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