xref: /titanic_41/usr/src/uts/common/fs/zfs/dmu_send.c (revision 6d24d8e5e4f6a2d14992d9392981d8a16d695aec)
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) 2011, 2014 by Delphix. All rights reserved.
25  * Copyright (c) 2014, 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 #include <sys/blkptr.h>
52 #include <sys/dsl_bookmark.h>
53 #include <sys/zfeature.h>
54 
55 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
56 int zfs_send_corrupt_data = B_FALSE;
57 
58 static char *dmu_recv_tag = "dmu_recv_tag";
59 static const char *recv_clone_name = "%recv";
60 
61 static int
62 dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
63 {
64 	dsl_dataset_t *ds = dsp->dsa_os->os_dsl_dataset;
65 	ssize_t resid; /* have to get resid to get detailed errno */
66 	ASSERT0(len % 8);
67 
68 	fletcher_4_incremental_native(buf, len, &dsp->dsa_zc);
69 	dsp->dsa_err = vn_rdwr(UIO_WRITE, dsp->dsa_vp,
70 	    (caddr_t)buf, len,
71 	    0, UIO_SYSSPACE, FAPPEND, RLIM64_INFINITY, CRED(), &resid);
72 
73 	mutex_enter(&ds->ds_sendstream_lock);
74 	*dsp->dsa_off += len;
75 	mutex_exit(&ds->ds_sendstream_lock);
76 
77 	return (dsp->dsa_err);
78 }
79 
80 static int
81 dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
82     uint64_t length)
83 {
84 	struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free);
85 
86 	/*
87 	 * When we receive a free record, dbuf_free_range() assumes
88 	 * that the receiving system doesn't have any dbufs in the range
89 	 * being freed.  This is always true because there is a one-record
90 	 * constraint: we only send one WRITE record for any given
91 	 * object+offset.  We know that the one-record constraint is
92 	 * true because we always send data in increasing order by
93 	 * object,offset.
94 	 *
95 	 * If the increasing-order constraint ever changes, we should find
96 	 * another way to assert that the one-record constraint is still
97 	 * satisfied.
98 	 */
99 	ASSERT(object > dsp->dsa_last_data_object ||
100 	    (object == dsp->dsa_last_data_object &&
101 	    offset > dsp->dsa_last_data_offset));
102 
103 	/*
104 	 * If we are doing a non-incremental send, then there can't
105 	 * be any data in the dataset we're receiving into.  Therefore
106 	 * a free record would simply be a no-op.  Save space by not
107 	 * sending it to begin with.
108 	 */
109 	if (!dsp->dsa_incremental)
110 		return (0);
111 
112 	if (length != -1ULL && offset + length < offset)
113 		length = -1ULL;
114 
115 	/*
116 	 * If there is a pending op, but it's not PENDING_FREE, push it out,
117 	 * since free block aggregation can only be done for blocks of the
118 	 * same type (i.e., DRR_FREE records can only be aggregated with
119 	 * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
120 	 * aggregated with other DRR_FREEOBJECTS records.
121 	 */
122 	if (dsp->dsa_pending_op != PENDING_NONE &&
123 	    dsp->dsa_pending_op != PENDING_FREE) {
124 		if (dump_bytes(dsp, dsp->dsa_drr,
125 		    sizeof (dmu_replay_record_t)) != 0)
126 			return (SET_ERROR(EINTR));
127 		dsp->dsa_pending_op = PENDING_NONE;
128 	}
129 
130 	if (dsp->dsa_pending_op == PENDING_FREE) {
131 		/*
132 		 * There should never be a PENDING_FREE if length is -1
133 		 * (because dump_dnode is the only place where this
134 		 * function is called with a -1, and only after flushing
135 		 * any pending record).
136 		 */
137 		ASSERT(length != -1ULL);
138 		/*
139 		 * Check to see whether this free block can be aggregated
140 		 * with pending one.
141 		 */
142 		if (drrf->drr_object == object && drrf->drr_offset +
143 		    drrf->drr_length == offset) {
144 			drrf->drr_length += length;
145 			return (0);
146 		} else {
147 			/* not a continuation.  Push out pending record */
148 			if (dump_bytes(dsp, dsp->dsa_drr,
149 			    sizeof (dmu_replay_record_t)) != 0)
150 				return (SET_ERROR(EINTR));
151 			dsp->dsa_pending_op = PENDING_NONE;
152 		}
153 	}
154 	/* create a FREE record and make it pending */
155 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
156 	dsp->dsa_drr->drr_type = DRR_FREE;
157 	drrf->drr_object = object;
158 	drrf->drr_offset = offset;
159 	drrf->drr_length = length;
160 	drrf->drr_toguid = dsp->dsa_toguid;
161 	if (length == -1ULL) {
162 		if (dump_bytes(dsp, dsp->dsa_drr,
163 		    sizeof (dmu_replay_record_t)) != 0)
164 			return (SET_ERROR(EINTR));
165 	} else {
166 		dsp->dsa_pending_op = PENDING_FREE;
167 	}
168 
169 	return (0);
170 }
171 
172 static int
173 dump_write(dmu_sendarg_t *dsp, dmu_object_type_t type,
174     uint64_t object, uint64_t offset, int blksz, const blkptr_t *bp, void *data)
175 {
176 	struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write);
177 
178 	/*
179 	 * We send data in increasing object, offset order.
180 	 * See comment in dump_free() for details.
181 	 */
182 	ASSERT(object > dsp->dsa_last_data_object ||
183 	    (object == dsp->dsa_last_data_object &&
184 	    offset > dsp->dsa_last_data_offset));
185 	dsp->dsa_last_data_object = object;
186 	dsp->dsa_last_data_offset = offset + blksz - 1;
187 
188 	/*
189 	 * If there is any kind of pending aggregation (currently either
190 	 * a grouping of free objects or free blocks), push it out to
191 	 * the stream, since aggregation can't be done across operations
192 	 * of different types.
193 	 */
194 	if (dsp->dsa_pending_op != PENDING_NONE) {
195 		if (dump_bytes(dsp, dsp->dsa_drr,
196 		    sizeof (dmu_replay_record_t)) != 0)
197 			return (SET_ERROR(EINTR));
198 		dsp->dsa_pending_op = PENDING_NONE;
199 	}
200 	/* write a DATA record */
201 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
202 	dsp->dsa_drr->drr_type = DRR_WRITE;
203 	drrw->drr_object = object;
204 	drrw->drr_type = type;
205 	drrw->drr_offset = offset;
206 	drrw->drr_length = blksz;
207 	drrw->drr_toguid = dsp->dsa_toguid;
208 	if (BP_IS_EMBEDDED(bp)) {
209 		/*
210 		 * There's no pre-computed checksum of embedded BP's, so
211 		 * (like fletcher4-checkummed blocks) userland will have
212 		 * to compute a dedup-capable checksum itself.
213 		 */
214 		drrw->drr_checksumtype = ZIO_CHECKSUM_OFF;
215 	} else {
216 		drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
217 		if (zio_checksum_table[drrw->drr_checksumtype].ci_dedup)
218 			drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
219 		DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
220 		DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
221 		DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
222 		drrw->drr_key.ddk_cksum = bp->blk_cksum;
223 	}
224 
225 	if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
226 		return (SET_ERROR(EINTR));
227 	if (dump_bytes(dsp, data, blksz) != 0)
228 		return (SET_ERROR(EINTR));
229 	return (0);
230 }
231 
232 static int
233 dump_write_embedded(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
234     int blksz, const blkptr_t *bp)
235 {
236 	char buf[BPE_PAYLOAD_SIZE];
237 	struct drr_write_embedded *drrw =
238 	    &(dsp->dsa_drr->drr_u.drr_write_embedded);
239 
240 	if (dsp->dsa_pending_op != PENDING_NONE) {
241 		if (dump_bytes(dsp, dsp->dsa_drr,
242 		    sizeof (dmu_replay_record_t)) != 0)
243 			return (EINTR);
244 		dsp->dsa_pending_op = PENDING_NONE;
245 	}
246 
247 	ASSERT(BP_IS_EMBEDDED(bp));
248 
249 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
250 	dsp->dsa_drr->drr_type = DRR_WRITE_EMBEDDED;
251 	drrw->drr_object = object;
252 	drrw->drr_offset = offset;
253 	drrw->drr_length = blksz;
254 	drrw->drr_toguid = dsp->dsa_toguid;
255 	drrw->drr_compression = BP_GET_COMPRESS(bp);
256 	drrw->drr_etype = BPE_GET_ETYPE(bp);
257 	drrw->drr_lsize = BPE_GET_LSIZE(bp);
258 	drrw->drr_psize = BPE_GET_PSIZE(bp);
259 
260 	decode_embedded_bp_compressed(bp, buf);
261 
262 	if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
263 		return (EINTR);
264 	if (dump_bytes(dsp, buf, P2ROUNDUP(drrw->drr_psize, 8)) != 0)
265 		return (EINTR);
266 	return (0);
267 }
268 
269 static int
270 dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data)
271 {
272 	struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill);
273 
274 	if (dsp->dsa_pending_op != PENDING_NONE) {
275 		if (dump_bytes(dsp, dsp->dsa_drr,
276 		    sizeof (dmu_replay_record_t)) != 0)
277 			return (SET_ERROR(EINTR));
278 		dsp->dsa_pending_op = PENDING_NONE;
279 	}
280 
281 	/* write a SPILL record */
282 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
283 	dsp->dsa_drr->drr_type = DRR_SPILL;
284 	drrs->drr_object = object;
285 	drrs->drr_length = blksz;
286 	drrs->drr_toguid = dsp->dsa_toguid;
287 
288 	if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)))
289 		return (SET_ERROR(EINTR));
290 	if (dump_bytes(dsp, data, blksz))
291 		return (SET_ERROR(EINTR));
292 	return (0);
293 }
294 
295 static int
296 dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
297 {
298 	struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
299 
300 	/* See comment in dump_free(). */
301 	if (!dsp->dsa_incremental)
302 		return (0);
303 
304 	/*
305 	 * If there is a pending op, but it's not PENDING_FREEOBJECTS,
306 	 * push it out, since free block aggregation can only be done for
307 	 * blocks of the same type (i.e., DRR_FREE records can only be
308 	 * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
309 	 * can only be aggregated with other DRR_FREEOBJECTS records.
310 	 */
311 	if (dsp->dsa_pending_op != PENDING_NONE &&
312 	    dsp->dsa_pending_op != PENDING_FREEOBJECTS) {
313 		if (dump_bytes(dsp, dsp->dsa_drr,
314 		    sizeof (dmu_replay_record_t)) != 0)
315 			return (SET_ERROR(EINTR));
316 		dsp->dsa_pending_op = PENDING_NONE;
317 	}
318 	if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) {
319 		/*
320 		 * See whether this free object array can be aggregated
321 		 * with pending one
322 		 */
323 		if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
324 			drrfo->drr_numobjs += numobjs;
325 			return (0);
326 		} else {
327 			/* can't be aggregated.  Push out pending record */
328 			if (dump_bytes(dsp, dsp->dsa_drr,
329 			    sizeof (dmu_replay_record_t)) != 0)
330 				return (SET_ERROR(EINTR));
331 			dsp->dsa_pending_op = PENDING_NONE;
332 		}
333 	}
334 
335 	/* write a FREEOBJECTS record */
336 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
337 	dsp->dsa_drr->drr_type = DRR_FREEOBJECTS;
338 	drrfo->drr_firstobj = firstobj;
339 	drrfo->drr_numobjs = numobjs;
340 	drrfo->drr_toguid = dsp->dsa_toguid;
341 
342 	dsp->dsa_pending_op = PENDING_FREEOBJECTS;
343 
344 	return (0);
345 }
346 
347 static int
348 dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
349 {
350 	struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object);
351 
352 	if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
353 		return (dump_freeobjects(dsp, object, 1));
354 
355 	if (dsp->dsa_pending_op != PENDING_NONE) {
356 		if (dump_bytes(dsp, dsp->dsa_drr,
357 		    sizeof (dmu_replay_record_t)) != 0)
358 			return (SET_ERROR(EINTR));
359 		dsp->dsa_pending_op = PENDING_NONE;
360 	}
361 
362 	/* write an OBJECT record */
363 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
364 	dsp->dsa_drr->drr_type = DRR_OBJECT;
365 	drro->drr_object = object;
366 	drro->drr_type = dnp->dn_type;
367 	drro->drr_bonustype = dnp->dn_bonustype;
368 	drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
369 	drro->drr_bonuslen = dnp->dn_bonuslen;
370 	drro->drr_checksumtype = dnp->dn_checksum;
371 	drro->drr_compress = dnp->dn_compress;
372 	drro->drr_toguid = dsp->dsa_toguid;
373 
374 	if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
375 		return (SET_ERROR(EINTR));
376 
377 	if (dump_bytes(dsp, DN_BONUS(dnp), P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0)
378 		return (SET_ERROR(EINTR));
379 
380 	/* Free anything past the end of the file. */
381 	if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) *
382 	    (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0)
383 		return (SET_ERROR(EINTR));
384 	if (dsp->dsa_err != 0)
385 		return (SET_ERROR(EINTR));
386 	return (0);
387 }
388 
389 static boolean_t
390 backup_do_embed(dmu_sendarg_t *dsp, const blkptr_t *bp)
391 {
392 	if (!BP_IS_EMBEDDED(bp))
393 		return (B_FALSE);
394 
395 	/*
396 	 * Compression function must be legacy, or explicitly enabled.
397 	 */
398 	if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS &&
399 	    !(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4)))
400 		return (B_FALSE);
401 
402 	/*
403 	 * Embed type must be explicitly enabled.
404 	 */
405 	switch (BPE_GET_ETYPE(bp)) {
406 	case BP_EMBEDDED_TYPE_DATA:
407 		if (dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)
408 			return (B_TRUE);
409 		break;
410 	default:
411 		return (B_FALSE);
412 	}
413 	return (B_FALSE);
414 }
415 
416 #define	BP_SPAN(dnp, level) \
417 	(((uint64_t)dnp->dn_datablkszsec) << (SPA_MINBLOCKSHIFT + \
418 	(level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)))
419 
420 /* ARGSUSED */
421 static int
422 backup_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
423     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
424 {
425 	dmu_sendarg_t *dsp = arg;
426 	dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
427 	int err = 0;
428 
429 	if (issig(JUSTLOOKING) && issig(FORREAL))
430 		return (SET_ERROR(EINTR));
431 
432 	if (zb->zb_object != DMU_META_DNODE_OBJECT &&
433 	    DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
434 		return (0);
435 	} else if (zb->zb_level == ZB_ZIL_LEVEL) {
436 		/*
437 		 * If we are sending a non-snapshot (which is allowed on
438 		 * read-only pools), it may have a ZIL, which must be ignored.
439 		 */
440 		return (0);
441 	} else if (BP_IS_HOLE(bp) &&
442 	    zb->zb_object == DMU_META_DNODE_OBJECT) {
443 		uint64_t span = BP_SPAN(dnp, zb->zb_level);
444 		uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
445 		err = dump_freeobjects(dsp, dnobj, span >> DNODE_SHIFT);
446 	} else if (BP_IS_HOLE(bp)) {
447 		uint64_t span = BP_SPAN(dnp, zb->zb_level);
448 		err = dump_free(dsp, zb->zb_object, zb->zb_blkid * span, span);
449 	} else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
450 		return (0);
451 	} else if (type == DMU_OT_DNODE) {
452 		dnode_phys_t *blk;
453 		int i;
454 		int blksz = BP_GET_LSIZE(bp);
455 		uint32_t aflags = ARC_WAIT;
456 		arc_buf_t *abuf;
457 
458 		if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
459 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
460 		    &aflags, zb) != 0)
461 			return (SET_ERROR(EIO));
462 
463 		blk = abuf->b_data;
464 		for (i = 0; i < blksz >> DNODE_SHIFT; i++) {
465 			uint64_t dnobj = (zb->zb_blkid <<
466 			    (DNODE_BLOCK_SHIFT - DNODE_SHIFT)) + i;
467 			err = dump_dnode(dsp, dnobj, blk+i);
468 			if (err != 0)
469 				break;
470 		}
471 		(void) arc_buf_remove_ref(abuf, &abuf);
472 	} else if (type == DMU_OT_SA) {
473 		uint32_t aflags = ARC_WAIT;
474 		arc_buf_t *abuf;
475 		int blksz = BP_GET_LSIZE(bp);
476 
477 		if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
478 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
479 		    &aflags, zb) != 0)
480 			return (SET_ERROR(EIO));
481 
482 		err = dump_spill(dsp, zb->zb_object, blksz, abuf->b_data);
483 		(void) arc_buf_remove_ref(abuf, &abuf);
484 	} else if (backup_do_embed(dsp, bp)) {
485 		/* it's an embedded level-0 block of a regular object */
486 		int blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
487 		err = dump_write_embedded(dsp, zb->zb_object,
488 		    zb->zb_blkid * blksz, blksz, bp);
489 	} else { /* it's a level-0 block of a regular object */
490 		uint32_t aflags = ARC_WAIT;
491 		arc_buf_t *abuf;
492 		int blksz = BP_GET_LSIZE(bp);
493 
494 		ASSERT3U(blksz, ==, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
495 		ASSERT0(zb->zb_level);
496 		if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
497 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
498 		    &aflags, zb) != 0) {
499 			if (zfs_send_corrupt_data) {
500 				/* Send a block filled with 0x"zfs badd bloc" */
501 				abuf = arc_buf_alloc(spa, blksz, &abuf,
502 				    ARC_BUFC_DATA);
503 				uint64_t *ptr;
504 				for (ptr = abuf->b_data;
505 				    (char *)ptr < (char *)abuf->b_data + blksz;
506 				    ptr++)
507 					*ptr = 0x2f5baddb10c;
508 			} else {
509 				return (SET_ERROR(EIO));
510 			}
511 		}
512 
513 		err = dump_write(dsp, type, zb->zb_object, zb->zb_blkid * blksz,
514 		    blksz, bp, abuf->b_data);
515 		(void) arc_buf_remove_ref(abuf, &abuf);
516 	}
517 
518 	ASSERT(err == 0 || err == EINTR);
519 	return (err);
520 }
521 
522 /*
523  * Releases dp using the specified tag.
524  */
525 static int
526 dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *ds,
527     zfs_bookmark_phys_t *fromzb, boolean_t is_clone, boolean_t embedok,
528     int outfd, vnode_t *vp, offset_t *off)
529 {
530 	objset_t *os;
531 	dmu_replay_record_t *drr;
532 	dmu_sendarg_t *dsp;
533 	int err;
534 	uint64_t fromtxg = 0;
535 	uint64_t featureflags = 0;
536 
537 	err = dmu_objset_from_ds(ds, &os);
538 	if (err != 0) {
539 		dsl_pool_rele(dp, tag);
540 		return (err);
541 	}
542 
543 	drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
544 	drr->drr_type = DRR_BEGIN;
545 	drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
546 	DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
547 	    DMU_SUBSTREAM);
548 
549 #ifdef _KERNEL
550 	if (dmu_objset_type(os) == DMU_OST_ZFS) {
551 		uint64_t version;
552 		if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0) {
553 			kmem_free(drr, sizeof (dmu_replay_record_t));
554 			dsl_pool_rele(dp, tag);
555 			return (SET_ERROR(EINVAL));
556 		}
557 		if (version >= ZPL_VERSION_SA) {
558 			featureflags |= DMU_BACKUP_FEATURE_SA_SPILL;
559 		}
560 	}
561 #endif
562 
563 	if (embedok &&
564 	    spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) {
565 		featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA;
566 		if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
567 			featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA_LZ4;
568 	} else {
569 		embedok = B_FALSE;
570 	}
571 
572 	DMU_SET_FEATUREFLAGS(drr->drr_u.drr_begin.drr_versioninfo,
573 	    featureflags);
574 
575 	drr->drr_u.drr_begin.drr_creation_time =
576 	    ds->ds_phys->ds_creation_time;
577 	drr->drr_u.drr_begin.drr_type = dmu_objset_type(os);
578 	if (is_clone)
579 		drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
580 	drr->drr_u.drr_begin.drr_toguid = ds->ds_phys->ds_guid;
581 	if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
582 		drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
583 
584 	if (fromzb != NULL) {
585 		drr->drr_u.drr_begin.drr_fromguid = fromzb->zbm_guid;
586 		fromtxg = fromzb->zbm_creation_txg;
587 	}
588 	dsl_dataset_name(ds, drr->drr_u.drr_begin.drr_toname);
589 	if (!dsl_dataset_is_snapshot(ds)) {
590 		(void) strlcat(drr->drr_u.drr_begin.drr_toname, "@--head--",
591 		    sizeof (drr->drr_u.drr_begin.drr_toname));
592 	}
593 
594 	dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP);
595 
596 	dsp->dsa_drr = drr;
597 	dsp->dsa_vp = vp;
598 	dsp->dsa_outfd = outfd;
599 	dsp->dsa_proc = curproc;
600 	dsp->dsa_os = os;
601 	dsp->dsa_off = off;
602 	dsp->dsa_toguid = ds->ds_phys->ds_guid;
603 	ZIO_SET_CHECKSUM(&dsp->dsa_zc, 0, 0, 0, 0);
604 	dsp->dsa_pending_op = PENDING_NONE;
605 	dsp->dsa_incremental = (fromzb != NULL);
606 	dsp->dsa_featureflags = featureflags;
607 
608 	mutex_enter(&ds->ds_sendstream_lock);
609 	list_insert_head(&ds->ds_sendstreams, dsp);
610 	mutex_exit(&ds->ds_sendstream_lock);
611 
612 	dsl_dataset_long_hold(ds, FTAG);
613 	dsl_pool_rele(dp, tag);
614 
615 	if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0) {
616 		err = dsp->dsa_err;
617 		goto out;
618 	}
619 
620 	err = traverse_dataset(ds, fromtxg, TRAVERSE_PRE | TRAVERSE_PREFETCH,
621 	    backup_cb, dsp);
622 
623 	if (dsp->dsa_pending_op != PENDING_NONE)
624 		if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0)
625 			err = SET_ERROR(EINTR);
626 
627 	if (err != 0) {
628 		if (err == EINTR && dsp->dsa_err != 0)
629 			err = dsp->dsa_err;
630 		goto out;
631 	}
632 
633 	bzero(drr, sizeof (dmu_replay_record_t));
634 	drr->drr_type = DRR_END;
635 	drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc;
636 	drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid;
637 
638 	if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0) {
639 		err = dsp->dsa_err;
640 		goto out;
641 	}
642 
643 out:
644 	mutex_enter(&ds->ds_sendstream_lock);
645 	list_remove(&ds->ds_sendstreams, dsp);
646 	mutex_exit(&ds->ds_sendstream_lock);
647 
648 	kmem_free(drr, sizeof (dmu_replay_record_t));
649 	kmem_free(dsp, sizeof (dmu_sendarg_t));
650 
651 	dsl_dataset_long_rele(ds, FTAG);
652 
653 	return (err);
654 }
655 
656 int
657 dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
658     boolean_t embedok, int outfd, vnode_t *vp, offset_t *off)
659 {
660 	dsl_pool_t *dp;
661 	dsl_dataset_t *ds;
662 	dsl_dataset_t *fromds = NULL;
663 	int err;
664 
665 	err = dsl_pool_hold(pool, FTAG, &dp);
666 	if (err != 0)
667 		return (err);
668 
669 	err = dsl_dataset_hold_obj(dp, tosnap, FTAG, &ds);
670 	if (err != 0) {
671 		dsl_pool_rele(dp, FTAG);
672 		return (err);
673 	}
674 
675 	if (fromsnap != 0) {
676 		zfs_bookmark_phys_t zb;
677 		boolean_t is_clone;
678 
679 		err = dsl_dataset_hold_obj(dp, fromsnap, FTAG, &fromds);
680 		if (err != 0) {
681 			dsl_dataset_rele(ds, FTAG);
682 			dsl_pool_rele(dp, FTAG);
683 			return (err);
684 		}
685 		if (!dsl_dataset_is_before(ds, fromds, 0))
686 			err = SET_ERROR(EXDEV);
687 		zb.zbm_creation_time = fromds->ds_phys->ds_creation_time;
688 		zb.zbm_creation_txg = fromds->ds_phys->ds_creation_txg;
689 		zb.zbm_guid = fromds->ds_phys->ds_guid;
690 		is_clone = (fromds->ds_dir != ds->ds_dir);
691 		dsl_dataset_rele(fromds, FTAG);
692 		err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone, embedok,
693 		    outfd, vp, off);
694 	} else {
695 		err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE, embedok,
696 		    outfd, vp, off);
697 	}
698 	dsl_dataset_rele(ds, FTAG);
699 	return (err);
700 }
701 
702 int
703 dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok,
704     int outfd, vnode_t *vp, offset_t *off)
705 {
706 	dsl_pool_t *dp;
707 	dsl_dataset_t *ds;
708 	int err;
709 	boolean_t owned = B_FALSE;
710 
711 	if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL)
712 		return (SET_ERROR(EINVAL));
713 
714 	err = dsl_pool_hold(tosnap, FTAG, &dp);
715 	if (err != 0)
716 		return (err);
717 
718 	if (strchr(tosnap, '@') == NULL && spa_writeable(dp->dp_spa)) {
719 		/*
720 		 * We are sending a filesystem or volume.  Ensure
721 		 * that it doesn't change by owning the dataset.
722 		 */
723 		err = dsl_dataset_own(dp, tosnap, FTAG, &ds);
724 		owned = B_TRUE;
725 	} else {
726 		err = dsl_dataset_hold(dp, tosnap, FTAG, &ds);
727 	}
728 	if (err != 0) {
729 		dsl_pool_rele(dp, FTAG);
730 		return (err);
731 	}
732 
733 	if (fromsnap != NULL) {
734 		zfs_bookmark_phys_t zb;
735 		boolean_t is_clone = B_FALSE;
736 		int fsnamelen = strchr(tosnap, '@') - tosnap;
737 
738 		/*
739 		 * If the fromsnap is in a different filesystem, then
740 		 * mark the send stream as a clone.
741 		 */
742 		if (strncmp(tosnap, fromsnap, fsnamelen) != 0 ||
743 		    (fromsnap[fsnamelen] != '@' &&
744 		    fromsnap[fsnamelen] != '#')) {
745 			is_clone = B_TRUE;
746 		}
747 
748 		if (strchr(fromsnap, '@')) {
749 			dsl_dataset_t *fromds;
750 			err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds);
751 			if (err == 0) {
752 				if (!dsl_dataset_is_before(ds, fromds, 0))
753 					err = SET_ERROR(EXDEV);
754 				zb.zbm_creation_time =
755 				    fromds->ds_phys->ds_creation_time;
756 				zb.zbm_creation_txg =
757 				    fromds->ds_phys->ds_creation_txg;
758 				zb.zbm_guid = fromds->ds_phys->ds_guid;
759 				is_clone = (ds->ds_dir != fromds->ds_dir);
760 				dsl_dataset_rele(fromds, FTAG);
761 			}
762 		} else {
763 			err = dsl_bookmark_lookup(dp, fromsnap, ds, &zb);
764 		}
765 		if (err != 0) {
766 			dsl_dataset_rele(ds, FTAG);
767 			dsl_pool_rele(dp, FTAG);
768 			return (err);
769 		}
770 		err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone, embedok,
771 		    outfd, vp, off);
772 	} else {
773 		err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE, embedok,
774 		    outfd, vp, off);
775 	}
776 	if (owned)
777 		dsl_dataset_disown(ds, FTAG);
778 	else
779 		dsl_dataset_rele(ds, FTAG);
780 	return (err);
781 }
782 
783 int
784 dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds, uint64_t *sizep)
785 {
786 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
787 	int err;
788 	uint64_t size;
789 
790 	ASSERT(dsl_pool_config_held(dp));
791 
792 	/* tosnap must be a snapshot */
793 	if (!dsl_dataset_is_snapshot(ds))
794 		return (SET_ERROR(EINVAL));
795 
796 	/*
797 	 * fromsnap must be an earlier snapshot from the same fs as tosnap,
798 	 * or the origin's fs.
799 	 */
800 	if (fromds != NULL && !dsl_dataset_is_before(ds, fromds, 0))
801 		return (SET_ERROR(EXDEV));
802 
803 	/* Get uncompressed size estimate of changed data. */
804 	if (fromds == NULL) {
805 		size = ds->ds_phys->ds_uncompressed_bytes;
806 	} else {
807 		uint64_t used, comp;
808 		err = dsl_dataset_space_written(fromds, ds,
809 		    &used, &comp, &size);
810 		if (err != 0)
811 			return (err);
812 	}
813 
814 	/*
815 	 * Assume that space (both on-disk and in-stream) is dominated by
816 	 * data.  We will adjust for indirect blocks and the copies property,
817 	 * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
818 	 */
819 
820 	/*
821 	 * Subtract out approximate space used by indirect blocks.
822 	 * Assume most space is used by data blocks (non-indirect, non-dnode).
823 	 * Assume all blocks are recordsize.  Assume ditto blocks and
824 	 * internal fragmentation counter out compression.
825 	 *
826 	 * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
827 	 * block, which we observe in practice.
828 	 */
829 	uint64_t recordsize;
830 	err = dsl_prop_get_int_ds(ds, "recordsize", &recordsize);
831 	if (err != 0)
832 		return (err);
833 	size -= size / recordsize * sizeof (blkptr_t);
834 
835 	/* Add in the space for the record associated with each block. */
836 	size += size / recordsize * sizeof (dmu_replay_record_t);
837 
838 	*sizep = size;
839 
840 	return (0);
841 }
842 
843 typedef struct dmu_recv_begin_arg {
844 	const char *drba_origin;
845 	dmu_recv_cookie_t *drba_cookie;
846 	cred_t *drba_cred;
847 	uint64_t drba_snapobj;
848 } dmu_recv_begin_arg_t;
849 
850 static int
851 recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds,
852     uint64_t fromguid)
853 {
854 	uint64_t val;
855 	int error;
856 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
857 
858 	/* temporary clone name must not exist */
859 	error = zap_lookup(dp->dp_meta_objset,
860 	    ds->ds_dir->dd_phys->dd_child_dir_zapobj, recv_clone_name,
861 	    8, 1, &val);
862 	if (error != ENOENT)
863 		return (error == 0 ? EBUSY : error);
864 
865 	/* new snapshot name must not exist */
866 	error = zap_lookup(dp->dp_meta_objset,
867 	    ds->ds_phys->ds_snapnames_zapobj, drba->drba_cookie->drc_tosnap,
868 	    8, 1, &val);
869 	if (error != ENOENT)
870 		return (error == 0 ? EEXIST : error);
871 
872 	/*
873 	 * Check snapshot limit before receiving. We'll recheck again at the
874 	 * end, but might as well abort before receiving if we're already over
875 	 * the limit.
876 	 *
877 	 * Note that we do not check the file system limit with
878 	 * dsl_dir_fscount_check because the temporary %clones don't count
879 	 * against that limit.
880 	 */
881 	error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT,
882 	    NULL, drba->drba_cred);
883 	if (error != 0)
884 		return (error);
885 
886 	if (fromguid != 0) {
887 		dsl_dataset_t *snap;
888 		uint64_t obj = ds->ds_phys->ds_prev_snap_obj;
889 
890 		/* Find snapshot in this dir that matches fromguid. */
891 		while (obj != 0) {
892 			error = dsl_dataset_hold_obj(dp, obj, FTAG,
893 			    &snap);
894 			if (error != 0)
895 				return (SET_ERROR(ENODEV));
896 			if (snap->ds_dir != ds->ds_dir) {
897 				dsl_dataset_rele(snap, FTAG);
898 				return (SET_ERROR(ENODEV));
899 			}
900 			if (snap->ds_phys->ds_guid == fromguid)
901 				break;
902 			obj = snap->ds_phys->ds_prev_snap_obj;
903 			dsl_dataset_rele(snap, FTAG);
904 		}
905 		if (obj == 0)
906 			return (SET_ERROR(ENODEV));
907 
908 		if (drba->drba_cookie->drc_force) {
909 			drba->drba_snapobj = obj;
910 		} else {
911 			/*
912 			 * If we are not forcing, there must be no
913 			 * changes since fromsnap.
914 			 */
915 			if (dsl_dataset_modified_since_snap(ds, snap)) {
916 				dsl_dataset_rele(snap, FTAG);
917 				return (SET_ERROR(ETXTBSY));
918 			}
919 			drba->drba_snapobj = ds->ds_prev->ds_object;
920 		}
921 
922 		dsl_dataset_rele(snap, FTAG);
923 	} else {
924 		/* if full, most recent snapshot must be $ORIGIN */
925 		if (ds->ds_phys->ds_prev_snap_txg >= TXG_INITIAL)
926 			return (SET_ERROR(ENODEV));
927 		drba->drba_snapobj = ds->ds_phys->ds_prev_snap_obj;
928 	}
929 
930 	return (0);
931 
932 }
933 
934 static int
935 dmu_recv_begin_check(void *arg, dmu_tx_t *tx)
936 {
937 	dmu_recv_begin_arg_t *drba = arg;
938 	dsl_pool_t *dp = dmu_tx_pool(tx);
939 	struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
940 	uint64_t fromguid = drrb->drr_fromguid;
941 	int flags = drrb->drr_flags;
942 	int error;
943 	uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
944 	dsl_dataset_t *ds;
945 	const char *tofs = drba->drba_cookie->drc_tofs;
946 
947 	/* already checked */
948 	ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
949 
950 	if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
951 	    DMU_COMPOUNDSTREAM ||
952 	    drrb->drr_type >= DMU_OST_NUMTYPES ||
953 	    ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL))
954 		return (SET_ERROR(EINVAL));
955 
956 	/* Verify pool version supports SA if SA_SPILL feature set */
957 	if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
958 	    spa_version(dp->dp_spa) < SPA_VERSION_SA)
959 		return (SET_ERROR(ENOTSUP));
960 
961 	/*
962 	 * The receiving code doesn't know how to translate a WRITE_EMBEDDED
963 	 * record to a plan WRITE record, so the pool must have the
964 	 * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
965 	 * records.  Same with WRITE_EMBEDDED records that use LZ4 compression.
966 	 */
967 	if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
968 	    !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
969 		return (SET_ERROR(ENOTSUP));
970 	if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4) &&
971 	    !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
972 		return (SET_ERROR(ENOTSUP));
973 
974 	error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
975 	if (error == 0) {
976 		/* target fs already exists; recv into temp clone */
977 
978 		/* Can't recv a clone into an existing fs */
979 		if (flags & DRR_FLAG_CLONE) {
980 			dsl_dataset_rele(ds, FTAG);
981 			return (SET_ERROR(EINVAL));
982 		}
983 
984 		error = recv_begin_check_existing_impl(drba, ds, fromguid);
985 		dsl_dataset_rele(ds, FTAG);
986 	} else if (error == ENOENT) {
987 		/* target fs does not exist; must be a full backup or clone */
988 		char buf[MAXNAMELEN];
989 
990 		/*
991 		 * If it's a non-clone incremental, we are missing the
992 		 * target fs, so fail the recv.
993 		 */
994 		if (fromguid != 0 && !(flags & DRR_FLAG_CLONE))
995 			return (SET_ERROR(ENOENT));
996 
997 		/* Open the parent of tofs */
998 		ASSERT3U(strlen(tofs), <, MAXNAMELEN);
999 		(void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1);
1000 		error = dsl_dataset_hold(dp, buf, FTAG, &ds);
1001 		if (error != 0)
1002 			return (error);
1003 
1004 		/*
1005 		 * Check filesystem and snapshot limits before receiving. We'll
1006 		 * recheck snapshot limits again at the end (we create the
1007 		 * filesystems and increment those counts during begin_sync).
1008 		 */
1009 		error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1010 		    ZFS_PROP_FILESYSTEM_LIMIT, NULL, drba->drba_cred);
1011 		if (error != 0) {
1012 			dsl_dataset_rele(ds, FTAG);
1013 			return (error);
1014 		}
1015 
1016 		error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1017 		    ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred);
1018 		if (error != 0) {
1019 			dsl_dataset_rele(ds, FTAG);
1020 			return (error);
1021 		}
1022 
1023 		if (drba->drba_origin != NULL) {
1024 			dsl_dataset_t *origin;
1025 			error = dsl_dataset_hold(dp, drba->drba_origin,
1026 			    FTAG, &origin);
1027 			if (error != 0) {
1028 				dsl_dataset_rele(ds, FTAG);
1029 				return (error);
1030 			}
1031 			if (!dsl_dataset_is_snapshot(origin)) {
1032 				dsl_dataset_rele(origin, FTAG);
1033 				dsl_dataset_rele(ds, FTAG);
1034 				return (SET_ERROR(EINVAL));
1035 			}
1036 			if (origin->ds_phys->ds_guid != fromguid) {
1037 				dsl_dataset_rele(origin, FTAG);
1038 				dsl_dataset_rele(ds, FTAG);
1039 				return (SET_ERROR(ENODEV));
1040 			}
1041 			dsl_dataset_rele(origin, FTAG);
1042 		}
1043 		dsl_dataset_rele(ds, FTAG);
1044 		error = 0;
1045 	}
1046 	return (error);
1047 }
1048 
1049 static void
1050 dmu_recv_begin_sync(void *arg, dmu_tx_t *tx)
1051 {
1052 	dmu_recv_begin_arg_t *drba = arg;
1053 	dsl_pool_t *dp = dmu_tx_pool(tx);
1054 	struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1055 	const char *tofs = drba->drba_cookie->drc_tofs;
1056 	dsl_dataset_t *ds, *newds;
1057 	uint64_t dsobj;
1058 	int error;
1059 	uint64_t crflags;
1060 
1061 	crflags = (drrb->drr_flags & DRR_FLAG_CI_DATA) ?
1062 	    DS_FLAG_CI_DATASET : 0;
1063 
1064 	error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1065 	if (error == 0) {
1066 		/* create temporary clone */
1067 		dsl_dataset_t *snap = NULL;
1068 		if (drba->drba_snapobj != 0) {
1069 			VERIFY0(dsl_dataset_hold_obj(dp,
1070 			    drba->drba_snapobj, FTAG, &snap));
1071 		}
1072 		dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name,
1073 		    snap, crflags, drba->drba_cred, tx);
1074 		dsl_dataset_rele(snap, FTAG);
1075 		dsl_dataset_rele(ds, FTAG);
1076 	} else {
1077 		dsl_dir_t *dd;
1078 		const char *tail;
1079 		dsl_dataset_t *origin = NULL;
1080 
1081 		VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail));
1082 
1083 		if (drba->drba_origin != NULL) {
1084 			VERIFY0(dsl_dataset_hold(dp, drba->drba_origin,
1085 			    FTAG, &origin));
1086 		}
1087 
1088 		/* Create new dataset. */
1089 		dsobj = dsl_dataset_create_sync(dd,
1090 		    strrchr(tofs, '/') + 1,
1091 		    origin, crflags, drba->drba_cred, tx);
1092 		if (origin != NULL)
1093 			dsl_dataset_rele(origin, FTAG);
1094 		dsl_dir_rele(dd, FTAG);
1095 		drba->drba_cookie->drc_newfs = B_TRUE;
1096 	}
1097 	VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds));
1098 
1099 	dmu_buf_will_dirty(newds->ds_dbuf, tx);
1100 	newds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
1101 
1102 	/*
1103 	 * If we actually created a non-clone, we need to create the
1104 	 * objset in our new dataset.
1105 	 */
1106 	if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) {
1107 		(void) dmu_objset_create_impl(dp->dp_spa,
1108 		    newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx);
1109 	}
1110 
1111 	drba->drba_cookie->drc_ds = newds;
1112 
1113 	spa_history_log_internal_ds(newds, "receive", tx, "");
1114 }
1115 
1116 /*
1117  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
1118  * succeeds; otherwise we will leak the holds on the datasets.
1119  */
1120 int
1121 dmu_recv_begin(char *tofs, char *tosnap, struct drr_begin *drrb,
1122     boolean_t force, char *origin, dmu_recv_cookie_t *drc)
1123 {
1124 	dmu_recv_begin_arg_t drba = { 0 };
1125 	dmu_replay_record_t *drr;
1126 
1127 	bzero(drc, sizeof (dmu_recv_cookie_t));
1128 	drc->drc_drrb = drrb;
1129 	drc->drc_tosnap = tosnap;
1130 	drc->drc_tofs = tofs;
1131 	drc->drc_force = force;
1132 	drc->drc_cred = CRED();
1133 
1134 	if (drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
1135 		drc->drc_byteswap = B_TRUE;
1136 	else if (drrb->drr_magic != DMU_BACKUP_MAGIC)
1137 		return (SET_ERROR(EINVAL));
1138 
1139 	drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
1140 	drr->drr_type = DRR_BEGIN;
1141 	drr->drr_u.drr_begin = *drc->drc_drrb;
1142 	if (drc->drc_byteswap) {
1143 		fletcher_4_incremental_byteswap(drr,
1144 		    sizeof (dmu_replay_record_t), &drc->drc_cksum);
1145 	} else {
1146 		fletcher_4_incremental_native(drr,
1147 		    sizeof (dmu_replay_record_t), &drc->drc_cksum);
1148 	}
1149 	kmem_free(drr, sizeof (dmu_replay_record_t));
1150 
1151 	if (drc->drc_byteswap) {
1152 		drrb->drr_magic = BSWAP_64(drrb->drr_magic);
1153 		drrb->drr_versioninfo = BSWAP_64(drrb->drr_versioninfo);
1154 		drrb->drr_creation_time = BSWAP_64(drrb->drr_creation_time);
1155 		drrb->drr_type = BSWAP_32(drrb->drr_type);
1156 		drrb->drr_toguid = BSWAP_64(drrb->drr_toguid);
1157 		drrb->drr_fromguid = BSWAP_64(drrb->drr_fromguid);
1158 	}
1159 
1160 	drba.drba_origin = origin;
1161 	drba.drba_cookie = drc;
1162 	drba.drba_cred = CRED();
1163 
1164 	return (dsl_sync_task(tofs, dmu_recv_begin_check, dmu_recv_begin_sync,
1165 	    &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1166 }
1167 
1168 struct restorearg {
1169 	int err;
1170 	boolean_t byteswap;
1171 	vnode_t *vp;
1172 	char *buf;
1173 	uint64_t voff;
1174 	int bufsize; /* amount of memory allocated for buf */
1175 	zio_cksum_t cksum;
1176 	avl_tree_t *guid_to_ds_map;
1177 };
1178 
1179 typedef struct guid_map_entry {
1180 	uint64_t	guid;
1181 	dsl_dataset_t	*gme_ds;
1182 	avl_node_t	avlnode;
1183 } guid_map_entry_t;
1184 
1185 static int
1186 guid_compare(const void *arg1, const void *arg2)
1187 {
1188 	const guid_map_entry_t *gmep1 = arg1;
1189 	const guid_map_entry_t *gmep2 = arg2;
1190 
1191 	if (gmep1->guid < gmep2->guid)
1192 		return (-1);
1193 	else if (gmep1->guid > gmep2->guid)
1194 		return (1);
1195 	return (0);
1196 }
1197 
1198 static void
1199 free_guid_map_onexit(void *arg)
1200 {
1201 	avl_tree_t *ca = arg;
1202 	void *cookie = NULL;
1203 	guid_map_entry_t *gmep;
1204 
1205 	while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
1206 		dsl_dataset_long_rele(gmep->gme_ds, gmep);
1207 		dsl_dataset_rele(gmep->gme_ds, gmep);
1208 		kmem_free(gmep, sizeof (guid_map_entry_t));
1209 	}
1210 	avl_destroy(ca);
1211 	kmem_free(ca, sizeof (avl_tree_t));
1212 }
1213 
1214 static void *
1215 restore_read(struct restorearg *ra, int len)
1216 {
1217 	void *rv;
1218 	int done = 0;
1219 
1220 	/* some things will require 8-byte alignment, so everything must */
1221 	ASSERT0(len % 8);
1222 
1223 	while (done < len) {
1224 		ssize_t resid;
1225 
1226 		ra->err = vn_rdwr(UIO_READ, ra->vp,
1227 		    (caddr_t)ra->buf + done, len - done,
1228 		    ra->voff, UIO_SYSSPACE, FAPPEND,
1229 		    RLIM64_INFINITY, CRED(), &resid);
1230 
1231 		if (resid == len - done)
1232 			ra->err = SET_ERROR(EINVAL);
1233 		ra->voff += len - done - resid;
1234 		done = len - resid;
1235 		if (ra->err != 0)
1236 			return (NULL);
1237 	}
1238 
1239 	ASSERT3U(done, ==, len);
1240 	rv = ra->buf;
1241 	if (ra->byteswap)
1242 		fletcher_4_incremental_byteswap(rv, len, &ra->cksum);
1243 	else
1244 		fletcher_4_incremental_native(rv, len, &ra->cksum);
1245 	return (rv);
1246 }
1247 
1248 static void
1249 backup_byteswap(dmu_replay_record_t *drr)
1250 {
1251 #define	DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
1252 #define	DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
1253 	drr->drr_type = BSWAP_32(drr->drr_type);
1254 	drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
1255 	switch (drr->drr_type) {
1256 	case DRR_BEGIN:
1257 		DO64(drr_begin.drr_magic);
1258 		DO64(drr_begin.drr_versioninfo);
1259 		DO64(drr_begin.drr_creation_time);
1260 		DO32(drr_begin.drr_type);
1261 		DO32(drr_begin.drr_flags);
1262 		DO64(drr_begin.drr_toguid);
1263 		DO64(drr_begin.drr_fromguid);
1264 		break;
1265 	case DRR_OBJECT:
1266 		DO64(drr_object.drr_object);
1267 		DO32(drr_object.drr_type);
1268 		DO32(drr_object.drr_bonustype);
1269 		DO32(drr_object.drr_blksz);
1270 		DO32(drr_object.drr_bonuslen);
1271 		DO64(drr_object.drr_toguid);
1272 		break;
1273 	case DRR_FREEOBJECTS:
1274 		DO64(drr_freeobjects.drr_firstobj);
1275 		DO64(drr_freeobjects.drr_numobjs);
1276 		DO64(drr_freeobjects.drr_toguid);
1277 		break;
1278 	case DRR_WRITE:
1279 		DO64(drr_write.drr_object);
1280 		DO32(drr_write.drr_type);
1281 		DO64(drr_write.drr_offset);
1282 		DO64(drr_write.drr_length);
1283 		DO64(drr_write.drr_toguid);
1284 		DO64(drr_write.drr_key.ddk_cksum.zc_word[0]);
1285 		DO64(drr_write.drr_key.ddk_cksum.zc_word[1]);
1286 		DO64(drr_write.drr_key.ddk_cksum.zc_word[2]);
1287 		DO64(drr_write.drr_key.ddk_cksum.zc_word[3]);
1288 		DO64(drr_write.drr_key.ddk_prop);
1289 		break;
1290 	case DRR_WRITE_BYREF:
1291 		DO64(drr_write_byref.drr_object);
1292 		DO64(drr_write_byref.drr_offset);
1293 		DO64(drr_write_byref.drr_length);
1294 		DO64(drr_write_byref.drr_toguid);
1295 		DO64(drr_write_byref.drr_refguid);
1296 		DO64(drr_write_byref.drr_refobject);
1297 		DO64(drr_write_byref.drr_refoffset);
1298 		DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[0]);
1299 		DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[1]);
1300 		DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[2]);
1301 		DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[3]);
1302 		DO64(drr_write_byref.drr_key.ddk_prop);
1303 		break;
1304 	case DRR_WRITE_EMBEDDED:
1305 		DO64(drr_write_embedded.drr_object);
1306 		DO64(drr_write_embedded.drr_offset);
1307 		DO64(drr_write_embedded.drr_length);
1308 		DO64(drr_write_embedded.drr_toguid);
1309 		DO32(drr_write_embedded.drr_lsize);
1310 		DO32(drr_write_embedded.drr_psize);
1311 		break;
1312 	case DRR_FREE:
1313 		DO64(drr_free.drr_object);
1314 		DO64(drr_free.drr_offset);
1315 		DO64(drr_free.drr_length);
1316 		DO64(drr_free.drr_toguid);
1317 		break;
1318 	case DRR_SPILL:
1319 		DO64(drr_spill.drr_object);
1320 		DO64(drr_spill.drr_length);
1321 		DO64(drr_spill.drr_toguid);
1322 		break;
1323 	case DRR_END:
1324 		DO64(drr_end.drr_checksum.zc_word[0]);
1325 		DO64(drr_end.drr_checksum.zc_word[1]);
1326 		DO64(drr_end.drr_checksum.zc_word[2]);
1327 		DO64(drr_end.drr_checksum.zc_word[3]);
1328 		DO64(drr_end.drr_toguid);
1329 		break;
1330 	}
1331 #undef DO64
1332 #undef DO32
1333 }
1334 
1335 static int
1336 restore_object(struct restorearg *ra, objset_t *os, struct drr_object *drro)
1337 {
1338 	int err;
1339 	dmu_tx_t *tx;
1340 	void *data = NULL;
1341 
1342 	if (drro->drr_type == DMU_OT_NONE ||
1343 	    !DMU_OT_IS_VALID(drro->drr_type) ||
1344 	    !DMU_OT_IS_VALID(drro->drr_bonustype) ||
1345 	    drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
1346 	    drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
1347 	    P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
1348 	    drro->drr_blksz < SPA_MINBLOCKSIZE ||
1349 	    drro->drr_blksz > SPA_MAXBLOCKSIZE ||
1350 	    drro->drr_bonuslen > DN_MAX_BONUSLEN) {
1351 		return (SET_ERROR(EINVAL));
1352 	}
1353 
1354 	err = dmu_object_info(os, drro->drr_object, NULL);
1355 
1356 	if (err != 0 && err != ENOENT)
1357 		return (SET_ERROR(EINVAL));
1358 
1359 	if (drro->drr_bonuslen) {
1360 		data = restore_read(ra, P2ROUNDUP(drro->drr_bonuslen, 8));
1361 		if (ra->err != 0)
1362 			return (ra->err);
1363 	}
1364 
1365 	if (err == ENOENT) {
1366 		/* currently free, want to be allocated */
1367 		tx = dmu_tx_create(os);
1368 		dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1369 		err = dmu_tx_assign(tx, TXG_WAIT);
1370 		if (err != 0) {
1371 			dmu_tx_abort(tx);
1372 			return (err);
1373 		}
1374 		err = dmu_object_claim(os, drro->drr_object,
1375 		    drro->drr_type, drro->drr_blksz,
1376 		    drro->drr_bonustype, drro->drr_bonuslen, tx);
1377 		dmu_tx_commit(tx);
1378 	} else {
1379 		/* currently allocated, want to be allocated */
1380 		err = dmu_object_reclaim(os, drro->drr_object,
1381 		    drro->drr_type, drro->drr_blksz,
1382 		    drro->drr_bonustype, drro->drr_bonuslen);
1383 	}
1384 	if (err != 0) {
1385 		return (SET_ERROR(EINVAL));
1386 	}
1387 
1388 	tx = dmu_tx_create(os);
1389 	dmu_tx_hold_bonus(tx, drro->drr_object);
1390 	err = dmu_tx_assign(tx, TXG_WAIT);
1391 	if (err != 0) {
1392 		dmu_tx_abort(tx);
1393 		return (err);
1394 	}
1395 
1396 	dmu_object_set_checksum(os, drro->drr_object, drro->drr_checksumtype,
1397 	    tx);
1398 	dmu_object_set_compress(os, drro->drr_object, drro->drr_compress, tx);
1399 
1400 	if (data != NULL) {
1401 		dmu_buf_t *db;
1402 
1403 		VERIFY(0 == dmu_bonus_hold(os, drro->drr_object, FTAG, &db));
1404 		dmu_buf_will_dirty(db, tx);
1405 
1406 		ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
1407 		bcopy(data, db->db_data, drro->drr_bonuslen);
1408 		if (ra->byteswap) {
1409 			dmu_object_byteswap_t byteswap =
1410 			    DMU_OT_BYTESWAP(drro->drr_bonustype);
1411 			dmu_ot_byteswap[byteswap].ob_func(db->db_data,
1412 			    drro->drr_bonuslen);
1413 		}
1414 		dmu_buf_rele(db, FTAG);
1415 	}
1416 	dmu_tx_commit(tx);
1417 	return (0);
1418 }
1419 
1420 /* ARGSUSED */
1421 static int
1422 restore_freeobjects(struct restorearg *ra, objset_t *os,
1423     struct drr_freeobjects *drrfo)
1424 {
1425 	uint64_t obj;
1426 
1427 	if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
1428 		return (SET_ERROR(EINVAL));
1429 
1430 	for (obj = drrfo->drr_firstobj;
1431 	    obj < drrfo->drr_firstobj + drrfo->drr_numobjs;
1432 	    (void) dmu_object_next(os, &obj, FALSE, 0)) {
1433 		int err;
1434 
1435 		if (dmu_object_info(os, obj, NULL) != 0)
1436 			continue;
1437 
1438 		err = dmu_free_long_object(os, obj);
1439 		if (err != 0)
1440 			return (err);
1441 	}
1442 	return (0);
1443 }
1444 
1445 static int
1446 restore_write(struct restorearg *ra, objset_t *os,
1447     struct drr_write *drrw)
1448 {
1449 	dmu_tx_t *tx;
1450 	void *data;
1451 	int err;
1452 
1453 	if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
1454 	    !DMU_OT_IS_VALID(drrw->drr_type))
1455 		return (SET_ERROR(EINVAL));
1456 
1457 	data = restore_read(ra, drrw->drr_length);
1458 	if (data == NULL)
1459 		return (ra->err);
1460 
1461 	if (dmu_object_info(os, drrw->drr_object, NULL) != 0)
1462 		return (SET_ERROR(EINVAL));
1463 
1464 	tx = dmu_tx_create(os);
1465 
1466 	dmu_tx_hold_write(tx, drrw->drr_object,
1467 	    drrw->drr_offset, drrw->drr_length);
1468 	err = dmu_tx_assign(tx, TXG_WAIT);
1469 	if (err != 0) {
1470 		dmu_tx_abort(tx);
1471 		return (err);
1472 	}
1473 	if (ra->byteswap) {
1474 		dmu_object_byteswap_t byteswap =
1475 		    DMU_OT_BYTESWAP(drrw->drr_type);
1476 		dmu_ot_byteswap[byteswap].ob_func(data, drrw->drr_length);
1477 	}
1478 	dmu_write(os, drrw->drr_object,
1479 	    drrw->drr_offset, drrw->drr_length, data, tx);
1480 	dmu_tx_commit(tx);
1481 	return (0);
1482 }
1483 
1484 /*
1485  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
1486  * streams to refer to a copy of the data that is already on the
1487  * system because it came in earlier in the stream.  This function
1488  * finds the earlier copy of the data, and uses that copy instead of
1489  * data from the stream to fulfill this write.
1490  */
1491 static int
1492 restore_write_byref(struct restorearg *ra, objset_t *os,
1493     struct drr_write_byref *drrwbr)
1494 {
1495 	dmu_tx_t *tx;
1496 	int err;
1497 	guid_map_entry_t gmesrch;
1498 	guid_map_entry_t *gmep;
1499 	avl_index_t where;
1500 	objset_t *ref_os = NULL;
1501 	dmu_buf_t *dbp;
1502 
1503 	if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
1504 		return (SET_ERROR(EINVAL));
1505 
1506 	/*
1507 	 * If the GUID of the referenced dataset is different from the
1508 	 * GUID of the target dataset, find the referenced dataset.
1509 	 */
1510 	if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
1511 		gmesrch.guid = drrwbr->drr_refguid;
1512 		if ((gmep = avl_find(ra->guid_to_ds_map, &gmesrch,
1513 		    &where)) == NULL) {
1514 			return (SET_ERROR(EINVAL));
1515 		}
1516 		if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
1517 			return (SET_ERROR(EINVAL));
1518 	} else {
1519 		ref_os = os;
1520 	}
1521 
1522 	err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
1523 	    drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH);
1524 	if (err != 0)
1525 		return (err);
1526 
1527 	tx = dmu_tx_create(os);
1528 
1529 	dmu_tx_hold_write(tx, drrwbr->drr_object,
1530 	    drrwbr->drr_offset, drrwbr->drr_length);
1531 	err = dmu_tx_assign(tx, TXG_WAIT);
1532 	if (err != 0) {
1533 		dmu_tx_abort(tx);
1534 		return (err);
1535 	}
1536 	dmu_write(os, drrwbr->drr_object,
1537 	    drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
1538 	dmu_buf_rele(dbp, FTAG);
1539 	dmu_tx_commit(tx);
1540 	return (0);
1541 }
1542 
1543 static int
1544 restore_write_embedded(struct restorearg *ra, objset_t *os,
1545     struct drr_write_embedded *drrwnp)
1546 {
1547 	dmu_tx_t *tx;
1548 	int err;
1549 	void *data;
1550 
1551 	if (drrwnp->drr_offset + drrwnp->drr_length < drrwnp->drr_offset)
1552 		return (EINVAL);
1553 
1554 	if (drrwnp->drr_psize > BPE_PAYLOAD_SIZE)
1555 		return (EINVAL);
1556 
1557 	if (drrwnp->drr_etype >= NUM_BP_EMBEDDED_TYPES)
1558 		return (EINVAL);
1559 	if (drrwnp->drr_compression >= ZIO_COMPRESS_FUNCTIONS)
1560 		return (EINVAL);
1561 
1562 	data = restore_read(ra, P2ROUNDUP(drrwnp->drr_psize, 8));
1563 	if (data == NULL)
1564 		return (ra->err);
1565 
1566 	tx = dmu_tx_create(os);
1567 
1568 	dmu_tx_hold_write(tx, drrwnp->drr_object,
1569 	    drrwnp->drr_offset, drrwnp->drr_length);
1570 	err = dmu_tx_assign(tx, TXG_WAIT);
1571 	if (err != 0) {
1572 		dmu_tx_abort(tx);
1573 		return (err);
1574 	}
1575 
1576 	dmu_write_embedded(os, drrwnp->drr_object,
1577 	    drrwnp->drr_offset, data, drrwnp->drr_etype,
1578 	    drrwnp->drr_compression, drrwnp->drr_lsize, drrwnp->drr_psize,
1579 	    ra->byteswap ^ ZFS_HOST_BYTEORDER, tx);
1580 
1581 	dmu_tx_commit(tx);
1582 	return (0);
1583 }
1584 
1585 static int
1586 restore_spill(struct restorearg *ra, objset_t *os, struct drr_spill *drrs)
1587 {
1588 	dmu_tx_t *tx;
1589 	void *data;
1590 	dmu_buf_t *db, *db_spill;
1591 	int err;
1592 
1593 	if (drrs->drr_length < SPA_MINBLOCKSIZE ||
1594 	    drrs->drr_length > SPA_MAXBLOCKSIZE)
1595 		return (SET_ERROR(EINVAL));
1596 
1597 	data = restore_read(ra, drrs->drr_length);
1598 	if (data == NULL)
1599 		return (ra->err);
1600 
1601 	if (dmu_object_info(os, drrs->drr_object, NULL) != 0)
1602 		return (SET_ERROR(EINVAL));
1603 
1604 	VERIFY(0 == dmu_bonus_hold(os, drrs->drr_object, FTAG, &db));
1605 	if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
1606 		dmu_buf_rele(db, FTAG);
1607 		return (err);
1608 	}
1609 
1610 	tx = dmu_tx_create(os);
1611 
1612 	dmu_tx_hold_spill(tx, db->db_object);
1613 
1614 	err = dmu_tx_assign(tx, TXG_WAIT);
1615 	if (err != 0) {
1616 		dmu_buf_rele(db, FTAG);
1617 		dmu_buf_rele(db_spill, FTAG);
1618 		dmu_tx_abort(tx);
1619 		return (err);
1620 	}
1621 	dmu_buf_will_dirty(db_spill, tx);
1622 
1623 	if (db_spill->db_size < drrs->drr_length)
1624 		VERIFY(0 == dbuf_spill_set_blksz(db_spill,
1625 		    drrs->drr_length, tx));
1626 	bcopy(data, db_spill->db_data, drrs->drr_length);
1627 
1628 	dmu_buf_rele(db, FTAG);
1629 	dmu_buf_rele(db_spill, FTAG);
1630 
1631 	dmu_tx_commit(tx);
1632 	return (0);
1633 }
1634 
1635 /* ARGSUSED */
1636 static int
1637 restore_free(struct restorearg *ra, objset_t *os,
1638     struct drr_free *drrf)
1639 {
1640 	int err;
1641 	uint64_t length = drrf->drr_length;
1642 	uint64_t offset = drrf->drr_offset;
1643 
1644 	if (length != -1ULL && offset + length < offset)
1645 		length = -1ULL;
1646 
1647 	if (dmu_object_info(os, drrf->drr_object, NULL) != 0)
1648 		return (SET_ERROR(EINVAL));
1649 
1650 	err = dmu_free_long_range(os, drrf->drr_object, offset, length);
1651 
1652 	return (err);
1653 }
1654 
1655 /* used to destroy the drc_ds on error */
1656 static void
1657 dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc)
1658 {
1659 	char name[MAXNAMELEN];
1660 	dsl_dataset_name(drc->drc_ds, name);
1661 	dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
1662 	(void) dsl_destroy_head(name);
1663 }
1664 
1665 /*
1666  * NB: callers *must* call dmu_recv_end() if this succeeds.
1667  */
1668 int
1669 dmu_recv_stream(dmu_recv_cookie_t *drc, vnode_t *vp, offset_t *voffp,
1670     int cleanup_fd, uint64_t *action_handlep)
1671 {
1672 	struct restorearg ra = { 0 };
1673 	dmu_replay_record_t *drr;
1674 	objset_t *os;
1675 	zio_cksum_t pcksum;
1676 	int featureflags;
1677 
1678 	ra.byteswap = drc->drc_byteswap;
1679 	ra.cksum = drc->drc_cksum;
1680 	ra.vp = vp;
1681 	ra.voff = *voffp;
1682 	ra.bufsize = 1<<20;
1683 	ra.buf = kmem_alloc(ra.bufsize, KM_SLEEP);
1684 
1685 	/* these were verified in dmu_recv_begin */
1686 	ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==,
1687 	    DMU_SUBSTREAM);
1688 	ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES);
1689 
1690 	/*
1691 	 * Open the objset we are modifying.
1692 	 */
1693 	VERIFY0(dmu_objset_from_ds(drc->drc_ds, &os));
1694 
1695 	ASSERT(drc->drc_ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT);
1696 
1697 	featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
1698 
1699 	/* if this stream is dedup'ed, set up the avl tree for guid mapping */
1700 	if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
1701 		minor_t minor;
1702 
1703 		if (cleanup_fd == -1) {
1704 			ra.err = SET_ERROR(EBADF);
1705 			goto out;
1706 		}
1707 		ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
1708 		if (ra.err != 0) {
1709 			cleanup_fd = -1;
1710 			goto out;
1711 		}
1712 
1713 		if (*action_handlep == 0) {
1714 			ra.guid_to_ds_map =
1715 			    kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
1716 			avl_create(ra.guid_to_ds_map, guid_compare,
1717 			    sizeof (guid_map_entry_t),
1718 			    offsetof(guid_map_entry_t, avlnode));
1719 			ra.err = zfs_onexit_add_cb(minor,
1720 			    free_guid_map_onexit, ra.guid_to_ds_map,
1721 			    action_handlep);
1722 			if (ra.err != 0)
1723 				goto out;
1724 		} else {
1725 			ra.err = zfs_onexit_cb_data(minor, *action_handlep,
1726 			    (void **)&ra.guid_to_ds_map);
1727 			if (ra.err != 0)
1728 				goto out;
1729 		}
1730 
1731 		drc->drc_guid_to_ds_map = ra.guid_to_ds_map;
1732 	}
1733 
1734 	/*
1735 	 * Read records and process them.
1736 	 */
1737 	pcksum = ra.cksum;
1738 	while (ra.err == 0 &&
1739 	    NULL != (drr = restore_read(&ra, sizeof (*drr)))) {
1740 		if (issig(JUSTLOOKING) && issig(FORREAL)) {
1741 			ra.err = SET_ERROR(EINTR);
1742 			goto out;
1743 		}
1744 
1745 		if (ra.byteswap)
1746 			backup_byteswap(drr);
1747 
1748 		switch (drr->drr_type) {
1749 		case DRR_OBJECT:
1750 		{
1751 			/*
1752 			 * We need to make a copy of the record header,
1753 			 * because restore_{object,write} may need to
1754 			 * restore_read(), which will invalidate drr.
1755 			 */
1756 			struct drr_object drro = drr->drr_u.drr_object;
1757 			ra.err = restore_object(&ra, os, &drro);
1758 			break;
1759 		}
1760 		case DRR_FREEOBJECTS:
1761 		{
1762 			struct drr_freeobjects drrfo =
1763 			    drr->drr_u.drr_freeobjects;
1764 			ra.err = restore_freeobjects(&ra, os, &drrfo);
1765 			break;
1766 		}
1767 		case DRR_WRITE:
1768 		{
1769 			struct drr_write drrw = drr->drr_u.drr_write;
1770 			ra.err = restore_write(&ra, os, &drrw);
1771 			break;
1772 		}
1773 		case DRR_WRITE_BYREF:
1774 		{
1775 			struct drr_write_byref drrwbr =
1776 			    drr->drr_u.drr_write_byref;
1777 			ra.err = restore_write_byref(&ra, os, &drrwbr);
1778 			break;
1779 		}
1780 		case DRR_WRITE_EMBEDDED:
1781 		{
1782 			struct drr_write_embedded drrwe =
1783 			    drr->drr_u.drr_write_embedded;
1784 			ra.err = restore_write_embedded(&ra, os, &drrwe);
1785 			break;
1786 		}
1787 		case DRR_FREE:
1788 		{
1789 			struct drr_free drrf = drr->drr_u.drr_free;
1790 			ra.err = restore_free(&ra, os, &drrf);
1791 			break;
1792 		}
1793 		case DRR_END:
1794 		{
1795 			struct drr_end drre = drr->drr_u.drr_end;
1796 			/*
1797 			 * We compare against the *previous* checksum
1798 			 * value, because the stored checksum is of
1799 			 * everything before the DRR_END record.
1800 			 */
1801 			if (!ZIO_CHECKSUM_EQUAL(drre.drr_checksum, pcksum))
1802 				ra.err = SET_ERROR(ECKSUM);
1803 			goto out;
1804 		}
1805 		case DRR_SPILL:
1806 		{
1807 			struct drr_spill drrs = drr->drr_u.drr_spill;
1808 			ra.err = restore_spill(&ra, os, &drrs);
1809 			break;
1810 		}
1811 		default:
1812 			ra.err = SET_ERROR(EINVAL);
1813 			goto out;
1814 		}
1815 		pcksum = ra.cksum;
1816 	}
1817 	ASSERT(ra.err != 0);
1818 
1819 out:
1820 	if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
1821 		zfs_onexit_fd_rele(cleanup_fd);
1822 
1823 	if (ra.err != 0) {
1824 		/*
1825 		 * destroy what we created, so we don't leave it in the
1826 		 * inconsistent restoring state.
1827 		 */
1828 		dmu_recv_cleanup_ds(drc);
1829 	}
1830 
1831 	kmem_free(ra.buf, ra.bufsize);
1832 	*voffp = ra.voff;
1833 	return (ra.err);
1834 }
1835 
1836 static int
1837 dmu_recv_end_check(void *arg, dmu_tx_t *tx)
1838 {
1839 	dmu_recv_cookie_t *drc = arg;
1840 	dsl_pool_t *dp = dmu_tx_pool(tx);
1841 	int error;
1842 
1843 	ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag);
1844 
1845 	if (!drc->drc_newfs) {
1846 		dsl_dataset_t *origin_head;
1847 
1848 		error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head);
1849 		if (error != 0)
1850 			return (error);
1851 		if (drc->drc_force) {
1852 			/*
1853 			 * We will destroy any snapshots in tofs (i.e. before
1854 			 * origin_head) that are after the origin (which is
1855 			 * the snap before drc_ds, because drc_ds can not
1856 			 * have any snaps of its own).
1857 			 */
1858 			uint64_t obj = origin_head->ds_phys->ds_prev_snap_obj;
1859 			while (obj != drc->drc_ds->ds_phys->ds_prev_snap_obj) {
1860 				dsl_dataset_t *snap;
1861 				error = dsl_dataset_hold_obj(dp, obj, FTAG,
1862 				    &snap);
1863 				if (error != 0)
1864 					return (error);
1865 				if (snap->ds_dir != origin_head->ds_dir)
1866 					error = SET_ERROR(EINVAL);
1867 				if (error == 0)  {
1868 					error = dsl_destroy_snapshot_check_impl(
1869 					    snap, B_FALSE);
1870 				}
1871 				obj = snap->ds_phys->ds_prev_snap_obj;
1872 				dsl_dataset_rele(snap, FTAG);
1873 				if (error != 0)
1874 					return (error);
1875 			}
1876 		}
1877 		error = dsl_dataset_clone_swap_check_impl(drc->drc_ds,
1878 		    origin_head, drc->drc_force, drc->drc_owner, tx);
1879 		if (error != 0) {
1880 			dsl_dataset_rele(origin_head, FTAG);
1881 			return (error);
1882 		}
1883 		error = dsl_dataset_snapshot_check_impl(origin_head,
1884 		    drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
1885 		dsl_dataset_rele(origin_head, FTAG);
1886 		if (error != 0)
1887 			return (error);
1888 
1889 		error = dsl_destroy_head_check_impl(drc->drc_ds, 1);
1890 	} else {
1891 		error = dsl_dataset_snapshot_check_impl(drc->drc_ds,
1892 		    drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
1893 	}
1894 	return (error);
1895 }
1896 
1897 static void
1898 dmu_recv_end_sync(void *arg, dmu_tx_t *tx)
1899 {
1900 	dmu_recv_cookie_t *drc = arg;
1901 	dsl_pool_t *dp = dmu_tx_pool(tx);
1902 
1903 	spa_history_log_internal_ds(drc->drc_ds, "finish receiving",
1904 	    tx, "snap=%s", drc->drc_tosnap);
1905 
1906 	if (!drc->drc_newfs) {
1907 		dsl_dataset_t *origin_head;
1908 
1909 		VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG,
1910 		    &origin_head));
1911 
1912 		if (drc->drc_force) {
1913 			/*
1914 			 * Destroy any snapshots of drc_tofs (origin_head)
1915 			 * after the origin (the snap before drc_ds).
1916 			 */
1917 			uint64_t obj = origin_head->ds_phys->ds_prev_snap_obj;
1918 			while (obj != drc->drc_ds->ds_phys->ds_prev_snap_obj) {
1919 				dsl_dataset_t *snap;
1920 				VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG,
1921 				    &snap));
1922 				ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir);
1923 				obj = snap->ds_phys->ds_prev_snap_obj;
1924 				dsl_destroy_snapshot_sync_impl(snap,
1925 				    B_FALSE, tx);
1926 				dsl_dataset_rele(snap, FTAG);
1927 			}
1928 		}
1929 		VERIFY3P(drc->drc_ds->ds_prev, ==,
1930 		    origin_head->ds_prev);
1931 
1932 		dsl_dataset_clone_swap_sync_impl(drc->drc_ds,
1933 		    origin_head, tx);
1934 		dsl_dataset_snapshot_sync_impl(origin_head,
1935 		    drc->drc_tosnap, tx);
1936 
1937 		/* set snapshot's creation time and guid */
1938 		dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx);
1939 		origin_head->ds_prev->ds_phys->ds_creation_time =
1940 		    drc->drc_drrb->drr_creation_time;
1941 		origin_head->ds_prev->ds_phys->ds_guid =
1942 		    drc->drc_drrb->drr_toguid;
1943 		origin_head->ds_prev->ds_phys->ds_flags &=
1944 		    ~DS_FLAG_INCONSISTENT;
1945 
1946 		dmu_buf_will_dirty(origin_head->ds_dbuf, tx);
1947 		origin_head->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1948 
1949 		dsl_dataset_rele(origin_head, FTAG);
1950 		dsl_destroy_head_sync_impl(drc->drc_ds, tx);
1951 
1952 		if (drc->drc_owner != NULL)
1953 			VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner);
1954 	} else {
1955 		dsl_dataset_t *ds = drc->drc_ds;
1956 
1957 		dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx);
1958 
1959 		/* set snapshot's creation time and guid */
1960 		dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
1961 		ds->ds_prev->ds_phys->ds_creation_time =
1962 		    drc->drc_drrb->drr_creation_time;
1963 		ds->ds_prev->ds_phys->ds_guid = drc->drc_drrb->drr_toguid;
1964 		ds->ds_prev->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1965 
1966 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
1967 		ds->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1968 	}
1969 	drc->drc_newsnapobj = drc->drc_ds->ds_phys->ds_prev_snap_obj;
1970 	/*
1971 	 * Release the hold from dmu_recv_begin.  This must be done before
1972 	 * we return to open context, so that when we free the dataset's dnode,
1973 	 * we can evict its bonus buffer.
1974 	 */
1975 	dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
1976 	drc->drc_ds = NULL;
1977 }
1978 
1979 static int
1980 add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj)
1981 {
1982 	dsl_pool_t *dp;
1983 	dsl_dataset_t *snapds;
1984 	guid_map_entry_t *gmep;
1985 	int err;
1986 
1987 	ASSERT(guid_map != NULL);
1988 
1989 	err = dsl_pool_hold(name, FTAG, &dp);
1990 	if (err != 0)
1991 		return (err);
1992 	gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP);
1993 	err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds);
1994 	if (err == 0) {
1995 		gmep->guid = snapds->ds_phys->ds_guid;
1996 		gmep->gme_ds = snapds;
1997 		avl_add(guid_map, gmep);
1998 		dsl_dataset_long_hold(snapds, gmep);
1999 	} else {
2000 		kmem_free(gmep, sizeof (*gmep));
2001 	}
2002 
2003 	dsl_pool_rele(dp, FTAG);
2004 	return (err);
2005 }
2006 
2007 static int dmu_recv_end_modified_blocks = 3;
2008 
2009 static int
2010 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
2011 {
2012 	int error;
2013 	char name[MAXNAMELEN];
2014 
2015 #ifdef _KERNEL
2016 	/*
2017 	 * We will be destroying the ds; make sure its origin is unmounted if
2018 	 * necessary.
2019 	 */
2020 	dsl_dataset_name(drc->drc_ds, name);
2021 	zfs_destroy_unmount_origin(name);
2022 #endif
2023 
2024 	error = dsl_sync_task(drc->drc_tofs,
2025 	    dmu_recv_end_check, dmu_recv_end_sync, drc,
2026 	    dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL);
2027 
2028 	if (error != 0)
2029 		dmu_recv_cleanup_ds(drc);
2030 	return (error);
2031 }
2032 
2033 static int
2034 dmu_recv_new_end(dmu_recv_cookie_t *drc)
2035 {
2036 	int error;
2037 
2038 	error = dsl_sync_task(drc->drc_tofs,
2039 	    dmu_recv_end_check, dmu_recv_end_sync, drc,
2040 	    dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL);
2041 
2042 	if (error != 0) {
2043 		dmu_recv_cleanup_ds(drc);
2044 	} else if (drc->drc_guid_to_ds_map != NULL) {
2045 		(void) add_ds_to_guidmap(drc->drc_tofs,
2046 		    drc->drc_guid_to_ds_map,
2047 		    drc->drc_newsnapobj);
2048 	}
2049 	return (error);
2050 }
2051 
2052 int
2053 dmu_recv_end(dmu_recv_cookie_t *drc, void *owner)
2054 {
2055 	drc->drc_owner = owner;
2056 
2057 	if (drc->drc_newfs)
2058 		return (dmu_recv_new_end(drc));
2059 	else
2060 		return (dmu_recv_existing_end(drc));
2061 }
2062 
2063 /*
2064  * Return TRUE if this objset is currently being received into.
2065  */
2066 boolean_t
2067 dmu_objset_is_receiving(objset_t *os)
2068 {
2069 	return (os->os_dsl_dataset != NULL &&
2070 	    os->os_dsl_dataset->ds_owner == dmu_recv_tag);
2071 }
2072