block/commit.c

   1 /*
   2  * Live block commit
   3  *
   4  * Copyright Red Hat, Inc. 2012
   5  *
   6  * Authors:
   7  *  Jeff Cody   <[email protected]>
   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 typedef struct {
  72     int ret;
  73 } CommitCompleteData;
  74
  75 static void commit_complete(BlockJob *job, void *opaque)
  76 {
  77     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
  78     CommitCompleteData *data = opaque;
  79     BlockDriverState *top = blk_bs(s->top);
  80     BlockDriverState *base = blk_bs(s->base);
  81     BlockDriverState *commit_top_bs = s->commit_top_bs;
  82     int ret = data->ret;
  83     bool remove_commit_top_bs = false;
  84
  85     /* Make sure commit_top_bs and top stay around until bdrv_replace_node() */
  86     bdrv_ref(top);
  87     bdrv_ref(commit_top_bs);
  88
  89     /* Remove base node parent that still uses BLK_PERM_WRITE/RESIZE before
  90      * the normal backing chain can be restored. */
  91     blk_unref(s->base);
  92
  93     if (!block_job_is_cancelled(&s->common) && ret == 0) {
  94         /* success */
  95         ret = bdrv_drop_intermediate(s->commit_top_bs, base,
  96                                      s->backing_file_str);
  97     } else {
  98         /* XXX Can (or should) we somehow keep 'consistent read' blocked even
  99          * after the failed/cancelled commit job is gone? If we already wrote
 100          * something to base, the intermediate images aren't valid any more. */
 101         remove_commit_top_bs = true;
 102     }
 103
 104     /* restore base open flags here if appropriate (e.g., change the base back
 105      * to r/o). These reopens do not need to be atomic, since we won't abort
 106      * even on failure here */
 107     if (s->base_flags != bdrv_get_flags(base)) {
 108         bdrv_reopen(base, s->base_flags, NULL);
 109     }
 110     g_free(s->backing_file_str);
 111     blk_unref(s->top);
 112
 113     /* If there is more than one reference to the job (e.g. if called from
 114      * block_job_finish_sync()), block_job_completed() won't free it and
 115      * therefore the blockers on the intermediate nodes remain. This would
 116      * cause bdrv_set_backing_hd() to fail. */
 117     block_job_remove_all_bdrv(job);
 118
 119     block_job_completed(&s->common, ret);
 120     g_free(data);
 121
 122     /* If bdrv_drop_intermediate() didn't already do that, remove the commit
 123      * filter driver from the backing chain. Do this as the final step so that
 124      * the 'consistent read' permission can be granted.  */
 125     if (remove_commit_top_bs) {
 126         bdrv_child_try_set_perm(commit_top_bs->backing, 0, BLK_PERM_ALL,
 127                                 &error_abort);
 128         bdrv_replace_node(commit_top_bs, backing_bs(commit_top_bs),
 129                           &error_abort);
 130     }
 131
 132     bdrv_unref(commit_top_bs);
 133     bdrv_unref(top);
 134 }
 135
 136 static void coroutine_fn commit_run(void *opaque)
 137 {
 138     CommitBlockJob *s = opaque;
 139     CommitCompleteData *data;
 140     int64_t offset;
 141     uint64_t delay_ns = 0;
 142     int ret = 0;
 143     int64_t n = 0; /* bytes */
 144     void *buf = NULL;
 145     int bytes_written = 0;
 146     int64_t len, base_len;
 147
 148     ret = len = blk_getlength(s->top);
 149     if (len < 0) {
 150         goto out;
 151     }
 152     block_job_progress_set_remaining(&s->common, len);
 153
 154     ret = base_len = blk_getlength(s->base);
 155     if (base_len < 0) {
 156         goto out;
 157     }
 158
 159     if (base_len < len) {
 160         ret = blk_truncate(s->base, len, PREALLOC_MODE_OFF, NULL);
 161         if (ret) {
 162             goto out;
 163         }
 164     }
 165
 166     buf = blk_blockalign(s->top, COMMIT_BUFFER_SIZE);
 167
 168     for (offset = 0; offset < len; offset += n) {
 169         bool copy;
 170
 171         /* Note that even when no rate limit is applied we need to yield
 172          * with no pending I/O here so that bdrv_drain_all() returns.
 173          */
 174         block_job_sleep_ns(&s->common, delay_ns);
 175         if (block_job_is_cancelled(&s->common)) {
 176             break;
 177         }
 178         /* Copy if allocated above the base */
 179         ret = bdrv_is_allocated_above(blk_bs(s->top), blk_bs(s->base),
 180                                       offset, COMMIT_BUFFER_SIZE, &n);
 181         copy = (ret == 1);
 182         trace_commit_one_iteration(s, offset, n, ret);
 183         if (copy) {
 184             ret = commit_populate(s->top, s->base, offset, n, buf);
 185             bytes_written += n;
 186         }
 187         if (ret < 0) {
 188             BlockErrorAction action =
 189                 block_job_error_action(&s->common, false, s->on_error, -ret);
 190             if (action == BLOCK_ERROR_ACTION_REPORT) {
 191                 goto out;
 192             } else {
 193                 n = 0;
 194                 continue;
 195             }
 196         }
 197         /* Publish progress */
 198         block_job_progress_update(&s->common, n);
 199
 200         if (copy) {
 201             delay_ns = block_job_ratelimit_get_delay(&s->common, n);
 202         } else {
 203             delay_ns = 0;
 204         }
 205     }
 206
 207     ret = 0;
 208
 209 out:
 210     qemu_vfree(buf);
 211
 212     data = g_malloc(sizeof(*data));
 213     data->ret = ret;
 214     block_job_defer_to_main_loop(&s->common, commit_complete, data);
 215 }
 216
 217 static const BlockJobDriver commit_job_driver = {
 218     .job_driver = {
 219         .instance_size = sizeof(CommitBlockJob),
 220     },
 221     .job_type      = JOB_TYPE_COMMIT,
 222     .start         = commit_run,
 223 };
 224
 225 static int coroutine_fn bdrv_commit_top_preadv(BlockDriverState *bs,
 226     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
 227 {
 228     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
 229 }
 230
 231 static void bdrv_commit_top_refresh_filename(BlockDriverState *bs, QDict *opts)
 232 {
 233     bdrv_refresh_filename(bs->backing->bs);
 234     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
 235             bs->backing->bs->filename);
 236 }
 237
 238 static void bdrv_commit_top_close(BlockDriverState *bs)
 239 {
 240 }
 241
 242 static void bdrv_commit_top_child_perm(BlockDriverState *bs, BdrvChild *c,
 243                                        const BdrvChildRole *role,
 244                                        BlockReopenQueue *reopen_queue,
 245                                        uint64_t perm, uint64_t shared,
 246                                        uint64_t *nperm, uint64_t *nshared)
 247 {
 248     *nperm = 0;
 249     *nshared = BLK_PERM_ALL;
 250 }
 251
 252 /* Dummy node that provides consistent read to its users without requiring it
 253  * from its backing file and that allows writes on the backing file chain. */
 254 static BlockDriver bdrv_commit_top = {
 255     .format_name                = "commit_top",
 256     .bdrv_co_preadv             = bdrv_commit_top_preadv,
 257     .bdrv_co_block_status       = bdrv_co_block_status_from_backing,
 258     .bdrv_refresh_filename      = bdrv_commit_top_refresh_filename,
 259     .bdrv_close                 = bdrv_commit_top_close,
 260     .bdrv_child_perm            = bdrv_commit_top_child_perm,
 261 };
 262
 263 void commit_start(const char *job_id, BlockDriverState *bs,
 264                   BlockDriverState *base, BlockDriverState *top, int64_t speed,
 265                   BlockdevOnError on_error, const char *backing_file_str,
 266                   const char *filter_node_name, Error **errp)
 267 {
 268     CommitBlockJob *s;
 269     int orig_base_flags;
 270     BlockDriverState *iter;
 271     BlockDriverState *commit_top_bs = NULL;
 272     Error *local_err = NULL;
 273     int ret;
 274
 275     assert(top != bs);
 276     if (top == base) {
 277         error_setg(errp, "Invalid files for merge: top and base are the same");
 278         return;
 279     }
 280
 281     s = block_job_create(job_id, &commit_job_driver, NULL, bs, 0, BLK_PERM_ALL,
 282                          speed, BLOCK_JOB_DEFAULT, NULL, NULL, errp);
 283     if (!s) {
 284         return;
 285     }
 286
 287     /* convert base to r/w, if necessary */
 288     orig_base_flags = bdrv_get_flags(base);
 289     if (!(orig_base_flags & BDRV_O_RDWR)) {
 290         bdrv_reopen(base, orig_base_flags | BDRV_O_RDWR, &local_err);
 291         if (local_err != NULL) {
 292             error_propagate(errp, local_err);
 293             goto fail;
 294         }
 295     }
 296
 297     /* Insert commit_top block node above top, so we can block consistent read
 298      * on the backing chain below it */
 299     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, filter_node_name, 0,
 300                                          errp);
 301     if (commit_top_bs == NULL) {
 302         goto fail;
 303     }
 304     if (!filter_node_name) {
 305         commit_top_bs->implicit = true;
 306     }
 307     commit_top_bs->total_sectors = top->total_sectors;
 308     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(top));
 309
 310     bdrv_set_backing_hd(commit_top_bs, top, &local_err);
 311     if (local_err) {
 312         bdrv_unref(commit_top_bs);
 313         commit_top_bs = NULL;
 314         error_propagate(errp, local_err);
 315         goto fail;
 316     }
 317     bdrv_replace_node(top, commit_top_bs, &local_err);
 318     if (local_err) {
 319         bdrv_unref(commit_top_bs);
 320         commit_top_bs = NULL;
 321         error_propagate(errp, local_err);
 322         goto fail;
 323     }
 324
 325     s->commit_top_bs = commit_top_bs;
 326     bdrv_unref(commit_top_bs);
 327
 328     /* Block all nodes between top and base, because they will
 329      * disappear from the chain after this operation. */
 330     assert(bdrv_chain_contains(top, base));
 331     for (iter = top; iter != base; iter = backing_bs(iter)) {
 332         /* XXX BLK_PERM_WRITE needs to be allowed so we don't block ourselves
 333          * at s->base (if writes are blocked for a node, they are also blocked
 334          * for its backing file). The other options would be a second filter
 335          * driver above s->base. */
 336         ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
 337                                  BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
 338                                  errp);
 339         if (ret < 0) {
 340             goto fail;
 341         }
 342     }
 343
 344     ret = block_job_add_bdrv(&s->common, "base", base, 0, BLK_PERM_ALL, errp);
 345     if (ret < 0) {
 346         goto fail;
 347     }
 348
 349     s->base = blk_new(BLK_PERM_CONSISTENT_READ
 350                       | BLK_PERM_WRITE
 351                       | BLK_PERM_RESIZE,
 352                       BLK_PERM_CONSISTENT_READ
 353                       | BLK_PERM_GRAPH_MOD
 354                       | BLK_PERM_WRITE_UNCHANGED);
 355     ret = blk_insert_bs(s->base, base, errp);
 356     if (ret < 0) {
 357         goto fail;
 358     }
 359
 360     /* Required permissions are already taken with block_job_add_bdrv() */
 361     s->top = blk_new(0, BLK_PERM_ALL);
 362     ret = blk_insert_bs(s->top, top, errp);
 363     if (ret < 0) {
 364         goto fail;
 365     }
 366
 367     s->base_flags = orig_base_flags;
 368     s->backing_file_str = g_strdup(backing_file_str);
 369     s->on_error = on_error;
 370
 371     trace_commit_start(bs, base, top, s);
 372     block_job_start(&s->common);
 373     return;
 374
 375 fail:
 376     if (s->base) {
 377         blk_unref(s->base);
 378     }
 379     if (s->top) {
 380         blk_unref(s->top);
 381     }
 382     if (commit_top_bs) {
 383         bdrv_replace_node(commit_top_bs, top, &error_abort);
 384     }
 385     block_job_early_fail(&s->common);
 386 }
 387
 388
 389 #define COMMIT_BUF_SIZE (2048 * BDRV_SECTOR_SIZE)
 390
 391 /* commit COW file into the raw image */
 392 int bdrv_commit(BlockDriverState *bs)
 393 {
 394     BlockBackend *src, *backing;
 395     BlockDriverState *backing_file_bs = NULL;
 396     BlockDriverState *commit_top_bs = NULL;
 397     BlockDriver *drv = bs->drv;
 398     int64_t offset, length, backing_length;
 399     int ro, open_flags;
 400     int64_t n;
 401     int ret = 0;
 402     uint8_t *buf = NULL;
 403     Error *local_err = NULL;
 404
 405     if (!drv)
 406         return -ENOMEDIUM;
 407
 408     if (!bs->backing) {
 409         return -ENOTSUP;
 410     }
 411
 412     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
 413         bdrv_op_is_blocked(bs->backing->bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL)) {
 414         return -EBUSY;
 415     }
 416
 417     ro = bs->backing->bs->read_only;
 418     open_flags =  bs->backing->bs->open_flags;
 419
 420     if (ro) {
 421         if (bdrv_reopen(bs->backing->bs, open_flags | BDRV_O_RDWR, NULL)) {
 422             return -EACCES;
 423         }
 424     }
 425
 426     src = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 427     backing = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
 428
 429     ret = blk_insert_bs(src, bs, &local_err);
 430     if (ret < 0) {
 431         error_report_err(local_err);
 432         goto ro_cleanup;
 433     }
 434
 435     /* Insert commit_top block node above backing, so we can write to it */
 436     backing_file_bs = backing_bs(bs);
 437
 438     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, NULL, BDRV_O_RDWR,
 439                                          &local_err);
 440     if (commit_top_bs == NULL) {
 441         error_report_err(local_err);
 442         goto ro_cleanup;
 443     }
 444     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(backing_file_bs));
 445
 446     bdrv_set_backing_hd(commit_top_bs, backing_file_bs, &error_abort);
 447     bdrv_set_backing_hd(bs, commit_top_bs, &error_abort);
 448
 449     ret = blk_insert_bs(backing, backing_file_bs, &local_err);
 450     if (ret < 0) {
 451         error_report_err(local_err);
 452         goto ro_cleanup;
 453     }
 454
 455     length = blk_getlength(src);
 456     if (length < 0) {
 457         ret = length;
 458         goto ro_cleanup;
 459     }
 460
 461     backing_length = blk_getlength(backing);
 462     if (backing_length < 0) {
 463         ret = backing_length;
 464         goto ro_cleanup;
 465     }
 466
 467     /* If our top snapshot is larger than the backing file image,
 468      * grow the backing file image if possible.  If not possible,
 469      * we must return an error */
 470     if (length > backing_length) {
 471         ret = blk_truncate(backing, length, PREALLOC_MODE_OFF, &local_err);
 472         if (ret < 0) {
 473             error_report_err(local_err);
 474             goto ro_cleanup;
 475         }
 476     }
 477
 478     /* blk_try_blockalign() for src will choose an alignment that works for
 479      * backing as well, so no need to compare the alignment manually. */
 480     buf = blk_try_blockalign(src, COMMIT_BUF_SIZE);
 481     if (buf == NULL) {
 482         ret = -ENOMEM;
 483         goto ro_cleanup;
 484     }
 485
 486     for (offset = 0; offset < length; offset += n) {
 487         ret = bdrv_is_allocated(bs, offset, COMMIT_BUF_SIZE, &n);
 488         if (ret < 0) {
 489             goto ro_cleanup;
 490         }
 491         if (ret) {
 492             ret = blk_pread(src, offset, buf, n);
 493             if (ret < 0) {
 494                 goto ro_cleanup;
 495             }
 496
 497             ret = blk_pwrite(backing, offset, buf, n, 0);
 498             if (ret < 0) {
 499                 goto ro_cleanup;
 500             }
 501         }
 502     }
 503
 504     if (drv->bdrv_make_empty) {
 505         ret = drv->bdrv_make_empty(bs);
 506         if (ret < 0) {
 507             goto ro_cleanup;
 508         }
 509         blk_flush(src);
 510     }
 511
 512     /*
 513      * Make sure all data we wrote to the backing device is actually
 514      * stable on disk.
 515      */
 516     blk_flush(backing);
 517
 518     ret = 0;
 519 ro_cleanup:
 520     qemu_vfree(buf);
 521
 522     blk_unref(backing);
 523     if (backing_file_bs) {
 524         bdrv_set_backing_hd(bs, backing_file_bs, &error_abort);
 525     }
 526     bdrv_unref(commit_top_bs);
 527     blk_unref(src);
 528
 529     if (ro) {
 530         /* ignoring error return here */
 531         bdrv_reopen(bs->backing->bs, open_flags & ~BDRV_O_RDWR, NULL);
 532     }
 533
 534     return ret;
 535 }