sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dmu_send.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
  24  * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
  25  * Copyright (c) 2014, Joyent, Inc. All rights reserved.
  26  * Copyright (c) 2012, Martin Matuska <mm@FreeBSD.org>. All rights reserved.
  27  * Copyright 2014 HybridCluster. All rights reserved.
  28  */
  29
  30 #include <sys/dmu.h>
  31 #include <sys/dmu_impl.h>
  32 #include <sys/dmu_tx.h>
  33 #include <sys/dbuf.h>
  34 #include <sys/dnode.h>
  35 #include <sys/zfs_context.h>
  36 #include <sys/dmu_objset.h>
  37 #include <sys/dmu_traverse.h>
  38 #include <sys/dsl_dataset.h>
  39 #include <sys/dsl_dir.h>
  40 #include <sys/dsl_prop.h>
  41 #include <sys/dsl_pool.h>
  42 #include <sys/dsl_synctask.h>
  43 #include <sys/zfs_ioctl.h>
  44 #include <sys/zap.h>
  45 #include <sys/zio_checksum.h>
  46 #include <sys/zfs_znode.h>
  47 #include <zfs_fletcher.h>
  48 #include <sys/avl.h>
  49 #include <sys/ddt.h>
  50 #include <sys/zfs_onexit.h>
  51 #include <sys/dmu_send.h>
  52 #include <sys/dsl_destroy.h>
  53 #include <sys/blkptr.h>
  54 #include <sys/dsl_bookmark.h>
  55 #include <sys/zfeature.h>
  56
  57 #ifdef __FreeBSD__
  58 #undef dump_write
  59 #define dump_write dmu_dump_write
  60 #endif
  61
  62 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
  63 int zfs_send_corrupt_data = B_FALSE;
  64
  65 static char *dmu_recv_tag = "dmu_recv_tag";
  66 static const char *recv_clone_name = "%recv";
  67
  68 static int
  69 dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
  70 {
  71         dsl_dataset_t *ds = dsp->dsa_os->os_dsl_dataset;
  72         struct uio auio;
  73         struct iovec aiov;
  74         ASSERT0(len % 8);
  75
  76         fletcher_4_incremental_native(buf, len, &dsp->dsa_zc);
  77         aiov.iov_base = buf;
  78         aiov.iov_len = len;
  79         auio.uio_iov = &aiov;
  80         auio.uio_iovcnt = 1;
  81         auio.uio_resid = len;
  82         auio.uio_segflg = UIO_SYSSPACE;
  83         auio.uio_rw = UIO_WRITE;
  84         auio.uio_offset = (off_t)-1;
  85         auio.uio_td = dsp->dsa_td;
  86 #ifdef _KERNEL
  87         if (dsp->dsa_fp->f_type == DTYPE_VNODE)
  88                 bwillwrite();
  89         dsp->dsa_err = fo_write(dsp->dsa_fp, &auio, dsp->dsa_td->td_ucred, 0,
  90             dsp->dsa_td);
  91 #else
  92         fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__);
  93         dsp->dsa_err = EOPNOTSUPP;
  94 #endif
  95         mutex_enter(&ds->ds_sendstream_lock);
  96         *dsp->dsa_off += len;
  97         mutex_exit(&ds->ds_sendstream_lock);
  98
  99         return (dsp->dsa_err);
 100 }
 101
 102 static int
 103 dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
 104     uint64_t length)
 105 {
 106         struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free);
 107
 108         /*
 109          * When we receive a free record, dbuf_free_range() assumes
 110          * that the receiving system doesn't have any dbufs in the range
 111          * being freed.  This is always true because there is a one-record
 112          * constraint: we only send one WRITE record for any given
 113          * object+offset.  We know that the one-record constraint is
 114          * true because we always send data in increasing order by
 115          * object,offset.
 116          *
 117          * If the increasing-order constraint ever changes, we should find
 118          * another way to assert that the one-record constraint is still
 119          * satisfied.
 120          */
 121         ASSERT(object > dsp->dsa_last_data_object ||
 122             (object == dsp->dsa_last_data_object &&
 123             offset > dsp->dsa_last_data_offset));
 124
 125         /*
 126          * If we are doing a non-incremental send, then there can't
 127          * be any data in the dataset we're receiving into.  Therefore
 128          * a free record would simply be a no-op.  Save space by not
 129          * sending it to begin with.
 130          */
 131         if (!dsp->dsa_incremental)
 132                 return (0);
 133
 134         if (length != -1ULL && offset + length < offset)
 135                 length = -1ULL;
 136
 137         /*
 138          * If there is a pending op, but it's not PENDING_FREE, push it out,
 139          * since free block aggregation can only be done for blocks of the
 140          * same type (i.e., DRR_FREE records can only be aggregated with
 141          * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
 142          * aggregated with other DRR_FREEOBJECTS records.
 143          */
 144         if (dsp->dsa_pending_op != PENDING_NONE &&
 145             dsp->dsa_pending_op != PENDING_FREE) {
 146                 if (dump_bytes(dsp, dsp->dsa_drr,
 147                     sizeof (dmu_replay_record_t)) != 0)
 148                         return (SET_ERROR(EINTR));
 149                 dsp->dsa_pending_op = PENDING_NONE;
 150         }
 151
 152         if (dsp->dsa_pending_op == PENDING_FREE) {
 153                 /*
 154                  * There should never be a PENDING_FREE if length is -1
 155                  * (because dump_dnode is the only place where this
 156                  * function is called with a -1, and only after flushing
 157                  * any pending record).
 158                  */
 159                 ASSERT(length != -1ULL);
 160                 /*
 161                  * Check to see whether this free block can be aggregated
 162                  * with pending one.
 163                  */
 164                 if (drrf->drr_object == object && drrf->drr_offset +
 165                     drrf->drr_length == offset) {
 166                         drrf->drr_length += length;
 167                         return (0);
 168                 } else {
 169                         /* not a continuation.  Push out pending record */
 170                         if (dump_bytes(dsp, dsp->dsa_drr,
 171                             sizeof (dmu_replay_record_t)) != 0)
 172                                 return (SET_ERROR(EINTR));
 173                         dsp->dsa_pending_op = PENDING_NONE;
 174                 }
 175         }
 176         /* create a FREE record and make it pending */
 177         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 178         dsp->dsa_drr->drr_type = DRR_FREE;
 179         drrf->drr_object = object;
 180         drrf->drr_offset = offset;
 181         drrf->drr_length = length;
 182         drrf->drr_toguid = dsp->dsa_toguid;
 183         if (length == -1ULL) {
 184                 if (dump_bytes(dsp, dsp->dsa_drr,
 185                     sizeof (dmu_replay_record_t)) != 0)
 186                         return (SET_ERROR(EINTR));
 187         } else {
 188                 dsp->dsa_pending_op = PENDING_FREE;
 189         }
 190
 191         return (0);
 192 }
 193
 194 static int
 195 dump_write(dmu_sendarg_t *dsp, dmu_object_type_t type,
 196     uint64_t object, uint64_t offset, int blksz, const blkptr_t *bp, void *data)
 197 {
 198         struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write);
 199
 200         /*
 201          * We send data in increasing object, offset order.
 202          * See comment in dump_free() for details.
 203          */
 204         ASSERT(object > dsp->dsa_last_data_object ||
 205             (object == dsp->dsa_last_data_object &&
 206             offset > dsp->dsa_last_data_offset));
 207         dsp->dsa_last_data_object = object;
 208         dsp->dsa_last_data_offset = offset + blksz - 1;
 209
 210         /*
 211          * If there is any kind of pending aggregation (currently either
 212          * a grouping of free objects or free blocks), push it out to
 213          * the stream, since aggregation can't be done across operations
 214          * of different types.
 215          */
 216         if (dsp->dsa_pending_op != PENDING_NONE) {
 217                 if (dump_bytes(dsp, dsp->dsa_drr,
 218                     sizeof (dmu_replay_record_t)) != 0)
 219                         return (SET_ERROR(EINTR));
 220                 dsp->dsa_pending_op = PENDING_NONE;
 221         }
 222         /* write a DATA record */
 223         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 224         dsp->dsa_drr->drr_type = DRR_WRITE;
 225         drrw->drr_object = object;
 226         drrw->drr_type = type;
 227         drrw->drr_offset = offset;
 228         drrw->drr_length = blksz;
 229         drrw->drr_toguid = dsp->dsa_toguid;
 230         if (bp == NULL || BP_IS_EMBEDDED(bp)) {
 231                 /*
 232                  * There's no pre-computed checksum for partial-block
 233                  * writes or embedded BP's, so (like
 234                  * fletcher4-checkummed blocks) userland will have to
 235                  * compute a dedup-capable checksum itself.
 236                  */
 237                 drrw->drr_checksumtype = ZIO_CHECKSUM_OFF;
 238         } else {
 239                 drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
 240                 if (zio_checksum_table[drrw->drr_checksumtype].ci_dedup)
 241                         drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
 242                 DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
 243                 DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
 244                 DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
 245                 drrw->drr_key.ddk_cksum = bp->blk_cksum;
 246         }
 247
 248         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
 249                 return (SET_ERROR(EINTR));
 250         if (dump_bytes(dsp, data, blksz) != 0)
 251                 return (SET_ERROR(EINTR));
 252         return (0);
 253 }
 254
 255 static int
 256 dump_write_embedded(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
 257     int blksz, const blkptr_t *bp)
 258 {
 259         char buf[BPE_PAYLOAD_SIZE];
 260         struct drr_write_embedded *drrw =
 261             &(dsp->dsa_drr->drr_u.drr_write_embedded);
 262
 263         if (dsp->dsa_pending_op != PENDING_NONE) {
 264                 if (dump_bytes(dsp, dsp->dsa_drr,
 265                     sizeof (dmu_replay_record_t)) != 0)
 266                         return (EINTR);
 267                 dsp->dsa_pending_op = PENDING_NONE;
 268         }
 269
 270         ASSERT(BP_IS_EMBEDDED(bp));
 271
 272         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 273         dsp->dsa_drr->drr_type = DRR_WRITE_EMBEDDED;
 274         drrw->drr_object = object;
 275         drrw->drr_offset = offset;
 276         drrw->drr_length = blksz;
 277         drrw->drr_toguid = dsp->dsa_toguid;
 278         drrw->drr_compression = BP_GET_COMPRESS(bp);
 279         drrw->drr_etype = BPE_GET_ETYPE(bp);
 280         drrw->drr_lsize = BPE_GET_LSIZE(bp);
 281         drrw->drr_psize = BPE_GET_PSIZE(bp);
 282
 283         decode_embedded_bp_compressed(bp, buf);
 284
 285         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
 286                 return (EINTR);
 287         if (dump_bytes(dsp, buf, P2ROUNDUP(drrw->drr_psize, 8)) != 0)
 288                 return (EINTR);
 289         return (0);
 290 }
 291
 292 static int
 293 dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data)
 294 {
 295         struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill);
 296
 297         if (dsp->dsa_pending_op != PENDING_NONE) {
 298                 if (dump_bytes(dsp, dsp->dsa_drr,
 299                     sizeof (dmu_replay_record_t)) != 0)
 300                         return (SET_ERROR(EINTR));
 301                 dsp->dsa_pending_op = PENDING_NONE;
 302         }
 303
 304         /* write a SPILL record */
 305         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 306         dsp->dsa_drr->drr_type = DRR_SPILL;
 307         drrs->drr_object = object;
 308         drrs->drr_length = blksz;
 309         drrs->drr_toguid = dsp->dsa_toguid;
 310
 311         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)))
 312                 return (SET_ERROR(EINTR));
 313         if (dump_bytes(dsp, data, blksz))
 314                 return (SET_ERROR(EINTR));
 315         return (0);
 316 }
 317
 318 static int
 319 dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
 320 {
 321         struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
 322
 323         /* See comment in dump_free(). */
 324         if (!dsp->dsa_incremental)
 325                 return (0);
 326
 327         /*
 328          * If there is a pending op, but it's not PENDING_FREEOBJECTS,
 329          * push it out, since free block aggregation can only be done for
 330          * blocks of the same type (i.e., DRR_FREE records can only be
 331          * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
 332          * can only be aggregated with other DRR_FREEOBJECTS records.
 333          */
 334         if (dsp->dsa_pending_op != PENDING_NONE &&
 335             dsp->dsa_pending_op != PENDING_FREEOBJECTS) {
 336                 if (dump_bytes(dsp, dsp->dsa_drr,
 337                     sizeof (dmu_replay_record_t)) != 0)
 338                         return (SET_ERROR(EINTR));
 339                 dsp->dsa_pending_op = PENDING_NONE;
 340         }
 341         if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) {
 342                 /*
 343                  * See whether this free object array can be aggregated
 344                  * with pending one
 345                  */
 346                 if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
 347                         drrfo->drr_numobjs += numobjs;
 348                         return (0);
 349                 } else {
 350                         /* can't be aggregated.  Push out pending record */
 351                         if (dump_bytes(dsp, dsp->dsa_drr,
 352                             sizeof (dmu_replay_record_t)) != 0)
 353                                 return (SET_ERROR(EINTR));
 354                         dsp->dsa_pending_op = PENDING_NONE;
 355                 }
 356         }
 357
 358         /* write a FREEOBJECTS record */
 359         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 360         dsp->dsa_drr->drr_type = DRR_FREEOBJECTS;
 361         drrfo->drr_firstobj = firstobj;
 362         drrfo->drr_numobjs = numobjs;
 363         drrfo->drr_toguid = dsp->dsa_toguid;
 364
 365         dsp->dsa_pending_op = PENDING_FREEOBJECTS;
 366
 367         return (0);
 368 }
 369
 370 static int
 371 dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
 372 {
 373         struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object);
 374
 375         if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
 376                 return (dump_freeobjects(dsp, object, 1));
 377
 378         if (dsp->dsa_pending_op != PENDING_NONE) {
 379                 if (dump_bytes(dsp, dsp->dsa_drr,
 380                     sizeof (dmu_replay_record_t)) != 0)
 381                         return (SET_ERROR(EINTR));
 382                 dsp->dsa_pending_op = PENDING_NONE;
 383         }
 384
 385         /* write an OBJECT record */
 386         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 387         dsp->dsa_drr->drr_type = DRR_OBJECT;
 388         drro->drr_object = object;
 389         drro->drr_type = dnp->dn_type;
 390         drro->drr_bonustype = dnp->dn_bonustype;
 391         drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 392         drro->drr_bonuslen = dnp->dn_bonuslen;
 393         drro->drr_checksumtype = dnp->dn_checksum;
 394         drro->drr_compress = dnp->dn_compress;
 395         drro->drr_toguid = dsp->dsa_toguid;
 396
 397         if (!(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
 398             drro->drr_blksz > SPA_OLD_MAXBLOCKSIZE)
 399                 drro->drr_blksz = SPA_OLD_MAXBLOCKSIZE;
 400
 401         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
 402                 return (SET_ERROR(EINTR));
 403
 404         if (dump_bytes(dsp, DN_BONUS(dnp), P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0)
 405                 return (SET_ERROR(EINTR));
 406
 407         /* Free anything past the end of the file. */
 408         if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) *
 409             (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0)
 410                 return (SET_ERROR(EINTR));
 411         if (dsp->dsa_err != 0)
 412                 return (SET_ERROR(EINTR));
 413         return (0);
 414 }
 415
 416 static boolean_t
 417 backup_do_embed(dmu_sendarg_t *dsp, const blkptr_t *bp)
 418 {
 419         if (!BP_IS_EMBEDDED(bp))
 420                 return (B_FALSE);
 421
 422         /*
 423          * Compression function must be legacy, or explicitly enabled.
 424          */
 425         if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS &&
 426             !(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4)))
 427                 return (B_FALSE);
 428
 429         /*
 430          * Embed type must be explicitly enabled.
 431          */
 432         switch (BPE_GET_ETYPE(bp)) {
 433         case BP_EMBEDDED_TYPE_DATA:
 434                 if (dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)
 435                         return (B_TRUE);
 436                 break;
 437         default:
 438                 return (B_FALSE);
 439         }
 440         return (B_FALSE);
 441 }
 442
 443 #define BP_SPAN(dnp, level) \
 444         (((uint64_t)dnp->dn_datablkszsec) << (SPA_MINBLOCKSHIFT + \
 445         (level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)))
 446
 447 /* ARGSUSED */
 448 static int
 449 backup_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 450     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
 451 {
 452         dmu_sendarg_t *dsp = arg;
 453         dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
 454         int err = 0;
 455
 456         if (issig(JUSTLOOKING) && issig(FORREAL))
 457                 return (SET_ERROR(EINTR));
 458
 459         if (zb->zb_object != DMU_META_DNODE_OBJECT &&
 460             DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
 461                 return (0);
 462         } else if (zb->zb_level == ZB_ZIL_LEVEL) {
 463                 /*
 464                  * If we are sending a non-snapshot (which is allowed on
 465                  * read-only pools), it may have a ZIL, which must be ignored.
 466                  */
 467                 return (0);
 468         } else if (BP_IS_HOLE(bp) &&
 469             zb->zb_object == DMU_META_DNODE_OBJECT) {
 470                 uint64_t span = BP_SPAN(dnp, zb->zb_level);
 471                 uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
 472                 err = dump_freeobjects(dsp, dnobj, span >> DNODE_SHIFT);
 473         } else if (BP_IS_HOLE(bp)) {
 474                 uint64_t span = BP_SPAN(dnp, zb->zb_level);
 475                 err = dump_free(dsp, zb->zb_object, zb->zb_blkid * span, span);
 476         } else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
 477                 return (0);
 478         } else if (type == DMU_OT_DNODE) {
 479                 dnode_phys_t *blk;
 480                 int i;
 481                 int blksz = BP_GET_LSIZE(bp);
 482                 arc_flags_t aflags = ARC_FLAG_WAIT;
 483                 arc_buf_t *abuf;
 484
 485                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 486                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 487                     &aflags, zb) != 0)
 488                         return (SET_ERROR(EIO));
 489
 490                 blk = abuf->b_data;
 491                 for (i = 0; i < blksz >> DNODE_SHIFT; i++) {
 492                         uint64_t dnobj = (zb->zb_blkid <<
 493                             (DNODE_BLOCK_SHIFT - DNODE_SHIFT)) + i;
 494                         err = dump_dnode(dsp, dnobj, blk+i);
 495                         if (err != 0)
 496                                 break;
 497                 }
 498                 (void) arc_buf_remove_ref(abuf, &abuf);
 499         } else if (type == DMU_OT_SA) {
 500                 arc_flags_t aflags = ARC_FLAG_WAIT;
 501                 arc_buf_t *abuf;
 502                 int blksz = BP_GET_LSIZE(bp);
 503
 504                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 505                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 506                     &aflags, zb) != 0)
 507                         return (SET_ERROR(EIO));
 508
 509                 err = dump_spill(dsp, zb->zb_object, blksz, abuf->b_data);
 510                 (void) arc_buf_remove_ref(abuf, &abuf);
 511         } else if (backup_do_embed(dsp, bp)) {
 512                 /* it's an embedded level-0 block of a regular object */
 513                 int blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 514                 err = dump_write_embedded(dsp, zb->zb_object,
 515                     zb->zb_blkid * blksz, blksz, bp);
 516         } else { /* it's a level-0 block of a regular object */
 517                 arc_flags_t aflags = ARC_FLAG_WAIT;
 518                 arc_buf_t *abuf;
 519                 int blksz = BP_GET_LSIZE(bp);
 520                 uint64_t offset;
 521
 522                 ASSERT3U(blksz, ==, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
 523                 ASSERT0(zb->zb_level);
 524                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 525                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 526                     &aflags, zb) != 0) {
 527                         if (zfs_send_corrupt_data) {
 528                                 /* Send a block filled with 0x"zfs badd bloc" */
 529                                 abuf = arc_buf_alloc(spa, blksz, &abuf,
 530                                     ARC_BUFC_DATA);
 531                                 uint64_t *ptr;
 532                                 for (ptr = abuf->b_data;
 533                                     (char *)ptr < (char *)abuf->b_data + blksz;
 534                                     ptr++)
 535                                         *ptr = 0x2f5baddb10c;
 536                         } else {
 537                                 return (SET_ERROR(EIO));
 538                         }
 539                 }
 540
 541                 offset = zb->zb_blkid * blksz;
 542
 543                 if (!(dsp->dsa_featureflags &
 544                     DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
 545                     blksz > SPA_OLD_MAXBLOCKSIZE) {
 546                         char *buf = abuf->b_data;
 547                         while (blksz > 0 && err == 0) {
 548                                 int n = MIN(blksz, SPA_OLD_MAXBLOCKSIZE);
 549                                 err = dump_write(dsp, type, zb->zb_object,
 550                                     offset, n, NULL, buf);
 551                                 offset += n;
 552                                 buf += n;
 553                                 blksz -= n;
 554                         }
 555                 } else {
 556                         err = dump_write(dsp, type, zb->zb_object,
 557                             offset, blksz, bp, abuf->b_data);
 558                 }
 559                 (void) arc_buf_remove_ref(abuf, &abuf);
 560         }
 561
 562         ASSERT(err == 0 || err == EINTR);
 563         return (err);
 564 }
 565
 566 /*
 567  * Releases dp using the specified tag.
 568  */
 569 static int
 570 dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *ds,
 571     zfs_bookmark_phys_t *fromzb, boolean_t is_clone, boolean_t embedok,
 572 #ifdef illumos
 573     boolean_t large_block_ok, int outfd, vnode_t *vp, offset_t *off)
 574 #else
 575     boolean_t large_block_ok, int outfd, struct file *fp, offset_t *off)
 576 #endif
 577 {
 578         objset_t *os;
 579         dmu_replay_record_t *drr;
 580         dmu_sendarg_t *dsp;
 581         int err;
 582         uint64_t fromtxg = 0;
 583         uint64_t featureflags = 0;
 584
 585         err = dmu_objset_from_ds(ds, &os);
 586         if (err != 0) {
 587                 dsl_pool_rele(dp, tag);
 588                 return (err);
 589         }
 590
 591         drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
 592         drr->drr_type = DRR_BEGIN;
 593         drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
 594         DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
 595             DMU_SUBSTREAM);
 596
 597 #ifdef _KERNEL
 598         if (dmu_objset_type(os) == DMU_OST_ZFS) {
 599                 uint64_t version;
 600                 if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0) {
 601                         kmem_free(drr, sizeof (dmu_replay_record_t));
 602                         dsl_pool_rele(dp, tag);
 603                         return (SET_ERROR(EINVAL));
 604                 }
 605                 if (version >= ZPL_VERSION_SA) {
 606                         featureflags |= DMU_BACKUP_FEATURE_SA_SPILL;
 607                 }
 608         }
 609 #endif
 610
 611         if (large_block_ok && ds->ds_large_blocks)
 612                 featureflags |= DMU_BACKUP_FEATURE_LARGE_BLOCKS;
 613         if (embedok &&
 614             spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) {
 615                 featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA;
 616                 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
 617                         featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA_LZ4;
 618         } else {
 619                 embedok = B_FALSE;
 620         }
 621
 622         DMU_SET_FEATUREFLAGS(drr->drr_u.drr_begin.drr_versioninfo,
 623             featureflags);
 624
 625         drr->drr_u.drr_begin.drr_creation_time =
 626             dsl_dataset_phys(ds)->ds_creation_time;
 627         drr->drr_u.drr_begin.drr_type = dmu_objset_type(os);
 628         if (is_clone)
 629                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
 630         drr->drr_u.drr_begin.drr_toguid = dsl_dataset_phys(ds)->ds_guid;
 631         if (dsl_dataset_phys(ds)->ds_flags & DS_FLAG_CI_DATASET)
 632                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
 633
 634         if (fromzb != NULL) {
 635                 drr->drr_u.drr_begin.drr_fromguid = fromzb->zbm_guid;
 636                 fromtxg = fromzb->zbm_creation_txg;
 637         }
 638         dsl_dataset_name(ds, drr->drr_u.drr_begin.drr_toname);
 639         if (!dsl_dataset_is_snapshot(ds)) {
 640                 (void) strlcat(drr->drr_u.drr_begin.drr_toname, "@--head--",
 641                     sizeof (drr->drr_u.drr_begin.drr_toname));
 642         }
 643
 644         dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP);
 645
 646         dsp->dsa_drr = drr;
 647         dsp->dsa_outfd = outfd;
 648         dsp->dsa_proc = curproc;
 649         dsp->dsa_td = curthread;
 650         dsp->dsa_fp = fp;
 651         dsp->dsa_os = os;
 652         dsp->dsa_off = off;
 653         dsp->dsa_toguid = dsl_dataset_phys(ds)->ds_guid;
 654         ZIO_SET_CHECKSUM(&dsp->dsa_zc, 0, 0, 0, 0);
 655         dsp->dsa_pending_op = PENDING_NONE;
 656         dsp->dsa_incremental = (fromzb != NULL);
 657         dsp->dsa_featureflags = featureflags;
 658
 659         mutex_enter(&ds->ds_sendstream_lock);
 660         list_insert_head(&ds->ds_sendstreams, dsp);
 661         mutex_exit(&ds->ds_sendstream_lock);
 662
 663         dsl_dataset_long_hold(ds, FTAG);
 664         dsl_pool_rele(dp, tag);
 665
 666         if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0) {
 667                 err = dsp->dsa_err;
 668                 goto out;
 669         }
 670
 671         err = traverse_dataset(ds, fromtxg, TRAVERSE_PRE | TRAVERSE_PREFETCH,
 672             backup_cb, dsp);
 673
 674         if (dsp->dsa_pending_op != PENDING_NONE)
 675                 if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0)
 676                         err = SET_ERROR(EINTR);
 677
 678         if (err != 0) {
 679                 if (err == EINTR && dsp->dsa_err != 0)
 680                         err = dsp->dsa_err;
 681                 goto out;
 682         }
 683
 684         bzero(drr, sizeof (dmu_replay_record_t));
 685         drr->drr_type = DRR_END;
 686         drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc;
 687         drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid;
 688
 689         if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0) {
 690                 err = dsp->dsa_err;
 691                 goto out;
 692         }
 693
 694 out:
 695         mutex_enter(&ds->ds_sendstream_lock);
 696         list_remove(&ds->ds_sendstreams, dsp);
 697         mutex_exit(&ds->ds_sendstream_lock);
 698
 699         kmem_free(drr, sizeof (dmu_replay_record_t));
 700         kmem_free(dsp, sizeof (dmu_sendarg_t));
 701
 702         dsl_dataset_long_rele(ds, FTAG);
 703
 704         return (err);
 705 }
 706
 707 int
 708 dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
 709     boolean_t embedok, boolean_t large_block_ok,
 710 #ifdef illumos
 711     int outfd, vnode_t *vp, offset_t *off)
 712 #else
 713     int outfd, struct file *fp, offset_t *off)
 714 #endif
 715 {
 716         dsl_pool_t *dp;
 717         dsl_dataset_t *ds;
 718         dsl_dataset_t *fromds = NULL;
 719         int err;
 720
 721         err = dsl_pool_hold(pool, FTAG, &dp);
 722         if (err != 0)
 723                 return (err);
 724
 725         err = dsl_dataset_hold_obj(dp, tosnap, FTAG, &ds);
 726         if (err != 0) {
 727                 dsl_pool_rele(dp, FTAG);
 728                 return (err);
 729         }
 730
 731         if (fromsnap != 0) {
 732                 zfs_bookmark_phys_t zb;
 733                 boolean_t is_clone;
 734
 735                 err = dsl_dataset_hold_obj(dp, fromsnap, FTAG, &fromds);
 736                 if (err != 0) {
 737                         dsl_dataset_rele(ds, FTAG);
 738                         dsl_pool_rele(dp, FTAG);
 739                         return (err);
 740                 }
 741                 if (!dsl_dataset_is_before(ds, fromds, 0))
 742                         err = SET_ERROR(EXDEV);
 743                 zb.zbm_creation_time =
 744                     dsl_dataset_phys(fromds)->ds_creation_time;
 745                 zb.zbm_creation_txg = dsl_dataset_phys(fromds)->ds_creation_txg;
 746                 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
 747                 is_clone = (fromds->ds_dir != ds->ds_dir);
 748                 dsl_dataset_rele(fromds, FTAG);
 749                 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
 750                     embedok, large_block_ok, outfd, fp, off);
 751         } else {
 752                 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
 753                     embedok, large_block_ok, outfd, fp, off);
 754         }
 755         dsl_dataset_rele(ds, FTAG);
 756         return (err);
 757 }
 758
 759 int
 760 dmu_send(const char *tosnap, const char *fromsnap,
 761     boolean_t embedok, boolean_t large_block_ok,
 762 #ifdef illumos
 763     int outfd, vnode_t *vp, offset_t *off)
 764 #else
 765     int outfd, struct file *fp, offset_t *off)
 766 #endif
 767 {
 768         dsl_pool_t *dp;
 769         dsl_dataset_t *ds;
 770         int err;
 771         boolean_t owned = B_FALSE;
 772
 773         if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL)
 774                 return (SET_ERROR(EINVAL));
 775
 776         err = dsl_pool_hold(tosnap, FTAG, &dp);
 777         if (err != 0)
 778                 return (err);
 779
 780         if (strchr(tosnap, '@') == NULL && spa_writeable(dp->dp_spa)) {
 781                 /*
 782                  * We are sending a filesystem or volume.  Ensure
 783                  * that it doesn't change by owning the dataset.
 784                  */
 785                 err = dsl_dataset_own(dp, tosnap, FTAG, &ds);
 786                 owned = B_TRUE;
 787         } else {
 788                 err = dsl_dataset_hold(dp, tosnap, FTAG, &ds);
 789         }
 790         if (err != 0) {
 791                 dsl_pool_rele(dp, FTAG);
 792                 return (err);
 793         }
 794
 795         if (fromsnap != NULL) {
 796                 zfs_bookmark_phys_t zb;
 797                 boolean_t is_clone = B_FALSE;
 798                 int fsnamelen = strchr(tosnap, '@') - tosnap;
 799
 800                 /*
 801                  * If the fromsnap is in a different filesystem, then
 802                  * mark the send stream as a clone.
 803                  */
 804                 if (strncmp(tosnap, fromsnap, fsnamelen) != 0 ||
 805                     (fromsnap[fsnamelen] != '@' &&
 806                     fromsnap[fsnamelen] != '#')) {
 807                         is_clone = B_TRUE;
 808                 }
 809
 810                 if (strchr(fromsnap, '@')) {
 811                         dsl_dataset_t *fromds;
 812                         err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds);
 813                         if (err == 0) {
 814                                 if (!dsl_dataset_is_before(ds, fromds, 0))
 815                                         err = SET_ERROR(EXDEV);
 816                                 zb.zbm_creation_time =
 817                                     dsl_dataset_phys(fromds)->ds_creation_time;
 818                                 zb.zbm_creation_txg =
 819                                     dsl_dataset_phys(fromds)->ds_creation_txg;
 820                                 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
 821                                 is_clone = (ds->ds_dir != fromds->ds_dir);
 822                                 dsl_dataset_rele(fromds, FTAG);
 823                         }
 824                 } else {
 825                         err = dsl_bookmark_lookup(dp, fromsnap, ds, &zb);
 826                 }
 827                 if (err != 0) {
 828                         dsl_dataset_rele(ds, FTAG);
 829                         dsl_pool_rele(dp, FTAG);
 830                         return (err);
 831                 }
 832                 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
 833                     embedok, large_block_ok, outfd, fp, off);
 834         } else {
 835                 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
 836                     embedok, large_block_ok, outfd, fp, off);
 837         }
 838         if (owned)
 839                 dsl_dataset_disown(ds, FTAG);
 840         else
 841                 dsl_dataset_rele(ds, FTAG);
 842         return (err);
 843 }
 844
 845 int
 846 dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds, uint64_t *sizep)
 847 {
 848         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 849         int err;
 850         uint64_t size;
 851
 852         ASSERT(dsl_pool_config_held(dp));
 853
 854         /* tosnap must be a snapshot */
 855         if (!dsl_dataset_is_snapshot(ds))
 856                 return (SET_ERROR(EINVAL));
 857
 858         /* fromsnap, if provided, must be a snapshot */
 859         if (fromds != NULL && !dsl_dataset_is_snapshot(fromds))
 860                 return (SET_ERROR(EINVAL));
 861
 862         /*
 863          * fromsnap must be an earlier snapshot from the same fs as tosnap,
 864          * or the origin's fs.
 865          */
 866         if (fromds != NULL && !dsl_dataset_is_before(ds, fromds, 0))
 867                 return (SET_ERROR(EXDEV));
 868
 869         /* Get uncompressed size estimate of changed data. */
 870         if (fromds == NULL) {
 871                 size = dsl_dataset_phys(ds)->ds_uncompressed_bytes;
 872         } else {
 873                 uint64_t used, comp;
 874                 err = dsl_dataset_space_written(fromds, ds,
 875                     &used, &comp, &size);
 876                 if (err != 0)
 877                         return (err);
 878         }
 879
 880         /*
 881          * Assume that space (both on-disk and in-stream) is dominated by
 882          * data.  We will adjust for indirect blocks and the copies property,
 883          * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
 884          */
 885
 886         /*
 887          * Subtract out approximate space used by indirect blocks.
 888          * Assume most space is used by data blocks (non-indirect, non-dnode).
 889          * Assume all blocks are recordsize.  Assume ditto blocks and
 890          * internal fragmentation counter out compression.
 891          *
 892          * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
 893          * block, which we observe in practice.
 894          */
 895         uint64_t recordsize;
 896         err = dsl_prop_get_int_ds(ds, "recordsize", &recordsize);
 897         if (err != 0)
 898                 return (err);
 899         size -= size / recordsize * sizeof (blkptr_t);
 900
 901         /* Add in the space for the record associated with each block. */
 902         size += size / recordsize * sizeof (dmu_replay_record_t);
 903
 904         *sizep = size;
 905
 906         return (0);
 907 }
 908
 909 typedef struct dmu_recv_begin_arg {
 910         const char *drba_origin;
 911         dmu_recv_cookie_t *drba_cookie;
 912         cred_t *drba_cred;
 913         uint64_t drba_snapobj;
 914 } dmu_recv_begin_arg_t;
 915
 916 static int
 917 recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds,
 918     uint64_t fromguid)
 919 {
 920         uint64_t val;
 921         int error;
 922         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 923
 924         /* temporary clone name must not exist */
 925         error = zap_lookup(dp->dp_meta_objset,
 926             dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, recv_clone_name,
 927             8, 1, &val);
 928         if (error != ENOENT)
 929                 return (error == 0 ? EBUSY : error);
 930
 931         /* new snapshot name must not exist */
 932         error = zap_lookup(dp->dp_meta_objset,
 933             dsl_dataset_phys(ds)->ds_snapnames_zapobj,
 934             drba->drba_cookie->drc_tosnap, 8, 1, &val);
 935         if (error != ENOENT)
 936                 return (error == 0 ? EEXIST : error);
 937
 938         /*
 939          * Check snapshot limit before receiving. We'll recheck again at the
 940          * end, but might as well abort before receiving if we're already over
 941          * the limit.
 942          *
 943          * Note that we do not check the file system limit with
 944          * dsl_dir_fscount_check because the temporary %clones don't count
 945          * against that limit.
 946          */
 947         error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT,
 948             NULL, drba->drba_cred);
 949         if (error != 0)
 950                 return (error);
 951
 952         if (fromguid != 0) {
 953                 dsl_dataset_t *snap;
 954                 uint64_t obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
 955
 956                 /* Find snapshot in this dir that matches fromguid. */
 957                 while (obj != 0) {
 958                         error = dsl_dataset_hold_obj(dp, obj, FTAG,
 959                             &snap);
 960                         if (error != 0)
 961                                 return (SET_ERROR(ENODEV));
 962                         if (snap->ds_dir != ds->ds_dir) {
 963                                 dsl_dataset_rele(snap, FTAG);
 964                                 return (SET_ERROR(ENODEV));
 965                         }
 966                         if (dsl_dataset_phys(snap)->ds_guid == fromguid)
 967                                 break;
 968                         obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
 969                         dsl_dataset_rele(snap, FTAG);
 970                 }
 971                 if (obj == 0)
 972                         return (SET_ERROR(ENODEV));
 973
 974                 if (drba->drba_cookie->drc_force) {
 975                         drba->drba_snapobj = obj;
 976                 } else {
 977                         /*
 978                          * If we are not forcing, there must be no
 979                          * changes since fromsnap.
 980                          */
 981                         if (dsl_dataset_modified_since_snap(ds, snap)) {
 982                                 dsl_dataset_rele(snap, FTAG);
 983                                 return (SET_ERROR(ETXTBSY));
 984                         }
 985                         drba->drba_snapobj = ds->ds_prev->ds_object;
 986                 }
 987
 988                 dsl_dataset_rele(snap, FTAG);
 989         } else {
 990                 /* if full, then must be forced */
 991                 if (!drba->drba_cookie->drc_force)
 992                         return (SET_ERROR(EEXIST));
 993                 /* start from $ORIGIN@$ORIGIN, if supported */
 994                 drba->drba_snapobj = dp->dp_origin_snap != NULL ?
 995                     dp->dp_origin_snap->ds_object : 0;
 996         }
 997
 998         return (0);
 999
1000 }
1001
1002 static int
1003 dmu_recv_begin_check(void *arg, dmu_tx_t *tx)
1004 {
1005         dmu_recv_begin_arg_t *drba = arg;
1006         dsl_pool_t *dp = dmu_tx_pool(tx);
1007         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1008         uint64_t fromguid = drrb->drr_fromguid;
1009         int flags = drrb->drr_flags;
1010         int error;
1011         uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1012         dsl_dataset_t *ds;
1013         const char *tofs = drba->drba_cookie->drc_tofs;
1014
1015         /* already checked */
1016         ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1017
1018         if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1019             DMU_COMPOUNDSTREAM ||
1020             drrb->drr_type >= DMU_OST_NUMTYPES ||
1021             ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL))
1022                 return (SET_ERROR(EINVAL));
1023
1024         /* Verify pool version supports SA if SA_SPILL feature set */
1025         if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1026             spa_version(dp->dp_spa) < SPA_VERSION_SA)
1027                 return (SET_ERROR(ENOTSUP));
1028
1029         /*
1030          * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1031          * record to a plan WRITE record, so the pool must have the
1032          * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1033          * records.  Same with WRITE_EMBEDDED records that use LZ4 compression.
1034          */
1035         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1036             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1037                 return (SET_ERROR(ENOTSUP));
1038         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4) &&
1039             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1040                 return (SET_ERROR(ENOTSUP));
1041
1042         /*
1043          * The receiving code doesn't know how to translate large blocks
1044          * to smaller ones, so the pool must have the LARGE_BLOCKS
1045          * feature enabled if the stream has LARGE_BLOCKS.
1046          */
1047         if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
1048             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS))
1049                 return (SET_ERROR(ENOTSUP));
1050
1051         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1052         if (error == 0) {
1053                 /* target fs already exists; recv into temp clone */
1054
1055                 /* Can't recv a clone into an existing fs */
1056                 if (flags & DRR_FLAG_CLONE) {
1057                         dsl_dataset_rele(ds, FTAG);
1058                         return (SET_ERROR(EINVAL));
1059                 }
1060
1061                 error = recv_begin_check_existing_impl(drba, ds, fromguid);
1062                 dsl_dataset_rele(ds, FTAG);
1063         } else if (error == ENOENT) {
1064                 /* target fs does not exist; must be a full backup or clone */
1065                 char buf[MAXNAMELEN];
1066
1067                 /*
1068                  * If it's a non-clone incremental, we are missing the
1069                  * target fs, so fail the recv.
1070                  */
1071                 if (fromguid != 0 && !(flags & DRR_FLAG_CLONE))
1072                         return (SET_ERROR(ENOENT));
1073
1074                 /* Open the parent of tofs */
1075                 ASSERT3U(strlen(tofs), <, MAXNAMELEN);
1076                 (void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1);
1077                 error = dsl_dataset_hold(dp, buf, FTAG, &ds);
1078                 if (error != 0)
1079                         return (error);
1080
1081                 /*
1082                  * Check filesystem and snapshot limits before receiving. We'll
1083                  * recheck snapshot limits again at the end (we create the
1084                  * filesystems and increment those counts during begin_sync).
1085                  */
1086                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1087                     ZFS_PROP_FILESYSTEM_LIMIT, NULL, drba->drba_cred);
1088                 if (error != 0) {
1089                         dsl_dataset_rele(ds, FTAG);
1090                         return (error);
1091                 }
1092
1093                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1094                     ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred);
1095                 if (error != 0) {
1096                         dsl_dataset_rele(ds, FTAG);
1097                         return (error);
1098                 }
1099
1100                 if (drba->drba_origin != NULL) {
1101                         dsl_dataset_t *origin;
1102                         error = dsl_dataset_hold(dp, drba->drba_origin,
1103                             FTAG, &origin);
1104                         if (error != 0) {
1105                                 dsl_dataset_rele(ds, FTAG);
1106                                 return (error);
1107                         }
1108                         if (!dsl_dataset_is_snapshot(origin)) {
1109                                 dsl_dataset_rele(origin, FTAG);
1110                                 dsl_dataset_rele(ds, FTAG);
1111                                 return (SET_ERROR(EINVAL));
1112                         }
1113                         if (dsl_dataset_phys(origin)->ds_guid != fromguid) {
1114                                 dsl_dataset_rele(origin, FTAG);
1115                                 dsl_dataset_rele(ds, FTAG);
1116                                 return (SET_ERROR(ENODEV));
1117                         }
1118                         dsl_dataset_rele(origin, FTAG);
1119                 }
1120                 dsl_dataset_rele(ds, FTAG);
1121                 error = 0;
1122         }
1123         return (error);
1124 }
1125
1126 static void
1127 dmu_recv_begin_sync(void *arg, dmu_tx_t *tx)
1128 {
1129         dmu_recv_begin_arg_t *drba = arg;
1130         dsl_pool_t *dp = dmu_tx_pool(tx);
1131         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1132         const char *tofs = drba->drba_cookie->drc_tofs;
1133         dsl_dataset_t *ds, *newds;
1134         uint64_t dsobj;
1135         int error;
1136         uint64_t crflags;
1137
1138         crflags = (drrb->drr_flags & DRR_FLAG_CI_DATA) ?
1139             DS_FLAG_CI_DATASET : 0;
1140
1141         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1142         if (error == 0) {
1143                 /* create temporary clone */
1144                 dsl_dataset_t *snap = NULL;
1145                 if (drba->drba_snapobj != 0) {
1146                         VERIFY0(dsl_dataset_hold_obj(dp,
1147                             drba->drba_snapobj, FTAG, &snap));
1148                 }
1149                 dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name,
1150                     snap, crflags, drba->drba_cred, tx);
1151                 if (drba->drba_snapobj != 0)
1152                         dsl_dataset_rele(snap, FTAG);
1153                 dsl_dataset_rele(ds, FTAG);
1154         } else {
1155                 dsl_dir_t *dd;
1156                 const char *tail;
1157                 dsl_dataset_t *origin = NULL;
1158
1159                 VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail));
1160
1161                 if (drba->drba_origin != NULL) {
1162                         VERIFY0(dsl_dataset_hold(dp, drba->drba_origin,
1163                             FTAG, &origin));
1164                 }
1165
1166                 /* Create new dataset. */
1167                 dsobj = dsl_dataset_create_sync(dd,
1168                     strrchr(tofs, '/') + 1,
1169                     origin, crflags, drba->drba_cred, tx);
1170                 if (origin != NULL)
1171                         dsl_dataset_rele(origin, FTAG);
1172                 dsl_dir_rele(dd, FTAG);
1173                 drba->drba_cookie->drc_newfs = B_TRUE;
1174         }
1175         VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds));
1176
1177         if ((DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
1178             DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
1179             !newds->ds_large_blocks) {
1180                 dsl_dataset_activate_large_blocks_sync_impl(dsobj, tx);
1181                 newds->ds_large_blocks = B_TRUE;
1182         }
1183
1184         dmu_buf_will_dirty(newds->ds_dbuf, tx);
1185         dsl_dataset_phys(newds)->ds_flags |= DS_FLAG_INCONSISTENT;
1186
1187         /*
1188          * If we actually created a non-clone, we need to create the
1189          * objset in our new dataset.
1190          */
1191         if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) {
1192                 (void) dmu_objset_create_impl(dp->dp_spa,
1193                     newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx);
1194         }
1195
1196         drba->drba_cookie->drc_ds = newds;
1197
1198         spa_history_log_internal_ds(newds, "receive", tx, "");
1199 }
1200
1201 /*
1202  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
1203  * succeeds; otherwise we will leak the holds on the datasets.
1204  */
1205 int
1206 dmu_recv_begin(char *tofs, char *tosnap, struct drr_begin *drrb,
1207     boolean_t force, char *origin, dmu_recv_cookie_t *drc)
1208 {
1209         dmu_recv_begin_arg_t drba = { 0 };
1210         dmu_replay_record_t *drr;
1211
1212         bzero(drc, sizeof (dmu_recv_cookie_t));
1213         drc->drc_drrb = drrb;
1214         drc->drc_tosnap = tosnap;
1215         drc->drc_tofs = tofs;
1216         drc->drc_force = force;
1217         drc->drc_cred = CRED();
1218
1219         if (drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
1220                 drc->drc_byteswap = B_TRUE;
1221         else if (drrb->drr_magic != DMU_BACKUP_MAGIC)
1222                 return (SET_ERROR(EINVAL));
1223
1224         drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
1225         drr->drr_type = DRR_BEGIN;
1226         drr->drr_u.drr_begin = *drc->drc_drrb;
1227         if (drc->drc_byteswap) {
1228                 fletcher_4_incremental_byteswap(drr,
1229                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1230         } else {
1231                 fletcher_4_incremental_native(drr,
1232                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1233         }
1234         kmem_free(drr, sizeof (dmu_replay_record_t));
1235
1236         if (drc->drc_byteswap) {
1237                 drrb->drr_magic = BSWAP_64(drrb->drr_magic);
1238                 drrb->drr_versioninfo = BSWAP_64(drrb->drr_versioninfo);
1239                 drrb->drr_creation_time = BSWAP_64(drrb->drr_creation_time);
1240                 drrb->drr_type = BSWAP_32(drrb->drr_type);
1241                 drrb->drr_toguid = BSWAP_64(drrb->drr_toguid);
1242                 drrb->drr_fromguid = BSWAP_64(drrb->drr_fromguid);
1243         }
1244
1245         drba.drba_origin = origin;
1246         drba.drba_cookie = drc;
1247         drba.drba_cred = CRED();
1248
1249         return (dsl_sync_task(tofs, dmu_recv_begin_check, dmu_recv_begin_sync,
1250             &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1251 }
1252
1253 struct restorearg {
1254         int err;
1255         boolean_t byteswap;
1256         kthread_t *td;
1257         struct file *fp;
1258         char *buf;
1259         uint64_t voff;
1260         int bufsize; /* amount of memory allocated for buf */
1261         zio_cksum_t cksum;
1262         avl_tree_t *guid_to_ds_map;
1263 };
1264
1265 typedef struct guid_map_entry {
1266         uint64_t        guid;
1267         dsl_dataset_t   *gme_ds;
1268         avl_node_t      avlnode;
1269 } guid_map_entry_t;
1270
1271 static int
1272 guid_compare(const void *arg1, const void *arg2)
1273 {
1274         const guid_map_entry_t *gmep1 = arg1;
1275         const guid_map_entry_t *gmep2 = arg2;
1276
1277         if (gmep1->guid < gmep2->guid)
1278                 return (-1);
1279         else if (gmep1->guid > gmep2->guid)
1280                 return (1);
1281         return (0);
1282 }
1283
1284 static void
1285 free_guid_map_onexit(void *arg)
1286 {
1287         avl_tree_t *ca = arg;
1288         void *cookie = NULL;
1289         guid_map_entry_t *gmep;
1290
1291         while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
1292                 dsl_dataset_long_rele(gmep->gme_ds, gmep);
1293                 dsl_dataset_rele(gmep->gme_ds, gmep);
1294                 kmem_free(gmep, sizeof (guid_map_entry_t));
1295         }
1296         avl_destroy(ca);
1297         kmem_free(ca, sizeof (avl_tree_t));
1298 }
1299
1300 static int
1301 restore_bytes(struct restorearg *ra, void *buf, int len, off_t off, ssize_t *resid)
1302 {
1303         struct uio auio;
1304         struct iovec aiov;
1305         int error;
1306
1307         aiov.iov_base = buf;
1308         aiov.iov_len = len;
1309         auio.uio_iov = &aiov;
1310         auio.uio_iovcnt = 1;
1311         auio.uio_resid = len;
1312         auio.uio_segflg = UIO_SYSSPACE;
1313         auio.uio_rw = UIO_READ;
1314         auio.uio_offset = off;
1315         auio.uio_td = ra->td;
1316 #ifdef _KERNEL
1317         error = fo_read(ra->fp, &auio, ra->td->td_ucred, FOF_OFFSET, ra->td);
1318 #else
1319         fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__);
1320         error = EOPNOTSUPP;
1321 #endif
1322         *resid = auio.uio_resid;
1323         return (error);
1324 }
1325
1326 static void *
1327 restore_read(struct restorearg *ra, int len, char *buf)
1328 {
1329         int done = 0;
1330
1331         if (buf == NULL)
1332                 buf = ra->buf;
1333
1334         /* some things will require 8-byte alignment, so everything must */
1335         ASSERT0(len % 8);
1336         ASSERT3U(len, <=, ra->bufsize);
1337
1338         while (done < len) {
1339                 ssize_t resid;
1340
1341                 ra->err = restore_bytes(ra, buf + done,
1342                     len - done, ra->voff, &resid);
1343
1344                 if (resid == len - done)
1345                         ra->err = SET_ERROR(EINVAL);
1346                 ra->voff += len - done - resid;
1347                 done = len - resid;
1348                 if (ra->err != 0)
1349                         return (NULL);
1350         }
1351
1352         ASSERT3U(done, ==, len);
1353         if (ra->byteswap)
1354                 fletcher_4_incremental_byteswap(buf, len, &ra->cksum);
1355         else
1356                 fletcher_4_incremental_native(buf, len, &ra->cksum);
1357         return (buf);
1358 }
1359
1360 static void
1361 backup_byteswap(dmu_replay_record_t *drr)
1362 {
1363 #define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
1364 #define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
1365         drr->drr_type = BSWAP_32(drr->drr_type);
1366         drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
1367         switch (drr->drr_type) {
1368         case DRR_BEGIN:
1369                 DO64(drr_begin.drr_magic);
1370                 DO64(drr_begin.drr_versioninfo);
1371                 DO64(drr_begin.drr_creation_time);
1372                 DO32(drr_begin.drr_type);
1373                 DO32(drr_begin.drr_flags);
1374                 DO64(drr_begin.drr_toguid);
1375                 DO64(drr_begin.drr_fromguid);
1376                 break;
1377         case DRR_OBJECT:
1378                 DO64(drr_object.drr_object);
1379                 DO32(drr_object.drr_type);
1380                 DO32(drr_object.drr_bonustype);
1381                 DO32(drr_object.drr_blksz);
1382                 DO32(drr_object.drr_bonuslen);
1383                 DO64(drr_object.drr_toguid);
1384                 break;
1385         case DRR_FREEOBJECTS:
1386                 DO64(drr_freeobjects.drr_firstobj);
1387                 DO64(drr_freeobjects.drr_numobjs);
1388                 DO64(drr_freeobjects.drr_toguid);
1389                 break;
1390         case DRR_WRITE:
1391                 DO64(drr_write.drr_object);
1392                 DO32(drr_write.drr_type);
1393                 DO64(drr_write.drr_offset);
1394                 DO64(drr_write.drr_length);
1395                 DO64(drr_write.drr_toguid);
1396                 DO64(drr_write.drr_key.ddk_cksum.zc_word[0]);
1397                 DO64(drr_write.drr_key.ddk_cksum.zc_word[1]);
1398                 DO64(drr_write.drr_key.ddk_cksum.zc_word[2]);
1399                 DO64(drr_write.drr_key.ddk_cksum.zc_word[3]);
1400                 DO64(drr_write.drr_key.ddk_prop);
1401                 break;
1402         case DRR_WRITE_BYREF:
1403                 DO64(drr_write_byref.drr_object);
1404                 DO64(drr_write_byref.drr_offset);
1405                 DO64(drr_write_byref.drr_length);
1406                 DO64(drr_write_byref.drr_toguid);
1407                 DO64(drr_write_byref.drr_refguid);
1408                 DO64(drr_write_byref.drr_refobject);
1409                 DO64(drr_write_byref.drr_refoffset);
1410                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[0]);
1411                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[1]);
1412                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[2]);
1413                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[3]);
1414                 DO64(drr_write_byref.drr_key.ddk_prop);
1415                 break;
1416         case DRR_WRITE_EMBEDDED:
1417                 DO64(drr_write_embedded.drr_object);
1418                 DO64(drr_write_embedded.drr_offset);
1419                 DO64(drr_write_embedded.drr_length);
1420                 DO64(drr_write_embedded.drr_toguid);
1421                 DO32(drr_write_embedded.drr_lsize);
1422                 DO32(drr_write_embedded.drr_psize);
1423                 break;
1424         case DRR_FREE:
1425                 DO64(drr_free.drr_object);
1426                 DO64(drr_free.drr_offset);
1427                 DO64(drr_free.drr_length);
1428                 DO64(drr_free.drr_toguid);
1429                 break;
1430         case DRR_SPILL:
1431                 DO64(drr_spill.drr_object);
1432                 DO64(drr_spill.drr_length);
1433                 DO64(drr_spill.drr_toguid);
1434                 break;
1435         case DRR_END:
1436                 DO64(drr_end.drr_checksum.zc_word[0]);
1437                 DO64(drr_end.drr_checksum.zc_word[1]);
1438                 DO64(drr_end.drr_checksum.zc_word[2]);
1439                 DO64(drr_end.drr_checksum.zc_word[3]);
1440                 DO64(drr_end.drr_toguid);
1441                 break;
1442         }
1443 #undef DO64
1444 #undef DO32
1445 }
1446
1447 static inline uint8_t
1448 deduce_nblkptr(dmu_object_type_t bonus_type, uint64_t bonus_size)
1449 {
1450         if (bonus_type == DMU_OT_SA) {
1451                 return (1);
1452         } else {
1453                 return (1 +
1454                     ((DN_MAX_BONUSLEN - bonus_size) >> SPA_BLKPTRSHIFT));
1455         }
1456 }
1457
1458 static int
1459 restore_object(struct restorearg *ra, objset_t *os, struct drr_object *drro)
1460 {
1461         dmu_object_info_t doi;
1462         dmu_tx_t *tx;
1463         void *data = NULL;
1464         uint64_t object;
1465         int err;
1466
1467         if (drro->drr_type == DMU_OT_NONE ||
1468             !DMU_OT_IS_VALID(drro->drr_type) ||
1469             !DMU_OT_IS_VALID(drro->drr_bonustype) ||
1470             drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
1471             drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
1472             P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
1473             drro->drr_blksz < SPA_MINBLOCKSIZE ||
1474             drro->drr_blksz > spa_maxblocksize(dmu_objset_spa(os)) ||
1475             drro->drr_bonuslen > DN_MAX_BONUSLEN) {
1476                 return (SET_ERROR(EINVAL));
1477         }
1478
1479         err = dmu_object_info(os, drro->drr_object, &doi);
1480
1481         if (err != 0 && err != ENOENT)
1482                 return (SET_ERROR(EINVAL));
1483         object = err == 0 ? drro->drr_object : DMU_NEW_OBJECT;
1484
1485         if (drro->drr_bonuslen) {
1486                 data = restore_read(ra, P2ROUNDUP(drro->drr_bonuslen, 8), NULL);
1487                 if (ra->err != 0)
1488                         return (ra->err);
1489         }
1490
1491         /*
1492          * If we are losing blkptrs or changing the block size this must
1493          * be a new file instance.  We must clear out the previous file
1494          * contents before we can change this type of metadata in the dnode.
1495          */
1496         if (err == 0) {
1497                 int nblkptr;
1498
1499                 nblkptr = deduce_nblkptr(drro->drr_bonustype,
1500                     drro->drr_bonuslen);
1501
1502                 if (drro->drr_blksz != doi.doi_data_block_size ||
1503                     nblkptr < doi.doi_nblkptr) {
1504                         err = dmu_free_long_range(os, drro->drr_object,
1505                             0, DMU_OBJECT_END);
1506                         if (err != 0)
1507                                 return (SET_ERROR(EINVAL));
1508                 }
1509         }
1510
1511         tx = dmu_tx_create(os);
1512         dmu_tx_hold_bonus(tx, object);
1513         err = dmu_tx_assign(tx, TXG_WAIT);
1514         if (err != 0) {
1515                 dmu_tx_abort(tx);
1516                 return (err);
1517         }
1518
1519         if (object == DMU_NEW_OBJECT) {
1520                 /* currently free, want to be allocated */
1521                 err = dmu_object_claim(os, drro->drr_object,
1522                     drro->drr_type, drro->drr_blksz,
1523                     drro->drr_bonustype, drro->drr_bonuslen, tx);
1524         } else if (drro->drr_type != doi.doi_type ||
1525             drro->drr_blksz != doi.doi_data_block_size ||
1526             drro->drr_bonustype != doi.doi_bonus_type ||
1527             drro->drr_bonuslen != doi.doi_bonus_size) {
1528                 /* currently allocated, but with different properties */
1529                 err = dmu_object_reclaim(os, drro->drr_object,
1530                     drro->drr_type, drro->drr_blksz,
1531                     drro->drr_bonustype, drro->drr_bonuslen, tx);
1532         }
1533         if (err != 0) {
1534                 dmu_tx_commit(tx);
1535                 return (SET_ERROR(EINVAL));
1536         }
1537
1538         dmu_object_set_checksum(os, drro->drr_object, drro->drr_checksumtype,
1539             tx);
1540         dmu_object_set_compress(os, drro->drr_object, drro->drr_compress, tx);
1541
1542         if (data != NULL) {
1543                 dmu_buf_t *db;
1544
1545                 VERIFY(0 == dmu_bonus_hold(os, drro->drr_object, FTAG, &db));
1546                 dmu_buf_will_dirty(db, tx);
1547
1548                 ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
1549                 bcopy(data, db->db_data, drro->drr_bonuslen);
1550                 if (ra->byteswap) {
1551                         dmu_object_byteswap_t byteswap =
1552                             DMU_OT_BYTESWAP(drro->drr_bonustype);
1553                         dmu_ot_byteswap[byteswap].ob_func(db->db_data,
1554                             drro->drr_bonuslen);
1555                 }
1556                 dmu_buf_rele(db, FTAG);
1557         }
1558         dmu_tx_commit(tx);
1559         return (0);
1560 }
1561
1562 /* ARGSUSED */
1563 static int
1564 restore_freeobjects(struct restorearg *ra, objset_t *os,
1565     struct drr_freeobjects *drrfo)
1566 {
1567         uint64_t obj;
1568
1569         if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
1570                 return (SET_ERROR(EINVAL));
1571
1572         for (obj = drrfo->drr_firstobj;
1573             obj < drrfo->drr_firstobj + drrfo->drr_numobjs;
1574             (void) dmu_object_next(os, &obj, FALSE, 0)) {
1575                 int err;
1576
1577                 if (dmu_object_info(os, obj, NULL) != 0)
1578                         continue;
1579
1580                 err = dmu_free_long_object(os, obj);
1581                 if (err != 0)
1582                         return (err);
1583         }
1584         return (0);
1585 }
1586
1587 static int
1588 restore_write(struct restorearg *ra, objset_t *os,
1589     struct drr_write *drrw)
1590 {
1591         dmu_tx_t *tx;
1592         void *data;
1593         int err;
1594
1595         if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
1596             !DMU_OT_IS_VALID(drrw->drr_type))
1597                 return (SET_ERROR(EINVAL));
1598
1599         if (dmu_object_info(os, drrw->drr_object, NULL) != 0)
1600                 return (SET_ERROR(EINVAL));
1601
1602         dmu_buf_t *bonus;
1603         if (dmu_bonus_hold(os, drrw->drr_object, FTAG, &bonus) != 0)
1604                 return (SET_ERROR(EINVAL));
1605
1606         arc_buf_t *abuf = dmu_request_arcbuf(bonus, drrw->drr_length);
1607
1608         data = restore_read(ra, drrw->drr_length, abuf->b_data);
1609         if (data == NULL) {
1610                 dmu_return_arcbuf(abuf);
1611                 dmu_buf_rele(bonus, FTAG);
1612                 return (ra->err);
1613         }
1614
1615         tx = dmu_tx_create(os);
1616
1617         dmu_tx_hold_write(tx, drrw->drr_object,
1618             drrw->drr_offset, drrw->drr_length);
1619         err = dmu_tx_assign(tx, TXG_WAIT);
1620         if (err != 0) {
1621                 dmu_return_arcbuf(abuf);
1622                 dmu_buf_rele(bonus, FTAG);
1623                 dmu_tx_abort(tx);
1624                 return (err);
1625         }
1626         if (ra->byteswap) {
1627                 dmu_object_byteswap_t byteswap =
1628                     DMU_OT_BYTESWAP(drrw->drr_type);
1629                 dmu_ot_byteswap[byteswap].ob_func(data, drrw->drr_length);
1630         }
1631         dmu_assign_arcbuf(bonus, drrw->drr_offset, abuf, tx);
1632         dmu_tx_commit(tx);
1633         dmu_buf_rele(bonus, FTAG);
1634         return (0);
1635 }
1636
1637 /*
1638  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
1639  * streams to refer to a copy of the data that is already on the
1640  * system because it came in earlier in the stream.  This function
1641  * finds the earlier copy of the data, and uses that copy instead of
1642  * data from the stream to fulfill this write.
1643  */
1644 static int
1645 restore_write_byref(struct restorearg *ra, objset_t *os,
1646     struct drr_write_byref *drrwbr)
1647 {
1648         dmu_tx_t *tx;
1649         int err;
1650         guid_map_entry_t gmesrch;
1651         guid_map_entry_t *gmep;
1652         avl_index_t where;
1653         objset_t *ref_os = NULL;
1654         dmu_buf_t *dbp;
1655
1656         if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
1657                 return (SET_ERROR(EINVAL));
1658
1659         /*
1660          * If the GUID of the referenced dataset is different from the
1661          * GUID of the target dataset, find the referenced dataset.
1662          */
1663         if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
1664                 gmesrch.guid = drrwbr->drr_refguid;
1665                 if ((gmep = avl_find(ra->guid_to_ds_map, &gmesrch,
1666                     &where)) == NULL) {
1667                         return (SET_ERROR(EINVAL));
1668                 }
1669                 if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
1670                         return (SET_ERROR(EINVAL));
1671         } else {
1672                 ref_os = os;
1673         }
1674
1675         err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
1676             drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH);
1677         if (err != 0)
1678                 return (err);
1679
1680         tx = dmu_tx_create(os);
1681
1682         dmu_tx_hold_write(tx, drrwbr->drr_object,
1683             drrwbr->drr_offset, drrwbr->drr_length);
1684         err = dmu_tx_assign(tx, TXG_WAIT);
1685         if (err != 0) {
1686                 dmu_tx_abort(tx);
1687                 return (err);
1688         }
1689         dmu_write(os, drrwbr->drr_object,
1690             drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
1691         dmu_buf_rele(dbp, FTAG);
1692         dmu_tx_commit(tx);
1693         return (0);
1694 }
1695
1696 static int
1697 restore_write_embedded(struct restorearg *ra, objset_t *os,
1698     struct drr_write_embedded *drrwnp)
1699 {
1700         dmu_tx_t *tx;
1701         int err;
1702         void *data;
1703
1704         if (drrwnp->drr_offset + drrwnp->drr_length < drrwnp->drr_offset)
1705                 return (EINVAL);
1706
1707         if (drrwnp->drr_psize > BPE_PAYLOAD_SIZE)
1708                 return (EINVAL);
1709
1710         if (drrwnp->drr_etype >= NUM_BP_EMBEDDED_TYPES)
1711                 return (EINVAL);
1712         if (drrwnp->drr_compression >= ZIO_COMPRESS_FUNCTIONS)
1713                 return (EINVAL);
1714
1715         data = restore_read(ra, P2ROUNDUP(drrwnp->drr_psize, 8), NULL);
1716         if (data == NULL)
1717                 return (ra->err);
1718
1719         tx = dmu_tx_create(os);
1720
1721         dmu_tx_hold_write(tx, drrwnp->drr_object,
1722             drrwnp->drr_offset, drrwnp->drr_length);
1723         err = dmu_tx_assign(tx, TXG_WAIT);
1724         if (err != 0) {
1725                 dmu_tx_abort(tx);
1726                 return (err);
1727         }
1728
1729         dmu_write_embedded(os, drrwnp->drr_object,
1730             drrwnp->drr_offset, data, drrwnp->drr_etype,
1731             drrwnp->drr_compression, drrwnp->drr_lsize, drrwnp->drr_psize,
1732             ra->byteswap ^ ZFS_HOST_BYTEORDER, tx);
1733
1734         dmu_tx_commit(tx);
1735         return (0);
1736 }
1737
1738 static int
1739 restore_spill(struct restorearg *ra, objset_t *os, struct drr_spill *drrs)
1740 {
1741         dmu_tx_t *tx;
1742         void *data;
1743         dmu_buf_t *db, *db_spill;
1744         int err;
1745
1746         if (drrs->drr_length < SPA_MINBLOCKSIZE ||
1747             drrs->drr_length > spa_maxblocksize(dmu_objset_spa(os)))
1748                 return (SET_ERROR(EINVAL));
1749
1750         data = restore_read(ra, drrs->drr_length, NULL);
1751         if (data == NULL)
1752                 return (ra->err);
1753
1754         if (dmu_object_info(os, drrs->drr_object, NULL) != 0)
1755                 return (SET_ERROR(EINVAL));
1756
1757         VERIFY(0 == dmu_bonus_hold(os, drrs->drr_object, FTAG, &db));
1758         if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
1759                 dmu_buf_rele(db, FTAG);
1760                 return (err);
1761         }
1762
1763         tx = dmu_tx_create(os);
1764
1765         dmu_tx_hold_spill(tx, db->db_object);
1766
1767         err = dmu_tx_assign(tx, TXG_WAIT);
1768         if (err != 0) {
1769                 dmu_buf_rele(db, FTAG);
1770                 dmu_buf_rele(db_spill, FTAG);
1771                 dmu_tx_abort(tx);
1772                 return (err);
1773         }
1774         dmu_buf_will_dirty(db_spill, tx);
1775
1776         if (db_spill->db_size < drrs->drr_length)
1777                 VERIFY(0 == dbuf_spill_set_blksz(db_spill,
1778                     drrs->drr_length, tx));
1779         bcopy(data, db_spill->db_data, drrs->drr_length);
1780
1781         dmu_buf_rele(db, FTAG);
1782         dmu_buf_rele(db_spill, FTAG);
1783
1784         dmu_tx_commit(tx);
1785         return (0);
1786 }
1787
1788 /* ARGSUSED */
1789 static int
1790 restore_free(struct restorearg *ra, objset_t *os,
1791     struct drr_free *drrf)
1792 {
1793         int err;
1794
1795         if (drrf->drr_length != -1ULL &&
1796             drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
1797                 return (SET_ERROR(EINVAL));
1798
1799         if (dmu_object_info(os, drrf->drr_object, NULL) != 0)
1800                 return (SET_ERROR(EINVAL));
1801
1802         err = dmu_free_long_range(os, drrf->drr_object,
1803             drrf->drr_offset, drrf->drr_length);
1804         return (err);
1805 }
1806
1807 /* used to destroy the drc_ds on error */
1808 static void
1809 dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc)
1810 {
1811         char name[MAXNAMELEN];
1812         dsl_dataset_name(drc->drc_ds, name);
1813         dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
1814         (void) dsl_destroy_head(name);
1815 }
1816
1817 /*
1818  * NB: callers *must* call dmu_recv_end() if this succeeds.
1819  */
1820 int
1821 dmu_recv_stream(dmu_recv_cookie_t *drc, struct file *fp, offset_t *voffp,
1822     int cleanup_fd, uint64_t *action_handlep)
1823 {
1824         struct restorearg ra = { 0 };
1825         dmu_replay_record_t *drr;
1826         objset_t *os;
1827         zio_cksum_t pcksum;
1828         int featureflags;
1829
1830         ra.byteswap = drc->drc_byteswap;
1831         ra.cksum = drc->drc_cksum;
1832         ra.td = curthread;
1833         ra.fp = fp;
1834         ra.voff = *voffp;
1835         ra.bufsize = SPA_MAXBLOCKSIZE;
1836         ra.buf = kmem_alloc(ra.bufsize, KM_SLEEP);
1837
1838         /* these were verified in dmu_recv_begin */
1839         ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==,
1840             DMU_SUBSTREAM);
1841         ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES);
1842
1843         /*
1844          * Open the objset we are modifying.
1845          */
1846         VERIFY0(dmu_objset_from_ds(drc->drc_ds, &os));
1847
1848         ASSERT(dsl_dataset_phys(drc->drc_ds)->ds_flags & DS_FLAG_INCONSISTENT);
1849
1850         featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
1851
1852         /* if this stream is dedup'ed, set up the avl tree for guid mapping */
1853         if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
1854                 minor_t minor;
1855
1856                 if (cleanup_fd == -1) {
1857                         ra.err = SET_ERROR(EBADF);
1858                         goto out;
1859                 }
1860                 ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
1861                 if (ra.err != 0) {
1862                         cleanup_fd = -1;
1863                         goto out;
1864                 }
1865
1866                 if (*action_handlep == 0) {
1867                         ra.guid_to_ds_map =
1868                             kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
1869                         avl_create(ra.guid_to_ds_map, guid_compare,
1870                             sizeof (guid_map_entry_t),
1871                             offsetof(guid_map_entry_t, avlnode));
1872                         ra.err = zfs_onexit_add_cb(minor,
1873                             free_guid_map_onexit, ra.guid_to_ds_map,
1874                             action_handlep);
1875                         if (ra.err != 0)
1876                                 goto out;
1877                 } else {
1878                         ra.err = zfs_onexit_cb_data(minor, *action_handlep,
1879                             (void **)&ra.guid_to_ds_map);
1880                         if (ra.err != 0)
1881                                 goto out;
1882                 }
1883
1884                 drc->drc_guid_to_ds_map = ra.guid_to_ds_map;
1885         }
1886
1887         /*
1888          * Read records and process them.
1889          */
1890         pcksum = ra.cksum;
1891         while (ra.err == 0 &&
1892             NULL != (drr = restore_read(&ra, sizeof (*drr), NULL))) {
1893                 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1894                         ra.err = SET_ERROR(EINTR);
1895                         goto out;
1896                 }
1897
1898                 if (ra.byteswap)
1899                         backup_byteswap(drr);
1900
1901                 switch (drr->drr_type) {
1902                 case DRR_OBJECT:
1903                 {
1904                         /*
1905                          * We need to make a copy of the record header,
1906                          * because restore_{object,write} may need to
1907                          * restore_read(), which will invalidate drr.
1908                          */
1909                         struct drr_object drro = drr->drr_u.drr_object;
1910                         ra.err = restore_object(&ra, os, &drro);
1911                         break;
1912                 }
1913                 case DRR_FREEOBJECTS:
1914                 {
1915                         struct drr_freeobjects drrfo =
1916                             drr->drr_u.drr_freeobjects;
1917                         ra.err = restore_freeobjects(&ra, os, &drrfo);
1918                         break;
1919                 }
1920                 case DRR_WRITE:
1921                 {
1922                         struct drr_write drrw = drr->drr_u.drr_write;
1923                         ra.err = restore_write(&ra, os, &drrw);
1924                         break;
1925                 }
1926                 case DRR_WRITE_BYREF:
1927                 {
1928                         struct drr_write_byref drrwbr =
1929                             drr->drr_u.drr_write_byref;
1930                         ra.err = restore_write_byref(&ra, os, &drrwbr);
1931                         break;
1932                 }
1933                 case DRR_WRITE_EMBEDDED:
1934                 {
1935                         struct drr_write_embedded drrwe =
1936                             drr->drr_u.drr_write_embedded;
1937                         ra.err = restore_write_embedded(&ra, os, &drrwe);
1938                         break;
1939                 }
1940                 case DRR_FREE:
1941                 {
1942                         struct drr_free drrf = drr->drr_u.drr_free;
1943                         ra.err = restore_free(&ra, os, &drrf);
1944                         break;
1945                 }
1946                 case DRR_END:
1947                 {
1948                         struct drr_end drre = drr->drr_u.drr_end;
1949                         /*
1950                          * We compare against the *previous* checksum
1951                          * value, because the stored checksum is of
1952                          * everything before the DRR_END record.
1953                          */
1954                         if (!ZIO_CHECKSUM_EQUAL(drre.drr_checksum, pcksum))
1955                                 ra.err = SET_ERROR(ECKSUM);
1956                         goto out;
1957                 }
1958                 case DRR_SPILL:
1959                 {
1960                         struct drr_spill drrs = drr->drr_u.drr_spill;
1961                         ra.err = restore_spill(&ra, os, &drrs);
1962                         break;
1963                 }
1964                 default:
1965                         ra.err = SET_ERROR(EINVAL);
1966                         goto out;
1967                 }
1968                 pcksum = ra.cksum;
1969         }
1970         ASSERT(ra.err != 0);
1971
1972 out:
1973         if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
1974                 zfs_onexit_fd_rele(cleanup_fd);
1975
1976         if (ra.err != 0) {
1977                 /*
1978                  * destroy what we created, so we don't leave it in the
1979                  * inconsistent restoring state.
1980                  */
1981                 dmu_recv_cleanup_ds(drc);
1982         }
1983
1984         kmem_free(ra.buf, ra.bufsize);
1985         *voffp = ra.voff;
1986         return (ra.err);
1987 }
1988
1989 static int
1990 dmu_recv_end_check(void *arg, dmu_tx_t *tx)
1991 {
1992         dmu_recv_cookie_t *drc = arg;
1993         dsl_pool_t *dp = dmu_tx_pool(tx);
1994         int error;
1995
1996         ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag);
1997
1998         if (!drc->drc_newfs) {
1999                 dsl_dataset_t *origin_head;
2000
2001                 error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head);
2002                 if (error != 0)
2003                         return (error);
2004                 if (drc->drc_force) {
2005                         /*
2006                          * We will destroy any snapshots in tofs (i.e. before
2007                          * origin_head) that are after the origin (which is
2008                          * the snap before drc_ds, because drc_ds can not
2009                          * have any snaps of its own).
2010                          */
2011                         uint64_t obj;
2012
2013                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
2014                         while (obj !=
2015                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
2016                                 dsl_dataset_t *snap;
2017                                 error = dsl_dataset_hold_obj(dp, obj, FTAG,
2018                                     &snap);
2019                                 if (error != 0)
2020                                         break;
2021                                 if (snap->ds_dir != origin_head->ds_dir)
2022                                         error = SET_ERROR(EINVAL);
2023                                 if (error == 0)  {
2024                                         error = dsl_destroy_snapshot_check_impl(
2025                                             snap, B_FALSE);
2026                                 }
2027                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
2028                                 dsl_dataset_rele(snap, FTAG);
2029                                 if (error != 0)
2030                                         break;
2031                         }
2032                         if (error != 0) {
2033                                 dsl_dataset_rele(origin_head, FTAG);
2034                                 return (error);
2035                         }
2036                 }
2037                 error = dsl_dataset_clone_swap_check_impl(drc->drc_ds,
2038                     origin_head, drc->drc_force, drc->drc_owner, tx);
2039                 if (error != 0) {
2040                         dsl_dataset_rele(origin_head, FTAG);
2041                         return (error);
2042                 }
2043                 error = dsl_dataset_snapshot_check_impl(origin_head,
2044                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
2045                 dsl_dataset_rele(origin_head, FTAG);
2046                 if (error != 0)
2047                         return (error);
2048
2049                 error = dsl_destroy_head_check_impl(drc->drc_ds, 1);
2050         } else {
2051                 error = dsl_dataset_snapshot_check_impl(drc->drc_ds,
2052                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
2053         }
2054         return (error);
2055 }
2056
2057 static void
2058 dmu_recv_end_sync(void *arg, dmu_tx_t *tx)
2059 {
2060         dmu_recv_cookie_t *drc = arg;
2061         dsl_pool_t *dp = dmu_tx_pool(tx);
2062
2063         spa_history_log_internal_ds(drc->drc_ds, "finish receiving",
2064             tx, "snap=%s", drc->drc_tosnap);
2065
2066         if (!drc->drc_newfs) {
2067                 dsl_dataset_t *origin_head;
2068
2069                 VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG,
2070                     &origin_head));
2071
2072                 if (drc->drc_force) {
2073                         /*
2074                          * Destroy any snapshots of drc_tofs (origin_head)
2075                          * after the origin (the snap before drc_ds).
2076                          */
2077                         uint64_t obj;
2078
2079                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
2080                         while (obj !=
2081                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
2082                                 dsl_dataset_t *snap;
2083                                 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG,
2084                                     &snap));
2085                                 ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir);
2086                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
2087                                 dsl_destroy_snapshot_sync_impl(snap,
2088                                     B_FALSE, tx);
2089                                 dsl_dataset_rele(snap, FTAG);
2090                         }
2091                 }
2092                 VERIFY3P(drc->drc_ds->ds_prev, ==,
2093                     origin_head->ds_prev);
2094
2095                 dsl_dataset_clone_swap_sync_impl(drc->drc_ds,
2096                     origin_head, tx);
2097                 dsl_dataset_snapshot_sync_impl(origin_head,
2098                     drc->drc_tosnap, tx);
2099
2100                 /* set snapshot's creation time and guid */
2101                 dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx);
2102                 dsl_dataset_phys(origin_head->ds_prev)->ds_creation_time =
2103                     drc->drc_drrb->drr_creation_time;
2104                 dsl_dataset_phys(origin_head->ds_prev)->ds_guid =
2105                     drc->drc_drrb->drr_toguid;
2106                 dsl_dataset_phys(origin_head->ds_prev)->ds_flags &=
2107                     ~DS_FLAG_INCONSISTENT;
2108
2109                 dmu_buf_will_dirty(origin_head->ds_dbuf, tx);
2110                 dsl_dataset_phys(origin_head)->ds_flags &=
2111                     ~DS_FLAG_INCONSISTENT;
2112
2113                 dsl_dataset_rele(origin_head, FTAG);
2114                 dsl_destroy_head_sync_impl(drc->drc_ds, tx);
2115
2116                 if (drc->drc_owner != NULL)
2117                         VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner);
2118         } else {
2119                 dsl_dataset_t *ds = drc->drc_ds;
2120
2121                 dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx);
2122
2123                 /* set snapshot's creation time and guid */
2124                 dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
2125                 dsl_dataset_phys(ds->ds_prev)->ds_creation_time =
2126                     drc->drc_drrb->drr_creation_time;
2127                 dsl_dataset_phys(ds->ds_prev)->ds_guid =
2128                     drc->drc_drrb->drr_toguid;
2129                 dsl_dataset_phys(ds->ds_prev)->ds_flags &=
2130                     ~DS_FLAG_INCONSISTENT;
2131
2132                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
2133                 dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
2134         }
2135         drc->drc_newsnapobj = dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj;
2136         /*
2137          * Release the hold from dmu_recv_begin.  This must be done before
2138          * we return to open context, so that when we free the dataset's dnode,
2139          * we can evict its bonus buffer.
2140          */
2141         dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2142         drc->drc_ds = NULL;
2143 }
2144
2145 static int
2146 add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj)
2147 {
2148         dsl_pool_t *dp;
2149         dsl_dataset_t *snapds;
2150         guid_map_entry_t *gmep;
2151         int err;
2152
2153         ASSERT(guid_map != NULL);
2154
2155         err = dsl_pool_hold(name, FTAG, &dp);
2156         if (err != 0)
2157                 return (err);
2158         gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP);
2159         err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds);
2160         if (err == 0) {
2161                 gmep->guid = dsl_dataset_phys(snapds)->ds_guid;
2162                 gmep->gme_ds = snapds;
2163                 avl_add(guid_map, gmep);
2164                 dsl_dataset_long_hold(snapds, gmep);
2165         } else
2166                 kmem_free(gmep, sizeof (*gmep));
2167
2168         dsl_pool_rele(dp, FTAG);
2169         return (err);
2170 }
2171
2172 static int dmu_recv_end_modified_blocks = 3;
2173
2174 static int
2175 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
2176 {
2177         int error;
2178         char name[MAXNAMELEN];
2179
2180 #ifdef _KERNEL
2181         /*
2182          * We will be destroying the ds; make sure its origin is unmounted if
2183          * necessary.
2184          */
2185         dsl_dataset_name(drc->drc_ds, name);
2186         zfs_destroy_unmount_origin(name);
2187 #endif
2188
2189         error = dsl_sync_task(drc->drc_tofs,
2190             dmu_recv_end_check, dmu_recv_end_sync, drc,
2191             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL);
2192
2193         if (error != 0)
2194                 dmu_recv_cleanup_ds(drc);
2195         return (error);
2196 }
2197
2198 static int
2199 dmu_recv_new_end(dmu_recv_cookie_t *drc)
2200 {
2201         int error;
2202
2203         error = dsl_sync_task(drc->drc_tofs,
2204             dmu_recv_end_check, dmu_recv_end_sync, drc,
2205             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL);
2206
2207         if (error != 0) {
2208                 dmu_recv_cleanup_ds(drc);
2209         } else if (drc->drc_guid_to_ds_map != NULL) {
2210                 (void) add_ds_to_guidmap(drc->drc_tofs,
2211                     drc->drc_guid_to_ds_map,
2212                     drc->drc_newsnapobj);
2213         }
2214         return (error);
2215 }
2216
2217 int
2218 dmu_recv_end(dmu_recv_cookie_t *drc, void *owner)
2219 {
2220         drc->drc_owner = owner;
2221
2222         if (drc->drc_newfs)
2223                 return (dmu_recv_new_end(drc));
2224         else
2225                 return (dmu_recv_existing_end(drc));
2226 }
2227
2228 /*
2229  * Return TRUE if this objset is currently being received into.
2230  */
2231 boolean_t
2232 dmu_objset_is_receiving(objset_t *os)
2233 {
2234         return (os->os_dsl_dataset != NULL &&
2235             os->os_dsl_dataset->ds_owner == dmu_recv_tag);
2236 }