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(Job *job, void *opaque)
  76 {
  77     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
  78     BlockJob *bjob = &s->common;
  79     CommitCompleteData *data = opaque;
  80     BlockDriverState *top = blk_bs(s->top);
  81     BlockDriverState *base = blk_bs(s->base);
  82     BlockDriverState *commit_top_bs = s->commit_top_bs;
  83     int ret = data->ret;
  84     bool remove_commit_top_bs = false;
  85
  86     /* Make sure commit_top_bs and top stay around until bdrv_replace_node() */
  87     bdrv_ref(top);
  88     bdrv_ref(commit_top_bs);
  89
  90     /* Remove base node parent that still uses BLK_PERM_WRITE/RESIZE before
  91      * the normal backing chain can be restored. */
  92     blk_unref(s->base);
  93
  94     if (!job_is_cancelled(job) && ret == 0) {
  95         /* success */
  96         ret = bdrv_drop_intermediate(s->commit_top_bs, base,
  97                                      s->backing_file_str);
  98     } else {
  99         /* XXX Can (or should) we somehow keep 'consistent read' blocked even
 100          * after the failed/cancelled commit job is gone? If we already wrote
 101          * something to base, the intermediate images aren't valid any more. */
 102         remove_commit_top_bs = true;
 103     }
 104
 105     /* restore base open flags here if appropriate (e.g., change the base back
 106      * to r/o). These reopens do not need to be atomic, since we won't abort
 107      * even on failure here */
 108     if (s->base_flags != bdrv_get_flags(base)) {
 109         bdrv_reopen(base, s->base_flags, NULL);
 110     }
 111     g_free(s->backing_file_str);
 112     blk_unref(s->top);
 113
 114     /* If there is more than one reference to the job (e.g. if called from
 115      * job_finish_sync()), job_completed() won't free it and therefore the
 116      * blockers on the intermediate nodes remain. This would cause
 117      * bdrv_set_backing_hd() to fail. */
 118     block_job_remove_all_bdrv(bjob);
 119
 120     job_completed(job, ret, NULL);
 121     g_free(data);
 122
 123     /* If bdrv_drop_intermediate() didn't already do that, remove the commit
 124      * filter driver from the backing chain. Do this as the final step so that
 125      * the 'consistent read' permission can be granted.  */
 126     if (remove_commit_top_bs) {
 127         bdrv_child_try_set_perm(commit_top_bs->backing, 0, BLK_PERM_ALL,
 128                                 &error_abort);
 129         bdrv_replace_node(commit_top_bs, backing_bs(commit_top_bs),
 130                           &error_abort);
 131     }
 132
 133     bdrv_unref(commit_top_bs);
 134     bdrv_unref(top);
 135 }
 136
 137 static void coroutine_fn commit_run(void *opaque)
 138 {
 139     CommitBlockJob *s = opaque;
 140     CommitCompleteData *data;
 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     data = g_malloc(sizeof(*data));
 214     data->ret = ret;
 215     job_defer_to_main_loop(&s->common.job, commit_complete, data);
 216 }
 217
 218 static const BlockJobDriver commit_job_driver = {
 219     .job_driver = {
 220         .instance_size = sizeof(CommitBlockJob),
 221         .job_type      = JOB_TYPE_COMMIT,
 222         .free          = block_job_free,
 223         .user_resume   = block_job_user_resume,
 224         .drain         = block_job_drain,
 225         .start         = commit_run,
 226     },
 227 };
 228
 229 static int coroutine_fn bdrv_commit_top_preadv(BlockDriverState *bs,
 230     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
 231 {
 232     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
 233 }
 234
 235 static void bdrv_commit_top_refresh_filename(BlockDriverState *bs, QDict *opts)
 236 {
 237     bdrv_refresh_filename(bs->backing->bs);
 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, int64_t speed,
 264                   BlockdevOnError on_error, const char *backing_file_str,
 265                   const char *filter_node_name, Error **errp)
 266 {
 267     CommitBlockJob *s;
 268     int orig_base_flags;
 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, JOB_DEFAULT, NULL, NULL, errp);
 282     if (!s) {
 283         return;
 284     }
 285
 286     /* convert base to r/w, if necessary */
 287     orig_base_flags = bdrv_get_flags(base);
 288     if (!(orig_base_flags & BDRV_O_RDWR)) {
 289         bdrv_reopen(base, orig_base_flags | BDRV_O_RDWR, &local_err);
 290         if (local_err != NULL) {
 291             error_propagate(errp, local_err);
 292             goto fail;
 293         }
 294     }
 295
 296     /* Insert commit_top block node above top, so we can block consistent read
 297      * on the backing chain below it */
 298     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, filter_node_name, 0,
 299                                          errp);
 300     if (commit_top_bs == NULL) {
 301         goto fail;
 302     }
 303     if (!filter_node_name) {
 304         commit_top_bs->implicit = true;
 305     }
 306     commit_top_bs->total_sectors = top->total_sectors;
 307     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(top));
 308
 309     bdrv_set_backing_hd(commit_top_bs, top, &local_err);
 310     if (local_err) {
 311         bdrv_unref(commit_top_bs);
 312         commit_top_bs = NULL;
 313         error_propagate(errp, local_err);
 314         goto fail;
 315     }
 316     bdrv_replace_node(top, commit_top_bs, &local_err);
 317     if (local_err) {
 318         bdrv_unref(commit_top_bs);
 319         commit_top_bs = NULL;
 320         error_propagate(errp, local_err);
 321         goto fail;
 322     }
 323
 324     s->commit_top_bs = commit_top_bs;
 325     bdrv_unref(commit_top_bs);
 326
 327     /* Block all nodes between top and base, because they will
 328      * disappear from the chain after this operation. */
 329     assert(bdrv_chain_contains(top, base));
 330     for (iter = top; iter != base; iter = backing_bs(iter)) {
 331         /* XXX BLK_PERM_WRITE needs to be allowed so we don't block ourselves
 332          * at s->base (if writes are blocked for a node, they are also blocked
 333          * for its backing file). The other options would be a second filter
 334          * driver above s->base. */
 335         ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
 336                                  BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
 337                                  errp);
 338         if (ret < 0) {
 339             goto fail;
 340         }
 341     }
 342
 343     ret = block_job_add_bdrv(&s->common, "base", base, 0, BLK_PERM_ALL, errp);
 344     if (ret < 0) {
 345         goto fail;
 346     }
 347
 348     s->base = blk_new(BLK_PERM_CONSISTENT_READ
 349                       | BLK_PERM_WRITE
 350                       | BLK_PERM_RESIZE,
 351                       BLK_PERM_CONSISTENT_READ
 352                       | BLK_PERM_GRAPH_MOD
 353                       | BLK_PERM_WRITE_UNCHANGED);
 354     ret = blk_insert_bs(s->base, base, errp);
 355     if (ret < 0) {
 356         goto fail;
 357     }
 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 }