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