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     BlockdevOnError on_error;
  40     int base_flags;
  41     char *backing_file_str;
  42 } CommitBlockJob;
  43
  44 static int coroutine_fn commit_populate(BlockBackend *bs, BlockBackend *base,
  45                                         int64_t offset, uint64_t bytes,
  46                                         void *buf)
  47 {
  48     int ret = 0;
  49     QEMUIOVector qiov;
  50     struct iovec iov = {
  51         .iov_base = buf,
  52         .iov_len = bytes,
  53     };
  54
  55     assert(bytes < SIZE_MAX);
  56     qemu_iovec_init_external(&qiov, &iov, 1);
  57
  58     ret = blk_co_preadv(bs, offset, qiov.size, &qiov, 0);
  59     if (ret < 0) {
  60         return ret;
  61     }
  62
  63     ret = blk_co_pwritev(base, offset, qiov.size, &qiov, 0);
  64     if (ret < 0) {
  65         return ret;
  66     }
  67
  68     return 0;
  69 }
  70
  71 typedef struct {
  72     int ret;
  73 } CommitCompleteData;
  74
  75 static void commit_complete(BlockJob *job, void *opaque)
  76 {
  77     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
  78     CommitCompleteData *data = opaque;
  79     BlockDriverState *top = blk_bs(s->top);
  80     BlockDriverState *base = blk_bs(s->base);
  81     BlockDriverState *commit_top_bs = s->commit_top_bs;
  82     int ret = data->ret;
  83     bool remove_commit_top_bs = false;
  84
  85     /* Make sure commit_top_bs and top stay around until bdrv_replace_node() */
  86     bdrv_ref(top);
  87     bdrv_ref(commit_top_bs);
  88
  89     /* Remove base node parent that still uses BLK_PERM_WRITE/RESIZE before
  90      * the normal backing chain can be restored. */
  91     blk_unref(s->base);
  92
  93     if (!block_job_is_cancelled(&s->common) && ret == 0) {
  94         /* success */
  95         ret = bdrv_drop_intermediate(s->commit_top_bs, base,
  96                                      s->backing_file_str);
  97     } else {
  98         /* XXX Can (or should) we somehow keep 'consistent read' blocked even
  99          * after the failed/cancelled commit job is gone? If we already wrote
 100          * something to base, the intermediate images aren't valid any more. */
 101         remove_commit_top_bs = true;
 102     }
 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_flags != bdrv_get_flags(base)) {
 108         bdrv_reopen(base, s->base_flags, NULL);
 109     }
 110     g_free(s->backing_file_str);
 111     blk_unref(s->top);
 112
 113     /* If there is more than one reference to the job (e.g. if called from
 114      * block_job_finish_sync()), block_job_completed() won't free it and
 115      * therefore the blockers on the intermediate nodes remain. This would
 116      * cause bdrv_set_backing_hd() to fail. */
 117     block_job_remove_all_bdrv(job);
 118
 119     block_job_completed(&s->common, ret);
 120     g_free(data);
 121
 122     /* If bdrv_drop_intermediate() didn't already do that, remove the commit
 123      * filter driver from the backing chain. Do this as the final step so that
 124      * the 'consistent read' permission can be granted.  */
 125     if (remove_commit_top_bs) {
 126         bdrv_child_try_set_perm(commit_top_bs->backing, 0, BLK_PERM_ALL,
 127                                 &error_abort);
 128         bdrv_replace_node(commit_top_bs, backing_bs(commit_top_bs),
 129                           &error_abort);
 130     }
 131
 132     bdrv_unref(commit_top_bs);
 133     bdrv_unref(top);
 134 }
 135
 136 static void coroutine_fn commit_run(void *opaque)
 137 {
 138     CommitBlockJob *s = opaque;
 139     CommitCompleteData *data;
 140     int64_t offset;
 141     uint64_t delay_ns = 0;
 142     int ret = 0;
 143     int64_t n = 0; /* bytes */
 144     void *buf = NULL;
 145     int bytes_written = 0;
 146     int64_t len, base_len;
 147
 148     ret = len = blk_getlength(s->top);
 149     if (len < 0) {
 150         goto out;
 151     }
 152     block_job_progress_set_remaining(&s->common, len);
 153
 154     ret = base_len = blk_getlength(s->base);
 155     if (base_len < 0) {
 156         goto out;
 157     }
 158
 159     if (base_len < len) {
 160         ret = blk_truncate(s->base, len, PREALLOC_MODE_OFF, NULL);
 161         if (ret) {
 162             goto out;
 163         }
 164     }
 165
 166     buf = blk_blockalign(s->top, COMMIT_BUFFER_SIZE);
 167
 168     for (offset = 0; offset < len; offset += n) {
 169         bool copy;
 170
 171         /* Note that even when no rate limit is applied we need to yield
 172          * with no pending I/O here so that bdrv_drain_all() returns.
 173          */
 174         block_job_sleep_ns(&s->common, delay_ns);
 175         if (block_job_is_cancelled(&s->common)) {
 176             break;
 177         }
 178         /* Copy if allocated above the base */
 179         ret = bdrv_is_allocated_above(blk_bs(s->top), blk_bs(s->base),
 180                                       offset, COMMIT_BUFFER_SIZE, &n);
 181         copy = (ret == 1);
 182         trace_commit_one_iteration(s, offset, n, ret);
 183         if (copy) {
 184             ret = commit_populate(s->top, s->base, offset, n, buf);
 185             bytes_written += n;
 186         }
 187         if (ret < 0) {
 188             BlockErrorAction action =
 189                 block_job_error_action(&s->common, false, s->on_error, -ret);
 190             if (action == BLOCK_ERROR_ACTION_REPORT) {
 191                 goto out;
 192             } else {
 193                 n = 0;
 194                 continue;
 195             }
 196         }
 197         /* Publish progress */
 198         block_job_progress_update(&s->common, n);
 199
 200         if (copy) {
 201             delay_ns = block_job_ratelimit_get_delay(&s->common, n);
 202         } else {
 203             delay_ns = 0;
 204         }
 205     }
 206
 207     ret = 0;
 208
 209 out:
 210     qemu_vfree(buf);
 211
 212     data = g_malloc(sizeof(*data));
 213     data->ret = ret;
 214     block_job_defer_to_main_loop(&s->common, commit_complete, data);
 215 }
 216
 217 static const BlockJobDriver commit_job_driver = {
 218     .instance_size = sizeof(CommitBlockJob),
 219     .job_type      = BLOCK_JOB_TYPE_COMMIT,
 220     .start         = commit_run,
 221 };
 222
 223 static int coroutine_fn bdrv_commit_top_preadv(BlockDriverState *bs,
 224     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
 225 {
 226     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
 227 }
 228
 229 static void bdrv_commit_top_refresh_filename(BlockDriverState *bs, QDict *opts)
 230 {
 231     bdrv_refresh_filename(bs->backing->bs);
 232     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
 233             bs->backing->bs->filename);
 234 }
 235
 236 static void bdrv_commit_top_close(BlockDriverState *bs)
 237 {
 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_close                 = bdrv_commit_top_close,
 258     .bdrv_child_perm            = bdrv_commit_top_child_perm,
 259 };
 260
 261 void commit_start(const char *job_id, BlockDriverState *bs,
 262                   BlockDriverState *base, BlockDriverState *top, 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, BLOCK_JOB_DEFAULT, 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
 358     /* Required permissions are already taken with block_job_add_bdrv() */
 359     s->top = blk_new(0, BLK_PERM_ALL);
 360     ret = blk_insert_bs(s->top, top, errp);
 361     if (ret < 0) {
 362         goto fail;
 363     }
 364
 365     s->base_flags = orig_base_flags;
 366     s->backing_file_str = g_strdup(backing_file_str);
 367     s->on_error = on_error;
 368
 369     trace_commit_start(bs, base, top, s);
 370     block_job_start(&s->common);
 371     return;
 372
 373 fail:
 374     if (s->base) {
 375         blk_unref(s->base);
 376     }
 377     if (s->top) {
 378         blk_unref(s->top);
 379     }
 380     if (commit_top_bs) {
 381         bdrv_replace_node(commit_top_bs, top, &error_abort);
 382     }
 383     block_job_early_fail(&s->common);
 384 }
 385
 386
 387 #define COMMIT_BUF_SIZE (2048 * BDRV_SECTOR_SIZE)
 388
 389 /* commit COW file into the raw image */
 390 int bdrv_commit(BlockDriverState *bs)
 391 {
 392     BlockBackend *src, *backing;
 393     BlockDriverState *backing_file_bs = NULL;
 394     BlockDriverState *commit_top_bs = NULL;
 395     BlockDriver *drv = bs->drv;
 396     int64_t offset, length, backing_length;
 397     int ro, open_flags;
 398     int64_t n;
 399     int ret = 0;
 400     uint8_t *buf = NULL;
 401     Error *local_err = NULL;
 402
 403     if (!drv)
 404         return -ENOMEDIUM;
 405
 406     if (!bs->backing) {
 407         return -ENOTSUP;
 408     }
 409
 410     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
 411         bdrv_op_is_blocked(bs->backing->bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL)) {
 412         return -EBUSY;
 413     }
 414
 415     ro = bs->backing->bs->read_only;
 416     open_flags =  bs->backing->bs->open_flags;
 417
 418     if (ro) {
 419         if (bdrv_reopen(bs->backing->bs, open_flags | BDRV_O_RDWR, NULL)) {
 420             return -EACCES;
 421         }
 422     }
 423
 424     src = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 425     backing = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
 426
 427     ret = blk_insert_bs(src, bs, &local_err);
 428     if (ret < 0) {
 429         error_report_err(local_err);
 430         goto ro_cleanup;
 431     }
 432
 433     /* Insert commit_top block node above backing, so we can write to it */
 434     backing_file_bs = backing_bs(bs);
 435
 436     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, NULL, BDRV_O_RDWR,
 437                                          &local_err);
 438     if (commit_top_bs == NULL) {
 439         error_report_err(local_err);
 440         goto ro_cleanup;
 441     }
 442     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(backing_file_bs));
 443
 444     bdrv_set_backing_hd(commit_top_bs, backing_file_bs, &error_abort);
 445     bdrv_set_backing_hd(bs, commit_top_bs, &error_abort);
 446
 447     ret = blk_insert_bs(backing, backing_file_bs, &local_err);
 448     if (ret < 0) {
 449         error_report_err(local_err);
 450         goto ro_cleanup;
 451     }
 452
 453     length = blk_getlength(src);
 454     if (length < 0) {
 455         ret = length;
 456         goto ro_cleanup;
 457     }
 458
 459     backing_length = blk_getlength(backing);
 460     if (backing_length < 0) {
 461         ret = backing_length;
 462         goto ro_cleanup;
 463     }
 464
 465     /* If our top snapshot is larger than the backing file image,
 466      * grow the backing file image if possible.  If not possible,
 467      * we must return an error */
 468     if (length > backing_length) {
 469         ret = blk_truncate(backing, length, PREALLOC_MODE_OFF, &local_err);
 470         if (ret < 0) {
 471             error_report_err(local_err);
 472             goto ro_cleanup;
 473         }
 474     }
 475
 476     /* blk_try_blockalign() for src will choose an alignment that works for
 477      * backing as well, so no need to compare the alignment manually. */
 478     buf = blk_try_blockalign(src, COMMIT_BUF_SIZE);
 479     if (buf == NULL) {
 480         ret = -ENOMEM;
 481         goto ro_cleanup;
 482     }
 483
 484     for (offset = 0; offset < length; offset += n) {
 485         ret = bdrv_is_allocated(bs, offset, COMMIT_BUF_SIZE, &n);
 486         if (ret < 0) {
 487             goto ro_cleanup;
 488         }
 489         if (ret) {
 490             ret = blk_pread(src, offset, buf, n);
 491             if (ret < 0) {
 492                 goto ro_cleanup;
 493             }
 494
 495             ret = blk_pwrite(backing, offset, buf, n, 0);
 496             if (ret < 0) {
 497                 goto ro_cleanup;
 498             }
 499         }
 500     }
 501
 502     if (drv->bdrv_make_empty) {
 503         ret = drv->bdrv_make_empty(bs);
 504         if (ret < 0) {
 505             goto ro_cleanup;
 506         }
 507         blk_flush(src);
 508     }
 509
 510     /*
 511      * Make sure all data we wrote to the backing device is actually
 512      * stable on disk.
 513      */
 514     blk_flush(backing);
 515
 516     ret = 0;
 517 ro_cleanup:
 518     qemu_vfree(buf);
 519
 520     blk_unref(backing);
 521     if (backing_file_bs) {
 522         bdrv_set_backing_hd(bs, backing_file_bs, &error_abort);
 523     }
 524     bdrv_unref(commit_top_bs);
 525     blk_unref(src);
 526
 527     if (ro) {
 528         /* ignoring error return here */
 529         bdrv_reopen(bs->backing->bs, open_flags & ~BDRV_O_RDWR, NULL);
 530     }
 531
 532     return ret;
 533 }