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