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 "qemu/ratelimit.h"
  22 #include "qemu/memalign.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_graph_rdlock_main_loop();
  52     bdrv_unfreeze_backing_chain(s->commit_top_bs, s->base_bs);
  53     s->chain_frozen = false;
  54     bdrv_graph_rdunlock_main_loop();
  55
  56     /* Remove base node parent that still uses BLK_PERM_WRITE/RESIZE before
  57      * the normal backing chain can be restored. */
  58     blk_unref(s->base);
  59     s->base = NULL;
  60
  61     /* FIXME: bdrv_drop_intermediate treats total failures and partial failures
  62      * identically. Further work is needed to disambiguate these cases. */
  63     return bdrv_drop_intermediate(s->commit_top_bs, s->base_bs,
  64                                   s->backing_file_str);
  65 }
  66
  67 static void commit_abort(Job *job)
  68 {
  69     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
  70     BlockDriverState *top_bs = blk_bs(s->top);
  71     BlockDriverState *commit_top_backing_bs;
  72
  73     if (s->chain_frozen) {
  74         bdrv_graph_rdlock_main_loop();
  75         bdrv_unfreeze_backing_chain(s->commit_top_bs, s->base_bs);
  76         bdrv_graph_rdunlock_main_loop();
  77     }
  78
  79     /* Make sure commit_top_bs and top stay around until bdrv_replace_node() */
  80     bdrv_ref(top_bs);
  81     bdrv_ref(s->commit_top_bs);
  82
  83     if (s->base) {
  84         blk_unref(s->base);
  85     }
  86
  87     /* free the blockers on the intermediate nodes so that bdrv_replace_nodes
  88      * can succeed */
  89     block_job_remove_all_bdrv(&s->common);
  90
  91     /* If bdrv_drop_intermediate() failed (or was not invoked), remove the
  92      * commit filter driver from the backing chain now. Do this as the final
  93      * step so that the 'consistent read' permission can be granted.
  94      *
  95      * XXX Can (or should) we somehow keep 'consistent read' blocked even
  96      * after the failed/cancelled commit job is gone? If we already wrote
  97      * something to base, the intermediate images aren't valid any more. */
  98     bdrv_graph_rdlock_main_loop();
  99     commit_top_backing_bs = s->commit_top_bs->backing->bs;
 100     bdrv_graph_rdunlock_main_loop();
 101
 102     bdrv_drained_begin(commit_top_backing_bs);
 103     bdrv_graph_wrlock(commit_top_backing_bs);
 104     bdrv_replace_node(s->commit_top_bs, commit_top_backing_bs, &error_abort);
 105     bdrv_graph_wrunlock();
 106     bdrv_drained_end(commit_top_backing_bs);
 107
 108     bdrv_unref(s->commit_top_bs);
 109     bdrv_unref(top_bs);
 110 }
 111
 112 static void commit_clean(Job *job)
 113 {
 114     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
 115
 116     /* restore base open flags here if appropriate (e.g., change the base back
 117      * to r/o). These reopens do not need to be atomic, since we won't abort
 118      * even on failure here */
 119     if (s->base_read_only) {
 120         bdrv_reopen_set_read_only(s->base_bs, true, NULL);
 121     }
 122
 123     g_free(s->backing_file_str);
 124     blk_unref(s->top);
 125 }
 126
 127 static int coroutine_fn commit_run(Job *job, Error **errp)
 128 {
 129     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
 130     int64_t offset;
 131     int ret = 0;
 132     int64_t n = 0; /* bytes */
 133     QEMU_AUTO_VFREE void *buf = NULL;
 134     int64_t len, base_len;
 135
 136     len = blk_co_getlength(s->top);
 137     if (len < 0) {
 138         return len;
 139     }
 140     job_progress_set_remaining(&s->common.job, len);
 141
 142     base_len = blk_co_getlength(s->base);
 143     if (base_len < 0) {
 144         return base_len;
 145     }
 146
 147     if (base_len < len) {
 148         ret = blk_co_truncate(s->base, len, false, PREALLOC_MODE_OFF, 0, NULL);
 149         if (ret) {
 150             return ret;
 151         }
 152     }
 153
 154     buf = blk_blockalign(s->top, COMMIT_BUFFER_SIZE);
 155
 156     for (offset = 0; offset < len; offset += n) {
 157         bool copy;
 158         bool error_in_source = true;
 159
 160         /* Note that even when no rate limit is applied we need to yield
 161          * with no pending I/O here so that bdrv_drain_all() returns.
 162          */
 163         block_job_ratelimit_sleep(&s->common);
 164         if (job_is_cancelled(&s->common.job)) {
 165             break;
 166         }
 167         /* Copy if allocated above the base */
 168         ret = blk_co_is_allocated_above(s->top, s->base_overlay, true,
 169                                         offset, COMMIT_BUFFER_SIZE, &n);
 170         copy = (ret > 0);
 171         trace_commit_one_iteration(s, offset, n, ret);
 172         if (copy) {
 173             assert(n < SIZE_MAX);
 174
 175             ret = blk_co_pread(s->top, offset, n, buf, 0);
 176             if (ret >= 0) {
 177                 ret = blk_co_pwrite(s->base, offset, n, buf, 0);
 178                 if (ret < 0) {
 179                     error_in_source = false;
 180                 }
 181             }
 182         }
 183         if (ret < 0) {
 184             BlockErrorAction action =
 185                 block_job_error_action(&s->common, s->on_error,
 186                                        error_in_source, -ret);
 187             if (action == BLOCK_ERROR_ACTION_REPORT) {
 188                 return ret;
 189             } else {
 190                 n = 0;
 191                 continue;
 192             }
 193         }
 194         /* Publish progress */
 195         job_progress_update(&s->common.job, n);
 196
 197         if (copy) {
 198             block_job_ratelimit_processed_bytes(&s->common, n);
 199         }
 200     }
 201
 202     return 0;
 203 }
 204
 205 static const BlockJobDriver commit_job_driver = {
 206     .job_driver = {
 207         .instance_size = sizeof(CommitBlockJob),
 208         .job_type      = JOB_TYPE_COMMIT,
 209         .free          = block_job_free,
 210         .user_resume   = block_job_user_resume,
 211         .run           = commit_run,
 212         .prepare       = commit_prepare,
 213         .abort         = commit_abort,
 214         .clean         = commit_clean
 215     },
 216 };
 217
 218 static int coroutine_fn GRAPH_RDLOCK
 219 bdrv_commit_top_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
 220                        QEMUIOVector *qiov, BdrvRequestFlags flags)
 221 {
 222     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
 223 }
 224
 225 static GRAPH_RDLOCK void bdrv_commit_top_refresh_filename(BlockDriverState *bs)
 226 {
 227     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
 228             bs->backing->bs->filename);
 229 }
 230
 231 static void bdrv_commit_top_child_perm(BlockDriverState *bs, BdrvChild *c,
 232                                        BdrvChildRole role,
 233                                        BlockReopenQueue *reopen_queue,
 234                                        uint64_t perm, uint64_t shared,
 235                                        uint64_t *nperm, uint64_t *nshared)
 236 {
 237     *nperm = 0;
 238     *nshared = BLK_PERM_ALL;
 239 }
 240
 241 /* Dummy node that provides consistent read to its users without requiring it
 242  * from its backing file and that allows writes on the backing file chain. */
 243 static BlockDriver bdrv_commit_top = {
 244     .format_name                = "commit_top",
 245     .bdrv_co_preadv             = bdrv_commit_top_preadv,
 246     .bdrv_refresh_filename      = bdrv_commit_top_refresh_filename,
 247     .bdrv_child_perm            = bdrv_commit_top_child_perm,
 248
 249     .is_filter                  = true,
 250     .filtered_child_is_backing  = true,
 251 };
 252
 253 void commit_start(const char *job_id, BlockDriverState *bs,
 254                   BlockDriverState *base, BlockDriverState *top,
 255                   int creation_flags, int64_t speed,
 256                   BlockdevOnError on_error, const char *backing_file_str,
 257                   const char *filter_node_name, Error **errp)
 258 {
 259     CommitBlockJob *s;
 260     BlockDriverState *iter;
 261     BlockDriverState *commit_top_bs = NULL;
 262     BlockDriverState *filtered_base;
 263     int64_t base_size, top_size;
 264     uint64_t base_perms, iter_shared_perms;
 265     int ret;
 266
 267     GLOBAL_STATE_CODE();
 268
 269     assert(top != bs);
 270     bdrv_graph_rdlock_main_loop();
 271     if (bdrv_skip_filters(top) == bdrv_skip_filters(base)) {
 272         error_setg(errp, "Invalid files for merge: top and base are the same");
 273         bdrv_graph_rdunlock_main_loop();
 274         return;
 275     }
 276     bdrv_graph_rdunlock_main_loop();
 277
 278     base_size = bdrv_getlength(base);
 279     if (base_size < 0) {
 280         error_setg_errno(errp, -base_size, "Could not inquire base image size");
 281         return;
 282     }
 283
 284     top_size = bdrv_getlength(top);
 285     if (top_size < 0) {
 286         error_setg_errno(errp, -top_size, "Could not inquire top image size");
 287         return;
 288     }
 289
 290     base_perms = BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE;
 291     if (base_size < top_size) {
 292         base_perms |= BLK_PERM_RESIZE;
 293     }
 294
 295     s = block_job_create(job_id, &commit_job_driver, NULL, bs, 0, BLK_PERM_ALL,
 296                          speed, creation_flags, NULL, NULL, errp);
 297     if (!s) {
 298         return;
 299     }
 300
 301     /* convert base to r/w, if necessary */
 302     s->base_read_only = bdrv_is_read_only(base);
 303     if (s->base_read_only) {
 304         if (bdrv_reopen_set_read_only(base, false, errp) != 0) {
 305             goto fail;
 306         }
 307     }
 308
 309     /* Insert commit_top block node above top, so we can block consistent read
 310      * on the backing chain below it */
 311     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, filter_node_name, 0,
 312                                          errp);
 313     if (commit_top_bs == NULL) {
 314         goto fail;
 315     }
 316     if (!filter_node_name) {
 317         commit_top_bs->implicit = true;
 318     }
 319
 320     /* So that we can always drop this node */
 321     commit_top_bs->never_freeze = true;
 322
 323     commit_top_bs->total_sectors = top->total_sectors;
 324
 325     ret = bdrv_append(commit_top_bs, top, errp);
 326     bdrv_unref(commit_top_bs); /* referenced by new parents or failed */
 327     if (ret < 0) {
 328         commit_top_bs = NULL;
 329         goto fail;
 330     }
 331
 332     s->commit_top_bs = commit_top_bs;
 333
 334     /*
 335      * Block all nodes between top and base, because they will
 336      * disappear from the chain after this operation.
 337      * Note that this assumes that the user is fine with removing all
 338      * nodes (including R/W filters) between top and base.  Assuring
 339      * this is the responsibility of the interface (i.e. whoever calls
 340      * commit_start()).
 341      */
 342     bdrv_graph_wrlock(top);
 343     s->base_overlay = bdrv_find_overlay(top, base);
 344     assert(s->base_overlay);
 345
 346     /*
 347      * The topmost node with
 348      * bdrv_skip_filters(filtered_base) == bdrv_skip_filters(base)
 349      */
 350     filtered_base = bdrv_cow_bs(s->base_overlay);
 351     assert(bdrv_skip_filters(filtered_base) == bdrv_skip_filters(base));
 352
 353     /*
 354      * XXX BLK_PERM_WRITE needs to be allowed so we don't block ourselves
 355      * at s->base (if writes are blocked for a node, they are also blocked
 356      * for its backing file). The other options would be a second filter
 357      * driver above s->base.
 358      */
 359     iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE;
 360
 361     for (iter = top; iter != base; iter = bdrv_filter_or_cow_bs(iter)) {
 362         if (iter == filtered_base) {
 363             /*
 364              * From here on, all nodes are filters on the base.  This
 365              * allows us to share BLK_PERM_CONSISTENT_READ.
 366              */
 367             iter_shared_perms |= BLK_PERM_CONSISTENT_READ;
 368         }
 369
 370         ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
 371                                  iter_shared_perms, errp);
 372         if (ret < 0) {
 373             bdrv_graph_wrunlock();
 374             goto fail;
 375         }
 376     }
 377
 378     if (bdrv_freeze_backing_chain(commit_top_bs, base, errp) < 0) {
 379         bdrv_graph_wrunlock();
 380         goto fail;
 381     }
 382     s->chain_frozen = true;
 383
 384     ret = block_job_add_bdrv(&s->common, "base", base, 0, BLK_PERM_ALL, errp);
 385     bdrv_graph_wrunlock();
 386
 387     if (ret < 0) {
 388         goto fail;
 389     }
 390
 391     s->base = blk_new(s->common.job.aio_context,
 392                       base_perms,
 393                       BLK_PERM_CONSISTENT_READ
 394                       | BLK_PERM_WRITE_UNCHANGED);
 395     ret = blk_insert_bs(s->base, base, errp);
 396     if (ret < 0) {
 397         goto fail;
 398     }
 399     blk_set_disable_request_queuing(s->base, true);
 400     s->base_bs = base;
 401
 402     /* Required permissions are already taken with block_job_add_bdrv() */
 403     s->top = blk_new(s->common.job.aio_context, 0, BLK_PERM_ALL);
 404     ret = blk_insert_bs(s->top, top, errp);
 405     if (ret < 0) {
 406         goto fail;
 407     }
 408     blk_set_disable_request_queuing(s->top, true);
 409
 410     s->backing_file_str = g_strdup(backing_file_str);
 411     s->on_error = on_error;
 412
 413     trace_commit_start(bs, base, top, s);
 414     job_start(&s->common.job);
 415     return;
 416
 417 fail:
 418     if (s->chain_frozen) {
 419         bdrv_graph_rdlock_main_loop();
 420         bdrv_unfreeze_backing_chain(commit_top_bs, base);
 421         bdrv_graph_rdunlock_main_loop();
 422     }
 423     if (s->base) {
 424         blk_unref(s->base);
 425     }
 426     if (s->top) {
 427         blk_unref(s->top);
 428     }
 429     if (s->base_read_only) {
 430         bdrv_reopen_set_read_only(base, true, NULL);
 431     }
 432     job_early_fail(&s->common.job);
 433     /* commit_top_bs has to be replaced after deleting the block job,
 434      * otherwise this would fail because of lack of permissions. */
 435     if (commit_top_bs) {
 436         bdrv_drained_begin(top);
 437         bdrv_graph_wrlock(top);
 438         bdrv_replace_node(commit_top_bs, top, &error_abort);
 439         bdrv_graph_wrunlock();
 440         bdrv_drained_end(top);
 441     }
 442 }
 443
 444
 445 #define COMMIT_BUF_SIZE (2048 * BDRV_SECTOR_SIZE)
 446
 447 /* commit COW file into the raw image */
 448 int bdrv_commit(BlockDriverState *bs)
 449 {
 450     BlockBackend *src, *backing;
 451     BlockDriverState *backing_file_bs = NULL;
 452     BlockDriverState *commit_top_bs = NULL;
 453     BlockDriver *drv = bs->drv;
 454     AioContext *ctx;
 455     int64_t offset, length, backing_length;
 456     int ro;
 457     int64_t n;
 458     int ret = 0;
 459     QEMU_AUTO_VFREE uint8_t *buf = NULL;
 460     Error *local_err = NULL;
 461
 462     GLOBAL_STATE_CODE();
 463     GRAPH_RDLOCK_GUARD_MAINLOOP();
 464
 465     if (!drv)
 466         return -ENOMEDIUM;
 467
 468     backing_file_bs = bdrv_cow_bs(bs);
 469
 470     if (!backing_file_bs) {
 471         return -ENOTSUP;
 472     }
 473
 474     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
 475         bdrv_op_is_blocked(backing_file_bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL))
 476     {
 477         return -EBUSY;
 478     }
 479
 480     ro = bdrv_is_read_only(backing_file_bs);
 481
 482     if (ro) {
 483         if (bdrv_reopen_set_read_only(backing_file_bs, false, NULL)) {
 484             return -EACCES;
 485         }
 486     }
 487
 488     ctx = bdrv_get_aio_context(bs);
 489     /* WRITE_UNCHANGED is required for bdrv_make_empty() */
 490     src = blk_new(ctx, BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED,
 491                   BLK_PERM_ALL);
 492     backing = blk_new(ctx, BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
 493
 494     ret = blk_insert_bs(src, bs, &local_err);
 495     if (ret < 0) {
 496         error_report_err(local_err);
 497         goto ro_cleanup;
 498     }
 499
 500     /* Insert commit_top block node above backing, so we can write to it */
 501     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, NULL, BDRV_O_RDWR,
 502                                          &local_err);
 503     if (commit_top_bs == NULL) {
 504         error_report_err(local_err);
 505         goto ro_cleanup;
 506     }
 507
 508     bdrv_set_backing_hd(commit_top_bs, backing_file_bs, &error_abort);
 509     bdrv_set_backing_hd(bs, commit_top_bs, &error_abort);
 510
 511     ret = blk_insert_bs(backing, backing_file_bs, &local_err);
 512     if (ret < 0) {
 513         error_report_err(local_err);
 514         goto ro_cleanup;
 515     }
 516
 517     length = blk_getlength(src);
 518     if (length < 0) {
 519         ret = length;
 520         goto ro_cleanup;
 521     }
 522
 523     backing_length = blk_getlength(backing);
 524     if (backing_length < 0) {
 525         ret = backing_length;
 526         goto ro_cleanup;
 527     }
 528
 529     /* If our top snapshot is larger than the backing file image,
 530      * grow the backing file image if possible.  If not possible,
 531      * we must return an error */
 532     if (length > backing_length) {
 533         ret = blk_truncate(backing, length, false, PREALLOC_MODE_OFF, 0,
 534                            &local_err);
 535         if (ret < 0) {
 536             error_report_err(local_err);
 537             goto ro_cleanup;
 538         }
 539     }
 540
 541     /* blk_try_blockalign() for src will choose an alignment that works for
 542      * backing as well, so no need to compare the alignment manually. */
 543     buf = blk_try_blockalign(src, COMMIT_BUF_SIZE);
 544     if (buf == NULL) {
 545         ret = -ENOMEM;
 546         goto ro_cleanup;
 547     }
 548
 549     for (offset = 0; offset < length; offset += n) {
 550         ret = bdrv_is_allocated(bs, offset, COMMIT_BUF_SIZE, &n);
 551         if (ret < 0) {
 552             goto ro_cleanup;
 553         }
 554         if (ret) {
 555             ret = blk_pread(src, offset, n, buf, 0);
 556             if (ret < 0) {
 557                 goto ro_cleanup;
 558             }
 559
 560             ret = blk_pwrite(backing, offset, n, buf, 0);
 561             if (ret < 0) {
 562                 goto ro_cleanup;
 563             }
 564         }
 565     }
 566
 567     ret = blk_make_empty(src, NULL);
 568     /* Ignore -ENOTSUP */
 569     if (ret < 0 && ret != -ENOTSUP) {
 570         goto ro_cleanup;
 571     }
 572
 573     blk_flush(src);
 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     blk_unref(backing);
 584     if (bdrv_cow_bs(bs) != backing_file_bs) {
 585         bdrv_set_backing_hd(bs, backing_file_bs, &error_abort);
 586     }
 587     bdrv_unref(commit_top_bs);
 588     blk_unref(src);
 589
 590     if (ro) {
 591         /* ignoring error return here */
 592         bdrv_reopen_set_read_only(backing_file_bs, true, NULL);
 593     }
 594
 595     return ret;
 596 }