1 /* 2 * QEMU backup 3 * 4 * Copyright (C) 2013 Proxmox Server Solutions 5 * 6 * Authors: 7 * Dietmar Maurer (dietmar@proxmox.com) 8 * 9 * This work is licensed under the terms of the GNU GPL, version 2 or later. 10 * See the COPYING file in the top-level directory. 11 * 12 */ 13 14 #include <stdio.h> 15 #include <errno.h> 16 #include <unistd.h> 17 18 #include "trace.h" 19 #include "block/block.h" 20 #include "block/block_int.h" 21 #include "block/blockjob.h" 22 #include "qapi/qmp/qerror.h" 23 #include "qemu/ratelimit.h" 24 #include "sysemu/block-backend.h" 25 26 #define BACKUP_CLUSTER_BITS 16 27 #define BACKUP_CLUSTER_SIZE (1 << BACKUP_CLUSTER_BITS) 28 #define BACKUP_SECTORS_PER_CLUSTER (BACKUP_CLUSTER_SIZE / BDRV_SECTOR_SIZE) 29 30 #define SLICE_TIME 100000000ULL /* ns */ 31 32 typedef struct CowRequest { 33 int64_t start; 34 int64_t end; 35 QLIST_ENTRY(CowRequest) list; 36 CoQueue wait_queue; /* coroutines blocked on this request */ 37 } CowRequest; 38 39 typedef struct BackupBlockJob { 40 BlockJob common; 41 BlockDriverState *target; 42 /* bitmap for sync=incremental */ 43 BdrvDirtyBitmap *sync_bitmap; 44 MirrorSyncMode sync_mode; 45 RateLimit limit; 46 BlockdevOnError on_source_error; 47 BlockdevOnError on_target_error; 48 CoRwlock flush_rwlock; 49 uint64_t sectors_read; 50 HBitmap *bitmap; 51 QLIST_HEAD(, CowRequest) inflight_reqs; 52 } BackupBlockJob; 53 54 /* See if in-flight requests overlap and wait for them to complete */ 55 static void coroutine_fn wait_for_overlapping_requests(BackupBlockJob *job, 56 int64_t start, 57 int64_t end) 58 { 59 CowRequest *req; 60 bool retry; 61 62 do { 63 retry = false; 64 QLIST_FOREACH(req, &job->inflight_reqs, list) { 65 if (end > req->start && start < req->end) { 66 qemu_co_queue_wait(&req->wait_queue); 67 retry = true; 68 break; 69 } 70 } 71 } while (retry); 72 } 73 74 /* Keep track of an in-flight request */ 75 static void cow_request_begin(CowRequest *req, BackupBlockJob *job, 76 int64_t start, int64_t end) 77 { 78 req->start = start; 79 req->end = end; 80 qemu_co_queue_init(&req->wait_queue); 81 QLIST_INSERT_HEAD(&job->inflight_reqs, req, list); 82 } 83 84 /* Forget about a completed request */ 85 static void cow_request_end(CowRequest *req) 86 { 87 QLIST_REMOVE(req, list); 88 qemu_co_queue_restart_all(&req->wait_queue); 89 } 90 91 static int coroutine_fn backup_do_cow(BlockDriverState *bs, 92 int64_t sector_num, int nb_sectors, 93 bool *error_is_read, 94 bool is_write_notifier) 95 { 96 BackupBlockJob *job = (BackupBlockJob *)bs->job; 97 CowRequest cow_request; 98 struct iovec iov; 99 QEMUIOVector bounce_qiov; 100 void *bounce_buffer = NULL; 101 int ret = 0; 102 int64_t start, end; 103 int n; 104 105 qemu_co_rwlock_rdlock(&job->flush_rwlock); 106 107 start = sector_num / BACKUP_SECTORS_PER_CLUSTER; 108 end = DIV_ROUND_UP(sector_num + nb_sectors, BACKUP_SECTORS_PER_CLUSTER); 109 110 trace_backup_do_cow_enter(job, start, sector_num, nb_sectors); 111 112 wait_for_overlapping_requests(job, start, end); 113 cow_request_begin(&cow_request, job, start, end); 114 115 for (; start < end; start++) { 116 if (hbitmap_get(job->bitmap, start)) { 117 trace_backup_do_cow_skip(job, start); 118 continue; /* already copied */ 119 } 120 121 trace_backup_do_cow_process(job, start); 122 123 n = MIN(BACKUP_SECTORS_PER_CLUSTER, 124 job->common.len / BDRV_SECTOR_SIZE - 125 start * BACKUP_SECTORS_PER_CLUSTER); 126 127 if (!bounce_buffer) { 128 bounce_buffer = qemu_blockalign(bs, BACKUP_CLUSTER_SIZE); 129 } 130 iov.iov_base = bounce_buffer; 131 iov.iov_len = n * BDRV_SECTOR_SIZE; 132 qemu_iovec_init_external(&bounce_qiov, &iov, 1); 133 134 if (is_write_notifier) { 135 ret = bdrv_co_no_copy_on_readv(bs, 136 start * BACKUP_SECTORS_PER_CLUSTER, 137 n, &bounce_qiov); 138 } else { 139 ret = bdrv_co_readv(bs, start * BACKUP_SECTORS_PER_CLUSTER, n, 140 &bounce_qiov); 141 } 142 if (ret < 0) { 143 trace_backup_do_cow_read_fail(job, start, ret); 144 if (error_is_read) { 145 *error_is_read = true; 146 } 147 goto out; 148 } 149 150 if (buffer_is_zero(iov.iov_base, iov.iov_len)) { 151 ret = bdrv_co_write_zeroes(job->target, 152 start * BACKUP_SECTORS_PER_CLUSTER, 153 n, BDRV_REQ_MAY_UNMAP); 154 } else { 155 ret = bdrv_co_writev(job->target, 156 start * BACKUP_SECTORS_PER_CLUSTER, n, 157 &bounce_qiov); 158 } 159 if (ret < 0) { 160 trace_backup_do_cow_write_fail(job, start, ret); 161 if (error_is_read) { 162 *error_is_read = false; 163 } 164 goto out; 165 } 166 167 hbitmap_set(job->bitmap, start, 1); 168 169 /* Publish progress, guest I/O counts as progress too. Note that the 170 * offset field is an opaque progress value, it is not a disk offset. 171 */ 172 job->sectors_read += n; 173 job->common.offset += n * BDRV_SECTOR_SIZE; 174 } 175 176 out: 177 if (bounce_buffer) { 178 qemu_vfree(bounce_buffer); 179 } 180 181 cow_request_end(&cow_request); 182 183 trace_backup_do_cow_return(job, sector_num, nb_sectors, ret); 184 185 qemu_co_rwlock_unlock(&job->flush_rwlock); 186 187 return ret; 188 } 189 190 static int coroutine_fn backup_before_write_notify( 191 NotifierWithReturn *notifier, 192 void *opaque) 193 { 194 BdrvTrackedRequest *req = opaque; 195 int64_t sector_num = req->offset >> BDRV_SECTOR_BITS; 196 int nb_sectors = req->bytes >> BDRV_SECTOR_BITS; 197 198 assert((req->offset & (BDRV_SECTOR_SIZE - 1)) == 0); 199 assert((req->bytes & (BDRV_SECTOR_SIZE - 1)) == 0); 200 201 return backup_do_cow(req->bs, sector_num, nb_sectors, NULL, true); 202 } 203 204 static void backup_set_speed(BlockJob *job, int64_t speed, Error **errp) 205 { 206 BackupBlockJob *s = container_of(job, BackupBlockJob, common); 207 208 if (speed < 0) { 209 error_setg(errp, QERR_INVALID_PARAMETER, "speed"); 210 return; 211 } 212 ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME); 213 } 214 215 static void backup_iostatus_reset(BlockJob *job) 216 { 217 BackupBlockJob *s = container_of(job, BackupBlockJob, common); 218 219 if (s->target->blk) { 220 blk_iostatus_reset(s->target->blk); 221 } 222 } 223 224 static const BlockJobDriver backup_job_driver = { 225 .instance_size = sizeof(BackupBlockJob), 226 .job_type = BLOCK_JOB_TYPE_BACKUP, 227 .set_speed = backup_set_speed, 228 .iostatus_reset = backup_iostatus_reset, 229 }; 230 231 static BlockErrorAction backup_error_action(BackupBlockJob *job, 232 bool read, int error) 233 { 234 if (read) { 235 return block_job_error_action(&job->common, job->common.bs, 236 job->on_source_error, true, error); 237 } else { 238 return block_job_error_action(&job->common, job->target, 239 job->on_target_error, false, error); 240 } 241 } 242 243 typedef struct { 244 int ret; 245 } BackupCompleteData; 246 247 static void backup_complete(BlockJob *job, void *opaque) 248 { 249 BackupBlockJob *s = container_of(job, BackupBlockJob, common); 250 BackupCompleteData *data = opaque; 251 252 bdrv_unref(s->target); 253 254 block_job_completed(job, data->ret); 255 g_free(data); 256 } 257 258 static bool coroutine_fn yield_and_check(BackupBlockJob *job) 259 { 260 if (block_job_is_cancelled(&job->common)) { 261 return true; 262 } 263 264 /* we need to yield so that bdrv_drain_all() returns. 265 * (without, VM does not reboot) 266 */ 267 if (job->common.speed) { 268 uint64_t delay_ns = ratelimit_calculate_delay(&job->limit, 269 job->sectors_read); 270 job->sectors_read = 0; 271 block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, delay_ns); 272 } else { 273 block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, 0); 274 } 275 276 if (block_job_is_cancelled(&job->common)) { 277 return true; 278 } 279 280 return false; 281 } 282 283 static int coroutine_fn backup_run_incremental(BackupBlockJob *job) 284 { 285 bool error_is_read; 286 int ret = 0; 287 int clusters_per_iter; 288 uint32_t granularity; 289 int64_t sector; 290 int64_t cluster; 291 int64_t end; 292 int64_t last_cluster = -1; 293 BlockDriverState *bs = job->common.bs; 294 HBitmapIter hbi; 295 296 granularity = bdrv_dirty_bitmap_granularity(job->sync_bitmap); 297 clusters_per_iter = MAX((granularity / BACKUP_CLUSTER_SIZE), 1); 298 bdrv_dirty_iter_init(job->sync_bitmap, &hbi); 299 300 /* Find the next dirty sector(s) */ 301 while ((sector = hbitmap_iter_next(&hbi)) != -1) { 302 cluster = sector / BACKUP_SECTORS_PER_CLUSTER; 303 304 /* Fake progress updates for any clusters we skipped */ 305 if (cluster != last_cluster + 1) { 306 job->common.offset += ((cluster - last_cluster - 1) * 307 BACKUP_CLUSTER_SIZE); 308 } 309 310 for (end = cluster + clusters_per_iter; cluster < end; cluster++) { 311 do { 312 if (yield_and_check(job)) { 313 return ret; 314 } 315 ret = backup_do_cow(bs, cluster * BACKUP_SECTORS_PER_CLUSTER, 316 BACKUP_SECTORS_PER_CLUSTER, &error_is_read, 317 false); 318 if ((ret < 0) && 319 backup_error_action(job, error_is_read, -ret) == 320 BLOCK_ERROR_ACTION_REPORT) { 321 return ret; 322 } 323 } while (ret < 0); 324 } 325 326 /* If the bitmap granularity is smaller than the backup granularity, 327 * we need to advance the iterator pointer to the next cluster. */ 328 if (granularity < BACKUP_CLUSTER_SIZE) { 329 bdrv_set_dirty_iter(&hbi, cluster * BACKUP_SECTORS_PER_CLUSTER); 330 } 331 332 last_cluster = cluster - 1; 333 } 334 335 /* Play some final catchup with the progress meter */ 336 end = DIV_ROUND_UP(job->common.len, BACKUP_CLUSTER_SIZE); 337 if (last_cluster + 1 < end) { 338 job->common.offset += ((end - last_cluster - 1) * BACKUP_CLUSTER_SIZE); 339 } 340 341 return ret; 342 } 343 344 static void coroutine_fn backup_run(void *opaque) 345 { 346 BackupBlockJob *job = opaque; 347 BackupCompleteData *data; 348 BlockDriverState *bs = job->common.bs; 349 BlockDriverState *target = job->target; 350 BlockdevOnError on_target_error = job->on_target_error; 351 NotifierWithReturn before_write = { 352 .notify = backup_before_write_notify, 353 }; 354 int64_t start, end; 355 int ret = 0; 356 357 QLIST_INIT(&job->inflight_reqs); 358 qemu_co_rwlock_init(&job->flush_rwlock); 359 360 start = 0; 361 end = DIV_ROUND_UP(job->common.len, BACKUP_CLUSTER_SIZE); 362 363 job->bitmap = hbitmap_alloc(end, 0); 364 365 bdrv_set_enable_write_cache(target, true); 366 if (target->blk) { 367 blk_set_on_error(target->blk, on_target_error, on_target_error); 368 blk_iostatus_enable(target->blk); 369 } 370 371 bdrv_add_before_write_notifier(bs, &before_write); 372 373 if (job->sync_mode == MIRROR_SYNC_MODE_NONE) { 374 while (!block_job_is_cancelled(&job->common)) { 375 /* Yield until the job is cancelled. We just let our before_write 376 * notify callback service CoW requests. */ 377 job->common.busy = false; 378 qemu_coroutine_yield(); 379 job->common.busy = true; 380 } 381 } else if (job->sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) { 382 ret = backup_run_incremental(job); 383 } else { 384 /* Both FULL and TOP SYNC_MODE's require copying.. */ 385 for (; start < end; start++) { 386 bool error_is_read; 387 if (yield_and_check(job)) { 388 break; 389 } 390 391 if (job->sync_mode == MIRROR_SYNC_MODE_TOP) { 392 int i, n; 393 int alloced = 0; 394 395 /* Check to see if these blocks are already in the 396 * backing file. */ 397 398 for (i = 0; i < BACKUP_SECTORS_PER_CLUSTER;) { 399 /* bdrv_is_allocated() only returns true/false based 400 * on the first set of sectors it comes across that 401 * are are all in the same state. 402 * For that reason we must verify each sector in the 403 * backup cluster length. We end up copying more than 404 * needed but at some point that is always the case. */ 405 alloced = 406 bdrv_is_allocated(bs, 407 start * BACKUP_SECTORS_PER_CLUSTER + i, 408 BACKUP_SECTORS_PER_CLUSTER - i, &n); 409 i += n; 410 411 if (alloced == 1 || n == 0) { 412 break; 413 } 414 } 415 416 /* If the above loop never found any sectors that are in 417 * the topmost image, skip this backup. */ 418 if (alloced == 0) { 419 continue; 420 } 421 } 422 /* FULL sync mode we copy the whole drive. */ 423 ret = backup_do_cow(bs, start * BACKUP_SECTORS_PER_CLUSTER, 424 BACKUP_SECTORS_PER_CLUSTER, &error_is_read, false); 425 if (ret < 0) { 426 /* Depending on error action, fail now or retry cluster */ 427 BlockErrorAction action = 428 backup_error_action(job, error_is_read, -ret); 429 if (action == BLOCK_ERROR_ACTION_REPORT) { 430 break; 431 } else { 432 start--; 433 continue; 434 } 435 } 436 } 437 } 438 439 notifier_with_return_remove(&before_write); 440 441 /* wait until pending backup_do_cow() calls have completed */ 442 qemu_co_rwlock_wrlock(&job->flush_rwlock); 443 qemu_co_rwlock_unlock(&job->flush_rwlock); 444 445 if (job->sync_bitmap) { 446 BdrvDirtyBitmap *bm; 447 if (ret < 0 || block_job_is_cancelled(&job->common)) { 448 /* Merge the successor back into the parent, delete nothing. */ 449 bm = bdrv_reclaim_dirty_bitmap(bs, job->sync_bitmap, NULL); 450 assert(bm); 451 } else { 452 /* Everything is fine, delete this bitmap and install the backup. */ 453 bm = bdrv_dirty_bitmap_abdicate(bs, job->sync_bitmap, NULL); 454 assert(bm); 455 } 456 } 457 hbitmap_free(job->bitmap); 458 459 if (target->blk) { 460 blk_iostatus_disable(target->blk); 461 } 462 bdrv_op_unblock_all(target, job->common.blocker); 463 464 data = g_malloc(sizeof(*data)); 465 data->ret = ret; 466 block_job_defer_to_main_loop(&job->common, backup_complete, data); 467 } 468 469 void backup_start(BlockDriverState *bs, BlockDriverState *target, 470 int64_t speed, MirrorSyncMode sync_mode, 471 BdrvDirtyBitmap *sync_bitmap, 472 BlockdevOnError on_source_error, 473 BlockdevOnError on_target_error, 474 BlockCompletionFunc *cb, void *opaque, 475 Error **errp) 476 { 477 int64_t len; 478 479 assert(bs); 480 assert(target); 481 assert(cb); 482 483 if (bs == target) { 484 error_setg(errp, "Source and target cannot be the same"); 485 return; 486 } 487 488 if ((on_source_error == BLOCKDEV_ON_ERROR_STOP || 489 on_source_error == BLOCKDEV_ON_ERROR_ENOSPC) && 490 (!bs->blk || !blk_iostatus_is_enabled(bs->blk))) { 491 error_setg(errp, QERR_INVALID_PARAMETER, "on-source-error"); 492 return; 493 } 494 495 if (!bdrv_is_inserted(bs)) { 496 error_setg(errp, "Device is not inserted: %s", 497 bdrv_get_device_name(bs)); 498 return; 499 } 500 501 if (!bdrv_is_inserted(target)) { 502 error_setg(errp, "Device is not inserted: %s", 503 bdrv_get_device_name(target)); 504 return; 505 } 506 507 if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_BACKUP_SOURCE, errp)) { 508 return; 509 } 510 511 if (bdrv_op_is_blocked(target, BLOCK_OP_TYPE_BACKUP_TARGET, errp)) { 512 return; 513 } 514 515 if (sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) { 516 if (!sync_bitmap) { 517 error_setg(errp, "must provide a valid bitmap name for " 518 "\"incremental\" sync mode"); 519 return; 520 } 521 522 /* Create a new bitmap, and freeze/disable this one. */ 523 if (bdrv_dirty_bitmap_create_successor(bs, sync_bitmap, errp) < 0) { 524 return; 525 } 526 } else if (sync_bitmap) { 527 error_setg(errp, 528 "a sync_bitmap was provided to backup_run, " 529 "but received an incompatible sync_mode (%s)", 530 MirrorSyncMode_lookup[sync_mode]); 531 return; 532 } 533 534 len = bdrv_getlength(bs); 535 if (len < 0) { 536 error_setg_errno(errp, -len, "unable to get length for '%s'", 537 bdrv_get_device_name(bs)); 538 goto error; 539 } 540 541 BackupBlockJob *job = block_job_create(&backup_job_driver, bs, speed, 542 cb, opaque, errp); 543 if (!job) { 544 goto error; 545 } 546 547 bdrv_op_block_all(target, job->common.blocker); 548 549 job->on_source_error = on_source_error; 550 job->on_target_error = on_target_error; 551 job->target = target; 552 job->sync_mode = sync_mode; 553 job->sync_bitmap = sync_mode == MIRROR_SYNC_MODE_INCREMENTAL ? 554 sync_bitmap : NULL; 555 job->common.len = len; 556 job->common.co = qemu_coroutine_create(backup_run); 557 qemu_coroutine_enter(job->common.co, job); 558 return; 559 560 error: 561 if (sync_bitmap) { 562 bdrv_reclaim_dirty_bitmap(bs, sync_bitmap, NULL); 563 } 564 } 565