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         }
 206     }
 207
 208     ret = 0;
 209
 210 out:
 211     qemu_vfree(buf);
 212
 213     data = g_malloc(sizeof(*data));
 214     data->ret = ret;
 215     block_job_defer_to_main_loop(&s->common, commit_complete, data);
 216 }
 217
 218 static void commit_set_speed(BlockJob *job, int64_t speed, Error **errp)
 219 {
 220     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
 221
 222     if (speed < 0) {
 223         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 224         return;
 225     }
 226     ratelimit_set_speed(&s->limit, speed, SLICE_TIME);
 227 }
 228
 229 static const BlockJobDriver commit_job_driver = {
 230     .instance_size = sizeof(CommitBlockJob),
 231     .job_type      = BLOCK_JOB_TYPE_COMMIT,
 232     .set_speed     = commit_set_speed,
 233     .start         = commit_run,
 234 };
 235
 236 static int coroutine_fn bdrv_commit_top_preadv(BlockDriverState *bs,
 237     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
 238 {
 239     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
 240 }
 241
 242 static void bdrv_commit_top_refresh_filename(BlockDriverState *bs, QDict *opts)
 243 {
 244     bdrv_refresh_filename(bs->backing->bs);
 245     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
 246             bs->backing->bs->filename);
 247 }
 248
 249 static void bdrv_commit_top_close(BlockDriverState *bs)
 250 {
 251 }
 252
 253 static void bdrv_commit_top_child_perm(BlockDriverState *bs, BdrvChild *c,
 254                                        const BdrvChildRole *role,
 255                                        BlockReopenQueue *reopen_queue,
 256                                        uint64_t perm, uint64_t shared,
 257                                        uint64_t *nperm, uint64_t *nshared)
 258 {
 259     *nperm = 0;
 260     *nshared = BLK_PERM_ALL;
 261 }
 262
 263 /* Dummy node that provides consistent read to its users without requiring it
 264  * from its backing file and that allows writes on the backing file chain. */
 265 static BlockDriver bdrv_commit_top = {
 266     .format_name                = "commit_top",
 267     .bdrv_co_preadv             = bdrv_commit_top_preadv,
 268     .bdrv_co_get_block_status   = bdrv_co_get_block_status_from_backing,
 269     .bdrv_refresh_filename      = bdrv_commit_top_refresh_filename,
 270     .bdrv_close                 = bdrv_commit_top_close,
 271     .bdrv_child_perm            = bdrv_commit_top_child_perm,
 272 };
 273
 274 void commit_start(const char *job_id, BlockDriverState *bs,
 275                   BlockDriverState *base, BlockDriverState *top, int64_t speed,
 276                   BlockdevOnError on_error, const char *backing_file_str,
 277                   const char *filter_node_name, Error **errp)
 278 {
 279     CommitBlockJob *s;
 280     int orig_base_flags;
 281     BlockDriverState *iter;
 282     BlockDriverState *commit_top_bs = NULL;
 283     Error *local_err = NULL;
 284     int ret;
 285
 286     assert(top != bs);
 287     if (top == base) {
 288         error_setg(errp, "Invalid files for merge: top and base are the same");
 289         return;
 290     }
 291
 292     s = block_job_create(job_id, &commit_job_driver, bs, 0, BLK_PERM_ALL,
 293                          speed, BLOCK_JOB_DEFAULT, NULL, NULL, errp);
 294     if (!s) {
 295         return;
 296     }
 297
 298     /* convert base to r/w, if necessary */
 299     orig_base_flags = bdrv_get_flags(base);
 300     if (!(orig_base_flags & BDRV_O_RDWR)) {
 301         bdrv_reopen(base, orig_base_flags | BDRV_O_RDWR, &local_err);
 302         if (local_err != NULL) {
 303             error_propagate(errp, local_err);
 304             goto fail;
 305         }
 306     }
 307
 308     /* Insert commit_top block node above top, so we can block consistent read
 309      * on the backing chain below it */
 310     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, filter_node_name, 0,
 311                                          errp);
 312     if (commit_top_bs == NULL) {
 313         goto fail;
 314     }
 315     if (!filter_node_name) {
 316         commit_top_bs->implicit = true;
 317     }
 318     commit_top_bs->total_sectors = top->total_sectors;
 319     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(top));
 320
 321     bdrv_set_backing_hd(commit_top_bs, top, &local_err);
 322     if (local_err) {
 323         bdrv_unref(commit_top_bs);
 324         commit_top_bs = NULL;
 325         error_propagate(errp, local_err);
 326         goto fail;
 327     }
 328     bdrv_replace_node(top, commit_top_bs, &local_err);
 329     if (local_err) {
 330         bdrv_unref(commit_top_bs);
 331         commit_top_bs = NULL;
 332         error_propagate(errp, local_err);
 333         goto fail;
 334     }
 335
 336     s->commit_top_bs = commit_top_bs;
 337     bdrv_unref(commit_top_bs);
 338
 339     /* Block all nodes between top and base, because they will
 340      * disappear from the chain after this operation. */
 341     assert(bdrv_chain_contains(top, base));
 342     for (iter = top; iter != base; iter = backing_bs(iter)) {
 343         /* XXX BLK_PERM_WRITE needs to be allowed so we don't block ourselves
 344          * at s->base (if writes are blocked for a node, they are also blocked
 345          * for its backing file). The other options would be a second filter
 346          * driver above s->base. */
 347         ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
 348                                  BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
 349                                  errp);
 350         if (ret < 0) {
 351             goto fail;
 352         }
 353     }
 354
 355     ret = block_job_add_bdrv(&s->common, "base", base, 0, BLK_PERM_ALL, errp);
 356     if (ret < 0) {
 357         goto fail;
 358     }
 359
 360     s->base = blk_new(BLK_PERM_CONSISTENT_READ
 361                       | BLK_PERM_WRITE
 362                       | BLK_PERM_RESIZE,
 363                       BLK_PERM_CONSISTENT_READ
 364                       | BLK_PERM_GRAPH_MOD
 365                       | BLK_PERM_WRITE_UNCHANGED);
 366     ret = blk_insert_bs(s->base, base, errp);
 367     if (ret < 0) {
 368         goto fail;
 369     }
 370
 371     /* Required permissions are already taken with block_job_add_bdrv() */
 372     s->top = blk_new(0, BLK_PERM_ALL);
 373     ret = blk_insert_bs(s->top, top, errp);
 374     if (ret < 0) {
 375         goto fail;
 376     }
 377
 378     s->base_flags = orig_base_flags;
 379     s->backing_file_str = g_strdup(backing_file_str);
 380     s->on_error = on_error;
 381
 382     trace_commit_start(bs, base, top, s);
 383     block_job_start(&s->common);
 384     return;
 385
 386 fail:
 387     if (s->base) {
 388         blk_unref(s->base);
 389     }
 390     if (s->top) {
 391         blk_unref(s->top);
 392     }
 393     if (commit_top_bs) {
 394         bdrv_replace_node(commit_top_bs, top, &error_abort);
 395     }
 396     block_job_early_fail(&s->common);
 397 }
 398
 399
 400 #define COMMIT_BUF_SIZE (2048 * BDRV_SECTOR_SIZE)
 401
 402 /* commit COW file into the raw image */
 403 int bdrv_commit(BlockDriverState *bs)
 404 {
 405     BlockBackend *src, *backing;
 406     BlockDriverState *backing_file_bs = NULL;
 407     BlockDriverState *commit_top_bs = NULL;
 408     BlockDriver *drv = bs->drv;
 409     int64_t offset, length, backing_length;
 410     int ro, open_flags;
 411     int64_t n;
 412     int ret = 0;
 413     uint8_t *buf = NULL;
 414     Error *local_err = NULL;
 415
 416     if (!drv)
 417         return -ENOMEDIUM;
 418
 419     if (!bs->backing) {
 420         return -ENOTSUP;
 421     }
 422
 423     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
 424         bdrv_op_is_blocked(bs->backing->bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL)) {
 425         return -EBUSY;
 426     }
 427
 428     ro = bs->backing->bs->read_only;
 429     open_flags =  bs->backing->bs->open_flags;
 430
 431     if (ro) {
 432         if (bdrv_reopen(bs->backing->bs, open_flags | BDRV_O_RDWR, NULL)) {
 433             return -EACCES;
 434         }
 435     }
 436
 437     src = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 438     backing = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
 439
 440     ret = blk_insert_bs(src, bs, &local_err);
 441     if (ret < 0) {
 442         error_report_err(local_err);
 443         goto ro_cleanup;
 444     }
 445
 446     /* Insert commit_top block node above backing, so we can write to it */
 447     backing_file_bs = backing_bs(bs);
 448
 449     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, NULL, BDRV_O_RDWR,
 450                                          &local_err);
 451     if (commit_top_bs == NULL) {
 452         error_report_err(local_err);
 453         goto ro_cleanup;
 454     }
 455     bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(backing_file_bs));
 456
 457     bdrv_set_backing_hd(commit_top_bs, backing_file_bs, &error_abort);
 458     bdrv_set_backing_hd(bs, commit_top_bs, &error_abort);
 459
 460     ret = blk_insert_bs(backing, backing_file_bs, &local_err);
 461     if (ret < 0) {
 462         error_report_err(local_err);
 463         goto ro_cleanup;
 464     }
 465
 466     length = blk_getlength(src);
 467     if (length < 0) {
 468         ret = length;
 469         goto ro_cleanup;
 470     }
 471
 472     backing_length = blk_getlength(backing);
 473     if (backing_length < 0) {
 474         ret = backing_length;
 475         goto ro_cleanup;
 476     }
 477
 478     /* If our top snapshot is larger than the backing file image,
 479      * grow the backing file image if possible.  If not possible,
 480      * we must return an error */
 481     if (length > backing_length) {
 482         ret = blk_truncate(backing, length, PREALLOC_MODE_OFF, &local_err);
 483         if (ret < 0) {
 484             error_report_err(local_err);
 485             goto ro_cleanup;
 486         }
 487     }
 488
 489     /* blk_try_blockalign() for src will choose an alignment that works for
 490      * backing as well, so no need to compare the alignment manually. */
 491     buf = blk_try_blockalign(src, COMMIT_BUF_SIZE);
 492     if (buf == NULL) {
 493         ret = -ENOMEM;
 494         goto ro_cleanup;
 495     }
 496
 497     for (offset = 0; offset < length; offset += n) {
 498         ret = bdrv_is_allocated(bs, offset, COMMIT_BUF_SIZE, &n);
 499         if (ret < 0) {
 500             goto ro_cleanup;
 501         }
 502         if (ret) {
 503             ret = blk_pread(src, offset, buf, n);
 504             if (ret < 0) {
 505                 goto ro_cleanup;
 506             }
 507
 508             ret = blk_pwrite(backing, offset, buf, n, 0);
 509             if (ret < 0) {
 510                 goto ro_cleanup;
 511             }
 512         }
 513     }
 514
 515     if (drv->bdrv_make_empty) {
 516         ret = drv->bdrv_make_empty(bs);
 517         if (ret < 0) {
 518             goto ro_cleanup;
 519         }
 520         blk_flush(src);
 521     }
 522
 523     /*
 524      * Make sure all data we wrote to the backing device is actually
 525      * stable on disk.
 526      */
 527     blk_flush(backing);
 528
 529     ret = 0;
 530 ro_cleanup:
 531     qemu_vfree(buf);
 532
 533     blk_unref(backing);
 534     if (backing_file_bs) {
 535         bdrv_set_backing_hd(bs, backing_file_bs, &error_abort);
 536     }
 537     bdrv_unref(commit_top_bs);
 538     blk_unref(src);
 539
 540     if (ro) {
 541         /* ignoring error return here */
 542         bdrv_reopen(bs->backing->bs, open_flags & ~BDRV_O_RDWR, NULL);
 543     }
 544
 545     return ret;
 546 }