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
  30 static coroutine_fn int block_copy_task_entry(AioTask *task);
  31
  32 typedef struct BlockCopyCallState {
  33     bool failed;
  34     bool error_is_read;
  35 } BlockCopyCallState;
  36
  37 typedef struct BlockCopyTask {
  38     AioTask task;
  39
  40     BlockCopyState *s;
  41     BlockCopyCallState *call_state;
  42     int64_t offset;
  43     int64_t bytes;
  44     bool zeroes;
  45     QLIST_ENTRY(BlockCopyTask) list;
  46     CoQueue wait_queue; /* coroutines blocked on this task */
  47 } BlockCopyTask;
  48
  49 static int64_t task_end(BlockCopyTask *task)
  50 {
  51     return task->offset + task->bytes;
  52 }
  53
  54 typedef struct BlockCopyState {
  55     /*
  56      * BdrvChild objects are not owned or managed by block-copy. They are
  57      * provided by block-copy user and user is responsible for appropriate
  58      * permissions on these children.
  59      */
  60     BdrvChild *source;
  61     BdrvChild *target;
  62     BdrvDirtyBitmap *copy_bitmap;
  63     int64_t in_flight_bytes;
  64     int64_t cluster_size;
  65     bool use_copy_range;
  66     int64_t copy_size;
  67     uint64_t len;
  68     QLIST_HEAD(, BlockCopyTask) tasks;
  69
  70     BdrvRequestFlags write_flags;
  71
  72     /*
  73      * skip_unallocated:
  74      *
  75      * Used by sync=top jobs, which first scan the source node for unallocated
  76      * areas and clear them in the copy_bitmap.  During this process, the bitmap
  77      * is thus not fully initialized: It may still have bits set for areas that
  78      * are unallocated and should actually not be copied.
  79      *
  80      * This is indicated by skip_unallocated.
  81      *
  82      * In this case, block_copy() will query the source’s allocation status,
  83      * skip unallocated regions, clear them in the copy_bitmap, and invoke
  84      * block_copy_reset_unallocated() every time it does.
  85      */
  86     bool skip_unallocated;
  87
  88     ProgressMeter *progress;
  89     /* progress_bytes_callback: called when some copying progress is done. */
  90     ProgressBytesCallbackFunc progress_bytes_callback;
  91     void *progress_opaque;
  92
  93     SharedResource *mem;
  94 } BlockCopyState;
  95
  96 static BlockCopyTask *find_conflicting_task(BlockCopyState *s,
  97                                             int64_t offset, int64_t bytes)
  98 {
  99     BlockCopyTask *t;
 100
 101     QLIST_FOREACH(t, &s->tasks, list) {
 102         if (offset + bytes > t->offset && offset < t->offset + t->bytes) {
 103             return t;
 104         }
 105     }
 106
 107     return NULL;
 108 }
 109
 110 /*
 111  * If there are no intersecting tasks return false. Otherwise, wait for the
 112  * first found intersecting tasks to finish and return true.
 113  */
 114 static bool coroutine_fn block_copy_wait_one(BlockCopyState *s, int64_t offset,
 115                                              int64_t bytes)
 116 {
 117     BlockCopyTask *task = find_conflicting_task(s, offset, bytes);
 118
 119     if (!task) {
 120         return false;
 121     }
 122
 123     qemu_co_queue_wait(&task->wait_queue, NULL);
 124
 125     return true;
 126 }
 127
 128 /*
 129  * Search for the first dirty area in offset/bytes range and create task at
 130  * the beginning of it.
 131  */
 132 static BlockCopyTask *block_copy_task_create(BlockCopyState *s,
 133                                              BlockCopyCallState *call_state,
 134                                              int64_t offset, int64_t bytes)
 135 {
 136     BlockCopyTask *task;
 137
 138     if (!bdrv_dirty_bitmap_next_dirty_area(s->copy_bitmap,
 139                                            offset, offset + bytes,
 140                                            s->copy_size, &offset, &bytes))
 141     {
 142         return NULL;
 143     }
 144
 145     /* region is dirty, so no existent tasks possible in it */
 146     assert(!find_conflicting_task(s, offset, bytes));
 147
 148     bdrv_reset_dirty_bitmap(s->copy_bitmap, offset, bytes);
 149     s->in_flight_bytes += bytes;
 150
 151     task = g_new(BlockCopyTask, 1);
 152     *task = (BlockCopyTask) {
 153         .task.func = block_copy_task_entry,
 154         .s = s,
 155         .call_state = call_state,
 156         .offset = offset,
 157         .bytes = bytes,
 158     };
 159     qemu_co_queue_init(&task->wait_queue);
 160     QLIST_INSERT_HEAD(&s->tasks, task, list);
 161
 162     return task;
 163 }
 164
 165 /*
 166  * block_copy_task_shrink
 167  *
 168  * Drop the tail of the task to be handled later. Set dirty bits back and
 169  * wake up all tasks waiting for us (may be some of them are not intersecting
 170  * with shrunk task)
 171  */
 172 static void coroutine_fn block_copy_task_shrink(BlockCopyTask *task,
 173                                                 int64_t new_bytes)
 174 {
 175     if (new_bytes == task->bytes) {
 176         return;
 177     }
 178
 179     assert(new_bytes > 0 && new_bytes < task->bytes);
 180
 181     task->s->in_flight_bytes -= task->bytes - new_bytes;
 182     bdrv_set_dirty_bitmap(task->s->copy_bitmap,
 183                           task->offset + new_bytes, task->bytes - new_bytes);
 184
 185     task->bytes = new_bytes;
 186     qemu_co_queue_restart_all(&task->wait_queue);
 187 }
 188
 189 static void coroutine_fn block_copy_task_end(BlockCopyTask *task, int ret)
 190 {
 191     task->s->in_flight_bytes -= task->bytes;
 192     if (ret < 0) {
 193         bdrv_set_dirty_bitmap(task->s->copy_bitmap, task->offset, task->bytes);
 194     }
 195     QLIST_REMOVE(task, list);
 196     qemu_co_queue_restart_all(&task->wait_queue);
 197 }
 198
 199 void block_copy_state_free(BlockCopyState *s)
 200 {
 201     if (!s) {
 202         return;
 203     }
 204
 205     bdrv_release_dirty_bitmap(s->copy_bitmap);
 206     shres_destroy(s->mem);
 207     g_free(s);
 208 }
 209
 210 static uint32_t block_copy_max_transfer(BdrvChild *source, BdrvChild *target)
 211 {
 212     return MIN_NON_ZERO(INT_MAX,
 213                         MIN_NON_ZERO(source->bs->bl.max_transfer,
 214                                      target->bs->bl.max_transfer));
 215 }
 216
 217 BlockCopyState *block_copy_state_new(BdrvChild *source, BdrvChild *target,
 218                                      int64_t cluster_size,
 219                                      BdrvRequestFlags write_flags, Error **errp)
 220 {
 221     BlockCopyState *s;
 222     BdrvDirtyBitmap *copy_bitmap;
 223
 224     copy_bitmap = bdrv_create_dirty_bitmap(source->bs, cluster_size, NULL,
 225                                            errp);
 226     if (!copy_bitmap) {
 227         return NULL;
 228     }
 229     bdrv_disable_dirty_bitmap(copy_bitmap);
 230
 231     s = g_new(BlockCopyState, 1);
 232     *s = (BlockCopyState) {
 233         .source = source,
 234         .target = target,
 235         .copy_bitmap = copy_bitmap,
 236         .cluster_size = cluster_size,
 237         .len = bdrv_dirty_bitmap_size(copy_bitmap),
 238         .write_flags = write_flags,
 239         .mem = shres_create(BLOCK_COPY_MAX_MEM),
 240     };
 241
 242     if (block_copy_max_transfer(source, target) < cluster_size) {
 243         /*
 244          * copy_range does not respect max_transfer. We don't want to bother
 245          * with requests smaller than block-copy cluster size, so fallback to
 246          * buffered copying (read and write respect max_transfer on their
 247          * behalf).
 248          */
 249         s->use_copy_range = false;
 250         s->copy_size = cluster_size;
 251     } else if (write_flags & BDRV_REQ_WRITE_COMPRESSED) {
 252         /* Compression supports only cluster-size writes and no copy-range. */
 253         s->use_copy_range = false;
 254         s->copy_size = cluster_size;
 255     } else {
 256         /*
 257          * We enable copy-range, but keep small copy_size, until first
 258          * successful copy_range (look at block_copy_do_copy).
 259          */
 260         s->use_copy_range = true;
 261         s->copy_size = MAX(s->cluster_size, BLOCK_COPY_MAX_BUFFER);
 262     }
 263
 264     QLIST_INIT(&s->tasks);
 265
 266     return s;
 267 }
 268
 269 void block_copy_set_progress_callback(
 270         BlockCopyState *s,
 271         ProgressBytesCallbackFunc progress_bytes_callback,
 272         void *progress_opaque)
 273 {
 274     s->progress_bytes_callback = progress_bytes_callback;
 275     s->progress_opaque = progress_opaque;
 276 }
 277
 278 void block_copy_set_progress_meter(BlockCopyState *s, ProgressMeter *pm)
 279 {
 280     s->progress = pm;
 281 }
 282
 283 /*
 284  * Takes ownership of @task
 285  *
 286  * If pool is NULL directly run the task, otherwise schedule it into the pool.
 287  *
 288  * Returns: task.func return code if pool is NULL
 289  *          otherwise -ECANCELED if pool status is bad
 290  *          otherwise 0 (successfully scheduled)
 291  */
 292 static coroutine_fn int block_copy_task_run(AioTaskPool *pool,
 293                                             BlockCopyTask *task)
 294 {
 295     if (!pool) {
 296         int ret = task->task.func(&task->task);
 297
 298         g_free(task);
 299         return ret;
 300     }
 301
 302     aio_task_pool_wait_slot(pool);
 303     if (aio_task_pool_status(pool) < 0) {
 304         co_put_to_shres(task->s->mem, task->bytes);
 305         block_copy_task_end(task, -ECANCELED);
 306         g_free(task);
 307         return -ECANCELED;
 308     }
 309
 310     aio_task_pool_start_task(pool, &task->task);
 311
 312     return 0;
 313 }
 314
 315 /*
 316  * block_copy_do_copy
 317  *
 318  * Do copy of cluster-aligned chunk. Requested region is allowed to exceed
 319  * s->len only to cover last cluster when s->len is not aligned to clusters.
 320  *
 321  * No sync here: nor bitmap neighter intersecting requests handling, only copy.
 322  *
 323  * Returns 0 on success.
 324  */
 325 static int coroutine_fn block_copy_do_copy(BlockCopyState *s,
 326                                            int64_t offset, int64_t bytes,
 327                                            bool zeroes, bool *error_is_read)
 328 {
 329     int ret;
 330     int64_t nbytes = MIN(offset + bytes, s->len) - offset;
 331     void *bounce_buffer = NULL;
 332
 333     assert(offset >= 0 && bytes > 0 && INT64_MAX - offset >= bytes);
 334     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
 335     assert(QEMU_IS_ALIGNED(bytes, s->cluster_size));
 336     assert(offset < s->len);
 337     assert(offset + bytes <= s->len ||
 338            offset + bytes == QEMU_ALIGN_UP(s->len, s->cluster_size));
 339     assert(nbytes < INT_MAX);
 340
 341     if (zeroes) {
 342         ret = bdrv_co_pwrite_zeroes(s->target, offset, nbytes, s->write_flags &
 343                                     ~BDRV_REQ_WRITE_COMPRESSED);
 344         if (ret < 0) {
 345             trace_block_copy_write_zeroes_fail(s, offset, ret);
 346             *error_is_read = false;
 347         }
 348         return ret;
 349     }
 350
 351     if (s->use_copy_range) {
 352         ret = bdrv_co_copy_range(s->source, offset, s->target, offset, nbytes,
 353                                  0, s->write_flags);
 354         if (ret < 0) {
 355             trace_block_copy_copy_range_fail(s, offset, ret);
 356             s->use_copy_range = false;
 357             s->copy_size = MAX(s->cluster_size, BLOCK_COPY_MAX_BUFFER);
 358             /* Fallback to read+write with allocated buffer */
 359         } else {
 360             if (s->use_copy_range) {
 361                 /*
 362                  * Successful copy-range. Now increase copy_size.  copy_range
 363                  * does not respect max_transfer (it's a TODO), so we factor
 364                  * that in here.
 365                  *
 366                  * Note: we double-check s->use_copy_range for the case when
 367                  * parallel block-copy request unsets it during previous
 368                  * bdrv_co_copy_range call.
 369                  */
 370                 s->copy_size =
 371                         MIN(MAX(s->cluster_size, BLOCK_COPY_MAX_COPY_RANGE),
 372                             QEMU_ALIGN_DOWN(block_copy_max_transfer(s->source,
 373                                                                     s->target),
 374                                             s->cluster_size));
 375             }
 376             goto out;
 377         }
 378     }
 379
 380     /*
 381      * In case of failed copy_range request above, we may proceed with buffered
 382      * request larger than BLOCK_COPY_MAX_BUFFER. Still, further requests will
 383      * be properly limited, so don't care too much. Moreover the most likely
 384      * case (copy_range is unsupported for the configuration, so the very first
 385      * copy_range request fails) is handled by setting large copy_size only
 386      * after first successful copy_range.
 387      */
 388
 389     bounce_buffer = qemu_blockalign(s->source->bs, nbytes);
 390
 391     ret = bdrv_co_pread(s->source, offset, nbytes, bounce_buffer, 0);
 392     if (ret < 0) {
 393         trace_block_copy_read_fail(s, offset, ret);
 394         *error_is_read = true;
 395         goto out;
 396     }
 397
 398     ret = bdrv_co_pwrite(s->target, offset, nbytes, bounce_buffer,
 399                          s->write_flags);
 400     if (ret < 0) {
 401         trace_block_copy_write_fail(s, offset, ret);
 402         *error_is_read = false;
 403         goto out;
 404     }
 405
 406 out:
 407     qemu_vfree(bounce_buffer);
 408
 409     return ret;
 410 }
 411
 412 static coroutine_fn int block_copy_task_entry(AioTask *task)
 413 {
 414     BlockCopyTask *t = container_of(task, BlockCopyTask, task);
 415     bool error_is_read = false;
 416     int ret;
 417
 418     ret = block_copy_do_copy(t->s, t->offset, t->bytes, t->zeroes,
 419                              &error_is_read);
 420     if (ret < 0 && !t->call_state->failed) {
 421         t->call_state->failed = true;
 422         t->call_state->error_is_read = error_is_read;
 423     } else {
 424         progress_work_done(t->s->progress, t->bytes);
 425         t->s->progress_bytes_callback(t->bytes, t->s->progress_opaque);
 426     }
 427     co_put_to_shres(t->s->mem, t->bytes);
 428     block_copy_task_end(t, ret);
 429
 430     return ret;
 431 }
 432
 433 static int block_copy_block_status(BlockCopyState *s, int64_t offset,
 434                                    int64_t bytes, int64_t *pnum)
 435 {
 436     int64_t num;
 437     BlockDriverState *base;
 438     int ret;
 439
 440     if (s->skip_unallocated && s->source->bs->backing) {
 441         base = s->source->bs->backing->bs;
 442     } else {
 443         base = NULL;
 444     }
 445
 446     ret = bdrv_block_status_above(s->source->bs, base, offset, bytes, &num,
 447                                   NULL, NULL);
 448     if (ret < 0 || num < s->cluster_size) {
 449         /*
 450          * On error or if failed to obtain large enough chunk just fallback to
 451          * copy one cluster.
 452          */
 453         num = s->cluster_size;
 454         ret = BDRV_BLOCK_ALLOCATED | BDRV_BLOCK_DATA;
 455     } else if (offset + num == s->len) {
 456         num = QEMU_ALIGN_UP(num, s->cluster_size);
 457     } else {
 458         num = QEMU_ALIGN_DOWN(num, s->cluster_size);
 459     }
 460
 461     *pnum = num;
 462     return ret;
 463 }
 464
 465 /*
 466  * Check if the cluster starting at offset is allocated or not.
 467  * return via pnum the number of contiguous clusters sharing this allocation.
 468  */
 469 static int block_copy_is_cluster_allocated(BlockCopyState *s, int64_t offset,
 470                                            int64_t *pnum)
 471 {
 472     BlockDriverState *bs = s->source->bs;
 473     int64_t count, total_count = 0;
 474     int64_t bytes = s->len - offset;
 475     int ret;
 476
 477     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
 478
 479     while (true) {
 480         ret = bdrv_is_allocated(bs, offset, bytes, &count);
 481         if (ret < 0) {
 482             return ret;
 483         }
 484
 485         total_count += count;
 486
 487         if (ret || count == 0) {
 488             /*
 489              * ret: partial segment(s) are considered allocated.
 490              * otherwise: unallocated tail is treated as an entire segment.
 491              */
 492             *pnum = DIV_ROUND_UP(total_count, s->cluster_size);
 493             return ret;
 494         }
 495
 496         /* Unallocated segment(s) with uncertain following segment(s) */
 497         if (total_count >= s->cluster_size) {
 498             *pnum = total_count / s->cluster_size;
 499             return 0;
 500         }
 501
 502         offset += count;
 503         bytes -= count;
 504     }
 505 }
 506
 507 /*
 508  * Reset bits in copy_bitmap starting at offset if they represent unallocated
 509  * data in the image. May reset subsequent contiguous bits.
 510  * @return 0 when the cluster at @offset was unallocated,
 511  *         1 otherwise, and -ret on error.
 512  */
 513 int64_t block_copy_reset_unallocated(BlockCopyState *s,
 514                                      int64_t offset, int64_t *count)
 515 {
 516     int ret;
 517     int64_t clusters, bytes;
 518
 519     ret = block_copy_is_cluster_allocated(s, offset, &clusters);
 520     if (ret < 0) {
 521         return ret;
 522     }
 523
 524     bytes = clusters * s->cluster_size;
 525
 526     if (!ret) {
 527         bdrv_reset_dirty_bitmap(s->copy_bitmap, offset, bytes);
 528         progress_set_remaining(s->progress,
 529                                bdrv_get_dirty_count(s->copy_bitmap) +
 530                                s->in_flight_bytes);
 531     }
 532
 533     *count = bytes;
 534     return ret;
 535 }
 536
 537 /*
 538  * block_copy_dirty_clusters
 539  *
 540  * Copy dirty clusters in @offset/@bytes range.
 541  * Returns 1 if dirty clusters found and successfully copied, 0 if no dirty
 542  * clusters found and -errno on failure.
 543  */
 544 static int coroutine_fn block_copy_dirty_clusters(BlockCopyState *s,
 545                                                   int64_t offset, int64_t bytes,
 546                                                   bool *error_is_read)
 547 {
 548     int ret = 0;
 549     bool found_dirty = false;
 550     int64_t end = offset + bytes;
 551     AioTaskPool *aio = NULL;
 552     BlockCopyCallState call_state = {false, false};
 553
 554     /*
 555      * block_copy() user is responsible for keeping source and target in same
 556      * aio context
 557      */
 558     assert(bdrv_get_aio_context(s->source->bs) ==
 559            bdrv_get_aio_context(s->target->bs));
 560
 561     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
 562     assert(QEMU_IS_ALIGNED(bytes, s->cluster_size));
 563
 564     while (bytes && aio_task_pool_status(aio) == 0) {
 565         BlockCopyTask *task;
 566         int64_t status_bytes;
 567
 568         task = block_copy_task_create(s, &call_state, offset, bytes);
 569         if (!task) {
 570             /* No more dirty bits in the bitmap */
 571             trace_block_copy_skip_range(s, offset, bytes);
 572             break;
 573         }
 574         if (task->offset > offset) {
 575             trace_block_copy_skip_range(s, offset, task->offset - offset);
 576         }
 577
 578         found_dirty = true;
 579
 580         ret = block_copy_block_status(s, task->offset, task->bytes,
 581                                       &status_bytes);
 582         assert(ret >= 0); /* never fail */
 583         if (status_bytes < task->bytes) {
 584             block_copy_task_shrink(task, status_bytes);
 585         }
 586         if (s->skip_unallocated && !(ret & BDRV_BLOCK_ALLOCATED)) {
 587             block_copy_task_end(task, 0);
 588             progress_set_remaining(s->progress,
 589                                    bdrv_get_dirty_count(s->copy_bitmap) +
 590                                    s->in_flight_bytes);
 591             trace_block_copy_skip_range(s, task->offset, task->bytes);
 592             offset = task_end(task);
 593             bytes = end - offset;
 594             g_free(task);
 595             continue;
 596         }
 597         task->zeroes = ret & BDRV_BLOCK_ZERO;
 598
 599         trace_block_copy_process(s, task->offset);
 600
 601         co_get_from_shres(s->mem, task->bytes);
 602
 603         offset = task_end(task);
 604         bytes = end - offset;
 605
 606         if (!aio && bytes) {
 607             aio = aio_task_pool_new(BLOCK_COPY_MAX_WORKERS);
 608         }
 609
 610         ret = block_copy_task_run(aio, task);
 611         if (ret < 0) {
 612             goto out;
 613         }
 614     }
 615
 616 out:
 617     if (aio) {
 618         aio_task_pool_wait_all(aio);
 619
 620         /*
 621          * We are not really interested in -ECANCELED returned from
 622          * block_copy_task_run. If it fails, it means some task already failed
 623          * for real reason, let's return first failure.
 624          * Still, assert that we don't rewrite failure by success.
 625          */
 626         assert(ret == 0 || aio_task_pool_status(aio) < 0);
 627         ret = aio_task_pool_status(aio);
 628
 629         aio_task_pool_free(aio);
 630     }
 631     if (error_is_read && ret < 0) {
 632         *error_is_read = call_state.error_is_read;
 633     }
 634
 635     return ret < 0 ? ret : found_dirty;
 636 }
 637
 638 /*
 639  * block_copy
 640  *
 641  * Copy requested region, accordingly to dirty bitmap.
 642  * Collaborate with parallel block_copy requests: if they succeed it will help
 643  * us. If they fail, we will retry not-copied regions. So, if we return error,
 644  * it means that some I/O operation failed in context of _this_ block_copy call,
 645  * not some parallel operation.
 646  */
 647 int coroutine_fn block_copy(BlockCopyState *s, int64_t offset, int64_t bytes,
 648                             bool *error_is_read)
 649 {
 650     int ret;
 651
 652     do {
 653         ret = block_copy_dirty_clusters(s, offset, bytes, error_is_read);
 654
 655         if (ret == 0) {
 656             ret = block_copy_wait_one(s, offset, bytes);
 657         }
 658
 659         /*
 660          * We retry in two cases:
 661          * 1. Some progress done
 662          *    Something was copied, which means that there were yield points
 663          *    and some new dirty bits may have appeared (due to failed parallel
 664          *    block-copy requests).
 665          * 2. We have waited for some intersecting block-copy request
 666          *    It may have failed and produced new dirty bits.
 667          */
 668     } while (ret > 0);
 669
 670     return ret;
 671 }
 672
 673 BdrvDirtyBitmap *block_copy_dirty_bitmap(BlockCopyState *s)
 674 {
 675     return s->copy_bitmap;
 676 }
 677
 678 void block_copy_set_skip_unallocated(BlockCopyState *s, bool skip)
 679 {
 680     s->skip_unallocated = skip;
 681 }