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 *commit_top_bs;
  40     BlockBackend *top;
  41     BlockBackend *base;
  42     BlockdevOnError on_error;
  43     int base_flags;
  44     char *backing_file_str;
  45 } CommitBlockJob;
  46
  47 static int coroutine_fn commit_populate(BlockBackend *bs, BlockBackend *base,
  48                                         int64_t offset, uint64_t bytes,
  49                                         void *buf)
  50 {
  51     int ret = 0;
  52     QEMUIOVector qiov;
  53     struct iovec iov = {
  54         .iov_base = buf,
  55         .iov_len = bytes,
  56     };
  57
  58     assert(bytes < SIZE_MAX);
  59     qemu_iovec_init_external(&qiov, &iov, 1);
  60
  61     ret = blk_co_preadv(bs, offset, qiov.size, &qiov, 0);
  62     if (ret < 0) {
  63         return ret;
  64     }
  65
  66     ret = blk_co_pwritev(base, offset, qiov.size, &qiov, 0);
  67     if (ret < 0) {
  68         return ret;
  69     }
  70
  71     return 0;
  72 }
  73
  74 typedef struct {
  75     int ret;
  76 } CommitCompleteData;
  77
  78 static void commit_complete(BlockJob *job, void *opaque)
  79 {
  80     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
  81     CommitCompleteData *data = opaque;
  82     BlockDriverState *top = blk_bs(s->top);
  83     BlockDriverState *base = blk_bs(s->base);
  84     BlockDriverState *commit_top_bs = s->commit_top_bs;
  85     int ret = data->ret;
  86     bool remove_commit_top_bs = false;
  87
  88     /* Make sure commit_top_bs and top stay around until bdrv_replace_node() */
  89     bdrv_ref(top);
  90     bdrv_ref(commit_top_bs);
  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(s->commit_top_bs, base,
  99                                      s->backing_file_str);
 100     } else {
 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     g_free(s->backing_file_str);
 114     blk_unref(s->top);
 115
 116     /* If there is more than one reference to the job (e.g. if called from
 117      * block_job_finish_sync()), block_job_completed() won't free it and
 118      * therefore the blockers on the intermediate nodes remain. This would
 119      * cause bdrv_set_backing_hd() to fail. */
 120     block_job_remove_all_bdrv(job);
 121
 122     block_job_completed(&s->common, ret);
 123     g_free(data);
 124
 125     /* If bdrv_drop_intermediate() didn't already do that, remove the commit
 126      * filter driver from the backing chain. Do this as the final step so that
 127      * the 'consistent read' permission can be granted.  */
 128     if (remove_commit_top_bs) {
 129         bdrv_child_try_set_perm(commit_top_bs->backing, 0, BLK_PERM_ALL,
 130                                 &error_abort);
 131         bdrv_replace_node(commit_top_bs, backing_bs(commit_top_bs),
 132                           &error_abort);
 133     }
 134
 135     bdrv_unref(commit_top_bs);
 136     bdrv_unref(top);
 137 }
 138
 139 static void coroutine_fn commit_run(void *opaque)
 140 {
 141     CommitBlockJob *s = opaque;
 142     CommitCompleteData *data;
 143     int64_t offset;
 144     uint64_t delay_ns = 0;
 145     int ret = 0;
 146     int64_t n = 0; /* bytes */
 147     void *buf = NULL;
 148     int bytes_written = 0;
 149     int64_t base_len;
 150
 151     ret = s->common.len = blk_getlength(s->top);
 152
 153     if (s->common.len < 0) {
 154         goto out;
 155     }
 156
 157     ret = base_len = blk_getlength(s->base);
 158     if (base_len < 0) {
 159         goto out;
 160     }
 161
 162     if (base_len < s->common.len) {
 163         ret = blk_truncate(s->base, s->common.len, PREALLOC_MODE_OFF, NULL);
 164         if (ret) {
 165             goto out;
 166         }
 167     }
 168
 169     buf = blk_blockalign(s->top, COMMIT_BUFFER_SIZE);
 170
 171     for (offset = 0; offset < s->common.len; offset += n) {
 172         bool copy;
 173
 174         /* Note that even when no rate limit is applied we need to yield
 175          * with no pending I/O here so that bdrv_drain_all() returns.
 176          */
 177         block_job_sleep_ns(&s->common, delay_ns);
 178         if (block_job_is_cancelled(&s->common)) {
 179             break;
 180         }
 181         /* Copy if allocated above the base */
 182         ret = bdrv_is_allocated_above(blk_bs(s->top), blk_bs(s->base),
 183                                       offset, COMMIT_BUFFER_SIZE, &n);
 184         copy = (ret == 1);
 185         trace_commit_one_iteration(s, offset, n, ret);
 186         if (copy) {
 187             ret = commit_populate(s->top, s->base, offset, n, buf);
 188             bytes_written += n;
 189         }
 190         if (ret < 0) {
 191             BlockErrorAction action =
 192                 block_job_error_action(&s->common, false, s->on_error, -ret);
 193             if (action == BLOCK_ERROR_ACTION_REPORT) {
 194                 goto out;
 195             } else {
 196                 n = 0;
 197                 continue;
 198             }
 199         }
 200         /* Publish progress */
 201         s->common.offset += n;
 202
 203         if (copy && s->common.speed) {
 204             delay_ns = ratelimit_calculate_delay(&s->limit, n);
 205         } else {
 206             delay_ns = 0;
 207         }
 208     }
 209
 210     ret = 0;
 211
 212 out:
 213     qemu_vfree(buf);
 214
 215     data = g_malloc(sizeof(*data));
 216     data->ret = ret;
 217     block_job_defer_to_main_loop(&s->common, commit_complete, data);
 218 }
 219
 220 static void commit_set_speed(BlockJob *job, int64_t speed, Error **errp)
 221 {
 222     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
 223
 224     if (speed < 0) {
 225         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 226         return;
 227     }
 228     ratelimit_set_speed(&s->limit, speed, SLICE_TIME);
 229 }
 230
 231 static const BlockJobDriver commit_job_driver = {
 232     .instance_size = sizeof(CommitBlockJob),
 233     .job_type      = BLOCK_JOB_TYPE_COMMIT,
 234     .set_speed     = commit_set_speed,
 235     .start         = commit_run,
 236 };
 237
 238 static int coroutine_fn bdrv_commit_top_preadv(BlockDriverState *bs,
 239     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
 240 {
 241     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
 242 }
 243
 244 static void bdrv_commit_top_refresh_filename(BlockDriverState *bs, QDict *opts)
 245 {
 246     bdrv_refresh_filename(bs->backing->bs);
 247     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
 248             bs->backing->bs->filename);
 249 }
 250
 251 static void bdrv_commit_top_close(BlockDriverState *bs)
 252 {
 253 }
 254
 255 static void bdrv_commit_top_child_perm(BlockDriverState *bs, BdrvChild *c,
 256                                        const BdrvChildRole *role,
 257                                        BlockReopenQueue *reopen_queue,
 258                                        uint64_t perm, uint64_t shared,
 259                                        uint64_t *nperm, uint64_t *nshared)
 260 {
 261     *nperm = 0;
 262     *nshared = BLK_PERM_ALL;
 263 }
 264
 265 /* Dummy node that provides consistent read to its users without requiring it
 266  * from its backing file and that allows writes on the backing file chain. */
 267 static BlockDriver bdrv_commit_top = {
 268     .format_name                = "commit_top",
 269     .bdrv_co_preadv             = bdrv_commit_top_preadv,
 270     .bdrv_co_block_status       = bdrv_co_block_status_from_backing,
 271     .bdrv_refresh_filename      = bdrv_commit_top_refresh_filename,
 272     .bdrv_close                 = bdrv_commit_top_close,
 273     .bdrv_child_perm            = bdrv_commit_top_child_perm,
 274 };
 275
 276 void commit_start(const char *job_id, BlockDriverState *bs,
 277                   BlockDriverState *base, BlockDriverState *top, int64_t speed,
 278                   BlockdevOnError on_error, const char *backing_file_str,
 279                   const char *filter_node_name, Error **errp)
 280 {
 281     CommitBlockJob *s;
 282     int orig_base_flags;
 283     BlockDriverState *iter;
 284     BlockDriverState *commit_top_bs = NULL;
 285     Error *local_err = NULL;
 286     int ret;
 287
 288     assert(top != bs);
 289     if (top == base) {
 290         error_setg(errp, "Invalid files for merge: top and base are the same");
 291         return;
 292     }
 293
 294     s = block_job_create(job_id, &commit_job_driver, NULL, bs, 0, BLK_PERM_ALL,
 295                          speed, BLOCK_JOB_DEFAULT, NULL, NULL, errp);
 296     if (!s) {
 297         return;
 298     }
 299
 300     /* convert base to r/w, if necessary */
 301     orig_base_flags = bdrv_get_flags(base);
 302     if (!(orig_base_flags & BDRV_O_RDWR)) {
 303         bdrv_reopen(base, orig_base_flags | BDRV_O_RDWR, &local_err);
 304         if (local_err != NULL) {
 305             error_propagate(errp, local_err);
 306             goto fail;
 307         }
 308     }
 309
 310     /* Insert commit_top block node above top, so we can block consistent read
 311      * on the backing chain below it */
 312     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, filter_node_name, 0,
 313                                          errp);
 314     if (commit_top_bs == NULL) {
 315         goto fail;
 316     }
 317     if (!filter_node_name) {
 318         commit_top_bs->implicit = true;
 319     }
 320     commit_top_bs->total_sectors = top->total_sectors;
 321     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(top));
 322
 323     bdrv_set_backing_hd(commit_top_bs, top, &local_err);
 324     if (local_err) {
 325         bdrv_unref(commit_top_bs);
 326         commit_top_bs = NULL;
 327         error_propagate(errp, local_err);
 328         goto fail;
 329     }
 330     bdrv_replace_node(top, commit_top_bs, &local_err);
 331     if (local_err) {
 332         bdrv_unref(commit_top_bs);
 333         commit_top_bs = NULL;
 334         error_propagate(errp, local_err);
 335         goto fail;
 336     }
 337
 338     s->commit_top_bs = commit_top_bs;
 339     bdrv_unref(commit_top_bs);
 340
 341     /* Block all nodes between top and base, because they will
 342      * disappear from the chain after this operation. */
 343     assert(bdrv_chain_contains(top, base));
 344     for (iter = top; iter != base; iter = backing_bs(iter)) {
 345         /* XXX BLK_PERM_WRITE needs to be allowed so we don't block ourselves
 346          * at s->base (if writes are blocked for a node, they are also blocked
 347          * for its backing file). The other options would be a second filter
 348          * driver above s->base. */
 349         ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
 350                                  BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
 351                                  errp);
 352         if (ret < 0) {
 353             goto fail;
 354         }
 355     }
 356
 357     ret = block_job_add_bdrv(&s->common, "base", base, 0, BLK_PERM_ALL, errp);
 358     if (ret < 0) {
 359         goto fail;
 360     }
 361
 362     s->base = blk_new(BLK_PERM_CONSISTENT_READ
 363                       | BLK_PERM_WRITE
 364                       | BLK_PERM_RESIZE,
 365                       BLK_PERM_CONSISTENT_READ
 366                       | BLK_PERM_GRAPH_MOD
 367                       | BLK_PERM_WRITE_UNCHANGED);
 368     ret = blk_insert_bs(s->base, base, errp);
 369     if (ret < 0) {
 370         goto fail;
 371     }
 372
 373     /* Required permissions are already taken with block_job_add_bdrv() */
 374     s->top = blk_new(0, BLK_PERM_ALL);
 375     ret = blk_insert_bs(s->top, top, errp);
 376     if (ret < 0) {
 377         goto fail;
 378     }
 379
 380     s->base_flags = orig_base_flags;
 381     s->backing_file_str = g_strdup(backing_file_str);
 382     s->on_error = on_error;
 383
 384     trace_commit_start(bs, base, top, s);
 385     block_job_start(&s->common);
 386     return;
 387
 388 fail:
 389     if (s->base) {
 390         blk_unref(s->base);
 391     }
 392     if (s->top) {
 393         blk_unref(s->top);
 394     }
 395     if (commit_top_bs) {
 396         bdrv_replace_node(commit_top_bs, top, &error_abort);
 397     }
 398     block_job_early_fail(&s->common);
 399 }
 400
 401
 402 #define COMMIT_BUF_SIZE (2048 * BDRV_SECTOR_SIZE)
 403
 404 /* commit COW file into the raw image */
 405 int bdrv_commit(BlockDriverState *bs)
 406 {
 407     BlockBackend *src, *backing;
 408     BlockDriverState *backing_file_bs = NULL;
 409     BlockDriverState *commit_top_bs = NULL;
 410     BlockDriver *drv = bs->drv;
 411     int64_t offset, length, backing_length;
 412     int ro, open_flags;
 413     int64_t n;
 414     int ret = 0;
 415     uint8_t *buf = NULL;
 416     Error *local_err = NULL;
 417
 418     if (!drv)
 419         return -ENOMEDIUM;
 420
 421     if (!bs->backing) {
 422         return -ENOTSUP;
 423     }
 424
 425     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
 426         bdrv_op_is_blocked(bs->backing->bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL)) {
 427         return -EBUSY;
 428     }
 429
 430     ro = bs->backing->bs->read_only;
 431     open_flags =  bs->backing->bs->open_flags;
 432
 433     if (ro) {
 434         if (bdrv_reopen(bs->backing->bs, open_flags | BDRV_O_RDWR, NULL)) {
 435             return -EACCES;
 436         }
 437     }
 438
 439     src = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 440     backing = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
 441
 442     ret = blk_insert_bs(src, bs, &local_err);
 443     if (ret < 0) {
 444         error_report_err(local_err);
 445         goto ro_cleanup;
 446     }
 447
 448     /* Insert commit_top block node above backing, so we can write to it */
 449     backing_file_bs = backing_bs(bs);
 450
 451     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, NULL, BDRV_O_RDWR,
 452                                          &local_err);
 453     if (commit_top_bs == NULL) {
 454         error_report_err(local_err);
 455         goto ro_cleanup;
 456     }
 457     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(backing_file_bs));
 458
 459     bdrv_set_backing_hd(commit_top_bs, backing_file_bs, &error_abort);
 460     bdrv_set_backing_hd(bs, commit_top_bs, &error_abort);
 461
 462     ret = blk_insert_bs(backing, backing_file_bs, &local_err);
 463     if (ret < 0) {
 464         error_report_err(local_err);
 465         goto ro_cleanup;
 466     }
 467
 468     length = blk_getlength(src);
 469     if (length < 0) {
 470         ret = length;
 471         goto ro_cleanup;
 472     }
 473
 474     backing_length = blk_getlength(backing);
 475     if (backing_length < 0) {
 476         ret = backing_length;
 477         goto ro_cleanup;
 478     }
 479
 480     /* If our top snapshot is larger than the backing file image,
 481      * grow the backing file image if possible.  If not possible,
 482      * we must return an error */
 483     if (length > backing_length) {
 484         ret = blk_truncate(backing, length, PREALLOC_MODE_OFF, &local_err);
 485         if (ret < 0) {
 486             error_report_err(local_err);
 487             goto ro_cleanup;
 488         }
 489     }
 490
 491     /* blk_try_blockalign() for src will choose an alignment that works for
 492      * backing as well, so no need to compare the alignment manually. */
 493     buf = blk_try_blockalign(src, COMMIT_BUF_SIZE);
 494     if (buf == NULL) {
 495         ret = -ENOMEM;
 496         goto ro_cleanup;
 497     }
 498
 499     for (offset = 0; offset < length; offset += n) {
 500         ret = bdrv_is_allocated(bs, offset, COMMIT_BUF_SIZE, &n);
 501         if (ret < 0) {
 502             goto ro_cleanup;
 503         }
 504         if (ret) {
 505             ret = blk_pread(src, offset, buf, n);
 506             if (ret < 0) {
 507                 goto ro_cleanup;
 508             }
 509
 510             ret = blk_pwrite(backing, offset, buf, n, 0);
 511             if (ret < 0) {
 512                 goto ro_cleanup;
 513             }
 514         }
 515     }
 516
 517     if (drv->bdrv_make_empty) {
 518         ret = drv->bdrv_make_empty(bs);
 519         if (ret < 0) {
 520             goto ro_cleanup;
 521         }
 522         blk_flush(src);
 523     }
 524
 525     /*
 526      * Make sure all data we wrote to the backing device is actually
 527      * stable on disk.
 528      */
 529     blk_flush(backing);
 530
 531     ret = 0;
 532 ro_cleanup:
 533     qemu_vfree(buf);
 534
 535     blk_unref(backing);
 536     if (backing_file_bs) {
 537         bdrv_set_backing_hd(bs, backing_file_bs, &error_abort);
 538     }
 539     bdrv_unref(commit_top_bs);
 540     blk_unref(src);
 541
 542     if (ro) {
 543         /* ignoring error return here */
 544         bdrv_reopen(bs->backing->bs, open_flags & ~BDRV_O_RDWR, NULL);
 545     }
 546
 547     return ret;
 548 }