block/block-copy.c

   1 /*
   2  * block_copy API
   3  *
   4  * Copyright (C) 2013 Proxmox Server Solutions
   5  * Copyright (c) 2019 Virtuozzo International GmbH.
   6  *
   7  * Authors:
   8  *  Dietmar Maurer (dietmar@proxmox.com)
   9  *  Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
  10  *
  11  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  12  * See the COPYING file in the top-level directory.
  13  */
  14
  15 #include "qemu/osdep.h"
  16
  17 #include "trace.h"
  18 #include "qapi/error.h"
  19 #include "block/block-copy.h"
  20 #include "sysemu/block-backend.h"
  21 #include "qemu/units.h"
  22 #include "qemu/coroutine.h"
  23 #include "block/aio_task.h"
  24
  25 #define BLOCK_COPY_MAX_COPY_RANGE (16 * MiB)
  26 #define BLOCK_COPY_MAX_BUFFER (1 * MiB)
  27 #define BLOCK_COPY_MAX_MEM (128 * MiB)
  28 #define BLOCK_COPY_MAX_WORKERS 64
  29 #define BLOCK_COPY_SLICE_TIME 100000000ULL /* ns */
  30
  31 static coroutine_fn int block_copy_task_entry(AioTask *task);
  32
  33 typedef struct BlockCopyCallState {
  34     /* IN parameters. Initialized in block_copy_async() and never changed. */
  35     BlockCopyState *s;
  36     int64_t offset;
  37     int64_t bytes;
  38     int max_workers;
  39     int64_t max_chunk;
  40     bool ignore_ratelimit;
  41     BlockCopyAsyncCallbackFunc cb;
  42     void *cb_opaque;
  43
  44     /* Coroutine where async block-copy is running */
  45     Coroutine *co;
  46
  47     /* To reference all call states from BlockCopyState */
  48     QLIST_ENTRY(BlockCopyCallState) list;
  49
  50     /* State */
  51     int ret;
  52     bool finished;
  53     QemuCoSleep sleep;
  54     bool cancelled;
  55
  56     /* OUT parameters */
  57     bool error_is_read;
  58 } BlockCopyCallState;
  59
  60 typedef struct BlockCopyTask {
  61     AioTask task;
  62
  63     BlockCopyState *s;
  64     BlockCopyCallState *call_state;
  65     int64_t offset;
  66     int64_t bytes;
  67     bool zeroes;
  68     QLIST_ENTRY(BlockCopyTask) list;
  69     CoQueue wait_queue; /* coroutines blocked on this task */
  70 } BlockCopyTask;
  71
  72 static int64_t task_end(BlockCopyTask *task)
  73 {
  74     return task->offset + task->bytes;
  75 }
  76
  77 typedef struct BlockCopyState {
  78     /*
  79      * BdrvChild objects are not owned or managed by block-copy. They are
  80      * provided by block-copy user and user is responsible for appropriate
  81      * permissions on these children.
  82      */
  83     BdrvChild *source;
  84     BdrvChild *target;
  85     BdrvDirtyBitmap *copy_bitmap;
  86     int64_t in_flight_bytes;
  87     int64_t cluster_size;
  88     bool use_copy_range;
  89     int64_t copy_size;
  90     uint64_t len;
  91     QLIST_HEAD(, BlockCopyTask) tasks; /* All tasks from all block-copy calls */
  92     QLIST_HEAD(, BlockCopyCallState) calls;
  93
  94     BdrvRequestFlags write_flags;
  95
  96     /*
  97      * skip_unallocated:
  98      *
  99      * Used by sync=top jobs, which first scan the source node for unallocated
 100      * areas and clear them in the copy_bitmap.  During this process, the bitmap
 101      * is thus not fully initialized: It may still have bits set for areas that
 102      * are unallocated and should actually not be copied.
 103      *
 104      * This is indicated by skip_unallocated.
 105      *
 106      * In this case, block_copy() will query the source’s allocation status,
 107      * skip unallocated regions, clear them in the copy_bitmap, and invoke
 108      * block_copy_reset_unallocated() every time it does.
 109      */
 110     bool skip_unallocated;
 111
 112     ProgressMeter *progress;
 113
 114     SharedResource *mem;
 115
 116     uint64_t speed;
 117     RateLimit rate_limit;
 118 } BlockCopyState;
 119
 120 static BlockCopyTask *find_conflicting_task(BlockCopyState *s,
 121                                             int64_t offset, int64_t bytes)
 122 {
 123     BlockCopyTask *t;
 124
 125     QLIST_FOREACH(t, &s->tasks, list) {
 126         if (offset + bytes > t->offset && offset < t->offset + t->bytes) {
 127             return t;
 128         }
 129     }
 130
 131     return NULL;
 132 }
 133
 134 /*
 135  * If there are no intersecting tasks return false. Otherwise, wait for the
 136  * first found intersecting tasks to finish and return true.
 137  */
 138 static bool coroutine_fn block_copy_wait_one(BlockCopyState *s, int64_t offset,
 139                                              int64_t bytes)
 140 {
 141     BlockCopyTask *task = find_conflicting_task(s, offset, bytes);
 142
 143     if (!task) {
 144         return false;
 145     }
 146
 147     qemu_co_queue_wait(&task->wait_queue, NULL);
 148
 149     return true;
 150 }
 151
 152 /*
 153  * Search for the first dirty area in offset/bytes range and create task at
 154  * the beginning of it.
 155  */
 156 static BlockCopyTask *block_copy_task_create(BlockCopyState *s,
 157                                              BlockCopyCallState *call_state,
 158                                              int64_t offset, int64_t bytes)
 159 {
 160     BlockCopyTask *task;
 161     int64_t max_chunk = MIN_NON_ZERO(s->copy_size, call_state->max_chunk);
 162
 163     if (!bdrv_dirty_bitmap_next_dirty_area(s->copy_bitmap,
 164                                            offset, offset + bytes,
 165                                            max_chunk, &offset, &bytes))
 166     {
 167         return NULL;
 168     }
 169
 170     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
 171     bytes = QEMU_ALIGN_UP(bytes, s->cluster_size);
 172
 173     /* region is dirty, so no existent tasks possible in it */
 174     assert(!find_conflicting_task(s, offset, bytes));
 175
 176     bdrv_reset_dirty_bitmap(s->copy_bitmap, offset, bytes);
 177     s->in_flight_bytes += bytes;
 178
 179     task = g_new(BlockCopyTask, 1);
 180     *task = (BlockCopyTask) {
 181         .task.func = block_copy_task_entry,
 182         .s = s,
 183         .call_state = call_state,
 184         .offset = offset,
 185         .bytes = bytes,
 186     };
 187     qemu_co_queue_init(&task->wait_queue);
 188     QLIST_INSERT_HEAD(&s->tasks, task, list);
 189
 190     return task;
 191 }
 192
 193 /*
 194  * block_copy_task_shrink
 195  *
 196  * Drop the tail of the task to be handled later. Set dirty bits back and
 197  * wake up all tasks waiting for us (may be some of them are not intersecting
 198  * with shrunk task)
 199  */
 200 static void coroutine_fn block_copy_task_shrink(BlockCopyTask *task,
 201                                                 int64_t new_bytes)
 202 {
 203     if (new_bytes == task->bytes) {
 204         return;
 205     }
 206
 207     assert(new_bytes > 0 && new_bytes < task->bytes);
 208
 209     task->s->in_flight_bytes -= task->bytes - new_bytes;
 210     bdrv_set_dirty_bitmap(task->s->copy_bitmap,
 211                           task->offset + new_bytes, task->bytes - new_bytes);
 212
 213     task->bytes = new_bytes;
 214     qemu_co_queue_restart_all(&task->wait_queue);
 215 }
 216
 217 static void coroutine_fn block_copy_task_end(BlockCopyTask *task, int ret)
 218 {
 219     task->s->in_flight_bytes -= task->bytes;
 220     if (ret < 0) {
 221         bdrv_set_dirty_bitmap(task->s->copy_bitmap, task->offset, task->bytes);
 222     }
 223     QLIST_REMOVE(task, list);
 224     qemu_co_queue_restart_all(&task->wait_queue);
 225 }
 226
 227 void block_copy_state_free(BlockCopyState *s)
 228 {
 229     if (!s) {
 230         return;
 231     }
 232
 233     ratelimit_destroy(&s->rate_limit);
 234     bdrv_release_dirty_bitmap(s->copy_bitmap);
 235     shres_destroy(s->mem);
 236     g_free(s);
 237 }
 238
 239 static uint32_t block_copy_max_transfer(BdrvChild *source, BdrvChild *target)
 240 {
 241     return MIN_NON_ZERO(INT_MAX,
 242                         MIN_NON_ZERO(source->bs->bl.max_transfer,
 243                                      target->bs->bl.max_transfer));
 244 }
 245
 246 BlockCopyState *block_copy_state_new(BdrvChild *source, BdrvChild *target,
 247                                      int64_t cluster_size, bool use_copy_range,
 248                                      BdrvRequestFlags write_flags, Error **errp)
 249 {
 250     BlockCopyState *s;
 251     BdrvDirtyBitmap *copy_bitmap;
 252
 253     copy_bitmap = bdrv_create_dirty_bitmap(source->bs, cluster_size, NULL,
 254                                            errp);
 255     if (!copy_bitmap) {
 256         return NULL;
 257     }
 258     bdrv_disable_dirty_bitmap(copy_bitmap);
 259
 260     s = g_new(BlockCopyState, 1);
 261     *s = (BlockCopyState) {
 262         .source = source,
 263         .target = target,
 264         .copy_bitmap = copy_bitmap,
 265         .cluster_size = cluster_size,
 266         .len = bdrv_dirty_bitmap_size(copy_bitmap),
 267         .write_flags = write_flags,
 268         .mem = shres_create(BLOCK_COPY_MAX_MEM),
 269     };
 270
 271     if (block_copy_max_transfer(source, target) < cluster_size) {
 272         /*
 273          * copy_range does not respect max_transfer. We don't want to bother
 274          * with requests smaller than block-copy cluster size, so fallback to
 275          * buffered copying (read and write respect max_transfer on their
 276          * behalf).
 277          */
 278         s->use_copy_range = false;
 279         s->copy_size = cluster_size;
 280     } else if (write_flags & BDRV_REQ_WRITE_COMPRESSED) {
 281         /* Compression supports only cluster-size writes and no copy-range. */
 282         s->use_copy_range = false;
 283         s->copy_size = cluster_size;
 284     } else {
 285         /*
 286          * We enable copy-range, but keep small copy_size, until first
 287          * successful copy_range (look at block_copy_do_copy).
 288          */
 289         s->use_copy_range = use_copy_range;
 290         s->copy_size = MAX(s->cluster_size, BLOCK_COPY_MAX_BUFFER);
 291     }
 292
 293     ratelimit_init(&s->rate_limit);
 294     QLIST_INIT(&s->tasks);
 295     QLIST_INIT(&s->calls);
 296
 297     return s;
 298 }
 299
 300 void block_copy_set_progress_meter(BlockCopyState *s, ProgressMeter *pm)
 301 {
 302     s->progress = pm;
 303 }
 304
 305 /*
 306  * Takes ownership of @task
 307  *
 308  * If pool is NULL directly run the task, otherwise schedule it into the pool.
 309  *
 310  * Returns: task.func return code if pool is NULL
 311  *          otherwise -ECANCELED if pool status is bad
 312  *          otherwise 0 (successfully scheduled)
 313  */
 314 static coroutine_fn int block_copy_task_run(AioTaskPool *pool,
 315                                             BlockCopyTask *task)
 316 {
 317     if (!pool) {
 318         int ret = task->task.func(&task->task);
 319
 320         g_free(task);
 321         return ret;
 322     }
 323
 324     aio_task_pool_wait_slot(pool);
 325     if (aio_task_pool_status(pool) < 0) {
 326         co_put_to_shres(task->s->mem, task->bytes);
 327         block_copy_task_end(task, -ECANCELED);
 328         g_free(task);
 329         return -ECANCELED;
 330     }
 331
 332     aio_task_pool_start_task(pool, &task->task);
 333
 334     return 0;
 335 }
 336
 337 /*
 338  * block_copy_do_copy
 339  *
 340  * Do copy of cluster-aligned chunk. Requested region is allowed to exceed
 341  * s->len only to cover last cluster when s->len is not aligned to clusters.
 342  *
 343  * No sync here: nor bitmap neighter intersecting requests handling, only copy.
 344  *
 345  * Returns 0 on success.
 346  */
 347 static int coroutine_fn block_copy_do_copy(BlockCopyState *s,
 348                                            int64_t offset, int64_t bytes,
 349                                            bool zeroes, bool *error_is_read)
 350 {
 351     int ret;
 352     int64_t nbytes = MIN(offset + bytes, s->len) - offset;
 353     void *bounce_buffer = NULL;
 354
 355     assert(offset >= 0 && bytes > 0 && INT64_MAX - offset >= bytes);
 356     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
 357     assert(QEMU_IS_ALIGNED(bytes, s->cluster_size));
 358     assert(offset < s->len);
 359     assert(offset + bytes <= s->len ||
 360            offset + bytes == QEMU_ALIGN_UP(s->len, s->cluster_size));
 361     assert(nbytes < INT_MAX);
 362
 363     if (zeroes) {
 364         ret = bdrv_co_pwrite_zeroes(s->target, offset, nbytes, s->write_flags &
 365                                     ~BDRV_REQ_WRITE_COMPRESSED);
 366         if (ret < 0) {
 367             trace_block_copy_write_zeroes_fail(s, offset, ret);
 368             *error_is_read = false;
 369         }
 370         return ret;
 371     }
 372
 373     if (s->use_copy_range) {
 374         ret = bdrv_co_copy_range(s->source, offset, s->target, offset, nbytes,
 375                                  0, s->write_flags);
 376         if (ret < 0) {
 377             trace_block_copy_copy_range_fail(s, offset, ret);
 378             s->use_copy_range = false;
 379             s->copy_size = MAX(s->cluster_size, BLOCK_COPY_MAX_BUFFER);
 380             /* Fallback to read+write with allocated buffer */
 381         } else {
 382             if (s->use_copy_range) {
 383                 /*
 384                  * Successful copy-range. Now increase copy_size.  copy_range
 385                  * does not respect max_transfer (it's a TODO), so we factor
 386                  * that in here.
 387                  *
 388                  * Note: we double-check s->use_copy_range for the case when
 389                  * parallel block-copy request unsets it during previous
 390                  * bdrv_co_copy_range call.
 391                  */
 392                 s->copy_size =
 393                         MIN(MAX(s->cluster_size, BLOCK_COPY_MAX_COPY_RANGE),
 394                             QEMU_ALIGN_DOWN(block_copy_max_transfer(s->source,
 395                                                                     s->target),
 396                                             s->cluster_size));
 397             }
 398             goto out;
 399         }
 400     }
 401
 402     /*
 403      * In case of failed copy_range request above, we may proceed with buffered
 404      * request larger than BLOCK_COPY_MAX_BUFFER. Still, further requests will
 405      * be properly limited, so don't care too much. Moreover the most likely
 406      * case (copy_range is unsupported for the configuration, so the very first
 407      * copy_range request fails) is handled by setting large copy_size only
 408      * after first successful copy_range.
 409      */
 410
 411     bounce_buffer = qemu_blockalign(s->source->bs, nbytes);
 412
 413     ret = bdrv_co_pread(s->source, offset, nbytes, bounce_buffer, 0);
 414     if (ret < 0) {
 415         trace_block_copy_read_fail(s, offset, ret);
 416         *error_is_read = true;
 417         goto out;
 418     }
 419
 420     ret = bdrv_co_pwrite(s->target, offset, nbytes, bounce_buffer,
 421                          s->write_flags);
 422     if (ret < 0) {
 423         trace_block_copy_write_fail(s, offset, ret);
 424         *error_is_read = false;
 425         goto out;
 426     }
 427
 428 out:
 429     qemu_vfree(bounce_buffer);
 430
 431     return ret;
 432 }
 433
 434 static coroutine_fn int block_copy_task_entry(AioTask *task)
 435 {
 436     BlockCopyTask *t = container_of(task, BlockCopyTask, task);
 437     bool error_is_read = false;
 438     int ret;
 439
 440     ret = block_copy_do_copy(t->s, t->offset, t->bytes, t->zeroes,
 441                              &error_is_read);
 442     if (ret < 0 && !t->call_state->ret) {
 443         t->call_state->ret = ret;
 444         t->call_state->error_is_read = error_is_read;
 445     } else {
 446         progress_work_done(t->s->progress, t->bytes);
 447     }
 448     co_put_to_shres(t->s->mem, t->bytes);
 449     block_copy_task_end(t, ret);
 450
 451     return ret;
 452 }
 453
 454 static int block_copy_block_status(BlockCopyState *s, int64_t offset,
 455                                    int64_t bytes, int64_t *pnum)
 456 {
 457     int64_t num;
 458     BlockDriverState *base;
 459     int ret;
 460
 461     if (s->skip_unallocated) {
 462         base = bdrv_backing_chain_next(s->source->bs);
 463     } else {
 464         base = NULL;
 465     }
 466
 467     ret = bdrv_block_status_above(s->source->bs, base, offset, bytes, &num,
 468                                   NULL, NULL);
 469     if (ret < 0 || num < s->cluster_size) {
 470         /*
 471          * On error or if failed to obtain large enough chunk just fallback to
 472          * copy one cluster.
 473          */
 474         num = s->cluster_size;
 475         ret = BDRV_BLOCK_ALLOCATED | BDRV_BLOCK_DATA;
 476     } else if (offset + num == s->len) {
 477         num = QEMU_ALIGN_UP(num, s->cluster_size);
 478     } else {
 479         num = QEMU_ALIGN_DOWN(num, s->cluster_size);
 480     }
 481
 482     *pnum = num;
 483     return ret;
 484 }
 485
 486 /*
 487  * Check if the cluster starting at offset is allocated or not.
 488  * return via pnum the number of contiguous clusters sharing this allocation.
 489  */
 490 static int block_copy_is_cluster_allocated(BlockCopyState *s, int64_t offset,
 491                                            int64_t *pnum)
 492 {
 493     BlockDriverState *bs = s->source->bs;
 494     int64_t count, total_count = 0;
 495     int64_t bytes = s->len - offset;
 496     int ret;
 497
 498     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
 499
 500     while (true) {
 501         ret = bdrv_is_allocated(bs, offset, bytes, &count);
 502         if (ret < 0) {
 503             return ret;
 504         }
 505
 506         total_count += count;
 507
 508         if (ret || count == 0) {
 509             /*
 510              * ret: partial segment(s) are considered allocated.
 511              * otherwise: unallocated tail is treated as an entire segment.
 512              */
 513             *pnum = DIV_ROUND_UP(total_count, s->cluster_size);
 514             return ret;
 515         }
 516
 517         /* Unallocated segment(s) with uncertain following segment(s) */
 518         if (total_count >= s->cluster_size) {
 519             *pnum = total_count / s->cluster_size;
 520             return 0;
 521         }
 522
 523         offset += count;
 524         bytes -= count;
 525     }
 526 }
 527
 528 /*
 529  * Reset bits in copy_bitmap starting at offset if they represent unallocated
 530  * data in the image. May reset subsequent contiguous bits.
 531  * @return 0 when the cluster at @offset was unallocated,
 532  *         1 otherwise, and -ret on error.
 533  */
 534 int64_t block_copy_reset_unallocated(BlockCopyState *s,
 535                                      int64_t offset, int64_t *count)
 536 {
 537     int ret;
 538     int64_t clusters, bytes;
 539
 540     ret = block_copy_is_cluster_allocated(s, offset, &clusters);
 541     if (ret < 0) {
 542         return ret;
 543     }
 544
 545     bytes = clusters * s->cluster_size;
 546
 547     if (!ret) {
 548         bdrv_reset_dirty_bitmap(s->copy_bitmap, offset, bytes);
 549         progress_set_remaining(s->progress,
 550                                bdrv_get_dirty_count(s->copy_bitmap) +
 551                                s->in_flight_bytes);
 552     }
 553
 554     *count = bytes;
 555     return ret;
 556 }
 557
 558 /*
 559  * block_copy_dirty_clusters
 560  *
 561  * Copy dirty clusters in @offset/@bytes range.
 562  * Returns 1 if dirty clusters found and successfully copied, 0 if no dirty
 563  * clusters found and -errno on failure.
 564  */
 565 static int coroutine_fn
 566 block_copy_dirty_clusters(BlockCopyCallState *call_state)
 567 {
 568     BlockCopyState *s = call_state->s;
 569     int64_t offset = call_state->offset;
 570     int64_t bytes = call_state->bytes;
 571
 572     int ret = 0;
 573     bool found_dirty = false;
 574     int64_t end = offset + bytes;
 575     AioTaskPool *aio = NULL;
 576
 577     /*
 578      * block_copy() user is responsible for keeping source and target in same
 579      * aio context
 580      */
 581     assert(bdrv_get_aio_context(s->source->bs) ==
 582            bdrv_get_aio_context(s->target->bs));
 583
 584     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
 585     assert(QEMU_IS_ALIGNED(bytes, s->cluster_size));
 586
 587     while (bytes && aio_task_pool_status(aio) == 0 && !call_state->cancelled) {
 588         BlockCopyTask *task;
 589         int64_t status_bytes;
 590
 591         task = block_copy_task_create(s, call_state, offset, bytes);
 592         if (!task) {
 593             /* No more dirty bits in the bitmap */
 594             trace_block_copy_skip_range(s, offset, bytes);
 595             break;
 596         }
 597         if (task->offset > offset) {
 598             trace_block_copy_skip_range(s, offset, task->offset - offset);
 599         }
 600
 601         found_dirty = true;
 602
 603         ret = block_copy_block_status(s, task->offset, task->bytes,
 604                                       &status_bytes);
 605         assert(ret >= 0); /* never fail */
 606         if (status_bytes < task->bytes) {
 607             block_copy_task_shrink(task, status_bytes);
 608         }
 609         if (s->skip_unallocated && !(ret & BDRV_BLOCK_ALLOCATED)) {
 610             block_copy_task_end(task, 0);
 611             progress_set_remaining(s->progress,
 612                                    bdrv_get_dirty_count(s->copy_bitmap) +
 613                                    s->in_flight_bytes);
 614             trace_block_copy_skip_range(s, task->offset, task->bytes);
 615             offset = task_end(task);
 616             bytes = end - offset;
 617             g_free(task);
 618             continue;
 619         }
 620         task->zeroes = ret & BDRV_BLOCK_ZERO;
 621
 622         if (s->speed) {
 623             if (!call_state->ignore_ratelimit) {
 624                 uint64_t ns = ratelimit_calculate_delay(&s->rate_limit, 0);
 625                 if (ns > 0) {
 626                     block_copy_task_end(task, -EAGAIN);
 627                     g_free(task);
 628                     qemu_co_sleep_ns_wakeable(&call_state->sleep,
 629                                               QEMU_CLOCK_REALTIME, ns);
 630                     continue;
 631                 }
 632             }
 633
 634             ratelimit_calculate_delay(&s->rate_limit, task->bytes);
 635         }
 636
 637         trace_block_copy_process(s, task->offset);
 638
 639         co_get_from_shres(s->mem, task->bytes);
 640
 641         offset = task_end(task);
 642         bytes = end - offset;
 643
 644         if (!aio && bytes) {
 645             aio = aio_task_pool_new(call_state->max_workers);
 646         }
 647
 648         ret = block_copy_task_run(aio, task);
 649         if (ret < 0) {
 650             goto out;
 651         }
 652     }
 653
 654 out:
 655     if (aio) {
 656         aio_task_pool_wait_all(aio);
 657
 658         /*
 659          * We are not really interested in -ECANCELED returned from
 660          * block_copy_task_run. If it fails, it means some task already failed
 661          * for real reason, let's return first failure.
 662          * Still, assert that we don't rewrite failure by success.
 663          *
 664          * Note: ret may be positive here because of block-status result.
 665          */
 666         assert(ret >= 0 || aio_task_pool_status(aio) < 0);
 667         ret = aio_task_pool_status(aio);
 668
 669         aio_task_pool_free(aio);
 670     }
 671
 672     return ret < 0 ? ret : found_dirty;
 673 }
 674
 675 void block_copy_kick(BlockCopyCallState *call_state)
 676 {
 677     qemu_co_sleep_wake(&call_state->sleep);
 678 }
 679
 680 /*
 681  * block_copy_common
 682  *
 683  * Copy requested region, accordingly to dirty bitmap.
 684  * Collaborate with parallel block_copy requests: if they succeed it will help
 685  * us. If they fail, we will retry not-copied regions. So, if we return error,
 686  * it means that some I/O operation failed in context of _this_ block_copy call,
 687  * not some parallel operation.
 688  */
 689 static int coroutine_fn block_copy_common(BlockCopyCallState *call_state)
 690 {
 691     int ret;
 692
 693     QLIST_INSERT_HEAD(&call_state->s->calls, call_state, list);
 694
 695     do {
 696         ret = block_copy_dirty_clusters(call_state);
 697
 698         if (ret == 0 && !call_state->cancelled) {
 699             ret = block_copy_wait_one(call_state->s, call_state->offset,
 700                                       call_state->bytes);
 701         }
 702
 703         /*
 704          * We retry in two cases:
 705          * 1. Some progress done
 706          *    Something was copied, which means that there were yield points
 707          *    and some new dirty bits may have appeared (due to failed parallel
 708          *    block-copy requests).
 709          * 2. We have waited for some intersecting block-copy request
 710          *    It may have failed and produced new dirty bits.
 711          */
 712     } while (ret > 0 && !call_state->cancelled);
 713
 714     call_state->finished = true;
 715
 716     if (call_state->cb) {
 717         call_state->cb(call_state->cb_opaque);
 718     }
 719
 720     QLIST_REMOVE(call_state, list);
 721
 722     return ret;
 723 }
 724
 725 int coroutine_fn block_copy(BlockCopyState *s, int64_t start, int64_t bytes,
 726                             bool ignore_ratelimit)
 727 {
 728     BlockCopyCallState call_state = {
 729         .s = s,
 730         .offset = start,
 731         .bytes = bytes,
 732         .ignore_ratelimit = ignore_ratelimit,
 733         .max_workers = BLOCK_COPY_MAX_WORKERS,
 734     };
 735
 736     return block_copy_common(&call_state);
 737 }
 738
 739 static void coroutine_fn block_copy_async_co_entry(void *opaque)
 740 {
 741     block_copy_common(opaque);
 742 }
 743
 744 BlockCopyCallState *block_copy_async(BlockCopyState *s,
 745                                      int64_t offset, int64_t bytes,
 746                                      int max_workers, int64_t max_chunk,
 747                                      BlockCopyAsyncCallbackFunc cb,
 748                                      void *cb_opaque)
 749 {
 750     BlockCopyCallState *call_state = g_new(BlockCopyCallState, 1);
 751
 752     *call_state = (BlockCopyCallState) {
 753         .s = s,
 754         .offset = offset,
 755         .bytes = bytes,
 756         .max_workers = max_workers,
 757         .max_chunk = max_chunk,
 758         .cb = cb,
 759         .cb_opaque = cb_opaque,
 760
 761         .co = qemu_coroutine_create(block_copy_async_co_entry, call_state),
 762     };
 763
 764     qemu_coroutine_enter(call_state->co);
 765
 766     return call_state;
 767 }
 768
 769 void block_copy_call_free(BlockCopyCallState *call_state)
 770 {
 771     if (!call_state) {
 772         return;
 773     }
 774
 775     assert(call_state->finished);
 776     g_free(call_state);
 777 }
 778
 779 bool block_copy_call_finished(BlockCopyCallState *call_state)
 780 {
 781     return call_state->finished;
 782 }
 783
 784 bool block_copy_call_succeeded(BlockCopyCallState *call_state)
 785 {
 786     return call_state->finished && !call_state->cancelled &&
 787         call_state->ret == 0;
 788 }
 789
 790 bool block_copy_call_failed(BlockCopyCallState *call_state)
 791 {
 792     return call_state->finished && !call_state->cancelled &&
 793         call_state->ret < 0;
 794 }
 795
 796 bool block_copy_call_cancelled(BlockCopyCallState *call_state)
 797 {
 798     return call_state->cancelled;
 799 }
 800
 801 int block_copy_call_status(BlockCopyCallState *call_state, bool *error_is_read)
 802 {
 803     assert(call_state->finished);
 804     if (error_is_read) {
 805         *error_is_read = call_state->error_is_read;
 806     }
 807     return call_state->ret;
 808 }
 809
 810 void block_copy_call_cancel(BlockCopyCallState *call_state)
 811 {
 812     call_state->cancelled = true;
 813     block_copy_kick(call_state);
 814 }
 815
 816 BdrvDirtyBitmap *block_copy_dirty_bitmap(BlockCopyState *s)
 817 {
 818     return s->copy_bitmap;
 819 }
 820
 821 void block_copy_set_skip_unallocated(BlockCopyState *s, bool skip)
 822 {
 823     s->skip_unallocated = skip;
 824 }
 825
 826 void block_copy_set_speed(BlockCopyState *s, uint64_t speed)
 827 {
 828     s->speed = speed;
 829     if (speed > 0) {
 830         ratelimit_set_speed(&s->rate_limit, speed, BLOCK_COPY_SLICE_TIME);
 831     }
 832
 833     /*
 834      * Note: it's good to kick all call states from here, but it should be done
 835      * only from a coroutine, to not crash if s->calls list changed while
 836      * entering one call. So for now, the only user of this function kicks its
 837      * only one call_state by hand.
 838      */
 839 }