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 "block/block_int-io.h"
  21 #include "block/dirty-bitmap.h"
  22 #include "block/reqlist.h"
  23 #include "sysemu/block-backend.h"
  24 #include "qemu/units.h"
  25 #include "qemu/co-shared-resource.h"
  26 #include "qemu/coroutine.h"
  27 #include "qemu/ratelimit.h"
  28 #include "block/aio_task.h"
  29 #include "qemu/error-report.h"
  30 #include "qemu/memalign.h"
  31
  32 #define BLOCK_COPY_MAX_COPY_RANGE (16 * MiB)
  33 #define BLOCK_COPY_MAX_BUFFER (1 * MiB)
  34 #define BLOCK_COPY_MAX_MEM (128 * MiB)
  35 #define BLOCK_COPY_MAX_WORKERS 64
  36 #define BLOCK_COPY_SLICE_TIME 100000000ULL /* ns */
  37 #define BLOCK_COPY_CLUSTER_SIZE_DEFAULT (1 << 16)
  38
  39 typedef enum {
  40     COPY_READ_WRITE_CLUSTER,
  41     COPY_READ_WRITE,
  42     COPY_WRITE_ZEROES,
  43     COPY_RANGE_SMALL,
  44     COPY_RANGE_FULL
  45 } BlockCopyMethod;
  46
  47 static coroutine_fn int block_copy_task_entry(AioTask *task);
  48
  49 typedef struct BlockCopyCallState {
  50     /* Fields initialized in block_copy_async() and never changed. */
  51     BlockCopyState *s;
  52     int64_t offset;
  53     int64_t bytes;
  54     int max_workers;
  55     int64_t max_chunk;
  56     bool ignore_ratelimit;
  57     BlockCopyAsyncCallbackFunc cb;
  58     void *cb_opaque;
  59     /* Coroutine where async block-copy is running */
  60     Coroutine *co;
  61
  62     /* Fields whose state changes throughout the execution */
  63     bool finished; /* atomic */
  64     QemuCoSleep sleep; /* TODO: protect API with a lock */
  65     bool cancelled; /* atomic */
  66     /* To reference all call states from BlockCopyState */
  67     QLIST_ENTRY(BlockCopyCallState) list;
  68
  69     /*
  70      * Fields that report information about return values and erros.
  71      * Protected by lock in BlockCopyState.
  72      */
  73     bool error_is_read;
  74     /*
  75      * @ret is set concurrently by tasks under mutex. Only set once by first
  76      * failed task (and untouched if no task failed).
  77      * After finishing (call_state->finished is true), it is not modified
  78      * anymore and may be safely read without mutex.
  79      */
  80     int ret;
  81 } BlockCopyCallState;
  82
  83 typedef struct BlockCopyTask {
  84     AioTask task;
  85
  86     /*
  87      * Fields initialized in block_copy_task_create()
  88      * and never changed.
  89      */
  90     BlockCopyState *s;
  91     BlockCopyCallState *call_state;
  92     /*
  93      * @method can also be set again in the while loop of
  94      * block_copy_dirty_clusters(), but it is never accessed concurrently
  95      * because the only other function that reads it is
  96      * block_copy_task_entry() and it is invoked afterwards in the same
  97      * iteration.
  98      */
  99     BlockCopyMethod method;
 100
 101     /*
 102      * Generally, req is protected by lock in BlockCopyState, Still req.offset
 103      * is only set on task creation, so may be read concurrently after creation.
 104      * req.bytes is changed at most once, and need only protecting the case of
 105      * parallel read while updating @bytes value in block_copy_task_shrink().
 106      */
 107     BlockReq req;
 108 } BlockCopyTask;
 109
 110 static int64_t task_end(BlockCopyTask *task)
 111 {
 112     return task->req.offset + task->req.bytes;
 113 }
 114
 115 typedef struct BlockCopyState {
 116     /*
 117      * BdrvChild objects are not owned or managed by block-copy. They are
 118      * provided by block-copy user and user is responsible for appropriate
 119      * permissions on these children.
 120      */
 121     BdrvChild *source;
 122     BdrvChild *target;
 123
 124     /*
 125      * Fields initialized in block_copy_state_new()
 126      * and never changed.
 127      */
 128     int64_t cluster_size;
 129     int64_t max_transfer;
 130     uint64_t len;
 131     BdrvRequestFlags write_flags;
 132
 133     /*
 134      * Fields whose state changes throughout the execution
 135      * Protected by lock.
 136      */
 137     CoMutex lock;
 138     int64_t in_flight_bytes;
 139     BlockCopyMethod method;
 140     BlockReqList reqs;
 141     QLIST_HEAD(, BlockCopyCallState) calls;
 142     /*
 143      * skip_unallocated:
 144      *
 145      * Used by sync=top jobs, which first scan the source node for unallocated
 146      * areas and clear them in the copy_bitmap.  During this process, the bitmap
 147      * is thus not fully initialized: It may still have bits set for areas that
 148      * are unallocated and should actually not be copied.
 149      *
 150      * This is indicated by skip_unallocated.
 151      *
 152      * In this case, block_copy() will query the source’s allocation status,
 153      * skip unallocated regions, clear them in the copy_bitmap, and invoke
 154      * block_copy_reset_unallocated() every time it does.
 155      */
 156     bool skip_unallocated; /* atomic */
 157     /* State fields that use a thread-safe API */
 158     BdrvDirtyBitmap *copy_bitmap;
 159     ProgressMeter *progress;
 160     SharedResource *mem;
 161     RateLimit rate_limit;
 162 } BlockCopyState;
 163
 164 /* Called with lock held */
 165 static int64_t block_copy_chunk_size(BlockCopyState *s)
 166 {
 167     switch (s->method) {
 168     case COPY_READ_WRITE_CLUSTER:
 169         return s->cluster_size;
 170     case COPY_READ_WRITE:
 171     case COPY_RANGE_SMALL:
 172         return MIN(MAX(s->cluster_size, BLOCK_COPY_MAX_BUFFER),
 173                    s->max_transfer);
 174     case COPY_RANGE_FULL:
 175         return MIN(MAX(s->cluster_size, BLOCK_COPY_MAX_COPY_RANGE),
 176                    s->max_transfer);
 177     default:
 178         /* Cannot have COPY_WRITE_ZEROES here.  */
 179         abort();
 180     }
 181 }
 182
 183 /*
 184  * Search for the first dirty area in offset/bytes range and create task at
 185  * the beginning of it.
 186  */
 187 static coroutine_fn BlockCopyTask *
 188 block_copy_task_create(BlockCopyState *s, BlockCopyCallState *call_state,
 189                        int64_t offset, int64_t bytes)
 190 {
 191     BlockCopyTask *task;
 192     int64_t max_chunk;
 193
 194     QEMU_LOCK_GUARD(&s->lock);
 195     max_chunk = MIN_NON_ZERO(block_copy_chunk_size(s), call_state->max_chunk);
 196     if (!bdrv_dirty_bitmap_next_dirty_area(s->copy_bitmap,
 197                                            offset, offset + bytes,
 198                                            max_chunk, &offset, &bytes))
 199     {
 200         return NULL;
 201     }
 202
 203     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
 204     bytes = QEMU_ALIGN_UP(bytes, s->cluster_size);
 205
 206     /* region is dirty, so no existent tasks possible in it */
 207     assert(!reqlist_find_conflict(&s->reqs, offset, bytes));
 208
 209     bdrv_reset_dirty_bitmap(s->copy_bitmap, offset, bytes);
 210     s->in_flight_bytes += bytes;
 211
 212     task = g_new(BlockCopyTask, 1);
 213     *task = (BlockCopyTask) {
 214         .task.func = block_copy_task_entry,
 215         .s = s,
 216         .call_state = call_state,
 217         .method = s->method,
 218     };
 219     reqlist_init_req(&s->reqs, &task->req, offset, bytes);
 220
 221     return task;
 222 }
 223
 224 /*
 225  * block_copy_task_shrink
 226  *
 227  * Drop the tail of the task to be handled later. Set dirty bits back and
 228  * wake up all tasks waiting for us (may be some of them are not intersecting
 229  * with shrunk task)
 230  */
 231 static void coroutine_fn block_copy_task_shrink(BlockCopyTask *task,
 232                                                 int64_t new_bytes)
 233 {
 234     QEMU_LOCK_GUARD(&task->s->lock);
 235     if (new_bytes == task->req.bytes) {
 236         return;
 237     }
 238
 239     assert(new_bytes > 0 && new_bytes < task->req.bytes);
 240
 241     task->s->in_flight_bytes -= task->req.bytes - new_bytes;
 242     bdrv_set_dirty_bitmap(task->s->copy_bitmap,
 243                           task->req.offset + new_bytes,
 244                           task->req.bytes - new_bytes);
 245
 246     reqlist_shrink_req(&task->req, new_bytes);
 247 }
 248
 249 static void coroutine_fn block_copy_task_end(BlockCopyTask *task, int ret)
 250 {
 251     QEMU_LOCK_GUARD(&task->s->lock);
 252     task->s->in_flight_bytes -= task->req.bytes;
 253     if (ret < 0) {
 254         bdrv_set_dirty_bitmap(task->s->copy_bitmap, task->req.offset,
 255                               task->req.bytes);
 256     }
 257     if (task->s->progress) {
 258         progress_set_remaining(task->s->progress,
 259                                bdrv_get_dirty_count(task->s->copy_bitmap) +
 260                                task->s->in_flight_bytes);
 261     }
 262     reqlist_remove_req(&task->req);
 263 }
 264
 265 void block_copy_state_free(BlockCopyState *s)
 266 {
 267     if (!s) {
 268         return;
 269     }
 270
 271     ratelimit_destroy(&s->rate_limit);
 272     bdrv_release_dirty_bitmap(s->copy_bitmap);
 273     shres_destroy(s->mem);
 274     g_free(s);
 275 }
 276
 277 static uint32_t block_copy_max_transfer(BdrvChild *source, BdrvChild *target)
 278 {
 279     return MIN_NON_ZERO(INT_MAX,
 280                         MIN_NON_ZERO(source->bs->bl.max_transfer,
 281                                      target->bs->bl.max_transfer));
 282 }
 283
 284 void block_copy_set_copy_opts(BlockCopyState *s, bool use_copy_range,
 285                               bool compress)
 286 {
 287     /* Keep BDRV_REQ_SERIALISING set (or not set) in block_copy_state_new() */
 288     s->write_flags = (s->write_flags & BDRV_REQ_SERIALISING) |
 289         (compress ? BDRV_REQ_WRITE_COMPRESSED : 0);
 290
 291     if (s->max_transfer < s->cluster_size) {
 292         /*
 293          * copy_range does not respect max_transfer. We don't want to bother
 294          * with requests smaller than block-copy cluster size, so fallback to
 295          * buffered copying (read and write respect max_transfer on their
 296          * behalf).
 297          */
 298         s->method = COPY_READ_WRITE_CLUSTER;
 299     } else if (compress) {
 300         /* Compression supports only cluster-size writes and no copy-range. */
 301         s->method = COPY_READ_WRITE_CLUSTER;
 302     } else {
 303         /*
 304          * If copy range enabled, start with COPY_RANGE_SMALL, until first
 305          * successful copy_range (look at block_copy_do_copy).
 306          */
 307         s->method = use_copy_range ? COPY_RANGE_SMALL : COPY_READ_WRITE;
 308     }
 309 }
 310
 311 static int64_t block_copy_calculate_cluster_size(BlockDriverState *target,
 312                                                  Error **errp)
 313 {
 314     int ret;
 315     BlockDriverInfo bdi;
 316     bool target_does_cow = bdrv_backing_chain_next(target);
 317
 318     /*
 319      * If there is no backing file on the target, we cannot rely on COW if our
 320      * backup cluster size is smaller than the target cluster size. Even for
 321      * targets with a backing file, try to avoid COW if possible.
 322      */
 323     ret = bdrv_get_info(target, &bdi);
 324     if (ret == -ENOTSUP && !target_does_cow) {
 325         /* Cluster size is not defined */
 326         warn_report("The target block device doesn't provide "
 327                     "information about the block size and it doesn't have a "
 328                     "backing file. The default block size of %u bytes is "
 329                     "used. If the actual block size of the target exceeds "
 330                     "this default, the backup may be unusable",
 331                     BLOCK_COPY_CLUSTER_SIZE_DEFAULT);
 332         return BLOCK_COPY_CLUSTER_SIZE_DEFAULT;
 333     } else if (ret < 0 && !target_does_cow) {
 334         error_setg_errno(errp, -ret,
 335             "Couldn't determine the cluster size of the target image, "
 336             "which has no backing file");
 337         error_append_hint(errp,
 338             "Aborting, since this may create an unusable destination image\n");
 339         return ret;
 340     } else if (ret < 0 && target_does_cow) {
 341         /* Not fatal; just trudge on ahead. */
 342         return BLOCK_COPY_CLUSTER_SIZE_DEFAULT;
 343     }
 344
 345     return MAX(BLOCK_COPY_CLUSTER_SIZE_DEFAULT, bdi.cluster_size);
 346 }
 347
 348 BlockCopyState *block_copy_state_new(BdrvChild *source, BdrvChild *target,
 349                                      const BdrvDirtyBitmap *bitmap,
 350                                      Error **errp)
 351 {
 352     ERRP_GUARD();
 353     BlockCopyState *s;
 354     int64_t cluster_size;
 355     BdrvDirtyBitmap *copy_bitmap;
 356     bool is_fleecing;
 357
 358     cluster_size = block_copy_calculate_cluster_size(target->bs, errp);
 359     if (cluster_size < 0) {
 360         return NULL;
 361     }
 362
 363     copy_bitmap = bdrv_create_dirty_bitmap(source->bs, cluster_size, NULL,
 364                                            errp);
 365     if (!copy_bitmap) {
 366         return NULL;
 367     }
 368     bdrv_disable_dirty_bitmap(copy_bitmap);
 369     if (bitmap) {
 370         if (!bdrv_merge_dirty_bitmap(copy_bitmap, bitmap, NULL, errp)) {
 371             error_prepend(errp, "Failed to merge bitmap '%s' to internal "
 372                           "copy-bitmap: ", bdrv_dirty_bitmap_name(bitmap));
 373             bdrv_release_dirty_bitmap(copy_bitmap);
 374             return NULL;
 375         }
 376     } else {
 377         bdrv_set_dirty_bitmap(copy_bitmap, 0,
 378                               bdrv_dirty_bitmap_size(copy_bitmap));
 379     }
 380
 381     /*
 382      * If source is in backing chain of target assume that target is going to be
 383      * used for "image fleecing", i.e. it should represent a kind of snapshot of
 384      * source at backup-start point in time. And target is going to be read by
 385      * somebody (for example, used as NBD export) during backup job.
 386      *
 387      * In this case, we need to add BDRV_REQ_SERIALISING write flag to avoid
 388      * intersection of backup writes and third party reads from target,
 389      * otherwise reading from target we may occasionally read already updated by
 390      * guest data.
 391      *
 392      * For more information see commit f8d59dfb40bb and test
 393      * tests/qemu-iotests/222
 394      */
 395     is_fleecing = bdrv_chain_contains(target->bs, source->bs);
 396
 397     s = g_new(BlockCopyState, 1);
 398     *s = (BlockCopyState) {
 399         .source = source,
 400         .target = target,
 401         .copy_bitmap = copy_bitmap,
 402         .cluster_size = cluster_size,
 403         .len = bdrv_dirty_bitmap_size(copy_bitmap),
 404         .write_flags = (is_fleecing ? BDRV_REQ_SERIALISING : 0),
 405         .mem = shres_create(BLOCK_COPY_MAX_MEM),
 406         .max_transfer = QEMU_ALIGN_DOWN(
 407                                     block_copy_max_transfer(source, target),
 408                                     cluster_size),
 409     };
 410
 411     block_copy_set_copy_opts(s, false, false);
 412
 413     ratelimit_init(&s->rate_limit);
 414     qemu_co_mutex_init(&s->lock);
 415     QLIST_INIT(&s->reqs);
 416     QLIST_INIT(&s->calls);
 417
 418     return s;
 419 }
 420
 421 /* Only set before running the job, no need for locking. */
 422 void block_copy_set_progress_meter(BlockCopyState *s, ProgressMeter *pm)
 423 {
 424     s->progress = pm;
 425 }
 426
 427 /*
 428  * Takes ownership of @task
 429  *
 430  * If pool is NULL directly run the task, otherwise schedule it into the pool.
 431  *
 432  * Returns: task.func return code if pool is NULL
 433  *          otherwise -ECANCELED if pool status is bad
 434  *          otherwise 0 (successfully scheduled)
 435  */
 436 static coroutine_fn int block_copy_task_run(AioTaskPool *pool,
 437                                             BlockCopyTask *task)
 438 {
 439     if (!pool) {
 440         int ret = task->task.func(&task->task);
 441
 442         g_free(task);
 443         return ret;
 444     }
 445
 446     aio_task_pool_wait_slot(pool);
 447     if (aio_task_pool_status(pool) < 0) {
 448         co_put_to_shres(task->s->mem, task->req.bytes);
 449         block_copy_task_end(task, -ECANCELED);
 450         g_free(task);
 451         return -ECANCELED;
 452     }
 453
 454     aio_task_pool_start_task(pool, &task->task);
 455
 456     return 0;
 457 }
 458
 459 /*
 460  * block_copy_do_copy
 461  *
 462  * Do copy of cluster-aligned chunk. Requested region is allowed to exceed
 463  * s->len only to cover last cluster when s->len is not aligned to clusters.
 464  *
 465  * No sync here: nor bitmap neighter intersecting requests handling, only copy.
 466  *
 467  * @method is an in-out argument, so that copy_range can be either extended to
 468  * a full-size buffer or disabled if the copy_range attempt fails.  The output
 469  * value of @method should be used for subsequent tasks.
 470  * Returns 0 on success.
 471  */
 472 static int coroutine_fn block_copy_do_copy(BlockCopyState *s,
 473                                            int64_t offset, int64_t bytes,
 474                                            BlockCopyMethod *method,
 475                                            bool *error_is_read)
 476 {
 477     int ret;
 478     int64_t nbytes = MIN(offset + bytes, s->len) - offset;
 479     void *bounce_buffer = NULL;
 480
 481     assert(offset >= 0 && bytes > 0 && INT64_MAX - offset >= bytes);
 482     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
 483     assert(QEMU_IS_ALIGNED(bytes, s->cluster_size));
 484     assert(offset < s->len);
 485     assert(offset + bytes <= s->len ||
 486            offset + bytes == QEMU_ALIGN_UP(s->len, s->cluster_size));
 487     assert(nbytes < INT_MAX);
 488
 489     switch (*method) {
 490     case COPY_WRITE_ZEROES:
 491         ret = bdrv_co_pwrite_zeroes(s->target, offset, nbytes, s->write_flags &
 492                                     ~BDRV_REQ_WRITE_COMPRESSED);
 493         if (ret < 0) {
 494             trace_block_copy_write_zeroes_fail(s, offset, ret);
 495             *error_is_read = false;
 496         }
 497         return ret;
 498
 499     case COPY_RANGE_SMALL:
 500     case COPY_RANGE_FULL:
 501         ret = bdrv_co_copy_range(s->source, offset, s->target, offset, nbytes,
 502                                  0, s->write_flags);
 503         if (ret >= 0) {
 504             /* Successful copy-range, increase chunk size.  */
 505             *method = COPY_RANGE_FULL;
 506             return 0;
 507         }
 508
 509         trace_block_copy_copy_range_fail(s, offset, ret);
 510         *method = COPY_READ_WRITE;
 511         /* Fall through to read+write with allocated buffer */
 512
 513     case COPY_READ_WRITE_CLUSTER:
 514     case COPY_READ_WRITE:
 515         /*
 516          * In case of failed copy_range request above, we may proceed with
 517          * buffered request larger than BLOCK_COPY_MAX_BUFFER.
 518          * Still, further requests will be properly limited, so don't care too
 519          * much. Moreover the most likely case (copy_range is unsupported for
 520          * the configuration, so the very first copy_range request fails)
 521          * is handled by setting large copy_size only after first successful
 522          * copy_range.
 523          */
 524
 525         bounce_buffer = qemu_blockalign(s->source->bs, nbytes);
 526
 527         ret = bdrv_co_pread(s->source, offset, nbytes, bounce_buffer, 0);
 528         if (ret < 0) {
 529             trace_block_copy_read_fail(s, offset, ret);
 530             *error_is_read = true;
 531             goto out;
 532         }
 533
 534         ret = bdrv_co_pwrite(s->target, offset, nbytes, bounce_buffer,
 535                              s->write_flags);
 536         if (ret < 0) {
 537             trace_block_copy_write_fail(s, offset, ret);
 538             *error_is_read = false;
 539             goto out;
 540         }
 541
 542     out:
 543         qemu_vfree(bounce_buffer);
 544         break;
 545
 546     default:
 547         abort();
 548     }
 549
 550     return ret;
 551 }
 552
 553 static coroutine_fn int block_copy_task_entry(AioTask *task)
 554 {
 555     BlockCopyTask *t = container_of(task, BlockCopyTask, task);
 556     BlockCopyState *s = t->s;
 557     bool error_is_read = false;
 558     BlockCopyMethod method = t->method;
 559     int ret;
 560
 561     ret = block_copy_do_copy(s, t->req.offset, t->req.bytes, &method,
 562                              &error_is_read);
 563
 564     WITH_QEMU_LOCK_GUARD(&s->lock) {
 565         if (s->method == t->method) {
 566             s->method = method;
 567         }
 568
 569         if (ret < 0) {
 570             if (!t->call_state->ret) {
 571                 t->call_state->ret = ret;
 572                 t->call_state->error_is_read = error_is_read;
 573             }
 574         } else if (s->progress) {
 575             progress_work_done(s->progress, t->req.bytes);
 576         }
 577     }
 578     co_put_to_shres(s->mem, t->req.bytes);
 579     block_copy_task_end(t, ret);
 580
 581     return ret;
 582 }
 583
 584 static coroutine_fn int block_copy_block_status(BlockCopyState *s,
 585                                                 int64_t offset,
 586                                                 int64_t bytes, int64_t *pnum)
 587 {
 588     int64_t num;
 589     BlockDriverState *base;
 590     int ret;
 591
 592     if (qatomic_read(&s->skip_unallocated)) {
 593         base = bdrv_backing_chain_next(s->source->bs);
 594     } else {
 595         base = NULL;
 596     }
 597
 598     ret = bdrv_co_block_status_above(s->source->bs, base, offset, bytes, &num,
 599                                      NULL, NULL);
 600     if (ret < 0 || num < s->cluster_size) {
 601         /*
 602          * On error or if failed to obtain large enough chunk just fallback to
 603          * copy one cluster.
 604          */
 605         num = s->cluster_size;
 606         ret = BDRV_BLOCK_ALLOCATED | BDRV_BLOCK_DATA;
 607     } else if (offset + num == s->len) {
 608         num = QEMU_ALIGN_UP(num, s->cluster_size);
 609     } else {
 610         num = QEMU_ALIGN_DOWN(num, s->cluster_size);
 611     }
 612
 613     *pnum = num;
 614     return ret;
 615 }
 616
 617 /*
 618  * Check if the cluster starting at offset is allocated or not.
 619  * return via pnum the number of contiguous clusters sharing this allocation.
 620  */
 621 static int coroutine_fn block_copy_is_cluster_allocated(BlockCopyState *s,
 622                                                         int64_t offset,
 623                                                         int64_t *pnum)
 624 {
 625     BlockDriverState *bs = s->source->bs;
 626     int64_t count, total_count = 0;
 627     int64_t bytes = s->len - offset;
 628     int ret;
 629
 630     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
 631
 632     while (true) {
 633         ret = bdrv_co_is_allocated(bs, offset, bytes, &count);
 634         if (ret < 0) {
 635             return ret;
 636         }
 637
 638         total_count += count;
 639
 640         if (ret || count == 0) {
 641             /*
 642              * ret: partial segment(s) are considered allocated.
 643              * otherwise: unallocated tail is treated as an entire segment.
 644              */
 645             *pnum = DIV_ROUND_UP(total_count, s->cluster_size);
 646             return ret;
 647         }
 648
 649         /* Unallocated segment(s) with uncertain following segment(s) */
 650         if (total_count >= s->cluster_size) {
 651             *pnum = total_count / s->cluster_size;
 652             return 0;
 653         }
 654
 655         offset += count;
 656         bytes -= count;
 657     }
 658 }
 659
 660 void block_copy_reset(BlockCopyState *s, int64_t offset, int64_t bytes)
 661 {
 662     QEMU_LOCK_GUARD(&s->lock);
 663
 664     bdrv_reset_dirty_bitmap(s->copy_bitmap, offset, bytes);
 665     if (s->progress) {
 666         progress_set_remaining(s->progress,
 667                                bdrv_get_dirty_count(s->copy_bitmap) +
 668                                s->in_flight_bytes);
 669     }
 670 }
 671
 672 /*
 673  * Reset bits in copy_bitmap starting at offset if they represent unallocated
 674  * data in the image. May reset subsequent contiguous bits.
 675  * @return 0 when the cluster at @offset was unallocated,
 676  *         1 otherwise, and -ret on error.
 677  */
 678 int64_t coroutine_fn block_copy_reset_unallocated(BlockCopyState *s,
 679                                                   int64_t offset,
 680                                                   int64_t *count)
 681 {
 682     int ret;
 683     int64_t clusters, bytes;
 684
 685     ret = block_copy_is_cluster_allocated(s, offset, &clusters);
 686     if (ret < 0) {
 687         return ret;
 688     }
 689
 690     bytes = clusters * s->cluster_size;
 691
 692     if (!ret) {
 693         block_copy_reset(s, offset, bytes);
 694     }
 695
 696     *count = bytes;
 697     return ret;
 698 }
 699
 700 /*
 701  * block_copy_dirty_clusters
 702  *
 703  * Copy dirty clusters in @offset/@bytes range.
 704  * Returns 1 if dirty clusters found and successfully copied, 0 if no dirty
 705  * clusters found and -errno on failure.
 706  */
 707 static int coroutine_fn
 708 block_copy_dirty_clusters(BlockCopyCallState *call_state)
 709 {
 710     BlockCopyState *s = call_state->s;
 711     int64_t offset = call_state->offset;
 712     int64_t bytes = call_state->bytes;
 713
 714     int ret = 0;
 715     bool found_dirty = false;
 716     int64_t end = offset + bytes;
 717     AioTaskPool *aio = NULL;
 718
 719     /*
 720      * block_copy() user is responsible for keeping source and target in same
 721      * aio context
 722      */
 723     assert(bdrv_get_aio_context(s->source->bs) ==
 724            bdrv_get_aio_context(s->target->bs));
 725
 726     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
 727     assert(QEMU_IS_ALIGNED(bytes, s->cluster_size));
 728
 729     while (bytes && aio_task_pool_status(aio) == 0 &&
 730            !qatomic_read(&call_state->cancelled)) {
 731         BlockCopyTask *task;
 732         int64_t status_bytes;
 733
 734         task = block_copy_task_create(s, call_state, offset, bytes);
 735         if (!task) {
 736             /* No more dirty bits in the bitmap */
 737             trace_block_copy_skip_range(s, offset, bytes);
 738             break;
 739         }
 740         if (task->req.offset > offset) {
 741             trace_block_copy_skip_range(s, offset, task->req.offset - offset);
 742         }
 743
 744         found_dirty = true;
 745
 746         ret = block_copy_block_status(s, task->req.offset, task->req.bytes,
 747                                       &status_bytes);
 748         assert(ret >= 0); /* never fail */
 749         if (status_bytes < task->req.bytes) {
 750             block_copy_task_shrink(task, status_bytes);
 751         }
 752         if (qatomic_read(&s->skip_unallocated) &&
 753             !(ret & BDRV_BLOCK_ALLOCATED)) {
 754             block_copy_task_end(task, 0);
 755             trace_block_copy_skip_range(s, task->req.offset, task->req.bytes);
 756             offset = task_end(task);
 757             bytes = end - offset;
 758             g_free(task);
 759             continue;
 760         }
 761         if (ret & BDRV_BLOCK_ZERO) {
 762             task->method = COPY_WRITE_ZEROES;
 763         }
 764
 765         if (!call_state->ignore_ratelimit) {
 766             uint64_t ns = ratelimit_calculate_delay(&s->rate_limit, 0);
 767             if (ns > 0) {
 768                 block_copy_task_end(task, -EAGAIN);
 769                 g_free(task);
 770                 qemu_co_sleep_ns_wakeable(&call_state->sleep,
 771                                           QEMU_CLOCK_REALTIME, ns);
 772                 continue;
 773             }
 774         }
 775
 776         ratelimit_calculate_delay(&s->rate_limit, task->req.bytes);
 777
 778         trace_block_copy_process(s, task->req.offset);
 779
 780         co_get_from_shres(s->mem, task->req.bytes);
 781
 782         offset = task_end(task);
 783         bytes = end - offset;
 784
 785         if (!aio && bytes) {
 786             aio = aio_task_pool_new(call_state->max_workers);
 787         }
 788
 789         ret = block_copy_task_run(aio, task);
 790         if (ret < 0) {
 791             goto out;
 792         }
 793     }
 794
 795 out:
 796     if (aio) {
 797         aio_task_pool_wait_all(aio);
 798
 799         /*
 800          * We are not really interested in -ECANCELED returned from
 801          * block_copy_task_run. If it fails, it means some task already failed
 802          * for real reason, let's return first failure.
 803          * Still, assert that we don't rewrite failure by success.
 804          *
 805          * Note: ret may be positive here because of block-status result.
 806          */
 807         assert(ret >= 0 || aio_task_pool_status(aio) < 0);
 808         ret = aio_task_pool_status(aio);
 809
 810         aio_task_pool_free(aio);
 811     }
 812
 813     return ret < 0 ? ret : found_dirty;
 814 }
 815
 816 void block_copy_kick(BlockCopyCallState *call_state)
 817 {
 818     qemu_co_sleep_wake(&call_state->sleep);
 819 }
 820
 821 /*
 822  * block_copy_common
 823  *
 824  * Copy requested region, accordingly to dirty bitmap.
 825  * Collaborate with parallel block_copy requests: if they succeed it will help
 826  * us. If they fail, we will retry not-copied regions. So, if we return error,
 827  * it means that some I/O operation failed in context of _this_ block_copy call,
 828  * not some parallel operation.
 829  */
 830 static int coroutine_fn block_copy_common(BlockCopyCallState *call_state)
 831 {
 832     int ret;
 833     BlockCopyState *s = call_state->s;
 834
 835     qemu_co_mutex_lock(&s->lock);
 836     QLIST_INSERT_HEAD(&s->calls, call_state, list);
 837     qemu_co_mutex_unlock(&s->lock);
 838
 839     do {
 840         ret = block_copy_dirty_clusters(call_state);
 841
 842         if (ret == 0 && !qatomic_read(&call_state->cancelled)) {
 843             WITH_QEMU_LOCK_GUARD(&s->lock) {
 844                 /*
 845                  * Check that there is no task we still need to
 846                  * wait to complete
 847                  */
 848                 ret = reqlist_wait_one(&s->reqs, call_state->offset,
 849                                        call_state->bytes, &s->lock);
 850                 if (ret == 0) {
 851                     /*
 852                      * No pending tasks, but check again the bitmap in this
 853                      * same critical section, since a task might have failed
 854                      * between this and the critical section in
 855                      * block_copy_dirty_clusters().
 856                      *
 857                      * reqlist_wait_one return value 0 also means that it
 858                      * didn't release the lock. So, we are still in the same
 859                      * critical section, not interrupted by any concurrent
 860                      * access to state.
 861                      */
 862                     ret = bdrv_dirty_bitmap_next_dirty(s->copy_bitmap,
 863                                                        call_state->offset,
 864                                                        call_state->bytes) >= 0;
 865                 }
 866             }
 867         }
 868
 869         /*
 870          * We retry in two cases:
 871          * 1. Some progress done
 872          *    Something was copied, which means that there were yield points
 873          *    and some new dirty bits may have appeared (due to failed parallel
 874          *    block-copy requests).
 875          * 2. We have waited for some intersecting block-copy request
 876          *    It may have failed and produced new dirty bits.
 877          */
 878     } while (ret > 0 && !qatomic_read(&call_state->cancelled));
 879
 880     qatomic_store_release(&call_state->finished, true);
 881
 882     if (call_state->cb) {
 883         call_state->cb(call_state->cb_opaque);
 884     }
 885
 886     qemu_co_mutex_lock(&s->lock);
 887     QLIST_REMOVE(call_state, list);
 888     qemu_co_mutex_unlock(&s->lock);
 889
 890     return ret;
 891 }
 892
 893 static void coroutine_fn block_copy_async_co_entry(void *opaque)
 894 {
 895     block_copy_common(opaque);
 896 }
 897
 898 int coroutine_fn block_copy(BlockCopyState *s, int64_t start, int64_t bytes,
 899                             bool ignore_ratelimit, uint64_t timeout_ns,
 900                             BlockCopyAsyncCallbackFunc cb,
 901                             void *cb_opaque)
 902 {
 903     int ret;
 904     BlockCopyCallState *call_state = g_new(BlockCopyCallState, 1);
 905
 906     *call_state = (BlockCopyCallState) {
 907         .s = s,
 908         .offset = start,
 909         .bytes = bytes,
 910         .ignore_ratelimit = ignore_ratelimit,
 911         .max_workers = BLOCK_COPY_MAX_WORKERS,
 912         .cb = cb,
 913         .cb_opaque = cb_opaque,
 914     };
 915
 916     ret = qemu_co_timeout(block_copy_async_co_entry, call_state, timeout_ns,
 917                           g_free);
 918     if (ret < 0) {
 919         assert(ret == -ETIMEDOUT);
 920         block_copy_call_cancel(call_state);
 921         /* call_state will be freed by running coroutine. */
 922         return ret;
 923     }
 924
 925     ret = call_state->ret;
 926     g_free(call_state);
 927
 928     return ret;
 929 }
 930
 931 BlockCopyCallState *block_copy_async(BlockCopyState *s,
 932                                      int64_t offset, int64_t bytes,
 933                                      int max_workers, int64_t max_chunk,
 934                                      BlockCopyAsyncCallbackFunc cb,
 935                                      void *cb_opaque)
 936 {
 937     BlockCopyCallState *call_state = g_new(BlockCopyCallState, 1);
 938
 939     *call_state = (BlockCopyCallState) {
 940         .s = s,
 941         .offset = offset,
 942         .bytes = bytes,
 943         .max_workers = max_workers,
 944         .max_chunk = max_chunk,
 945         .cb = cb,
 946         .cb_opaque = cb_opaque,
 947
 948         .co = qemu_coroutine_create(block_copy_async_co_entry, call_state),
 949     };
 950
 951     qemu_coroutine_enter(call_state->co);
 952
 953     return call_state;
 954 }
 955
 956 void block_copy_call_free(BlockCopyCallState *call_state)
 957 {
 958     if (!call_state) {
 959         return;
 960     }
 961
 962     assert(qatomic_read(&call_state->finished));
 963     g_free(call_state);
 964 }
 965
 966 bool block_copy_call_finished(BlockCopyCallState *call_state)
 967 {
 968     return qatomic_read(&call_state->finished);
 969 }
 970
 971 bool block_copy_call_succeeded(BlockCopyCallState *call_state)
 972 {
 973     return qatomic_load_acquire(&call_state->finished) &&
 974            !qatomic_read(&call_state->cancelled) &&
 975            call_state->ret == 0;
 976 }
 977
 978 bool block_copy_call_failed(BlockCopyCallState *call_state)
 979 {
 980     return qatomic_load_acquire(&call_state->finished) &&
 981            !qatomic_read(&call_state->cancelled) &&
 982            call_state->ret < 0;
 983 }
 984
 985 bool block_copy_call_cancelled(BlockCopyCallState *call_state)
 986 {
 987     return qatomic_read(&call_state->cancelled);
 988 }
 989
 990 int block_copy_call_status(BlockCopyCallState *call_state, bool *error_is_read)
 991 {
 992     assert(qatomic_load_acquire(&call_state->finished));
 993     if (error_is_read) {
 994         *error_is_read = call_state->error_is_read;
 995     }
 996     return call_state->ret;
 997 }
 998
 999 /*
1000  * Note that cancelling and finishing are racy.
1001  * User can cancel a block-copy that is already finished.
1002  */
1003 void block_copy_call_cancel(BlockCopyCallState *call_state)
1004 {
1005     qatomic_set(&call_state->cancelled, true);
1006     block_copy_kick(call_state);
1007 }
1008
1009 BdrvDirtyBitmap *block_copy_dirty_bitmap(BlockCopyState *s)
1010 {
1011     return s->copy_bitmap;
1012 }
1013
1014 int64_t block_copy_cluster_size(BlockCopyState *s)
1015 {
1016     return s->cluster_size;
1017 }
1018
1019 void block_copy_set_skip_unallocated(BlockCopyState *s, bool skip)
1020 {
1021     qatomic_set(&s->skip_unallocated, skip);
1022 }
1023
1024 void block_copy_set_speed(BlockCopyState *s, uint64_t speed)
1025 {
1026     ratelimit_set_speed(&s->rate_limit, speed, BLOCK_COPY_SLICE_TIME);
1027
1028     /*
1029      * Note: it's good to kick all call states from here, but it should be done
1030      * only from a coroutine, to not crash if s->calls list changed while
1031      * entering one call. So for now, the only user of this function kicks its
1032      * only one call_state by hand.
1033      */
1034 }