block/backup.c

   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 "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     QLIST_HEAD(, CowRequest) inflight_reqs;
  51 } BackupBlockJob;
  52
  53 /* Size of a cluster in sectors, instead of bytes. */
  54 static inline int64_t cluster_size_sectors(BackupBlockJob *job)
  55 {
  56   return job->cluster_size / BDRV_SECTOR_SIZE;
  57 }
  58
  59 /* See if in-flight requests overlap and wait for them to complete */
  60 static void coroutine_fn wait_for_overlapping_requests(BackupBlockJob *job,
  61                                                        int64_t start,
  62                                                        int64_t end)
  63 {
  64     CowRequest *req;
  65     bool retry;
  66
  67     do {
  68         retry = false;
  69         QLIST_FOREACH(req, &job->inflight_reqs, list) {
  70             if (end > req->start && start < req->end) {
  71                 qemu_co_queue_wait(&req->wait_queue);
  72                 retry = true;
  73                 break;
  74             }
  75         }
  76     } while (retry);
  77 }
  78
  79 /* Keep track of an in-flight request */
  80 static void cow_request_begin(CowRequest *req, BackupBlockJob *job,
  81                                      int64_t start, int64_t end)
  82 {
  83     req->start = start;
  84     req->end = end;
  85     qemu_co_queue_init(&req->wait_queue);
  86     QLIST_INSERT_HEAD(&job->inflight_reqs, req, list);
  87 }
  88
  89 /* Forget about a completed request */
  90 static void cow_request_end(CowRequest *req)
  91 {
  92     QLIST_REMOVE(req, list);
  93     qemu_co_queue_restart_all(&req->wait_queue);
  94 }
  95
  96 static int coroutine_fn backup_do_cow(BlockDriverState *bs,
  97                                       int64_t sector_num, int nb_sectors,
  98                                       bool *error_is_read,
  99                                       bool is_write_notifier)
 100 {
 101     BackupBlockJob *job = (BackupBlockJob *)bs->job;
 102     CowRequest cow_request;
 103     struct iovec iov;
 104     QEMUIOVector bounce_qiov;
 105     void *bounce_buffer = NULL;
 106     int ret = 0;
 107     int64_t sectors_per_cluster = cluster_size_sectors(job);
 108     int64_t start, end;
 109     int n;
 110
 111     qemu_co_rwlock_rdlock(&job->flush_rwlock);
 112
 113     start = sector_num / sectors_per_cluster;
 114     end = DIV_ROUND_UP(sector_num + nb_sectors, sectors_per_cluster);
 115
 116     trace_backup_do_cow_enter(job, start, sector_num, nb_sectors);
 117
 118     wait_for_overlapping_requests(job, start, end);
 119     cow_request_begin(&cow_request, job, start, end);
 120
 121     for (; start < end; start++) {
 122         if (test_bit(start, job->done_bitmap)) {
 123             trace_backup_do_cow_skip(job, start);
 124             continue; /* already copied */
 125         }
 126
 127         trace_backup_do_cow_process(job, start);
 128
 129         n = MIN(sectors_per_cluster,
 130                 job->common.len / BDRV_SECTOR_SIZE -
 131                 start * sectors_per_cluster);
 132
 133         if (!bounce_buffer) {
 134             bounce_buffer = qemu_blockalign(bs, job->cluster_size);
 135         }
 136         iov.iov_base = bounce_buffer;
 137         iov.iov_len = n * BDRV_SECTOR_SIZE;
 138         qemu_iovec_init_external(&bounce_qiov, &iov, 1);
 139
 140         if (is_write_notifier) {
 141             ret = bdrv_co_readv_no_serialising(bs,
 142                                            start * sectors_per_cluster,
 143                                            n, &bounce_qiov);
 144         } else {
 145             ret = bdrv_co_readv(bs, start * sectors_per_cluster, n,
 146                                 &bounce_qiov);
 147         }
 148         if (ret < 0) {
 149             trace_backup_do_cow_read_fail(job, start, ret);
 150             if (error_is_read) {
 151                 *error_is_read = true;
 152             }
 153             goto out;
 154         }
 155
 156         if (buffer_is_zero(iov.iov_base, iov.iov_len)) {
 157             ret = bdrv_co_write_zeroes(job->target,
 158                                        start * sectors_per_cluster,
 159                                        n, BDRV_REQ_MAY_UNMAP);
 160         } else {
 161             ret = bdrv_co_writev(job->target,
 162                                  start * sectors_per_cluster, n,
 163                                  &bounce_qiov);
 164         }
 165         if (ret < 0) {
 166             trace_backup_do_cow_write_fail(job, start, ret);
 167             if (error_is_read) {
 168                 *error_is_read = false;
 169             }
 170             goto out;
 171         }
 172
 173         set_bit(start, job->done_bitmap);
 174
 175         /* Publish progress, guest I/O counts as progress too.  Note that the
 176          * offset field is an opaque progress value, it is not a disk offset.
 177          */
 178         job->sectors_read += n;
 179         job->common.offset += n * BDRV_SECTOR_SIZE;
 180     }
 181
 182 out:
 183     if (bounce_buffer) {
 184         qemu_vfree(bounce_buffer);
 185     }
 186
 187     cow_request_end(&cow_request);
 188
 189     trace_backup_do_cow_return(job, sector_num, nb_sectors, ret);
 190
 191     qemu_co_rwlock_unlock(&job->flush_rwlock);
 192
 193     return ret;
 194 }
 195
 196 static int coroutine_fn backup_before_write_notify(
 197         NotifierWithReturn *notifier,
 198         void *opaque)
 199 {
 200     BdrvTrackedRequest *req = opaque;
 201     int64_t sector_num = req->offset >> BDRV_SECTOR_BITS;
 202     int nb_sectors = req->bytes >> BDRV_SECTOR_BITS;
 203
 204     assert((req->offset & (BDRV_SECTOR_SIZE - 1)) == 0);
 205     assert((req->bytes & (BDRV_SECTOR_SIZE - 1)) == 0);
 206
 207     return backup_do_cow(req->bs, sector_num, nb_sectors, NULL, true);
 208 }
 209
 210 static void backup_set_speed(BlockJob *job, int64_t speed, Error **errp)
 211 {
 212     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 213
 214     if (speed < 0) {
 215         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 216         return;
 217     }
 218     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 219 }
 220
 221 static void backup_cleanup_sync_bitmap(BackupBlockJob *job, int ret)
 222 {
 223     BdrvDirtyBitmap *bm;
 224     BlockDriverState *bs = job->common.bs;
 225
 226     if (ret < 0 || block_job_is_cancelled(&job->common)) {
 227         /* Merge the successor back into the parent, delete nothing. */
 228         bm = bdrv_reclaim_dirty_bitmap(bs, job->sync_bitmap, NULL);
 229         assert(bm);
 230     } else {
 231         /* Everything is fine, delete this bitmap and install the backup. */
 232         bm = bdrv_dirty_bitmap_abdicate(bs, job->sync_bitmap, NULL);
 233         assert(bm);
 234     }
 235 }
 236
 237 static void backup_commit(BlockJob *job)
 238 {
 239     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 240     if (s->sync_bitmap) {
 241         backup_cleanup_sync_bitmap(s, 0);
 242     }
 243 }
 244
 245 static void backup_abort(BlockJob *job)
 246 {
 247     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 248     if (s->sync_bitmap) {
 249         backup_cleanup_sync_bitmap(s, -1);
 250     }
 251 }
 252
 253 static const BlockJobDriver backup_job_driver = {
 254     .instance_size  = sizeof(BackupBlockJob),
 255     .job_type       = BLOCK_JOB_TYPE_BACKUP,
 256     .set_speed      = backup_set_speed,
 257     .commit         = backup_commit,
 258     .abort          = backup_abort,
 259 };
 260
 261 static BlockErrorAction backup_error_action(BackupBlockJob *job,
 262                                             bool read, int error)
 263 {
 264     if (read) {
 265         return block_job_error_action(&job->common, job->on_source_error,
 266                                       true, error);
 267     } else {
 268         return block_job_error_action(&job->common, job->on_target_error,
 269                                       false, error);
 270     }
 271 }
 272
 273 typedef struct {
 274     int ret;
 275 } BackupCompleteData;
 276
 277 static void backup_complete(BlockJob *job, void *opaque)
 278 {
 279     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 280     BackupCompleteData *data = opaque;
 281
 282     bdrv_unref(s->target);
 283
 284     block_job_completed(job, data->ret);
 285     g_free(data);
 286 }
 287
 288 static bool coroutine_fn yield_and_check(BackupBlockJob *job)
 289 {
 290     if (block_job_is_cancelled(&job->common)) {
 291         return true;
 292     }
 293
 294     /* we need to yield so that bdrv_drain_all() returns.
 295      * (without, VM does not reboot)
 296      */
 297     if (job->common.speed) {
 298         uint64_t delay_ns = ratelimit_calculate_delay(&job->limit,
 299                                                       job->sectors_read);
 300         job->sectors_read = 0;
 301         block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, delay_ns);
 302     } else {
 303         block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, 0);
 304     }
 305
 306     if (block_job_is_cancelled(&job->common)) {
 307         return true;
 308     }
 309
 310     return false;
 311 }
 312
 313 static int coroutine_fn backup_run_incremental(BackupBlockJob *job)
 314 {
 315     bool error_is_read;
 316     int ret = 0;
 317     int clusters_per_iter;
 318     uint32_t granularity;
 319     int64_t sector;
 320     int64_t cluster;
 321     int64_t end;
 322     int64_t last_cluster = -1;
 323     int64_t sectors_per_cluster = cluster_size_sectors(job);
 324     BlockDriverState *bs = job->common.bs;
 325     HBitmapIter hbi;
 326
 327     granularity = bdrv_dirty_bitmap_granularity(job->sync_bitmap);
 328     clusters_per_iter = MAX((granularity / job->cluster_size), 1);
 329     bdrv_dirty_iter_init(job->sync_bitmap, &hbi);
 330
 331     /* Find the next dirty sector(s) */
 332     while ((sector = hbitmap_iter_next(&hbi)) != -1) {
 333         cluster = sector / sectors_per_cluster;
 334
 335         /* Fake progress updates for any clusters we skipped */
 336         if (cluster != last_cluster + 1) {
 337             job->common.offset += ((cluster - last_cluster - 1) *
 338                                    job->cluster_size);
 339         }
 340
 341         for (end = cluster + clusters_per_iter; cluster < end; cluster++) {
 342             do {
 343                 if (yield_and_check(job)) {
 344                     return ret;
 345                 }
 346                 ret = backup_do_cow(bs, cluster * sectors_per_cluster,
 347                                     sectors_per_cluster, &error_is_read,
 348                                     false);
 349                 if ((ret < 0) &&
 350                     backup_error_action(job, error_is_read, -ret) ==
 351                     BLOCK_ERROR_ACTION_REPORT) {
 352                     return ret;
 353                 }
 354             } while (ret < 0);
 355         }
 356
 357         /* If the bitmap granularity is smaller than the backup granularity,
 358          * we need to advance the iterator pointer to the next cluster. */
 359         if (granularity < job->cluster_size) {
 360             bdrv_set_dirty_iter(&hbi, cluster * sectors_per_cluster);
 361         }
 362
 363         last_cluster = cluster - 1;
 364     }
 365
 366     /* Play some final catchup with the progress meter */
 367     end = DIV_ROUND_UP(job->common.len, job->cluster_size);
 368     if (last_cluster + 1 < end) {
 369         job->common.offset += ((end - last_cluster - 1) * job->cluster_size);
 370     }
 371
 372     return ret;
 373 }
 374
 375 static void coroutine_fn backup_run(void *opaque)
 376 {
 377     BackupBlockJob *job = opaque;
 378     BackupCompleteData *data;
 379     BlockDriverState *bs = job->common.bs;
 380     BlockDriverState *target = job->target;
 381     NotifierWithReturn before_write = {
 382         .notify = backup_before_write_notify,
 383     };
 384     int64_t start, end;
 385     int64_t sectors_per_cluster = cluster_size_sectors(job);
 386     int ret = 0;
 387
 388     QLIST_INIT(&job->inflight_reqs);
 389     qemu_co_rwlock_init(&job->flush_rwlock);
 390
 391     start = 0;
 392     end = DIV_ROUND_UP(job->common.len, job->cluster_size);
 393
 394     job->done_bitmap = bitmap_new(end);
 395
 396     bdrv_add_before_write_notifier(bs, &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(bs, 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(&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     int ret;
 489
 490     assert(bs);
 491     assert(target);
 492     assert(cb);
 493
 494     if (bs == target) {
 495         error_setg(errp, "Source and target cannot be the same");
 496         return;
 497     }
 498
 499     if (!bdrv_is_inserted(bs)) {
 500         error_setg(errp, "Device is not inserted: %s",
 501                    bdrv_get_device_name(bs));
 502         return;
 503     }
 504
 505     if (!bdrv_is_inserted(target)) {
 506         error_setg(errp, "Device is not inserted: %s",
 507                    bdrv_get_device_name(target));
 508         return;
 509     }
 510
 511     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_BACKUP_SOURCE, errp)) {
 512         return;
 513     }
 514
 515     if (bdrv_op_is_blocked(target, BLOCK_OP_TYPE_BACKUP_TARGET, errp)) {
 516         return;
 517     }
 518
 519     if (sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
 520         if (!sync_bitmap) {
 521             error_setg(errp, "must provide a valid bitmap name for "
 522                              "\"incremental\" sync mode");
 523             return;
 524         }
 525
 526         /* Create a new bitmap, and freeze/disable this one. */
 527         if (bdrv_dirty_bitmap_create_successor(bs, sync_bitmap, errp) < 0) {
 528             return;
 529         }
 530     } else if (sync_bitmap) {
 531         error_setg(errp,
 532                    "a sync_bitmap was provided to backup_run, "
 533                    "but received an incompatible sync_mode (%s)",
 534                    MirrorSyncMode_lookup[sync_mode]);
 535         return;
 536     }
 537
 538     len = bdrv_getlength(bs);
 539     if (len < 0) {
 540         error_setg_errno(errp, -len, "unable to get length for '%s'",
 541                          bdrv_get_device_name(bs));
 542         goto error;
 543     }
 544
 545     BackupBlockJob *job = block_job_create(&backup_job_driver, bs, speed,
 546                                            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 }