sys/cddl/contrib/opensolaris/uts/common/fs/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  * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
  24  * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
  25  * Copyright (c) 2014, Joyent, Inc. All rights reserved.
  26  * Copyright (c) 2012, Martin Matuska <mm@FreeBSD.org>. All rights reserved.
  27  * Copyright 2014 HybridCluster. All rights reserved.
  28  */
  29
  30 #include <sys/dmu.h>
  31 #include <sys/dmu_impl.h>
  32 #include <sys/dmu_tx.h>
  33 #include <sys/dbuf.h>
  34 #include <sys/dnode.h>
  35 #include <sys/zfs_context.h>
  36 #include <sys/dmu_objset.h>
  37 #include <sys/dmu_traverse.h>
  38 #include <sys/dsl_dataset.h>
  39 #include <sys/dsl_dir.h>
  40 #include <sys/dsl_prop.h>
  41 #include <sys/dsl_pool.h>
  42 #include <sys/dsl_synctask.h>
  43 #include <sys/zfs_ioctl.h>
  44 #include <sys/zap.h>
  45 #include <sys/zio_checksum.h>
  46 #include <sys/zfs_znode.h>
  47 #include <zfs_fletcher.h>
  48 #include <sys/avl.h>
  49 #include <sys/ddt.h>
  50 #include <sys/zfs_onexit.h>
  51 #include <sys/dmu_send.h>
  52 #include <sys/dsl_destroy.h>
  53 #include <sys/blkptr.h>
  54 #include <sys/dsl_bookmark.h>
  55 #include <sys/zfeature.h>
  56
  57 #ifdef __FreeBSD__
  58 #undef dump_write
  59 #define dump_write dmu_dump_write
  60 #endif
  61
  62 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
  63 int zfs_send_corrupt_data = B_FALSE;
  64
  65 static char *dmu_recv_tag = "dmu_recv_tag";
  66 static const char *recv_clone_name = "%recv";
  67
  68 static int
  69 dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
  70 {
  71         dsl_dataset_t *ds = dsp->dsa_os->os_dsl_dataset;
  72         struct uio auio;
  73         struct iovec aiov;
  74         ASSERT0(len % 8);
  75
  76         fletcher_4_incremental_native(buf, len, &dsp->dsa_zc);
  77         aiov.iov_base = buf;
  78         aiov.iov_len = len;
  79         auio.uio_iov = &aiov;
  80         auio.uio_iovcnt = 1;
  81         auio.uio_resid = len;
  82         auio.uio_segflg = UIO_SYSSPACE;
  83         auio.uio_rw = UIO_WRITE;
  84         auio.uio_offset = (off_t)-1;
  85         auio.uio_td = dsp->dsa_td;
  86 #ifdef _KERNEL
  87         if (dsp->dsa_fp->f_type == DTYPE_VNODE)
  88                 bwillwrite();
  89         dsp->dsa_err = fo_write(dsp->dsa_fp, &auio, dsp->dsa_td->td_ucred, 0,
  90             dsp->dsa_td);
  91 #else
  92         fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__);
  93         dsp->dsa_err = EOPNOTSUPP;
  94 #endif
  95         mutex_enter(&ds->ds_sendstream_lock);
  96         *dsp->dsa_off += len;
  97         mutex_exit(&ds->ds_sendstream_lock);
  98
  99         return (dsp->dsa_err);
 100 }
 101
 102 static int
 103 dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
 104     uint64_t length)
 105 {
 106         struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free);
 107
 108         /*
 109          * When we receive a free record, dbuf_free_range() assumes
 110          * that the receiving system doesn't have any dbufs in the range
 111          * being freed.  This is always true because there is a one-record
 112          * constraint: we only send one WRITE record for any given
 113          * object+offset.  We know that the one-record constraint is
 114          * true because we always send data in increasing order by
 115          * object,offset.
 116          *
 117          * If the increasing-order constraint ever changes, we should find
 118          * another way to assert that the one-record constraint is still
 119          * satisfied.
 120          */
 121         ASSERT(object > dsp->dsa_last_data_object ||
 122             (object == dsp->dsa_last_data_object &&
 123             offset > dsp->dsa_last_data_offset));
 124
 125         /*
 126          * If we are doing a non-incremental send, then there can't
 127          * be any data in the dataset we're receiving into.  Therefore
 128          * a free record would simply be a no-op.  Save space by not
 129          * sending it to begin with.
 130          */
 131         if (!dsp->dsa_incremental)
 132                 return (0);
 133
 134         if (length != -1ULL && offset + length < offset)
 135                 length = -1ULL;
 136
 137         /*
 138          * If there is a pending op, but it's not PENDING_FREE, push it out,
 139          * since free block aggregation can only be done for blocks of the
 140          * same type (i.e., DRR_FREE records can only be aggregated with
 141          * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
 142          * aggregated with other DRR_FREEOBJECTS records.
 143          */
 144         if (dsp->dsa_pending_op != PENDING_NONE &&
 145             dsp->dsa_pending_op != PENDING_FREE) {
 146                 if (dump_bytes(dsp, dsp->dsa_drr,
 147                     sizeof (dmu_replay_record_t)) != 0)
 148                         return (SET_ERROR(EINTR));
 149                 dsp->dsa_pending_op = PENDING_NONE;
 150         }
 151
 152         if (dsp->dsa_pending_op == PENDING_FREE) {
 153                 /*
 154                  * There should never be a PENDING_FREE if length is -1
 155                  * (because dump_dnode is the only place where this
 156                  * function is called with a -1, and only after flushing
 157                  * any pending record).
 158                  */
 159                 ASSERT(length != -1ULL);
 160                 /*
 161                  * Check to see whether this free block can be aggregated
 162                  * with pending one.
 163                  */
 164                 if (drrf->drr_object == object && drrf->drr_offset +
 165                     drrf->drr_length == offset) {
 166                         drrf->drr_length += length;
 167                         return (0);
 168                 } else {
 169                         /* not a continuation.  Push out pending record */
 170                         if (dump_bytes(dsp, dsp->dsa_drr,
 171                             sizeof (dmu_replay_record_t)) != 0)
 172                                 return (SET_ERROR(EINTR));
 173                         dsp->dsa_pending_op = PENDING_NONE;
 174                 }
 175         }
 176         /* create a FREE record and make it pending */
 177         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 178         dsp->dsa_drr->drr_type = DRR_FREE;
 179         drrf->drr_object = object;
 180         drrf->drr_offset = offset;
 181         drrf->drr_length = length;
 182         drrf->drr_toguid = dsp->dsa_toguid;
 183         if (length == -1ULL) {
 184                 if (dump_bytes(dsp, dsp->dsa_drr,
 185                     sizeof (dmu_replay_record_t)) != 0)
 186                         return (SET_ERROR(EINTR));
 187         } else {
 188                 dsp->dsa_pending_op = PENDING_FREE;
 189         }
 190
 191         return (0);
 192 }
 193
 194 static int
 195 dump_write(dmu_sendarg_t *dsp, dmu_object_type_t type,
 196     uint64_t object, uint64_t offset, int blksz, const blkptr_t *bp, void *data)
 197 {
 198         struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write);
 199
 200         /*
 201          * We send data in increasing object, offset order.
 202          * See comment in dump_free() for details.
 203          */
 204         ASSERT(object > dsp->dsa_last_data_object ||
 205             (object == dsp->dsa_last_data_object &&
 206             offset > dsp->dsa_last_data_offset));
 207         dsp->dsa_last_data_object = object;
 208         dsp->dsa_last_data_offset = offset + blksz - 1;
 209
 210         /*
 211          * If there is any kind of pending aggregation (currently either
 212          * a grouping of free objects or free blocks), push it out to
 213          * the stream, since aggregation can't be done across operations
 214          * of different types.
 215          */
 216         if (dsp->dsa_pending_op != PENDING_NONE) {
 217                 if (dump_bytes(dsp, dsp->dsa_drr,
 218                     sizeof (dmu_replay_record_t)) != 0)
 219                         return (SET_ERROR(EINTR));
 220                 dsp->dsa_pending_op = PENDING_NONE;
 221         }
 222         /* write a DATA record */
 223         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 224         dsp->dsa_drr->drr_type = DRR_WRITE;
 225         drrw->drr_object = object;
 226         drrw->drr_type = type;
 227         drrw->drr_offset = offset;
 228         drrw->drr_length = blksz;
 229         drrw->drr_toguid = dsp->dsa_toguid;
 230         if (bp == NULL || BP_IS_EMBEDDED(bp)) {
 231                 /*
 232                  * There's no pre-computed checksum for partial-block
 233                  * writes or embedded BP's, so (like
 234                  * fletcher4-checkummed blocks) userland will have to
 235                  * compute a dedup-capable checksum itself.
 236                  */
 237                 drrw->drr_checksumtype = ZIO_CHECKSUM_OFF;
 238         } else {
 239                 drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
 240                 if (zio_checksum_table[drrw->drr_checksumtype].ci_dedup)
 241                         drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
 242                 DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
 243                 DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
 244                 DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
 245                 drrw->drr_key.ddk_cksum = bp->blk_cksum;
 246         }
 247
 248         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
 249                 return (SET_ERROR(EINTR));
 250         if (dump_bytes(dsp, data, blksz) != 0)
 251                 return (SET_ERROR(EINTR));
 252         return (0);
 253 }
 254
 255 static int
 256 dump_write_embedded(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
 257     int blksz, const blkptr_t *bp)
 258 {
 259         char buf[BPE_PAYLOAD_SIZE];
 260         struct drr_write_embedded *drrw =
 261             &(dsp->dsa_drr->drr_u.drr_write_embedded);
 262
 263         if (dsp->dsa_pending_op != PENDING_NONE) {
 264                 if (dump_bytes(dsp, dsp->dsa_drr,
 265                     sizeof (dmu_replay_record_t)) != 0)
 266                         return (EINTR);
 267                 dsp->dsa_pending_op = PENDING_NONE;
 268         }
 269
 270         ASSERT(BP_IS_EMBEDDED(bp));
 271
 272         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 273         dsp->dsa_drr->drr_type = DRR_WRITE_EMBEDDED;
 274         drrw->drr_object = object;
 275         drrw->drr_offset = offset;
 276         drrw->drr_length = blksz;
 277         drrw->drr_toguid = dsp->dsa_toguid;
 278         drrw->drr_compression = BP_GET_COMPRESS(bp);
 279         drrw->drr_etype = BPE_GET_ETYPE(bp);
 280         drrw->drr_lsize = BPE_GET_LSIZE(bp);
 281         drrw->drr_psize = BPE_GET_PSIZE(bp);
 282
 283         decode_embedded_bp_compressed(bp, buf);
 284
 285         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
 286                 return (EINTR);
 287         if (dump_bytes(dsp, buf, P2ROUNDUP(drrw->drr_psize, 8)) != 0)
 288                 return (EINTR);
 289         return (0);
 290 }
 291
 292 static int
 293 dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data)
 294 {
 295         struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill);
 296
 297         if (dsp->dsa_pending_op != PENDING_NONE) {
 298                 if (dump_bytes(dsp, dsp->dsa_drr,
 299                     sizeof (dmu_replay_record_t)) != 0)
 300                         return (SET_ERROR(EINTR));
 301                 dsp->dsa_pending_op = PENDING_NONE;
 302         }
 303
 304         /* write a SPILL record */
 305         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 306         dsp->dsa_drr->drr_type = DRR_SPILL;
 307         drrs->drr_object = object;
 308         drrs->drr_length = blksz;
 309         drrs->drr_toguid = dsp->dsa_toguid;
 310
 311         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)))
 312                 return (SET_ERROR(EINTR));
 313         if (dump_bytes(dsp, data, blksz))
 314                 return (SET_ERROR(EINTR));
 315         return (0);
 316 }
 317
 318 static int
 319 dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
 320 {
 321         struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
 322
 323         /* See comment in dump_free(). */
 324         if (!dsp->dsa_incremental)
 325                 return (0);
 326
 327         /*
 328          * If there is a pending op, but it's not PENDING_FREEOBJECTS,
 329          * push it out, since free block aggregation can only be done for
 330          * blocks of the same type (i.e., DRR_FREE records can only be
 331          * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
 332          * can only be aggregated with other DRR_FREEOBJECTS records.
 333          */
 334         if (dsp->dsa_pending_op != PENDING_NONE &&
 335             dsp->dsa_pending_op != PENDING_FREEOBJECTS) {
 336                 if (dump_bytes(dsp, dsp->dsa_drr,
 337                     sizeof (dmu_replay_record_t)) != 0)
 338                         return (SET_ERROR(EINTR));
 339                 dsp->dsa_pending_op = PENDING_NONE;
 340         }
 341         if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) {
 342                 /*
 343                  * See whether this free object array can be aggregated
 344                  * with pending one
 345                  */
 346                 if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
 347                         drrfo->drr_numobjs += numobjs;
 348                         return (0);
 349                 } else {
 350                         /* can't be aggregated.  Push out pending record */
 351                         if (dump_bytes(dsp, dsp->dsa_drr,
 352                             sizeof (dmu_replay_record_t)) != 0)
 353                                 return (SET_ERROR(EINTR));
 354                         dsp->dsa_pending_op = PENDING_NONE;
 355                 }
 356         }
 357
 358         /* write a FREEOBJECTS record */
 359         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 360         dsp->dsa_drr->drr_type = DRR_FREEOBJECTS;
 361         drrfo->drr_firstobj = firstobj;
 362         drrfo->drr_numobjs = numobjs;
 363         drrfo->drr_toguid = dsp->dsa_toguid;
 364
 365         dsp->dsa_pending_op = PENDING_FREEOBJECTS;
 366
 367         return (0);
 368 }
 369
 370 static int
 371 dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
 372 {
 373         struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object);
 374
 375         if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
 376                 return (dump_freeobjects(dsp, object, 1));
 377
 378         if (dsp->dsa_pending_op != PENDING_NONE) {
 379                 if (dump_bytes(dsp, dsp->dsa_drr,
 380                     sizeof (dmu_replay_record_t)) != 0)
 381                         return (SET_ERROR(EINTR));
 382                 dsp->dsa_pending_op = PENDING_NONE;
 383         }
 384
 385         /* write an OBJECT record */
 386         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 387         dsp->dsa_drr->drr_type = DRR_OBJECT;
 388         drro->drr_object = object;
 389         drro->drr_type = dnp->dn_type;
 390         drro->drr_bonustype = dnp->dn_bonustype;
 391         drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 392         drro->drr_bonuslen = dnp->dn_bonuslen;
 393         drro->drr_checksumtype = dnp->dn_checksum;
 394         drro->drr_compress = dnp->dn_compress;
 395         drro->drr_toguid = dsp->dsa_toguid;
 396
 397         if (!(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
 398             drro->drr_blksz > SPA_OLD_MAXBLOCKSIZE)
 399                 drro->drr_blksz = SPA_OLD_MAXBLOCKSIZE;
 400
 401         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
 402                 return (SET_ERROR(EINTR));
 403
 404         if (dump_bytes(dsp, DN_BONUS(dnp), P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0)
 405                 return (SET_ERROR(EINTR));
 406
 407         /* Free anything past the end of the file. */
 408         if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) *
 409             (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0)
 410                 return (SET_ERROR(EINTR));
 411         if (dsp->dsa_err != 0)
 412                 return (SET_ERROR(EINTR));
 413         return (0);
 414 }
 415
 416 static boolean_t
 417 backup_do_embed(dmu_sendarg_t *dsp, const blkptr_t *bp)
 418 {
 419         if (!BP_IS_EMBEDDED(bp))
 420                 return (B_FALSE);
 421
 422         /*
 423          * Compression function must be legacy, or explicitly enabled.
 424          */
 425         if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS &&
 426             !(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4)))
 427                 return (B_FALSE);
 428
 429         /*
 430          * Embed type must be explicitly enabled.
 431          */
 432         switch (BPE_GET_ETYPE(bp)) {
 433         case BP_EMBEDDED_TYPE_DATA:
 434                 if (dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)
 435                         return (B_TRUE);
 436                 break;
 437         default:
 438                 return (B_FALSE);
 439         }
 440         return (B_FALSE);
 441 }
 442
 443 #define BP_SPAN(dnp, level) \
 444         (((uint64_t)dnp->dn_datablkszsec) << (SPA_MINBLOCKSHIFT + \
 445         (level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)))
 446
 447 /* ARGSUSED */
 448 static int
 449 backup_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 450     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
 451 {
 452         dmu_sendarg_t *dsp = arg;
 453         dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
 454         int err = 0;
 455
 456         if (issig(JUSTLOOKING) && issig(FORREAL))
 457                 return (SET_ERROR(EINTR));
 458
 459         if (zb->zb_object != DMU_META_DNODE_OBJECT &&
 460             DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
 461                 return (0);
 462         } else if (zb->zb_level == ZB_ZIL_LEVEL) {
 463                 /*
 464                  * If we are sending a non-snapshot (which is allowed on
 465                  * read-only pools), it may have a ZIL, which must be ignored.
 466                  */
 467                 return (0);
 468         } else if (BP_IS_HOLE(bp) &&
 469             zb->zb_object == DMU_META_DNODE_OBJECT) {
 470                 uint64_t span = BP_SPAN(dnp, zb->zb_level);
 471                 uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
 472                 err = dump_freeobjects(dsp, dnobj, span >> DNODE_SHIFT);
 473         } else if (BP_IS_HOLE(bp)) {
 474                 uint64_t span = BP_SPAN(dnp, zb->zb_level);
 475                 err = dump_free(dsp, zb->zb_object, zb->zb_blkid * span, span);
 476         } else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
 477                 return (0);
 478         } else if (type == DMU_OT_DNODE) {
 479                 dnode_phys_t *blk;
 480                 int i;
 481                 int blksz = BP_GET_LSIZE(bp);
 482                 arc_flags_t aflags = ARC_FLAG_WAIT;
 483                 arc_buf_t *abuf;
 484
 485                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 486                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 487                     &aflags, zb) != 0)
 488                         return (SET_ERROR(EIO));
 489
 490                 blk = abuf->b_data;
 491                 for (i = 0; i < blksz >> DNODE_SHIFT; i++) {
 492                         uint64_t dnobj = (zb->zb_blkid <<
 493                             (DNODE_BLOCK_SHIFT - DNODE_SHIFT)) + i;
 494                         err = dump_dnode(dsp, dnobj, blk+i);
 495                         if (err != 0)
 496                                 break;
 497                 }
 498                 (void) arc_buf_remove_ref(abuf, &abuf);
 499         } else if (type == DMU_OT_SA) {
 500                 arc_flags_t aflags = ARC_FLAG_WAIT;
 501                 arc_buf_t *abuf;
 502                 int blksz = BP_GET_LSIZE(bp);
 503
 504                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 505                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 506                     &aflags, zb) != 0)
 507                         return (SET_ERROR(EIO));
 508
 509                 err = dump_spill(dsp, zb->zb_object, blksz, abuf->b_data);
 510                 (void) arc_buf_remove_ref(abuf, &abuf);
 511         } else if (backup_do_embed(dsp, bp)) {
 512                 /* it's an embedded level-0 block of a regular object */
 513                 int blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 514                 err = dump_write_embedded(dsp, zb->zb_object,
 515                     zb->zb_blkid * blksz, blksz, bp);
 516         } else { /* it's a level-0 block of a regular object */
 517                 arc_flags_t aflags = ARC_FLAG_WAIT;
 518                 arc_buf_t *abuf;
 519                 int blksz = BP_GET_LSIZE(bp);
 520                 uint64_t offset;
 521
 522                 ASSERT3U(blksz, ==, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 523                 ASSERT0(zb->zb_level);
 524                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 525                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 526                     &aflags, zb) != 0) {
 527                         if (zfs_send_corrupt_data) {
 528                                 /* Send a block filled with 0x"zfs badd bloc" */
 529                                 abuf = arc_buf_alloc(spa, blksz, &abuf,
 530                                     ARC_BUFC_DATA);
 531                                 uint64_t *ptr;
 532                                 for (ptr = abuf->b_data;
 533                                     (char *)ptr < (char *)abuf->b_data + blksz;
 534                                     ptr++)
 535                                         *ptr = 0x2f5baddb10c;
 536                         } else {
 537                                 return (SET_ERROR(EIO));
 538                         }
 539                 }
 540
 541                 offset = zb->zb_blkid * blksz;
 542
 543                 if (!(dsp->dsa_featureflags &
 544                     DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
 545                     blksz > SPA_OLD_MAXBLOCKSIZE) {
 546                         char *buf = abuf->b_data;
 547                         while (blksz > 0 && err == 0) {
 548                                 int n = MIN(blksz, SPA_OLD_MAXBLOCKSIZE);
 549                                 err = dump_write(dsp, type, zb->zb_object,
 550                                     offset, n, NULL, buf);
 551                                 offset += n;
 552                                 buf += n;
 553                                 blksz -= n;
 554                         }
 555                 } else {
 556                         err = dump_write(dsp, type, zb->zb_object,
 557                             offset, blksz, bp, abuf->b_data);
 558                 }
 559                 (void) arc_buf_remove_ref(abuf, &abuf);
 560         }
 561
 562         ASSERT(err == 0 || err == EINTR);
 563         return (err);
 564 }
 565
 566 /*
 567  * Releases dp using the specified tag.
 568  */
 569 static int
 570 dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *ds,
 571     zfs_bookmark_phys_t *fromzb, boolean_t is_clone, boolean_t embedok,
 572 #ifdef illumos
 573     boolean_t large_block_ok, int outfd, vnode_t *vp, offset_t *off)
 574 #else
 575     boolean_t large_block_ok, int outfd, struct file *fp, offset_t *off)
 576 #endif
 577 {
 578         objset_t *os;
 579         dmu_replay_record_t *drr;
 580         dmu_sendarg_t *dsp;
 581         int err;
 582         uint64_t fromtxg = 0;
 583         uint64_t featureflags = 0;
 584
 585         err = dmu_objset_from_ds(ds, &os);
 586         if (err != 0) {
 587                 dsl_pool_rele(dp, tag);
 588                 return (err);
 589         }
 590
 591         drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
 592         drr->drr_type = DRR_BEGIN;
 593         drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
 594         DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
 595             DMU_SUBSTREAM);
 596
 597 #ifdef _KERNEL
 598         if (dmu_objset_type(os) == DMU_OST_ZFS) {
 599                 uint64_t version;
 600                 if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0) {
 601                         kmem_free(drr, sizeof (dmu_replay_record_t));
 602                         dsl_pool_rele(dp, tag);
 603                         return (SET_ERROR(EINVAL));
 604                 }
 605                 if (version >= ZPL_VERSION_SA) {
 606                         featureflags |= DMU_BACKUP_FEATURE_SA_SPILL;
 607                 }
 608         }
 609 #endif
 610
 611         if (large_block_ok && ds->ds_large_blocks)
 612                 featureflags |= DMU_BACKUP_FEATURE_LARGE_BLOCKS;
 613         if (embedok &&
 614             spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) {
 615                 featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA;
 616                 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
 617                         featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA_LZ4;
 618         } else {
 619                 embedok = B_FALSE;
 620         }
 621
 622         DMU_SET_FEATUREFLAGS(drr->drr_u.drr_begin.drr_versioninfo,
 623             featureflags);
 624
 625         drr->drr_u.drr_begin.drr_creation_time =
 626             dsl_dataset_phys(ds)->ds_creation_time;
 627         drr->drr_u.drr_begin.drr_type = dmu_objset_type(os);
 628         if (is_clone)
 629                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
 630         drr->drr_u.drr_begin.drr_toguid = dsl_dataset_phys(ds)->ds_guid;
 631         if (dsl_dataset_phys(ds)->ds_flags & DS_FLAG_CI_DATASET)
 632                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
 633
 634         if (fromzb != NULL) {
 635                 drr->drr_u.drr_begin.drr_fromguid = fromzb->zbm_guid;
 636                 fromtxg = fromzb->zbm_creation_txg;
 637         }
 638         dsl_dataset_name(ds, drr->drr_u.drr_begin.drr_toname);
 639         if (!dsl_dataset_is_snapshot(ds)) {
 640                 (void) strlcat(drr->drr_u.drr_begin.drr_toname, "@--head--",
 641                     sizeof (drr->drr_u.drr_begin.drr_toname));
 642         }
 643
 644         dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP);
 645
 646         dsp->dsa_drr = drr;
 647         dsp->dsa_outfd = outfd;
 648         dsp->dsa_proc = curproc;
 649         dsp->dsa_td = curthread;
 650         dsp->dsa_fp = fp;
 651         dsp->dsa_os = os;
 652         dsp->dsa_off = off;
 653         dsp->dsa_toguid = dsl_dataset_phys(ds)->ds_guid;
 654         ZIO_SET_CHECKSUM(&dsp->dsa_zc, 0, 0, 0, 0);
 655         dsp->dsa_pending_op = PENDING_NONE;
 656         dsp->dsa_incremental = (fromzb != NULL);
 657         dsp->dsa_featureflags = featureflags;
 658
 659         mutex_enter(&ds->ds_sendstream_lock);
 660         list_insert_head(&ds->ds_sendstreams, dsp);
 661         mutex_exit(&ds->ds_sendstream_lock);
 662
 663         dsl_dataset_long_hold(ds, FTAG);
 664         dsl_pool_rele(dp, tag);
 665
 666         if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0) {
 667                 err = dsp->dsa_err;
 668                 goto out;
 669         }
 670
 671         err = traverse_dataset(ds, fromtxg, TRAVERSE_PRE | TRAVERSE_PREFETCH,
 672             backup_cb, dsp);
 673
 674         if (dsp->dsa_pending_op != PENDING_NONE)
 675                 if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0)
 676                         err = SET_ERROR(EINTR);
 677
 678         if (err != 0) {
 679                 if (err == EINTR && dsp->dsa_err != 0)
 680                         err = dsp->dsa_err;
 681                 goto out;
 682         }
 683
 684         bzero(drr, sizeof (dmu_replay_record_t));
 685         drr->drr_type = DRR_END;
 686         drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc;
 687         drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid;
 688
 689         if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0) {
 690                 err = dsp->dsa_err;
 691                 goto out;
 692         }
 693
 694 out:
 695         mutex_enter(&ds->ds_sendstream_lock);
 696         list_remove(&ds->ds_sendstreams, dsp);
 697         mutex_exit(&ds->ds_sendstream_lock);
 698
 699         kmem_free(drr, sizeof (dmu_replay_record_t));
 700         kmem_free(dsp, sizeof (dmu_sendarg_t));
 701
 702         dsl_dataset_long_rele(ds, FTAG);
 703
 704         return (err);
 705 }
 706
 707 int
 708 dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
 709     boolean_t embedok, boolean_t large_block_ok,
 710 #ifdef illumos
 711     int outfd, vnode_t *vp, offset_t *off)
 712 #else
 713     int outfd, struct file *fp, offset_t *off)
 714 #endif
 715 {
 716         dsl_pool_t *dp;
 717         dsl_dataset_t *ds;
 718         dsl_dataset_t *fromds = NULL;
 719         int err;
 720
 721         err = dsl_pool_hold(pool, FTAG, &dp);
 722         if (err != 0)
 723                 return (err);
 724
 725         err = dsl_dataset_hold_obj(dp, tosnap, FTAG, &ds);
 726         if (err != 0) {
 727                 dsl_pool_rele(dp, FTAG);
 728                 return (err);
 729         }
 730
 731         if (fromsnap != 0) {
 732                 zfs_bookmark_phys_t zb;
 733                 boolean_t is_clone;
 734
 735                 err = dsl_dataset_hold_obj(dp, fromsnap, FTAG, &fromds);
 736                 if (err != 0) {
 737                         dsl_dataset_rele(ds, FTAG);
 738                         dsl_pool_rele(dp, FTAG);
 739                         return (err);
 740                 }
 741                 if (!dsl_dataset_is_before(ds, fromds, 0))
 742                         err = SET_ERROR(EXDEV);
 743                 zb.zbm_creation_time =
 744                     dsl_dataset_phys(fromds)->ds_creation_time;
 745                 zb.zbm_creation_txg = dsl_dataset_phys(fromds)->ds_creation_txg;
 746                 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
 747                 is_clone = (fromds->ds_dir != ds->ds_dir);
 748                 dsl_dataset_rele(fromds, FTAG);
 749                 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
 750                     embedok, large_block_ok, outfd, fp, off);
 751         } else {
 752                 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
 753                     embedok, large_block_ok, outfd, fp, off);
 754         }
 755         dsl_dataset_rele(ds, FTAG);
 756         return (err);
 757 }
 758
 759 int
 760 dmu_send(const char *tosnap, const char *fromsnap,
 761     boolean_t embedok, boolean_t large_block_ok,
 762 #ifdef illumos
 763     int outfd, vnode_t *vp, offset_t *off)
 764 #else
 765     int outfd, struct file *fp, offset_t *off)
 766 #endif
 767 {
 768         dsl_pool_t *dp;
 769         dsl_dataset_t *ds;
 770         int err;
 771         boolean_t owned = B_FALSE;
 772
 773         if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL)
 774                 return (SET_ERROR(EINVAL));
 775
 776         err = dsl_pool_hold(tosnap, FTAG, &dp);
 777         if (err != 0)
 778                 return (err);
 779
 780         if (strchr(tosnap, '@') == NULL && spa_writeable(dp->dp_spa)) {
 781                 /*
 782                  * We are sending a filesystem or volume.  Ensure
 783                  * that it doesn't change by owning the dataset.
 784                  */
 785                 err = dsl_dataset_own(dp, tosnap, FTAG, &ds);
 786                 owned = B_TRUE;
 787         } else {
 788                 err = dsl_dataset_hold(dp, tosnap, FTAG, &ds);
 789         }
 790         if (err != 0) {
 791                 dsl_pool_rele(dp, FTAG);
 792                 return (err);
 793         }
 794
 795         if (fromsnap != NULL) {
 796                 zfs_bookmark_phys_t zb;
 797                 boolean_t is_clone = B_FALSE;
 798                 int fsnamelen = strchr(tosnap, '@') - tosnap;
 799
 800                 /*
 801                  * If the fromsnap is in a different filesystem, then
 802                  * mark the send stream as a clone.
 803                  */
 804                 if (strncmp(tosnap, fromsnap, fsnamelen) != 0 ||
 805                     (fromsnap[fsnamelen] != '@' &&
 806                     fromsnap[fsnamelen] != '#')) {
 807                         is_clone = B_TRUE;
 808                 }
 809
 810                 if (strchr(fromsnap, '@')) {
 811                         dsl_dataset_t *fromds;
 812                         err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds);
 813                         if (err == 0) {
 814                                 if (!dsl_dataset_is_before(ds, fromds, 0))
 815                                         err = SET_ERROR(EXDEV);
 816                                 zb.zbm_creation_time =
 817                                     dsl_dataset_phys(fromds)->ds_creation_time;
 818                                 zb.zbm_creation_txg =
 819                                     dsl_dataset_phys(fromds)->ds_creation_txg;
 820                                 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
 821                                 is_clone = (ds->ds_dir != fromds->ds_dir);
 822                                 dsl_dataset_rele(fromds, FTAG);
 823                         }
 824                 } else {
 825                         err = dsl_bookmark_lookup(dp, fromsnap, ds, &zb);
 826                 }
 827                 if (err != 0) {
 828                         dsl_dataset_rele(ds, FTAG);
 829                         dsl_pool_rele(dp, FTAG);
 830                         return (err);
 831                 }
 832                 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
 833                     embedok, large_block_ok, outfd, fp, off);
 834         } else {
 835                 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
 836                     embedok, large_block_ok, outfd, fp, off);
 837         }
 838         if (owned)
 839                 dsl_dataset_disown(ds, FTAG);
 840         else
 841                 dsl_dataset_rele(ds, FTAG);
 842         return (err);
 843 }
 844
 845 int
 846 dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds, uint64_t *sizep)
 847 {
 848         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 849         int err;
 850         uint64_t size;
 851
 852         ASSERT(dsl_pool_config_held(dp));
 853
 854         /* tosnap must be a snapshot */
 855         if (!dsl_dataset_is_snapshot(ds))
 856                 return (SET_ERROR(EINVAL));
 857
 858         /*
 859          * fromsnap must be an earlier snapshot from the same fs as tosnap,
 860          * or the origin's fs.
 861          */
 862         if (fromds != NULL && !dsl_dataset_is_before(ds, fromds, 0))
 863                 return (SET_ERROR(EXDEV));
 864
 865         /* Get uncompressed size estimate of changed data. */
 866         if (fromds == NULL) {
 867                 size = dsl_dataset_phys(ds)->ds_uncompressed_bytes;
 868         } else {
 869                 uint64_t used, comp;
 870                 err = dsl_dataset_space_written(fromds, ds,
 871                     &used, &comp, &size);
 872                 if (err != 0)
 873                         return (err);
 874         }
 875
 876         /*
 877          * Assume that space (both on-disk and in-stream) is dominated by
 878          * data.  We will adjust for indirect blocks and the copies property,
 879          * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
 880          */
 881
 882         /*
 883          * Subtract out approximate space used by indirect blocks.
 884          * Assume most space is used by data blocks (non-indirect, non-dnode).
 885          * Assume all blocks are recordsize.  Assume ditto blocks and
 886          * internal fragmentation counter out compression.
 887          *
 888          * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
 889          * block, which we observe in practice.
 890          */
 891         uint64_t recordsize;
 892         err = dsl_prop_get_int_ds(ds, "recordsize", &recordsize);
 893         if (err != 0)
 894                 return (err);
 895         size -= size / recordsize * sizeof (blkptr_t);
 896
 897         /* Add in the space for the record associated with each block. */
 898         size += size / recordsize * sizeof (dmu_replay_record_t);
 899
 900         *sizep = size;
 901
 902         return (0);
 903 }
 904
 905 typedef struct dmu_recv_begin_arg {
 906         const char *drba_origin;
 907         dmu_recv_cookie_t *drba_cookie;
 908         cred_t *drba_cred;
 909         uint64_t drba_snapobj;
 910 } dmu_recv_begin_arg_t;
 911
 912 static int
 913 recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds,
 914     uint64_t fromguid)
 915 {
 916         uint64_t val;
 917         int error;
 918         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 919
 920         /* temporary clone name must not exist */
 921         error = zap_lookup(dp->dp_meta_objset,
 922             dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, recv_clone_name,
 923             8, 1, &val);
 924         if (error != ENOENT)
 925                 return (error == 0 ? EBUSY : error);
 926
 927         /* new snapshot name must not exist */
 928         error = zap_lookup(dp->dp_meta_objset,
 929             dsl_dataset_phys(ds)->ds_snapnames_zapobj,
 930             drba->drba_cookie->drc_tosnap, 8, 1, &val);
 931         if (error != ENOENT)
 932                 return (error == 0 ? EEXIST : error);
 933
 934         /*
 935          * Check snapshot limit before receiving. We'll recheck again at the
 936          * end, but might as well abort before receiving if we're already over
 937          * the limit.
 938          *
 939          * Note that we do not check the file system limit with
 940          * dsl_dir_fscount_check because the temporary %clones don't count
 941          * against that limit.
 942          */
 943         error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT,
 944             NULL, drba->drba_cred);
 945         if (error != 0)
 946                 return (error);
 947
 948         if (fromguid != 0) {
 949                 dsl_dataset_t *snap;
 950                 uint64_t obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
 951
 952                 /* Find snapshot in this dir that matches fromguid. */
 953                 while (obj != 0) {
 954                         error = dsl_dataset_hold_obj(dp, obj, FTAG,
 955                             &snap);
 956                         if (error != 0)
 957                                 return (SET_ERROR(ENODEV));
 958                         if (snap->ds_dir != ds->ds_dir) {
 959                                 dsl_dataset_rele(snap, FTAG);
 960                                 return (SET_ERROR(ENODEV));
 961                         }
 962                         if (dsl_dataset_phys(snap)->ds_guid == fromguid)
 963                                 break;
 964                         obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
 965                         dsl_dataset_rele(snap, FTAG);
 966                 }
 967                 if (obj == 0)
 968                         return (SET_ERROR(ENODEV));
 969
 970                 if (drba->drba_cookie->drc_force) {
 971                         drba->drba_snapobj = obj;
 972                 } else {
 973                         /*
 974                          * If we are not forcing, there must be no
 975                          * changes since fromsnap.
 976                          */
 977                         if (dsl_dataset_modified_since_snap(ds, snap)) {
 978                                 dsl_dataset_rele(snap, FTAG);
 979                                 return (SET_ERROR(ETXTBSY));
 980                         }
 981                         drba->drba_snapobj = ds->ds_prev->ds_object;
 982                 }
 983
 984                 dsl_dataset_rele(snap, FTAG);
 985         } else {
 986                 /* if full, most recent snapshot must be $ORIGIN */
 987                 if (dsl_dataset_phys(ds)->ds_prev_snap_txg >= TXG_INITIAL)
 988                         return (SET_ERROR(ENODEV));
 989                 drba->drba_snapobj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
 990         }
 991
 992         return (0);
 993
 994 }
 995
 996 static int
 997 dmu_recv_begin_check(void *arg, dmu_tx_t *tx)
 998 {
 999         dmu_recv_begin_arg_t *drba = arg;
1000         dsl_pool_t *dp = dmu_tx_pool(tx);
1001         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1002         uint64_t fromguid = drrb->drr_fromguid;
1003         int flags = drrb->drr_flags;
1004         int error;
1005         uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1006         dsl_dataset_t *ds;
1007         const char *tofs = drba->drba_cookie->drc_tofs;
1008
1009         /* already checked */
1010         ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1011
1012         if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1013             DMU_COMPOUNDSTREAM ||
1014             drrb->drr_type >= DMU_OST_NUMTYPES ||
1015             ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL))
1016                 return (SET_ERROR(EINVAL));
1017
1018         /* Verify pool version supports SA if SA_SPILL feature set */
1019         if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1020             spa_version(dp->dp_spa) < SPA_VERSION_SA)
1021                 return (SET_ERROR(ENOTSUP));
1022
1023         /*
1024          * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1025          * record to a plan WRITE record, so the pool must have the
1026          * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1027          * records.  Same with WRITE_EMBEDDED records that use LZ4 compression.
1028          */
1029         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1030             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1031                 return (SET_ERROR(ENOTSUP));
1032         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4) &&
1033             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1034                 return (SET_ERROR(ENOTSUP));
1035
1036         /*
1037          * The receiving code doesn't know how to translate large blocks
1038          * to smaller ones, so the pool must have the LARGE_BLOCKS
1039          * feature enabled if the stream has LARGE_BLOCKS.
1040          */
1041         if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
1042             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS))
1043                 return (SET_ERROR(ENOTSUP));
1044
1045         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1046         if (error == 0) {
1047                 /* target fs already exists; recv into temp clone */
1048
1049                 /* Can't recv a clone into an existing fs */
1050                 if (flags & DRR_FLAG_CLONE) {
1051                         dsl_dataset_rele(ds, FTAG);
1052                         return (SET_ERROR(EINVAL));
1053                 }
1054
1055                 error = recv_begin_check_existing_impl(drba, ds, fromguid);
1056                 dsl_dataset_rele(ds, FTAG);
1057         } else if (error == ENOENT) {
1058                 /* target fs does not exist; must be a full backup or clone */
1059                 char buf[MAXNAMELEN];
1060
1061                 /*
1062                  * If it's a non-clone incremental, we are missing the
1063                  * target fs, so fail the recv.
1064                  */
1065                 if (fromguid != 0 && !(flags & DRR_FLAG_CLONE))
1066                         return (SET_ERROR(ENOENT));
1067
1068                 /* Open the parent of tofs */
1069                 ASSERT3U(strlen(tofs), <, MAXNAMELEN);
1070                 (void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1);
1071                 error = dsl_dataset_hold(dp, buf, FTAG, &ds);
1072                 if (error != 0)
1073                         return (error);
1074
1075                 /*
1076                  * Check filesystem and snapshot limits before receiving. We'll
1077                  * recheck snapshot limits again at the end (we create the
1078                  * filesystems and increment those counts during begin_sync).
1079                  */
1080                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1081                     ZFS_PROP_FILESYSTEM_LIMIT, NULL, drba->drba_cred);
1082                 if (error != 0) {
1083                         dsl_dataset_rele(ds, FTAG);
1084                         return (error);
1085                 }
1086
1087                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1088                     ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred);
1089                 if (error != 0) {
1090                         dsl_dataset_rele(ds, FTAG);
1091                         return (error);
1092                 }
1093
1094                 if (drba->drba_origin != NULL) {
1095                         dsl_dataset_t *origin;
1096                         error = dsl_dataset_hold(dp, drba->drba_origin,
1097                             FTAG, &origin);
1098                         if (error != 0) {
1099                                 dsl_dataset_rele(ds, FTAG);
1100                                 return (error);
1101                         }
1102                         if (!dsl_dataset_is_snapshot(origin)) {
1103                                 dsl_dataset_rele(origin, FTAG);
1104                                 dsl_dataset_rele(ds, FTAG);
1105                                 return (SET_ERROR(EINVAL));
1106                         }
1107                         if (dsl_dataset_phys(origin)->ds_guid != fromguid) {
1108                                 dsl_dataset_rele(origin, FTAG);
1109                                 dsl_dataset_rele(ds, FTAG);
1110                                 return (SET_ERROR(ENODEV));
1111                         }
1112                         dsl_dataset_rele(origin, FTAG);
1113                 }
1114                 dsl_dataset_rele(ds, FTAG);
1115                 error = 0;
1116         }
1117         return (error);
1118 }
1119
1120 static void
1121 dmu_recv_begin_sync(void *arg, dmu_tx_t *tx)
1122 {
1123         dmu_recv_begin_arg_t *drba = arg;
1124         dsl_pool_t *dp = dmu_tx_pool(tx);
1125         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1126         const char *tofs = drba->drba_cookie->drc_tofs;
1127         dsl_dataset_t *ds, *newds;
1128         uint64_t dsobj;
1129         int error;
1130         uint64_t crflags;
1131
1132         crflags = (drrb->drr_flags & DRR_FLAG_CI_DATA) ?
1133             DS_FLAG_CI_DATASET : 0;
1134
1135         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1136         if (error == 0) {
1137                 /* create temporary clone */
1138                 dsl_dataset_t *snap = NULL;
1139                 if (drba->drba_snapobj != 0) {
1140                         VERIFY0(dsl_dataset_hold_obj(dp,
1141                             drba->drba_snapobj, FTAG, &snap));
1142                 }
1143                 dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name,
1144                     snap, crflags, drba->drba_cred, tx);
1145                 if (drba->drba_snapobj != 0)
1146                         dsl_dataset_rele(snap, FTAG);
1147                 dsl_dataset_rele(ds, FTAG);
1148         } else {
1149                 dsl_dir_t *dd;
1150                 const char *tail;
1151                 dsl_dataset_t *origin = NULL;
1152
1153                 VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail));
1154
1155                 if (drba->drba_origin != NULL) {
1156                         VERIFY0(dsl_dataset_hold(dp, drba->drba_origin,
1157                             FTAG, &origin));
1158                 }
1159
1160                 /* Create new dataset. */
1161                 dsobj = dsl_dataset_create_sync(dd,
1162                     strrchr(tofs, '/') + 1,
1163                     origin, crflags, drba->drba_cred, tx);
1164                 if (origin != NULL)
1165                         dsl_dataset_rele(origin, FTAG);
1166                 dsl_dir_rele(dd, FTAG);
1167                 drba->drba_cookie->drc_newfs = B_TRUE;
1168         }
1169         VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds));
1170
1171         if ((DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
1172             DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
1173             !newds->ds_large_blocks) {
1174                 dsl_dataset_activate_large_blocks_sync_impl(dsobj, tx);
1175                 newds->ds_large_blocks = B_TRUE;
1176         }
1177
1178         dmu_buf_will_dirty(newds->ds_dbuf, tx);
1179         dsl_dataset_phys(newds)->ds_flags |= DS_FLAG_INCONSISTENT;
1180
1181         /*
1182          * If we actually created a non-clone, we need to create the
1183          * objset in our new dataset.
1184          */
1185         if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) {
1186                 (void) dmu_objset_create_impl(dp->dp_spa,
1187                     newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx);
1188         }
1189
1190         drba->drba_cookie->drc_ds = newds;
1191
1192         spa_history_log_internal_ds(newds, "receive", tx, "");
1193 }
1194
1195 /*
1196  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
1197  * succeeds; otherwise we will leak the holds on the datasets.
1198  */
1199 int
1200 dmu_recv_begin(char *tofs, char *tosnap, struct drr_begin *drrb,
1201     boolean_t force, char *origin, dmu_recv_cookie_t *drc)
1202 {
1203         dmu_recv_begin_arg_t drba = { 0 };
1204         dmu_replay_record_t *drr;
1205
1206         bzero(drc, sizeof (dmu_recv_cookie_t));
1207         drc->drc_drrb = drrb;
1208         drc->drc_tosnap = tosnap;
1209         drc->drc_tofs = tofs;
1210         drc->drc_force = force;
1211         drc->drc_cred = CRED();
1212
1213         if (drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
1214                 drc->drc_byteswap = B_TRUE;
1215         else if (drrb->drr_magic != DMU_BACKUP_MAGIC)
1216                 return (SET_ERROR(EINVAL));
1217
1218         drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
1219         drr->drr_type = DRR_BEGIN;
1220         drr->drr_u.drr_begin = *drc->drc_drrb;
1221         if (drc->drc_byteswap) {
1222                 fletcher_4_incremental_byteswap(drr,
1223                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1224         } else {
1225                 fletcher_4_incremental_native(drr,
1226                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1227         }
1228         kmem_free(drr, sizeof (dmu_replay_record_t));
1229
1230         if (drc->drc_byteswap) {
1231                 drrb->drr_magic = BSWAP_64(drrb->drr_magic);
1232                 drrb->drr_versioninfo = BSWAP_64(drrb->drr_versioninfo);
1233                 drrb->drr_creation_time = BSWAP_64(drrb->drr_creation_time);
1234                 drrb->drr_type = BSWAP_32(drrb->drr_type);
1235                 drrb->drr_toguid = BSWAP_64(drrb->drr_toguid);
1236                 drrb->drr_fromguid = BSWAP_64(drrb->drr_fromguid);
1237         }
1238
1239         drba.drba_origin = origin;
1240         drba.drba_cookie = drc;
1241         drba.drba_cred = CRED();
1242
1243         return (dsl_sync_task(tofs, dmu_recv_begin_check, dmu_recv_begin_sync,
1244             &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1245 }
1246
1247 struct restorearg {
1248         int err;
1249         boolean_t byteswap;
1250         kthread_t *td;
1251         struct file *fp;
1252         char *buf;
1253         uint64_t voff;
1254         int bufsize; /* amount of memory allocated for buf */
1255         zio_cksum_t cksum;
1256         avl_tree_t *guid_to_ds_map;
1257 };
1258
1259 typedef struct guid_map_entry {
1260         uint64_t        guid;
1261         dsl_dataset_t   *gme_ds;
1262         avl_node_t      avlnode;
1263 } guid_map_entry_t;
1264
1265 static int
1266 guid_compare(const void *arg1, const void *arg2)
1267 {
1268         const guid_map_entry_t *gmep1 = arg1;
1269         const guid_map_entry_t *gmep2 = arg2;
1270
1271         if (gmep1->guid < gmep2->guid)
1272                 return (-1);
1273         else if (gmep1->guid > gmep2->guid)
1274                 return (1);
1275         return (0);
1276 }
1277
1278 static void
1279 free_guid_map_onexit(void *arg)
1280 {
1281         avl_tree_t *ca = arg;
1282         void *cookie = NULL;
1283         guid_map_entry_t *gmep;
1284
1285         while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
1286                 dsl_dataset_long_rele(gmep->gme_ds, gmep);
1287                 dsl_dataset_rele(gmep->gme_ds, gmep);
1288                 kmem_free(gmep, sizeof (guid_map_entry_t));
1289         }
1290         avl_destroy(ca);
1291         kmem_free(ca, sizeof (avl_tree_t));
1292 }
1293
1294 static int
1295 restore_bytes(struct restorearg *ra, void *buf, int len, off_t off, ssize_t *resid)
1296 {
1297         struct uio auio;
1298         struct iovec aiov;
1299         int error;
1300
1301         aiov.iov_base = buf;
1302         aiov.iov_len = len;
1303         auio.uio_iov = &aiov;
1304         auio.uio_iovcnt = 1;
1305         auio.uio_resid = len;
1306         auio.uio_segflg = UIO_SYSSPACE;
1307         auio.uio_rw = UIO_READ;
1308         auio.uio_offset = off;
1309         auio.uio_td = ra->td;
1310 #ifdef _KERNEL
1311         error = fo_read(ra->fp, &auio, ra->td->td_ucred, FOF_OFFSET, ra->td);
1312 #else
1313         fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__);
1314         error = EOPNOTSUPP;
1315 #endif
1316         *resid = auio.uio_resid;
1317         return (error);
1318 }
1319
1320 static void *
1321 restore_read(struct restorearg *ra, int len, char *buf)
1322 {
1323         int done = 0;
1324
1325         if (buf == NULL)
1326                 buf = ra->buf;
1327
1328         /* some things will require 8-byte alignment, so everything must */
1329         ASSERT0(len % 8);
1330         ASSERT3U(len, <=, ra->bufsize);
1331
1332         while (done < len) {
1333                 ssize_t resid;
1334
1335                 ra->err = restore_bytes(ra, buf + done,
1336                     len - done, ra->voff, &resid);
1337
1338                 if (resid == len - done)
1339                         ra->err = SET_ERROR(EINVAL);
1340                 ra->voff += len - done - resid;
1341                 done = len - resid;
1342                 if (ra->err != 0)
1343                         return (NULL);
1344         }
1345
1346         ASSERT3U(done, ==, len);
1347         if (ra->byteswap)
1348                 fletcher_4_incremental_byteswap(buf, len, &ra->cksum);
1349         else
1350                 fletcher_4_incremental_native(buf, len, &ra->cksum);
1351         return (buf);
1352 }
1353
1354 static void
1355 backup_byteswap(dmu_replay_record_t *drr)
1356 {
1357 #define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
1358 #define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
1359         drr->drr_type = BSWAP_32(drr->drr_type);
1360         drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
1361         switch (drr->drr_type) {
1362         case DRR_BEGIN:
1363                 DO64(drr_begin.drr_magic);
1364                 DO64(drr_begin.drr_versioninfo);
1365                 DO64(drr_begin.drr_creation_time);
1366                 DO32(drr_begin.drr_type);
1367                 DO32(drr_begin.drr_flags);
1368                 DO64(drr_begin.drr_toguid);
1369                 DO64(drr_begin.drr_fromguid);
1370                 break;
1371         case DRR_OBJECT:
1372                 DO64(drr_object.drr_object);
1373                 DO32(drr_object.drr_type);
1374                 DO32(drr_object.drr_bonustype);
1375                 DO32(drr_object.drr_blksz);
1376                 DO32(drr_object.drr_bonuslen);
1377                 DO64(drr_object.drr_toguid);
1378                 break;
1379         case DRR_FREEOBJECTS:
1380                 DO64(drr_freeobjects.drr_firstobj);
1381                 DO64(drr_freeobjects.drr_numobjs);
1382                 DO64(drr_freeobjects.drr_toguid);
1383                 break;
1384         case DRR_WRITE:
1385                 DO64(drr_write.drr_object);
1386                 DO32(drr_write.drr_type);
1387                 DO64(drr_write.drr_offset);
1388                 DO64(drr_write.drr_length);
1389                 DO64(drr_write.drr_toguid);
1390                 DO64(drr_write.drr_key.ddk_cksum.zc_word[0]);
1391                 DO64(drr_write.drr_key.ddk_cksum.zc_word[1]);
1392                 DO64(drr_write.drr_key.ddk_cksum.zc_word[2]);
1393                 DO64(drr_write.drr_key.ddk_cksum.zc_word[3]);
1394                 DO64(drr_write.drr_key.ddk_prop);
1395                 break;
1396         case DRR_WRITE_BYREF:
1397                 DO64(drr_write_byref.drr_object);
1398                 DO64(drr_write_byref.drr_offset);
1399                 DO64(drr_write_byref.drr_length);
1400                 DO64(drr_write_byref.drr_toguid);
1401                 DO64(drr_write_byref.drr_refguid);
1402                 DO64(drr_write_byref.drr_refobject);
1403                 DO64(drr_write_byref.drr_refoffset);
1404                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[0]);
1405                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[1]);
1406                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[2]);
1407                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[3]);
1408                 DO64(drr_write_byref.drr_key.ddk_prop);
1409                 break;
1410         case DRR_WRITE_EMBEDDED:
1411                 DO64(drr_write_embedded.drr_object);
1412                 DO64(drr_write_embedded.drr_offset);
1413                 DO64(drr_write_embedded.drr_length);
1414                 DO64(drr_write_embedded.drr_toguid);
1415                 DO32(drr_write_embedded.drr_lsize);
1416                 DO32(drr_write_embedded.drr_psize);
1417                 break;
1418         case DRR_FREE:
1419                 DO64(drr_free.drr_object);
1420                 DO64(drr_free.drr_offset);
1421                 DO64(drr_free.drr_length);
1422                 DO64(drr_free.drr_toguid);
1423                 break;
1424         case DRR_SPILL:
1425                 DO64(drr_spill.drr_object);
1426                 DO64(drr_spill.drr_length);
1427                 DO64(drr_spill.drr_toguid);
1428                 break;
1429         case DRR_END:
1430                 DO64(drr_end.drr_checksum.zc_word[0]);
1431                 DO64(drr_end.drr_checksum.zc_word[1]);
1432                 DO64(drr_end.drr_checksum.zc_word[2]);
1433                 DO64(drr_end.drr_checksum.zc_word[3]);
1434                 DO64(drr_end.drr_toguid);
1435                 break;
1436         }
1437 #undef DO64
1438 #undef DO32
1439 }
1440
1441 static inline uint8_t
1442 deduce_nblkptr(dmu_object_type_t bonus_type, uint64_t bonus_size)
1443 {
1444         if (bonus_type == DMU_OT_SA) {
1445                 return (1);
1446         } else {
1447                 return (1 +
1448                     ((DN_MAX_BONUSLEN - bonus_size) >> SPA_BLKPTRSHIFT));
1449         }
1450 }
1451
1452 static int
1453 restore_object(struct restorearg *ra, objset_t *os, struct drr_object *drro)
1454 {
1455         dmu_object_info_t doi;
1456         dmu_tx_t *tx;
1457         void *data = NULL;
1458         uint64_t object;
1459         int err;
1460
1461         if (drro->drr_type == DMU_OT_NONE ||
1462             !DMU_OT_IS_VALID(drro->drr_type) ||
1463             !DMU_OT_IS_VALID(drro->drr_bonustype) ||
1464             drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
1465             drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
1466             P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
1467             drro->drr_blksz < SPA_MINBLOCKSIZE ||
1468             drro->drr_blksz > spa_maxblocksize(dmu_objset_spa(os)) ||
1469             drro->drr_bonuslen > DN_MAX_BONUSLEN) {
1470                 return (SET_ERROR(EINVAL));
1471         }
1472
1473         err = dmu_object_info(os, drro->drr_object, &doi);
1474
1475         if (err != 0 && err != ENOENT)
1476                 return (SET_ERROR(EINVAL));
1477         object = err == 0 ? drro->drr_object : DMU_NEW_OBJECT;
1478
1479         if (drro->drr_bonuslen) {
1480                 data = restore_read(ra, P2ROUNDUP(drro->drr_bonuslen, 8), NULL);
1481                 if (ra->err != 0)
1482                         return (ra->err);
1483         }
1484
1485         /*
1486          * If we are losing blkptrs or changing the block size this must
1487          * be a new file instance.  We must clear out the previous file
1488          * contents before we can change this type of metadata in the dnode.
1489          */
1490         if (err == 0) {
1491                 int nblkptr;
1492
1493                 nblkptr = deduce_nblkptr(drro->drr_bonustype,
1494                     drro->drr_bonuslen);
1495
1496                 if (drro->drr_blksz != doi.doi_data_block_size ||
1497                     nblkptr < doi.doi_nblkptr) {
1498                         err = dmu_free_long_range(os, drro->drr_object,
1499                             0, DMU_OBJECT_END);
1500                         if (err != 0)
1501                                 return (SET_ERROR(EINVAL));
1502                 }
1503         }
1504
1505         tx = dmu_tx_create(os);
1506         dmu_tx_hold_bonus(tx, object);
1507         err = dmu_tx_assign(tx, TXG_WAIT);
1508         if (err != 0) {
1509                 dmu_tx_abort(tx);
1510                 return (err);
1511         }
1512
1513         if (object == DMU_NEW_OBJECT) {
1514                 /* currently free, want to be allocated */
1515                 err = dmu_object_claim(os, drro->drr_object,
1516                     drro->drr_type, drro->drr_blksz,
1517                     drro->drr_bonustype, drro->drr_bonuslen, tx);
1518         } else if (drro->drr_type != doi.doi_type ||
1519             drro->drr_blksz != doi.doi_data_block_size ||
1520             drro->drr_bonustype != doi.doi_bonus_type ||
1521             drro->drr_bonuslen != doi.doi_bonus_size) {
1522                 /* currently allocated, but with different properties */
1523                 err = dmu_object_reclaim(os, drro->drr_object,
1524                     drro->drr_type, drro->drr_blksz,
1525                     drro->drr_bonustype, drro->drr_bonuslen, tx);
1526         }
1527         if (err != 0) {
1528                 dmu_tx_commit(tx);
1529                 return (SET_ERROR(EINVAL));
1530         }
1531
1532         dmu_object_set_checksum(os, drro->drr_object, drro->drr_checksumtype,
1533             tx);
1534         dmu_object_set_compress(os, drro->drr_object, drro->drr_compress, tx);
1535
1536         if (data != NULL) {
1537                 dmu_buf_t *db;
1538
1539                 VERIFY(0 == dmu_bonus_hold(os, drro->drr_object, FTAG, &db));
1540                 dmu_buf_will_dirty(db, tx);
1541
1542                 ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
1543                 bcopy(data, db->db_data, drro->drr_bonuslen);
1544                 if (ra->byteswap) {
1545                         dmu_object_byteswap_t byteswap =
1546                             DMU_OT_BYTESWAP(drro->drr_bonustype);
1547                         dmu_ot_byteswap[byteswap].ob_func(db->db_data,
1548                             drro->drr_bonuslen);
1549                 }
1550                 dmu_buf_rele(db, FTAG);
1551         }
1552         dmu_tx_commit(tx);
1553         return (0);
1554 }
1555
1556 /* ARGSUSED */
1557 static int
1558 restore_freeobjects(struct restorearg *ra, objset_t *os,
1559     struct drr_freeobjects *drrfo)
1560 {
1561         uint64_t obj;
1562
1563         if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
1564                 return (SET_ERROR(EINVAL));
1565
1566         for (obj = drrfo->drr_firstobj;
1567             obj < drrfo->drr_firstobj + drrfo->drr_numobjs;
1568             (void) dmu_object_next(os, &obj, FALSE, 0)) {
1569                 int err;
1570
1571                 if (dmu_object_info(os, obj, NULL) != 0)
1572                         continue;
1573
1574                 err = dmu_free_long_object(os, obj);
1575                 if (err != 0)
1576                         return (err);
1577         }
1578         return (0);
1579 }
1580
1581 static int
1582 restore_write(struct restorearg *ra, objset_t *os,
1583     struct drr_write *drrw)
1584 {
1585         dmu_tx_t *tx;
1586         void *data;
1587         int err;
1588
1589         if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
1590             !DMU_OT_IS_VALID(drrw->drr_type))
1591                 return (SET_ERROR(EINVAL));
1592
1593         if (dmu_object_info(os, drrw->drr_object, NULL) != 0)
1594                 return (SET_ERROR(EINVAL));
1595
1596         dmu_buf_t *bonus;
1597         if (dmu_bonus_hold(os, drrw->drr_object, FTAG, &bonus) != 0)
1598                 return (SET_ERROR(EINVAL));
1599
1600         arc_buf_t *abuf = dmu_request_arcbuf(bonus, drrw->drr_length);
1601
1602         data = restore_read(ra, drrw->drr_length, abuf->b_data);
1603         if (data == NULL) {
1604                 dmu_return_arcbuf(abuf);
1605                 dmu_buf_rele(bonus, FTAG);
1606                 return (ra->err);
1607         }
1608
1609         tx = dmu_tx_create(os);
1610
1611         dmu_tx_hold_write(tx, drrw->drr_object,
1612             drrw->drr_offset, drrw->drr_length);
1613         err = dmu_tx_assign(tx, TXG_WAIT);
1614         if (err != 0) {
1615                 dmu_return_arcbuf(abuf);
1616                 dmu_buf_rele(bonus, FTAG);
1617                 dmu_tx_abort(tx);
1618                 return (err);
1619         }
1620         if (ra->byteswap) {
1621                 dmu_object_byteswap_t byteswap =
1622                     DMU_OT_BYTESWAP(drrw->drr_type);
1623                 dmu_ot_byteswap[byteswap].ob_func(data, drrw->drr_length);
1624         }
1625         dmu_assign_arcbuf(bonus, drrw->drr_offset, abuf, tx);
1626         dmu_tx_commit(tx);
1627         dmu_buf_rele(bonus, FTAG);
1628         return (0);
1629 }
1630
1631 /*
1632  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
1633  * streams to refer to a copy of the data that is already on the
1634  * system because it came in earlier in the stream.  This function
1635  * finds the earlier copy of the data, and uses that copy instead of
1636  * data from the stream to fulfill this write.
1637  */
1638 static int
1639 restore_write_byref(struct restorearg *ra, objset_t *os,
1640     struct drr_write_byref *drrwbr)
1641 {
1642         dmu_tx_t *tx;
1643         int err;
1644         guid_map_entry_t gmesrch;
1645         guid_map_entry_t *gmep;
1646         avl_index_t where;
1647         objset_t *ref_os = NULL;
1648         dmu_buf_t *dbp;
1649
1650         if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
1651                 return (SET_ERROR(EINVAL));
1652
1653         /*
1654          * If the GUID of the referenced dataset is different from the
1655          * GUID of the target dataset, find the referenced dataset.
1656          */
1657         if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
1658                 gmesrch.guid = drrwbr->drr_refguid;
1659                 if ((gmep = avl_find(ra->guid_to_ds_map, &gmesrch,
1660                     &where)) == NULL) {
1661                         return (SET_ERROR(EINVAL));
1662                 }
1663                 if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
1664                         return (SET_ERROR(EINVAL));
1665         } else {
1666                 ref_os = os;
1667         }
1668
1669         err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
1670             drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH);
1671         if (err != 0)
1672                 return (err);
1673
1674         tx = dmu_tx_create(os);
1675
1676         dmu_tx_hold_write(tx, drrwbr->drr_object,
1677             drrwbr->drr_offset, drrwbr->drr_length);
1678         err = dmu_tx_assign(tx, TXG_WAIT);
1679         if (err != 0) {
1680                 dmu_tx_abort(tx);
1681                 return (err);
1682         }
1683         dmu_write(os, drrwbr->drr_object,
1684             drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
1685         dmu_buf_rele(dbp, FTAG);
1686         dmu_tx_commit(tx);
1687         return (0);
1688 }
1689
1690 static int
1691 restore_write_embedded(struct restorearg *ra, objset_t *os,
1692     struct drr_write_embedded *drrwnp)
1693 {
1694         dmu_tx_t *tx;
1695         int err;
1696         void *data;
1697
1698         if (drrwnp->drr_offset + drrwnp->drr_length < drrwnp->drr_offset)
1699                 return (EINVAL);
1700
1701         if (drrwnp->drr_psize > BPE_PAYLOAD_SIZE)
1702                 return (EINVAL);
1703
1704         if (drrwnp->drr_etype >= NUM_BP_EMBEDDED_TYPES)
1705                 return (EINVAL);
1706         if (drrwnp->drr_compression >= ZIO_COMPRESS_FUNCTIONS)
1707                 return (EINVAL);
1708
1709         data = restore_read(ra, P2ROUNDUP(drrwnp->drr_psize, 8), NULL);
1710         if (data == NULL)
1711                 return (ra->err);
1712
1713         tx = dmu_tx_create(os);
1714
1715         dmu_tx_hold_write(tx, drrwnp->drr_object,
1716             drrwnp->drr_offset, drrwnp->drr_length);
1717         err = dmu_tx_assign(tx, TXG_WAIT);
1718         if (err != 0) {
1719                 dmu_tx_abort(tx);
1720                 return (err);
1721         }
1722
1723         dmu_write_embedded(os, drrwnp->drr_object,
1724             drrwnp->drr_offset, data, drrwnp->drr_etype,
1725             drrwnp->drr_compression, drrwnp->drr_lsize, drrwnp->drr_psize,
1726             ra->byteswap ^ ZFS_HOST_BYTEORDER, tx);
1727
1728         dmu_tx_commit(tx);
1729         return (0);
1730 }
1731
1732 static int
1733 restore_spill(struct restorearg *ra, objset_t *os, struct drr_spill *drrs)
1734 {
1735         dmu_tx_t *tx;
1736         void *data;
1737         dmu_buf_t *db, *db_spill;
1738         int err;
1739
1740         if (drrs->drr_length < SPA_MINBLOCKSIZE ||
1741             drrs->drr_length > spa_maxblocksize(dmu_objset_spa(os)))
1742                 return (SET_ERROR(EINVAL));
1743
1744         data = restore_read(ra, drrs->drr_length, NULL);
1745         if (data == NULL)
1746                 return (ra->err);
1747
1748         if (dmu_object_info(os, drrs->drr_object, NULL) != 0)
1749                 return (SET_ERROR(EINVAL));
1750
1751         VERIFY(0 == dmu_bonus_hold(os, drrs->drr_object, FTAG, &db));
1752         if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
1753                 dmu_buf_rele(db, FTAG);
1754                 return (err);
1755         }
1756
1757         tx = dmu_tx_create(os);
1758
1759         dmu_tx_hold_spill(tx, db->db_object);
1760
1761         err = dmu_tx_assign(tx, TXG_WAIT);
1762         if (err != 0) {
1763                 dmu_buf_rele(db, FTAG);
1764                 dmu_buf_rele(db_spill, FTAG);
1765                 dmu_tx_abort(tx);
1766                 return (err);
1767         }
1768         dmu_buf_will_dirty(db_spill, tx);
1769
1770         if (db_spill->db_size < drrs->drr_length)
1771                 VERIFY(0 == dbuf_spill_set_blksz(db_spill,
1772                     drrs->drr_length, tx));
1773         bcopy(data, db_spill->db_data, drrs->drr_length);
1774
1775         dmu_buf_rele(db, FTAG);
1776         dmu_buf_rele(db_spill, FTAG);
1777
1778         dmu_tx_commit(tx);
1779         return (0);
1780 }
1781
1782 /* ARGSUSED */
1783 static int
1784 restore_free(struct restorearg *ra, objset_t *os,
1785     struct drr_free *drrf)
1786 {
1787         int err;
1788
1789         if (drrf->drr_length != -1ULL &&
1790             drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
1791                 return (SET_ERROR(EINVAL));
1792
1793         if (dmu_object_info(os, drrf->drr_object, NULL) != 0)
1794                 return (SET_ERROR(EINVAL));
1795
1796         err = dmu_free_long_range(os, drrf->drr_object,
1797             drrf->drr_offset, drrf->drr_length);
1798         return (err);
1799 }
1800
1801 /* used to destroy the drc_ds on error */
1802 static void
1803 dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc)
1804 {
1805         char name[MAXNAMELEN];
1806         dsl_dataset_name(drc->drc_ds, name);
1807         dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
1808         (void) dsl_destroy_head(name);
1809 }
1810
1811 /*
1812  * NB: callers *must* call dmu_recv_end() if this succeeds.
1813  */
1814 int
1815 dmu_recv_stream(dmu_recv_cookie_t *drc, struct file *fp, offset_t *voffp,
1816     int cleanup_fd, uint64_t *action_handlep)
1817 {
1818         struct restorearg ra = { 0 };
1819         dmu_replay_record_t *drr;
1820         objset_t *os;
1821         zio_cksum_t pcksum;
1822         int featureflags;
1823
1824         ra.byteswap = drc->drc_byteswap;
1825         ra.cksum = drc->drc_cksum;
1826         ra.td = curthread;
1827         ra.fp = fp;
1828         ra.voff = *voffp;
1829         ra.bufsize = SPA_MAXBLOCKSIZE;
1830         ra.buf = kmem_alloc(ra.bufsize, KM_SLEEP);
1831
1832         /* these were verified in dmu_recv_begin */
1833         ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==,
1834             DMU_SUBSTREAM);
1835         ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES);
1836
1837         /*
1838          * Open the objset we are modifying.
1839          */
1840         VERIFY0(dmu_objset_from_ds(drc->drc_ds, &os));
1841
1842         ASSERT(dsl_dataset_phys(drc->drc_ds)->ds_flags & DS_FLAG_INCONSISTENT);
1843
1844         featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
1845
1846         /* if this stream is dedup'ed, set up the avl tree for guid mapping */
1847         if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
1848                 minor_t minor;
1849
1850                 if (cleanup_fd == -1) {
1851                         ra.err = SET_ERROR(EBADF);
1852                         goto out;
1853                 }
1854                 ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
1855                 if (ra.err != 0) {
1856                         cleanup_fd = -1;
1857                         goto out;
1858                 }
1859
1860                 if (*action_handlep == 0) {
1861                         ra.guid_to_ds_map =
1862                             kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
1863                         avl_create(ra.guid_to_ds_map, guid_compare,
1864                             sizeof (guid_map_entry_t),
1865                             offsetof(guid_map_entry_t, avlnode));
1866                         ra.err = zfs_onexit_add_cb(minor,
1867                             free_guid_map_onexit, ra.guid_to_ds_map,
1868                             action_handlep);
1869                         if (ra.err != 0)
1870                                 goto out;
1871                 } else {
1872                         ra.err = zfs_onexit_cb_data(minor, *action_handlep,
1873                             (void **)&ra.guid_to_ds_map);
1874                         if (ra.err != 0)
1875                                 goto out;
1876                 }
1877
1878                 drc->drc_guid_to_ds_map = ra.guid_to_ds_map;
1879         }
1880
1881         /*
1882          * Read records and process them.
1883          */
1884         pcksum = ra.cksum;
1885         while (ra.err == 0 &&
1886             NULL != (drr = restore_read(&ra, sizeof (*drr), NULL))) {
1887                 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1888                         ra.err = SET_ERROR(EINTR);
1889                         goto out;
1890                 }
1891
1892                 if (ra.byteswap)
1893                         backup_byteswap(drr);
1894
1895                 switch (drr->drr_type) {
1896                 case DRR_OBJECT:
1897                 {
1898                         /*
1899                          * We need to make a copy of the record header,
1900                          * because restore_{object,write} may need to
1901                          * restore_read(), which will invalidate drr.
1902                          */
1903                         struct drr_object drro = drr->drr_u.drr_object;
1904                         ra.err = restore_object(&ra, os, &drro);
1905                         break;
1906                 }
1907                 case DRR_FREEOBJECTS:
1908                 {
1909                         struct drr_freeobjects drrfo =
1910                             drr->drr_u.drr_freeobjects;
1911                         ra.err = restore_freeobjects(&ra, os, &drrfo);
1912                         break;
1913                 }
1914                 case DRR_WRITE:
1915                 {
1916                         struct drr_write drrw = drr->drr_u.drr_write;
1917                         ra.err = restore_write(&ra, os, &drrw);
1918                         break;
1919                 }
1920                 case DRR_WRITE_BYREF:
1921                 {
1922                         struct drr_write_byref drrwbr =
1923                             drr->drr_u.drr_write_byref;
1924                         ra.err = restore_write_byref(&ra, os, &drrwbr);
1925                         break;
1926                 }
1927                 case DRR_WRITE_EMBEDDED:
1928                 {
1929                         struct drr_write_embedded drrwe =
1930                             drr->drr_u.drr_write_embedded;
1931                         ra.err = restore_write_embedded(&ra, os, &drrwe);
1932                         break;
1933                 }
1934                 case DRR_FREE:
1935                 {
1936                         struct drr_free drrf = drr->drr_u.drr_free;
1937                         ra.err = restore_free(&ra, os, &drrf);
1938                         break;
1939                 }
1940                 case DRR_END:
1941                 {
1942                         struct drr_end drre = drr->drr_u.drr_end;
1943                         /*
1944                          * We compare against the *previous* checksum
1945                          * value, because the stored checksum is of
1946                          * everything before the DRR_END record.
1947                          */
1948                         if (!ZIO_CHECKSUM_EQUAL(drre.drr_checksum, pcksum))
1949                                 ra.err = SET_ERROR(ECKSUM);
1950                         goto out;
1951                 }
1952                 case DRR_SPILL:
1953                 {
1954                         struct drr_spill drrs = drr->drr_u.drr_spill;
1955                         ra.err = restore_spill(&ra, os, &drrs);
1956                         break;
1957                 }
1958                 default:
1959                         ra.err = SET_ERROR(EINVAL);
1960                         goto out;
1961                 }
1962                 pcksum = ra.cksum;
1963         }
1964         ASSERT(ra.err != 0);
1965
1966 out:
1967         if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
1968                 zfs_onexit_fd_rele(cleanup_fd);
1969
1970         if (ra.err != 0) {
1971                 /*
1972                  * destroy what we created, so we don't leave it in the
1973                  * inconsistent restoring state.
1974                  */
1975                 dmu_recv_cleanup_ds(drc);
1976         }
1977
1978         kmem_free(ra.buf, ra.bufsize);
1979         *voffp = ra.voff;
1980         return (ra.err);
1981 }
1982
1983 static int
1984 dmu_recv_end_check(void *arg, dmu_tx_t *tx)
1985 {
1986         dmu_recv_cookie_t *drc = arg;
1987         dsl_pool_t *dp = dmu_tx_pool(tx);
1988         int error;
1989
1990         ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag);
1991
1992         if (!drc->drc_newfs) {
1993                 dsl_dataset_t *origin_head;
1994
1995                 error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head);
1996                 if (error != 0)
1997                         return (error);
1998                 if (drc->drc_force) {
1999                         /*
2000                          * We will destroy any snapshots in tofs (i.e. before
2001                          * origin_head) that are after the origin (which is
2002                          * the snap before drc_ds, because drc_ds can not
2003                          * have any snaps of its own).
2004                          */
2005                         uint64_t obj;
2006
2007                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
2008                         while (obj !=
2009                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
2010                                 dsl_dataset_t *snap;
2011                                 error = dsl_dataset_hold_obj(dp, obj, FTAG,
2012                                     &snap);
2013                                 if (error != 0)
2014                                         break;
2015                                 if (snap->ds_dir != origin_head->ds_dir)
2016                                         error = SET_ERROR(EINVAL);
2017                                 if (error == 0)  {
2018                                         error = dsl_destroy_snapshot_check_impl(
2019                                             snap, B_FALSE);
2020                                 }
2021                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
2022                                 dsl_dataset_rele(snap, FTAG);
2023                                 if (error != 0)
2024                                         break;
2025                         }
2026                         if (error != 0) {
2027                                 dsl_dataset_rele(origin_head, FTAG);
2028                                 return (error);
2029                         }
2030                 }
2031                 error = dsl_dataset_clone_swap_check_impl(drc->drc_ds,
2032                     origin_head, drc->drc_force, drc->drc_owner, tx);
2033                 if (error != 0) {
2034                         dsl_dataset_rele(origin_head, FTAG);
2035                         return (error);
2036                 }
2037                 error = dsl_dataset_snapshot_check_impl(origin_head,
2038                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
2039                 dsl_dataset_rele(origin_head, FTAG);
2040                 if (error != 0)
2041                         return (error);
2042
2043                 error = dsl_destroy_head_check_impl(drc->drc_ds, 1);
2044         } else {
2045                 error = dsl_dataset_snapshot_check_impl(drc->drc_ds,
2046                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
2047         }
2048         return (error);
2049 }
2050
2051 static void
2052 dmu_recv_end_sync(void *arg, dmu_tx_t *tx)
2053 {
2054         dmu_recv_cookie_t *drc = arg;
2055         dsl_pool_t *dp = dmu_tx_pool(tx);
2056
2057         spa_history_log_internal_ds(drc->drc_ds, "finish receiving",
2058             tx, "snap=%s", drc->drc_tosnap);
2059
2060         if (!drc->drc_newfs) {
2061                 dsl_dataset_t *origin_head;
2062
2063                 VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG,
2064                     &origin_head));
2065
2066                 if (drc->drc_force) {
2067                         /*
2068                          * Destroy any snapshots of drc_tofs (origin_head)
2069                          * after the origin (the snap before drc_ds).
2070                          */
2071                         uint64_t obj;
2072
2073                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
2074                         while (obj !=
2075                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
2076                                 dsl_dataset_t *snap;
2077                                 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG,
2078                                     &snap));
2079                                 ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir);
2080                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
2081                                 dsl_destroy_snapshot_sync_impl(snap,
2082                                     B_FALSE, tx);
2083                                 dsl_dataset_rele(snap, FTAG);
2084                         }
2085                 }
2086                 VERIFY3P(drc->drc_ds->ds_prev, ==,
2087                     origin_head->ds_prev);
2088
2089                 dsl_dataset_clone_swap_sync_impl(drc->drc_ds,
2090                     origin_head, tx);
2091                 dsl_dataset_snapshot_sync_impl(origin_head,
2092                     drc->drc_tosnap, tx);
2093
2094                 /* set snapshot's creation time and guid */
2095                 dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx);
2096                 dsl_dataset_phys(origin_head->ds_prev)->ds_creation_time =
2097                     drc->drc_drrb->drr_creation_time;
2098                 dsl_dataset_phys(origin_head->ds_prev)->ds_guid =
2099                     drc->drc_drrb->drr_toguid;
2100                 dsl_dataset_phys(origin_head->ds_prev)->ds_flags &=
2101                     ~DS_FLAG_INCONSISTENT;
2102
2103                 dmu_buf_will_dirty(origin_head->ds_dbuf, tx);
2104                 dsl_dataset_phys(origin_head)->ds_flags &=
2105                     ~DS_FLAG_INCONSISTENT;
2106
2107                 dsl_dataset_rele(origin_head, FTAG);
2108                 dsl_destroy_head_sync_impl(drc->drc_ds, tx);
2109
2110                 if (drc->drc_owner != NULL)
2111                         VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner);
2112         } else {
2113                 dsl_dataset_t *ds = drc->drc_ds;
2114
2115                 dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx);
2116
2117                 /* set snapshot's creation time and guid */
2118                 dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
2119                 dsl_dataset_phys(ds->ds_prev)->ds_creation_time =
2120                     drc->drc_drrb->drr_creation_time;
2121                 dsl_dataset_phys(ds->ds_prev)->ds_guid =
2122                     drc->drc_drrb->drr_toguid;
2123                 dsl_dataset_phys(ds->ds_prev)->ds_flags &=
2124                     ~DS_FLAG_INCONSISTENT;
2125
2126                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
2127                 dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
2128         }
2129         drc->drc_newsnapobj = dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj;
2130         /*
2131          * Release the hold from dmu_recv_begin.  This must be done before
2132          * we return to open context, so that when we free the dataset's dnode,
2133          * we can evict its bonus buffer.
2134          */
2135         dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2136         drc->drc_ds = NULL;
2137 }
2138
2139 static int
2140 add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj)
2141 {
2142         dsl_pool_t *dp;
2143         dsl_dataset_t *snapds;
2144         guid_map_entry_t *gmep;
2145         int err;
2146
2147         ASSERT(guid_map != NULL);
2148
2149         err = dsl_pool_hold(name, FTAG, &dp);
2150         if (err != 0)
2151                 return (err);
2152         gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP);
2153         err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds);
2154         if (err == 0) {
2155                 gmep->guid = dsl_dataset_phys(snapds)->ds_guid;
2156                 gmep->gme_ds = snapds;
2157                 avl_add(guid_map, gmep);
2158                 dsl_dataset_long_hold(snapds, gmep);
2159         } else
2160                 kmem_free(gmep, sizeof (*gmep));
2161
2162         dsl_pool_rele(dp, FTAG);
2163         return (err);
2164 }
2165
2166 static int dmu_recv_end_modified_blocks = 3;
2167
2168 static int
2169 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
2170 {
2171         int error;
2172         char name[MAXNAMELEN];
2173
2174 #ifdef _KERNEL
2175         /*
2176          * We will be destroying the ds; make sure its origin is unmounted if
2177          * necessary.
2178          */
2179         dsl_dataset_name(drc->drc_ds, name);
2180         zfs_destroy_unmount_origin(name);
2181 #endif
2182
2183         error = dsl_sync_task(drc->drc_tofs,
2184             dmu_recv_end_check, dmu_recv_end_sync, drc,
2185             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL);
2186
2187         if (error != 0)
2188                 dmu_recv_cleanup_ds(drc);
2189         return (error);
2190 }
2191
2192 static int
2193 dmu_recv_new_end(dmu_recv_cookie_t *drc)
2194 {
2195         int error;
2196
2197         error = dsl_sync_task(drc->drc_tofs,
2198             dmu_recv_end_check, dmu_recv_end_sync, drc,
2199             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL);
2200
2201         if (error != 0) {
2202                 dmu_recv_cleanup_ds(drc);
2203         } else if (drc->drc_guid_to_ds_map != NULL) {
2204                 (void) add_ds_to_guidmap(drc->drc_tofs,
2205                     drc->drc_guid_to_ds_map,
2206                     drc->drc_newsnapobj);
2207         }
2208         return (error);
2209 }
2210
2211 int
2212 dmu_recv_end(dmu_recv_cookie_t *drc, void *owner)
2213 {
2214         drc->drc_owner = owner;
2215
2216         if (drc->drc_newfs)
2217                 return (dmu_recv_new_end(drc));
2218         else
2219                 return (dmu_recv_existing_end(drc));
2220 }
2221
2222 /*
2223  * Return TRUE if this objset is currently being received into.
2224  */
2225 boolean_t
2226 dmu_objset_is_receiving(objset_t *os)
2227 {
2228         return (os->os_dsl_dataset != NULL &&
2229             os->os_dsl_dataset->ds_owner == dmu_recv_tag);
2230 }