xref: /titanic_44/usr/src/uts/common/fs/zfs/dmu_send.c (revision 86d7016b0051dd58772baafe5b5bcee51d560b05)
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) 2012 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 
50 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
51 int zfs_send_corrupt_data = B_FALSE;
52 
53 static char *dmu_recv_tag = "dmu_recv_tag";
54 
55 static int
56 dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
57 {
58 	dsl_dataset_t *ds = dsp->dsa_os->os_dsl_dataset;
59 	ssize_t resid; /* have to get resid to get detailed errno */
60 	ASSERT3U(len % 8, ==, 0);
61 
62 	fletcher_4_incremental_native(buf, len, &dsp->dsa_zc);
63 	dsp->dsa_err = vn_rdwr(UIO_WRITE, dsp->dsa_vp,
64 	    (caddr_t)buf, len,
65 	    0, UIO_SYSSPACE, FAPPEND, RLIM64_INFINITY, CRED(), &resid);
66 
67 	mutex_enter(&ds->ds_sendstream_lock);
68 	*dsp->dsa_off += len;
69 	mutex_exit(&ds->ds_sendstream_lock);
70 
71 	return (dsp->dsa_err);
72 }
73 
74 static int
75 dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
76     uint64_t length)
77 {
78 	struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free);
79 
80 	if (length != -1ULL && offset + length < offset)
81 		length = -1ULL;
82 
83 	/*
84 	 * If there is a pending op, but it's not PENDING_FREE, push it out,
85 	 * since free block aggregation can only be done for blocks of the
86 	 * same type (i.e., DRR_FREE records can only be aggregated with
87 	 * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
88 	 * aggregated with other DRR_FREEOBJECTS records.
89 	 */
90 	if (dsp->dsa_pending_op != PENDING_NONE &&
91 	    dsp->dsa_pending_op != PENDING_FREE) {
92 		if (dump_bytes(dsp, dsp->dsa_drr,
93 		    sizeof (dmu_replay_record_t)) != 0)
94 			return (EINTR);
95 		dsp->dsa_pending_op = PENDING_NONE;
96 	}
97 
98 	if (dsp->dsa_pending_op == PENDING_FREE) {
99 		/*
100 		 * There should never be a PENDING_FREE if length is -1
101 		 * (because dump_dnode is the only place where this
102 		 * function is called with a -1, and only after flushing
103 		 * any pending record).
104 		 */
105 		ASSERT(length != -1ULL);
106 		/*
107 		 * Check to see whether this free block can be aggregated
108 		 * with pending one.
109 		 */
110 		if (drrf->drr_object == object && drrf->drr_offset +
111 		    drrf->drr_length == offset) {
112 			drrf->drr_length += length;
113 			return (0);
114 		} else {
115 			/* not a continuation.  Push out pending record */
116 			if (dump_bytes(dsp, dsp->dsa_drr,
117 			    sizeof (dmu_replay_record_t)) != 0)
118 				return (EINTR);
119 			dsp->dsa_pending_op = PENDING_NONE;
120 		}
121 	}
122 	/* create a FREE record and make it pending */
123 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
124 	dsp->dsa_drr->drr_type = DRR_FREE;
125 	drrf->drr_object = object;
126 	drrf->drr_offset = offset;
127 	drrf->drr_length = length;
128 	drrf->drr_toguid = dsp->dsa_toguid;
129 	if (length == -1ULL) {
130 		if (dump_bytes(dsp, dsp->dsa_drr,
131 		    sizeof (dmu_replay_record_t)) != 0)
132 			return (EINTR);
133 	} else {
134 		dsp->dsa_pending_op = PENDING_FREE;
135 	}
136 
137 	return (0);
138 }
139 
140 static int
141 dump_data(dmu_sendarg_t *dsp, dmu_object_type_t type,
142     uint64_t object, uint64_t offset, int blksz, const blkptr_t *bp, void *data)
143 {
144 	struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write);
145 
146 
147 	/*
148 	 * If there is any kind of pending aggregation (currently either
149 	 * a grouping of free objects or free blocks), push it out to
150 	 * the stream, since aggregation can't be done across operations
151 	 * of different types.
152 	 */
153 	if (dsp->dsa_pending_op != PENDING_NONE) {
154 		if (dump_bytes(dsp, dsp->dsa_drr,
155 		    sizeof (dmu_replay_record_t)) != 0)
156 			return (EINTR);
157 		dsp->dsa_pending_op = PENDING_NONE;
158 	}
159 	/* write a DATA record */
160 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
161 	dsp->dsa_drr->drr_type = DRR_WRITE;
162 	drrw->drr_object = object;
163 	drrw->drr_type = type;
164 	drrw->drr_offset = offset;
165 	drrw->drr_length = blksz;
166 	drrw->drr_toguid = dsp->dsa_toguid;
167 	drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
168 	if (zio_checksum_table[drrw->drr_checksumtype].ci_dedup)
169 		drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
170 	DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
171 	DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
172 	DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
173 	drrw->drr_key.ddk_cksum = bp->blk_cksum;
174 
175 	if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
176 		return (EINTR);
177 	if (dump_bytes(dsp, data, blksz) != 0)
178 		return (EINTR);
179 	return (0);
180 }
181 
182 static int
183 dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data)
184 {
185 	struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill);
186 
187 	if (dsp->dsa_pending_op != PENDING_NONE) {
188 		if (dump_bytes(dsp, dsp->dsa_drr,
189 		    sizeof (dmu_replay_record_t)) != 0)
190 			return (EINTR);
191 		dsp->dsa_pending_op = PENDING_NONE;
192 	}
193 
194 	/* write a SPILL record */
195 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
196 	dsp->dsa_drr->drr_type = DRR_SPILL;
197 	drrs->drr_object = object;
198 	drrs->drr_length = blksz;
199 	drrs->drr_toguid = dsp->dsa_toguid;
200 
201 	if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)))
202 		return (EINTR);
203 	if (dump_bytes(dsp, data, blksz))
204 		return (EINTR);
205 	return (0);
206 }
207 
208 static int
209 dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
210 {
211 	struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
212 
213 	/*
214 	 * If there is a pending op, but it's not PENDING_FREEOBJECTS,
215 	 * push it out, since free block aggregation can only be done for
216 	 * blocks of the same type (i.e., DRR_FREE records can only be
217 	 * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
218 	 * can only be aggregated with other DRR_FREEOBJECTS records.
219 	 */
220 	if (dsp->dsa_pending_op != PENDING_NONE &&
221 	    dsp->dsa_pending_op != PENDING_FREEOBJECTS) {
222 		if (dump_bytes(dsp, dsp->dsa_drr,
223 		    sizeof (dmu_replay_record_t)) != 0)
224 			return (EINTR);
225 		dsp->dsa_pending_op = PENDING_NONE;
226 	}
227 	if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) {
228 		/*
229 		 * See whether this free object array can be aggregated
230 		 * with pending one
231 		 */
232 		if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
233 			drrfo->drr_numobjs += numobjs;
234 			return (0);
235 		} else {
236 			/* can't be aggregated.  Push out pending record */
237 			if (dump_bytes(dsp, dsp->dsa_drr,
238 			    sizeof (dmu_replay_record_t)) != 0)
239 				return (EINTR);
240 			dsp->dsa_pending_op = PENDING_NONE;
241 		}
242 	}
243 
244 	/* write a FREEOBJECTS record */
245 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
246 	dsp->dsa_drr->drr_type = DRR_FREEOBJECTS;
247 	drrfo->drr_firstobj = firstobj;
248 	drrfo->drr_numobjs = numobjs;
249 	drrfo->drr_toguid = dsp->dsa_toguid;
250 
251 	dsp->dsa_pending_op = PENDING_FREEOBJECTS;
252 
253 	return (0);
254 }
255 
256 static int
257 dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
258 {
259 	struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object);
260 
261 	if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
262 		return (dump_freeobjects(dsp, object, 1));
263 
264 	if (dsp->dsa_pending_op != PENDING_NONE) {
265 		if (dump_bytes(dsp, dsp->dsa_drr,
266 		    sizeof (dmu_replay_record_t)) != 0)
267 			return (EINTR);
268 		dsp->dsa_pending_op = PENDING_NONE;
269 	}
270 
271 	/* write an OBJECT record */
272 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
273 	dsp->dsa_drr->drr_type = DRR_OBJECT;
274 	drro->drr_object = object;
275 	drro->drr_type = dnp->dn_type;
276 	drro->drr_bonustype = dnp->dn_bonustype;
277 	drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
278 	drro->drr_bonuslen = dnp->dn_bonuslen;
279 	drro->drr_checksumtype = dnp->dn_checksum;
280 	drro->drr_compress = dnp->dn_compress;
281 	drro->drr_toguid = dsp->dsa_toguid;
282 
283 	if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
284 		return (EINTR);
285 
286 	if (dump_bytes(dsp, DN_BONUS(dnp), P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0)
287 		return (EINTR);
288 
289 	/* free anything past the end of the file */
290 	if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) *
291 	    (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL))
292 		return (EINTR);
293 	if (dsp->dsa_err)
294 		return (EINTR);
295 	return (0);
296 }
297 
298 #define	BP_SPAN(dnp, level) \
299 	(((uint64_t)dnp->dn_datablkszsec) << (SPA_MINBLOCKSHIFT + \
300 	(level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)))
301 
302 /* ARGSUSED */
303 static int
304 backup_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, arc_buf_t *pbuf,
305     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
306 {
307 	dmu_sendarg_t *dsp = arg;
308 	dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
309 	int err = 0;
310 
311 	if (issig(JUSTLOOKING) && issig(FORREAL))
312 		return (EINTR);
313 
314 	if (zb->zb_object != DMU_META_DNODE_OBJECT &&
315 	    DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
316 		return (0);
317 	} else if (bp == NULL && zb->zb_object == DMU_META_DNODE_OBJECT) {
318 		uint64_t span = BP_SPAN(dnp, zb->zb_level);
319 		uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
320 		err = dump_freeobjects(dsp, dnobj, span >> DNODE_SHIFT);
321 	} else if (bp == NULL) {
322 		uint64_t span = BP_SPAN(dnp, zb->zb_level);
323 		err = dump_free(dsp, zb->zb_object, zb->zb_blkid * span, span);
324 	} else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
325 		return (0);
326 	} else if (type == DMU_OT_DNODE) {
327 		dnode_phys_t *blk;
328 		int i;
329 		int blksz = BP_GET_LSIZE(bp);
330 		uint32_t aflags = ARC_WAIT;
331 		arc_buf_t *abuf;
332 
333 		if (dsl_read(NULL, spa, bp, pbuf,
334 		    arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
335 		    ZIO_FLAG_CANFAIL, &aflags, zb) != 0)
336 			return (EIO);
337 
338 		blk = abuf->b_data;
339 		for (i = 0; i < blksz >> DNODE_SHIFT; i++) {
340 			uint64_t dnobj = (zb->zb_blkid <<
341 			    (DNODE_BLOCK_SHIFT - DNODE_SHIFT)) + i;
342 			err = dump_dnode(dsp, dnobj, blk+i);
343 			if (err)
344 				break;
345 		}
346 		(void) arc_buf_remove_ref(abuf, &abuf);
347 	} else if (type == DMU_OT_SA) {
348 		uint32_t aflags = ARC_WAIT;
349 		arc_buf_t *abuf;
350 		int blksz = BP_GET_LSIZE(bp);
351 
352 		if (arc_read_nolock(NULL, spa, bp,
353 		    arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
354 		    ZIO_FLAG_CANFAIL, &aflags, zb) != 0)
355 			return (EIO);
356 
357 		err = dump_spill(dsp, zb->zb_object, blksz, abuf->b_data);
358 		(void) arc_buf_remove_ref(abuf, &abuf);
359 	} else { /* it's a level-0 block of a regular object */
360 		uint32_t aflags = ARC_WAIT;
361 		arc_buf_t *abuf;
362 		int blksz = BP_GET_LSIZE(bp);
363 
364 		if (dsl_read(NULL, spa, bp, pbuf,
365 		    arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
366 		    ZIO_FLAG_CANFAIL, &aflags, zb) != 0) {
367 			if (zfs_send_corrupt_data) {
368 				/* Send a block filled with 0x"zfs badd bloc" */
369 				abuf = arc_buf_alloc(spa, blksz, &abuf,
370 				    ARC_BUFC_DATA);
371 				uint64_t *ptr;
372 				for (ptr = abuf->b_data;
373 				    (char *)ptr < (char *)abuf->b_data + blksz;
374 				    ptr++)
375 					*ptr = 0x2f5baddb10c;
376 			} else {
377 				return (EIO);
378 			}
379 		}
380 
381 		err = dump_data(dsp, type, zb->zb_object, zb->zb_blkid * blksz,
382 		    blksz, bp, abuf->b_data);
383 		(void) arc_buf_remove_ref(abuf, &abuf);
384 	}
385 
386 	ASSERT(err == 0 || err == EINTR);
387 	return (err);
388 }
389 
390 /*
391  * Return TRUE if 'earlier' is an earlier snapshot in 'later's timeline.
392  * For example, they could both be snapshots of the same filesystem, and
393  * 'earlier' is before 'later'.  Or 'earlier' could be the origin of
394  * 'later's filesystem.  Or 'earlier' could be an older snapshot in the origin's
395  * filesystem.  Or 'earlier' could be the origin's origin.
396  */
397 static boolean_t
398 is_before(dsl_dataset_t *later, dsl_dataset_t *earlier)
399 {
400 	dsl_pool_t *dp = later->ds_dir->dd_pool;
401 	int error;
402 	boolean_t ret;
403 	dsl_dataset_t *origin;
404 
405 	if (earlier->ds_phys->ds_creation_txg >=
406 	    later->ds_phys->ds_creation_txg)
407 		return (B_FALSE);
408 
409 	if (later->ds_dir == earlier->ds_dir)
410 		return (B_TRUE);
411 	if (!dsl_dir_is_clone(later->ds_dir))
412 		return (B_FALSE);
413 
414 	rw_enter(&dp->dp_config_rwlock, RW_READER);
415 	if (later->ds_dir->dd_phys->dd_origin_obj == earlier->ds_object) {
416 		rw_exit(&dp->dp_config_rwlock);
417 		return (B_TRUE);
418 	}
419 	error = dsl_dataset_hold_obj(dp,
420 	    later->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin);
421 	rw_exit(&dp->dp_config_rwlock);
422 	if (error != 0)
423 		return (B_FALSE);
424 	ret = is_before(origin, earlier);
425 	dsl_dataset_rele(origin, FTAG);
426 	return (ret);
427 }
428 
429 int
430 dmu_send(objset_t *tosnap, objset_t *fromsnap, int outfd, vnode_t *vp,
431     offset_t *off)
432 {
433 	dsl_dataset_t *ds = tosnap->os_dsl_dataset;
434 	dsl_dataset_t *fromds = fromsnap ? fromsnap->os_dsl_dataset : NULL;
435 	dmu_replay_record_t *drr;
436 	dmu_sendarg_t *dsp;
437 	int err;
438 	uint64_t fromtxg = 0;
439 
440 	/* tosnap must be a snapshot */
441 	if (ds->ds_phys->ds_next_snap_obj == 0)
442 		return (EINVAL);
443 
444 	/*
445 	 * fromsnap must be an earlier snapshot from the same fs as tosnap,
446 	 * or the origin's fs.
447 	 */
448 	if (fromds != NULL && !is_before(ds, fromds))
449 		return (EXDEV);
450 
451 	drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
452 	drr->drr_type = DRR_BEGIN;
453 	drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
454 	DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
455 	    DMU_SUBSTREAM);
456 
457 #ifdef _KERNEL
458 	if (dmu_objset_type(tosnap) == DMU_OST_ZFS) {
459 		uint64_t version;
460 		if (zfs_get_zplprop(tosnap, ZFS_PROP_VERSION, &version) != 0) {
461 			kmem_free(drr, sizeof (dmu_replay_record_t));
462 			return (EINVAL);
463 		}
464 		if (version == ZPL_VERSION_SA) {
465 			DMU_SET_FEATUREFLAGS(
466 			    drr->drr_u.drr_begin.drr_versioninfo,
467 			    DMU_BACKUP_FEATURE_SA_SPILL);
468 		}
469 	}
470 #endif
471 
472 	drr->drr_u.drr_begin.drr_creation_time =
473 	    ds->ds_phys->ds_creation_time;
474 	drr->drr_u.drr_begin.drr_type = tosnap->os_phys->os_type;
475 	if (fromds != NULL && ds->ds_dir != fromds->ds_dir)
476 		drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
477 	drr->drr_u.drr_begin.drr_toguid = ds->ds_phys->ds_guid;
478 	if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
479 		drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
480 
481 	if (fromds)
482 		drr->drr_u.drr_begin.drr_fromguid = fromds->ds_phys->ds_guid;
483 	dsl_dataset_name(ds, drr->drr_u.drr_begin.drr_toname);
484 
485 	if (fromds)
486 		fromtxg = fromds->ds_phys->ds_creation_txg;
487 
488 	dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP);
489 
490 	dsp->dsa_drr = drr;
491 	dsp->dsa_vp = vp;
492 	dsp->dsa_outfd = outfd;
493 	dsp->dsa_proc = curproc;
494 	dsp->dsa_os = tosnap;
495 	dsp->dsa_off = off;
496 	dsp->dsa_toguid = ds->ds_phys->ds_guid;
497 	ZIO_SET_CHECKSUM(&dsp->dsa_zc, 0, 0, 0, 0);
498 	dsp->dsa_pending_op = PENDING_NONE;
499 
500 	mutex_enter(&ds->ds_sendstream_lock);
501 	list_insert_head(&ds->ds_sendstreams, dsp);
502 	mutex_exit(&ds->ds_sendstream_lock);
503 
504 	if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0) {
505 		err = dsp->dsa_err;
506 		goto out;
507 	}
508 
509 	err = traverse_dataset(ds, fromtxg, TRAVERSE_PRE | TRAVERSE_PREFETCH,
510 	    backup_cb, dsp);
511 
512 	if (dsp->dsa_pending_op != PENDING_NONE)
513 		if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0)
514 			err = EINTR;
515 
516 	if (err) {
517 		if (err == EINTR && dsp->dsa_err)
518 			err = dsp->dsa_err;
519 		goto out;
520 	}
521 
522 	bzero(drr, sizeof (dmu_replay_record_t));
523 	drr->drr_type = DRR_END;
524 	drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc;
525 	drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid;
526 
527 	if (dump_bytes(dsp, drr, sizeof (dmu_replay_record_t)) != 0) {
528 		err = dsp->dsa_err;
529 		goto out;
530 	}
531 
532 out:
533 	mutex_enter(&ds->ds_sendstream_lock);
534 	list_remove(&ds->ds_sendstreams, dsp);
535 	mutex_exit(&ds->ds_sendstream_lock);
536 
537 	kmem_free(drr, sizeof (dmu_replay_record_t));
538 	kmem_free(dsp, sizeof (dmu_sendarg_t));
539 
540 	return (err);
541 }
542 
543 int
544 dmu_send_estimate(objset_t *tosnap, objset_t *fromsnap, uint64_t *sizep)
545 {
546 	dsl_dataset_t *ds = tosnap->os_dsl_dataset;
547 	dsl_dataset_t *fromds = fromsnap ? fromsnap->os_dsl_dataset : NULL;
548 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
549 	int err;
550 	uint64_t size;
551 
552 	/* tosnap must be a snapshot */
553 	if (ds->ds_phys->ds_next_snap_obj == 0)
554 		return (EINVAL);
555 
556 	/*
557 	 * fromsnap must be an earlier snapshot from the same fs as tosnap,
558 	 * or the origin's fs.
559 	 */
560 	if (fromds != NULL && !is_before(ds, fromds))
561 		return (EXDEV);
562 
563 	/* Get uncompressed size estimate of changed data. */
564 	if (fromds == NULL) {
565 		size = ds->ds_phys->ds_uncompressed_bytes;
566 	} else {
567 		uint64_t used, comp;
568 		err = dsl_dataset_space_written(fromds, ds,
569 		    &used, &comp, &size);
570 		if (err)
571 			return (err);
572 	}
573 
574 	/*
575 	 * Assume that space (both on-disk and in-stream) is dominated by
576 	 * data.  We will adjust for indirect blocks and the copies property,
577 	 * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
578 	 */
579 
580 	/*
581 	 * Subtract out approximate space used by indirect blocks.
582 	 * Assume most space is used by data blocks (non-indirect, non-dnode).
583 	 * Assume all blocks are recordsize.  Assume ditto blocks and
584 	 * internal fragmentation counter out compression.
585 	 *
586 	 * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
587 	 * block, which we observe in practice.
588 	 */
589 	uint64_t recordsize;
590 	rw_enter(&dp->dp_config_rwlock, RW_READER);
591 	err = dsl_prop_get_ds(ds, "recordsize",
592 	    sizeof (recordsize), 1, &recordsize, NULL);
593 	rw_exit(&dp->dp_config_rwlock);
594 	if (err)
595 		return (err);
596 	size -= size / recordsize * sizeof (blkptr_t);
597 
598 	/* Add in the space for the record associated with each block. */
599 	size += size / recordsize * sizeof (dmu_replay_record_t);
600 
601 	*sizep = size;
602 
603 	return (0);
604 }
605 
606 struct recvbeginsyncarg {
607 	const char *tofs;
608 	const char *tosnap;
609 	dsl_dataset_t *origin;
610 	uint64_t fromguid;
611 	dmu_objset_type_t type;
612 	void *tag;
613 	boolean_t force;
614 	uint64_t dsflags;
615 	char clonelastname[MAXNAMELEN];
616 	dsl_dataset_t *ds; /* the ds to recv into; returned from the syncfunc */
617 	cred_t *cr;
618 };
619 
620 /* ARGSUSED */
621 static int
622 recv_new_check(void *arg1, void *arg2, dmu_tx_t *tx)
623 {
624 	dsl_dir_t *dd = arg1;
625 	struct recvbeginsyncarg *rbsa = arg2;
626 	objset_t *mos = dd->dd_pool->dp_meta_objset;
627 	uint64_t val;
628 	int err;
629 
630 	err = zap_lookup(mos, dd->dd_phys->dd_child_dir_zapobj,
631 	    strrchr(rbsa->tofs, '/') + 1, sizeof (uint64_t), 1, &val);
632 
633 	if (err != ENOENT)
634 		return (err ? err : EEXIST);
635 
636 	if (rbsa->origin) {
637 		/* make sure it's a snap in the same pool */
638 		if (rbsa->origin->ds_dir->dd_pool != dd->dd_pool)
639 			return (EXDEV);
640 		if (!dsl_dataset_is_snapshot(rbsa->origin))
641 			return (EINVAL);
642 		if (rbsa->origin->ds_phys->ds_guid != rbsa->fromguid)
643 			return (ENODEV);
644 	}
645 
646 	return (0);
647 }
648 
649 static void
650 recv_new_sync(void *arg1, void *arg2, dmu_tx_t *tx)
651 {
652 	dsl_dir_t *dd = arg1;
653 	struct recvbeginsyncarg *rbsa = arg2;
654 	uint64_t flags = DS_FLAG_INCONSISTENT | rbsa->dsflags;
655 	uint64_t dsobj;
656 
657 	/* Create and open new dataset. */
658 	dsobj = dsl_dataset_create_sync(dd, strrchr(rbsa->tofs, '/') + 1,
659 	    rbsa->origin, flags, rbsa->cr, tx);
660 	VERIFY(0 == dsl_dataset_own_obj(dd->dd_pool, dsobj,
661 	    B_TRUE, dmu_recv_tag, &rbsa->ds));
662 
663 	if (rbsa->origin == NULL) {
664 		(void) dmu_objset_create_impl(dd->dd_pool->dp_spa,
665 		    rbsa->ds, &rbsa->ds->ds_phys->ds_bp, rbsa->type, tx);
666 	}
667 
668 	spa_history_log_internal_ds(rbsa->ds, "receive new", tx, "");
669 }
670 
671 /* ARGSUSED */
672 static int
673 recv_existing_check(void *arg1, void *arg2, dmu_tx_t *tx)
674 {
675 	dsl_dataset_t *ds = arg1;
676 	struct recvbeginsyncarg *rbsa = arg2;
677 	int err;
678 	uint64_t val;
679 
680 	/* must not have any changes since most recent snapshot */
681 	if (!rbsa->force && dsl_dataset_modified_since_lastsnap(ds))
682 		return (ETXTBSY);
683 
684 	/* new snapshot name must not exist */
685 	err = zap_lookup(ds->ds_dir->dd_pool->dp_meta_objset,
686 	    ds->ds_phys->ds_snapnames_zapobj, rbsa->tosnap, 8, 1, &val);
687 	if (err == 0)
688 		return (EEXIST);
689 	if (err != ENOENT)
690 		return (err);
691 
692 	if (rbsa->fromguid) {
693 		/* if incremental, most recent snapshot must match fromguid */
694 		if (ds->ds_prev == NULL)
695 			return (ENODEV);
696 
697 		/*
698 		 * most recent snapshot must match fromguid, or there are no
699 		 * changes since the fromguid one
700 		 */
701 		if (ds->ds_prev->ds_phys->ds_guid != rbsa->fromguid) {
702 			uint64_t birth = ds->ds_prev->ds_phys->ds_bp.blk_birth;
703 			uint64_t obj = ds->ds_prev->ds_phys->ds_prev_snap_obj;
704 			while (obj != 0) {
705 				dsl_dataset_t *snap;
706 				err = dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
707 				    obj, FTAG, &snap);
708 				if (err)
709 					return (ENODEV);
710 				if (snap->ds_phys->ds_creation_txg < birth) {
711 					dsl_dataset_rele(snap, FTAG);
712 					return (ENODEV);
713 				}
714 				if (snap->ds_phys->ds_guid == rbsa->fromguid) {
715 					dsl_dataset_rele(snap, FTAG);
716 					break; /* it's ok */
717 				}
718 				obj = snap->ds_phys->ds_prev_snap_obj;
719 				dsl_dataset_rele(snap, FTAG);
720 			}
721 			if (obj == 0)
722 				return (ENODEV);
723 		}
724 	} else {
725 		/* if full, most recent snapshot must be $ORIGIN */
726 		if (ds->ds_phys->ds_prev_snap_txg >= TXG_INITIAL)
727 			return (ENODEV);
728 	}
729 
730 	/* temporary clone name must not exist */
731 	err = zap_lookup(ds->ds_dir->dd_pool->dp_meta_objset,
732 	    ds->ds_dir->dd_phys->dd_child_dir_zapobj,
733 	    rbsa->clonelastname, 8, 1, &val);
734 	if (err == 0)
735 		return (EEXIST);
736 	if (err != ENOENT)
737 		return (err);
738 
739 	return (0);
740 }
741 
742 /* ARGSUSED */
743 static void
744 recv_existing_sync(void *arg1, void *arg2, dmu_tx_t *tx)
745 {
746 	dsl_dataset_t *ohds = arg1;
747 	struct recvbeginsyncarg *rbsa = arg2;
748 	dsl_pool_t *dp = ohds->ds_dir->dd_pool;
749 	dsl_dataset_t *cds;
750 	uint64_t flags = DS_FLAG_INCONSISTENT | rbsa->dsflags;
751 	uint64_t dsobj;
752 
753 	/* create and open the temporary clone */
754 	dsobj = dsl_dataset_create_sync(ohds->ds_dir, rbsa->clonelastname,
755 	    ohds->ds_prev, flags, rbsa->cr, tx);
756 	VERIFY(0 == dsl_dataset_own_obj(dp, dsobj, B_TRUE, dmu_recv_tag, &cds));
757 
758 	/*
759 	 * If we actually created a non-clone, we need to create the
760 	 * objset in our new dataset.
761 	 */
762 	if (BP_IS_HOLE(dsl_dataset_get_blkptr(cds))) {
763 		(void) dmu_objset_create_impl(dp->dp_spa,
764 		    cds, dsl_dataset_get_blkptr(cds), rbsa->type, tx);
765 	}
766 
767 	rbsa->ds = cds;
768 
769 	spa_history_log_internal_ds(cds, "receive over existing", tx, "");
770 }
771 
772 static boolean_t
773 dmu_recv_verify_features(dsl_dataset_t *ds, struct drr_begin *drrb)
774 {
775 	int featureflags;
776 
777 	featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
778 
779 	/* Verify pool version supports SA if SA_SPILL feature set */
780 	return ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
781 	    (spa_version(dsl_dataset_get_spa(ds)) < SPA_VERSION_SA));
782 }
783 
784 /*
785  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
786  * succeeds; otherwise we will leak the holds on the datasets.
787  */
788 int
789 dmu_recv_begin(char *tofs, char *tosnap, char *top_ds, struct drr_begin *drrb,
790     boolean_t force, objset_t *origin, dmu_recv_cookie_t *drc)
791 {
792 	int err = 0;
793 	boolean_t byteswap;
794 	struct recvbeginsyncarg rbsa = { 0 };
795 	uint64_t versioninfo;
796 	int flags;
797 	dsl_dataset_t *ds;
798 
799 	if (drrb->drr_magic == DMU_BACKUP_MAGIC)
800 		byteswap = FALSE;
801 	else if (drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
802 		byteswap = TRUE;
803 	else
804 		return (EINVAL);
805 
806 	rbsa.tofs = tofs;
807 	rbsa.tosnap = tosnap;
808 	rbsa.origin = origin ? origin->os_dsl_dataset : NULL;
809 	rbsa.fromguid = drrb->drr_fromguid;
810 	rbsa.type = drrb->drr_type;
811 	rbsa.tag = FTAG;
812 	rbsa.dsflags = 0;
813 	rbsa.cr = CRED();
814 	versioninfo = drrb->drr_versioninfo;
815 	flags = drrb->drr_flags;
816 
817 	if (byteswap) {
818 		rbsa.type = BSWAP_32(rbsa.type);
819 		rbsa.fromguid = BSWAP_64(rbsa.fromguid);
820 		versioninfo = BSWAP_64(versioninfo);
821 		flags = BSWAP_32(flags);
822 	}
823 
824 	if (DMU_GET_STREAM_HDRTYPE(versioninfo) == DMU_COMPOUNDSTREAM ||
825 	    rbsa.type >= DMU_OST_NUMTYPES ||
826 	    ((flags & DRR_FLAG_CLONE) && origin == NULL))
827 		return (EINVAL);
828 
829 	if (flags & DRR_FLAG_CI_DATA)
830 		rbsa.dsflags = DS_FLAG_CI_DATASET;
831 
832 	bzero(drc, sizeof (dmu_recv_cookie_t));
833 	drc->drc_drrb = drrb;
834 	drc->drc_tosnap = tosnap;
835 	drc->drc_top_ds = top_ds;
836 	drc->drc_force = force;
837 
838 	/*
839 	 * Process the begin in syncing context.
840 	 */
841 
842 	/* open the dataset we are logically receiving into */
843 	err = dsl_dataset_hold(tofs, dmu_recv_tag, &ds);
844 	if (err == 0) {
845 		if (dmu_recv_verify_features(ds, drrb)) {
846 			dsl_dataset_rele(ds, dmu_recv_tag);
847 			return (ENOTSUP);
848 		}
849 		/* target fs already exists; recv into temp clone */
850 
851 		/* Can't recv a clone into an existing fs */
852 		if (flags & DRR_FLAG_CLONE) {
853 			dsl_dataset_rele(ds, dmu_recv_tag);
854 			return (EINVAL);
855 		}
856 
857 		/* must not have an incremental recv already in progress */
858 		if (!mutex_tryenter(&ds->ds_recvlock)) {
859 			dsl_dataset_rele(ds, dmu_recv_tag);
860 			return (EBUSY);
861 		}
862 
863 		/* tmp clone name is: tofs/%tosnap" */
864 		(void) snprintf(rbsa.clonelastname, sizeof (rbsa.clonelastname),
865 		    "%%%s", tosnap);
866 		rbsa.force = force;
867 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
868 		    recv_existing_check, recv_existing_sync, ds, &rbsa, 5);
869 		if (err) {
870 			mutex_exit(&ds->ds_recvlock);
871 			dsl_dataset_rele(ds, dmu_recv_tag);
872 			return (err);
873 		}
874 		drc->drc_logical_ds = ds;
875 		drc->drc_real_ds = rbsa.ds;
876 	} else if (err == ENOENT) {
877 		/* target fs does not exist; must be a full backup or clone */
878 		char *cp;
879 
880 		/*
881 		 * If it's a non-clone incremental, we are missing the
882 		 * target fs, so fail the recv.
883 		 */
884 		if (rbsa.fromguid && !(flags & DRR_FLAG_CLONE))
885 			return (ENOENT);
886 
887 		/* Open the parent of tofs */
888 		cp = strrchr(tofs, '/');
889 		*cp = '\0';
890 		err = dsl_dataset_hold(tofs, FTAG, &ds);
891 		*cp = '/';
892 		if (err)
893 			return (err);
894 
895 		if (dmu_recv_verify_features(ds, drrb)) {
896 			dsl_dataset_rele(ds, FTAG);
897 			return (ENOTSUP);
898 		}
899 
900 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
901 		    recv_new_check, recv_new_sync, ds->ds_dir, &rbsa, 5);
902 		dsl_dataset_rele(ds, FTAG);
903 		if (err)
904 			return (err);
905 		drc->drc_logical_ds = drc->drc_real_ds = rbsa.ds;
906 		drc->drc_newfs = B_TRUE;
907 	}
908 
909 	return (err);
910 }
911 
912 struct restorearg {
913 	int err;
914 	int byteswap;
915 	vnode_t *vp;
916 	char *buf;
917 	uint64_t voff;
918 	int bufsize; /* amount of memory allocated for buf */
919 	zio_cksum_t cksum;
920 	avl_tree_t *guid_to_ds_map;
921 };
922 
923 typedef struct guid_map_entry {
924 	uint64_t	guid;
925 	dsl_dataset_t	*gme_ds;
926 	avl_node_t	avlnode;
927 } guid_map_entry_t;
928 
929 static int
930 guid_compare(const void *arg1, const void *arg2)
931 {
932 	const guid_map_entry_t *gmep1 = arg1;
933 	const guid_map_entry_t *gmep2 = arg2;
934 
935 	if (gmep1->guid < gmep2->guid)
936 		return (-1);
937 	else if (gmep1->guid > gmep2->guid)
938 		return (1);
939 	return (0);
940 }
941 
942 static void
943 free_guid_map_onexit(void *arg)
944 {
945 	avl_tree_t *ca = arg;
946 	void *cookie = NULL;
947 	guid_map_entry_t *gmep;
948 
949 	while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
950 		dsl_dataset_rele(gmep->gme_ds, ca);
951 		kmem_free(gmep, sizeof (guid_map_entry_t));
952 	}
953 	avl_destroy(ca);
954 	kmem_free(ca, sizeof (avl_tree_t));
955 }
956 
957 static void *
958 restore_read(struct restorearg *ra, int len)
959 {
960 	void *rv;
961 	int done = 0;
962 
963 	/* some things will require 8-byte alignment, so everything must */
964 	ASSERT3U(len % 8, ==, 0);
965 
966 	while (done < len) {
967 		ssize_t resid;
968 
969 		ra->err = vn_rdwr(UIO_READ, ra->vp,
970 		    (caddr_t)ra->buf + done, len - done,
971 		    ra->voff, UIO_SYSSPACE, FAPPEND,
972 		    RLIM64_INFINITY, CRED(), &resid);
973 
974 		if (resid == len - done)
975 			ra->err = EINVAL;
976 		ra->voff += len - done - resid;
977 		done = len - resid;
978 		if (ra->err)
979 			return (NULL);
980 	}
981 
982 	ASSERT3U(done, ==, len);
983 	rv = ra->buf;
984 	if (ra->byteswap)
985 		fletcher_4_incremental_byteswap(rv, len, &ra->cksum);
986 	else
987 		fletcher_4_incremental_native(rv, len, &ra->cksum);
988 	return (rv);
989 }
990 
991 static void
992 backup_byteswap(dmu_replay_record_t *drr)
993 {
994 #define	DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
995 #define	DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
996 	drr->drr_type = BSWAP_32(drr->drr_type);
997 	drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
998 	switch (drr->drr_type) {
999 	case DRR_BEGIN:
1000 		DO64(drr_begin.drr_magic);
1001 		DO64(drr_begin.drr_versioninfo);
1002 		DO64(drr_begin.drr_creation_time);
1003 		DO32(drr_begin.drr_type);
1004 		DO32(drr_begin.drr_flags);
1005 		DO64(drr_begin.drr_toguid);
1006 		DO64(drr_begin.drr_fromguid);
1007 		break;
1008 	case DRR_OBJECT:
1009 		DO64(drr_object.drr_object);
1010 		/* DO64(drr_object.drr_allocation_txg); */
1011 		DO32(drr_object.drr_type);
1012 		DO32(drr_object.drr_bonustype);
1013 		DO32(drr_object.drr_blksz);
1014 		DO32(drr_object.drr_bonuslen);
1015 		DO64(drr_object.drr_toguid);
1016 		break;
1017 	case DRR_FREEOBJECTS:
1018 		DO64(drr_freeobjects.drr_firstobj);
1019 		DO64(drr_freeobjects.drr_numobjs);
1020 		DO64(drr_freeobjects.drr_toguid);
1021 		break;
1022 	case DRR_WRITE:
1023 		DO64(drr_write.drr_object);
1024 		DO32(drr_write.drr_type);
1025 		DO64(drr_write.drr_offset);
1026 		DO64(drr_write.drr_length);
1027 		DO64(drr_write.drr_toguid);
1028 		DO64(drr_write.drr_key.ddk_cksum.zc_word[0]);
1029 		DO64(drr_write.drr_key.ddk_cksum.zc_word[1]);
1030 		DO64(drr_write.drr_key.ddk_cksum.zc_word[2]);
1031 		DO64(drr_write.drr_key.ddk_cksum.zc_word[3]);
1032 		DO64(drr_write.drr_key.ddk_prop);
1033 		break;
1034 	case DRR_WRITE_BYREF:
1035 		DO64(drr_write_byref.drr_object);
1036 		DO64(drr_write_byref.drr_offset);
1037 		DO64(drr_write_byref.drr_length);
1038 		DO64(drr_write_byref.drr_toguid);
1039 		DO64(drr_write_byref.drr_refguid);
1040 		DO64(drr_write_byref.drr_refobject);
1041 		DO64(drr_write_byref.drr_refoffset);
1042 		DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[0]);
1043 		DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[1]);
1044 		DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[2]);
1045 		DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[3]);
1046 		DO64(drr_write_byref.drr_key.ddk_prop);
1047 		break;
1048 	case DRR_FREE:
1049 		DO64(drr_free.drr_object);
1050 		DO64(drr_free.drr_offset);
1051 		DO64(drr_free.drr_length);
1052 		DO64(drr_free.drr_toguid);
1053 		break;
1054 	case DRR_SPILL:
1055 		DO64(drr_spill.drr_object);
1056 		DO64(drr_spill.drr_length);
1057 		DO64(drr_spill.drr_toguid);
1058 		break;
1059 	case DRR_END:
1060 		DO64(drr_end.drr_checksum.zc_word[0]);
1061 		DO64(drr_end.drr_checksum.zc_word[1]);
1062 		DO64(drr_end.drr_checksum.zc_word[2]);
1063 		DO64(drr_end.drr_checksum.zc_word[3]);
1064 		DO64(drr_end.drr_toguid);
1065 		break;
1066 	}
1067 #undef DO64
1068 #undef DO32
1069 }
1070 
1071 static int
1072 restore_object(struct restorearg *ra, objset_t *os, struct drr_object *drro)
1073 {
1074 	int err;
1075 	dmu_tx_t *tx;
1076 	void *data = NULL;
1077 
1078 	if (drro->drr_type == DMU_OT_NONE ||
1079 	    !DMU_OT_IS_VALID(drro->drr_type) ||
1080 	    !DMU_OT_IS_VALID(drro->drr_bonustype) ||
1081 	    drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
1082 	    drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
1083 	    P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
1084 	    drro->drr_blksz < SPA_MINBLOCKSIZE ||
1085 	    drro->drr_blksz > SPA_MAXBLOCKSIZE ||
1086 	    drro->drr_bonuslen > DN_MAX_BONUSLEN) {
1087 		return (EINVAL);
1088 	}
1089 
1090 	err = dmu_object_info(os, drro->drr_object, NULL);
1091 
1092 	if (err != 0 && err != ENOENT)
1093 		return (EINVAL);
1094 
1095 	if (drro->drr_bonuslen) {
1096 		data = restore_read(ra, P2ROUNDUP(drro->drr_bonuslen, 8));
1097 		if (ra->err)
1098 			return (ra->err);
1099 	}
1100 
1101 	if (err == ENOENT) {
1102 		/* currently free, want to be allocated */
1103 		tx = dmu_tx_create(os);
1104 		dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1105 		err = dmu_tx_assign(tx, TXG_WAIT);
1106 		if (err) {
1107 			dmu_tx_abort(tx);
1108 			return (err);
1109 		}
1110 		err = dmu_object_claim(os, drro->drr_object,
1111 		    drro->drr_type, drro->drr_blksz,
1112 		    drro->drr_bonustype, drro->drr_bonuslen, tx);
1113 		dmu_tx_commit(tx);
1114 	} else {
1115 		/* currently allocated, want to be allocated */
1116 		err = dmu_object_reclaim(os, drro->drr_object,
1117 		    drro->drr_type, drro->drr_blksz,
1118 		    drro->drr_bonustype, drro->drr_bonuslen);
1119 	}
1120 	if (err) {
1121 		return (EINVAL);
1122 	}
1123 
1124 	tx = dmu_tx_create(os);
1125 	dmu_tx_hold_bonus(tx, drro->drr_object);
1126 	err = dmu_tx_assign(tx, TXG_WAIT);
1127 	if (err) {
1128 		dmu_tx_abort(tx);
1129 		return (err);
1130 	}
1131 
1132 	dmu_object_set_checksum(os, drro->drr_object, drro->drr_checksumtype,
1133 	    tx);
1134 	dmu_object_set_compress(os, drro->drr_object, drro->drr_compress, tx);
1135 
1136 	if (data != NULL) {
1137 		dmu_buf_t *db;
1138 
1139 		VERIFY(0 == dmu_bonus_hold(os, drro->drr_object, FTAG, &db));
1140 		dmu_buf_will_dirty(db, tx);
1141 
1142 		ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
1143 		bcopy(data, db->db_data, drro->drr_bonuslen);
1144 		if (ra->byteswap) {
1145 			dmu_object_byteswap_t byteswap =
1146 			    DMU_OT_BYTESWAP(drro->drr_bonustype);
1147 			dmu_ot_byteswap[byteswap].ob_func(db->db_data,
1148 			    drro->drr_bonuslen);
1149 		}
1150 		dmu_buf_rele(db, FTAG);
1151 	}
1152 	dmu_tx_commit(tx);
1153 	return (0);
1154 }
1155 
1156 /* ARGSUSED */
1157 static int
1158 restore_freeobjects(struct restorearg *ra, objset_t *os,
1159     struct drr_freeobjects *drrfo)
1160 {
1161 	uint64_t obj;
1162 
1163 	if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
1164 		return (EINVAL);
1165 
1166 	for (obj = drrfo->drr_firstobj;
1167 	    obj < drrfo->drr_firstobj + drrfo->drr_numobjs;
1168 	    (void) dmu_object_next(os, &obj, FALSE, 0)) {
1169 		int err;
1170 
1171 		if (dmu_object_info(os, obj, NULL) != 0)
1172 			continue;
1173 
1174 		err = dmu_free_object(os, obj);
1175 		if (err)
1176 			return (err);
1177 	}
1178 	return (0);
1179 }
1180 
1181 static int
1182 restore_write(struct restorearg *ra, objset_t *os,
1183     struct drr_write *drrw)
1184 {
1185 	dmu_tx_t *tx;
1186 	void *data;
1187 	int err;
1188 
1189 	if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
1190 	    !DMU_OT_IS_VALID(drrw->drr_type))
1191 		return (EINVAL);
1192 
1193 	data = restore_read(ra, drrw->drr_length);
1194 	if (data == NULL)
1195 		return (ra->err);
1196 
1197 	if (dmu_object_info(os, drrw->drr_object, NULL) != 0)
1198 		return (EINVAL);
1199 
1200 	tx = dmu_tx_create(os);
1201 
1202 	dmu_tx_hold_write(tx, drrw->drr_object,
1203 	    drrw->drr_offset, drrw->drr_length);
1204 	err = dmu_tx_assign(tx, TXG_WAIT);
1205 	if (err) {
1206 		dmu_tx_abort(tx);
1207 		return (err);
1208 	}
1209 	if (ra->byteswap) {
1210 		dmu_object_byteswap_t byteswap =
1211 		    DMU_OT_BYTESWAP(drrw->drr_type);
1212 		dmu_ot_byteswap[byteswap].ob_func(data, drrw->drr_length);
1213 	}
1214 	dmu_write(os, drrw->drr_object,
1215 	    drrw->drr_offset, drrw->drr_length, data, tx);
1216 	dmu_tx_commit(tx);
1217 	return (0);
1218 }
1219 
1220 /*
1221  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
1222  * streams to refer to a copy of the data that is already on the
1223  * system because it came in earlier in the stream.  This function
1224  * finds the earlier copy of the data, and uses that copy instead of
1225  * data from the stream to fulfill this write.
1226  */
1227 static int
1228 restore_write_byref(struct restorearg *ra, objset_t *os,
1229     struct drr_write_byref *drrwbr)
1230 {
1231 	dmu_tx_t *tx;
1232 	int err;
1233 	guid_map_entry_t gmesrch;
1234 	guid_map_entry_t *gmep;
1235 	avl_index_t	where;
1236 	objset_t *ref_os = NULL;
1237 	dmu_buf_t *dbp;
1238 
1239 	if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
1240 		return (EINVAL);
1241 
1242 	/*
1243 	 * If the GUID of the referenced dataset is different from the
1244 	 * GUID of the target dataset, find the referenced dataset.
1245 	 */
1246 	if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
1247 		gmesrch.guid = drrwbr->drr_refguid;
1248 		if ((gmep = avl_find(ra->guid_to_ds_map, &gmesrch,
1249 		    &where)) == NULL) {
1250 			return (EINVAL);
1251 		}
1252 		if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
1253 			return (EINVAL);
1254 	} else {
1255 		ref_os = os;
1256 	}
1257 
1258 	if (err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
1259 	    drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH))
1260 		return (err);
1261 
1262 	tx = dmu_tx_create(os);
1263 
1264 	dmu_tx_hold_write(tx, drrwbr->drr_object,
1265 	    drrwbr->drr_offset, drrwbr->drr_length);
1266 	err = dmu_tx_assign(tx, TXG_WAIT);
1267 	if (err) {
1268 		dmu_tx_abort(tx);
1269 		return (err);
1270 	}
1271 	dmu_write(os, drrwbr->drr_object,
1272 	    drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
1273 	dmu_buf_rele(dbp, FTAG);
1274 	dmu_tx_commit(tx);
1275 	return (0);
1276 }
1277 
1278 static int
1279 restore_spill(struct restorearg *ra, objset_t *os, struct drr_spill *drrs)
1280 {
1281 	dmu_tx_t *tx;
1282 	void *data;
1283 	dmu_buf_t *db, *db_spill;
1284 	int err;
1285 
1286 	if (drrs->drr_length < SPA_MINBLOCKSIZE ||
1287 	    drrs->drr_length > SPA_MAXBLOCKSIZE)
1288 		return (EINVAL);
1289 
1290 	data = restore_read(ra, drrs->drr_length);
1291 	if (data == NULL)
1292 		return (ra->err);
1293 
1294 	if (dmu_object_info(os, drrs->drr_object, NULL) != 0)
1295 		return (EINVAL);
1296 
1297 	VERIFY(0 == dmu_bonus_hold(os, drrs->drr_object, FTAG, &db));
1298 	if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
1299 		dmu_buf_rele(db, FTAG);
1300 		return (err);
1301 	}
1302 
1303 	tx = dmu_tx_create(os);
1304 
1305 	dmu_tx_hold_spill(tx, db->db_object);
1306 
1307 	err = dmu_tx_assign(tx, TXG_WAIT);
1308 	if (err) {
1309 		dmu_buf_rele(db, FTAG);
1310 		dmu_buf_rele(db_spill, FTAG);
1311 		dmu_tx_abort(tx);
1312 		return (err);
1313 	}
1314 	dmu_buf_will_dirty(db_spill, tx);
1315 
1316 	if (db_spill->db_size < drrs->drr_length)
1317 		VERIFY(0 == dbuf_spill_set_blksz(db_spill,
1318 		    drrs->drr_length, tx));
1319 	bcopy(data, db_spill->db_data, drrs->drr_length);
1320 
1321 	dmu_buf_rele(db, FTAG);
1322 	dmu_buf_rele(db_spill, FTAG);
1323 
1324 	dmu_tx_commit(tx);
1325 	return (0);
1326 }
1327 
1328 /* ARGSUSED */
1329 static int
1330 restore_free(struct restorearg *ra, objset_t *os,
1331     struct drr_free *drrf)
1332 {
1333 	int err;
1334 
1335 	if (drrf->drr_length != -1ULL &&
1336 	    drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
1337 		return (EINVAL);
1338 
1339 	if (dmu_object_info(os, drrf->drr_object, NULL) != 0)
1340 		return (EINVAL);
1341 
1342 	err = dmu_free_long_range(os, drrf->drr_object,
1343 	    drrf->drr_offset, drrf->drr_length);
1344 	return (err);
1345 }
1346 
1347 /*
1348  * NB: callers *must* call dmu_recv_end() if this succeeds.
1349  */
1350 int
1351 dmu_recv_stream(dmu_recv_cookie_t *drc, vnode_t *vp, offset_t *voffp,
1352     int cleanup_fd, uint64_t *action_handlep)
1353 {
1354 	struct restorearg ra = { 0 };
1355 	dmu_replay_record_t *drr;
1356 	objset_t *os;
1357 	zio_cksum_t pcksum;
1358 	int featureflags;
1359 
1360 	if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
1361 		ra.byteswap = TRUE;
1362 
1363 	{
1364 		/* compute checksum of drr_begin record */
1365 		dmu_replay_record_t *drr;
1366 		drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
1367 
1368 		drr->drr_type = DRR_BEGIN;
1369 		drr->drr_u.drr_begin = *drc->drc_drrb;
1370 		if (ra.byteswap) {
1371 			fletcher_4_incremental_byteswap(drr,
1372 			    sizeof (dmu_replay_record_t), &ra.cksum);
1373 		} else {
1374 			fletcher_4_incremental_native(drr,
1375 			    sizeof (dmu_replay_record_t), &ra.cksum);
1376 		}
1377 		kmem_free(drr, sizeof (dmu_replay_record_t));
1378 	}
1379 
1380 	if (ra.byteswap) {
1381 		struct drr_begin *drrb = drc->drc_drrb;
1382 		drrb->drr_magic = BSWAP_64(drrb->drr_magic);
1383 		drrb->drr_versioninfo = BSWAP_64(drrb->drr_versioninfo);
1384 		drrb->drr_creation_time = BSWAP_64(drrb->drr_creation_time);
1385 		drrb->drr_type = BSWAP_32(drrb->drr_type);
1386 		drrb->drr_toguid = BSWAP_64(drrb->drr_toguid);
1387 		drrb->drr_fromguid = BSWAP_64(drrb->drr_fromguid);
1388 	}
1389 
1390 	ra.vp = vp;
1391 	ra.voff = *voffp;
1392 	ra.bufsize = 1<<20;
1393 	ra.buf = kmem_alloc(ra.bufsize, KM_SLEEP);
1394 
1395 	/* these were verified in dmu_recv_begin */
1396 	ASSERT(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo) ==
1397 	    DMU_SUBSTREAM);
1398 	ASSERT(drc->drc_drrb->drr_type < DMU_OST_NUMTYPES);
1399 
1400 	/*
1401 	 * Open the objset we are modifying.
1402 	 */
1403 	VERIFY(dmu_objset_from_ds(drc->drc_real_ds, &os) == 0);
1404 
1405 	ASSERT(drc->drc_real_ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT);
1406 
1407 	featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
1408 
1409 	/* if this stream is dedup'ed, set up the avl tree for guid mapping */
1410 	if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
1411 		minor_t minor;
1412 
1413 		if (cleanup_fd == -1) {
1414 			ra.err = EBADF;
1415 			goto out;
1416 		}
1417 		ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
1418 		if (ra.err) {
1419 			cleanup_fd = -1;
1420 			goto out;
1421 		}
1422 
1423 		if (*action_handlep == 0) {
1424 			ra.guid_to_ds_map =
1425 			    kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
1426 			avl_create(ra.guid_to_ds_map, guid_compare,
1427 			    sizeof (guid_map_entry_t),
1428 			    offsetof(guid_map_entry_t, avlnode));
1429 			ra.err = zfs_onexit_add_cb(minor,
1430 			    free_guid_map_onexit, ra.guid_to_ds_map,
1431 			    action_handlep);
1432 			if (ra.err)
1433 				goto out;
1434 		} else {
1435 			ra.err = zfs_onexit_cb_data(minor, *action_handlep,
1436 			    (void **)&ra.guid_to_ds_map);
1437 			if (ra.err)
1438 				goto out;
1439 		}
1440 
1441 		drc->drc_guid_to_ds_map = ra.guid_to_ds_map;
1442 	}
1443 
1444 	/*
1445 	 * Read records and process them.
1446 	 */
1447 	pcksum = ra.cksum;
1448 	while (ra.err == 0 &&
1449 	    NULL != (drr = restore_read(&ra, sizeof (*drr)))) {
1450 		if (issig(JUSTLOOKING) && issig(FORREAL)) {
1451 			ra.err = EINTR;
1452 			goto out;
1453 		}
1454 
1455 		if (ra.byteswap)
1456 			backup_byteswap(drr);
1457 
1458 		switch (drr->drr_type) {
1459 		case DRR_OBJECT:
1460 		{
1461 			/*
1462 			 * We need to make a copy of the record header,
1463 			 * because restore_{object,write} may need to
1464 			 * restore_read(), which will invalidate drr.
1465 			 */
1466 			struct drr_object drro = drr->drr_u.drr_object;
1467 			ra.err = restore_object(&ra, os, &drro);
1468 			break;
1469 		}
1470 		case DRR_FREEOBJECTS:
1471 		{
1472 			struct drr_freeobjects drrfo =
1473 			    drr->drr_u.drr_freeobjects;
1474 			ra.err = restore_freeobjects(&ra, os, &drrfo);
1475 			break;
1476 		}
1477 		case DRR_WRITE:
1478 		{
1479 			struct drr_write drrw = drr->drr_u.drr_write;
1480 			ra.err = restore_write(&ra, os, &drrw);
1481 			break;
1482 		}
1483 		case DRR_WRITE_BYREF:
1484 		{
1485 			struct drr_write_byref drrwbr =
1486 			    drr->drr_u.drr_write_byref;
1487 			ra.err = restore_write_byref(&ra, os, &drrwbr);
1488 			break;
1489 		}
1490 		case DRR_FREE:
1491 		{
1492 			struct drr_free drrf = drr->drr_u.drr_free;
1493 			ra.err = restore_free(&ra, os, &drrf);
1494 			break;
1495 		}
1496 		case DRR_END:
1497 		{
1498 			struct drr_end drre = drr->drr_u.drr_end;
1499 			/*
1500 			 * We compare against the *previous* checksum
1501 			 * value, because the stored checksum is of
1502 			 * everything before the DRR_END record.
1503 			 */
1504 			if (!ZIO_CHECKSUM_EQUAL(drre.drr_checksum, pcksum))
1505 				ra.err = ECKSUM;
1506 			goto out;
1507 		}
1508 		case DRR_SPILL:
1509 		{
1510 			struct drr_spill drrs = drr->drr_u.drr_spill;
1511 			ra.err = restore_spill(&ra, os, &drrs);
1512 			break;
1513 		}
1514 		default:
1515 			ra.err = EINVAL;
1516 			goto out;
1517 		}
1518 		pcksum = ra.cksum;
1519 	}
1520 	ASSERT(ra.err != 0);
1521 
1522 out:
1523 	if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
1524 		zfs_onexit_fd_rele(cleanup_fd);
1525 
1526 	if (ra.err != 0) {
1527 		/*
1528 		 * destroy what we created, so we don't leave it in the
1529 		 * inconsistent restoring state.
1530 		 */
1531 		txg_wait_synced(drc->drc_real_ds->ds_dir->dd_pool, 0);
1532 
1533 		(void) dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag,
1534 		    B_FALSE);
1535 		if (drc->drc_real_ds != drc->drc_logical_ds) {
1536 			mutex_exit(&drc->drc_logical_ds->ds_recvlock);
1537 			dsl_dataset_rele(drc->drc_logical_ds, dmu_recv_tag);
1538 		}
1539 	}
1540 
1541 	kmem_free(ra.buf, ra.bufsize);
1542 	*voffp = ra.voff;
1543 	return (ra.err);
1544 }
1545 
1546 struct recvendsyncarg {
1547 	char *tosnap;
1548 	uint64_t creation_time;
1549 	uint64_t toguid;
1550 };
1551 
1552 static int
1553 recv_end_check(void *arg1, void *arg2, dmu_tx_t *tx)
1554 {
1555 	dsl_dataset_t *ds = arg1;
1556 	struct recvendsyncarg *resa = arg2;
1557 
1558 	return (dsl_dataset_snapshot_check(ds, resa->tosnap, tx));
1559 }
1560 
1561 static void
1562 recv_end_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1563 {
1564 	dsl_dataset_t *ds = arg1;
1565 	struct recvendsyncarg *resa = arg2;
1566 
1567 	dsl_dataset_snapshot_sync(ds, resa->tosnap, tx);
1568 
1569 	/* set snapshot's creation time and guid */
1570 	dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
1571 	ds->ds_prev->ds_phys->ds_creation_time = resa->creation_time;
1572 	ds->ds_prev->ds_phys->ds_guid = resa->toguid;
1573 	ds->ds_prev->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1574 
1575 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1576 	ds->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1577 	spa_history_log_internal_ds(ds, "finished receiving", tx, "");
1578 }
1579 
1580 static int
1581 add_ds_to_guidmap(avl_tree_t *guid_map, dsl_dataset_t *ds)
1582 {
1583 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1584 	uint64_t snapobj = ds->ds_phys->ds_prev_snap_obj;
1585 	dsl_dataset_t *snapds;
1586 	guid_map_entry_t *gmep;
1587 	int err;
1588 
1589 	ASSERT(guid_map != NULL);
1590 
1591 	rw_enter(&dp->dp_config_rwlock, RW_READER);
1592 	err = dsl_dataset_hold_obj(dp, snapobj, guid_map, &snapds);
1593 	if (err == 0) {
1594 		gmep = kmem_alloc(sizeof (guid_map_entry_t), KM_SLEEP);
1595 		gmep->guid = snapds->ds_phys->ds_guid;
1596 		gmep->gme_ds = snapds;
1597 		avl_add(guid_map, gmep);
1598 	}
1599 
1600 	rw_exit(&dp->dp_config_rwlock);
1601 	return (err);
1602 }
1603 
1604 static int
1605 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
1606 {
1607 	struct recvendsyncarg resa;
1608 	dsl_dataset_t *ds = drc->drc_logical_ds;
1609 	int err, myerr;
1610 
1611 	/*
1612 	 * XXX hack; seems the ds is still dirty and dsl_pool_zil_clean()
1613 	 * expects it to have a ds_user_ptr (and zil), but clone_swap()
1614 	 * can close it.
1615 	 */
1616 	txg_wait_synced(ds->ds_dir->dd_pool, 0);
1617 
1618 	if (dsl_dataset_tryown(ds, FALSE, dmu_recv_tag)) {
1619 		err = dsl_dataset_clone_swap(drc->drc_real_ds, ds,
1620 		    drc->drc_force);
1621 		if (err)
1622 			goto out;
1623 	} else {
1624 		mutex_exit(&ds->ds_recvlock);
1625 		dsl_dataset_rele(ds, dmu_recv_tag);
1626 		(void) dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag,
1627 		    B_FALSE);
1628 		return (EBUSY);
1629 	}
1630 
1631 	resa.creation_time = drc->drc_drrb->drr_creation_time;
1632 	resa.toguid = drc->drc_drrb->drr_toguid;
1633 	resa.tosnap = drc->drc_tosnap;
1634 
1635 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1636 	    recv_end_check, recv_end_sync, ds, &resa, 3);
1637 	if (err) {
1638 		/* swap back */
1639 		(void) dsl_dataset_clone_swap(drc->drc_real_ds, ds, B_TRUE);
1640 	}
1641 
1642 out:
1643 	mutex_exit(&ds->ds_recvlock);
1644 	if (err == 0 && drc->drc_guid_to_ds_map != NULL)
1645 		(void) add_ds_to_guidmap(drc->drc_guid_to_ds_map, ds);
1646 	dsl_dataset_disown(ds, dmu_recv_tag);
1647 	myerr = dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag, B_FALSE);
1648 	ASSERT3U(myerr, ==, 0);
1649 	return (err);
1650 }
1651 
1652 static int
1653 dmu_recv_new_end(dmu_recv_cookie_t *drc)
1654 {
1655 	struct recvendsyncarg resa;
1656 	dsl_dataset_t *ds = drc->drc_logical_ds;
1657 	int err;
1658 
1659 	/*
1660 	 * XXX hack; seems the ds is still dirty and dsl_pool_zil_clean()
1661 	 * expects it to have a ds_user_ptr (and zil), but clone_swap()
1662 	 * can close it.
1663 	 */
1664 	txg_wait_synced(ds->ds_dir->dd_pool, 0);
1665 
1666 	resa.creation_time = drc->drc_drrb->drr_creation_time;
1667 	resa.toguid = drc->drc_drrb->drr_toguid;
1668 	resa.tosnap = drc->drc_tosnap;
1669 
1670 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1671 	    recv_end_check, recv_end_sync, ds, &resa, 3);
1672 	if (err) {
1673 		/* clean up the fs we just recv'd into */
1674 		(void) dsl_dataset_destroy(ds, dmu_recv_tag, B_FALSE);
1675 	} else {
1676 		if (drc->drc_guid_to_ds_map != NULL)
1677 			(void) add_ds_to_guidmap(drc->drc_guid_to_ds_map, ds);
1678 		/* release the hold from dmu_recv_begin */
1679 		dsl_dataset_disown(ds, dmu_recv_tag);
1680 	}
1681 	return (err);
1682 }
1683 
1684 int
1685 dmu_recv_end(dmu_recv_cookie_t *drc)
1686 {
1687 	if (drc->drc_logical_ds != drc->drc_real_ds)
1688 		return (dmu_recv_existing_end(drc));
1689 	else
1690 		return (dmu_recv_new_end(drc));
1691 }
1692