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