block/blklogwrites.c

   1 /*
   2  * Write logging blk driver based on blkverify and blkdebug.
   3  *
   4  * Copyright (c) 2017 Tuomas Tynkkynen <tuomas@tuxera.com>
   5  * Copyright (c) 2018 Aapo Vienamo <aapo@tuxera.com>
   6  * Copyright (c) 2018 Ari Sundholm <ari@tuxera.com>
   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     uint64_t update_interval;
  59 } BDRVBlkLogWritesState;
  60
  61 static QemuOptsList runtime_opts = {
  62     .name = "blklogwrites",
  63     .head = QTAILQ_HEAD_INITIALIZER(runtime_opts.head),
  64     .desc = {
  65         {
  66             .name = "log-append",
  67             .type = QEMU_OPT_BOOL,
  68             .help = "Append to an existing log",
  69         },
  70         {
  71             .name = "log-sector-size",
  72             .type = QEMU_OPT_SIZE,
  73             .help = "Log sector size",
  74         },
  75         {
  76             .name = "log-super-update-interval",
  77             .type = QEMU_OPT_NUMBER,
  78             .help = "Log superblock update interval (# of write requests)",
  79         },
  80         { /* end of list */ }
  81     },
  82 };
  83
  84 static inline uint32_t blk_log_writes_log2(uint32_t value)
  85 {
  86     assert(value > 0);
  87     return 31 - clz32(value);
  88 }
  89
  90 static inline bool blk_log_writes_sector_size_valid(uint32_t sector_size)
  91 {
  92     return is_power_of_2(sector_size) &&
  93         sector_size >= sizeof(struct log_write_super) &&
  94         sector_size >= sizeof(struct log_write_entry) &&
  95         sector_size < (1ull << 24);
  96 }
  97
  98 static uint64_t blk_log_writes_find_cur_log_sector(BdrvChild *log,
  99                                                    uint32_t sector_size,
 100                                                    uint64_t nr_entries,
 101                                                    Error **errp)
 102 {
 103     uint64_t cur_sector = 1;
 104     uint64_t cur_idx = 0;
 105     uint32_t sector_bits = blk_log_writes_log2(sector_size);
 106     struct log_write_entry cur_entry;
 107
 108     while (cur_idx < nr_entries) {
 109         int read_ret = bdrv_pread(log, cur_sector << sector_bits, &cur_entry,
 110                                   sizeof(cur_entry));
 111         if (read_ret < 0) {
 112             error_setg_errno(errp, -read_ret,
 113                              "Failed to read log entry %"PRIu64, cur_idx);
 114             return (uint64_t)-1ull;
 115         }
 116
 117         if (cur_entry.flags & ~cpu_to_le64(LOG_FLAG_MASK)) {
 118             error_setg(errp, "Invalid flags 0x%"PRIx64" in log entry %"PRIu64,
 119                        le64_to_cpu(cur_entry.flags), cur_idx);
 120             return (uint64_t)-1ull;
 121         }
 122
 123         /* Account for the sector of the entry itself */
 124         ++cur_sector;
 125
 126         /*
 127          * Account for the data of the write.
 128          * For discards, this data is not present.
 129          */
 130         if (!(cur_entry.flags & cpu_to_le64(LOG_DISCARD_FLAG))) {
 131             cur_sector += le64_to_cpu(cur_entry.nr_sectors);
 132         }
 133
 134         ++cur_idx;
 135     }
 136
 137     return cur_sector;
 138 }
 139
 140 static int blk_log_writes_open(BlockDriverState *bs, QDict *options, int flags,
 141                                Error **errp)
 142 {
 143     BDRVBlkLogWritesState *s = bs->opaque;
 144     QemuOpts *opts;
 145     Error *local_err = NULL;
 146     int ret;
 147     uint64_t log_sector_size;
 148     bool log_append;
 149
 150     opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
 151     qemu_opts_absorb_qdict(opts, options, &local_err);
 152     if (local_err) {
 153         ret = -EINVAL;
 154         error_propagate(errp, local_err);
 155         goto fail;
 156     }
 157
 158     /* Open the file */
 159     bs->file = bdrv_open_child(NULL, options, "file", 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     /* Open the log file */
 168     s->log_file = bdrv_open_child(NULL, options, "log", bs, &child_file, false,
 169                                   &local_err);
 170     if (local_err) {
 171         ret = -EINVAL;
 172         error_propagate(errp, local_err);
 173         goto fail;
 174     }
 175
 176     log_append = qemu_opt_get_bool(opts, "log-append", false);
 177
 178     if (log_append) {
 179         struct log_write_super log_sb = { 0, 0, 0, 0 };
 180
 181         if (qemu_opt_find(opts, "log-sector-size")) {
 182             ret = -EINVAL;
 183             error_setg(errp, "log-append and log-sector-size are mutually "
 184                        "exclusive");
 185             goto fail_log;
 186         }
 187
 188         /* Read log superblock or fake one for an empty log */
 189         if (!bdrv_getlength(s->log_file->bs)) {
 190             log_sb.magic      = cpu_to_le64(WRITE_LOG_MAGIC);
 191             log_sb.version    = cpu_to_le64(WRITE_LOG_VERSION);
 192             log_sb.nr_entries = cpu_to_le64(0);
 193             log_sb.sectorsize = cpu_to_le32(BDRV_SECTOR_SIZE);
 194         } else {
 195             ret = bdrv_pread(s->log_file, 0, &log_sb, sizeof(log_sb));
 196             if (ret < 0) {
 197                 error_setg_errno(errp, -ret, "Could not read log superblock");
 198                 goto fail_log;
 199             }
 200         }
 201
 202         if (log_sb.magic != cpu_to_le64(WRITE_LOG_MAGIC)) {
 203             ret = -EINVAL;
 204             error_setg(errp, "Invalid log superblock magic");
 205             goto fail_log;
 206         }
 207
 208         if (log_sb.version != cpu_to_le64(WRITE_LOG_VERSION)) {
 209             ret = -EINVAL;
 210             error_setg(errp, "Unsupported log version %"PRIu64,
 211                        le64_to_cpu(log_sb.version));
 212             goto fail_log;
 213         }
 214
 215         log_sector_size = le32_to_cpu(log_sb.sectorsize);
 216         s->cur_log_sector = 1;
 217         s->nr_entries = 0;
 218
 219         if (blk_log_writes_sector_size_valid(log_sector_size)) {
 220             s->cur_log_sector =
 221                 blk_log_writes_find_cur_log_sector(s->log_file, log_sector_size,
 222                                     le64_to_cpu(log_sb.nr_entries), &local_err);
 223             if (local_err) {
 224                 ret = -EINVAL;
 225                 error_propagate(errp, local_err);
 226                 goto fail_log;
 227             }
 228
 229             s->nr_entries = le64_to_cpu(log_sb.nr_entries);
 230         }
 231     } else {
 232         log_sector_size = qemu_opt_get_size(opts, "log-sector-size",
 233                                             BDRV_SECTOR_SIZE);
 234         s->cur_log_sector = 1;
 235         s->nr_entries = 0;
 236     }
 237
 238     if (!blk_log_writes_sector_size_valid(log_sector_size)) {
 239         ret = -EINVAL;
 240         error_setg(errp, "Invalid log sector size %"PRIu64, log_sector_size);
 241         goto fail_log;
 242     }
 243
 244     s->sectorsize = log_sector_size;
 245     s->sectorbits = blk_log_writes_log2(log_sector_size);
 246     s->update_interval = qemu_opt_get_number(opts, "log-super-update-interval",
 247                                              4096);
 248     if (!s->update_interval) {
 249         ret = -EINVAL;
 250         error_setg(errp, "Invalid log superblock update interval %"PRIu64,
 251                    s->update_interval);
 252         goto fail_log;
 253     }
 254
 255     ret = 0;
 256 fail_log:
 257     if (ret < 0) {
 258         bdrv_unref_child(bs, s->log_file);
 259         s->log_file = NULL;
 260     }
 261 fail:
 262     if (ret < 0) {
 263         bdrv_unref_child(bs, bs->file);
 264         bs->file = NULL;
 265     }
 266     qemu_opts_del(opts);
 267     return ret;
 268 }
 269
 270 static void blk_log_writes_close(BlockDriverState *bs)
 271 {
 272     BDRVBlkLogWritesState *s = bs->opaque;
 273
 274     bdrv_unref_child(bs, s->log_file);
 275     s->log_file = NULL;
 276 }
 277
 278 static int64_t blk_log_writes_getlength(BlockDriverState *bs)
 279 {
 280     return bdrv_getlength(bs->file->bs);
 281 }
 282
 283 static void blk_log_writes_refresh_filename(BlockDriverState *bs,
 284                                             QDict *options)
 285 {
 286     BDRVBlkLogWritesState *s = bs->opaque;
 287
 288     /* bs->file->bs has already been refreshed */
 289     bdrv_refresh_filename(s->log_file->bs);
 290
 291     if (bs->file->bs->full_open_options
 292         && s->log_file->bs->full_open_options)
 293     {
 294         QDict *opts = qdict_new();
 295         qdict_put_str(opts, "driver", "blklogwrites");
 296
 297         qobject_ref(bs->file->bs->full_open_options);
 298         qdict_put_obj(opts, "file", QOBJECT(bs->file->bs->full_open_options));
 299         qobject_ref(s->log_file->bs->full_open_options);
 300         qdict_put_obj(opts, "log",
 301                       QOBJECT(s->log_file->bs->full_open_options));
 302         qdict_put_int(opts, "log-sector-size", s->sectorsize);
 303
 304         bs->full_open_options = opts;
 305     }
 306 }
 307
 308 static void blk_log_writes_child_perm(BlockDriverState *bs, BdrvChild *c,
 309                                       const BdrvChildRole *role,
 310                                       BlockReopenQueue *ro_q,
 311                                       uint64_t perm, uint64_t shrd,
 312                                       uint64_t *nperm, uint64_t *nshrd)
 313 {
 314     if (!c) {
 315         *nperm = perm & DEFAULT_PERM_PASSTHROUGH;
 316         *nshrd = (shrd & DEFAULT_PERM_PASSTHROUGH) | DEFAULT_PERM_UNCHANGED;
 317         return;
 318     }
 319
 320     if (!strcmp(c->name, "log")) {
 321         bdrv_format_default_perms(bs, c, role, ro_q, perm, shrd, nperm, nshrd);
 322     } else {
 323         bdrv_filter_default_perms(bs, c, role, ro_q, perm, shrd, nperm, nshrd);
 324     }
 325 }
 326
 327 static void blk_log_writes_refresh_limits(BlockDriverState *bs, Error **errp)
 328 {
 329     BDRVBlkLogWritesState *s = bs->opaque;
 330     bs->bl.request_alignment = s->sectorsize;
 331 }
 332
 333 static int coroutine_fn
 334 blk_log_writes_co_preadv(BlockDriverState *bs, uint64_t offset, uint64_t bytes,
 335                          QEMUIOVector *qiov, int flags)
 336 {
 337     return bdrv_co_preadv(bs->file, offset, bytes, qiov, flags);
 338 }
 339
 340 typedef struct BlkLogWritesFileReq {
 341     BlockDriverState *bs;
 342     uint64_t offset;
 343     uint64_t bytes;
 344     int file_flags;
 345     QEMUIOVector *qiov;
 346     int (*func)(struct BlkLogWritesFileReq *r);
 347     int file_ret;
 348 } BlkLogWritesFileReq;
 349
 350 typedef struct {
 351     BlockDriverState *bs;
 352     QEMUIOVector *qiov;
 353     struct log_write_entry entry;
 354     uint64_t zero_size;
 355     int log_ret;
 356 } BlkLogWritesLogReq;
 357
 358 static void coroutine_fn blk_log_writes_co_do_log(BlkLogWritesLogReq *lr)
 359 {
 360     BDRVBlkLogWritesState *s = lr->bs->opaque;
 361     uint64_t cur_log_offset = s->cur_log_sector << s->sectorbits;
 362
 363     s->nr_entries++;
 364     s->cur_log_sector +=
 365             ROUND_UP(lr->qiov->size, s->sectorsize) >> s->sectorbits;
 366
 367     lr->log_ret = bdrv_co_pwritev(s->log_file, cur_log_offset, lr->qiov->size,
 368                                   lr->qiov, 0);
 369
 370     /* Logging for the "write zeroes" operation */
 371     if (lr->log_ret == 0 && lr->zero_size) {
 372         cur_log_offset = s->cur_log_sector << s->sectorbits;
 373         s->cur_log_sector +=
 374                 ROUND_UP(lr->zero_size, s->sectorsize) >> s->sectorbits;
 375
 376         lr->log_ret = bdrv_co_pwrite_zeroes(s->log_file, cur_log_offset,
 377                                             lr->zero_size, 0);
 378     }
 379
 380     /* Update super block on flush or every update interval */
 381     if (lr->log_ret == 0 && ((lr->entry.flags & LOG_FLUSH_FLAG)
 382         || (s->nr_entries % s->update_interval == 0)))
 383     {
 384         struct log_write_super super = {
 385             .magic      = cpu_to_le64(WRITE_LOG_MAGIC),
 386             .version    = cpu_to_le64(WRITE_LOG_VERSION),
 387             .nr_entries = cpu_to_le64(s->nr_entries),
 388             .sectorsize = cpu_to_le32(s->sectorsize),
 389         };
 390         void *zeroes = g_malloc0(s->sectorsize - sizeof(super));
 391         QEMUIOVector qiov;
 392
 393         qemu_iovec_init(&qiov, 2);
 394         qemu_iovec_add(&qiov, &super, sizeof(super));
 395         qemu_iovec_add(&qiov, zeroes, s->sectorsize - sizeof(super));
 396
 397         lr->log_ret =
 398             bdrv_co_pwritev(s->log_file, 0, s->sectorsize, &qiov, 0);
 399         if (lr->log_ret == 0) {
 400             lr->log_ret = bdrv_co_flush(s->log_file->bs);
 401         }
 402         qemu_iovec_destroy(&qiov);
 403         g_free(zeroes);
 404     }
 405 }
 406
 407 static void coroutine_fn blk_log_writes_co_do_file(BlkLogWritesFileReq *fr)
 408 {
 409     fr->file_ret = fr->func(fr);
 410 }
 411
 412 static int coroutine_fn
 413 blk_log_writes_co_log(BlockDriverState *bs, uint64_t offset, uint64_t bytes,
 414                       QEMUIOVector *qiov, int flags,
 415                       int (*file_func)(BlkLogWritesFileReq *r),
 416                       uint64_t entry_flags, bool is_zero_write)
 417 {
 418     QEMUIOVector log_qiov;
 419     size_t niov = qiov ? qiov->niov : 0;
 420     BDRVBlkLogWritesState *s = bs->opaque;
 421     BlkLogWritesFileReq fr = {
 422         .bs         = bs,
 423         .offset     = offset,
 424         .bytes      = bytes,
 425         .file_flags = flags,
 426         .qiov       = qiov,
 427         .func       = file_func,
 428     };
 429     BlkLogWritesLogReq lr = {
 430         .bs             = bs,
 431         .qiov           = &log_qiov,
 432         .entry = {
 433             .sector     = cpu_to_le64(offset >> s->sectorbits),
 434             .nr_sectors = cpu_to_le64(bytes >> s->sectorbits),
 435             .flags      = cpu_to_le64(entry_flags),
 436             .data_len   = 0,
 437         },
 438         .zero_size = is_zero_write ? bytes : 0,
 439     };
 440     void *zeroes = g_malloc0(s->sectorsize - sizeof(lr.entry));
 441
 442     assert((1 << s->sectorbits) == s->sectorsize);
 443     assert(bs->bl.request_alignment == s->sectorsize);
 444     assert(QEMU_IS_ALIGNED(offset, bs->bl.request_alignment));
 445     assert(QEMU_IS_ALIGNED(bytes, bs->bl.request_alignment));
 446
 447     qemu_iovec_init(&log_qiov, niov + 2);
 448     qemu_iovec_add(&log_qiov, &lr.entry, sizeof(lr.entry));
 449     qemu_iovec_add(&log_qiov, zeroes, s->sectorsize - sizeof(lr.entry));
 450     if (qiov) {
 451         qemu_iovec_concat(&log_qiov, qiov, 0, qiov->size);
 452     }
 453
 454     blk_log_writes_co_do_file(&fr);
 455     blk_log_writes_co_do_log(&lr);
 456
 457     qemu_iovec_destroy(&log_qiov);
 458     g_free(zeroes);
 459
 460     if (lr.log_ret < 0) {
 461         return lr.log_ret;
 462     }
 463
 464     return fr.file_ret;
 465 }
 466
 467 static int coroutine_fn
 468 blk_log_writes_co_do_file_pwritev(BlkLogWritesFileReq *fr)
 469 {
 470     return bdrv_co_pwritev(fr->bs->file, fr->offset, fr->bytes,
 471                            fr->qiov, fr->file_flags);
 472 }
 473
 474 static int coroutine_fn
 475 blk_log_writes_co_do_file_pwrite_zeroes(BlkLogWritesFileReq *fr)
 476 {
 477     return bdrv_co_pwrite_zeroes(fr->bs->file, fr->offset, fr->bytes,
 478                                  fr->file_flags);
 479 }
 480
 481 static int coroutine_fn blk_log_writes_co_do_file_flush(BlkLogWritesFileReq *fr)
 482 {
 483     return bdrv_co_flush(fr->bs->file->bs);
 484 }
 485
 486 static int coroutine_fn
 487 blk_log_writes_co_do_file_pdiscard(BlkLogWritesFileReq *fr)
 488 {
 489     return bdrv_co_pdiscard(fr->bs->file, fr->offset, fr->bytes);
 490 }
 491
 492 static int coroutine_fn
 493 blk_log_writes_co_pwritev(BlockDriverState *bs, uint64_t offset, uint64_t bytes,
 494                           QEMUIOVector *qiov, int flags)
 495 {
 496     return blk_log_writes_co_log(bs, offset, bytes, qiov, flags,
 497                                  blk_log_writes_co_do_file_pwritev, 0, false);
 498 }
 499
 500 static int coroutine_fn
 501 blk_log_writes_co_pwrite_zeroes(BlockDriverState *bs, int64_t offset, int bytes,
 502                                 BdrvRequestFlags flags)
 503 {
 504     return blk_log_writes_co_log(bs, offset, bytes, NULL, flags,
 505                                  blk_log_writes_co_do_file_pwrite_zeroes, 0,
 506                                  true);
 507 }
 508
 509 static int coroutine_fn blk_log_writes_co_flush_to_disk(BlockDriverState *bs)
 510 {
 511     return blk_log_writes_co_log(bs, 0, 0, NULL, 0,
 512                                  blk_log_writes_co_do_file_flush,
 513                                  LOG_FLUSH_FLAG, false);
 514 }
 515
 516 static int coroutine_fn
 517 blk_log_writes_co_pdiscard(BlockDriverState *bs, int64_t offset, int count)
 518 {
 519     return blk_log_writes_co_log(bs, offset, count, NULL, 0,
 520                                  blk_log_writes_co_do_file_pdiscard,
 521                                  LOG_DISCARD_FLAG, false);
 522 }
 523
 524 static BlockDriver bdrv_blk_log_writes = {
 525     .format_name            = "blklogwrites",
 526     .instance_size          = sizeof(BDRVBlkLogWritesState),
 527
 528     .bdrv_open              = blk_log_writes_open,
 529     .bdrv_close             = blk_log_writes_close,
 530     .bdrv_getlength         = blk_log_writes_getlength,
 531     .bdrv_refresh_filename  = blk_log_writes_refresh_filename,
 532     .bdrv_child_perm        = blk_log_writes_child_perm,
 533     .bdrv_refresh_limits    = blk_log_writes_refresh_limits,
 534
 535     .bdrv_co_preadv         = blk_log_writes_co_preadv,
 536     .bdrv_co_pwritev        = blk_log_writes_co_pwritev,
 537     .bdrv_co_pwrite_zeroes  = blk_log_writes_co_pwrite_zeroes,
 538     .bdrv_co_flush_to_disk  = blk_log_writes_co_flush_to_disk,
 539     .bdrv_co_pdiscard       = blk_log_writes_co_pdiscard,
 540     .bdrv_co_block_status   = bdrv_co_block_status_from_file,
 541
 542     .is_filter              = true,
 543 };
 544
 545 static void bdrv_blk_log_writes_init(void)
 546 {
 547     bdrv_register(&bdrv_blk_log_writes);
 548 }
 549
 550 block_init(bdrv_blk_log_writes_init);