xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu_send.c (revision 46b592853d0f4f11781b6b0a7533f267c6aee132)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/dmu.h>
27 #include <sys/dmu_impl.h>
28 #include <sys/dmu_tx.h>
29 #include <sys/dbuf.h>
30 #include <sys/dnode.h>
31 #include <sys/zfs_context.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dmu_traverse.h>
34 #include <sys/dsl_dataset.h>
35 #include <sys/dsl_dir.h>
36 #include <sys/dsl_pool.h>
37 #include <sys/dsl_synctask.h>
38 #include <sys/zfs_ioctl.h>
39 #include <sys/zap.h>
40 #include <sys/zio_checksum.h>
41 
42 static char *dmu_recv_tag = "dmu_recv_tag";
43 
44 struct backuparg {
45 	dmu_replay_record_t *drr;
46 	vnode_t *vp;
47 	offset_t *off;
48 	objset_t *os;
49 	zio_cksum_t zc;
50 	int err;
51 };
52 
53 static int
54 dump_bytes(struct backuparg *ba, void *buf, int len)
55 {
56 	ssize_t resid; /* have to get resid to get detailed errno */
57 	ASSERT3U(len % 8, ==, 0);
58 
59 	fletcher_4_incremental_native(buf, len, &ba->zc);
60 	ba->err = vn_rdwr(UIO_WRITE, ba->vp,
61 	    (caddr_t)buf, len,
62 	    0, UIO_SYSSPACE, FAPPEND, RLIM64_INFINITY, CRED(), &resid);
63 	*ba->off += len;
64 	return (ba->err);
65 }
66 
67 static int
68 dump_free(struct backuparg *ba, uint64_t object, uint64_t offset,
69     uint64_t length)
70 {
71 	/* write a FREE record */
72 	bzero(ba->drr, sizeof (dmu_replay_record_t));
73 	ba->drr->drr_type = DRR_FREE;
74 	ba->drr->drr_u.drr_free.drr_object = object;
75 	ba->drr->drr_u.drr_free.drr_offset = offset;
76 	ba->drr->drr_u.drr_free.drr_length = length;
77 
78 	if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)))
79 		return (EINTR);
80 	return (0);
81 }
82 
83 static int
84 dump_data(struct backuparg *ba, dmu_object_type_t type,
85     uint64_t object, uint64_t offset, int blksz, void *data)
86 {
87 	/* write a DATA record */
88 	bzero(ba->drr, sizeof (dmu_replay_record_t));
89 	ba->drr->drr_type = DRR_WRITE;
90 	ba->drr->drr_u.drr_write.drr_object = object;
91 	ba->drr->drr_u.drr_write.drr_type = type;
92 	ba->drr->drr_u.drr_write.drr_offset = offset;
93 	ba->drr->drr_u.drr_write.drr_length = blksz;
94 
95 	if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)))
96 		return (EINTR);
97 	if (dump_bytes(ba, data, blksz))
98 		return (EINTR);
99 	return (0);
100 }
101 
102 static int
103 dump_freeobjects(struct backuparg *ba, uint64_t firstobj, uint64_t numobjs)
104 {
105 	/* write a FREEOBJECTS record */
106 	bzero(ba->drr, sizeof (dmu_replay_record_t));
107 	ba->drr->drr_type = DRR_FREEOBJECTS;
108 	ba->drr->drr_u.drr_freeobjects.drr_firstobj = firstobj;
109 	ba->drr->drr_u.drr_freeobjects.drr_numobjs = numobjs;
110 
111 	if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)))
112 		return (EINTR);
113 	return (0);
114 }
115 
116 static int
117 dump_dnode(struct backuparg *ba, uint64_t object, dnode_phys_t *dnp)
118 {
119 	if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
120 		return (dump_freeobjects(ba, object, 1));
121 
122 	/* write an OBJECT record */
123 	bzero(ba->drr, sizeof (dmu_replay_record_t));
124 	ba->drr->drr_type = DRR_OBJECT;
125 	ba->drr->drr_u.drr_object.drr_object = object;
126 	ba->drr->drr_u.drr_object.drr_type = dnp->dn_type;
127 	ba->drr->drr_u.drr_object.drr_bonustype = dnp->dn_bonustype;
128 	ba->drr->drr_u.drr_object.drr_blksz =
129 	    dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
130 	ba->drr->drr_u.drr_object.drr_bonuslen = dnp->dn_bonuslen;
131 	ba->drr->drr_u.drr_object.drr_checksum = dnp->dn_checksum;
132 	ba->drr->drr_u.drr_object.drr_compress = dnp->dn_compress;
133 
134 	if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)))
135 		return (EINTR);
136 
137 	if (dump_bytes(ba, DN_BONUS(dnp), P2ROUNDUP(dnp->dn_bonuslen, 8)))
138 		return (EINTR);
139 
140 	/* free anything past the end of the file */
141 	if (dump_free(ba, object, (dnp->dn_maxblkid + 1) *
142 	    (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL))
143 		return (EINTR);
144 	if (ba->err)
145 		return (EINTR);
146 	return (0);
147 }
148 
149 #define	BP_SPAN(dnp, level) \
150 	(((uint64_t)dnp->dn_datablkszsec) << (SPA_MINBLOCKSHIFT + \
151 	(level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)))
152 
153 /* ARGSUSED */
154 static int
155 backup_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
156     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
157 {
158 	struct backuparg *ba = arg;
159 	dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
160 	int err = 0;
161 
162 	if (issig(JUSTLOOKING) && issig(FORREAL))
163 		return (EINTR);
164 
165 	if (zb->zb_object != DMU_META_DNODE_OBJECT &&
166 	    DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
167 		return (0);
168 	} else if (bp == NULL && zb->zb_object == DMU_META_DNODE_OBJECT) {
169 		uint64_t span = BP_SPAN(dnp, zb->zb_level);
170 		uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
171 		err = dump_freeobjects(ba, dnobj, span >> DNODE_SHIFT);
172 	} else if (bp == NULL) {
173 		uint64_t span = BP_SPAN(dnp, zb->zb_level);
174 		err = dump_free(ba, zb->zb_object, zb->zb_blkid * span, span);
175 	} else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
176 		return (0);
177 	} else if (type == DMU_OT_DNODE) {
178 		dnode_phys_t *blk;
179 		int i;
180 		int blksz = BP_GET_LSIZE(bp);
181 		uint32_t aflags = ARC_WAIT;
182 		arc_buf_t *abuf;
183 
184 		if (arc_read_nolock(NULL, spa, bp,
185 		    arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
186 		    ZIO_FLAG_CANFAIL, &aflags, zb) != 0)
187 			return (EIO);
188 
189 		blk = abuf->b_data;
190 		for (i = 0; i < blksz >> DNODE_SHIFT; i++) {
191 			uint64_t dnobj = (zb->zb_blkid <<
192 			    (DNODE_BLOCK_SHIFT - DNODE_SHIFT)) + i;
193 			err = dump_dnode(ba, dnobj, blk+i);
194 			if (err)
195 				break;
196 		}
197 		(void) arc_buf_remove_ref(abuf, &abuf);
198 	} else { /* it's a level-0 block of a regular object */
199 		uint32_t aflags = ARC_WAIT;
200 		arc_buf_t *abuf;
201 		int blksz = BP_GET_LSIZE(bp);
202 
203 		if (arc_read_nolock(NULL, spa, bp,
204 		    arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
205 		    ZIO_FLAG_CANFAIL, &aflags, zb) != 0)
206 			return (EIO);
207 
208 		err = dump_data(ba, type, zb->zb_object, zb->zb_blkid * blksz,
209 		    blksz, abuf->b_data);
210 		(void) arc_buf_remove_ref(abuf, &abuf);
211 	}
212 
213 	ASSERT(err == 0 || err == EINTR);
214 	return (err);
215 }
216 
217 int
218 dmu_sendbackup(objset_t *tosnap, objset_t *fromsnap, boolean_t fromorigin,
219     vnode_t *vp, offset_t *off)
220 {
221 	dsl_dataset_t *ds = tosnap->os_dsl_dataset;
222 	dsl_dataset_t *fromds = fromsnap ? fromsnap->os_dsl_dataset : NULL;
223 	dmu_replay_record_t *drr;
224 	struct backuparg ba;
225 	int err;
226 	uint64_t fromtxg = 0;
227 
228 	/* tosnap must be a snapshot */
229 	if (ds->ds_phys->ds_next_snap_obj == 0)
230 		return (EINVAL);
231 
232 	/* fromsnap must be an earlier snapshot from the same fs as tosnap */
233 	if (fromds && (ds->ds_dir != fromds->ds_dir ||
234 	    fromds->ds_phys->ds_creation_txg >= ds->ds_phys->ds_creation_txg))
235 		return (EXDEV);
236 
237 	if (fromorigin) {
238 		dsl_pool_t *dp = ds->ds_dir->dd_pool;
239 
240 		if (fromsnap)
241 			return (EINVAL);
242 
243 		if (dsl_dir_is_clone(ds->ds_dir)) {
244 			rw_enter(&dp->dp_config_rwlock, RW_READER);
245 			err = dsl_dataset_hold_obj(dp,
246 			    ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &fromds);
247 			rw_exit(&dp->dp_config_rwlock);
248 			if (err)
249 				return (err);
250 		} else {
251 			fromorigin = B_FALSE;
252 		}
253 	}
254 
255 
256 	drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
257 	drr->drr_type = DRR_BEGIN;
258 	drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
259 	drr->drr_u.drr_begin.drr_version = DMU_BACKUP_STREAM_VERSION;
260 	drr->drr_u.drr_begin.drr_creation_time =
261 	    ds->ds_phys->ds_creation_time;
262 	drr->drr_u.drr_begin.drr_type = tosnap->os_phys->os_type;
263 	if (fromorigin)
264 		drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
265 	drr->drr_u.drr_begin.drr_toguid = ds->ds_phys->ds_guid;
266 	if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
267 		drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
268 
269 	if (fromds)
270 		drr->drr_u.drr_begin.drr_fromguid = fromds->ds_phys->ds_guid;
271 	dsl_dataset_name(ds, drr->drr_u.drr_begin.drr_toname);
272 
273 	if (fromds)
274 		fromtxg = fromds->ds_phys->ds_creation_txg;
275 	if (fromorigin)
276 		dsl_dataset_rele(fromds, FTAG);
277 
278 	ba.drr = drr;
279 	ba.vp = vp;
280 	ba.os = tosnap;
281 	ba.off = off;
282 	ZIO_SET_CHECKSUM(&ba.zc, 0, 0, 0, 0);
283 
284 	if (dump_bytes(&ba, drr, sizeof (dmu_replay_record_t))) {
285 		kmem_free(drr, sizeof (dmu_replay_record_t));
286 		return (ba.err);
287 	}
288 
289 	err = traverse_dataset(ds, fromtxg, TRAVERSE_PRE | TRAVERSE_PREFETCH,
290 	    backup_cb, &ba);
291 
292 	if (err) {
293 		if (err == EINTR && ba.err)
294 			err = ba.err;
295 		kmem_free(drr, sizeof (dmu_replay_record_t));
296 		return (err);
297 	}
298 
299 	bzero(drr, sizeof (dmu_replay_record_t));
300 	drr->drr_type = DRR_END;
301 	drr->drr_u.drr_end.drr_checksum = ba.zc;
302 
303 	if (dump_bytes(&ba, drr, sizeof (dmu_replay_record_t))) {
304 		kmem_free(drr, sizeof (dmu_replay_record_t));
305 		return (ba.err);
306 	}
307 
308 	kmem_free(drr, sizeof (dmu_replay_record_t));
309 
310 	return (0);
311 }
312 
313 struct recvbeginsyncarg {
314 	const char *tofs;
315 	const char *tosnap;
316 	dsl_dataset_t *origin;
317 	uint64_t fromguid;
318 	dmu_objset_type_t type;
319 	void *tag;
320 	boolean_t force;
321 	uint64_t dsflags;
322 	char clonelastname[MAXNAMELEN];
323 	dsl_dataset_t *ds; /* the ds to recv into; returned from the syncfunc */
324 };
325 
326 /* ARGSUSED */
327 static int
328 recv_new_check(void *arg1, void *arg2, dmu_tx_t *tx)
329 {
330 	dsl_dir_t *dd = arg1;
331 	struct recvbeginsyncarg *rbsa = arg2;
332 	objset_t *mos = dd->dd_pool->dp_meta_objset;
333 	uint64_t val;
334 	int err;
335 
336 	err = zap_lookup(mos, dd->dd_phys->dd_child_dir_zapobj,
337 	    strrchr(rbsa->tofs, '/') + 1, sizeof (uint64_t), 1, &val);
338 
339 	if (err != ENOENT)
340 		return (err ? err : EEXIST);
341 
342 	if (rbsa->origin) {
343 		/* make sure it's a snap in the same pool */
344 		if (rbsa->origin->ds_dir->dd_pool != dd->dd_pool)
345 			return (EXDEV);
346 		if (!dsl_dataset_is_snapshot(rbsa->origin))
347 			return (EINVAL);
348 		if (rbsa->origin->ds_phys->ds_guid != rbsa->fromguid)
349 			return (ENODEV);
350 	}
351 
352 	return (0);
353 }
354 
355 static void
356 recv_new_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
357 {
358 	dsl_dir_t *dd = arg1;
359 	struct recvbeginsyncarg *rbsa = arg2;
360 	uint64_t flags = DS_FLAG_INCONSISTENT | rbsa->dsflags;
361 	uint64_t dsobj;
362 
363 	/* Create and open new dataset. */
364 	dsobj = dsl_dataset_create_sync(dd, strrchr(rbsa->tofs, '/') + 1,
365 	    rbsa->origin, flags, cr, tx);
366 	VERIFY(0 == dsl_dataset_own_obj(dd->dd_pool, dsobj,
367 	    B_TRUE, dmu_recv_tag, &rbsa->ds));
368 
369 	if (rbsa->origin == NULL) {
370 		(void) dmu_objset_create_impl(dd->dd_pool->dp_spa,
371 		    rbsa->ds, &rbsa->ds->ds_phys->ds_bp, rbsa->type, tx);
372 	}
373 
374 	spa_history_internal_log(LOG_DS_REPLAY_FULL_SYNC,
375 	    dd->dd_pool->dp_spa, tx, cr, "dataset = %lld", dsobj);
376 }
377 
378 /* ARGSUSED */
379 static int
380 recv_existing_check(void *arg1, void *arg2, dmu_tx_t *tx)
381 {
382 	dsl_dataset_t *ds = arg1;
383 	struct recvbeginsyncarg *rbsa = arg2;
384 	int err;
385 	uint64_t val;
386 
387 	/* must not have any changes since most recent snapshot */
388 	if (!rbsa->force && dsl_dataset_modified_since_lastsnap(ds))
389 		return (ETXTBSY);
390 
391 	if (rbsa->fromguid) {
392 		/* if incremental, most recent snapshot must match fromguid */
393 		if (ds->ds_prev == NULL)
394 			return (ENODEV);
395 		if (ds->ds_prev->ds_phys->ds_guid != rbsa->fromguid)
396 			return (ENODEV);
397 	} else {
398 		/* if full, most recent snapshot must be $ORIGIN */
399 		if (ds->ds_phys->ds_prev_snap_txg >= TXG_INITIAL)
400 			return (ENODEV);
401 	}
402 
403 	/* temporary clone name must not exist */
404 	err = zap_lookup(ds->ds_dir->dd_pool->dp_meta_objset,
405 	    ds->ds_dir->dd_phys->dd_child_dir_zapobj,
406 	    rbsa->clonelastname, 8, 1, &val);
407 	if (err == 0)
408 		return (EEXIST);
409 	if (err != ENOENT)
410 		return (err);
411 
412 	/* new snapshot name must not exist */
413 	err = zap_lookup(ds->ds_dir->dd_pool->dp_meta_objset,
414 	    ds->ds_phys->ds_snapnames_zapobj, rbsa->tosnap, 8, 1, &val);
415 	if (err == 0)
416 		return (EEXIST);
417 	if (err != ENOENT)
418 		return (err);
419 	return (0);
420 }
421 
422 /* ARGSUSED */
423 static void
424 recv_existing_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
425 {
426 	dsl_dataset_t *ohds = arg1;
427 	struct recvbeginsyncarg *rbsa = arg2;
428 	dsl_pool_t *dp = ohds->ds_dir->dd_pool;
429 	dsl_dataset_t *cds;
430 	uint64_t flags = DS_FLAG_INCONSISTENT | rbsa->dsflags;
431 	uint64_t dsobj;
432 
433 	/* create and open the temporary clone */
434 	dsobj = dsl_dataset_create_sync(ohds->ds_dir, rbsa->clonelastname,
435 	    ohds->ds_prev, flags, cr, tx);
436 	VERIFY(0 == dsl_dataset_own_obj(dp, dsobj, B_TRUE, dmu_recv_tag, &cds));
437 
438 	/*
439 	 * If we actually created a non-clone, we need to create the
440 	 * objset in our new dataset.
441 	 */
442 	if (BP_IS_HOLE(dsl_dataset_get_blkptr(cds))) {
443 		(void) dmu_objset_create_impl(dp->dp_spa,
444 		    cds, dsl_dataset_get_blkptr(cds), rbsa->type, tx);
445 	}
446 
447 	/* copy the refquota from the target fs to the clone */
448 	if (ohds->ds_quota > 0)
449 		dsl_dataset_set_quota_sync(cds, &ohds->ds_quota, cr, tx);
450 
451 	rbsa->ds = cds;
452 
453 	spa_history_internal_log(LOG_DS_REPLAY_INC_SYNC,
454 	    dp->dp_spa, tx, cr, "dataset = %lld", dsobj);
455 }
456 
457 /*
458  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
459  * succeeds; otherwise we will leak the holds on the datasets.
460  */
461 int
462 dmu_recv_begin(char *tofs, char *tosnap, struct drr_begin *drrb,
463     boolean_t force, objset_t *origin, dmu_recv_cookie_t *drc)
464 {
465 	int err = 0;
466 	boolean_t byteswap;
467 	struct recvbeginsyncarg rbsa = { 0 };
468 	uint64_t version;
469 	int flags;
470 	dsl_dataset_t *ds;
471 
472 	if (drrb->drr_magic == DMU_BACKUP_MAGIC)
473 		byteswap = FALSE;
474 	else if (drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
475 		byteswap = TRUE;
476 	else
477 		return (EINVAL);
478 
479 	rbsa.tofs = tofs;
480 	rbsa.tosnap = tosnap;
481 	rbsa.origin = origin ? origin->os_dsl_dataset : NULL;
482 	rbsa.fromguid = drrb->drr_fromguid;
483 	rbsa.type = drrb->drr_type;
484 	rbsa.tag = FTAG;
485 	rbsa.dsflags = 0;
486 	version = drrb->drr_version;
487 	flags = drrb->drr_flags;
488 
489 	if (byteswap) {
490 		rbsa.type = BSWAP_32(rbsa.type);
491 		rbsa.fromguid = BSWAP_64(rbsa.fromguid);
492 		version = BSWAP_64(version);
493 		flags = BSWAP_32(flags);
494 	}
495 
496 	if (version != DMU_BACKUP_STREAM_VERSION ||
497 	    rbsa.type >= DMU_OST_NUMTYPES ||
498 	    ((flags & DRR_FLAG_CLONE) && origin == NULL))
499 		return (EINVAL);
500 
501 	if (flags & DRR_FLAG_CI_DATA)
502 		rbsa.dsflags = DS_FLAG_CI_DATASET;
503 
504 	bzero(drc, sizeof (dmu_recv_cookie_t));
505 	drc->drc_drrb = drrb;
506 	drc->drc_tosnap = tosnap;
507 	drc->drc_force = force;
508 
509 	/*
510 	 * Process the begin in syncing context.
511 	 */
512 
513 	/* open the dataset we are logically receiving into */
514 	err = dsl_dataset_hold(tofs, dmu_recv_tag, &ds);
515 	if (err == 0) {
516 		/* target fs already exists; recv into temp clone */
517 
518 		/* Can't recv a clone into an existing fs */
519 		if (flags & DRR_FLAG_CLONE) {
520 			dsl_dataset_rele(ds, dmu_recv_tag);
521 			return (EINVAL);
522 		}
523 
524 		/* must not have an incremental recv already in progress */
525 		if (!mutex_tryenter(&ds->ds_recvlock)) {
526 			dsl_dataset_rele(ds, dmu_recv_tag);
527 			return (EBUSY);
528 		}
529 
530 		/* tmp clone name is: tofs/%tosnap" */
531 		(void) snprintf(rbsa.clonelastname, sizeof (rbsa.clonelastname),
532 		    "%%%s", tosnap);
533 		rbsa.force = force;
534 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
535 		    recv_existing_check, recv_existing_sync, ds, &rbsa, 5);
536 		if (err) {
537 			mutex_exit(&ds->ds_recvlock);
538 			dsl_dataset_rele(ds, dmu_recv_tag);
539 			return (err);
540 		}
541 		drc->drc_logical_ds = ds;
542 		drc->drc_real_ds = rbsa.ds;
543 	} else if (err == ENOENT) {
544 		/* target fs does not exist; must be a full backup or clone */
545 		char *cp;
546 
547 		/*
548 		 * If it's a non-clone incremental, we are missing the
549 		 * target fs, so fail the recv.
550 		 */
551 		if (rbsa.fromguid && !(flags & DRR_FLAG_CLONE))
552 			return (ENOENT);
553 
554 		/* Open the parent of tofs */
555 		cp = strrchr(tofs, '/');
556 		*cp = '\0';
557 		err = dsl_dataset_hold(tofs, FTAG, &ds);
558 		*cp = '/';
559 		if (err)
560 			return (err);
561 
562 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
563 		    recv_new_check, recv_new_sync, ds->ds_dir, &rbsa, 5);
564 		dsl_dataset_rele(ds, FTAG);
565 		if (err)
566 			return (err);
567 		drc->drc_logical_ds = drc->drc_real_ds = rbsa.ds;
568 		drc->drc_newfs = B_TRUE;
569 	}
570 
571 	return (err);
572 }
573 
574 struct restorearg {
575 	int err;
576 	int byteswap;
577 	vnode_t *vp;
578 	char *buf;
579 	uint64_t voff;
580 	int bufsize; /* amount of memory allocated for buf */
581 	zio_cksum_t cksum;
582 };
583 
584 static void *
585 restore_read(struct restorearg *ra, int len)
586 {
587 	void *rv;
588 	int done = 0;
589 
590 	/* some things will require 8-byte alignment, so everything must */
591 	ASSERT3U(len % 8, ==, 0);
592 
593 	while (done < len) {
594 		ssize_t resid;
595 
596 		ra->err = vn_rdwr(UIO_READ, ra->vp,
597 		    (caddr_t)ra->buf + done, len - done,
598 		    ra->voff, UIO_SYSSPACE, FAPPEND,
599 		    RLIM64_INFINITY, CRED(), &resid);
600 
601 		if (resid == len - done)
602 			ra->err = EINVAL;
603 		ra->voff += len - done - resid;
604 		done = len - resid;
605 		if (ra->err)
606 			return (NULL);
607 	}
608 
609 	ASSERT3U(done, ==, len);
610 	rv = ra->buf;
611 	if (ra->byteswap)
612 		fletcher_4_incremental_byteswap(rv, len, &ra->cksum);
613 	else
614 		fletcher_4_incremental_native(rv, len, &ra->cksum);
615 	return (rv);
616 }
617 
618 static void
619 backup_byteswap(dmu_replay_record_t *drr)
620 {
621 #define	DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
622 #define	DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
623 	drr->drr_type = BSWAP_32(drr->drr_type);
624 	drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
625 	switch (drr->drr_type) {
626 	case DRR_BEGIN:
627 		DO64(drr_begin.drr_magic);
628 		DO64(drr_begin.drr_version);
629 		DO64(drr_begin.drr_creation_time);
630 		DO32(drr_begin.drr_type);
631 		DO32(drr_begin.drr_flags);
632 		DO64(drr_begin.drr_toguid);
633 		DO64(drr_begin.drr_fromguid);
634 		break;
635 	case DRR_OBJECT:
636 		DO64(drr_object.drr_object);
637 		/* DO64(drr_object.drr_allocation_txg); */
638 		DO32(drr_object.drr_type);
639 		DO32(drr_object.drr_bonustype);
640 		DO32(drr_object.drr_blksz);
641 		DO32(drr_object.drr_bonuslen);
642 		break;
643 	case DRR_FREEOBJECTS:
644 		DO64(drr_freeobjects.drr_firstobj);
645 		DO64(drr_freeobjects.drr_numobjs);
646 		break;
647 	case DRR_WRITE:
648 		DO64(drr_write.drr_object);
649 		DO32(drr_write.drr_type);
650 		DO64(drr_write.drr_offset);
651 		DO64(drr_write.drr_length);
652 		break;
653 	case DRR_FREE:
654 		DO64(drr_free.drr_object);
655 		DO64(drr_free.drr_offset);
656 		DO64(drr_free.drr_length);
657 		break;
658 	case DRR_END:
659 		DO64(drr_end.drr_checksum.zc_word[0]);
660 		DO64(drr_end.drr_checksum.zc_word[1]);
661 		DO64(drr_end.drr_checksum.zc_word[2]);
662 		DO64(drr_end.drr_checksum.zc_word[3]);
663 		break;
664 	}
665 #undef DO64
666 #undef DO32
667 }
668 
669 static int
670 restore_object(struct restorearg *ra, objset_t *os, struct drr_object *drro)
671 {
672 	int err;
673 	dmu_tx_t *tx;
674 	void *data = NULL;
675 
676 	if (drro->drr_type == DMU_OT_NONE ||
677 	    drro->drr_type >= DMU_OT_NUMTYPES ||
678 	    drro->drr_bonustype >= DMU_OT_NUMTYPES ||
679 	    drro->drr_checksum >= ZIO_CHECKSUM_FUNCTIONS ||
680 	    drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
681 	    P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
682 	    drro->drr_blksz < SPA_MINBLOCKSIZE ||
683 	    drro->drr_blksz > SPA_MAXBLOCKSIZE ||
684 	    drro->drr_bonuslen > DN_MAX_BONUSLEN) {
685 		return (EINVAL);
686 	}
687 
688 	err = dmu_object_info(os, drro->drr_object, NULL);
689 
690 	if (err != 0 && err != ENOENT)
691 		return (EINVAL);
692 
693 	if (drro->drr_bonuslen) {
694 		data = restore_read(ra, P2ROUNDUP(drro->drr_bonuslen, 8));
695 		if (ra->err)
696 			return (ra->err);
697 	}
698 
699 	if (err == ENOENT) {
700 		/* currently free, want to be allocated */
701 		tx = dmu_tx_create(os);
702 		dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
703 		err = dmu_tx_assign(tx, TXG_WAIT);
704 		if (err) {
705 			dmu_tx_abort(tx);
706 			return (err);
707 		}
708 		err = dmu_object_claim(os, drro->drr_object,
709 		    drro->drr_type, drro->drr_blksz,
710 		    drro->drr_bonustype, drro->drr_bonuslen, tx);
711 		dmu_tx_commit(tx);
712 	} else {
713 		/* currently allocated, want to be allocated */
714 		err = dmu_object_reclaim(os, drro->drr_object,
715 		    drro->drr_type, drro->drr_blksz,
716 		    drro->drr_bonustype, drro->drr_bonuslen);
717 	}
718 	if (err)
719 		return (EINVAL);
720 
721 	tx = dmu_tx_create(os);
722 	dmu_tx_hold_bonus(tx, drro->drr_object);
723 	err = dmu_tx_assign(tx, TXG_WAIT);
724 	if (err) {
725 		dmu_tx_abort(tx);
726 		return (err);
727 	}
728 
729 	dmu_object_set_checksum(os, drro->drr_object, drro->drr_checksum, tx);
730 	dmu_object_set_compress(os, drro->drr_object, drro->drr_compress, tx);
731 
732 	if (data != NULL) {
733 		dmu_buf_t *db;
734 
735 		VERIFY(0 == dmu_bonus_hold(os, drro->drr_object, FTAG, &db));
736 		dmu_buf_will_dirty(db, tx);
737 
738 		ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
739 		bcopy(data, db->db_data, drro->drr_bonuslen);
740 		if (ra->byteswap) {
741 			dmu_ot[drro->drr_bonustype].ot_byteswap(db->db_data,
742 			    drro->drr_bonuslen);
743 		}
744 		dmu_buf_rele(db, FTAG);
745 	}
746 	dmu_tx_commit(tx);
747 	return (0);
748 }
749 
750 /* ARGSUSED */
751 static int
752 restore_freeobjects(struct restorearg *ra, objset_t *os,
753     struct drr_freeobjects *drrfo)
754 {
755 	uint64_t obj;
756 
757 	if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
758 		return (EINVAL);
759 
760 	for (obj = drrfo->drr_firstobj;
761 	    obj < drrfo->drr_firstobj + drrfo->drr_numobjs;
762 	    (void) dmu_object_next(os, &obj, FALSE, 0)) {
763 		int err;
764 
765 		if (dmu_object_info(os, obj, NULL) != 0)
766 			continue;
767 
768 		err = dmu_free_object(os, obj);
769 		if (err)
770 			return (err);
771 	}
772 	return (0);
773 }
774 
775 static int
776 restore_write(struct restorearg *ra, objset_t *os,
777     struct drr_write *drrw)
778 {
779 	dmu_tx_t *tx;
780 	void *data;
781 	int err;
782 
783 	if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
784 	    drrw->drr_type >= DMU_OT_NUMTYPES)
785 		return (EINVAL);
786 
787 	data = restore_read(ra, drrw->drr_length);
788 	if (data == NULL)
789 		return (ra->err);
790 
791 	if (dmu_object_info(os, drrw->drr_object, NULL) != 0)
792 		return (EINVAL);
793 
794 	tx = dmu_tx_create(os);
795 
796 	dmu_tx_hold_write(tx, drrw->drr_object,
797 	    drrw->drr_offset, drrw->drr_length);
798 	err = dmu_tx_assign(tx, TXG_WAIT);
799 	if (err) {
800 		dmu_tx_abort(tx);
801 		return (err);
802 	}
803 	if (ra->byteswap)
804 		dmu_ot[drrw->drr_type].ot_byteswap(data, drrw->drr_length);
805 	dmu_write(os, drrw->drr_object,
806 	    drrw->drr_offset, drrw->drr_length, data, tx);
807 	dmu_tx_commit(tx);
808 	return (0);
809 }
810 
811 /* ARGSUSED */
812 static int
813 restore_free(struct restorearg *ra, objset_t *os,
814     struct drr_free *drrf)
815 {
816 	int err;
817 
818 	if (drrf->drr_length != -1ULL &&
819 	    drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
820 		return (EINVAL);
821 
822 	if (dmu_object_info(os, drrf->drr_object, NULL) != 0)
823 		return (EINVAL);
824 
825 	err = dmu_free_long_range(os, drrf->drr_object,
826 	    drrf->drr_offset, drrf->drr_length);
827 	return (err);
828 }
829 
830 /*
831  * NB: callers *must* call dmu_recv_end() if this succeeds.
832  */
833 int
834 dmu_recv_stream(dmu_recv_cookie_t *drc, vnode_t *vp, offset_t *voffp)
835 {
836 	struct restorearg ra = { 0 };
837 	dmu_replay_record_t *drr;
838 	objset_t *os;
839 	zio_cksum_t pcksum;
840 
841 	if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
842 		ra.byteswap = TRUE;
843 
844 	{
845 		/* compute checksum of drr_begin record */
846 		dmu_replay_record_t *drr;
847 		drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
848 
849 		drr->drr_type = DRR_BEGIN;
850 		drr->drr_u.drr_begin = *drc->drc_drrb;
851 		if (ra.byteswap) {
852 			fletcher_4_incremental_byteswap(drr,
853 			    sizeof (dmu_replay_record_t), &ra.cksum);
854 		} else {
855 			fletcher_4_incremental_native(drr,
856 			    sizeof (dmu_replay_record_t), &ra.cksum);
857 		}
858 		kmem_free(drr, sizeof (dmu_replay_record_t));
859 	}
860 
861 	if (ra.byteswap) {
862 		struct drr_begin *drrb = drc->drc_drrb;
863 		drrb->drr_magic = BSWAP_64(drrb->drr_magic);
864 		drrb->drr_version = BSWAP_64(drrb->drr_version);
865 		drrb->drr_creation_time = BSWAP_64(drrb->drr_creation_time);
866 		drrb->drr_type = BSWAP_32(drrb->drr_type);
867 		drrb->drr_toguid = BSWAP_64(drrb->drr_toguid);
868 		drrb->drr_fromguid = BSWAP_64(drrb->drr_fromguid);
869 	}
870 
871 	ra.vp = vp;
872 	ra.voff = *voffp;
873 	ra.bufsize = 1<<20;
874 	ra.buf = kmem_alloc(ra.bufsize, KM_SLEEP);
875 
876 	/* these were verified in dmu_recv_begin */
877 	ASSERT(drc->drc_drrb->drr_version == DMU_BACKUP_STREAM_VERSION);
878 	ASSERT(drc->drc_drrb->drr_type < DMU_OST_NUMTYPES);
879 
880 	/*
881 	 * Open the objset we are modifying.
882 	 */
883 	VERIFY(dmu_objset_from_ds(drc->drc_real_ds, &os) == 0);
884 
885 	ASSERT(drc->drc_real_ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT);
886 
887 	/*
888 	 * Read records and process them.
889 	 */
890 	pcksum = ra.cksum;
891 	while (ra.err == 0 &&
892 	    NULL != (drr = restore_read(&ra, sizeof (*drr)))) {
893 		if (issig(JUSTLOOKING) && issig(FORREAL)) {
894 			ra.err = EINTR;
895 			goto out;
896 		}
897 
898 		if (ra.byteswap)
899 			backup_byteswap(drr);
900 
901 		switch (drr->drr_type) {
902 		case DRR_OBJECT:
903 		{
904 			/*
905 			 * We need to make a copy of the record header,
906 			 * because restore_{object,write} may need to
907 			 * restore_read(), which will invalidate drr.
908 			 */
909 			struct drr_object drro = drr->drr_u.drr_object;
910 			ra.err = restore_object(&ra, os, &drro);
911 			break;
912 		}
913 		case DRR_FREEOBJECTS:
914 		{
915 			struct drr_freeobjects drrfo =
916 			    drr->drr_u.drr_freeobjects;
917 			ra.err = restore_freeobjects(&ra, os, &drrfo);
918 			break;
919 		}
920 		case DRR_WRITE:
921 		{
922 			struct drr_write drrw = drr->drr_u.drr_write;
923 			ra.err = restore_write(&ra, os, &drrw);
924 			break;
925 		}
926 		case DRR_FREE:
927 		{
928 			struct drr_free drrf = drr->drr_u.drr_free;
929 			ra.err = restore_free(&ra, os, &drrf);
930 			break;
931 		}
932 		case DRR_END:
933 		{
934 			struct drr_end drre = drr->drr_u.drr_end;
935 			/*
936 			 * We compare against the *previous* checksum
937 			 * value, because the stored checksum is of
938 			 * everything before the DRR_END record.
939 			 */
940 			if (!ZIO_CHECKSUM_EQUAL(drre.drr_checksum, pcksum))
941 				ra.err = ECKSUM;
942 			goto out;
943 		}
944 		default:
945 			ra.err = EINVAL;
946 			goto out;
947 		}
948 		pcksum = ra.cksum;
949 	}
950 	ASSERT(ra.err != 0);
951 
952 out:
953 	if (ra.err != 0) {
954 		/*
955 		 * destroy what we created, so we don't leave it in the
956 		 * inconsistent restoring state.
957 		 */
958 		txg_wait_synced(drc->drc_real_ds->ds_dir->dd_pool, 0);
959 
960 		(void) dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag,
961 		    B_FALSE);
962 		if (drc->drc_real_ds != drc->drc_logical_ds) {
963 			mutex_exit(&drc->drc_logical_ds->ds_recvlock);
964 			dsl_dataset_rele(drc->drc_logical_ds, dmu_recv_tag);
965 		}
966 	}
967 
968 	kmem_free(ra.buf, ra.bufsize);
969 	*voffp = ra.voff;
970 	return (ra.err);
971 }
972 
973 struct recvendsyncarg {
974 	char *tosnap;
975 	uint64_t creation_time;
976 	uint64_t toguid;
977 };
978 
979 static int
980 recv_end_check(void *arg1, void *arg2, dmu_tx_t *tx)
981 {
982 	dsl_dataset_t *ds = arg1;
983 	struct recvendsyncarg *resa = arg2;
984 
985 	return (dsl_dataset_snapshot_check(ds, resa->tosnap, tx));
986 }
987 
988 static void
989 recv_end_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
990 {
991 	dsl_dataset_t *ds = arg1;
992 	struct recvendsyncarg *resa = arg2;
993 
994 	dsl_dataset_snapshot_sync(ds, resa->tosnap, cr, tx);
995 
996 	/* set snapshot's creation time and guid */
997 	dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
998 	ds->ds_prev->ds_phys->ds_creation_time = resa->creation_time;
999 	ds->ds_prev->ds_phys->ds_guid = resa->toguid;
1000 	ds->ds_prev->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1001 
1002 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1003 	ds->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1004 }
1005 
1006 static int
1007 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
1008 {
1009 	struct recvendsyncarg resa;
1010 	dsl_dataset_t *ds = drc->drc_logical_ds;
1011 	int err;
1012 
1013 	/*
1014 	 * XXX hack; seems the ds is still dirty and dsl_pool_zil_clean()
1015 	 * expects it to have a ds_user_ptr (and zil), but clone_swap()
1016 	 * can close it.
1017 	 */
1018 	txg_wait_synced(ds->ds_dir->dd_pool, 0);
1019 
1020 	if (dsl_dataset_tryown(ds, FALSE, dmu_recv_tag)) {
1021 		err = dsl_dataset_clone_swap(drc->drc_real_ds, ds,
1022 		    drc->drc_force);
1023 		if (err)
1024 			goto out;
1025 	} else {
1026 		mutex_exit(&ds->ds_recvlock);
1027 		dsl_dataset_rele(ds, dmu_recv_tag);
1028 		(void) dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag,
1029 		    B_FALSE);
1030 		return (EBUSY);
1031 	}
1032 
1033 	resa.creation_time = drc->drc_drrb->drr_creation_time;
1034 	resa.toguid = drc->drc_drrb->drr_toguid;
1035 	resa.tosnap = drc->drc_tosnap;
1036 
1037 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1038 	    recv_end_check, recv_end_sync, ds, &resa, 3);
1039 	if (err) {
1040 		/* swap back */
1041 		(void) dsl_dataset_clone_swap(drc->drc_real_ds, ds, B_TRUE);
1042 	}
1043 
1044 out:
1045 	mutex_exit(&ds->ds_recvlock);
1046 	dsl_dataset_disown(ds, dmu_recv_tag);
1047 	(void) dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag, B_FALSE);
1048 	return (err);
1049 }
1050 
1051 static int
1052 dmu_recv_new_end(dmu_recv_cookie_t *drc)
1053 {
1054 	struct recvendsyncarg resa;
1055 	dsl_dataset_t *ds = drc->drc_logical_ds;
1056 	int err;
1057 
1058 	/*
1059 	 * XXX hack; seems the ds is still dirty and dsl_pool_zil_clean()
1060 	 * expects it to have a ds_user_ptr (and zil), but clone_swap()
1061 	 * can close it.
1062 	 */
1063 	txg_wait_synced(ds->ds_dir->dd_pool, 0);
1064 
1065 	resa.creation_time = drc->drc_drrb->drr_creation_time;
1066 	resa.toguid = drc->drc_drrb->drr_toguid;
1067 	resa.tosnap = drc->drc_tosnap;
1068 
1069 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1070 	    recv_end_check, recv_end_sync, ds, &resa, 3);
1071 	if (err) {
1072 		/* clean up the fs we just recv'd into */
1073 		(void) dsl_dataset_destroy(ds, dmu_recv_tag, B_FALSE);
1074 	} else {
1075 		/* release the hold from dmu_recv_begin */
1076 		dsl_dataset_disown(ds, dmu_recv_tag);
1077 	}
1078 	return (err);
1079 }
1080 
1081 int
1082 dmu_recv_end(dmu_recv_cookie_t *drc)
1083 {
1084 	if (drc->drc_logical_ds != drc->drc_real_ds)
1085 		return (dmu_recv_existing_end(drc));
1086 	else
1087 		return (dmu_recv_new_end(drc));
1088 }
1089