xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu_send.c (revision 4a3b1d5b615ff6e54da1cc17f331e1ac794c5191)
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 (c) 2011 by Delphix. All rights reserved.
24  */
25 /*
26  * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
27  * Copyright (c) 2011 by Delphix. All rights reserved.
28  */
29 
30 #include <sys/dmu.h>
31 #include <sys/dmu_impl.h>
32 #include <sys/dmu_tx.h>
33 #include <sys/dbuf.h>
34 #include <sys/dnode.h>
35 #include <sys/zfs_context.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/dmu_traverse.h>
38 #include <sys/dsl_dataset.h>
39 #include <sys/dsl_dir.h>
40 #include <sys/dsl_prop.h>
41 #include <sys/dsl_pool.h>
42 #include <sys/dsl_synctask.h>
43 #include <sys/zfs_ioctl.h>
44 #include <sys/zap.h>
45 #include <sys/zio_checksum.h>
46 #include <sys/zfs_znode.h>
47 #include <zfs_fletcher.h>
48 #include <sys/avl.h>
49 #include <sys/ddt.h>
50 #include <sys/zfs_onexit.h>
51 
52 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
53 int zfs_send_corrupt_data = B_FALSE;
54 
55 static char *dmu_recv_tag = "dmu_recv_tag";
56 
57 /*
58  * The list of data whose inclusion in a send stream can be pending from
59  * one call to backup_cb to another.  Multiple calls to dump_free() and
60  * dump_freeobjects() can be aggregated into a single DRR_FREE or
61  * DRR_FREEOBJECTS replay record.
62  */
63 typedef enum {
64 	PENDING_NONE,
65 	PENDING_FREE,
66 	PENDING_FREEOBJECTS
67 } pendop_t;
68 
69 struct backuparg {
70 	dmu_replay_record_t *drr;
71 	vnode_t *vp;
72 	offset_t *off;
73 	objset_t *os;
74 	zio_cksum_t zc;
75 	uint64_t toguid;
76 	int err;
77 	pendop_t pending_op;
78 };
79 
80 static int
81 dump_bytes(struct backuparg *ba, void *buf, int len)
82 {
83 	ssize_t resid; /* have to get resid to get detailed errno */
84 	ASSERT3U(len % 8, ==, 0);
85 
86 	fletcher_4_incremental_native(buf, len, &ba->zc);
87 	ba->err = vn_rdwr(UIO_WRITE, ba->vp,
88 	    (caddr_t)buf, len,
89 	    0, UIO_SYSSPACE, FAPPEND, RLIM64_INFINITY, CRED(), &resid);
90 	*ba->off += len;
91 	return (ba->err);
92 }
93 
94 static int
95 dump_free(struct backuparg *ba, uint64_t object, uint64_t offset,
96     uint64_t length)
97 {
98 	struct drr_free *drrf = &(ba->drr->drr_u.drr_free);
99 
100 	/*
101 	 * If there is a pending op, but it's not PENDING_FREE, push it out,
102 	 * since free block aggregation can only be done for blocks of the
103 	 * same type (i.e., DRR_FREE records can only be aggregated with
104 	 * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
105 	 * aggregated with other DRR_FREEOBJECTS records.
106 	 */
107 	if (ba->pending_op != PENDING_NONE && ba->pending_op != PENDING_FREE) {
108 		if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
109 			return (EINTR);
110 		ba->pending_op = PENDING_NONE;
111 	}
112 
113 	if (ba->pending_op == PENDING_FREE) {
114 		/*
115 		 * There should never be a PENDING_FREE if length is -1
116 		 * (because dump_dnode is the only place where this
117 		 * function is called with a -1, and only after flushing
118 		 * any pending record).
119 		 */
120 		ASSERT(length != -1ULL);
121 		/*
122 		 * Check to see whether this free block can be aggregated
123 		 * with pending one.
124 		 */
125 		if (drrf->drr_object == object && drrf->drr_offset +
126 		    drrf->drr_length == offset) {
127 			drrf->drr_length += length;
128 			return (0);
129 		} else {
130 			/* not a continuation.  Push out pending record */
131 			if (dump_bytes(ba, ba->drr,
132 			    sizeof (dmu_replay_record_t)) != 0)
133 				return (EINTR);
134 			ba->pending_op = PENDING_NONE;
135 		}
136 	}
137 	/* create a FREE record and make it pending */
138 	bzero(ba->drr, sizeof (dmu_replay_record_t));
139 	ba->drr->drr_type = DRR_FREE;
140 	drrf->drr_object = object;
141 	drrf->drr_offset = offset;
142 	drrf->drr_length = length;
143 	drrf->drr_toguid = ba->toguid;
144 	if (length == -1ULL) {
145 		if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
146 			return (EINTR);
147 	} else {
148 		ba->pending_op = PENDING_FREE;
149 	}
150 
151 	return (0);
152 }
153 
154 static int
155 dump_data(struct backuparg *ba, dmu_object_type_t type,
156     uint64_t object, uint64_t offset, int blksz, const blkptr_t *bp, void *data)
157 {
158 	struct drr_write *drrw = &(ba->drr->drr_u.drr_write);
159 
160 
161 	/*
162 	 * If there is any kind of pending aggregation (currently either
163 	 * a grouping of free objects or free blocks), push it out to
164 	 * the stream, since aggregation can't be done across operations
165 	 * of different types.
166 	 */
167 	if (ba->pending_op != PENDING_NONE) {
168 		if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
169 			return (EINTR);
170 		ba->pending_op = PENDING_NONE;
171 	}
172 	/* write a DATA record */
173 	bzero(ba->drr, sizeof (dmu_replay_record_t));
174 	ba->drr->drr_type = DRR_WRITE;
175 	drrw->drr_object = object;
176 	drrw->drr_type = type;
177 	drrw->drr_offset = offset;
178 	drrw->drr_length = blksz;
179 	drrw->drr_toguid = ba->toguid;
180 	drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
181 	if (zio_checksum_table[drrw->drr_checksumtype].ci_dedup)
182 		drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
183 	DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
184 	DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
185 	DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
186 	drrw->drr_key.ddk_cksum = bp->blk_cksum;
187 
188 	if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
189 		return (EINTR);
190 	if (dump_bytes(ba, data, blksz) != 0)
191 		return (EINTR);
192 	return (0);
193 }
194 
195 static int
196 dump_spill(struct backuparg *ba, uint64_t object, int blksz, void *data)
197 {
198 	struct drr_spill *drrs = &(ba->drr->drr_u.drr_spill);
199 
200 	if (ba->pending_op != PENDING_NONE) {
201 		if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
202 			return (EINTR);
203 		ba->pending_op = PENDING_NONE;
204 	}
205 
206 	/* write a SPILL record */
207 	bzero(ba->drr, sizeof (dmu_replay_record_t));
208 	ba->drr->drr_type = DRR_SPILL;
209 	drrs->drr_object = object;
210 	drrs->drr_length = blksz;
211 	drrs->drr_toguid = ba->toguid;
212 
213 	if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)))
214 		return (EINTR);
215 	if (dump_bytes(ba, data, blksz))
216 		return (EINTR);
217 	return (0);
218 }
219 
220 static int
221 dump_freeobjects(struct backuparg *ba, uint64_t firstobj, uint64_t numobjs)
222 {
223 	struct drr_freeobjects *drrfo = &(ba->drr->drr_u.drr_freeobjects);
224 
225 	/*
226 	 * If there is a pending op, but it's not PENDING_FREEOBJECTS,
227 	 * push it out, since free block aggregation can only be done for
228 	 * blocks of the same type (i.e., DRR_FREE records can only be
229 	 * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
230 	 * can only be aggregated with other DRR_FREEOBJECTS records.
231 	 */
232 	if (ba->pending_op != PENDING_NONE &&
233 	    ba->pending_op != PENDING_FREEOBJECTS) {
234 		if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
235 			return (EINTR);
236 		ba->pending_op = PENDING_NONE;
237 	}
238 	if (ba->pending_op == PENDING_FREEOBJECTS) {
239 		/*
240 		 * See whether this free object array can be aggregated
241 		 * with pending one
242 		 */
243 		if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
244 			drrfo->drr_numobjs += numobjs;
245 			return (0);
246 		} else {
247 			/* can't be aggregated.  Push out pending record */
248 			if (dump_bytes(ba, ba->drr,
249 			    sizeof (dmu_replay_record_t)) != 0)
250 				return (EINTR);
251 			ba->pending_op = PENDING_NONE;
252 		}
253 	}
254 
255 	/* write a FREEOBJECTS record */
256 	bzero(ba->drr, sizeof (dmu_replay_record_t));
257 	ba->drr->drr_type = DRR_FREEOBJECTS;
258 	drrfo->drr_firstobj = firstobj;
259 	drrfo->drr_numobjs = numobjs;
260 	drrfo->drr_toguid = ba->toguid;
261 
262 	ba->pending_op = PENDING_FREEOBJECTS;
263 
264 	return (0);
265 }
266 
267 static int
268 dump_dnode(struct backuparg *ba, uint64_t object, dnode_phys_t *dnp)
269 {
270 	struct drr_object *drro = &(ba->drr->drr_u.drr_object);
271 
272 	if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
273 		return (dump_freeobjects(ba, object, 1));
274 
275 	if (ba->pending_op != PENDING_NONE) {
276 		if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
277 			return (EINTR);
278 		ba->pending_op = PENDING_NONE;
279 	}
280 
281 	/* write an OBJECT record */
282 	bzero(ba->drr, sizeof (dmu_replay_record_t));
283 	ba->drr->drr_type = DRR_OBJECT;
284 	drro->drr_object = object;
285 	drro->drr_type = dnp->dn_type;
286 	drro->drr_bonustype = dnp->dn_bonustype;
287 	drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
288 	drro->drr_bonuslen = dnp->dn_bonuslen;
289 	drro->drr_checksumtype = dnp->dn_checksum;
290 	drro->drr_compress = dnp->dn_compress;
291 	drro->drr_toguid = ba->toguid;
292 
293 	if (dump_bytes(ba, ba->drr, sizeof (dmu_replay_record_t)) != 0)
294 		return (EINTR);
295 
296 	if (dump_bytes(ba, DN_BONUS(dnp), P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0)
297 		return (EINTR);
298 
299 	/* free anything past the end of the file */
300 	if (dump_free(ba, object, (dnp->dn_maxblkid + 1) *
301 	    (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL))
302 		return (EINTR);
303 	if (ba->err)
304 		return (EINTR);
305 	return (0);
306 }
307 
308 #define	BP_SPAN(dnp, level) \
309 	(((uint64_t)dnp->dn_datablkszsec) << (SPA_MINBLOCKSHIFT + \
310 	(level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)))
311 
312 /* ARGSUSED */
313 static int
314 backup_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, arc_buf_t *pbuf,
315     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
316 {
317 	struct backuparg *ba = arg;
318 	dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
319 	int err = 0;
320 
321 	if (issig(JUSTLOOKING) && issig(FORREAL))
322 		return (EINTR);
323 
324 	if (zb->zb_object != DMU_META_DNODE_OBJECT &&
325 	    DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
326 		return (0);
327 	} else if (bp == NULL && zb->zb_object == DMU_META_DNODE_OBJECT) {
328 		uint64_t span = BP_SPAN(dnp, zb->zb_level);
329 		uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
330 		err = dump_freeobjects(ba, dnobj, span >> DNODE_SHIFT);
331 	} else if (bp == NULL) {
332 		uint64_t span = BP_SPAN(dnp, zb->zb_level);
333 		err = dump_free(ba, zb->zb_object, zb->zb_blkid * span, span);
334 	} else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
335 		return (0);
336 	} else if (type == DMU_OT_DNODE) {
337 		dnode_phys_t *blk;
338 		int i;
339 		int blksz = BP_GET_LSIZE(bp);
340 		uint32_t aflags = ARC_WAIT;
341 		arc_buf_t *abuf;
342 
343 		if (dsl_read(NULL, spa, bp, pbuf,
344 		    arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
345 		    ZIO_FLAG_CANFAIL, &aflags, zb) != 0)
346 			return (EIO);
347 
348 		blk = abuf->b_data;
349 		for (i = 0; i < blksz >> DNODE_SHIFT; i++) {
350 			uint64_t dnobj = (zb->zb_blkid <<
351 			    (DNODE_BLOCK_SHIFT - DNODE_SHIFT)) + i;
352 			err = dump_dnode(ba, dnobj, blk+i);
353 			if (err)
354 				break;
355 		}
356 		(void) arc_buf_remove_ref(abuf, &abuf);
357 	} else if (type == DMU_OT_SA) {
358 		uint32_t aflags = ARC_WAIT;
359 		arc_buf_t *abuf;
360 		int blksz = BP_GET_LSIZE(bp);
361 
362 		if (arc_read_nolock(NULL, spa, bp,
363 		    arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
364 		    ZIO_FLAG_CANFAIL, &aflags, zb) != 0)
365 			return (EIO);
366 
367 		err = dump_spill(ba, zb->zb_object, blksz, abuf->b_data);
368 		(void) arc_buf_remove_ref(abuf, &abuf);
369 	} else { /* it's a level-0 block of a regular object */
370 		uint32_t aflags = ARC_WAIT;
371 		arc_buf_t *abuf;
372 		int blksz = BP_GET_LSIZE(bp);
373 
374 		if (dsl_read(NULL, spa, bp, pbuf,
375 		    arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
376 		    ZIO_FLAG_CANFAIL, &aflags, zb) != 0) {
377 			if (zfs_send_corrupt_data) {
378 				/* Send a block filled with 0x"zfs badd bloc" */
379 				abuf = arc_buf_alloc(spa, blksz, &abuf,
380 				    ARC_BUFC_DATA);
381 				uint64_t *ptr;
382 				for (ptr = abuf->b_data;
383 				    (char *)ptr < (char *)abuf->b_data + blksz;
384 				    ptr++)
385 					*ptr = 0x2f5baddb10c;
386 			} else {
387 				return (EIO);
388 			}
389 		}
390 
391 		err = dump_data(ba, type, zb->zb_object, zb->zb_blkid * blksz,
392 		    blksz, bp, abuf->b_data);
393 		(void) arc_buf_remove_ref(abuf, &abuf);
394 	}
395 
396 	ASSERT(err == 0 || err == EINTR);
397 	return (err);
398 }
399 
400 int
401 dmu_sendbackup(objset_t *tosnap, objset_t *fromsnap, boolean_t fromorigin,
402     vnode_t *vp, offset_t *off)
403 {
404 	dsl_dataset_t *ds = tosnap->os_dsl_dataset;
405 	dsl_dataset_t *fromds = fromsnap ? fromsnap->os_dsl_dataset : NULL;
406 	dmu_replay_record_t *drr;
407 	struct backuparg ba;
408 	int err;
409 	uint64_t fromtxg = 0;
410 
411 	/* tosnap must be a snapshot */
412 	if (ds->ds_phys->ds_next_snap_obj == 0)
413 		return (EINVAL);
414 
415 	/* fromsnap must be an earlier snapshot from the same fs as tosnap */
416 	if (fromds && (ds->ds_dir != fromds->ds_dir ||
417 	    fromds->ds_phys->ds_creation_txg >= ds->ds_phys->ds_creation_txg))
418 		return (EXDEV);
419 
420 	if (fromorigin) {
421 		dsl_pool_t *dp = ds->ds_dir->dd_pool;
422 
423 		if (fromsnap)
424 			return (EINVAL);
425 
426 		if (dsl_dir_is_clone(ds->ds_dir)) {
427 			rw_enter(&dp->dp_config_rwlock, RW_READER);
428 			err = dsl_dataset_hold_obj(dp,
429 			    ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &fromds);
430 			rw_exit(&dp->dp_config_rwlock);
431 			if (err)
432 				return (err);
433 		} else {
434 			fromorigin = B_FALSE;
435 		}
436 	}
437 
438 
439 	drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
440 	drr->drr_type = DRR_BEGIN;
441 	drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
442 	DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
443 	    DMU_SUBSTREAM);
444 
445 #ifdef _KERNEL
446 	if (dmu_objset_type(tosnap) == DMU_OST_ZFS) {
447 		uint64_t version;
448 		if (zfs_get_zplprop(tosnap, ZFS_PROP_VERSION, &version) != 0)
449 			return (EINVAL);
450 		if (version == ZPL_VERSION_SA) {
451 			DMU_SET_FEATUREFLAGS(
452 			    drr->drr_u.drr_begin.drr_versioninfo,
453 			    DMU_BACKUP_FEATURE_SA_SPILL);
454 		}
455 	}
456 #endif
457 
458 	drr->drr_u.drr_begin.drr_creation_time =
459 	    ds->ds_phys->ds_creation_time;
460 	drr->drr_u.drr_begin.drr_type = tosnap->os_phys->os_type;
461 	if (fromorigin)
462 		drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
463 	drr->drr_u.drr_begin.drr_toguid = ds->ds_phys->ds_guid;
464 	if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
465 		drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
466 
467 	if (fromds)
468 		drr->drr_u.drr_begin.drr_fromguid = fromds->ds_phys->ds_guid;
469 	dsl_dataset_name(ds, drr->drr_u.drr_begin.drr_toname);
470 
471 	if (fromds)
472 		fromtxg = fromds->ds_phys->ds_creation_txg;
473 	if (fromorigin)
474 		dsl_dataset_rele(fromds, FTAG);
475 
476 	ba.drr = drr;
477 	ba.vp = vp;
478 	ba.os = tosnap;
479 	ba.off = off;
480 	ba.toguid = ds->ds_phys->ds_guid;
481 	ZIO_SET_CHECKSUM(&ba.zc, 0, 0, 0, 0);
482 	ba.pending_op = PENDING_NONE;
483 
484 	if (dump_bytes(&ba, drr, sizeof (dmu_replay_record_t)) != 0) {
485 		kmem_free(drr, sizeof (dmu_replay_record_t));
486 		return (ba.err);
487 	}
488 
489 	err = traverse_dataset(ds, fromtxg, TRAVERSE_PRE | TRAVERSE_PREFETCH,
490 	    backup_cb, &ba);
491 
492 	if (ba.pending_op != PENDING_NONE)
493 		if (dump_bytes(&ba, drr, sizeof (dmu_replay_record_t)) != 0)
494 			err = EINTR;
495 
496 	if (err) {
497 		if (err == EINTR && ba.err)
498 			err = ba.err;
499 		kmem_free(drr, sizeof (dmu_replay_record_t));
500 		return (err);
501 	}
502 
503 	bzero(drr, sizeof (dmu_replay_record_t));
504 	drr->drr_type = DRR_END;
505 	drr->drr_u.drr_end.drr_checksum = ba.zc;
506 	drr->drr_u.drr_end.drr_toguid = ba.toguid;
507 
508 	if (dump_bytes(&ba, drr, sizeof (dmu_replay_record_t)) != 0) {
509 		kmem_free(drr, sizeof (dmu_replay_record_t));
510 		return (ba.err);
511 	}
512 
513 	kmem_free(drr, sizeof (dmu_replay_record_t));
514 
515 	return (0);
516 }
517 
518 int
519 dmu_send_estimate(objset_t *tosnap, objset_t *fromsnap, boolean_t fromorigin,
520     uint64_t *sizep)
521 {
522 	dsl_dataset_t *ds = tosnap->os_dsl_dataset;
523 	dsl_dataset_t *fromds = fromsnap ? fromsnap->os_dsl_dataset : NULL;
524 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
525 	int err;
526 	uint64_t size;
527 
528 	/* tosnap must be a snapshot */
529 	if (ds->ds_phys->ds_next_snap_obj == 0)
530 		return (EINVAL);
531 
532 	/* fromsnap must be an earlier snapshot from the same fs as tosnap */
533 	if (fromds && (ds->ds_dir != fromds->ds_dir ||
534 	    fromds->ds_phys->ds_creation_txg >= ds->ds_phys->ds_creation_txg))
535 		return (EXDEV);
536 
537 	if (fromorigin) {
538 		if (fromsnap)
539 			return (EINVAL);
540 
541 		if (dsl_dir_is_clone(ds->ds_dir)) {
542 			rw_enter(&dp->dp_config_rwlock, RW_READER);
543 			err = dsl_dataset_hold_obj(dp,
544 			    ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &fromds);
545 			rw_exit(&dp->dp_config_rwlock);
546 			if (err)
547 				return (err);
548 		} else {
549 			fromorigin = B_FALSE;
550 		}
551 	}
552 
553 	/* Get uncompressed size estimate of changed data. */
554 	if (fromds == NULL) {
555 		size = ds->ds_phys->ds_uncompressed_bytes;
556 	} else {
557 		uint64_t used, comp;
558 		err = dsl_dataset_space_written(fromds, ds,
559 		    &used, &comp, &size);
560 		if (fromorigin)
561 			dsl_dataset_rele(fromds, FTAG);
562 		if (err)
563 			return (err);
564 	}
565 
566 	/*
567 	 * Assume that space (both on-disk and in-stream) is dominated by
568 	 * data.  We will adjust for indirect blocks and the copies property,
569 	 * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
570 	 */
571 
572 	/*
573 	 * Subtract out approximate space used by indirect blocks.
574 	 * Assume most space is used by data blocks (non-indirect, non-dnode).
575 	 * Assume all blocks are recordsize.  Assume ditto blocks and
576 	 * internal fragmentation counter out compression.
577 	 *
578 	 * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
579 	 * block, which we observe in practice.
580 	 */
581 	uint64_t recordsize;
582 	rw_enter(&dp->dp_config_rwlock, RW_READER);
583 	err = dsl_prop_get_ds(ds, "recordsize",
584 	    sizeof (recordsize), 1, &recordsize, NULL);
585 	rw_exit(&dp->dp_config_rwlock);
586 	if (err)
587 		return (err);
588 	size -= size / recordsize * sizeof (blkptr_t);
589 
590 	/* Add in the space for the record associated with each block. */
591 	size += size / recordsize * sizeof (dmu_replay_record_t);
592 
593 	*sizep = size;
594 
595 	return (0);
596 }
597 
598 struct recvbeginsyncarg {
599 	const char *tofs;
600 	const char *tosnap;
601 	dsl_dataset_t *origin;
602 	uint64_t fromguid;
603 	dmu_objset_type_t type;
604 	void *tag;
605 	boolean_t force;
606 	uint64_t dsflags;
607 	char clonelastname[MAXNAMELEN];
608 	dsl_dataset_t *ds; /* the ds to recv into; returned from the syncfunc */
609 	cred_t *cr;
610 };
611 
612 /* ARGSUSED */
613 static int
614 recv_new_check(void *arg1, void *arg2, dmu_tx_t *tx)
615 {
616 	dsl_dir_t *dd = arg1;
617 	struct recvbeginsyncarg *rbsa = arg2;
618 	objset_t *mos = dd->dd_pool->dp_meta_objset;
619 	uint64_t val;
620 	int err;
621 
622 	err = zap_lookup(mos, dd->dd_phys->dd_child_dir_zapobj,
623 	    strrchr(rbsa->tofs, '/') + 1, sizeof (uint64_t), 1, &val);
624 
625 	if (err != ENOENT)
626 		return (err ? err : EEXIST);
627 
628 	if (rbsa->origin) {
629 		/* make sure it's a snap in the same pool */
630 		if (rbsa->origin->ds_dir->dd_pool != dd->dd_pool)
631 			return (EXDEV);
632 		if (!dsl_dataset_is_snapshot(rbsa->origin))
633 			return (EINVAL);
634 		if (rbsa->origin->ds_phys->ds_guid != rbsa->fromguid)
635 			return (ENODEV);
636 	}
637 
638 	return (0);
639 }
640 
641 static void
642 recv_new_sync(void *arg1, void *arg2, dmu_tx_t *tx)
643 {
644 	dsl_dir_t *dd = arg1;
645 	struct recvbeginsyncarg *rbsa = arg2;
646 	uint64_t flags = DS_FLAG_INCONSISTENT | rbsa->dsflags;
647 	uint64_t dsobj;
648 
649 	/* Create and open new dataset. */
650 	dsobj = dsl_dataset_create_sync(dd, strrchr(rbsa->tofs, '/') + 1,
651 	    rbsa->origin, flags, rbsa->cr, tx);
652 	VERIFY(0 == dsl_dataset_own_obj(dd->dd_pool, dsobj,
653 	    B_TRUE, dmu_recv_tag, &rbsa->ds));
654 
655 	if (rbsa->origin == NULL) {
656 		(void) dmu_objset_create_impl(dd->dd_pool->dp_spa,
657 		    rbsa->ds, &rbsa->ds->ds_phys->ds_bp, rbsa->type, tx);
658 	}
659 
660 	spa_history_log_internal(LOG_DS_REPLAY_FULL_SYNC,
661 	    dd->dd_pool->dp_spa, tx, "dataset = %lld", dsobj);
662 }
663 
664 /* ARGSUSED */
665 static int
666 recv_existing_check(void *arg1, void *arg2, dmu_tx_t *tx)
667 {
668 	dsl_dataset_t *ds = arg1;
669 	struct recvbeginsyncarg *rbsa = arg2;
670 	int err;
671 	uint64_t val;
672 
673 	/* must not have any changes since most recent snapshot */
674 	if (!rbsa->force && dsl_dataset_modified_since_lastsnap(ds))
675 		return (ETXTBSY);
676 
677 	/* new snapshot name must not exist */
678 	err = zap_lookup(ds->ds_dir->dd_pool->dp_meta_objset,
679 	    ds->ds_phys->ds_snapnames_zapobj, rbsa->tosnap, 8, 1, &val);
680 	if (err == 0)
681 		return (EEXIST);
682 	if (err != ENOENT)
683 		return (err);
684 
685 	if (rbsa->fromguid) {
686 		/* if incremental, most recent snapshot must match fromguid */
687 		if (ds->ds_prev == NULL)
688 			return (ENODEV);
689 
690 		/*
691 		 * most recent snapshot must match fromguid, or there are no
692 		 * changes since the fromguid one
693 		 */
694 		if (ds->ds_prev->ds_phys->ds_guid != rbsa->fromguid) {
695 			uint64_t birth = ds->ds_prev->ds_phys->ds_bp.blk_birth;
696 			uint64_t obj = ds->ds_prev->ds_phys->ds_prev_snap_obj;
697 			while (obj != 0) {
698 				dsl_dataset_t *snap;
699 				err = dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
700 				    obj, FTAG, &snap);
701 				if (err)
702 					return (ENODEV);
703 				if (snap->ds_phys->ds_creation_txg < birth) {
704 					dsl_dataset_rele(snap, FTAG);
705 					return (ENODEV);
706 				}
707 				if (snap->ds_phys->ds_guid == rbsa->fromguid) {
708 					dsl_dataset_rele(snap, FTAG);
709 					break; /* it's ok */
710 				}
711 				obj = snap->ds_phys->ds_prev_snap_obj;
712 				dsl_dataset_rele(snap, FTAG);
713 			}
714 			if (obj == 0)
715 				return (ENODEV);
716 		}
717 	} else {
718 		/* if full, most recent snapshot must be $ORIGIN */
719 		if (ds->ds_phys->ds_prev_snap_txg >= TXG_INITIAL)
720 			return (ENODEV);
721 	}
722 
723 	/* temporary clone name must not exist */
724 	err = zap_lookup(ds->ds_dir->dd_pool->dp_meta_objset,
725 	    ds->ds_dir->dd_phys->dd_child_dir_zapobj,
726 	    rbsa->clonelastname, 8, 1, &val);
727 	if (err == 0)
728 		return (EEXIST);
729 	if (err != ENOENT)
730 		return (err);
731 
732 	return (0);
733 }
734 
735 /* ARGSUSED */
736 static void
737 recv_existing_sync(void *arg1, void *arg2, dmu_tx_t *tx)
738 {
739 	dsl_dataset_t *ohds = arg1;
740 	struct recvbeginsyncarg *rbsa = arg2;
741 	dsl_pool_t *dp = ohds->ds_dir->dd_pool;
742 	dsl_dataset_t *cds;
743 	uint64_t flags = DS_FLAG_INCONSISTENT | rbsa->dsflags;
744 	uint64_t dsobj;
745 
746 	/* create and open the temporary clone */
747 	dsobj = dsl_dataset_create_sync(ohds->ds_dir, rbsa->clonelastname,
748 	    ohds->ds_prev, flags, rbsa->cr, tx);
749 	VERIFY(0 == dsl_dataset_own_obj(dp, dsobj, B_TRUE, dmu_recv_tag, &cds));
750 
751 	/*
752 	 * If we actually created a non-clone, we need to create the
753 	 * objset in our new dataset.
754 	 */
755 	if (BP_IS_HOLE(dsl_dataset_get_blkptr(cds))) {
756 		(void) dmu_objset_create_impl(dp->dp_spa,
757 		    cds, dsl_dataset_get_blkptr(cds), rbsa->type, tx);
758 	}
759 
760 	rbsa->ds = cds;
761 
762 	spa_history_log_internal(LOG_DS_REPLAY_INC_SYNC,
763 	    dp->dp_spa, tx, "dataset = %lld", dsobj);
764 }
765 
766 static boolean_t
767 dmu_recv_verify_features(dsl_dataset_t *ds, struct drr_begin *drrb)
768 {
769 	int featureflags;
770 
771 	featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
772 
773 	/* Verify pool version supports SA if SA_SPILL feature set */
774 	return ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
775 	    (spa_version(dsl_dataset_get_spa(ds)) < SPA_VERSION_SA));
776 }
777 
778 /*
779  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
780  * succeeds; otherwise we will leak the holds on the datasets.
781  */
782 int
783 dmu_recv_begin(char *tofs, char *tosnap, char *top_ds, struct drr_begin *drrb,
784     boolean_t force, objset_t *origin, dmu_recv_cookie_t *drc)
785 {
786 	int err = 0;
787 	boolean_t byteswap;
788 	struct recvbeginsyncarg rbsa = { 0 };
789 	uint64_t versioninfo;
790 	int flags;
791 	dsl_dataset_t *ds;
792 
793 	if (drrb->drr_magic == DMU_BACKUP_MAGIC)
794 		byteswap = FALSE;
795 	else if (drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
796 		byteswap = TRUE;
797 	else
798 		return (EINVAL);
799 
800 	rbsa.tofs = tofs;
801 	rbsa.tosnap = tosnap;
802 	rbsa.origin = origin ? origin->os_dsl_dataset : NULL;
803 	rbsa.fromguid = drrb->drr_fromguid;
804 	rbsa.type = drrb->drr_type;
805 	rbsa.tag = FTAG;
806 	rbsa.dsflags = 0;
807 	rbsa.cr = CRED();
808 	versioninfo = drrb->drr_versioninfo;
809 	flags = drrb->drr_flags;
810 
811 	if (byteswap) {
812 		rbsa.type = BSWAP_32(rbsa.type);
813 		rbsa.fromguid = BSWAP_64(rbsa.fromguid);
814 		versioninfo = BSWAP_64(versioninfo);
815 		flags = BSWAP_32(flags);
816 	}
817 
818 	if (DMU_GET_STREAM_HDRTYPE(versioninfo) == DMU_COMPOUNDSTREAM ||
819 	    rbsa.type >= DMU_OST_NUMTYPES ||
820 	    ((flags & DRR_FLAG_CLONE) && origin == NULL))
821 		return (EINVAL);
822 
823 	if (flags & DRR_FLAG_CI_DATA)
824 		rbsa.dsflags = DS_FLAG_CI_DATASET;
825 
826 	bzero(drc, sizeof (dmu_recv_cookie_t));
827 	drc->drc_drrb = drrb;
828 	drc->drc_tosnap = tosnap;
829 	drc->drc_top_ds = top_ds;
830 	drc->drc_force = force;
831 
832 	/*
833 	 * Process the begin in syncing context.
834 	 */
835 
836 	/* open the dataset we are logically receiving into */
837 	err = dsl_dataset_hold(tofs, dmu_recv_tag, &ds);
838 	if (err == 0) {
839 		if (dmu_recv_verify_features(ds, drrb)) {
840 			dsl_dataset_rele(ds, dmu_recv_tag);
841 			return (ENOTSUP);
842 		}
843 		/* target fs already exists; recv into temp clone */
844 
845 		/* Can't recv a clone into an existing fs */
846 		if (flags & DRR_FLAG_CLONE) {
847 			dsl_dataset_rele(ds, dmu_recv_tag);
848 			return (EINVAL);
849 		}
850 
851 		/* must not have an incremental recv already in progress */
852 		if (!mutex_tryenter(&ds->ds_recvlock)) {
853 			dsl_dataset_rele(ds, dmu_recv_tag);
854 			return (EBUSY);
855 		}
856 
857 		/* tmp clone name is: tofs/%tosnap" */
858 		(void) snprintf(rbsa.clonelastname, sizeof (rbsa.clonelastname),
859 		    "%%%s", tosnap);
860 		rbsa.force = force;
861 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
862 		    recv_existing_check, recv_existing_sync, ds, &rbsa, 5);
863 		if (err) {
864 			mutex_exit(&ds->ds_recvlock);
865 			dsl_dataset_rele(ds, dmu_recv_tag);
866 			return (err);
867 		}
868 		drc->drc_logical_ds = ds;
869 		drc->drc_real_ds = rbsa.ds;
870 	} else if (err == ENOENT) {
871 		/* target fs does not exist; must be a full backup or clone */
872 		char *cp;
873 
874 		/*
875 		 * If it's a non-clone incremental, we are missing the
876 		 * target fs, so fail the recv.
877 		 */
878 		if (rbsa.fromguid && !(flags & DRR_FLAG_CLONE))
879 			return (ENOENT);
880 
881 		/* Open the parent of tofs */
882 		cp = strrchr(tofs, '/');
883 		*cp = '\0';
884 		err = dsl_dataset_hold(tofs, FTAG, &ds);
885 		*cp = '/';
886 		if (err)
887 			return (err);
888 
889 		if (dmu_recv_verify_features(ds, drrb)) {
890 			dsl_dataset_rele(ds, FTAG);
891 			return (ENOTSUP);
892 		}
893 
894 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
895 		    recv_new_check, recv_new_sync, ds->ds_dir, &rbsa, 5);
896 		dsl_dataset_rele(ds, FTAG);
897 		if (err)
898 			return (err);
899 		drc->drc_logical_ds = drc->drc_real_ds = rbsa.ds;
900 		drc->drc_newfs = B_TRUE;
901 	}
902 
903 	return (err);
904 }
905 
906 struct restorearg {
907 	int err;
908 	int byteswap;
909 	vnode_t *vp;
910 	char *buf;
911 	uint64_t voff;
912 	int bufsize; /* amount of memory allocated for buf */
913 	zio_cksum_t cksum;
914 	avl_tree_t *guid_to_ds_map;
915 };
916 
917 typedef struct guid_map_entry {
918 	uint64_t	guid;
919 	dsl_dataset_t	*gme_ds;
920 	avl_node_t	avlnode;
921 } guid_map_entry_t;
922 
923 static int
924 guid_compare(const void *arg1, const void *arg2)
925 {
926 	const guid_map_entry_t *gmep1 = arg1;
927 	const guid_map_entry_t *gmep2 = arg2;
928 
929 	if (gmep1->guid < gmep2->guid)
930 		return (-1);
931 	else if (gmep1->guid > gmep2->guid)
932 		return (1);
933 	return (0);
934 }
935 
936 static void
937 free_guid_map_onexit(void *arg)
938 {
939 	avl_tree_t *ca = arg;
940 	void *cookie = NULL;
941 	guid_map_entry_t *gmep;
942 
943 	while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
944 		dsl_dataset_rele(gmep->gme_ds, ca);
945 		kmem_free(gmep, sizeof (guid_map_entry_t));
946 	}
947 	avl_destroy(ca);
948 	kmem_free(ca, sizeof (avl_tree_t));
949 }
950 
951 static void *
952 restore_read(struct restorearg *ra, int len)
953 {
954 	void *rv;
955 	int done = 0;
956 
957 	/* some things will require 8-byte alignment, so everything must */
958 	ASSERT3U(len % 8, ==, 0);
959 
960 	while (done < len) {
961 		ssize_t resid;
962 
963 		ra->err = vn_rdwr(UIO_READ, ra->vp,
964 		    (caddr_t)ra->buf + done, len - done,
965 		    ra->voff, UIO_SYSSPACE, FAPPEND,
966 		    RLIM64_INFINITY, CRED(), &resid);
967 
968 		if (resid == len - done)
969 			ra->err = EINVAL;
970 		ra->voff += len - done - resid;
971 		done = len - resid;
972 		if (ra->err)
973 			return (NULL);
974 	}
975 
976 	ASSERT3U(done, ==, len);
977 	rv = ra->buf;
978 	if (ra->byteswap)
979 		fletcher_4_incremental_byteswap(rv, len, &ra->cksum);
980 	else
981 		fletcher_4_incremental_native(rv, len, &ra->cksum);
982 	return (rv);
983 }
984 
985 static void
986 backup_byteswap(dmu_replay_record_t *drr)
987 {
988 #define	DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
989 #define	DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
990 	drr->drr_type = BSWAP_32(drr->drr_type);
991 	drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
992 	switch (drr->drr_type) {
993 	case DRR_BEGIN:
994 		DO64(drr_begin.drr_magic);
995 		DO64(drr_begin.drr_versioninfo);
996 		DO64(drr_begin.drr_creation_time);
997 		DO32(drr_begin.drr_type);
998 		DO32(drr_begin.drr_flags);
999 		DO64(drr_begin.drr_toguid);
1000 		DO64(drr_begin.drr_fromguid);
1001 		break;
1002 	case DRR_OBJECT:
1003 		DO64(drr_object.drr_object);
1004 		/* DO64(drr_object.drr_allocation_txg); */
1005 		DO32(drr_object.drr_type);
1006 		DO32(drr_object.drr_bonustype);
1007 		DO32(drr_object.drr_blksz);
1008 		DO32(drr_object.drr_bonuslen);
1009 		DO64(drr_object.drr_toguid);
1010 		break;
1011 	case DRR_FREEOBJECTS:
1012 		DO64(drr_freeobjects.drr_firstobj);
1013 		DO64(drr_freeobjects.drr_numobjs);
1014 		DO64(drr_freeobjects.drr_toguid);
1015 		break;
1016 	case DRR_WRITE:
1017 		DO64(drr_write.drr_object);
1018 		DO32(drr_write.drr_type);
1019 		DO64(drr_write.drr_offset);
1020 		DO64(drr_write.drr_length);
1021 		DO64(drr_write.drr_toguid);
1022 		DO64(drr_write.drr_key.ddk_cksum.zc_word[0]);
1023 		DO64(drr_write.drr_key.ddk_cksum.zc_word[1]);
1024 		DO64(drr_write.drr_key.ddk_cksum.zc_word[2]);
1025 		DO64(drr_write.drr_key.ddk_cksum.zc_word[3]);
1026 		DO64(drr_write.drr_key.ddk_prop);
1027 		break;
1028 	case DRR_WRITE_BYREF:
1029 		DO64(drr_write_byref.drr_object);
1030 		DO64(drr_write_byref.drr_offset);
1031 		DO64(drr_write_byref.drr_length);
1032 		DO64(drr_write_byref.drr_toguid);
1033 		DO64(drr_write_byref.drr_refguid);
1034 		DO64(drr_write_byref.drr_refobject);
1035 		DO64(drr_write_byref.drr_refoffset);
1036 		DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[0]);
1037 		DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[1]);
1038 		DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[2]);
1039 		DO64(drr_write_byref.drr_key.ddk_cksum.zc_word[3]);
1040 		DO64(drr_write_byref.drr_key.ddk_prop);
1041 		break;
1042 	case DRR_FREE:
1043 		DO64(drr_free.drr_object);
1044 		DO64(drr_free.drr_offset);
1045 		DO64(drr_free.drr_length);
1046 		DO64(drr_free.drr_toguid);
1047 		break;
1048 	case DRR_SPILL:
1049 		DO64(drr_spill.drr_object);
1050 		DO64(drr_spill.drr_length);
1051 		DO64(drr_spill.drr_toguid);
1052 		break;
1053 	case DRR_END:
1054 		DO64(drr_end.drr_checksum.zc_word[0]);
1055 		DO64(drr_end.drr_checksum.zc_word[1]);
1056 		DO64(drr_end.drr_checksum.zc_word[2]);
1057 		DO64(drr_end.drr_checksum.zc_word[3]);
1058 		DO64(drr_end.drr_toguid);
1059 		break;
1060 	}
1061 #undef DO64
1062 #undef DO32
1063 }
1064 
1065 static int
1066 restore_object(struct restorearg *ra, objset_t *os, struct drr_object *drro)
1067 {
1068 	int err;
1069 	dmu_tx_t *tx;
1070 	void *data = NULL;
1071 
1072 	if (drro->drr_type == DMU_OT_NONE ||
1073 	    drro->drr_type >= DMU_OT_NUMTYPES ||
1074 	    drro->drr_bonustype >= DMU_OT_NUMTYPES ||
1075 	    drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
1076 	    drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
1077 	    P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
1078 	    drro->drr_blksz < SPA_MINBLOCKSIZE ||
1079 	    drro->drr_blksz > SPA_MAXBLOCKSIZE ||
1080 	    drro->drr_bonuslen > DN_MAX_BONUSLEN) {
1081 		return (EINVAL);
1082 	}
1083 
1084 	err = dmu_object_info(os, drro->drr_object, NULL);
1085 
1086 	if (err != 0 && err != ENOENT)
1087 		return (EINVAL);
1088 
1089 	if (drro->drr_bonuslen) {
1090 		data = restore_read(ra, P2ROUNDUP(drro->drr_bonuslen, 8));
1091 		if (ra->err)
1092 			return (ra->err);
1093 	}
1094 
1095 	if (err == ENOENT) {
1096 		/* currently free, want to be allocated */
1097 		tx = dmu_tx_create(os);
1098 		dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1099 		err = dmu_tx_assign(tx, TXG_WAIT);
1100 		if (err) {
1101 			dmu_tx_abort(tx);
1102 			return (err);
1103 		}
1104 		err = dmu_object_claim(os, drro->drr_object,
1105 		    drro->drr_type, drro->drr_blksz,
1106 		    drro->drr_bonustype, drro->drr_bonuslen, tx);
1107 		dmu_tx_commit(tx);
1108 	} else {
1109 		/* currently allocated, want to be allocated */
1110 		err = dmu_object_reclaim(os, drro->drr_object,
1111 		    drro->drr_type, drro->drr_blksz,
1112 		    drro->drr_bonustype, drro->drr_bonuslen);
1113 	}
1114 	if (err) {
1115 		return (EINVAL);
1116 	}
1117 
1118 	tx = dmu_tx_create(os);
1119 	dmu_tx_hold_bonus(tx, drro->drr_object);
1120 	err = dmu_tx_assign(tx, TXG_WAIT);
1121 	if (err) {
1122 		dmu_tx_abort(tx);
1123 		return (err);
1124 	}
1125 
1126 	dmu_object_set_checksum(os, drro->drr_object, drro->drr_checksumtype,
1127 	    tx);
1128 	dmu_object_set_compress(os, drro->drr_object, drro->drr_compress, tx);
1129 
1130 	if (data != NULL) {
1131 		dmu_buf_t *db;
1132 
1133 		VERIFY(0 == dmu_bonus_hold(os, drro->drr_object, FTAG, &db));
1134 		dmu_buf_will_dirty(db, tx);
1135 
1136 		ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
1137 		bcopy(data, db->db_data, drro->drr_bonuslen);
1138 		if (ra->byteswap) {
1139 			dmu_ot[drro->drr_bonustype].ot_byteswap(db->db_data,
1140 			    drro->drr_bonuslen);
1141 		}
1142 		dmu_buf_rele(db, FTAG);
1143 	}
1144 	dmu_tx_commit(tx);
1145 	return (0);
1146 }
1147 
1148 /* ARGSUSED */
1149 static int
1150 restore_freeobjects(struct restorearg *ra, objset_t *os,
1151     struct drr_freeobjects *drrfo)
1152 {
1153 	uint64_t obj;
1154 
1155 	if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
1156 		return (EINVAL);
1157 
1158 	for (obj = drrfo->drr_firstobj;
1159 	    obj < drrfo->drr_firstobj + drrfo->drr_numobjs;
1160 	    (void) dmu_object_next(os, &obj, FALSE, 0)) {
1161 		int err;
1162 
1163 		if (dmu_object_info(os, obj, NULL) != 0)
1164 			continue;
1165 
1166 		err = dmu_free_object(os, obj);
1167 		if (err)
1168 			return (err);
1169 	}
1170 	return (0);
1171 }
1172 
1173 static int
1174 restore_write(struct restorearg *ra, objset_t *os,
1175     struct drr_write *drrw)
1176 {
1177 	dmu_tx_t *tx;
1178 	void *data;
1179 	int err;
1180 
1181 	if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
1182 	    drrw->drr_type >= DMU_OT_NUMTYPES)
1183 		return (EINVAL);
1184 
1185 	data = restore_read(ra, drrw->drr_length);
1186 	if (data == NULL)
1187 		return (ra->err);
1188 
1189 	if (dmu_object_info(os, drrw->drr_object, NULL) != 0)
1190 		return (EINVAL);
1191 
1192 	tx = dmu_tx_create(os);
1193 
1194 	dmu_tx_hold_write(tx, drrw->drr_object,
1195 	    drrw->drr_offset, drrw->drr_length);
1196 	err = dmu_tx_assign(tx, TXG_WAIT);
1197 	if (err) {
1198 		dmu_tx_abort(tx);
1199 		return (err);
1200 	}
1201 	if (ra->byteswap)
1202 		dmu_ot[drrw->drr_type].ot_byteswap(data, drrw->drr_length);
1203 	dmu_write(os, drrw->drr_object,
1204 	    drrw->drr_offset, drrw->drr_length, data, tx);
1205 	dmu_tx_commit(tx);
1206 	return (0);
1207 }
1208 
1209 /*
1210  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
1211  * streams to refer to a copy of the data that is already on the
1212  * system because it came in earlier in the stream.  This function
1213  * finds the earlier copy of the data, and uses that copy instead of
1214  * data from the stream to fulfill this write.
1215  */
1216 static int
1217 restore_write_byref(struct restorearg *ra, objset_t *os,
1218     struct drr_write_byref *drrwbr)
1219 {
1220 	dmu_tx_t *tx;
1221 	int err;
1222 	guid_map_entry_t gmesrch;
1223 	guid_map_entry_t *gmep;
1224 	avl_index_t	where;
1225 	objset_t *ref_os = NULL;
1226 	dmu_buf_t *dbp;
1227 
1228 	if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
1229 		return (EINVAL);
1230 
1231 	/*
1232 	 * If the GUID of the referenced dataset is different from the
1233 	 * GUID of the target dataset, find the referenced dataset.
1234 	 */
1235 	if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
1236 		gmesrch.guid = drrwbr->drr_refguid;
1237 		if ((gmep = avl_find(ra->guid_to_ds_map, &gmesrch,
1238 		    &where)) == NULL) {
1239 			return (EINVAL);
1240 		}
1241 		if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
1242 			return (EINVAL);
1243 	} else {
1244 		ref_os = os;
1245 	}
1246 
1247 	if (err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
1248 	    drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH))
1249 		return (err);
1250 
1251 	tx = dmu_tx_create(os);
1252 
1253 	dmu_tx_hold_write(tx, drrwbr->drr_object,
1254 	    drrwbr->drr_offset, drrwbr->drr_length);
1255 	err = dmu_tx_assign(tx, TXG_WAIT);
1256 	if (err) {
1257 		dmu_tx_abort(tx);
1258 		return (err);
1259 	}
1260 	dmu_write(os, drrwbr->drr_object,
1261 	    drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
1262 	dmu_buf_rele(dbp, FTAG);
1263 	dmu_tx_commit(tx);
1264 	return (0);
1265 }
1266 
1267 static int
1268 restore_spill(struct restorearg *ra, objset_t *os, struct drr_spill *drrs)
1269 {
1270 	dmu_tx_t *tx;
1271 	void *data;
1272 	dmu_buf_t *db, *db_spill;
1273 	int err;
1274 
1275 	if (drrs->drr_length < SPA_MINBLOCKSIZE ||
1276 	    drrs->drr_length > SPA_MAXBLOCKSIZE)
1277 		return (EINVAL);
1278 
1279 	data = restore_read(ra, drrs->drr_length);
1280 	if (data == NULL)
1281 		return (ra->err);
1282 
1283 	if (dmu_object_info(os, drrs->drr_object, NULL) != 0)
1284 		return (EINVAL);
1285 
1286 	VERIFY(0 == dmu_bonus_hold(os, drrs->drr_object, FTAG, &db));
1287 	if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
1288 		dmu_buf_rele(db, FTAG);
1289 		return (err);
1290 	}
1291 
1292 	tx = dmu_tx_create(os);
1293 
1294 	dmu_tx_hold_spill(tx, db->db_object);
1295 
1296 	err = dmu_tx_assign(tx, TXG_WAIT);
1297 	if (err) {
1298 		dmu_buf_rele(db, FTAG);
1299 		dmu_buf_rele(db_spill, FTAG);
1300 		dmu_tx_abort(tx);
1301 		return (err);
1302 	}
1303 	dmu_buf_will_dirty(db_spill, tx);
1304 
1305 	if (db_spill->db_size < drrs->drr_length)
1306 		VERIFY(0 == dbuf_spill_set_blksz(db_spill,
1307 		    drrs->drr_length, tx));
1308 	bcopy(data, db_spill->db_data, drrs->drr_length);
1309 
1310 	dmu_buf_rele(db, FTAG);
1311 	dmu_buf_rele(db_spill, FTAG);
1312 
1313 	dmu_tx_commit(tx);
1314 	return (0);
1315 }
1316 
1317 /* ARGSUSED */
1318 static int
1319 restore_free(struct restorearg *ra, objset_t *os,
1320     struct drr_free *drrf)
1321 {
1322 	int err;
1323 
1324 	if (drrf->drr_length != -1ULL &&
1325 	    drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
1326 		return (EINVAL);
1327 
1328 	if (dmu_object_info(os, drrf->drr_object, NULL) != 0)
1329 		return (EINVAL);
1330 
1331 	err = dmu_free_long_range(os, drrf->drr_object,
1332 	    drrf->drr_offset, drrf->drr_length);
1333 	return (err);
1334 }
1335 
1336 /*
1337  * NB: callers *must* call dmu_recv_end() if this succeeds.
1338  */
1339 int
1340 dmu_recv_stream(dmu_recv_cookie_t *drc, vnode_t *vp, offset_t *voffp,
1341     int cleanup_fd, uint64_t *action_handlep)
1342 {
1343 	struct restorearg ra = { 0 };
1344 	dmu_replay_record_t *drr;
1345 	objset_t *os;
1346 	zio_cksum_t pcksum;
1347 	int featureflags;
1348 
1349 	if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC))
1350 		ra.byteswap = TRUE;
1351 
1352 	{
1353 		/* compute checksum of drr_begin record */
1354 		dmu_replay_record_t *drr;
1355 		drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
1356 
1357 		drr->drr_type = DRR_BEGIN;
1358 		drr->drr_u.drr_begin = *drc->drc_drrb;
1359 		if (ra.byteswap) {
1360 			fletcher_4_incremental_byteswap(drr,
1361 			    sizeof (dmu_replay_record_t), &ra.cksum);
1362 		} else {
1363 			fletcher_4_incremental_native(drr,
1364 			    sizeof (dmu_replay_record_t), &ra.cksum);
1365 		}
1366 		kmem_free(drr, sizeof (dmu_replay_record_t));
1367 	}
1368 
1369 	if (ra.byteswap) {
1370 		struct drr_begin *drrb = drc->drc_drrb;
1371 		drrb->drr_magic = BSWAP_64(drrb->drr_magic);
1372 		drrb->drr_versioninfo = BSWAP_64(drrb->drr_versioninfo);
1373 		drrb->drr_creation_time = BSWAP_64(drrb->drr_creation_time);
1374 		drrb->drr_type = BSWAP_32(drrb->drr_type);
1375 		drrb->drr_toguid = BSWAP_64(drrb->drr_toguid);
1376 		drrb->drr_fromguid = BSWAP_64(drrb->drr_fromguid);
1377 	}
1378 
1379 	ra.vp = vp;
1380 	ra.voff = *voffp;
1381 	ra.bufsize = 1<<20;
1382 	ra.buf = kmem_alloc(ra.bufsize, KM_SLEEP);
1383 
1384 	/* these were verified in dmu_recv_begin */
1385 	ASSERT(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo) ==
1386 	    DMU_SUBSTREAM);
1387 	ASSERT(drc->drc_drrb->drr_type < DMU_OST_NUMTYPES);
1388 
1389 	/*
1390 	 * Open the objset we are modifying.
1391 	 */
1392 	VERIFY(dmu_objset_from_ds(drc->drc_real_ds, &os) == 0);
1393 
1394 	ASSERT(drc->drc_real_ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT);
1395 
1396 	featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
1397 
1398 	/* if this stream is dedup'ed, set up the avl tree for guid mapping */
1399 	if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
1400 		minor_t minor;
1401 
1402 		if (cleanup_fd == -1) {
1403 			ra.err = EBADF;
1404 			goto out;
1405 		}
1406 		ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
1407 		if (ra.err) {
1408 			cleanup_fd = -1;
1409 			goto out;
1410 		}
1411 
1412 		if (*action_handlep == 0) {
1413 			ra.guid_to_ds_map =
1414 			    kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
1415 			avl_create(ra.guid_to_ds_map, guid_compare,
1416 			    sizeof (guid_map_entry_t),
1417 			    offsetof(guid_map_entry_t, avlnode));
1418 			ra.err = zfs_onexit_add_cb(minor,
1419 			    free_guid_map_onexit, ra.guid_to_ds_map,
1420 			    action_handlep);
1421 			if (ra.err)
1422 				goto out;
1423 		} else {
1424 			ra.err = zfs_onexit_cb_data(minor, *action_handlep,
1425 			    (void **)&ra.guid_to_ds_map);
1426 			if (ra.err)
1427 				goto out;
1428 		}
1429 
1430 		drc->drc_guid_to_ds_map = ra.guid_to_ds_map;
1431 	}
1432 
1433 	/*
1434 	 * Read records and process them.
1435 	 */
1436 	pcksum = ra.cksum;
1437 	while (ra.err == 0 &&
1438 	    NULL != (drr = restore_read(&ra, sizeof (*drr)))) {
1439 		if (issig(JUSTLOOKING) && issig(FORREAL)) {
1440 			ra.err = EINTR;
1441 			goto out;
1442 		}
1443 
1444 		if (ra.byteswap)
1445 			backup_byteswap(drr);
1446 
1447 		switch (drr->drr_type) {
1448 		case DRR_OBJECT:
1449 		{
1450 			/*
1451 			 * We need to make a copy of the record header,
1452 			 * because restore_{object,write} may need to
1453 			 * restore_read(), which will invalidate drr.
1454 			 */
1455 			struct drr_object drro = drr->drr_u.drr_object;
1456 			ra.err = restore_object(&ra, os, &drro);
1457 			break;
1458 		}
1459 		case DRR_FREEOBJECTS:
1460 		{
1461 			struct drr_freeobjects drrfo =
1462 			    drr->drr_u.drr_freeobjects;
1463 			ra.err = restore_freeobjects(&ra, os, &drrfo);
1464 			break;
1465 		}
1466 		case DRR_WRITE:
1467 		{
1468 			struct drr_write drrw = drr->drr_u.drr_write;
1469 			ra.err = restore_write(&ra, os, &drrw);
1470 			break;
1471 		}
1472 		case DRR_WRITE_BYREF:
1473 		{
1474 			struct drr_write_byref drrwbr =
1475 			    drr->drr_u.drr_write_byref;
1476 			ra.err = restore_write_byref(&ra, os, &drrwbr);
1477 			break;
1478 		}
1479 		case DRR_FREE:
1480 		{
1481 			struct drr_free drrf = drr->drr_u.drr_free;
1482 			ra.err = restore_free(&ra, os, &drrf);
1483 			break;
1484 		}
1485 		case DRR_END:
1486 		{
1487 			struct drr_end drre = drr->drr_u.drr_end;
1488 			/*
1489 			 * We compare against the *previous* checksum
1490 			 * value, because the stored checksum is of
1491 			 * everything before the DRR_END record.
1492 			 */
1493 			if (!ZIO_CHECKSUM_EQUAL(drre.drr_checksum, pcksum))
1494 				ra.err = ECKSUM;
1495 			goto out;
1496 		}
1497 		case DRR_SPILL:
1498 		{
1499 			struct drr_spill drrs = drr->drr_u.drr_spill;
1500 			ra.err = restore_spill(&ra, os, &drrs);
1501 			break;
1502 		}
1503 		default:
1504 			ra.err = EINVAL;
1505 			goto out;
1506 		}
1507 		pcksum = ra.cksum;
1508 	}
1509 	ASSERT(ra.err != 0);
1510 
1511 out:
1512 	if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
1513 		zfs_onexit_fd_rele(cleanup_fd);
1514 
1515 	if (ra.err != 0) {
1516 		/*
1517 		 * destroy what we created, so we don't leave it in the
1518 		 * inconsistent restoring state.
1519 		 */
1520 		txg_wait_synced(drc->drc_real_ds->ds_dir->dd_pool, 0);
1521 
1522 		(void) dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag,
1523 		    B_FALSE);
1524 		if (drc->drc_real_ds != drc->drc_logical_ds) {
1525 			mutex_exit(&drc->drc_logical_ds->ds_recvlock);
1526 			dsl_dataset_rele(drc->drc_logical_ds, dmu_recv_tag);
1527 		}
1528 	}
1529 
1530 	kmem_free(ra.buf, ra.bufsize);
1531 	*voffp = ra.voff;
1532 	return (ra.err);
1533 }
1534 
1535 struct recvendsyncarg {
1536 	char *tosnap;
1537 	uint64_t creation_time;
1538 	uint64_t toguid;
1539 };
1540 
1541 static int
1542 recv_end_check(void *arg1, void *arg2, dmu_tx_t *tx)
1543 {
1544 	dsl_dataset_t *ds = arg1;
1545 	struct recvendsyncarg *resa = arg2;
1546 
1547 	return (dsl_dataset_snapshot_check(ds, resa->tosnap, tx));
1548 }
1549 
1550 static void
1551 recv_end_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1552 {
1553 	dsl_dataset_t *ds = arg1;
1554 	struct recvendsyncarg *resa = arg2;
1555 
1556 	dsl_dataset_snapshot_sync(ds, resa->tosnap, tx);
1557 
1558 	/* set snapshot's creation time and guid */
1559 	dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
1560 	ds->ds_prev->ds_phys->ds_creation_time = resa->creation_time;
1561 	ds->ds_prev->ds_phys->ds_guid = resa->toguid;
1562 	ds->ds_prev->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1563 
1564 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1565 	ds->ds_phys->ds_flags &= ~DS_FLAG_INCONSISTENT;
1566 }
1567 
1568 static int
1569 add_ds_to_guidmap(avl_tree_t *guid_map, dsl_dataset_t *ds)
1570 {
1571 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1572 	uint64_t snapobj = ds->ds_phys->ds_prev_snap_obj;
1573 	dsl_dataset_t *snapds;
1574 	guid_map_entry_t *gmep;
1575 	int err;
1576 
1577 	ASSERT(guid_map != NULL);
1578 
1579 	rw_enter(&dp->dp_config_rwlock, RW_READER);
1580 	err = dsl_dataset_hold_obj(dp, snapobj, guid_map, &snapds);
1581 	if (err == 0) {
1582 		gmep = kmem_alloc(sizeof (guid_map_entry_t), KM_SLEEP);
1583 		gmep->guid = snapds->ds_phys->ds_guid;
1584 		gmep->gme_ds = snapds;
1585 		avl_add(guid_map, gmep);
1586 	}
1587 
1588 	rw_exit(&dp->dp_config_rwlock);
1589 	return (err);
1590 }
1591 
1592 static int
1593 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
1594 {
1595 	struct recvendsyncarg resa;
1596 	dsl_dataset_t *ds = drc->drc_logical_ds;
1597 	int err, myerr;
1598 
1599 	/*
1600 	 * XXX hack; seems the ds is still dirty and dsl_pool_zil_clean()
1601 	 * expects it to have a ds_user_ptr (and zil), but clone_swap()
1602 	 * can close it.
1603 	 */
1604 	txg_wait_synced(ds->ds_dir->dd_pool, 0);
1605 
1606 	if (dsl_dataset_tryown(ds, FALSE, dmu_recv_tag)) {
1607 		err = dsl_dataset_clone_swap(drc->drc_real_ds, ds,
1608 		    drc->drc_force);
1609 		if (err)
1610 			goto out;
1611 	} else {
1612 		mutex_exit(&ds->ds_recvlock);
1613 		dsl_dataset_rele(ds, dmu_recv_tag);
1614 		(void) dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag,
1615 		    B_FALSE);
1616 		return (EBUSY);
1617 	}
1618 
1619 	resa.creation_time = drc->drc_drrb->drr_creation_time;
1620 	resa.toguid = drc->drc_drrb->drr_toguid;
1621 	resa.tosnap = drc->drc_tosnap;
1622 
1623 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1624 	    recv_end_check, recv_end_sync, ds, &resa, 3);
1625 	if (err) {
1626 		/* swap back */
1627 		(void) dsl_dataset_clone_swap(drc->drc_real_ds, ds, B_TRUE);
1628 	}
1629 
1630 out:
1631 	mutex_exit(&ds->ds_recvlock);
1632 	if (err == 0 && drc->drc_guid_to_ds_map != NULL)
1633 		(void) add_ds_to_guidmap(drc->drc_guid_to_ds_map, ds);
1634 	dsl_dataset_disown(ds, dmu_recv_tag);
1635 	myerr = dsl_dataset_destroy(drc->drc_real_ds, dmu_recv_tag, B_FALSE);
1636 	ASSERT3U(myerr, ==, 0);
1637 	return (err);
1638 }
1639 
1640 static int
1641 dmu_recv_new_end(dmu_recv_cookie_t *drc)
1642 {
1643 	struct recvendsyncarg resa;
1644 	dsl_dataset_t *ds = drc->drc_logical_ds;
1645 	int err;
1646 
1647 	/*
1648 	 * XXX hack; seems the ds is still dirty and dsl_pool_zil_clean()
1649 	 * expects it to have a ds_user_ptr (and zil), but clone_swap()
1650 	 * can close it.
1651 	 */
1652 	txg_wait_synced(ds->ds_dir->dd_pool, 0);
1653 
1654 	resa.creation_time = drc->drc_drrb->drr_creation_time;
1655 	resa.toguid = drc->drc_drrb->drr_toguid;
1656 	resa.tosnap = drc->drc_tosnap;
1657 
1658 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1659 	    recv_end_check, recv_end_sync, ds, &resa, 3);
1660 	if (err) {
1661 		/* clean up the fs we just recv'd into */
1662 		(void) dsl_dataset_destroy(ds, dmu_recv_tag, B_FALSE);
1663 	} else {
1664 		if (drc->drc_guid_to_ds_map != NULL)
1665 			(void) add_ds_to_guidmap(drc->drc_guid_to_ds_map, ds);
1666 		/* release the hold from dmu_recv_begin */
1667 		dsl_dataset_disown(ds, dmu_recv_tag);
1668 	}
1669 	return (err);
1670 }
1671 
1672 int
1673 dmu_recv_end(dmu_recv_cookie_t *drc)
1674 {
1675 	if (drc->drc_logical_ds != drc->drc_real_ds)
1676 		return (dmu_recv_existing_end(drc));
1677 	else
1678 		return (dmu_recv_new_end(drc));
1679 }
1680