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_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     int ret = 0;
 120     int64_t n = 0; /* bytes */
 121     QEMU_AUTO_VFREE void *buf = NULL;
 122     int64_t len, base_len;
 123
 124     len = blk_co_getlength(s->top);
 125     if (len < 0) {
 126         return len;
 127     }
 128     job_progress_set_remaining(&s->common.job, len);
 129
 130     base_len = blk_co_getlength(s->base);
 131     if (base_len < 0) {
 132         return base_len;
 133     }
 134
 135     if (base_len < len) {
 136         ret = blk_co_truncate(s->base, len, false, PREALLOC_MODE_OFF, 0, NULL);
 137         if (ret) {
 138             return ret;
 139         }
 140     }
 141
 142     buf = blk_blockalign(s->top, COMMIT_BUFFER_SIZE);
 143
 144     for (offset = 0; offset < len; offset += n) {
 145         bool copy;
 146         bool error_in_source = true;
 147
 148         /* Note that even when no rate limit is applied we need to yield
 149          * with no pending I/O here so that bdrv_drain_all() returns.
 150          */
 151         block_job_ratelimit_sleep(&s->common);
 152         if (job_is_cancelled(&s->common.job)) {
 153             break;
 154         }
 155         /* Copy if allocated above the base */
 156         ret = blk_co_is_allocated_above(s->top, s->base_overlay, true,
 157                                         offset, COMMIT_BUFFER_SIZE, &n);
 158         copy = (ret > 0);
 159         trace_commit_one_iteration(s, offset, n, ret);
 160         if (copy) {
 161             assert(n < SIZE_MAX);
 162
 163             ret = blk_co_pread(s->top, offset, n, buf, 0);
 164             if (ret >= 0) {
 165                 ret = blk_co_pwrite(s->base, offset, n, buf, 0);
 166                 if (ret < 0) {
 167                     error_in_source = false;
 168                 }
 169             }
 170         }
 171         if (ret < 0) {
 172             BlockErrorAction action =
 173                 block_job_error_action(&s->common, s->on_error,
 174                                        error_in_source, -ret);
 175             if (action == BLOCK_ERROR_ACTION_REPORT) {
 176                 return ret;
 177             } else {
 178                 n = 0;
 179                 continue;
 180             }
 181         }
 182         /* Publish progress */
 183         job_progress_update(&s->common.job, n);
 184
 185         if (copy) {
 186             block_job_ratelimit_processed_bytes(&s->common, n);
 187         }
 188     }
 189
 190     return 0;
 191 }
 192
 193 static const BlockJobDriver commit_job_driver = {
 194     .job_driver = {
 195         .instance_size = sizeof(CommitBlockJob),
 196         .job_type      = JOB_TYPE_COMMIT,
 197         .free          = block_job_free,
 198         .user_resume   = block_job_user_resume,
 199         .run           = commit_run,
 200         .prepare       = commit_prepare,
 201         .abort         = commit_abort,
 202         .clean         = commit_clean
 203     },
 204 };
 205
 206 static int coroutine_fn GRAPH_RDLOCK
 207 bdrv_commit_top_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
 208                        QEMUIOVector *qiov, BdrvRequestFlags flags)
 209 {
 210     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
 211 }
 212
 213 static void bdrv_commit_top_refresh_filename(BlockDriverState *bs)
 214 {
 215     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
 216             bs->backing->bs->filename);
 217 }
 218
 219 static void bdrv_commit_top_child_perm(BlockDriverState *bs, BdrvChild *c,
 220                                        BdrvChildRole role,
 221                                        BlockReopenQueue *reopen_queue,
 222                                        uint64_t perm, uint64_t shared,
 223                                        uint64_t *nperm, uint64_t *nshared)
 224 {
 225     *nperm = 0;
 226     *nshared = BLK_PERM_ALL;
 227 }
 228
 229 /* Dummy node that provides consistent read to its users without requiring it
 230  * from its backing file and that allows writes on the backing file chain. */
 231 static BlockDriver bdrv_commit_top = {
 232     .format_name                = "commit_top",
 233     .bdrv_co_preadv             = bdrv_commit_top_preadv,
 234     .bdrv_refresh_filename      = bdrv_commit_top_refresh_filename,
 235     .bdrv_child_perm            = bdrv_commit_top_child_perm,
 236
 237     .is_filter                  = true,
 238     .filtered_child_is_backing  = true,
 239 };
 240
 241 void commit_start(const char *job_id, BlockDriverState *bs,
 242                   BlockDriverState *base, BlockDriverState *top,
 243                   int creation_flags, int64_t speed,
 244                   BlockdevOnError on_error, const char *backing_file_str,
 245                   const char *filter_node_name, Error **errp)
 246 {
 247     CommitBlockJob *s;
 248     BlockDriverState *iter;
 249     BlockDriverState *commit_top_bs = NULL;
 250     BlockDriverState *filtered_base;
 251     int64_t base_size, top_size;
 252     uint64_t base_perms, iter_shared_perms;
 253     int ret;
 254
 255     GLOBAL_STATE_CODE();
 256
 257     assert(top != bs);
 258     bdrv_graph_rdlock_main_loop();
 259     if (bdrv_skip_filters(top) == bdrv_skip_filters(base)) {
 260         error_setg(errp, "Invalid files for merge: top and base are the same");
 261         bdrv_graph_rdunlock_main_loop();
 262         return;
 263     }
 264     bdrv_graph_rdunlock_main_loop();
 265
 266     base_size = bdrv_getlength(base);
 267     if (base_size < 0) {
 268         error_setg_errno(errp, -base_size, "Could not inquire base image size");
 269         return;
 270     }
 271
 272     top_size = bdrv_getlength(top);
 273     if (top_size < 0) {
 274         error_setg_errno(errp, -top_size, "Could not inquire top image size");
 275         return;
 276     }
 277
 278     base_perms = BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE;
 279     if (base_size < top_size) {
 280         base_perms |= BLK_PERM_RESIZE;
 281     }
 282
 283     s = block_job_create(job_id, &commit_job_driver, NULL, bs, 0, BLK_PERM_ALL,
 284                          speed, creation_flags, NULL, NULL, errp);
 285     if (!s) {
 286         return;
 287     }
 288
 289     /* convert base to r/w, if necessary */
 290     s->base_read_only = bdrv_is_read_only(base);
 291     if (s->base_read_only) {
 292         if (bdrv_reopen_set_read_only(base, false, errp) != 0) {
 293             goto fail;
 294         }
 295     }
 296
 297     /* Insert commit_top block node above top, so we can block consistent read
 298      * on the backing chain below it */
 299     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, filter_node_name, 0,
 300                                          errp);
 301     if (commit_top_bs == NULL) {
 302         goto fail;
 303     }
 304     if (!filter_node_name) {
 305         commit_top_bs->implicit = true;
 306     }
 307
 308     /* So that we can always drop this node */
 309     commit_top_bs->never_freeze = true;
 310
 311     commit_top_bs->total_sectors = top->total_sectors;
 312
 313     ret = bdrv_append(commit_top_bs, top, errp);
 314     bdrv_unref(commit_top_bs); /* referenced by new parents or failed */
 315     if (ret < 0) {
 316         commit_top_bs = NULL;
 317         goto fail;
 318     }
 319
 320     s->commit_top_bs = commit_top_bs;
 321
 322     /*
 323      * Block all nodes between top and base, because they will
 324      * disappear from the chain after this operation.
 325      * Note that this assumes that the user is fine with removing all
 326      * nodes (including R/W filters) between top and base.  Assuring
 327      * this is the responsibility of the interface (i.e. whoever calls
 328      * commit_start()).
 329      */
 330     bdrv_graph_wrlock(top);
 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             bdrv_graph_wrunlock();
 362             goto fail;
 363         }
 364     }
 365
 366     if (bdrv_freeze_backing_chain(commit_top_bs, base, errp) < 0) {
 367         bdrv_graph_wrunlock();
 368         goto fail;
 369     }
 370     s->chain_frozen = true;
 371
 372     ret = block_job_add_bdrv(&s->common, "base", base, 0, BLK_PERM_ALL, errp);
 373     bdrv_graph_wrunlock();
 374
 375     if (ret < 0) {
 376         goto fail;
 377     }
 378
 379     s->base = blk_new(s->common.job.aio_context,
 380                       base_perms,
 381                       BLK_PERM_CONSISTENT_READ
 382                       | BLK_PERM_WRITE_UNCHANGED);
 383     ret = blk_insert_bs(s->base, base, errp);
 384     if (ret < 0) {
 385         goto fail;
 386     }
 387     blk_set_disable_request_queuing(s->base, true);
 388     s->base_bs = base;
 389
 390     /* Required permissions are already taken with block_job_add_bdrv() */
 391     s->top = blk_new(s->common.job.aio_context, 0, BLK_PERM_ALL);
 392     ret = blk_insert_bs(s->top, top, errp);
 393     if (ret < 0) {
 394         goto fail;
 395     }
 396     blk_set_disable_request_queuing(s->top, true);
 397
 398     s->backing_file_str = g_strdup(backing_file_str);
 399     s->on_error = on_error;
 400
 401     trace_commit_start(bs, base, top, s);
 402     job_start(&s->common.job);
 403     return;
 404
 405 fail:
 406     if (s->chain_frozen) {
 407         bdrv_unfreeze_backing_chain(commit_top_bs, base);
 408     }
 409     if (s->base) {
 410         blk_unref(s->base);
 411     }
 412     if (s->top) {
 413         blk_unref(s->top);
 414     }
 415     if (s->base_read_only) {
 416         bdrv_reopen_set_read_only(base, true, NULL);
 417     }
 418     job_early_fail(&s->common.job);
 419     /* commit_top_bs has to be replaced after deleting the block job,
 420      * otherwise this would fail because of lack of permissions. */
 421     if (commit_top_bs) {
 422         bdrv_replace_node(commit_top_bs, top, &error_abort);
 423     }
 424 }
 425
 426
 427 #define COMMIT_BUF_SIZE (2048 * BDRV_SECTOR_SIZE)
 428
 429 /* commit COW file into the raw image */
 430 int bdrv_commit(BlockDriverState *bs)
 431 {
 432     BlockBackend *src, *backing;
 433     BlockDriverState *backing_file_bs = NULL;
 434     BlockDriverState *commit_top_bs = NULL;
 435     BlockDriver *drv = bs->drv;
 436     AioContext *ctx;
 437     int64_t offset, length, backing_length;
 438     int ro;
 439     int64_t n;
 440     int ret = 0;
 441     QEMU_AUTO_VFREE uint8_t *buf = NULL;
 442     Error *local_err = NULL;
 443
 444     GLOBAL_STATE_CODE();
 445     GRAPH_RDLOCK_GUARD_MAINLOOP();
 446
 447     if (!drv)
 448         return -ENOMEDIUM;
 449
 450     backing_file_bs = bdrv_cow_bs(bs);
 451
 452     if (!backing_file_bs) {
 453         return -ENOTSUP;
 454     }
 455
 456     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
 457         bdrv_op_is_blocked(backing_file_bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL))
 458     {
 459         return -EBUSY;
 460     }
 461
 462     ro = bdrv_is_read_only(backing_file_bs);
 463
 464     if (ro) {
 465         if (bdrv_reopen_set_read_only(backing_file_bs, false, NULL)) {
 466             return -EACCES;
 467         }
 468     }
 469
 470     ctx = bdrv_get_aio_context(bs);
 471     /* WRITE_UNCHANGED is required for bdrv_make_empty() */
 472     src = blk_new(ctx, BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED,
 473                   BLK_PERM_ALL);
 474     backing = blk_new(ctx, BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
 475
 476     ret = blk_insert_bs(src, bs, &local_err);
 477     if (ret < 0) {
 478         error_report_err(local_err);
 479         goto ro_cleanup;
 480     }
 481
 482     /* Insert commit_top block node above backing, so we can write to it */
 483     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, NULL, BDRV_O_RDWR,
 484                                          &local_err);
 485     if (commit_top_bs == NULL) {
 486         error_report_err(local_err);
 487         goto ro_cleanup;
 488     }
 489
 490     bdrv_set_backing_hd(commit_top_bs, backing_file_bs, &error_abort);
 491     bdrv_set_backing_hd(bs, commit_top_bs, &error_abort);
 492
 493     ret = blk_insert_bs(backing, backing_file_bs, &local_err);
 494     if (ret < 0) {
 495         error_report_err(local_err);
 496         goto ro_cleanup;
 497     }
 498
 499     length = blk_getlength(src);
 500     if (length < 0) {
 501         ret = length;
 502         goto ro_cleanup;
 503     }
 504
 505     backing_length = blk_getlength(backing);
 506     if (backing_length < 0) {
 507         ret = backing_length;
 508         goto ro_cleanup;
 509     }
 510
 511     /* If our top snapshot is larger than the backing file image,
 512      * grow the backing file image if possible.  If not possible,
 513      * we must return an error */
 514     if (length > backing_length) {
 515         ret = blk_truncate(backing, length, false, PREALLOC_MODE_OFF, 0,
 516                            &local_err);
 517         if (ret < 0) {
 518             error_report_err(local_err);
 519             goto ro_cleanup;
 520         }
 521     }
 522
 523     /* blk_try_blockalign() for src will choose an alignment that works for
 524      * backing as well, so no need to compare the alignment manually. */
 525     buf = blk_try_blockalign(src, COMMIT_BUF_SIZE);
 526     if (buf == NULL) {
 527         ret = -ENOMEM;
 528         goto ro_cleanup;
 529     }
 530
 531     for (offset = 0; offset < length; offset += n) {
 532         ret = bdrv_is_allocated(bs, offset, COMMIT_BUF_SIZE, &n);
 533         if (ret < 0) {
 534             goto ro_cleanup;
 535         }
 536         if (ret) {
 537             ret = blk_pread(src, offset, n, buf, 0);
 538             if (ret < 0) {
 539                 goto ro_cleanup;
 540             }
 541
 542             ret = blk_pwrite(backing, offset, n, buf, 0);
 543             if (ret < 0) {
 544                 goto ro_cleanup;
 545             }
 546         }
 547     }
 548
 549     ret = blk_make_empty(src, NULL);
 550     /* Ignore -ENOTSUP */
 551     if (ret < 0 && ret != -ENOTSUP) {
 552         goto ro_cleanup;
 553     }
 554
 555     blk_flush(src);
 556
 557     /*
 558      * Make sure all data we wrote to the backing device is actually
 559      * stable on disk.
 560      */
 561     blk_flush(backing);
 562
 563     ret = 0;
 564 ro_cleanup:
 565     blk_unref(backing);
 566     if (bdrv_cow_bs(bs) != backing_file_bs) {
 567         bdrv_set_backing_hd(bs, backing_file_bs, &error_abort);
 568     }
 569     bdrv_unref(commit_top_bs);
 570     blk_unref(src);
 571
 572     if (ro) {
 573         /* ignoring error return here */
 574         bdrv_reopen_set_read_only(backing_file_bs, true, NULL);
 575     }
 576
 577     return ret;
 578 }