xref: /titanic_50/usr/src/uts/common/fs/zfs/dmu_send.c (revision 72102e7461c97dc268d21d9dd8f02da45f174acd)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24  * Copyright (c) 2013 by Delphix. All rights reserved.
25  * Copyright (c) 2012, Joyent, Inc. 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 #include <sys/dmu_send.h>
50 #include <sys/dsl_destroy.h>
51 
52 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
53 int zfs_send_corrupt_data = B_FALSE;
54 
55 static char *dmu_recv_tag = "dmu_recv_tag";
56 static const char *recv_clone_name = "%recv";
57 
58 static int
59 dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
60 {
61 	dsl_dataset_t *ds = dsp->dsa_os->os_dsl_dataset;
62 	ssize_t resid; /* have to get resid to get detailed errno */
63 	ASSERT0(len % 8);
64 
65 	fletcher_4_incremental_native(buf, len, &dsp->dsa_zc);
66 	dsp->dsa_err = vn_rdwr(UIO_WRITE, dsp->dsa_vp,
67 	    (caddr_t)buf, len,
68 	    0, UIO_SYSSPACE, FAPPEND, RLIM64_INFINITY, CRED(), &resid);
69 
70 	mutex_enter(&ds->ds_sendstream_lock);
71 	*dsp->dsa_off += len;
72 	mutex_exit(&ds->ds_sendstream_lock);
73 
74 	return (dsp->dsa_err);
75 }
76 
77 static int
78 dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
79     uint64_t length)
80 {
81 	struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free);
82 
83 	/*
84 	 * When we receive a free record, dbuf_free_range() assumes
85 	 * that the receiving system doesn't have any dbufs in the range
86 	 * being freed.  This is always true because there is a one-record
87 	 * constraint: we only send one WRITE record for any given
88 	 * object+offset.  We know that the one-record constraint is
89 	 * true because we always send data in increasing order by
90 	 * object,offset.
91 	 *
92 	 * If the increasing-order constraint ever changes, we should find
93 	 * another way to assert that the one-record constraint is still
94 	 * satisfied.
95 	 */
96 	ASSERT(object > dsp->dsa_last_data_object ||
97 	    (object == dsp->dsa_last_data_object &&
98 	    offset > dsp->dsa_last_data_offset));
99 
100 	/*
101 	 * If we are doing a non-incremental send, then there can't
102 	 * be any data in the dataset we're receiving into.  Therefore
103 	 * a free record would simply be a no-op.  Save space by not
104 	 * sending it to begin with.
105 	 */
106 	if (!dsp->dsa_incremental)
107 		return (0);
108 
109 	if (length != -1ULL && offset + length < offset)
110 		length = -1ULL;
111 
112 	/*
113 	 * If there is a pending op, but it's not PENDING_FREE, push it out,
114 	 * since free block aggregation can only be done for blocks of the
115 	 * same type (i.e., DRR_FREE records can only be aggregated with
116 	 * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
117 	 * aggregated with other DRR_FREEOBJECTS records.
118 	 */
119 	if (dsp->dsa_pending_op != PENDING_NONE &&
120 	    dsp->dsa_pending_op != PENDING_FREE) {
121 		if (dump_bytes(dsp, dsp->dsa_drr,
122 		    sizeof (dmu_replay_record_t)) != 0)
123 			return (SET_ERROR(EINTR));
124 		dsp->dsa_pending_op = PENDING_NONE;
125 	}
126 
127 	if (dsp->dsa_pending_op == PENDING_FREE) {
128 		/*
129 		 * There should never be a PENDING_FREE if length is -1
130 		 * (because dump_dnode is the only place where this
131 		 * function is called with a -1, and only after flushing
132 		 * any pending record).
133 		 */
134 		ASSERT(length != -1ULL);
135 		/*
136 		 * Check to see whether this free block can be aggregated
137 		 * with pending one.
138 		 */
139 		if (drrf->drr_object == object && drrf->drr_offset +
140 		    drrf->drr_length == offset) {
141 			drrf->drr_length += length;
142 			return (0);
143 		} else {
144 			/* not a continuation.  Push out pending record */
145 			if (dump_bytes(dsp, dsp->dsa_drr,
146 			    sizeof (dmu_replay_record_t)) != 0)
147 				return (SET_ERROR(EINTR));
148 			dsp->dsa_pending_op = PENDING_NONE;
149 		}
150 	}
151 	/* create a FREE record and make it pending */
152 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
153 	dsp->dsa_drr->drr_type = DRR_FREE;
154 	drrf->drr_object = object;
155 	drrf->drr_offset = offset;
156 	drrf->drr_length = length;
157 	drrf->drr_toguid = dsp->dsa_toguid;
158 	if (length == -1ULL) {
159 		if (dump_bytes(dsp, dsp->dsa_drr,
160 		    sizeof (dmu_replay_record_t)) != 0)
161 			return (SET_ERROR(EINTR));
162 	} else {
163 		dsp->dsa_pending_op = PENDING_FREE;
164 	}
165 
166 	return (0);
167 }
168 
169 static int
170 dump_data(dmu_sendarg_t *dsp, dmu_object_type_t type,
171     uint64_t object, uint64_t offset, int blksz, const blkptr_t *bp, void *data)
172 {
173 	struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write);
174 
175 	/*
176 	 * We send data in increasing object, offset order.
177 	 * See comment in dump_free() for details.
178 	 */
179 	ASSERT(object > dsp->dsa_last_data_object ||
180 	    (object == dsp->dsa_last_data_object &&
181 	    offset > dsp->dsa_last_data_offset));
182 	dsp->dsa_last_data_object = object;
183 	dsp->dsa_last_data_offset = offset + blksz - 1;
184 
185 	/*
186 	 * If there is any kind of pending aggregation (currently either
187 	 * a grouping of free objects or free blocks), push it out to
188 	 * the stream, since aggregation can't be done across operations
189 	 * of different types.
190 	 */
191 	if (dsp->dsa_pending_op != PENDING_NONE) {
192 		if (dump_bytes(dsp, dsp->dsa_drr,
193 		    sizeof (dmu_replay_record_t)) != 0)
194 			return (SET_ERROR(EINTR));
195 		dsp->dsa_pending_op = PENDING_NONE;
196 	}
197 	/* write a DATA record */
198 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
199 	dsp->dsa_drr->drr_type = DRR_WRITE;
200 	drrw->drr_object = object;
201 	drrw->drr_type = type;
202 	drrw->drr_offset = offset;
203 	drrw->drr_length = blksz;
204 	drrw->drr_toguid = dsp->dsa_toguid;
205 	drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
206 	if (zio_checksum_table[drrw->drr_checksumtype].ci_dedup)
207 		drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
208 	DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
209 	DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
210 	DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
211 	drrw->drr_key.ddk_cksum = bp->blk_cksum;
212 
213 	if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
214 		return (SET_ERROR(EINTR));
215 	if (dump_bytes(dsp, data, blksz) != 0)
216 		return (SET_ERROR(EINTR));
217 	return (0);
218 }
219 
220 static int
221 dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data)
222 {
223 	struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill);
224 
225 	if (dsp->dsa_pending_op != PENDING_NONE) {
226 		if (dump_bytes(dsp, dsp->dsa_drr,
227 		    sizeof (dmu_replay_record_t)) != 0)
228 			return (SET_ERROR(EINTR));
229 		dsp->dsa_pending_op = PENDING_NONE;
230 	}
231 
232 	/* write a SPILL record */
233 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
234 	dsp->dsa_drr->drr_type = DRR_SPILL;
235 	drrs->drr_object = object;
236 	drrs->drr_length = blksz;
237 	drrs->drr_toguid = dsp->dsa_toguid;
238 
239 	if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)))
240 		return (SET_ERROR(EINTR));
241 	if (dump_bytes(dsp, data, blksz))
242 		return (SET_ERROR(EINTR));
243 	return (0);
244 }
245 
246 static int
247 dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
248 {
249 	struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
250 
251 	/* See comment in dump_free(). */
252 	if (!dsp->dsa_incremental)
253 		return (0);
254 
255 	/*
256 	 * If there is a pending op, but it's not PENDING_FREEOBJECTS,
257 	 * push it out, since free block aggregation can only be done for
258 	 * blocks of the same type (i.e., DRR_FREE records can only be
259 	 * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
260 	 * can only be aggregated with other DRR_FREEOBJECTS records.
261 	 */
262 	if (dsp->dsa_pending_op != PENDING_NONE &&
263 	    dsp->dsa_pending_op != PENDING_FREEOBJECTS) {
264 		if (dump_bytes(dsp, dsp->dsa_drr,
265 		    sizeof (dmu_replay_record_t)) != 0)
266 			return (SET_ERROR(EINTR));
267 		dsp->dsa_pending_op = PENDING_NONE;
268 	}
269 	if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) {
270 		/*
271 		 * See whether this free object array can be aggregated
272 		 * with pending one
273 		 */
274 		if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
275 			drrfo->drr_numobjs += numobjs;
276 			return (0);
277 		} else {
278 			/* can't be aggregated.  Push out pending record */
279 			if (dump_bytes(dsp, dsp->dsa_drr,
280 			    sizeof (dmu_replay_record_t)) != 0)
281 				return (SET_ERROR(EINTR));
282 			dsp->dsa_pending_op = PENDING_NONE;
283 		}
284 	}
285 
286 	/* write a FREEOBJECTS record */
287 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
288 	dsp->dsa_drr->drr_type = DRR_FREEOBJECTS;
289 	drrfo->drr_firstobj = firstobj;
290 	drrfo->drr_numobjs = numobjs;
291 	drrfo->drr_toguid = dsp->dsa_toguid;
292 
293 	dsp->dsa_pending_op = PENDING_FREEOBJECTS;
294 
295 	return (0);
296 }
297 
298 static int
299 dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
300 {
301 	struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object);
302 
303 	if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
304 		return (dump_freeobjects(dsp, object, 1));
305 
306 	if (dsp->dsa_pending_op != PENDING_NONE) {
307 		if (dump_bytes(dsp, dsp->dsa_drr,
308 		    sizeof (dmu_replay_record_t)) != 0)
309 			return (SET_ERROR(EINTR));
310 		dsp->dsa_pending_op = PENDING_NONE;
311 	}
312 
313 	/* write an OBJECT record */
314 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
315 	dsp->dsa_drr->drr_type = DRR_OBJECT;
316 	drro->drr_object = object;
317 	drro->drr_type = dnp->dn_type;
318 	drro->drr_bonustype = dnp->dn_bonustype;
319 	drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
320 	drro->drr_bonuslen = dnp->dn_bonuslen;
321 	drro->drr_checksumtype = dnp->dn_checksum;
322 	drro->drr_compress = dnp->dn_compress;
323 	drro->drr_toguid = dsp->dsa_toguid;
324 
325 	if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
326 		return (SET_ERROR(EINTR));
327 
328 	if (dump_bytes(dsp, DN_BONUS(dnp), P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0)
329 		return (SET_ERROR(EINTR));
330 
331 	/* Free anything past the end of the file. */
332 	if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) *
333 	    (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0)
334 		return (SET_ERROR(EINTR));
335 	if (dsp->dsa_err != 0)
336 		return (SET_ERROR(EINTR));
337 	return (0);
338 }
339 
340 #define	BP_SPAN(dnp, level) \
341 	(((uint64_t)dnp->dn_datablkszsec) << (SPA_MINBLOCKSHIFT + \
342 	(level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)))
343 
344 /* ARGSUSED */
345 static int
346 backup_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
347     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
348 {
349 	dmu_sendarg_t *dsp = arg;
350 	dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
351 	int err = 0;
352 
353 	if (issig(JUSTLOOKING) && issig(FORREAL))
354 		return (SET_ERROR(EINTR));
355 
356 	if (zb->zb_object != DMU_META_DNODE_OBJECT &&
357 	    DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
358 		return (0);
359 	} else if (bp == NULL && zb->zb_object == DMU_META_DNODE_OBJECT) {
360 		uint64_t span = BP_SPAN(dnp, zb->zb_level);
361 		uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
362 		err = dump_freeobjects(dsp, dnobj, span >> DNODE_SHIFT);
363 	} else if (bp == NULL) {
364 		uint64_t span = BP_SPAN(dnp, zb->zb_level);
365 		err = dump_free(dsp, zb->zb_object, zb->zb_blkid * span, span);
366 	} else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
367 		return (0);
368 	} else if (type == DMU_OT_DNODE) {
369 		dnode_phys_t *blk;
370 		int i;
371 		int blksz = BP_GET_LSIZE(bp);
372 		uint32_t aflags = ARC_WAIT;
373 		arc_buf_t *abuf;
374 
375 		if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
376 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
377 		    &aflags, zb) != 0)
378 			return (SET_ERROR(EIO));
379 
380 		blk = abuf->b_data;
381 		for (i = 0; i < blksz >> DNODE_SHIFT; i++) {
382 			uint64_t dnobj = (zb->zb_blkid <<
383 			    (DNODE_BLOCK_SHIFT - DNODE_SHIFT)) + i;
384 			err = dump_dnode(dsp, dnobj, blk+i);
385 			if (err != 0)
386 				break;
387 		}
388 		(void) arc_buf_remove_ref(abuf, &abuf);
389 	} else if (type == DMU_OT_SA) {
390 		uint32_t aflags = ARC_WAIT;
391 		arc_buf_t *abuf;
392 		int blksz = BP_GET_LSIZE(bp);
393 
394 		if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
395 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
396 		    &aflags, zb) != 0)
397 			return (SET_ERROR(EIO));
398 
399 		err = dump_spill(dsp, zb->zb_object, blksz, abuf->b_data);
400 		(void) arc_buf_remove_ref(abuf, &abuf);
401 	} else { /* it's a level-0 block of a regular object */
402 		uint32_t aflags = ARC_WAIT;
403 		arc_buf_t *abuf;
404 		int blksz = BP_GET_LSIZE(bp);
405 
406 		if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
407 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
408 		    &aflags, zb) != 0) {
409 			if (zfs_send_corrupt_data) {
410 				/* Send a block filled with 0x"zfs badd bloc" */
411 				abuf = arc_buf_alloc(spa, blksz, &abuf,
412 				    ARC_BUFC_DATA);
413 				uint64_t *ptr;
414 				for (ptr = abuf->b_data;
415 				    (char *)ptr < (char *)abuf->b_data + blksz;
416 				    ptr++)
417 					*ptr = 0x2f5baddb10c;
418 			} else {
419 				return (SET_ERROR(EIO));
420 			}
421 		}
422 
423 		err = dump_data(dsp, type, zb->zb_object, zb->zb_blkid * blksz,
424 		    blksz, bp, abuf->b_data);
425 		(void) arc_buf_remove_ref(abuf, &abuf);
426 	}
427 
428 	ASSERT(err == 0 || err == EINTR);
429 	return (err);
430 }
431 
432 /*
433  * Releases dp, ds, and fromds, using the specified tag.
434  */
435 static int
436 dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *ds,
437     dsl_dataset_t *fromds, int outfd, vnode_t *vp, offset_t *off)
438 {
439 	objset_t *os;
440 	dmu_replay_record_t *drr;
441 	dmu_sendarg_t *dsp;
442 	int err;
443 	uint64_t fromtxg = 0;
444 
445 	if (fromds != NULL && !dsl_dataset_is_before(ds, fromds)) {
446 		dsl_dataset_rele(fromds, tag);
447 		dsl_dataset_rele(ds, tag);
448 		dsl_pool_rele(dp, tag);
449 		return (SET_ERROR(EXDEV));
450 	}
451 
452 	err = dmu_objset_from_ds(ds, &os);
453 	if (err != 0) {
454 		if (fromds != NULL)
455 			dsl_dataset_rele(fromds, tag);
456 		dsl_dataset_rele(ds, tag);
457 		dsl_pool_rele(dp, tag);
458 		return (err);
459 	}
460 
461 	drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
462 	drr->drr_type = DRR_BEGIN;
463 	drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
464 	DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
465 	    DMU_SUBSTREAM);
466 
467 #ifdef _KERNEL
468 	if (dmu_objset_type(os) == DMU_OST_ZFS) {
469 		uint64_t version;
470 		if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0) {
471 			kmem_free(drr, sizeof (dmu_replay_record_t));
472 			if (fromds != NULL)
473 				dsl_dataset_rele(fromds, tag);
474 			dsl_dataset_rele(ds, tag);
475 			dsl_pool_rele(dp, tag);
476 			return (SET_ERROR(EINVAL));
477 		}
478 		if (version >= ZPL_VERSION_SA) {
479 			DMU_SET_FEATUREFLAGS(
480 			    drr->drr_u.drr_begin.drr_versioninfo,
481 			    DMU_BACKUP_FEATURE_SA_SPILL);
482 		}
483 	}
484 #endif
485 
486 	drr->drr_u.drr_begin.drr_creation_time =
487 	    ds->ds_phys->ds_creation_time;
488 	drr->drr_u.drr_begin.drr_type = dmu_objset_type(os);
489 	if (fromds != NULL && ds->ds_dir != fromds->ds_dir)
490 		drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
491 	drr->drr_u.drr_begin.drr_toguid = ds->ds_phys->ds_guid;
492 	if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
493 		drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
494 
495 	if (fromds != NULL)
496 		drr->drr_u.drr_begin.drr_fromguid = fromds->ds_phys->ds_guid;
497 	dsl_dataset_name(ds, drr->drr_u.drr_begin.drr_toname);
498 
499 	if (fromds != NULL) {
500 		fromtxg = fromds->ds_phys->ds_creation_txg;
501 		dsl_dataset_rele(fromds, tag);
502 		fromds = NULL;
503 	}
504 
505 	dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP);
506 
507 	dsp->dsa_drr = drr;
508 	dsp->dsa_vp = vp;
509 	dsp->dsa_outfd = outfd;
510 	dsp->dsa_proc = curproc;
511 	dsp->dsa_os = os;
512 	dsp->dsa_off = off;
513 	dsp->dsa_toguid = ds->ds_phys->ds_guid;
514 	ZIO_SET_CHECKSUM(&dsp->dsa_zc, 0, 0, 0, 0);
515 	dsp->dsa_pending_op = PENDING_NONE;
516 	dsp->dsa_incremental = (fromtxg != 0);
517 
518 	mutex_enter(&ds->ds_sendstream_lock);
519 	list_insert_head(&ds->ds_sendstreams, dsp);
520 	mutex_exit(&ds->ds_sendstream_lock);
521 
522 	dsl_dataset_long_hold(ds, FTAG);
523 	dsl_pool_rele(dp, tag);
524 
525 	if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0) {
526 		err = dsp->dsa_err;
527 		goto out;
528 	}
529 
530 	err = traverse_dataset(ds, fromtxg, TRAVERSE_PRE | TRAVERSE_PREFETCH,
531 	    backup_cb, dsp);
532 
533 	if (dsp->dsa_pending_op != PENDING_NONE)
534 		if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0)
535 			err = SET_ERROR(EINTR);
536 
537 	if (err != 0) {
538 		if (err == EINTR && dsp->dsa_err != 0)
539 			err = dsp->dsa_err;
540 		goto out;
541 	}
542 
543 	bzero(drr, sizeof (dmu_replay_record_t));
544 	drr->drr_type = DRR_END;
545 	drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc;
546 	drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid;
547 
548 	if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0) {
549 		err = dsp->dsa_err;
550 		goto out;
551 	}
552 
553 out:
554 	mutex_enter(&ds->ds_sendstream_lock);
555 	list_remove(&ds->ds_sendstreams, dsp);
556 	mutex_exit(&ds->ds_sendstream_lock);
557 
558 	kmem_free(drr, sizeof (dmu_replay_record_t));
559 	kmem_free(dsp, sizeof (dmu_sendarg_t));
560 
561 	dsl_dataset_long_rele(ds, FTAG);
562 	dsl_dataset_rele(ds, tag);
563 
564 	return (err);
565 }
566 
567 int
568 dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
569     int outfd, vnode_t *vp, offset_t *off)
570 {
571 	dsl_pool_t *dp;
572 	dsl_dataset_t *ds;
573 	dsl_dataset_t *fromds = NULL;
574 	int err;
575 
576 	err = dsl_pool_hold(pool, FTAG, &dp);
577 	if (err != 0)
578 		return (err);
579 
580 	err = dsl_dataset_hold_obj(dp, tosnap, FTAG, &ds);
581 	if (err != 0) {
582 		dsl_pool_rele(dp, FTAG);
583 		return (err);
584 	}
585 
586 	if (fromsnap != 0) {
587 		err = dsl_dataset_hold_obj(dp, fromsnap, FTAG, &fromds);
588 		if (err != 0) {
589 			dsl_dataset_rele(ds, FTAG);
590 			dsl_pool_rele(dp, FTAG);
591 			return (err);
592 		}
593 	}
594 
595 	return (dmu_send_impl(FTAG, dp, ds, fromds, outfd, vp, off));
596 }
597 
598 int
599 dmu_send(const char *tosnap, const char *fromsnap,
600     int outfd, vnode_t *vp, offset_t *off)
601 {
602 	dsl_pool_t *dp;
603 	dsl_dataset_t *ds;
604 	dsl_dataset_t *fromds = NULL;
605 	int err;
606 
607 	if (strchr(tosnap, '@') == NULL)
608 		return (SET_ERROR(EINVAL));
609 	if (fromsnap != NULL && strchr(fromsnap, '@') == NULL)
610 		return (SET_ERROR(EINVAL));
611 
612 	err = dsl_pool_hold(tosnap, FTAG, &dp);
613 	if (err != 0)
614 		return (err);
615 
616 	err = dsl_dataset_hold(dp, tosnap, FTAG, &ds);
617 	if (err != 0) {
618 		dsl_pool_rele(dp, FTAG);
619 		return (err);
620 	}
621 
622 	if (fromsnap != NULL) {
623 		err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds);
624 		if (err != 0) {
625 			dsl_dataset_rele(ds, FTAG);
626 			dsl_pool_rele(dp, FTAG);
627 			return (err);
628 		}
629 	}
630 	return (dmu_send_impl(FTAG, dp, ds, fromds, outfd, vp, off));
631 }
632 
633 int
634 dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds, uint64_t *sizep)
635 {
636 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
637 	int err;
638 	uint64_t size;
639 
640 	ASSERT(dsl_pool_config_held(dp));
641 
642 	/* tosnap must be a snapshot */
643 	if (!dsl_dataset_is_snapshot(ds))
644 		return (SET_ERROR(EINVAL));
645 
646 	/*
647 	 * fromsnap must be an earlier snapshot from the same fs as tosnap,
648 	 * or the origin's fs.
649 	 */
650 	if (fromds != NULL && !dsl_dataset_is_before(ds, fromds))
651 		return (SET_ERROR(EXDEV));
652 
653 	/* Get uncompressed size estimate of changed data. */
654 	if (fromds == NULL) {
655 		size = ds->ds_phys->ds_uncompressed_bytes;
656 	} else {
657 		uint64_t used, comp;
658 		err = dsl_dataset_space_written(fromds, ds,
659 		    &used, &comp, &size);
660 		if (err != 0)
661 			return (err);
662 	}
663 
664 	/*
665 	 * Assume that space (both on-disk and in-stream) is dominated by
666 	 * data.  We will adjust for indirect blocks and the copies property,
667 	 * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
668 	 */
669 
670 	/*
671 	 * Subtract out approximate space used by indirect blocks.
672 	 * Assume most space is used by data blocks (non-indirect, non-dnode).
673 	 * Assume all blocks are recordsize.  Assume ditto blocks and
674 	 * internal fragmentation counter out compression.
675 	 *
676 	 * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
677 	 * block, which we observe in practice.
678 	 */
679 	uint64_t recordsize;
680 	err = dsl_prop_get_int_ds(ds, "recordsize", &recordsize);
681 	if (err != 0)
682 		return (err);
683 	size -= size / recordsize * sizeof (blkptr_t);
684 
685 	/* Add in the space for the record associated with each block. */
686 	size += size / recordsize * sizeof (dmu_replay_record_t);
687 
688 	*sizep = size;
689 
690 	return (0);
691 }
692 
693 typedef struct dmu_recv_begin_arg {
694 	const char *drba_origin;
695 	dmu_recv_cookie_t *drba_cookie;
696 	cred_t *drba_cred;
697 	uint64_t drba_snapobj;
698 } dmu_recv_begin_arg_t;
699 
700 static int
701 recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds,
702     uint64_t fromguid)
703 {
704 	uint64_t val;
705 	int error;
706 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
707 
708 	/* temporary clone name must not exist */
709 	error = zap_lookup(dp->dp_meta_objset,
710 	    ds->ds_dir->dd_phys->dd_child_dir_zapobj, recv_clone_name,
711 	    8, 1, &val);
712 	if (error != ENOENT)
713 		return (error == 0 ? EBUSY : error);
714 
715 	/* new snapshot name must not exist */
716 	error = zap_lookup(dp->dp_meta_objset,
717 	    ds->ds_phys->ds_snapnames_zapobj, drba->drba_cookie->drc_tosnap,
718 	    8, 1, &val);
719 	if (error != ENOENT)
720 		return (error == 0 ? EEXIST : error);
721 
722 	if (fromguid != 0) {
723 		dsl_dataset_t *snap;
724 		uint64_t obj = ds->ds_phys->ds_prev_snap_obj;
725 
726 		/* Find snapshot in this dir that matches fromguid. */
727 		while (obj != 0) {
728 			error = dsl_dataset_hold_obj(dp, obj, FTAG,
729 			    &snap);
730 			if (error != 0)
731 				return (SET_ERROR(ENODEV));
732 			if (snap->ds_dir != ds->ds_dir) {
733 				dsl_dataset_rele(snap, FTAG);
734 				return (SET_ERROR(ENODEV));
735 			}
736 			if (snap->ds_phys->ds_guid == fromguid)
737 				break;
738 			obj = snap->ds_phys->ds_prev_snap_obj;
739 			dsl_dataset_rele(snap, FTAG);
740 		}
741 		if (obj == 0)
742 			return (SET_ERROR(ENODEV));
743 
744 		if (drba->drba_cookie->drc_force) {
745 			drba->drba_snapobj = obj;
746 		} else {
747 			/*
748 			 * If we are not forcing, there must be no
749 			 * changes since fromsnap.
750 			 */
751 			if (dsl_dataset_modified_since_snap(ds, snap)) {
752 				dsl_dataset_rele(snap, FTAG);
753 				return (SET_ERROR(ETXTBSY));
754 			}
755 			drba->drba_snapobj = ds->ds_prev->ds_object;
756 		}
757 
758 		dsl_dataset_rele(snap, FTAG);
759 	} else {
760 		/* if full, most recent snapshot must be $ORIGIN */
761 		if (ds->ds_phys->ds_prev_snap_txg >= TXG_INITIAL)
762 			return (SET_ERROR(ENODEV));
763 		drba->drba_snapobj = ds->ds_phys->ds_prev_snap_obj;
764 	}
765 
766 	return (0);
767 
768 }
769 
770 static int
771 dmu_recv_begin_check(void *arg, dmu_tx_t *tx)
772 {
773 	dmu_recv_begin_arg_t *drba = arg;
774 	dsl_pool_t *dp = dmu_tx_pool(tx);
775 	struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
776 	uint64_t fromguid = drrb->drr_fromguid;
777 	int flags = drrb->drr_flags;
778 	int error;
779 	dsl_dataset_t *ds;
780 	const char *tofs = drba->drba_cookie->drc_tofs;
781 
782 	/* already checked */
783 	ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
784 
785 	if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
786 	    DMU_COMPOUNDSTREAM ||
787 	    drrb->drr_type >= DMU_OST_NUMTYPES ||
788 	    ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL))
789 		return (SET_ERROR(EINVAL));
790 
791 	/* Verify pool version supports SA if SA_SPILL feature set */
792 	if ((DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
793 	    DMU_BACKUP_FEATURE_SA_SPILL) &&
794 	    spa_version(dp->dp_spa) < SPA_VERSION_SA) {
795 		return (SET_ERROR(ENOTSUP));
796 	}
797 
798 	error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
799 	if (error == 0) {
800 		/* target fs already exists; recv into temp clone */
801 
802 		/* Can't recv a clone into an existing fs */
803 		if (flags & DRR_FLAG_CLONE) {
804 			dsl_dataset_rele(ds, FTAG);
805 			return (SET_ERROR(EINVAL));
806 		}
807 
808 		error = recv_begin_check_existing_impl(drba, ds, fromguid);
809 		dsl_dataset_rele(ds, FTAG);
810 	} else if (error == ENOENT) {
811 		/* target fs does not exist; must be a full backup or clone */
812 		char buf[MAXNAMELEN];
813 
814 		/*
815 		 * If it's a non-clone incremental, we are missing the
816 		 * target fs, so fail the recv.
817 		 */
818 		if (fromguid != 0 && !(flags & DRR_FLAG_CLONE))
819 			return (SET_ERROR(ENOENT));
820 
821 		/* Open the parent of tofs */
822 		ASSERT3U(strlen(tofs), <, MAXNAMELEN);
823 		(void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1);
824 		error = dsl_dataset_hold(dp, buf, FTAG, &ds);
825 		if (error != 0)
826 			return (error);
827 
828 		if (drba->drba_origin != NULL) {
829 			dsl_dataset_t *origin;
830 			error = dsl_dataset_hold(dp, drba->drba_origin,
831 			    FTAG, &origin);
832 			if (error != 0) {
833 				dsl_dataset_rele(ds, FTAG);
834 				return (error);
835 			}
836 			if (!dsl_dataset_is_snapshot(origin)) {
837 				dsl_dataset_rele(origin, FTAG);
838 				dsl_dataset_rele(ds, FTAG);
839 				return (SET_ERROR(EINVAL));
840 			}
841 			if (origin->ds_phys->ds_guid != fromguid) {
842 				dsl_dataset_rele(origin, FTAG);
843 				dsl_dataset_rele(ds, FTAG);
844 				return (SET_ERROR(ENODEV));
845 			}
846 			dsl_dataset_rele(origin, FTAG);
847 		}
848 		dsl_dataset_rele(ds, FTAG);
849 		error = 0;
850 	}
851 	return (error);
852 }
853 
854 static void
855 dmu_recv_begin_sync(void *arg, dmu_tx_t *tx)
856 {
857 	dmu_recv_begin_arg_t *drba = arg;
858 	dsl_pool_t *dp = dmu_tx_pool(tx);
859 	struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
860 	const char *tofs = drba->drba_cookie->drc_tofs;
861 	dsl_dataset_t *ds, *newds;
862 	uint64_t dsobj;
863 	int error;
864 	uint64_t crflags;
865 
866 	crflags = (drrb->drr_flags & DRR_FLAG_CI_DATA) ?
867 	    DS_FLAG_CI_DATASET : 0;
868 
869 	error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
870 	if (error == 0) {
871 		/* create temporary clone */
872 		dsl_dataset_t *snap = NULL;
873 		if (drba->drba_snapobj != 0) {
874 			VERIFY0(dsl_dataset_hold_obj(dp,
875 			    drba->drba_snapobj, FTAG, &snap));
876 		}
877 		dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name,
878 		    snap, crflags, drba->drba_cred, tx);
879 		dsl_dataset_rele(snap, FTAG);
880 		dsl_dataset_rele(ds, FTAG);
881 	} else {
882 		dsl_dir_t *dd;
883 		const char *tail;
884 		dsl_dataset_t *origin = NULL;
885 
886 		VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail));
887 
888 		if (drba->drba_origin != NULL) {
889 			VERIFY0(dsl_dataset_hold(dp, drba->drba_origin,
890 			    FTAG, &origin));
891 		}
892 
893 		/* Create new dataset. */
894 		dsobj = dsl_dataset_create_sync(dd,
895 		    strrchr(tofs, '/') + 1,
896 		    origin, crflags, drba->drba_cred, tx);
897 		if (origin != NULL)
898 			dsl_dataset_rele(origin, FTAG);
899 		dsl_dir_rele(dd, FTAG);
900 		drba->drba_cookie->drc_newfs = B_TRUE;
901 	}
902 	VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds));
903 
904 	dmu_buf_will_dirty(newds->ds_dbuf, tx);
905 	newds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
906 
907 	/*
908 	 * If we actually created a non-clone, we need to create the
909 	 * objset in our new dataset.
910 	 */
911 	if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) {
912 		(void) dmu_objset_create_impl(dp->dp_spa,
913 		    newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx);
914 	}
915 
916 	drba->drba_cookie->drc_ds = newds;
917 
918 	spa_history_log_internal_ds(newds, "receive", tx, "");
919 }
920 
921 /*
922  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
923  * succeeds; otherwise we will leak the holds on the datasets.
924  */
925 int
926 dmu_recv_begin(char *tofs, char *tosnap, struct drr_begin *drrb,
927     boolean_t force, char *origin, dmu_recv_cookie_t *drc)
928 {
929 	dmu_recv_begin_arg_t drba = { 0 };
930 	dmu_replay_record_t *drr;
931 
932 	bzero(drc, sizeof (dmu_recv_cookie_t));
933 	drc->drc_drrb = drrb;
934 	drc->drc_tosnap = tosnap;
935 	drc->drc_tofs = tofs;
936 	drc->drc_force = force;
937 
938 	if (drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
939 		drc->drc_byteswap = B_TRUE;
940 	else if (drrb->drr_magic != DMU_BACKUP_MAGIC)
941 		return (SET_ERROR(EINVAL));
942 
943 	drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
944 	drr->drr_type = DRR_BEGIN;
945 	drr->drr_u.drr_begin = *drc->drc_drrb;
946 	if (drc->drc_byteswap) {
947 		fletcher_4_incremental_byteswap(drr,
948 		    sizeof (dmu_replay_record_t), &drc->drc_cksum);
949 	} else {
950 		fletcher_4_incremental_native(drr,
951 		    sizeof (dmu_replay_record_t), &drc->drc_cksum);
952 	}
953 	kmem_free(drr, sizeof (dmu_replay_record_t));
954 
955 	if (drc->drc_byteswap) {
956 		drrb->drr_magic = BSWAP_64(drrb->drr_magic);
957 		drrb->drr_versioninfo = BSWAP_64(drrb->drr_versioninfo);
958 		drrb->drr_creation_time = BSWAP_64(drrb->drr_creation_time);
959 		drrb->drr_type = BSWAP_32(drrb->drr_type);
960 		drrb->drr_toguid = BSWAP_64(drrb->drr_toguid);
961 		drrb->drr_fromguid = BSWAP_64(drrb->drr_fromguid);
962 	}
963 
964 	drba.drba_origin = origin;
965 	drba.drba_cookie = drc;
966 	drba.drba_cred = CRED();
967 
968 	return (dsl_sync_task(tofs, dmu_recv_begin_check, dmu_recv_begin_sync,
969 	    &drba, 5));
970 }
971 
972 struct restorearg {
973 	int err;
974 	boolean_t byteswap;
975 	vnode_t *vp;
976 	char *buf;
977 	uint64_t voff;
978 	int bufsize; /* amount of memory allocated for buf */
979 	zio_cksum_t cksum;
980 	avl_tree_t *guid_to_ds_map;
981 };
982 
983 typedef struct guid_map_entry {
984 	uint64_t	guid;
985 	dsl_dataset_t	*gme_ds;
986 	avl_node_t	avlnode;
987 } guid_map_entry_t;
988 
989 static int
990 guid_compare(const void *arg1, const void *arg2)
991 {
992 	const guid_map_entry_t *gmep1 = arg1;
993 	const guid_map_entry_t *gmep2 = arg2;
994 
995 	if (gmep1->guid < gmep2->guid)
996 		return (-1);
997 	else if (gmep1->guid > gmep2->guid)
998 		return (1);
999 	return (0);
1000 }
1001 
1002 static void
1003 free_guid_map_onexit(void *arg)
1004 {
1005 	avl_tree_t *ca = arg;
1006 	void *cookie = NULL;
1007 	guid_map_entry_t *gmep;
1008 
1009 	while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
1010 		dsl_dataset_long_rele(gmep->gme_ds, gmep);
1011 		dsl_dataset_rele(gmep->gme_ds, gmep);
1012 		kmem_free(gmep, sizeof (guid_map_entry_t));
1013 	}
1014 	avl_destroy(ca);
1015 	kmem_free(ca, sizeof (avl_tree_t));
1016 }
1017 
1018 static void *
1019 restore_read(struct restorearg *ra, int len)
1020 {
1021 	void *rv;
1022 	int done = 0;
1023 
1024 	/* some things will require 8-byte alignment, so everything must */
1025 	ASSERT0(len % 8);
1026 
1027 	while (done < len) {
1028 		ssize_t resid;
1029 
1030 		ra->err = vn_rdwr(UIO_READ, ra->vp,
1031 		    (caddr_t)ra->buf + done, len - done,
1032 		    ra->voff, UIO_SYSSPACE, FAPPEND,
1033 		    RLIM64_INFINITY, CRED(), &resid);
1034 
1035 		if (resid == len - done)
1036 			ra->err = SET_ERROR(EINVAL);
1037 		ra->voff += len - done - resid;
1038 		done = len - resid;
1039 		if (ra->err != 0)
1040 			return (NULL);
1041 	}
1042 
1043 	ASSERT3U(done, ==, len);
1044 	rv = ra->buf;
1045 	if (ra->byteswap)
1046 		fletcher_4_incremental_byteswap(rv, len, &ra->cksum);
1047 	else
1048 		fletcher_4_incremental_native(rv, len, &ra->cksum);
1049 	return (rv);
1050 }
1051 
1052 static void
1053 backup_byteswap(dmu_replay_record_t *drr)
1054 {
1055 #define	DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
1056 #define	DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
1057 	drr->drr_type = BSWAP_32(drr->drr_type);
1058 	drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
1059 	switch (drr->drr_type) {
1060 	case DRR_BEGIN:
1061 		DO64(drr_begin.drr_magic);
1062 		DO64(drr_begin.drr_versioninfo);
1063 		DO64(drr_begin.drr_creation_time);
1064 		DO32(drr_begin.drr_type);
1065 		DO32(drr_begin.drr_flags);
1066 		DO64(drr_begin.drr_toguid);
1067 		DO64(drr_begin.drr_fromguid);
1068 		break;
1069 	case DRR_OBJECT:
1070 		DO64(drr_object.drr_object);
1071 		/* DO64(drr_object.drr_allocation_txg); */
1072 		DO32(drr_object.drr_type);
1073 		DO32(drr_object.drr_bonustype);
1074 		DO32(drr_object.drr_blksz);
1075 		DO32(drr_object.drr_bonuslen);
1076 		DO64(drr_object.drr_toguid);
1077 		break;
1078 	case DRR_FREEOBJECTS:
1079 		DO64(drr_freeobjects.drr_firstobj);
1080 		DO64(drr_freeobjects.drr_numobjs);
1081 		DO64(drr_freeobjects.drr_toguid);
1082 		break;
1083 	case DRR_WRITE:
1084 		DO64(drr_write.drr_object);
1085 		DO32(drr_write.drr_type);
1086 		DO64(drr_write.drr_offset);
1087 		DO64(drr_write.drr_length);
1088 		DO64(drr_write.drr_toguid);
1089 		DO64(drr_write.drr_key.ddk_cksum.zc_word[0]);
1090 		DO64(drr_write.drr_key.ddk_cksum.zc_word[1]);
1091 		DO64(drr_write.drr_key.ddk_cksum.zc_word[2]);
1092 		DO64(drr_write.drr_key.ddk_cksum.zc_word[3]);
1093 		DO64(drr_write.drr_key.ddk_prop);
1094 		break;
1095 	case DRR_WRITE_BYREF:
1096 		DO64(drr_write_byref.drr_object);
1097 		DO64(drr_write_byref.drr_offset);
1098 		DO64(drr_write_byref.drr_length);
1099 		DO64(drr_write_byref.drr_toguid);
1100 		DO64(drr_write_byref.drr_refguid);
1101 		DO64(drr_write_byref.drr_refobject);
1102 		DO64(drr_write_byref.drr_refoffset);
1103 		DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[0]);
1104 		DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[1]);
1105 		DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[2]);
1106 		DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[3]);
1107 		DO64(drr_write_byref.drr_key.ddk_prop);
1108 		break;
1109 	case DRR_FREE:
1110 		DO64(drr_free.drr_object);
1111 		DO64(drr_free.drr_offset);
1112 		DO64(drr_free.drr_length);
1113 		DO64(drr_free.drr_toguid);
1114 		break;
1115 	case DRR_SPILL:
1116 		DO64(drr_spill.drr_object);
1117 		DO64(drr_spill.drr_length);
1118 		DO64(drr_spill.drr_toguid);
1119 		break;
1120 	case DRR_END:
1121 		DO64(drr_end.drr_checksum.zc_word[0]);
1122 		DO64(drr_end.drr_checksum.zc_word[1]);
1123 		DO64(drr_end.drr_checksum.zc_word[2]);
1124 		DO64(drr_end.drr_checksum.zc_word[3]);
1125 		DO64(drr_end.drr_toguid);
1126 		break;
1127 	}
1128 #undef DO64
1129 #undef DO32
1130 }
1131 
1132 static int
1133 restore_object(struct restorearg *ra, objset_t *os, struct drr_object *drro)
1134 {
1135 	int err;
1136 	dmu_tx_t *tx;
1137 	void *data = NULL;
1138 
1139 	if (drro->drr_type == DMU_OT_NONE ||
1140 	    !DMU_OT_IS_VALID(drro->drr_type) ||
1141 	    !DMU_OT_IS_VALID(drro->drr_bonustype) ||
1142 	    drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
1143 	    drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
1144 	    P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
1145 	    drro->drr_blksz < SPA_MINBLOCKSIZE ||
1146 	    drro->drr_blksz > SPA_MAXBLOCKSIZE ||
1147 	    drro->drr_bonuslen > DN_MAX_BONUSLEN) {
1148 		return (SET_ERROR(EINVAL));
1149 	}
1150 
1151 	err = dmu_object_info(os, drro->drr_object, NULL);
1152 
1153 	if (err != 0 && err != ENOENT)
1154 		return (SET_ERROR(EINVAL));
1155 
1156 	if (drro->drr_bonuslen) {
1157 		data = restore_read(ra, P2ROUNDUP(drro->drr_bonuslen, 8));
1158 		if (ra->err != 0)
1159 			return (ra->err);
1160 	}
1161 
1162 	if (err == ENOENT) {
1163 		/* currently free, want to be allocated */
1164 		tx = dmu_tx_create(os);
1165 		dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1166 		err = dmu_tx_assign(tx, TXG_WAIT);
1167 		if (err != 0) {
1168 			dmu_tx_abort(tx);
1169 			return (err);
1170 		}
1171 		err = dmu_object_claim(os, drro->drr_object,
1172 		    drro->drr_type, drro->drr_blksz,
1173 		    drro->drr_bonustype, drro->drr_bonuslen, tx);
1174 		dmu_tx_commit(tx);
1175 	} else {
1176 		/* currently allocated, want to be allocated */
1177 		err = dmu_object_reclaim(os, drro->drr_object,
1178 		    drro->drr_type, drro->drr_blksz,
1179 		    drro->drr_bonustype, drro->drr_bonuslen);
1180 	}
1181 	if (err != 0) {
1182 		return (SET_ERROR(EINVAL));
1183 	}
1184 
1185 	tx = dmu_tx_create(os);
1186 	dmu_tx_hold_bonus(tx, drro->drr_object);
1187 	err = dmu_tx_assign(tx, TXG_WAIT);
1188 	if (err != 0) {
1189 		dmu_tx_abort(tx);
1190 		return (err);
1191 	}
1192 
1193 	dmu_object_set_checksum(os, drro->drr_object, drro->drr_checksumtype,
1194 	    tx);
1195 	dmu_object_set_compress(os, drro->drr_object, drro->drr_compress, tx);
1196 
1197 	if (data != NULL) {
1198 		dmu_buf_t *db;
1199 
1200 		VERIFY(0 == dmu_bonus_hold(os, drro->drr_object, FTAG, &db));
1201 		dmu_buf_will_dirty(db, tx);
1202 
1203 		ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
1204 		bcopy(data, db->db_data, drro->drr_bonuslen);
1205 		if (ra->byteswap) {
1206 			dmu_object_byteswap_t byteswap =
1207 			    DMU_OT_BYTESWAP(drro->drr_bonustype);
1208 			dmu_ot_byteswap[byteswap].ob_func(db->db_data,
1209 			    drro->drr_bonuslen);
1210 		}
1211 		dmu_buf_rele(db, FTAG);
1212 	}
1213 	dmu_tx_commit(tx);
1214 	return (0);
1215 }
1216 
1217 /* ARGSUSED */
1218 static int
1219 restore_freeobjects(struct restorearg *ra, objset_t *os,
1220     struct drr_freeobjects *drrfo)
1221 {
1222 	uint64_t obj;
1223 
1224 	if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
1225 		return (SET_ERROR(EINVAL));
1226 
1227 	for (obj = drrfo->drr_firstobj;
1228 	    obj < drrfo->drr_firstobj + drrfo->drr_numobjs;
1229 	    (void) dmu_object_next(os, &obj, FALSE, 0)) {
1230 		int err;
1231 
1232 		if (dmu_object_info(os, obj, NULL) != 0)
1233 			continue;
1234 
1235 		err = dmu_free_long_object(os, obj);
1236 		if (err != 0)
1237 			return (err);
1238 	}
1239 	return (0);
1240 }
1241 
1242 static int
1243 restore_write(struct restorearg *ra, objset_t *os,
1244     struct drr_write *drrw)
1245 {
1246 	dmu_tx_t *tx;
1247 	void *data;
1248 	int err;
1249 
1250 	if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
1251 	    !DMU_OT_IS_VALID(drrw->drr_type))
1252 		return (SET_ERROR(EINVAL));
1253 
1254 	data = restore_read(ra, drrw->drr_length);
1255 	if (data == NULL)
1256 		return (ra->err);
1257 
1258 	if (dmu_object_info(os, drrw->drr_object, NULL) != 0)
1259 		return (SET_ERROR(EINVAL));
1260 
1261 	tx = dmu_tx_create(os);
1262 
1263 	dmu_tx_hold_write(tx, drrw->drr_object,
1264 	    drrw->drr_offset, drrw->drr_length);
1265 	err = dmu_tx_assign(tx, TXG_WAIT);
1266 	if (err != 0) {
1267 		dmu_tx_abort(tx);
1268 		return (err);
1269 	}
1270 	if (ra->byteswap) {
1271 		dmu_object_byteswap_t byteswap =
1272 		    DMU_OT_BYTESWAP(drrw->drr_type);
1273 		dmu_ot_byteswap[byteswap].ob_func(data, drrw->drr_length);
1274 	}
1275 	dmu_write(os, drrw->drr_object,
1276 	    drrw->drr_offset, drrw->drr_length, data, tx);
1277 	dmu_tx_commit(tx);
1278 	return (0);
1279 }
1280 
1281 /*
1282  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
1283  * streams to refer to a copy of the data that is already on the
1284  * system because it came in earlier in the stream.  This function
1285  * finds the earlier copy of the data, and uses that copy instead of
1286  * data from the stream to fulfill this write.
1287  */
1288 static int
1289 restore_write_byref(struct restorearg *ra, objset_t *os,
1290     struct drr_write_byref *drrwbr)
1291 {
1292 	dmu_tx_t *tx;
1293 	int err;
1294 	guid_map_entry_t gmesrch;
1295 	guid_map_entry_t *gmep;
1296 	avl_index_t	where;
1297 	objset_t *ref_os = NULL;
1298 	dmu_buf_t *dbp;
1299 
1300 	if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
1301 		return (SET_ERROR(EINVAL));
1302 
1303 	/*
1304 	 * If the GUID of the referenced dataset is different from the
1305 	 * GUID of the target dataset, find the referenced dataset.
1306 	 */
1307 	if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
1308 		gmesrch.guid = drrwbr->drr_refguid;
1309 		if ((gmep = avl_find(ra->guid_to_ds_map, &gmesrch,
1310 		    &where)) == NULL) {
1311 			return (SET_ERROR(EINVAL));
1312 		}
1313 		if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
1314 			return (SET_ERROR(EINVAL));
1315 	} else {
1316 		ref_os = os;
1317 	}
1318 
1319 	if (err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
1320 	    drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH))
1321 		return (err);
1322 
1323 	tx = dmu_tx_create(os);
1324 
1325 	dmu_tx_hold_write(tx, drrwbr->drr_object,
1326 	    drrwbr->drr_offset, drrwbr->drr_length);
1327 	err = dmu_tx_assign(tx, TXG_WAIT);
1328 	if (err != 0) {
1329 		dmu_tx_abort(tx);
1330 		return (err);
1331 	}
1332 	dmu_write(os, drrwbr->drr_object,
1333 	    drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
1334 	dmu_buf_rele(dbp, FTAG);
1335 	dmu_tx_commit(tx);
1336 	return (0);
1337 }
1338 
1339 static int
1340 restore_spill(struct restorearg *ra, objset_t *os, struct drr_spill *drrs)
1341 {
1342 	dmu_tx_t *tx;
1343 	void *data;
1344 	dmu_buf_t *db, *db_spill;
1345 	int err;
1346 
1347 	if (drrs->drr_length < SPA_MINBLOCKSIZE ||
1348 	    drrs->drr_length > SPA_MAXBLOCKSIZE)
1349 		return (SET_ERROR(EINVAL));
1350 
1351 	data = restore_read(ra, drrs->drr_length);
1352 	if (data == NULL)
1353 		return (ra->err);
1354 
1355 	if (dmu_object_info(os, drrs->drr_object, NULL) != 0)
1356 		return (SET_ERROR(EINVAL));
1357 
1358 	VERIFY(0 == dmu_bonus_hold(os, drrs->drr_object, FTAG, &db));
1359 	if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
1360 		dmu_buf_rele(db, FTAG);
1361 		return (err);
1362 	}
1363 
1364 	tx = dmu_tx_create(os);
1365 
1366 	dmu_tx_hold_spill(tx, db->db_object);
1367 
1368 	err = dmu_tx_assign(tx, TXG_WAIT);
1369 	if (err != 0) {
1370 		dmu_buf_rele(db, FTAG);
1371 		dmu_buf_rele(db_spill, FTAG);
1372 		dmu_tx_abort(tx);
1373 		return (err);
1374 	}
1375 	dmu_buf_will_dirty(db_spill, tx);
1376 
1377 	if (db_spill->db_size < drrs->drr_length)
1378 		VERIFY(0 == dbuf_spill_set_blksz(db_spill,
1379 		    drrs->drr_length, tx));
1380 	bcopy(data, db_spill->db_data, drrs->drr_length);
1381 
1382 	dmu_buf_rele(db, FTAG);
1383 	dmu_buf_rele(db_spill, FTAG);
1384 
1385 	dmu_tx_commit(tx);
1386 	return (0);
1387 }
1388 
1389 /* ARGSUSED */
1390 static int
1391 restore_free(struct restorearg *ra, objset_t *os,
1392     struct drr_free *drrf)
1393 {
1394 	int err;
1395 
1396 	if (drrf->drr_length != -1ULL &&
1397 	    drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
1398 		return (SET_ERROR(EINVAL));
1399 
1400 	if (dmu_object_info(os, drrf->drr_object, NULL) != 0)
1401 		return (SET_ERROR(EINVAL));
1402 
1403 	err = dmu_free_long_range(os, drrf->drr_object,
1404 	    drrf->drr_offset, drrf->drr_length);
1405 	return (err);
1406 }
1407 
1408 /* used to destroy the drc_ds on error */
1409 static void
1410 dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc)
1411 {
1412 	char name[MAXNAMELEN];
1413 	dsl_dataset_name(drc->drc_ds, name);
1414 	dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
1415 	(void) dsl_destroy_head(name);
1416 }
1417 
1418 /*
1419  * NB: callers *must* call dmu_recv_end() if this succeeds.
1420  */
1421 int
1422 dmu_recv_stream(dmu_recv_cookie_t *drc, vnode_t *vp, offset_t *voffp,
1423     int cleanup_fd, uint64_t *action_handlep)
1424 {
1425 	struct restorearg ra = { 0 };
1426 	dmu_replay_record_t *drr;
1427 	objset_t *os;
1428 	zio_cksum_t pcksum;
1429 	int featureflags;
1430 
1431 	ra.byteswap = drc->drc_byteswap;
1432 	ra.cksum = drc->drc_cksum;
1433 	ra.vp = vp;
1434 	ra.voff = *voffp;
1435 	ra.bufsize = 1<<20;
1436 	ra.buf = kmem_alloc(ra.bufsize, KM_SLEEP);
1437 
1438 	/* these were verified in dmu_recv_begin */
1439 	ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==,
1440 	    DMU_SUBSTREAM);
1441 	ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES);
1442 
1443 	/*
1444 	 * Open the objset we are modifying.
1445 	 */
1446 	VERIFY0(dmu_objset_from_ds(drc->drc_ds, &os));
1447 
1448 	ASSERT(drc->drc_ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT);
1449 
1450 	featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
1451 
1452 	/* if this stream is dedup'ed, set up the avl tree for guid mapping */
1453 	if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
1454 		minor_t minor;
1455 
1456 		if (cleanup_fd == -1) {
1457 			ra.err = SET_ERROR(EBADF);
1458 			goto out;
1459 		}
1460 		ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
1461 		if (ra.err != 0) {
1462 			cleanup_fd = -1;
1463 			goto out;
1464 		}
1465 
1466 		if (*action_handlep == 0) {
1467 			ra.guid_to_ds_map =
1468 			    kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
1469 			avl_create(ra.guid_to_ds_map, guid_compare,
1470 			    sizeof (guid_map_entry_t),
1471 			    offsetof(guid_map_entry_t, avlnode));
1472 			ra.err = zfs_onexit_add_cb(minor,
1473 			    free_guid_map_onexit, ra.guid_to_ds_map,
1474 			    action_handlep);
1475 			if (ra.err != 0)
1476 				goto out;
1477 		} else {
1478 			ra.err = zfs_onexit_cb_data(minor, *action_handlep,
1479 			    (void **)&ra.guid_to_ds_map);
1480 			if (ra.err != 0)
1481 				goto out;
1482 		}
1483 
1484 		drc->drc_guid_to_ds_map = ra.guid_to_ds_map;
1485 	}
1486 
1487 	/*
1488 	 * Read records and process them.
1489 	 */
1490 	pcksum = ra.cksum;
1491 	while (ra.err == 0 &&
1492 	    NULL != (drr = restore_read(&ra, sizeof (*drr)))) {
1493 		if (issig(JUSTLOOKING) && issig(FORREAL)) {
1494 			ra.err = SET_ERROR(EINTR);
1495 			goto out;
1496 		}
1497 
1498 		if (ra.byteswap)
1499 			backup_byteswap(drr);
1500 
1501 		switch (drr->drr_type) {
1502 		case DRR_OBJECT:
1503 		{
1504 			/*
1505 			 * We need to make a copy of the record header,
1506 			 * because restore_{object,write} may need to
1507 			 * restore_read(), which will invalidate drr.
1508 			 */
1509 			struct drr_object drro = drr->drr_u.drr_object;
1510 			ra.err = restore_object(&ra, os, &drro);
1511 			break;
1512 		}
1513 		case DRR_FREEOBJECTS:
1514 		{
1515 			struct drr_freeobjects drrfo =
1516 			    drr->drr_u.drr_freeobjects;
1517 			ra.err = restore_freeobjects(&ra, os, &drrfo);
1518 			break;
1519 		}
1520 		case DRR_WRITE:
1521 		{
1522 			struct drr_write drrw = drr->drr_u.drr_write;
1523 			ra.err = restore_write(&ra, os, &drrw);
1524 			break;
1525 		}
1526 		case DRR_WRITE_BYREF:
1527 		{
1528 			struct drr_write_byref drrwbr =
1529 			    drr->drr_u.drr_write_byref;
1530 			ra.err = restore_write_byref(&ra, os, &drrwbr);
1531 			break;
1532 		}
1533 		case DRR_FREE:
1534 		{
1535 			struct drr_free drrf = drr->drr_u.drr_free;
1536 			ra.err = restore_free(&ra, os, &drrf);
1537 			break;
1538 		}
1539 		case DRR_END:
1540 		{
1541 			struct drr_end drre = drr->drr_u.drr_end;
1542 			/*
1543 			 * We compare against the *previous* checksum
1544 			 * value, because the stored checksum is of
1545 			 * everything before the DRR_END record.
1546 			 */
1547 			if (!ZIO_CHECKSUM_EQUAL(drre.drr_checksum, pcksum))
1548 				ra.err = SET_ERROR(ECKSUM);
1549 			goto out;
1550 		}
1551 		case DRR_SPILL:
1552 		{
1553 			struct drr_spill drrs = drr->drr_u.drr_spill;
1554 			ra.err = restore_spill(&ra, os, &drrs);
1555 			break;
1556 		}
1557 		default:
1558 			ra.err = SET_ERROR(EINVAL);
1559 			goto out;
1560 		}
1561 		pcksum = ra.cksum;
1562 	}
1563 	ASSERT(ra.err != 0);
1564 
1565 out:
1566 	if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
1567 		zfs_onexit_fd_rele(cleanup_fd);
1568 
1569 	if (ra.err != 0) {
1570 		/*
1571 		 * destroy what we created, so we don't leave it in the
1572 		 * inconsistent restoring state.
1573 		 */
1574 		dmu_recv_cleanup_ds(drc);
1575 	}
1576 
1577 	kmem_free(ra.buf, ra.bufsize);
1578 	*voffp = ra.voff;
1579 	return (ra.err);
1580 }
1581 
1582 static int
1583 dmu_recv_end_check(void *arg, dmu_tx_t *tx)
1584 {
1585 	dmu_recv_cookie_t *drc = arg;
1586 	dsl_pool_t *dp = dmu_tx_pool(tx);
1587 	int error;
1588 
1589 	ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag);
1590 
1591 	if (!drc->drc_newfs) {
1592 		dsl_dataset_t *origin_head;
1593 
1594 		error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head);
1595 		if (error != 0)
1596 			return (error);
1597 		if (drc->drc_force) {
1598 			/*
1599 			 * We will destroy any snapshots in tofs (i.e. before
1600 			 * origin_head) that are after the origin (which is
1601 			 * the snap before drc_ds, because drc_ds can not
1602 			 * have any snaps of its own).
1603 			 */
1604 			uint64_t obj = origin_head->ds_phys->ds_prev_snap_obj;
1605 			while (obj != drc->drc_ds->ds_phys->ds_prev_snap_obj) {
1606 				dsl_dataset_t *snap;
1607 				error = dsl_dataset_hold_obj(dp, obj, FTAG,
1608 				    &snap);
1609 				if (error != 0)
1610 					return (error);
1611 				if (snap->ds_dir != origin_head->ds_dir)
1612 					error = SET_ERROR(EINVAL);
1613 				if (error == 0)  {
1614 					error = dsl_destroy_snapshot_check_impl(
1615 					    snap, B_FALSE);
1616 				}
1617 				obj = snap->ds_phys->ds_prev_snap_obj;
1618 				dsl_dataset_rele(snap, FTAG);
1619 				if (error != 0)
1620 					return (error);
1621 			}
1622 		}
1623 		error = dsl_dataset_clone_swap_check_impl(drc->drc_ds,
1624 		    origin_head, drc->drc_force, drc->drc_owner, tx);
1625 		if (error != 0) {
1626 			dsl_dataset_rele(origin_head, FTAG);
1627 			return (error);
1628 		}
1629 		error = dsl_dataset_snapshot_check_impl(origin_head,
1630 		    drc->drc_tosnap, tx, B_TRUE);
1631 		dsl_dataset_rele(origin_head, FTAG);
1632 		if (error != 0)
1633 			return (error);
1634 
1635 		error = dsl_destroy_head_check_impl(drc->drc_ds, 1);
1636 	} else {
1637 		error = dsl_dataset_snapshot_check_impl(drc->drc_ds,
1638 		    drc->drc_tosnap, tx, B_TRUE);
1639 	}
1640 	return (error);
1641 }
1642 
1643 static void
1644 dmu_recv_end_sync(void *arg, dmu_tx_t *tx)
1645 {
1646 	dmu_recv_cookie_t *drc = arg;
1647 	dsl_pool_t *dp = dmu_tx_pool(tx);
1648 
1649 	spa_history_log_internal_ds(drc->drc_ds, "finish receiving",
1650 	    tx, "snap=%s", drc->drc_tosnap);
1651 
1652 	if (!drc->drc_newfs) {
1653 		dsl_dataset_t *origin_head;
1654 
1655 		VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG,
1656 		    &origin_head));
1657 
1658 		if (drc->drc_force) {
1659 			/*
1660 			 * Destroy any snapshots of drc_tofs (origin_head)
1661 			 * after the origin (the snap before drc_ds).
1662 			 */
1663 			uint64_t obj = origin_head->ds_phys->ds_prev_snap_obj;
1664 			while (obj != drc->drc_ds->ds_phys->ds_prev_snap_obj) {
1665 				dsl_dataset_t *snap;
1666 				VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG,
1667 				    &snap));
1668 				ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir);
1669 				obj = snap->ds_phys->ds_prev_snap_obj;
1670 				dsl_destroy_snapshot_sync_impl(snap,
1671 				    B_FALSE, tx);
1672 				dsl_dataset_rele(snap, FTAG);
1673 			}
1674 		}
1675 		VERIFY3P(drc->drc_ds->ds_prev, ==,
1676 		    origin_head->ds_prev);
1677 
1678 		dsl_dataset_clone_swap_sync_impl(drc->drc_ds,
1679 		    origin_head, tx);
1680 		dsl_dataset_snapshot_sync_impl(origin_head,
1681 		    drc->drc_tosnap, tx);
1682 
1683 		/* set snapshot's creation time and guid */
1684 		dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx);
1685 		origin_head->ds_prev->ds_phys->ds_creation_time =
1686 		    drc->drc_drrb->drr_creation_time;
1687 		origin_head->ds_prev->ds_phys->ds_guid =
1688 		    drc->drc_drrb->drr_toguid;
1689 		origin_head->ds_prev->ds_phys->ds_flags &=
1690 		    ~DS_FLAG_INCONSISTENT;
1691 
1692 		dmu_buf_will_dirty(origin_head->ds_dbuf, tx);
1693 		origin_head->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1694 
1695 		dsl_dataset_rele(origin_head, FTAG);
1696 		dsl_destroy_head_sync_impl(drc->drc_ds, tx);
1697 
1698 		if (drc->drc_owner != NULL)
1699 			VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner);
1700 	} else {
1701 		dsl_dataset_t *ds = drc->drc_ds;
1702 
1703 		dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx);
1704 
1705 		/* set snapshot's creation time and guid */
1706 		dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
1707 		ds->ds_prev->ds_phys->ds_creation_time =
1708 		    drc->drc_drrb->drr_creation_time;
1709 		ds->ds_prev->ds_phys->ds_guid = drc->drc_drrb->drr_toguid;
1710 		ds->ds_prev->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1711 
1712 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
1713 		ds->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1714 	}
1715 	drc->drc_newsnapobj = drc->drc_ds->ds_phys->ds_prev_snap_obj;
1716 	/*
1717 	 * Release the hold from dmu_recv_begin.  This must be done before
1718 	 * we return to open context, so that when we free the dataset's dnode,
1719 	 * we can evict its bonus buffer.
1720 	 */
1721 	dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
1722 	drc->drc_ds = NULL;
1723 }
1724 
1725 static int
1726 add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj)
1727 {
1728 	dsl_pool_t *dp;
1729 	dsl_dataset_t *snapds;
1730 	guid_map_entry_t *gmep;
1731 	int err;
1732 
1733 	ASSERT(guid_map != NULL);
1734 
1735 	err = dsl_pool_hold(name, FTAG, &dp);
1736 	if (err != 0)
1737 		return (err);
1738 	gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP);
1739 	err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds);
1740 	if (err == 0) {
1741 		gmep->guid = snapds->ds_phys->ds_guid;
1742 		gmep->gme_ds = snapds;
1743 		avl_add(guid_map, gmep);
1744 		dsl_dataset_long_hold(snapds, gmep);
1745 	} else {
1746 		kmem_free(gmep, sizeof (*gmep));
1747 	}
1748 
1749 	dsl_pool_rele(dp, FTAG);
1750 	return (err);
1751 }
1752 
1753 static int dmu_recv_end_modified_blocks = 3;
1754 
1755 static int
1756 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
1757 {
1758 	int error;
1759 	char name[MAXNAMELEN];
1760 
1761 #ifdef _KERNEL
1762 	/*
1763 	 * We will be destroying the ds; make sure its origin is unmounted if
1764 	 * necessary.
1765 	 */
1766 	dsl_dataset_name(drc->drc_ds, name);
1767 	zfs_destroy_unmount_origin(name);
1768 #endif
1769 
1770 	error = dsl_sync_task(drc->drc_tofs,
1771 	    dmu_recv_end_check, dmu_recv_end_sync, drc,
1772 	    dmu_recv_end_modified_blocks);
1773 
1774 	if (error != 0)
1775 		dmu_recv_cleanup_ds(drc);
1776 	return (error);
1777 }
1778 
1779 static int
1780 dmu_recv_new_end(dmu_recv_cookie_t *drc)
1781 {
1782 	int error;
1783 
1784 	error = dsl_sync_task(drc->drc_tofs,
1785 	    dmu_recv_end_check, dmu_recv_end_sync, drc,
1786 	    dmu_recv_end_modified_blocks);
1787 
1788 	if (error != 0) {
1789 		dmu_recv_cleanup_ds(drc);
1790 	} else if (drc->drc_guid_to_ds_map != NULL) {
1791 		(void) add_ds_to_guidmap(drc->drc_tofs,
1792 		    drc->drc_guid_to_ds_map,
1793 		    drc->drc_newsnapobj);
1794 	}
1795 	return (error);
1796 }
1797 
1798 int
1799 dmu_recv_end(dmu_recv_cookie_t *drc, void *owner)
1800 {
1801 	drc->drc_owner = owner;
1802 
1803 	if (drc->drc_newfs)
1804 		return (dmu_recv_new_end(drc));
1805 	else
1806 		return (dmu_recv_existing_end(drc));
1807 }
1808 
1809 /*
1810  * Return TRUE if this objset is currently being received into.
1811  */
1812 boolean_t
1813 dmu_objset_is_receiving(objset_t *os)
1814 {
1815 	return (os->os_dsl_dataset != NULL &&
1816 	    os->os_dsl_dataset->ds_owner == dmu_recv_tag);
1817 }
1818