| 1 | /* |
| 2 | * Block driver for Hyper-V VHDX Images |
| 3 | * |
| 4 | * Copyright (c) 2013 Red Hat, Inc., |
| 5 | * |
| 6 | * Authors: |
| 7 | * Jeff Cody <jcody@redhat.com> |
| 8 | * |
| 9 | * This is based on the "VHDX Format Specification v1.00", published 8/25/2012 |
| 10 | * by Microsoft: |
| 11 | * https://www.microsoft.com/en-us/download/details.aspx?id=34750 |
| 12 | * |
| 13 | * This file covers the functionality of the metadata log writing, parsing, and |
| 14 | * replay. |
| 15 | * |
| 16 | * This work is licensed under the terms of the GNU LGPL, version 2 or later. |
| 17 | * See the COPYING.LIB file in the top-level directory. |
| 18 | * |
| 19 | */ |
| 20 | |
| 21 | #include "qemu/osdep.h" |
| 22 | #include "qapi/error.h" |
| 23 | #include "block/block-io.h" |
| 24 | #include "block/block_int.h" |
| 25 | #include "qemu/error-report.h" |
| 26 | #include "qemu/bswap.h" |
| 27 | #include "qemu/memalign.h" |
| 28 | #include "vhdx.h" |
| 29 | |
| 30 | |
| 31 | typedef struct VHDXLogSequence { |
| 32 | bool valid; |
| 33 | uint32_t count; |
| 34 | VHDXLogEntries log; |
| 35 | VHDXLogEntryHeader hdr; |
| 36 | } VHDXLogSequence; |
| 37 | |
| 38 | typedef struct VHDXLogDescEntries { |
| 39 | VHDXLogEntryHeader hdr; |
| 40 | VHDXLogDescriptor desc[]; |
| 41 | } VHDXLogDescEntries; |
| 42 | |
| 43 | static const MSGUID zero_guid = { 0 }; |
| 44 | |
| 45 | /* The log located on the disk is circular buffer containing |
| 46 | * sectors of 4096 bytes each. |
| 47 | * |
| 48 | * It is assumed for the read/write functions below that the |
| 49 | * circular buffer scheme uses a 'one sector open' to indicate |
| 50 | * the buffer is full. Given the validation methods used for each |
| 51 | * sector, this method should be compatible with other methods that |
| 52 | * do not waste a sector. |
| 53 | */ |
| 54 | |
| 55 | |
| 56 | /* Allow peeking at the hdr entry at the beginning of the current |
| 57 | * read index, without advancing the read index */ |
| 58 | static int GRAPH_RDLOCK |
| 59 | vhdx_log_peek_hdr(BlockDriverState *bs, VHDXLogEntries *log, |
| 60 | VHDXLogEntryHeader *hdr) |
| 61 | { |
| 62 | int ret = 0; |
| 63 | uint64_t offset; |
| 64 | uint32_t read; |
| 65 | |
| 66 | assert(hdr != NULL); |
| 67 | |
| 68 | /* peek is only supported on sector boundaries */ |
| 69 | if (log->read % VHDX_LOG_SECTOR_SIZE) { |
| 70 | ret = -EFAULT; |
| 71 | goto exit; |
| 72 | } |
| 73 | |
| 74 | read = log->read; |
| 75 | /* we are guaranteed that a) log sectors are 4096 bytes, |
| 76 | * and b) the log length is a multiple of 1MB. So, there |
| 77 | * is always a round number of sectors in the buffer */ |
| 78 | if ((read + sizeof(VHDXLogEntryHeader)) > log->length) { |
| 79 | read = 0; |
| 80 | } |
| 81 | |
| 82 | if (read == log->write) { |
| 83 | ret = -EINVAL; |
| 84 | goto exit; |
| 85 | } |
| 86 | |
| 87 | offset = log->offset + read; |
| 88 | |
| 89 | ret = bdrv_pread(bs->file, offset, sizeof(VHDXLogEntryHeader), hdr, 0); |
| 90 | if (ret < 0) { |
| 91 | goto exit; |
| 92 | } |
| 93 | vhdx_log_entry_hdr_le_import(hdr); |
| 94 | |
| 95 | exit: |
| 96 | return ret; |
| 97 | } |
| 98 | |
| 99 | /* Index increment for log, based on sector boundaries */ |
| 100 | static int vhdx_log_inc_idx(uint32_t idx, uint64_t length) |
| 101 | { |
| 102 | idx += VHDX_LOG_SECTOR_SIZE; |
| 103 | /* we are guaranteed that a) log sectors are 4096 bytes, |
| 104 | * and b) the log length is a multiple of 1MB. So, there |
| 105 | * is always a round number of sectors in the buffer */ |
| 106 | return idx >= length ? 0 : idx; |
| 107 | } |
| 108 | |
| 109 | |
| 110 | /* Reset the log to empty */ |
| 111 | static void GRAPH_RDLOCK vhdx_log_reset(BlockDriverState *bs, BDRVVHDXState *s) |
| 112 | { |
| 113 | MSGUID guid = { 0 }; |
| 114 | s->log.read = s->log.write = 0; |
| 115 | /* a log guid of 0 indicates an empty log to any parser of v0 |
| 116 | * VHDX logs */ |
| 117 | vhdx_update_headers(bs, s, false, &guid); |
| 118 | } |
| 119 | |
| 120 | /* Reads num_sectors from the log (all log sectors are 4096 bytes), |
| 121 | * into buffer 'buffer'. Upon return, *sectors_read will contain |
| 122 | * the number of sectors successfully read. |
| 123 | * |
| 124 | * It is assumed that 'buffer' is already allocated, and of sufficient |
| 125 | * size (i.e. >= 4096*num_sectors). |
| 126 | * |
| 127 | * If 'peek' is true, then the tail (read) pointer for the circular buffer is |
| 128 | * not modified. |
| 129 | * |
| 130 | * 0 is returned on success, -errno otherwise. */ |
| 131 | static int GRAPH_RDLOCK |
| 132 | vhdx_log_read_sectors(BlockDriverState *bs, VHDXLogEntries *log, |
| 133 | uint32_t *sectors_read, void *buffer, |
| 134 | uint32_t num_sectors, bool peek) |
| 135 | { |
| 136 | int ret = 0; |
| 137 | uint64_t offset; |
| 138 | uint32_t read; |
| 139 | |
| 140 | read = log->read; |
| 141 | |
| 142 | *sectors_read = 0; |
| 143 | while (num_sectors) { |
| 144 | if (read == log->write) { |
| 145 | /* empty */ |
| 146 | break; |
| 147 | } |
| 148 | offset = log->offset + read; |
| 149 | |
| 150 | ret = bdrv_pread(bs->file, offset, VHDX_LOG_SECTOR_SIZE, buffer, 0); |
| 151 | if (ret < 0) { |
| 152 | goto exit; |
| 153 | } |
| 154 | read = vhdx_log_inc_idx(read, log->length); |
| 155 | |
| 156 | *sectors_read = *sectors_read + 1; |
| 157 | num_sectors--; |
| 158 | } |
| 159 | |
| 160 | exit: |
| 161 | if (!peek) { |
| 162 | log->read = read; |
| 163 | } |
| 164 | return ret; |
| 165 | } |
| 166 | |
| 167 | /* Writes num_sectors to the log (all log sectors are 4096 bytes), |
| 168 | * from buffer 'buffer'. Upon return, *sectors_written will contain |
| 169 | * the number of sectors successfully written. |
| 170 | * |
| 171 | * It is assumed that 'buffer' is at least 4096*num_sectors large. |
| 172 | * |
| 173 | * 0 is returned on success, -errno otherwise */ |
| 174 | static int coroutine_fn GRAPH_RDLOCK |
| 175 | vhdx_log_write_sectors(BlockDriverState *bs, VHDXLogEntries *log, |
| 176 | uint32_t *sectors_written, void *buffer, |
| 177 | uint32_t num_sectors) |
| 178 | { |
| 179 | int ret = 0; |
| 180 | uint64_t offset; |
| 181 | uint32_t write; |
| 182 | void *buffer_tmp; |
| 183 | BDRVVHDXState *s = bs->opaque; |
| 184 | |
| 185 | ret = vhdx_user_visible_write(bs, s); |
| 186 | if (ret < 0) { |
| 187 | goto exit; |
| 188 | } |
| 189 | |
| 190 | write = log->write; |
| 191 | |
| 192 | buffer_tmp = buffer; |
| 193 | while (num_sectors) { |
| 194 | |
| 195 | offset = log->offset + write; |
| 196 | write = vhdx_log_inc_idx(write, log->length); |
| 197 | if (write == log->read) { |
| 198 | /* full */ |
| 199 | break; |
| 200 | } |
| 201 | ret = bdrv_co_pwrite(bs->file, offset, VHDX_LOG_SECTOR_SIZE, buffer_tmp, 0); |
| 202 | if (ret < 0) { |
| 203 | goto exit; |
| 204 | } |
| 205 | buffer_tmp += VHDX_LOG_SECTOR_SIZE; |
| 206 | |
| 207 | log->write = write; |
| 208 | *sectors_written = *sectors_written + 1; |
| 209 | num_sectors--; |
| 210 | } |
| 211 | |
| 212 | exit: |
| 213 | return ret; |
| 214 | } |
| 215 | |
| 216 | |
| 217 | /* Validates a log entry header */ |
| 218 | static bool vhdx_log_hdr_is_valid(VHDXLogEntries *log, VHDXLogEntryHeader *hdr, |
| 219 | BDRVVHDXState *s) |
| 220 | { |
| 221 | int valid = false; |
| 222 | |
| 223 | if (hdr->signature != VHDX_LOG_SIGNATURE) { |
| 224 | goto exit; |
| 225 | } |
| 226 | |
| 227 | /* if the individual entry length is larger than the whole log |
| 228 | * buffer, that is obviously invalid */ |
| 229 | if (log->length < hdr->entry_length) { |
| 230 | goto exit; |
| 231 | } |
| 232 | |
| 233 | /* length of entire entry must be in units of 4KB (log sector size) */ |
| 234 | if (hdr->entry_length % (VHDX_LOG_SECTOR_SIZE)) { |
| 235 | goto exit; |
| 236 | } |
| 237 | |
| 238 | /* per spec, sequence # must be > 0 */ |
| 239 | if (hdr->sequence_number == 0) { |
| 240 | goto exit; |
| 241 | } |
| 242 | |
| 243 | /* log entries are only valid if they match the file-wide log guid |
| 244 | * found in the active header */ |
| 245 | if (!guid_eq(hdr->log_guid, s->headers[s->curr_header]->log_guid)) { |
| 246 | goto exit; |
| 247 | } |
| 248 | |
| 249 | if (hdr->descriptor_count * sizeof(VHDXLogDescriptor) > hdr->entry_length) { |
| 250 | goto exit; |
| 251 | } |
| 252 | |
| 253 | valid = true; |
| 254 | |
| 255 | exit: |
| 256 | return valid; |
| 257 | } |
| 258 | |
| 259 | /* |
| 260 | * Given a log header, this will validate that the descriptors and the |
| 261 | * corresponding data sectors (if applicable) |
| 262 | * |
| 263 | * Validation consists of: |
| 264 | * 1. Making sure the sequence numbers matches the entry header |
| 265 | * 2. Verifying a valid signature ('zero' or 'desc' for descriptors) |
| 266 | * 3. File offset field is a multiple of 4KB |
| 267 | * 4. If a data descriptor, the corresponding data sector |
| 268 | * has its signature ('data') and matching sequence number |
| 269 | * |
| 270 | * @desc: the data buffer containing the descriptor |
| 271 | * @hdr: the log entry header |
| 272 | * |
| 273 | * Returns true if valid |
| 274 | */ |
| 275 | static bool vhdx_log_desc_is_valid(VHDXLogDescriptor *desc, |
| 276 | VHDXLogEntryHeader *hdr) |
| 277 | { |
| 278 | bool ret = false; |
| 279 | |
| 280 | if (desc->sequence_number != hdr->sequence_number) { |
| 281 | goto exit; |
| 282 | } |
| 283 | if (desc->file_offset % VHDX_LOG_SECTOR_SIZE) { |
| 284 | goto exit; |
| 285 | } |
| 286 | |
| 287 | if (desc->signature == VHDX_LOG_ZERO_SIGNATURE) { |
| 288 | if (desc->zero_length % VHDX_LOG_SECTOR_SIZE == 0) { |
| 289 | /* valid */ |
| 290 | ret = true; |
| 291 | } |
| 292 | } else if (desc->signature == VHDX_LOG_DESC_SIGNATURE) { |
| 293 | /* valid */ |
| 294 | ret = true; |
| 295 | } |
| 296 | |
| 297 | exit: |
| 298 | return ret; |
| 299 | } |
| 300 | |
| 301 | |
| 302 | /* Prior to sector data for a log entry, there is the header |
| 303 | * and the descriptors referenced in the header: |
| 304 | * |
| 305 | * [] = 4KB sector |
| 306 | * |
| 307 | * [ hdr, desc ][ desc ][ ... ][ data ][ ... ] |
| 308 | * |
| 309 | * The first sector in a log entry has a 64 byte header, and |
| 310 | * up to 126 32-byte descriptors. If more descriptors than |
| 311 | * 126 are required, then subsequent sectors can have up to 128 |
| 312 | * descriptors. Each sector is 4KB. Data follows the descriptor |
| 313 | * sectors. |
| 314 | * |
| 315 | * This will return the number of sectors needed to encompass |
| 316 | * the passed number of descriptors in desc_cnt. |
| 317 | * |
| 318 | * This will never return 0, even if desc_cnt is 0. |
| 319 | */ |
| 320 | static int vhdx_compute_desc_sectors(uint32_t desc_cnt) |
| 321 | { |
| 322 | uint32_t desc_sectors; |
| 323 | |
| 324 | desc_cnt += 2; /* account for header in first sector */ |
| 325 | desc_sectors = desc_cnt / 128; |
| 326 | if (desc_cnt % 128) { |
| 327 | desc_sectors++; |
| 328 | } |
| 329 | |
| 330 | return desc_sectors; |
| 331 | } |
| 332 | |
| 333 | |
| 334 | /* Reads the log header, and subsequent descriptors (if any). This |
| 335 | * will allocate all the space for buffer, which must be NULL when |
| 336 | * passed into this function. Each descriptor will also be validated, |
| 337 | * and error returned if any are invalid. */ |
| 338 | static int GRAPH_RDLOCK |
| 339 | vhdx_log_read_desc(BlockDriverState *bs, BDRVVHDXState *s, VHDXLogEntries *log, |
| 340 | VHDXLogDescEntries **buffer, bool convert_endian) |
| 341 | { |
| 342 | int ret = 0; |
| 343 | uint32_t desc_sectors; |
| 344 | uint32_t sectors_read; |
| 345 | VHDXLogEntryHeader hdr; |
| 346 | VHDXLogDescEntries *desc_entries = NULL; |
| 347 | VHDXLogDescriptor desc; |
| 348 | int i; |
| 349 | |
| 350 | assert(*buffer == NULL); |
| 351 | |
| 352 | ret = vhdx_log_peek_hdr(bs, log, &hdr); |
| 353 | if (ret < 0) { |
| 354 | goto exit; |
| 355 | } |
| 356 | |
| 357 | if (vhdx_log_hdr_is_valid(log, &hdr, s) == false) { |
| 358 | ret = -EINVAL; |
| 359 | goto exit; |
| 360 | } |
| 361 | |
| 362 | desc_sectors = vhdx_compute_desc_sectors(hdr.descriptor_count); |
| 363 | desc_entries = qemu_try_blockalign(bs->file->bs, |
| 364 | desc_sectors * VHDX_LOG_SECTOR_SIZE); |
| 365 | if (desc_entries == NULL) { |
| 366 | ret = -ENOMEM; |
| 367 | goto exit; |
| 368 | } |
| 369 | |
| 370 | ret = vhdx_log_read_sectors(bs, log, §ors_read, desc_entries, |
| 371 | desc_sectors, false); |
| 372 | if (ret < 0) { |
| 373 | goto free_and_exit; |
| 374 | } |
| 375 | if (sectors_read != desc_sectors) { |
| 376 | ret = -EINVAL; |
| 377 | goto free_and_exit; |
| 378 | } |
| 379 | |
| 380 | /* put in proper endianness, and validate each desc */ |
| 381 | for (i = 0; i < hdr.descriptor_count; i++) { |
| 382 | desc = desc_entries->desc[i]; |
| 383 | vhdx_log_desc_le_import(&desc); |
| 384 | if (convert_endian) { |
| 385 | desc_entries->desc[i] = desc; |
| 386 | } |
| 387 | if (vhdx_log_desc_is_valid(&desc, &hdr) == false) { |
| 388 | ret = -EINVAL; |
| 389 | goto free_and_exit; |
| 390 | } |
| 391 | } |
| 392 | if (convert_endian) { |
| 393 | desc_entries->hdr = hdr; |
| 394 | } |
| 395 | |
| 396 | *buffer = desc_entries; |
| 397 | goto exit; |
| 398 | |
| 399 | free_and_exit: |
| 400 | qemu_vfree(desc_entries); |
| 401 | exit: |
| 402 | return ret; |
| 403 | } |
| 404 | |
| 405 | |
| 406 | /* Flushes the descriptor described by desc to the VHDX image file. |
| 407 | * If the descriptor is a data descriptor, than 'data' must be non-NULL, |
| 408 | * and >= 4096 bytes (VHDX_LOG_SECTOR_SIZE), containing the data to be |
| 409 | * written. |
| 410 | * |
| 411 | * Verification is performed to make sure the sequence numbers of a data |
| 412 | * descriptor match the sequence number in the desc. |
| 413 | * |
| 414 | * For a zero descriptor, it may describe multiple sectors to fill with zeroes. |
| 415 | * In this case, it should be noted that zeroes are written to disk, and the |
| 416 | * image file is not extended as a sparse file. */ |
| 417 | static int GRAPH_RDLOCK |
| 418 | vhdx_log_flush_desc(BlockDriverState *bs, VHDXLogDescriptor *desc, |
| 419 | VHDXLogDataSector *data) |
| 420 | { |
| 421 | int ret = 0; |
| 422 | uint64_t seq, file_offset; |
| 423 | uint32_t offset = 0; |
| 424 | void *buffer = NULL; |
| 425 | uint64_t count = 1; |
| 426 | int i; |
| 427 | |
| 428 | buffer = qemu_blockalign(bs, VHDX_LOG_SECTOR_SIZE); |
| 429 | |
| 430 | if (desc->signature == VHDX_LOG_DESC_SIGNATURE) { |
| 431 | /* data sector */ |
| 432 | if (data == NULL) { |
| 433 | ret = -EFAULT; |
| 434 | goto exit; |
| 435 | } |
| 436 | |
| 437 | /* The sequence number of the data sector must match that |
| 438 | * in the descriptor */ |
| 439 | seq = data->sequence_high; |
| 440 | seq <<= 32; |
| 441 | seq |= data->sequence_low & 0xffffffff; |
| 442 | |
| 443 | if (seq != desc->sequence_number) { |
| 444 | ret = -EINVAL; |
| 445 | goto exit; |
| 446 | } |
| 447 | |
| 448 | /* Each data sector is in total 4096 bytes, however the first |
| 449 | * 8 bytes, and last 4 bytes, are located in the descriptor */ |
| 450 | memcpy(buffer, &desc->leading_bytes, 8); |
| 451 | offset += 8; |
| 452 | |
| 453 | memcpy(buffer+offset, data->data, 4084); |
| 454 | offset += 4084; |
| 455 | |
| 456 | memcpy(buffer+offset, &desc->trailing_bytes, 4); |
| 457 | |
| 458 | } else if (desc->signature == VHDX_LOG_ZERO_SIGNATURE) { |
| 459 | /* write 'count' sectors of sector */ |
| 460 | memset(buffer, 0, VHDX_LOG_SECTOR_SIZE); |
| 461 | count = desc->zero_length / VHDX_LOG_SECTOR_SIZE; |
| 462 | } else { |
| 463 | error_report("Invalid VHDX log descriptor entry signature 0x%" PRIx32, |
| 464 | desc->signature); |
| 465 | ret = -EINVAL; |
| 466 | goto exit; |
| 467 | } |
| 468 | |
| 469 | file_offset = desc->file_offset; |
| 470 | |
| 471 | /* count is only > 1 if we are writing zeroes */ |
| 472 | for (i = 0; i < count; i++) { |
| 473 | ret = bdrv_pwrite_sync(bs->file, file_offset, VHDX_LOG_SECTOR_SIZE, |
| 474 | buffer, 0); |
| 475 | if (ret < 0) { |
| 476 | goto exit; |
| 477 | } |
| 478 | file_offset += VHDX_LOG_SECTOR_SIZE; |
| 479 | } |
| 480 | |
| 481 | exit: |
| 482 | qemu_vfree(buffer); |
| 483 | return ret; |
| 484 | } |
| 485 | |
| 486 | /* Flush the entire log (as described by 'logs') to the VHDX image |
| 487 | * file, and then set the log to 'empty' status once complete. |
| 488 | * |
| 489 | * The log entries should be validate prior to flushing */ |
| 490 | static int GRAPH_RDLOCK |
| 491 | vhdx_log_flush(BlockDriverState *bs, BDRVVHDXState *s, VHDXLogSequence *logs) |
| 492 | { |
| 493 | int ret = 0; |
| 494 | int i; |
| 495 | uint32_t cnt, sectors_read; |
| 496 | uint64_t new_file_size; |
| 497 | void *data = NULL; |
| 498 | int64_t file_length; |
| 499 | VHDXLogDescEntries *desc_entries = NULL; |
| 500 | VHDXLogEntryHeader hdr_tmp = { 0 }; |
| 501 | |
| 502 | cnt = logs->count; |
| 503 | |
| 504 | data = qemu_blockalign(bs, VHDX_LOG_SECTOR_SIZE); |
| 505 | |
| 506 | ret = vhdx_user_visible_write(bs, s); |
| 507 | if (ret < 0) { |
| 508 | goto exit; |
| 509 | } |
| 510 | |
| 511 | /* each iteration represents one log sequence, which may span multiple |
| 512 | * sectors */ |
| 513 | while (cnt--) { |
| 514 | ret = vhdx_log_peek_hdr(bs, &logs->log, &hdr_tmp); |
| 515 | if (ret < 0) { |
| 516 | goto exit; |
| 517 | } |
| 518 | file_length = bdrv_getlength(bs->file->bs); |
| 519 | if (file_length < 0) { |
| 520 | ret = file_length; |
| 521 | goto exit; |
| 522 | } |
| 523 | /* if the log shows a FlushedFileOffset larger than our current file |
| 524 | * size, then that means the file has been truncated / corrupted, and |
| 525 | * we must refused to open it / use it */ |
| 526 | if (hdr_tmp.flushed_file_offset > file_length) { |
| 527 | ret = -EINVAL; |
| 528 | goto exit; |
| 529 | } |
| 530 | |
| 531 | ret = vhdx_log_read_desc(bs, s, &logs->log, &desc_entries, true); |
| 532 | if (ret < 0) { |
| 533 | goto exit; |
| 534 | } |
| 535 | |
| 536 | for (i = 0; i < desc_entries->hdr.descriptor_count; i++) { |
| 537 | if (desc_entries->desc[i].signature == VHDX_LOG_DESC_SIGNATURE) { |
| 538 | /* data sector, so read a sector to flush */ |
| 539 | ret = vhdx_log_read_sectors(bs, &logs->log, §ors_read, |
| 540 | data, 1, false); |
| 541 | if (ret < 0) { |
| 542 | goto exit; |
| 543 | } |
| 544 | if (sectors_read != 1) { |
| 545 | ret = -EINVAL; |
| 546 | goto exit; |
| 547 | } |
| 548 | vhdx_log_data_le_import(data); |
| 549 | } |
| 550 | |
| 551 | ret = vhdx_log_flush_desc(bs, &desc_entries->desc[i], data); |
| 552 | if (ret < 0) { |
| 553 | goto exit; |
| 554 | } |
| 555 | } |
| 556 | if (file_length < desc_entries->hdr.last_file_offset) { |
| 557 | new_file_size = desc_entries->hdr.last_file_offset; |
| 558 | if (new_file_size % (1 * MiB)) { |
| 559 | /* round up to nearest 1MB boundary */ |
| 560 | new_file_size = QEMU_ALIGN_UP(new_file_size, MiB); |
| 561 | if (new_file_size > INT64_MAX) { |
| 562 | ret = -EINVAL; |
| 563 | goto exit; |
| 564 | } |
| 565 | ret = bdrv_truncate(bs->file, new_file_size, false, |
| 566 | PREALLOC_MODE_OFF, 0, NULL); |
| 567 | if (ret < 0) { |
| 568 | goto exit; |
| 569 | } |
| 570 | } |
| 571 | } |
| 572 | qemu_vfree(desc_entries); |
| 573 | desc_entries = NULL; |
| 574 | } |
| 575 | |
| 576 | ret = bdrv_flush(bs); |
| 577 | if (ret < 0) { |
| 578 | goto exit; |
| 579 | } |
| 580 | /* once the log is fully flushed, indicate that we have an empty log |
| 581 | * now. This also sets the log guid to 0, to indicate an empty log */ |
| 582 | vhdx_log_reset(bs, s); |
| 583 | |
| 584 | exit: |
| 585 | qemu_vfree(data); |
| 586 | qemu_vfree(desc_entries); |
| 587 | return ret; |
| 588 | } |
| 589 | |
| 590 | static int GRAPH_RDLOCK |
| 591 | vhdx_validate_log_entry(BlockDriverState *bs, BDRVVHDXState *s, |
| 592 | VHDXLogEntries *log, uint64_t seq, |
| 593 | bool *valid, VHDXLogEntryHeader *entry) |
| 594 | { |
| 595 | int ret = 0; |
| 596 | VHDXLogEntryHeader hdr; |
| 597 | void *buffer = NULL; |
| 598 | uint32_t i, desc_sectors, total_sectors, crc; |
| 599 | uint32_t sectors_read = 0; |
| 600 | VHDXLogDescEntries *desc_buffer = NULL; |
| 601 | |
| 602 | *valid = false; |
| 603 | |
| 604 | ret = vhdx_log_peek_hdr(bs, log, &hdr); |
| 605 | if (ret < 0) { |
| 606 | goto inc_and_exit; |
| 607 | } |
| 608 | |
| 609 | if (vhdx_log_hdr_is_valid(log, &hdr, s) == false) { |
| 610 | goto inc_and_exit; |
| 611 | } |
| 612 | |
| 613 | if (seq > 0) { |
| 614 | if (hdr.sequence_number != seq + 1) { |
| 615 | goto inc_and_exit; |
| 616 | } |
| 617 | } |
| 618 | |
| 619 | desc_sectors = vhdx_compute_desc_sectors(hdr.descriptor_count); |
| 620 | |
| 621 | /* Read all log sectors, and calculate log checksum */ |
| 622 | |
| 623 | total_sectors = hdr.entry_length / VHDX_LOG_SECTOR_SIZE; |
| 624 | |
| 625 | |
| 626 | /* read_desc() will increment the read idx */ |
| 627 | ret = vhdx_log_read_desc(bs, s, log, &desc_buffer, false); |
| 628 | if (ret < 0) { |
| 629 | goto free_and_exit; |
| 630 | } |
| 631 | |
| 632 | crc = vhdx_checksum_calc(0xffffffff, (void *)desc_buffer, |
| 633 | desc_sectors * VHDX_LOG_SECTOR_SIZE, 4); |
| 634 | crc ^= 0xffffffff; |
| 635 | |
| 636 | buffer = qemu_blockalign(bs, VHDX_LOG_SECTOR_SIZE); |
| 637 | if (total_sectors > desc_sectors) { |
| 638 | for (i = 0; i < total_sectors - desc_sectors; i++) { |
| 639 | sectors_read = 0; |
| 640 | ret = vhdx_log_read_sectors(bs, log, §ors_read, buffer, |
| 641 | 1, false); |
| 642 | if (ret < 0 || sectors_read != 1) { |
| 643 | goto free_and_exit; |
| 644 | } |
| 645 | crc = vhdx_checksum_calc(crc, buffer, VHDX_LOG_SECTOR_SIZE, -1); |
| 646 | crc ^= 0xffffffff; |
| 647 | } |
| 648 | } |
| 649 | crc ^= 0xffffffff; |
| 650 | if (crc != hdr.checksum) { |
| 651 | goto free_and_exit; |
| 652 | } |
| 653 | |
| 654 | *valid = true; |
| 655 | *entry = hdr; |
| 656 | goto free_and_exit; |
| 657 | |
| 658 | inc_and_exit: |
| 659 | log->read = vhdx_log_inc_idx(log->read, log->length); |
| 660 | |
| 661 | free_and_exit: |
| 662 | qemu_vfree(buffer); |
| 663 | qemu_vfree(desc_buffer); |
| 664 | return ret; |
| 665 | } |
| 666 | |
| 667 | /* Search through the log circular buffer, and find the valid, active |
| 668 | * log sequence, if any exists |
| 669 | * */ |
| 670 | static int GRAPH_RDLOCK |
| 671 | vhdx_log_search(BlockDriverState *bs, BDRVVHDXState *s, VHDXLogSequence *logs) |
| 672 | { |
| 673 | int ret = 0; |
| 674 | uint32_t tail; |
| 675 | bool seq_valid = false; |
| 676 | VHDXLogSequence candidate = { 0 }; |
| 677 | VHDXLogEntryHeader hdr = { 0 }; |
| 678 | VHDXLogEntries curr_log; |
| 679 | |
| 680 | memcpy(&curr_log, &s->log, sizeof(VHDXLogEntries)); |
| 681 | curr_log.write = curr_log.length; /* assume log is full */ |
| 682 | curr_log.read = 0; |
| 683 | |
| 684 | |
| 685 | /* now we will go through the whole log sector by sector, until |
| 686 | * we find a valid, active log sequence, or reach the end of the |
| 687 | * log buffer */ |
| 688 | for (;;) { |
| 689 | uint64_t curr_seq = 0; |
| 690 | VHDXLogSequence current = { 0 }; |
| 691 | |
| 692 | tail = curr_log.read; |
| 693 | |
| 694 | ret = vhdx_validate_log_entry(bs, s, &curr_log, curr_seq, |
| 695 | &seq_valid, &hdr); |
| 696 | if (ret < 0) { |
| 697 | goto exit; |
| 698 | } |
| 699 | |
| 700 | if (seq_valid) { |
| 701 | current.valid = true; |
| 702 | current.log = curr_log; |
| 703 | current.log.read = tail; |
| 704 | current.log.write = curr_log.read; |
| 705 | current.count = 1; |
| 706 | current.hdr = hdr; |
| 707 | |
| 708 | |
| 709 | for (;;) { |
| 710 | ret = vhdx_validate_log_entry(bs, s, &curr_log, curr_seq, |
| 711 | &seq_valid, &hdr); |
| 712 | if (ret < 0) { |
| 713 | goto exit; |
| 714 | } |
| 715 | if (seq_valid == false) { |
| 716 | break; |
| 717 | } |
| 718 | current.log.write = curr_log.read; |
| 719 | current.count++; |
| 720 | |
| 721 | curr_seq = hdr.sequence_number; |
| 722 | } |
| 723 | } |
| 724 | |
| 725 | if (current.valid) { |
| 726 | if (candidate.valid == false || |
| 727 | current.hdr.sequence_number > candidate.hdr.sequence_number) { |
| 728 | candidate = current; |
| 729 | } |
| 730 | } |
| 731 | |
| 732 | if (curr_log.read < tail) { |
| 733 | break; |
| 734 | } |
| 735 | } |
| 736 | |
| 737 | *logs = candidate; |
| 738 | |
| 739 | if (candidate.valid) { |
| 740 | /* this is the next sequence number, for writes */ |
| 741 | s->log.sequence = candidate.hdr.sequence_number + 1; |
| 742 | } |
| 743 | |
| 744 | |
| 745 | exit: |
| 746 | return ret; |
| 747 | } |
| 748 | |
| 749 | /* Parse the replay log. Per the VHDX spec, if the log is present |
| 750 | * it must be replayed prior to opening the file, even read-only. |
| 751 | * |
| 752 | * If read-only, we must replay the log in RAM (or refuse to open |
| 753 | * a dirty VHDX file read-only) */ |
| 754 | int vhdx_parse_log(BlockDriverState *bs, BDRVVHDXState *s, bool *flushed, |
| 755 | Error **errp) |
| 756 | { |
| 757 | int ret = 0; |
| 758 | VHDXHeader *hdr; |
| 759 | VHDXLogSequence logs = { 0 }; |
| 760 | |
| 761 | hdr = s->headers[s->curr_header]; |
| 762 | |
| 763 | *flushed = false; |
| 764 | |
| 765 | /* s->log.hdr is freed in vhdx_close() */ |
| 766 | if (s->log.hdr == NULL) { |
| 767 | s->log.hdr = qemu_blockalign(bs, sizeof(VHDXLogEntryHeader)); |
| 768 | } |
| 769 | |
| 770 | s->log.offset = hdr->log_offset; |
| 771 | s->log.length = hdr->log_length; |
| 772 | |
| 773 | if (s->log.offset < VHDX_LOG_MIN_SIZE || |
| 774 | s->log.offset % VHDX_LOG_MIN_SIZE) { |
| 775 | ret = -EINVAL; |
| 776 | goto exit; |
| 777 | } |
| 778 | |
| 779 | /* per spec, only log version of 0 is supported */ |
| 780 | if (hdr->log_version != 0) { |
| 781 | ret = -EINVAL; |
| 782 | goto exit; |
| 783 | } |
| 784 | |
| 785 | /* If either the log guid, or log length is zero, |
| 786 | * then a replay log is not present */ |
| 787 | if (guid_eq(hdr->log_guid, zero_guid)) { |
| 788 | goto exit; |
| 789 | } |
| 790 | |
| 791 | if (hdr->log_length == 0) { |
| 792 | goto exit; |
| 793 | } |
| 794 | |
| 795 | if (hdr->log_length % VHDX_LOG_MIN_SIZE) { |
| 796 | ret = -EINVAL; |
| 797 | goto exit; |
| 798 | } |
| 799 | |
| 800 | |
| 801 | /* The log is present, we need to find if and where there is an active |
| 802 | * sequence of valid entries present in the log. */ |
| 803 | |
| 804 | ret = vhdx_log_search(bs, s, &logs); |
| 805 | if (ret < 0) { |
| 806 | goto exit; |
| 807 | } |
| 808 | |
| 809 | if (logs.valid) { |
| 810 | if (bdrv_is_read_only(bs)) { |
| 811 | bdrv_refresh_filename(bs); |
| 812 | ret = -EPERM; |
| 813 | error_setg(errp, |
| 814 | "VHDX image file '%s' opened read-only, but " |
| 815 | "contains a log that needs to be replayed", |
| 816 | bs->filename); |
| 817 | error_append_hint(errp, "To replay the log, run:\n" |
| 818 | "qemu-img check -r all '%s'\n", |
| 819 | bs->filename); |
| 820 | goto exit; |
| 821 | } |
| 822 | /* now flush the log */ |
| 823 | ret = vhdx_log_flush(bs, s, &logs); |
| 824 | if (ret < 0) { |
| 825 | goto exit; |
| 826 | } |
| 827 | *flushed = true; |
| 828 | } |
| 829 | |
| 830 | |
| 831 | exit: |
| 832 | return ret; |
| 833 | } |
| 834 | |
| 835 | |
| 836 | |
| 837 | static void vhdx_log_raw_to_le_sector(VHDXLogDescriptor *desc, |
| 838 | VHDXLogDataSector *sector, void *data, |
| 839 | uint64_t seq) |
| 840 | { |
| 841 | /* 8 + 4084 + 4 = 4096, 1 log sector */ |
| 842 | memcpy(&desc->leading_bytes, data, 8); |
| 843 | data += 8; |
| 844 | desc->leading_bytes = cpu_to_le64(desc->leading_bytes); |
| 845 | memcpy(sector->data, data, 4084); |
| 846 | data += 4084; |
| 847 | memcpy(&desc->trailing_bytes, data, 4); |
| 848 | desc->trailing_bytes = cpu_to_le32(desc->trailing_bytes); |
| 849 | data += 4; |
| 850 | |
| 851 | sector->sequence_high = (uint32_t) (seq >> 32); |
| 852 | sector->sequence_low = (uint32_t) (seq & 0xffffffff); |
| 853 | sector->data_signature = VHDX_LOG_DATA_SIGNATURE; |
| 854 | |
| 855 | vhdx_log_desc_le_export(desc); |
| 856 | vhdx_log_data_le_export(sector); |
| 857 | } |
| 858 | |
| 859 | |
| 860 | static int coroutine_fn GRAPH_RDLOCK |
| 861 | vhdx_log_write(BlockDriverState *bs, BDRVVHDXState *s, |
| 862 | void *data, uint32_t length, uint64_t offset) |
| 863 | { |
| 864 | int ret = 0; |
| 865 | void *buffer = NULL; |
| 866 | void *merged_sector = NULL; |
| 867 | void *data_tmp, *sector_write; |
| 868 | unsigned int i; |
| 869 | int sector_offset; |
| 870 | uint32_t desc_sectors, sectors, total_length; |
| 871 | uint32_t sectors_written = 0; |
| 872 | uint32_t aligned_length; |
| 873 | uint32_t leading_length = 0; |
| 874 | uint32_t trailing_length = 0; |
| 875 | uint32_t partial_sectors = 0; |
| 876 | uint32_t bytes_written = 0; |
| 877 | uint64_t file_offset; |
| 878 | int64_t file_length; |
| 879 | VHDXHeader *header; |
| 880 | VHDXLogEntryHeader new_hdr; |
| 881 | VHDXLogDescriptor *new_desc = NULL; |
| 882 | VHDXLogDataSector *data_sector = NULL; |
| 883 | MSGUID new_guid = { 0 }; |
| 884 | |
| 885 | header = s->headers[s->curr_header]; |
| 886 | |
| 887 | /* need to have offset read data, and be on 4096 byte boundary */ |
| 888 | |
| 889 | if (length > header->log_length) { |
| 890 | /* no log present. we could create a log here instead of failing */ |
| 891 | ret = -EINVAL; |
| 892 | goto exit; |
| 893 | } |
| 894 | |
| 895 | if (guid_eq(header->log_guid, zero_guid)) { |
| 896 | vhdx_guid_generate(&new_guid); |
| 897 | vhdx_update_headers(bs, s, false, &new_guid); |
| 898 | } else { |
| 899 | /* currently, we require that the log be flushed after |
| 900 | * every write. */ |
| 901 | ret = -ENOTSUP; |
| 902 | goto exit; |
| 903 | } |
| 904 | |
| 905 | /* 0 is an invalid sequence number, but may also represent the first |
| 906 | * log write (or a wrapped seq) */ |
| 907 | if (s->log.sequence == 0) { |
| 908 | s->log.sequence = 1; |
| 909 | } |
| 910 | |
| 911 | sector_offset = offset % VHDX_LOG_SECTOR_SIZE; |
| 912 | file_offset = QEMU_ALIGN_DOWN(offset, VHDX_LOG_SECTOR_SIZE); |
| 913 | |
| 914 | aligned_length = length; |
| 915 | |
| 916 | /* add in the unaligned head and tail bytes */ |
| 917 | if (sector_offset) { |
| 918 | leading_length = (VHDX_LOG_SECTOR_SIZE - sector_offset); |
| 919 | leading_length = leading_length > length ? length : leading_length; |
| 920 | aligned_length -= leading_length; |
| 921 | partial_sectors++; |
| 922 | } |
| 923 | |
| 924 | sectors = aligned_length / VHDX_LOG_SECTOR_SIZE; |
| 925 | trailing_length = aligned_length - (sectors * VHDX_LOG_SECTOR_SIZE); |
| 926 | if (trailing_length) { |
| 927 | partial_sectors++; |
| 928 | } |
| 929 | |
| 930 | sectors += partial_sectors; |
| 931 | |
| 932 | file_length = bdrv_co_getlength(bs->file->bs); |
| 933 | if (file_length < 0) { |
| 934 | ret = file_length; |
| 935 | goto exit; |
| 936 | } |
| 937 | |
| 938 | /* sectors is now how many sectors the data itself takes, not |
| 939 | * including the header and descriptor metadata */ |
| 940 | |
| 941 | new_hdr = (VHDXLogEntryHeader) { |
| 942 | .signature = VHDX_LOG_SIGNATURE, |
| 943 | .tail = s->log.tail, |
| 944 | .sequence_number = s->log.sequence, |
| 945 | .descriptor_count = sectors, |
| 946 | .reserved = 0, |
| 947 | .flushed_file_offset = file_length, |
| 948 | .last_file_offset = file_length, |
| 949 | .log_guid = header->log_guid, |
| 950 | }; |
| 951 | |
| 952 | |
| 953 | desc_sectors = vhdx_compute_desc_sectors(new_hdr.descriptor_count); |
| 954 | |
| 955 | total_length = (desc_sectors + sectors) * VHDX_LOG_SECTOR_SIZE; |
| 956 | new_hdr.entry_length = total_length; |
| 957 | |
| 958 | vhdx_log_entry_hdr_le_export(&new_hdr); |
| 959 | |
| 960 | buffer = qemu_blockalign(bs, total_length); |
| 961 | memcpy(buffer, &new_hdr, sizeof(new_hdr)); |
| 962 | |
| 963 | new_desc = buffer + sizeof(new_hdr); |
| 964 | data_sector = buffer + (desc_sectors * VHDX_LOG_SECTOR_SIZE); |
| 965 | data_tmp = data; |
| 966 | |
| 967 | /* All log sectors are 4KB, so for any partial sectors we must |
| 968 | * merge the data with preexisting data from the final file |
| 969 | * destination */ |
| 970 | merged_sector = qemu_blockalign(bs, VHDX_LOG_SECTOR_SIZE); |
| 971 | |
| 972 | for (i = 0; i < sectors; i++) { |
| 973 | new_desc->signature = VHDX_LOG_DESC_SIGNATURE; |
| 974 | new_desc->sequence_number = s->log.sequence; |
| 975 | new_desc->file_offset = file_offset; |
| 976 | |
| 977 | if (i == 0 && leading_length) { |
| 978 | /* partial sector at the front of the buffer */ |
| 979 | ret = bdrv_co_pread(bs->file, file_offset, VHDX_LOG_SECTOR_SIZE, |
| 980 | merged_sector, 0); |
| 981 | if (ret < 0) { |
| 982 | goto exit; |
| 983 | } |
| 984 | memcpy(merged_sector + sector_offset, data_tmp, leading_length); |
| 985 | bytes_written = leading_length; |
| 986 | sector_write = merged_sector; |
| 987 | } else if (i == sectors - 1 && trailing_length) { |
| 988 | /* partial sector at the end of the buffer */ |
| 989 | ret = bdrv_co_pread(bs->file, file_offset + trailing_length, |
| 990 | VHDX_LOG_SECTOR_SIZE - trailing_length, |
| 991 | merged_sector + trailing_length, 0); |
| 992 | if (ret < 0) { |
| 993 | goto exit; |
| 994 | } |
| 995 | memcpy(merged_sector, data_tmp, trailing_length); |
| 996 | bytes_written = trailing_length; |
| 997 | sector_write = merged_sector; |
| 998 | } else { |
| 999 | bytes_written = VHDX_LOG_SECTOR_SIZE; |
| 1000 | sector_write = data_tmp; |
| 1001 | } |
| 1002 | |
| 1003 | /* populate the raw sector data into the proper structures, |
| 1004 | * as well as update the descriptor, and convert to proper |
| 1005 | * endianness */ |
| 1006 | vhdx_log_raw_to_le_sector(new_desc, data_sector, sector_write, |
| 1007 | s->log.sequence); |
| 1008 | |
| 1009 | data_tmp += bytes_written; |
| 1010 | data_sector++; |
| 1011 | new_desc++; |
| 1012 | file_offset += VHDX_LOG_SECTOR_SIZE; |
| 1013 | } |
| 1014 | |
| 1015 | /* checksum covers entire entry, from the log header through the |
| 1016 | * last data sector */ |
| 1017 | vhdx_update_checksum(buffer, total_length, |
| 1018 | offsetof(VHDXLogEntryHeader, checksum)); |
| 1019 | |
| 1020 | /* now write to the log */ |
| 1021 | ret = vhdx_log_write_sectors(bs, &s->log, §ors_written, buffer, |
| 1022 | desc_sectors + sectors); |
| 1023 | if (ret < 0) { |
| 1024 | goto exit; |
| 1025 | } |
| 1026 | |
| 1027 | if (sectors_written != desc_sectors + sectors) { |
| 1028 | /* instead of failing, we could flush the log here */ |
| 1029 | ret = -EINVAL; |
| 1030 | goto exit; |
| 1031 | } |
| 1032 | |
| 1033 | s->log.sequence++; |
| 1034 | /* write new tail */ |
| 1035 | s->log.tail = s->log.write; |
| 1036 | |
| 1037 | exit: |
| 1038 | qemu_vfree(buffer); |
| 1039 | qemu_vfree(merged_sector); |
| 1040 | return ret; |
| 1041 | } |
| 1042 | |
| 1043 | /* Perform a log write, and then immediately flush the entire log */ |
| 1044 | int coroutine_fn |
| 1045 | vhdx_log_write_and_flush(BlockDriverState *bs, BDRVVHDXState *s, |
| 1046 | void *data, uint32_t length, uint64_t offset) |
| 1047 | { |
| 1048 | int ret = 0; |
| 1049 | VHDXLogSequence logs = { .valid = true, |
| 1050 | .count = 1, |
| 1051 | .hdr = { 0 } }; |
| 1052 | |
| 1053 | |
| 1054 | /* Make sure data written (new and/or changed blocks) is stable |
| 1055 | * on disk, before creating log entry */ |
| 1056 | ret = bdrv_co_flush(bs); |
| 1057 | if (ret < 0) { |
| 1058 | goto exit; |
| 1059 | } |
| 1060 | |
| 1061 | ret = vhdx_log_write(bs, s, data, length, offset); |
| 1062 | if (ret < 0) { |
| 1063 | goto exit; |
| 1064 | } |
| 1065 | logs.log = s->log; |
| 1066 | |
| 1067 | /* Make sure log is stable on disk */ |
| 1068 | ret = bdrv_co_flush(bs); |
| 1069 | if (ret < 0) { |
| 1070 | goto exit; |
| 1071 | } |
| 1072 | |
| 1073 | ret = vhdx_log_flush(bs, s, &logs); |
| 1074 | if (ret < 0) { |
| 1075 | goto exit; |
| 1076 | } |
| 1077 | |
| 1078 | s->log = logs.log; |
| 1079 | |
| 1080 | exit: |
| 1081 | return ret; |
| 1082 | } |
| 1083 |