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