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  * Copyright (c) 2012, Joyent, Inc. All rights reserved.
  27  */
  28
  29 #include <sys/dmu.h>
  30 #include <sys/dmu_impl.h>
  31 #include <sys/dmu_tx.h>
  32 #include <sys/dbuf.h>
  33 #include <sys/dnode.h>
  34 #include <sys/zfs_context.h>
  35 #include <sys/dmu_objset.h>
  36 #include <sys/dmu_traverse.h>
  37 #include <sys/dsl_dataset.h>
  38 #include <sys/dsl_dir.h>
  39 #include <sys/dsl_prop.h>
  40 #include <sys/dsl_pool.h>
  41 #include <sys/dsl_synctask.h>
  42 #include <sys/zfs_ioctl.h>
  43 #include <sys/zap.h>
  44 #include <sys/zio_checksum.h>
  45 #include <sys/zfs_znode.h>
  46 #include <zfs_fletcher.h>
  47 #include <sys/avl.h>
  48 #include <sys/ddt.h>
  49 #include <sys/zfs_onexit.h>
  50
  51 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
  52 int zfs_send_corrupt_data = B_FALSE;
  53
  54 static char *dmu_recv_tag = "dmu_recv_tag";
  55
  56 static int
  57 dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
  58 {
  59         dsl_dataset_t *ds = dsp->dsa_os->os_dsl_dataset;
  60         ssize_t resid; /* have to get resid to get detailed errno */
  61         ASSERT3U(len % 8, ==, 0);
  62
  63         fletcher_4_incremental_native(buf, len, &dsp->dsa_zc);
  64         dsp->dsa_err = vn_rdwr(UIO_WRITE, dsp->dsa_vp,
  65             (caddr_t)buf, len,
  66             0, UIO_SYSSPACE, FAPPEND, RLIM64_INFINITY, CRED(), &resid);
  67
  68         mutex_enter(&ds->ds_sendstream_lock);
  69         *dsp->dsa_off += len;
  70         mutex_exit(&ds->ds_sendstream_lock);
  71
  72         return (dsp->dsa_err);
  73 }
  74
  75 static int
  76 dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
  77     uint64_t length)
  78 {
  79         struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free);
  80
  81         if (length != -1ULL && offset + length < offset)
  82                 length = -1ULL;
  83
  84         /*
  85          * If there is a pending op, but it's not PENDING_FREE, push it out,
  86          * since free block aggregation can only be done for blocks of the
  87          * same type (i.e., DRR_FREE records can only be aggregated with
  88          * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
  89          * aggregated with other DRR_FREEOBJECTS records.
  90          */
  91         if (dsp->dsa_pending_op != PENDING_NONE &&
  92             dsp->dsa_pending_op != PENDING_FREE) {
  93                 if (dump_bytes(dsp, dsp->dsa_drr,
  94                     sizeof (dmu_replay_record_t)) != 0)
  95                         return (EINTR);
  96                 dsp->dsa_pending_op = PENDING_NONE;
  97         }
  98
  99         if (dsp->dsa_pending_op == PENDING_FREE) {
 100                 /*
 101                  * There should never be a PENDING_FREE if length is -1
 102                  * (because dump_dnode is the only place where this
 103                  * function is called with a -1, and only after flushing
 104                  * any pending record).
 105                  */
 106                 ASSERT(length != -1ULL);
 107                 /*
 108                  * Check to see whether this free block can be aggregated
 109                  * with pending one.
 110                  */
 111                 if (drrf->drr_object == object && drrf->drr_offset +
 112                     drrf->drr_length == offset) {
 113                         drrf->drr_length += length;
 114                         return (0);
 115                 } else {
 116                         /* not a continuation.  Push out pending record */
 117                         if (dump_bytes(dsp, dsp->dsa_drr,
 118                             sizeof (dmu_replay_record_t)) != 0)
 119                                 return (EINTR);
 120                         dsp->dsa_pending_op = PENDING_NONE;
 121                 }
 122         }
 123         /* create a FREE record and make it pending */
 124         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 125         dsp->dsa_drr->drr_type = DRR_FREE;
 126         drrf->drr_object = object;
 127         drrf->drr_offset = offset;
 128         drrf->drr_length = length;
 129         drrf->drr_toguid = dsp->dsa_toguid;
 130         if (length == -1ULL) {
 131                 if (dump_bytes(dsp, dsp->dsa_drr,
 132                     sizeof (dmu_replay_record_t)) != 0)
 133                         return (EINTR);
 134         } else {
 135                 dsp->dsa_pending_op = PENDING_FREE;
 136         }
 137
 138         return (0);
 139 }
 140
 141 static int
 142 dump_data(dmu_sendarg_t *dsp, dmu_object_type_t type,
 143     uint64_t object, uint64_t offset, int blksz, const blkptr_t *bp, void *data)
 144 {
 145         struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write);
 146
 147
 148         /*
 149          * If there is any kind of pending aggregation (currently either
 150          * a grouping of free objects or free blocks), push it out to
 151          * the stream, since aggregation can't be done across operations
 152          * of different types.
 153          */
 154         if (dsp->dsa_pending_op != PENDING_NONE) {
 155                 if (dump_bytes(dsp, dsp->dsa_drr,
 156                     sizeof (dmu_replay_record_t)) != 0)
 157                         return (EINTR);
 158                 dsp->dsa_pending_op = PENDING_NONE;
 159         }
 160         /* write a DATA record */
 161         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 162         dsp->dsa_drr->drr_type = DRR_WRITE;
 163         drrw->drr_object = object;
 164         drrw->drr_type = type;
 165         drrw->drr_offset = offset;
 166         drrw->drr_length = blksz;
 167         drrw->drr_toguid = dsp->dsa_toguid;
 168         drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
 169         if (zio_checksum_table[drrw->drr_checksumtype].ci_dedup)
 170                 drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
 171         DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
 172         DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
 173         DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
 174         drrw->drr_key.ddk_cksum = bp->blk_cksum;
 175
 176         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
 177                 return (EINTR);
 178         if (dump_bytes(dsp, data, blksz) != 0)
 179                 return (EINTR);
 180         return (0);
 181 }
 182
 183 static int
 184 dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data)
 185 {
 186         struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill);
 187
 188         if (dsp->dsa_pending_op != PENDING_NONE) {
 189                 if (dump_bytes(dsp, dsp->dsa_drr,
 190                     sizeof (dmu_replay_record_t)) != 0)
 191                         return (EINTR);
 192                 dsp->dsa_pending_op = PENDING_NONE;
 193         }
 194
 195         /* write a SPILL record */
 196         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 197         dsp->dsa_drr->drr_type = DRR_SPILL;
 198         drrs->drr_object = object;
 199         drrs->drr_length = blksz;
 200         drrs->drr_toguid = dsp->dsa_toguid;
 201
 202         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)))
 203                 return (EINTR);
 204         if (dump_bytes(dsp, data, blksz))
 205                 return (EINTR);
 206         return (0);
 207 }
 208
 209 static int
 210 dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
 211 {
 212         struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
 213
 214         /*
 215          * If there is a pending op, but it's not PENDING_FREEOBJECTS,
 216          * push it out, since free block aggregation can only be done for
 217          * blocks of the same type (i.e., DRR_FREE records can only be
 218          * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
 219          * can only be aggregated with other DRR_FREEOBJECTS records.
 220          */
 221         if (dsp->dsa_pending_op != PENDING_NONE &&
 222             dsp->dsa_pending_op != PENDING_FREEOBJECTS) {
 223                 if (dump_bytes(dsp, dsp->dsa_drr,
 224                     sizeof (dmu_replay_record_t)) != 0)
 225                         return (EINTR);
 226                 dsp->dsa_pending_op = PENDING_NONE;
 227         }
 228         if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) {
 229                 /*
 230                  * See whether this free object array can be aggregated
 231                  * with pending one
 232                  */
 233                 if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
 234                         drrfo->drr_numobjs += numobjs;
 235                         return (0);
 236                 } else {
 237                         /* can't be aggregated.  Push out pending record */
 238                         if (dump_bytes(dsp, dsp->dsa_drr,
 239                             sizeof (dmu_replay_record_t)) != 0)
 240                                 return (EINTR);
 241                         dsp->dsa_pending_op = PENDING_NONE;
 242                 }
 243         }
 244
 245         /* write a FREEOBJECTS record */
 246         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 247         dsp->dsa_drr->drr_type = DRR_FREEOBJECTS;
 248         drrfo->drr_firstobj = firstobj;
 249         drrfo->drr_numobjs = numobjs;
 250         drrfo->drr_toguid = dsp->dsa_toguid;
 251
 252         dsp->dsa_pending_op = PENDING_FREEOBJECTS;
 253
 254         return (0);
 255 }
 256
 257 static int
 258 dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
 259 {
 260         struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object);
 261
 262         if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
 263                 return (dump_freeobjects(dsp, object, 1));
 264
 265         if (dsp->dsa_pending_op != PENDING_NONE) {
 266                 if (dump_bytes(dsp, dsp->dsa_drr,
 267                     sizeof (dmu_replay_record_t)) != 0)
 268                         return (EINTR);
 269                 dsp->dsa_pending_op = PENDING_NONE;
 270         }
 271
 272         /* write an OBJECT record */
 273         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 274         dsp->dsa_drr->drr_type = DRR_OBJECT;
 275         drro->drr_object = object;
 276         drro->drr_type = dnp->dn_type;
 277         drro->drr_bonustype = dnp->dn_bonustype;
 278         drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 279         drro->drr_bonuslen = dnp->dn_bonuslen;
 280         drro->drr_checksumtype = dnp->dn_checksum;
 281         drro->drr_compress = dnp->dn_compress;
 282         drro->drr_toguid = dsp->dsa_toguid;
 283
 284         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
 285                 return (EINTR);
 286
 287         if (dump_bytes(dsp, DN_BONUS(dnp), P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0)
 288                 return (EINTR);
 289
 290         /* free anything past the end of the file */
 291         if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) *
 292             (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL))
 293                 return (EINTR);
 294         if (dsp->dsa_err)
 295                 return (EINTR);
 296         return (0);
 297 }
 298
 299 #define BP_SPAN(dnp, level) \
 300         (((uint64_t)dnp->dn_datablkszsec) << (SPA_MINBLOCKSHIFT + \
 301         (level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)))
 302
 303 /* ARGSUSED */
 304 static int
 305 backup_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, arc_buf_t *pbuf,
 306     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
 307 {
 308         dmu_sendarg_t *dsp = arg;
 309         dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
 310         int err = 0;
 311
 312         if (issig(JUSTLOOKING) && issig(FORREAL))
 313                 return (EINTR);
 314
 315         if (zb->zb_object != DMU_META_DNODE_OBJECT &&
 316             DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
 317                 return (0);
 318         } else if (bp == NULL && zb->zb_object == DMU_META_DNODE_OBJECT) {
 319                 uint64_t span = BP_SPAN(dnp, zb->zb_level);
 320                 uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
 321                 err = dump_freeobjects(dsp, dnobj, span >> DNODE_SHIFT);
 322         } else if (bp == NULL) {
 323                 uint64_t span = BP_SPAN(dnp, zb->zb_level);
 324                 err = dump_free(dsp, zb->zb_object, zb->zb_blkid * span, span);
 325         } else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
 326                 return (0);
 327         } else if (type == DMU_OT_DNODE) {
 328                 dnode_phys_t *blk;
 329                 int i;
 330                 int blksz = BP_GET_LSIZE(bp);
 331                 uint32_t aflags = ARC_WAIT;
 332                 arc_buf_t *abuf;
 333
 334                 if (dsl_read(NULL, spa, bp, pbuf,
 335                     arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
 336                     ZIO_FLAG_CANFAIL, &aflags, zb) != 0)
 337                         return (EIO);
 338
 339                 blk = abuf->b_data;
 340                 for (i = 0; i < blksz >> DNODE_SHIFT; i++) {
 341                         uint64_t dnobj = (zb->zb_blkid <<
 342                             (DNODE_BLOCK_SHIFT - DNODE_SHIFT)) + i;
 343                         err = dump_dnode(dsp, dnobj, blk+i);
 344                         if (err)
 345                                 break;
 346                 }
 347                 (void) arc_buf_remove_ref(abuf, &abuf);
 348         } else if (type == DMU_OT_SA) {
 349                 uint32_t aflags = ARC_WAIT;
 350                 arc_buf_t *abuf;
 351                 int blksz = BP_GET_LSIZE(bp);
 352
 353                 if (arc_read_nolock(NULL, spa, bp,
 354                     arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
 355                     ZIO_FLAG_CANFAIL, &aflags, zb) != 0)
 356                         return (EIO);
 357
 358                 err = dump_spill(dsp, zb->zb_object, blksz, abuf->b_data);
 359                 (void) arc_buf_remove_ref(abuf, &abuf);
 360         } else { /* it's a level-0 block of a regular object */
 361                 uint32_t aflags = ARC_WAIT;
 362                 arc_buf_t *abuf;
 363                 int blksz = BP_GET_LSIZE(bp);
 364
 365                 if (dsl_read(NULL, spa, bp, pbuf,
 366                     arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
 367                     ZIO_FLAG_CANFAIL, &aflags, zb) != 0) {
 368                         if (zfs_send_corrupt_data) {
 369                                 uint64_t *ptr;
 370                                 /* Send a block filled with 0x"zfs badd bloc" */
 371                                 abuf = arc_buf_alloc(spa, blksz, &abuf,
 372                                     ARC_BUFC_DATA);
 373                                 for (ptr = abuf->b_data;
 374                                     (char *)ptr < (char *)abuf->b_data + blksz;
 375                                     ptr++)
 376                                         *ptr = 0x2f5baddb10c;
 377                         } else {
 378                                 return (EIO);
 379                         }
 380                 }
 381
 382                 err = dump_data(dsp, type, zb->zb_object, zb->zb_blkid * blksz,
 383                     blksz, bp, abuf->b_data);
 384                 (void) arc_buf_remove_ref(abuf, &abuf);
 385         }
 386
 387         ASSERT(err == 0 || err == EINTR);
 388         return (err);
 389 }
 390
 391 int
 392 dmu_send(objset_t *tosnap, objset_t *fromsnap, boolean_t fromorigin,
 393     int outfd, vnode_t *vp, offset_t *off)
 394 {
 395         dsl_dataset_t *ds = tosnap->os_dsl_dataset;
 396         dsl_dataset_t *fromds = fromsnap ? fromsnap->os_dsl_dataset : NULL;
 397         dmu_replay_record_t *drr;
 398         dmu_sendarg_t *dsp;
 399         int err;
 400         uint64_t fromtxg = 0;
 401
 402         /* tosnap must be a snapshot */
 403         if (ds->ds_phys->ds_next_snap_obj == 0)
 404                 return (EINVAL);
 405
 406         /* fromsnap must be an earlier snapshot from the same fs as tosnap */
 407         if (fromds && (ds->ds_dir != fromds->ds_dir ||
 408             fromds->ds_phys->ds_creation_txg >= ds->ds_phys->ds_creation_txg))
 409                 return (EXDEV);
 410
 411         if (fromorigin) {
 412                 dsl_pool_t *dp = ds->ds_dir->dd_pool;
 413
 414                 if (fromsnap)
 415                         return (EINVAL);
 416
 417                 if (dsl_dir_is_clone(ds->ds_dir)) {
 418                         rw_enter(&dp->dp_config_rwlock, RW_READER);
 419                         err = dsl_dataset_hold_obj(dp,
 420                             ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &fromds);
 421                         rw_exit(&dp->dp_config_rwlock);
 422                         if (err)
 423                                 return (err);
 424                 } else {
 425                         fromorigin = B_FALSE;
 426                 }
 427         }
 428
 429
 430         drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
 431         drr->drr_type = DRR_BEGIN;
 432         drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
 433         DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
 434             DMU_SUBSTREAM);
 435
 436 #ifdef _KERNEL
 437         if (dmu_objset_type(tosnap) == DMU_OST_ZFS) {
 438                 uint64_t version;
 439                 if (zfs_get_zplprop(tosnap, ZFS_PROP_VERSION, &version) != 0) {
 440                         kmem_free(drr, sizeof (dmu_replay_record_t));
 441                         return (EINVAL);
 442                 }
 443                 if (version == ZPL_VERSION_SA) {
 444                         DMU_SET_FEATUREFLAGS(
 445                             drr->drr_u.drr_begin.drr_versioninfo,
 446                             DMU_BACKUP_FEATURE_SA_SPILL);
 447                 }
 448         }
 449 #endif
 450
 451         drr->drr_u.drr_begin.drr_creation_time =
 452             ds->ds_phys->ds_creation_time;
 453         drr->drr_u.drr_begin.drr_type = tosnap->os_phys->os_type;
 454         if (fromorigin)
 455                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
 456         drr->drr_u.drr_begin.drr_toguid = ds->ds_phys->ds_guid;
 457         if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
 458                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
 459
 460         if (fromds)
 461                 drr->drr_u.drr_begin.drr_fromguid = fromds->ds_phys->ds_guid;
 462         dsl_dataset_name(ds, drr->drr_u.drr_begin.drr_toname);
 463
 464         if (fromds)
 465                 fromtxg = fromds->ds_phys->ds_creation_txg;
 466         if (fromorigin)
 467                 dsl_dataset_rele(fromds, FTAG);
 468
 469         dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP);
 470
 471         dsp->dsa_drr = drr;
 472         dsp->dsa_vp = vp;
 473         dsp->dsa_outfd = outfd;
 474         dsp->dsa_proc = curproc;
 475         dsp->dsa_os = tosnap;
 476         dsp->dsa_off = off;
 477         dsp->dsa_toguid = ds->ds_phys->ds_guid;
 478         ZIO_SET_CHECKSUM(&dsp->dsa_zc, 0, 0, 0, 0);
 479         dsp->dsa_pending_op = PENDING_NONE;
 480
 481         mutex_enter(&ds->ds_sendstream_lock);
 482         list_insert_head(&ds->ds_sendstreams, dsp);
 483         mutex_exit(&ds->ds_sendstream_lock);
 484
 485         if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0) {
 486                 err = dsp->dsa_err;
 487                 goto out;
 488         }
 489
 490         err = traverse_dataset(ds, fromtxg, TRAVERSE_PRE | TRAVERSE_PREFETCH,
 491             backup_cb, dsp);
 492
 493         if (dsp->dsa_pending_op != PENDING_NONE)
 494                 if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0)
 495                         err = EINTR;
 496
 497         if (err) {
 498                 if (err == EINTR && dsp->dsa_err)
 499                         err = dsp->dsa_err;
 500                 goto out;
 501         }
 502
 503         bzero(drr, sizeof (dmu_replay_record_t));
 504         drr->drr_type = DRR_END;
 505         drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc;
 506         drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid;
 507
 508         if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0) {
 509                 err = dsp->dsa_err;
 510                 goto out;
 511         }
 512
 513 out:
 514         mutex_enter(&ds->ds_sendstream_lock);
 515         list_remove(&ds->ds_sendstreams, dsp);
 516         mutex_exit(&ds->ds_sendstream_lock);
 517
 518         kmem_free(drr, sizeof (dmu_replay_record_t));
 519         kmem_free(dsp, sizeof (dmu_sendarg_t));
 520
 521         return (err);
 522 }
 523
 524 int
 525 dmu_send_estimate(objset_t *tosnap, objset_t *fromsnap, boolean_t fromorigin,
 526     uint64_t *sizep)
 527 {
 528         dsl_dataset_t *ds = tosnap->os_dsl_dataset;
 529         dsl_dataset_t *fromds = fromsnap ? fromsnap->os_dsl_dataset : NULL;
 530         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 531         int err;
 532         uint64_t size, recordsize;
 533
 534         /* tosnap must be a snapshot */
 535         if (ds->ds_phys->ds_next_snap_obj == 0)
 536                 return (EINVAL);
 537
 538         /* fromsnap must be an earlier snapshot from the same fs as tosnap */
 539         if (fromds && (ds->ds_dir != fromds->ds_dir ||
 540             fromds->ds_phys->ds_creation_txg >= ds->ds_phys->ds_creation_txg))
 541                 return (EXDEV);
 542
 543         if (fromorigin) {
 544                 if (fromsnap)
 545                         return (EINVAL);
 546
 547                 if (dsl_dir_is_clone(ds->ds_dir)) {
 548                         rw_enter(&dp->dp_config_rwlock, RW_READER);
 549                         err = dsl_dataset_hold_obj(dp,
 550                             ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &fromds);
 551                         rw_exit(&dp->dp_config_rwlock);
 552                         if (err)
 553                                 return (err);
 554                 } else {
 555                         fromorigin = B_FALSE;
 556                 }
 557         }
 558
 559         /* Get uncompressed size estimate of changed data. */
 560         if (fromds == NULL) {
 561                 size = ds->ds_phys->ds_uncompressed_bytes;
 562         } else {
 563                 uint64_t used, comp;
 564                 err = dsl_dataset_space_written(fromds, ds,
 565                     &used, &comp, &size);
 566                 if (fromorigin)
 567                         dsl_dataset_rele(fromds, FTAG);
 568                 if (err)
 569                         return (err);
 570         }
 571
 572         /*
 573          * Assume that space (both on-disk and in-stream) is dominated by
 574          * data.  We will adjust for indirect blocks and the copies property,
 575          * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
 576          */
 577
 578         /*
 579          * Subtract out approximate space used by indirect blocks.
 580          * Assume most space is used by data blocks (non-indirect, non-dnode).
 581          * Assume all blocks are recordsize.  Assume ditto blocks and
 582          * internal fragmentation counter out compression.
 583          *
 584          * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
 585          * block, which we observe in practice.
 586          */
 587         rw_enter(&dp->dp_config_rwlock, RW_READER);
 588         err = dsl_prop_get_ds(ds, "recordsize",
 589             sizeof (recordsize), 1, &recordsize, NULL);
 590         rw_exit(&dp->dp_config_rwlock);
 591         if (err)
 592                 return (err);
 593         size -= size / recordsize * sizeof (blkptr_t);
 594
 595         /* Add in the space for the record associated with each block. */
 596         size += size / recordsize * sizeof (dmu_replay_record_t);
 597
 598         *sizep = size;
 599
 600         return (0);
 601 }
 602
 603 struct recvbeginsyncarg {
 604         const char *tofs;
 605         const char *tosnap;
 606         dsl_dataset_t *origin;
 607         uint64_t fromguid;
 608         dmu_objset_type_t type;
 609         void *tag;
 610         boolean_t force;
 611         uint64_t dsflags;
 612         char clonelastname[MAXNAMELEN];
 613         dsl_dataset_t *ds; /* the ds to recv into; returned from the syncfunc */
 614         cred_t *cr;
 615 };
 616
 617 /* ARGSUSED */
 618 static int
 619 recv_new_check(void *arg1, void *arg2, dmu_tx_t *tx)
 620 {
 621         dsl_dir_t *dd = arg1;
 622         struct recvbeginsyncarg *rbsa = arg2;
 623         objset_t *mos = dd->dd_pool->dp_meta_objset;
 624         uint64_t val;
 625         int err;
 626
 627         err = zap_lookup(mos, dd->dd_phys->dd_child_dir_zapobj,
 628             strrchr(rbsa->tofs, '/') + 1, sizeof (uint64_t), 1, &val);
 629
 630         if (err != ENOENT)
 631                 return (err ? err : EEXIST);
 632
 633         if (rbsa->origin) {
 634                 /* make sure it's a snap in the same pool */
 635                 if (rbsa->origin->ds_dir->dd_pool != dd->dd_pool)
 636                         return (EXDEV);
 637                 if (!dsl_dataset_is_snapshot(rbsa->origin))
 638                         return (EINVAL);
 639                 if (rbsa->origin->ds_phys->ds_guid != rbsa->fromguid)
 640                         return (ENODEV);
 641         }
 642
 643         return (0);
 644 }
 645
 646 static void
 647 recv_new_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 648 {
 649         dsl_dir_t *dd = arg1;
 650         struct recvbeginsyncarg *rbsa = arg2;
 651         uint64_t flags = DS_FLAG_INCONSISTENT | rbsa->dsflags;
 652         uint64_t dsobj;
 653
 654         /* Create and open new dataset. */
 655         dsobj = dsl_dataset_create_sync(dd, strrchr(rbsa->tofs, '/') + 1,
 656             rbsa->origin, flags, rbsa->cr, tx);
 657         VERIFY(0 == dsl_dataset_own_obj(dd->dd_pool, dsobj,
 658             B_TRUE, dmu_recv_tag, &rbsa->ds));
 659
 660         if (rbsa->origin == NULL) {
 661                 (void) dmu_objset_create_impl(dd->dd_pool->dp_spa,
 662                     rbsa->ds, &rbsa->ds->ds_phys->ds_bp, rbsa->type, tx);
 663         }
 664
 665         spa_history_log_internal(LOG_DS_REPLAY_FULL_SYNC,
 666             dd->dd_pool->dp_spa, tx, "dataset = %lld", dsobj);
 667 }
 668
 669 /* ARGSUSED */
 670 static int
 671 recv_existing_check(void *arg1, void *arg2, dmu_tx_t *tx)
 672 {
 673         dsl_dataset_t *ds = arg1;
 674         struct recvbeginsyncarg *rbsa = arg2;
 675         int err;
 676         uint64_t val;
 677
 678         /* must not have any changes since most recent snapshot */
 679         if (!rbsa->force && dsl_dataset_modified_since_lastsnap(ds))
 680                 return (ETXTBSY);
 681
 682         /* new snapshot name must not exist */
 683         err = zap_lookup(ds->ds_dir->dd_pool->dp_meta_objset,
 684             ds->ds_phys->ds_snapnames_zapobj, rbsa->tosnap, 8, 1, &val);
 685         if (err == 0)
 686                 return (EEXIST);
 687         if (err != ENOENT)
 688                 return (err);
 689
 690         if (rbsa->fromguid) {
 691                 /* if incremental, most recent snapshot must match fromguid */
 692                 if (ds->ds_prev == NULL)
 693                         return (ENODEV);
 694
 695                 /*
 696                  * most recent snapshot must match fromguid, or there are no
 697                  * changes since the fromguid one
 698                  */
 699                 if (ds->ds_prev->ds_phys->ds_guid != rbsa->fromguid) {
 700                         uint64_t birth = ds->ds_prev->ds_phys->ds_bp.blk_birth;
 701                         uint64_t obj = ds->ds_prev->ds_phys->ds_prev_snap_obj;
 702                         while (obj != 0) {
 703                                 dsl_dataset_t *snap;
 704                                 err = dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
 705                                     obj, FTAG, &snap);
 706                                 if (err)
 707                                         return (ENODEV);
 708                                 if (snap->ds_phys->ds_creation_txg < birth) {
 709                                         dsl_dataset_rele(snap, FTAG);
 710                                         return (ENODEV);
 711                                 }
 712                                 if (snap->ds_phys->ds_guid == rbsa->fromguid) {
 713                                         dsl_dataset_rele(snap, FTAG);
 714                                         break; /* it's ok */
 715                                 }
 716                                 obj = snap->ds_phys->ds_prev_snap_obj;
 717                                 dsl_dataset_rele(snap, FTAG);
 718                         }
 719                         if (obj == 0)
 720                                 return (ENODEV);
 721                 }
 722         } else {
 723                 /* if full, most recent snapshot must be $ORIGIN */
 724                 if (ds->ds_phys->ds_prev_snap_txg >= TXG_INITIAL)
 725                         return (ENODEV);
 726         }
 727
 728         /* temporary clone name must not exist */
 729         err = zap_lookup(ds->ds_dir->dd_pool->dp_meta_objset,
 730             ds->ds_dir->dd_phys->dd_child_dir_zapobj,
 731             rbsa->clonelastname, 8, 1, &val);
 732         if (err == 0)
 733                 return (EEXIST);
 734         if (err != ENOENT)
 735                 return (err);
 736
 737         return (0);
 738 }
 739
 740 /* ARGSUSED */
 741 static void
 742 recv_existing_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 743 {
 744         dsl_dataset_t *ohds = arg1;
 745         struct recvbeginsyncarg *rbsa = arg2;
 746         dsl_pool_t *dp = ohds->ds_dir->dd_pool;
 747         dsl_dataset_t *cds;
 748         uint64_t flags = DS_FLAG_INCONSISTENT | rbsa->dsflags;
 749         uint64_t dsobj;
 750
 751         /* create and open the temporary clone */
 752         dsobj = dsl_dataset_create_sync(ohds->ds_dir, rbsa->clonelastname,
 753             ohds->ds_prev, flags, rbsa->cr, tx);
 754         VERIFY(0 == dsl_dataset_own_obj(dp, dsobj, B_TRUE, dmu_recv_tag, &cds));
 755
 756         /*
 757          * If we actually created a non-clone, we need to create the
 758          * objset in our new dataset.
 759          */
 760         if (BP_IS_HOLE(dsl_dataset_get_blkptr(cds))) {
 761                 (void) dmu_objset_create_impl(dp->dp_spa,
 762                     cds, dsl_dataset_get_blkptr(cds), rbsa->type, tx);
 763         }
 764
 765         rbsa->ds = cds;
 766
 767         spa_history_log_internal(LOG_DS_REPLAY_INC_SYNC,
 768             dp->dp_spa, tx, "dataset = %lld", dsobj);
 769 }
 770
 771 static boolean_t
 772 dmu_recv_verify_features(dsl_dataset_t *ds, struct drr_begin *drrb)
 773 {
 774         int featureflags;
 775
 776         featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
 777
 778         /* Verify pool version supports SA if SA_SPILL feature set */
 779         return ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
 780             (spa_version(dsl_dataset_get_spa(ds)) < SPA_VERSION_SA));
 781 }
 782
 783 /*
 784  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
 785  * succeeds; otherwise we will leak the holds on the datasets.
 786  */
 787 int
 788 dmu_recv_begin(char *tofs, char *tosnap, char *top_ds, struct drr_begin *drrb,
 789     boolean_t force, objset_t *origin, dmu_recv_cookie_t *drc)
 790 {
 791         int err = 0;
 792         boolean_t byteswap;
 793         struct recvbeginsyncarg rbsa = { 0 };
 794         uint64_t versioninfo;
 795         int flags;
 796         dsl_dataset_t *ds;
 797
 798         if (drrb->drr_magic == DMU_BACKUP_MAGIC)
 799                 byteswap = FALSE;
 800         else if (drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
 801                 byteswap = TRUE;
 802         else
 803                 return (EINVAL);
 804
 805         rbsa.tofs = tofs;
 806         rbsa.tosnap = tosnap;
 807         rbsa.origin = origin ? origin->os_dsl_dataset : NULL;
 808         rbsa.fromguid = drrb->drr_fromguid;
 809         rbsa.type = drrb->drr_type;
 810         rbsa.tag = FTAG;
 811         rbsa.dsflags = 0;
 812         rbsa.cr = CRED();
 813         versioninfo = drrb->drr_versioninfo;
 814         flags = drrb->drr_flags;
 815
 816         if (byteswap) {
 817                 rbsa.type = BSWAP_32(rbsa.type);
 818                 rbsa.fromguid = BSWAP_64(rbsa.fromguid);
 819                 versioninfo = BSWAP_64(versioninfo);
 820                 flags = BSWAP_32(flags);
 821         }
 822
 823         if (DMU_GET_STREAM_HDRTYPE(versioninfo) == DMU_COMPOUNDSTREAM ||
 824             rbsa.type >= DMU_OST_NUMTYPES ||
 825             ((flags & DRR_FLAG_CLONE) && origin == NULL))
 826                 return (EINVAL);
 827
 828         if (flags & DRR_FLAG_CI_DATA)
 829                 rbsa.dsflags = DS_FLAG_CI_DATASET;
 830
 831         bzero(drc, sizeof (dmu_recv_cookie_t));
 832         drc->drc_drrb = drrb;
 833         drc->drc_tosnap = tosnap;
 834         drc->drc_top_ds = top_ds;
 835         drc->drc_force = force;
 836
 837         /*
 838          * Process the begin in syncing context.
 839          */
 840
 841         /* open the dataset we are logically receiving into */
 842         err = dsl_dataset_hold(tofs, dmu_recv_tag, &ds);
 843         if (err == 0) {
 844                 if (dmu_recv_verify_features(ds, drrb)) {
 845                         dsl_dataset_rele(ds, dmu_recv_tag);
 846                         return (ENOTSUP);
 847                 }
 848                 /* target fs already exists; recv into temp clone */
 849
 850                 /* Can't recv a clone into an existing fs */
 851                 if (flags & DRR_FLAG_CLONE) {
 852                         dsl_dataset_rele(ds, dmu_recv_tag);
 853                         return (EINVAL);
 854                 }
 855
 856                 /* must not have an incremental recv already in progress */
 857                 if (!mutex_tryenter(&ds->ds_recvlock)) {
 858                         dsl_dataset_rele(ds, dmu_recv_tag);
 859                         return (EBUSY);
 860                 }
 861
 862                 /* tmp clone name is: tofs/%tosnap" */
 863                 (void) snprintf(rbsa.clonelastname, sizeof (rbsa.clonelastname),
 864                     "%%%s", tosnap);
 865                 rbsa.force = force;
 866                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
 867                     recv_existing_check, recv_existing_sync, ds, &rbsa, 5);
 868                 if (err) {
 869                         mutex_exit(&ds->ds_recvlock);
 870                         dsl_dataset_rele(ds, dmu_recv_tag);
 871                         return (err);
 872                 }
 873                 drc->drc_logical_ds = ds;
 874                 drc->drc_real_ds = rbsa.ds;
 875         } else if (err == ENOENT) {
 876                 /* target fs does not exist; must be a full backup or clone */
 877                 char *cp;
 878
 879                 /*
 880                  * If it's a non-clone incremental, we are missing the
 881                  * target fs, so fail the recv.
 882                  */
 883                 if (rbsa.fromguid && !(flags & DRR_FLAG_CLONE))
 884                         return (ENOENT);
 885
 886                 /* Open the parent of tofs */
 887                 cp = strrchr(tofs, '/');
 888                 *cp = '\0';
 889                 err = dsl_dataset_hold(tofs, FTAG, &ds);
 890                 *cp = '/';
 891                 if (err)
 892                         return (err);
 893
 894                 if (dmu_recv_verify_features(ds, drrb)) {
 895                         dsl_dataset_rele(ds, FTAG);
 896                         return (ENOTSUP);
 897                 }
 898
 899                 err = dsl_sync_task_do(ds->ds_dir->dd_pool,
 900                     recv_new_check, recv_new_sync, ds->ds_dir, &rbsa, 5);
 901                 dsl_dataset_rele(ds, FTAG);
 902                 if (err)
 903                         return (err);
 904                 drc->drc_logical_ds = drc->drc_real_ds = rbsa.ds;
 905                 drc->drc_newfs = B_TRUE;
 906         }
 907
 908         return (err);
 909 }
 910
 911 struct restorearg {
 912         int err;
 913         int byteswap;
 914         vnode_t *vp;
 915         char *buf;
 916         uint64_t voff;
 917         int bufsize; /* amount of memory allocated for buf */
 918         zio_cksum_t cksum;
 919         avl_tree_t *guid_to_ds_map;
 920 };
 921
 922 typedef struct guid_map_entry {
 923         uint64_t        guid;
 924         dsl_dataset_t   *gme_ds;
 925         avl_node_t      avlnode;
 926 } guid_map_entry_t;
 927
 928 static int
 929 guid_compare(const void *arg1, const void *arg2)
 930 {
 931         const guid_map_entry_t *gmep1 = arg1;
 932         const guid_map_entry_t *gmep2 = arg2;
 933
 934         if (gmep1->guid < gmep2->guid)
 935                 return (-1);
 936         else if (gmep1->guid > gmep2->guid)
 937                 return (1);
 938         return (0);
 939 }
 940
 941 static void
 942 free_guid_map_onexit(void *arg)
 943 {
 944         avl_tree_t *ca = arg;
 945         void *cookie = NULL;
 946         guid_map_entry_t *gmep;
 947
 948         while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
 949                 dsl_dataset_rele(gmep->gme_ds, ca);
 950                 kmem_free(gmep, sizeof (guid_map_entry_t));
 951         }
 952         avl_destroy(ca);
 953         kmem_free(ca, sizeof (avl_tree_t));
 954 }
 955
 956 static void *
 957 restore_read(struct restorearg *ra, int len)
 958 {
 959         void *rv;
 960         int done = 0;
 961
 962         /* some things will require 8-byte alignment, so everything must */
 963         ASSERT3U(len % 8, ==, 0);
 964
 965         while (done < len) {
 966                 ssize_t resid;
 967
 968                 ra->err = vn_rdwr(UIO_READ, ra->vp,
 969                     (caddr_t)ra->buf + done, len - done,
 970                     ra->voff, UIO_SYSSPACE, FAPPEND,
 971                     RLIM64_INFINITY, CRED(), &resid);
 972
 973                 if (resid == len - done)
 974                         ra->err = EINVAL;
 975                 ra->voff += len - done - resid;
 976                 done = len - resid;
 977                 if (ra->err)
 978                         return (NULL);
 979         }
 980
 981         ASSERT3U(done, ==, len);
 982         rv = ra->buf;
 983         if (ra->byteswap)
 984                 fletcher_4_incremental_byteswap(rv, len, &ra->cksum);
 985         else
 986                 fletcher_4_incremental_native(rv, len, &ra->cksum);
 987         return (rv);
 988 }
 989
 990 noinline static void
 991 backup_byteswap(dmu_replay_record_t *drr)
 992 {
 993 #define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
 994 #define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
 995         drr->drr_type = BSWAP_32(drr->drr_type);
 996         drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
 997         switch (drr->drr_type) {
 998         case DRR_BEGIN:
 999                 DO64(drr_begin.drr_magic);
1000                 DO64(drr_begin.drr_versioninfo);
1001                 DO64(drr_begin.drr_creation_time);
1002                 DO32(drr_begin.drr_type);
1003                 DO32(drr_begin.drr_flags);
1004                 DO64(drr_begin.drr_toguid);
1005                 DO64(drr_begin.drr_fromguid);
1006                 break;
1007         case DRR_OBJECT:
1008                 DO64(drr_object.drr_object);
1009                 /* DO64(drr_object.drr_allocation_txg); */
1010                 DO32(drr_object.drr_type);
1011                 DO32(drr_object.drr_bonustype);
1012                 DO32(drr_object.drr_blksz);
1013                 DO32(drr_object.drr_bonuslen);
1014                 DO64(drr_object.drr_toguid);
1015                 break;
1016         case DRR_FREEOBJECTS:
1017                 DO64(drr_freeobjects.drr_firstobj);
1018                 DO64(drr_freeobjects.drr_numobjs);
1019                 DO64(drr_freeobjects.drr_toguid);
1020                 break;
1021         case DRR_WRITE:
1022                 DO64(drr_write.drr_object);
1023                 DO32(drr_write.drr_type);
1024                 DO64(drr_write.drr_offset);
1025                 DO64(drr_write.drr_length);
1026                 DO64(drr_write.drr_toguid);
1027                 DO64(drr_write.drr_key.ddk_cksum.zc_word[0]);
1028                 DO64(drr_write.drr_key.ddk_cksum.zc_word[1]);
1029                 DO64(drr_write.drr_key.ddk_cksum.zc_word[2]);
1030                 DO64(drr_write.drr_key.ddk_cksum.zc_word[3]);
1031                 DO64(drr_write.drr_key.ddk_prop);
1032                 break;
1033         case DRR_WRITE_BYREF:
1034                 DO64(drr_write_byref.drr_object);
1035                 DO64(drr_write_byref.drr_offset);
1036                 DO64(drr_write_byref.drr_length);
1037                 DO64(drr_write_byref.drr_toguid);
1038                 DO64(drr_write_byref.drr_refguid);
1039                 DO64(drr_write_byref.drr_refobject);
1040                 DO64(drr_write_byref.drr_refoffset);
1041                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[0]);
1042                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[1]);
1043                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[2]);
1044                 DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[3]);
1045                 DO64(drr_write_byref.drr_key.ddk_prop);
1046                 break;
1047         case DRR_FREE:
1048                 DO64(drr_free.drr_object);
1049                 DO64(drr_free.drr_offset);
1050                 DO64(drr_free.drr_length);
1051                 DO64(drr_free.drr_toguid);
1052                 break;
1053         case DRR_SPILL:
1054                 DO64(drr_spill.drr_object);
1055                 DO64(drr_spill.drr_length);
1056                 DO64(drr_spill.drr_toguid);
1057                 break;
1058         case DRR_END:
1059                 DO64(drr_end.drr_checksum.zc_word[0]);
1060                 DO64(drr_end.drr_checksum.zc_word[1]);
1061                 DO64(drr_end.drr_checksum.zc_word[2]);
1062                 DO64(drr_end.drr_checksum.zc_word[3]);
1063                 DO64(drr_end.drr_toguid);
1064                 break;
1065         default:
1066                 break;
1067         }
1068 #undef DO64
1069 #undef DO32
1070 }
1071
1072 noinline static int
1073 restore_object(struct restorearg *ra, objset_t *os, struct drr_object *drro)
1074 {
1075         int err;
1076         dmu_tx_t *tx;
1077         void *data = NULL;
1078
1079         if (drro->drr_type == DMU_OT_NONE ||
1080             drro->drr_type >= DMU_OT_NUMTYPES ||
1081             drro->drr_bonustype >= DMU_OT_NUMTYPES ||
1082             drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
1083             drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
1084             P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
1085             drro->drr_blksz < SPA_MINBLOCKSIZE ||
1086             drro->drr_blksz > SPA_MAXBLOCKSIZE ||
1087             drro->drr_bonuslen > DN_MAX_BONUSLEN) {
1088                 return (EINVAL);
1089         }
1090
1091         err = dmu_object_info(os, drro->drr_object, NULL);
1092
1093         if (err != 0 && err != ENOENT)
1094                 return (EINVAL);
1095
1096         if (drro->drr_bonuslen) {
1097                 data = restore_read(ra, P2ROUNDUP(drro->drr_bonuslen, 8));
1098                 if (ra->err)
1099                         return (ra->err);
1100         }
1101
1102         if (err == ENOENT) {
1103                 /* currently free, want to be allocated */
1104                 tx = dmu_tx_create(os);
1105                 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1106                 err = dmu_tx_assign(tx, TXG_WAIT);
1107                 if (err) {
1108                         dmu_tx_abort(tx);
1109                         return (err);
1110                 }
1111                 err = dmu_object_claim(os, drro->drr_object,
1112                     drro->drr_type, drro->drr_blksz,
1113                     drro->drr_bonustype, drro->drr_bonuslen, tx);
1114                 dmu_tx_commit(tx);
1115         } else {
1116                 /* currently allocated, want to be allocated */
1117                 err = dmu_object_reclaim(os, drro->drr_object,
1118                     drro->drr_type, drro->drr_blksz,
1119                     drro->drr_bonustype, drro->drr_bonuslen);
1120         }
1121         if (err) {
1122                 return (EINVAL);
1123         }
1124
1125         tx = dmu_tx_create(os);
1126         dmu_tx_hold_bonus(tx, drro->drr_object);
1127         err = dmu_tx_assign(tx, TXG_WAIT);
1128         if (err) {
1129                 dmu_tx_abort(tx);
1130                 return (err);
1131         }
1132
1133         dmu_object_set_checksum(os, drro->drr_object, drro->drr_checksumtype,
1134             tx);
1135         dmu_object_set_compress(os, drro->drr_object, drro->drr_compress, tx);
1136
1137         if (data != NULL) {
1138                 dmu_buf_t *db;
1139
1140                 VERIFY(0 == dmu_bonus_hold(os, drro->drr_object, FTAG, &db));
1141                 dmu_buf_will_dirty(db, tx);
1142
1143                 ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
1144                 bcopy(data, db->db_data, drro->drr_bonuslen);
1145                 if (ra->byteswap) {
1146                         dmu_ot[drro->drr_bonustype].ot_byteswap(db->db_data,
1147                             drro->drr_bonuslen);
1148                 }
1149                 dmu_buf_rele(db, FTAG);
1150         }
1151         dmu_tx_commit(tx);
1152         return (0);
1153 }
1154
1155 /* ARGSUSED */
1156 noinline static int
1157 restore_freeobjects(struct restorearg *ra, objset_t *os,
1158     struct drr_freeobjects *drrfo)
1159 {
1160         uint64_t obj;
1161
1162         if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
1163                 return (EINVAL);
1164
1165         for (obj = drrfo->drr_firstobj;
1166             obj < drrfo->drr_firstobj + drrfo->drr_numobjs;
1167             (void) dmu_object_next(os, &obj, FALSE, 0)) {
1168                 int err;
1169
1170                 if (dmu_object_info(os, obj, NULL) != 0)
1171                         continue;
1172
1173                 err = dmu_free_object(os, obj);
1174                 if (err)
1175                         return (err);
1176         }
1177         return (0);
1178 }
1179
1180 noinline static int
1181 restore_write(struct restorearg *ra, objset_t *os,
1182     struct drr_write *drrw)
1183 {
1184         dmu_tx_t *tx;
1185         void *data;
1186         int err;
1187
1188         if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
1189             drrw->drr_type >= DMU_OT_NUMTYPES)
1190                 return (EINVAL);
1191
1192         data = restore_read(ra, drrw->drr_length);
1193         if (data == NULL)
1194                 return (ra->err);
1195
1196         if (dmu_object_info(os, drrw->drr_object, NULL) != 0)
1197                 return (EINVAL);
1198
1199         tx = dmu_tx_create(os);
1200
1201         dmu_tx_hold_write(tx, drrw->drr_object,
1202             drrw->drr_offset, drrw->drr_length);
1203         err = dmu_tx_assign(tx, TXG_WAIT);
1204         if (err) {
1205                 dmu_tx_abort(tx);
1206                 return (err);
1207         }
1208         if (ra->byteswap)
1209                 dmu_ot[drrw->drr_type].ot_byteswap(data, drrw->drr_length);
1210         dmu_write(os, drrw->drr_object,
1211             drrw->drr_offset, drrw->drr_length, data, tx);
1212         dmu_tx_commit(tx);
1213         return (0);
1214 }
1215
1216 /*
1217  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
1218  * streams to refer to a copy of the data that is already on the
1219  * system because it came in earlier in the stream.  This function
1220  * finds the earlier copy of the data, and uses that copy instead of
1221  * data from the stream to fulfill this write.
1222  */
1223 static int
1224 restore_write_byref(struct restorearg *ra, objset_t *os,
1225     struct drr_write_byref *drrwbr)
1226 {
1227         dmu_tx_t *tx;
1228         int err;
1229         guid_map_entry_t gmesrch;
1230         guid_map_entry_t *gmep;
1231         avl_index_t     where;
1232         objset_t *ref_os = NULL;
1233         dmu_buf_t *dbp;
1234
1235         if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
1236                 return (EINVAL);
1237
1238         /*
1239          * If the GUID of the referenced dataset is different from the
1240          * GUID of the target dataset, find the referenced dataset.
1241          */
1242         if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
1243                 gmesrch.guid = drrwbr->drr_refguid;
1244                 if ((gmep = avl_find(ra->guid_to_ds_map, &gmesrch,
1245                     &where)) == NULL) {
1246                         return (EINVAL);
1247                 }
1248                 if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
1249                         return (EINVAL);
1250         } else {
1251                 ref_os = os;
1252         }
1253
1254         err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
1255             drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH);
1256         if (err)
1257                 return (err);
1258
1259         tx = dmu_tx_create(os);
1260
1261         dmu_tx_hold_write(tx, drrwbr->drr_object,
1262             drrwbr->drr_offset, drrwbr->drr_length);
1263         err = dmu_tx_assign(tx, TXG_WAIT);
1264         if (err) {
1265                 dmu_tx_abort(tx);
1266                 return (err);
1267         }
1268         dmu_write(os, drrwbr->drr_object,
1269             drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
1270         dmu_buf_rele(dbp, FTAG);
1271         dmu_tx_commit(tx);
1272         return (0);
1273 }
1274
1275 static int
1276 restore_spill(struct restorearg *ra, objset_t *os, struct drr_spill *drrs)
1277 {
1278         dmu_tx_t *tx;
1279         void *data;
1280         dmu_buf_t *db, *db_spill;
1281         int err;
1282
1283         if (drrs->drr_length < SPA_MINBLOCKSIZE ||
1284             drrs->drr_length > SPA_MAXBLOCKSIZE)
1285                 return (EINVAL);
1286
1287         data = restore_read(ra, drrs->drr_length);
1288         if (data == NULL)
1289                 return (ra->err);
1290
1291         if (dmu_object_info(os, drrs->drr_object, NULL) != 0)
1292                 return (EINVAL);
1293
1294         VERIFY(0 == dmu_bonus_hold(os, drrs->drr_object, FTAG, &db));
1295         if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
1296                 dmu_buf_rele(db, FTAG);
1297                 return (err);
1298         }
1299
1300         tx = dmu_tx_create(os);
1301
1302         dmu_tx_hold_spill(tx, db->db_object);
1303
1304         err = dmu_tx_assign(tx, TXG_WAIT);
1305         if (err) {
1306                 dmu_buf_rele(db, FTAG);
1307                 dmu_buf_rele(db_spill, FTAG);
1308                 dmu_tx_abort(tx);
1309                 return (err);
1310         }
1311         dmu_buf_will_dirty(db_spill, tx);
1312
1313         if (db_spill->db_size < drrs->drr_length)
1314                 VERIFY(0 == dbuf_spill_set_blksz(db_spill,
1315                     drrs->drr_length, tx));
1316         bcopy(data, db_spill->db_data, drrs->drr_length);
1317
1318         dmu_buf_rele(db, FTAG);
1319         dmu_buf_rele(db_spill, FTAG);
1320
1321         dmu_tx_commit(tx);
1322         return (0);
1323 }
1324
1325 /* ARGSUSED */
1326 noinline static int
1327 restore_free(struct restorearg *ra, objset_t *os,
1328     struct drr_free *drrf)
1329 {
1330         int err;
1331
1332         if (drrf->drr_length != -1ULL &&
1333             drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
1334                 return (EINVAL);
1335
1336         if (dmu_object_info(os, drrf->drr_object, NULL) != 0)
1337                 return (EINVAL);
1338
1339         err = dmu_free_long_range(os, drrf->drr_object,
1340             drrf->drr_offset, drrf->drr_length);
1341         return (err);
1342 }
1343
1344 /*
1345  * NB: callers *must* call dmu_recv_end() if this succeeds.
1346  */
1347 int
1348 dmu_recv_stream(dmu_recv_cookie_t *drc, vnode_t *vp, offset_t *voffp,
1349     int cleanup_fd, uint64_t *action_handlep)
1350 {
1351         struct restorearg ra = { 0 };
1352         dmu_replay_record_t *drr;
1353         objset_t *os;
1354         zio_cksum_t pcksum;
1355         int featureflags;
1356
1357         if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
1358                 ra.byteswap = TRUE;
1359
1360         {
1361                 /* compute checksum of drr_begin record */
1362                 dmu_replay_record_t *drr;
1363                 drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
1364
1365                 drr->drr_type = DRR_BEGIN;
1366                 drr->drr_u.drr_begin = *drc->drc_drrb;
1367                 if (ra.byteswap) {
1368                         fletcher_4_incremental_byteswap(drr,
1369                             sizeof (dmu_replay_record_t), &ra.cksum);
1370                 } else {
1371                         fletcher_4_incremental_native(drr,
1372                             sizeof (dmu_replay_record_t), &ra.cksum);
1373                 }
1374                 kmem_free(drr, sizeof (dmu_replay_record_t));
1375         }
1376
1377         if (ra.byteswap) {
1378                 struct drr_begin *drrb = drc->drc_drrb;
1379                 drrb->drr_magic = BSWAP_64(drrb->drr_magic);
1380                 drrb->drr_versioninfo = BSWAP_64(drrb->drr_versioninfo);
1381                 drrb->drr_creation_time = BSWAP_64(drrb->drr_creation_time);
1382                 drrb->drr_type = BSWAP_32(drrb->drr_type);
1383                 drrb->drr_toguid = BSWAP_64(drrb->drr_toguid);
1384                 drrb->drr_fromguid = BSWAP_64(drrb->drr_fromguid);
1385         }
1386
1387         ra.vp = vp;
1388         ra.voff = *voffp;
1389         ra.bufsize = 1<<20;
1390         ra.buf = vmem_alloc(ra.bufsize, KM_SLEEP);
1391
1392         /* these were verified in dmu_recv_begin */
1393         ASSERT(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo) ==
1394             DMU_SUBSTREAM);
1395         ASSERT(drc->drc_drrb->drr_type < DMU_OST_NUMTYPES);
1396
1397         /*
1398          * Open the objset we are modifying.
1399          */
1400         VERIFY(dmu_objset_from_ds(drc->drc_real_ds, &os) == 0);
1401
1402         ASSERT(drc->drc_real_ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT);
1403
1404         featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
1405
1406         /* if this stream is dedup'ed, set up the avl tree for guid mapping */
1407         if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
1408                 minor_t minor;
1409
1410                 if (cleanup_fd == -1) {
1411                         ra.err = EBADF;
1412                         goto out;
1413                 }
1414                 ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
1415                 if (ra.err) {
1416                         cleanup_fd = -1;
1417                         goto out;
1418                 }
1419
1420                 if (*action_handlep == 0) {
1421                         ra.guid_to_ds_map =
1422                             kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
1423                         avl_create(ra.guid_to_ds_map, guid_compare,
1424                             sizeof (guid_map_entry_t),
1425                             offsetof(guid_map_entry_t, avlnode));
1426                         ra.err = zfs_onexit_add_cb(minor,
1427                             free_guid_map_onexit, ra.guid_to_ds_map,
1428                             action_handlep);
1429                         if (ra.err)
1430                                 goto out;
1431                 } else {
1432                         ra.err = zfs_onexit_cb_data(minor, *action_handlep,
1433                             (void **)&ra.guid_to_ds_map);
1434                         if (ra.err)
1435                                 goto out;
1436                 }
1437
1438                 drc->drc_guid_to_ds_map = ra.guid_to_ds_map;
1439         }
1440
1441         /*
1442          * Read records and process them.
1443          */
1444         pcksum = ra.cksum;
1445         while (ra.err == 0 &&
1446             NULL != (drr = restore_read(&ra, sizeof (*drr)))) {
1447                 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1448                         ra.err = EINTR;
1449                         goto out;
1450                 }
1451
1452                 if (ra.byteswap)
1453                         backup_byteswap(drr);
1454
1455                 switch (drr->drr_type) {
1456                 case DRR_OBJECT:
1457                 {
1458                         /*
1459                          * We need to make a copy of the record header,
1460                          * because restore_{object,write} may need to
1461                          * restore_read(), which will invalidate drr.
1462                          */
1463                         struct drr_object drro = drr->drr_u.drr_object;
1464                         ra.err = restore_object(&ra, os, &drro);
1465                         break;
1466                 }
1467                 case DRR_FREEOBJECTS:
1468                 {
1469                         struct drr_freeobjects drrfo =
1470                             drr->drr_u.drr_freeobjects;
1471                         ra.err = restore_freeobjects(&ra, os, &drrfo);
1472                         break;
1473                 }
1474                 case DRR_WRITE:
1475                 {
1476                         struct drr_write drrw = drr->drr_u.drr_write;
1477                         ra.err = restore_write(&ra, os, &drrw);
1478                         break;
1479                 }
1480                 case DRR_WRITE_BYREF:
1481                 {
1482                         struct drr_write_byref drrwbr =
1483                             drr->drr_u.drr_write_byref;
1484                         ra.err = restore_write_byref(&ra, os, &drrwbr);
1485                         break;
1486                 }
1487                 case DRR_FREE:
1488                 {
1489                         struct drr_free drrf = drr->drr_u.drr_free;
1490                         ra.err = restore_free(&ra, os, &drrf);
1491                         break;
1492                 }
1493                 case DRR_END:
1494                 {
1495                         struct drr_end drre = drr->drr_u.drr_end;
1496                         /*
1497                          * We compare against the *previous* checksum
1498                          * value, because the stored checksum is of
1499                          * everything before the DRR_END record.
1500                          */
1501                         if (!ZIO_CHECKSUM_EQUAL(drre.drr_checksum, pcksum))
1502                                 ra.err = ECKSUM;
1503                         goto out;
1504                 }
1505                 case DRR_SPILL:
1506                 {
1507                         struct drr_spill drrs = drr->drr_u.drr_spill;
1508                         ra.err = restore_spill(&ra, os, &drrs);
1509                         break;
1510                 }
1511                 default:
1512                         ra.err = EINVAL;
1513                         goto out;
1514                 }
1515                 pcksum = ra.cksum;
1516         }
1517         ASSERT(ra.err != 0);
1518
1519 out:
1520         if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
1521                 zfs_onexit_fd_rele(cleanup_fd);
1522
1523         if (ra.err != 0) {
1524                 /*
1525                  * destroy what we created, so we don't leave it in the
1526                  * inconsistent restoring state.
1527                  */
1528                 txg_wait_synced(drc->drc_real_ds->ds_dir->dd_pool, 0);
1529
1530                 (void) dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag,
1531                     B_FALSE);
1532                 if (drc->drc_real_ds != drc->drc_logical_ds) {
1533                         mutex_exit(&drc->drc_logical_ds->ds_recvlock);
1534                         dsl_dataset_rele(drc->drc_logical_ds, dmu_recv_tag);
1535                 }
1536         }
1537
1538         vmem_free(ra.buf, ra.bufsize);
1539         *voffp = ra.voff;
1540         return (ra.err);
1541 }
1542
1543 struct recvendsyncarg {
1544         char *tosnap;
1545         uint64_t creation_time;
1546         uint64_t toguid;
1547 };
1548
1549 static int
1550 recv_end_check(void *arg1, void *arg2, dmu_tx_t *tx)
1551 {
1552         dsl_dataset_t *ds = arg1;
1553         struct recvendsyncarg *resa = arg2;
1554
1555         return (dsl_dataset_snapshot_check(ds, resa->tosnap, tx));
1556 }
1557
1558 static void
1559 recv_end_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1560 {
1561         dsl_dataset_t *ds = arg1;
1562         struct recvendsyncarg *resa = arg2;
1563
1564         dsl_dataset_snapshot_sync(ds, resa->tosnap, tx);
1565
1566         /* set snapshot's creation time and guid */
1567         dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
1568         ds->ds_prev->ds_phys->ds_creation_time = resa->creation_time;
1569         ds->ds_prev->ds_phys->ds_guid = resa->toguid;
1570         ds->ds_prev->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1571
1572         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1573         ds->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1574 }
1575
1576 static int
1577 add_ds_to_guidmap(avl_tree_t *guid_map, dsl_dataset_t *ds)
1578 {
1579         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1580         uint64_t snapobj = ds->ds_phys->ds_prev_snap_obj;
1581         dsl_dataset_t *snapds;
1582         guid_map_entry_t *gmep;
1583         int err;
1584
1585         ASSERT(guid_map != NULL);
1586
1587         rw_enter(&dp->dp_config_rwlock, RW_READER);
1588         err = dsl_dataset_hold_obj(dp, snapobj, guid_map, &snapds);
1589         if (err == 0) {
1590                 gmep = kmem_alloc(sizeof (guid_map_entry_t), KM_SLEEP);
1591                 gmep->guid = snapds->ds_phys->ds_guid;
1592                 gmep->gme_ds = snapds;
1593                 avl_add(guid_map, gmep);
1594         }
1595
1596         rw_exit(&dp->dp_config_rwlock);
1597         return (err);
1598 }
1599
1600 static int
1601 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
1602 {
1603         struct recvendsyncarg resa;
1604         dsl_dataset_t *ds = drc->drc_logical_ds;
1605         int err, myerr;
1606
1607         /*
1608          * XXX hack; seems the ds is still dirty and dsl_pool_zil_clean()
1609          * expects it to have a ds_user_ptr (and zil), but clone_swap()
1610          * can close it.
1611          */
1612         txg_wait_synced(ds->ds_dir->dd_pool, 0);
1613
1614         if (dsl_dataset_tryown(ds, FALSE, dmu_recv_tag)) {
1615                 err = dsl_dataset_clone_swap(drc->drc_real_ds, ds,
1616                     drc->drc_force);
1617                 if (err)
1618                         goto out;
1619         } else {
1620                 mutex_exit(&ds->ds_recvlock);
1621                 dsl_dataset_rele(ds, dmu_recv_tag);
1622                 (void) dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag,
1623                     B_FALSE);
1624                 return (EBUSY);
1625         }
1626
1627         resa.creation_time = drc->drc_drrb->drr_creation_time;
1628         resa.toguid = drc->drc_drrb->drr_toguid;
1629         resa.tosnap = drc->drc_tosnap;
1630
1631         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1632             recv_end_check, recv_end_sync, ds, &resa, 3);
1633         if (err) {
1634                 /* swap back */
1635                 (void) dsl_dataset_clone_swap(drc->drc_real_ds, ds, B_TRUE);
1636         }
1637
1638 out:
1639         mutex_exit(&ds->ds_recvlock);
1640         if (err == 0 && drc->drc_guid_to_ds_map != NULL)
1641                 (void) add_ds_to_guidmap(drc->drc_guid_to_ds_map, ds);
1642         dsl_dataset_disown(ds, dmu_recv_tag);
1643         myerr = dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag, B_FALSE);
1644         ASSERT3U(myerr, ==, 0);
1645         return (err);
1646 }
1647
1648 static int
1649 dmu_recv_new_end(dmu_recv_cookie_t *drc)
1650 {
1651         struct recvendsyncarg resa;
1652         dsl_dataset_t *ds = drc->drc_logical_ds;
1653         int err;
1654
1655         /*
1656          * XXX hack; seems the ds is still dirty and dsl_pool_zil_clean()
1657          * expects it to have a ds_user_ptr (and zil), but clone_swap()
1658          * can close it.
1659          */
1660         txg_wait_synced(ds->ds_dir->dd_pool, 0);
1661
1662         resa.creation_time = drc->drc_drrb->drr_creation_time;
1663         resa.toguid = drc->drc_drrb->drr_toguid;
1664         resa.tosnap = drc->drc_tosnap;
1665
1666         err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1667             recv_end_check, recv_end_sync, ds, &resa, 3);
1668         if (err) {
1669                 /* clean up the fs we just recv'd into */
1670                 (void) dsl_dataset_destroy(ds, dmu_recv_tag, B_FALSE);
1671         } else {
1672                 if (drc->drc_guid_to_ds_map != NULL)
1673                         (void) add_ds_to_guidmap(drc->drc_guid_to_ds_map, ds);
1674                 /* release the hold from dmu_recv_begin */
1675                 dsl_dataset_disown(ds, dmu_recv_tag);
1676         }
1677         return (err);
1678 }
1679
1680 int
1681 dmu_recv_end(dmu_recv_cookie_t *drc)
1682 {
1683         if (drc->drc_logical_ds != drc->drc_real_ds)
1684                 return (dmu_recv_existing_end(drc));
1685         else
1686                 return (dmu_recv_new_end(drc));
1687 }