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     BlockdevOnError on_error;
  41     bool base_read_only;
  42     bool chain_frozen;
  43     char *backing_file_str;
  44 } CommitBlockJob;
  45
  46 static int coroutine_fn commit_populate(BlockBackend *bs, BlockBackend *base,
  47                                         int64_t offset, uint64_t bytes,
  48                                         void *buf)
  49 {
  50     int ret = 0;
  51     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, bytes);
  52
  53     assert(bytes < SIZE_MAX);
  54
  55     ret = blk_co_preadv(bs, offset, qiov.size, &qiov, 0);
  56     if (ret < 0) {
  57         return ret;
  58     }
  59
  60     ret = blk_co_pwritev(base, offset, qiov.size, &qiov, 0);
  61     if (ret < 0) {
  62         return ret;
  63     }
  64
  65     return 0;
  66 }
  67
  68 static int commit_prepare(Job *job)
  69 {
  70     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
  71
  72     bdrv_unfreeze_backing_chain(s->commit_top_bs, s->base_bs);
  73     s->chain_frozen = false;
  74
  75     /* Remove base node parent that still uses BLK_PERM_WRITE/RESIZE before
  76      * the normal backing chain can be restored. */
  77     blk_unref(s->base);
  78     s->base = NULL;
  79
  80     /* FIXME: bdrv_drop_intermediate treats total failures and partial failures
  81      * identically. Further work is needed to disambiguate these cases. */
  82     return bdrv_drop_intermediate(s->commit_top_bs, s->base_bs,
  83                                   s->backing_file_str);
  84 }
  85
  86 static void commit_abort(Job *job)
  87 {
  88     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
  89     BlockDriverState *top_bs = blk_bs(s->top);
  90
  91     if (s->chain_frozen) {
  92         bdrv_unfreeze_backing_chain(s->commit_top_bs, s->base_bs);
  93     }
  94
  95     /* Make sure commit_top_bs and top stay around until bdrv_replace_node() */
  96     bdrv_ref(top_bs);
  97     bdrv_ref(s->commit_top_bs);
  98
  99     if (s->base) {
 100         blk_unref(s->base);
 101     }
 102
 103     /* free the blockers on the intermediate nodes so that bdrv_replace_nodes
 104      * can succeed */
 105     block_job_remove_all_bdrv(&s->common);
 106
 107     /* If bdrv_drop_intermediate() failed (or was not invoked), remove the
 108      * commit filter driver from the backing chain now. Do this as the final
 109      * step so that the 'consistent read' permission can be granted.
 110      *
 111      * XXX Can (or should) we somehow keep 'consistent read' blocked even
 112      * after the failed/cancelled commit job is gone? If we already wrote
 113      * something to base, the intermediate images aren't valid any more. */
 114     bdrv_child_try_set_perm(s->commit_top_bs->backing, 0, BLK_PERM_ALL,
 115                             &error_abort);
 116     bdrv_replace_node(s->commit_top_bs, backing_bs(s->commit_top_bs),
 117                       &error_abort);
 118
 119     bdrv_unref(s->commit_top_bs);
 120     bdrv_unref(top_bs);
 121 }
 122
 123 static void commit_clean(Job *job)
 124 {
 125     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
 126
 127     /* restore base open flags here if appropriate (e.g., change the base back
 128      * to r/o). These reopens do not need to be atomic, since we won't abort
 129      * even on failure here */
 130     if (s->base_read_only) {
 131         bdrv_reopen_set_read_only(s->base_bs, true, NULL);
 132     }
 133
 134     g_free(s->backing_file_str);
 135     blk_unref(s->top);
 136 }
 137
 138 static int coroutine_fn commit_run(Job *job, Error **errp)
 139 {
 140     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
 141     int64_t offset;
 142     uint64_t delay_ns = 0;
 143     int ret = 0;
 144     int64_t n = 0; /* bytes */
 145     void *buf = NULL;
 146     int bytes_written = 0;
 147     int64_t len, base_len;
 148
 149     ret = len = blk_getlength(s->top);
 150     if (len < 0) {
 151         goto out;
 152     }
 153     job_progress_set_remaining(&s->common.job, len);
 154
 155     ret = base_len = blk_getlength(s->base);
 156     if (base_len < 0) {
 157         goto out;
 158     }
 159
 160     if (base_len < len) {
 161         ret = blk_truncate(s->base, len, PREALLOC_MODE_OFF, NULL);
 162         if (ret) {
 163             goto out;
 164         }
 165     }
 166
 167     buf = blk_blockalign(s->top, COMMIT_BUFFER_SIZE);
 168
 169     for (offset = 0; offset < len; offset += n) {
 170         bool copy;
 171
 172         /* Note that even when no rate limit is applied we need to yield
 173          * with no pending I/O here so that bdrv_drain_all() returns.
 174          */
 175         job_sleep_ns(&s->common.job, delay_ns);
 176         if (job_is_cancelled(&s->common.job)) {
 177             break;
 178         }
 179         /* Copy if allocated above the base */
 180         ret = bdrv_is_allocated_above(blk_bs(s->top), blk_bs(s->base),
 181                                       offset, COMMIT_BUFFER_SIZE, &n);
 182         copy = (ret == 1);
 183         trace_commit_one_iteration(s, offset, n, ret);
 184         if (copy) {
 185             ret = commit_populate(s->top, s->base, offset, n, buf);
 186             bytes_written += n;
 187         }
 188         if (ret < 0) {
 189             BlockErrorAction action =
 190                 block_job_error_action(&s->common, false, s->on_error, -ret);
 191             if (action == BLOCK_ERROR_ACTION_REPORT) {
 192                 goto out;
 193             } else {
 194                 n = 0;
 195                 continue;
 196             }
 197         }
 198         /* Publish progress */
 199         job_progress_update(&s->common.job, n);
 200
 201         if (copy) {
 202             delay_ns = block_job_ratelimit_get_delay(&s->common, n);
 203         } else {
 204             delay_ns = 0;
 205         }
 206     }
 207
 208     ret = 0;
 209
 210 out:
 211     qemu_vfree(buf);
 212
 213     return ret;
 214 }
 215
 216 static const BlockJobDriver commit_job_driver = {
 217     .job_driver = {
 218         .instance_size = sizeof(CommitBlockJob),
 219         .job_type      = JOB_TYPE_COMMIT,
 220         .free          = block_job_free,
 221         .user_resume   = block_job_user_resume,
 222         .drain         = block_job_drain,
 223         .run           = commit_run,
 224         .prepare       = commit_prepare,
 225         .abort         = commit_abort,
 226         .clean         = commit_clean
 227     },
 228 };
 229
 230 static int coroutine_fn bdrv_commit_top_preadv(BlockDriverState *bs,
 231     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
 232 {
 233     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
 234 }
 235
 236 static void bdrv_commit_top_refresh_filename(BlockDriverState *bs)
 237 {
 238     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
 239             bs->backing->bs->filename);
 240 }
 241
 242 static void bdrv_commit_top_child_perm(BlockDriverState *bs, BdrvChild *c,
 243                                        const BdrvChildRole *role,
 244                                        BlockReopenQueue *reopen_queue,
 245                                        uint64_t perm, uint64_t shared,
 246                                        uint64_t *nperm, uint64_t *nshared)
 247 {
 248     *nperm = 0;
 249     *nshared = BLK_PERM_ALL;
 250 }
 251
 252 /* Dummy node that provides consistent read to its users without requiring it
 253  * from its backing file and that allows writes on the backing file chain. */
 254 static BlockDriver bdrv_commit_top = {
 255     .format_name                = "commit_top",
 256     .bdrv_co_preadv             = bdrv_commit_top_preadv,
 257     .bdrv_co_block_status       = bdrv_co_block_status_from_backing,
 258     .bdrv_refresh_filename      = bdrv_commit_top_refresh_filename,
 259     .bdrv_child_perm            = bdrv_commit_top_child_perm,
 260 };
 261
 262 void commit_start(const char *job_id, BlockDriverState *bs,
 263                   BlockDriverState *base, BlockDriverState *top,
 264                   int creation_flags, int64_t speed,
 265                   BlockdevOnError on_error, const char *backing_file_str,
 266                   const char *filter_node_name, Error **errp)
 267 {
 268     CommitBlockJob *s;
 269     BlockDriverState *iter;
 270     BlockDriverState *commit_top_bs = NULL;
 271     Error *local_err = NULL;
 272     int ret;
 273
 274     assert(top != bs);
 275     if (top == base) {
 276         error_setg(errp, "Invalid files for merge: top and base are the same");
 277         return;
 278     }
 279
 280     s = block_job_create(job_id, &commit_job_driver, NULL, bs, 0, BLK_PERM_ALL,
 281                          speed, creation_flags, NULL, NULL, errp);
 282     if (!s) {
 283         return;
 284     }
 285
 286     /* convert base to r/w, if necessary */
 287     s->base_read_only = bdrv_is_read_only(base);
 288     if (s->base_read_only) {
 289         if (bdrv_reopen_set_read_only(base, false, errp) != 0) {
 290             goto fail;
 291         }
 292     }
 293
 294     /* Insert commit_top block node above top, so we can block consistent read
 295      * on the backing chain below it */
 296     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, filter_node_name, 0,
 297                                          errp);
 298     if (commit_top_bs == NULL) {
 299         goto fail;
 300     }
 301     if (!filter_node_name) {
 302         commit_top_bs->implicit = true;
 303     }
 304     commit_top_bs->total_sectors = top->total_sectors;
 305     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(top));
 306
 307     bdrv_set_backing_hd(commit_top_bs, top, &local_err);
 308     if (local_err) {
 309         bdrv_unref(commit_top_bs);
 310         commit_top_bs = NULL;
 311         error_propagate(errp, local_err);
 312         goto fail;
 313     }
 314     bdrv_replace_node(top, commit_top_bs, &local_err);
 315     if (local_err) {
 316         bdrv_unref(commit_top_bs);
 317         commit_top_bs = NULL;
 318         error_propagate(errp, local_err);
 319         goto fail;
 320     }
 321
 322     s->commit_top_bs = commit_top_bs;
 323     bdrv_unref(commit_top_bs);
 324
 325     /* Block all nodes between top and base, because they will
 326      * disappear from the chain after this operation. */
 327     assert(bdrv_chain_contains(top, base));
 328     for (iter = top; iter != base; iter = backing_bs(iter)) {
 329         /* XXX BLK_PERM_WRITE needs to be allowed so we don't block ourselves
 330          * at s->base (if writes are blocked for a node, they are also blocked
 331          * for its backing file). The other options would be a second filter
 332          * driver above s->base. */
 333         ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
 334                                  BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
 335                                  errp);
 336         if (ret < 0) {
 337             goto fail;
 338         }
 339     }
 340
 341     if (bdrv_freeze_backing_chain(commit_top_bs, base, errp) < 0) {
 342         goto fail;
 343     }
 344     s->chain_frozen = true;
 345
 346     ret = block_job_add_bdrv(&s->common, "base", base, 0, BLK_PERM_ALL, errp);
 347     if (ret < 0) {
 348         goto fail;
 349     }
 350
 351     s->base = blk_new(BLK_PERM_CONSISTENT_READ
 352                       | BLK_PERM_WRITE
 353                       | BLK_PERM_RESIZE,
 354                       BLK_PERM_CONSISTENT_READ
 355                       | BLK_PERM_GRAPH_MOD
 356                       | BLK_PERM_WRITE_UNCHANGED);
 357     ret = blk_insert_bs(s->base, base, errp);
 358     if (ret < 0) {
 359         goto fail;
 360     }
 361     s->base_bs = base;
 362
 363     /* Required permissions are already taken with block_job_add_bdrv() */
 364     s->top = blk_new(0, BLK_PERM_ALL);
 365     ret = blk_insert_bs(s->top, top, errp);
 366     if (ret < 0) {
 367         goto fail;
 368     }
 369
 370     s->backing_file_str = g_strdup(backing_file_str);
 371     s->on_error = on_error;
 372
 373     trace_commit_start(bs, base, top, s);
 374     job_start(&s->common.job);
 375     return;
 376
 377 fail:
 378     if (s->chain_frozen) {
 379         bdrv_unfreeze_backing_chain(commit_top_bs, base);
 380     }
 381     if (s->base) {
 382         blk_unref(s->base);
 383     }
 384     if (s->top) {
 385         blk_unref(s->top);
 386     }
 387     job_early_fail(&s->common.job);
 388     /* commit_top_bs has to be replaced after deleting the block job,
 389      * otherwise this would fail because of lack of permissions. */
 390     if (commit_top_bs) {
 391         bdrv_replace_node(commit_top_bs, top, &error_abort);
 392     }
 393 }
 394
 395
 396 #define COMMIT_BUF_SIZE (2048 * BDRV_SECTOR_SIZE)
 397
 398 /* commit COW file into the raw image */
 399 int bdrv_commit(BlockDriverState *bs)
 400 {
 401     BlockBackend *src, *backing;
 402     BlockDriverState *backing_file_bs = NULL;
 403     BlockDriverState *commit_top_bs = NULL;
 404     BlockDriver *drv = bs->drv;
 405     int64_t offset, length, backing_length;
 406     int ro;
 407     int64_t n;
 408     int ret = 0;
 409     uint8_t *buf = NULL;
 410     Error *local_err = NULL;
 411
 412     if (!drv)
 413         return -ENOMEDIUM;
 414
 415     if (!bs->backing) {
 416         return -ENOTSUP;
 417     }
 418
 419     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
 420         bdrv_op_is_blocked(bs->backing->bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL)) {
 421         return -EBUSY;
 422     }
 423
 424     ro = bs->backing->bs->read_only;
 425
 426     if (ro) {
 427         if (bdrv_reopen_set_read_only(bs->backing->bs, false, NULL)) {
 428             return -EACCES;
 429         }
 430     }
 431
 432     src = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 433     backing = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
 434
 435     ret = blk_insert_bs(src, bs, &local_err);
 436     if (ret < 0) {
 437         error_report_err(local_err);
 438         goto ro_cleanup;
 439     }
 440
 441     /* Insert commit_top block node above backing, so we can write to it */
 442     backing_file_bs = backing_bs(bs);
 443
 444     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, NULL, BDRV_O_RDWR,
 445                                          &local_err);
 446     if (commit_top_bs == NULL) {
 447         error_report_err(local_err);
 448         goto ro_cleanup;
 449     }
 450     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(backing_file_bs));
 451
 452     bdrv_set_backing_hd(commit_top_bs, backing_file_bs, &error_abort);
 453     bdrv_set_backing_hd(bs, commit_top_bs, &error_abort);
 454
 455     ret = blk_insert_bs(backing, backing_file_bs, &local_err);
 456     if (ret < 0) {
 457         error_report_err(local_err);
 458         goto ro_cleanup;
 459     }
 460
 461     length = blk_getlength(src);
 462     if (length < 0) {
 463         ret = length;
 464         goto ro_cleanup;
 465     }
 466
 467     backing_length = blk_getlength(backing);
 468     if (backing_length < 0) {
 469         ret = backing_length;
 470         goto ro_cleanup;
 471     }
 472
 473     /* If our top snapshot is larger than the backing file image,
 474      * grow the backing file image if possible.  If not possible,
 475      * we must return an error */
 476     if (length > backing_length) {
 477         ret = blk_truncate(backing, length, PREALLOC_MODE_OFF, &local_err);
 478         if (ret < 0) {
 479             error_report_err(local_err);
 480             goto ro_cleanup;
 481         }
 482     }
 483
 484     /* blk_try_blockalign() for src will choose an alignment that works for
 485      * backing as well, so no need to compare the alignment manually. */
 486     buf = blk_try_blockalign(src, COMMIT_BUF_SIZE);
 487     if (buf == NULL) {
 488         ret = -ENOMEM;
 489         goto ro_cleanup;
 490     }
 491
 492     for (offset = 0; offset < length; offset += n) {
 493         ret = bdrv_is_allocated(bs, offset, COMMIT_BUF_SIZE, &n);
 494         if (ret < 0) {
 495             goto ro_cleanup;
 496         }
 497         if (ret) {
 498             ret = blk_pread(src, offset, buf, n);
 499             if (ret < 0) {
 500                 goto ro_cleanup;
 501             }
 502
 503             ret = blk_pwrite(backing, offset, buf, n, 0);
 504             if (ret < 0) {
 505                 goto ro_cleanup;
 506             }
 507         }
 508     }
 509
 510     if (drv->bdrv_make_empty) {
 511         ret = drv->bdrv_make_empty(bs);
 512         if (ret < 0) {
 513             goto ro_cleanup;
 514         }
 515         blk_flush(src);
 516     }
 517
 518     /*
 519      * Make sure all data we wrote to the backing device is actually
 520      * stable on disk.
 521      */
 522     blk_flush(backing);
 523
 524     ret = 0;
 525 ro_cleanup:
 526     qemu_vfree(buf);
 527
 528     blk_unref(backing);
 529     if (backing_file_bs) {
 530         bdrv_set_backing_hd(bs, backing_file_bs, &error_abort);
 531     }
 532     bdrv_unref(commit_top_bs);
 533     blk_unref(src);
 534
 535     if (ro) {
 536         /* ignoring error return here */
 537         bdrv_reopen_set_read_only(bs->backing->bs, true, NULL);
 538     }
 539
 540     return ret;
 541 }