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