block/blklogwrites.c

   1 /*
   2  * Write logging blk driver based on blkverify and blkdebug.
   3  *
   4  * Copyright (c) 2017 Tuomas Tynkkynen <[email protected]>
   5  * Copyright (c) 2018 Aapo Vienamo <[email protected]>
   6  * Copyright (c) 2018 Ari Sundholm <[email protected]>
   7  *
   8  * This work is licensed under the terms of the GNU GPL, version 2 or later.
   9  * See the COPYING file in the top-level directory.
  10  */
  11
  12 #include "qemu/osdep.h"
  13 #include "qapi/error.h"
  14 #include "qemu/sockets.h" /* for EINPROGRESS on Windows */
  15 #include "block/block_int.h"
  16 #include "qapi/qmp/qdict.h"
  17 #include "qapi/qmp/qstring.h"
  18 #include "qemu/cutils.h"
  19 #include "qemu/option.h"
  20
  21 /* Disk format stuff - taken from Linux drivers/md/dm-log-writes.c */
  22
  23 #define LOG_FLUSH_FLAG   (1 << 0)
  24 #define LOG_FUA_FLAG     (1 << 1)
  25 #define LOG_DISCARD_FLAG (1 << 2)
  26 #define LOG_MARK_FLAG    (1 << 3)
  27 #define LOG_FLAG_MASK    (LOG_FLUSH_FLAG \
  28                          | LOG_FUA_FLAG \
  29                          | LOG_DISCARD_FLAG \
  30                          | LOG_MARK_FLAG)
  31
  32 #define WRITE_LOG_VERSION 1ULL
  33 #define WRITE_LOG_MAGIC 0x6a736677736872ULL
  34
  35 /* All fields are little-endian. */
  36 struct log_write_super {
  37     uint64_t magic;
  38     uint64_t version;
  39     uint64_t nr_entries;
  40     uint32_t sectorsize;
  41 } QEMU_PACKED;
  42
  43 struct log_write_entry {
  44     uint64_t sector;
  45     uint64_t nr_sectors;
  46     uint64_t flags;
  47     uint64_t data_len;
  48 } QEMU_PACKED;
  49
  50 /* End of disk format structures. */
  51
  52 typedef struct {
  53     BdrvChild *log_file;
  54     uint32_t sectorsize;
  55     uint32_t sectorbits;
  56     uint64_t cur_log_sector;
  57     uint64_t nr_entries;
  58 } BDRVBlkLogWritesState;
  59
  60 static QemuOptsList runtime_opts = {
  61     .name = "blklogwrites",
  62     .head = QTAILQ_HEAD_INITIALIZER(runtime_opts.head),
  63     .desc = {
  64         {
  65             .name = "log-append",
  66             .type = QEMU_OPT_BOOL,
  67             .help = "Append to an existing log",
  68         },
  69         {
  70             .name = "log-sector-size",
  71             .type = QEMU_OPT_SIZE,
  72             .help = "Log sector size",
  73         },
  74         { /* end of list */ }
  75     },
  76 };
  77
  78 static inline uint32_t blk_log_writes_log2(uint32_t value)
  79 {
  80     assert(value > 0);
  81     return 31 - clz32(value);
  82 }
  83
  84 static inline bool blk_log_writes_sector_size_valid(uint32_t sector_size)
  85 {
  86     return sector_size < (1ull << 24) && is_power_of_2(sector_size);
  87 }
  88
  89 static uint64_t blk_log_writes_find_cur_log_sector(BdrvChild *log,
  90                                                    uint32_t sector_size,
  91                                                    uint64_t nr_entries,
  92                                                    Error **errp)
  93 {
  94     uint64_t cur_sector = 1;
  95     uint64_t cur_idx = 0;
  96     uint32_t sector_bits = blk_log_writes_log2(sector_size);
  97     struct log_write_entry cur_entry;
  98
  99     while (cur_idx < nr_entries) {
 100         int read_ret = bdrv_pread(log, cur_sector << sector_bits, &cur_entry,
 101                                   sizeof(cur_entry));
 102         if (read_ret < 0) {
 103             error_setg_errno(errp, -read_ret,
 104                              "Failed to read log entry %"PRIu64, cur_idx);
 105             return (uint64_t)-1ull;
 106         }
 107
 108         if (cur_entry.flags & ~cpu_to_le64(LOG_FLAG_MASK)) {
 109             error_setg(errp, "Invalid flags 0x%"PRIx64" in log entry %"PRIu64,
 110                        le64_to_cpu(cur_entry.flags), cur_idx);
 111             return (uint64_t)-1ull;
 112         }
 113
 114         /* Account for the sector of the entry itself */
 115         ++cur_sector;
 116
 117         /*
 118          * Account for the data of the write.
 119          * For discards, this data is not present.
 120          */
 121         if (!(cur_entry.flags & cpu_to_le64(LOG_DISCARD_FLAG))) {
 122             cur_sector += le64_to_cpu(cur_entry.nr_sectors);
 123         }
 124
 125         ++cur_idx;
 126     }
 127
 128     return cur_sector;
 129 }
 130
 131 static int blk_log_writes_open(BlockDriverState *bs, QDict *options, int flags,
 132                                Error **errp)
 133 {
 134     BDRVBlkLogWritesState *s = bs->opaque;
 135     QemuOpts *opts;
 136     Error *local_err = NULL;
 137     int ret;
 138     uint64_t log_sector_size;
 139     bool log_append;
 140
 141     opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
 142     qemu_opts_absorb_qdict(opts, options, &local_err);
 143     if (local_err) {
 144         ret = -EINVAL;
 145         error_propagate(errp, local_err);
 146         goto fail;
 147     }
 148
 149     /* Open the file */
 150     bs->file = bdrv_open_child(NULL, options, "file", bs, &child_file, false,
 151                                &local_err);
 152     if (local_err) {
 153         ret = -EINVAL;
 154         error_propagate(errp, local_err);
 155         goto fail;
 156     }
 157
 158     /* Open the log file */
 159     s->log_file = bdrv_open_child(NULL, options, "log", bs, &child_file, false,
 160                                   &local_err);
 161     if (local_err) {
 162         ret = -EINVAL;
 163         error_propagate(errp, local_err);
 164         goto fail;
 165     }
 166
 167     log_append = qemu_opt_get_bool(opts, "log-append", false);
 168
 169     if (log_append) {
 170         struct log_write_super log_sb = { 0, 0, 0, 0 };
 171
 172         if (qemu_opt_find(opts, "log-sector-size")) {
 173             ret = -EINVAL;
 174             error_setg(errp, "log-append and log-sector-size are mutually "
 175                        "exclusive");
 176             goto fail_log;
 177         }
 178
 179         /* Read log superblock or fake one for an empty log */
 180         if (!bdrv_getlength(s->log_file->bs)) {
 181             log_sb.magic      = cpu_to_le64(WRITE_LOG_MAGIC);
 182             log_sb.version    = cpu_to_le64(WRITE_LOG_VERSION);
 183             log_sb.nr_entries = cpu_to_le64(0);
 184             log_sb.sectorsize = cpu_to_le32(BDRV_SECTOR_SIZE);
 185         } else {
 186             ret = bdrv_pread(s->log_file, 0, &log_sb, sizeof(log_sb));
 187             if (ret < 0) {
 188                 error_setg_errno(errp, -ret, "Could not read log superblock");
 189                 goto fail_log;
 190             }
 191         }
 192
 193         if (log_sb.magic != cpu_to_le64(WRITE_LOG_MAGIC)) {
 194             ret = -EINVAL;
 195             error_setg(errp, "Invalid log superblock magic");
 196             goto fail_log;
 197         }
 198
 199         if (log_sb.version != cpu_to_le64(WRITE_LOG_VERSION)) {
 200             ret = -EINVAL;
 201             error_setg(errp, "Unsupported log version %"PRIu64,
 202                        le64_to_cpu(log_sb.version));
 203             goto fail_log;
 204         }
 205
 206         log_sector_size = le32_to_cpu(log_sb.sectorsize);
 207         s->cur_log_sector = 1;
 208         s->nr_entries = 0;
 209
 210         if (blk_log_writes_sector_size_valid(log_sector_size)) {
 211             s->cur_log_sector =
 212                 blk_log_writes_find_cur_log_sector(s->log_file, log_sector_size,
 213                                     le64_to_cpu(log_sb.nr_entries), &local_err);
 214             if (local_err) {
 215                 ret = -EINVAL;
 216                 error_propagate(errp, local_err);
 217                 goto fail_log;
 218             }
 219
 220             s->nr_entries = le64_to_cpu(log_sb.nr_entries);
 221         }
 222     } else {
 223         log_sector_size = qemu_opt_get_size(opts, "log-sector-size",
 224                                             BDRV_SECTOR_SIZE);
 225         s->cur_log_sector = 1;
 226         s->nr_entries = 0;
 227     }
 228
 229     if (!blk_log_writes_sector_size_valid(log_sector_size)) {
 230         ret = -EINVAL;
 231         error_setg(errp, "Invalid log sector size %"PRIu64, log_sector_size);
 232         goto fail_log;
 233     }
 234
 235     s->sectorsize = log_sector_size;
 236     s->sectorbits = blk_log_writes_log2(log_sector_size);
 237
 238     ret = 0;
 239 fail_log:
 240     if (ret < 0) {
 241         bdrv_unref_child(bs, s->log_file);
 242         s->log_file = NULL;
 243     }
 244 fail:
 245     if (ret < 0) {
 246         bdrv_unref_child(bs, bs->file);
 247         bs->file = NULL;
 248     }
 249     qemu_opts_del(opts);
 250     return ret;
 251 }
 252
 253 static void blk_log_writes_close(BlockDriverState *bs)
 254 {
 255     BDRVBlkLogWritesState *s = bs->opaque;
 256
 257     bdrv_unref_child(bs, s->log_file);
 258     s->log_file = NULL;
 259 }
 260
 261 static int64_t blk_log_writes_getlength(BlockDriverState *bs)
 262 {
 263     return bdrv_getlength(bs->file->bs);
 264 }
 265
 266 static void blk_log_writes_refresh_filename(BlockDriverState *bs,
 267                                             QDict *options)
 268 {
 269     BDRVBlkLogWritesState *s = bs->opaque;
 270
 271     /* bs->file->bs has already been refreshed */
 272     bdrv_refresh_filename(s->log_file->bs);
 273
 274     if (bs->file->bs->full_open_options
 275         && s->log_file->bs->full_open_options)
 276     {
 277         QDict *opts = qdict_new();
 278         qdict_put_str(opts, "driver", "blklogwrites");
 279
 280         qobject_ref(bs->file->bs->full_open_options);
 281         qdict_put_obj(opts, "file", QOBJECT(bs->file->bs->full_open_options));
 282         qobject_ref(s->log_file->bs->full_open_options);
 283         qdict_put_obj(opts, "log",
 284                       QOBJECT(s->log_file->bs->full_open_options));
 285         qdict_put_int(opts, "log-sector-size", s->sectorsize);
 286
 287         bs->full_open_options = opts;
 288     }
 289 }
 290
 291 static void blk_log_writes_child_perm(BlockDriverState *bs, BdrvChild *c,
 292                                       const BdrvChildRole *role,
 293                                       BlockReopenQueue *ro_q,
 294                                       uint64_t perm, uint64_t shrd,
 295                                       uint64_t *nperm, uint64_t *nshrd)
 296 {
 297     if (!c) {
 298         *nperm = perm & DEFAULT_PERM_PASSTHROUGH;
 299         *nshrd = (shrd & DEFAULT_PERM_PASSTHROUGH) | DEFAULT_PERM_UNCHANGED;
 300         return;
 301     }
 302
 303     if (!strcmp(c->name, "log")) {
 304         bdrv_format_default_perms(bs, c, role, ro_q, perm, shrd, nperm, nshrd);
 305     } else {
 306         bdrv_filter_default_perms(bs, c, role, ro_q, perm, shrd, nperm, nshrd);
 307     }
 308 }
 309
 310 static void blk_log_writes_refresh_limits(BlockDriverState *bs, Error **errp)
 311 {
 312     BDRVBlkLogWritesState *s = bs->opaque;
 313     bs->bl.request_alignment = s->sectorsize;
 314 }
 315
 316 static int coroutine_fn
 317 blk_log_writes_co_preadv(BlockDriverState *bs, uint64_t offset, uint64_t bytes,
 318                          QEMUIOVector *qiov, int flags)
 319 {
 320     return bdrv_co_preadv(bs->file, offset, bytes, qiov, flags);
 321 }
 322
 323 typedef struct BlkLogWritesFileReq {
 324     BlockDriverState *bs;
 325     uint64_t offset;
 326     uint64_t bytes;
 327     int file_flags;
 328     QEMUIOVector *qiov;
 329     int (*func)(struct BlkLogWritesFileReq *r);
 330     int file_ret;
 331 } BlkLogWritesFileReq;
 332
 333 typedef struct {
 334     BlockDriverState *bs;
 335     QEMUIOVector *qiov;
 336     struct log_write_entry entry;
 337     uint64_t zero_size;
 338     int log_ret;
 339 } BlkLogWritesLogReq;
 340
 341 static void coroutine_fn blk_log_writes_co_do_log(BlkLogWritesLogReq *lr)
 342 {
 343     BDRVBlkLogWritesState *s = lr->bs->opaque;
 344     uint64_t cur_log_offset = s->cur_log_sector << s->sectorbits;
 345
 346     s->nr_entries++;
 347     s->cur_log_sector +=
 348             ROUND_UP(lr->qiov->size, s->sectorsize) >> s->sectorbits;
 349
 350     lr->log_ret = bdrv_co_pwritev(s->log_file, cur_log_offset, lr->qiov->size,
 351                                   lr->qiov, 0);
 352
 353     /* Logging for the "write zeroes" operation */
 354     if (lr->log_ret == 0 && lr->zero_size) {
 355         cur_log_offset = s->cur_log_sector << s->sectorbits;
 356         s->cur_log_sector +=
 357                 ROUND_UP(lr->zero_size, s->sectorsize) >> s->sectorbits;
 358
 359         lr->log_ret = bdrv_co_pwrite_zeroes(s->log_file, cur_log_offset,
 360                                             lr->zero_size, 0);
 361     }
 362
 363     /* Update super block on flush */
 364     if (lr->log_ret == 0 && lr->entry.flags & LOG_FLUSH_FLAG) {
 365         struct log_write_super super = {
 366             .magic      = cpu_to_le64(WRITE_LOG_MAGIC),
 367             .version    = cpu_to_le64(WRITE_LOG_VERSION),
 368             .nr_entries = cpu_to_le64(s->nr_entries),
 369             .sectorsize = cpu_to_le32(s->sectorsize),
 370         };
 371         void *zeroes = g_malloc0(s->sectorsize - sizeof(super));
 372         QEMUIOVector qiov;
 373
 374         qemu_iovec_init(&qiov, 2);
 375         qemu_iovec_add(&qiov, &super, sizeof(super));
 376         qemu_iovec_add(&qiov, zeroes, s->sectorsize - sizeof(super));
 377
 378         lr->log_ret =
 379             bdrv_co_pwritev(s->log_file, 0, s->sectorsize, &qiov, 0);
 380         if (lr->log_ret == 0) {
 381             lr->log_ret = bdrv_co_flush(s->log_file->bs);
 382         }
 383         qemu_iovec_destroy(&qiov);
 384         g_free(zeroes);
 385     }
 386 }
 387
 388 static void coroutine_fn blk_log_writes_co_do_file(BlkLogWritesFileReq *fr)
 389 {
 390     fr->file_ret = fr->func(fr);
 391 }
 392
 393 static int coroutine_fn
 394 blk_log_writes_co_log(BlockDriverState *bs, uint64_t offset, uint64_t bytes,
 395                       QEMUIOVector *qiov, int flags,
 396                       int (*file_func)(BlkLogWritesFileReq *r),
 397                       uint64_t entry_flags, bool is_zero_write)
 398 {
 399     QEMUIOVector log_qiov;
 400     size_t niov = qiov ? qiov->niov : 0;
 401     BDRVBlkLogWritesState *s = bs->opaque;
 402     BlkLogWritesFileReq fr = {
 403         .bs         = bs,
 404         .offset     = offset,
 405         .bytes      = bytes,
 406         .file_flags = flags,
 407         .qiov       = qiov,
 408         .func       = file_func,
 409     };
 410     BlkLogWritesLogReq lr = {
 411         .bs             = bs,
 412         .qiov           = &log_qiov,
 413         .entry = {
 414             .sector     = cpu_to_le64(offset >> s->sectorbits),
 415             .nr_sectors = cpu_to_le64(bytes >> s->sectorbits),
 416             .flags      = cpu_to_le64(entry_flags),
 417             .data_len   = 0,
 418         },
 419         .zero_size = is_zero_write ? bytes : 0,
 420     };
 421     void *zeroes = g_malloc0(s->sectorsize - sizeof(lr.entry));
 422
 423     assert((1 << s->sectorbits) == s->sectorsize);
 424     assert(bs->bl.request_alignment == s->sectorsize);
 425     assert(QEMU_IS_ALIGNED(offset, bs->bl.request_alignment));
 426     assert(QEMU_IS_ALIGNED(bytes, bs->bl.request_alignment));
 427
 428     qemu_iovec_init(&log_qiov, niov + 2);
 429     qemu_iovec_add(&log_qiov, &lr.entry, sizeof(lr.entry));
 430     qemu_iovec_add(&log_qiov, zeroes, s->sectorsize - sizeof(lr.entry));
 431     if (qiov) {
 432         qemu_iovec_concat(&log_qiov, qiov, 0, qiov->size);
 433     }
 434
 435     blk_log_writes_co_do_file(&fr);
 436     blk_log_writes_co_do_log(&lr);
 437
 438     qemu_iovec_destroy(&log_qiov);
 439     g_free(zeroes);
 440
 441     if (lr.log_ret < 0) {
 442         return lr.log_ret;
 443     }
 444
 445     return fr.file_ret;
 446 }
 447
 448 static int coroutine_fn
 449 blk_log_writes_co_do_file_pwritev(BlkLogWritesFileReq *fr)
 450 {
 451     return bdrv_co_pwritev(fr->bs->file, fr->offset, fr->bytes,
 452                            fr->qiov, fr->file_flags);
 453 }
 454
 455 static int coroutine_fn
 456 blk_log_writes_co_do_file_pwrite_zeroes(BlkLogWritesFileReq *fr)
 457 {
 458     return bdrv_co_pwrite_zeroes(fr->bs->file, fr->offset, fr->bytes,
 459                                  fr->file_flags);
 460 }
 461
 462 static int coroutine_fn blk_log_writes_co_do_file_flush(BlkLogWritesFileReq *fr)
 463 {
 464     return bdrv_co_flush(fr->bs->file->bs);
 465 }
 466
 467 static int coroutine_fn
 468 blk_log_writes_co_do_file_pdiscard(BlkLogWritesFileReq *fr)
 469 {
 470     return bdrv_co_pdiscard(fr->bs->file->bs, fr->offset, fr->bytes);
 471 }
 472
 473 static int coroutine_fn
 474 blk_log_writes_co_pwritev(BlockDriverState *bs, uint64_t offset, uint64_t bytes,
 475                           QEMUIOVector *qiov, int flags)
 476 {
 477     return blk_log_writes_co_log(bs, offset, bytes, qiov, flags,
 478                                  blk_log_writes_co_do_file_pwritev, 0, false);
 479 }
 480
 481 static int coroutine_fn
 482 blk_log_writes_co_pwrite_zeroes(BlockDriverState *bs, int64_t offset, int bytes,
 483                                 BdrvRequestFlags flags)
 484 {
 485     return blk_log_writes_co_log(bs, offset, bytes, NULL, flags,
 486                                  blk_log_writes_co_do_file_pwrite_zeroes, 0,
 487                                  true);
 488 }
 489
 490 static int coroutine_fn blk_log_writes_co_flush_to_disk(BlockDriverState *bs)
 491 {
 492     return blk_log_writes_co_log(bs, 0, 0, NULL, 0,
 493                                  blk_log_writes_co_do_file_flush,
 494                                  LOG_FLUSH_FLAG, false);
 495 }
 496
 497 static int coroutine_fn
 498 blk_log_writes_co_pdiscard(BlockDriverState *bs, int64_t offset, int count)
 499 {
 500     return blk_log_writes_co_log(bs, offset, count, NULL, 0,
 501                                  blk_log_writes_co_do_file_pdiscard,
 502                                  LOG_DISCARD_FLAG, false);
 503 }
 504
 505 static BlockDriver bdrv_blk_log_writes = {
 506     .format_name            = "blklogwrites",
 507     .instance_size          = sizeof(BDRVBlkLogWritesState),
 508
 509     .bdrv_open              = blk_log_writes_open,
 510     .bdrv_close             = blk_log_writes_close,
 511     .bdrv_getlength         = blk_log_writes_getlength,
 512     .bdrv_refresh_filename  = blk_log_writes_refresh_filename,
 513     .bdrv_child_perm        = blk_log_writes_child_perm,
 514     .bdrv_refresh_limits    = blk_log_writes_refresh_limits,
 515
 516     .bdrv_co_preadv         = blk_log_writes_co_preadv,
 517     .bdrv_co_pwritev        = blk_log_writes_co_pwritev,
 518     .bdrv_co_pwrite_zeroes  = blk_log_writes_co_pwrite_zeroes,
 519     .bdrv_co_flush_to_disk  = blk_log_writes_co_flush_to_disk,
 520     .bdrv_co_pdiscard       = blk_log_writes_co_pdiscard,
 521     .bdrv_co_block_status   = bdrv_co_block_status_from_file,
 522
 523     .is_filter              = true,
 524 };
 525
 526 static void bdrv_blk_log_writes_init(void)
 527 {
 528     bdrv_register(&bdrv_blk_log_writes);
 529 }
 530
 531 block_init(bdrv_blk_log_writes_init);