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     int n = 0; /* sectors */
 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, 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 * BDRV_SECTOR_SIZE) {
 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 / BDRV_SECTOR_SIZE,
 187                                       COMMIT_BUFFER_SIZE / BDRV_SECTOR_SIZE,
 188                                       &n);
 189         copy = (ret == 1);
 190         trace_commit_one_iteration(s, offset, n * BDRV_SECTOR_SIZE, ret);
 191         if (copy) {
 192             ret = commit_populate(s->top, s->base, offset,
 193                                   n * BDRV_SECTOR_SIZE, buf);
 194             bytes_written += n * BDRV_SECTOR_SIZE;
 195         }
 196         if (ret < 0) {
 197             BlockErrorAction action =
 198                 block_job_error_action(&s->common, false, s->on_error, -ret);
 199             if (action == BLOCK_ERROR_ACTION_REPORT) {
 200                 goto out;
 201             } else {
 202                 n = 0;
 203                 continue;
 204             }
 205         }
 206         /* Publish progress */
 207         s->common.offset += n * BDRV_SECTOR_SIZE;
 208
 209         if (copy && s->common.speed) {
 210             delay_ns = ratelimit_calculate_delay(&s->limit,
 211                                                  n * BDRV_SECTOR_SIZE);
 212         }
 213     }
 214
 215     ret = 0;
 216
 217 out:
 218     qemu_vfree(buf);
 219
 220     data = g_malloc(sizeof(*data));
 221     data->ret = ret;
 222     block_job_defer_to_main_loop(&s->common, commit_complete, data);
 223 }
 224
 225 static void commit_set_speed(BlockJob *job, int64_t speed, Error **errp)
 226 {
 227     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
 228
 229     if (speed < 0) {
 230         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 231         return;
 232     }
 233     ratelimit_set_speed(&s->limit, speed, SLICE_TIME);
 234 }
 235
 236 static const BlockJobDriver commit_job_driver = {
 237     .instance_size = sizeof(CommitBlockJob),
 238     .job_type      = BLOCK_JOB_TYPE_COMMIT,
 239     .set_speed     = commit_set_speed,
 240     .start         = commit_run,
 241 };
 242
 243 static int coroutine_fn bdrv_commit_top_preadv(BlockDriverState *bs,
 244     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
 245 {
 246     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
 247 }
 248
 249 static int64_t coroutine_fn bdrv_commit_top_get_block_status(
 250     BlockDriverState *bs, int64_t sector_num, int nb_sectors, int *pnum,
 251     BlockDriverState **file)
 252 {
 253     *pnum = nb_sectors;
 254     *file = bs->backing->bs;
 255     return BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID |
 256            (sector_num << BDRV_SECTOR_BITS);
 257 }
 258
 259 static void bdrv_commit_top_refresh_filename(BlockDriverState *bs, QDict *opts)
 260 {
 261     bdrv_refresh_filename(bs->backing->bs);
 262     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
 263             bs->backing->bs->filename);
 264 }
 265
 266 static void bdrv_commit_top_close(BlockDriverState *bs)
 267 {
 268 }
 269
 270 static void bdrv_commit_top_child_perm(BlockDriverState *bs, BdrvChild *c,
 271                                        const BdrvChildRole *role,
 272                                        uint64_t perm, uint64_t shared,
 273                                        uint64_t *nperm, uint64_t *nshared)
 274 {
 275     *nperm = 0;
 276     *nshared = BLK_PERM_ALL;
 277 }
 278
 279 /* Dummy node that provides consistent read to its users without requiring it
 280  * from its backing file and that allows writes on the backing file chain. */
 281 static BlockDriver bdrv_commit_top = {
 282     .format_name                = "commit_top",
 283     .bdrv_co_preadv             = bdrv_commit_top_preadv,
 284     .bdrv_co_get_block_status   = bdrv_commit_top_get_block_status,
 285     .bdrv_refresh_filename      = bdrv_commit_top_refresh_filename,
 286     .bdrv_close                 = bdrv_commit_top_close,
 287     .bdrv_child_perm            = bdrv_commit_top_child_perm,
 288 };
 289
 290 void commit_start(const char *job_id, BlockDriverState *bs,
 291                   BlockDriverState *base, BlockDriverState *top, int64_t speed,
 292                   BlockdevOnError on_error, const char *backing_file_str,
 293                   const char *filter_node_name, Error **errp)
 294 {
 295     CommitBlockJob *s;
 296     BlockReopenQueue *reopen_queue = NULL;
 297     int orig_overlay_flags;
 298     int orig_base_flags;
 299     BlockDriverState *iter;
 300     BlockDriverState *overlay_bs;
 301     BlockDriverState *commit_top_bs = NULL;
 302     Error *local_err = NULL;
 303     int ret;
 304
 305     assert(top != bs);
 306     if (top == base) {
 307         error_setg(errp, "Invalid files for merge: top and base are the same");
 308         return;
 309     }
 310
 311     overlay_bs = bdrv_find_overlay(bs, top);
 312
 313     if (overlay_bs == NULL) {
 314         error_setg(errp, "Could not find overlay image for %s:", top->filename);
 315         return;
 316     }
 317
 318     s = block_job_create(job_id, &commit_job_driver, bs, 0, BLK_PERM_ALL,
 319                          speed, BLOCK_JOB_DEFAULT, NULL, NULL, errp);
 320     if (!s) {
 321         return;
 322     }
 323
 324     orig_base_flags    = bdrv_get_flags(base);
 325     orig_overlay_flags = bdrv_get_flags(overlay_bs);
 326
 327     /* convert base & overlay_bs to r/w, if necessary */
 328     if (!(orig_base_flags & BDRV_O_RDWR)) {
 329         reopen_queue = bdrv_reopen_queue(reopen_queue, base, NULL,
 330                                          orig_base_flags | BDRV_O_RDWR);
 331     }
 332     if (!(orig_overlay_flags & BDRV_O_RDWR)) {
 333         reopen_queue = bdrv_reopen_queue(reopen_queue, overlay_bs, NULL,
 334                                          orig_overlay_flags | BDRV_O_RDWR);
 335     }
 336     if (reopen_queue) {
 337         bdrv_reopen_multiple(bdrv_get_aio_context(bs), reopen_queue, &local_err);
 338         if (local_err != NULL) {
 339             error_propagate(errp, local_err);
 340             goto fail;
 341         }
 342     }
 343
 344     /* Insert commit_top block node above top, so we can block consistent read
 345      * on the backing chain below it */
 346     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, filter_node_name, 0,
 347                                          errp);
 348     if (commit_top_bs == NULL) {
 349         goto fail;
 350     }
 351     commit_top_bs->total_sectors = top->total_sectors;
 352     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(top));
 353
 354     bdrv_set_backing_hd(commit_top_bs, top, &local_err);
 355     if (local_err) {
 356         bdrv_unref(commit_top_bs);
 357         commit_top_bs = NULL;
 358         error_propagate(errp, local_err);
 359         goto fail;
 360     }
 361     bdrv_set_backing_hd(overlay_bs, commit_top_bs, &local_err);
 362     if (local_err) {
 363         bdrv_unref(commit_top_bs);
 364         commit_top_bs = NULL;
 365         error_propagate(errp, local_err);
 366         goto fail;
 367     }
 368
 369     s->commit_top_bs = commit_top_bs;
 370     bdrv_unref(commit_top_bs);
 371
 372     /* Block all nodes between top and base, because they will
 373      * disappear from the chain after this operation. */
 374     assert(bdrv_chain_contains(top, base));
 375     for (iter = top; iter != base; iter = backing_bs(iter)) {
 376         /* XXX BLK_PERM_WRITE needs to be allowed so we don't block ourselves
 377          * at s->base (if writes are blocked for a node, they are also blocked
 378          * for its backing file). The other options would be a second filter
 379          * driver above s->base. */
 380         ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
 381                                  BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
 382                                  errp);
 383         if (ret < 0) {
 384             goto fail;
 385         }
 386     }
 387
 388     ret = block_job_add_bdrv(&s->common, "base", base, 0, BLK_PERM_ALL, errp);
 389     if (ret < 0) {
 390         goto fail;
 391     }
 392
 393     /* overlay_bs must be blocked because it needs to be modified to
 394      * update the backing image string. */
 395     ret = block_job_add_bdrv(&s->common, "overlay of top", overlay_bs,
 396                              BLK_PERM_GRAPH_MOD, BLK_PERM_ALL, errp);
 397     if (ret < 0) {
 398         goto fail;
 399     }
 400
 401     s->base = blk_new(BLK_PERM_CONSISTENT_READ
 402                       | BLK_PERM_WRITE
 403                       | BLK_PERM_RESIZE,
 404                       BLK_PERM_CONSISTENT_READ
 405                       | BLK_PERM_GRAPH_MOD
 406                       | BLK_PERM_WRITE_UNCHANGED);
 407     ret = blk_insert_bs(s->base, base, errp);
 408     if (ret < 0) {
 409         goto fail;
 410     }
 411
 412     /* Required permissions are already taken with block_job_add_bdrv() */
 413     s->top = blk_new(0, BLK_PERM_ALL);
 414     ret = blk_insert_bs(s->top, top, errp);
 415     if (ret < 0) {
 416         goto fail;
 417     }
 418
 419     s->active = bs;
 420
 421     s->base_flags          = orig_base_flags;
 422     s->orig_overlay_flags  = orig_overlay_flags;
 423
 424     s->backing_file_str = g_strdup(backing_file_str);
 425
 426     s->on_error = on_error;
 427
 428     trace_commit_start(bs, base, top, s);
 429     block_job_start(&s->common);
 430     return;
 431
 432 fail:
 433     if (s->base) {
 434         blk_unref(s->base);
 435     }
 436     if (s->top) {
 437         blk_unref(s->top);
 438     }
 439     if (commit_top_bs) {
 440         bdrv_set_backing_hd(overlay_bs, top, &error_abort);
 441     }
 442     block_job_early_fail(&s->common);
 443 }
 444
 445
 446 #define COMMIT_BUF_SIZE (2048 * BDRV_SECTOR_SIZE)
 447
 448 /* commit COW file into the raw image */
 449 int bdrv_commit(BlockDriverState *bs)
 450 {
 451     BlockBackend *src, *backing;
 452     BlockDriverState *backing_file_bs = NULL;
 453     BlockDriverState *commit_top_bs = NULL;
 454     BlockDriver *drv = bs->drv;
 455     int64_t offset, length, backing_length;
 456     int ro, open_flags;
 457     int64_t n;
 458     int ret = 0;
 459     uint8_t *buf = NULL;
 460     Error *local_err = NULL;
 461
 462     if (!drv)
 463         return -ENOMEDIUM;
 464
 465     if (!bs->backing) {
 466         return -ENOTSUP;
 467     }
 468
 469     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
 470         bdrv_op_is_blocked(bs->backing->bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL)) {
 471         return -EBUSY;
 472     }
 473
 474     ro = bs->backing->bs->read_only;
 475     open_flags =  bs->backing->bs->open_flags;
 476
 477     if (ro) {
 478         if (bdrv_reopen(bs->backing->bs, open_flags | BDRV_O_RDWR, NULL)) {
 479             return -EACCES;
 480         }
 481     }
 482
 483     src = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 484     backing = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
 485
 486     ret = blk_insert_bs(src, bs, &local_err);
 487     if (ret < 0) {
 488         error_report_err(local_err);
 489         goto ro_cleanup;
 490     }
 491
 492     /* Insert commit_top block node above backing, so we can write to it */
 493     backing_file_bs = backing_bs(bs);
 494
 495     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, NULL, BDRV_O_RDWR,
 496                                          &local_err);
 497     if (commit_top_bs == NULL) {
 498         error_report_err(local_err);
 499         goto ro_cleanup;
 500     }
 501     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(backing_file_bs));
 502
 503     bdrv_set_backing_hd(commit_top_bs, backing_file_bs, &error_abort);
 504     bdrv_set_backing_hd(bs, commit_top_bs, &error_abort);
 505
 506     ret = blk_insert_bs(backing, backing_file_bs, &local_err);
 507     if (ret < 0) {
 508         error_report_err(local_err);
 509         goto ro_cleanup;
 510     }
 511
 512     length = blk_getlength(src);
 513     if (length < 0) {
 514         ret = length;
 515         goto ro_cleanup;
 516     }
 517
 518     backing_length = blk_getlength(backing);
 519     if (backing_length < 0) {
 520         ret = backing_length;
 521         goto ro_cleanup;
 522     }
 523
 524     /* If our top snapshot is larger than the backing file image,
 525      * grow the backing file image if possible.  If not possible,
 526      * we must return an error */
 527     if (length > backing_length) {
 528         ret = blk_truncate(backing, length, &local_err);
 529         if (ret < 0) {
 530             error_report_err(local_err);
 531             goto ro_cleanup;
 532         }
 533     }
 534
 535     /* blk_try_blockalign() for src will choose an alignment that works for
 536      * backing as well, so no need to compare the alignment manually. */
 537     buf = blk_try_blockalign(src, COMMIT_BUF_SIZE);
 538     if (buf == NULL) {
 539         ret = -ENOMEM;
 540         goto ro_cleanup;
 541     }
 542
 543     for (offset = 0; offset < length; offset += n) {
 544         ret = bdrv_is_allocated(bs, offset, COMMIT_BUF_SIZE, &n);
 545         if (ret < 0) {
 546             goto ro_cleanup;
 547         }
 548         if (ret) {
 549             ret = blk_pread(src, offset, buf, n);
 550             if (ret < 0) {
 551                 goto ro_cleanup;
 552             }
 553
 554             ret = blk_pwrite(backing, offset, buf, n, 0);
 555             if (ret < 0) {
 556                 goto ro_cleanup;
 557             }
 558         }
 559     }
 560
 561     if (drv->bdrv_make_empty) {
 562         ret = drv->bdrv_make_empty(bs);
 563         if (ret < 0) {
 564             goto ro_cleanup;
 565         }
 566         blk_flush(src);
 567     }
 568
 569     /*
 570      * Make sure all data we wrote to the backing device is actually
 571      * stable on disk.
 572      */
 573     blk_flush(backing);
 574
 575     ret = 0;
 576 ro_cleanup:
 577     qemu_vfree(buf);
 578
 579     blk_unref(backing);
 580     if (backing_file_bs) {
 581         bdrv_set_backing_hd(bs, backing_file_bs, &error_abort);
 582     }
 583     bdrv_unref(commit_top_bs);
 584     blk_unref(src);
 585
 586     if (ro) {
 587         /* ignoring error return here */
 588         bdrv_reopen(bs->backing->bs, open_flags & ~BDRV_O_RDWR, NULL);
 589     }
 590
 591     return ret;
 592 }