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 "qemu/memalign.h" 24 #include "sysemu/block-backend.h" 25 26 enum { 27 /* 28 * Size of data buffer for populating the image file. This should be large 29 * enough to process multiple clusters in a single call, so that populating 30 * contiguous regions of the image is efficient. 31 */ 32 COMMIT_BUFFER_SIZE = 512 * 1024, /* in bytes */ 33 }; 34 35 typedef struct CommitBlockJob { 36 BlockJob common; 37 BlockDriverState *commit_top_bs; 38 BlockBackend *top; 39 BlockBackend *base; 40 BlockDriverState *base_bs; 41 BlockDriverState *base_overlay; 42 BlockdevOnError on_error; 43 bool base_read_only; 44 bool chain_frozen; 45 char *backing_file_str; 46 } CommitBlockJob; 47 48 static int commit_prepare(Job *job) 49 { 50 CommitBlockJob *s = container_of(job, CommitBlockJob, common.job); 51 52 bdrv_unfreeze_backing_chain(s->commit_top_bs, s->base_bs); 53 s->chain_frozen = false; 54 55 /* Remove base node parent that still uses BLK_PERM_WRITE/RESIZE before 56 * the normal backing chain can be restored. */ 57 blk_unref(s->base); 58 s->base = NULL; 59 60 /* FIXME: bdrv_drop_intermediate treats total failures and partial failures 61 * identically. Further work is needed to disambiguate these cases. */ 62 return bdrv_drop_intermediate(s->commit_top_bs, s->base_bs, 63 s->backing_file_str); 64 } 65 66 static void commit_abort(Job *job) 67 { 68 CommitBlockJob *s = container_of(job, CommitBlockJob, common.job); 69 BlockDriverState *top_bs = blk_bs(s->top); 70 71 if (s->chain_frozen) { 72 bdrv_unfreeze_backing_chain(s->commit_top_bs, s->base_bs); 73 } 74 75 /* Make sure commit_top_bs and top stay around until bdrv_replace_node() */ 76 bdrv_ref(top_bs); 77 bdrv_ref(s->commit_top_bs); 78 79 if (s->base) { 80 blk_unref(s->base); 81 } 82 83 /* free the blockers on the intermediate nodes so that bdrv_replace_nodes 84 * can succeed */ 85 block_job_remove_all_bdrv(&s->common); 86 87 /* If bdrv_drop_intermediate() failed (or was not invoked), remove the 88 * commit filter driver from the backing chain now. Do this as the final 89 * step so that the 'consistent read' permission can be granted. 90 * 91 * XXX Can (or should) we somehow keep 'consistent read' blocked even 92 * after the failed/cancelled commit job is gone? If we already wrote 93 * something to base, the intermediate images aren't valid any more. */ 94 bdrv_replace_node(s->commit_top_bs, s->commit_top_bs->backing->bs, 95 &error_abort); 96 97 bdrv_unref(s->commit_top_bs); 98 bdrv_unref(top_bs); 99 } 100 101 static void commit_clean(Job *job) 102 { 103 CommitBlockJob *s = container_of(job, CommitBlockJob, common.job); 104 105 /* restore base open flags here if appropriate (e.g., change the base back 106 * to r/o). These reopens do not need to be atomic, since we won't abort 107 * even on failure here */ 108 if (s->base_read_only) { 109 bdrv_reopen_set_read_only(s->base_bs, true, NULL); 110 } 111 112 g_free(s->backing_file_str); 113 blk_unref(s->top); 114 } 115 116 static int coroutine_fn commit_run(Job *job, Error **errp) 117 { 118 CommitBlockJob *s = container_of(job, CommitBlockJob, common.job); 119 int64_t offset; 120 uint64_t delay_ns = 0; 121 int ret = 0; 122 int64_t n = 0; /* bytes */ 123 QEMU_AUTO_VFREE void *buf = NULL; 124 int64_t len, base_len; 125 126 len = blk_co_getlength(s->top); 127 if (len < 0) { 128 return len; 129 } 130 job_progress_set_remaining(&s->common.job, len); 131 132 base_len = blk_co_getlength(s->base); 133 if (base_len < 0) { 134 return base_len; 135 } 136 137 if (base_len < len) { 138 ret = blk_co_truncate(s->base, len, false, PREALLOC_MODE_OFF, 0, NULL); 139 if (ret) { 140 return ret; 141 } 142 } 143 144 buf = blk_blockalign(s->top, COMMIT_BUFFER_SIZE); 145 146 for (offset = 0; offset < len; offset += n) { 147 bool copy; 148 bool error_in_source = true; 149 150 /* Note that even when no rate limit is applied we need to yield 151 * with no pending I/O here so that bdrv_drain_all() returns. 152 */ 153 job_sleep_ns(&s->common.job, delay_ns); 154 if (job_is_cancelled(&s->common.job)) { 155 break; 156 } 157 /* Copy if allocated above the base */ 158 ret = blk_co_is_allocated_above(s->top, s->base_overlay, true, 159 offset, COMMIT_BUFFER_SIZE, &n); 160 copy = (ret > 0); 161 trace_commit_one_iteration(s, offset, n, ret); 162 if (copy) { 163 assert(n < SIZE_MAX); 164 165 ret = blk_co_pread(s->top, offset, n, buf, 0); 166 if (ret >= 0) { 167 ret = blk_co_pwrite(s->base, offset, n, buf, 0); 168 if (ret < 0) { 169 error_in_source = false; 170 } 171 } 172 } 173 if (ret < 0) { 174 BlockErrorAction action = 175 block_job_error_action(&s->common, s->on_error, 176 error_in_source, -ret); 177 if (action == BLOCK_ERROR_ACTION_REPORT) { 178 return ret; 179 } else { 180 n = 0; 181 continue; 182 } 183 } 184 /* Publish progress */ 185 job_progress_update(&s->common.job, n); 186 187 if (copy) { 188 delay_ns = block_job_ratelimit_get_delay(&s->common, n); 189 } else { 190 delay_ns = 0; 191 } 192 } 193 194 return 0; 195 } 196 197 static const BlockJobDriver commit_job_driver = { 198 .job_driver = { 199 .instance_size = sizeof(CommitBlockJob), 200 .job_type = JOB_TYPE_COMMIT, 201 .free = block_job_free, 202 .user_resume = block_job_user_resume, 203 .run = commit_run, 204 .prepare = commit_prepare, 205 .abort = commit_abort, 206 .clean = commit_clean 207 }, 208 }; 209 210 static int coroutine_fn GRAPH_RDLOCK 211 bdrv_commit_top_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes, 212 QEMUIOVector *qiov, BdrvRequestFlags flags) 213 { 214 return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags); 215 } 216 217 static void bdrv_commit_top_refresh_filename(BlockDriverState *bs) 218 { 219 pstrcpy(bs->exact_filename, sizeof(bs->exact_filename), 220 bs->backing->bs->filename); 221 } 222 223 static void bdrv_commit_top_child_perm(BlockDriverState *bs, BdrvChild *c, 224 BdrvChildRole role, 225 BlockReopenQueue *reopen_queue, 226 uint64_t perm, uint64_t shared, 227 uint64_t *nperm, uint64_t *nshared) 228 { 229 *nperm = 0; 230 *nshared = BLK_PERM_ALL; 231 } 232 233 /* Dummy node that provides consistent read to its users without requiring it 234 * from its backing file and that allows writes on the backing file chain. */ 235 static BlockDriver bdrv_commit_top = { 236 .format_name = "commit_top", 237 .bdrv_co_preadv = bdrv_commit_top_preadv, 238 .bdrv_refresh_filename = bdrv_commit_top_refresh_filename, 239 .bdrv_child_perm = bdrv_commit_top_child_perm, 240 241 .is_filter = true, 242 .filtered_child_is_backing = true, 243 }; 244 245 void commit_start(const char *job_id, BlockDriverState *bs, 246 BlockDriverState *base, BlockDriverState *top, 247 int creation_flags, int64_t speed, 248 BlockdevOnError on_error, const char *backing_file_str, 249 const char *filter_node_name, Error **errp) 250 { 251 CommitBlockJob *s; 252 BlockDriverState *iter; 253 BlockDriverState *commit_top_bs = NULL; 254 BlockDriverState *filtered_base; 255 int64_t base_size, top_size; 256 uint64_t base_perms, iter_shared_perms; 257 int ret; 258 259 GLOBAL_STATE_CODE(); 260 261 assert(top != bs); 262 if (bdrv_skip_filters(top) == bdrv_skip_filters(base)) { 263 error_setg(errp, "Invalid files for merge: top and base are the same"); 264 return; 265 } 266 267 base_size = bdrv_getlength(base); 268 if (base_size < 0) { 269 error_setg_errno(errp, -base_size, "Could not inquire base image size"); 270 return; 271 } 272 273 top_size = bdrv_getlength(top); 274 if (top_size < 0) { 275 error_setg_errno(errp, -top_size, "Could not inquire top image size"); 276 return; 277 } 278 279 base_perms = BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE; 280 if (base_size < top_size) { 281 base_perms |= BLK_PERM_RESIZE; 282 } 283 284 s = block_job_create(job_id, &commit_job_driver, NULL, bs, 0, BLK_PERM_ALL, 285 speed, creation_flags, NULL, NULL, errp); 286 if (!s) { 287 return; 288 } 289 290 /* convert base to r/w, if necessary */ 291 s->base_read_only = bdrv_is_read_only(base); 292 if (s->base_read_only) { 293 if (bdrv_reopen_set_read_only(base, false, errp) != 0) { 294 goto fail; 295 } 296 } 297 298 /* Insert commit_top block node above top, so we can block consistent read 299 * on the backing chain below it */ 300 commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, filter_node_name, 0, 301 errp); 302 if (commit_top_bs == NULL) { 303 goto fail; 304 } 305 if (!filter_node_name) { 306 commit_top_bs->implicit = true; 307 } 308 309 /* So that we can always drop this node */ 310 commit_top_bs->never_freeze = true; 311 312 commit_top_bs->total_sectors = top->total_sectors; 313 314 ret = bdrv_append(commit_top_bs, top, errp); 315 bdrv_unref(commit_top_bs); /* referenced by new parents or failed */ 316 if (ret < 0) { 317 commit_top_bs = NULL; 318 goto fail; 319 } 320 321 s->commit_top_bs = commit_top_bs; 322 323 /* 324 * Block all nodes between top and base, because they will 325 * disappear from the chain after this operation. 326 * Note that this assumes that the user is fine with removing all 327 * nodes (including R/W filters) between top and base. Assuring 328 * this is the responsibility of the interface (i.e. whoever calls 329 * commit_start()). 330 */ 331 s->base_overlay = bdrv_find_overlay(top, base); 332 assert(s->base_overlay); 333 334 /* 335 * The topmost node with 336 * bdrv_skip_filters(filtered_base) == bdrv_skip_filters(base) 337 */ 338 filtered_base = bdrv_cow_bs(s->base_overlay); 339 assert(bdrv_skip_filters(filtered_base) == bdrv_skip_filters(base)); 340 341 /* 342 * XXX BLK_PERM_WRITE needs to be allowed so we don't block ourselves 343 * at s->base (if writes are blocked for a node, they are also blocked 344 * for its backing file). The other options would be a second filter 345 * driver above s->base. 346 */ 347 iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE; 348 349 for (iter = top; iter != base; iter = bdrv_filter_or_cow_bs(iter)) { 350 if (iter == filtered_base) { 351 /* 352 * From here on, all nodes are filters on the base. This 353 * allows us to share BLK_PERM_CONSISTENT_READ. 354 */ 355 iter_shared_perms |= BLK_PERM_CONSISTENT_READ; 356 } 357 358 ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0, 359 iter_shared_perms, errp); 360 if (ret < 0) { 361 goto fail; 362 } 363 } 364 365 if (bdrv_freeze_backing_chain(commit_top_bs, base, errp) < 0) { 366 goto fail; 367 } 368 s->chain_frozen = true; 369 370 ret = block_job_add_bdrv(&s->common, "base", base, 0, BLK_PERM_ALL, errp); 371 if (ret < 0) { 372 goto fail; 373 } 374 375 s->base = blk_new(s->common.job.aio_context, 376 base_perms, 377 BLK_PERM_CONSISTENT_READ 378 | BLK_PERM_WRITE_UNCHANGED); 379 ret = blk_insert_bs(s->base, base, errp); 380 if (ret < 0) { 381 goto fail; 382 } 383 blk_set_disable_request_queuing(s->base, true); 384 s->base_bs = base; 385 386 /* Required permissions are already taken with block_job_add_bdrv() */ 387 s->top = blk_new(s->common.job.aio_context, 0, BLK_PERM_ALL); 388 ret = blk_insert_bs(s->top, top, errp); 389 if (ret < 0) { 390 goto fail; 391 } 392 blk_set_disable_request_queuing(s->top, true); 393 394 s->backing_file_str = g_strdup(backing_file_str); 395 s->on_error = on_error; 396 397 trace_commit_start(bs, base, top, s); 398 job_start(&s->common.job); 399 return; 400 401 fail: 402 if (s->chain_frozen) { 403 bdrv_unfreeze_backing_chain(commit_top_bs, base); 404 } 405 if (s->base) { 406 blk_unref(s->base); 407 } 408 if (s->top) { 409 blk_unref(s->top); 410 } 411 if (s->base_read_only) { 412 bdrv_reopen_set_read_only(base, true, NULL); 413 } 414 job_early_fail(&s->common.job); 415 /* commit_top_bs has to be replaced after deleting the block job, 416 * otherwise this would fail because of lack of permissions. */ 417 if (commit_top_bs) { 418 bdrv_replace_node(commit_top_bs, top, &error_abort); 419 } 420 } 421 422 423 #define COMMIT_BUF_SIZE (2048 * BDRV_SECTOR_SIZE) 424 425 /* commit COW file into the raw image */ 426 int bdrv_commit(BlockDriverState *bs) 427 { 428 BlockBackend *src, *backing; 429 BlockDriverState *backing_file_bs = NULL; 430 BlockDriverState *commit_top_bs = NULL; 431 BlockDriver *drv = bs->drv; 432 AioContext *ctx; 433 int64_t offset, length, backing_length; 434 int ro; 435 int64_t n; 436 int ret = 0; 437 QEMU_AUTO_VFREE uint8_t *buf = NULL; 438 Error *local_err = NULL; 439 440 GLOBAL_STATE_CODE(); 441 442 if (!drv) 443 return -ENOMEDIUM; 444 445 backing_file_bs = bdrv_cow_bs(bs); 446 447 if (!backing_file_bs) { 448 return -ENOTSUP; 449 } 450 451 if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) || 452 bdrv_op_is_blocked(backing_file_bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL)) 453 { 454 return -EBUSY; 455 } 456 457 ro = bdrv_is_read_only(backing_file_bs); 458 459 if (ro) { 460 if (bdrv_reopen_set_read_only(backing_file_bs, false, NULL)) { 461 return -EACCES; 462 } 463 } 464 465 ctx = bdrv_get_aio_context(bs); 466 /* WRITE_UNCHANGED is required for bdrv_make_empty() */ 467 src = blk_new(ctx, BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED, 468 BLK_PERM_ALL); 469 backing = blk_new(ctx, BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL); 470 471 ret = blk_insert_bs(src, bs, &local_err); 472 if (ret < 0) { 473 error_report_err(local_err); 474 goto ro_cleanup; 475 } 476 477 /* Insert commit_top block node above backing, so we can write to it */ 478 commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, NULL, BDRV_O_RDWR, 479 &local_err); 480 if (commit_top_bs == NULL) { 481 error_report_err(local_err); 482 goto ro_cleanup; 483 } 484 485 bdrv_set_backing_hd(commit_top_bs, backing_file_bs, &error_abort); 486 bdrv_set_backing_hd(bs, commit_top_bs, &error_abort); 487 488 ret = blk_insert_bs(backing, backing_file_bs, &local_err); 489 if (ret < 0) { 490 error_report_err(local_err); 491 goto ro_cleanup; 492 } 493 494 length = blk_getlength(src); 495 if (length < 0) { 496 ret = length; 497 goto ro_cleanup; 498 } 499 500 backing_length = blk_getlength(backing); 501 if (backing_length < 0) { 502 ret = backing_length; 503 goto ro_cleanup; 504 } 505 506 /* If our top snapshot is larger than the backing file image, 507 * grow the backing file image if possible. If not possible, 508 * we must return an error */ 509 if (length > backing_length) { 510 ret = blk_truncate(backing, length, false, PREALLOC_MODE_OFF, 0, 511 &local_err); 512 if (ret < 0) { 513 error_report_err(local_err); 514 goto ro_cleanup; 515 } 516 } 517 518 /* blk_try_blockalign() for src will choose an alignment that works for 519 * backing as well, so no need to compare the alignment manually. */ 520 buf = blk_try_blockalign(src, COMMIT_BUF_SIZE); 521 if (buf == NULL) { 522 ret = -ENOMEM; 523 goto ro_cleanup; 524 } 525 526 for (offset = 0; offset < length; offset += n) { 527 ret = bdrv_is_allocated(bs, offset, COMMIT_BUF_SIZE, &n); 528 if (ret < 0) { 529 goto ro_cleanup; 530 } 531 if (ret) { 532 ret = blk_pread(src, offset, n, buf, 0); 533 if (ret < 0) { 534 goto ro_cleanup; 535 } 536 537 ret = blk_pwrite(backing, offset, n, buf, 0); 538 if (ret < 0) { 539 goto ro_cleanup; 540 } 541 } 542 } 543 544 ret = blk_make_empty(src, NULL); 545 /* Ignore -ENOTSUP */ 546 if (ret < 0 && ret != -ENOTSUP) { 547 goto ro_cleanup; 548 } 549 550 blk_flush(src); 551 552 /* 553 * Make sure all data we wrote to the backing device is actually 554 * stable on disk. 555 */ 556 blk_flush(backing); 557 558 ret = 0; 559 ro_cleanup: 560 blk_unref(backing); 561 if (bdrv_cow_bs(bs) != backing_file_bs) { 562 bdrv_set_backing_hd(bs, backing_file_bs, &error_abort); 563 } 564 bdrv_unref(commit_top_bs); 565 blk_unref(src); 566 567 if (ro) { 568 /* ignoring error return here */ 569 bdrv_reopen_set_read_only(backing_file_bs, true, NULL); 570 } 571 572 return ret; 573 } 574