block/copy-before-write.c

   1 /*
   2  * copy-before-write filter driver
   3  *
   4  * The driver performs Copy-Before-Write (CBW) operation: it is injected above
   5  * some node, and before each write it copies _old_ data to the target node.
   6  *
   7  * Copyright (c) 2018-2021 Virtuozzo International GmbH.
   8  *
   9  * Author:
  10  *  Sementsov-Ogievskiy Vladimir <vsementsov@virtuozzo.com>
  11  *
  12  * This program is free software; you can redistribute it and/or modify
  13  * it under the terms of the GNU General Public License as published by
  14  * the Free Software Foundation; either version 2 of the License, or
  15  * (at your option) any later version.
  16  *
  17  * This program is distributed in the hope that it will be useful,
  18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  * GNU General Public License for more details.
  21  *
  22  * You should have received a copy of the GNU General Public License
  23  * along with this program. If not, see <http://www.gnu.org/licenses/>.
  24  */
  25
  26 #include "qemu/osdep.h"
  27 #include "qapi/qmp/qjson.h"
  28
  29 #include "sysemu/block-backend.h"
  30 #include "qemu/cutils.h"
  31 #include "qapi/error.h"
  32 #include "block/block_int.h"
  33 #include "block/qdict.h"
  34 #include "block/block-copy.h"
  35 #include "block/dirty-bitmap.h"
  36
  37 #include "block/copy-before-write.h"
  38 #include "block/reqlist.h"
  39
  40 #include "qapi/qapi-visit-block-core.h"
  41
  42 typedef struct BDRVCopyBeforeWriteState {
  43     BlockCopyState *bcs;
  44     BdrvChild *target;
  45     OnCbwError on_cbw_error;
  46     uint32_t cbw_timeout_ns;
  47
  48     /*
  49      * @lock: protects access to @access_bitmap, @done_bitmap and
  50      * @frozen_read_reqs
  51      */
  52     CoMutex lock;
  53
  54     /*
  55      * @access_bitmap: represents areas allowed for reading by fleecing user.
  56      * Reading from non-dirty areas leads to -EACCES.
  57      */
  58     BdrvDirtyBitmap *access_bitmap;
  59
  60     /*
  61      * @done_bitmap: represents areas that was successfully copied to @target by
  62      * copy-before-write operations.
  63      */
  64     BdrvDirtyBitmap *done_bitmap;
  65
  66     /*
  67      * @frozen_read_reqs: current read requests for fleecing user in bs->file
  68      * node. These areas must not be rewritten by guest.
  69      */
  70     BlockReqList frozen_read_reqs;
  71
  72     /*
  73      * @snapshot_error is normally zero. But on first copy-before-write failure
  74      * when @on_cbw_error == ON_CBW_ERROR_BREAK_SNAPSHOT, @snapshot_error takes
  75      * value of this error (<0). After that all in-flight and further
  76      * snapshot-API requests will fail with that error.
  77      */
  78     int snapshot_error;
  79 } BDRVCopyBeforeWriteState;
  80
  81 static int coroutine_fn GRAPH_RDLOCK
  82 cbw_co_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
  83               QEMUIOVector *qiov, BdrvRequestFlags flags)
  84 {
  85     return bdrv_co_preadv(bs->file, offset, bytes, qiov, flags);
  86 }
  87
  88 static void block_copy_cb(void *opaque)
  89 {
  90     BlockDriverState *bs = opaque;
  91
  92     bdrv_dec_in_flight(bs);
  93 }
  94
  95 /*
  96  * Do copy-before-write operation.
  97  *
  98  * On failure guest request must be failed too.
  99  *
 100  * On success, we also wait for all in-flight fleecing read requests in source
 101  * node, and it's guaranteed that after cbw_do_copy_before_write() successful
 102  * return there are no such requests and they will never appear.
 103  */
 104 static coroutine_fn int cbw_do_copy_before_write(BlockDriverState *bs,
 105         uint64_t offset, uint64_t bytes, BdrvRequestFlags flags)
 106 {
 107     BDRVCopyBeforeWriteState *s = bs->opaque;
 108     int ret;
 109     uint64_t off, end;
 110     int64_t cluster_size = block_copy_cluster_size(s->bcs);
 111
 112     if (flags & BDRV_REQ_WRITE_UNCHANGED) {
 113         return 0;
 114     }
 115
 116     if (s->snapshot_error) {
 117         return 0;
 118     }
 119
 120     off = QEMU_ALIGN_DOWN(offset, cluster_size);
 121     end = QEMU_ALIGN_UP(offset + bytes, cluster_size);
 122
 123     /*
 124      * Increase in_flight, so that in case of timed-out block-copy, the
 125      * remaining background block_copy() request (which can't be immediately
 126      * cancelled by timeout) is presented in bs->in_flight. This way we are
 127      * sure that on bs close() we'll previously wait for all timed-out but yet
 128      * running block_copy calls.
 129      */
 130     bdrv_inc_in_flight(bs);
 131     ret = block_copy(s->bcs, off, end - off, true, s->cbw_timeout_ns,
 132                      block_copy_cb, bs);
 133     if (ret < 0 && s->on_cbw_error == ON_CBW_ERROR_BREAK_GUEST_WRITE) {
 134         return ret;
 135     }
 136
 137     WITH_QEMU_LOCK_GUARD(&s->lock) {
 138         if (ret < 0) {
 139             assert(s->on_cbw_error == ON_CBW_ERROR_BREAK_SNAPSHOT);
 140             if (!s->snapshot_error) {
 141                 s->snapshot_error = ret;
 142             }
 143         } else {
 144             bdrv_set_dirty_bitmap(s->done_bitmap, off, end - off);
 145         }
 146         reqlist_wait_all(&s->frozen_read_reqs, off, end - off, &s->lock);
 147     }
 148
 149     return 0;
 150 }
 151
 152 static int coroutine_fn GRAPH_RDLOCK
 153 cbw_co_pdiscard(BlockDriverState *bs, int64_t offset, int64_t bytes)
 154 {
 155     int ret = cbw_do_copy_before_write(bs, offset, bytes, 0);
 156     if (ret < 0) {
 157         return ret;
 158     }
 159
 160     return bdrv_co_pdiscard(bs->file, offset, bytes);
 161 }
 162
 163 static int coroutine_fn GRAPH_RDLOCK
 164 cbw_co_pwrite_zeroes(BlockDriverState *bs, int64_t offset, int64_t bytes,
 165                      BdrvRequestFlags flags)
 166 {
 167     int ret = cbw_do_copy_before_write(bs, offset, bytes, flags);
 168     if (ret < 0) {
 169         return ret;
 170     }
 171
 172     return bdrv_co_pwrite_zeroes(bs->file, offset, bytes, flags);
 173 }
 174
 175 static coroutine_fn GRAPH_RDLOCK
 176 int cbw_co_pwritev(BlockDriverState *bs, int64_t offset, int64_t bytes,
 177                    QEMUIOVector *qiov, BdrvRequestFlags flags)
 178 {
 179     int ret = cbw_do_copy_before_write(bs, offset, bytes, flags);
 180     if (ret < 0) {
 181         return ret;
 182     }
 183
 184     return bdrv_co_pwritev(bs->file, offset, bytes, qiov, flags);
 185 }
 186
 187 static int coroutine_fn GRAPH_RDLOCK cbw_co_flush(BlockDriverState *bs)
 188 {
 189     if (!bs->file) {
 190         return 0;
 191     }
 192
 193     return bdrv_co_flush(bs->file->bs);
 194 }
 195
 196 /*
 197  * If @offset not accessible - return NULL.
 198  *
 199  * Otherwise, set @pnum to some bytes that accessible from @file (@file is set
 200  * to bs->file or to s->target). Return newly allocated BlockReq object that
 201  * should be than passed to cbw_snapshot_read_unlock().
 202  *
 203  * It's guaranteed that guest writes will not interact in the region until
 204  * cbw_snapshot_read_unlock() called.
 205  */
 206 static coroutine_fn BlockReq *
 207 cbw_snapshot_read_lock(BlockDriverState *bs, int64_t offset, int64_t bytes,
 208                        int64_t *pnum, BdrvChild **file)
 209 {
 210     BDRVCopyBeforeWriteState *s = bs->opaque;
 211     BlockReq *req = g_new(BlockReq, 1);
 212     bool done;
 213
 214     QEMU_LOCK_GUARD(&s->lock);
 215
 216     if (s->snapshot_error) {
 217         g_free(req);
 218         return NULL;
 219     }
 220
 221     if (bdrv_dirty_bitmap_next_zero(s->access_bitmap, offset, bytes) != -1) {
 222         g_free(req);
 223         return NULL;
 224     }
 225
 226     done = bdrv_dirty_bitmap_status(s->done_bitmap, offset, bytes, pnum);
 227     if (done) {
 228         /*
 229          * Special invalid BlockReq, that is handled in
 230          * cbw_snapshot_read_unlock(). We don't need to lock something to read
 231          * from s->target.
 232          */
 233         *req = (BlockReq) {.offset = -1, .bytes = -1};
 234         *file = s->target;
 235     } else {
 236         reqlist_init_req(&s->frozen_read_reqs, req, offset, bytes);
 237         *file = bs->file;
 238     }
 239
 240     return req;
 241 }
 242
 243 static coroutine_fn void
 244 cbw_snapshot_read_unlock(BlockDriverState *bs, BlockReq *req)
 245 {
 246     BDRVCopyBeforeWriteState *s = bs->opaque;
 247
 248     if (req->offset == -1 && req->bytes == -1) {
 249         g_free(req);
 250         return;
 251     }
 252
 253     QEMU_LOCK_GUARD(&s->lock);
 254
 255     reqlist_remove_req(req);
 256     g_free(req);
 257 }
 258
 259 static int coroutine_fn GRAPH_RDLOCK
 260 cbw_co_preadv_snapshot(BlockDriverState *bs, int64_t offset, int64_t bytes,
 261                        QEMUIOVector *qiov, size_t qiov_offset)
 262 {
 263     BlockReq *req;
 264     BdrvChild *file;
 265     int ret;
 266
 267     /* TODO: upgrade to async loop using AioTask */
 268     while (bytes) {
 269         int64_t cur_bytes;
 270
 271         req = cbw_snapshot_read_lock(bs, offset, bytes, &cur_bytes, &file);
 272         if (!req) {
 273             return -EACCES;
 274         }
 275
 276         ret = bdrv_co_preadv_part(file, offset, cur_bytes,
 277                                   qiov, qiov_offset, 0);
 278         cbw_snapshot_read_unlock(bs, req);
 279         if (ret < 0) {
 280             return ret;
 281         }
 282
 283         bytes -= cur_bytes;
 284         offset += cur_bytes;
 285         qiov_offset += cur_bytes;
 286     }
 287
 288     return 0;
 289 }
 290
 291 static int coroutine_fn GRAPH_RDLOCK
 292 cbw_co_snapshot_block_status(BlockDriverState *bs,
 293                              bool want_zero, int64_t offset, int64_t bytes,
 294                              int64_t *pnum, int64_t *map,
 295                              BlockDriverState **file)
 296 {
 297     BDRVCopyBeforeWriteState *s = bs->opaque;
 298     BlockReq *req;
 299     int ret;
 300     int64_t cur_bytes;
 301     BdrvChild *child;
 302
 303     req = cbw_snapshot_read_lock(bs, offset, bytes, &cur_bytes, &child);
 304     if (!req) {
 305         return -EACCES;
 306     }
 307
 308     ret = bdrv_block_status(child->bs, offset, cur_bytes, pnum, map, file);
 309     if (child == s->target) {
 310         /*
 311          * We refer to s->target only for areas that we've written to it.
 312          * And we can not report unallocated blocks in s->target: this will
 313          * break generic block-status-above logic, that will go to
 314          * copy-before-write filtered child in this case.
 315          */
 316         assert(ret & BDRV_BLOCK_ALLOCATED);
 317     }
 318
 319     cbw_snapshot_read_unlock(bs, req);
 320
 321     return ret;
 322 }
 323
 324 static int coroutine_fn GRAPH_RDLOCK
 325 cbw_co_pdiscard_snapshot(BlockDriverState *bs, int64_t offset, int64_t bytes)
 326 {
 327     BDRVCopyBeforeWriteState *s = bs->opaque;
 328
 329     WITH_QEMU_LOCK_GUARD(&s->lock) {
 330         bdrv_reset_dirty_bitmap(s->access_bitmap, offset, bytes);
 331     }
 332
 333     block_copy_reset(s->bcs, offset, bytes);
 334
 335     return bdrv_co_pdiscard(s->target, offset, bytes);
 336 }
 337
 338 static void cbw_refresh_filename(BlockDriverState *bs)
 339 {
 340     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
 341             bs->file->bs->filename);
 342 }
 343
 344 static void GRAPH_RDLOCK
 345 cbw_child_perm(BlockDriverState *bs, BdrvChild *c, BdrvChildRole role,
 346                BlockReopenQueue *reopen_queue,
 347                uint64_t perm, uint64_t shared,
 348                uint64_t *nperm, uint64_t *nshared)
 349 {
 350     if (!(role & BDRV_CHILD_FILTERED)) {
 351         /*
 352          * Target child
 353          *
 354          * Share write to target (child_file), to not interfere
 355          * with guest writes to its disk which may be in target backing chain.
 356          * Can't resize during a backup block job because we check the size
 357          * only upfront.
 358          */
 359         *nshared = BLK_PERM_ALL & ~BLK_PERM_RESIZE;
 360         *nperm = BLK_PERM_WRITE;
 361     } else {
 362         /* Source child */
 363         bdrv_default_perms(bs, c, role, reopen_queue,
 364                            perm, shared, nperm, nshared);
 365
 366         if (!QLIST_EMPTY(&bs->parents)) {
 367             if (perm & BLK_PERM_WRITE) {
 368                 *nperm = *nperm | BLK_PERM_CONSISTENT_READ;
 369             }
 370             *nshared &= ~(BLK_PERM_WRITE | BLK_PERM_RESIZE);
 371         }
 372     }
 373 }
 374
 375 static BlockdevOptions *cbw_parse_options(QDict *options, Error **errp)
 376 {
 377     BlockdevOptions *opts = NULL;
 378     Visitor *v = NULL;
 379
 380     qdict_put_str(options, "driver", "copy-before-write");
 381
 382     v = qobject_input_visitor_new_flat_confused(options, errp);
 383     if (!v) {
 384         goto out;
 385     }
 386
 387     visit_type_BlockdevOptions(v, NULL, &opts, errp);
 388     if (!opts) {
 389         goto out;
 390     }
 391
 392     /*
 393      * Delete options which we are going to parse through BlockdevOptions
 394      * object for original options.
 395      */
 396     qdict_extract_subqdict(options, NULL, "bitmap");
 397     qdict_del(options, "on-cbw-error");
 398     qdict_del(options, "cbw-timeout");
 399
 400 out:
 401     visit_free(v);
 402     qdict_del(options, "driver");
 403
 404     return opts;
 405 }
 406
 407 static int cbw_open(BlockDriverState *bs, QDict *options, int flags,
 408                     Error **errp)
 409 {
 410     BDRVCopyBeforeWriteState *s = bs->opaque;
 411     BdrvDirtyBitmap *bitmap = NULL;
 412     int64_t cluster_size;
 413     g_autoptr(BlockdevOptions) full_opts = NULL;
 414     BlockdevOptionsCbw *opts;
 415     AioContext *ctx;
 416     int ret;
 417
 418     full_opts = cbw_parse_options(options, errp);
 419     if (!full_opts) {
 420         return -EINVAL;
 421     }
 422     assert(full_opts->driver == BLOCKDEV_DRIVER_COPY_BEFORE_WRITE);
 423     opts = &full_opts->u.copy_before_write;
 424
 425     ret = bdrv_open_file_child(NULL, options, "file", bs, errp);
 426     if (ret < 0) {
 427         return ret;
 428     }
 429
 430     s->target = bdrv_open_child(NULL, options, "target", bs, &child_of_bds,
 431                                 BDRV_CHILD_DATA, false, errp);
 432     if (!s->target) {
 433         return -EINVAL;
 434     }
 435
 436     ctx = bdrv_get_aio_context(bs);
 437     aio_context_acquire(ctx);
 438
 439     if (opts->bitmap) {
 440         bitmap = block_dirty_bitmap_lookup(opts->bitmap->node,
 441                                            opts->bitmap->name, NULL, errp);
 442         if (!bitmap) {
 443             ret = -EINVAL;
 444             goto out;
 445         }
 446     }
 447     s->on_cbw_error = opts->has_on_cbw_error ? opts->on_cbw_error :
 448             ON_CBW_ERROR_BREAK_GUEST_WRITE;
 449     s->cbw_timeout_ns = opts->has_cbw_timeout ?
 450         opts->cbw_timeout * NANOSECONDS_PER_SECOND : 0;
 451
 452     bs->total_sectors = bs->file->bs->total_sectors;
 453     bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED |
 454             (BDRV_REQ_FUA & bs->file->bs->supported_write_flags);
 455     bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
 456             ((BDRV_REQ_FUA | BDRV_REQ_MAY_UNMAP | BDRV_REQ_NO_FALLBACK) &
 457              bs->file->bs->supported_zero_flags);
 458
 459     s->bcs = block_copy_state_new(bs->file, s->target, bitmap, errp);
 460     if (!s->bcs) {
 461         error_prepend(errp, "Cannot create block-copy-state: ");
 462         ret = -EINVAL;
 463         goto out;
 464     }
 465
 466     cluster_size = block_copy_cluster_size(s->bcs);
 467
 468     s->done_bitmap = bdrv_create_dirty_bitmap(bs, cluster_size, NULL, errp);
 469     if (!s->done_bitmap) {
 470         ret = -EINVAL;
 471         goto out;
 472     }
 473     bdrv_disable_dirty_bitmap(s->done_bitmap);
 474
 475     /* s->access_bitmap starts equal to bcs bitmap */
 476     s->access_bitmap = bdrv_create_dirty_bitmap(bs, cluster_size, NULL, errp);
 477     if (!s->access_bitmap) {
 478         ret = -EINVAL;
 479         goto out;
 480     }
 481     bdrv_disable_dirty_bitmap(s->access_bitmap);
 482     bdrv_dirty_bitmap_merge_internal(s->access_bitmap,
 483                                      block_copy_dirty_bitmap(s->bcs), NULL,
 484                                      true);
 485
 486     qemu_co_mutex_init(&s->lock);
 487     QLIST_INIT(&s->frozen_read_reqs);
 488
 489     ret = 0;
 490 out:
 491     aio_context_release(ctx);
 492     return ret;
 493 }
 494
 495 static void cbw_close(BlockDriverState *bs)
 496 {
 497     BDRVCopyBeforeWriteState *s = bs->opaque;
 498
 499     bdrv_release_dirty_bitmap(s->access_bitmap);
 500     bdrv_release_dirty_bitmap(s->done_bitmap);
 501
 502     block_copy_state_free(s->bcs);
 503     s->bcs = NULL;
 504 }
 505
 506 static BlockDriver bdrv_cbw_filter = {
 507     .format_name = "copy-before-write",
 508     .instance_size = sizeof(BDRVCopyBeforeWriteState),
 509
 510     .bdrv_open                  = cbw_open,
 511     .bdrv_close                 = cbw_close,
 512
 513     .bdrv_co_preadv             = cbw_co_preadv,
 514     .bdrv_co_pwritev            = cbw_co_pwritev,
 515     .bdrv_co_pwrite_zeroes      = cbw_co_pwrite_zeroes,
 516     .bdrv_co_pdiscard           = cbw_co_pdiscard,
 517     .bdrv_co_flush              = cbw_co_flush,
 518
 519     .bdrv_co_preadv_snapshot       = cbw_co_preadv_snapshot,
 520     .bdrv_co_pdiscard_snapshot     = cbw_co_pdiscard_snapshot,
 521     .bdrv_co_snapshot_block_status = cbw_co_snapshot_block_status,
 522
 523     .bdrv_refresh_filename      = cbw_refresh_filename,
 524
 525     .bdrv_child_perm            = cbw_child_perm,
 526
 527     .is_filter = true,
 528 };
 529
 530 BlockDriverState *bdrv_cbw_append(BlockDriverState *source,
 531                                   BlockDriverState *target,
 532                                   const char *filter_node_name,
 533                                   BlockCopyState **bcs,
 534                                   Error **errp)
 535 {
 536     BDRVCopyBeforeWriteState *state;
 537     BlockDriverState *top;
 538     QDict *opts;
 539
 540     assert(source->total_sectors == target->total_sectors);
 541     GLOBAL_STATE_CODE();
 542
 543     opts = qdict_new();
 544     qdict_put_str(opts, "driver", "copy-before-write");
 545     if (filter_node_name) {
 546         qdict_put_str(opts, "node-name", filter_node_name);
 547     }
 548     qdict_put_str(opts, "file", bdrv_get_node_name(source));
 549     qdict_put_str(opts, "target", bdrv_get_node_name(target));
 550
 551     top = bdrv_insert_node(source, opts, BDRV_O_RDWR, errp);
 552     if (!top) {
 553         return NULL;
 554     }
 555
 556     state = top->opaque;
 557     *bcs = state->bcs;
 558
 559     return top;
 560 }
 561
 562 void bdrv_cbw_drop(BlockDriverState *bs)
 563 {
 564     GLOBAL_STATE_CODE();
 565     bdrv_drop_filter(bs, &error_abort);
 566     bdrv_unref(bs);
 567 }
 568
 569 static void cbw_init(void)
 570 {
 571     bdrv_register(&bdrv_cbw_filter);
 572 }
 573
 574 block_init(cbw_init);