block/commit.c

   1 /*
   2  * Live block commit
   3  *
   4  * Copyright Red Hat, Inc. 2012
   5  *
   6  * Authors:
   7  *  Jeff Cody   <[email protected]>
   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     int base_flags;
  42     char *backing_file_str;
  43 } CommitBlockJob;
  44
  45 static int coroutine_fn commit_populate(BlockBackend *bs, BlockBackend *base,
  46                                         int64_t offset, uint64_t bytes,
  47                                         void *buf)
  48 {
  49     int ret = 0;
  50     QEMUIOVector qiov;
  51     struct iovec iov = {
  52         .iov_base = buf,
  53         .iov_len = bytes,
  54     };
  55
  56     assert(bytes < SIZE_MAX);
  57     qemu_iovec_init_external(&qiov, &iov, 1);
  58
  59     ret = blk_co_preadv(bs, offset, qiov.size, &qiov, 0);
  60     if (ret < 0) {
  61         return ret;
  62     }
  63
  64     ret = blk_co_pwritev(base, offset, qiov.size, &qiov, 0);
  65     if (ret < 0) {
  66         return ret;
  67     }
  68
  69     return 0;
  70 }
  71
  72 static int commit_prepare(Job *job)
  73 {
  74     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
  75
  76     /* Remove base node parent that still uses BLK_PERM_WRITE/RESIZE before
  77      * the normal backing chain can be restored. */
  78     blk_unref(s->base);
  79     s->base = NULL;
  80
  81     /* FIXME: bdrv_drop_intermediate treats total failures and partial failures
  82      * identically. Further work is needed to disambiguate these cases. */
  83     return bdrv_drop_intermediate(s->commit_top_bs, s->base_bs,
  84                                   s->backing_file_str);
  85 }
  86
  87 static void commit_abort(Job *job)
  88 {
  89     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
  90     BlockDriverState *top_bs = blk_bs(s->top);
  91
  92     /* Make sure commit_top_bs and top stay around until bdrv_replace_node() */
  93     bdrv_ref(top_bs);
  94     bdrv_ref(s->commit_top_bs);
  95
  96     if (s->base) {
  97         blk_unref(s->base);
  98     }
  99
 100     /* free the blockers on the intermediate nodes so that bdrv_replace_nodes
 101      * can succeed */
 102     block_job_remove_all_bdrv(&s->common);
 103
 104     /* If bdrv_drop_intermediate() failed (or was not invoked), remove the
 105      * commit filter driver from the backing chain now. Do this as the final
 106      * step so that the 'consistent read' permission can be granted.
 107      *
 108      * XXX Can (or should) we somehow keep 'consistent read' blocked even
 109      * after the failed/cancelled commit job is gone? If we already wrote
 110      * something to base, the intermediate images aren't valid any more. */
 111     bdrv_child_try_set_perm(s->commit_top_bs->backing, 0, BLK_PERM_ALL,
 112                             &error_abort);
 113     bdrv_replace_node(s->commit_top_bs, backing_bs(s->commit_top_bs),
 114                       &error_abort);
 115
 116     bdrv_unref(s->commit_top_bs);
 117     bdrv_unref(top_bs);
 118 }
 119
 120 static void commit_clean(Job *job)
 121 {
 122     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
 123
 124     /* restore base open flags here if appropriate (e.g., change the base back
 125      * to r/o). These reopens do not need to be atomic, since we won't abort
 126      * even on failure here */
 127     if (s->base_flags != bdrv_get_flags(s->base_bs)) {
 128         bdrv_reopen(s->base_bs, s->base_flags, NULL);
 129     }
 130
 131     g_free(s->backing_file_str);
 132     blk_unref(s->top);
 133 }
 134
 135 static int coroutine_fn commit_run(Job *job, Error **errp)
 136 {
 137     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
 138     int64_t offset;
 139     uint64_t delay_ns = 0;
 140     int ret = 0;
 141     int64_t n = 0; /* bytes */
 142     void *buf = NULL;
 143     int bytes_written = 0;
 144     int64_t len, base_len;
 145
 146     ret = len = blk_getlength(s->top);
 147     if (len < 0) {
 148         goto out;
 149     }
 150     job_progress_set_remaining(&s->common.job, len);
 151
 152     ret = base_len = blk_getlength(s->base);
 153     if (base_len < 0) {
 154         goto out;
 155     }
 156
 157     if (base_len < len) {
 158         ret = blk_truncate(s->base, len, PREALLOC_MODE_OFF, NULL);
 159         if (ret) {
 160             goto out;
 161         }
 162     }
 163
 164     buf = blk_blockalign(s->top, COMMIT_BUFFER_SIZE);
 165
 166     for (offset = 0; offset < len; offset += n) {
 167         bool copy;
 168
 169         /* Note that even when no rate limit is applied we need to yield
 170          * with no pending I/O here so that bdrv_drain_all() returns.
 171          */
 172         job_sleep_ns(&s->common.job, delay_ns);
 173         if (job_is_cancelled(&s->common.job)) {
 174             break;
 175         }
 176         /* Copy if allocated above the base */
 177         ret = bdrv_is_allocated_above(blk_bs(s->top), blk_bs(s->base),
 178                                       offset, COMMIT_BUFFER_SIZE, &n);
 179         copy = (ret == 1);
 180         trace_commit_one_iteration(s, offset, n, ret);
 181         if (copy) {
 182             ret = commit_populate(s->top, s->base, offset, n, buf);
 183             bytes_written += n;
 184         }
 185         if (ret < 0) {
 186             BlockErrorAction action =
 187                 block_job_error_action(&s->common, false, s->on_error, -ret);
 188             if (action == BLOCK_ERROR_ACTION_REPORT) {
 189                 goto out;
 190             } else {
 191                 n = 0;
 192                 continue;
 193             }
 194         }
 195         /* Publish progress */
 196         job_progress_update(&s->common.job, n);
 197
 198         if (copy) {
 199             delay_ns = block_job_ratelimit_get_delay(&s->common, n);
 200         } else {
 201             delay_ns = 0;
 202         }
 203     }
 204
 205     ret = 0;
 206
 207 out:
 208     qemu_vfree(buf);
 209
 210     return ret;
 211 }
 212
 213 static const BlockJobDriver commit_job_driver = {
 214     .job_driver = {
 215         .instance_size = sizeof(CommitBlockJob),
 216         .job_type      = JOB_TYPE_COMMIT,
 217         .free          = block_job_free,
 218         .user_resume   = block_job_user_resume,
 219         .drain         = block_job_drain,
 220         .run           = commit_run,
 221         .prepare       = commit_prepare,
 222         .abort         = commit_abort,
 223         .clean         = commit_clean
 224     },
 225 };
 226
 227 static int coroutine_fn bdrv_commit_top_preadv(BlockDriverState *bs,
 228     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
 229 {
 230     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
 231 }
 232
 233 static void bdrv_commit_top_refresh_filename(BlockDriverState *bs, QDict *opts)
 234 {
 235     bdrv_refresh_filename(bs->backing->bs);
 236     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
 237             bs->backing->bs->filename);
 238 }
 239
 240 static void bdrv_commit_top_child_perm(BlockDriverState *bs, BdrvChild *c,
 241                                        const BdrvChildRole *role,
 242                                        BlockReopenQueue *reopen_queue,
 243                                        uint64_t perm, uint64_t shared,
 244                                        uint64_t *nperm, uint64_t *nshared)
 245 {
 246     *nperm = 0;
 247     *nshared = BLK_PERM_ALL;
 248 }
 249
 250 /* Dummy node that provides consistent read to its users without requiring it
 251  * from its backing file and that allows writes on the backing file chain. */
 252 static BlockDriver bdrv_commit_top = {
 253     .format_name                = "commit_top",
 254     .bdrv_co_preadv             = bdrv_commit_top_preadv,
 255     .bdrv_co_block_status       = bdrv_co_block_status_from_backing,
 256     .bdrv_refresh_filename      = bdrv_commit_top_refresh_filename,
 257     .bdrv_child_perm            = bdrv_commit_top_child_perm,
 258 };
 259
 260 void commit_start(const char *job_id, BlockDriverState *bs,
 261                   BlockDriverState *base, BlockDriverState *top,
 262                   int creation_flags, int64_t speed,
 263                   BlockdevOnError on_error, const char *backing_file_str,
 264                   const char *filter_node_name, Error **errp)
 265 {
 266     CommitBlockJob *s;
 267     int orig_base_flags;
 268     BlockDriverState *iter;
 269     BlockDriverState *commit_top_bs = NULL;
 270     Error *local_err = NULL;
 271     int ret;
 272
 273     assert(top != bs);
 274     if (top == base) {
 275         error_setg(errp, "Invalid files for merge: top and base are the same");
 276         return;
 277     }
 278
 279     s = block_job_create(job_id, &commit_job_driver, NULL, bs, 0, BLK_PERM_ALL,
 280                          speed, creation_flags, NULL, NULL, errp);
 281     if (!s) {
 282         return;
 283     }
 284
 285     /* convert base to r/w, if necessary */
 286     orig_base_flags = bdrv_get_flags(base);
 287     if (!(orig_base_flags & BDRV_O_RDWR)) {
 288         bdrv_reopen(base, orig_base_flags | BDRV_O_RDWR, &local_err);
 289         if (local_err != NULL) {
 290             error_propagate(errp, local_err);
 291             goto fail;
 292         }
 293     }
 294
 295     /* Insert commit_top block node above top, so we can block consistent read
 296      * on the backing chain below it */
 297     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, filter_node_name, 0,
 298                                          errp);
 299     if (commit_top_bs == NULL) {
 300         goto fail;
 301     }
 302     if (!filter_node_name) {
 303         commit_top_bs->implicit = true;
 304     }
 305     commit_top_bs->total_sectors = top->total_sectors;
 306     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(top));
 307
 308     bdrv_set_backing_hd(commit_top_bs, top, &local_err);
 309     if (local_err) {
 310         bdrv_unref(commit_top_bs);
 311         commit_top_bs = NULL;
 312         error_propagate(errp, local_err);
 313         goto fail;
 314     }
 315     bdrv_replace_node(top, commit_top_bs, &local_err);
 316     if (local_err) {
 317         bdrv_unref(commit_top_bs);
 318         commit_top_bs = NULL;
 319         error_propagate(errp, local_err);
 320         goto fail;
 321     }
 322
 323     s->commit_top_bs = commit_top_bs;
 324     bdrv_unref(commit_top_bs);
 325
 326     /* Block all nodes between top and base, because they will
 327      * disappear from the chain after this operation. */
 328     assert(bdrv_chain_contains(top, base));
 329     for (iter = top; iter != base; iter = backing_bs(iter)) {
 330         /* XXX BLK_PERM_WRITE needs to be allowed so we don't block ourselves
 331          * at s->base (if writes are blocked for a node, they are also blocked
 332          * for its backing file). The other options would be a second filter
 333          * driver above s->base. */
 334         ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
 335                                  BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
 336                                  errp);
 337         if (ret < 0) {
 338             goto fail;
 339         }
 340     }
 341
 342     ret = block_job_add_bdrv(&s->common, "base", base, 0, BLK_PERM_ALL, errp);
 343     if (ret < 0) {
 344         goto fail;
 345     }
 346
 347     s->base = blk_new(BLK_PERM_CONSISTENT_READ
 348                       | BLK_PERM_WRITE
 349                       | BLK_PERM_RESIZE,
 350                       BLK_PERM_CONSISTENT_READ
 351                       | BLK_PERM_GRAPH_MOD
 352                       | BLK_PERM_WRITE_UNCHANGED);
 353     ret = blk_insert_bs(s->base, base, errp);
 354     if (ret < 0) {
 355         goto fail;
 356     }
 357     s->base_bs = base;
 358
 359     /* Required permissions are already taken with block_job_add_bdrv() */
 360     s->top = blk_new(0, BLK_PERM_ALL);
 361     ret = blk_insert_bs(s->top, top, errp);
 362     if (ret < 0) {
 363         goto fail;
 364     }
 365
 366     s->base_flags = orig_base_flags;
 367     s->backing_file_str = g_strdup(backing_file_str);
 368     s->on_error = on_error;
 369
 370     trace_commit_start(bs, base, top, s);
 371     job_start(&s->common.job);
 372     return;
 373
 374 fail:
 375     if (s->base) {
 376         blk_unref(s->base);
 377     }
 378     if (s->top) {
 379         blk_unref(s->top);
 380     }
 381     if (commit_top_bs) {
 382         bdrv_replace_node(commit_top_bs, top, &error_abort);
 383     }
 384     job_early_fail(&s->common.job);
 385 }
 386
 387
 388 #define COMMIT_BUF_SIZE (2048 * BDRV_SECTOR_SIZE)
 389
 390 /* commit COW file into the raw image */
 391 int bdrv_commit(BlockDriverState *bs)
 392 {
 393     BlockBackend *src, *backing;
 394     BlockDriverState *backing_file_bs = NULL;
 395     BlockDriverState *commit_top_bs = NULL;
 396     BlockDriver *drv = bs->drv;
 397     int64_t offset, length, backing_length;
 398     int ro, open_flags;
 399     int64_t n;
 400     int ret = 0;
 401     uint8_t *buf = NULL;
 402     Error *local_err = NULL;
 403
 404     if (!drv)
 405         return -ENOMEDIUM;
 406
 407     if (!bs->backing) {
 408         return -ENOTSUP;
 409     }
 410
 411     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
 412         bdrv_op_is_blocked(bs->backing->bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL)) {
 413         return -EBUSY;
 414     }
 415
 416     ro = bs->backing->bs->read_only;
 417     open_flags =  bs->backing->bs->open_flags;
 418
 419     if (ro) {
 420         if (bdrv_reopen(bs->backing->bs, open_flags | BDRV_O_RDWR, NULL)) {
 421             return -EACCES;
 422         }
 423     }
 424
 425     src = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 426     backing = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
 427
 428     ret = blk_insert_bs(src, bs, &local_err);
 429     if (ret < 0) {
 430         error_report_err(local_err);
 431         goto ro_cleanup;
 432     }
 433
 434     /* Insert commit_top block node above backing, so we can write to it */
 435     backing_file_bs = backing_bs(bs);
 436
 437     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, NULL, BDRV_O_RDWR,
 438                                          &local_err);
 439     if (commit_top_bs == NULL) {
 440         error_report_err(local_err);
 441         goto ro_cleanup;
 442     }
 443     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(backing_file_bs));
 444
 445     bdrv_set_backing_hd(commit_top_bs, backing_file_bs, &error_abort);
 446     bdrv_set_backing_hd(bs, commit_top_bs, &error_abort);
 447
 448     ret = blk_insert_bs(backing, backing_file_bs, &local_err);
 449     if (ret < 0) {
 450         error_report_err(local_err);
 451         goto ro_cleanup;
 452     }
 453
 454     length = blk_getlength(src);
 455     if (length < 0) {
 456         ret = length;
 457         goto ro_cleanup;
 458     }
 459
 460     backing_length = blk_getlength(backing);
 461     if (backing_length < 0) {
 462         ret = backing_length;
 463         goto ro_cleanup;
 464     }
 465
 466     /* If our top snapshot is larger than the backing file image,
 467      * grow the backing file image if possible.  If not possible,
 468      * we must return an error */
 469     if (length > backing_length) {
 470         ret = blk_truncate(backing, length, PREALLOC_MODE_OFF, &local_err);
 471         if (ret < 0) {
 472             error_report_err(local_err);
 473             goto ro_cleanup;
 474         }
 475     }
 476
 477     /* blk_try_blockalign() for src will choose an alignment that works for
 478      * backing as well, so no need to compare the alignment manually. */
 479     buf = blk_try_blockalign(src, COMMIT_BUF_SIZE);
 480     if (buf == NULL) {
 481         ret = -ENOMEM;
 482         goto ro_cleanup;
 483     }
 484
 485     for (offset = 0; offset < length; offset += n) {
 486         ret = bdrv_is_allocated(bs, offset, COMMIT_BUF_SIZE, &n);
 487         if (ret < 0) {
 488             goto ro_cleanup;
 489         }
 490         if (ret) {
 491             ret = blk_pread(src, offset, buf, n);
 492             if (ret < 0) {
 493                 goto ro_cleanup;
 494             }
 495
 496             ret = blk_pwrite(backing, offset, buf, n, 0);
 497             if (ret < 0) {
 498                 goto ro_cleanup;
 499             }
 500         }
 501     }
 502
 503     if (drv->bdrv_make_empty) {
 504         ret = drv->bdrv_make_empty(bs);
 505         if (ret < 0) {
 506             goto ro_cleanup;
 507         }
 508         blk_flush(src);
 509     }
 510
 511     /*
 512      * Make sure all data we wrote to the backing device is actually
 513      * stable on disk.
 514      */
 515     blk_flush(backing);
 516
 517     ret = 0;
 518 ro_cleanup:
 519     qemu_vfree(buf);
 520
 521     blk_unref(backing);
 522     if (backing_file_bs) {
 523         bdrv_set_backing_hd(bs, backing_file_bs, &error_abort);
 524     }
 525     bdrv_unref(commit_top_bs);
 526     blk_unref(src);
 527
 528     if (ro) {
 529         /* ignoring error return here */
 530         bdrv_reopen(bs->backing->bs, open_flags & ~BDRV_O_RDWR, NULL);
 531     }
 532
 533     return ret;
 534 }