module/zfs/dmu_send.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  */
  24 /*
  25  * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
  26  */
  27
  28 #include <sys/dmu.h>
  29 #include <sys/dmu_impl.h>
  30 #include <sys/dmu_tx.h>
  31 #include <sys/dbuf.h>
  32 #include <sys/dnode.h>
  33 #include <sys/zfs_context.h>
  34 #include <sys/dmu_objset.h>
  35 #include <sys/dmu_traverse.h>
  36 #include <sys/dsl_dataset.h>
  37 #include <sys/dsl_dir.h>
  38 #include <sys/dsl_prop.h>
  39 #include <sys/dsl_pool.h>
  40 #include <sys/dsl_synctask.h>
  41 #include <sys/zfs_ioctl.h>
  42 #include <sys/zap.h>
  43 #include <sys/zio_checksum.h>
  44 #include <sys/zfs_znode.h>
  45 #include <zfs_fletcher.h>
  46 #include <sys/avl.h>
  47 #include <sys/ddt.h>
  48 #include <sys/zfs_onexit.h>
  49
  50 static char *dmu_recv_tag = "dmu_recv_tag";
  51
  52 /*
  53  * The list of data whose inclusion in a send stream can be pending from
  54  * one call to backup_cb to another.  Multiple calls to dump_free() and
  55  * dump_freeobjects() can be aggregated into a single DRR_FREE or
  56  * DRR_FREEOBJECTS replay record.
  57  */
  58 typedef enum {
  59         PENDING_NONE,
  60         PENDING_FREE,
  61         PENDING_FREEOBJECTS
  62 } pendop_t;
  63
  64 struct backuparg {
  65         dmu_replay_record_t *drr;
  66         vnode_t *vp;
  67         offset_t *off;
  68         objset_t *os;
  69         zio_cksum_t zc;
  70         uint64_t toguid;
  71         int err;
  72         pendop_t pending_op;
  73 };
  74
  75 static int
  76 dump_bytes(struct backuparg *ba, void *buf, int len)
  77 {
  78         ssize_t resid; /* have to get resid to get detailed errno */
  79         ASSERT3U(len % 8, ==, 0);
  80
  81         fletcher_4_incremental_native(buf, len, &ba->zc);
  82         ba->err = vn_rdwr(UIO_WRITE, ba->vp,
  83             (caddr_t)buf, len,
  84             0, UIO_SYSSPACE, FAPPEND, RLIM64_INFINITY, CRED(), &resid);
  85         *ba->off += len;
  86         return (ba->err);
  87 }
  88
  89 static int
  90 dump_free(struct backuparg *ba, uint64_t object, uint64_t offset,
  91     uint64_t length)
  92 {
  93         struct drr_free *drrf = &(ba->drr->drr_u.drr_free);
  94
  95         /*
  96          * If there is a pending op, but it's not PENDING_FREE, push it out,
  97          * since free block aggregation can only be done for blocks of the
  98          * same type (i.e., DRR_FREE records can only be aggregated with
  99          * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
 100          * aggregated with other DRR_FREEOBJECTS records.
 101          */
 102         if (ba->pending_op != PENDING_NONE && ba->pending_op != PENDING_FREE) {
 103                 if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
 104                         return (EINTR);
 105                 ba->pending_op = PENDING_NONE;
 106         }
 107
 108         if (ba->pending_op == PENDING_FREE) {
 109                 /*
 110                  * There should never be a PENDING_FREE if length is -1
 111                  * (because dump_dnode is the only place where this
 112                  * function is called with a -1, and only after flushing
 113                  * any pending record).
 114                  */
 115                 ASSERT(length != -1ULL);
 116                 /*
 117                  * Check to see whether this free block can be aggregated
 118                  * with pending one.
 119                  */
 120                 if (drrf->drr_object == object && drrf->drr_offset +
 121                     drrf->drr_length == offset) {
 122                         drrf->drr_length += length;
 123                         return (0);
 124                 } else {
 125                         /* not a continuation.  Push out pending record */
 126                         if (dump_bytes(ba, ba->drr,
 127                             sizeof (dmu_replay_record_t)) != 0)
 128                                 return (EINTR);
 129                         ba->pending_op = PENDING_NONE;
 130                 }
 131         }
 132         /* create a FREE record and make it pending */
 133         bzero(ba->drr, sizeof (dmu_replay_record_t));
 134         ba->drr->drr_type = DRR_FREE;
 135         drrf->drr_object = object;
 136         drrf->drr_offset = offset;
 137         drrf->drr_length = length;
 138         drrf->drr_toguid = ba->toguid;
 139         if (length == -1ULL) {
 140                 if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
 141                         return (EINTR);
 142         } else {
 143                 ba->pending_op = PENDING_FREE;
 144         }
 145
 146         return (0);
 147 }
 148
 149 static int
 150 dump_data(struct backuparg *ba, dmu_object_type_t type,
 151     uint64_t object, uint64_t offset, int blksz, const blkptr_t *bp, void *data)
 152 {
 153         struct drr_write *drrw = &(ba->drr->drr_u.drr_write);
 154
 155
 156         /*
 157          * If there is any kind of pending aggregation (currently either
 158          * a grouping of free objects or free blocks), push it out to
 159          * the stream, since aggregation can't be done across operations
 160          * of different types.
 161          */
 162         if (ba->pending_op != PENDING_NONE) {
 163                 if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
 164                         return (EINTR);
 165                 ba->pending_op = PENDING_NONE;
 166         }
 167         /* write a DATA record */
 168         bzero(ba->drr, sizeof (dmu_replay_record_t));
 169         ba->drr->drr_type = DRR_WRITE;
 170         drrw->drr_object = object;
 171         drrw->drr_type = type;
 172         drrw->drr_offset = offset;
 173         drrw->drr_length = blksz;
 174         drrw->drr_toguid = ba->toguid;
 175         drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
 176         if (zio_checksum_table[drrw->drr_checksumtype].ci_dedup)
 177                 drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
 178         DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
 179         DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
 180         DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
 181         drrw->drr_key.ddk_cksum = bp->blk_cksum;
 182
 183         if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
 184                 return (EINTR);
 185         if (dump_bytes(ba, data, blksz) != 0)
 186                 return (EINTR);
 187         return (0);
 188 }
 189
 190 static int
 191 dump_spill(struct backuparg *ba, uint64_t object, int blksz, void *data)
 192 {
 193         struct drr_spill *drrs = &(ba->drr->drr_u.drr_spill);
 194
 195         if (ba->pending_op != PENDING_NONE) {
 196                 if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
 197                         return (EINTR);
 198                 ba->pending_op = PENDING_NONE;
 199         }
 200
 201         /* write a SPILL record */
 202         bzero(ba->drr, sizeof (dmu_replay_record_t));
 203         ba->drr->drr_type = DRR_SPILL;
 204         drrs->drr_object = object;
 205         drrs->drr_length = blksz;
 206         drrs->drr_toguid = ba->toguid;
 207
 208         if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)))
 209                 return (EINTR);
 210         if (dump_bytes(ba, data, blksz))
 211                 return (EINTR);
 212         return (0);
 213 }
 214
 215 static int
 216 dump_freeobjects(struct backuparg *ba, uint64_t firstobj, uint64_t numobjs)
 217 {
 218         struct drr_freeobjects *drrfo = &(ba->drr->drr_u.drr_freeobjects);
 219
 220         /*
 221          * If there is a pending op, but it's not PENDING_FREEOBJECTS,
 222          * push it out, since free block aggregation can only be done for
 223          * blocks of the same type (i.e., DRR_FREE records can only be
 224          * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
 225          * can only be aggregated with other DRR_FREEOBJECTS records.
 226          */
 227         if (ba->pending_op != PENDING_NONE &&
 228             ba->pending_op != PENDING_FREEOBJECTS) {
 229                 if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
 230                         return (EINTR);
 231                 ba->pending_op = PENDING_NONE;
 232         }
 233         if (ba->pending_op == PENDING_FREEOBJECTS) {
 234                 /*
 235                  * See whether this free object array can be aggregated
 236                  * with pending one
 237                  */
 238                 if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
 239                         drrfo->drr_numobjs += numobjs;
 240                         return (0);
 241                 } else {
 242                         /* can't be aggregated.  Push out pending record */
 243                         if (dump_bytes(ba, ba->drr,
 244                             sizeof (dmu_replay_record_t)) != 0)
 245                                 return (EINTR);
 246                         ba->pending_op = PENDING_NONE;
 247                 }
 248         }
 249
 250         /* write a FREEOBJECTS record */
 251         bzero(ba->drr, sizeof (dmu_replay_record_t));
 252         ba->drr->drr_type = DRR_FREEOBJECTS;
 253         drrfo->drr_firstobj = firstobj;
 254         drrfo->drr_numobjs = numobjs;
 255         drrfo->drr_toguid = ba->toguid;
 256
 257         ba->pending_op = PENDING_FREEOBJECTS;
 258
 259         return (0);
 260 }
 261
 262 static int
 263 dump_dnode(struct backuparg *ba, uint64_t object, dnode_phys_t *dnp)
 264 {
 265         struct drr_object *drro = &(ba->drr->drr_u.drr_object);
 266
 267         if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
 268                 return (dump_freeobjects(ba, object, 1));
 269
 270         if (ba->pending_op != PENDING_NONE) {
 271                 if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
 272                         return (EINTR);
 273                 ba->pending_op = PENDING_NONE;
 274         }
 275
 276         /* write an OBJECT record */
 277         bzero(ba->drr, sizeof (dmu_replay_record_t));
 278         ba->drr->drr_type = DRR_OBJECT;
 279         drro->drr_object = object;
 280         drro->drr_type = dnp->dn_type;
 281         drro->drr_bonustype = dnp->dn_bonustype;
 282         drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 283         drro->drr_bonuslen = dnp->dn_bonuslen;
 284         drro->drr_checksumtype = dnp->dn_checksum;
 285         drro->drr_compress = dnp->dn_compress;
 286         drro->drr_toguid = ba->toguid;
 287
 288         if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
 289                 return (EINTR);
 290
 291         if (dump_bytes(ba, DN_BONUS(dnp), P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0)
 292                 return (EINTR);
 293
 294         /* free anything past the end of the file */
 295         if (dump_free(ba, object, (dnp->dn_maxblkid + 1) *
 296             (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL))
 297                 return (EINTR);
 298         if (ba->err)
 299                 return (EINTR);
 300         return (0);
 301 }
 302
 303 #define BP_SPAN(dnp, level) \
 304         (((uint64_t)dnp->dn_datablkszsec) << (SPA_MINBLOCKSHIFT + \
 305         (level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)))
 306
 307 /* ARGSUSED */
 308 static int
 309 backup_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, arc_buf_t *pbuf,
 310     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
 311 {
 312         struct backuparg *ba = arg;
 313         dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
 314         int err = 0;
 315
 316         if (issig(JUSTLOOKING) && issig(FORREAL))
 317                 return (EINTR);
 318
 319         if (zb->zb_object != DMU_META_DNODE_OBJECT &&
 320             DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
 321                 return (0);
 322         } else if (bp == NULL && zb->zb_object == DMU_META_DNODE_OBJECT) {
 323                 uint64_t span = BP_SPAN(dnp, zb->zb_level);
 324                 uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
 325                 err = dump_freeobjects(ba, dnobj, span >> DNODE_SHIFT);
 326         } else if (bp == NULL) {
 327                 uint64_t span = BP_SPAN(dnp, zb->zb_level);
 328                 err = dump_free(ba, zb->zb_object, zb->zb_blkid * span, span);
 329         } else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
 330                 return (0);
 331         } else if (type == DMU_OT_DNODE) {
 332                 dnode_phys_t *blk;
 333                 int i;
 334                 int blksz = BP_GET_LSIZE(bp);
 335                 uint32_t aflags = ARC_WAIT;
 336                 arc_buf_t *abuf;
 337
 338                 if (dsl_read(NULL, spa, bp, pbuf,
 339                     arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
 340                     ZIO_FLAG_CANFAIL, &aflags, zb) != 0)
 341                         return (EIO);
 342
 343                 blk = abuf->b_data;
 344                 for (i = 0; i < blksz >> DNODE_SHIFT; i++) {
 345                         uint64_t dnobj = (zb->zb_blkid <<
 346                             (DNODE_BLOCK_SHIFT - DNODE_SHIFT)) + i;
 347                         err = dump_dnode(ba, dnobj, blk+i);
 348                         if (err)
 349                                 break;
 350                 }
 351                 (void) arc_buf_remove_ref(abuf, &abuf);
 352         } else if (type == DMU_OT_SA) {
 353                 uint32_t aflags = ARC_WAIT;
 354                 arc_buf_t *abuf;
 355                 int blksz = BP_GET_LSIZE(bp);
 356
 357                 if (arc_read_nolock(NULL, spa, bp,
 358                     arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
 359                     ZIO_FLAG_CANFAIL, &aflags, zb) != 0)
 360                         return (EIO);
 361
 362                 err = dump_spill(ba, zb->zb_object, blksz, abuf->b_data);
 363                 (void) arc_buf_remove_ref(abuf, &abuf);
 364         } else { /* it's a level-0 block of a regular object */
 365                 uint32_t aflags = ARC_WAIT;
 366                 arc_buf_t *abuf;
 367                 int blksz = BP_GET_LSIZE(bp);
 368
 369                 if (dsl_read(NULL, spa, bp, pbuf,
 370                     arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
 371                     ZIO_FLAG_CANFAIL, &aflags, zb) != 0)
 372                         return (EIO);
 373
 374                 err = dump_data(ba, type, zb->zb_object, zb->zb_blkid * blksz,
 375                     blksz, bp, abuf->b_data);
 376                 (void) arc_buf_remove_ref(abuf, &abuf);
 377         }
 378
 379         ASSERT(err == 0 || err == EINTR);
 380         return (err);
 381 }
 382
 383 int
 384 dmu_sendbackup(objset_t *tosnap, objset_t *fromsnap, boolean_t fromorigin,
 385     vnode_t *vp, offset_t *off)
 386 {
 387         dsl_dataset_t *ds = tosnap->os_dsl_dataset;
 388         dsl_dataset_t *fromds = fromsnap ? fromsnap->os_dsl_dataset : NULL;
 389         dmu_replay_record_t *drr;
 390         struct backuparg ba;
 391         int err;
 392         uint64_t fromtxg = 0;
 393
 394         /* tosnap must be a snapshot */
 395         if (ds->ds_phys->ds_next_snap_obj == 0)
 396                 return (EINVAL);
 397
 398         /* fromsnap must be an earlier snapshot from the same fs as tosnap */
 399         if (fromds && (ds->ds_dir != fromds->ds_dir ||
 400             fromds->ds_phys->ds_creation_txg >= ds->ds_phys->ds_creation_txg))
 401                 return (EXDEV);
 402
 403         if (fromorigin) {
 404                 dsl_pool_t *dp = ds->ds_dir->dd_pool;
 405
 406                 if (fromsnap)
 407                         return (EINVAL);
 408
 409                 if (dsl_dir_is_clone(ds->ds_dir)) {
 410                         rw_enter(&dp->dp_config_rwlock, RW_READER);
 411                         err = dsl_dataset_hold_obj(dp,
 412                             ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &fromds);
 413                         rw_exit(&dp->dp_config_rwlock);
 414                         if (err)
 415                                 return (err);
 416                 } else {
 417                         fromorigin = B_FALSE;
 418                 }
 419         }
 420
 421
 422         drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
 423         drr->drr_type = DRR_BEGIN;
 424         drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
 425         DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
 426             DMU_SUBSTREAM);
 427
 428 #ifdef _KERNEL
 429         if (dmu_objset_type(tosnap) == DMU_OST_ZFS) {
 430                 uint64_t version;
 431                 if (zfs_get_zplprop(tosnap, ZFS_PROP_VERSION, &version) != 0)
 432                         return (EINVAL);
 433                 if (version == ZPL_VERSION_SA) {
 434                         DMU_SET_FEATUREFLAGS(
 435                             drr->drr_u.drr_begin.drr_versioninfo,
 436                             DMU_BACKUP_FEATURE_SA_SPILL);
 437                 }
 438         }
 439 #endif
 440
 441         drr->drr_u.drr_begin.drr_creation_time =
 442             ds->ds_phys->ds_creation_time;
 443         drr->drr_u.drr_begin.drr_type = tosnap->os_phys->os_type;
 444         if (fromorigin)
 445                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
 446         drr->drr_u.drr_begin.drr_toguid = ds->ds_phys->ds_guid;
 447         if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
 448                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
 449
 450         if (fromds)
 451                 drr->drr_u.drr_begin.drr_fromguid = fromds->ds_phys->ds_guid;
 452         dsl_dataset_name(ds, drr->drr_u.drr_begin.drr_toname);
 453
 454         if (fromds)
 455                 fromtxg = fromds->ds_phys->ds_creation_txg;
 456         if (fromorigin)
 457                 dsl_dataset_rele(fromds, FTAG);
 458
 459         ba.drr = drr;
 460         ba.vp = vp;
 461         ba.os = tosnap;
 462         ba.off = off;
 463         ba.toguid = ds->ds_phys->ds_guid;
 464         ZIO_SET_CHECKSUM(&ba.zc, 0, 0, 0, 0);
 465         ba.pending_op = PENDING_NONE;
 466
 467         if (dump_bytes(&ba, drr, sizeof (dmu_replay_record_t)) != 0) {
 468                 kmem_free(drr, sizeof (dmu_replay_record_t));
 469                 return (ba.err);
 470         }
 471
 472         err = traverse_dataset(ds, fromtxg, TRAVERSE_PRE | TRAVERSE_PREFETCH,
 473             backup_cb, &ba);
 474
 475         if (ba.pending_op != PENDING_NONE)
 476                 if (dump_bytes(&ba, drr, sizeof (dmu_replay_record_t)) != 0)
 477                         err = EINTR;
 478
 479         if (err) {
 480                 if (err == EINTR && ba.err)
 481                         err = ba.err;
 482                 kmem_free(drr, sizeof (dmu_replay_record_t));
 483                 return (err);
 484         }
 485
 486         bzero(drr, sizeof (dmu_replay_record_t));
 487         drr->drr_type = DRR_END;
 488         drr->drr_u.drr_end.drr_checksum = ba.zc;
 489         drr->drr_u.drr_end.drr_toguid = ba.toguid;
 490
 491         if (dump_bytes(&ba, drr, sizeof (dmu_replay_record_t)) != 0) {
 492                 kmem_free(drr, sizeof (dmu_replay_record_t));
 493                 return (ba.err);
 494         }
 495
 496         kmem_free(drr, sizeof (dmu_replay_record_t));
 497
 498         return (0);
 499 }
 500
 501 struct recvbeginsyncarg {
 502         const char *tofs;
 503         const char *tosnap;
 504         dsl_dataset_t *origin;
 505         uint64_t fromguid;
 506         dmu_objset_type_t type;
 507         void *tag;
 508         boolean_t force;
 509         uint64_t dsflags;
 510         char clonelastname[MAXNAMELEN];
 511         dsl_dataset_t *ds; /* the ds to recv into; returned from the syncfunc */
 512         cred_t *cr;
 513 };
 514
 515 /* ARGSUSED */
 516 static int
 517 recv_new_check(void *arg1, void *arg2, dmu_tx_t *tx)
 518 {
 519         dsl_dir_t *dd = arg1;
 520         struct recvbeginsyncarg *rbsa = arg2;
 521         objset_t *mos = dd->dd_pool->dp_meta_objset;
 522         uint64_t val;
 523         int err;
 524
 525         err = zap_lookup(mos, dd->dd_phys->dd_child_dir_zapobj,
 526             strrchr(rbsa->tofs, '/') + 1, sizeof (uint64_t), 1, &val);
 527
 528         if (err != ENOENT)
 529                 return (err ? err : EEXIST);
 530
 531         if (rbsa->origin) {
 532                 /* make sure it's a snap in the same pool */
 533                 if (rbsa->origin->ds_dir->dd_pool != dd->dd_pool)
 534                         return (EXDEV);
 535                 if (!dsl_dataset_is_snapshot(rbsa->origin))
 536                         return (EINVAL);
 537                 if (rbsa->origin->ds_phys->ds_guid != rbsa->fromguid)
 538                         return (ENODEV);
 539         }
 540
 541         return (0);
 542 }
 543
 544 static void
 545 recv_new_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 546 {
 547         dsl_dir_t *dd = arg1;
 548         struct recvbeginsyncarg *rbsa = arg2;
 549         uint64_t flags = DS_FLAG_INCONSISTENT | rbsa->dsflags;
 550         uint64_t dsobj;
 551
 552         /* Create and open new dataset. */
 553         dsobj = dsl_dataset_create_sync(dd, strrchr(rbsa->tofs, '/') + 1,
 554             rbsa->origin, flags, rbsa->cr, tx);
 555         VERIFY(0 == dsl_dataset_own_obj(dd->dd_pool, dsobj,
 556             B_TRUE, dmu_recv_tag, &rbsa->ds));
 557
 558         if (rbsa->origin == NULL) {
 559                 (void) dmu_objset_create_impl(dd->dd_pool->dp_spa,
 560                     rbsa->ds, &rbsa->ds->ds_phys->ds_bp, rbsa->type, tx);
 561         }
 562
 563         spa_history_log_internal(LOG_DS_REPLAY_FULL_SYNC,
 564             dd->dd_pool->dp_spa, tx, "dataset = %lld", dsobj);
 565 }
 566
 567 /* ARGSUSED */
 568 static int
 569 recv_existing_check(void *arg1, void *arg2, dmu_tx_t *tx)
 570 {
 571         dsl_dataset_t *ds = arg1;
 572         struct recvbeginsyncarg *rbsa = arg2;
 573         int err;
 574         uint64_t val;
 575
 576         /* must not have any changes since most recent snapshot */
 577         if (!rbsa->force && dsl_dataset_modified_since_lastsnap(ds))
 578                 return (ETXTBSY);
 579
 580         /* new snapshot name must not exist */
 581         err = zap_lookup(ds->ds_dir->dd_pool->dp_meta_objset,
 582             ds->ds_phys->ds_snapnames_zapobj, rbsa->tosnap, 8, 1, &val);
 583         if (err == 0)
 584                 return (EEXIST);
 585         if (err != ENOENT)
 586                 return (err);
 587
 588         if (rbsa->fromguid) {
 589                 /* if incremental, most recent snapshot must match fromguid */
 590                 if (ds->ds_prev == NULL)
 591                         return (ENODEV);
 592
 593                 /*
 594                  * most recent snapshot must match fromguid, or there are no
 595                  * changes since the fromguid one
 596                  */
 597                 if (ds->ds_prev->ds_phys->ds_guid != rbsa->fromguid) {
 598                         uint64_t birth = ds->ds_prev->ds_phys->ds_bp.blk_birth;
 599                         uint64_t obj = ds->ds_prev->ds_phys->ds_prev_snap_obj;
 600                         while (obj != 0) {
 601                                 dsl_dataset_t *snap;
 602                                 err = dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
 603                                     obj, FTAG, &snap);
 604                                 if (err)
 605                                         return (ENODEV);
 606                                 if (snap->ds_phys->ds_creation_txg < birth) {
 607                                         dsl_dataset_rele(snap, FTAG);
 608                                         return (ENODEV);
 609                                 }
 610                                 if (snap->ds_phys->ds_guid == rbsa->fromguid) {
 611                                         dsl_dataset_rele(snap, FTAG);
 612                                         break; /* it's ok */
 613                                 }
 614                                 obj = snap->ds_phys->ds_prev_snap_obj;
 615                                 dsl_dataset_rele(snap, FTAG);
 616                         }
 617                         if (obj == 0)
 618                                 return (ENODEV);
 619                 }
 620         } else {
 621                 /* if full, most recent snapshot must be $ORIGIN */
 622                 if (ds->ds_phys->ds_prev_snap_txg >= TXG_INITIAL)
 623                         return (ENODEV);
 624         }
 625
 626         /* temporary clone name must not exist */
 627         err = zap_lookup(ds->ds_dir->dd_pool->dp_meta_objset,
 628             ds->ds_dir->dd_phys->dd_child_dir_zapobj,
 629             rbsa->clonelastname, 8, 1, &val);
 630         if (err == 0)
 631                 return (EEXIST);
 632         if (err != ENOENT)
 633                 return (err);
 634
 635         return (0);
 636 }
 637
 638 /* ARGSUSED */
 639 static void
 640 recv_existing_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 641 {
 642         dsl_dataset_t *ohds = arg1;
 643         struct recvbeginsyncarg *rbsa = arg2;
 644         dsl_pool_t *dp = ohds->ds_dir->dd_pool;
 645         dsl_dataset_t *cds;
 646         uint64_t flags = DS_FLAG_INCONSISTENT | rbsa->dsflags;
 647         uint64_t dsobj;
 648
 649         /* create and open the temporary clone */
 650         dsobj = dsl_dataset_create_sync(ohds->ds_dir, rbsa->clonelastname,
 651             ohds->ds_prev, flags, rbsa->cr, tx);
 652         VERIFY(0 == dsl_dataset_own_obj(dp, dsobj, B_TRUE, dmu_recv_tag, &cds));
 653
 654         /*
 655          * If we actually created a non-clone, we need to create the
 656          * objset in our new dataset.
 657          */
 658         if (BP_IS_HOLE(dsl_dataset_get_blkptr(cds))) {
 659                 (void) dmu_objset_create_impl(dp->dp_spa,
 660                     cds, dsl_dataset_get_blkptr(cds), rbsa->type, tx);
 661         }
 662
 663         rbsa->ds = cds;
 664
 665         spa_history_log_internal(LOG_DS_REPLAY_INC_SYNC,
 666             dp->dp_spa, tx, "dataset = %lld", dsobj);
 667 }
 668
 669 static boolean_t
 670 dmu_recv_verify_features(dsl_dataset_t *ds, struct drr_begin *drrb)
 671 {
 672         int featureflags;
 673
 674         featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
 675
 676         /* Verify pool version supports SA if SA_SPILL feature set */
 677         return ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
 678             (spa_version(dsl_dataset_get_spa(ds)) < SPA_VERSION_SA));
 679 }
 680
 681 /*
 682  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
 683  * succeeds; otherwise we will leak the holds on the datasets.
 684  */
 685 int
 686 dmu_recv_begin(char *tofs, char *tosnap, char *top_ds, struct drr_begin *drrb,
 687     boolean_t force, objset_t *origin, dmu_recv_cookie_t *drc)
 688 {
 689         int err = 0;
 690         boolean_t byteswap;
 691         struct recvbeginsyncarg rbsa = { 0 };
 692         uint64_t versioninfo;
 693         int flags;
 694         dsl_dataset_t *ds;
 695
 696         if (drrb->drr_magic == DMU_BACKUP_MAGIC)
 697                 byteswap = FALSE;
 698         else if (drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
 699                 byteswap = TRUE;
 700         else
 701                 return (EINVAL);
 702
 703         rbsa.tofs = tofs;
 704         rbsa.tosnap = tosnap;
 705         rbsa.origin = origin ? origin->os_dsl_dataset : NULL;
 706         rbsa.fromguid = drrb->drr_fromguid;
 707         rbsa.type = drrb->drr_type;
 708         rbsa.tag = FTAG;
 709         rbsa.dsflags = 0;
 710         rbsa.cr = CRED();
 711         versioninfo = drrb->drr_versioninfo;
 712         flags = drrb->drr_flags;
 713
 714         if (byteswap) {
 715                 rbsa.type = BSWAP_32(rbsa.type);
 716                 rbsa.fromguid = BSWAP_64(rbsa.fromguid);
 717                 versioninfo = BSWAP_64(versioninfo);
 718                 flags = BSWAP_32(flags);
 719         }
 720
 721         if (DMU_GET_STREAM_HDRTYPE(versioninfo) == DMU_COMPOUNDSTREAM ||
 722             rbsa.type >= DMU_OST_NUMTYPES ||
 723             ((flags & DRR_FLAG_CLONE) && origin == NULL))
 724                 return (EINVAL);
 725
 726         if (flags & DRR_FLAG_CI_DATA)
 727                 rbsa.dsflags = DS_FLAG_CI_DATASET;
 728
 729         bzero(drc, sizeof (dmu_recv_cookie_t));
 730         drc->drc_drrb = drrb;
 731         drc->drc_tosnap = tosnap;
 732         drc->drc_top_ds = top_ds;
 733         drc->drc_force = force;
 734
 735         /*
 736          * Process the begin in syncing context.
 737          */
 738
 739         /* open the dataset we are logically receiving into */
 740         err = dsl_dataset_hold(tofs, dmu_recv_tag, &ds);
 741         if (err == 0) {
 742                 if (dmu_recv_verify_features(ds, drrb)) {
 743                         dsl_dataset_rele(ds, dmu_recv_tag);
 744                         return (ENOTSUP);
 745                 }
 746                 /* target fs already exists; recv into temp clone */
 747
 748                 /* Can't recv a clone into an existing fs */
 749                 if (flags & DRR_FLAG_CLONE) {
 750                         dsl_dataset_rele(ds, dmu_recv_tag);
 751                         return (EINVAL);
 752                 }
 753
 754                 /* must not have an incremental recv already in progress */
 755                 if (!mutex_tryenter(&ds->ds_recvlock)) {
 756                         dsl_dataset_rele(ds, dmu_recv_tag);
 757                         return (EBUSY);
 758                 }
 759
 760                 /* tmp clone name is: tofs/%tosnap" */
 761                 (void) snprintf(rbsa.clonelastname, sizeof (rbsa.clonelastname),
 762                     "%%%s", tosnap);
 763                 rbsa.force = force;
 764                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
 765                     recv_existing_check, recv_existing_sync, ds, &rbsa, 5);
 766                 if (err) {
 767                         mutex_exit(&ds->ds_recvlock);
 768                         dsl_dataset_rele(ds, dmu_recv_tag);
 769                         return (err);
 770                 }
 771                 drc->drc_logical_ds = ds;
 772                 drc->drc_real_ds = rbsa.ds;
 773         } else if (err == ENOENT) {
 774                 /* target fs does not exist; must be a full backup or clone */
 775                 char *cp;
 776
 777                 /*
 778                  * If it's a non-clone incremental, we are missing the
 779                  * target fs, so fail the recv.
 780                  */
 781                 if (rbsa.fromguid && !(flags & DRR_FLAG_CLONE))
 782                         return (ENOENT);
 783
 784                 /* Open the parent of tofs */
 785                 cp = strrchr(tofs, '/');
 786                 *cp = '\0';
 787                 err = dsl_dataset_hold(tofs, FTAG, &ds);
 788                 *cp = '/';
 789                 if (err)
 790                         return (err);
 791
 792                 if (dmu_recv_verify_features(ds, drrb)) {
 793                         dsl_dataset_rele(ds, FTAG);
 794                         return (ENOTSUP);
 795                 }
 796
 797                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
 798                     recv_new_check, recv_new_sync, ds->ds_dir, &rbsa, 5);
 799                 dsl_dataset_rele(ds, FTAG);
 800                 if (err)
 801                         return (err);
 802                 drc->drc_logical_ds = drc->drc_real_ds = rbsa.ds;
 803                 drc->drc_newfs = B_TRUE;
 804         }
 805
 806         return (err);
 807 }
 808
 809 struct restorearg {
 810         int err;
 811         int byteswap;
 812         vnode_t *vp;
 813         char *buf;
 814         uint64_t voff;
 815         int bufsize; /* amount of memory allocated for buf */
 816         zio_cksum_t cksum;
 817         avl_tree_t *guid_to_ds_map;
 818 };
 819
 820 typedef struct guid_map_entry {
 821         uint64_t        guid;
 822         dsl_dataset_t   *gme_ds;
 823         avl_node_t      avlnode;
 824 } guid_map_entry_t;
 825
 826 static int
 827 guid_compare(const void *arg1, const void *arg2)
 828 {
 829         const guid_map_entry_t *gmep1 = arg1;
 830         const guid_map_entry_t *gmep2 = arg2;
 831
 832         if (gmep1->guid < gmep2->guid)
 833                 return (-1);
 834         else if (gmep1->guid > gmep2->guid)
 835                 return (1);
 836         return (0);
 837 }
 838
 839 static void
 840 free_guid_map_onexit(void *arg)
 841 {
 842         avl_tree_t *ca = arg;
 843         void *cookie = NULL;
 844         guid_map_entry_t *gmep;
 845
 846         while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
 847                 dsl_dataset_rele(gmep->gme_ds, ca);
 848                 kmem_free(gmep, sizeof (guid_map_entry_t));
 849         }
 850         avl_destroy(ca);
 851         kmem_free(ca, sizeof (avl_tree_t));
 852 }
 853
 854 static void *
 855 restore_read(struct restorearg *ra, int len)
 856 {
 857         void *rv;
 858         int done = 0;
 859
 860         /* some things will require 8-byte alignment, so everything must */
 861         ASSERT3U(len % 8, ==, 0);
 862
 863         while (done < len) {
 864                 ssize_t resid;
 865
 866                 ra->err = vn_rdwr(UIO_READ, ra->vp,
 867                     (caddr_t)ra->buf + done, len - done,
 868                     ra->voff, UIO_SYSSPACE, FAPPEND,
 869                     RLIM64_INFINITY, CRED(), &resid);
 870
 871                 if (resid == len - done)
 872                         ra->err = EINVAL;
 873                 ra->voff += len - done - resid;
 874                 done = len - resid;
 875                 if (ra->err)
 876                         return (NULL);
 877         }
 878
 879         ASSERT3U(done, ==, len);
 880         rv = ra->buf;
 881         if (ra->byteswap)
 882                 fletcher_4_incremental_byteswap(rv, len, &ra->cksum);
 883         else
 884                 fletcher_4_incremental_native(rv, len, &ra->cksum);
 885         return (rv);
 886 }
 887
 888 noinline static void
 889 backup_byteswap(dmu_replay_record_t *drr)
 890 {
 891 #define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
 892 #define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
 893         drr->drr_type = BSWAP_32(drr->drr_type);
 894         drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
 895         switch (drr->drr_type) {
 896         case DRR_BEGIN:
 897                 DO64(drr_begin.drr_magic);
 898                 DO64(drr_begin.drr_versioninfo);
 899                 DO64(drr_begin.drr_creation_time);
 900                 DO32(drr_begin.drr_type);
 901                 DO32(drr_begin.drr_flags);
 902                 DO64(drr_begin.drr_toguid);
 903                 DO64(drr_begin.drr_fromguid);
 904                 break;
 905         case DRR_OBJECT:
 906                 DO64(drr_object.drr_object);
 907                 /* DO64(drr_object.drr_allocation_txg); */
 908                 DO32(drr_object.drr_type);
 909                 DO32(drr_object.drr_bonustype);
 910                 DO32(drr_object.drr_blksz);
 911                 DO32(drr_object.drr_bonuslen);
 912                 DO64(drr_object.drr_toguid);
 913                 break;
 914         case DRR_FREEOBJECTS:
 915                 DO64(drr_freeobjects.drr_firstobj);
 916                 DO64(drr_freeobjects.drr_numobjs);
 917                 DO64(drr_freeobjects.drr_toguid);
 918                 break;
 919         case DRR_WRITE:
 920                 DO64(drr_write.drr_object);
 921                 DO32(drr_write.drr_type);
 922                 DO64(drr_write.drr_offset);
 923                 DO64(drr_write.drr_length);
 924                 DO64(drr_write.drr_toguid);
 925                 DO64(drr_write.drr_key.ddk_cksum.zc_word[0]);
 926                 DO64(drr_write.drr_key.ddk_cksum.zc_word[1]);
 927                 DO64(drr_write.drr_key.ddk_cksum.zc_word[2]);
 928                 DO64(drr_write.drr_key.ddk_cksum.zc_word[3]);
 929                 DO64(drr_write.drr_key.ddk_prop);
 930                 break;
 931         case DRR_WRITE_BYREF:
 932                 DO64(drr_write_byref.drr_object);
 933                 DO64(drr_write_byref.drr_offset);
 934                 DO64(drr_write_byref.drr_length);
 935                 DO64(drr_write_byref.drr_toguid);
 936                 DO64(drr_write_byref.drr_refguid);
 937                 DO64(drr_write_byref.drr_refobject);
 938                 DO64(drr_write_byref.drr_refoffset);
 939                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[0]);
 940                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[1]);
 941                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[2]);
 942                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[3]);
 943                 DO64(drr_write_byref.drr_key.ddk_prop);
 944                 break;
 945         case DRR_FREE:
 946                 DO64(drr_free.drr_object);
 947                 DO64(drr_free.drr_offset);
 948                 DO64(drr_free.drr_length);
 949                 DO64(drr_free.drr_toguid);
 950                 break;
 951         case DRR_SPILL:
 952                 DO64(drr_spill.drr_object);
 953                 DO64(drr_spill.drr_length);
 954                 DO64(drr_spill.drr_toguid);
 955                 break;
 956         case DRR_END:
 957                 DO64(drr_end.drr_checksum.zc_word[0]);
 958                 DO64(drr_end.drr_checksum.zc_word[1]);
 959                 DO64(drr_end.drr_checksum.zc_word[2]);
 960                 DO64(drr_end.drr_checksum.zc_word[3]);
 961                 DO64(drr_end.drr_toguid);
 962                 break;
 963         default:
 964                 break;
 965         }
 966 #undef DO64
 967 #undef DO32
 968 }
 969
 970 noinline static int
 971 restore_object(struct restorearg *ra, objset_t *os, struct drr_object *drro)
 972 {
 973         int err;
 974         dmu_tx_t *tx;
 975         void *data = NULL;
 976
 977         if (drro->drr_type == DMU_OT_NONE ||
 978             drro->drr_type >= DMU_OT_NUMTYPES ||
 979             drro->drr_bonustype >= DMU_OT_NUMTYPES ||
 980             drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
 981             drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
 982             P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
 983             drro->drr_blksz < SPA_MINBLOCKSIZE ||
 984             drro->drr_blksz > SPA_MAXBLOCKSIZE ||
 985             drro->drr_bonuslen > DN_MAX_BONUSLEN) {
 986                 return (EINVAL);
 987         }
 988
 989         err = dmu_object_info(os, drro->drr_object, NULL);
 990
 991         if (err != 0 && err != ENOENT)
 992                 return (EINVAL);
 993
 994         if (drro->drr_bonuslen) {
 995                 data = restore_read(ra, P2ROUNDUP(drro->drr_bonuslen, 8));
 996                 if (ra->err)
 997                         return (ra->err);
 998         }
 999
1000         if (err == ENOENT) {
1001                 /* currently free, want to be allocated */
1002                 tx = dmu_tx_create(os);
1003                 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1004                 err = dmu_tx_assign(tx, TXG_WAIT);
1005                 if (err) {
1006                         dmu_tx_abort(tx);
1007                         return (err);
1008                 }
1009                 err = dmu_object_claim(os, drro->drr_object,
1010                     drro->drr_type, drro->drr_blksz,
1011                     drro->drr_bonustype, drro->drr_bonuslen, tx);
1012                 dmu_tx_commit(tx);
1013         } else {
1014                 /* currently allocated, want to be allocated */
1015                 err = dmu_object_reclaim(os, drro->drr_object,
1016                     drro->drr_type, drro->drr_blksz,
1017                     drro->drr_bonustype, drro->drr_bonuslen);
1018         }
1019         if (err) {
1020                 return (EINVAL);
1021         }
1022
1023         tx = dmu_tx_create(os);
1024         dmu_tx_hold_bonus(tx, drro->drr_object);
1025         err = dmu_tx_assign(tx, TXG_WAIT);
1026         if (err) {
1027                 dmu_tx_abort(tx);
1028                 return (err);
1029         }
1030
1031         dmu_object_set_checksum(os, drro->drr_object, drro->drr_checksumtype,
1032             tx);
1033         dmu_object_set_compress(os, drro->drr_object, drro->drr_compress, tx);
1034
1035         if (data != NULL) {
1036                 dmu_buf_t *db;
1037
1038                 VERIFY(0 == dmu_bonus_hold(os, drro->drr_object, FTAG, &db));
1039                 dmu_buf_will_dirty(db, tx);
1040
1041                 ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
1042                 bcopy(data, db->db_data, drro->drr_bonuslen);
1043                 if (ra->byteswap) {
1044                         dmu_ot[drro->drr_bonustype].ot_byteswap(db->db_data,
1045                             drro->drr_bonuslen);
1046                 }
1047                 dmu_buf_rele(db, FTAG);
1048         }
1049         dmu_tx_commit(tx);
1050         return (0);
1051 }
1052
1053 /* ARGSUSED */
1054 noinline static int
1055 restore_freeobjects(struct restorearg *ra, objset_t *os,
1056     struct drr_freeobjects *drrfo)
1057 {
1058         uint64_t obj;
1059
1060         if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
1061                 return (EINVAL);
1062
1063         for (obj = drrfo->drr_firstobj;
1064             obj < drrfo->drr_firstobj + drrfo->drr_numobjs;
1065             (void) dmu_object_next(os, &obj, FALSE, 0)) {
1066                 int err;
1067
1068                 if (dmu_object_info(os, obj, NULL) != 0)
1069                         continue;
1070
1071                 err = dmu_free_object(os, obj);
1072                 if (err)
1073                         return (err);
1074         }
1075         return (0);
1076 }
1077
1078 noinline static int
1079 restore_write(struct restorearg *ra, objset_t *os,
1080     struct drr_write *drrw)
1081 {
1082         dmu_tx_t *tx;
1083         void *data;
1084         int err;
1085
1086         if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
1087             drrw->drr_type >= DMU_OT_NUMTYPES)
1088                 return (EINVAL);
1089
1090         data = restore_read(ra, drrw->drr_length);
1091         if (data == NULL)
1092                 return (ra->err);
1093
1094         if (dmu_object_info(os, drrw->drr_object, NULL) != 0)
1095                 return (EINVAL);
1096
1097         tx = dmu_tx_create(os);
1098
1099         dmu_tx_hold_write(tx, drrw->drr_object,
1100             drrw->drr_offset, drrw->drr_length);
1101         err = dmu_tx_assign(tx, TXG_WAIT);
1102         if (err) {
1103                 dmu_tx_abort(tx);
1104                 return (err);
1105         }
1106         if (ra->byteswap)
1107                 dmu_ot[drrw->drr_type].ot_byteswap(data, drrw->drr_length);
1108         dmu_write(os, drrw->drr_object,
1109             drrw->drr_offset, drrw->drr_length, data, tx);
1110         dmu_tx_commit(tx);
1111         return (0);
1112 }
1113
1114 /*
1115  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
1116  * streams to refer to a copy of the data that is already on the
1117  * system because it came in earlier in the stream.  This function
1118  * finds the earlier copy of the data, and uses that copy instead of
1119  * data from the stream to fulfill this write.
1120  */
1121 static int
1122 restore_write_byref(struct restorearg *ra, objset_t *os,
1123     struct drr_write_byref *drrwbr)
1124 {
1125         dmu_tx_t *tx;
1126         int err;
1127         guid_map_entry_t gmesrch;
1128         guid_map_entry_t *gmep;
1129         avl_index_t     where;
1130         objset_t *ref_os = NULL;
1131         dmu_buf_t *dbp;
1132
1133         if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
1134                 return (EINVAL);
1135
1136         /*
1137          * If the GUID of the referenced dataset is different from the
1138          * GUID of the target dataset, find the referenced dataset.
1139          */
1140         if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
1141                 gmesrch.guid = drrwbr->drr_refguid;
1142                 if ((gmep = avl_find(ra->guid_to_ds_map, &gmesrch,
1143                     &where)) == NULL) {
1144                         return (EINVAL);
1145                 }
1146                 if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
1147                         return (EINVAL);
1148         } else {
1149                 ref_os = os;
1150         }
1151
1152         err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
1153             drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH);
1154         if (err)
1155                 return (err);
1156
1157         tx = dmu_tx_create(os);
1158
1159         dmu_tx_hold_write(tx, drrwbr->drr_object,
1160             drrwbr->drr_offset, drrwbr->drr_length);
1161         err = dmu_tx_assign(tx, TXG_WAIT);
1162         if (err) {
1163                 dmu_tx_abort(tx);
1164                 return (err);
1165         }
1166         dmu_write(os, drrwbr->drr_object,
1167             drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
1168         dmu_buf_rele(dbp, FTAG);
1169         dmu_tx_commit(tx);
1170         return (0);
1171 }
1172
1173 static int
1174 restore_spill(struct restorearg *ra, objset_t *os, struct drr_spill *drrs)
1175 {
1176         dmu_tx_t *tx;
1177         void *data;
1178         dmu_buf_t *db, *db_spill;
1179         int err;
1180
1181         if (drrs->drr_length < SPA_MINBLOCKSIZE ||
1182             drrs->drr_length > SPA_MAXBLOCKSIZE)
1183                 return (EINVAL);
1184
1185         data = restore_read(ra, drrs->drr_length);
1186         if (data == NULL)
1187                 return (ra->err);
1188
1189         if (dmu_object_info(os, drrs->drr_object, NULL) != 0)
1190                 return (EINVAL);
1191
1192         VERIFY(0 == dmu_bonus_hold(os, drrs->drr_object, FTAG, &db));
1193         if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
1194                 dmu_buf_rele(db, FTAG);
1195                 return (err);
1196         }
1197
1198         tx = dmu_tx_create(os);
1199
1200         dmu_tx_hold_spill(tx, db->db_object);
1201
1202         err = dmu_tx_assign(tx, TXG_WAIT);
1203         if (err) {
1204                 dmu_buf_rele(db, FTAG);
1205                 dmu_buf_rele(db_spill, FTAG);
1206                 dmu_tx_abort(tx);
1207                 return (err);
1208         }
1209         dmu_buf_will_dirty(db_spill, tx);
1210
1211         if (db_spill->db_size < drrs->drr_length)
1212                 VERIFY(0 == dbuf_spill_set_blksz(db_spill,
1213                     drrs->drr_length, tx));
1214         bcopy(data, db_spill->db_data, drrs->drr_length);
1215
1216         dmu_buf_rele(db, FTAG);
1217         dmu_buf_rele(db_spill, FTAG);
1218
1219         dmu_tx_commit(tx);
1220         return (0);
1221 }
1222
1223 /* ARGSUSED */
1224 noinline static int
1225 restore_free(struct restorearg *ra, objset_t *os,
1226     struct drr_free *drrf)
1227 {
1228         int err;
1229
1230         if (drrf->drr_length != -1ULL &&
1231             drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
1232                 return (EINVAL);
1233
1234         if (dmu_object_info(os, drrf->drr_object, NULL) != 0)
1235                 return (EINVAL);
1236
1237         err = dmu_free_long_range(os, drrf->drr_object,
1238             drrf->drr_offset, drrf->drr_length);
1239         return (err);
1240 }
1241
1242 /*
1243  * NB: callers *must* call dmu_recv_end() if this succeeds.
1244  */
1245 int
1246 dmu_recv_stream(dmu_recv_cookie_t *drc, vnode_t *vp, offset_t *voffp,
1247     int cleanup_fd, uint64_t *action_handlep)
1248 {
1249         struct restorearg ra = { 0 };
1250         dmu_replay_record_t *drr;
1251         objset_t *os;
1252         zio_cksum_t pcksum;
1253         int featureflags;
1254
1255         if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
1256                 ra.byteswap = TRUE;
1257
1258         {
1259                 /* compute checksum of drr_begin record */
1260                 dmu_replay_record_t *drr;
1261                 drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
1262
1263                 drr->drr_type = DRR_BEGIN;
1264                 drr->drr_u.drr_begin = *drc->drc_drrb;
1265                 if (ra.byteswap) {
1266                         fletcher_4_incremental_byteswap(drr,
1267                             sizeof (dmu_replay_record_t), &ra.cksum);
1268                 } else {
1269                         fletcher_4_incremental_native(drr,
1270                             sizeof (dmu_replay_record_t), &ra.cksum);
1271                 }
1272                 kmem_free(drr, sizeof (dmu_replay_record_t));
1273         }
1274
1275         if (ra.byteswap) {
1276                 struct drr_begin *drrb = drc->drc_drrb;
1277                 drrb->drr_magic = BSWAP_64(drrb->drr_magic);
1278                 drrb->drr_versioninfo = BSWAP_64(drrb->drr_versioninfo);
1279                 drrb->drr_creation_time = BSWAP_64(drrb->drr_creation_time);
1280                 drrb->drr_type = BSWAP_32(drrb->drr_type);
1281                 drrb->drr_toguid = BSWAP_64(drrb->drr_toguid);
1282                 drrb->drr_fromguid = BSWAP_64(drrb->drr_fromguid);
1283         }
1284
1285         ra.vp = vp;
1286         ra.voff = *voffp;
1287         ra.bufsize = 1<<20;
1288         ra.buf = vmem_alloc(ra.bufsize, KM_SLEEP);
1289
1290         /* these were verified in dmu_recv_begin */
1291         ASSERT(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo) ==
1292             DMU_SUBSTREAM);
1293         ASSERT(drc->drc_drrb->drr_type < DMU_OST_NUMTYPES);
1294
1295         /*
1296          * Open the objset we are modifying.
1297          */
1298         VERIFY(dmu_objset_from_ds(drc->drc_real_ds, &os) == 0);
1299
1300         ASSERT(drc->drc_real_ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT);
1301
1302         featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
1303
1304         /* if this stream is dedup'ed, set up the avl tree for guid mapping */
1305         if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
1306                 minor_t minor;
1307
1308                 if (cleanup_fd == -1) {
1309                         ra.err = EBADF;
1310                         goto out;
1311                 }
1312                 ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
1313                 if (ra.err) {
1314                         cleanup_fd = -1;
1315                         goto out;
1316                 }
1317
1318                 if (*action_handlep == 0) {
1319                         ra.guid_to_ds_map =
1320                             kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
1321                         avl_create(ra.guid_to_ds_map, guid_compare,
1322                             sizeof (guid_map_entry_t),
1323                             offsetof(guid_map_entry_t, avlnode));
1324                         ra.err = zfs_onexit_add_cb(minor,
1325                             free_guid_map_onexit, ra.guid_to_ds_map,
1326                             action_handlep);
1327                         if (ra.err)
1328                                 goto out;
1329                 } else {
1330                         ra.err = zfs_onexit_cb_data(minor, *action_handlep,
1331                             (void **)&ra.guid_to_ds_map);
1332                         if (ra.err)
1333                                 goto out;
1334                 }
1335
1336                 drc->drc_guid_to_ds_map = ra.guid_to_ds_map;
1337         }
1338
1339         /*
1340          * Read records and process them.
1341          */
1342         pcksum = ra.cksum;
1343         while (ra.err == 0 &&
1344             NULL != (drr = restore_read(&ra, sizeof (*drr)))) {
1345                 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1346                         ra.err = EINTR;
1347                         goto out;
1348                 }
1349
1350                 if (ra.byteswap)
1351                         backup_byteswap(drr);
1352
1353                 switch (drr->drr_type) {
1354                 case DRR_OBJECT:
1355                 {
1356                         /*
1357                          * We need to make a copy of the record header,
1358                          * because restore_{object,write} may need to
1359                          * restore_read(), which will invalidate drr.
1360                          */
1361                         struct drr_object drro = drr->drr_u.drr_object;
1362                         ra.err = restore_object(&ra, os, &drro);
1363                         break;
1364                 }
1365                 case DRR_FREEOBJECTS:
1366                 {
1367                         struct drr_freeobjects drrfo =
1368                             drr->drr_u.drr_freeobjects;
1369                         ra.err = restore_freeobjects(&ra, os, &drrfo);
1370                         break;
1371                 }
1372                 case DRR_WRITE:
1373                 {
1374                         struct drr_write drrw = drr->drr_u.drr_write;
1375                         ra.err = restore_write(&ra, os, &drrw);
1376                         break;
1377                 }
1378                 case DRR_WRITE_BYREF:
1379                 {
1380                         struct drr_write_byref drrwbr =
1381                             drr->drr_u.drr_write_byref;
1382                         ra.err = restore_write_byref(&ra, os, &drrwbr);
1383                         break;
1384                 }
1385                 case DRR_FREE:
1386                 {
1387                         struct drr_free drrf = drr->drr_u.drr_free;
1388                         ra.err = restore_free(&ra, os, &drrf);
1389                         break;
1390                 }
1391                 case DRR_END:
1392                 {
1393                         struct drr_end drre = drr->drr_u.drr_end;
1394                         /*
1395                          * We compare against the *previous* checksum
1396                          * value, because the stored checksum is of
1397                          * everything before the DRR_END record.
1398                          */
1399                         if (!ZIO_CHECKSUM_EQUAL(drre.drr_checksum, pcksum))
1400                                 ra.err = ECKSUM;
1401                         goto out;
1402                 }
1403                 case DRR_SPILL:
1404                 {
1405                         struct drr_spill drrs = drr->drr_u.drr_spill;
1406                         ra.err = restore_spill(&ra, os, &drrs);
1407                         break;
1408                 }
1409                 default:
1410                         ra.err = EINVAL;
1411                         goto out;
1412                 }
1413                 pcksum = ra.cksum;
1414         }
1415         ASSERT(ra.err != 0);
1416
1417 out:
1418         if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
1419                 zfs_onexit_fd_rele(cleanup_fd);
1420
1421         if (ra.err != 0) {
1422                 /*
1423                  * destroy what we created, so we don't leave it in the
1424                  * inconsistent restoring state.
1425                  */
1426                 txg_wait_synced(drc->drc_real_ds->ds_dir->dd_pool, 0);
1427
1428                 (void) dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag,
1429                     B_FALSE);
1430                 if (drc->drc_real_ds != drc->drc_logical_ds) {
1431                         mutex_exit(&drc->drc_logical_ds->ds_recvlock);
1432                         dsl_dataset_rele(drc->drc_logical_ds, dmu_recv_tag);
1433                 }
1434         }
1435
1436         vmem_free(ra.buf, ra.bufsize);
1437         *voffp = ra.voff;
1438         return (ra.err);
1439 }
1440
1441 struct recvendsyncarg {
1442         char *tosnap;
1443         uint64_t creation_time;
1444         uint64_t toguid;
1445 };
1446
1447 static int
1448 recv_end_check(void *arg1, void *arg2, dmu_tx_t *tx)
1449 {
1450         dsl_dataset_t *ds = arg1;
1451         struct recvendsyncarg *resa = arg2;
1452
1453         return (dsl_dataset_snapshot_check(ds, resa->tosnap, tx));
1454 }
1455
1456 static void
1457 recv_end_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1458 {
1459         dsl_dataset_t *ds = arg1;
1460         struct recvendsyncarg *resa = arg2;
1461
1462         dsl_dataset_snapshot_sync(ds, resa->tosnap, tx);
1463
1464         /* set snapshot's creation time and guid */
1465         dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
1466         ds->ds_prev->ds_phys->ds_creation_time = resa->creation_time;
1467         ds->ds_prev->ds_phys->ds_guid = resa->toguid;
1468         ds->ds_prev->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1469
1470         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1471         ds->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1472 }
1473
1474 static int
1475 add_ds_to_guidmap(avl_tree_t *guid_map, dsl_dataset_t *ds)
1476 {
1477         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1478         uint64_t snapobj = ds->ds_phys->ds_prev_snap_obj;
1479         dsl_dataset_t *snapds;
1480         guid_map_entry_t *gmep;
1481         int err;
1482
1483         ASSERT(guid_map != NULL);
1484
1485         rw_enter(&dp->dp_config_rwlock, RW_READER);
1486         err = dsl_dataset_hold_obj(dp, snapobj, guid_map, &snapds);
1487         if (err == 0) {
1488                 gmep = kmem_alloc(sizeof (guid_map_entry_t), KM_SLEEP);
1489                 gmep->guid = snapds->ds_phys->ds_guid;
1490                 gmep->gme_ds = snapds;
1491                 avl_add(guid_map, gmep);
1492         }
1493
1494         rw_exit(&dp->dp_config_rwlock);
1495         return (err);
1496 }
1497
1498 static int
1499 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
1500 {
1501         struct recvendsyncarg resa;
1502         dsl_dataset_t *ds = drc->drc_logical_ds;
1503         int err;
1504
1505         /*
1506          * XXX hack; seems the ds is still dirty and dsl_pool_zil_clean()
1507          * expects it to have a ds_user_ptr (and zil), but clone_swap()
1508          * can close it.
1509          */
1510         txg_wait_synced(ds->ds_dir->dd_pool, 0);
1511
1512         if (dsl_dataset_tryown(ds, FALSE, dmu_recv_tag)) {
1513                 err = dsl_dataset_clone_swap(drc->drc_real_ds, ds,
1514                     drc->drc_force);
1515                 if (err)
1516                         goto out;
1517         } else {
1518                 mutex_exit(&ds->ds_recvlock);
1519                 dsl_dataset_rele(ds, dmu_recv_tag);
1520                 (void) dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag,
1521                     B_FALSE);
1522                 return (EBUSY);
1523         }
1524
1525         resa.creation_time = drc->drc_drrb->drr_creation_time;
1526         resa.toguid = drc->drc_drrb->drr_toguid;
1527         resa.tosnap = drc->drc_tosnap;
1528
1529         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1530             recv_end_check, recv_end_sync, ds, &resa, 3);
1531         if (err) {
1532                 /* swap back */
1533                 (void) dsl_dataset_clone_swap(drc->drc_real_ds, ds, B_TRUE);
1534         }
1535
1536 out:
1537         mutex_exit(&ds->ds_recvlock);
1538         if (err == 0 && drc->drc_guid_to_ds_map != NULL)
1539                 (void) add_ds_to_guidmap(drc->drc_guid_to_ds_map, ds);
1540         dsl_dataset_disown(ds, dmu_recv_tag);
1541         (void) dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag, B_FALSE);
1542         return (err);
1543 }
1544
1545 static int
1546 dmu_recv_new_end(dmu_recv_cookie_t *drc)
1547 {
1548         struct recvendsyncarg resa;
1549         dsl_dataset_t *ds = drc->drc_logical_ds;
1550         int err;
1551
1552         /*
1553          * XXX hack; seems the ds is still dirty and dsl_pool_zil_clean()
1554          * expects it to have a ds_user_ptr (and zil), but clone_swap()
1555          * can close it.
1556          */
1557         txg_wait_synced(ds->ds_dir->dd_pool, 0);
1558
1559         resa.creation_time = drc->drc_drrb->drr_creation_time;
1560         resa.toguid = drc->drc_drrb->drr_toguid;
1561         resa.tosnap = drc->drc_tosnap;
1562
1563         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1564             recv_end_check, recv_end_sync, ds, &resa, 3);
1565         if (err) {
1566                 /* clean up the fs we just recv'd into */
1567                 (void) dsl_dataset_destroy(ds, dmu_recv_tag, B_FALSE);
1568         } else {
1569                 if (drc->drc_guid_to_ds_map != NULL)
1570                         (void) add_ds_to_guidmap(drc->drc_guid_to_ds_map, ds);
1571                 /* release the hold from dmu_recv_begin */
1572                 dsl_dataset_disown(ds, dmu_recv_tag);
1573         }
1574         return (err);
1575 }
1576
1577 int
1578 dmu_recv_end(dmu_recv_cookie_t *drc)
1579 {
1580         if (drc->drc_logical_ds != drc->drc_real_ds)
1581                 return (dmu_recv_existing_end(drc));
1582         else
1583                 return (dmu_recv_new_end(drc));
1584 }