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 typedef struct CommitBlockJob {
  35     BlockJob common;
  36     BlockDriverState *commit_top_bs;
  37     BlockBackend *top;
  38     BlockBackend *base;
  39     BlockDriverState *base_bs;
  40     BlockDriverState *base_overlay;
  41     BlockdevOnError on_error;
  42     bool base_read_only;
  43     bool chain_frozen;
  44     char *backing_file_str;
  45 } CommitBlockJob;
  46
  47 static int commit_prepare(Job *job)
  48 {
  49     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
  50
  51     bdrv_unfreeze_backing_chain(s->commit_top_bs, s->base_bs);
  52     s->chain_frozen = false;
  53
  54     /* Remove base node parent that still uses BLK_PERM_WRITE/RESIZE before
  55      * the normal backing chain can be restored. */
  56     blk_unref(s->base);
  57     s->base = NULL;
  58
  59     /* FIXME: bdrv_drop_intermediate treats total failures and partial failures
  60      * identically. Further work is needed to disambiguate these cases. */
  61     return bdrv_drop_intermediate(s->commit_top_bs, s->base_bs,
  62                                   s->backing_file_str);
  63 }
  64
  65 static void commit_abort(Job *job)
  66 {
  67     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
  68     BlockDriverState *top_bs = blk_bs(s->top);
  69
  70     if (s->chain_frozen) {
  71         bdrv_unfreeze_backing_chain(s->commit_top_bs, s->base_bs);
  72     }
  73
  74     /* Make sure commit_top_bs and top stay around until bdrv_replace_node() */
  75     bdrv_ref(top_bs);
  76     bdrv_ref(s->commit_top_bs);
  77
  78     if (s->base) {
  79         blk_unref(s->base);
  80     }
  81
  82     /* free the blockers on the intermediate nodes so that bdrv_replace_nodes
  83      * can succeed */
  84     block_job_remove_all_bdrv(&s->common);
  85
  86     /* If bdrv_drop_intermediate() failed (or was not invoked), remove the
  87      * commit filter driver from the backing chain now. Do this as the final
  88      * step so that the 'consistent read' permission can be granted.
  89      *
  90      * XXX Can (or should) we somehow keep 'consistent read' blocked even
  91      * after the failed/cancelled commit job is gone? If we already wrote
  92      * something to base, the intermediate images aren't valid any more. */
  93     bdrv_replace_node(s->commit_top_bs, s->commit_top_bs->backing->bs,
  94                       &error_abort);
  95
  96     bdrv_unref(s->commit_top_bs);
  97     bdrv_unref(top_bs);
  98 }
  99
 100 static void commit_clean(Job *job)
 101 {
 102     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
 103
 104     /* restore base open flags here if appropriate (e.g., change the base back
 105      * to r/o). These reopens do not need to be atomic, since we won't abort
 106      * even on failure here */
 107     if (s->base_read_only) {
 108         bdrv_reopen_set_read_only(s->base_bs, true, NULL);
 109     }
 110
 111     g_free(s->backing_file_str);
 112     blk_unref(s->top);
 113 }
 114
 115 static int coroutine_fn commit_run(Job *job, Error **errp)
 116 {
 117     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
 118     int64_t offset;
 119     uint64_t delay_ns = 0;
 120     int ret = 0;
 121     int64_t n = 0; /* bytes */
 122     void *buf = NULL;
 123     int64_t len, base_len;
 124
 125     ret = len = blk_getlength(s->top);
 126     if (len < 0) {
 127         goto out;
 128     }
 129     job_progress_set_remaining(&s->common.job, len);
 130
 131     ret = base_len = blk_getlength(s->base);
 132     if (base_len < 0) {
 133         goto out;
 134     }
 135
 136     if (base_len < len) {
 137         ret = blk_truncate(s->base, len, false, PREALLOC_MODE_OFF, 0, NULL);
 138         if (ret) {
 139             goto out;
 140         }
 141     }
 142
 143     buf = blk_blockalign(s->top, COMMIT_BUFFER_SIZE);
 144
 145     for (offset = 0; offset < len; offset += n) {
 146         bool copy;
 147         bool error_in_source = true;
 148
 149         /* Note that even when no rate limit is applied we need to yield
 150          * with no pending I/O here so that bdrv_drain_all() returns.
 151          */
 152         job_sleep_ns(&s->common.job, delay_ns);
 153         if (job_is_cancelled(&s->common.job)) {
 154             break;
 155         }
 156         /* Copy if allocated above the base */
 157         ret = bdrv_is_allocated_above(blk_bs(s->top), s->base_overlay, true,
 158                                       offset, COMMIT_BUFFER_SIZE, &n);
 159         copy = (ret > 0);
 160         trace_commit_one_iteration(s, offset, n, ret);
 161         if (copy) {
 162             assert(n < SIZE_MAX);
 163
 164             ret = blk_co_pread(s->top, offset, n, buf, 0);
 165             if (ret >= 0) {
 166                 ret = blk_co_pwrite(s->base, offset, n, buf, 0);
 167                 if (ret < 0) {
 168                     error_in_source = false;
 169                 }
 170             }
 171         }
 172         if (ret < 0) {
 173             BlockErrorAction action =
 174                 block_job_error_action(&s->common, s->on_error,
 175                                        error_in_source, -ret);
 176             if (action == BLOCK_ERROR_ACTION_REPORT) {
 177                 goto out;
 178             } else {
 179                 n = 0;
 180                 continue;
 181             }
 182         }
 183         /* Publish progress */
 184         job_progress_update(&s->common.job, n);
 185
 186         if (copy) {
 187             delay_ns = block_job_ratelimit_get_delay(&s->common, n);
 188         } else {
 189             delay_ns = 0;
 190         }
 191     }
 192
 193     ret = 0;
 194
 195 out:
 196     qemu_vfree(buf);
 197
 198     return ret;
 199 }
 200
 201 static const BlockJobDriver commit_job_driver = {
 202     .job_driver = {
 203         .instance_size = sizeof(CommitBlockJob),
 204         .job_type      = JOB_TYPE_COMMIT,
 205         .free          = block_job_free,
 206         .user_resume   = block_job_user_resume,
 207         .run           = commit_run,
 208         .prepare       = commit_prepare,
 209         .abort         = commit_abort,
 210         .clean         = commit_clean
 211     },
 212 };
 213
 214 static int coroutine_fn bdrv_commit_top_preadv(BlockDriverState *bs,
 215     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
 216 {
 217     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
 218 }
 219
 220 static void bdrv_commit_top_refresh_filename(BlockDriverState *bs)
 221 {
 222     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
 223             bs->backing->bs->filename);
 224 }
 225
 226 static void bdrv_commit_top_child_perm(BlockDriverState *bs, BdrvChild *c,
 227                                        BdrvChildRole role,
 228                                        BlockReopenQueue *reopen_queue,
 229                                        uint64_t perm, uint64_t shared,
 230                                        uint64_t *nperm, uint64_t *nshared)
 231 {
 232     *nperm = 0;
 233     *nshared = BLK_PERM_ALL;
 234 }
 235
 236 /* Dummy node that provides consistent read to its users without requiring it
 237  * from its backing file and that allows writes on the backing file chain. */
 238 static BlockDriver bdrv_commit_top = {
 239     .format_name                = "commit_top",
 240     .bdrv_co_preadv             = bdrv_commit_top_preadv,
 241     .bdrv_refresh_filename      = bdrv_commit_top_refresh_filename,
 242     .bdrv_child_perm            = bdrv_commit_top_child_perm,
 243
 244     .is_filter                  = true,
 245 };
 246
 247 void commit_start(const char *job_id, BlockDriverState *bs,
 248                   BlockDriverState *base, BlockDriverState *top,
 249                   int creation_flags, int64_t speed,
 250                   BlockdevOnError on_error, const char *backing_file_str,
 251                   const char *filter_node_name, Error **errp)
 252 {
 253     CommitBlockJob *s;
 254     BlockDriverState *iter;
 255     BlockDriverState *commit_top_bs = NULL;
 256     BlockDriverState *filtered_base;
 257     int64_t base_size, top_size;
 258     uint64_t base_perms, iter_shared_perms;
 259     int ret;
 260
 261     assert(top != bs);
 262     if (bdrv_skip_filters(top) == bdrv_skip_filters(base)) {
 263         error_setg(errp, "Invalid files for merge: top and base are the same");
 264         return;
 265     }
 266
 267     base_size = bdrv_getlength(base);
 268     if (base_size < 0) {
 269         error_setg_errno(errp, -base_size, "Could not inquire base image size");
 270         return;
 271     }
 272
 273     top_size = bdrv_getlength(top);
 274     if (top_size < 0) {
 275         error_setg_errno(errp, -top_size, "Could not inquire top image size");
 276         return;
 277     }
 278
 279     base_perms = BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE;
 280     if (base_size < top_size) {
 281         base_perms |= BLK_PERM_RESIZE;
 282     }
 283
 284     s = block_job_create(job_id, &commit_job_driver, NULL, bs, 0, BLK_PERM_ALL,
 285                          speed, creation_flags, NULL, NULL, errp);
 286     if (!s) {
 287         return;
 288     }
 289
 290     /* convert base to r/w, if necessary */
 291     s->base_read_only = bdrv_is_read_only(base);
 292     if (s->base_read_only) {
 293         if (bdrv_reopen_set_read_only(base, false, errp) != 0) {
 294             goto fail;
 295         }
 296     }
 297
 298     /* Insert commit_top block node above top, so we can block consistent read
 299      * on the backing chain below it */
 300     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, filter_node_name, 0,
 301                                          errp);
 302     if (commit_top_bs == NULL) {
 303         goto fail;
 304     }
 305     if (!filter_node_name) {
 306         commit_top_bs->implicit = true;
 307     }
 308
 309     /* So that we can always drop this node */
 310     commit_top_bs->never_freeze = true;
 311
 312     commit_top_bs->total_sectors = top->total_sectors;
 313
 314     ret = bdrv_append(commit_top_bs, top, errp);
 315     bdrv_unref(commit_top_bs); /* referenced by new parents or failed */
 316     if (ret < 0) {
 317         commit_top_bs = NULL;
 318         goto fail;
 319     }
 320
 321     s->commit_top_bs = commit_top_bs;
 322
 323     /*
 324      * Block all nodes between top and base, because they will
 325      * disappear from the chain after this operation.
 326      * Note that this assumes that the user is fine with removing all
 327      * nodes (including R/W filters) between top and base.  Assuring
 328      * this is the responsibility of the interface (i.e. whoever calls
 329      * commit_start()).
 330      */
 331     s->base_overlay = bdrv_find_overlay(top, base);
 332     assert(s->base_overlay);
 333
 334     /*
 335      * The topmost node with
 336      * bdrv_skip_filters(filtered_base) == bdrv_skip_filters(base)
 337      */
 338     filtered_base = bdrv_cow_bs(s->base_overlay);
 339     assert(bdrv_skip_filters(filtered_base) == bdrv_skip_filters(base));
 340
 341     /*
 342      * XXX BLK_PERM_WRITE needs to be allowed so we don't block ourselves
 343      * at s->base (if writes are blocked for a node, they are also blocked
 344      * for its backing file). The other options would be a second filter
 345      * driver above s->base.
 346      */
 347     iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE;
 348
 349     for (iter = top; iter != base; iter = bdrv_filter_or_cow_bs(iter)) {
 350         if (iter == filtered_base) {
 351             /*
 352              * From here on, all nodes are filters on the base.  This
 353              * allows us to share BLK_PERM_CONSISTENT_READ.
 354              */
 355             iter_shared_perms |= BLK_PERM_CONSISTENT_READ;
 356         }
 357
 358         ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
 359                                  iter_shared_perms, errp);
 360         if (ret < 0) {
 361             goto fail;
 362         }
 363     }
 364
 365     if (bdrv_freeze_backing_chain(commit_top_bs, base, errp) < 0) {
 366         goto fail;
 367     }
 368     s->chain_frozen = true;
 369
 370     ret = block_job_add_bdrv(&s->common, "base", base, 0, BLK_PERM_ALL, errp);
 371     if (ret < 0) {
 372         goto fail;
 373     }
 374
 375     s->base = blk_new(s->common.job.aio_context,
 376                       base_perms,
 377                       BLK_PERM_CONSISTENT_READ
 378                       | BLK_PERM_GRAPH_MOD
 379                       | BLK_PERM_WRITE_UNCHANGED);
 380     ret = blk_insert_bs(s->base, base, errp);
 381     if (ret < 0) {
 382         goto fail;
 383     }
 384     blk_set_disable_request_queuing(s->base, true);
 385     s->base_bs = base;
 386
 387     /* Required permissions are already taken with block_job_add_bdrv() */
 388     s->top = blk_new(s->common.job.aio_context, 0, BLK_PERM_ALL);
 389     ret = blk_insert_bs(s->top, top, errp);
 390     if (ret < 0) {
 391         goto fail;
 392     }
 393     blk_set_disable_request_queuing(s->top, true);
 394
 395     s->backing_file_str = g_strdup(backing_file_str);
 396     s->on_error = on_error;
 397
 398     trace_commit_start(bs, base, top, s);
 399     job_start(&s->common.job);
 400     return;
 401
 402 fail:
 403     if (s->chain_frozen) {
 404         bdrv_unfreeze_backing_chain(commit_top_bs, base);
 405     }
 406     if (s->base) {
 407         blk_unref(s->base);
 408     }
 409     if (s->top) {
 410         blk_unref(s->top);
 411     }
 412     if (s->base_read_only) {
 413         bdrv_reopen_set_read_only(base, true, NULL);
 414     }
 415     job_early_fail(&s->common.job);
 416     /* commit_top_bs has to be replaced after deleting the block job,
 417      * otherwise this would fail because of lack of permissions. */
 418     if (commit_top_bs) {
 419         bdrv_replace_node(commit_top_bs, top, &error_abort);
 420     }
 421 }
 422
 423
 424 #define COMMIT_BUF_SIZE (2048 * BDRV_SECTOR_SIZE)
 425
 426 /* commit COW file into the raw image */
 427 int bdrv_commit(BlockDriverState *bs)
 428 {
 429     BlockBackend *src, *backing;
 430     BlockDriverState *backing_file_bs = NULL;
 431     BlockDriverState *commit_top_bs = NULL;
 432     BlockDriver *drv = bs->drv;
 433     AioContext *ctx;
 434     int64_t offset, length, backing_length;
 435     int ro;
 436     int64_t n;
 437     int ret = 0;
 438     uint8_t *buf = NULL;
 439     Error *local_err = NULL;
 440
 441     if (!drv)
 442         return -ENOMEDIUM;
 443
 444     backing_file_bs = bdrv_cow_bs(bs);
 445
 446     if (!backing_file_bs) {
 447         return -ENOTSUP;
 448     }
 449
 450     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
 451         bdrv_op_is_blocked(backing_file_bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL))
 452     {
 453         return -EBUSY;
 454     }
 455
 456     ro = bdrv_is_read_only(backing_file_bs);
 457
 458     if (ro) {
 459         if (bdrv_reopen_set_read_only(backing_file_bs, false, NULL)) {
 460             return -EACCES;
 461         }
 462     }
 463
 464     ctx = bdrv_get_aio_context(bs);
 465     /* WRITE_UNCHANGED is required for bdrv_make_empty() */
 466     src = blk_new(ctx, BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED,
 467                   BLK_PERM_ALL);
 468     backing = blk_new(ctx, BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
 469
 470     ret = blk_insert_bs(src, bs, &local_err);
 471     if (ret < 0) {
 472         error_report_err(local_err);
 473         goto ro_cleanup;
 474     }
 475
 476     /* Insert commit_top block node above backing, so we can write to it */
 477     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, NULL, BDRV_O_RDWR,
 478                                          &local_err);
 479     if (commit_top_bs == NULL) {
 480         error_report_err(local_err);
 481         goto ro_cleanup;
 482     }
 483
 484     bdrv_set_backing_hd(commit_top_bs, backing_file_bs, &error_abort);
 485     bdrv_set_backing_hd(bs, commit_top_bs, &error_abort);
 486
 487     ret = blk_insert_bs(backing, backing_file_bs, &local_err);
 488     if (ret < 0) {
 489         error_report_err(local_err);
 490         goto ro_cleanup;
 491     }
 492
 493     length = blk_getlength(src);
 494     if (length < 0) {
 495         ret = length;
 496         goto ro_cleanup;
 497     }
 498
 499     backing_length = blk_getlength(backing);
 500     if (backing_length < 0) {
 501         ret = backing_length;
 502         goto ro_cleanup;
 503     }
 504
 505     /* If our top snapshot is larger than the backing file image,
 506      * grow the backing file image if possible.  If not possible,
 507      * we must return an error */
 508     if (length > backing_length) {
 509         ret = blk_truncate(backing, length, false, PREALLOC_MODE_OFF, 0,
 510                            &local_err);
 511         if (ret < 0) {
 512             error_report_err(local_err);
 513             goto ro_cleanup;
 514         }
 515     }
 516
 517     /* blk_try_blockalign() for src will choose an alignment that works for
 518      * backing as well, so no need to compare the alignment manually. */
 519     buf = blk_try_blockalign(src, COMMIT_BUF_SIZE);
 520     if (buf == NULL) {
 521         ret = -ENOMEM;
 522         goto ro_cleanup;
 523     }
 524
 525     for (offset = 0; offset < length; offset += n) {
 526         ret = bdrv_is_allocated(bs, offset, COMMIT_BUF_SIZE, &n);
 527         if (ret < 0) {
 528             goto ro_cleanup;
 529         }
 530         if (ret) {
 531             ret = blk_pread(src, offset, buf, n);
 532             if (ret < 0) {
 533                 goto ro_cleanup;
 534             }
 535
 536             ret = blk_pwrite(backing, offset, buf, n, 0);
 537             if (ret < 0) {
 538                 goto ro_cleanup;
 539             }
 540         }
 541     }
 542
 543     ret = blk_make_empty(src, NULL);
 544     /* Ignore -ENOTSUP */
 545     if (ret < 0 && ret != -ENOTSUP) {
 546         goto ro_cleanup;
 547     }
 548
 549     blk_flush(src);
 550
 551     /*
 552      * Make sure all data we wrote to the backing device is actually
 553      * stable on disk.
 554      */
 555     blk_flush(backing);
 556
 557     ret = 0;
 558 ro_cleanup:
 559     qemu_vfree(buf);
 560
 561     blk_unref(backing);
 562     if (bdrv_cow_bs(bs) != backing_file_bs) {
 563         bdrv_set_backing_hd(bs, backing_file_bs, &error_abort);
 564     }
 565     bdrv_unref(commit_top_bs);
 566     blk_unref(src);
 567
 568     if (ro) {
 569         /* ignoring error return here */
 570         bdrv_reopen_set_read_only(backing_file_bs, true, NULL);
 571     }
 572
 573     return ret;
 574 }