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 #define SLICE_TIME 100000000ULL /* ns */
  35
  36 typedef struct CommitBlockJob {
  37     BlockJob common;
  38     RateLimit limit;
  39     BlockDriverState *active;
  40     BlockDriverState *commit_top_bs;
  41     BlockBackend *top;
  42     BlockBackend *base;
  43     BlockdevOnError on_error;
  44     int base_flags;
  45     int orig_overlay_flags;
  46     char *backing_file_str;
  47 } CommitBlockJob;
  48
  49 static int coroutine_fn commit_populate(BlockBackend *bs, BlockBackend *base,
  50                                         int64_t offset, uint64_t bytes,
  51                                         void *buf)
  52 {
  53     int ret = 0;
  54     QEMUIOVector qiov;
  55     struct iovec iov = {
  56         .iov_base = buf,
  57         .iov_len = bytes,
  58     };
  59
  60     assert(bytes < SIZE_MAX);
  61     qemu_iovec_init_external(&qiov, &iov, 1);
  62
  63     ret = blk_co_preadv(bs, offset, qiov.size, &qiov, 0);
  64     if (ret < 0) {
  65         return ret;
  66     }
  67
  68     ret = blk_co_pwritev(base, offset, qiov.size, &qiov, 0);
  69     if (ret < 0) {
  70         return ret;
  71     }
  72
  73     return 0;
  74 }
  75
  76 typedef struct {
  77     int ret;
  78 } CommitCompleteData;
  79
  80 static void commit_complete(BlockJob *job, void *opaque)
  81 {
  82     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
  83     CommitCompleteData *data = opaque;
  84     BlockDriverState *active = s->active;
  85     BlockDriverState *top = blk_bs(s->top);
  86     BlockDriverState *base = blk_bs(s->base);
  87     BlockDriverState *overlay_bs = bdrv_find_overlay(active, s->commit_top_bs);
  88     int ret = data->ret;
  89     bool remove_commit_top_bs = false;
  90
  91     /* Make sure overlay_bs and top stay around until bdrv_set_backing_hd() */
  92     bdrv_ref(top);
  93     if (overlay_bs) {
  94         bdrv_ref(overlay_bs);
  95     }
  96
  97     /* Remove base node parent that still uses BLK_PERM_WRITE/RESIZE before
  98      * the normal backing chain can be restored. */
  99     blk_unref(s->base);
 100
 101     if (!block_job_is_cancelled(&s->common) && ret == 0) {
 102         /* success */
 103         ret = bdrv_drop_intermediate(active, s->commit_top_bs, base,
 104                                      s->backing_file_str);
 105     } else if (overlay_bs) {
 106         /* XXX Can (or should) we somehow keep 'consistent read' blocked even
 107          * after the failed/cancelled commit job is gone? If we already wrote
 108          * something to base, the intermediate images aren't valid any more. */
 109         remove_commit_top_bs = true;
 110     }
 111
 112     /* restore base open flags here if appropriate (e.g., change the base back
 113      * to r/o). These reopens do not need to be atomic, since we won't abort
 114      * even on failure here */
 115     if (s->base_flags != bdrv_get_flags(base)) {
 116         bdrv_reopen(base, s->base_flags, NULL);
 117     }
 118     if (overlay_bs && s->orig_overlay_flags != bdrv_get_flags(overlay_bs)) {
 119         bdrv_reopen(overlay_bs, s->orig_overlay_flags, NULL);
 120     }
 121     g_free(s->backing_file_str);
 122     blk_unref(s->top);
 123
 124     /* If there is more than one reference to the job (e.g. if called from
 125      * block_job_finish_sync()), block_job_completed() won't free it and
 126      * therefore the blockers on the intermediate nodes remain. This would
 127      * cause bdrv_set_backing_hd() to fail. */
 128     block_job_remove_all_bdrv(job);
 129
 130     block_job_completed(&s->common, ret);
 131     g_free(data);
 132
 133     /* If bdrv_drop_intermediate() didn't already do that, remove the commit
 134      * filter driver from the backing chain. Do this as the final step so that
 135      * the 'consistent read' permission can be granted.  */
 136     if (remove_commit_top_bs) {
 137         bdrv_set_backing_hd(overlay_bs, top, &error_abort);
 138     }
 139
 140     bdrv_unref(overlay_bs);
 141     bdrv_unref(top);
 142 }
 143
 144 static void coroutine_fn commit_run(void *opaque)
 145 {
 146     CommitBlockJob *s = opaque;
 147     CommitCompleteData *data;
 148     int64_t offset;
 149     uint64_t delay_ns = 0;
 150     int ret = 0;
 151     int64_t n = 0; /* bytes */
 152     void *buf = NULL;
 153     int bytes_written = 0;
 154     int64_t base_len;
 155
 156     ret = s->common.len = blk_getlength(s->top);
 157
 158     if (s->common.len < 0) {
 159         goto out;
 160     }
 161
 162     ret = base_len = blk_getlength(s->base);
 163     if (base_len < 0) {
 164         goto out;
 165     }
 166
 167     if (base_len < s->common.len) {
 168         ret = blk_truncate(s->base, s->common.len, PREALLOC_MODE_OFF, NULL);
 169         if (ret) {
 170             goto out;
 171         }
 172     }
 173
 174     buf = blk_blockalign(s->top, COMMIT_BUFFER_SIZE);
 175
 176     for (offset = 0; offset < s->common.len; offset += n) {
 177         bool copy;
 178
 179         /* Note that even when no rate limit is applied we need to yield
 180          * with no pending I/O here so that bdrv_drain_all() returns.
 181          */
 182         block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 183         if (block_job_is_cancelled(&s->common)) {
 184             break;
 185         }
 186         /* Copy if allocated above the base */
 187         ret = bdrv_is_allocated_above(blk_bs(s->top), blk_bs(s->base),
 188                                       offset, COMMIT_BUFFER_SIZE, &n);
 189         copy = (ret == 1);
 190         trace_commit_one_iteration(s, offset, n, ret);
 191         if (copy) {
 192             ret = commit_populate(s->top, s->base, offset, n, buf);
 193             bytes_written += n;
 194         }
 195         if (ret < 0) {
 196             BlockErrorAction action =
 197                 block_job_error_action(&s->common, false, s->on_error, -ret);
 198             if (action == BLOCK_ERROR_ACTION_REPORT) {
 199                 goto out;
 200             } else {
 201                 n = 0;
 202                 continue;
 203             }
 204         }
 205         /* Publish progress */
 206         s->common.offset += n;
 207
 208         if (copy && s->common.speed) {
 209             delay_ns = ratelimit_calculate_delay(&s->limit, n);
 210         }
 211     }
 212
 213     ret = 0;
 214
 215 out:
 216     qemu_vfree(buf);
 217
 218     data = g_malloc(sizeof(*data));
 219     data->ret = ret;
 220     block_job_defer_to_main_loop(&s->common, commit_complete, data);
 221 }
 222
 223 static void commit_set_speed(BlockJob *job, int64_t speed, Error **errp)
 224 {
 225     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
 226
 227     if (speed < 0) {
 228         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 229         return;
 230     }
 231     ratelimit_set_speed(&s->limit, speed, SLICE_TIME);
 232 }
 233
 234 static const BlockJobDriver commit_job_driver = {
 235     .instance_size = sizeof(CommitBlockJob),
 236     .job_type      = BLOCK_JOB_TYPE_COMMIT,
 237     .set_speed     = commit_set_speed,
 238     .start         = commit_run,
 239 };
 240
 241 static int coroutine_fn bdrv_commit_top_preadv(BlockDriverState *bs,
 242     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
 243 {
 244     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
 245 }
 246
 247 static int64_t coroutine_fn bdrv_commit_top_get_block_status(
 248     BlockDriverState *bs, int64_t sector_num, int nb_sectors, int *pnum,
 249     BlockDriverState **file)
 250 {
 251     *pnum = nb_sectors;
 252     *file = bs->backing->bs;
 253     return BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID |
 254            (sector_num << BDRV_SECTOR_BITS);
 255 }
 256
 257 static void bdrv_commit_top_refresh_filename(BlockDriverState *bs, QDict *opts)
 258 {
 259     bdrv_refresh_filename(bs->backing->bs);
 260     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
 261             bs->backing->bs->filename);
 262 }
 263
 264 static void bdrv_commit_top_close(BlockDriverState *bs)
 265 {
 266 }
 267
 268 static void bdrv_commit_top_child_perm(BlockDriverState *bs, BdrvChild *c,
 269                                        const BdrvChildRole *role,
 270                                        uint64_t perm, uint64_t shared,
 271                                        uint64_t *nperm, uint64_t *nshared)
 272 {
 273     *nperm = 0;
 274     *nshared = BLK_PERM_ALL;
 275 }
 276
 277 /* Dummy node that provides consistent read to its users without requiring it
 278  * from its backing file and that allows writes on the backing file chain. */
 279 static BlockDriver bdrv_commit_top = {
 280     .format_name                = "commit_top",
 281     .bdrv_co_preadv             = bdrv_commit_top_preadv,
 282     .bdrv_co_get_block_status   = bdrv_commit_top_get_block_status,
 283     .bdrv_refresh_filename      = bdrv_commit_top_refresh_filename,
 284     .bdrv_close                 = bdrv_commit_top_close,
 285     .bdrv_child_perm            = bdrv_commit_top_child_perm,
 286 };
 287
 288 void commit_start(const char *job_id, BlockDriverState *bs,
 289                   BlockDriverState *base, BlockDriverState *top, int64_t speed,
 290                   BlockdevOnError on_error, const char *backing_file_str,
 291                   const char *filter_node_name, Error **errp)
 292 {
 293     CommitBlockJob *s;
 294     BlockReopenQueue *reopen_queue = NULL;
 295     int orig_overlay_flags;
 296     int orig_base_flags;
 297     BlockDriverState *iter;
 298     BlockDriverState *overlay_bs;
 299     BlockDriverState *commit_top_bs = NULL;
 300     Error *local_err = NULL;
 301     int ret;
 302
 303     assert(top != bs);
 304     if (top == base) {
 305         error_setg(errp, "Invalid files for merge: top and base are the same");
 306         return;
 307     }
 308
 309     overlay_bs = bdrv_find_overlay(bs, top);
 310
 311     if (overlay_bs == NULL) {
 312         error_setg(errp, "Could not find overlay image for %s:", top->filename);
 313         return;
 314     }
 315
 316     s = block_job_create(job_id, &commit_job_driver, bs, 0, BLK_PERM_ALL,
 317                          speed, BLOCK_JOB_DEFAULT, NULL, NULL, errp);
 318     if (!s) {
 319         return;
 320     }
 321
 322     orig_base_flags    = bdrv_get_flags(base);
 323     orig_overlay_flags = bdrv_get_flags(overlay_bs);
 324
 325     /* convert base & overlay_bs to r/w, if necessary */
 326     if (!(orig_base_flags & BDRV_O_RDWR)) {
 327         reopen_queue = bdrv_reopen_queue(reopen_queue, base, NULL,
 328                                          orig_base_flags | BDRV_O_RDWR);
 329     }
 330     if (!(orig_overlay_flags & BDRV_O_RDWR)) {
 331         reopen_queue = bdrv_reopen_queue(reopen_queue, overlay_bs, NULL,
 332                                          orig_overlay_flags | BDRV_O_RDWR);
 333     }
 334     if (reopen_queue) {
 335         bdrv_reopen_multiple(bdrv_get_aio_context(bs), reopen_queue, &local_err);
 336         if (local_err != NULL) {
 337             error_propagate(errp, local_err);
 338             goto fail;
 339         }
 340     }
 341
 342     /* Insert commit_top block node above top, so we can block consistent read
 343      * on the backing chain below it */
 344     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, filter_node_name, 0,
 345                                          errp);
 346     if (commit_top_bs == NULL) {
 347         goto fail;
 348     }
 349     commit_top_bs->total_sectors = top->total_sectors;
 350     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(top));
 351
 352     bdrv_set_backing_hd(commit_top_bs, top, &local_err);
 353     if (local_err) {
 354         bdrv_unref(commit_top_bs);
 355         commit_top_bs = NULL;
 356         error_propagate(errp, local_err);
 357         goto fail;
 358     }
 359     bdrv_set_backing_hd(overlay_bs, commit_top_bs, &local_err);
 360     if (local_err) {
 361         bdrv_unref(commit_top_bs);
 362         commit_top_bs = NULL;
 363         error_propagate(errp, local_err);
 364         goto fail;
 365     }
 366
 367     s->commit_top_bs = commit_top_bs;
 368     bdrv_unref(commit_top_bs);
 369
 370     /* Block all nodes between top and base, because they will
 371      * disappear from the chain after this operation. */
 372     assert(bdrv_chain_contains(top, base));
 373     for (iter = top; iter != base; iter = backing_bs(iter)) {
 374         /* XXX BLK_PERM_WRITE needs to be allowed so we don't block ourselves
 375          * at s->base (if writes are blocked for a node, they are also blocked
 376          * for its backing file). The other options would be a second filter
 377          * driver above s->base. */
 378         ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
 379                                  BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
 380                                  errp);
 381         if (ret < 0) {
 382             goto fail;
 383         }
 384     }
 385
 386     ret = block_job_add_bdrv(&s->common, "base", base, 0, BLK_PERM_ALL, errp);
 387     if (ret < 0) {
 388         goto fail;
 389     }
 390
 391     /* overlay_bs must be blocked because it needs to be modified to
 392      * update the backing image string. */
 393     ret = block_job_add_bdrv(&s->common, "overlay of top", overlay_bs,
 394                              BLK_PERM_GRAPH_MOD, BLK_PERM_ALL, errp);
 395     if (ret < 0) {
 396         goto fail;
 397     }
 398
 399     s->base = blk_new(BLK_PERM_CONSISTENT_READ
 400                       | BLK_PERM_WRITE
 401                       | BLK_PERM_RESIZE,
 402                       BLK_PERM_CONSISTENT_READ
 403                       | BLK_PERM_GRAPH_MOD
 404                       | BLK_PERM_WRITE_UNCHANGED);
 405     ret = blk_insert_bs(s->base, base, errp);
 406     if (ret < 0) {
 407         goto fail;
 408     }
 409
 410     /* Required permissions are already taken with block_job_add_bdrv() */
 411     s->top = blk_new(0, BLK_PERM_ALL);
 412     ret = blk_insert_bs(s->top, top, errp);
 413     if (ret < 0) {
 414         goto fail;
 415     }
 416
 417     s->active = bs;
 418
 419     s->base_flags          = orig_base_flags;
 420     s->orig_overlay_flags  = orig_overlay_flags;
 421
 422     s->backing_file_str = g_strdup(backing_file_str);
 423
 424     s->on_error = on_error;
 425
 426     trace_commit_start(bs, base, top, s);
 427     block_job_start(&s->common);
 428     return;
 429
 430 fail:
 431     if (s->base) {
 432         blk_unref(s->base);
 433     }
 434     if (s->top) {
 435         blk_unref(s->top);
 436     }
 437     if (commit_top_bs) {
 438         bdrv_set_backing_hd(overlay_bs, top, &error_abort);
 439     }
 440     block_job_early_fail(&s->common);
 441 }
 442
 443
 444 #define COMMIT_BUF_SIZE (2048 * BDRV_SECTOR_SIZE)
 445
 446 /* commit COW file into the raw image */
 447 int bdrv_commit(BlockDriverState *bs)
 448 {
 449     BlockBackend *src, *backing;
 450     BlockDriverState *backing_file_bs = NULL;
 451     BlockDriverState *commit_top_bs = NULL;
 452     BlockDriver *drv = bs->drv;
 453     int64_t offset, length, backing_length;
 454     int ro, open_flags;
 455     int64_t n;
 456     int ret = 0;
 457     uint8_t *buf = NULL;
 458     Error *local_err = NULL;
 459
 460     if (!drv)
 461         return -ENOMEDIUM;
 462
 463     if (!bs->backing) {
 464         return -ENOTSUP;
 465     }
 466
 467     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
 468         bdrv_op_is_blocked(bs->backing->bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL)) {
 469         return -EBUSY;
 470     }
 471
 472     ro = bs->backing->bs->read_only;
 473     open_flags =  bs->backing->bs->open_flags;
 474
 475     if (ro) {
 476         if (bdrv_reopen(bs->backing->bs, open_flags | BDRV_O_RDWR, NULL)) {
 477             return -EACCES;
 478         }
 479     }
 480
 481     src = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 482     backing = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
 483
 484     ret = blk_insert_bs(src, bs, &local_err);
 485     if (ret < 0) {
 486         error_report_err(local_err);
 487         goto ro_cleanup;
 488     }
 489
 490     /* Insert commit_top block node above backing, so we can write to it */
 491     backing_file_bs = backing_bs(bs);
 492
 493     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, NULL, BDRV_O_RDWR,
 494                                          &local_err);
 495     if (commit_top_bs == NULL) {
 496         error_report_err(local_err);
 497         goto ro_cleanup;
 498     }
 499     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(backing_file_bs));
 500
 501     bdrv_set_backing_hd(commit_top_bs, backing_file_bs, &error_abort);
 502     bdrv_set_backing_hd(bs, commit_top_bs, &error_abort);
 503
 504     ret = blk_insert_bs(backing, backing_file_bs, &local_err);
 505     if (ret < 0) {
 506         error_report_err(local_err);
 507         goto ro_cleanup;
 508     }
 509
 510     length = blk_getlength(src);
 511     if (length < 0) {
 512         ret = length;
 513         goto ro_cleanup;
 514     }
 515
 516     backing_length = blk_getlength(backing);
 517     if (backing_length < 0) {
 518         ret = backing_length;
 519         goto ro_cleanup;
 520     }
 521
 522     /* If our top snapshot is larger than the backing file image,
 523      * grow the backing file image if possible.  If not possible,
 524      * we must return an error */
 525     if (length > backing_length) {
 526         ret = blk_truncate(backing, length, PREALLOC_MODE_OFF, &local_err);
 527         if (ret < 0) {
 528             error_report_err(local_err);
 529             goto ro_cleanup;
 530         }
 531     }
 532
 533     /* blk_try_blockalign() for src will choose an alignment that works for
 534      * backing as well, so no need to compare the alignment manually. */
 535     buf = blk_try_blockalign(src, COMMIT_BUF_SIZE);
 536     if (buf == NULL) {
 537         ret = -ENOMEM;
 538         goto ro_cleanup;
 539     }
 540
 541     for (offset = 0; offset < length; offset += n) {
 542         ret = bdrv_is_allocated(bs, offset, COMMIT_BUF_SIZE, &n);
 543         if (ret < 0) {
 544             goto ro_cleanup;
 545         }
 546         if (ret) {
 547             ret = blk_pread(src, offset, buf, n);
 548             if (ret < 0) {
 549                 goto ro_cleanup;
 550             }
 551
 552             ret = blk_pwrite(backing, offset, buf, n, 0);
 553             if (ret < 0) {
 554                 goto ro_cleanup;
 555             }
 556         }
 557     }
 558
 559     if (drv->bdrv_make_empty) {
 560         ret = drv->bdrv_make_empty(bs);
 561         if (ret < 0) {
 562             goto ro_cleanup;
 563         }
 564         blk_flush(src);
 565     }
 566
 567     /*
 568      * Make sure all data we wrote to the backing device is actually
 569      * stable on disk.
 570      */
 571     blk_flush(backing);
 572
 573     ret = 0;
 574 ro_cleanup:
 575     qemu_vfree(buf);
 576
 577     blk_unref(backing);
 578     if (backing_file_bs) {
 579         bdrv_set_backing_hd(bs, backing_file_bs, &error_abort);
 580     }
 581     bdrv_unref(commit_top_bs);
 582     blk_unref(src);
 583
 584     if (ro) {
 585         /* ignoring error return here */
 586         bdrv_reopen(bs->backing->bs, open_flags & ~BDRV_O_RDWR, NULL);
 587     }
 588
 589     return ret;
 590 }