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 sector_num, int nb_sectors,
  51                                         void *buf)
  52 {
  53     int ret = 0;
  54     QEMUIOVector qiov;
  55     struct iovec iov = {
  56         .iov_base = buf,
  57         .iov_len = nb_sectors * BDRV_SECTOR_SIZE,
  58     };
  59
  60     qemu_iovec_init_external(&qiov, &iov, 1);
  61
  62     ret = blk_co_preadv(bs, sector_num * BDRV_SECTOR_SIZE,
  63                         qiov.size, &qiov, 0);
  64     if (ret < 0) {
  65         return ret;
  66     }
  67
  68     ret = blk_co_pwritev(base, sector_num * BDRV_SECTOR_SIZE,
  69                          qiov.size, &qiov, 0);
  70     if (ret < 0) {
  71         return ret;
  72     }
  73
  74     return 0;
  75 }
  76
  77 typedef struct {
  78     int ret;
  79 } CommitCompleteData;
  80
  81 static void commit_complete(BlockJob *job, void *opaque)
  82 {
  83     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
  84     CommitCompleteData *data = opaque;
  85     BlockDriverState *active = s->active;
  86     BlockDriverState *top = blk_bs(s->top);
  87     BlockDriverState *base = blk_bs(s->base);
  88     BlockDriverState *overlay_bs = bdrv_find_overlay(active, s->commit_top_bs);
  89     int ret = data->ret;
  90     bool remove_commit_top_bs = false;
  91
  92     /* Make sure overlay_bs and top stay around until bdrv_set_backing_hd() */
  93     bdrv_ref(top);
  94     if (overlay_bs) {
  95         bdrv_ref(overlay_bs);
  96     }
  97
  98     /* Remove base node parent that still uses BLK_PERM_WRITE/RESIZE before
  99      * the normal backing chain can be restored. */
 100     blk_unref(s->base);
 101
 102     if (!block_job_is_cancelled(&s->common) && ret == 0) {
 103         /* success */
 104         ret = bdrv_drop_intermediate(active, s->commit_top_bs, base,
 105                                      s->backing_file_str);
 106     } else if (overlay_bs) {
 107         /* XXX Can (or should) we somehow keep 'consistent read' blocked even
 108          * after the failed/cancelled commit job is gone? If we already wrote
 109          * something to base, the intermediate images aren't valid any more. */
 110         remove_commit_top_bs = true;
 111     }
 112
 113     /* restore base open flags here if appropriate (e.g., change the base back
 114      * to r/o). These reopens do not need to be atomic, since we won't abort
 115      * even on failure here */
 116     if (s->base_flags != bdrv_get_flags(base)) {
 117         bdrv_reopen(base, s->base_flags, NULL);
 118     }
 119     if (overlay_bs && s->orig_overlay_flags != bdrv_get_flags(overlay_bs)) {
 120         bdrv_reopen(overlay_bs, s->orig_overlay_flags, NULL);
 121     }
 122     g_free(s->backing_file_str);
 123     blk_unref(s->top);
 124
 125     /* If there is more than one reference to the job (e.g. if called from
 126      * block_job_finish_sync()), block_job_completed() won't free it and
 127      * therefore the blockers on the intermediate nodes remain. This would
 128      * cause bdrv_set_backing_hd() to fail. */
 129     block_job_remove_all_bdrv(job);
 130
 131     block_job_completed(&s->common, ret);
 132     g_free(data);
 133
 134     /* If bdrv_drop_intermediate() didn't already do that, remove the commit
 135      * filter driver from the backing chain. Do this as the final step so that
 136      * the 'consistent read' permission can be granted.  */
 137     if (remove_commit_top_bs) {
 138         bdrv_set_backing_hd(overlay_bs, top, &error_abort);
 139     }
 140
 141     bdrv_unref(overlay_bs);
 142     bdrv_unref(top);
 143 }
 144
 145 static void coroutine_fn commit_run(void *opaque)
 146 {
 147     CommitBlockJob *s = opaque;
 148     CommitCompleteData *data;
 149     int64_t sector_num, end;
 150     uint64_t delay_ns = 0;
 151     int ret = 0;
 152     int n = 0;
 153     void *buf = NULL;
 154     int bytes_written = 0;
 155     int64_t base_len;
 156
 157     ret = s->common.len = blk_getlength(s->top);
 158
 159
 160     if (s->common.len < 0) {
 161         goto out;
 162     }
 163
 164     ret = base_len = blk_getlength(s->base);
 165     if (base_len < 0) {
 166         goto out;
 167     }
 168
 169     if (base_len < s->common.len) {
 170         ret = blk_truncate(s->base, s->common.len, NULL);
 171         if (ret) {
 172             goto out;
 173         }
 174     }
 175
 176     end = s->common.len >> BDRV_SECTOR_BITS;
 177     buf = blk_blockalign(s->top, COMMIT_BUFFER_SIZE);
 178
 179     for (sector_num = 0; sector_num < end; sector_num += n) {
 180         bool copy;
 181
 182         /* Note that even when no rate limit is applied we need to yield
 183          * with no pending I/O here so that bdrv_drain_all() returns.
 184          */
 185         block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 186         if (block_job_is_cancelled(&s->common)) {
 187             break;
 188         }
 189         /* Copy if allocated above the base */
 190         ret = bdrv_is_allocated_above(blk_bs(s->top), blk_bs(s->base),
 191                                       sector_num,
 192                                       COMMIT_BUFFER_SIZE / BDRV_SECTOR_SIZE,
 193                                       &n);
 194         copy = (ret == 1);
 195         trace_commit_one_iteration(s, sector_num, n, ret);
 196         if (copy) {
 197             ret = commit_populate(s->top, s->base, sector_num, n, buf);
 198             bytes_written += n * BDRV_SECTOR_SIZE;
 199         }
 200         if (ret < 0) {
 201             BlockErrorAction action =
 202                 block_job_error_action(&s->common, false, s->on_error, -ret);
 203             if (action == BLOCK_ERROR_ACTION_REPORT) {
 204                 goto out;
 205             } else {
 206                 n = 0;
 207                 continue;
 208             }
 209         }
 210         /* Publish progress */
 211         s->common.offset += n * BDRV_SECTOR_SIZE;
 212
 213         if (copy && s->common.speed) {
 214             delay_ns = ratelimit_calculate_delay(&s->limit, n);
 215         }
 216     }
 217
 218     ret = 0;
 219
 220 out:
 221     qemu_vfree(buf);
 222
 223     data = g_malloc(sizeof(*data));
 224     data->ret = ret;
 225     block_job_defer_to_main_loop(&s->common, commit_complete, data);
 226 }
 227
 228 static void commit_set_speed(BlockJob *job, int64_t speed, Error **errp)
 229 {
 230     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
 231
 232     if (speed < 0) {
 233         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 234         return;
 235     }
 236     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 237 }
 238
 239 static const BlockJobDriver commit_job_driver = {
 240     .instance_size = sizeof(CommitBlockJob),
 241     .job_type      = BLOCK_JOB_TYPE_COMMIT,
 242     .set_speed     = commit_set_speed,
 243     .start         = commit_run,
 244 };
 245
 246 static int coroutine_fn bdrv_commit_top_preadv(BlockDriverState *bs,
 247     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
 248 {
 249     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
 250 }
 251
 252 static int64_t coroutine_fn bdrv_commit_top_get_block_status(
 253     BlockDriverState *bs, int64_t sector_num, int nb_sectors, int *pnum,
 254     BlockDriverState **file)
 255 {
 256     *pnum = nb_sectors;
 257     *file = bs->backing->bs;
 258     return BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID | BDRV_BLOCK_DATA |
 259            (sector_num << BDRV_SECTOR_BITS);
 260 }
 261
 262 static void bdrv_commit_top_refresh_filename(BlockDriverState *bs, QDict *opts)
 263 {
 264     bdrv_refresh_filename(bs->backing->bs);
 265     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
 266             bs->backing->bs->filename);
 267 }
 268
 269 static void bdrv_commit_top_close(BlockDriverState *bs)
 270 {
 271 }
 272
 273 static void bdrv_commit_top_child_perm(BlockDriverState *bs, BdrvChild *c,
 274                                        const BdrvChildRole *role,
 275                                        uint64_t perm, uint64_t shared,
 276                                        uint64_t *nperm, uint64_t *nshared)
 277 {
 278     *nperm = 0;
 279     *nshared = BLK_PERM_ALL;
 280 }
 281
 282 /* Dummy node that provides consistent read to its users without requiring it
 283  * from its backing file and that allows writes on the backing file chain. */
 284 static BlockDriver bdrv_commit_top = {
 285     .format_name                = "commit_top",
 286     .bdrv_co_preadv             = bdrv_commit_top_preadv,
 287     .bdrv_co_get_block_status   = bdrv_commit_top_get_block_status,
 288     .bdrv_refresh_filename      = bdrv_commit_top_refresh_filename,
 289     .bdrv_close                 = bdrv_commit_top_close,
 290     .bdrv_child_perm            = bdrv_commit_top_child_perm,
 291 };
 292
 293 void commit_start(const char *job_id, BlockDriverState *bs,
 294                   BlockDriverState *base, BlockDriverState *top, int64_t speed,
 295                   BlockdevOnError on_error, const char *backing_file_str,
 296                   const char *filter_node_name, Error **errp)
 297 {
 298     CommitBlockJob *s;
 299     BlockReopenQueue *reopen_queue = NULL;
 300     int orig_overlay_flags;
 301     int orig_base_flags;
 302     BlockDriverState *iter;
 303     BlockDriverState *overlay_bs;
 304     BlockDriverState *commit_top_bs = NULL;
 305     Error *local_err = NULL;
 306     int ret;
 307
 308     assert(top != bs);
 309     if (top == base) {
 310         error_setg(errp, "Invalid files for merge: top and base are the same");
 311         return;
 312     }
 313
 314     overlay_bs = bdrv_find_overlay(bs, top);
 315
 316     if (overlay_bs == NULL) {
 317         error_setg(errp, "Could not find overlay image for %s:", top->filename);
 318         return;
 319     }
 320
 321     s = block_job_create(job_id, &commit_job_driver, bs, 0, BLK_PERM_ALL,
 322                          speed, BLOCK_JOB_DEFAULT, NULL, NULL, errp);
 323     if (!s) {
 324         return;
 325     }
 326
 327     orig_base_flags    = bdrv_get_flags(base);
 328     orig_overlay_flags = bdrv_get_flags(overlay_bs);
 329
 330     /* convert base & overlay_bs to r/w, if necessary */
 331     if (!(orig_base_flags & BDRV_O_RDWR)) {
 332         reopen_queue = bdrv_reopen_queue(reopen_queue, base, NULL,
 333                                          orig_base_flags | BDRV_O_RDWR);
 334     }
 335     if (!(orig_overlay_flags & BDRV_O_RDWR)) {
 336         reopen_queue = bdrv_reopen_queue(reopen_queue, overlay_bs, NULL,
 337                                          orig_overlay_flags | BDRV_O_RDWR);
 338     }
 339     if (reopen_queue) {
 340         bdrv_reopen_multiple(bdrv_get_aio_context(bs), reopen_queue, &local_err);
 341         if (local_err != NULL) {
 342             error_propagate(errp, local_err);
 343             goto fail;
 344         }
 345     }
 346
 347     /* Insert commit_top block node above top, so we can block consistent read
 348      * on the backing chain below it */
 349     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, filter_node_name, 0,
 350                                          errp);
 351     if (commit_top_bs == NULL) {
 352         goto fail;
 353     }
 354     if (!filter_node_name) {
 355         commit_top_bs->implicit = true;
 356     }
 357     commit_top_bs->total_sectors = top->total_sectors;
 358     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(top));
 359
 360     bdrv_set_backing_hd(commit_top_bs, top, &local_err);
 361     if (local_err) {
 362         bdrv_unref(commit_top_bs);
 363         commit_top_bs = NULL;
 364         error_propagate(errp, local_err);
 365         goto fail;
 366     }
 367     bdrv_set_backing_hd(overlay_bs, commit_top_bs, &local_err);
 368     if (local_err) {
 369         bdrv_unref(commit_top_bs);
 370         commit_top_bs = NULL;
 371         error_propagate(errp, local_err);
 372         goto fail;
 373     }
 374
 375     s->commit_top_bs = commit_top_bs;
 376     bdrv_unref(commit_top_bs);
 377
 378     /* Block all nodes between top and base, because they will
 379      * disappear from the chain after this operation. */
 380     assert(bdrv_chain_contains(top, base));
 381     for (iter = top; iter != base; iter = backing_bs(iter)) {
 382         /* XXX BLK_PERM_WRITE needs to be allowed so we don't block ourselves
 383          * at s->base (if writes are blocked for a node, they are also blocked
 384          * for its backing file). The other options would be a second filter
 385          * driver above s->base. */
 386         ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
 387                                  BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
 388                                  errp);
 389         if (ret < 0) {
 390             goto fail;
 391         }
 392     }
 393
 394     ret = block_job_add_bdrv(&s->common, "base", base, 0, BLK_PERM_ALL, errp);
 395     if (ret < 0) {
 396         goto fail;
 397     }
 398
 399     /* overlay_bs must be blocked because it needs to be modified to
 400      * update the backing image string. */
 401     ret = block_job_add_bdrv(&s->common, "overlay of top", overlay_bs,
 402                              BLK_PERM_GRAPH_MOD, BLK_PERM_ALL, errp);
 403     if (ret < 0) {
 404         goto fail;
 405     }
 406
 407     s->base = blk_new(BLK_PERM_CONSISTENT_READ
 408                       | BLK_PERM_WRITE
 409                       | BLK_PERM_RESIZE,
 410                       BLK_PERM_CONSISTENT_READ
 411                       | BLK_PERM_GRAPH_MOD
 412                       | BLK_PERM_WRITE_UNCHANGED);
 413     ret = blk_insert_bs(s->base, base, errp);
 414     if (ret < 0) {
 415         goto fail;
 416     }
 417
 418     /* Required permissions are already taken with block_job_add_bdrv() */
 419     s->top = blk_new(0, BLK_PERM_ALL);
 420     ret = blk_insert_bs(s->top, top, errp);
 421     if (ret < 0) {
 422         goto fail;
 423     }
 424
 425     s->active = bs;
 426
 427     s->base_flags          = orig_base_flags;
 428     s->orig_overlay_flags  = orig_overlay_flags;
 429
 430     s->backing_file_str = g_strdup(backing_file_str);
 431
 432     s->on_error = on_error;
 433
 434     trace_commit_start(bs, base, top, s);
 435     block_job_start(&s->common);
 436     return;
 437
 438 fail:
 439     if (s->base) {
 440         blk_unref(s->base);
 441     }
 442     if (s->top) {
 443         blk_unref(s->top);
 444     }
 445     if (commit_top_bs) {
 446         bdrv_set_backing_hd(overlay_bs, top, &error_abort);
 447     }
 448     block_job_unref(&s->common);
 449 }
 450
 451
 452 #define COMMIT_BUF_SECTORS 2048
 453
 454 /* commit COW file into the raw image */
 455 int bdrv_commit(BlockDriverState *bs)
 456 {
 457     BlockBackend *src, *backing;
 458     BlockDriverState *backing_file_bs = NULL;
 459     BlockDriverState *commit_top_bs = NULL;
 460     BlockDriver *drv = bs->drv;
 461     int64_t sector, total_sectors, length, backing_length;
 462     int n, ro, open_flags;
 463     int ret = 0;
 464     uint8_t *buf = NULL;
 465     Error *local_err = NULL;
 466
 467     if (!drv)
 468         return -ENOMEDIUM;
 469
 470     if (!bs->backing) {
 471         return -ENOTSUP;
 472     }
 473
 474     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
 475         bdrv_op_is_blocked(bs->backing->bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL)) {
 476         return -EBUSY;
 477     }
 478
 479     ro = bs->backing->bs->read_only;
 480     open_flags =  bs->backing->bs->open_flags;
 481
 482     if (ro) {
 483         if (bdrv_reopen(bs->backing->bs, open_flags | BDRV_O_RDWR, NULL)) {
 484             return -EACCES;
 485         }
 486     }
 487
 488     src = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 489     backing = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
 490
 491     ret = blk_insert_bs(src, bs, &local_err);
 492     if (ret < 0) {
 493         error_report_err(local_err);
 494         goto ro_cleanup;
 495     }
 496
 497     /* Insert commit_top block node above backing, so we can write to it */
 498     backing_file_bs = backing_bs(bs);
 499
 500     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, NULL, BDRV_O_RDWR,
 501                                          &local_err);
 502     if (commit_top_bs == NULL) {
 503         error_report_err(local_err);
 504         goto ro_cleanup;
 505     }
 506     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(backing_file_bs));
 507
 508     bdrv_set_backing_hd(commit_top_bs, backing_file_bs, &error_abort);
 509     bdrv_set_backing_hd(bs, commit_top_bs, &error_abort);
 510
 511     ret = blk_insert_bs(backing, backing_file_bs, &local_err);
 512     if (ret < 0) {
 513         error_report_err(local_err);
 514         goto ro_cleanup;
 515     }
 516
 517     length = blk_getlength(src);
 518     if (length < 0) {
 519         ret = length;
 520         goto ro_cleanup;
 521     }
 522
 523     backing_length = blk_getlength(backing);
 524     if (backing_length < 0) {
 525         ret = backing_length;
 526         goto ro_cleanup;
 527     }
 528
 529     /* If our top snapshot is larger than the backing file image,
 530      * grow the backing file image if possible.  If not possible,
 531      * we must return an error */
 532     if (length > backing_length) {
 533         ret = blk_truncate(backing, length, &local_err);
 534         if (ret < 0) {
 535             error_report_err(local_err);
 536             goto ro_cleanup;
 537         }
 538     }
 539
 540     total_sectors = length >> BDRV_SECTOR_BITS;
 541
 542     /* blk_try_blockalign() for src will choose an alignment that works for
 543      * backing as well, so no need to compare the alignment manually. */
 544     buf = blk_try_blockalign(src, COMMIT_BUF_SECTORS * BDRV_SECTOR_SIZE);
 545     if (buf == NULL) {
 546         ret = -ENOMEM;
 547         goto ro_cleanup;
 548     }
 549
 550     for (sector = 0; sector < total_sectors; sector += n) {
 551         ret = bdrv_is_allocated(bs, sector, COMMIT_BUF_SECTORS, &n);
 552         if (ret < 0) {
 553             goto ro_cleanup;
 554         }
 555         if (ret) {
 556             ret = blk_pread(src, sector * BDRV_SECTOR_SIZE, buf,
 557                             n * BDRV_SECTOR_SIZE);
 558             if (ret < 0) {
 559                 goto ro_cleanup;
 560             }
 561
 562             ret = blk_pwrite(backing, sector * BDRV_SECTOR_SIZE, buf,
 563                              n * BDRV_SECTOR_SIZE, 0);
 564             if (ret < 0) {
 565                 goto ro_cleanup;
 566             }
 567         }
 568     }
 569
 570     if (drv->bdrv_make_empty) {
 571         ret = drv->bdrv_make_empty(bs);
 572         if (ret < 0) {
 573             goto ro_cleanup;
 574         }
 575         blk_flush(src);
 576     }
 577
 578     /*
 579      * Make sure all data we wrote to the backing device is actually
 580      * stable on disk.
 581      */
 582     blk_flush(backing);
 583
 584     ret = 0;
 585 ro_cleanup:
 586     qemu_vfree(buf);
 587
 588     blk_unref(backing);
 589     if (backing_file_bs) {
 590         bdrv_set_backing_hd(bs, backing_file_bs, &error_abort);
 591     }
 592     bdrv_unref(commit_top_bs);
 593     blk_unref(src);
 594
 595     if (ro) {
 596         /* ignoring error return here */
 597         bdrv_reopen(bs->backing->bs, open_flags & ~BDRV_O_RDWR, NULL);
 598     }
 599
 600     return ret;
 601 }