module/zfs/dmu_send.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2011 by Delphix. All rights reserved.
  24  * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
  25  * Copyright (c) 2011 by Delphix. All rights reserved.
  26  */
  27
  28 #include <sys/dmu.h>
  29 #include <sys/dmu_impl.h>
  30 #include <sys/dmu_tx.h>
  31 #include <sys/dbuf.h>
  32 #include <sys/dnode.h>
  33 #include <sys/zfs_context.h>
  34 #include <sys/dmu_objset.h>
  35 #include <sys/dmu_traverse.h>
  36 #include <sys/dsl_dataset.h>
  37 #include <sys/dsl_dir.h>
  38 #include <sys/dsl_prop.h>
  39 #include <sys/dsl_pool.h>
  40 #include <sys/dsl_synctask.h>
  41 #include <sys/zfs_ioctl.h>
  42 #include <sys/zap.h>
  43 #include <sys/zio_checksum.h>
  44 #include <sys/zfs_znode.h>
  45 #include <zfs_fletcher.h>
  46 #include <sys/avl.h>
  47 #include <sys/ddt.h>
  48 #include <sys/zfs_onexit.h>
  49
  50 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
  51 int zfs_send_corrupt_data = B_FALSE;
  52
  53 static char *dmu_recv_tag = "dmu_recv_tag";
  54
  55 /*
  56  * The list of data whose inclusion in a send stream can be pending from
  57  * one call to backup_cb to another.  Multiple calls to dump_free() and
  58  * dump_freeobjects() can be aggregated into a single DRR_FREE or
  59  * DRR_FREEOBJECTS replay record.
  60  */
  61 typedef enum {
  62         PENDING_NONE,
  63         PENDING_FREE,
  64         PENDING_FREEOBJECTS
  65 } pendop_t;
  66
  67 struct backuparg {
  68         dmu_replay_record_t *drr;
  69         vnode_t *vp;
  70         offset_t *off;
  71         objset_t *os;
  72         zio_cksum_t zc;
  73         uint64_t toguid;
  74         int err;
  75         pendop_t pending_op;
  76 };
  77
  78 static int
  79 dump_bytes(struct backuparg *ba, void *buf, int len)
  80 {
  81         ssize_t resid; /* have to get resid to get detailed errno */
  82         ASSERT3U(len % 8, ==, 0);
  83
  84         fletcher_4_incremental_native(buf, len, &ba->zc);
  85         ba->err = vn_rdwr(UIO_WRITE, ba->vp,
  86             (caddr_t)buf, len,
  87             0, UIO_SYSSPACE, FAPPEND, RLIM64_INFINITY, CRED(), &resid);
  88         *ba->off += len;
  89         return (ba->err);
  90 }
  91
  92 static int
  93 dump_free(struct backuparg *ba, uint64_t object, uint64_t offset,
  94     uint64_t length)
  95 {
  96         struct drr_free *drrf = &(ba->drr->drr_u.drr_free);
  97
  98         if (length != -1ULL && offset + length < offset)
  99                 length = -1ULL;
 100
 101         /*
 102          * If there is a pending op, but it's not PENDING_FREE, push it out,
 103          * since free block aggregation can only be done for blocks of the
 104          * same type (i.e., DRR_FREE records can only be aggregated with
 105          * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
 106          * aggregated with other DRR_FREEOBJECTS records.
 107          */
 108         if (ba->pending_op != PENDING_NONE && ba->pending_op != PENDING_FREE) {
 109                 if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
 110                         return (EINTR);
 111                 ba->pending_op = PENDING_NONE;
 112         }
 113
 114         if (ba->pending_op == PENDING_FREE) {
 115                 /*
 116                  * There should never be a PENDING_FREE if length is -1
 117                  * (because dump_dnode is the only place where this
 118                  * function is called with a -1, and only after flushing
 119                  * any pending record).
 120                  */
 121                 ASSERT(length != -1ULL);
 122                 /*
 123                  * Check to see whether this free block can be aggregated
 124                  * with pending one.
 125                  */
 126                 if (drrf->drr_object == object && drrf->drr_offset +
 127                     drrf->drr_length == offset) {
 128                         drrf->drr_length += length;
 129                         return (0);
 130                 } else {
 131                         /* not a continuation.  Push out pending record */
 132                         if (dump_bytes(ba, ba->drr,
 133                             sizeof (dmu_replay_record_t)) != 0)
 134                                 return (EINTR);
 135                         ba->pending_op = PENDING_NONE;
 136                 }
 137         }
 138         /* create a FREE record and make it pending */
 139         bzero(ba->drr, sizeof (dmu_replay_record_t));
 140         ba->drr->drr_type = DRR_FREE;
 141         drrf->drr_object = object;
 142         drrf->drr_offset = offset;
 143         drrf->drr_length = length;
 144         drrf->drr_toguid = ba->toguid;
 145         if (length == -1ULL) {
 146                 if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
 147                         return (EINTR);
 148         } else {
 149                 ba->pending_op = PENDING_FREE;
 150         }
 151
 152         return (0);
 153 }
 154
 155 static int
 156 dump_data(struct backuparg *ba, dmu_object_type_t type,
 157     uint64_t object, uint64_t offset, int blksz, const blkptr_t *bp, void *data)
 158 {
 159         struct drr_write *drrw = &(ba->drr->drr_u.drr_write);
 160
 161
 162         /*
 163          * If there is any kind of pending aggregation (currently either
 164          * a grouping of free objects or free blocks), push it out to
 165          * the stream, since aggregation can't be done across operations
 166          * of different types.
 167          */
 168         if (ba->pending_op != PENDING_NONE) {
 169                 if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
 170                         return (EINTR);
 171                 ba->pending_op = PENDING_NONE;
 172         }
 173         /* write a DATA record */
 174         bzero(ba->drr, sizeof (dmu_replay_record_t));
 175         ba->drr->drr_type = DRR_WRITE;
 176         drrw->drr_object = object;
 177         drrw->drr_type = type;
 178         drrw->drr_offset = offset;
 179         drrw->drr_length = blksz;
 180         drrw->drr_toguid = ba->toguid;
 181         drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
 182         if (zio_checksum_table[drrw->drr_checksumtype].ci_dedup)
 183                 drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
 184         DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
 185         DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
 186         DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
 187         drrw->drr_key.ddk_cksum = bp->blk_cksum;
 188
 189         if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
 190                 return (EINTR);
 191         if (dump_bytes(ba, data, blksz) != 0)
 192                 return (EINTR);
 193         return (0);
 194 }
 195
 196 static int
 197 dump_spill(struct backuparg *ba, uint64_t object, int blksz, void *data)
 198 {
 199         struct drr_spill *drrs = &(ba->drr->drr_u.drr_spill);
 200
 201         if (ba->pending_op != PENDING_NONE) {
 202                 if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
 203                         return (EINTR);
 204                 ba->pending_op = PENDING_NONE;
 205         }
 206
 207         /* write a SPILL record */
 208         bzero(ba->drr, sizeof (dmu_replay_record_t));
 209         ba->drr->drr_type = DRR_SPILL;
 210         drrs->drr_object = object;
 211         drrs->drr_length = blksz;
 212         drrs->drr_toguid = ba->toguid;
 213
 214         if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)))
 215                 return (EINTR);
 216         if (dump_bytes(ba, data, blksz))
 217                 return (EINTR);
 218         return (0);
 219 }
 220
 221 static int
 222 dump_freeobjects(struct backuparg *ba, uint64_t firstobj, uint64_t numobjs)
 223 {
 224         struct drr_freeobjects *drrfo = &(ba->drr->drr_u.drr_freeobjects);
 225
 226         /*
 227          * If there is a pending op, but it's not PENDING_FREEOBJECTS,
 228          * push it out, since free block aggregation can only be done for
 229          * blocks of the same type (i.e., DRR_FREE records can only be
 230          * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
 231          * can only be aggregated with other DRR_FREEOBJECTS records.
 232          */
 233         if (ba->pending_op != PENDING_NONE &&
 234             ba->pending_op != PENDING_FREEOBJECTS) {
 235                 if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
 236                         return (EINTR);
 237                 ba->pending_op = PENDING_NONE;
 238         }
 239         if (ba->pending_op == PENDING_FREEOBJECTS) {
 240                 /*
 241                  * See whether this free object array can be aggregated
 242                  * with pending one
 243                  */
 244                 if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
 245                         drrfo->drr_numobjs += numobjs;
 246                         return (0);
 247                 } else {
 248                         /* can't be aggregated.  Push out pending record */
 249                         if (dump_bytes(ba, ba->drr,
 250                             sizeof (dmu_replay_record_t)) != 0)
 251                                 return (EINTR);
 252                         ba->pending_op = PENDING_NONE;
 253                 }
 254         }
 255
 256         /* write a FREEOBJECTS record */
 257         bzero(ba->drr, sizeof (dmu_replay_record_t));
 258         ba->drr->drr_type = DRR_FREEOBJECTS;
 259         drrfo->drr_firstobj = firstobj;
 260         drrfo->drr_numobjs = numobjs;
 261         drrfo->drr_toguid = ba->toguid;
 262
 263         ba->pending_op = PENDING_FREEOBJECTS;
 264
 265         return (0);
 266 }
 267
 268 static int
 269 dump_dnode(struct backuparg *ba, uint64_t object, dnode_phys_t *dnp)
 270 {
 271         struct drr_object *drro = &(ba->drr->drr_u.drr_object);
 272
 273         if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
 274                 return (dump_freeobjects(ba, object, 1));
 275
 276         if (ba->pending_op != PENDING_NONE) {
 277                 if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
 278                         return (EINTR);
 279                 ba->pending_op = PENDING_NONE;
 280         }
 281
 282         /* write an OBJECT record */
 283         bzero(ba->drr, sizeof (dmu_replay_record_t));
 284         ba->drr->drr_type = DRR_OBJECT;
 285         drro->drr_object = object;
 286         drro->drr_type = dnp->dn_type;
 287         drro->drr_bonustype = dnp->dn_bonustype;
 288         drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 289         drro->drr_bonuslen = dnp->dn_bonuslen;
 290         drro->drr_checksumtype = dnp->dn_checksum;
 291         drro->drr_compress = dnp->dn_compress;
 292         drro->drr_toguid = ba->toguid;
 293
 294         if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
 295                 return (EINTR);
 296
 297         if (dump_bytes(ba, DN_BONUS(dnp), P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0)
 298                 return (EINTR);
 299
 300         /* free anything past the end of the file */
 301         if (dump_free(ba, object, (dnp->dn_maxblkid + 1) *
 302             (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL))
 303                 return (EINTR);
 304         if (ba->err)
 305                 return (EINTR);
 306         return (0);
 307 }
 308
 309 #define BP_SPAN(dnp, level) \
 310         (((uint64_t)dnp->dn_datablkszsec) << (SPA_MINBLOCKSHIFT + \
 311         (level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)))
 312
 313 /* ARGSUSED */
 314 static int
 315 backup_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, arc_buf_t *pbuf,
 316     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
 317 {
 318         struct backuparg *ba = arg;
 319         dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
 320         int err = 0;
 321
 322         if (issig(JUSTLOOKING) && issig(FORREAL))
 323                 return (EINTR);
 324
 325         if (zb->zb_object != DMU_META_DNODE_OBJECT &&
 326             DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
 327                 return (0);
 328         } else if (bp == NULL && zb->zb_object == DMU_META_DNODE_OBJECT) {
 329                 uint64_t span = BP_SPAN(dnp, zb->zb_level);
 330                 uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
 331                 err = dump_freeobjects(ba, dnobj, span >> DNODE_SHIFT);
 332         } else if (bp == NULL) {
 333                 uint64_t span = BP_SPAN(dnp, zb->zb_level);
 334                 err = dump_free(ba, zb->zb_object, zb->zb_blkid * span, span);
 335         } else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
 336                 return (0);
 337         } else if (type == DMU_OT_DNODE) {
 338                 dnode_phys_t *blk;
 339                 int i;
 340                 int blksz = BP_GET_LSIZE(bp);
 341                 uint32_t aflags = ARC_WAIT;
 342                 arc_buf_t *abuf;
 343
 344                 if (dsl_read(NULL, spa, bp, pbuf,
 345                     arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
 346                     ZIO_FLAG_CANFAIL, &aflags, zb) != 0)
 347                         return (EIO);
 348
 349                 blk = abuf->b_data;
 350                 for (i = 0; i < blksz >> DNODE_SHIFT; i++) {
 351                         uint64_t dnobj = (zb->zb_blkid <<
 352                             (DNODE_BLOCK_SHIFT - DNODE_SHIFT)) + i;
 353                         err = dump_dnode(ba, dnobj, blk+i);
 354                         if (err)
 355                                 break;
 356                 }
 357                 (void) arc_buf_remove_ref(abuf, &abuf);
 358         } else if (type == DMU_OT_SA) {
 359                 uint32_t aflags = ARC_WAIT;
 360                 arc_buf_t *abuf;
 361                 int blksz = BP_GET_LSIZE(bp);
 362
 363                 if (arc_read_nolock(NULL, spa, bp,
 364                     arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
 365                     ZIO_FLAG_CANFAIL, &aflags, zb) != 0)
 366                         return (EIO);
 367
 368                 err = dump_spill(ba, zb->zb_object, blksz, abuf->b_data);
 369                 (void) arc_buf_remove_ref(abuf, &abuf);
 370         } else { /* it's a level-0 block of a regular object */
 371                 uint32_t aflags = ARC_WAIT;
 372                 arc_buf_t *abuf;
 373                 int blksz = BP_GET_LSIZE(bp);
 374
 375                 if (dsl_read(NULL, spa, bp, pbuf,
 376                     arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
 377                     ZIO_FLAG_CANFAIL, &aflags, zb) != 0) {
 378                         if (zfs_send_corrupt_data) {
 379                                 uint64_t *ptr;
 380                                 /* Send a block filled with 0x"zfs badd bloc" */
 381                                 abuf = arc_buf_alloc(spa, blksz, &abuf,
 382                                     ARC_BUFC_DATA);
 383                                 for (ptr = abuf->b_data;
 384                                     (char *)ptr < (char *)abuf->b_data + blksz;
 385                                     ptr++)
 386                                         *ptr = 0x2f5baddb10c;
 387                         } else {
 388                                 return (EIO);
 389                         }
 390                 }
 391
 392                 err = dump_data(ba, type, zb->zb_object, zb->zb_blkid * blksz,
 393                     blksz, bp, abuf->b_data);
 394                 (void) arc_buf_remove_ref(abuf, &abuf);
 395         }
 396
 397         ASSERT(err == 0 || err == EINTR);
 398         return (err);
 399 }
 400
 401 int
 402 dmu_sendbackup(objset_t *tosnap, objset_t *fromsnap, boolean_t fromorigin,
 403     vnode_t *vp, offset_t *off)
 404 {
 405         dsl_dataset_t *ds = tosnap->os_dsl_dataset;
 406         dsl_dataset_t *fromds = fromsnap ? fromsnap->os_dsl_dataset : NULL;
 407         dmu_replay_record_t *drr;
 408         struct backuparg ba;
 409         int err;
 410         uint64_t fromtxg = 0;
 411
 412         /* tosnap must be a snapshot */
 413         if (ds->ds_phys->ds_next_snap_obj == 0)
 414                 return (EINVAL);
 415
 416         /* fromsnap must be an earlier snapshot from the same fs as tosnap */
 417         if (fromds && (ds->ds_dir != fromds->ds_dir ||
 418             fromds->ds_phys->ds_creation_txg >= ds->ds_phys->ds_creation_txg))
 419                 return (EXDEV);
 420
 421         if (fromorigin) {
 422                 dsl_pool_t *dp = ds->ds_dir->dd_pool;
 423
 424                 if (fromsnap)
 425                         return (EINVAL);
 426
 427                 if (dsl_dir_is_clone(ds->ds_dir)) {
 428                         rw_enter(&dp->dp_config_rwlock, RW_READER);
 429                         err = dsl_dataset_hold_obj(dp,
 430                             ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &fromds);
 431                         rw_exit(&dp->dp_config_rwlock);
 432                         if (err)
 433                                 return (err);
 434                 } else {
 435                         fromorigin = B_FALSE;
 436                 }
 437         }
 438
 439
 440         drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
 441         drr->drr_type = DRR_BEGIN;
 442         drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
 443         DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
 444             DMU_SUBSTREAM);
 445
 446 #ifdef _KERNEL
 447         if (dmu_objset_type(tosnap) == DMU_OST_ZFS) {
 448                 uint64_t version;
 449                 if (zfs_get_zplprop(tosnap, ZFS_PROP_VERSION, &version) != 0)
 450                         return (EINVAL);
 451                 if (version == ZPL_VERSION_SA) {
 452                         DMU_SET_FEATUREFLAGS(
 453                             drr->drr_u.drr_begin.drr_versioninfo,
 454                             DMU_BACKUP_FEATURE_SA_SPILL);
 455                 }
 456         }
 457 #endif
 458
 459         drr->drr_u.drr_begin.drr_creation_time =
 460             ds->ds_phys->ds_creation_time;
 461         drr->drr_u.drr_begin.drr_type = tosnap->os_phys->os_type;
 462         if (fromorigin)
 463                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
 464         drr->drr_u.drr_begin.drr_toguid = ds->ds_phys->ds_guid;
 465         if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
 466                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
 467
 468         if (fromds)
 469                 drr->drr_u.drr_begin.drr_fromguid = fromds->ds_phys->ds_guid;
 470         dsl_dataset_name(ds, drr->drr_u.drr_begin.drr_toname);
 471
 472         if (fromds)
 473                 fromtxg = fromds->ds_phys->ds_creation_txg;
 474         if (fromorigin)
 475                 dsl_dataset_rele(fromds, FTAG);
 476
 477         ba.drr = drr;
 478         ba.vp = vp;
 479         ba.os = tosnap;
 480         ba.off = off;
 481         ba.toguid = ds->ds_phys->ds_guid;
 482         ZIO_SET_CHECKSUM(&ba.zc, 0, 0, 0, 0);
 483         ba.pending_op = PENDING_NONE;
 484
 485         if (dump_bytes(&ba, drr, sizeof (dmu_replay_record_t)) != 0) {
 486                 kmem_free(drr, sizeof (dmu_replay_record_t));
 487                 return (ba.err);
 488         }
 489
 490         err = traverse_dataset(ds, fromtxg, TRAVERSE_PRE | TRAVERSE_PREFETCH,
 491             backup_cb, &ba);
 492
 493         if (ba.pending_op != PENDING_NONE)
 494                 if (dump_bytes(&ba, drr, sizeof (dmu_replay_record_t)) != 0)
 495                         err = EINTR;
 496
 497         if (err) {
 498                 if (err == EINTR && ba.err)
 499                         err = ba.err;
 500                 kmem_free(drr, sizeof (dmu_replay_record_t));
 501                 return (err);
 502         }
 503
 504         bzero(drr, sizeof (dmu_replay_record_t));
 505         drr->drr_type = DRR_END;
 506         drr->drr_u.drr_end.drr_checksum = ba.zc;
 507         drr->drr_u.drr_end.drr_toguid = ba.toguid;
 508
 509         if (dump_bytes(&ba, drr, sizeof (dmu_replay_record_t)) != 0) {
 510                 kmem_free(drr, sizeof (dmu_replay_record_t));
 511                 return (ba.err);
 512         }
 513
 514         kmem_free(drr, sizeof (dmu_replay_record_t));
 515
 516         return (0);
 517 }
 518
 519 int
 520 dmu_send_estimate(objset_t *tosnap, objset_t *fromsnap, boolean_t fromorigin,
 521     uint64_t *sizep)
 522 {
 523         dsl_dataset_t *ds = tosnap->os_dsl_dataset;
 524         dsl_dataset_t *fromds = fromsnap ? fromsnap->os_dsl_dataset : NULL;
 525         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 526         int err;
 527         uint64_t size, recordsize;
 528
 529         /* tosnap must be a snapshot */
 530         if (ds->ds_phys->ds_next_snap_obj == 0)
 531                 return (EINVAL);
 532
 533         /* fromsnap must be an earlier snapshot from the same fs as tosnap */
 534         if (fromds && (ds->ds_dir != fromds->ds_dir ||
 535             fromds->ds_phys->ds_creation_txg >= ds->ds_phys->ds_creation_txg))
 536                 return (EXDEV);
 537
 538         if (fromorigin) {
 539                 if (fromsnap)
 540                         return (EINVAL);
 541
 542                 if (dsl_dir_is_clone(ds->ds_dir)) {
 543                         rw_enter(&dp->dp_config_rwlock, RW_READER);
 544                         err = dsl_dataset_hold_obj(dp,
 545                             ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &fromds);
 546                         rw_exit(&dp->dp_config_rwlock);
 547                         if (err)
 548                                 return (err);
 549                 } else {
 550                         fromorigin = B_FALSE;
 551                 }
 552         }
 553
 554         /* Get uncompressed size estimate of changed data. */
 555         if (fromds == NULL) {
 556                 size = ds->ds_phys->ds_uncompressed_bytes;
 557         } else {
 558                 uint64_t used, comp;
 559                 err = dsl_dataset_space_written(fromds, ds,
 560                     &used, &comp, &size);
 561                 if (fromorigin)
 562                         dsl_dataset_rele(fromds, FTAG);
 563                 if (err)
 564                         return (err);
 565         }
 566
 567         /*
 568          * Assume that space (both on-disk and in-stream) is dominated by
 569          * data.  We will adjust for indirect blocks and the copies property,
 570          * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
 571          */
 572
 573         /*
 574          * Subtract out approximate space used by indirect blocks.
 575          * Assume most space is used by data blocks (non-indirect, non-dnode).
 576          * Assume all blocks are recordsize.  Assume ditto blocks and
 577          * internal fragmentation counter out compression.
 578          *
 579          * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
 580          * block, which we observe in practice.
 581          */
 582         rw_enter(&dp->dp_config_rwlock, RW_READER);
 583         err = dsl_prop_get_ds(ds, "recordsize",
 584             sizeof (recordsize), 1, &recordsize, NULL);
 585         rw_exit(&dp->dp_config_rwlock);
 586         if (err)
 587                 return (err);
 588         size -= size / recordsize * sizeof (blkptr_t);
 589
 590         /* Add in the space for the record associated with each block. */
 591         size += size / recordsize * sizeof (dmu_replay_record_t);
 592
 593         *sizep = size;
 594
 595         return (0);
 596 }
 597
 598 struct recvbeginsyncarg {
 599         const char *tofs;
 600         const char *tosnap;
 601         dsl_dataset_t *origin;
 602         uint64_t fromguid;
 603         dmu_objset_type_t type;
 604         void *tag;
 605         boolean_t force;
 606         uint64_t dsflags;
 607         char clonelastname[MAXNAMELEN];
 608         dsl_dataset_t *ds; /* the ds to recv into; returned from the syncfunc */
 609         cred_t *cr;
 610 };
 611
 612 /* ARGSUSED */
 613 static int
 614 recv_new_check(void *arg1, void *arg2, dmu_tx_t *tx)
 615 {
 616         dsl_dir_t *dd = arg1;
 617         struct recvbeginsyncarg *rbsa = arg2;
 618         objset_t *mos = dd->dd_pool->dp_meta_objset;
 619         uint64_t val;
 620         int err;
 621
 622         err = zap_lookup(mos, dd->dd_phys->dd_child_dir_zapobj,
 623             strrchr(rbsa->tofs, '/') + 1, sizeof (uint64_t), 1, &val);
 624
 625         if (err != ENOENT)
 626                 return (err ? err : EEXIST);
 627
 628         if (rbsa->origin) {
 629                 /* make sure it's a snap in the same pool */
 630                 if (rbsa->origin->ds_dir->dd_pool != dd->dd_pool)
 631                         return (EXDEV);
 632                 if (!dsl_dataset_is_snapshot(rbsa->origin))
 633                         return (EINVAL);
 634                 if (rbsa->origin->ds_phys->ds_guid != rbsa->fromguid)
 635                         return (ENODEV);
 636         }
 637
 638         return (0);
 639 }
 640
 641 static void
 642 recv_new_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 643 {
 644         dsl_dir_t *dd = arg1;
 645         struct recvbeginsyncarg *rbsa = arg2;
 646         uint64_t flags = DS_FLAG_INCONSISTENT | rbsa->dsflags;
 647         uint64_t dsobj;
 648
 649         /* Create and open new dataset. */
 650         dsobj = dsl_dataset_create_sync(dd, strrchr(rbsa->tofs, '/') + 1,
 651             rbsa->origin, flags, rbsa->cr, tx);
 652         VERIFY(0 == dsl_dataset_own_obj(dd->dd_pool, dsobj,
 653             B_TRUE, dmu_recv_tag, &rbsa->ds));
 654
 655         if (rbsa->origin == NULL) {
 656                 (void) dmu_objset_create_impl(dd->dd_pool->dp_spa,
 657                     rbsa->ds, &rbsa->ds->ds_phys->ds_bp, rbsa->type, tx);
 658         }
 659
 660         spa_history_log_internal(LOG_DS_REPLAY_FULL_SYNC,
 661             dd->dd_pool->dp_spa, tx, "dataset = %lld", dsobj);
 662 }
 663
 664 /* ARGSUSED */
 665 static int
 666 recv_existing_check(void *arg1, void *arg2, dmu_tx_t *tx)
 667 {
 668         dsl_dataset_t *ds = arg1;
 669         struct recvbeginsyncarg *rbsa = arg2;
 670         int err;
 671         uint64_t val;
 672
 673         /* must not have any changes since most recent snapshot */
 674         if (!rbsa->force && dsl_dataset_modified_since_lastsnap(ds))
 675                 return (ETXTBSY);
 676
 677         /* new snapshot name must not exist */
 678         err = zap_lookup(ds->ds_dir->dd_pool->dp_meta_objset,
 679             ds->ds_phys->ds_snapnames_zapobj, rbsa->tosnap, 8, 1, &val);
 680         if (err == 0)
 681                 return (EEXIST);
 682         if (err != ENOENT)
 683                 return (err);
 684
 685         if (rbsa->fromguid) {
 686                 /* if incremental, most recent snapshot must match fromguid */
 687                 if (ds->ds_prev == NULL)
 688                         return (ENODEV);
 689
 690                 /*
 691                  * most recent snapshot must match fromguid, or there are no
 692                  * changes since the fromguid one
 693                  */
 694                 if (ds->ds_prev->ds_phys->ds_guid != rbsa->fromguid) {
 695                         uint64_t birth = ds->ds_prev->ds_phys->ds_bp.blk_birth;
 696                         uint64_t obj = ds->ds_prev->ds_phys->ds_prev_snap_obj;
 697                         while (obj != 0) {
 698                                 dsl_dataset_t *snap;
 699                                 err = dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
 700                                     obj, FTAG, &snap);
 701                                 if (err)
 702                                         return (ENODEV);
 703                                 if (snap->ds_phys->ds_creation_txg < birth) {
 704                                         dsl_dataset_rele(snap, FTAG);
 705                                         return (ENODEV);
 706                                 }
 707                                 if (snap->ds_phys->ds_guid == rbsa->fromguid) {
 708                                         dsl_dataset_rele(snap, FTAG);
 709                                         break; /* it's ok */
 710                                 }
 711                                 obj = snap->ds_phys->ds_prev_snap_obj;
 712                                 dsl_dataset_rele(snap, FTAG);
 713                         }
 714                         if (obj == 0)
 715                                 return (ENODEV);
 716                 }
 717         } else {
 718                 /* if full, most recent snapshot must be $ORIGIN */
 719                 if (ds->ds_phys->ds_prev_snap_txg >= TXG_INITIAL)
 720                         return (ENODEV);
 721         }
 722
 723         /* temporary clone name must not exist */
 724         err = zap_lookup(ds->ds_dir->dd_pool->dp_meta_objset,
 725             ds->ds_dir->dd_phys->dd_child_dir_zapobj,
 726             rbsa->clonelastname, 8, 1, &val);
 727         if (err == 0)
 728                 return (EEXIST);
 729         if (err != ENOENT)
 730                 return (err);
 731
 732         return (0);
 733 }
 734
 735 /* ARGSUSED */
 736 static void
 737 recv_existing_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 738 {
 739         dsl_dataset_t *ohds = arg1;
 740         struct recvbeginsyncarg *rbsa = arg2;
 741         dsl_pool_t *dp = ohds->ds_dir->dd_pool;
 742         dsl_dataset_t *cds;
 743         uint64_t flags = DS_FLAG_INCONSISTENT | rbsa->dsflags;
 744         uint64_t dsobj;
 745
 746         /* create and open the temporary clone */
 747         dsobj = dsl_dataset_create_sync(ohds->ds_dir, rbsa->clonelastname,
 748             ohds->ds_prev, flags, rbsa->cr, tx);
 749         VERIFY(0 == dsl_dataset_own_obj(dp, dsobj, B_TRUE, dmu_recv_tag, &cds));
 750
 751         /*
 752          * If we actually created a non-clone, we need to create the
 753          * objset in our new dataset.
 754          */
 755         if (BP_IS_HOLE(dsl_dataset_get_blkptr(cds))) {
 756                 (void) dmu_objset_create_impl(dp->dp_spa,
 757                     cds, dsl_dataset_get_blkptr(cds), rbsa->type, tx);
 758         }
 759
 760         rbsa->ds = cds;
 761
 762         spa_history_log_internal(LOG_DS_REPLAY_INC_SYNC,
 763             dp->dp_spa, tx, "dataset = %lld", dsobj);
 764 }
 765
 766 static boolean_t
 767 dmu_recv_verify_features(dsl_dataset_t *ds, struct drr_begin *drrb)
 768 {
 769         int featureflags;
 770
 771         featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
 772
 773         /* Verify pool version supports SA if SA_SPILL feature set */
 774         return ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
 775             (spa_version(dsl_dataset_get_spa(ds)) < SPA_VERSION_SA));
 776 }
 777
 778 /*
 779  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
 780  * succeeds; otherwise we will leak the holds on the datasets.
 781  */
 782 int
 783 dmu_recv_begin(char *tofs, char *tosnap, char *top_ds, struct drr_begin *drrb,
 784     boolean_t force, objset_t *origin, dmu_recv_cookie_t *drc)
 785 {
 786         int err = 0;
 787         boolean_t byteswap;
 788         struct recvbeginsyncarg rbsa = { 0 };
 789         uint64_t versioninfo;
 790         int flags;
 791         dsl_dataset_t *ds;
 792
 793         if (drrb->drr_magic == DMU_BACKUP_MAGIC)
 794                 byteswap = FALSE;
 795         else if (drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
 796                 byteswap = TRUE;
 797         else
 798                 return (EINVAL);
 799
 800         rbsa.tofs = tofs;
 801         rbsa.tosnap = tosnap;
 802         rbsa.origin = origin ? origin->os_dsl_dataset : NULL;
 803         rbsa.fromguid = drrb->drr_fromguid;
 804         rbsa.type = drrb->drr_type;
 805         rbsa.tag = FTAG;
 806         rbsa.dsflags = 0;
 807         rbsa.cr = CRED();
 808         versioninfo = drrb->drr_versioninfo;
 809         flags = drrb->drr_flags;
 810
 811         if (byteswap) {
 812                 rbsa.type = BSWAP_32(rbsa.type);
 813                 rbsa.fromguid = BSWAP_64(rbsa.fromguid);
 814                 versioninfo = BSWAP_64(versioninfo);
 815                 flags = BSWAP_32(flags);
 816         }
 817
 818         if (DMU_GET_STREAM_HDRTYPE(versioninfo) == DMU_COMPOUNDSTREAM ||
 819             rbsa.type >= DMU_OST_NUMTYPES ||
 820             ((flags & DRR_FLAG_CLONE) && origin == NULL))
 821                 return (EINVAL);
 822
 823         if (flags & DRR_FLAG_CI_DATA)
 824                 rbsa.dsflags = DS_FLAG_CI_DATASET;
 825
 826         bzero(drc, sizeof (dmu_recv_cookie_t));
 827         drc->drc_drrb = drrb;
 828         drc->drc_tosnap = tosnap;
 829         drc->drc_top_ds = top_ds;
 830         drc->drc_force = force;
 831
 832         /*
 833          * Process the begin in syncing context.
 834          */
 835
 836         /* open the dataset we are logically receiving into */
 837         err = dsl_dataset_hold(tofs, dmu_recv_tag, &ds);
 838         if (err == 0) {
 839                 if (dmu_recv_verify_features(ds, drrb)) {
 840                         dsl_dataset_rele(ds, dmu_recv_tag);
 841                         return (ENOTSUP);
 842                 }
 843                 /* target fs already exists; recv into temp clone */
 844
 845                 /* Can't recv a clone into an existing fs */
 846                 if (flags & DRR_FLAG_CLONE) {
 847                         dsl_dataset_rele(ds, dmu_recv_tag);
 848                         return (EINVAL);
 849                 }
 850
 851                 /* must not have an incremental recv already in progress */
 852                 if (!mutex_tryenter(&ds->ds_recvlock)) {
 853                         dsl_dataset_rele(ds, dmu_recv_tag);
 854                         return (EBUSY);
 855                 }
 856
 857                 /* tmp clone name is: tofs/%tosnap" */
 858                 (void) snprintf(rbsa.clonelastname, sizeof (rbsa.clonelastname),
 859                     "%%%s", tosnap);
 860                 rbsa.force = force;
 861                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
 862                     recv_existing_check, recv_existing_sync, ds, &rbsa, 5);
 863                 if (err) {
 864                         mutex_exit(&ds->ds_recvlock);
 865                         dsl_dataset_rele(ds, dmu_recv_tag);
 866                         return (err);
 867                 }
 868                 drc->drc_logical_ds = ds;
 869                 drc->drc_real_ds = rbsa.ds;
 870         } else if (err == ENOENT) {
 871                 /* target fs does not exist; must be a full backup or clone */
 872                 char *cp;
 873
 874                 /*
 875                  * If it's a non-clone incremental, we are missing the
 876                  * target fs, so fail the recv.
 877                  */
 878                 if (rbsa.fromguid && !(flags & DRR_FLAG_CLONE))
 879                         return (ENOENT);
 880
 881                 /* Open the parent of tofs */
 882                 cp = strrchr(tofs, '/');
 883                 *cp = '\0';
 884                 err = dsl_dataset_hold(tofs, FTAG, &ds);
 885                 *cp = '/';
 886                 if (err)
 887                         return (err);
 888
 889                 if (dmu_recv_verify_features(ds, drrb)) {
 890                         dsl_dataset_rele(ds, FTAG);
 891                         return (ENOTSUP);
 892                 }
 893
 894                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
 895                     recv_new_check, recv_new_sync, ds->ds_dir, &rbsa, 5);
 896                 dsl_dataset_rele(ds, FTAG);
 897                 if (err)
 898                         return (err);
 899                 drc->drc_logical_ds = drc->drc_real_ds = rbsa.ds;
 900                 drc->drc_newfs = B_TRUE;
 901         }
 902
 903         return (err);
 904 }
 905
 906 struct restorearg {
 907         int err;
 908         int byteswap;
 909         vnode_t *vp;
 910         char *buf;
 911         uint64_t voff;
 912         int bufsize; /* amount of memory allocated for buf */
 913         zio_cksum_t cksum;
 914         avl_tree_t *guid_to_ds_map;
 915 };
 916
 917 typedef struct guid_map_entry {
 918         uint64_t        guid;
 919         dsl_dataset_t   *gme_ds;
 920         avl_node_t      avlnode;
 921 } guid_map_entry_t;
 922
 923 static int
 924 guid_compare(const void *arg1, const void *arg2)
 925 {
 926         const guid_map_entry_t *gmep1 = arg1;
 927         const guid_map_entry_t *gmep2 = arg2;
 928
 929         if (gmep1->guid < gmep2->guid)
 930                 return (-1);
 931         else if (gmep1->guid > gmep2->guid)
 932                 return (1);
 933         return (0);
 934 }
 935
 936 static void
 937 free_guid_map_onexit(void *arg)
 938 {
 939         avl_tree_t *ca = arg;
 940         void *cookie = NULL;
 941         guid_map_entry_t *gmep;
 942
 943         while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
 944                 dsl_dataset_rele(gmep->gme_ds, ca);
 945                 kmem_free(gmep, sizeof (guid_map_entry_t));
 946         }
 947         avl_destroy(ca);
 948         kmem_free(ca, sizeof (avl_tree_t));
 949 }
 950
 951 static void *
 952 restore_read(struct restorearg *ra, int len)
 953 {
 954         void *rv;
 955         int done = 0;
 956
 957         /* some things will require 8-byte alignment, so everything must */
 958         ASSERT3U(len % 8, ==, 0);
 959
 960         while (done < len) {
 961                 ssize_t resid;
 962
 963                 ra->err = vn_rdwr(UIO_READ, ra->vp,
 964                     (caddr_t)ra->buf + done, len - done,
 965                     ra->voff, UIO_SYSSPACE, FAPPEND,
 966                     RLIM64_INFINITY, CRED(), &resid);
 967
 968                 if (resid == len - done)
 969                         ra->err = EINVAL;
 970                 ra->voff += len - done - resid;
 971                 done = len - resid;
 972                 if (ra->err)
 973                         return (NULL);
 974         }
 975
 976         ASSERT3U(done, ==, len);
 977         rv = ra->buf;
 978         if (ra->byteswap)
 979                 fletcher_4_incremental_byteswap(rv, len, &ra->cksum);
 980         else
 981                 fletcher_4_incremental_native(rv, len, &ra->cksum);
 982         return (rv);
 983 }
 984
 985 noinline static void
 986 backup_byteswap(dmu_replay_record_t *drr)
 987 {
 988 #define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
 989 #define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
 990         drr->drr_type = BSWAP_32(drr->drr_type);
 991         drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
 992         switch (drr->drr_type) {
 993         case DRR_BEGIN:
 994                 DO64(drr_begin.drr_magic);
 995                 DO64(drr_begin.drr_versioninfo);
 996                 DO64(drr_begin.drr_creation_time);
 997                 DO32(drr_begin.drr_type);
 998                 DO32(drr_begin.drr_flags);
 999                 DO64(drr_begin.drr_toguid);
1000                 DO64(drr_begin.drr_fromguid);
1001                 break;
1002         case DRR_OBJECT:
1003                 DO64(drr_object.drr_object);
1004                 /* DO64(drr_object.drr_allocation_txg); */
1005                 DO32(drr_object.drr_type);
1006                 DO32(drr_object.drr_bonustype);
1007                 DO32(drr_object.drr_blksz);
1008                 DO32(drr_object.drr_bonuslen);
1009                 DO64(drr_object.drr_toguid);
1010                 break;
1011         case DRR_FREEOBJECTS:
1012                 DO64(drr_freeobjects.drr_firstobj);
1013                 DO64(drr_freeobjects.drr_numobjs);
1014                 DO64(drr_freeobjects.drr_toguid);
1015                 break;
1016         case DRR_WRITE:
1017                 DO64(drr_write.drr_object);
1018                 DO32(drr_write.drr_type);
1019                 DO64(drr_write.drr_offset);
1020                 DO64(drr_write.drr_length);
1021                 DO64(drr_write.drr_toguid);
1022                 DO64(drr_write.drr_key.ddk_cksum.zc_word[0]);
1023                 DO64(drr_write.drr_key.ddk_cksum.zc_word[1]);
1024                 DO64(drr_write.drr_key.ddk_cksum.zc_word[2]);
1025                 DO64(drr_write.drr_key.ddk_cksum.zc_word[3]);
1026                 DO64(drr_write.drr_key.ddk_prop);
1027                 break;
1028         case DRR_WRITE_BYREF:
1029                 DO64(drr_write_byref.drr_object);
1030                 DO64(drr_write_byref.drr_offset);
1031                 DO64(drr_write_byref.drr_length);
1032                 DO64(drr_write_byref.drr_toguid);
1033                 DO64(drr_write_byref.drr_refguid);
1034                 DO64(drr_write_byref.drr_refobject);
1035                 DO64(drr_write_byref.drr_refoffset);
1036                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[0]);
1037                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[1]);
1038                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[2]);
1039                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[3]);
1040                 DO64(drr_write_byref.drr_key.ddk_prop);
1041                 break;
1042         case DRR_FREE:
1043                 DO64(drr_free.drr_object);
1044                 DO64(drr_free.drr_offset);
1045                 DO64(drr_free.drr_length);
1046                 DO64(drr_free.drr_toguid);
1047                 break;
1048         case DRR_SPILL:
1049                 DO64(drr_spill.drr_object);
1050                 DO64(drr_spill.drr_length);
1051                 DO64(drr_spill.drr_toguid);
1052                 break;
1053         case DRR_END:
1054                 DO64(drr_end.drr_checksum.zc_word[0]);
1055                 DO64(drr_end.drr_checksum.zc_word[1]);
1056                 DO64(drr_end.drr_checksum.zc_word[2]);
1057                 DO64(drr_end.drr_checksum.zc_word[3]);
1058                 DO64(drr_end.drr_toguid);
1059                 break;
1060         default:
1061                 break;
1062         }
1063 #undef DO64
1064 #undef DO32
1065 }
1066
1067 noinline static int
1068 restore_object(struct restorearg *ra, objset_t *os, struct drr_object *drro)
1069 {
1070         int err;
1071         dmu_tx_t *tx;
1072         void *data = NULL;
1073
1074         if (drro->drr_type == DMU_OT_NONE ||
1075             drro->drr_type >= DMU_OT_NUMTYPES ||
1076             drro->drr_bonustype >= DMU_OT_NUMTYPES ||
1077             drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
1078             drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
1079             P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
1080             drro->drr_blksz < SPA_MINBLOCKSIZE ||
1081             drro->drr_blksz > SPA_MAXBLOCKSIZE ||
1082             drro->drr_bonuslen > DN_MAX_BONUSLEN) {
1083                 return (EINVAL);
1084         }
1085
1086         err = dmu_object_info(os, drro->drr_object, NULL);
1087
1088         if (err != 0 && err != ENOENT)
1089                 return (EINVAL);
1090
1091         if (drro->drr_bonuslen) {
1092                 data = restore_read(ra, P2ROUNDUP(drro->drr_bonuslen, 8));
1093                 if (ra->err)
1094                         return (ra->err);
1095         }
1096
1097         if (err == ENOENT) {
1098                 /* currently free, want to be allocated */
1099                 tx = dmu_tx_create(os);
1100                 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1101                 err = dmu_tx_assign(tx, TXG_WAIT);
1102                 if (err) {
1103                         dmu_tx_abort(tx);
1104                         return (err);
1105                 }
1106                 err = dmu_object_claim(os, drro->drr_object,
1107                     drro->drr_type, drro->drr_blksz,
1108                     drro->drr_bonustype, drro->drr_bonuslen, tx);
1109                 dmu_tx_commit(tx);
1110         } else {
1111                 /* currently allocated, want to be allocated */
1112                 err = dmu_object_reclaim(os, drro->drr_object,
1113                     drro->drr_type, drro->drr_blksz,
1114                     drro->drr_bonustype, drro->drr_bonuslen);
1115         }
1116         if (err) {
1117                 return (EINVAL);
1118         }
1119
1120         tx = dmu_tx_create(os);
1121         dmu_tx_hold_bonus(tx, drro->drr_object);
1122         err = dmu_tx_assign(tx, TXG_WAIT);
1123         if (err) {
1124                 dmu_tx_abort(tx);
1125                 return (err);
1126         }
1127
1128         dmu_object_set_checksum(os, drro->drr_object, drro->drr_checksumtype,
1129             tx);
1130         dmu_object_set_compress(os, drro->drr_object, drro->drr_compress, tx);
1131
1132         if (data != NULL) {
1133                 dmu_buf_t *db;
1134
1135                 VERIFY(0 == dmu_bonus_hold(os, drro->drr_object, FTAG, &db));
1136                 dmu_buf_will_dirty(db, tx);
1137
1138                 ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
1139                 bcopy(data, db->db_data, drro->drr_bonuslen);
1140                 if (ra->byteswap) {
1141                         dmu_ot[drro->drr_bonustype].ot_byteswap(db->db_data,
1142                             drro->drr_bonuslen);
1143                 }
1144                 dmu_buf_rele(db, FTAG);
1145         }
1146         dmu_tx_commit(tx);
1147         return (0);
1148 }
1149
1150 /* ARGSUSED */
1151 noinline static int
1152 restore_freeobjects(struct restorearg *ra, objset_t *os,
1153     struct drr_freeobjects *drrfo)
1154 {
1155         uint64_t obj;
1156
1157         if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
1158                 return (EINVAL);
1159
1160         for (obj = drrfo->drr_firstobj;
1161             obj < drrfo->drr_firstobj + drrfo->drr_numobjs;
1162             (void) dmu_object_next(os, &obj, FALSE, 0)) {
1163                 int err;
1164
1165                 if (dmu_object_info(os, obj, NULL) != 0)
1166                         continue;
1167
1168                 err = dmu_free_object(os, obj);
1169                 if (err)
1170                         return (err);
1171         }
1172         return (0);
1173 }
1174
1175 noinline static int
1176 restore_write(struct restorearg *ra, objset_t *os,
1177     struct drr_write *drrw)
1178 {
1179         dmu_tx_t *tx;
1180         void *data;
1181         int err;
1182
1183         if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
1184             drrw->drr_type >= DMU_OT_NUMTYPES)
1185                 return (EINVAL);
1186
1187         data = restore_read(ra, drrw->drr_length);
1188         if (data == NULL)
1189                 return (ra->err);
1190
1191         if (dmu_object_info(os, drrw->drr_object, NULL) != 0)
1192                 return (EINVAL);
1193
1194         tx = dmu_tx_create(os);
1195
1196         dmu_tx_hold_write(tx, drrw->drr_object,
1197             drrw->drr_offset, drrw->drr_length);
1198         err = dmu_tx_assign(tx, TXG_WAIT);
1199         if (err) {
1200                 dmu_tx_abort(tx);
1201                 return (err);
1202         }
1203         if (ra->byteswap)
1204                 dmu_ot[drrw->drr_type].ot_byteswap(data, drrw->drr_length);
1205         dmu_write(os, drrw->drr_object,
1206             drrw->drr_offset, drrw->drr_length, data, tx);
1207         dmu_tx_commit(tx);
1208         return (0);
1209 }
1210
1211 /*
1212  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
1213  * streams to refer to a copy of the data that is already on the
1214  * system because it came in earlier in the stream.  This function
1215  * finds the earlier copy of the data, and uses that copy instead of
1216  * data from the stream to fulfill this write.
1217  */
1218 static int
1219 restore_write_byref(struct restorearg *ra, objset_t *os,
1220     struct drr_write_byref *drrwbr)
1221 {
1222         dmu_tx_t *tx;
1223         int err;
1224         guid_map_entry_t gmesrch;
1225         guid_map_entry_t *gmep;
1226         avl_index_t     where;
1227         objset_t *ref_os = NULL;
1228         dmu_buf_t *dbp;
1229
1230         if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
1231                 return (EINVAL);
1232
1233         /*
1234          * If the GUID of the referenced dataset is different from the
1235          * GUID of the target dataset, find the referenced dataset.
1236          */
1237         if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
1238                 gmesrch.guid = drrwbr->drr_refguid;
1239                 if ((gmep = avl_find(ra->guid_to_ds_map, &gmesrch,
1240                     &where)) == NULL) {
1241                         return (EINVAL);
1242                 }
1243                 if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
1244                         return (EINVAL);
1245         } else {
1246                 ref_os = os;
1247         }
1248
1249         err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
1250             drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH);
1251         if (err)
1252                 return (err);
1253
1254         tx = dmu_tx_create(os);
1255
1256         dmu_tx_hold_write(tx, drrwbr->drr_object,
1257             drrwbr->drr_offset, drrwbr->drr_length);
1258         err = dmu_tx_assign(tx, TXG_WAIT);
1259         if (err) {
1260                 dmu_tx_abort(tx);
1261                 return (err);
1262         }
1263         dmu_write(os, drrwbr->drr_object,
1264             drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
1265         dmu_buf_rele(dbp, FTAG);
1266         dmu_tx_commit(tx);
1267         return (0);
1268 }
1269
1270 static int
1271 restore_spill(struct restorearg *ra, objset_t *os, struct drr_spill *drrs)
1272 {
1273         dmu_tx_t *tx;
1274         void *data;
1275         dmu_buf_t *db, *db_spill;
1276         int err;
1277
1278         if (drrs->drr_length < SPA_MINBLOCKSIZE ||
1279             drrs->drr_length > SPA_MAXBLOCKSIZE)
1280                 return (EINVAL);
1281
1282         data = restore_read(ra, drrs->drr_length);
1283         if (data == NULL)
1284                 return (ra->err);
1285
1286         if (dmu_object_info(os, drrs->drr_object, NULL) != 0)
1287                 return (EINVAL);
1288
1289         VERIFY(0 == dmu_bonus_hold(os, drrs->drr_object, FTAG, &db));
1290         if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
1291                 dmu_buf_rele(db, FTAG);
1292                 return (err);
1293         }
1294
1295         tx = dmu_tx_create(os);
1296
1297         dmu_tx_hold_spill(tx, db->db_object);
1298
1299         err = dmu_tx_assign(tx, TXG_WAIT);
1300         if (err) {
1301                 dmu_buf_rele(db, FTAG);
1302                 dmu_buf_rele(db_spill, FTAG);
1303                 dmu_tx_abort(tx);
1304                 return (err);
1305         }
1306         dmu_buf_will_dirty(db_spill, tx);
1307
1308         if (db_spill->db_size < drrs->drr_length)
1309                 VERIFY(0 == dbuf_spill_set_blksz(db_spill,
1310                     drrs->drr_length, tx));
1311         bcopy(data, db_spill->db_data, drrs->drr_length);
1312
1313         dmu_buf_rele(db, FTAG);
1314         dmu_buf_rele(db_spill, FTAG);
1315
1316         dmu_tx_commit(tx);
1317         return (0);
1318 }
1319
1320 /* ARGSUSED */
1321 noinline static int
1322 restore_free(struct restorearg *ra, objset_t *os,
1323     struct drr_free *drrf)
1324 {
1325         int err;
1326
1327         if (drrf->drr_length != -1ULL &&
1328             drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
1329                 return (EINVAL);
1330
1331         if (dmu_object_info(os, drrf->drr_object, NULL) != 0)
1332                 return (EINVAL);
1333
1334         err = dmu_free_long_range(os, drrf->drr_object,
1335             drrf->drr_offset, drrf->drr_length);
1336         return (err);
1337 }
1338
1339 /*
1340  * NB: callers *must* call dmu_recv_end() if this succeeds.
1341  */
1342 int
1343 dmu_recv_stream(dmu_recv_cookie_t *drc, vnode_t *vp, offset_t *voffp,
1344     int cleanup_fd, uint64_t *action_handlep)
1345 {
1346         struct restorearg ra = { 0 };
1347         dmu_replay_record_t *drr;
1348         objset_t *os;
1349         zio_cksum_t pcksum;
1350         int featureflags;
1351
1352         if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
1353                 ra.byteswap = TRUE;
1354
1355         {
1356                 /* compute checksum of drr_begin record */
1357                 dmu_replay_record_t *drr;
1358                 drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
1359
1360                 drr->drr_type = DRR_BEGIN;
1361                 drr->drr_u.drr_begin = *drc->drc_drrb;
1362                 if (ra.byteswap) {
1363                         fletcher_4_incremental_byteswap(drr,
1364                             sizeof (dmu_replay_record_t), &ra.cksum);
1365                 } else {
1366                         fletcher_4_incremental_native(drr,
1367                             sizeof (dmu_replay_record_t), &ra.cksum);
1368                 }
1369                 kmem_free(drr, sizeof (dmu_replay_record_t));
1370         }
1371
1372         if (ra.byteswap) {
1373                 struct drr_begin *drrb = drc->drc_drrb;
1374                 drrb->drr_magic = BSWAP_64(drrb->drr_magic);
1375                 drrb->drr_versioninfo = BSWAP_64(drrb->drr_versioninfo);
1376                 drrb->drr_creation_time = BSWAP_64(drrb->drr_creation_time);
1377                 drrb->drr_type = BSWAP_32(drrb->drr_type);
1378                 drrb->drr_toguid = BSWAP_64(drrb->drr_toguid);
1379                 drrb->drr_fromguid = BSWAP_64(drrb->drr_fromguid);
1380         }
1381
1382         ra.vp = vp;
1383         ra.voff = *voffp;
1384         ra.bufsize = 1<<20;
1385         ra.buf = vmem_alloc(ra.bufsize, KM_SLEEP);
1386
1387         /* these were verified in dmu_recv_begin */
1388         ASSERT(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo) ==
1389             DMU_SUBSTREAM);
1390         ASSERT(drc->drc_drrb->drr_type < DMU_OST_NUMTYPES);
1391
1392         /*
1393          * Open the objset we are modifying.
1394          */
1395         VERIFY(dmu_objset_from_ds(drc->drc_real_ds, &os) == 0);
1396
1397         ASSERT(drc->drc_real_ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT);
1398
1399         featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
1400
1401         /* if this stream is dedup'ed, set up the avl tree for guid mapping */
1402         if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
1403                 minor_t minor;
1404
1405                 if (cleanup_fd == -1) {
1406                         ra.err = EBADF;
1407                         goto out;
1408                 }
1409                 ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
1410                 if (ra.err) {
1411                         cleanup_fd = -1;
1412                         goto out;
1413                 }
1414
1415                 if (*action_handlep == 0) {
1416                         ra.guid_to_ds_map =
1417                             kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
1418                         avl_create(ra.guid_to_ds_map, guid_compare,
1419                             sizeof (guid_map_entry_t),
1420                             offsetof(guid_map_entry_t, avlnode));
1421                         ra.err = zfs_onexit_add_cb(minor,
1422                             free_guid_map_onexit, ra.guid_to_ds_map,
1423                             action_handlep);
1424                         if (ra.err)
1425                                 goto out;
1426                 } else {
1427                         ra.err = zfs_onexit_cb_data(minor, *action_handlep,
1428                             (void **)&ra.guid_to_ds_map);
1429                         if (ra.err)
1430                                 goto out;
1431                 }
1432
1433                 drc->drc_guid_to_ds_map = ra.guid_to_ds_map;
1434         }
1435
1436         /*
1437          * Read records and process them.
1438          */
1439         pcksum = ra.cksum;
1440         while (ra.err == 0 &&
1441             NULL != (drr = restore_read(&ra, sizeof (*drr)))) {
1442                 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1443                         ra.err = EINTR;
1444                         goto out;
1445                 }
1446
1447                 if (ra.byteswap)
1448                         backup_byteswap(drr);
1449
1450                 switch (drr->drr_type) {
1451                 case DRR_OBJECT:
1452                 {
1453                         /*
1454                          * We need to make a copy of the record header,
1455                          * because restore_{object,write} may need to
1456                          * restore_read(), which will invalidate drr.
1457                          */
1458                         struct drr_object drro = drr->drr_u.drr_object;
1459                         ra.err = restore_object(&ra, os, &drro);
1460                         break;
1461                 }
1462                 case DRR_FREEOBJECTS:
1463                 {
1464                         struct drr_freeobjects drrfo =
1465                             drr->drr_u.drr_freeobjects;
1466                         ra.err = restore_freeobjects(&ra, os, &drrfo);
1467                         break;
1468                 }
1469                 case DRR_WRITE:
1470                 {
1471                         struct drr_write drrw = drr->drr_u.drr_write;
1472                         ra.err = restore_write(&ra, os, &drrw);
1473                         break;
1474                 }
1475                 case DRR_WRITE_BYREF:
1476                 {
1477                         struct drr_write_byref drrwbr =
1478                             drr->drr_u.drr_write_byref;
1479                         ra.err = restore_write_byref(&ra, os, &drrwbr);
1480                         break;
1481                 }
1482                 case DRR_FREE:
1483                 {
1484                         struct drr_free drrf = drr->drr_u.drr_free;
1485                         ra.err = restore_free(&ra, os, &drrf);
1486                         break;
1487                 }
1488                 case DRR_END:
1489                 {
1490                         struct drr_end drre = drr->drr_u.drr_end;
1491                         /*
1492                          * We compare against the *previous* checksum
1493                          * value, because the stored checksum is of
1494                          * everything before the DRR_END record.
1495                          */
1496                         if (!ZIO_CHECKSUM_EQUAL(drre.drr_checksum, pcksum))
1497                                 ra.err = ECKSUM;
1498                         goto out;
1499                 }
1500                 case DRR_SPILL:
1501                 {
1502                         struct drr_spill drrs = drr->drr_u.drr_spill;
1503                         ra.err = restore_spill(&ra, os, &drrs);
1504                         break;
1505                 }
1506                 default:
1507                         ra.err = EINVAL;
1508                         goto out;
1509                 }
1510                 pcksum = ra.cksum;
1511         }
1512         ASSERT(ra.err != 0);
1513
1514 out:
1515         if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
1516                 zfs_onexit_fd_rele(cleanup_fd);
1517
1518         if (ra.err != 0) {
1519                 /*
1520                  * destroy what we created, so we don't leave it in the
1521                  * inconsistent restoring state.
1522                  */
1523                 txg_wait_synced(drc->drc_real_ds->ds_dir->dd_pool, 0);
1524
1525                 (void) dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag,
1526                     B_FALSE);
1527                 if (drc->drc_real_ds != drc->drc_logical_ds) {
1528                         mutex_exit(&drc->drc_logical_ds->ds_recvlock);
1529                         dsl_dataset_rele(drc->drc_logical_ds, dmu_recv_tag);
1530                 }
1531         }
1532
1533         vmem_free(ra.buf, ra.bufsize);
1534         *voffp = ra.voff;
1535         return (ra.err);
1536 }
1537
1538 struct recvendsyncarg {
1539         char *tosnap;
1540         uint64_t creation_time;
1541         uint64_t toguid;
1542 };
1543
1544 static int
1545 recv_end_check(void *arg1, void *arg2, dmu_tx_t *tx)
1546 {
1547         dsl_dataset_t *ds = arg1;
1548         struct recvendsyncarg *resa = arg2;
1549
1550         return (dsl_dataset_snapshot_check(ds, resa->tosnap, tx));
1551 }
1552
1553 static void
1554 recv_end_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1555 {
1556         dsl_dataset_t *ds = arg1;
1557         struct recvendsyncarg *resa = arg2;
1558
1559         dsl_dataset_snapshot_sync(ds, resa->tosnap, tx);
1560
1561         /* set snapshot's creation time and guid */
1562         dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
1563         ds->ds_prev->ds_phys->ds_creation_time = resa->creation_time;
1564         ds->ds_prev->ds_phys->ds_guid = resa->toguid;
1565         ds->ds_prev->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1566
1567         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1568         ds->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1569 }
1570
1571 static int
1572 add_ds_to_guidmap(avl_tree_t *guid_map, dsl_dataset_t *ds)
1573 {
1574         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1575         uint64_t snapobj = ds->ds_phys->ds_prev_snap_obj;
1576         dsl_dataset_t *snapds;
1577         guid_map_entry_t *gmep;
1578         int err;
1579
1580         ASSERT(guid_map != NULL);
1581
1582         rw_enter(&dp->dp_config_rwlock, RW_READER);
1583         err = dsl_dataset_hold_obj(dp, snapobj, guid_map, &snapds);
1584         if (err == 0) {
1585                 gmep = kmem_alloc(sizeof (guid_map_entry_t), KM_SLEEP);
1586                 gmep->guid = snapds->ds_phys->ds_guid;
1587                 gmep->gme_ds = snapds;
1588                 avl_add(guid_map, gmep);
1589         }
1590
1591         rw_exit(&dp->dp_config_rwlock);
1592         return (err);
1593 }
1594
1595 static int
1596 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
1597 {
1598         struct recvendsyncarg resa;
1599         dsl_dataset_t *ds = drc->drc_logical_ds;
1600         int err, myerr;
1601
1602         /*
1603          * XXX hack; seems the ds is still dirty and dsl_pool_zil_clean()
1604          * expects it to have a ds_user_ptr (and zil), but clone_swap()
1605          * can close it.
1606          */
1607         txg_wait_synced(ds->ds_dir->dd_pool, 0);
1608
1609         if (dsl_dataset_tryown(ds, FALSE, dmu_recv_tag)) {
1610                 err = dsl_dataset_clone_swap(drc->drc_real_ds, ds,
1611                     drc->drc_force);
1612                 if (err)
1613                         goto out;
1614         } else {
1615                 mutex_exit(&ds->ds_recvlock);
1616                 dsl_dataset_rele(ds, dmu_recv_tag);
1617                 (void) dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag,
1618                     B_FALSE);
1619                 return (EBUSY);
1620         }
1621
1622         resa.creation_time = drc->drc_drrb->drr_creation_time;
1623         resa.toguid = drc->drc_drrb->drr_toguid;
1624         resa.tosnap = drc->drc_tosnap;
1625
1626         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1627             recv_end_check, recv_end_sync, ds, &resa, 3);
1628         if (err) {
1629                 /* swap back */
1630                 (void) dsl_dataset_clone_swap(drc->drc_real_ds, ds, B_TRUE);
1631         }
1632
1633 out:
1634         mutex_exit(&ds->ds_recvlock);
1635         if (err == 0 && drc->drc_guid_to_ds_map != NULL)
1636                 (void) add_ds_to_guidmap(drc->drc_guid_to_ds_map, ds);
1637         dsl_dataset_disown(ds, dmu_recv_tag);
1638         myerr = dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag, B_FALSE);
1639         ASSERT3U(myerr, ==, 0);
1640         return (err);
1641 }
1642
1643 static int
1644 dmu_recv_new_end(dmu_recv_cookie_t *drc)
1645 {
1646         struct recvendsyncarg resa;
1647         dsl_dataset_t *ds = drc->drc_logical_ds;
1648         int err;
1649
1650         /*
1651          * XXX hack; seems the ds is still dirty and dsl_pool_zil_clean()
1652          * expects it to have a ds_user_ptr (and zil), but clone_swap()
1653          * can close it.
1654          */
1655         txg_wait_synced(ds->ds_dir->dd_pool, 0);
1656
1657         resa.creation_time = drc->drc_drrb->drr_creation_time;
1658         resa.toguid = drc->drc_drrb->drr_toguid;
1659         resa.tosnap = drc->drc_tosnap;
1660
1661         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1662             recv_end_check, recv_end_sync, ds, &resa, 3);
1663         if (err) {
1664                 /* clean up the fs we just recv'd into */
1665                 (void) dsl_dataset_destroy(ds, dmu_recv_tag, B_FALSE);
1666         } else {
1667                 if (drc->drc_guid_to_ds_map != NULL)
1668                         (void) add_ds_to_guidmap(drc->drc_guid_to_ds_map, ds);
1669                 /* release the hold from dmu_recv_begin */
1670                 dsl_dataset_disown(ds, dmu_recv_tag);
1671         }
1672         return (err);
1673 }
1674
1675 int
1676 dmu_recv_end(dmu_recv_cookie_t *drc)
1677 {
1678         if (drc->drc_logical_ds != drc->drc_real_ds)
1679                 return (dmu_recv_existing_end(drc));
1680         else
1681                 return (dmu_recv_new_end(drc));
1682 }