xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu_send.c (revision f9ead4a57883f3ef04ef20d83cc47987d98c0687)
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 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/dmu.h>
27 #include <sys/dmu_impl.h>
28 #include <sys/dmu_tx.h>
29 #include <sys/dbuf.h>
30 #include <sys/dnode.h>
31 #include <sys/zfs_context.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dmu_traverse.h>
34 #include <sys/dsl_dataset.h>
35 #include <sys/dsl_dir.h>
36 #include <sys/dsl_prop.h>
37 #include <sys/dsl_pool.h>
38 #include <sys/dsl_synctask.h>
39 #include <sys/zfs_ioctl.h>
40 #include <sys/zap.h>
41 #include <sys/zio_checksum.h>
42 #include <sys/avl.h>
43 
44 static char *dmu_recv_tag = "dmu_recv_tag";
45 
46 /*
47  * The list of data whose inclusion in a send stream can be pending from
48  * one call to backup_cb to another.  Multiple calls to dump_free() and
49  * dump_freeobjects() can be aggregated into a single DRR_FREE or
50  * DRR_FREEOBJECTS replay record.
51  */
52 typedef enum {
53 	PENDING_NONE,
54 	PENDING_FREE,
55 	PENDING_FREEOBJECTS
56 } pendop_t;
57 
58 struct backuparg {
59 	dmu_replay_record_t *drr;
60 	vnode_t *vp;
61 	offset_t *off;
62 	objset_t *os;
63 	zio_cksum_t zc;
64 	uint64_t toguid;
65 	int err;
66 	pendop_t pending_op;
67 };
68 
69 static int
70 dump_bytes(struct backuparg *ba, void *buf, int len)
71 {
72 	ssize_t resid; /* have to get resid to get detailed errno */
73 	ASSERT3U(len % 8, ==, 0);
74 
75 	fletcher_4_incremental_native(buf, len, &ba->zc);
76 	ba->err = vn_rdwr(UIO_WRITE, ba->vp,
77 	    (caddr_t)buf, len,
78 	    0, UIO_SYSSPACE, FAPPEND, RLIM64_INFINITY, CRED(), &resid);
79 	*ba->off += len;
80 	return (ba->err);
81 }
82 
83 static int
84 dump_free(struct backuparg *ba, uint64_t object, uint64_t offset,
85     uint64_t length)
86 {
87 	struct drr_free *drrf = &(ba->drr->drr_u.drr_free);
88 
89 	/*
90 	 * If there is a pending op, but it's not PENDING_FREE, push it out,
91 	 * since free block aggregation can only be done for blocks of the
92 	 * same type (i.e., DRR_FREE records can only be aggregated with
93 	 * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
94 	 * aggregated with other DRR_FREEOBJECTS records.
95 	 */
96 	if (ba->pending_op != PENDING_NONE && ba->pending_op != PENDING_FREE) {
97 		if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
98 			return (EINTR);
99 		ba->pending_op = PENDING_NONE;
100 	}
101 
102 	if (ba->pending_op == PENDING_FREE) {
103 		/*
104 		 * There should never be a PENDING_FREE if length is -1
105 		 * (because dump_dnode is the only place where this
106 		 * function is called with a -1, and only after flushing
107 		 * any pending record).
108 		 */
109 		ASSERT(length != -1ULL);
110 		/*
111 		 * Check to see whether this free block can be aggregated
112 		 * with pending one.
113 		 */
114 		if (drrf->drr_object == object && drrf->drr_offset +
115 		    drrf->drr_length == offset) {
116 			drrf->drr_length += length;
117 			return (0);
118 		} else {
119 			/* not a continuation.  Push out pending record */
120 			if (dump_bytes(ba, ba->drr,
121 			    sizeof (dmu_replay_record_t)) != 0)
122 				return (EINTR);
123 			ba->pending_op = PENDING_NONE;
124 		}
125 	}
126 	/* create a FREE record and make it pending */
127 	bzero(ba->drr, sizeof (dmu_replay_record_t));
128 	ba->drr->drr_type = DRR_FREE;
129 	drrf->drr_object = object;
130 	drrf->drr_offset = offset;
131 	drrf->drr_length = length;
132 	drrf->drr_toguid = ba->toguid;
133 	if (length == -1ULL) {
134 		if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
135 			return (EINTR);
136 	} else {
137 		ba->pending_op = PENDING_FREE;
138 	}
139 
140 	return (0);
141 }
142 
143 static int
144 dump_data(struct backuparg *ba, dmu_object_type_t type,
145     uint64_t object, uint64_t offset, int blksz, void *data)
146 {
147 	struct drr_write *drrw = &(ba->drr->drr_u.drr_write);
148 
149 	/*
150 	 * If there is any kind of pending aggregation (currently either
151 	 * a grouping of free objects or free blocks), push it out to
152 	 * the stream, since aggregation can't be done across operations
153 	 * of different types.
154 	 */
155 	if (ba->pending_op != PENDING_NONE) {
156 		if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
157 			return (EINTR);
158 		ba->pending_op = PENDING_NONE;
159 	}
160 	/* write a DATA record */
161 	bzero(ba->drr, sizeof (dmu_replay_record_t));
162 	ba->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 = ba->toguid;
168 
169 	if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
170 		return (EINTR);
171 	if (dump_bytes(ba, data, blksz) != 0)
172 		return (EINTR);
173 	return (0);
174 }
175 
176 static int
177 dump_freeobjects(struct backuparg *ba, uint64_t firstobj, uint64_t numobjs)
178 {
179 	struct drr_freeobjects *drrfo = &(ba->drr->drr_u.drr_freeobjects);
180 
181 	/*
182 	 * If there is a pending op, but it's not PENDING_FREEOBJECTS,
183 	 * push it out, since free block aggregation can only be done for
184 	 * blocks of the same type (i.e., DRR_FREE records can only be
185 	 * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
186 	 * can only be aggregated with other DRR_FREEOBJECTS records.
187 	 */
188 	if (ba->pending_op != PENDING_NONE &&
189 	    ba->pending_op != PENDING_FREEOBJECTS) {
190 		if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
191 			return (EINTR);
192 		ba->pending_op = PENDING_NONE;
193 	}
194 	if (ba->pending_op == PENDING_FREEOBJECTS) {
195 		/*
196 		 * See whether this free object array can be aggregated
197 		 * with pending one
198 		 */
199 		if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
200 			drrfo->drr_numobjs += numobjs;
201 			return (0);
202 		} else {
203 			/* can't be aggregated.  Push out pending record */
204 			if (dump_bytes(ba, ba->drr,
205 			    sizeof (dmu_replay_record_t)) != 0)
206 				return (EINTR);
207 			ba->pending_op = PENDING_NONE;
208 		}
209 	}
210 
211 	/* write a FREEOBJECTS record */
212 	bzero(ba->drr, sizeof (dmu_replay_record_t));
213 	ba->drr->drr_type = DRR_FREEOBJECTS;
214 	drrfo->drr_firstobj = firstobj;
215 	drrfo->drr_numobjs = numobjs;
216 	drrfo->drr_toguid = ba->toguid;
217 
218 	ba->pending_op = PENDING_FREEOBJECTS;
219 
220 	return (0);
221 }
222 
223 static int
224 dump_dnode(struct backuparg *ba, uint64_t object, dnode_phys_t *dnp)
225 {
226 	struct drr_object *drro = &(ba->drr->drr_u.drr_object);
227 
228 	if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
229 		return (dump_freeobjects(ba, object, 1));
230 
231 	if (ba->pending_op != PENDING_NONE) {
232 		if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
233 			return (EINTR);
234 		ba->pending_op = PENDING_NONE;
235 	}
236 
237 	/* write an OBJECT record */
238 	bzero(ba->drr, sizeof (dmu_replay_record_t));
239 	ba->drr->drr_type = DRR_OBJECT;
240 	drro->drr_object = object;
241 	drro->drr_type = dnp->dn_type;
242 	drro->drr_bonustype = dnp->dn_bonustype;
243 	drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
244 	drro->drr_bonuslen = dnp->dn_bonuslen;
245 	drro->drr_checksumtype = dnp->dn_checksum;
246 	drro->drr_compress = dnp->dn_compress;
247 	drro->drr_toguid = ba->toguid;
248 
249 	if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
250 		return (EINTR);
251 
252 	if (dump_bytes(ba, DN_BONUS(dnp), P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0)
253 		return (EINTR);
254 
255 	/* free anything past the end of the file */
256 	if (dump_free(ba, object, (dnp->dn_maxblkid + 1) *
257 	    (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL))
258 		return (EINTR);
259 	if (ba->err)
260 		return (EINTR);
261 	return (0);
262 }
263 
264 #define	BP_SPAN(dnp, level) \
265 	(((uint64_t)dnp->dn_datablkszsec) << (SPA_MINBLOCKSHIFT + \
266 	(level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)))
267 
268 /* ARGSUSED */
269 static int
270 backup_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
271     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
272 {
273 	struct backuparg *ba = arg;
274 	dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
275 	int err = 0;
276 
277 	if (issig(JUSTLOOKING) && issig(FORREAL))
278 		return (EINTR);
279 
280 	if (zb->zb_object != DMU_META_DNODE_OBJECT &&
281 	    DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
282 		return (0);
283 	} else if (bp == NULL && zb->zb_object == DMU_META_DNODE_OBJECT) {
284 		uint64_t span = BP_SPAN(dnp, zb->zb_level);
285 		uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
286 		err = dump_freeobjects(ba, dnobj, span >> DNODE_SHIFT);
287 	} else if (bp == NULL) {
288 		uint64_t span = BP_SPAN(dnp, zb->zb_level);
289 		err = dump_free(ba, zb->zb_object, zb->zb_blkid * span, span);
290 	} else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
291 		return (0);
292 	} else if (type == DMU_OT_DNODE) {
293 		dnode_phys_t *blk;
294 		int i;
295 		int blksz = BP_GET_LSIZE(bp);
296 		uint32_t aflags = ARC_WAIT;
297 		arc_buf_t *abuf;
298 
299 		if (arc_read_nolock(NULL, spa, bp,
300 		    arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
301 		    ZIO_FLAG_CANFAIL, &aflags, zb) != 0)
302 			return (EIO);
303 
304 		blk = abuf->b_data;
305 		for (i = 0; i < blksz >> DNODE_SHIFT; i++) {
306 			uint64_t dnobj = (zb->zb_blkid <<
307 			    (DNODE_BLOCK_SHIFT - DNODE_SHIFT)) + i;
308 			err = dump_dnode(ba, dnobj, blk+i);
309 			if (err)
310 				break;
311 		}
312 		(void) arc_buf_remove_ref(abuf, &abuf);
313 	} else { /* it's a level-0 block of a regular object */
314 		uint32_t aflags = ARC_WAIT;
315 		arc_buf_t *abuf;
316 		int blksz = BP_GET_LSIZE(bp);
317 
318 		if (arc_read_nolock(NULL, spa, bp,
319 		    arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
320 		    ZIO_FLAG_CANFAIL, &aflags, zb) != 0)
321 			return (EIO);
322 
323 		err = dump_data(ba, type, zb->zb_object, zb->zb_blkid * blksz,
324 		    blksz, abuf->b_data);
325 		(void) arc_buf_remove_ref(abuf, &abuf);
326 	}
327 
328 	ASSERT(err == 0 || err == EINTR);
329 	return (err);
330 }
331 
332 int
333 dmu_sendbackup(objset_t *tosnap, objset_t *fromsnap, boolean_t fromorigin,
334     vnode_t *vp, offset_t *off)
335 {
336 	dsl_dataset_t *ds = tosnap->os_dsl_dataset;
337 	dsl_dataset_t *fromds = fromsnap ? fromsnap->os_dsl_dataset : NULL;
338 	dmu_replay_record_t *drr;
339 	struct backuparg ba;
340 	int err;
341 	uint64_t fromtxg = 0;
342 
343 	/* tosnap must be a snapshot */
344 	if (ds->ds_phys->ds_next_snap_obj == 0)
345 		return (EINVAL);
346 
347 	/* fromsnap must be an earlier snapshot from the same fs as tosnap */
348 	if (fromds && (ds->ds_dir != fromds->ds_dir ||
349 	    fromds->ds_phys->ds_creation_txg >= ds->ds_phys->ds_creation_txg))
350 		return (EXDEV);
351 
352 	if (fromorigin) {
353 		dsl_pool_t *dp = ds->ds_dir->dd_pool;
354 
355 		if (fromsnap)
356 			return (EINVAL);
357 
358 		if (dsl_dir_is_clone(ds->ds_dir)) {
359 			rw_enter(&dp->dp_config_rwlock, RW_READER);
360 			err = dsl_dataset_hold_obj(dp,
361 			    ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &fromds);
362 			rw_exit(&dp->dp_config_rwlock);
363 			if (err)
364 				return (err);
365 		} else {
366 			fromorigin = B_FALSE;
367 		}
368 	}
369 
370 
371 	drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
372 	drr->drr_type = DRR_BEGIN;
373 	drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
374 	DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
375 	    DMU_SUBSTREAM);
376 	drr->drr_u.drr_begin.drr_creation_time =
377 	    ds->ds_phys->ds_creation_time;
378 	drr->drr_u.drr_begin.drr_type = tosnap->os_phys->os_type;
379 	if (fromorigin)
380 		drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
381 	drr->drr_u.drr_begin.drr_toguid = ds->ds_phys->ds_guid;
382 	if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
383 		drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
384 
385 	if (fromds)
386 		drr->drr_u.drr_begin.drr_fromguid = fromds->ds_phys->ds_guid;
387 	dsl_dataset_name(ds, drr->drr_u.drr_begin.drr_toname);
388 
389 	if (fromds)
390 		fromtxg = fromds->ds_phys->ds_creation_txg;
391 	if (fromorigin)
392 		dsl_dataset_rele(fromds, FTAG);
393 
394 	ba.drr = drr;
395 	ba.vp = vp;
396 	ba.os = tosnap;
397 	ba.off = off;
398 	ba.toguid = ds->ds_phys->ds_guid;
399 	ZIO_SET_CHECKSUM(&ba.zc, 0, 0, 0, 0);
400 	ba.pending_op = PENDING_NONE;
401 
402 	if (dump_bytes(&ba, drr, sizeof (dmu_replay_record_t)) != 0) {
403 		kmem_free(drr, sizeof (dmu_replay_record_t));
404 		return (ba.err);
405 	}
406 
407 	err = traverse_dataset(ds, fromtxg, TRAVERSE_PRE | TRAVERSE_PREFETCH,
408 	    backup_cb, &ba);
409 
410 	if (ba.pending_op != PENDING_NONE)
411 		if (dump_bytes(&ba, drr, sizeof (dmu_replay_record_t)) != 0)
412 			err = EINTR;
413 
414 	if (err) {
415 		if (err == EINTR && ba.err)
416 			err = ba.err;
417 		kmem_free(drr, sizeof (dmu_replay_record_t));
418 		return (err);
419 	}
420 
421 	bzero(drr, sizeof (dmu_replay_record_t));
422 	drr->drr_type = DRR_END;
423 	drr->drr_u.drr_end.drr_checksum = ba.zc;
424 	drr->drr_u.drr_end.drr_toguid = ba.toguid;
425 
426 	if (dump_bytes(&ba, drr, sizeof (dmu_replay_record_t)) != 0) {
427 		kmem_free(drr, sizeof (dmu_replay_record_t));
428 		return (ba.err);
429 	}
430 
431 	kmem_free(drr, sizeof (dmu_replay_record_t));
432 
433 	return (0);
434 }
435 
436 struct recvbeginsyncarg {
437 	const char *tofs;
438 	const char *tosnap;
439 	dsl_dataset_t *origin;
440 	uint64_t fromguid;
441 	dmu_objset_type_t type;
442 	void *tag;
443 	boolean_t force;
444 	uint64_t dsflags;
445 	char clonelastname[MAXNAMELEN];
446 	dsl_dataset_t *ds; /* the ds to recv into; returned from the syncfunc */
447 };
448 
449 /* ARGSUSED */
450 static int
451 recv_new_check(void *arg1, void *arg2, dmu_tx_t *tx)
452 {
453 	dsl_dir_t *dd = arg1;
454 	struct recvbeginsyncarg *rbsa = arg2;
455 	objset_t *mos = dd->dd_pool->dp_meta_objset;
456 	uint64_t val;
457 	int err;
458 
459 	err = zap_lookup(mos, dd->dd_phys->dd_child_dir_zapobj,
460 	    strrchr(rbsa->tofs, '/') + 1, sizeof (uint64_t), 1, &val);
461 
462 	if (err != ENOENT)
463 		return (err ? err : EEXIST);
464 
465 	if (rbsa->origin) {
466 		/* make sure it's a snap in the same pool */
467 		if (rbsa->origin->ds_dir->dd_pool != dd->dd_pool)
468 			return (EXDEV);
469 		if (!dsl_dataset_is_snapshot(rbsa->origin))
470 			return (EINVAL);
471 		if (rbsa->origin->ds_phys->ds_guid != rbsa->fromguid)
472 			return (ENODEV);
473 	}
474 
475 	return (0);
476 }
477 
478 static void
479 recv_new_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
480 {
481 	dsl_dir_t *dd = arg1;
482 	struct recvbeginsyncarg *rbsa = arg2;
483 	uint64_t flags = DS_FLAG_INCONSISTENT | rbsa->dsflags;
484 	uint64_t dsobj;
485 
486 	/* Create and open new dataset. */
487 	dsobj = dsl_dataset_create_sync(dd, strrchr(rbsa->tofs, '/') + 1,
488 	    rbsa->origin, flags, cr, tx);
489 	VERIFY(0 == dsl_dataset_own_obj(dd->dd_pool, dsobj,
490 	    B_TRUE, dmu_recv_tag, &rbsa->ds));
491 
492 	if (rbsa->origin == NULL) {
493 		(void) dmu_objset_create_impl(dd->dd_pool->dp_spa,
494 		    rbsa->ds, &rbsa->ds->ds_phys->ds_bp, rbsa->type, tx);
495 	}
496 
497 	spa_history_internal_log(LOG_DS_REPLAY_FULL_SYNC,
498 	    dd->dd_pool->dp_spa, tx, cr, "dataset = %lld", dsobj);
499 }
500 
501 /* ARGSUSED */
502 static int
503 recv_existing_check(void *arg1, void *arg2, dmu_tx_t *tx)
504 {
505 	dsl_dataset_t *ds = arg1;
506 	struct recvbeginsyncarg *rbsa = arg2;
507 	int err;
508 	uint64_t val;
509 
510 	/* must not have any changes since most recent snapshot */
511 	if (!rbsa->force && dsl_dataset_modified_since_lastsnap(ds))
512 		return (ETXTBSY);
513 
514 	if (rbsa->fromguid) {
515 		/* if incremental, most recent snapshot must match fromguid */
516 		if (ds->ds_prev == NULL)
517 			return (ENODEV);
518 
519 		/*
520 		 * most recent snapshot must match fromguid, or there are no
521 		 * changes since the fromguid one
522 		 */
523 		if (ds->ds_prev->ds_phys->ds_guid != rbsa->fromguid) {
524 			uint64_t birth = ds->ds_prev->ds_phys->ds_bp.blk_birth;
525 			uint64_t obj = ds->ds_prev->ds_phys->ds_prev_snap_obj;
526 			while (obj != 0) {
527 				dsl_dataset_t *snap;
528 				err = dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
529 				    obj, FTAG, &snap);
530 				if (err)
531 					return (ENODEV);
532 				if (snap->ds_phys->ds_creation_txg < birth) {
533 					dsl_dataset_rele(snap, FTAG);
534 					return (ENODEV);
535 				}
536 				if (snap->ds_phys->ds_guid == rbsa->fromguid) {
537 					dsl_dataset_rele(snap, FTAG);
538 					break; /* it's ok */
539 				}
540 				obj = snap->ds_phys->ds_prev_snap_obj;
541 				dsl_dataset_rele(snap, FTAG);
542 			}
543 			if (obj == 0)
544 				return (ENODEV);
545 		}
546 	} else {
547 		/* if full, most recent snapshot must be $ORIGIN */
548 		if (ds->ds_phys->ds_prev_snap_txg >= TXG_INITIAL)
549 			return (ENODEV);
550 	}
551 
552 	/* temporary clone name must not exist */
553 	err = zap_lookup(ds->ds_dir->dd_pool->dp_meta_objset,
554 	    ds->ds_dir->dd_phys->dd_child_dir_zapobj,
555 	    rbsa->clonelastname, 8, 1, &val);
556 	if (err == 0)
557 		return (EEXIST);
558 	if (err != ENOENT)
559 		return (err);
560 
561 	/* new snapshot name must not exist */
562 	err = zap_lookup(ds->ds_dir->dd_pool->dp_meta_objset,
563 	    ds->ds_phys->ds_snapnames_zapobj, rbsa->tosnap, 8, 1, &val);
564 	if (err == 0)
565 		return (EEXIST);
566 	if (err != ENOENT)
567 		return (err);
568 	return (0);
569 }
570 
571 /* ARGSUSED */
572 static void
573 recv_existing_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
574 {
575 	dsl_dataset_t *ohds = arg1;
576 	struct recvbeginsyncarg *rbsa = arg2;
577 	dsl_pool_t *dp = ohds->ds_dir->dd_pool;
578 	dsl_dataset_t *cds;
579 	uint64_t flags = DS_FLAG_INCONSISTENT | rbsa->dsflags;
580 	uint64_t dsobj;
581 
582 	/* create and open the temporary clone */
583 	dsobj = dsl_dataset_create_sync(ohds->ds_dir, rbsa->clonelastname,
584 	    ohds->ds_prev, flags, cr, tx);
585 	VERIFY(0 == dsl_dataset_own_obj(dp, dsobj, B_TRUE, dmu_recv_tag, &cds));
586 
587 	/*
588 	 * If we actually created a non-clone, we need to create the
589 	 * objset in our new dataset.
590 	 */
591 	if (BP_IS_HOLE(dsl_dataset_get_blkptr(cds))) {
592 		(void) dmu_objset_create_impl(dp->dp_spa,
593 		    cds, dsl_dataset_get_blkptr(cds), rbsa->type, tx);
594 	}
595 
596 	rbsa->ds = cds;
597 
598 	spa_history_internal_log(LOG_DS_REPLAY_INC_SYNC,
599 	    dp->dp_spa, tx, cr, "dataset = %lld", dsobj);
600 }
601 
602 /*
603  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
604  * succeeds; otherwise we will leak the holds on the datasets.
605  */
606 int
607 dmu_recv_begin(char *tofs, char *tosnap, char *top_ds, struct drr_begin *drrb,
608     boolean_t force, objset_t *origin, dmu_recv_cookie_t *drc)
609 {
610 	int err = 0;
611 	boolean_t byteswap;
612 	struct recvbeginsyncarg rbsa = { 0 };
613 	uint64_t versioninfo;
614 	int flags;
615 	dsl_dataset_t *ds;
616 
617 	if (drrb->drr_magic == DMU_BACKUP_MAGIC)
618 		byteswap = FALSE;
619 	else if (drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
620 		byteswap = TRUE;
621 	else
622 		return (EINVAL);
623 
624 	rbsa.tofs = tofs;
625 	rbsa.tosnap = tosnap;
626 	rbsa.origin = origin ? origin->os_dsl_dataset : NULL;
627 	rbsa.fromguid = drrb->drr_fromguid;
628 	rbsa.type = drrb->drr_type;
629 	rbsa.tag = FTAG;
630 	rbsa.dsflags = 0;
631 	versioninfo = drrb->drr_versioninfo;
632 	flags = drrb->drr_flags;
633 
634 	if (byteswap) {
635 		rbsa.type = BSWAP_32(rbsa.type);
636 		rbsa.fromguid = BSWAP_64(rbsa.fromguid);
637 		versioninfo = BSWAP_64(versioninfo);
638 		flags = BSWAP_32(flags);
639 	}
640 
641 	if (DMU_GET_STREAM_HDRTYPE(versioninfo) == DMU_COMPOUNDSTREAM ||
642 	    rbsa.type >= DMU_OST_NUMTYPES ||
643 	    ((flags & DRR_FLAG_CLONE) && origin == NULL))
644 		return (EINVAL);
645 
646 	if (flags & DRR_FLAG_CI_DATA)
647 		rbsa.dsflags = DS_FLAG_CI_DATASET;
648 
649 	bzero(drc, sizeof (dmu_recv_cookie_t));
650 	drc->drc_drrb = drrb;
651 	drc->drc_tosnap = tosnap;
652 	drc->drc_top_ds = top_ds;
653 	drc->drc_force = force;
654 
655 	/*
656 	 * Process the begin in syncing context.
657 	 */
658 
659 	/* open the dataset we are logically receiving into */
660 	err = dsl_dataset_hold(tofs, dmu_recv_tag, &ds);
661 	if (err == 0) {
662 		/* target fs already exists; recv into temp clone */
663 
664 		/* Can't recv a clone into an existing fs */
665 		if (flags & DRR_FLAG_CLONE) {
666 			dsl_dataset_rele(ds, dmu_recv_tag);
667 			return (EINVAL);
668 		}
669 
670 		/* must not have an incremental recv already in progress */
671 		if (!mutex_tryenter(&ds->ds_recvlock)) {
672 			dsl_dataset_rele(ds, dmu_recv_tag);
673 			return (EBUSY);
674 		}
675 
676 		/* tmp clone name is: tofs/%tosnap" */
677 		(void) snprintf(rbsa.clonelastname, sizeof (rbsa.clonelastname),
678 		    "%%%s", tosnap);
679 		rbsa.force = force;
680 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
681 		    recv_existing_check, recv_existing_sync, ds, &rbsa, 5);
682 		if (err) {
683 			mutex_exit(&ds->ds_recvlock);
684 			dsl_dataset_rele(ds, dmu_recv_tag);
685 			return (err);
686 		}
687 		drc->drc_logical_ds = ds;
688 		drc->drc_real_ds = rbsa.ds;
689 	} else if (err == ENOENT) {
690 		/* target fs does not exist; must be a full backup or clone */
691 		char *cp;
692 
693 		/*
694 		 * If it's a non-clone incremental, we are missing the
695 		 * target fs, so fail the recv.
696 		 */
697 		if (rbsa.fromguid && !(flags & DRR_FLAG_CLONE))
698 			return (ENOENT);
699 
700 		/* Open the parent of tofs */
701 		cp = strrchr(tofs, '/');
702 		*cp = '\0';
703 		err = dsl_dataset_hold(tofs, FTAG, &ds);
704 		*cp = '/';
705 		if (err)
706 			return (err);
707 
708 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
709 		    recv_new_check, recv_new_sync, ds->ds_dir, &rbsa, 5);
710 		dsl_dataset_rele(ds, FTAG);
711 		if (err)
712 			return (err);
713 		drc->drc_logical_ds = drc->drc_real_ds = rbsa.ds;
714 		drc->drc_newfs = B_TRUE;
715 	}
716 
717 	return (err);
718 }
719 
720 struct restorearg {
721 	int err;
722 	int byteswap;
723 	vnode_t *vp;
724 	char *buf;
725 	uint64_t voff;
726 	int bufsize; /* amount of memory allocated for buf */
727 	zio_cksum_t cksum;
728 	avl_tree_t guid_to_ds_map;
729 };
730 
731 typedef struct guid_map_entry {
732 	uint64_t	guid;
733 	dsl_dataset_t	*gme_ds;
734 	avl_node_t	avlnode;
735 } guid_map_entry_t;
736 
737 static int
738 guid_compare(const void *arg1, const void *arg2)
739 {
740 	const guid_map_entry_t *gmep1 = arg1;
741 	const guid_map_entry_t *gmep2 = arg2;
742 
743 	if (gmep1->guid < gmep2->guid)
744 		return (-1);
745 	else if (gmep1->guid > gmep2->guid)
746 		return (1);
747 	return (0);
748 }
749 
750 /*
751  * This function is a callback used by dmu_objset_find() (which
752  * enumerates the object sets) to build an avl tree that maps guids
753  * to datasets.  The resulting table is used when processing DRR_WRITE_BYREF
754  * send stream records.  These records, which are used in dedup'ed
755  * streams, do not contain data themselves, but refer to a copy
756  * of the data block that has already been written because it was
757  * earlier in the stream.  That previous copy is identified by the
758  * guid of the dataset with the referenced data.
759  */
760 int
761 find_ds_by_guid(const char *name, void *arg)
762 {
763 	avl_tree_t *guid_map = arg;
764 	dsl_dataset_t *ds, *snapds;
765 	guid_map_entry_t *gmep;
766 	dsl_pool_t *dp;
767 	int err;
768 	uint64_t lastobj, firstobj;
769 
770 	if (dsl_dataset_hold(name, FTAG, &ds) != 0)
771 		return (0);
772 
773 	dp = ds->ds_dir->dd_pool;
774 	rw_enter(&dp->dp_config_rwlock, RW_READER);
775 	firstobj = ds->ds_dir->dd_phys->dd_origin_obj;
776 	lastobj = ds->ds_phys->ds_prev_snap_obj;
777 
778 	while (lastobj != firstobj) {
779 		err = dsl_dataset_hold_obj(dp, lastobj, guid_map, &snapds);
780 		if (err) {
781 			/*
782 			 * Skip this snapshot and move on. It's not
783 			 * clear why this would ever happen, but the
784 			 * remainder of the snapshot streadm can be
785 			 * processed.
786 			 */
787 			rw_exit(&dp->dp_config_rwlock);
788 			dsl_dataset_rele(ds, FTAG);
789 			return (0);
790 		}
791 
792 		gmep = kmem_alloc(sizeof (guid_map_entry_t), KM_SLEEP);
793 		gmep->guid = snapds->ds_phys->ds_guid;
794 		gmep->gme_ds = snapds;
795 		avl_add(guid_map, gmep);
796 		lastobj = snapds->ds_phys->ds_prev_snap_obj;
797 	}
798 
799 	rw_exit(&dp->dp_config_rwlock);
800 	dsl_dataset_rele(ds, FTAG);
801 
802 	return (0);
803 }
804 
805 static void *
806 restore_read(struct restorearg *ra, int len)
807 {
808 	void *rv;
809 	int done = 0;
810 
811 	/* some things will require 8-byte alignment, so everything must */
812 	ASSERT3U(len % 8, ==, 0);
813 
814 	while (done < len) {
815 		ssize_t resid;
816 
817 		ra->err = vn_rdwr(UIO_READ, ra->vp,
818 		    (caddr_t)ra->buf + done, len - done,
819 		    ra->voff, UIO_SYSSPACE, FAPPEND,
820 		    RLIM64_INFINITY, CRED(), &resid);
821 
822 		if (resid == len - done)
823 			ra->err = EINVAL;
824 		ra->voff += len - done - resid;
825 		done = len - resid;
826 		if (ra->err)
827 			return (NULL);
828 	}
829 
830 	ASSERT3U(done, ==, len);
831 	rv = ra->buf;
832 	if (ra->byteswap)
833 		fletcher_4_incremental_byteswap(rv, len, &ra->cksum);
834 	else
835 		fletcher_4_incremental_native(rv, len, &ra->cksum);
836 	return (rv);
837 }
838 
839 static void
840 backup_byteswap(dmu_replay_record_t *drr)
841 {
842 #define	DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
843 #define	DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
844 	drr->drr_type = BSWAP_32(drr->drr_type);
845 	drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
846 	switch (drr->drr_type) {
847 	case DRR_BEGIN:
848 		DO64(drr_begin.drr_magic);
849 		DO64(drr_begin.drr_versioninfo);
850 		DO64(drr_begin.drr_creation_time);
851 		DO32(drr_begin.drr_type);
852 		DO32(drr_begin.drr_flags);
853 		DO64(drr_begin.drr_toguid);
854 		DO64(drr_begin.drr_fromguid);
855 		break;
856 	case DRR_OBJECT:
857 		DO64(drr_object.drr_object);
858 		/* DO64(drr_object.drr_allocation_txg); */
859 		DO32(drr_object.drr_type);
860 		DO32(drr_object.drr_bonustype);
861 		DO32(drr_object.drr_blksz);
862 		DO32(drr_object.drr_bonuslen);
863 		DO64(drr_object.drr_toguid);
864 		break;
865 	case DRR_FREEOBJECTS:
866 		DO64(drr_freeobjects.drr_firstobj);
867 		DO64(drr_freeobjects.drr_numobjs);
868 		DO64(drr_freeobjects.drr_toguid);
869 		break;
870 	case DRR_WRITE:
871 		DO64(drr_write.drr_object);
872 		DO32(drr_write.drr_type);
873 		DO64(drr_write.drr_offset);
874 		DO64(drr_write.drr_length);
875 		DO64(drr_write.drr_toguid);
876 		DO64(drr_write.drr_blkcksum.zc_word[0]);
877 		DO64(drr_write.drr_blkcksum.zc_word[1]);
878 		DO64(drr_write.drr_blkcksum.zc_word[2]);
879 		DO64(drr_write.drr_blkcksum.zc_word[3]);
880 		break;
881 	case DRR_WRITE_BYREF:
882 		DO64(drr_write_byref.drr_object);
883 		DO64(drr_write_byref.drr_offset);
884 		DO64(drr_write_byref.drr_length);
885 		DO64(drr_write_byref.drr_toguid);
886 		DO64(drr_write_byref.drr_refguid);
887 		DO64(drr_write_byref.drr_refobject);
888 		DO64(drr_write_byref.drr_refoffset);
889 		DO64(drr_write_byref.drr_blkcksum.zc_word[0]);
890 		DO64(drr_write_byref.drr_blkcksum.zc_word[1]);
891 		DO64(drr_write_byref.drr_blkcksum.zc_word[2]);
892 		DO64(drr_write_byref.drr_blkcksum.zc_word[3]);
893 		break;
894 	case DRR_FREE:
895 		DO64(drr_free.drr_object);
896 		DO64(drr_free.drr_offset);
897 		DO64(drr_free.drr_length);
898 		DO64(drr_free.drr_toguid);
899 		break;
900 	case DRR_END:
901 		DO64(drr_end.drr_checksum.zc_word[0]);
902 		DO64(drr_end.drr_checksum.zc_word[1]);
903 		DO64(drr_end.drr_checksum.zc_word[2]);
904 		DO64(drr_end.drr_checksum.zc_word[3]);
905 		DO64(drr_end.drr_toguid);
906 		break;
907 	}
908 #undef DO64
909 #undef DO32
910 }
911 
912 static int
913 restore_object(struct restorearg *ra, objset_t *os, struct drr_object *drro)
914 {
915 	int err;
916 	dmu_tx_t *tx;
917 	void *data = NULL;
918 
919 	if (drro->drr_type == DMU_OT_NONE ||
920 	    drro->drr_type >= DMU_OT_NUMTYPES ||
921 	    drro->drr_bonustype >= DMU_OT_NUMTYPES ||
922 	    drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
923 	    drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
924 	    P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
925 	    drro->drr_blksz < SPA_MINBLOCKSIZE ||
926 	    drro->drr_blksz > SPA_MAXBLOCKSIZE ||
927 	    drro->drr_bonuslen > DN_MAX_BONUSLEN) {
928 		return (EINVAL);
929 	}
930 
931 	err = dmu_object_info(os, drro->drr_object, NULL);
932 
933 	if (err != 0 && err != ENOENT)
934 		return (EINVAL);
935 
936 	if (drro->drr_bonuslen) {
937 		data = restore_read(ra, P2ROUNDUP(drro->drr_bonuslen, 8));
938 		if (ra->err)
939 			return (ra->err);
940 	}
941 
942 	if (err == ENOENT) {
943 		/* currently free, want to be allocated */
944 		tx = dmu_tx_create(os);
945 		dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
946 		err = dmu_tx_assign(tx, TXG_WAIT);
947 		if (err) {
948 			dmu_tx_abort(tx);
949 			return (err);
950 		}
951 		err = dmu_object_claim(os, drro->drr_object,
952 		    drro->drr_type, drro->drr_blksz,
953 		    drro->drr_bonustype, drro->drr_bonuslen, tx);
954 		dmu_tx_commit(tx);
955 	} else {
956 		/* currently allocated, want to be allocated */
957 		err = dmu_object_reclaim(os, drro->drr_object,
958 		    drro->drr_type, drro->drr_blksz,
959 		    drro->drr_bonustype, drro->drr_bonuslen);
960 	}
961 	if (err)
962 		return (EINVAL);
963 
964 	tx = dmu_tx_create(os);
965 	dmu_tx_hold_bonus(tx, drro->drr_object);
966 	err = dmu_tx_assign(tx, TXG_WAIT);
967 	if (err) {
968 		dmu_tx_abort(tx);
969 		return (err);
970 	}
971 
972 	dmu_object_set_checksum(os, drro->drr_object, drro->drr_checksumtype,
973 	    tx);
974 	dmu_object_set_compress(os, drro->drr_object, drro->drr_compress, tx);
975 
976 	if (data != NULL) {
977 		dmu_buf_t *db;
978 
979 		VERIFY(0 == dmu_bonus_hold(os, drro->drr_object, FTAG, &db));
980 		dmu_buf_will_dirty(db, tx);
981 
982 		ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
983 		bcopy(data, db->db_data, drro->drr_bonuslen);
984 		if (ra->byteswap) {
985 			dmu_ot[drro->drr_bonustype].ot_byteswap(db->db_data,
986 			    drro->drr_bonuslen);
987 		}
988 		dmu_buf_rele(db, FTAG);
989 	}
990 	dmu_tx_commit(tx);
991 	return (0);
992 }
993 
994 /* ARGSUSED */
995 static int
996 restore_freeobjects(struct restorearg *ra, objset_t *os,
997     struct drr_freeobjects *drrfo)
998 {
999 	uint64_t obj;
1000 
1001 	if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
1002 		return (EINVAL);
1003 
1004 	for (obj = drrfo->drr_firstobj;
1005 	    obj < drrfo->drr_firstobj + drrfo->drr_numobjs;
1006 	    (void) dmu_object_next(os, &obj, FALSE, 0)) {
1007 		int err;
1008 
1009 		if (dmu_object_info(os, obj, NULL) != 0)
1010 			continue;
1011 
1012 		err = dmu_free_object(os, obj);
1013 		if (err)
1014 			return (err);
1015 	}
1016 	return (0);
1017 }
1018 
1019 static int
1020 restore_write(struct restorearg *ra, objset_t *os,
1021     struct drr_write *drrw)
1022 {
1023 	dmu_tx_t *tx;
1024 	void *data;
1025 	int err;
1026 
1027 	if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
1028 	    drrw->drr_type >= DMU_OT_NUMTYPES)
1029 		return (EINVAL);
1030 
1031 	data = restore_read(ra, drrw->drr_length);
1032 	if (data == NULL)
1033 		return (ra->err);
1034 
1035 	if (dmu_object_info(os, drrw->drr_object, NULL) != 0)
1036 		return (EINVAL);
1037 
1038 	tx = dmu_tx_create(os);
1039 
1040 	dmu_tx_hold_write(tx, drrw->drr_object,
1041 	    drrw->drr_offset, drrw->drr_length);
1042 	err = dmu_tx_assign(tx, TXG_WAIT);
1043 	if (err) {
1044 		dmu_tx_abort(tx);
1045 		return (err);
1046 	}
1047 	if (ra->byteswap)
1048 		dmu_ot[drrw->drr_type].ot_byteswap(data, drrw->drr_length);
1049 	dmu_write(os, drrw->drr_object,
1050 	    drrw->drr_offset, drrw->drr_length, data, tx);
1051 	dmu_tx_commit(tx);
1052 	return (0);
1053 }
1054 
1055 /*
1056  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
1057  * streams to refer to a copy of the data that is already on the
1058  * system because it came in earlier in the stream.  This function
1059  * finds the earlier copy of the data, and uses that copy instead of
1060  * data from the stream to fulfill this write.
1061  */
1062 static int
1063 restore_write_byref(struct restorearg *ra, objset_t *os,
1064     struct drr_write_byref *drrwbr)
1065 {
1066 	dmu_tx_t *tx;
1067 	int err;
1068 	guid_map_entry_t gmesrch;
1069 	guid_map_entry_t *gmep;
1070 	avl_index_t	where;
1071 	objset_t *ref_os = NULL;
1072 	dmu_buf_t *dbp;
1073 
1074 	if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
1075 		return (EINVAL);
1076 
1077 	/*
1078 	 * If the GUID of the referenced dataset is different from the
1079 	 * GUID of the target dataset, find the referenced dataset.
1080 	 */
1081 	if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
1082 		gmesrch.guid = drrwbr->drr_refguid;
1083 		if ((gmep = avl_find(&ra->guid_to_ds_map, &gmesrch,
1084 		    &where)) == NULL) {
1085 			return (EINVAL);
1086 		}
1087 		if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
1088 			return (EINVAL);
1089 	} else {
1090 		ref_os = os;
1091 	}
1092 
1093 	if (err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
1094 	    drrwbr->drr_refoffset, FTAG, &dbp))
1095 		return (err);
1096 
1097 	tx = dmu_tx_create(os);
1098 
1099 	dmu_tx_hold_write(tx, drrwbr->drr_object,
1100 	    drrwbr->drr_offset, drrwbr->drr_length);
1101 	err = dmu_tx_assign(tx, TXG_WAIT);
1102 	if (err) {
1103 		dmu_tx_abort(tx);
1104 		return (err);
1105 	}
1106 	dmu_write(os, drrwbr->drr_object,
1107 	    drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
1108 	dmu_buf_rele(dbp, FTAG);
1109 	dmu_tx_commit(tx);
1110 	return (0);
1111 }
1112 
1113 /* ARGSUSED */
1114 static int
1115 restore_free(struct restorearg *ra, objset_t *os,
1116     struct drr_free *drrf)
1117 {
1118 	int err;
1119 
1120 	if (drrf->drr_length != -1ULL &&
1121 	    drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
1122 		return (EINVAL);
1123 
1124 	if (dmu_object_info(os, drrf->drr_object, NULL) != 0)
1125 		return (EINVAL);
1126 
1127 	err = dmu_free_long_range(os, drrf->drr_object,
1128 	    drrf->drr_offset, drrf->drr_length);
1129 	return (err);
1130 }
1131 
1132 /*
1133  * NB: callers *must* call dmu_recv_end() if this succeeds.
1134  */
1135 int
1136 dmu_recv_stream(dmu_recv_cookie_t *drc, vnode_t *vp, offset_t *voffp)
1137 {
1138 	struct restorearg ra = { 0 };
1139 	dmu_replay_record_t *drr;
1140 	objset_t *os;
1141 	zio_cksum_t pcksum;
1142 	guid_map_entry_t *gmep;
1143 	int featureflags;
1144 
1145 	if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
1146 		ra.byteswap = TRUE;
1147 
1148 	{
1149 		/* compute checksum of drr_begin record */
1150 		dmu_replay_record_t *drr;
1151 		drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
1152 
1153 		drr->drr_type = DRR_BEGIN;
1154 		drr->drr_u.drr_begin = *drc->drc_drrb;
1155 		if (ra.byteswap) {
1156 			fletcher_4_incremental_byteswap(drr,
1157 			    sizeof (dmu_replay_record_t), &ra.cksum);
1158 		} else {
1159 			fletcher_4_incremental_native(drr,
1160 			    sizeof (dmu_replay_record_t), &ra.cksum);
1161 		}
1162 		kmem_free(drr, sizeof (dmu_replay_record_t));
1163 	}
1164 
1165 	if (ra.byteswap) {
1166 		struct drr_begin *drrb = drc->drc_drrb;
1167 		drrb->drr_magic = BSWAP_64(drrb->drr_magic);
1168 		drrb->drr_versioninfo = BSWAP_64(drrb->drr_versioninfo);
1169 		drrb->drr_creation_time = BSWAP_64(drrb->drr_creation_time);
1170 		drrb->drr_type = BSWAP_32(drrb->drr_type);
1171 		drrb->drr_toguid = BSWAP_64(drrb->drr_toguid);
1172 		drrb->drr_fromguid = BSWAP_64(drrb->drr_fromguid);
1173 	}
1174 
1175 	ra.vp = vp;
1176 	ra.voff = *voffp;
1177 	ra.bufsize = 1<<20;
1178 	ra.buf = kmem_alloc(ra.bufsize, KM_SLEEP);
1179 
1180 	/* these were verified in dmu_recv_begin */
1181 	ASSERT(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo) ==
1182 	    DMU_SUBSTREAM);
1183 	ASSERT(drc->drc_drrb->drr_type < DMU_OST_NUMTYPES);
1184 
1185 	/*
1186 	 * Open the objset we are modifying.
1187 	 */
1188 	VERIFY(dmu_objset_from_ds(drc->drc_real_ds, &os) == 0);
1189 
1190 	ASSERT(drc->drc_real_ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT);
1191 
1192 	featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
1193 
1194 	/* if this stream is dedup'ed, set up the avl tree for guid mapping */
1195 	if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
1196 		avl_create(&ra.guid_to_ds_map, guid_compare,
1197 		    sizeof (guid_map_entry_t),
1198 		    offsetof(guid_map_entry_t, avlnode));
1199 		(void) dmu_objset_find(drc->drc_top_ds, find_ds_by_guid,
1200 		    (void *)&ra.guid_to_ds_map,
1201 		    DS_FIND_CHILDREN);
1202 	}
1203 
1204 	/*
1205 	 * Read records and process them.
1206 	 */
1207 	pcksum = ra.cksum;
1208 	while (ra.err == 0 &&
1209 	    NULL != (drr = restore_read(&ra, sizeof (*drr)))) {
1210 		if (issig(JUSTLOOKING) && issig(FORREAL)) {
1211 			ra.err = EINTR;
1212 			goto out;
1213 		}
1214 
1215 		if (ra.byteswap)
1216 			backup_byteswap(drr);
1217 
1218 		switch (drr->drr_type) {
1219 		case DRR_OBJECT:
1220 		{
1221 			/*
1222 			 * We need to make a copy of the record header,
1223 			 * because restore_{object,write} may need to
1224 			 * restore_read(), which will invalidate drr.
1225 			 */
1226 			struct drr_object drro = drr->drr_u.drr_object;
1227 			ra.err = restore_object(&ra, os, &drro);
1228 			break;
1229 		}
1230 		case DRR_FREEOBJECTS:
1231 		{
1232 			struct drr_freeobjects drrfo =
1233 			    drr->drr_u.drr_freeobjects;
1234 			ra.err = restore_freeobjects(&ra, os, &drrfo);
1235 			break;
1236 		}
1237 		case DRR_WRITE:
1238 		{
1239 			struct drr_write drrw = drr->drr_u.drr_write;
1240 			ra.err = restore_write(&ra, os, &drrw);
1241 			break;
1242 		}
1243 		case DRR_WRITE_BYREF:
1244 		{
1245 			struct drr_write_byref drrwbr =
1246 			    drr->drr_u.drr_write_byref;
1247 			ra.err = restore_write_byref(&ra, os, &drrwbr);
1248 			break;
1249 		}
1250 		case DRR_FREE:
1251 		{
1252 			struct drr_free drrf = drr->drr_u.drr_free;
1253 			ra.err = restore_free(&ra, os, &drrf);
1254 			break;
1255 		}
1256 		case DRR_END:
1257 		{
1258 			struct drr_end drre = drr->drr_u.drr_end;
1259 			/*
1260 			 * We compare against the *previous* checksum
1261 			 * value, because the stored checksum is of
1262 			 * everything before the DRR_END record.
1263 			 */
1264 			if (!ZIO_CHECKSUM_EQUAL(drre.drr_checksum, pcksum))
1265 				ra.err = ECKSUM;
1266 			goto out;
1267 		}
1268 		default:
1269 			ra.err = EINVAL;
1270 			goto out;
1271 		}
1272 		pcksum = ra.cksum;
1273 	}
1274 	ASSERT(ra.err != 0);
1275 
1276 out:
1277 	if (ra.err != 0) {
1278 		/*
1279 		 * destroy what we created, so we don't leave it in the
1280 		 * inconsistent restoring state.
1281 		 */
1282 		txg_wait_synced(drc->drc_real_ds->ds_dir->dd_pool, 0);
1283 
1284 		(void) dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag,
1285 		    B_FALSE);
1286 		if (drc->drc_real_ds != drc->drc_logical_ds) {
1287 			mutex_exit(&drc->drc_logical_ds->ds_recvlock);
1288 			dsl_dataset_rele(drc->drc_logical_ds, dmu_recv_tag);
1289 		}
1290 	}
1291 
1292 	if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
1293 		void *cookie = NULL;
1294 
1295 		while (gmep = avl_destroy_nodes(&ra.guid_to_ds_map, &cookie)) {
1296 			dsl_dataset_rele(gmep->gme_ds, &ra.guid_to_ds_map);
1297 			kmem_free(gmep, sizeof (guid_map_entry_t));
1298 		}
1299 		avl_destroy(&ra.guid_to_ds_map);
1300 	}
1301 
1302 	kmem_free(ra.buf, ra.bufsize);
1303 	*voffp = ra.voff;
1304 	return (ra.err);
1305 }
1306 
1307 struct recvendsyncarg {
1308 	char *tosnap;
1309 	uint64_t creation_time;
1310 	uint64_t toguid;
1311 };
1312 
1313 static int
1314 recv_end_check(void *arg1, void *arg2, dmu_tx_t *tx)
1315 {
1316 	dsl_dataset_t *ds = arg1;
1317 	struct recvendsyncarg *resa = arg2;
1318 
1319 	return (dsl_dataset_snapshot_check(ds, resa->tosnap, tx));
1320 }
1321 
1322 static void
1323 recv_end_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
1324 {
1325 	dsl_dataset_t *ds = arg1;
1326 	struct recvendsyncarg *resa = arg2;
1327 
1328 	dsl_dataset_snapshot_sync(ds, resa->tosnap, cr, tx);
1329 
1330 	/* set snapshot's creation time and guid */
1331 	dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
1332 	ds->ds_prev->ds_phys->ds_creation_time = resa->creation_time;
1333 	ds->ds_prev->ds_phys->ds_guid = resa->toguid;
1334 	ds->ds_prev->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1335 
1336 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1337 	ds->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1338 }
1339 
1340 static int
1341 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
1342 {
1343 	struct recvendsyncarg resa;
1344 	dsl_dataset_t *ds = drc->drc_logical_ds;
1345 	int err;
1346 
1347 	/*
1348 	 * XXX hack; seems the ds is still dirty and dsl_pool_zil_clean()
1349 	 * expects it to have a ds_user_ptr (and zil), but clone_swap()
1350 	 * can close it.
1351 	 */
1352 	txg_wait_synced(ds->ds_dir->dd_pool, 0);
1353 
1354 	if (dsl_dataset_tryown(ds, FALSE, dmu_recv_tag)) {
1355 		err = dsl_dataset_clone_swap(drc->drc_real_ds, ds,
1356 		    drc->drc_force);
1357 		if (err)
1358 			goto out;
1359 	} else {
1360 		mutex_exit(&ds->ds_recvlock);
1361 		dsl_dataset_rele(ds, dmu_recv_tag);
1362 		(void) dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag,
1363 		    B_FALSE);
1364 		return (EBUSY);
1365 	}
1366 
1367 	resa.creation_time = drc->drc_drrb->drr_creation_time;
1368 	resa.toguid = drc->drc_drrb->drr_toguid;
1369 	resa.tosnap = drc->drc_tosnap;
1370 
1371 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1372 	    recv_end_check, recv_end_sync, ds, &resa, 3);
1373 	if (err) {
1374 		/* swap back */
1375 		(void) dsl_dataset_clone_swap(drc->drc_real_ds, ds, B_TRUE);
1376 	}
1377 
1378 out:
1379 	mutex_exit(&ds->ds_recvlock);
1380 	dsl_dataset_disown(ds, dmu_recv_tag);
1381 	(void) dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag, B_FALSE);
1382 	return (err);
1383 }
1384 
1385 static int
1386 dmu_recv_new_end(dmu_recv_cookie_t *drc)
1387 {
1388 	struct recvendsyncarg resa;
1389 	dsl_dataset_t *ds = drc->drc_logical_ds;
1390 	int err;
1391 
1392 	/*
1393 	 * XXX hack; seems the ds is still dirty and dsl_pool_zil_clean()
1394 	 * expects it to have a ds_user_ptr (and zil), but clone_swap()
1395 	 * can close it.
1396 	 */
1397 	txg_wait_synced(ds->ds_dir->dd_pool, 0);
1398 
1399 	resa.creation_time = drc->drc_drrb->drr_creation_time;
1400 	resa.toguid = drc->drc_drrb->drr_toguid;
1401 	resa.tosnap = drc->drc_tosnap;
1402 
1403 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1404 	    recv_end_check, recv_end_sync, ds, &resa, 3);
1405 	if (err) {
1406 		/* clean up the fs we just recv'd into */
1407 		(void) dsl_dataset_destroy(ds, dmu_recv_tag, B_FALSE);
1408 	} else {
1409 		/* release the hold from dmu_recv_begin */
1410 		dsl_dataset_disown(ds, dmu_recv_tag);
1411 	}
1412 	return (err);
1413 }
1414 
1415 int
1416 dmu_recv_end(dmu_recv_cookie_t *drc)
1417 {
1418 	if (drc->drc_logical_ds != drc->drc_real_ds)
1419 		return (dmu_recv_existing_end(drc));
1420 	else
1421 		return (dmu_recv_new_end(drc));
1422 }
1423