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                 uint32_t aflags = ARC_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                 uint32_t aflags = ARC_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                 uint32_t aflags = ARC_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             ds->ds_phys->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 = ds->ds_phys->ds_guid;
 631         if (ds->ds_phys->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 = ds->ds_phys->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 = fromds->ds_phys->ds_creation_time;
 744                 zb.zbm_creation_txg = fromds->ds_phys->ds_creation_txg;
 745                 zb.zbm_guid = fromds->ds_phys->ds_guid;
 746                 is_clone = (fromds->ds_dir != ds->ds_dir);
 747                 dsl_dataset_rele(fromds, FTAG);
 748                 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
 749                     embedok, large_block_ok, outfd, fp, off);
 750         } else {
 751                 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
 752                     embedok, large_block_ok, outfd, fp, off);
 753         }
 754         dsl_dataset_rele(ds, FTAG);
 755         return (err);
 756 }
 757
 758 int
 759 dmu_send(const char *tosnap, const char *fromsnap,
 760     boolean_t embedok, boolean_t large_block_ok,
 761 #ifdef illumos
 762     int outfd, vnode_t *vp, offset_t *off)
 763 #else
 764     int outfd, struct file *fp, offset_t *off)
 765 #endif
 766 {
 767         dsl_pool_t *dp;
 768         dsl_dataset_t *ds;
 769         int err;
 770         boolean_t owned = B_FALSE;
 771
 772         if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL)
 773                 return (SET_ERROR(EINVAL));
 774
 775         err = dsl_pool_hold(tosnap, FTAG, &dp);
 776         if (err != 0)
 777                 return (err);
 778
 779         if (strchr(tosnap, '@') == NULL && spa_writeable(dp->dp_spa)) {
 780                 /*
 781                  * We are sending a filesystem or volume.  Ensure
 782                  * that it doesn't change by owning the dataset.
 783                  */
 784                 err = dsl_dataset_own(dp, tosnap, FTAG, &ds);
 785                 owned = B_TRUE;
 786         } else {
 787                 err = dsl_dataset_hold(dp, tosnap, FTAG, &ds);
 788         }
 789         if (err != 0) {
 790                 dsl_pool_rele(dp, FTAG);
 791                 return (err);
 792         }
 793
 794         if (fromsnap != NULL) {
 795                 zfs_bookmark_phys_t zb;
 796                 boolean_t is_clone = B_FALSE;
 797                 int fsnamelen = strchr(tosnap, '@') - tosnap;
 798
 799                 /*
 800                  * If the fromsnap is in a different filesystem, then
 801                  * mark the send stream as a clone.
 802                  */
 803                 if (strncmp(tosnap, fromsnap, fsnamelen) != 0 ||
 804                     (fromsnap[fsnamelen] != '@' &&
 805                     fromsnap[fsnamelen] != '#')) {
 806                         is_clone = B_TRUE;
 807                 }
 808
 809                 if (strchr(fromsnap, '@')) {
 810                         dsl_dataset_t *fromds;
 811                         err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds);
 812                         if (err == 0) {
 813                                 if (!dsl_dataset_is_before(ds, fromds, 0))
 814                                         err = SET_ERROR(EXDEV);
 815                                 zb.zbm_creation_time =
 816                                     fromds->ds_phys->ds_creation_time;
 817                                 zb.zbm_creation_txg =
 818                                     fromds->ds_phys->ds_creation_txg;
 819                                 zb.zbm_guid = fromds->ds_phys->ds_guid;
 820                                 is_clone = (ds->ds_dir != fromds->ds_dir);
 821                                 dsl_dataset_rele(fromds, FTAG);
 822                         }
 823                 } else {
 824                         err = dsl_bookmark_lookup(dp, fromsnap, ds, &zb);
 825                 }
 826                 if (err != 0) {
 827                         dsl_dataset_rele(ds, FTAG);
 828                         dsl_pool_rele(dp, FTAG);
 829                         return (err);
 830                 }
 831                 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
 832                     embedok, large_block_ok, outfd, fp, off);
 833         } else {
 834                 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
 835                     embedok, large_block_ok, outfd, fp, off);
 836         }
 837         if (owned)
 838                 dsl_dataset_disown(ds, FTAG);
 839         else
 840                 dsl_dataset_rele(ds, FTAG);
 841         return (err);
 842 }
 843
 844 int
 845 dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds, uint64_t *sizep)
 846 {
 847         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 848         int err;
 849         uint64_t size;
 850
 851         ASSERT(dsl_pool_config_held(dp));
 852
 853         /* tosnap must be a snapshot */
 854         if (!dsl_dataset_is_snapshot(ds))
 855                 return (SET_ERROR(EINVAL));
 856
 857         /*
 858          * fromsnap must be an earlier snapshot from the same fs as tosnap,
 859          * or the origin's fs.
 860          */
 861         if (fromds != NULL && !dsl_dataset_is_before(ds, fromds, 0))
 862                 return (SET_ERROR(EXDEV));
 863
 864         /* Get uncompressed size estimate of changed data. */
 865         if (fromds == NULL) {
 866                 size = ds->ds_phys->ds_uncompressed_bytes;
 867         } else {
 868                 uint64_t used, comp;
 869                 err = dsl_dataset_space_written(fromds, ds,
 870                     &used, &comp, &size);
 871                 if (err != 0)
 872                         return (err);
 873         }
 874
 875         /*
 876          * Assume that space (both on-disk and in-stream) is dominated by
 877          * data.  We will adjust for indirect blocks and the copies property,
 878          * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
 879          */
 880
 881         /*
 882          * Subtract out approximate space used by indirect blocks.
 883          * Assume most space is used by data blocks (non-indirect, non-dnode).
 884          * Assume all blocks are recordsize.  Assume ditto blocks and
 885          * internal fragmentation counter out compression.
 886          *
 887          * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
 888          * block, which we observe in practice.
 889          */
 890         uint64_t recordsize;
 891         err = dsl_prop_get_int_ds(ds, "recordsize", &recordsize);
 892         if (err != 0)
 893                 return (err);
 894         size -= size / recordsize * sizeof (blkptr_t);
 895
 896         /* Add in the space for the record associated with each block. */
 897         size += size / recordsize * sizeof (dmu_replay_record_t);
 898
 899         *sizep = size;
 900
 901         return (0);
 902 }
 903
 904 typedef struct dmu_recv_begin_arg {
 905         const char *drba_origin;
 906         dmu_recv_cookie_t *drba_cookie;
 907         cred_t *drba_cred;
 908         uint64_t drba_snapobj;
 909 } dmu_recv_begin_arg_t;
 910
 911 static int
 912 recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds,
 913     uint64_t fromguid)
 914 {
 915         uint64_t val;
 916         int error;
 917         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 918
 919         /* temporary clone name must not exist */
 920         error = zap_lookup(dp->dp_meta_objset,
 921             ds->ds_dir->dd_phys->dd_child_dir_zapobj, recv_clone_name,
 922             8, 1, &val);
 923         if (error != ENOENT)
 924                 return (error == 0 ? EBUSY : error);
 925
 926         /* new snapshot name must not exist */
 927         error = zap_lookup(dp->dp_meta_objset,
 928             ds->ds_phys->ds_snapnames_zapobj, drba->drba_cookie->drc_tosnap,
 929             8, 1, &val);
 930         if (error != ENOENT)
 931                 return (error == 0 ? EEXIST : error);
 932
 933         /*
 934          * Check snapshot limit before receiving. We'll recheck again at the
 935          * end, but might as well abort before receiving if we're already over
 936          * the limit.
 937          *
 938          * Note that we do not check the file system limit with
 939          * dsl_dir_fscount_check because the temporary %clones don't count
 940          * against that limit.
 941          */
 942         error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT,
 943             NULL, drba->drba_cred);
 944         if (error != 0)
 945                 return (error);
 946
 947         if (fromguid != 0) {
 948                 dsl_dataset_t *snap;
 949                 uint64_t obj = ds->ds_phys->ds_prev_snap_obj;
 950
 951                 /* Find snapshot in this dir that matches fromguid. */
 952                 while (obj != 0) {
 953                         error = dsl_dataset_hold_obj(dp, obj, FTAG,
 954                             &snap);
 955                         if (error != 0)
 956                                 return (SET_ERROR(ENODEV));
 957                         if (snap->ds_dir != ds->ds_dir) {
 958                                 dsl_dataset_rele(snap, FTAG);
 959                                 return (SET_ERROR(ENODEV));
 960                         }
 961                         if (snap->ds_phys->ds_guid == fromguid)
 962                                 break;
 963                         obj = snap->ds_phys->ds_prev_snap_obj;
 964                         dsl_dataset_rele(snap, FTAG);
 965                 }
 966                 if (obj == 0)
 967                         return (SET_ERROR(ENODEV));
 968
 969                 if (drba->drba_cookie->drc_force) {
 970                         drba->drba_snapobj = obj;
 971                 } else {
 972                         /*
 973                          * If we are not forcing, there must be no
 974                          * changes since fromsnap.
 975                          */
 976                         if (dsl_dataset_modified_since_snap(ds, snap)) {
 977                                 dsl_dataset_rele(snap, FTAG);
 978                                 return (SET_ERROR(ETXTBSY));
 979                         }
 980                         drba->drba_snapobj = ds->ds_prev->ds_object;
 981                 }
 982
 983                 dsl_dataset_rele(snap, FTAG);
 984         } else {
 985                 /* if full, most recent snapshot must be $ORIGIN */
 986                 if (ds->ds_phys->ds_prev_snap_txg >= TXG_INITIAL)
 987                         return (SET_ERROR(ENODEV));
 988                 drba->drba_snapobj = ds->ds_phys->ds_prev_snap_obj;
 989         }
 990
 991         return (0);
 992
 993 }
 994
 995 static int
 996 dmu_recv_begin_check(void *arg, dmu_tx_t *tx)
 997 {
 998         dmu_recv_begin_arg_t *drba = arg;
 999         dsl_pool_t *dp = dmu_tx_pool(tx);
1000         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1001         uint64_t fromguid = drrb->drr_fromguid;
1002         int flags = drrb->drr_flags;
1003         int error;
1004         uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1005         dsl_dataset_t *ds;
1006         const char *tofs = drba->drba_cookie->drc_tofs;
1007
1008         /* already checked */
1009         ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1010
1011         if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1012             DMU_COMPOUNDSTREAM ||
1013             drrb->drr_type >= DMU_OST_NUMTYPES ||
1014             ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL))
1015                 return (SET_ERROR(EINVAL));
1016
1017         /* Verify pool version supports SA if SA_SPILL feature set */
1018         if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1019             spa_version(dp->dp_spa) < SPA_VERSION_SA)
1020                 return (SET_ERROR(ENOTSUP));
1021
1022         /*
1023          * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1024          * record to a plan WRITE record, so the pool must have the
1025          * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1026          * records.  Same with WRITE_EMBEDDED records that use LZ4 compression.
1027          */
1028         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1029             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1030                 return (SET_ERROR(ENOTSUP));
1031         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4) &&
1032             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1033                 return (SET_ERROR(ENOTSUP));
1034
1035         /*
1036          * The receiving code doesn't know how to translate large blocks
1037          * to smaller ones, so the pool must have the LARGE_BLOCKS
1038          * feature enabled if the stream has LARGE_BLOCKS.
1039          */
1040         if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
1041             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS))
1042                 return (SET_ERROR(ENOTSUP));
1043
1044         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1045         if (error == 0) {
1046                 /* target fs already exists; recv into temp clone */
1047
1048                 /* Can't recv a clone into an existing fs */
1049                 if (flags & DRR_FLAG_CLONE) {
1050                         dsl_dataset_rele(ds, FTAG);
1051                         return (SET_ERROR(EINVAL));
1052                 }
1053
1054                 error = recv_begin_check_existing_impl(drba, ds, fromguid);
1055                 dsl_dataset_rele(ds, FTAG);
1056         } else if (error == ENOENT) {
1057                 /* target fs does not exist; must be a full backup or clone */
1058                 char buf[MAXNAMELEN];
1059
1060                 /*
1061                  * If it's a non-clone incremental, we are missing the
1062                  * target fs, so fail the recv.
1063                  */
1064                 if (fromguid != 0 && !(flags & DRR_FLAG_CLONE))
1065                         return (SET_ERROR(ENOENT));
1066
1067                 /* Open the parent of tofs */
1068                 ASSERT3U(strlen(tofs), <, MAXNAMELEN);
1069                 (void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1);
1070                 error = dsl_dataset_hold(dp, buf, FTAG, &ds);
1071                 if (error != 0)
1072                         return (error);
1073
1074                 /*
1075                  * Check filesystem and snapshot limits before receiving. We'll
1076                  * recheck snapshot limits again at the end (we create the
1077                  * filesystems and increment those counts during begin_sync).
1078                  */
1079                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1080                     ZFS_PROP_FILESYSTEM_LIMIT, NULL, drba->drba_cred);
1081                 if (error != 0) {
1082                         dsl_dataset_rele(ds, FTAG);
1083                         return (error);
1084                 }
1085
1086                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1087                     ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred);
1088                 if (error != 0) {
1089                         dsl_dataset_rele(ds, FTAG);
1090                         return (error);
1091                 }
1092
1093                 if (drba->drba_origin != NULL) {
1094                         dsl_dataset_t *origin;
1095                         error = dsl_dataset_hold(dp, drba->drba_origin,
1096                             FTAG, &origin);
1097                         if (error != 0) {
1098                                 dsl_dataset_rele(ds, FTAG);
1099                                 return (error);
1100                         }
1101                         if (!dsl_dataset_is_snapshot(origin)) {
1102                                 dsl_dataset_rele(origin, FTAG);
1103                                 dsl_dataset_rele(ds, FTAG);
1104                                 return (SET_ERROR(EINVAL));
1105                         }
1106                         if (origin->ds_phys->ds_guid != fromguid) {
1107                                 dsl_dataset_rele(origin, FTAG);
1108                                 dsl_dataset_rele(ds, FTAG);
1109                                 return (SET_ERROR(ENODEV));
1110                         }
1111                         dsl_dataset_rele(origin, FTAG);
1112                 }
1113                 dsl_dataset_rele(ds, FTAG);
1114                 error = 0;
1115         }
1116         return (error);
1117 }
1118
1119 static void
1120 dmu_recv_begin_sync(void *arg, dmu_tx_t *tx)
1121 {
1122         dmu_recv_begin_arg_t *drba = arg;
1123         dsl_pool_t *dp = dmu_tx_pool(tx);
1124         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1125         const char *tofs = drba->drba_cookie->drc_tofs;
1126         dsl_dataset_t *ds, *newds;
1127         uint64_t dsobj;
1128         int error;
1129         uint64_t crflags;
1130
1131         crflags = (drrb->drr_flags & DRR_FLAG_CI_DATA) ?
1132             DS_FLAG_CI_DATASET : 0;
1133
1134         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1135         if (error == 0) {
1136                 /* create temporary clone */
1137                 dsl_dataset_t *snap = NULL;
1138                 if (drba->drba_snapobj != 0) {
1139                         VERIFY0(dsl_dataset_hold_obj(dp,
1140                             drba->drba_snapobj, FTAG, &snap));
1141                 }
1142                 dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name,
1143                     snap, crflags, drba->drba_cred, tx);
1144                 dsl_dataset_rele(snap, FTAG);
1145                 dsl_dataset_rele(ds, FTAG);
1146         } else {
1147                 dsl_dir_t *dd;
1148                 const char *tail;
1149                 dsl_dataset_t *origin = NULL;
1150
1151                 VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail));
1152
1153                 if (drba->drba_origin != NULL) {
1154                         VERIFY0(dsl_dataset_hold(dp, drba->drba_origin,
1155                             FTAG, &origin));
1156                 }
1157
1158                 /* Create new dataset. */
1159                 dsobj = dsl_dataset_create_sync(dd,
1160                     strrchr(tofs, '/') + 1,
1161                     origin, crflags, drba->drba_cred, tx);
1162                 if (origin != NULL)
1163                         dsl_dataset_rele(origin, FTAG);
1164                 dsl_dir_rele(dd, FTAG);
1165                 drba->drba_cookie->drc_newfs = B_TRUE;
1166         }
1167         VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds));
1168
1169         if ((DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
1170             DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
1171             !newds->ds_large_blocks) {
1172                 dsl_dataset_activate_large_blocks_sync_impl(dsobj, tx);
1173                 newds->ds_large_blocks = B_TRUE;
1174         }
1175
1176         dmu_buf_will_dirty(newds->ds_dbuf, tx);
1177         newds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
1178
1179         /*
1180          * If we actually created a non-clone, we need to create the
1181          * objset in our new dataset.
1182          */
1183         if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) {
1184                 (void) dmu_objset_create_impl(dp->dp_spa,
1185                     newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx);
1186         }
1187
1188         drba->drba_cookie->drc_ds = newds;
1189
1190         spa_history_log_internal_ds(newds, "receive", tx, "");
1191 }
1192
1193 /*
1194  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
1195  * succeeds; otherwise we will leak the holds on the datasets.
1196  */
1197 int
1198 dmu_recv_begin(char *tofs, char *tosnap, struct drr_begin *drrb,
1199     boolean_t force, char *origin, dmu_recv_cookie_t *drc)
1200 {
1201         dmu_recv_begin_arg_t drba = { 0 };
1202         dmu_replay_record_t *drr;
1203
1204         bzero(drc, sizeof (dmu_recv_cookie_t));
1205         drc->drc_drrb = drrb;
1206         drc->drc_tosnap = tosnap;
1207         drc->drc_tofs = tofs;
1208         drc->drc_force = force;
1209         drc->drc_cred = CRED();
1210
1211         if (drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
1212                 drc->drc_byteswap = B_TRUE;
1213         else if (drrb->drr_magic != DMU_BACKUP_MAGIC)
1214                 return (SET_ERROR(EINVAL));
1215
1216         drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
1217         drr->drr_type = DRR_BEGIN;
1218         drr->drr_u.drr_begin = *drc->drc_drrb;
1219         if (drc->drc_byteswap) {
1220                 fletcher_4_incremental_byteswap(drr,
1221                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1222         } else {
1223                 fletcher_4_incremental_native(drr,
1224                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1225         }
1226         kmem_free(drr, sizeof (dmu_replay_record_t));
1227
1228         if (drc->drc_byteswap) {
1229                 drrb->drr_magic = BSWAP_64(drrb->drr_magic);
1230                 drrb->drr_versioninfo = BSWAP_64(drrb->drr_versioninfo);
1231                 drrb->drr_creation_time = BSWAP_64(drrb->drr_creation_time);
1232                 drrb->drr_type = BSWAP_32(drrb->drr_type);
1233                 drrb->drr_toguid = BSWAP_64(drrb->drr_toguid);
1234                 drrb->drr_fromguid = BSWAP_64(drrb->drr_fromguid);
1235         }
1236
1237         drba.drba_origin = origin;
1238         drba.drba_cookie = drc;
1239         drba.drba_cred = CRED();
1240
1241         return (dsl_sync_task(tofs, dmu_recv_begin_check, dmu_recv_begin_sync,
1242             &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1243 }
1244
1245 struct restorearg {
1246         int err;
1247         boolean_t byteswap;
1248         kthread_t *td;
1249         struct file *fp;
1250         char *buf;
1251         uint64_t voff;
1252         int bufsize; /* amount of memory allocated for buf */
1253         zio_cksum_t cksum;
1254         avl_tree_t *guid_to_ds_map;
1255 };
1256
1257 typedef struct guid_map_entry {
1258         uint64_t        guid;
1259         dsl_dataset_t   *gme_ds;
1260         avl_node_t      avlnode;
1261 } guid_map_entry_t;
1262
1263 static int
1264 guid_compare(const void *arg1, const void *arg2)
1265 {
1266         const guid_map_entry_t *gmep1 = arg1;
1267         const guid_map_entry_t *gmep2 = arg2;
1268
1269         if (gmep1->guid < gmep2->guid)
1270                 return (-1);
1271         else if (gmep1->guid > gmep2->guid)
1272                 return (1);
1273         return (0);
1274 }
1275
1276 static void
1277 free_guid_map_onexit(void *arg)
1278 {
1279         avl_tree_t *ca = arg;
1280         void *cookie = NULL;
1281         guid_map_entry_t *gmep;
1282
1283         while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
1284                 dsl_dataset_long_rele(gmep->gme_ds, gmep);
1285                 dsl_dataset_rele(gmep->gme_ds, gmep);
1286                 kmem_free(gmep, sizeof (guid_map_entry_t));
1287         }
1288         avl_destroy(ca);
1289         kmem_free(ca, sizeof (avl_tree_t));
1290 }
1291
1292 static int
1293 restore_bytes(struct restorearg *ra, void *buf, int len, off_t off, ssize_t *resid)
1294 {
1295         struct uio auio;
1296         struct iovec aiov;
1297         int error;
1298
1299         aiov.iov_base = buf;
1300         aiov.iov_len = len;
1301         auio.uio_iov = &aiov;
1302         auio.uio_iovcnt = 1;
1303         auio.uio_resid = len;
1304         auio.uio_segflg = UIO_SYSSPACE;
1305         auio.uio_rw = UIO_READ;
1306         auio.uio_offset = off;
1307         auio.uio_td = ra->td;
1308 #ifdef _KERNEL
1309         error = fo_read(ra->fp, &auio, ra->td->td_ucred, FOF_OFFSET, ra->td);
1310 #else
1311         fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__);
1312         error = EOPNOTSUPP;
1313 #endif
1314         *resid = auio.uio_resid;
1315         return (error);
1316 }
1317
1318 static void *
1319 restore_read(struct restorearg *ra, int len, char *buf)
1320 {
1321         int done = 0;
1322
1323         if (buf == NULL)
1324                 buf = ra->buf;
1325
1326         /* some things will require 8-byte alignment, so everything must */
1327         ASSERT0(len % 8);
1328         ASSERT3U(len, <=, ra->bufsize);
1329
1330         while (done < len) {
1331                 ssize_t resid;
1332
1333                 ra->err = restore_bytes(ra, buf + done,
1334                     len - done, ra->voff, &resid);
1335
1336                 if (resid == len - done)
1337                         ra->err = SET_ERROR(EINVAL);
1338                 ra->voff += len - done - resid;
1339                 done = len - resid;
1340                 if (ra->err != 0)
1341                         return (NULL);
1342         }
1343
1344         ASSERT3U(done, ==, len);
1345         if (ra->byteswap)
1346                 fletcher_4_incremental_byteswap(buf, len, &ra->cksum);
1347         else
1348                 fletcher_4_incremental_native(buf, len, &ra->cksum);
1349         return (buf);
1350 }
1351
1352 static void
1353 backup_byteswap(dmu_replay_record_t *drr)
1354 {
1355 #define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
1356 #define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
1357         drr->drr_type = BSWAP_32(drr->drr_type);
1358         drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
1359         switch (drr->drr_type) {
1360         case DRR_BEGIN:
1361                 DO64(drr_begin.drr_magic);
1362                 DO64(drr_begin.drr_versioninfo);
1363                 DO64(drr_begin.drr_creation_time);
1364                 DO32(drr_begin.drr_type);
1365                 DO32(drr_begin.drr_flags);
1366                 DO64(drr_begin.drr_toguid);
1367                 DO64(drr_begin.drr_fromguid);
1368                 break;
1369         case DRR_OBJECT:
1370                 DO64(drr_object.drr_object);
1371                 DO32(drr_object.drr_type);
1372                 DO32(drr_object.drr_bonustype);
1373                 DO32(drr_object.drr_blksz);
1374                 DO32(drr_object.drr_bonuslen);
1375                 DO64(drr_object.drr_toguid);
1376                 break;
1377         case DRR_FREEOBJECTS:
1378                 DO64(drr_freeobjects.drr_firstobj);
1379                 DO64(drr_freeobjects.drr_numobjs);
1380                 DO64(drr_freeobjects.drr_toguid);
1381                 break;
1382         case DRR_WRITE:
1383                 DO64(drr_write.drr_object);
1384                 DO32(drr_write.drr_type);
1385                 DO64(drr_write.drr_offset);
1386                 DO64(drr_write.drr_length);
1387                 DO64(drr_write.drr_toguid);
1388                 DO64(drr_write.drr_key.ddk_cksum.zc_word[0]);
1389                 DO64(drr_write.drr_key.ddk_cksum.zc_word[1]);
1390                 DO64(drr_write.drr_key.ddk_cksum.zc_word[2]);
1391                 DO64(drr_write.drr_key.ddk_cksum.zc_word[3]);
1392                 DO64(drr_write.drr_key.ddk_prop);
1393                 break;
1394         case DRR_WRITE_BYREF:
1395                 DO64(drr_write_byref.drr_object);
1396                 DO64(drr_write_byref.drr_offset);
1397                 DO64(drr_write_byref.drr_length);
1398                 DO64(drr_write_byref.drr_toguid);
1399                 DO64(drr_write_byref.drr_refguid);
1400                 DO64(drr_write_byref.drr_refobject);
1401                 DO64(drr_write_byref.drr_refoffset);
1402                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[0]);
1403                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[1]);
1404                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[2]);
1405                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[3]);
1406                 DO64(drr_write_byref.drr_key.ddk_prop);
1407                 break;
1408         case DRR_WRITE_EMBEDDED:
1409                 DO64(drr_write_embedded.drr_object);
1410                 DO64(drr_write_embedded.drr_offset);
1411                 DO64(drr_write_embedded.drr_length);
1412                 DO64(drr_write_embedded.drr_toguid);
1413                 DO32(drr_write_embedded.drr_lsize);
1414                 DO32(drr_write_embedded.drr_psize);
1415                 break;
1416         case DRR_FREE:
1417                 DO64(drr_free.drr_object);
1418                 DO64(drr_free.drr_offset);
1419                 DO64(drr_free.drr_length);
1420                 DO64(drr_free.drr_toguid);
1421                 break;
1422         case DRR_SPILL:
1423                 DO64(drr_spill.drr_object);
1424                 DO64(drr_spill.drr_length);
1425                 DO64(drr_spill.drr_toguid);
1426                 break;
1427         case DRR_END:
1428                 DO64(drr_end.drr_checksum.zc_word[0]);
1429                 DO64(drr_end.drr_checksum.zc_word[1]);
1430                 DO64(drr_end.drr_checksum.zc_word[2]);
1431                 DO64(drr_end.drr_checksum.zc_word[3]);
1432                 DO64(drr_end.drr_toguid);
1433                 break;
1434         }
1435 #undef DO64
1436 #undef DO32
1437 }
1438
1439 static inline uint8_t
1440 deduce_nblkptr(dmu_object_type_t bonus_type, uint64_t bonus_size)
1441 {
1442         if (bonus_type == DMU_OT_SA) {
1443                 return (1);
1444         } else {
1445                 return (1 +
1446                     ((DN_MAX_BONUSLEN - bonus_size) >> SPA_BLKPTRSHIFT));
1447         }
1448 }
1449
1450 static int
1451 restore_object(struct restorearg *ra, objset_t *os, struct drr_object *drro)
1452 {
1453         dmu_object_info_t doi;
1454         dmu_tx_t *tx;
1455         void *data = NULL;
1456         uint64_t object;
1457         int err;
1458
1459         if (drro->drr_type == DMU_OT_NONE ||
1460             !DMU_OT_IS_VALID(drro->drr_type) ||
1461             !DMU_OT_IS_VALID(drro->drr_bonustype) ||
1462             drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
1463             drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
1464             P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
1465             drro->drr_blksz < SPA_MINBLOCKSIZE ||
1466             drro->drr_blksz > spa_maxblocksize(dmu_objset_spa(os)) ||
1467             drro->drr_bonuslen > DN_MAX_BONUSLEN) {
1468                 return (SET_ERROR(EINVAL));
1469         }
1470
1471         err = dmu_object_info(os, drro->drr_object, &doi);
1472
1473         if (err != 0 && err != ENOENT)
1474                 return (SET_ERROR(EINVAL));
1475         object = err == 0 ? drro->drr_object : DMU_NEW_OBJECT;
1476
1477         if (drro->drr_bonuslen) {
1478                 data = restore_read(ra, P2ROUNDUP(drro->drr_bonuslen, 8), NULL);
1479                 if (ra->err != 0)
1480                         return (ra->err);
1481         }
1482
1483         /*
1484          * If we are losing blkptrs or changing the block size this must
1485          * be a new file instance.  We must clear out the previous file
1486          * contents before we can change this type of metadata in the dnode.
1487          */
1488         if (err == 0) {
1489                 int nblkptr;
1490
1491                 nblkptr = deduce_nblkptr(drro->drr_bonustype,
1492                     drro->drr_bonuslen);
1493
1494                 if (drro->drr_blksz != doi.doi_data_block_size ||
1495                     nblkptr < doi.doi_nblkptr) {
1496                         err = dmu_free_long_range(os, drro->drr_object,
1497                             0, DMU_OBJECT_END);
1498                         if (err != 0)
1499                                 return (SET_ERROR(EINVAL));
1500                 }
1501         }
1502
1503         tx = dmu_tx_create(os);
1504         dmu_tx_hold_bonus(tx, object);
1505         err = dmu_tx_assign(tx, TXG_WAIT);
1506         if (err != 0) {
1507                 dmu_tx_abort(tx);
1508                 return (err);
1509         }
1510
1511         if (object == DMU_NEW_OBJECT) {
1512                 /* currently free, want to be allocated */
1513                 err = dmu_object_claim(os, drro->drr_object,
1514                     drro->drr_type, drro->drr_blksz,
1515                     drro->drr_bonustype, drro->drr_bonuslen, tx);
1516         } else if (drro->drr_type != doi.doi_type ||
1517             drro->drr_blksz != doi.doi_data_block_size ||
1518             drro->drr_bonustype != doi.doi_bonus_type ||
1519             drro->drr_bonuslen != doi.doi_bonus_size) {
1520                 /* currently allocated, but with different properties */
1521                 err = dmu_object_reclaim(os, drro->drr_object,
1522                     drro->drr_type, drro->drr_blksz,
1523                     drro->drr_bonustype, drro->drr_bonuslen, tx);
1524         }
1525         if (err != 0) {
1526                 dmu_tx_commit(tx);
1527                 return (SET_ERROR(EINVAL));
1528         }
1529
1530         dmu_object_set_checksum(os, drro->drr_object, drro->drr_checksumtype,
1531             tx);
1532         dmu_object_set_compress(os, drro->drr_object, drro->drr_compress, tx);
1533
1534         if (data != NULL) {
1535                 dmu_buf_t *db;
1536
1537                 VERIFY(0 == dmu_bonus_hold(os, drro->drr_object, FTAG, &db));
1538                 dmu_buf_will_dirty(db, tx);
1539
1540                 ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
1541                 bcopy(data, db->db_data, drro->drr_bonuslen);
1542                 if (ra->byteswap) {
1543                         dmu_object_byteswap_t byteswap =
1544                             DMU_OT_BYTESWAP(drro->drr_bonustype);
1545                         dmu_ot_byteswap[byteswap].ob_func(db->db_data,
1546                             drro->drr_bonuslen);
1547                 }
1548                 dmu_buf_rele(db, FTAG);
1549         }
1550         dmu_tx_commit(tx);
1551         return (0);
1552 }
1553
1554 /* ARGSUSED */
1555 static int
1556 restore_freeobjects(struct restorearg *ra, objset_t *os,
1557     struct drr_freeobjects *drrfo)
1558 {
1559         uint64_t obj;
1560
1561         if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
1562                 return (SET_ERROR(EINVAL));
1563
1564         for (obj = drrfo->drr_firstobj;
1565             obj < drrfo->drr_firstobj + drrfo->drr_numobjs;
1566             (void) dmu_object_next(os, &obj, FALSE, 0)) {
1567                 int err;
1568
1569                 if (dmu_object_info(os, obj, NULL) != 0)
1570                         continue;
1571
1572                 err = dmu_free_long_object(os, obj);
1573                 if (err != 0)
1574                         return (err);
1575         }
1576         return (0);
1577 }
1578
1579 static int
1580 restore_write(struct restorearg *ra, objset_t *os,
1581     struct drr_write *drrw)
1582 {
1583         dmu_tx_t *tx;
1584         void *data;
1585         int err;
1586
1587         if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
1588             !DMU_OT_IS_VALID(drrw->drr_type))
1589                 return (SET_ERROR(EINVAL));
1590
1591         if (dmu_object_info(os, drrw->drr_object, NULL) != 0)
1592                 return (SET_ERROR(EINVAL));
1593
1594         dmu_buf_t *bonus;
1595         if (dmu_bonus_hold(os, drrw->drr_object, FTAG, &bonus) != 0)
1596                 return (SET_ERROR(EINVAL));
1597
1598         arc_buf_t *abuf = dmu_request_arcbuf(bonus, drrw->drr_length);
1599
1600         data = restore_read(ra, drrw->drr_length, abuf->b_data);
1601         if (data == NULL) {
1602                 dmu_return_arcbuf(abuf);
1603                 dmu_buf_rele(bonus, FTAG);
1604                 return (ra->err);
1605         }
1606
1607         tx = dmu_tx_create(os);
1608
1609         dmu_tx_hold_write(tx, drrw->drr_object,
1610             drrw->drr_offset, drrw->drr_length);
1611         err = dmu_tx_assign(tx, TXG_WAIT);
1612         if (err != 0) {
1613                 dmu_return_arcbuf(abuf);
1614                 dmu_buf_rele(bonus, FTAG);
1615                 dmu_tx_abort(tx);
1616                 return (err);
1617         }
1618         if (ra->byteswap) {
1619                 dmu_object_byteswap_t byteswap =
1620                     DMU_OT_BYTESWAP(drrw->drr_type);
1621                 dmu_ot_byteswap[byteswap].ob_func(data, drrw->drr_length);
1622         }
1623         dmu_assign_arcbuf(bonus, drrw->drr_offset, abuf, tx);
1624         dmu_tx_commit(tx);
1625         dmu_buf_rele(bonus, FTAG);
1626         return (0);
1627 }
1628
1629 /*
1630  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
1631  * streams to refer to a copy of the data that is already on the
1632  * system because it came in earlier in the stream.  This function
1633  * finds the earlier copy of the data, and uses that copy instead of
1634  * data from the stream to fulfill this write.
1635  */
1636 static int
1637 restore_write_byref(struct restorearg *ra, objset_t *os,
1638     struct drr_write_byref *drrwbr)
1639 {
1640         dmu_tx_t *tx;
1641         int err;
1642         guid_map_entry_t gmesrch;
1643         guid_map_entry_t *gmep;
1644         avl_index_t where;
1645         objset_t *ref_os = NULL;
1646         dmu_buf_t *dbp;
1647
1648         if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
1649                 return (SET_ERROR(EINVAL));
1650
1651         /*
1652          * If the GUID of the referenced dataset is different from the
1653          * GUID of the target dataset, find the referenced dataset.
1654          */
1655         if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
1656                 gmesrch.guid = drrwbr->drr_refguid;
1657                 if ((gmep = avl_find(ra->guid_to_ds_map, &gmesrch,
1658                     &where)) == NULL) {
1659                         return (SET_ERROR(EINVAL));
1660                 }
1661                 if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
1662                         return (SET_ERROR(EINVAL));
1663         } else {
1664                 ref_os = os;
1665         }
1666
1667         err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
1668             drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH);
1669         if (err != 0)
1670                 return (err);
1671
1672         tx = dmu_tx_create(os);
1673
1674         dmu_tx_hold_write(tx, drrwbr->drr_object,
1675             drrwbr->drr_offset, drrwbr->drr_length);
1676         err = dmu_tx_assign(tx, TXG_WAIT);
1677         if (err != 0) {
1678                 dmu_tx_abort(tx);
1679                 return (err);
1680         }
1681         dmu_write(os, drrwbr->drr_object,
1682             drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
1683         dmu_buf_rele(dbp, FTAG);
1684         dmu_tx_commit(tx);
1685         return (0);
1686 }
1687
1688 static int
1689 restore_write_embedded(struct restorearg *ra, objset_t *os,
1690     struct drr_write_embedded *drrwnp)
1691 {
1692         dmu_tx_t *tx;
1693         int err;
1694         void *data;
1695
1696         if (drrwnp->drr_offset + drrwnp->drr_length < drrwnp->drr_offset)
1697                 return (EINVAL);
1698
1699         if (drrwnp->drr_psize > BPE_PAYLOAD_SIZE)
1700                 return (EINVAL);
1701
1702         if (drrwnp->drr_etype >= NUM_BP_EMBEDDED_TYPES)
1703                 return (EINVAL);
1704         if (drrwnp->drr_compression >= ZIO_COMPRESS_FUNCTIONS)
1705                 return (EINVAL);
1706
1707         data = restore_read(ra, P2ROUNDUP(drrwnp->drr_psize, 8), NULL);
1708         if (data == NULL)
1709                 return (ra->err);
1710
1711         tx = dmu_tx_create(os);
1712
1713         dmu_tx_hold_write(tx, drrwnp->drr_object,
1714             drrwnp->drr_offset, drrwnp->drr_length);
1715         err = dmu_tx_assign(tx, TXG_WAIT);
1716         if (err != 0) {
1717                 dmu_tx_abort(tx);
1718                 return (err);
1719         }
1720
1721         dmu_write_embedded(os, drrwnp->drr_object,
1722             drrwnp->drr_offset, data, drrwnp->drr_etype,
1723             drrwnp->drr_compression, drrwnp->drr_lsize, drrwnp->drr_psize,
1724             ra->byteswap ^ ZFS_HOST_BYTEORDER, tx);
1725
1726         dmu_tx_commit(tx);
1727         return (0);
1728 }
1729
1730 static int
1731 restore_spill(struct restorearg *ra, objset_t *os, struct drr_spill *drrs)
1732 {
1733         dmu_tx_t *tx;
1734         void *data;
1735         dmu_buf_t *db, *db_spill;
1736         int err;
1737
1738         if (drrs->drr_length < SPA_MINBLOCKSIZE ||
1739             drrs->drr_length > spa_maxblocksize(dmu_objset_spa(os)))
1740                 return (SET_ERROR(EINVAL));
1741
1742         data = restore_read(ra, drrs->drr_length, NULL);
1743         if (data == NULL)
1744                 return (ra->err);
1745
1746         if (dmu_object_info(os, drrs->drr_object, NULL) != 0)
1747                 return (SET_ERROR(EINVAL));
1748
1749         VERIFY(0 == dmu_bonus_hold(os, drrs->drr_object, FTAG, &db));
1750         if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
1751                 dmu_buf_rele(db, FTAG);
1752                 return (err);
1753         }
1754
1755         tx = dmu_tx_create(os);
1756
1757         dmu_tx_hold_spill(tx, db->db_object);
1758
1759         err = dmu_tx_assign(tx, TXG_WAIT);
1760         if (err != 0) {
1761                 dmu_buf_rele(db, FTAG);
1762                 dmu_buf_rele(db_spill, FTAG);
1763                 dmu_tx_abort(tx);
1764                 return (err);
1765         }
1766         dmu_buf_will_dirty(db_spill, tx);
1767
1768         if (db_spill->db_size < drrs->drr_length)
1769                 VERIFY(0 == dbuf_spill_set_blksz(db_spill,
1770                     drrs->drr_length, tx));
1771         bcopy(data, db_spill->db_data, drrs->drr_length);
1772
1773         dmu_buf_rele(db, FTAG);
1774         dmu_buf_rele(db_spill, FTAG);
1775
1776         dmu_tx_commit(tx);
1777         return (0);
1778 }
1779
1780 /* ARGSUSED */
1781 static int
1782 restore_free(struct restorearg *ra, objset_t *os,
1783     struct drr_free *drrf)
1784 {
1785         int err;
1786
1787         if (drrf->drr_length != -1ULL &&
1788             drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
1789                 return (SET_ERROR(EINVAL));
1790
1791         if (dmu_object_info(os, drrf->drr_object, NULL) != 0)
1792                 return (SET_ERROR(EINVAL));
1793
1794         err = dmu_free_long_range(os, drrf->drr_object,
1795             drrf->drr_offset, drrf->drr_length);
1796         return (err);
1797 }
1798
1799 /* used to destroy the drc_ds on error */
1800 static void
1801 dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc)
1802 {
1803         char name[MAXNAMELEN];
1804         dsl_dataset_name(drc->drc_ds, name);
1805         dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
1806         (void) dsl_destroy_head(name);
1807 }
1808
1809 /*
1810  * NB: callers *must* call dmu_recv_end() if this succeeds.
1811  */
1812 int
1813 dmu_recv_stream(dmu_recv_cookie_t *drc, struct file *fp, offset_t *voffp,
1814     int cleanup_fd, uint64_t *action_handlep)
1815 {
1816         struct restorearg ra = { 0 };
1817         dmu_replay_record_t *drr;
1818         objset_t *os;
1819         zio_cksum_t pcksum;
1820         int featureflags;
1821
1822         ra.byteswap = drc->drc_byteswap;
1823         ra.cksum = drc->drc_cksum;
1824         ra.td = curthread;
1825         ra.fp = fp;
1826         ra.voff = *voffp;
1827         ra.bufsize = SPA_MAXBLOCKSIZE;
1828         ra.buf = kmem_alloc(ra.bufsize, KM_SLEEP);
1829
1830         /* these were verified in dmu_recv_begin */
1831         ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==,
1832             DMU_SUBSTREAM);
1833         ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES);
1834
1835         /*
1836          * Open the objset we are modifying.
1837          */
1838         VERIFY0(dmu_objset_from_ds(drc->drc_ds, &os));
1839
1840         ASSERT(drc->drc_ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT);
1841
1842         featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
1843
1844         /* if this stream is dedup'ed, set up the avl tree for guid mapping */
1845         if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
1846                 minor_t minor;
1847
1848                 if (cleanup_fd == -1) {
1849                         ra.err = SET_ERROR(EBADF);
1850                         goto out;
1851                 }
1852                 ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
1853                 if (ra.err != 0) {
1854                         cleanup_fd = -1;
1855                         goto out;
1856                 }
1857
1858                 if (*action_handlep == 0) {
1859                         ra.guid_to_ds_map =
1860                             kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
1861                         avl_create(ra.guid_to_ds_map, guid_compare,
1862                             sizeof (guid_map_entry_t),
1863                             offsetof(guid_map_entry_t, avlnode));
1864                         ra.err = zfs_onexit_add_cb(minor,
1865                             free_guid_map_onexit, ra.guid_to_ds_map,
1866                             action_handlep);
1867                         if (ra.err != 0)
1868                                 goto out;
1869                 } else {
1870                         ra.err = zfs_onexit_cb_data(minor, *action_handlep,
1871                             (void **)&ra.guid_to_ds_map);
1872                         if (ra.err != 0)
1873                                 goto out;
1874                 }
1875
1876                 drc->drc_guid_to_ds_map = ra.guid_to_ds_map;
1877         }
1878
1879         /*
1880          * Read records and process them.
1881          */
1882         pcksum = ra.cksum;
1883         while (ra.err == 0 &&
1884             NULL != (drr = restore_read(&ra, sizeof (*drr), NULL))) {
1885                 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1886                         ra.err = SET_ERROR(EINTR);
1887                         goto out;
1888                 }
1889
1890                 if (ra.byteswap)
1891                         backup_byteswap(drr);
1892
1893                 switch (drr->drr_type) {
1894                 case DRR_OBJECT:
1895                 {
1896                         /*
1897                          * We need to make a copy of the record header,
1898                          * because restore_{object,write} may need to
1899                          * restore_read(), which will invalidate drr.
1900                          */
1901                         struct drr_object drro = drr->drr_u.drr_object;
1902                         ra.err = restore_object(&ra, os, &drro);
1903                         break;
1904                 }
1905                 case DRR_FREEOBJECTS:
1906                 {
1907                         struct drr_freeobjects drrfo =
1908                             drr->drr_u.drr_freeobjects;
1909                         ra.err = restore_freeobjects(&ra, os, &drrfo);
1910                         break;
1911                 }
1912                 case DRR_WRITE:
1913                 {
1914                         struct drr_write drrw = drr->drr_u.drr_write;
1915                         ra.err = restore_write(&ra, os, &drrw);
1916                         break;
1917                 }
1918                 case DRR_WRITE_BYREF:
1919                 {
1920                         struct drr_write_byref drrwbr =
1921                             drr->drr_u.drr_write_byref;
1922                         ra.err = restore_write_byref(&ra, os, &drrwbr);
1923                         break;
1924                 }
1925                 case DRR_WRITE_EMBEDDED:
1926                 {
1927                         struct drr_write_embedded drrwe =
1928                             drr->drr_u.drr_write_embedded;
1929                         ra.err = restore_write_embedded(&ra, os, &drrwe);
1930                         break;
1931                 }
1932                 case DRR_FREE:
1933                 {
1934                         struct drr_free drrf = drr->drr_u.drr_free;
1935                         ra.err = restore_free(&ra, os, &drrf);
1936                         break;
1937                 }
1938                 case DRR_END:
1939                 {
1940                         struct drr_end drre = drr->drr_u.drr_end;
1941                         /*
1942                          * We compare against the *previous* checksum
1943                          * value, because the stored checksum is of
1944                          * everything before the DRR_END record.
1945                          */
1946                         if (!ZIO_CHECKSUM_EQUAL(drre.drr_checksum, pcksum))
1947                                 ra.err = SET_ERROR(ECKSUM);
1948                         goto out;
1949                 }
1950                 case DRR_SPILL:
1951                 {
1952                         struct drr_spill drrs = drr->drr_u.drr_spill;
1953                         ra.err = restore_spill(&ra, os, &drrs);
1954                         break;
1955                 }
1956                 default:
1957                         ra.err = SET_ERROR(EINVAL);
1958                         goto out;
1959                 }
1960                 pcksum = ra.cksum;
1961         }
1962         ASSERT(ra.err != 0);
1963
1964 out:
1965         if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
1966                 zfs_onexit_fd_rele(cleanup_fd);
1967
1968         if (ra.err != 0) {
1969                 /*
1970                  * destroy what we created, so we don't leave it in the
1971                  * inconsistent restoring state.
1972                  */
1973                 dmu_recv_cleanup_ds(drc);
1974         }
1975
1976         kmem_free(ra.buf, ra.bufsize);
1977         *voffp = ra.voff;
1978         return (ra.err);
1979 }
1980
1981 static int
1982 dmu_recv_end_check(void *arg, dmu_tx_t *tx)
1983 {
1984         dmu_recv_cookie_t *drc = arg;
1985         dsl_pool_t *dp = dmu_tx_pool(tx);
1986         int error;
1987
1988         ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag);
1989
1990         if (!drc->drc_newfs) {
1991                 dsl_dataset_t *origin_head;
1992
1993                 error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head);
1994                 if (error != 0)
1995                         return (error);
1996                 if (drc->drc_force) {
1997                         /*
1998                          * We will destroy any snapshots in tofs (i.e. before
1999                          * origin_head) that are after the origin (which is
2000                          * the snap before drc_ds, because drc_ds can not
2001                          * have any snaps of its own).
2002                          */
2003                         uint64_t obj = origin_head->ds_phys->ds_prev_snap_obj;
2004                         while (obj != drc->drc_ds->ds_phys->ds_prev_snap_obj) {
2005                                 dsl_dataset_t *snap;
2006                                 error = dsl_dataset_hold_obj(dp, obj, FTAG,
2007                                     &snap);
2008                                 if (error != 0)
2009                                         return (error);
2010                                 if (snap->ds_dir != origin_head->ds_dir)
2011                                         error = SET_ERROR(EINVAL);
2012                                 if (error == 0)  {
2013                                         error = dsl_destroy_snapshot_check_impl(
2014                                             snap, B_FALSE);
2015                                 }
2016                                 obj = snap->ds_phys->ds_prev_snap_obj;
2017                                 dsl_dataset_rele(snap, FTAG);
2018                                 if (error != 0)
2019                                         return (error);
2020                         }
2021                 }
2022                 error = dsl_dataset_clone_swap_check_impl(drc->drc_ds,
2023                     origin_head, drc->drc_force, drc->drc_owner, tx);
2024                 if (error != 0) {
2025                         dsl_dataset_rele(origin_head, FTAG);
2026                         return (error);
2027                 }
2028                 error = dsl_dataset_snapshot_check_impl(origin_head,
2029                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
2030                 dsl_dataset_rele(origin_head, FTAG);
2031                 if (error != 0)
2032                         return (error);
2033
2034                 error = dsl_destroy_head_check_impl(drc->drc_ds, 1);
2035         } else {
2036                 error = dsl_dataset_snapshot_check_impl(drc->drc_ds,
2037                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
2038         }
2039         return (error);
2040 }
2041
2042 static void
2043 dmu_recv_end_sync(void *arg, dmu_tx_t *tx)
2044 {
2045         dmu_recv_cookie_t *drc = arg;
2046         dsl_pool_t *dp = dmu_tx_pool(tx);
2047
2048         spa_history_log_internal_ds(drc->drc_ds, "finish receiving",
2049             tx, "snap=%s", drc->drc_tosnap);
2050
2051         if (!drc->drc_newfs) {
2052                 dsl_dataset_t *origin_head;
2053
2054                 VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG,
2055                     &origin_head));
2056
2057                 if (drc->drc_force) {
2058                         /*
2059                          * Destroy any snapshots of drc_tofs (origin_head)
2060                          * after the origin (the snap before drc_ds).
2061                          */
2062                         uint64_t obj = origin_head->ds_phys->ds_prev_snap_obj;
2063                         while (obj != drc->drc_ds->ds_phys->ds_prev_snap_obj) {
2064                                 dsl_dataset_t *snap;
2065                                 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG,
2066                                     &snap));
2067                                 ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir);
2068                                 obj = snap->ds_phys->ds_prev_snap_obj;
2069                                 dsl_destroy_snapshot_sync_impl(snap,
2070                                     B_FALSE, tx);
2071                                 dsl_dataset_rele(snap, FTAG);
2072                         }
2073                 }
2074                 VERIFY3P(drc->drc_ds->ds_prev, ==,
2075                     origin_head->ds_prev);
2076
2077                 dsl_dataset_clone_swap_sync_impl(drc->drc_ds,
2078                     origin_head, tx);
2079                 dsl_dataset_snapshot_sync_impl(origin_head,
2080                     drc->drc_tosnap, tx);
2081
2082                 /* set snapshot's creation time and guid */
2083                 dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx);
2084                 origin_head->ds_prev->ds_phys->ds_creation_time =
2085                     drc->drc_drrb->drr_creation_time;
2086                 origin_head->ds_prev->ds_phys->ds_guid =
2087                     drc->drc_drrb->drr_toguid;
2088                 origin_head->ds_prev->ds_phys->ds_flags &=
2089                     ~DS_FLAG_INCONSISTENT;
2090
2091                 dmu_buf_will_dirty(origin_head->ds_dbuf, tx);
2092                 origin_head->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
2093
2094                 dsl_dataset_rele(origin_head, FTAG);
2095                 dsl_destroy_head_sync_impl(drc->drc_ds, tx);
2096
2097                 if (drc->drc_owner != NULL)
2098                         VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner);
2099         } else {
2100                 dsl_dataset_t *ds = drc->drc_ds;
2101
2102                 dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx);
2103
2104                 /* set snapshot's creation time and guid */
2105                 dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
2106                 ds->ds_prev->ds_phys->ds_creation_time =
2107                     drc->drc_drrb->drr_creation_time;
2108                 ds->ds_prev->ds_phys->ds_guid = drc->drc_drrb->drr_toguid;
2109                 ds->ds_prev->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
2110
2111                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
2112                 ds->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
2113         }
2114         drc->drc_newsnapobj = drc->drc_ds->ds_phys->ds_prev_snap_obj;
2115         /*
2116          * Release the hold from dmu_recv_begin.  This must be done before
2117          * we return to open context, so that when we free the dataset's dnode,
2118          * we can evict its bonus buffer.
2119          */
2120         dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2121         drc->drc_ds = NULL;
2122 }
2123
2124 static int
2125 add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj)
2126 {
2127         dsl_pool_t *dp;
2128         dsl_dataset_t *snapds;
2129         guid_map_entry_t *gmep;
2130         int err;
2131
2132         ASSERT(guid_map != NULL);
2133
2134         err = dsl_pool_hold(name, FTAG, &dp);
2135         if (err != 0)
2136                 return (err);
2137         gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP);
2138         err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds);
2139         if (err == 0) {
2140                 gmep->guid = snapds->ds_phys->ds_guid;
2141                 gmep->gme_ds = snapds;
2142                 avl_add(guid_map, gmep);
2143                 dsl_dataset_long_hold(snapds, gmep);
2144         } else
2145                 kmem_free(gmep, sizeof (*gmep));
2146
2147         dsl_pool_rele(dp, FTAG);
2148         return (err);
2149 }
2150
2151 static int dmu_recv_end_modified_blocks = 3;
2152
2153 static int
2154 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
2155 {
2156         int error;
2157         char name[MAXNAMELEN];
2158
2159 #ifdef _KERNEL
2160         /*
2161          * We will be destroying the ds; make sure its origin is unmounted if
2162          * necessary.
2163          */
2164         dsl_dataset_name(drc->drc_ds, name);
2165         zfs_destroy_unmount_origin(name);
2166 #endif
2167
2168         error = dsl_sync_task(drc->drc_tofs,
2169             dmu_recv_end_check, dmu_recv_end_sync, drc,
2170             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL);
2171
2172         if (error != 0)
2173                 dmu_recv_cleanup_ds(drc);
2174         return (error);
2175 }
2176
2177 static int
2178 dmu_recv_new_end(dmu_recv_cookie_t *drc)
2179 {
2180         int error;
2181
2182         error = dsl_sync_task(drc->drc_tofs,
2183             dmu_recv_end_check, dmu_recv_end_sync, drc,
2184             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL);
2185
2186         if (error != 0) {
2187                 dmu_recv_cleanup_ds(drc);
2188         } else if (drc->drc_guid_to_ds_map != NULL) {
2189                 (void) add_ds_to_guidmap(drc->drc_tofs,
2190                     drc->drc_guid_to_ds_map,
2191                     drc->drc_newsnapobj);
2192         }
2193         return (error);
2194 }
2195
2196 int
2197 dmu_recv_end(dmu_recv_cookie_t *drc, void *owner)
2198 {
2199         drc->drc_owner = owner;
2200
2201         if (drc->drc_newfs)
2202                 return (dmu_recv_new_end(drc));
2203         else
2204                 return (dmu_recv_existing_end(drc));
2205 }
2206
2207 /*
2208  * Return TRUE if this objset is currently being received into.
2209  */
2210 boolean_t
2211 dmu_objset_is_receiving(objset_t *os)
2212 {
2213         return (os->os_dsl_dataset != NULL &&
2214             os->os_dsl_dataset->ds_owner == dmu_recv_tag);
2215 }