
2947 .has_compressed = extent->compressed,
2948 .cluster_size = extent->cluster_sectors * BDRV_SECTOR_SIZE,
2949 .has_cluster_size = !extent->flat,
2950 };
2951
2952 return info;
2953}
2954
Kevin Wolfb9b10c32023-02-03 16:21:50 +01002955static int coroutine_fn GRAPH_RDLOCK
2956vmdk_co_check(BlockDriverState *bs, BdrvCheckResult *result, BdrvCheckMode fix)
Peter Lievenf43aa8e2014-01-29 09:34:16 +01002957{
2958 BDRVVmdkState *s = bs->opaque;
2959 VmdkExtent *extent = NULL;
2960 int64_t sector_num = 0;
Paolo Bonzini17362392023-06-01 13:51:45 +02002961 int64_t total_sectors = bdrv_co_nb_sectors(bs);
Peter Lievenf43aa8e2014-01-29 09:34:16 +01002962 int ret;
2963 uint64_t cluster_offset;
2964
2965 if (fix) {
2966 return -ENOTSUP;
2967 }
2968
2969 for (;;) {
2970 if (sector_num >= total_sectors) {
2971 return 0;
2972 }
2973 extent = find_extent(s, sector_num, extent);
2974 if (!extent) {
2975 fprintf(stderr,
2976 "ERROR: could not find extent for sector %" PRId64 "\n",
2977 sector_num);
Fam Zheng0e51b9b2017-08-04 22:09:42 +08002978 ret = -EINVAL;
Peter Lievenf43aa8e2014-01-29 09:34:16 +01002979 break;
2980 }
2981 ret = get_cluster_offset(bs, extent, NULL,
2982 sector_num << BDRV_SECTOR_BITS,
Fam Zhengc6ac36e2014-07-30 14:39:10 +08002983 false, &cluster_offset, 0, 0);
Peter Lievenf43aa8e2014-01-29 09:34:16 +01002984 if (ret == VMDK_ERROR) {
2985 fprintf(stderr,
2986 "ERROR: could not get cluster_offset for sector %"
2987 PRId64 "\n", sector_num);
2988 break;
2989 }
Fam Zheng0e51b9b2017-08-04 22:09:42 +08002990 if (ret == VMDK_OK) {
Paolo Bonzini0af02bd2023-06-01 13:51:44 +02002991 int64_t extent_len = bdrv_co_getlength(extent->file->bs);
Fam Zheng0e51b9b2017-08-04 22:09:42 +08002992 if (extent_len < 0) {
2993 fprintf(stderr,
2994 "ERROR: could not get extent file length for sector %"
2995 PRId64 "\n", sector_num);
2996 ret = extent_len;
2997 break;
2998 }
2999 if (cluster_offset >= extent_len) {
3000 fprintf(stderr,
3001 "ERROR: cluster offset for sector %"
3002 PRId64 " points after EOF\n", sector_num);
3003 ret = -EINVAL;
3004 break;
3005 }
Peter Lievenf43aa8e2014-01-29 09:34:16 +01003006 }
3007 sector_num += extent->cluster_sectors;
3008 }
3009
3010 result->corruptions++;
Fam Zheng0e51b9b2017-08-04 22:09:42 +08003011 return ret;
Peter Lievenf43aa8e2014-01-29 09:34:16 +01003012}
3013
Kevin Wolf35744992023-09-29 16:51:54 +02003014static ImageInfoSpecific * GRAPH_RDLOCK
3015vmdk_get_specific_info(BlockDriverState *bs, Error **errp)
Fam Zhengf4c129a2013-10-31 10:06:23 +08003016{
3017 int i;
3018 BDRVVmdkState *s = bs->opaque;
3019 ImageInfoSpecific *spec_info = g_new0(ImageInfoSpecific, 1);
Hanna Reitz456e7512022-06-20 18:26:55 +02003020 VmdkExtentInfoList **tail;
Fam Zhengf4c129a2013-10-31 10:06:23 +08003021
3022 *spec_info = (ImageInfoSpecific){
Eric Blake6a8f9662015-10-26 16:34:54 -06003023 .type = IMAGE_INFO_SPECIFIC_KIND_VMDK,
Eric Blake32bafa82016-03-17 16:48:37 -06003024 .u = {
3025 .vmdk.data = g_new0(ImageInfoSpecificVmdk, 1),
Fam Zhengf4c129a2013-10-31 10:06:23 +08003026 },
3027 };
3028
Eric Blake32bafa82016-03-17 16:48:37 -06003029 *spec_info->u.vmdk.data = (ImageInfoSpecificVmdk) {
Fam Zhengf4c129a2013-10-31 10:06:23 +08003030 .create_type = g_strdup(s->create_type),
3031 .cid = s->cid,
3032 .parent_cid = s->parent_cid,
3033 };
3034
Eric Blakec3033fd2021-01-13 16:10:12 -06003035 tail = &spec_info->u.vmdk.data->extents;
Fam Zhengf4c129a2013-10-31 10:06:23 +08003036 for (i = 0; i < s->num_extents; i++) {
Eric Blakec3033fd2021-01-13 16:10:12 -06003037 QAPI_LIST_APPEND(tail, vmdk_get_extent_info(&s->extents[i]));
Fam Zhengf4c129a2013-10-31 10:06:23 +08003038 }
3039
3040 return spec_info;
3041}
3042
Fam Zheng5f583302014-11-14 12:09:21 +08003043static bool vmdk_extents_type_eq(const VmdkExtent *a, const VmdkExtent *b)
3044{
3045 return a->flat == b->flat &&
3046 a->compressed == b->compressed &&
3047 (a->flat || a->cluster_sectors == b->cluster_sectors);
3048}
3049
Emanuele Giuseppe Esposito3d47eb02023-01-13 21:42:08 +01003050static int coroutine_fn
3051vmdk_co_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
Fam Zheng74fe1882014-05-06 21:08:45 +08003052{
3053 int i;
3054 BDRVVmdkState *s = bs->opaque;
3055 assert(s->num_extents);
Fam Zheng5f583302014-11-14 12:09:21 +08003056
3057 /* See if we have multiple extents but they have different cases */
3058 for (i = 1; i < s->num_extents; i++) {
3059 if (!vmdk_extents_type_eq(&s->extents[0], &s->extents[i])) {
3060 return -ENOTSUP;
3061 }
3062 }
Fam Zheng74fe1882014-05-06 21:08:45 +08003063 bdi->needs_compressed_writes = s->extents[0].compressed;
3064 if (!s->extents[0].flat) {
3065 bdi->cluster_size = s->extents[0].cluster_sectors << BDRV_SECTOR_BITS;
3066 }
Fam Zheng74fe1882014-05-06 21:08:45 +08003067 return 0;
3068}
3069
Kevin Wolf004915a2023-10-27 17:53:26 +02003070static void GRAPH_RDLOCK
3071vmdk_gather_child_options(BlockDriverState *bs, QDict *target,
3072 bool backing_overridden)
Max Reitzabc521a2019-02-01 20:29:26 +01003073{
3074 /* No children but file and backing can be explicitly specified (TODO) */
3075 qdict_put(target, "file",
3076 qobject_ref(bs->file->bs->full_open_options));
3077
3078 if (backing_overridden) {
3079 if (bs->backing) {
3080 qdict_put(target, "backing",
3081 qobject_ref(bs->backing->bs->full_open_options));
3082 } else {
3083 qdict_put_null(target, "backing");
3084 }
3085 }
3086}
3087
Chunyan Liu5820f1d2014-06-05 17:21:09 +08003088static QemuOptsList vmdk_create_opts = {
3089 .name = "vmdk-create-opts",
3090 .head = QTAILQ_HEAD_INITIALIZER(vmdk_create_opts.head),
3091 .desc = {
3092 {
3093 .name = BLOCK_OPT_SIZE,
3094 .type = QEMU_OPT_SIZE,
3095 .help = "Virtual disk size"
3096 },
3097 {
3098 .name = BLOCK_OPT_ADAPTER_TYPE,
3099 .type = QEMU_OPT_STRING,
3100 .help = "Virtual adapter type, can be one of "
3101 "ide (default), lsilogic, buslogic or legacyESX"
3102 },
3103 {
3104 .name = BLOCK_OPT_BACKING_FILE,
3105 .type = QEMU_OPT_STRING,
3106 .help = "File name of a base image"
3107 },
3108 {
Eric Blaked51a8142020-07-06 15:39:48 -05003109 .name = BLOCK_OPT_BACKING_FMT,
3110 .type = QEMU_OPT_STRING,
3111 .help = "Must be 'vmdk' if present",
3112 },
3113 {
Chunyan Liu5820f1d2014-06-05 17:21:09 +08003114 .name = BLOCK_OPT_COMPAT6,
3115 .type = QEMU_OPT_BOOL,
3116 .help = "VMDK version 6 image",
3117 .def_value_str = "off"
3118 },
3119 {
Janne Karhunenf2499242016-05-03 02:43:30 -07003120 .name = BLOCK_OPT_HWVERSION,
3121 .type = QEMU_OPT_STRING,
3122 .help = "VMDK hardware version",
3123 .def_value_str = "undefined"
3124 },
3125 {
Thomas Weißschuhf3d43df2021-09-13 15:04:19 +02003126 .name = BLOCK_OPT_TOOLSVERSION,
3127 .type = QEMU_OPT_STRING,
3128 .help = "VMware guest tools version",
3129 },
3130 {
Chunyan Liu5820f1d2014-06-05 17:21:09 +08003131 .name = BLOCK_OPT_SUBFMT,
3132 .type = QEMU_OPT_STRING,
3133 .help =
3134 "VMDK flat extent format, can be one of "
3135 "{monolithicSparse (default) | monolithicFlat | twoGbMaxExtentSparse | twoGbMaxExtentFlat | streamOptimized} "
3136 },
3137 {
3138 .name = BLOCK_OPT_ZEROED_GRAIN,
3139 .type = QEMU_OPT_BOOL,
3140 .help = "Enable efficient zero writes "
3141 "using the zeroed-grain GTE feature"
3142 },
3143 { /* end of list */ }
3144 }
Kevin Wolf0e7e1982009-05-18 16:42:10 +02003145};
3146
Anthony Liguori5efa9d52009-05-09 17:03:42 -05003147static BlockDriver bdrv_vmdk = {
Fam Zhengda7a50f2013-07-01 11:33:17 +08003148 .format_name = "vmdk",
3149 .instance_size = sizeof(BDRVVmdkState),
3150 .bdrv_probe = vmdk_probe,
3151 .bdrv_open = vmdk_open,
Paolo Bonzini2fd61632018-03-01 17:36:19 +01003152 .bdrv_co_check = vmdk_co_check,
Fam Zhengda7a50f2013-07-01 11:33:17 +08003153 .bdrv_reopen_prepare = vmdk_reopen_prepare,
Hanna Reitz6d17e282022-03-14 17:27:18 +01003154 .bdrv_reopen_commit = vmdk_reopen_commit,
3155 .bdrv_reopen_abort = vmdk_reopen_abort,
Max Reitz69dca432020-05-13 13:05:39 +02003156 .bdrv_child_perm = bdrv_default_perms,
Kevin Wolff10cc242016-04-25 17:34:41 +02003157 .bdrv_co_preadv = vmdk_co_preadv,
Kevin Wolf37b1d7d2016-04-26 13:39:11 +02003158 .bdrv_co_pwritev = vmdk_co_pwritev,
Pavel Butsykinb2c622d2016-07-22 11:17:45 +03003159 .bdrv_co_pwritev_compressed = vmdk_co_pwritev_compressed,
Eric Blakea620f2a2016-06-01 15:10:12 -06003160 .bdrv_co_pwrite_zeroes = vmdk_co_pwrite_zeroes,
Fam Zhengda7a50f2013-07-01 11:33:17 +08003161 .bdrv_close = vmdk_close,
Stefan Hajnocziefc75e22018-01-18 13:43:45 +01003162 .bdrv_co_create_opts = vmdk_co_create_opts,
Fam Zheng30153722018-05-15 23:36:32 +08003163 .bdrv_co_create = vmdk_co_create,
Eric Blakec72080b2018-02-13 14:26:58 -06003164 .bdrv_co_block_status = vmdk_co_block_status,
Emanuele Giuseppe Esposito82618d72023-01-13 21:42:07 +01003165 .bdrv_co_get_allocated_file_size = vmdk_co_get_allocated_file_size,
Fam Zhengda7a50f2013-07-01 11:33:17 +08003166 .bdrv_has_zero_init = vmdk_has_zero_init,
Fam Zhengf4c129a2013-10-31 10:06:23 +08003167 .bdrv_get_specific_info = vmdk_get_specific_info,
Kevin Wolfd34682c2013-12-11 19:26:16 +01003168 .bdrv_refresh_limits = vmdk_refresh_limits,
Emanuele Giuseppe Esposito3d47eb02023-01-13 21:42:08 +01003169 .bdrv_co_get_info = vmdk_co_get_info,
Max Reitzabc521a2019-02-01 20:29:26 +01003170 .bdrv_gather_child_options = vmdk_gather_child_options,
Kevin Wolf0e7e1982009-05-18 16:42:10 +02003171
Max Reitzd67066d2020-05-13 13:05:12 +02003172 .is_format = true,
Kevin Wolf8ee79e72014-06-04 15:09:35 +02003173 .supports_backing = true,
Chunyan Liu5820f1d2014-06-05 17:21:09 +08003174 .create_opts = &vmdk_create_opts,
bellardea2384d2004-08-01 21:59:26 +00003175};
Anthony Liguori5efa9d52009-05-09 17:03:42 -05003176
3177static void bdrv_vmdk_init(void)
3178{
3179 bdrv_register(&bdrv_vmdk);
3180}
3181
3182block_init(bdrv_vmdk_init);