block/backup.c

   1 /*
   2  * QEMU backup
   3  *
   4  * Copyright (C) 2013 Proxmox Server Solutions
   5  *
   6  * Authors:
   7  *  Dietmar Maurer ([email protected])
   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 "qemu/osdep.h"
  15
  16 #include "trace.h"
  17 #include "block/block.h"
  18 #include "block/block_int.h"
  19 #include "block/blockjob.h"
  20 #include "qapi/error.h"
  21 #include "qapi/qmp/qerror.h"
  22 #include "qemu/ratelimit.h"
  23 #include "qemu/cutils.h"
  24 #include "sysemu/block-backend.h"
  25 #include "qemu/bitmap.h"
  26
  27 #define BACKUP_CLUSTER_SIZE_DEFAULT (1 << 16)
  28 #define SLICE_TIME 100000000ULL /* ns */
  29
  30 typedef struct CowRequest {
  31     int64_t start;
  32     int64_t end;
  33     QLIST_ENTRY(CowRequest) list;
  34     CoQueue wait_queue; /* coroutines blocked on this request */
  35 } CowRequest;
  36
  37 typedef struct BackupBlockJob {
  38     BlockJob common;
  39     BlockDriverState *target;
  40     /* bitmap for sync=incremental */
  41     BdrvDirtyBitmap *sync_bitmap;
  42     MirrorSyncMode sync_mode;
  43     RateLimit limit;
  44     BlockdevOnError on_source_error;
  45     BlockdevOnError on_target_error;
  46     CoRwlock flush_rwlock;
  47     uint64_t sectors_read;
  48     unsigned long *done_bitmap;
  49     int64_t cluster_size;
  50     NotifierWithReturn before_write;
  51     QLIST_HEAD(, CowRequest) inflight_reqs;
  52 } BackupBlockJob;
  53
  54 /* Size of a cluster in sectors, instead of bytes. */
  55 static inline int64_t cluster_size_sectors(BackupBlockJob *job)
  56 {
  57   return job->cluster_size / BDRV_SECTOR_SIZE;
  58 }
  59
  60 /* See if in-flight requests overlap and wait for them to complete */
  61 static void coroutine_fn wait_for_overlapping_requests(BackupBlockJob *job,
  62                                                        int64_t start,
  63                                                        int64_t end)
  64 {
  65     CowRequest *req;
  66     bool retry;
  67
  68     do {
  69         retry = false;
  70         QLIST_FOREACH(req, &job->inflight_reqs, list) {
  71             if (end > req->start && start < req->end) {
  72                 qemu_co_queue_wait(&req->wait_queue);
  73                 retry = true;
  74                 break;
  75             }
  76         }
  77     } while (retry);
  78 }
  79
  80 /* Keep track of an in-flight request */
  81 static void cow_request_begin(CowRequest *req, BackupBlockJob *job,
  82                                      int64_t start, int64_t end)
  83 {
  84     req->start = start;
  85     req->end = end;
  86     qemu_co_queue_init(&req->wait_queue);
  87     QLIST_INSERT_HEAD(&job->inflight_reqs, req, list);
  88 }
  89
  90 /* Forget about a completed request */
  91 static void cow_request_end(CowRequest *req)
  92 {
  93     QLIST_REMOVE(req, list);
  94     qemu_co_queue_restart_all(&req->wait_queue);
  95 }
  96
  97 static int coroutine_fn backup_do_cow(BackupBlockJob *job,
  98                                       int64_t sector_num, int nb_sectors,
  99                                       bool *error_is_read,
 100                                       bool is_write_notifier)
 101 {
 102     BlockDriverState *bs = job->common.bs;
 103     CowRequest cow_request;
 104     struct iovec iov;
 105     QEMUIOVector bounce_qiov;
 106     void *bounce_buffer = NULL;
 107     int ret = 0;
 108     int64_t sectors_per_cluster = cluster_size_sectors(job);
 109     int64_t start, end;
 110     int n;
 111
 112     qemu_co_rwlock_rdlock(&job->flush_rwlock);
 113
 114     start = sector_num / sectors_per_cluster;
 115     end = DIV_ROUND_UP(sector_num + nb_sectors, sectors_per_cluster);
 116
 117     trace_backup_do_cow_enter(job, start, sector_num, nb_sectors);
 118
 119     wait_for_overlapping_requests(job, start, end);
 120     cow_request_begin(&cow_request, job, start, end);
 121
 122     for (; start < end; start++) {
 123         if (test_bit(start, job->done_bitmap)) {
 124             trace_backup_do_cow_skip(job, start);
 125             continue; /* already copied */
 126         }
 127
 128         trace_backup_do_cow_process(job, start);
 129
 130         n = MIN(sectors_per_cluster,
 131                 job->common.len / BDRV_SECTOR_SIZE -
 132                 start * sectors_per_cluster);
 133
 134         if (!bounce_buffer) {
 135             bounce_buffer = qemu_blockalign(bs, job->cluster_size);
 136         }
 137         iov.iov_base = bounce_buffer;
 138         iov.iov_len = n * BDRV_SECTOR_SIZE;
 139         qemu_iovec_init_external(&bounce_qiov, &iov, 1);
 140
 141         if (is_write_notifier) {
 142             ret = bdrv_co_readv_no_serialising(bs,
 143                                            start * sectors_per_cluster,
 144                                            n, &bounce_qiov);
 145         } else {
 146             ret = bdrv_co_readv(bs, start * sectors_per_cluster, n,
 147                                 &bounce_qiov);
 148         }
 149         if (ret < 0) {
 150             trace_backup_do_cow_read_fail(job, start, ret);
 151             if (error_is_read) {
 152                 *error_is_read = true;
 153             }
 154             goto out;
 155         }
 156
 157         if (buffer_is_zero(iov.iov_base, iov.iov_len)) {
 158             ret = bdrv_co_write_zeroes(job->target,
 159                                        start * sectors_per_cluster,
 160                                        n, BDRV_REQ_MAY_UNMAP);
 161         } else {
 162             ret = bdrv_co_writev(job->target,
 163                                  start * sectors_per_cluster, n,
 164                                  &bounce_qiov);
 165         }
 166         if (ret < 0) {
 167             trace_backup_do_cow_write_fail(job, start, ret);
 168             if (error_is_read) {
 169                 *error_is_read = false;
 170             }
 171             goto out;
 172         }
 173
 174         set_bit(start, job->done_bitmap);
 175
 176         /* Publish progress, guest I/O counts as progress too.  Note that the
 177          * offset field is an opaque progress value, it is not a disk offset.
 178          */
 179         job->sectors_read += n;
 180         job->common.offset += n * BDRV_SECTOR_SIZE;
 181     }
 182
 183 out:
 184     if (bounce_buffer) {
 185         qemu_vfree(bounce_buffer);
 186     }
 187
 188     cow_request_end(&cow_request);
 189
 190     trace_backup_do_cow_return(job, sector_num, nb_sectors, ret);
 191
 192     qemu_co_rwlock_unlock(&job->flush_rwlock);
 193
 194     return ret;
 195 }
 196
 197 static int coroutine_fn backup_before_write_notify(
 198         NotifierWithReturn *notifier,
 199         void *opaque)
 200 {
 201     BackupBlockJob *job = container_of(notifier, BackupBlockJob, before_write);
 202     BdrvTrackedRequest *req = opaque;
 203     int64_t sector_num = req->offset >> BDRV_SECTOR_BITS;
 204     int nb_sectors = req->bytes >> BDRV_SECTOR_BITS;
 205
 206     assert(req->bs == job->common.bs);
 207     assert((req->offset & (BDRV_SECTOR_SIZE - 1)) == 0);
 208     assert((req->bytes & (BDRV_SECTOR_SIZE - 1)) == 0);
 209
 210     return backup_do_cow(job, sector_num, nb_sectors, NULL, true);
 211 }
 212
 213 static void backup_set_speed(BlockJob *job, int64_t speed, Error **errp)
 214 {
 215     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 216
 217     if (speed < 0) {
 218         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 219         return;
 220     }
 221     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 222 }
 223
 224 static void backup_cleanup_sync_bitmap(BackupBlockJob *job, int ret)
 225 {
 226     BdrvDirtyBitmap *bm;
 227     BlockDriverState *bs = job->common.bs;
 228
 229     if (ret < 0 || block_job_is_cancelled(&job->common)) {
 230         /* Merge the successor back into the parent, delete nothing. */
 231         bm = bdrv_reclaim_dirty_bitmap(bs, job->sync_bitmap, NULL);
 232         assert(bm);
 233     } else {
 234         /* Everything is fine, delete this bitmap and install the backup. */
 235         bm = bdrv_dirty_bitmap_abdicate(bs, job->sync_bitmap, NULL);
 236         assert(bm);
 237     }
 238 }
 239
 240 static void backup_commit(BlockJob *job)
 241 {
 242     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 243     if (s->sync_bitmap) {
 244         backup_cleanup_sync_bitmap(s, 0);
 245     }
 246 }
 247
 248 static void backup_abort(BlockJob *job)
 249 {
 250     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 251     if (s->sync_bitmap) {
 252         backup_cleanup_sync_bitmap(s, -1);
 253     }
 254 }
 255
 256 static const BlockJobDriver backup_job_driver = {
 257     .instance_size  = sizeof(BackupBlockJob),
 258     .job_type       = BLOCK_JOB_TYPE_BACKUP,
 259     .set_speed      = backup_set_speed,
 260     .commit         = backup_commit,
 261     .abort          = backup_abort,
 262 };
 263
 264 static BlockErrorAction backup_error_action(BackupBlockJob *job,
 265                                             bool read, int error)
 266 {
 267     if (read) {
 268         return block_job_error_action(&job->common, job->on_source_error,
 269                                       true, error);
 270     } else {
 271         return block_job_error_action(&job->common, job->on_target_error,
 272                                       false, error);
 273     }
 274 }
 275
 276 typedef struct {
 277     int ret;
 278 } BackupCompleteData;
 279
 280 static void backup_complete(BlockJob *job, void *opaque)
 281 {
 282     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 283     BackupCompleteData *data = opaque;
 284
 285     bdrv_unref(s->target);
 286
 287     block_job_completed(job, data->ret);
 288     g_free(data);
 289 }
 290
 291 static bool coroutine_fn yield_and_check(BackupBlockJob *job)
 292 {
 293     if (block_job_is_cancelled(&job->common)) {
 294         return true;
 295     }
 296
 297     /* we need to yield so that bdrv_drain_all() returns.
 298      * (without, VM does not reboot)
 299      */
 300     if (job->common.speed) {
 301         uint64_t delay_ns = ratelimit_calculate_delay(&job->limit,
 302                                                       job->sectors_read);
 303         job->sectors_read = 0;
 304         block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, delay_ns);
 305     } else {
 306         block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, 0);
 307     }
 308
 309     if (block_job_is_cancelled(&job->common)) {
 310         return true;
 311     }
 312
 313     return false;
 314 }
 315
 316 static int coroutine_fn backup_run_incremental(BackupBlockJob *job)
 317 {
 318     bool error_is_read;
 319     int ret = 0;
 320     int clusters_per_iter;
 321     uint32_t granularity;
 322     int64_t sector;
 323     int64_t cluster;
 324     int64_t end;
 325     int64_t last_cluster = -1;
 326     int64_t sectors_per_cluster = cluster_size_sectors(job);
 327     HBitmapIter hbi;
 328
 329     granularity = bdrv_dirty_bitmap_granularity(job->sync_bitmap);
 330     clusters_per_iter = MAX((granularity / job->cluster_size), 1);
 331     bdrv_dirty_iter_init(job->sync_bitmap, &hbi);
 332
 333     /* Find the next dirty sector(s) */
 334     while ((sector = hbitmap_iter_next(&hbi)) != -1) {
 335         cluster = sector / sectors_per_cluster;
 336
 337         /* Fake progress updates for any clusters we skipped */
 338         if (cluster != last_cluster + 1) {
 339             job->common.offset += ((cluster - last_cluster - 1) *
 340                                    job->cluster_size);
 341         }
 342
 343         for (end = cluster + clusters_per_iter; cluster < end; cluster++) {
 344             do {
 345                 if (yield_and_check(job)) {
 346                     return ret;
 347                 }
 348                 ret = backup_do_cow(job, cluster * sectors_per_cluster,
 349                                     sectors_per_cluster, &error_is_read,
 350                                     false);
 351                 if ((ret < 0) &&
 352                     backup_error_action(job, error_is_read, -ret) ==
 353                     BLOCK_ERROR_ACTION_REPORT) {
 354                     return ret;
 355                 }
 356             } while (ret < 0);
 357         }
 358
 359         /* If the bitmap granularity is smaller than the backup granularity,
 360          * we need to advance the iterator pointer to the next cluster. */
 361         if (granularity < job->cluster_size) {
 362             bdrv_set_dirty_iter(&hbi, cluster * sectors_per_cluster);
 363         }
 364
 365         last_cluster = cluster - 1;
 366     }
 367
 368     /* Play some final catchup with the progress meter */
 369     end = DIV_ROUND_UP(job->common.len, job->cluster_size);
 370     if (last_cluster + 1 < end) {
 371         job->common.offset += ((end - last_cluster - 1) * job->cluster_size);
 372     }
 373
 374     return ret;
 375 }
 376
 377 static void coroutine_fn backup_run(void *opaque)
 378 {
 379     BackupBlockJob *job = opaque;
 380     BackupCompleteData *data;
 381     BlockDriverState *bs = job->common.bs;
 382     BlockDriverState *target = job->target;
 383     int64_t start, end;
 384     int64_t sectors_per_cluster = cluster_size_sectors(job);
 385     int ret = 0;
 386
 387     QLIST_INIT(&job->inflight_reqs);
 388     qemu_co_rwlock_init(&job->flush_rwlock);
 389
 390     start = 0;
 391     end = DIV_ROUND_UP(job->common.len, job->cluster_size);
 392
 393     job->done_bitmap = bitmap_new(end);
 394
 395     job->before_write.notify = backup_before_write_notify;
 396     bdrv_add_before_write_notifier(bs, &job->before_write);
 397
 398     if (job->sync_mode == MIRROR_SYNC_MODE_NONE) {
 399         while (!block_job_is_cancelled(&job->common)) {
 400             /* Yield until the job is cancelled.  We just let our before_write
 401              * notify callback service CoW requests. */
 402             job->common.busy = false;
 403             qemu_coroutine_yield();
 404             job->common.busy = true;
 405         }
 406     } else if (job->sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
 407         ret = backup_run_incremental(job);
 408     } else {
 409         /* Both FULL and TOP SYNC_MODE's require copying.. */
 410         for (; start < end; start++) {
 411             bool error_is_read;
 412             if (yield_and_check(job)) {
 413                 break;
 414             }
 415
 416             if (job->sync_mode == MIRROR_SYNC_MODE_TOP) {
 417                 int i, n;
 418                 int alloced = 0;
 419
 420                 /* Check to see if these blocks are already in the
 421                  * backing file. */
 422
 423                 for (i = 0; i < sectors_per_cluster;) {
 424                     /* bdrv_is_allocated() only returns true/false based
 425                      * on the first set of sectors it comes across that
 426                      * are are all in the same state.
 427                      * For that reason we must verify each sector in the
 428                      * backup cluster length.  We end up copying more than
 429                      * needed but at some point that is always the case. */
 430                     alloced =
 431                         bdrv_is_allocated(bs,
 432                                 start * sectors_per_cluster + i,
 433                                 sectors_per_cluster - i, &n);
 434                     i += n;
 435
 436                     if (alloced == 1 || n == 0) {
 437                         break;
 438                     }
 439                 }
 440
 441                 /* If the above loop never found any sectors that are in
 442                  * the topmost image, skip this backup. */
 443                 if (alloced == 0) {
 444                     continue;
 445                 }
 446             }
 447             /* FULL sync mode we copy the whole drive. */
 448             ret = backup_do_cow(job, start * sectors_per_cluster,
 449                                 sectors_per_cluster, &error_is_read, false);
 450             if (ret < 0) {
 451                 /* Depending on error action, fail now or retry cluster */
 452                 BlockErrorAction action =
 453                     backup_error_action(job, error_is_read, -ret);
 454                 if (action == BLOCK_ERROR_ACTION_REPORT) {
 455                     break;
 456                 } else {
 457                     start--;
 458                     continue;
 459                 }
 460             }
 461         }
 462     }
 463
 464     notifier_with_return_remove(&job->before_write);
 465
 466     /* wait until pending backup_do_cow() calls have completed */
 467     qemu_co_rwlock_wrlock(&job->flush_rwlock);
 468     qemu_co_rwlock_unlock(&job->flush_rwlock);
 469     g_free(job->done_bitmap);
 470
 471     bdrv_op_unblock_all(target, job->common.blocker);
 472
 473     data = g_malloc(sizeof(*data));
 474     data->ret = ret;
 475     block_job_defer_to_main_loop(&job->common, backup_complete, data);
 476 }
 477
 478 void backup_start(BlockDriverState *bs, BlockDriverState *target,
 479                   int64_t speed, MirrorSyncMode sync_mode,
 480                   BdrvDirtyBitmap *sync_bitmap,
 481                   BlockdevOnError on_source_error,
 482                   BlockdevOnError on_target_error,
 483                   BlockCompletionFunc *cb, void *opaque,
 484                   BlockJobTxn *txn, Error **errp)
 485 {
 486     int64_t len;
 487     BlockDriverInfo bdi;
 488     BackupBlockJob *job = NULL;
 489     int ret;
 490
 491     assert(bs);
 492     assert(target);
 493     assert(cb);
 494
 495     if (bs == target) {
 496         error_setg(errp, "Source and target cannot be the same");
 497         return;
 498     }
 499
 500     if (!bdrv_is_inserted(bs)) {
 501         error_setg(errp, "Device is not inserted: %s",
 502                    bdrv_get_device_name(bs));
 503         return;
 504     }
 505
 506     if (!bdrv_is_inserted(target)) {
 507         error_setg(errp, "Device is not inserted: %s",
 508                    bdrv_get_device_name(target));
 509         return;
 510     }
 511
 512     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_BACKUP_SOURCE, errp)) {
 513         return;
 514     }
 515
 516     if (bdrv_op_is_blocked(target, BLOCK_OP_TYPE_BACKUP_TARGET, errp)) {
 517         return;
 518     }
 519
 520     if (sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
 521         if (!sync_bitmap) {
 522             error_setg(errp, "must provide a valid bitmap name for "
 523                              "\"incremental\" sync mode");
 524             return;
 525         }
 526
 527         /* Create a new bitmap, and freeze/disable this one. */
 528         if (bdrv_dirty_bitmap_create_successor(bs, sync_bitmap, errp) < 0) {
 529             return;
 530         }
 531     } else if (sync_bitmap) {
 532         error_setg(errp,
 533                    "a sync_bitmap was provided to backup_run, "
 534                    "but received an incompatible sync_mode (%s)",
 535                    MirrorSyncMode_lookup[sync_mode]);
 536         return;
 537     }
 538
 539     len = bdrv_getlength(bs);
 540     if (len < 0) {
 541         error_setg_errno(errp, -len, "unable to get length for '%s'",
 542                          bdrv_get_device_name(bs));
 543         goto error;
 544     }
 545
 546     job = block_job_create(&backup_job_driver, bs, speed, cb, opaque, errp);
 547     if (!job) {
 548         goto error;
 549     }
 550
 551     job->on_source_error = on_source_error;
 552     job->on_target_error = on_target_error;
 553     job->target = target;
 554     job->sync_mode = sync_mode;
 555     job->sync_bitmap = sync_mode == MIRROR_SYNC_MODE_INCREMENTAL ?
 556                        sync_bitmap : NULL;
 557
 558     /* If there is no backing file on the target, we cannot rely on COW if our
 559      * backup cluster size is smaller than the target cluster size. Even for
 560      * targets with a backing file, try to avoid COW if possible. */
 561     ret = bdrv_get_info(job->target, &bdi);
 562     if (ret < 0 && !target->backing) {
 563         error_setg_errno(errp, -ret,
 564             "Couldn't determine the cluster size of the target image, "
 565             "which has no backing file");
 566         error_append_hint(errp,
 567             "Aborting, since this may create an unusable destination image\n");
 568         goto error;
 569     } else if (ret < 0 && target->backing) {
 570         /* Not fatal; just trudge on ahead. */
 571         job->cluster_size = BACKUP_CLUSTER_SIZE_DEFAULT;
 572     } else {
 573         job->cluster_size = MAX(BACKUP_CLUSTER_SIZE_DEFAULT, bdi.cluster_size);
 574     }
 575
 576     bdrv_op_block_all(target, job->common.blocker);
 577     job->common.len = len;
 578     job->common.co = qemu_coroutine_create(backup_run);
 579     block_job_txn_add_job(txn, &job->common);
 580     qemu_coroutine_enter(job->common.co, job);
 581     return;
 582
 583  error:
 584     if (sync_bitmap) {
 585         bdrv_reclaim_dirty_bitmap(bs, sync_bitmap, NULL);
 586     }
 587     if (job) {
 588         block_job_unref(&job->common);
 589     }
 590 }