block/commit.c

   1 /*
   2  * Live block commit
   3  *
   4  * Copyright Red Hat, Inc. 2012
   5  *
   6  * Authors:
   7  *  Jeff Cody   <jcody@redhat.com>
   8  *  Based on stream.c by Stefan Hajnoczi
   9  *
  10  * This work is licensed under the terms of the GNU LGPL, version 2 or later.
  11  * See the COPYING.LIB file in the top-level directory.
  12  *
  13  */
  14
  15 #include "qemu/osdep.h"
  16 #include "qemu/cutils.h"
  17 #include "trace.h"
  18 #include "block/block_int.h"
  19 #include "block/blockjob_int.h"
  20 #include "qapi/error.h"
  21 #include "qapi/qmp/qerror.h"
  22 #include "qemu/ratelimit.h"
  23 #include "sysemu/block-backend.h"
  24
  25 enum {
  26     /*
  27      * Size of data buffer for populating the image file.  This should be large
  28      * enough to process multiple clusters in a single call, so that populating
  29      * contiguous regions of the image is efficient.
  30      */
  31     COMMIT_BUFFER_SIZE = 512 * 1024, /* in bytes */
  32 };
  33
  34 #define SLICE_TIME 100000000ULL /* ns */
  35
  36 typedef struct CommitBlockJob {
  37     BlockJob common;
  38     RateLimit limit;
  39     BlockDriverState *active;
  40     BlockDriverState *commit_top_bs;
  41     BlockBackend *top;
  42     BlockBackend *base;
  43     BlockdevOnError on_error;
  44     int base_flags;
  45     int orig_overlay_flags;
  46     char *backing_file_str;
  47 } CommitBlockJob;
  48
  49 static int coroutine_fn commit_populate(BlockBackend *bs, BlockBackend *base,
  50                                         int64_t sector_num, int nb_sectors,
  51                                         void *buf)
  52 {
  53     int ret = 0;
  54     QEMUIOVector qiov;
  55     struct iovec iov = {
  56         .iov_base = buf,
  57         .iov_len = nb_sectors * BDRV_SECTOR_SIZE,
  58     };
  59
  60     qemu_iovec_init_external(&qiov, &iov, 1);
  61
  62     ret = blk_co_preadv(bs, sector_num * BDRV_SECTOR_SIZE,
  63                         qiov.size, &qiov, 0);
  64     if (ret < 0) {
  65         return ret;
  66     }
  67
  68     ret = blk_co_pwritev(base, sector_num * BDRV_SECTOR_SIZE,
  69                          qiov.size, &qiov, 0);
  70     if (ret < 0) {
  71         return ret;
  72     }
  73
  74     return 0;
  75 }
  76
  77 typedef struct {
  78     int ret;
  79 } CommitCompleteData;
  80
  81 static void commit_complete(BlockJob *job, void *opaque)
  82 {
  83     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
  84     CommitCompleteData *data = opaque;
  85     BlockDriverState *active = s->active;
  86     BlockDriverState *top = blk_bs(s->top);
  87     BlockDriverState *base = blk_bs(s->base);
  88     BlockDriverState *overlay_bs = bdrv_find_overlay(active, s->commit_top_bs);
  89     int ret = data->ret;
  90     bool remove_commit_top_bs = false;
  91
  92     /* Make sure overlay_bs and top stay around until bdrv_set_backing_hd() */
  93     bdrv_ref(top);
  94     bdrv_ref(overlay_bs);
  95
  96     /* Remove base node parent that still uses BLK_PERM_WRITE/RESIZE before
  97      * the normal backing chain can be restored. */
  98     blk_unref(s->base);
  99
 100     if (!block_job_is_cancelled(&s->common) && ret == 0) {
 101         /* success */
 102         ret = bdrv_drop_intermediate(active, s->commit_top_bs, base,
 103                                      s->backing_file_str);
 104     } else if (overlay_bs) {
 105         /* XXX Can (or should) we somehow keep 'consistent read' blocked even
 106          * after the failed/cancelled commit job is gone? If we already wrote
 107          * something to base, the intermediate images aren't valid any more. */
 108         remove_commit_top_bs = true;
 109     }
 110
 111     /* restore base open flags here if appropriate (e.g., change the base back
 112      * to r/o). These reopens do not need to be atomic, since we won't abort
 113      * even on failure here */
 114     if (s->base_flags != bdrv_get_flags(base)) {
 115         bdrv_reopen(base, s->base_flags, NULL);
 116     }
 117     if (overlay_bs && s->orig_overlay_flags != bdrv_get_flags(overlay_bs)) {
 118         bdrv_reopen(overlay_bs, s->orig_overlay_flags, NULL);
 119     }
 120     g_free(s->backing_file_str);
 121     blk_unref(s->top);
 122
 123     /* If there is more than one reference to the job (e.g. if called from
 124      * block_job_finish_sync()), block_job_completed() won't free it and
 125      * therefore the blockers on the intermediate nodes remain. This would
 126      * cause bdrv_set_backing_hd() to fail. */
 127     block_job_remove_all_bdrv(job);
 128
 129     block_job_completed(&s->common, ret);
 130     g_free(data);
 131
 132     /* If bdrv_drop_intermediate() didn't already do that, remove the commit
 133      * filter driver from the backing chain. Do this as the final step so that
 134      * the 'consistent read' permission can be granted.  */
 135     if (remove_commit_top_bs) {
 136         bdrv_set_backing_hd(overlay_bs, top, &error_abort);
 137     }
 138
 139     bdrv_unref(overlay_bs);
 140     bdrv_unref(top);
 141 }
 142
 143 static void coroutine_fn commit_run(void *opaque)
 144 {
 145     CommitBlockJob *s = opaque;
 146     CommitCompleteData *data;
 147     int64_t sector_num, end;
 148     uint64_t delay_ns = 0;
 149     int ret = 0;
 150     int n = 0;
 151     void *buf = NULL;
 152     int bytes_written = 0;
 153     int64_t base_len;
 154
 155     ret = s->common.len = blk_getlength(s->top);
 156
 157
 158     if (s->common.len < 0) {
 159         goto out;
 160     }
 161
 162     ret = base_len = blk_getlength(s->base);
 163     if (base_len < 0) {
 164         goto out;
 165     }
 166
 167     if (base_len < s->common.len) {
 168         ret = blk_truncate(s->base, s->common.len, NULL);
 169         if (ret) {
 170             goto out;
 171         }
 172     }
 173
 174     end = s->common.len >> BDRV_SECTOR_BITS;
 175     buf = blk_blockalign(s->top, COMMIT_BUFFER_SIZE);
 176
 177     for (sector_num = 0; sector_num < end; sector_num += n) {
 178         bool copy;
 179
 180         /* Note that even when no rate limit is applied we need to yield
 181          * with no pending I/O here so that bdrv_drain_all() returns.
 182          */
 183         block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 184         if (block_job_is_cancelled(&s->common)) {
 185             break;
 186         }
 187         /* Copy if allocated above the base */
 188         ret = bdrv_is_allocated_above(blk_bs(s->top), blk_bs(s->base),
 189                                       sector_num,
 190                                       COMMIT_BUFFER_SIZE / BDRV_SECTOR_SIZE,
 191                                       &n);
 192         copy = (ret == 1);
 193         trace_commit_one_iteration(s, sector_num * BDRV_SECTOR_SIZE,
 194                                    n * BDRV_SECTOR_SIZE, ret);
 195         if (copy) {
 196             ret = commit_populate(s->top, s->base, sector_num, n, buf);
 197             bytes_written += n * BDRV_SECTOR_SIZE;
 198         }
 199         if (ret < 0) {
 200             BlockErrorAction action =
 201                 block_job_error_action(&s->common, false, s->on_error, -ret);
 202             if (action == BLOCK_ERROR_ACTION_REPORT) {
 203                 goto out;
 204             } else {
 205                 n = 0;
 206                 continue;
 207             }
 208         }
 209         /* Publish progress */
 210         s->common.offset += n * BDRV_SECTOR_SIZE;
 211
 212         if (copy && s->common.speed) {
 213             delay_ns = ratelimit_calculate_delay(&s->limit,
 214                                                  n * BDRV_SECTOR_SIZE);
 215         }
 216     }
 217
 218     ret = 0;
 219
 220 out:
 221     qemu_vfree(buf);
 222
 223     data = g_malloc(sizeof(*data));
 224     data->ret = ret;
 225     block_job_defer_to_main_loop(&s->common, commit_complete, data);
 226 }
 227
 228 static void commit_set_speed(BlockJob *job, int64_t speed, Error **errp)
 229 {
 230     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
 231
 232     if (speed < 0) {
 233         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 234         return;
 235     }
 236     ratelimit_set_speed(&s->limit, speed, SLICE_TIME);
 237 }
 238
 239 static const BlockJobDriver commit_job_driver = {
 240     .instance_size = sizeof(CommitBlockJob),
 241     .job_type      = BLOCK_JOB_TYPE_COMMIT,
 242     .set_speed     = commit_set_speed,
 243     .start         = commit_run,
 244 };
 245
 246 static int coroutine_fn bdrv_commit_top_preadv(BlockDriverState *bs,
 247     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
 248 {
 249     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
 250 }
 251
 252 static int64_t coroutine_fn bdrv_commit_top_get_block_status(
 253     BlockDriverState *bs, int64_t sector_num, int nb_sectors, int *pnum,
 254     BlockDriverState **file)
 255 {
 256     *pnum = nb_sectors;
 257     *file = bs->backing->bs;
 258     return BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID |
 259            (sector_num << BDRV_SECTOR_BITS);
 260 }
 261
 262 static void bdrv_commit_top_refresh_filename(BlockDriverState *bs, QDict *opts)
 263 {
 264     bdrv_refresh_filename(bs->backing->bs);
 265     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
 266             bs->backing->bs->filename);
 267 }
 268
 269 static void bdrv_commit_top_close(BlockDriverState *bs)
 270 {
 271 }
 272
 273 static void bdrv_commit_top_child_perm(BlockDriverState *bs, BdrvChild *c,
 274                                        const BdrvChildRole *role,
 275                                        uint64_t perm, uint64_t shared,
 276                                        uint64_t *nperm, uint64_t *nshared)
 277 {
 278     *nperm = 0;
 279     *nshared = BLK_PERM_ALL;
 280 }
 281
 282 /* Dummy node that provides consistent read to its users without requiring it
 283  * from its backing file and that allows writes on the backing file chain. */
 284 static BlockDriver bdrv_commit_top = {
 285     .format_name                = "commit_top",
 286     .bdrv_co_preadv             = bdrv_commit_top_preadv,
 287     .bdrv_co_get_block_status   = bdrv_commit_top_get_block_status,
 288     .bdrv_refresh_filename      = bdrv_commit_top_refresh_filename,
 289     .bdrv_close                 = bdrv_commit_top_close,
 290     .bdrv_child_perm            = bdrv_commit_top_child_perm,
 291 };
 292
 293 void commit_start(const char *job_id, BlockDriverState *bs,
 294                   BlockDriverState *base, BlockDriverState *top, int64_t speed,
 295                   BlockdevOnError on_error, const char *backing_file_str,
 296                   const char *filter_node_name, Error **errp)
 297 {
 298     CommitBlockJob *s;
 299     BlockReopenQueue *reopen_queue = NULL;
 300     int orig_overlay_flags;
 301     int orig_base_flags;
 302     BlockDriverState *iter;
 303     BlockDriverState *overlay_bs;
 304     BlockDriverState *commit_top_bs = NULL;
 305     Error *local_err = NULL;
 306     int ret;
 307
 308     assert(top != bs);
 309     if (top == base) {
 310         error_setg(errp, "Invalid files for merge: top and base are the same");
 311         return;
 312     }
 313
 314     overlay_bs = bdrv_find_overlay(bs, top);
 315
 316     if (overlay_bs == NULL) {
 317         error_setg(errp, "Could not find overlay image for %s:", top->filename);
 318         return;
 319     }
 320
 321     s = block_job_create(job_id, &commit_job_driver, bs, 0, BLK_PERM_ALL,
 322                          speed, BLOCK_JOB_DEFAULT, NULL, NULL, errp);
 323     if (!s) {
 324         return;
 325     }
 326
 327     orig_base_flags    = bdrv_get_flags(base);
 328     orig_overlay_flags = bdrv_get_flags(overlay_bs);
 329
 330     /* convert base & overlay_bs to r/w, if necessary */
 331     if (!(orig_base_flags & BDRV_O_RDWR)) {
 332         reopen_queue = bdrv_reopen_queue(reopen_queue, base, NULL,
 333                                          orig_base_flags | BDRV_O_RDWR);
 334     }
 335     if (!(orig_overlay_flags & BDRV_O_RDWR)) {
 336         reopen_queue = bdrv_reopen_queue(reopen_queue, overlay_bs, NULL,
 337                                          orig_overlay_flags | BDRV_O_RDWR);
 338     }
 339     if (reopen_queue) {
 340         bdrv_reopen_multiple(bdrv_get_aio_context(bs), reopen_queue, &local_err);
 341         if (local_err != NULL) {
 342             error_propagate(errp, local_err);
 343             goto fail;
 344         }
 345     }
 346
 347     /* Insert commit_top block node above top, so we can block consistent read
 348      * on the backing chain below it */
 349     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, filter_node_name, 0,
 350                                          errp);
 351     if (commit_top_bs == NULL) {
 352         goto fail;
 353     }
 354     commit_top_bs->total_sectors = top->total_sectors;
 355     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(top));
 356
 357     bdrv_set_backing_hd(commit_top_bs, top, &local_err);
 358     if (local_err) {
 359         bdrv_unref(commit_top_bs);
 360         commit_top_bs = NULL;
 361         error_propagate(errp, local_err);
 362         goto fail;
 363     }
 364     bdrv_set_backing_hd(overlay_bs, commit_top_bs, &local_err);
 365     if (local_err) {
 366         bdrv_unref(commit_top_bs);
 367         commit_top_bs = NULL;
 368         error_propagate(errp, local_err);
 369         goto fail;
 370     }
 371
 372     s->commit_top_bs = commit_top_bs;
 373     bdrv_unref(commit_top_bs);
 374
 375     /* Block all nodes between top and base, because they will
 376      * disappear from the chain after this operation. */
 377     assert(bdrv_chain_contains(top, base));
 378     for (iter = top; iter != base; iter = backing_bs(iter)) {
 379         /* XXX BLK_PERM_WRITE needs to be allowed so we don't block ourselves
 380          * at s->base (if writes are blocked for a node, they are also blocked
 381          * for its backing file). The other options would be a second filter
 382          * driver above s->base. */
 383         ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
 384                                  BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
 385                                  errp);
 386         if (ret < 0) {
 387             goto fail;
 388         }
 389     }
 390
 391     ret = block_job_add_bdrv(&s->common, "base", base, 0, BLK_PERM_ALL, errp);
 392     if (ret < 0) {
 393         goto fail;
 394     }
 395
 396     /* overlay_bs must be blocked because it needs to be modified to
 397      * update the backing image string. */
 398     ret = block_job_add_bdrv(&s->common, "overlay of top", overlay_bs,
 399                              BLK_PERM_GRAPH_MOD, BLK_PERM_ALL, errp);
 400     if (ret < 0) {
 401         goto fail;
 402     }
 403
 404     s->base = blk_new(BLK_PERM_CONSISTENT_READ
 405                       | BLK_PERM_WRITE
 406                       | BLK_PERM_RESIZE,
 407                       BLK_PERM_CONSISTENT_READ
 408                       | BLK_PERM_GRAPH_MOD
 409                       | BLK_PERM_WRITE_UNCHANGED);
 410     ret = blk_insert_bs(s->base, base, errp);
 411     if (ret < 0) {
 412         goto fail;
 413     }
 414
 415     /* Required permissions are already taken with block_job_add_bdrv() */
 416     s->top = blk_new(0, BLK_PERM_ALL);
 417     ret = blk_insert_bs(s->top, top, errp);
 418     if (ret < 0) {
 419         goto fail;
 420     }
 421
 422     s->active = bs;
 423
 424     s->base_flags          = orig_base_flags;
 425     s->orig_overlay_flags  = orig_overlay_flags;
 426
 427     s->backing_file_str = g_strdup(backing_file_str);
 428
 429     s->on_error = on_error;
 430
 431     trace_commit_start(bs, base, top, s);
 432     block_job_start(&s->common);
 433     return;
 434
 435 fail:
 436     if (s->base) {
 437         blk_unref(s->base);
 438     }
 439     if (s->top) {
 440         blk_unref(s->top);
 441     }
 442     if (commit_top_bs) {
 443         bdrv_set_backing_hd(overlay_bs, top, &error_abort);
 444     }
 445     block_job_early_fail(&s->common);
 446 }
 447
 448
 449 #define COMMIT_BUF_SECTORS 2048
 450
 451 /* commit COW file into the raw image */
 452 int bdrv_commit(BlockDriverState *bs)
 453 {
 454     BlockBackend *src, *backing;
 455     BlockDriverState *backing_file_bs = NULL;
 456     BlockDriverState *commit_top_bs = NULL;
 457     BlockDriver *drv = bs->drv;
 458     int64_t sector, total_sectors, length, backing_length;
 459     int n, ro, open_flags;
 460     int ret = 0;
 461     uint8_t *buf = NULL;
 462     Error *local_err = NULL;
 463
 464     if (!drv)
 465         return -ENOMEDIUM;
 466
 467     if (!bs->backing) {
 468         return -ENOTSUP;
 469     }
 470
 471     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
 472         bdrv_op_is_blocked(bs->backing->bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL)) {
 473         return -EBUSY;
 474     }
 475
 476     ro = bs->backing->bs->read_only;
 477     open_flags =  bs->backing->bs->open_flags;
 478
 479     if (ro) {
 480         if (bdrv_reopen(bs->backing->bs, open_flags | BDRV_O_RDWR, NULL)) {
 481             return -EACCES;
 482         }
 483     }
 484
 485     src = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 486     backing = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
 487
 488     ret = blk_insert_bs(src, bs, &local_err);
 489     if (ret < 0) {
 490         error_report_err(local_err);
 491         goto ro_cleanup;
 492     }
 493
 494     /* Insert commit_top block node above backing, so we can write to it */
 495     backing_file_bs = backing_bs(bs);
 496
 497     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, NULL, BDRV_O_RDWR,
 498                                          &local_err);
 499     if (commit_top_bs == NULL) {
 500         error_report_err(local_err);
 501         goto ro_cleanup;
 502     }
 503     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(backing_file_bs));
 504
 505     bdrv_set_backing_hd(commit_top_bs, backing_file_bs, &error_abort);
 506     bdrv_set_backing_hd(bs, commit_top_bs, &error_abort);
 507
 508     ret = blk_insert_bs(backing, backing_file_bs, &local_err);
 509     if (ret < 0) {
 510         error_report_err(local_err);
 511         goto ro_cleanup;
 512     }
 513
 514     length = blk_getlength(src);
 515     if (length < 0) {
 516         ret = length;
 517         goto ro_cleanup;
 518     }
 519
 520     backing_length = blk_getlength(backing);
 521     if (backing_length < 0) {
 522         ret = backing_length;
 523         goto ro_cleanup;
 524     }
 525
 526     /* If our top snapshot is larger than the backing file image,
 527      * grow the backing file image if possible.  If not possible,
 528      * we must return an error */
 529     if (length > backing_length) {
 530         ret = blk_truncate(backing, length, &local_err);
 531         if (ret < 0) {
 532             error_report_err(local_err);
 533             goto ro_cleanup;
 534         }
 535     }
 536
 537     total_sectors = length >> BDRV_SECTOR_BITS;
 538
 539     /* blk_try_blockalign() for src will choose an alignment that works for
 540      * backing as well, so no need to compare the alignment manually. */
 541     buf = blk_try_blockalign(src, COMMIT_BUF_SECTORS * BDRV_SECTOR_SIZE);
 542     if (buf == NULL) {
 543         ret = -ENOMEM;
 544         goto ro_cleanup;
 545     }
 546
 547     for (sector = 0; sector < total_sectors; sector += n) {
 548         ret = bdrv_is_allocated(bs, sector, COMMIT_BUF_SECTORS, &n);
 549         if (ret < 0) {
 550             goto ro_cleanup;
 551         }
 552         if (ret) {
 553             ret = blk_pread(src, sector * BDRV_SECTOR_SIZE, buf,
 554                             n * BDRV_SECTOR_SIZE);
 555             if (ret < 0) {
 556                 goto ro_cleanup;
 557             }
 558
 559             ret = blk_pwrite(backing, sector * BDRV_SECTOR_SIZE, buf,
 560                              n * BDRV_SECTOR_SIZE, 0);
 561             if (ret < 0) {
 562                 goto ro_cleanup;
 563             }
 564         }
 565     }
 566
 567     if (drv->bdrv_make_empty) {
 568         ret = drv->bdrv_make_empty(bs);
 569         if (ret < 0) {
 570             goto ro_cleanup;
 571         }
 572         blk_flush(src);
 573     }
 574
 575     /*
 576      * Make sure all data we wrote to the backing device is actually
 577      * stable on disk.
 578      */
 579     blk_flush(backing);
 580
 581     ret = 0;
 582 ro_cleanup:
 583     qemu_vfree(buf);
 584
 585     blk_unref(backing);
 586     if (backing_file_bs) {
 587         bdrv_set_backing_hd(bs, backing_file_bs, &error_abort);
 588     }
 589     bdrv_unref(commit_top_bs);
 590     blk_unref(src);
 591
 592     if (ro) {
 593         /* ignoring error return here */
 594         bdrv_reopen(bs->backing->bs, open_flags & ~BDRV_O_RDWR, NULL);
 595     }
 596
 597     return ret;
 598 }