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