module/zfs/dmu_send.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2011 by Delphix. All rights reserved.
  24  * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
  25  * Copyright (c) 2014, Joyent, Inc. All rights reserved.
  26  * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
  27  */
  28
  29 #include <sys/dmu.h>
  30 #include <sys/dmu_impl.h>
  31 #include <sys/dmu_tx.h>
  32 #include <sys/dbuf.h>
  33 #include <sys/dnode.h>
  34 #include <sys/zfs_context.h>
  35 #include <sys/dmu_objset.h>
  36 #include <sys/dmu_traverse.h>
  37 #include <sys/dsl_dataset.h>
  38 #include <sys/dsl_dir.h>
  39 #include <sys/dsl_prop.h>
  40 #include <sys/dsl_pool.h>
  41 #include <sys/dsl_synctask.h>
  42 #include <sys/spa_impl.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 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
  58 int zfs_send_corrupt_data = B_FALSE;
  59
  60 static char *dmu_recv_tag = "dmu_recv_tag";
  61 static const char *recv_clone_name = "%recv";
  62
  63 typedef struct dump_bytes_io {
  64         dmu_sendarg_t   *dbi_dsp;
  65         void            *dbi_buf;
  66         int             dbi_len;
  67 } dump_bytes_io_t;
  68
  69 static void
  70 dump_bytes_strategy(void *arg)
  71 {
  72         dump_bytes_io_t *dbi = (dump_bytes_io_t *)arg;
  73         dmu_sendarg_t *dsp = dbi->dbi_dsp;
  74         dsl_dataset_t *ds = dsp->dsa_os->os_dsl_dataset;
  75         ssize_t resid; /* have to get resid to get detailed errno */
  76         ASSERT0(dbi->dbi_len % 8);
  77
  78         fletcher_4_incremental_native(dbi->dbi_buf, dbi->dbi_len, &dsp->dsa_zc);
  79         dsp->dsa_err = vn_rdwr(UIO_WRITE, dsp->dsa_vp,
  80             (caddr_t)dbi->dbi_buf, dbi->dbi_len,
  81             0, UIO_SYSSPACE, FAPPEND, RLIM64_INFINITY, CRED(), &resid);
  82
  83         mutex_enter(&ds->ds_sendstream_lock);
  84         *dsp->dsa_off += dbi->dbi_len;
  85         mutex_exit(&ds->ds_sendstream_lock);
  86 }
  87
  88 static int
  89 dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
  90 {
  91         dump_bytes_io_t dbi;
  92
  93         dbi.dbi_dsp = dsp;
  94         dbi.dbi_buf = buf;
  95         dbi.dbi_len = len;
  96
  97         /*
  98          * The vn_rdwr() call is performed in a taskq to ensure that there is
  99          * always enough stack space to write safely to the target filesystem.
 100          * The ZIO_TYPE_FREE threads are used because there can be a lot of
 101          * them and they are used in vdev_file.c for a similar purpose.
 102          */
 103         spa_taskq_dispatch_sync(dmu_objset_spa(dsp->dsa_os), ZIO_TYPE_FREE,
 104             ZIO_TASKQ_ISSUE, dump_bytes_strategy, &dbi, TQ_SLEEP);
 105
 106         return (dsp->dsa_err);
 107 }
 108
 109 static int
 110 dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
 111     uint64_t length)
 112 {
 113         struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free);
 114
 115         /*
 116          * When we receive a free record, dbuf_free_range() assumes
 117          * that the receiving system doesn't have any dbufs in the range
 118          * being freed.  This is always true because there is a one-record
 119          * constraint: we only send one WRITE record for any given
 120          * object+offset.  We know that the one-record constraint is
 121          * true because we always send data in increasing order by
 122          * object,offset.
 123          *
 124          * If the increasing-order constraint ever changes, we should find
 125          * another way to assert that the one-record constraint is still
 126          * satisfied.
 127          */
 128         ASSERT(object > dsp->dsa_last_data_object ||
 129             (object == dsp->dsa_last_data_object &&
 130             offset > dsp->dsa_last_data_offset));
 131
 132         /*
 133          * If we are doing a non-incremental send, then there can't
 134          * be any data in the dataset we're receiving into.  Therefore
 135          * a free record would simply be a no-op.  Save space by not
 136          * sending it to begin with.
 137          */
 138         if (!dsp->dsa_incremental)
 139                 return (0);
 140
 141         if (length != -1ULL && offset + length < offset)
 142                 length = -1ULL;
 143
 144         /*
 145          * If there is a pending op, but it's not PENDING_FREE, push it out,
 146          * since free block aggregation can only be done for blocks of the
 147          * same type (i.e., DRR_FREE records can only be aggregated with
 148          * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
 149          * aggregated with other DRR_FREEOBJECTS records.
 150          */
 151         if (dsp->dsa_pending_op != PENDING_NONE &&
 152             dsp->dsa_pending_op != PENDING_FREE) {
 153                 if (dump_bytes(dsp, dsp->dsa_drr,
 154                     sizeof (dmu_replay_record_t)) != 0)
 155                         return (SET_ERROR(EINTR));
 156                 dsp->dsa_pending_op = PENDING_NONE;
 157         }
 158
 159         if (dsp->dsa_pending_op == PENDING_FREE) {
 160                 /*
 161                  * There should never be a PENDING_FREE if length is -1
 162                  * (because dump_dnode is the only place where this
 163                  * function is called with a -1, and only after flushing
 164                  * any pending record).
 165                  */
 166                 ASSERT(length != -1ULL);
 167                 /*
 168                  * Check to see whether this free block can be aggregated
 169                  * with pending one.
 170                  */
 171                 if (drrf->drr_object == object && drrf->drr_offset +
 172                     drrf->drr_length == offset) {
 173                         drrf->drr_length += length;
 174                         return (0);
 175                 } else {
 176                         /* not a continuation.  Push out pending record */
 177                         if (dump_bytes(dsp, dsp->dsa_drr,
 178                             sizeof (dmu_replay_record_t)) != 0)
 179                                 return (SET_ERROR(EINTR));
 180                         dsp->dsa_pending_op = PENDING_NONE;
 181                 }
 182         }
 183         /* create a FREE record and make it pending */
 184         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 185         dsp->dsa_drr->drr_type = DRR_FREE;
 186         drrf->drr_object = object;
 187         drrf->drr_offset = offset;
 188         drrf->drr_length = length;
 189         drrf->drr_toguid = dsp->dsa_toguid;
 190         if (length == -1ULL) {
 191                 if (dump_bytes(dsp, dsp->dsa_drr,
 192                     sizeof (dmu_replay_record_t)) != 0)
 193                         return (SET_ERROR(EINTR));
 194         } else {
 195                 dsp->dsa_pending_op = PENDING_FREE;
 196         }
 197
 198         return (0);
 199 }
 200
 201 static int
 202 dump_write(dmu_sendarg_t *dsp, dmu_object_type_t type,
 203     uint64_t object, uint64_t offset, int blksz, const blkptr_t *bp, void *data)
 204 {
 205         struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write);
 206
 207         /*
 208          * We send data in increasing object, offset order.
 209          * See comment in dump_free() for details.
 210          */
 211         ASSERT(object > dsp->dsa_last_data_object ||
 212             (object == dsp->dsa_last_data_object &&
 213             offset > dsp->dsa_last_data_offset));
 214         dsp->dsa_last_data_object = object;
 215         dsp->dsa_last_data_offset = offset + blksz - 1;
 216
 217         /*
 218          * If there is any kind of pending aggregation (currently either
 219          * a grouping of free objects or free blocks), push it out to
 220          * the stream, since aggregation can't be done across operations
 221          * of different types.
 222          */
 223         if (dsp->dsa_pending_op != PENDING_NONE) {
 224                 if (dump_bytes(dsp, dsp->dsa_drr,
 225                     sizeof (dmu_replay_record_t)) != 0)
 226                         return (SET_ERROR(EINTR));
 227                 dsp->dsa_pending_op = PENDING_NONE;
 228         }
 229         /* write a DATA record */
 230         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 231         dsp->dsa_drr->drr_type = DRR_WRITE;
 232         drrw->drr_object = object;
 233         drrw->drr_type = type;
 234         drrw->drr_offset = offset;
 235         drrw->drr_length = blksz;
 236         drrw->drr_toguid = dsp->dsa_toguid;
 237         if (BP_IS_EMBEDDED(bp)) {
 238                 /*
 239                  * There's no pre-computed checksum of embedded BP's, so
 240                  * (like fletcher4-checkummed blocks) userland will have
 241                  * to compute a dedup-capable checksum itself.
 242                  */
 243                 drrw->drr_checksumtype = ZIO_CHECKSUM_OFF;
 244         } else {
 245                 drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
 246                 if (zio_checksum_table[drrw->drr_checksumtype].ci_dedup)
 247                         drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
 248                 DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
 249                 DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
 250                 DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
 251                 drrw->drr_key.ddk_cksum = bp->blk_cksum;
 252         }
 253
 254         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
 255                 return (SET_ERROR(EINTR));
 256         if (dump_bytes(dsp, data, blksz) != 0)
 257                 return (SET_ERROR(EINTR));
 258         return (0);
 259 }
 260
 261 static int
 262 dump_write_embedded(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
 263     int blksz, const blkptr_t *bp)
 264 {
 265         char buf[BPE_PAYLOAD_SIZE];
 266         struct drr_write_embedded *drrw =
 267             &(dsp->dsa_drr->drr_u.drr_write_embedded);
 268
 269         if (dsp->dsa_pending_op != PENDING_NONE) {
 270                 if (dump_bytes(dsp, dsp->dsa_drr,
 271                     sizeof (dmu_replay_record_t)) != 0)
 272                         return (EINTR);
 273                 dsp->dsa_pending_op = PENDING_NONE;
 274         }
 275
 276         ASSERT(BP_IS_EMBEDDED(bp));
 277
 278         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 279         dsp->dsa_drr->drr_type = DRR_WRITE_EMBEDDED;
 280         drrw->drr_object = object;
 281         drrw->drr_offset = offset;
 282         drrw->drr_length = blksz;
 283         drrw->drr_toguid = dsp->dsa_toguid;
 284         drrw->drr_compression = BP_GET_COMPRESS(bp);
 285         drrw->drr_etype = BPE_GET_ETYPE(bp);
 286         drrw->drr_lsize = BPE_GET_LSIZE(bp);
 287         drrw->drr_psize = BPE_GET_PSIZE(bp);
 288
 289         decode_embedded_bp_compressed(bp, buf);
 290
 291         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
 292                 return (EINTR);
 293         if (dump_bytes(dsp, buf, P2ROUNDUP(drrw->drr_psize, 8)) != 0)
 294                 return (EINTR);
 295         return (0);
 296 }
 297
 298 static int
 299 dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data)
 300 {
 301         struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill);
 302
 303         if (dsp->dsa_pending_op != PENDING_NONE) {
 304                 if (dump_bytes(dsp, dsp->dsa_drr,
 305                     sizeof (dmu_replay_record_t)) != 0)
 306                         return (SET_ERROR(EINTR));
 307                 dsp->dsa_pending_op = PENDING_NONE;
 308         }
 309
 310         /* write a SPILL record */
 311         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 312         dsp->dsa_drr->drr_type = DRR_SPILL;
 313         drrs->drr_object = object;
 314         drrs->drr_length = blksz;
 315         drrs->drr_toguid = dsp->dsa_toguid;
 316
 317         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)))
 318                 return (SET_ERROR(EINTR));
 319         if (dump_bytes(dsp, data, blksz))
 320                 return (SET_ERROR(EINTR));
 321         return (0);
 322 }
 323
 324 static int
 325 dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
 326 {
 327         struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
 328
 329         /* See comment in dump_free(). */
 330         if (!dsp->dsa_incremental)
 331                 return (0);
 332
 333         /*
 334          * If there is a pending op, but it's not PENDING_FREEOBJECTS,
 335          * push it out, since free block aggregation can only be done for
 336          * blocks of the same type (i.e., DRR_FREE records can only be
 337          * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
 338          * can only be aggregated with other DRR_FREEOBJECTS records.
 339          */
 340         if (dsp->dsa_pending_op != PENDING_NONE &&
 341             dsp->dsa_pending_op != PENDING_FREEOBJECTS) {
 342                 if (dump_bytes(dsp, dsp->dsa_drr,
 343                     sizeof (dmu_replay_record_t)) != 0)
 344                         return (SET_ERROR(EINTR));
 345                 dsp->dsa_pending_op = PENDING_NONE;
 346         }
 347         if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) {
 348                 /*
 349                  * See whether this free object array can be aggregated
 350                  * with pending one
 351                  */
 352                 if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
 353                         drrfo->drr_numobjs += numobjs;
 354                         return (0);
 355                 } else {
 356                         /* can't be aggregated.  Push out pending record */
 357                         if (dump_bytes(dsp, dsp->dsa_drr,
 358                             sizeof (dmu_replay_record_t)) != 0)
 359                                 return (SET_ERROR(EINTR));
 360                         dsp->dsa_pending_op = PENDING_NONE;
 361                 }
 362         }
 363
 364         /* write a FREEOBJECTS record */
 365         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 366         dsp->dsa_drr->drr_type = DRR_FREEOBJECTS;
 367         drrfo->drr_firstobj = firstobj;
 368         drrfo->drr_numobjs = numobjs;
 369         drrfo->drr_toguid = dsp->dsa_toguid;
 370
 371         dsp->dsa_pending_op = PENDING_FREEOBJECTS;
 372
 373         return (0);
 374 }
 375
 376 static int
 377 dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
 378 {
 379         struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object);
 380
 381         if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
 382                 return (dump_freeobjects(dsp, object, 1));
 383
 384         if (dsp->dsa_pending_op != PENDING_NONE) {
 385                 if (dump_bytes(dsp, dsp->dsa_drr,
 386                     sizeof (dmu_replay_record_t)) != 0)
 387                         return (SET_ERROR(EINTR));
 388                 dsp->dsa_pending_op = PENDING_NONE;
 389         }
 390
 391         /* write an OBJECT record */
 392         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 393         dsp->dsa_drr->drr_type = DRR_OBJECT;
 394         drro->drr_object = object;
 395         drro->drr_type = dnp->dn_type;
 396         drro->drr_bonustype = dnp->dn_bonustype;
 397         drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 398         drro->drr_bonuslen = dnp->dn_bonuslen;
 399         drro->drr_checksumtype = dnp->dn_checksum;
 400         drro->drr_compress = dnp->dn_compress;
 401         drro->drr_toguid = dsp->dsa_toguid;
 402
 403         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
 404                 return (SET_ERROR(EINTR));
 405
 406         if (dump_bytes(dsp, DN_BONUS(dnp), P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0)
 407                 return (SET_ERROR(EINTR));
 408
 409         /* Free anything past the end of the file. */
 410         if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) *
 411             (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0)
 412                 return (SET_ERROR(EINTR));
 413         if (dsp->dsa_err != 0)
 414                 return (SET_ERROR(EINTR));
 415         return (0);
 416 }
 417
 418 static boolean_t
 419 backup_do_embed(dmu_sendarg_t *dsp, const blkptr_t *bp)
 420 {
 421         if (!BP_IS_EMBEDDED(bp))
 422                 return (B_FALSE);
 423
 424         /*
 425          * Compression function must be legacy, or explicitly enabled.
 426          */
 427         if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS &&
 428             !(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4)))
 429                 return (B_FALSE);
 430
 431         /*
 432          * Embed type must be explicitly enabled.
 433          */
 434         switch (BPE_GET_ETYPE(bp)) {
 435         case BP_EMBEDDED_TYPE_DATA:
 436                 if (dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)
 437                         return (B_TRUE);
 438                 break;
 439         default:
 440                 return (B_FALSE);
 441         }
 442         return (B_FALSE);
 443 }
 444
 445 #define BP_SPAN(dnp, level) \
 446         (((uint64_t)dnp->dn_datablkszsec) << (SPA_MINBLOCKSHIFT + \
 447         (level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)))
 448
 449 /* ARGSUSED */
 450 static int
 451 backup_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 452     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
 453 {
 454         dmu_sendarg_t *dsp = arg;
 455         dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
 456         int err = 0;
 457
 458         if (issig(JUSTLOOKING) && issig(FORREAL))
 459                 return (SET_ERROR(EINTR));
 460
 461         if (zb->zb_object != DMU_META_DNODE_OBJECT &&
 462             DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
 463                 return (0);
 464         } else if (zb->zb_level == ZB_ZIL_LEVEL) {
 465                 /*
 466                  * If we are sending a non-snapshot (which is allowed on
 467                  * read-only pools), it may have a ZIL, which must be ignored.
 468                  */
 469                 return (0);
 470         } else if (BP_IS_HOLE(bp) &&
 471             zb->zb_object == DMU_META_DNODE_OBJECT) {
 472                 uint64_t span = BP_SPAN(dnp, zb->zb_level);
 473                 uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
 474                 err = dump_freeobjects(dsp, dnobj, span >> DNODE_SHIFT);
 475         } else if (BP_IS_HOLE(bp)) {
 476                 uint64_t span = BP_SPAN(dnp, zb->zb_level);
 477                 err = dump_free(dsp, zb->zb_object, zb->zb_blkid * span, span);
 478         } else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
 479                 return (0);
 480         } else if (type == DMU_OT_DNODE) {
 481                 dnode_phys_t *blk;
 482                 int i;
 483                 int blksz = BP_GET_LSIZE(bp);
 484                 uint32_t aflags = ARC_WAIT;
 485                 arc_buf_t *abuf;
 486
 487                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 488                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 489                     &aflags, zb) != 0)
 490                         return (SET_ERROR(EIO));
 491
 492                 blk = abuf->b_data;
 493                 for (i = 0; i < blksz >> DNODE_SHIFT; i++) {
 494                         uint64_t dnobj = (zb->zb_blkid <<
 495                             (DNODE_BLOCK_SHIFT - DNODE_SHIFT)) + i;
 496                         err = dump_dnode(dsp, dnobj, blk+i);
 497                         if (err != 0)
 498                                 break;
 499                 }
 500                 (void) arc_buf_remove_ref(abuf, &abuf);
 501         } else if (type == DMU_OT_SA) {
 502                 uint32_t aflags = ARC_WAIT;
 503                 arc_buf_t *abuf;
 504                 int blksz = BP_GET_LSIZE(bp);
 505
 506                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 507                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 508                     &aflags, zb) != 0)
 509                         return (SET_ERROR(EIO));
 510
 511                 err = dump_spill(dsp, zb->zb_object, blksz, abuf->b_data);
 512                 (void) arc_buf_remove_ref(abuf, &abuf);
 513         } else if (backup_do_embed(dsp, bp)) {
 514                 /* it's an embedded level-0 block of a regular object */
 515                 int blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 516                 err = dump_write_embedded(dsp, zb->zb_object,
 517                     zb->zb_blkid * blksz, blksz, bp);
 518         } else { /* it's a level-0 block of a regular object */
 519                 uint32_t aflags = ARC_WAIT;
 520                 arc_buf_t *abuf;
 521                 int blksz = BP_GET_LSIZE(bp);
 522
 523                 ASSERT3U(blksz, ==, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 524                 ASSERT0(zb->zb_level);
 525                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 526                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 527                     &aflags, zb) != 0) {
 528                         if (zfs_send_corrupt_data) {
 529                                 uint64_t *ptr;
 530                                 /* Send a block filled with 0x"zfs badd bloc" */
 531                                 abuf = arc_buf_alloc(spa, blksz, &abuf,
 532                                     ARC_BUFC_DATA);
 533                                 for (ptr = abuf->b_data;
 534                                     (char *)ptr < (char *)abuf->b_data + blksz;
 535                                     ptr++)
 536                                         *ptr = 0x2f5baddb10cULL;
 537                         } else {
 538                                 return (SET_ERROR(EIO));
 539                         }
 540                 }
 541
 542                 err = dump_write(dsp, type, zb->zb_object, zb->zb_blkid * blksz,
 543                     blksz, bp, abuf->b_data);
 544                 (void) arc_buf_remove_ref(abuf, &abuf);
 545         }
 546
 547         ASSERT(err == 0 || err == EINTR);
 548         return (err);
 549 }
 550
 551 /*
 552  * Releases dp using the specified tag.
 553  */
 554 static int
 555 dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *ds,
 556     zfs_bookmark_phys_t *fromzb, boolean_t is_clone, boolean_t embedok,
 557     int outfd, vnode_t *vp, offset_t *off)
 558 {
 559         objset_t *os;
 560         dmu_replay_record_t *drr;
 561         dmu_sendarg_t *dsp;
 562         int err;
 563         uint64_t fromtxg = 0;
 564         uint64_t featureflags = 0;
 565
 566         err = dmu_objset_from_ds(ds, &os);
 567         if (err != 0) {
 568                 dsl_pool_rele(dp, tag);
 569                 return (err);
 570         }
 571
 572         drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
 573         drr->drr_type = DRR_BEGIN;
 574         drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
 575         DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
 576             DMU_SUBSTREAM);
 577
 578 #ifdef _KERNEL
 579         if (dmu_objset_type(os) == DMU_OST_ZFS) {
 580                 uint64_t version;
 581                 if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0) {
 582                         kmem_free(drr, sizeof (dmu_replay_record_t));
 583                         dsl_pool_rele(dp, tag);
 584                         return (SET_ERROR(EINVAL));
 585                 }
 586                 if (version >= ZPL_VERSION_SA) {
 587                         featureflags |= DMU_BACKUP_FEATURE_SA_SPILL;
 588                 }
 589         }
 590 #endif
 591
 592         if (embedok &&
 593             spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) {
 594                 featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA;
 595                 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
 596                         featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA_LZ4;
 597         } else {
 598                 embedok = B_FALSE;
 599         }
 600
 601         DMU_SET_FEATUREFLAGS(drr->drr_u.drr_begin.drr_versioninfo,
 602             featureflags);
 603
 604         drr->drr_u.drr_begin.drr_creation_time =
 605             dsl_dataset_phys(ds)->ds_creation_time;
 606         drr->drr_u.drr_begin.drr_type = dmu_objset_type(os);
 607         if (is_clone)
 608                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
 609         drr->drr_u.drr_begin.drr_toguid = dsl_dataset_phys(ds)->ds_guid;
 610         if (dsl_dataset_phys(ds)->ds_flags & DS_FLAG_CI_DATASET)
 611                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
 612
 613         if (fromzb != NULL) {
 614                 drr->drr_u.drr_begin.drr_fromguid = fromzb->zbm_guid;
 615                 fromtxg = fromzb->zbm_creation_txg;
 616         }
 617         dsl_dataset_name(ds, drr->drr_u.drr_begin.drr_toname);
 618         if (!ds->ds_is_snapshot) {
 619                 (void) strlcat(drr->drr_u.drr_begin.drr_toname, "@--head--",
 620                     sizeof (drr->drr_u.drr_begin.drr_toname));
 621         }
 622
 623         dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP);
 624
 625         dsp->dsa_drr = drr;
 626         dsp->dsa_vp = vp;
 627         dsp->dsa_outfd = outfd;
 628         dsp->dsa_proc = curproc;
 629         dsp->dsa_os = os;
 630         dsp->dsa_off = off;
 631         dsp->dsa_toguid = dsl_dataset_phys(ds)->ds_guid;
 632         ZIO_SET_CHECKSUM(&dsp->dsa_zc, 0, 0, 0, 0);
 633         dsp->dsa_pending_op = PENDING_NONE;
 634         dsp->dsa_incremental = (fromzb != NULL);
 635         dsp->dsa_featureflags = featureflags;
 636
 637         mutex_enter(&ds->ds_sendstream_lock);
 638         list_insert_head(&ds->ds_sendstreams, dsp);
 639         mutex_exit(&ds->ds_sendstream_lock);
 640
 641         dsl_dataset_long_hold(ds, FTAG);
 642         dsl_pool_rele(dp, tag);
 643
 644         if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0) {
 645                 err = dsp->dsa_err;
 646                 goto out;
 647         }
 648
 649         err = traverse_dataset(ds, fromtxg, TRAVERSE_PRE | TRAVERSE_PREFETCH,
 650             backup_cb, dsp);
 651
 652         if (dsp->dsa_pending_op != PENDING_NONE)
 653                 if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0)
 654                         err = SET_ERROR(EINTR);
 655
 656         if (err != 0) {
 657                 if (err == EINTR && dsp->dsa_err != 0)
 658                         err = dsp->dsa_err;
 659                 goto out;
 660         }
 661
 662         bzero(drr, sizeof (dmu_replay_record_t));
 663         drr->drr_type = DRR_END;
 664         drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc;
 665         drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid;
 666
 667         if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0) {
 668                 err = dsp->dsa_err;
 669                 goto out;
 670         }
 671
 672 out:
 673         mutex_enter(&ds->ds_sendstream_lock);
 674         list_remove(&ds->ds_sendstreams, dsp);
 675         mutex_exit(&ds->ds_sendstream_lock);
 676
 677         kmem_free(drr, sizeof (dmu_replay_record_t));
 678         kmem_free(dsp, sizeof (dmu_sendarg_t));
 679
 680         dsl_dataset_long_rele(ds, FTAG);
 681
 682         return (err);
 683 }
 684
 685 int
 686 dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
 687     boolean_t embedok, int outfd, vnode_t *vp, offset_t *off)
 688 {
 689         dsl_pool_t *dp;
 690         dsl_dataset_t *ds;
 691         dsl_dataset_t *fromds = NULL;
 692         int err;
 693
 694         err = dsl_pool_hold(pool, FTAG, &dp);
 695         if (err != 0)
 696                 return (err);
 697
 698         err = dsl_dataset_hold_obj(dp, tosnap, FTAG, &ds);
 699         if (err != 0) {
 700                 dsl_pool_rele(dp, FTAG);
 701                 return (err);
 702         }
 703
 704         if (fromsnap != 0) {
 705                 zfs_bookmark_phys_t zb;
 706                 boolean_t is_clone;
 707
 708                 err = dsl_dataset_hold_obj(dp, fromsnap, FTAG, &fromds);
 709                 if (err != 0) {
 710                         dsl_dataset_rele(ds, FTAG);
 711                         dsl_pool_rele(dp, FTAG);
 712                         return (err);
 713                 }
 714                 if (!dsl_dataset_is_before(ds, fromds, 0))
 715                         err = SET_ERROR(EXDEV);
 716                 zb.zbm_creation_time =
 717                     dsl_dataset_phys(fromds)->ds_creation_time;
 718                 zb.zbm_creation_txg = dsl_dataset_phys(fromds)->ds_creation_txg;
 719                 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
 720                 is_clone = (fromds->ds_dir != ds->ds_dir);
 721                 dsl_dataset_rele(fromds, FTAG);
 722                 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone, embedok,
 723                     outfd, vp, off);
 724         } else {
 725                 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE, embedok,
 726                     outfd, vp, off);
 727         }
 728         dsl_dataset_rele(ds, FTAG);
 729         return (err);
 730 }
 731
 732 int
 733 dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok,
 734     int outfd, vnode_t *vp, offset_t *off)
 735 {
 736         dsl_pool_t *dp;
 737         dsl_dataset_t *ds;
 738         int err;
 739         boolean_t owned = B_FALSE;
 740
 741         if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL)
 742                 return (SET_ERROR(EINVAL));
 743
 744         err = dsl_pool_hold(tosnap, FTAG, &dp);
 745         if (err != 0)
 746                 return (err);
 747
 748         if (strchr(tosnap, '@') == NULL && spa_writeable(dp->dp_spa)) {
 749                 /*
 750                  * We are sending a filesystem or volume.  Ensure
 751                  * that it doesn't change by owning the dataset.
 752                  */
 753                 err = dsl_dataset_own(dp, tosnap, FTAG, &ds);
 754                 owned = B_TRUE;
 755         } else {
 756                 err = dsl_dataset_hold(dp, tosnap, FTAG, &ds);
 757         }
 758         if (err != 0) {
 759                 dsl_pool_rele(dp, FTAG);
 760                 return (err);
 761         }
 762
 763         if (fromsnap != NULL) {
 764                 zfs_bookmark_phys_t zb;
 765                 boolean_t is_clone = B_FALSE;
 766                 int fsnamelen = strchr(tosnap, '@') - tosnap;
 767
 768                 /*
 769                  * If the fromsnap is in a different filesystem, then
 770                  * mark the send stream as a clone.
 771                  */
 772                 if (strncmp(tosnap, fromsnap, fsnamelen) != 0 ||
 773                     (fromsnap[fsnamelen] != '@' &&
 774                     fromsnap[fsnamelen] != '#')) {
 775                         is_clone = B_TRUE;
 776                 }
 777
 778                 if (strchr(fromsnap, '@')) {
 779                         dsl_dataset_t *fromds;
 780                         err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds);
 781                         if (err == 0) {
 782                                 if (!dsl_dataset_is_before(ds, fromds, 0))
 783                                         err = SET_ERROR(EXDEV);
 784                                 zb.zbm_creation_time =
 785                                     dsl_dataset_phys(fromds)->ds_creation_time;
 786                                 zb.zbm_creation_txg =
 787                                     dsl_dataset_phys(fromds)->ds_creation_txg;
 788                                 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
 789                                 is_clone = (ds->ds_dir != fromds->ds_dir);
 790                                 dsl_dataset_rele(fromds, FTAG);
 791                         }
 792                 } else {
 793                         err = dsl_bookmark_lookup(dp, fromsnap, ds, &zb);
 794                 }
 795                 if (err != 0) {
 796                         dsl_dataset_rele(ds, FTAG);
 797                         dsl_pool_rele(dp, FTAG);
 798                         return (err);
 799                 }
 800                 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone, embedok,
 801                     outfd, vp, off);
 802         } else {
 803                 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE, embedok,
 804                     outfd, vp, off);
 805         }
 806         if (owned)
 807                 dsl_dataset_disown(ds, FTAG);
 808         else
 809                 dsl_dataset_rele(ds, FTAG);
 810         return (err);
 811 }
 812
 813 int
 814 dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds, uint64_t *sizep)
 815 {
 816         int err;
 817         uint64_t size, recordsize;
 818         ASSERTV(dsl_pool_t *dp = ds->ds_dir->dd_pool);
 819
 820         ASSERT(dsl_pool_config_held(dp));
 821
 822         /* tosnap must be a snapshot */
 823         if (!ds->ds_is_snapshot)
 824                 return (SET_ERROR(EINVAL));
 825
 826         /*
 827          * fromsnap must be an earlier snapshot from the same fs as tosnap,
 828          * or the origin's fs.
 829          */
 830         if (fromds != NULL && !dsl_dataset_is_before(ds, fromds, 0))
 831                 return (SET_ERROR(EXDEV));
 832
 833         /* Get uncompressed size estimate of changed data. */
 834         if (fromds == NULL) {
 835                 size = dsl_dataset_phys(ds)->ds_uncompressed_bytes;
 836         } else {
 837                 uint64_t used, comp;
 838                 err = dsl_dataset_space_written(fromds, ds,
 839                     &used, &comp, &size);
 840                 if (err != 0)
 841                         return (err);
 842         }
 843
 844         /*
 845          * Assume that space (both on-disk and in-stream) is dominated by
 846          * data.  We will adjust for indirect blocks and the copies property,
 847          * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
 848          */
 849
 850         /*
 851          * Subtract out approximate space used by indirect blocks.
 852          * Assume most space is used by data blocks (non-indirect, non-dnode).
 853          * Assume all blocks are recordsize.  Assume ditto blocks and
 854          * internal fragmentation counter out compression.
 855          *
 856          * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
 857          * block, which we observe in practice.
 858          */
 859         err = dsl_prop_get_int_ds(ds, "recordsize", &recordsize);
 860         if (err != 0)
 861                 return (err);
 862         size -= size / recordsize * sizeof (blkptr_t);
 863
 864         /* Add in the space for the record associated with each block. */
 865         size += size / recordsize * sizeof (dmu_replay_record_t);
 866
 867         *sizep = size;
 868
 869         return (0);
 870 }
 871
 872 typedef struct dmu_recv_begin_arg {
 873         const char *drba_origin;
 874         dmu_recv_cookie_t *drba_cookie;
 875         cred_t *drba_cred;
 876         uint64_t drba_snapobj;
 877 } dmu_recv_begin_arg_t;
 878
 879 static int
 880 recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds,
 881     uint64_t fromguid)
 882 {
 883         uint64_t val;
 884         int error;
 885         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 886
 887         /* temporary clone name must not exist */
 888         error = zap_lookup(dp->dp_meta_objset,
 889             dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, recv_clone_name,
 890             8, 1, &val);
 891         if (error != ENOENT)
 892                 return (error == 0 ? EBUSY : error);
 893
 894         /* new snapshot name must not exist */
 895         error = zap_lookup(dp->dp_meta_objset,
 896             dsl_dataset_phys(ds)->ds_snapnames_zapobj,
 897             drba->drba_cookie->drc_tosnap, 8, 1, &val);
 898         if (error != ENOENT)
 899                 return (error == 0 ? EEXIST : error);
 900
 901         /*
 902          * Check snapshot limit before receiving. We'll recheck again at the
 903          * end, but might as well abort before receiving if we're already over
 904          * the limit.
 905          *
 906          * Note that we do not check the file system limit with
 907          * dsl_dir_fscount_check because the temporary %clones don't count
 908          * against that limit.
 909          */
 910         error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT,
 911             NULL, drba->drba_cred);
 912         if (error != 0)
 913                 return (error);
 914
 915         if (fromguid != 0) {
 916                 dsl_dataset_t *snap;
 917                 uint64_t obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
 918
 919                 /* Find snapshot in this dir that matches fromguid. */
 920                 while (obj != 0) {
 921                         error = dsl_dataset_hold_obj(dp, obj, FTAG,
 922                             &snap);
 923                         if (error != 0)
 924                                 return (SET_ERROR(ENODEV));
 925                         if (snap->ds_dir != ds->ds_dir) {
 926                                 dsl_dataset_rele(snap, FTAG);
 927                                 return (SET_ERROR(ENODEV));
 928                         }
 929                         if (dsl_dataset_phys(snap)->ds_guid == fromguid)
 930                                 break;
 931                         obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
 932                         dsl_dataset_rele(snap, FTAG);
 933                 }
 934                 if (obj == 0)
 935                         return (SET_ERROR(ENODEV));
 936
 937                 if (drba->drba_cookie->drc_force) {
 938                         drba->drba_snapobj = obj;
 939                 } else {
 940                         /*
 941                          * If we are not forcing, there must be no
 942                          * changes since fromsnap.
 943                          */
 944                         if (dsl_dataset_modified_since_snap(ds, snap)) {
 945                                 dsl_dataset_rele(snap, FTAG);
 946                                 return (SET_ERROR(ETXTBSY));
 947                         }
 948                         drba->drba_snapobj = ds->ds_prev->ds_object;
 949                 }
 950
 951                 dsl_dataset_rele(snap, FTAG);
 952         } else {
 953                 /* if full, most recent snapshot must be $ORIGIN */
 954                 if (dsl_dataset_phys(ds)->ds_prev_snap_txg >= TXG_INITIAL)
 955                         return (SET_ERROR(ENODEV));
 956                 drba->drba_snapobj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
 957         }
 958
 959         return (0);
 960
 961 }
 962
 963 static int
 964 dmu_recv_begin_check(void *arg, dmu_tx_t *tx)
 965 {
 966         dmu_recv_begin_arg_t *drba = arg;
 967         dsl_pool_t *dp = dmu_tx_pool(tx);
 968         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
 969         uint64_t fromguid = drrb->drr_fromguid;
 970         int flags = drrb->drr_flags;
 971         int error;
 972         uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
 973         dsl_dataset_t *ds;
 974         const char *tofs = drba->drba_cookie->drc_tofs;
 975
 976         /* already checked */
 977         ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
 978
 979         if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
 980             DMU_COMPOUNDSTREAM ||
 981             drrb->drr_type >= DMU_OST_NUMTYPES ||
 982             ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL))
 983                 return (SET_ERROR(EINVAL));
 984
 985         /* Verify pool version supports SA if SA_SPILL feature set */
 986         if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
 987             spa_version(dp->dp_spa) < SPA_VERSION_SA)
 988                 return (SET_ERROR(ENOTSUP));
 989
 990         /*
 991          * The receiving code doesn't know how to translate a WRITE_EMBEDDED
 992          * record to a plan WRITE record, so the pool must have the
 993          * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
 994          * records.  Same with WRITE_EMBEDDED records that use LZ4 compression.
 995          */
 996         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
 997             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
 998                 return (SET_ERROR(ENOTSUP));
 999         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4) &&
1000             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1001                 return (SET_ERROR(ENOTSUP));
1002
1003         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1004         if (error == 0) {
1005                 /* target fs already exists; recv into temp clone */
1006
1007                 /* Can't recv a clone into an existing fs */
1008                 if (flags & DRR_FLAG_CLONE) {
1009                         dsl_dataset_rele(ds, FTAG);
1010                         return (SET_ERROR(EINVAL));
1011                 }
1012
1013                 error = recv_begin_check_existing_impl(drba, ds, fromguid);
1014                 dsl_dataset_rele(ds, FTAG);
1015         } else if (error == ENOENT) {
1016                 /* target fs does not exist; must be a full backup or clone */
1017                 char buf[MAXNAMELEN];
1018
1019                 /*
1020                  * If it's a non-clone incremental, we are missing the
1021                  * target fs, so fail the recv.
1022                  */
1023                 if (fromguid != 0 && !(flags & DRR_FLAG_CLONE))
1024                         return (SET_ERROR(ENOENT));
1025
1026                 /* Open the parent of tofs */
1027                 ASSERT3U(strlen(tofs), <, MAXNAMELEN);
1028                 (void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1);
1029                 error = dsl_dataset_hold(dp, buf, FTAG, &ds);
1030                 if (error != 0)
1031                         return (error);
1032
1033                 /*
1034                  * Check filesystem and snapshot limits before receiving. We'll
1035                  * recheck snapshot limits again at the end (we create the
1036                  * filesystems and increment those counts during begin_sync).
1037                  */
1038                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1039                     ZFS_PROP_FILESYSTEM_LIMIT, NULL, drba->drba_cred);
1040                 if (error != 0) {
1041                         dsl_dataset_rele(ds, FTAG);
1042                         return (error);
1043                 }
1044
1045                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1046                     ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred);
1047                 if (error != 0) {
1048                         dsl_dataset_rele(ds, FTAG);
1049                         return (error);
1050                 }
1051
1052                 if (drba->drba_origin != NULL) {
1053                         dsl_dataset_t *origin;
1054                         error = dsl_dataset_hold(dp, drba->drba_origin,
1055                             FTAG, &origin);
1056                         if (error != 0) {
1057                                 dsl_dataset_rele(ds, FTAG);
1058                                 return (error);
1059                         }
1060                         if (!origin->ds_is_snapshot) {
1061                                 dsl_dataset_rele(origin, FTAG);
1062                                 dsl_dataset_rele(ds, FTAG);
1063                                 return (SET_ERROR(EINVAL));
1064                         }
1065                         if (dsl_dataset_phys(origin)->ds_guid != fromguid) {
1066                                 dsl_dataset_rele(origin, FTAG);
1067                                 dsl_dataset_rele(ds, FTAG);
1068                                 return (SET_ERROR(ENODEV));
1069                         }
1070                         dsl_dataset_rele(origin, FTAG);
1071                 }
1072                 dsl_dataset_rele(ds, FTAG);
1073                 error = 0;
1074         }
1075         return (error);
1076 }
1077
1078 static void
1079 dmu_recv_begin_sync(void *arg, dmu_tx_t *tx)
1080 {
1081         dmu_recv_begin_arg_t *drba = arg;
1082         dsl_pool_t *dp = dmu_tx_pool(tx);
1083         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1084         const char *tofs = drba->drba_cookie->drc_tofs;
1085         dsl_dataset_t *ds, *newds;
1086         uint64_t dsobj;
1087         int error;
1088         uint64_t crflags;
1089
1090         crflags = (drrb->drr_flags & DRR_FLAG_CI_DATA) ?
1091             DS_FLAG_CI_DATASET : 0;
1092
1093         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1094         if (error == 0) {
1095                 /* create temporary clone */
1096                 dsl_dataset_t *snap = NULL;
1097                 if (drba->drba_snapobj != 0) {
1098                         VERIFY0(dsl_dataset_hold_obj(dp,
1099                             drba->drba_snapobj, FTAG, &snap));
1100                 }
1101                 dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name,
1102                     snap, crflags, drba->drba_cred, tx);
1103                 dsl_dataset_rele(snap, FTAG);
1104                 dsl_dataset_rele(ds, FTAG);
1105         } else {
1106                 dsl_dir_t *dd;
1107                 const char *tail;
1108                 dsl_dataset_t *origin = NULL;
1109
1110                 VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail));
1111
1112                 if (drba->drba_origin != NULL) {
1113                         VERIFY0(dsl_dataset_hold(dp, drba->drba_origin,
1114                             FTAG, &origin));
1115                 }
1116
1117                 /* Create new dataset. */
1118                 dsobj = dsl_dataset_create_sync(dd,
1119                     strrchr(tofs, '/') + 1,
1120                     origin, crflags, drba->drba_cred, tx);
1121                 if (origin != NULL)
1122                         dsl_dataset_rele(origin, FTAG);
1123                 dsl_dir_rele(dd, FTAG);
1124                 drba->drba_cookie->drc_newfs = B_TRUE;
1125         }
1126         VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds));
1127
1128         dmu_buf_will_dirty(newds->ds_dbuf, tx);
1129         dsl_dataset_phys(newds)->ds_flags |= DS_FLAG_INCONSISTENT;
1130
1131         /*
1132          * If we actually created a non-clone, we need to create the
1133          * objset in our new dataset.
1134          */
1135         if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) {
1136                 (void) dmu_objset_create_impl(dp->dp_spa,
1137                     newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx);
1138         }
1139
1140         drba->drba_cookie->drc_ds = newds;
1141
1142         spa_history_log_internal_ds(newds, "receive", tx, "");
1143 }
1144
1145 /*
1146  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
1147  * succeeds; otherwise we will leak the holds on the datasets.
1148  */
1149 int
1150 dmu_recv_begin(char *tofs, char *tosnap, struct drr_begin *drrb,
1151     boolean_t force, char *origin, dmu_recv_cookie_t *drc)
1152 {
1153         dmu_recv_begin_arg_t drba = { 0 };
1154         dmu_replay_record_t *drr;
1155
1156         bzero(drc, sizeof (dmu_recv_cookie_t));
1157         drc->drc_drrb = drrb;
1158         drc->drc_tosnap = tosnap;
1159         drc->drc_tofs = tofs;
1160         drc->drc_force = force;
1161         drc->drc_cred = CRED();
1162
1163         if (drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
1164                 drc->drc_byteswap = B_TRUE;
1165         else if (drrb->drr_magic != DMU_BACKUP_MAGIC)
1166                 return (SET_ERROR(EINVAL));
1167
1168         drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
1169         drr->drr_type = DRR_BEGIN;
1170         drr->drr_u.drr_begin = *drc->drc_drrb;
1171         if (drc->drc_byteswap) {
1172                 fletcher_4_incremental_byteswap(drr,
1173                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1174         } else {
1175                 fletcher_4_incremental_native(drr,
1176                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1177         }
1178         kmem_free(drr, sizeof (dmu_replay_record_t));
1179
1180         if (drc->drc_byteswap) {
1181                 drrb->drr_magic = BSWAP_64(drrb->drr_magic);
1182                 drrb->drr_versioninfo = BSWAP_64(drrb->drr_versioninfo);
1183                 drrb->drr_creation_time = BSWAP_64(drrb->drr_creation_time);
1184                 drrb->drr_type = BSWAP_32(drrb->drr_type);
1185                 drrb->drr_toguid = BSWAP_64(drrb->drr_toguid);
1186                 drrb->drr_fromguid = BSWAP_64(drrb->drr_fromguid);
1187         }
1188
1189         drba.drba_origin = origin;
1190         drba.drba_cookie = drc;
1191         drba.drba_cred = CRED();
1192
1193         return (dsl_sync_task(tofs, dmu_recv_begin_check, dmu_recv_begin_sync,
1194             &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1195 }
1196
1197 struct restorearg {
1198         int err;
1199         boolean_t byteswap;
1200         vnode_t *vp;
1201         char *buf;
1202         uint64_t voff;
1203         int bufsize; /* amount of memory allocated for buf */
1204         zio_cksum_t cksum;
1205         avl_tree_t *guid_to_ds_map;
1206 };
1207
1208 typedef struct guid_map_entry {
1209         uint64_t        guid;
1210         dsl_dataset_t   *gme_ds;
1211         avl_node_t      avlnode;
1212 } guid_map_entry_t;
1213
1214 static int
1215 guid_compare(const void *arg1, const void *arg2)
1216 {
1217         const guid_map_entry_t *gmep1 = arg1;
1218         const guid_map_entry_t *gmep2 = arg2;
1219
1220         if (gmep1->guid < gmep2->guid)
1221                 return (-1);
1222         else if (gmep1->guid > gmep2->guid)
1223                 return (1);
1224         return (0);
1225 }
1226
1227 static void
1228 free_guid_map_onexit(void *arg)
1229 {
1230         avl_tree_t *ca = arg;
1231         void *cookie = NULL;
1232         guid_map_entry_t *gmep;
1233
1234         while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
1235                 dsl_dataset_long_rele(gmep->gme_ds, gmep);
1236                 dsl_dataset_rele(gmep->gme_ds, gmep);
1237                 kmem_free(gmep, sizeof (guid_map_entry_t));
1238         }
1239         avl_destroy(ca);
1240         kmem_free(ca, sizeof (avl_tree_t));
1241 }
1242
1243 static void *
1244 restore_read(struct restorearg *ra, int len, char *buf)
1245 {
1246         int done = 0;
1247
1248         if (buf == NULL)
1249                 buf = ra->buf;
1250
1251         /* some things will require 8-byte alignment, so everything must */
1252         ASSERT0(len % 8);
1253
1254         while (done < len) {
1255                 ssize_t resid;
1256
1257                 ra->err = vn_rdwr(UIO_READ, ra->vp,
1258                     buf + done, len - done,
1259                     ra->voff, UIO_SYSSPACE, FAPPEND,
1260                     RLIM64_INFINITY, CRED(), &resid);
1261
1262                 if (resid == len - done)
1263                         ra->err = SET_ERROR(EINVAL);
1264                 ra->voff += len - done - resid;
1265                 done = len - resid;
1266                 if (ra->err != 0)
1267                         return (NULL);
1268         }
1269
1270         ASSERT3U(done, ==, len);
1271         if (ra->byteswap)
1272                 fletcher_4_incremental_byteswap(buf, len, &ra->cksum);
1273         else
1274                 fletcher_4_incremental_native(buf, len, &ra->cksum);
1275         return (buf);
1276 }
1277
1278 noinline static void
1279 backup_byteswap(dmu_replay_record_t *drr)
1280 {
1281 #define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
1282 #define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
1283         drr->drr_type = BSWAP_32(drr->drr_type);
1284         drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
1285         switch (drr->drr_type) {
1286         case DRR_BEGIN:
1287                 DO64(drr_begin.drr_magic);
1288                 DO64(drr_begin.drr_versioninfo);
1289                 DO64(drr_begin.drr_creation_time);
1290                 DO32(drr_begin.drr_type);
1291                 DO32(drr_begin.drr_flags);
1292                 DO64(drr_begin.drr_toguid);
1293                 DO64(drr_begin.drr_fromguid);
1294                 break;
1295         case DRR_OBJECT:
1296                 DO64(drr_object.drr_object);
1297                 DO32(drr_object.drr_type);
1298                 DO32(drr_object.drr_bonustype);
1299                 DO32(drr_object.drr_blksz);
1300                 DO32(drr_object.drr_bonuslen);
1301                 DO64(drr_object.drr_toguid);
1302                 break;
1303         case DRR_FREEOBJECTS:
1304                 DO64(drr_freeobjects.drr_firstobj);
1305                 DO64(drr_freeobjects.drr_numobjs);
1306                 DO64(drr_freeobjects.drr_toguid);
1307                 break;
1308         case DRR_WRITE:
1309                 DO64(drr_write.drr_object);
1310                 DO32(drr_write.drr_type);
1311                 DO64(drr_write.drr_offset);
1312                 DO64(drr_write.drr_length);
1313                 DO64(drr_write.drr_toguid);
1314                 DO64(drr_write.drr_key.ddk_cksum.zc_word[0]);
1315                 DO64(drr_write.drr_key.ddk_cksum.zc_word[1]);
1316                 DO64(drr_write.drr_key.ddk_cksum.zc_word[2]);
1317                 DO64(drr_write.drr_key.ddk_cksum.zc_word[3]);
1318                 DO64(drr_write.drr_key.ddk_prop);
1319                 break;
1320         case DRR_WRITE_BYREF:
1321                 DO64(drr_write_byref.drr_object);
1322                 DO64(drr_write_byref.drr_offset);
1323                 DO64(drr_write_byref.drr_length);
1324                 DO64(drr_write_byref.drr_toguid);
1325                 DO64(drr_write_byref.drr_refguid);
1326                 DO64(drr_write_byref.drr_refobject);
1327                 DO64(drr_write_byref.drr_refoffset);
1328                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[0]);
1329                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[1]);
1330                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[2]);
1331                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[3]);
1332                 DO64(drr_write_byref.drr_key.ddk_prop);
1333                 break;
1334         case DRR_WRITE_EMBEDDED:
1335                 DO64(drr_write_embedded.drr_object);
1336                 DO64(drr_write_embedded.drr_offset);
1337                 DO64(drr_write_embedded.drr_length);
1338                 DO64(drr_write_embedded.drr_toguid);
1339                 DO32(drr_write_embedded.drr_lsize);
1340                 DO32(drr_write_embedded.drr_psize);
1341                 break;
1342         case DRR_FREE:
1343                 DO64(drr_free.drr_object);
1344                 DO64(drr_free.drr_offset);
1345                 DO64(drr_free.drr_length);
1346                 DO64(drr_free.drr_toguid);
1347                 break;
1348         case DRR_SPILL:
1349                 DO64(drr_spill.drr_object);
1350                 DO64(drr_spill.drr_length);
1351                 DO64(drr_spill.drr_toguid);
1352                 break;
1353         case DRR_END:
1354                 DO64(drr_end.drr_checksum.zc_word[0]);
1355                 DO64(drr_end.drr_checksum.zc_word[1]);
1356                 DO64(drr_end.drr_checksum.zc_word[2]);
1357                 DO64(drr_end.drr_checksum.zc_word[3]);
1358                 DO64(drr_end.drr_toguid);
1359                 break;
1360         default:
1361                 break;
1362         }
1363 #undef DO64
1364 #undef DO32
1365 }
1366
1367 static inline uint8_t
1368 deduce_nblkptr(dmu_object_type_t bonus_type, uint64_t bonus_size)
1369 {
1370         if (bonus_type == DMU_OT_SA) {
1371                 return (1);
1372         } else {
1373                 return (1 +
1374                     ((DN_MAX_BONUSLEN - bonus_size) >> SPA_BLKPTRSHIFT));
1375         }
1376 }
1377
1378 noinline static int
1379 restore_object(struct restorearg *ra, objset_t *os, struct drr_object *drro)
1380 {
1381         dmu_object_info_t doi;
1382         dmu_tx_t *tx;
1383         void *data = NULL;
1384         uint64_t object;
1385         int err;
1386
1387         if (drro->drr_type == DMU_OT_NONE ||
1388             !DMU_OT_IS_VALID(drro->drr_type) ||
1389             !DMU_OT_IS_VALID(drro->drr_bonustype) ||
1390             drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
1391             drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
1392             P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
1393             drro->drr_blksz < SPA_MINBLOCKSIZE ||
1394             drro->drr_blksz > SPA_MAXBLOCKSIZE ||
1395             drro->drr_bonuslen > DN_MAX_BONUSLEN) {
1396                 return (SET_ERROR(EINVAL));
1397         }
1398
1399         err = dmu_object_info(os, drro->drr_object, &doi);
1400
1401         if (err != 0 && err != ENOENT)
1402                 return (SET_ERROR(EINVAL));
1403         object = err == 0 ? drro->drr_object : DMU_NEW_OBJECT;
1404
1405         if (drro->drr_bonuslen) {
1406                 data = restore_read(ra, P2ROUNDUP(drro->drr_bonuslen, 8), NULL);
1407                 if (ra->err != 0)
1408                         return (ra->err);
1409         }
1410
1411         /*
1412          * If we are losing blkptrs or changing the block size this must
1413          * be a new file instance.  We must clear out the previous file
1414          * contents before we can change this type of metadata in the dnode.
1415          */
1416         if (err == 0) {
1417                 int nblkptr;
1418
1419                 nblkptr = deduce_nblkptr(drro->drr_bonustype,
1420                     drro->drr_bonuslen);
1421
1422                 if (drro->drr_blksz != doi.doi_data_block_size ||
1423                     nblkptr < doi.doi_nblkptr) {
1424                         err = dmu_free_long_range(os, drro->drr_object,
1425                             0, DMU_OBJECT_END);
1426                         if (err != 0)
1427                                 return (SET_ERROR(EINVAL));
1428                 }
1429         }
1430
1431         tx = dmu_tx_create(os);
1432         dmu_tx_hold_bonus(tx, object);
1433         err = dmu_tx_assign(tx, TXG_WAIT);
1434         if (err != 0) {
1435                 dmu_tx_abort(tx);
1436                 return (err);
1437         }
1438
1439         if (object == DMU_NEW_OBJECT) {
1440                 /* currently free, want to be allocated */
1441                 err = dmu_object_claim(os, drro->drr_object,
1442                     drro->drr_type, drro->drr_blksz,
1443                     drro->drr_bonustype, drro->drr_bonuslen, tx);
1444         } else if (drro->drr_type != doi.doi_type ||
1445             drro->drr_blksz != doi.doi_data_block_size ||
1446             drro->drr_bonustype != doi.doi_bonus_type ||
1447             drro->drr_bonuslen != doi.doi_bonus_size) {
1448                 /* currently allocated, but with different properties */
1449                 err = dmu_object_reclaim(os, drro->drr_object,
1450                     drro->drr_type, drro->drr_blksz,
1451                     drro->drr_bonustype, drro->drr_bonuslen, tx);
1452         }
1453         if (err != 0) {
1454                 dmu_tx_commit(tx);
1455                 return (SET_ERROR(EINVAL));
1456         }
1457
1458         dmu_object_set_checksum(os, drro->drr_object, drro->drr_checksumtype,
1459             tx);
1460         dmu_object_set_compress(os, drro->drr_object, drro->drr_compress, tx);
1461
1462         if (data != NULL) {
1463                 dmu_buf_t *db;
1464
1465                 VERIFY(0 == dmu_bonus_hold(os, drro->drr_object, FTAG, &db));
1466                 dmu_buf_will_dirty(db, tx);
1467
1468                 ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
1469                 bcopy(data, db->db_data, drro->drr_bonuslen);
1470                 if (ra->byteswap) {
1471                         dmu_object_byteswap_t byteswap =
1472                             DMU_OT_BYTESWAP(drro->drr_bonustype);
1473                         dmu_ot_byteswap[byteswap].ob_func(db->db_data,
1474                             drro->drr_bonuslen);
1475                 }
1476                 dmu_buf_rele(db, FTAG);
1477         }
1478         dmu_tx_commit(tx);
1479         return (0);
1480 }
1481
1482 /* ARGSUSED */
1483 noinline static int
1484 restore_freeobjects(struct restorearg *ra, objset_t *os,
1485     struct drr_freeobjects *drrfo)
1486 {
1487         uint64_t obj;
1488
1489         if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
1490                 return (SET_ERROR(EINVAL));
1491
1492         for (obj = drrfo->drr_firstobj;
1493             obj < drrfo->drr_firstobj + drrfo->drr_numobjs;
1494             (void) dmu_object_next(os, &obj, FALSE, 0)) {
1495                 int err;
1496
1497                 if (dmu_object_info(os, obj, NULL) != 0)
1498                         continue;
1499
1500                 err = dmu_free_long_object(os, obj);
1501                 if (err != 0)
1502                         return (err);
1503         }
1504         return (0);
1505 }
1506
1507 noinline static int
1508 restore_write(struct restorearg *ra, objset_t *os,
1509     struct drr_write *drrw)
1510 {
1511         dmu_tx_t *tx;
1512         dmu_buf_t *bonus;
1513         arc_buf_t *abuf;
1514         void *data;
1515         int err;
1516
1517         if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
1518             !DMU_OT_IS_VALID(drrw->drr_type))
1519                 return (SET_ERROR(EINVAL));
1520
1521         if (dmu_object_info(os, drrw->drr_object, NULL) != 0)
1522                 return (SET_ERROR(EINVAL));
1523
1524         if (dmu_bonus_hold(os, drrw->drr_object, FTAG, &bonus) != 0)
1525                 return (SET_ERROR(EINVAL));
1526
1527         abuf = dmu_request_arcbuf(bonus, drrw->drr_length);
1528
1529         data = restore_read(ra, drrw->drr_length, abuf->b_data);
1530         if (data == NULL) {
1531                 dmu_return_arcbuf(abuf);
1532                 dmu_buf_rele(bonus, FTAG);
1533                 return (ra->err);
1534         }
1535
1536         tx = dmu_tx_create(os);
1537
1538         dmu_tx_hold_write(tx, drrw->drr_object,
1539             drrw->drr_offset, drrw->drr_length);
1540         err = dmu_tx_assign(tx, TXG_WAIT);
1541         if (err != 0) {
1542                 dmu_return_arcbuf(abuf);
1543                 dmu_buf_rele(bonus, FTAG);
1544                 dmu_tx_abort(tx);
1545                 return (err);
1546         }
1547         if (ra->byteswap) {
1548                 dmu_object_byteswap_t byteswap =
1549                     DMU_OT_BYTESWAP(drrw->drr_type);
1550                 dmu_ot_byteswap[byteswap].ob_func(data, drrw->drr_length);
1551         }
1552         dmu_assign_arcbuf(bonus, drrw->drr_offset, abuf, tx);
1553         dmu_tx_commit(tx);
1554         dmu_buf_rele(bonus, FTAG);
1555         return (0);
1556 }
1557
1558 /*
1559  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
1560  * streams to refer to a copy of the data that is already on the
1561  * system because it came in earlier in the stream.  This function
1562  * finds the earlier copy of the data, and uses that copy instead of
1563  * data from the stream to fulfill this write.
1564  */
1565 static int
1566 restore_write_byref(struct restorearg *ra, objset_t *os,
1567     struct drr_write_byref *drrwbr)
1568 {
1569         dmu_tx_t *tx;
1570         int err;
1571         guid_map_entry_t gmesrch;
1572         guid_map_entry_t *gmep;
1573         avl_index_t where;
1574         objset_t *ref_os = NULL;
1575         dmu_buf_t *dbp;
1576
1577         if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
1578                 return (SET_ERROR(EINVAL));
1579
1580         /*
1581          * If the GUID of the referenced dataset is different from the
1582          * GUID of the target dataset, find the referenced dataset.
1583          */
1584         if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
1585                 gmesrch.guid = drrwbr->drr_refguid;
1586                 if ((gmep = avl_find(ra->guid_to_ds_map, &gmesrch,
1587                     &where)) == NULL) {
1588                         return (SET_ERROR(EINVAL));
1589                 }
1590                 if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
1591                         return (SET_ERROR(EINVAL));
1592         } else {
1593                 ref_os = os;
1594         }
1595
1596         err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
1597             drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH);
1598         if (err != 0)
1599                 return (err);
1600
1601         tx = dmu_tx_create(os);
1602
1603         dmu_tx_hold_write(tx, drrwbr->drr_object,
1604             drrwbr->drr_offset, drrwbr->drr_length);
1605         err = dmu_tx_assign(tx, TXG_WAIT);
1606         if (err != 0) {
1607                 dmu_tx_abort(tx);
1608                 return (err);
1609         }
1610         dmu_write(os, drrwbr->drr_object,
1611             drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
1612         dmu_buf_rele(dbp, FTAG);
1613         dmu_tx_commit(tx);
1614         return (0);
1615 }
1616
1617 static int
1618 restore_write_embedded(struct restorearg *ra, objset_t *os,
1619     struct drr_write_embedded *drrwnp)
1620 {
1621         dmu_tx_t *tx;
1622         int err;
1623         void *data;
1624
1625         if (drrwnp->drr_offset + drrwnp->drr_length < drrwnp->drr_offset)
1626                 return (EINVAL);
1627
1628         if (drrwnp->drr_psize > BPE_PAYLOAD_SIZE)
1629                 return (EINVAL);
1630
1631         if (drrwnp->drr_etype >= NUM_BP_EMBEDDED_TYPES)
1632                 return (EINVAL);
1633         if (drrwnp->drr_compression >= ZIO_COMPRESS_FUNCTIONS)
1634                 return (EINVAL);
1635
1636         data = restore_read(ra, P2ROUNDUP(drrwnp->drr_psize, 8), NULL);
1637         if (data == NULL)
1638                 return (ra->err);
1639
1640         tx = dmu_tx_create(os);
1641
1642         dmu_tx_hold_write(tx, drrwnp->drr_object,
1643             drrwnp->drr_offset, drrwnp->drr_length);
1644         err = dmu_tx_assign(tx, TXG_WAIT);
1645         if (err != 0) {
1646                 dmu_tx_abort(tx);
1647                 return (err);
1648         }
1649
1650         dmu_write_embedded(os, drrwnp->drr_object,
1651             drrwnp->drr_offset, data, drrwnp->drr_etype,
1652             drrwnp->drr_compression, drrwnp->drr_lsize, drrwnp->drr_psize,
1653             ra->byteswap ^ ZFS_HOST_BYTEORDER, tx);
1654
1655         dmu_tx_commit(tx);
1656         return (0);
1657 }
1658
1659 static int
1660 restore_spill(struct restorearg *ra, objset_t *os, struct drr_spill *drrs)
1661 {
1662         dmu_tx_t *tx;
1663         void *data;
1664         dmu_buf_t *db, *db_spill;
1665         int err;
1666
1667         if (drrs->drr_length < SPA_MINBLOCKSIZE ||
1668             drrs->drr_length > SPA_MAXBLOCKSIZE)
1669                 return (SET_ERROR(EINVAL));
1670
1671         data = restore_read(ra, drrs->drr_length, NULL);
1672         if (data == NULL)
1673                 return (ra->err);
1674
1675         if (dmu_object_info(os, drrs->drr_object, NULL) != 0)
1676                 return (SET_ERROR(EINVAL));
1677
1678         VERIFY(0 == dmu_bonus_hold(os, drrs->drr_object, FTAG, &db));
1679         if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
1680                 dmu_buf_rele(db, FTAG);
1681                 return (err);
1682         }
1683
1684         tx = dmu_tx_create(os);
1685
1686         dmu_tx_hold_spill(tx, db->db_object);
1687
1688         err = dmu_tx_assign(tx, TXG_WAIT);
1689         if (err != 0) {
1690                 dmu_buf_rele(db, FTAG);
1691                 dmu_buf_rele(db_spill, FTAG);
1692                 dmu_tx_abort(tx);
1693                 return (err);
1694         }
1695         dmu_buf_will_dirty(db_spill, tx);
1696
1697         if (db_spill->db_size < drrs->drr_length)
1698                 VERIFY(0 == dbuf_spill_set_blksz(db_spill,
1699                     drrs->drr_length, tx));
1700         bcopy(data, db_spill->db_data, drrs->drr_length);
1701
1702         dmu_buf_rele(db, FTAG);
1703         dmu_buf_rele(db_spill, FTAG);
1704
1705         dmu_tx_commit(tx);
1706         return (0);
1707 }
1708
1709 /* ARGSUSED */
1710 noinline static int
1711 restore_free(struct restorearg *ra, objset_t *os,
1712     struct drr_free *drrf)
1713 {
1714         int err;
1715
1716         if (drrf->drr_length != -1ULL &&
1717             drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
1718                 return (SET_ERROR(EINVAL));
1719
1720         if (dmu_object_info(os, drrf->drr_object, NULL) != 0)
1721                 return (SET_ERROR(EINVAL));
1722
1723         err = dmu_free_long_range(os, drrf->drr_object,
1724             drrf->drr_offset, drrf->drr_length);
1725         return (err);
1726 }
1727
1728 /* used to destroy the drc_ds on error */
1729 static void
1730 dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc)
1731 {
1732         char name[MAXNAMELEN];
1733         dsl_dataset_name(drc->drc_ds, name);
1734         dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
1735         (void) dsl_destroy_head(name);
1736 }
1737
1738 /*
1739  * NB: callers *must* call dmu_recv_end() if this succeeds.
1740  */
1741 int
1742 dmu_recv_stream(dmu_recv_cookie_t *drc, vnode_t *vp, offset_t *voffp,
1743     int cleanup_fd, uint64_t *action_handlep)
1744 {
1745         struct restorearg ra = { 0 };
1746         dmu_replay_record_t *drr;
1747         objset_t *os;
1748         zio_cksum_t pcksum;
1749         int featureflags;
1750
1751         ra.byteswap = drc->drc_byteswap;
1752         ra.cksum = drc->drc_cksum;
1753         ra.vp = vp;
1754         ra.voff = *voffp;
1755         ra.bufsize = 1<<20;
1756         ra.buf = vmem_alloc(ra.bufsize, KM_SLEEP);
1757
1758         /* these were verified in dmu_recv_begin */
1759         ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==,
1760             DMU_SUBSTREAM);
1761         ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES);
1762
1763         /*
1764          * Open the objset we are modifying.
1765          */
1766         VERIFY0(dmu_objset_from_ds(drc->drc_ds, &os));
1767
1768         ASSERT(dsl_dataset_phys(drc->drc_ds)->ds_flags & DS_FLAG_INCONSISTENT);
1769
1770         featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
1771
1772         /* if this stream is dedup'ed, set up the avl tree for guid mapping */
1773         if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
1774                 minor_t minor;
1775
1776                 if (cleanup_fd == -1) {
1777                         ra.err = SET_ERROR(EBADF);
1778                         goto out;
1779                 }
1780                 ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
1781                 if (ra.err != 0) {
1782                         cleanup_fd = -1;
1783                         goto out;
1784                 }
1785
1786                 if (*action_handlep == 0) {
1787                         ra.guid_to_ds_map =
1788                             kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
1789                         avl_create(ra.guid_to_ds_map, guid_compare,
1790                             sizeof (guid_map_entry_t),
1791                             offsetof(guid_map_entry_t, avlnode));
1792                         ra.err = zfs_onexit_add_cb(minor,
1793                             free_guid_map_onexit, ra.guid_to_ds_map,
1794                             action_handlep);
1795                         if (ra.err != 0)
1796                                 goto out;
1797                 } else {
1798                         ra.err = zfs_onexit_cb_data(minor, *action_handlep,
1799                             (void **)&ra.guid_to_ds_map);
1800                         if (ra.err != 0)
1801                                 goto out;
1802                 }
1803
1804                 drc->drc_guid_to_ds_map = ra.guid_to_ds_map;
1805         }
1806
1807         /*
1808          * Read records and process them.
1809          */
1810         pcksum = ra.cksum;
1811         while (ra.err == 0 &&
1812             NULL != (drr = restore_read(&ra, sizeof (*drr), NULL))) {
1813                 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1814                         ra.err = SET_ERROR(EINTR);
1815                         goto out;
1816                 }
1817
1818                 if (ra.byteswap)
1819                         backup_byteswap(drr);
1820
1821                 switch (drr->drr_type) {
1822                 case DRR_OBJECT:
1823                 {
1824                         /*
1825                          * We need to make a copy of the record header,
1826                          * because restore_{object,write} may need to
1827                          * restore_read(), which will invalidate drr.
1828                          */
1829                         struct drr_object drro = drr->drr_u.drr_object;
1830                         ra.err = restore_object(&ra, os, &drro);
1831                         break;
1832                 }
1833                 case DRR_FREEOBJECTS:
1834                 {
1835                         struct drr_freeobjects drrfo =
1836                             drr->drr_u.drr_freeobjects;
1837                         ra.err = restore_freeobjects(&ra, os, &drrfo);
1838                         break;
1839                 }
1840                 case DRR_WRITE:
1841                 {
1842                         struct drr_write drrw = drr->drr_u.drr_write;
1843                         ra.err = restore_write(&ra, os, &drrw);
1844                         break;
1845                 }
1846                 case DRR_WRITE_BYREF:
1847                 {
1848                         struct drr_write_byref drrwbr =
1849                             drr->drr_u.drr_write_byref;
1850                         ra.err = restore_write_byref(&ra, os, &drrwbr);
1851                         break;
1852                 }
1853                 case DRR_WRITE_EMBEDDED:
1854                 {
1855                         struct drr_write_embedded drrwe =
1856                             drr->drr_u.drr_write_embedded;
1857                         ra.err = restore_write_embedded(&ra, os, &drrwe);
1858                         break;
1859                 }
1860                 case DRR_FREE:
1861                 {
1862                         struct drr_free drrf = drr->drr_u.drr_free;
1863                         ra.err = restore_free(&ra, os, &drrf);
1864                         break;
1865                 }
1866                 case DRR_END:
1867                 {
1868                         struct drr_end drre = drr->drr_u.drr_end;
1869                         /*
1870                          * We compare against the *previous* checksum
1871                          * value, because the stored checksum is of
1872                          * everything before the DRR_END record.
1873                          */
1874                         if (!ZIO_CHECKSUM_EQUAL(drre.drr_checksum, pcksum))
1875                                 ra.err = SET_ERROR(ECKSUM);
1876                         goto out;
1877                 }
1878                 case DRR_SPILL:
1879                 {
1880                         struct drr_spill drrs = drr->drr_u.drr_spill;
1881                         ra.err = restore_spill(&ra, os, &drrs);
1882                         break;
1883                 }
1884                 default:
1885                         ra.err = SET_ERROR(EINVAL);
1886                         goto out;
1887                 }
1888                 pcksum = ra.cksum;
1889         }
1890         ASSERT(ra.err != 0);
1891
1892 out:
1893         if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
1894                 zfs_onexit_fd_rele(cleanup_fd);
1895
1896         if (ra.err != 0) {
1897                 /*
1898                  * destroy what we created, so we don't leave it in the
1899                  * inconsistent restoring state.
1900                  */
1901                 dmu_recv_cleanup_ds(drc);
1902         }
1903
1904         vmem_free(ra.buf, ra.bufsize);
1905         *voffp = ra.voff;
1906         return (ra.err);
1907 }
1908
1909 static int
1910 dmu_recv_end_check(void *arg, dmu_tx_t *tx)
1911 {
1912         dmu_recv_cookie_t *drc = arg;
1913         dsl_pool_t *dp = dmu_tx_pool(tx);
1914         int error;
1915
1916         ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag);
1917
1918         if (!drc->drc_newfs) {
1919                 dsl_dataset_t *origin_head;
1920
1921                 error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head);
1922                 if (error != 0)
1923                         return (error);
1924                 if (drc->drc_force) {
1925                         /*
1926                          * We will destroy any snapshots in tofs (i.e. before
1927                          * origin_head) that are after the origin (which is
1928                          * the snap before drc_ds, because drc_ds can not
1929                          * have any snaps of its own).
1930                          */
1931                         uint64_t obj;
1932
1933                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
1934                         while (obj !=
1935                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
1936                                 dsl_dataset_t *snap;
1937                                 error = dsl_dataset_hold_obj(dp, obj, FTAG,
1938                                     &snap);
1939                                 if (error != 0)
1940                                         return (error);
1941                                 if (snap->ds_dir != origin_head->ds_dir)
1942                                         error = SET_ERROR(EINVAL);
1943                                 if (error == 0)  {
1944                                         error = dsl_destroy_snapshot_check_impl(
1945                                             snap, B_FALSE);
1946                                 }
1947                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
1948                                 dsl_dataset_rele(snap, FTAG);
1949                                 if (error != 0)
1950                                         return (error);
1951                         }
1952                 }
1953                 error = dsl_dataset_clone_swap_check_impl(drc->drc_ds,
1954                     origin_head, drc->drc_force, drc->drc_owner, tx);
1955                 if (error != 0) {
1956                         dsl_dataset_rele(origin_head, FTAG);
1957                         return (error);
1958                 }
1959                 error = dsl_dataset_snapshot_check_impl(origin_head,
1960                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
1961                 dsl_dataset_rele(origin_head, FTAG);
1962                 if (error != 0)
1963                         return (error);
1964
1965                 error = dsl_destroy_head_check_impl(drc->drc_ds, 1);
1966         } else {
1967                 error = dsl_dataset_snapshot_check_impl(drc->drc_ds,
1968                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
1969         }
1970         return (error);
1971 }
1972
1973 static void
1974 dmu_recv_end_sync(void *arg, dmu_tx_t *tx)
1975 {
1976         dmu_recv_cookie_t *drc = arg;
1977         dsl_pool_t *dp = dmu_tx_pool(tx);
1978
1979         spa_history_log_internal_ds(drc->drc_ds, "finish receiving",
1980             tx, "snap=%s", drc->drc_tosnap);
1981
1982         if (!drc->drc_newfs) {
1983                 dsl_dataset_t *origin_head;
1984
1985                 VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG,
1986                     &origin_head));
1987
1988                 if (drc->drc_force) {
1989                         /*
1990                          * Destroy any snapshots of drc_tofs (origin_head)
1991                          * after the origin (the snap before drc_ds).
1992                          */
1993                         uint64_t obj;
1994
1995                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
1996                         while (obj !=
1997                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
1998                                 dsl_dataset_t *snap;
1999                                 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG,
2000                                     &snap));
2001                                 ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir);
2002                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
2003                                 dsl_destroy_snapshot_sync_impl(snap,
2004                                     B_FALSE, tx);
2005                                 dsl_dataset_rele(snap, FTAG);
2006                         }
2007                 }
2008                 VERIFY3P(drc->drc_ds->ds_prev, ==,
2009                     origin_head->ds_prev);
2010
2011                 dsl_dataset_clone_swap_sync_impl(drc->drc_ds,
2012                     origin_head, tx);
2013                 dsl_dataset_snapshot_sync_impl(origin_head,
2014                     drc->drc_tosnap, tx);
2015
2016                 /* set snapshot's creation time and guid */
2017                 dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx);
2018                 dsl_dataset_phys(origin_head->ds_prev)->ds_creation_time =
2019                     drc->drc_drrb->drr_creation_time;
2020                 dsl_dataset_phys(origin_head->ds_prev)->ds_guid =
2021                     drc->drc_drrb->drr_toguid;
2022                 dsl_dataset_phys(origin_head->ds_prev)->ds_flags &=
2023                     ~DS_FLAG_INCONSISTENT;
2024
2025                 dmu_buf_will_dirty(origin_head->ds_dbuf, tx);
2026                 dsl_dataset_phys(origin_head)->ds_flags &=
2027                     ~DS_FLAG_INCONSISTENT;
2028
2029                 dsl_dataset_rele(origin_head, FTAG);
2030                 dsl_destroy_head_sync_impl(drc->drc_ds, tx);
2031
2032                 if (drc->drc_owner != NULL)
2033                         VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner);
2034         } else {
2035                 dsl_dataset_t *ds = drc->drc_ds;
2036
2037                 dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx);
2038
2039                 /* set snapshot's creation time and guid */
2040                 dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
2041                 dsl_dataset_phys(ds->ds_prev)->ds_creation_time =
2042                     drc->drc_drrb->drr_creation_time;
2043                 dsl_dataset_phys(ds->ds_prev)->ds_guid =
2044                     drc->drc_drrb->drr_toguid;
2045                 dsl_dataset_phys(ds->ds_prev)->ds_flags &=
2046                     ~DS_FLAG_INCONSISTENT;
2047
2048                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
2049                 dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
2050         }
2051         drc->drc_newsnapobj = dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj;
2052         /*
2053          * Release the hold from dmu_recv_begin.  This must be done before
2054          * we return to open context, so that when we free the dataset's dnode,
2055          * we can evict its bonus buffer.
2056          */
2057         dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2058         drc->drc_ds = NULL;
2059 }
2060
2061 static int
2062 add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj)
2063 {
2064         dsl_pool_t *dp;
2065         dsl_dataset_t *snapds;
2066         guid_map_entry_t *gmep;
2067         int err;
2068
2069         ASSERT(guid_map != NULL);
2070
2071         err = dsl_pool_hold(name, FTAG, &dp);
2072         if (err != 0)
2073                 return (err);
2074         gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP);
2075         err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds);
2076         if (err == 0) {
2077                 gmep->guid = dsl_dataset_phys(snapds)->ds_guid;
2078                 gmep->gme_ds = snapds;
2079                 avl_add(guid_map, gmep);
2080                 dsl_dataset_long_hold(snapds, gmep);
2081         } else {
2082                 kmem_free(gmep, sizeof (*gmep));
2083         }
2084
2085         dsl_pool_rele(dp, FTAG);
2086         return (err);
2087 }
2088
2089 static int dmu_recv_end_modified_blocks = 3;
2090
2091 static int
2092 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
2093 {
2094         int error;
2095
2096 #ifdef _KERNEL
2097         char *name;
2098
2099         /*
2100          * We will be destroying the ds; make sure its origin is unmounted if
2101          * necessary.
2102          */
2103         name = kmem_alloc(MAXNAMELEN, KM_SLEEP);
2104         dsl_dataset_name(drc->drc_ds, name);
2105         zfs_destroy_unmount_origin(name);
2106         kmem_free(name, MAXNAMELEN);
2107 #endif
2108
2109         error = dsl_sync_task(drc->drc_tofs,
2110             dmu_recv_end_check, dmu_recv_end_sync, drc,
2111             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL);
2112
2113         if (error != 0)
2114                 dmu_recv_cleanup_ds(drc);
2115         return (error);
2116 }
2117
2118 static int
2119 dmu_recv_new_end(dmu_recv_cookie_t *drc)
2120 {
2121         int error;
2122
2123         error = dsl_sync_task(drc->drc_tofs,
2124             dmu_recv_end_check, dmu_recv_end_sync, drc,
2125             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL);
2126
2127         if (error != 0) {
2128                 dmu_recv_cleanup_ds(drc);
2129         } else if (drc->drc_guid_to_ds_map != NULL) {
2130                 (void) add_ds_to_guidmap(drc->drc_tofs,
2131                     drc->drc_guid_to_ds_map,
2132                     drc->drc_newsnapobj);
2133         }
2134         return (error);
2135 }
2136
2137 int
2138 dmu_recv_end(dmu_recv_cookie_t *drc, void *owner)
2139 {
2140         drc->drc_owner = owner;
2141
2142         if (drc->drc_newfs)
2143                 return (dmu_recv_new_end(drc));
2144         else
2145                 return (dmu_recv_existing_end(drc));
2146 }
2147
2148 /*
2149  * Return TRUE if this objset is currently being received into.
2150  */
2151 boolean_t
2152 dmu_objset_is_receiving(objset_t *os)
2153 {
2154         return (os->os_dsl_dataset != NULL &&
2155             os->os_dsl_dataset->ds_owner == dmu_recv_tag);
2156 }
2157
2158 #if defined(_KERNEL)
2159 module_param(zfs_send_corrupt_data, int, 0644);
2160 MODULE_PARM_DESC(zfs_send_corrupt_data, "Allow sending corrupt data");
2161 #endif