xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu_send.c (revision 43da549e0419e2943d8bd143ebfb1db11a16c569)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24  * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
25  * Copyright (c) 2014, Joyent, Inc. All rights reserved.
26  * Copyright 2014 HybridCluster. All rights reserved.
27  * Copyright 2016 RackTop Systems.
28  * Copyright (c) 2014 Integros [integros.com]
29  */
30 
31 #include <sys/dmu.h>
32 #include <sys/dmu_impl.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/dbuf.h>
35 #include <sys/dnode.h>
36 #include <sys/zfs_context.h>
37 #include <sys/dmu_objset.h>
38 #include <sys/dmu_traverse.h>
39 #include <sys/dsl_dataset.h>
40 #include <sys/dsl_dir.h>
41 #include <sys/dsl_prop.h>
42 #include <sys/dsl_pool.h>
43 #include <sys/dsl_synctask.h>
44 #include <sys/zfs_ioctl.h>
45 #include <sys/zap.h>
46 #include <sys/zio_checksum.h>
47 #include <sys/zfs_znode.h>
48 #include <zfs_fletcher.h>
49 #include <sys/avl.h>
50 #include <sys/ddt.h>
51 #include <sys/zfs_onexit.h>
52 #include <sys/dmu_send.h>
53 #include <sys/dsl_destroy.h>
54 #include <sys/blkptr.h>
55 #include <sys/dsl_bookmark.h>
56 #include <sys/zfeature.h>
57 #include <sys/bqueue.h>
58 
59 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
60 int zfs_send_corrupt_data = B_FALSE;
61 int zfs_send_queue_length = 16 * 1024 * 1024;
62 int zfs_recv_queue_length = 16 * 1024 * 1024;
63 /* Set this tunable to FALSE to disable setting of DRR_FLAG_FREERECORDS */
64 int zfs_send_set_freerecords_bit = B_TRUE;
65 
66 static char *dmu_recv_tag = "dmu_recv_tag";
67 const char *recv_clone_name = "%recv";
68 
69 #define	BP_SPAN(datablkszsec, indblkshift, level) \
70 	(((uint64_t)datablkszsec) << (SPA_MINBLOCKSHIFT + \
71 	(level) * (indblkshift - SPA_BLKPTRSHIFT)))
72 
73 static void byteswap_record(dmu_replay_record_t *drr);
74 
75 struct send_thread_arg {
76 	bqueue_t	q;
77 	dsl_dataset_t	*ds;		/* Dataset to traverse */
78 	uint64_t	fromtxg;	/* Traverse from this txg */
79 	int		flags;		/* flags to pass to traverse_dataset */
80 	int		error_code;
81 	boolean_t	cancel;
82 	zbookmark_phys_t resume;
83 };
84 
85 struct send_block_record {
86 	boolean_t		eos_marker; /* Marks the end of the stream */
87 	blkptr_t		bp;
88 	zbookmark_phys_t	zb;
89 	uint8_t			indblkshift;
90 	uint16_t		datablkszsec;
91 	bqueue_node_t		ln;
92 };
93 
94 static int
95 dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
96 {
97 	dsl_dataset_t *ds = dmu_objset_ds(dsp->dsa_os);
98 	ssize_t resid; /* have to get resid to get detailed errno */
99 
100 	/*
101 	 * The code does not rely on this (len being a multiple of 8).  We keep
102 	 * this assertion because of the corresponding assertion in
103 	 * receive_read().  Keeping this assertion ensures that we do not
104 	 * inadvertently break backwards compatibility (causing the assertion
105 	 * in receive_read() to trigger on old software).
106 	 *
107 	 * Removing the assertions could be rolled into a new feature that uses
108 	 * data that isn't 8-byte aligned; if the assertions were removed, a
109 	 * feature flag would have to be added.
110 	 */
111 
112 	ASSERT0(len % 8);
113 
114 	dsp->dsa_err = vn_rdwr(UIO_WRITE, dsp->dsa_vp,
115 	    (caddr_t)buf, len,
116 	    0, UIO_SYSSPACE, FAPPEND, RLIM64_INFINITY, CRED(), &resid);
117 
118 	mutex_enter(&ds->ds_sendstream_lock);
119 	*dsp->dsa_off += len;
120 	mutex_exit(&ds->ds_sendstream_lock);
121 
122 	return (dsp->dsa_err);
123 }
124 
125 /*
126  * For all record types except BEGIN, fill in the checksum (overlaid in
127  * drr_u.drr_checksum.drr_checksum).  The checksum verifies everything
128  * up to the start of the checksum itself.
129  */
130 static int
131 dump_record(dmu_sendarg_t *dsp, void *payload, int payload_len)
132 {
133 	ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
134 	    ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
135 	fletcher_4_incremental_native(dsp->dsa_drr,
136 	    offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
137 	    &dsp->dsa_zc);
138 	if (dsp->dsa_drr->drr_type != DRR_BEGIN) {
139 		ASSERT(ZIO_CHECKSUM_IS_ZERO(&dsp->dsa_drr->drr_u.
140 		    drr_checksum.drr_checksum));
141 		dsp->dsa_drr->drr_u.drr_checksum.drr_checksum = dsp->dsa_zc;
142 	}
143 	fletcher_4_incremental_native(&dsp->dsa_drr->
144 	    drr_u.drr_checksum.drr_checksum,
145 	    sizeof (zio_cksum_t), &dsp->dsa_zc);
146 	if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
147 		return (SET_ERROR(EINTR));
148 	if (payload_len != 0) {
149 		fletcher_4_incremental_native(payload, payload_len,
150 		    &dsp->dsa_zc);
151 		if (dump_bytes(dsp, payload, payload_len) != 0)
152 			return (SET_ERROR(EINTR));
153 	}
154 	return (0);
155 }
156 
157 /*
158  * Fill in the drr_free struct, or perform aggregation if the previous record is
159  * also a free record, and the two are adjacent.
160  *
161  * Note that we send free records even for a full send, because we want to be
162  * able to receive a full send as a clone, which requires a list of all the free
163  * and freeobject records that were generated on the source.
164  */
165 static int
166 dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
167     uint64_t length)
168 {
169 	struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free);
170 
171 	/*
172 	 * When we receive a free record, dbuf_free_range() assumes
173 	 * that the receiving system doesn't have any dbufs in the range
174 	 * being freed.  This is always true because there is a one-record
175 	 * constraint: we only send one WRITE record for any given
176 	 * object,offset.  We know that the one-record constraint is
177 	 * true because we always send data in increasing order by
178 	 * object,offset.
179 	 *
180 	 * If the increasing-order constraint ever changes, we should find
181 	 * another way to assert that the one-record constraint is still
182 	 * satisfied.
183 	 */
184 	ASSERT(object > dsp->dsa_last_data_object ||
185 	    (object == dsp->dsa_last_data_object &&
186 	    offset > dsp->dsa_last_data_offset));
187 
188 	if (length != -1ULL && offset + length < offset)
189 		length = -1ULL;
190 
191 	/*
192 	 * If there is a pending op, but it's not PENDING_FREE, push it out,
193 	 * since free block aggregation can only be done for blocks of the
194 	 * same type (i.e., DRR_FREE records can only be aggregated with
195 	 * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
196 	 * aggregated with other DRR_FREEOBJECTS records.
197 	 */
198 	if (dsp->dsa_pending_op != PENDING_NONE &&
199 	    dsp->dsa_pending_op != PENDING_FREE) {
200 		if (dump_record(dsp, NULL, 0) != 0)
201 			return (SET_ERROR(EINTR));
202 		dsp->dsa_pending_op = PENDING_NONE;
203 	}
204 
205 	if (dsp->dsa_pending_op == PENDING_FREE) {
206 		/*
207 		 * There should never be a PENDING_FREE if length is -1
208 		 * (because dump_dnode is the only place where this
209 		 * function is called with a -1, and only after flushing
210 		 * any pending record).
211 		 */
212 		ASSERT(length != -1ULL);
213 		/*
214 		 * Check to see whether this free block can be aggregated
215 		 * with pending one.
216 		 */
217 		if (drrf->drr_object == object && drrf->drr_offset +
218 		    drrf->drr_length == offset) {
219 			drrf->drr_length += length;
220 			return (0);
221 		} else {
222 			/* not a continuation.  Push out pending record */
223 			if (dump_record(dsp, NULL, 0) != 0)
224 				return (SET_ERROR(EINTR));
225 			dsp->dsa_pending_op = PENDING_NONE;
226 		}
227 	}
228 	/* create a FREE record and make it pending */
229 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
230 	dsp->dsa_drr->drr_type = DRR_FREE;
231 	drrf->drr_object = object;
232 	drrf->drr_offset = offset;
233 	drrf->drr_length = length;
234 	drrf->drr_toguid = dsp->dsa_toguid;
235 	if (length == -1ULL) {
236 		if (dump_record(dsp, NULL, 0) != 0)
237 			return (SET_ERROR(EINTR));
238 	} else {
239 		dsp->dsa_pending_op = PENDING_FREE;
240 	}
241 
242 	return (0);
243 }
244 
245 static int
246 dump_write(dmu_sendarg_t *dsp, dmu_object_type_t type,
247     uint64_t object, uint64_t offset, int blksz, const blkptr_t *bp, void *data)
248 {
249 	struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write);
250 
251 	/*
252 	 * We send data in increasing object, offset order.
253 	 * See comment in dump_free() for details.
254 	 */
255 	ASSERT(object > dsp->dsa_last_data_object ||
256 	    (object == dsp->dsa_last_data_object &&
257 	    offset > dsp->dsa_last_data_offset));
258 	dsp->dsa_last_data_object = object;
259 	dsp->dsa_last_data_offset = offset + blksz - 1;
260 
261 	/*
262 	 * If there is any kind of pending aggregation (currently either
263 	 * a grouping of free objects or free blocks), push it out to
264 	 * the stream, since aggregation can't be done across operations
265 	 * of different types.
266 	 */
267 	if (dsp->dsa_pending_op != PENDING_NONE) {
268 		if (dump_record(dsp, NULL, 0) != 0)
269 			return (SET_ERROR(EINTR));
270 		dsp->dsa_pending_op = PENDING_NONE;
271 	}
272 	/* write a WRITE record */
273 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
274 	dsp->dsa_drr->drr_type = DRR_WRITE;
275 	drrw->drr_object = object;
276 	drrw->drr_type = type;
277 	drrw->drr_offset = offset;
278 	drrw->drr_length = blksz;
279 	drrw->drr_toguid = dsp->dsa_toguid;
280 	if (bp == NULL || BP_IS_EMBEDDED(bp)) {
281 		/*
282 		 * There's no pre-computed checksum for partial-block
283 		 * writes or embedded BP's, so (like
284 		 * fletcher4-checkummed blocks) userland will have to
285 		 * compute a dedup-capable checksum itself.
286 		 */
287 		drrw->drr_checksumtype = ZIO_CHECKSUM_OFF;
288 	} else {
289 		drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
290 		if (zio_checksum_table[drrw->drr_checksumtype].ci_flags &
291 		    ZCHECKSUM_FLAG_DEDUP)
292 			drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
293 		DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
294 		DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
295 		DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
296 		drrw->drr_key.ddk_cksum = bp->blk_cksum;
297 	}
298 
299 	if (dump_record(dsp, data, blksz) != 0)
300 		return (SET_ERROR(EINTR));
301 	return (0);
302 }
303 
304 static int
305 dump_write_embedded(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
306     int blksz, const blkptr_t *bp)
307 {
308 	char buf[BPE_PAYLOAD_SIZE];
309 	struct drr_write_embedded *drrw =
310 	    &(dsp->dsa_drr->drr_u.drr_write_embedded);
311 
312 	if (dsp->dsa_pending_op != PENDING_NONE) {
313 		if (dump_record(dsp, NULL, 0) != 0)
314 			return (EINTR);
315 		dsp->dsa_pending_op = PENDING_NONE;
316 	}
317 
318 	ASSERT(BP_IS_EMBEDDED(bp));
319 
320 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
321 	dsp->dsa_drr->drr_type = DRR_WRITE_EMBEDDED;
322 	drrw->drr_object = object;
323 	drrw->drr_offset = offset;
324 	drrw->drr_length = blksz;
325 	drrw->drr_toguid = dsp->dsa_toguid;
326 	drrw->drr_compression = BP_GET_COMPRESS(bp);
327 	drrw->drr_etype = BPE_GET_ETYPE(bp);
328 	drrw->drr_lsize = BPE_GET_LSIZE(bp);
329 	drrw->drr_psize = BPE_GET_PSIZE(bp);
330 
331 	decode_embedded_bp_compressed(bp, buf);
332 
333 	if (dump_record(dsp, buf, P2ROUNDUP(drrw->drr_psize, 8)) != 0)
334 		return (EINTR);
335 	return (0);
336 }
337 
338 static int
339 dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data)
340 {
341 	struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill);
342 
343 	if (dsp->dsa_pending_op != PENDING_NONE) {
344 		if (dump_record(dsp, NULL, 0) != 0)
345 			return (SET_ERROR(EINTR));
346 		dsp->dsa_pending_op = PENDING_NONE;
347 	}
348 
349 	/* write a SPILL record */
350 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
351 	dsp->dsa_drr->drr_type = DRR_SPILL;
352 	drrs->drr_object = object;
353 	drrs->drr_length = blksz;
354 	drrs->drr_toguid = dsp->dsa_toguid;
355 
356 	if (dump_record(dsp, data, blksz) != 0)
357 		return (SET_ERROR(EINTR));
358 	return (0);
359 }
360 
361 static int
362 dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
363 {
364 	struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
365 
366 	/*
367 	 * If there is a pending op, but it's not PENDING_FREEOBJECTS,
368 	 * push it out, since free block aggregation can only be done for
369 	 * blocks of the same type (i.e., DRR_FREE records can only be
370 	 * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
371 	 * can only be aggregated with other DRR_FREEOBJECTS records.
372 	 */
373 	if (dsp->dsa_pending_op != PENDING_NONE &&
374 	    dsp->dsa_pending_op != PENDING_FREEOBJECTS) {
375 		if (dump_record(dsp, NULL, 0) != 0)
376 			return (SET_ERROR(EINTR));
377 		dsp->dsa_pending_op = PENDING_NONE;
378 	}
379 	if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) {
380 		/*
381 		 * See whether this free object array can be aggregated
382 		 * with pending one
383 		 */
384 		if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
385 			drrfo->drr_numobjs += numobjs;
386 			return (0);
387 		} else {
388 			/* can't be aggregated.  Push out pending record */
389 			if (dump_record(dsp, NULL, 0) != 0)
390 				return (SET_ERROR(EINTR));
391 			dsp->dsa_pending_op = PENDING_NONE;
392 		}
393 	}
394 
395 	/* write a FREEOBJECTS record */
396 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
397 	dsp->dsa_drr->drr_type = DRR_FREEOBJECTS;
398 	drrfo->drr_firstobj = firstobj;
399 	drrfo->drr_numobjs = numobjs;
400 	drrfo->drr_toguid = dsp->dsa_toguid;
401 
402 	dsp->dsa_pending_op = PENDING_FREEOBJECTS;
403 
404 	return (0);
405 }
406 
407 static int
408 dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
409 {
410 	struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object);
411 
412 	if (object < dsp->dsa_resume_object) {
413 		/*
414 		 * Note: when resuming, we will visit all the dnodes in
415 		 * the block of dnodes that we are resuming from.  In
416 		 * this case it's unnecessary to send the dnodes prior to
417 		 * the one we are resuming from.  We should be at most one
418 		 * block's worth of dnodes behind the resume point.
419 		 */
420 		ASSERT3U(dsp->dsa_resume_object - object, <,
421 		    1 << (DNODE_BLOCK_SHIFT - DNODE_SHIFT));
422 		return (0);
423 	}
424 
425 	if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
426 		return (dump_freeobjects(dsp, object, 1));
427 
428 	if (dsp->dsa_pending_op != PENDING_NONE) {
429 		if (dump_record(dsp, NULL, 0) != 0)
430 			return (SET_ERROR(EINTR));
431 		dsp->dsa_pending_op = PENDING_NONE;
432 	}
433 
434 	/* write an OBJECT record */
435 	bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
436 	dsp->dsa_drr->drr_type = DRR_OBJECT;
437 	drro->drr_object = object;
438 	drro->drr_type = dnp->dn_type;
439 	drro->drr_bonustype = dnp->dn_bonustype;
440 	drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
441 	drro->drr_bonuslen = dnp->dn_bonuslen;
442 	drro->drr_checksumtype = dnp->dn_checksum;
443 	drro->drr_compress = dnp->dn_compress;
444 	drro->drr_toguid = dsp->dsa_toguid;
445 
446 	if (!(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
447 	    drro->drr_blksz > SPA_OLD_MAXBLOCKSIZE)
448 		drro->drr_blksz = SPA_OLD_MAXBLOCKSIZE;
449 
450 	if (dump_record(dsp, DN_BONUS(dnp),
451 	    P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0) {
452 		return (SET_ERROR(EINTR));
453 	}
454 
455 	/* Free anything past the end of the file. */
456 	if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) *
457 	    (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0)
458 		return (SET_ERROR(EINTR));
459 	if (dsp->dsa_err != 0)
460 		return (SET_ERROR(EINTR));
461 	return (0);
462 }
463 
464 static boolean_t
465 backup_do_embed(dmu_sendarg_t *dsp, const blkptr_t *bp)
466 {
467 	if (!BP_IS_EMBEDDED(bp))
468 		return (B_FALSE);
469 
470 	/*
471 	 * Compression function must be legacy, or explicitly enabled.
472 	 */
473 	if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS &&
474 	    !(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4)))
475 		return (B_FALSE);
476 
477 	/*
478 	 * Embed type must be explicitly enabled.
479 	 */
480 	switch (BPE_GET_ETYPE(bp)) {
481 	case BP_EMBEDDED_TYPE_DATA:
482 		if (dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)
483 			return (B_TRUE);
484 		break;
485 	default:
486 		return (B_FALSE);
487 	}
488 	return (B_FALSE);
489 }
490 
491 /*
492  * This is the callback function to traverse_dataset that acts as the worker
493  * thread for dmu_send_impl.
494  */
495 /*ARGSUSED*/
496 static int
497 send_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
498     const zbookmark_phys_t *zb, const struct dnode_phys *dnp, void *arg)
499 {
500 	struct send_thread_arg *sta = arg;
501 	struct send_block_record *record;
502 	uint64_t record_size;
503 	int err = 0;
504 
505 	ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
506 	    zb->zb_object >= sta->resume.zb_object);
507 
508 	if (sta->cancel)
509 		return (SET_ERROR(EINTR));
510 
511 	if (bp == NULL) {
512 		ASSERT3U(zb->zb_level, ==, ZB_DNODE_LEVEL);
513 		return (0);
514 	} else if (zb->zb_level < 0) {
515 		return (0);
516 	}
517 
518 	record = kmem_zalloc(sizeof (struct send_block_record), KM_SLEEP);
519 	record->eos_marker = B_FALSE;
520 	record->bp = *bp;
521 	record->zb = *zb;
522 	record->indblkshift = dnp->dn_indblkshift;
523 	record->datablkszsec = dnp->dn_datablkszsec;
524 	record_size = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
525 	bqueue_enqueue(&sta->q, record, record_size);
526 
527 	return (err);
528 }
529 
530 /*
531  * This function kicks off the traverse_dataset.  It also handles setting the
532  * error code of the thread in case something goes wrong, and pushes the End of
533  * Stream record when the traverse_dataset call has finished.  If there is no
534  * dataset to traverse, the thread immediately pushes End of Stream marker.
535  */
536 static void
537 send_traverse_thread(void *arg)
538 {
539 	struct send_thread_arg *st_arg = arg;
540 	int err;
541 	struct send_block_record *data;
542 
543 	if (st_arg->ds != NULL) {
544 		err = traverse_dataset_resume(st_arg->ds,
545 		    st_arg->fromtxg, &st_arg->resume,
546 		    st_arg->flags, send_cb, st_arg);
547 
548 		if (err != EINTR)
549 			st_arg->error_code = err;
550 	}
551 	data = kmem_zalloc(sizeof (*data), KM_SLEEP);
552 	data->eos_marker = B_TRUE;
553 	bqueue_enqueue(&st_arg->q, data, 1);
554 }
555 
556 /*
557  * This function actually handles figuring out what kind of record needs to be
558  * dumped, reading the data (which has hopefully been prefetched), and calling
559  * the appropriate helper function.
560  */
561 static int
562 do_dump(dmu_sendarg_t *dsa, struct send_block_record *data)
563 {
564 	dsl_dataset_t *ds = dmu_objset_ds(dsa->dsa_os);
565 	const blkptr_t *bp = &data->bp;
566 	const zbookmark_phys_t *zb = &data->zb;
567 	uint8_t indblkshift = data->indblkshift;
568 	uint16_t dblkszsec = data->datablkszsec;
569 	spa_t *spa = ds->ds_dir->dd_pool->dp_spa;
570 	dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
571 	int err = 0;
572 
573 	ASSERT3U(zb->zb_level, >=, 0);
574 
575 	ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
576 	    zb->zb_object >= dsa->dsa_resume_object);
577 
578 	if (zb->zb_object != DMU_META_DNODE_OBJECT &&
579 	    DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
580 		return (0);
581 	} else if (BP_IS_HOLE(bp) &&
582 	    zb->zb_object == DMU_META_DNODE_OBJECT) {
583 		uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
584 		uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
585 		err = dump_freeobjects(dsa, dnobj, span >> DNODE_SHIFT);
586 	} else if (BP_IS_HOLE(bp)) {
587 		uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
588 		uint64_t offset = zb->zb_blkid * span;
589 		err = dump_free(dsa, zb->zb_object, offset, span);
590 	} else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
591 		return (0);
592 	} else if (type == DMU_OT_DNODE) {
593 		int blksz = BP_GET_LSIZE(bp);
594 		arc_flags_t aflags = ARC_FLAG_WAIT;
595 		arc_buf_t *abuf;
596 
597 		ASSERT0(zb->zb_level);
598 
599 		if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
600 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
601 		    &aflags, zb) != 0)
602 			return (SET_ERROR(EIO));
603 
604 		dnode_phys_t *blk = abuf->b_data;
605 		uint64_t dnobj = zb->zb_blkid * (blksz >> DNODE_SHIFT);
606 		for (int i = 0; i < blksz >> DNODE_SHIFT; i++) {
607 			err = dump_dnode(dsa, dnobj + i, blk + i);
608 			if (err != 0)
609 				break;
610 		}
611 		(void) arc_buf_remove_ref(abuf, &abuf);
612 	} else if (type == DMU_OT_SA) {
613 		arc_flags_t aflags = ARC_FLAG_WAIT;
614 		arc_buf_t *abuf;
615 		int blksz = BP_GET_LSIZE(bp);
616 
617 		if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
618 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
619 		    &aflags, zb) != 0)
620 			return (SET_ERROR(EIO));
621 
622 		err = dump_spill(dsa, zb->zb_object, blksz, abuf->b_data);
623 		(void) arc_buf_remove_ref(abuf, &abuf);
624 	} else if (backup_do_embed(dsa, bp)) {
625 		/* it's an embedded level-0 block of a regular object */
626 		int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
627 		ASSERT0(zb->zb_level);
628 		err = dump_write_embedded(dsa, zb->zb_object,
629 		    zb->zb_blkid * blksz, blksz, bp);
630 	} else {
631 		/* it's a level-0 block of a regular object */
632 		arc_flags_t aflags = ARC_FLAG_WAIT;
633 		arc_buf_t *abuf;
634 		int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
635 		uint64_t offset;
636 
637 		ASSERT0(zb->zb_level);
638 		ASSERT(zb->zb_object > dsa->dsa_resume_object ||
639 		    (zb->zb_object == dsa->dsa_resume_object &&
640 		    zb->zb_blkid * blksz >= dsa->dsa_resume_offset));
641 
642 		if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
643 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
644 		    &aflags, zb) != 0) {
645 			if (zfs_send_corrupt_data) {
646 				/* Send a block filled with 0x"zfs badd bloc" */
647 				abuf = arc_buf_alloc(spa, blksz, &abuf,
648 				    ARC_BUFC_DATA);
649 				uint64_t *ptr;
650 				for (ptr = abuf->b_data;
651 				    (char *)ptr < (char *)abuf->b_data + blksz;
652 				    ptr++)
653 					*ptr = 0x2f5baddb10cULL;
654 			} else {
655 				return (SET_ERROR(EIO));
656 			}
657 		}
658 
659 		offset = zb->zb_blkid * blksz;
660 
661 		if (!(dsa->dsa_featureflags &
662 		    DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
663 		    blksz > SPA_OLD_MAXBLOCKSIZE) {
664 			char *buf = abuf->b_data;
665 			while (blksz > 0 && err == 0) {
666 				int n = MIN(blksz, SPA_OLD_MAXBLOCKSIZE);
667 				err = dump_write(dsa, type, zb->zb_object,
668 				    offset, n, NULL, buf);
669 				offset += n;
670 				buf += n;
671 				blksz -= n;
672 			}
673 		} else {
674 			err = dump_write(dsa, type, zb->zb_object,
675 			    offset, blksz, bp, abuf->b_data);
676 		}
677 		(void) arc_buf_remove_ref(abuf, &abuf);
678 	}
679 
680 	ASSERT(err == 0 || err == EINTR);
681 	return (err);
682 }
683 
684 /*
685  * Pop the new data off the queue, and free the old data.
686  */
687 static struct send_block_record *
688 get_next_record(bqueue_t *bq, struct send_block_record *data)
689 {
690 	struct send_block_record *tmp = bqueue_dequeue(bq);
691 	kmem_free(data, sizeof (*data));
692 	return (tmp);
693 }
694 
695 /*
696  * Actually do the bulk of the work in a zfs send.
697  *
698  * Note: Releases dp using the specified tag.
699  */
700 static int
701 dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *to_ds,
702     zfs_bookmark_phys_t *ancestor_zb,
703     boolean_t is_clone, boolean_t embedok, boolean_t large_block_ok, int outfd,
704     uint64_t resumeobj, uint64_t resumeoff,
705     vnode_t *vp, offset_t *off)
706 {
707 	objset_t *os;
708 	dmu_replay_record_t *drr;
709 	dmu_sendarg_t *dsp;
710 	int err;
711 	uint64_t fromtxg = 0;
712 	uint64_t featureflags = 0;
713 	struct send_thread_arg to_arg = { 0 };
714 
715 	err = dmu_objset_from_ds(to_ds, &os);
716 	if (err != 0) {
717 		dsl_pool_rele(dp, tag);
718 		return (err);
719 	}
720 
721 	drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
722 	drr->drr_type = DRR_BEGIN;
723 	drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
724 	DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
725 	    DMU_SUBSTREAM);
726 
727 #ifdef _KERNEL
728 	if (dmu_objset_type(os) == DMU_OST_ZFS) {
729 		uint64_t version;
730 		if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0) {
731 			kmem_free(drr, sizeof (dmu_replay_record_t));
732 			dsl_pool_rele(dp, tag);
733 			return (SET_ERROR(EINVAL));
734 		}
735 		if (version >= ZPL_VERSION_SA) {
736 			featureflags |= DMU_BACKUP_FEATURE_SA_SPILL;
737 		}
738 	}
739 #endif
740 
741 	if (large_block_ok && to_ds->ds_feature_inuse[SPA_FEATURE_LARGE_BLOCKS])
742 		featureflags |= DMU_BACKUP_FEATURE_LARGE_BLOCKS;
743 	if (embedok &&
744 	    spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) {
745 		featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA;
746 		if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
747 			featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA_LZ4;
748 	}
749 
750 	if (resumeobj != 0 || resumeoff != 0) {
751 		featureflags |= DMU_BACKUP_FEATURE_RESUMING;
752 	}
753 
754 	DMU_SET_FEATUREFLAGS(drr->drr_u.drr_begin.drr_versioninfo,
755 	    featureflags);
756 
757 	drr->drr_u.drr_begin.drr_creation_time =
758 	    dsl_dataset_phys(to_ds)->ds_creation_time;
759 	drr->drr_u.drr_begin.drr_type = dmu_objset_type(os);
760 	if (is_clone)
761 		drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
762 	drr->drr_u.drr_begin.drr_toguid = dsl_dataset_phys(to_ds)->ds_guid;
763 	if (dsl_dataset_phys(to_ds)->ds_flags & DS_FLAG_CI_DATASET)
764 		drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
765 	if (zfs_send_set_freerecords_bit)
766 		drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_FREERECORDS;
767 
768 	if (ancestor_zb != NULL) {
769 		drr->drr_u.drr_begin.drr_fromguid =
770 		    ancestor_zb->zbm_guid;
771 		fromtxg = ancestor_zb->zbm_creation_txg;
772 	}
773 	dsl_dataset_name(to_ds, drr->drr_u.drr_begin.drr_toname);
774 	if (!to_ds->ds_is_snapshot) {
775 		(void) strlcat(drr->drr_u.drr_begin.drr_toname, "@--head--",
776 		    sizeof (drr->drr_u.drr_begin.drr_toname));
777 	}
778 
779 	dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP);
780 
781 	dsp->dsa_drr = drr;
782 	dsp->dsa_vp = vp;
783 	dsp->dsa_outfd = outfd;
784 	dsp->dsa_proc = curproc;
785 	dsp->dsa_os = os;
786 	dsp->dsa_off = off;
787 	dsp->dsa_toguid = dsl_dataset_phys(to_ds)->ds_guid;
788 	dsp->dsa_pending_op = PENDING_NONE;
789 	dsp->dsa_featureflags = featureflags;
790 	dsp->dsa_resume_object = resumeobj;
791 	dsp->dsa_resume_offset = resumeoff;
792 
793 	mutex_enter(&to_ds->ds_sendstream_lock);
794 	list_insert_head(&to_ds->ds_sendstreams, dsp);
795 	mutex_exit(&to_ds->ds_sendstream_lock);
796 
797 	dsl_dataset_long_hold(to_ds, FTAG);
798 	dsl_pool_rele(dp, tag);
799 
800 	void *payload = NULL;
801 	size_t payload_len = 0;
802 	if (resumeobj != 0 || resumeoff != 0) {
803 		dmu_object_info_t to_doi;
804 		err = dmu_object_info(os, resumeobj, &to_doi);
805 		if (err != 0)
806 			goto out;
807 		SET_BOOKMARK(&to_arg.resume, to_ds->ds_object, resumeobj, 0,
808 		    resumeoff / to_doi.doi_data_block_size);
809 
810 		nvlist_t *nvl = fnvlist_alloc();
811 		fnvlist_add_uint64(nvl, "resume_object", resumeobj);
812 		fnvlist_add_uint64(nvl, "resume_offset", resumeoff);
813 		payload = fnvlist_pack(nvl, &payload_len);
814 		drr->drr_payloadlen = payload_len;
815 		fnvlist_free(nvl);
816 	}
817 
818 	err = dump_record(dsp, payload, payload_len);
819 	fnvlist_pack_free(payload, payload_len);
820 	if (err != 0) {
821 		err = dsp->dsa_err;
822 		goto out;
823 	}
824 
825 	err = bqueue_init(&to_arg.q, zfs_send_queue_length,
826 	    offsetof(struct send_block_record, ln));
827 	to_arg.error_code = 0;
828 	to_arg.cancel = B_FALSE;
829 	to_arg.ds = to_ds;
830 	to_arg.fromtxg = fromtxg;
831 	to_arg.flags = TRAVERSE_PRE | TRAVERSE_PREFETCH;
832 	(void) thread_create(NULL, 0, send_traverse_thread, &to_arg, 0, curproc,
833 	    TS_RUN, minclsyspri);
834 
835 	struct send_block_record *to_data;
836 	to_data = bqueue_dequeue(&to_arg.q);
837 
838 	while (!to_data->eos_marker && err == 0) {
839 		err = do_dump(dsp, to_data);
840 		to_data = get_next_record(&to_arg.q, to_data);
841 		if (issig(JUSTLOOKING) && issig(FORREAL))
842 			err = EINTR;
843 	}
844 
845 	if (err != 0) {
846 		to_arg.cancel = B_TRUE;
847 		while (!to_data->eos_marker) {
848 			to_data = get_next_record(&to_arg.q, to_data);
849 		}
850 	}
851 	kmem_free(to_data, sizeof (*to_data));
852 
853 	bqueue_destroy(&to_arg.q);
854 
855 	if (err == 0 && to_arg.error_code != 0)
856 		err = to_arg.error_code;
857 
858 	if (err != 0)
859 		goto out;
860 
861 	if (dsp->dsa_pending_op != PENDING_NONE)
862 		if (dump_record(dsp, NULL, 0) != 0)
863 			err = SET_ERROR(EINTR);
864 
865 	if (err != 0) {
866 		if (err == EINTR && dsp->dsa_err != 0)
867 			err = dsp->dsa_err;
868 		goto out;
869 	}
870 
871 	bzero(drr, sizeof (dmu_replay_record_t));
872 	drr->drr_type = DRR_END;
873 	drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc;
874 	drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid;
875 
876 	if (dump_record(dsp, NULL, 0) != 0)
877 		err = dsp->dsa_err;
878 
879 out:
880 	mutex_enter(&to_ds->ds_sendstream_lock);
881 	list_remove(&to_ds->ds_sendstreams, dsp);
882 	mutex_exit(&to_ds->ds_sendstream_lock);
883 
884 	kmem_free(drr, sizeof (dmu_replay_record_t));
885 	kmem_free(dsp, sizeof (dmu_sendarg_t));
886 
887 	dsl_dataset_long_rele(to_ds, FTAG);
888 
889 	return (err);
890 }
891 
892 int
893 dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
894     boolean_t embedok, boolean_t large_block_ok,
895     int outfd, vnode_t *vp, offset_t *off)
896 {
897 	dsl_pool_t *dp;
898 	dsl_dataset_t *ds;
899 	dsl_dataset_t *fromds = NULL;
900 	int err;
901 
902 	err = dsl_pool_hold(pool, FTAG, &dp);
903 	if (err != 0)
904 		return (err);
905 
906 	err = dsl_dataset_hold_obj(dp, tosnap, FTAG, &ds);
907 	if (err != 0) {
908 		dsl_pool_rele(dp, FTAG);
909 		return (err);
910 	}
911 
912 	if (fromsnap != 0) {
913 		zfs_bookmark_phys_t zb;
914 		boolean_t is_clone;
915 
916 		err = dsl_dataset_hold_obj(dp, fromsnap, FTAG, &fromds);
917 		if (err != 0) {
918 			dsl_dataset_rele(ds, FTAG);
919 			dsl_pool_rele(dp, FTAG);
920 			return (err);
921 		}
922 		if (!dsl_dataset_is_before(ds, fromds, 0))
923 			err = SET_ERROR(EXDEV);
924 		zb.zbm_creation_time =
925 		    dsl_dataset_phys(fromds)->ds_creation_time;
926 		zb.zbm_creation_txg = dsl_dataset_phys(fromds)->ds_creation_txg;
927 		zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
928 		is_clone = (fromds->ds_dir != ds->ds_dir);
929 		dsl_dataset_rele(fromds, FTAG);
930 		err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
931 		    embedok, large_block_ok, outfd, 0, 0, vp, off);
932 	} else {
933 		err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
934 		    embedok, large_block_ok, outfd, 0, 0, vp, off);
935 	}
936 	dsl_dataset_rele(ds, FTAG);
937 	return (err);
938 }
939 
940 int
941 dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok,
942     boolean_t large_block_ok, int outfd, uint64_t resumeobj, uint64_t resumeoff,
943     vnode_t *vp, offset_t *off)
944 {
945 	dsl_pool_t *dp;
946 	dsl_dataset_t *ds;
947 	int err;
948 	boolean_t owned = B_FALSE;
949 
950 	if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL)
951 		return (SET_ERROR(EINVAL));
952 
953 	err = dsl_pool_hold(tosnap, FTAG, &dp);
954 	if (err != 0)
955 		return (err);
956 
957 	if (strchr(tosnap, '@') == NULL && spa_writeable(dp->dp_spa)) {
958 		/*
959 		 * We are sending a filesystem or volume.  Ensure
960 		 * that it doesn't change by owning the dataset.
961 		 */
962 		err = dsl_dataset_own(dp, tosnap, FTAG, &ds);
963 		owned = B_TRUE;
964 	} else {
965 		err = dsl_dataset_hold(dp, tosnap, FTAG, &ds);
966 	}
967 	if (err != 0) {
968 		dsl_pool_rele(dp, FTAG);
969 		return (err);
970 	}
971 
972 	if (fromsnap != NULL) {
973 		zfs_bookmark_phys_t zb;
974 		boolean_t is_clone = B_FALSE;
975 		int fsnamelen = strchr(tosnap, '@') - tosnap;
976 
977 		/*
978 		 * If the fromsnap is in a different filesystem, then
979 		 * mark the send stream as a clone.
980 		 */
981 		if (strncmp(tosnap, fromsnap, fsnamelen) != 0 ||
982 		    (fromsnap[fsnamelen] != '@' &&
983 		    fromsnap[fsnamelen] != '#')) {
984 			is_clone = B_TRUE;
985 		}
986 
987 		if (strchr(fromsnap, '@')) {
988 			dsl_dataset_t *fromds;
989 			err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds);
990 			if (err == 0) {
991 				if (!dsl_dataset_is_before(ds, fromds, 0))
992 					err = SET_ERROR(EXDEV);
993 				zb.zbm_creation_time =
994 				    dsl_dataset_phys(fromds)->ds_creation_time;
995 				zb.zbm_creation_txg =
996 				    dsl_dataset_phys(fromds)->ds_creation_txg;
997 				zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
998 				is_clone = (ds->ds_dir != fromds->ds_dir);
999 				dsl_dataset_rele(fromds, FTAG);
1000 			}
1001 		} else {
1002 			err = dsl_bookmark_lookup(dp, fromsnap, ds, &zb);
1003 		}
1004 		if (err != 0) {
1005 			dsl_dataset_rele(ds, FTAG);
1006 			dsl_pool_rele(dp, FTAG);
1007 			return (err);
1008 		}
1009 		err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
1010 		    embedok, large_block_ok,
1011 		    outfd, resumeobj, resumeoff, vp, off);
1012 	} else {
1013 		err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
1014 		    embedok, large_block_ok,
1015 		    outfd, resumeobj, resumeoff, vp, off);
1016 	}
1017 	if (owned)
1018 		dsl_dataset_disown(ds, FTAG);
1019 	else
1020 		dsl_dataset_rele(ds, FTAG);
1021 	return (err);
1022 }
1023 
1024 static int
1025 dmu_adjust_send_estimate_for_indirects(dsl_dataset_t *ds, uint64_t size,
1026     uint64_t *sizep)
1027 {
1028 	int err;
1029 	/*
1030 	 * Assume that space (both on-disk and in-stream) is dominated by
1031 	 * data.  We will adjust for indirect blocks and the copies property,
1032 	 * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
1033 	 */
1034 
1035 	/*
1036 	 * Subtract out approximate space used by indirect blocks.
1037 	 * Assume most space is used by data blocks (non-indirect, non-dnode).
1038 	 * Assume all blocks are recordsize.  Assume ditto blocks and
1039 	 * internal fragmentation counter out compression.
1040 	 *
1041 	 * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
1042 	 * block, which we observe in practice.
1043 	 */
1044 	uint64_t recordsize;
1045 	err = dsl_prop_get_int_ds(ds, "recordsize", &recordsize);
1046 	if (err != 0)
1047 		return (err);
1048 	size -= size / recordsize * sizeof (blkptr_t);
1049 
1050 	/* Add in the space for the record associated with each block. */
1051 	size += size / recordsize * sizeof (dmu_replay_record_t);
1052 
1053 	*sizep = size;
1054 
1055 	return (0);
1056 }
1057 
1058 int
1059 dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds, uint64_t *sizep)
1060 {
1061 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1062 	int err;
1063 	uint64_t size;
1064 
1065 	ASSERT(dsl_pool_config_held(dp));
1066 
1067 	/* tosnap must be a snapshot */
1068 	if (!ds->ds_is_snapshot)
1069 		return (SET_ERROR(EINVAL));
1070 
1071 	/* fromsnap, if provided, must be a snapshot */
1072 	if (fromds != NULL && !fromds->ds_is_snapshot)
1073 		return (SET_ERROR(EINVAL));
1074 
1075 	/*
1076 	 * fromsnap must be an earlier snapshot from the same fs as tosnap,
1077 	 * or the origin's fs.
1078 	 */
1079 	if (fromds != NULL && !dsl_dataset_is_before(ds, fromds, 0))
1080 		return (SET_ERROR(EXDEV));
1081 
1082 	/* Get uncompressed size estimate of changed data. */
1083 	if (fromds == NULL) {
1084 		size = dsl_dataset_phys(ds)->ds_uncompressed_bytes;
1085 	} else {
1086 		uint64_t used, comp;
1087 		err = dsl_dataset_space_written(fromds, ds,
1088 		    &used, &comp, &size);
1089 		if (err != 0)
1090 			return (err);
1091 	}
1092 
1093 	err = dmu_adjust_send_estimate_for_indirects(ds, size, sizep);
1094 	return (err);
1095 }
1096 
1097 /*
1098  * Simple callback used to traverse the blocks of a snapshot and sum their
1099  * uncompressed size
1100  */
1101 /* ARGSUSED */
1102 static int
1103 dmu_calculate_send_traversal(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
1104     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
1105 {
1106 	uint64_t *spaceptr = arg;
1107 	if (bp != NULL && !BP_IS_HOLE(bp)) {
1108 		*spaceptr += BP_GET_UCSIZE(bp);
1109 	}
1110 	return (0);
1111 }
1112 
1113 /*
1114  * Given a desination snapshot and a TXG, calculate the approximate size of a
1115  * send stream sent from that TXG. from_txg may be zero, indicating that the
1116  * whole snapshot will be sent.
1117  */
1118 int
1119 dmu_send_estimate_from_txg(dsl_dataset_t *ds, uint64_t from_txg,
1120     uint64_t *sizep)
1121 {
1122 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1123 	int err;
1124 	uint64_t size = 0;
1125 
1126 	ASSERT(dsl_pool_config_held(dp));
1127 
1128 	/* tosnap must be a snapshot */
1129 	if (!dsl_dataset_is_snapshot(ds))
1130 		return (SET_ERROR(EINVAL));
1131 
1132 	/* verify that from_txg is before the provided snapshot was taken */
1133 	if (from_txg >= dsl_dataset_phys(ds)->ds_creation_txg) {
1134 		return (SET_ERROR(EXDEV));
1135 	}
1136 
1137 	/*
1138 	 * traverse the blocks of the snapshot with birth times after
1139 	 * from_txg, summing their uncompressed size
1140 	 */
1141 	err = traverse_dataset(ds, from_txg, TRAVERSE_POST,
1142 	    dmu_calculate_send_traversal, &size);
1143 	if (err)
1144 		return (err);
1145 
1146 	err = dmu_adjust_send_estimate_for_indirects(ds, size, sizep);
1147 	return (err);
1148 }
1149 
1150 typedef struct dmu_recv_begin_arg {
1151 	const char *drba_origin;
1152 	dmu_recv_cookie_t *drba_cookie;
1153 	cred_t *drba_cred;
1154 	uint64_t drba_snapobj;
1155 } dmu_recv_begin_arg_t;
1156 
1157 static int
1158 recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds,
1159     uint64_t fromguid)
1160 {
1161 	uint64_t val;
1162 	int error;
1163 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1164 
1165 	/* temporary clone name must not exist */
1166 	error = zap_lookup(dp->dp_meta_objset,
1167 	    dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, recv_clone_name,
1168 	    8, 1, &val);
1169 	if (error != ENOENT)
1170 		return (error == 0 ? EBUSY : error);
1171 
1172 	/* new snapshot name must not exist */
1173 	error = zap_lookup(dp->dp_meta_objset,
1174 	    dsl_dataset_phys(ds)->ds_snapnames_zapobj,
1175 	    drba->drba_cookie->drc_tosnap, 8, 1, &val);
1176 	if (error != ENOENT)
1177 		return (error == 0 ? EEXIST : error);
1178 
1179 	/*
1180 	 * Check snapshot limit before receiving. We'll recheck again at the
1181 	 * end, but might as well abort before receiving if we're already over
1182 	 * the limit.
1183 	 *
1184 	 * Note that we do not check the file system limit with
1185 	 * dsl_dir_fscount_check because the temporary %clones don't count
1186 	 * against that limit.
1187 	 */
1188 	error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT,
1189 	    NULL, drba->drba_cred);
1190 	if (error != 0)
1191 		return (error);
1192 
1193 	if (fromguid != 0) {
1194 		dsl_dataset_t *snap;
1195 		uint64_t obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
1196 
1197 		/* Find snapshot in this dir that matches fromguid. */
1198 		while (obj != 0) {
1199 			error = dsl_dataset_hold_obj(dp, obj, FTAG,
1200 			    &snap);
1201 			if (error != 0)
1202 				return (SET_ERROR(ENODEV));
1203 			if (snap->ds_dir != ds->ds_dir) {
1204 				dsl_dataset_rele(snap, FTAG);
1205 				return (SET_ERROR(ENODEV));
1206 			}
1207 			if (dsl_dataset_phys(snap)->ds_guid == fromguid)
1208 				break;
1209 			obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
1210 			dsl_dataset_rele(snap, FTAG);
1211 		}
1212 		if (obj == 0)
1213 			return (SET_ERROR(ENODEV));
1214 
1215 		if (drba->drba_cookie->drc_force) {
1216 			drba->drba_snapobj = obj;
1217 		} else {
1218 			/*
1219 			 * If we are not forcing, there must be no
1220 			 * changes since fromsnap.
1221 			 */
1222 			if (dsl_dataset_modified_since_snap(ds, snap)) {
1223 				dsl_dataset_rele(snap, FTAG);
1224 				return (SET_ERROR(ETXTBSY));
1225 			}
1226 			drba->drba_snapobj = ds->ds_prev->ds_object;
1227 		}
1228 
1229 		dsl_dataset_rele(snap, FTAG);
1230 	} else {
1231 		/* if full, then must be forced */
1232 		if (!drba->drba_cookie->drc_force)
1233 			return (SET_ERROR(EEXIST));
1234 		/* start from $ORIGIN@$ORIGIN, if supported */
1235 		drba->drba_snapobj = dp->dp_origin_snap != NULL ?
1236 		    dp->dp_origin_snap->ds_object : 0;
1237 	}
1238 
1239 	return (0);
1240 
1241 }
1242 
1243 static int
1244 dmu_recv_begin_check(void *arg, dmu_tx_t *tx)
1245 {
1246 	dmu_recv_begin_arg_t *drba = arg;
1247 	dsl_pool_t *dp = dmu_tx_pool(tx);
1248 	struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1249 	uint64_t fromguid = drrb->drr_fromguid;
1250 	int flags = drrb->drr_flags;
1251 	int error;
1252 	uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1253 	dsl_dataset_t *ds;
1254 	const char *tofs = drba->drba_cookie->drc_tofs;
1255 
1256 	/* already checked */
1257 	ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1258 	ASSERT(!(featureflags & DMU_BACKUP_FEATURE_RESUMING));
1259 
1260 	if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1261 	    DMU_COMPOUNDSTREAM ||
1262 	    drrb->drr_type >= DMU_OST_NUMTYPES ||
1263 	    ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL))
1264 		return (SET_ERROR(EINVAL));
1265 
1266 	/* Verify pool version supports SA if SA_SPILL feature set */
1267 	if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1268 	    spa_version(dp->dp_spa) < SPA_VERSION_SA)
1269 		return (SET_ERROR(ENOTSUP));
1270 
1271 	if (drba->drba_cookie->drc_resumable &&
1272 	    !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EXTENSIBLE_DATASET))
1273 		return (SET_ERROR(ENOTSUP));
1274 
1275 	/*
1276 	 * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1277 	 * record to a plan WRITE record, so the pool must have the
1278 	 * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1279 	 * records.  Same with WRITE_EMBEDDED records that use LZ4 compression.
1280 	 */
1281 	if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1282 	    !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1283 		return (SET_ERROR(ENOTSUP));
1284 	if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4) &&
1285 	    !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1286 		return (SET_ERROR(ENOTSUP));
1287 
1288 	/*
1289 	 * The receiving code doesn't know how to translate large blocks
1290 	 * to smaller ones, so the pool must have the LARGE_BLOCKS
1291 	 * feature enabled if the stream has LARGE_BLOCKS.
1292 	 */
1293 	if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
1294 	    !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS))
1295 		return (SET_ERROR(ENOTSUP));
1296 
1297 	error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1298 	if (error == 0) {
1299 		/* target fs already exists; recv into temp clone */
1300 
1301 		/* Can't recv a clone into an existing fs */
1302 		if (flags & DRR_FLAG_CLONE || drba->drba_origin) {
1303 			dsl_dataset_rele(ds, FTAG);
1304 			return (SET_ERROR(EINVAL));
1305 		}
1306 
1307 		error = recv_begin_check_existing_impl(drba, ds, fromguid);
1308 		dsl_dataset_rele(ds, FTAG);
1309 	} else if (error == ENOENT) {
1310 		/* target fs does not exist; must be a full backup or clone */
1311 		char buf[MAXNAMELEN];
1312 
1313 		/*
1314 		 * If it's a non-clone incremental, we are missing the
1315 		 * target fs, so fail the recv.
1316 		 */
1317 		if (fromguid != 0 && !(flags & DRR_FLAG_CLONE ||
1318 		    drba->drba_origin))
1319 			return (SET_ERROR(ENOENT));
1320 
1321 		/*
1322 		 * If we're receiving a full send as a clone, and it doesn't
1323 		 * contain all the necessary free records and freeobject
1324 		 * records, reject it.
1325 		 */
1326 		if (fromguid == 0 && drba->drba_origin &&
1327 		    !(flags & DRR_FLAG_FREERECORDS))
1328 			return (SET_ERROR(EINVAL));
1329 
1330 		/* Open the parent of tofs */
1331 		ASSERT3U(strlen(tofs), <, MAXNAMELEN);
1332 		(void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1);
1333 		error = dsl_dataset_hold(dp, buf, FTAG, &ds);
1334 		if (error != 0)
1335 			return (error);
1336 
1337 		/*
1338 		 * Check filesystem and snapshot limits before receiving. We'll
1339 		 * recheck snapshot limits again at the end (we create the
1340 		 * filesystems and increment those counts during begin_sync).
1341 		 */
1342 		error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1343 		    ZFS_PROP_FILESYSTEM_LIMIT, NULL, drba->drba_cred);
1344 		if (error != 0) {
1345 			dsl_dataset_rele(ds, FTAG);
1346 			return (error);
1347 		}
1348 
1349 		error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1350 		    ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred);
1351 		if (error != 0) {
1352 			dsl_dataset_rele(ds, FTAG);
1353 			return (error);
1354 		}
1355 
1356 		if (drba->drba_origin != NULL) {
1357 			dsl_dataset_t *origin;
1358 			error = dsl_dataset_hold(dp, drba->drba_origin,
1359 			    FTAG, &origin);
1360 			if (error != 0) {
1361 				dsl_dataset_rele(ds, FTAG);
1362 				return (error);
1363 			}
1364 			if (!origin->ds_is_snapshot) {
1365 				dsl_dataset_rele(origin, FTAG);
1366 				dsl_dataset_rele(ds, FTAG);
1367 				return (SET_ERROR(EINVAL));
1368 			}
1369 			if (dsl_dataset_phys(origin)->ds_guid != fromguid &&
1370 			    fromguid != 0) {
1371 				dsl_dataset_rele(origin, FTAG);
1372 				dsl_dataset_rele(ds, FTAG);
1373 				return (SET_ERROR(ENODEV));
1374 			}
1375 			dsl_dataset_rele(origin, FTAG);
1376 		}
1377 		dsl_dataset_rele(ds, FTAG);
1378 		error = 0;
1379 	}
1380 	return (error);
1381 }
1382 
1383 static void
1384 dmu_recv_begin_sync(void *arg, dmu_tx_t *tx)
1385 {
1386 	dmu_recv_begin_arg_t *drba = arg;
1387 	dsl_pool_t *dp = dmu_tx_pool(tx);
1388 	objset_t *mos = dp->dp_meta_objset;
1389 	struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1390 	const char *tofs = drba->drba_cookie->drc_tofs;
1391 	dsl_dataset_t *ds, *newds;
1392 	uint64_t dsobj;
1393 	int error;
1394 	uint64_t crflags = 0;
1395 
1396 	if (drrb->drr_flags & DRR_FLAG_CI_DATA)
1397 		crflags |= DS_FLAG_CI_DATASET;
1398 
1399 	error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1400 	if (error == 0) {
1401 		/* create temporary clone */
1402 		dsl_dataset_t *snap = NULL;
1403 		if (drba->drba_snapobj != 0) {
1404 			VERIFY0(dsl_dataset_hold_obj(dp,
1405 			    drba->drba_snapobj, FTAG, &snap));
1406 		}
1407 		dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name,
1408 		    snap, crflags, drba->drba_cred, tx);
1409 		if (drba->drba_snapobj != 0)
1410 			dsl_dataset_rele(snap, FTAG);
1411 		dsl_dataset_rele(ds, FTAG);
1412 	} else {
1413 		dsl_dir_t *dd;
1414 		const char *tail;
1415 		dsl_dataset_t *origin = NULL;
1416 
1417 		VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail));
1418 
1419 		if (drba->drba_origin != NULL) {
1420 			VERIFY0(dsl_dataset_hold(dp, drba->drba_origin,
1421 			    FTAG, &origin));
1422 		}
1423 
1424 		/* Create new dataset. */
1425 		dsobj = dsl_dataset_create_sync(dd,
1426 		    strrchr(tofs, '/') + 1,
1427 		    origin, crflags, drba->drba_cred, tx);
1428 		if (origin != NULL)
1429 			dsl_dataset_rele(origin, FTAG);
1430 		dsl_dir_rele(dd, FTAG);
1431 		drba->drba_cookie->drc_newfs = B_TRUE;
1432 	}
1433 	VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds));
1434 
1435 	if (drba->drba_cookie->drc_resumable) {
1436 		dsl_dataset_zapify(newds, tx);
1437 		if (drrb->drr_fromguid != 0) {
1438 			VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_FROMGUID,
1439 			    8, 1, &drrb->drr_fromguid, tx));
1440 		}
1441 		VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TOGUID,
1442 		    8, 1, &drrb->drr_toguid, tx));
1443 		VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TONAME,
1444 		    1, strlen(drrb->drr_toname) + 1, drrb->drr_toname, tx));
1445 		uint64_t one = 1;
1446 		uint64_t zero = 0;
1447 		VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OBJECT,
1448 		    8, 1, &one, tx));
1449 		VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OFFSET,
1450 		    8, 1, &zero, tx));
1451 		VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_BYTES,
1452 		    8, 1, &zero, tx));
1453 		if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
1454 		    DMU_BACKUP_FEATURE_EMBED_DATA) {
1455 			VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_EMBEDOK,
1456 			    8, 1, &one, tx));
1457 		}
1458 	}
1459 
1460 	dmu_buf_will_dirty(newds->ds_dbuf, tx);
1461 	dsl_dataset_phys(newds)->ds_flags |= DS_FLAG_INCONSISTENT;
1462 
1463 	/*
1464 	 * If we actually created a non-clone, we need to create the
1465 	 * objset in our new dataset.
1466 	 */
1467 	if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) {
1468 		(void) dmu_objset_create_impl(dp->dp_spa,
1469 		    newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx);
1470 	}
1471 
1472 	drba->drba_cookie->drc_ds = newds;
1473 
1474 	spa_history_log_internal_ds(newds, "receive", tx, "");
1475 }
1476 
1477 static int
1478 dmu_recv_resume_begin_check(void *arg, dmu_tx_t *tx)
1479 {
1480 	dmu_recv_begin_arg_t *drba = arg;
1481 	dsl_pool_t *dp = dmu_tx_pool(tx);
1482 	struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1483 	int error;
1484 	uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1485 	dsl_dataset_t *ds;
1486 	const char *tofs = drba->drba_cookie->drc_tofs;
1487 
1488 	/* already checked */
1489 	ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1490 	ASSERT(featureflags & DMU_BACKUP_FEATURE_RESUMING);
1491 
1492 	if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1493 	    DMU_COMPOUNDSTREAM ||
1494 	    drrb->drr_type >= DMU_OST_NUMTYPES)
1495 		return (SET_ERROR(EINVAL));
1496 
1497 	/* Verify pool version supports SA if SA_SPILL feature set */
1498 	if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1499 	    spa_version(dp->dp_spa) < SPA_VERSION_SA)
1500 		return (SET_ERROR(ENOTSUP));
1501 
1502 	/*
1503 	 * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1504 	 * record to a plain WRITE record, so the pool must have the
1505 	 * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1506 	 * records.  Same with WRITE_EMBEDDED records that use LZ4 compression.
1507 	 */
1508 	if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1509 	    !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1510 		return (SET_ERROR(ENOTSUP));
1511 	if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4) &&
1512 	    !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1513 		return (SET_ERROR(ENOTSUP));
1514 
1515 	char recvname[ZFS_MAXNAMELEN];
1516 
1517 	(void) snprintf(recvname, sizeof (recvname), "%s/%s",
1518 	    tofs, recv_clone_name);
1519 
1520 	if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
1521 		/* %recv does not exist; continue in tofs */
1522 		error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1523 		if (error != 0)
1524 			return (error);
1525 	}
1526 
1527 	/* check that ds is marked inconsistent */
1528 	if (!DS_IS_INCONSISTENT(ds)) {
1529 		dsl_dataset_rele(ds, FTAG);
1530 		return (SET_ERROR(EINVAL));
1531 	}
1532 
1533 	/* check that there is resuming data, and that the toguid matches */
1534 	if (!dsl_dataset_is_zapified(ds)) {
1535 		dsl_dataset_rele(ds, FTAG);
1536 		return (SET_ERROR(EINVAL));
1537 	}
1538 	uint64_t val;
1539 	error = zap_lookup(dp->dp_meta_objset, ds->ds_object,
1540 	    DS_FIELD_RESUME_TOGUID, sizeof (val), 1, &val);
1541 	if (error != 0 || drrb->drr_toguid != val) {
1542 		dsl_dataset_rele(ds, FTAG);
1543 		return (SET_ERROR(EINVAL));
1544 	}
1545 
1546 	/*
1547 	 * Check if the receive is still running.  If so, it will be owned.
1548 	 * Note that nothing else can own the dataset (e.g. after the receive
1549 	 * fails) because it will be marked inconsistent.
1550 	 */
1551 	if (dsl_dataset_has_owner(ds)) {
1552 		dsl_dataset_rele(ds, FTAG);
1553 		return (SET_ERROR(EBUSY));
1554 	}
1555 
1556 	/* There should not be any snapshots of this fs yet. */
1557 	if (ds->ds_prev != NULL && ds->ds_prev->ds_dir == ds->ds_dir) {
1558 		dsl_dataset_rele(ds, FTAG);
1559 		return (SET_ERROR(EINVAL));
1560 	}
1561 
1562 	/*
1563 	 * Note: resume point will be checked when we process the first WRITE
1564 	 * record.
1565 	 */
1566 
1567 	/* check that the origin matches */
1568 	val = 0;
1569 	(void) zap_lookup(dp->dp_meta_objset, ds->ds_object,
1570 	    DS_FIELD_RESUME_FROMGUID, sizeof (val), 1, &val);
1571 	if (drrb->drr_fromguid != val) {
1572 		dsl_dataset_rele(ds, FTAG);
1573 		return (SET_ERROR(EINVAL));
1574 	}
1575 
1576 	dsl_dataset_rele(ds, FTAG);
1577 	return (0);
1578 }
1579 
1580 static void
1581 dmu_recv_resume_begin_sync(void *arg, dmu_tx_t *tx)
1582 {
1583 	dmu_recv_begin_arg_t *drba = arg;
1584 	dsl_pool_t *dp = dmu_tx_pool(tx);
1585 	const char *tofs = drba->drba_cookie->drc_tofs;
1586 	dsl_dataset_t *ds;
1587 	uint64_t dsobj;
1588 	char recvname[ZFS_MAXNAMELEN];
1589 
1590 	(void) snprintf(recvname, sizeof (recvname), "%s/%s",
1591 	    tofs, recv_clone_name);
1592 
1593 	if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
1594 		/* %recv does not exist; continue in tofs */
1595 		VERIFY0(dsl_dataset_hold(dp, tofs, FTAG, &ds));
1596 		drba->drba_cookie->drc_newfs = B_TRUE;
1597 	}
1598 
1599 	/* clear the inconsistent flag so that we can own it */
1600 	ASSERT(DS_IS_INCONSISTENT(ds));
1601 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1602 	dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
1603 	dsobj = ds->ds_object;
1604 	dsl_dataset_rele(ds, FTAG);
1605 
1606 	VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &ds));
1607 
1608 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1609 	dsl_dataset_phys(ds)->ds_flags |= DS_FLAG_INCONSISTENT;
1610 
1611 	ASSERT(!BP_IS_HOLE(dsl_dataset_get_blkptr(ds)));
1612 
1613 	drba->drba_cookie->drc_ds = ds;
1614 
1615 	spa_history_log_internal_ds(ds, "resume receive", tx, "");
1616 }
1617 
1618 /*
1619  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
1620  * succeeds; otherwise we will leak the holds on the datasets.
1621  */
1622 int
1623 dmu_recv_begin(char *tofs, char *tosnap, dmu_replay_record_t *drr_begin,
1624     boolean_t force, boolean_t resumable, char *origin, dmu_recv_cookie_t *drc)
1625 {
1626 	dmu_recv_begin_arg_t drba = { 0 };
1627 
1628 	bzero(drc, sizeof (dmu_recv_cookie_t));
1629 	drc->drc_drr_begin = drr_begin;
1630 	drc->drc_drrb = &drr_begin->drr_u.drr_begin;
1631 	drc->drc_tosnap = tosnap;
1632 	drc->drc_tofs = tofs;
1633 	drc->drc_force = force;
1634 	drc->drc_resumable = resumable;
1635 	drc->drc_cred = CRED();
1636 
1637 	if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC)) {
1638 		drc->drc_byteswap = B_TRUE;
1639 		fletcher_4_incremental_byteswap(drr_begin,
1640 		    sizeof (dmu_replay_record_t), &drc->drc_cksum);
1641 		byteswap_record(drr_begin);
1642 	} else if (drc->drc_drrb->drr_magic == DMU_BACKUP_MAGIC) {
1643 		fletcher_4_incremental_native(drr_begin,
1644 		    sizeof (dmu_replay_record_t), &drc->drc_cksum);
1645 	} else {
1646 		return (SET_ERROR(EINVAL));
1647 	}
1648 
1649 	drba.drba_origin = origin;
1650 	drba.drba_cookie = drc;
1651 	drba.drba_cred = CRED();
1652 
1653 	if (DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo) &
1654 	    DMU_BACKUP_FEATURE_RESUMING) {
1655 		return (dsl_sync_task(tofs,
1656 		    dmu_recv_resume_begin_check, dmu_recv_resume_begin_sync,
1657 		    &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1658 	} else  {
1659 		return (dsl_sync_task(tofs,
1660 		    dmu_recv_begin_check, dmu_recv_begin_sync,
1661 		    &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1662 	}
1663 }
1664 
1665 struct receive_record_arg {
1666 	dmu_replay_record_t header;
1667 	void *payload; /* Pointer to a buffer containing the payload */
1668 	/*
1669 	 * If the record is a write, pointer to the arc_buf_t containing the
1670 	 * payload.
1671 	 */
1672 	arc_buf_t *write_buf;
1673 	int payload_size;
1674 	uint64_t bytes_read; /* bytes read from stream when record created */
1675 	boolean_t eos_marker; /* Marks the end of the stream */
1676 	bqueue_node_t node;
1677 };
1678 
1679 struct receive_writer_arg {
1680 	objset_t *os;
1681 	boolean_t byteswap;
1682 	bqueue_t q;
1683 
1684 	/*
1685 	 * These three args are used to signal to the main thread that we're
1686 	 * done.
1687 	 */
1688 	kmutex_t mutex;
1689 	kcondvar_t cv;
1690 	boolean_t done;
1691 
1692 	int err;
1693 	/* A map from guid to dataset to help handle dedup'd streams. */
1694 	avl_tree_t *guid_to_ds_map;
1695 	boolean_t resumable;
1696 	uint64_t last_object, last_offset;
1697 	uint64_t bytes_read; /* bytes read when current record created */
1698 };
1699 
1700 struct objlist {
1701 	list_t list; /* List of struct receive_objnode. */
1702 	/*
1703 	 * Last object looked up. Used to assert that objects are being looked
1704 	 * up in ascending order.
1705 	 */
1706 	uint64_t last_lookup;
1707 };
1708 
1709 struct receive_objnode {
1710 	list_node_t node;
1711 	uint64_t object;
1712 };
1713 
1714 struct receive_arg  {
1715 	objset_t *os;
1716 	vnode_t *vp; /* The vnode to read the stream from */
1717 	uint64_t voff; /* The current offset in the stream */
1718 	uint64_t bytes_read;
1719 	/*
1720 	 * A record that has had its payload read in, but hasn't yet been handed
1721 	 * off to the worker thread.
1722 	 */
1723 	struct receive_record_arg *rrd;
1724 	/* A record that has had its header read in, but not its payload. */
1725 	struct receive_record_arg *next_rrd;
1726 	zio_cksum_t cksum;
1727 	zio_cksum_t prev_cksum;
1728 	int err;
1729 	boolean_t byteswap;
1730 	/* Sorted list of objects not to issue prefetches for. */
1731 	struct objlist ignore_objlist;
1732 };
1733 
1734 typedef struct guid_map_entry {
1735 	uint64_t	guid;
1736 	dsl_dataset_t	*gme_ds;
1737 	avl_node_t	avlnode;
1738 } guid_map_entry_t;
1739 
1740 static int
1741 guid_compare(const void *arg1, const void *arg2)
1742 {
1743 	const guid_map_entry_t *gmep1 = arg1;
1744 	const guid_map_entry_t *gmep2 = arg2;
1745 
1746 	if (gmep1->guid < gmep2->guid)
1747 		return (-1);
1748 	else if (gmep1->guid > gmep2->guid)
1749 		return (1);
1750 	return (0);
1751 }
1752 
1753 static void
1754 free_guid_map_onexit(void *arg)
1755 {
1756 	avl_tree_t *ca = arg;
1757 	void *cookie = NULL;
1758 	guid_map_entry_t *gmep;
1759 
1760 	while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
1761 		dsl_dataset_long_rele(gmep->gme_ds, gmep);
1762 		dsl_dataset_rele(gmep->gme_ds, gmep);
1763 		kmem_free(gmep, sizeof (guid_map_entry_t));
1764 	}
1765 	avl_destroy(ca);
1766 	kmem_free(ca, sizeof (avl_tree_t));
1767 }
1768 
1769 static int
1770 receive_read(struct receive_arg *ra, int len, void *buf)
1771 {
1772 	int done = 0;
1773 
1774 	/*
1775 	 * The code doesn't rely on this (lengths being multiples of 8).  See
1776 	 * comment in dump_bytes.
1777 	 */
1778 	ASSERT0(len % 8);
1779 
1780 	while (done < len) {
1781 		ssize_t resid;
1782 
1783 		ra->err = vn_rdwr(UIO_READ, ra->vp,
1784 		    (char *)buf + done, len - done,
1785 		    ra->voff, UIO_SYSSPACE, FAPPEND,
1786 		    RLIM64_INFINITY, CRED(), &resid);
1787 
1788 		if (resid == len - done) {
1789 			/*
1790 			 * Note: ECKSUM indicates that the receive
1791 			 * was interrupted and can potentially be resumed.
1792 			 */
1793 			ra->err = SET_ERROR(ECKSUM);
1794 		}
1795 		ra->voff += len - done - resid;
1796 		done = len - resid;
1797 		if (ra->err != 0)
1798 			return (ra->err);
1799 	}
1800 
1801 	ra->bytes_read += len;
1802 
1803 	ASSERT3U(done, ==, len);
1804 	return (0);
1805 }
1806 
1807 static void
1808 byteswap_record(dmu_replay_record_t *drr)
1809 {
1810 #define	DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
1811 #define	DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
1812 	drr->drr_type = BSWAP_32(drr->drr_type);
1813 	drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
1814 
1815 	switch (drr->drr_type) {
1816 	case DRR_BEGIN:
1817 		DO64(drr_begin.drr_magic);
1818 		DO64(drr_begin.drr_versioninfo);
1819 		DO64(drr_begin.drr_creation_time);
1820 		DO32(drr_begin.drr_type);
1821 		DO32(drr_begin.drr_flags);
1822 		DO64(drr_begin.drr_toguid);
1823 		DO64(drr_begin.drr_fromguid);
1824 		break;
1825 	case DRR_OBJECT:
1826 		DO64(drr_object.drr_object);
1827 		DO32(drr_object.drr_type);
1828 		DO32(drr_object.drr_bonustype);
1829 		DO32(drr_object.drr_blksz);
1830 		DO32(drr_object.drr_bonuslen);
1831 		DO64(drr_object.drr_toguid);
1832 		break;
1833 	case DRR_FREEOBJECTS:
1834 		DO64(drr_freeobjects.drr_firstobj);
1835 		DO64(drr_freeobjects.drr_numobjs);
1836 		DO64(drr_freeobjects.drr_toguid);
1837 		break;
1838 	case DRR_WRITE:
1839 		DO64(drr_write.drr_object);
1840 		DO32(drr_write.drr_type);
1841 		DO64(drr_write.drr_offset);
1842 		DO64(drr_write.drr_length);
1843 		DO64(drr_write.drr_toguid);
1844 		ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write.drr_key.ddk_cksum);
1845 		DO64(drr_write.drr_key.ddk_prop);
1846 		break;
1847 	case DRR_WRITE_BYREF:
1848 		DO64(drr_write_byref.drr_object);
1849 		DO64(drr_write_byref.drr_offset);
1850 		DO64(drr_write_byref.drr_length);
1851 		DO64(drr_write_byref.drr_toguid);
1852 		DO64(drr_write_byref.drr_refguid);
1853 		DO64(drr_write_byref.drr_refobject);
1854 		DO64(drr_write_byref.drr_refoffset);
1855 		ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write_byref.
1856 		    drr_key.ddk_cksum);
1857 		DO64(drr_write_byref.drr_key.ddk_prop);
1858 		break;
1859 	case DRR_WRITE_EMBEDDED:
1860 		DO64(drr_write_embedded.drr_object);
1861 		DO64(drr_write_embedded.drr_offset);
1862 		DO64(drr_write_embedded.drr_length);
1863 		DO64(drr_write_embedded.drr_toguid);
1864 		DO32(drr_write_embedded.drr_lsize);
1865 		DO32(drr_write_embedded.drr_psize);
1866 		break;
1867 	case DRR_FREE:
1868 		DO64(drr_free.drr_object);
1869 		DO64(drr_free.drr_offset);
1870 		DO64(drr_free.drr_length);
1871 		DO64(drr_free.drr_toguid);
1872 		break;
1873 	case DRR_SPILL:
1874 		DO64(drr_spill.drr_object);
1875 		DO64(drr_spill.drr_length);
1876 		DO64(drr_spill.drr_toguid);
1877 		break;
1878 	case DRR_END:
1879 		DO64(drr_end.drr_toguid);
1880 		ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_end.drr_checksum);
1881 		break;
1882 	}
1883 
1884 	if (drr->drr_type != DRR_BEGIN) {
1885 		ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_checksum.drr_checksum);
1886 	}
1887 
1888 #undef DO64
1889 #undef DO32
1890 }
1891 
1892 static inline uint8_t
1893 deduce_nblkptr(dmu_object_type_t bonus_type, uint64_t bonus_size)
1894 {
1895 	if (bonus_type == DMU_OT_SA) {
1896 		return (1);
1897 	} else {
1898 		return (1 +
1899 		    ((DN_MAX_BONUSLEN - bonus_size) >> SPA_BLKPTRSHIFT));
1900 	}
1901 }
1902 
1903 static void
1904 save_resume_state(struct receive_writer_arg *rwa,
1905     uint64_t object, uint64_t offset, dmu_tx_t *tx)
1906 {
1907 	int txgoff = dmu_tx_get_txg(tx) & TXG_MASK;
1908 
1909 	if (!rwa->resumable)
1910 		return;
1911 
1912 	/*
1913 	 * We use ds_resume_bytes[] != 0 to indicate that we need to
1914 	 * update this on disk, so it must not be 0.
1915 	 */
1916 	ASSERT(rwa->bytes_read != 0);
1917 
1918 	/*
1919 	 * We only resume from write records, which have a valid
1920 	 * (non-meta-dnode) object number.
1921 	 */
1922 	ASSERT(object != 0);
1923 
1924 	/*
1925 	 * For resuming to work correctly, we must receive records in order,
1926 	 * sorted by object,offset.  This is checked by the callers, but
1927 	 * assert it here for good measure.
1928 	 */
1929 	ASSERT3U(object, >=, rwa->os->os_dsl_dataset->ds_resume_object[txgoff]);
1930 	ASSERT(object != rwa->os->os_dsl_dataset->ds_resume_object[txgoff] ||
1931 	    offset >= rwa->os->os_dsl_dataset->ds_resume_offset[txgoff]);
1932 	ASSERT3U(rwa->bytes_read, >=,
1933 	    rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff]);
1934 
1935 	rwa->os->os_dsl_dataset->ds_resume_object[txgoff] = object;
1936 	rwa->os->os_dsl_dataset->ds_resume_offset[txgoff] = offset;
1937 	rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff] = rwa->bytes_read;
1938 }
1939 
1940 static int
1941 receive_object(struct receive_writer_arg *rwa, struct drr_object *drro,
1942     void *data)
1943 {
1944 	dmu_object_info_t doi;
1945 	dmu_tx_t *tx;
1946 	uint64_t object;
1947 	int err;
1948 
1949 	if (drro->drr_type == DMU_OT_NONE ||
1950 	    !DMU_OT_IS_VALID(drro->drr_type) ||
1951 	    !DMU_OT_IS_VALID(drro->drr_bonustype) ||
1952 	    drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
1953 	    drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
1954 	    P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
1955 	    drro->drr_blksz < SPA_MINBLOCKSIZE ||
1956 	    drro->drr_blksz > spa_maxblocksize(dmu_objset_spa(rwa->os)) ||
1957 	    drro->drr_bonuslen > DN_MAX_BONUSLEN) {
1958 		return (SET_ERROR(EINVAL));
1959 	}
1960 
1961 	err = dmu_object_info(rwa->os, drro->drr_object, &doi);
1962 
1963 	if (err != 0 && err != ENOENT)
1964 		return (SET_ERROR(EINVAL));
1965 	object = err == 0 ? drro->drr_object : DMU_NEW_OBJECT;
1966 
1967 	/*
1968 	 * If we are losing blkptrs or changing the block size this must
1969 	 * be a new file instance.  We must clear out the previous file
1970 	 * contents before we can change this type of metadata in the dnode.
1971 	 */
1972 	if (err == 0) {
1973 		int nblkptr;
1974 
1975 		nblkptr = deduce_nblkptr(drro->drr_bonustype,
1976 		    drro->drr_bonuslen);
1977 
1978 		if (drro->drr_blksz != doi.doi_data_block_size ||
1979 		    nblkptr < doi.doi_nblkptr) {
1980 			err = dmu_free_long_range(rwa->os, drro->drr_object,
1981 			    0, DMU_OBJECT_END);
1982 			if (err != 0)
1983 				return (SET_ERROR(EINVAL));
1984 		}
1985 	}
1986 
1987 	tx = dmu_tx_create(rwa->os);
1988 	dmu_tx_hold_bonus(tx, object);
1989 	err = dmu_tx_assign(tx, TXG_WAIT);
1990 	if (err != 0) {
1991 		dmu_tx_abort(tx);
1992 		return (err);
1993 	}
1994 
1995 	if (object == DMU_NEW_OBJECT) {
1996 		/* currently free, want to be allocated */
1997 		err = dmu_object_claim(rwa->os, drro->drr_object,
1998 		    drro->drr_type, drro->drr_blksz,
1999 		    drro->drr_bonustype, drro->drr_bonuslen, tx);
2000 	} else if (drro->drr_type != doi.doi_type ||
2001 	    drro->drr_blksz != doi.doi_data_block_size ||
2002 	    drro->drr_bonustype != doi.doi_bonus_type ||
2003 	    drro->drr_bonuslen != doi.doi_bonus_size) {
2004 		/* currently allocated, but with different properties */
2005 		err = dmu_object_reclaim(rwa->os, drro->drr_object,
2006 		    drro->drr_type, drro->drr_blksz,
2007 		    drro->drr_bonustype, drro->drr_bonuslen, tx);
2008 	}
2009 	if (err != 0) {
2010 		dmu_tx_commit(tx);
2011 		return (SET_ERROR(EINVAL));
2012 	}
2013 
2014 	dmu_object_set_checksum(rwa->os, drro->drr_object,
2015 	    drro->drr_checksumtype, tx);
2016 	dmu_object_set_compress(rwa->os, drro->drr_object,
2017 	    drro->drr_compress, tx);
2018 
2019 	if (data != NULL) {
2020 		dmu_buf_t *db;
2021 
2022 		VERIFY0(dmu_bonus_hold(rwa->os, drro->drr_object, FTAG, &db));
2023 		dmu_buf_will_dirty(db, tx);
2024 
2025 		ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
2026 		bcopy(data, db->db_data, drro->drr_bonuslen);
2027 		if (rwa->byteswap) {
2028 			dmu_object_byteswap_t byteswap =
2029 			    DMU_OT_BYTESWAP(drro->drr_bonustype);
2030 			dmu_ot_byteswap[byteswap].ob_func(db->db_data,
2031 			    drro->drr_bonuslen);
2032 		}
2033 		dmu_buf_rele(db, FTAG);
2034 	}
2035 	dmu_tx_commit(tx);
2036 
2037 	return (0);
2038 }
2039 
2040 /* ARGSUSED */
2041 static int
2042 receive_freeobjects(struct receive_writer_arg *rwa,
2043     struct drr_freeobjects *drrfo)
2044 {
2045 	uint64_t obj;
2046 	int next_err = 0;
2047 
2048 	if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
2049 		return (SET_ERROR(EINVAL));
2050 
2051 	for (obj = drrfo->drr_firstobj;
2052 	    obj < drrfo->drr_firstobj + drrfo->drr_numobjs && next_err == 0;
2053 	    next_err = dmu_object_next(rwa->os, &obj, FALSE, 0)) {
2054 		int err;
2055 
2056 		if (dmu_object_info(rwa->os, obj, NULL) != 0)
2057 			continue;
2058 
2059 		err = dmu_free_long_object(rwa->os, obj);
2060 		if (err != 0)
2061 			return (err);
2062 	}
2063 	if (next_err != ESRCH)
2064 		return (next_err);
2065 	return (0);
2066 }
2067 
2068 static int
2069 receive_write(struct receive_writer_arg *rwa, struct drr_write *drrw,
2070     arc_buf_t *abuf)
2071 {
2072 	dmu_tx_t *tx;
2073 	int err;
2074 
2075 	if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
2076 	    !DMU_OT_IS_VALID(drrw->drr_type))
2077 		return (SET_ERROR(EINVAL));
2078 
2079 	/*
2080 	 * For resuming to work, records must be in increasing order
2081 	 * by (object, offset).
2082 	 */
2083 	if (drrw->drr_object < rwa->last_object ||
2084 	    (drrw->drr_object == rwa->last_object &&
2085 	    drrw->drr_offset < rwa->last_offset)) {
2086 		return (SET_ERROR(EINVAL));
2087 	}
2088 	rwa->last_object = drrw->drr_object;
2089 	rwa->last_offset = drrw->drr_offset;
2090 
2091 	if (dmu_object_info(rwa->os, drrw->drr_object, NULL) != 0)
2092 		return (SET_ERROR(EINVAL));
2093 
2094 	tx = dmu_tx_create(rwa->os);
2095 
2096 	dmu_tx_hold_write(tx, drrw->drr_object,
2097 	    drrw->drr_offset, drrw->drr_length);
2098 	err = dmu_tx_assign(tx, TXG_WAIT);
2099 	if (err != 0) {
2100 		dmu_tx_abort(tx);
2101 		return (err);
2102 	}
2103 	if (rwa->byteswap) {
2104 		dmu_object_byteswap_t byteswap =
2105 		    DMU_OT_BYTESWAP(drrw->drr_type);
2106 		dmu_ot_byteswap[byteswap].ob_func(abuf->b_data,
2107 		    drrw->drr_length);
2108 	}
2109 
2110 	dmu_buf_t *bonus;
2111 	if (dmu_bonus_hold(rwa->os, drrw->drr_object, FTAG, &bonus) != 0)
2112 		return (SET_ERROR(EINVAL));
2113 	dmu_assign_arcbuf(bonus, drrw->drr_offset, abuf, tx);
2114 
2115 	/*
2116 	 * Note: If the receive fails, we want the resume stream to start
2117 	 * with the same record that we last successfully received (as opposed
2118 	 * to the next record), so that we can verify that we are
2119 	 * resuming from the correct location.
2120 	 */
2121 	save_resume_state(rwa, drrw->drr_object, drrw->drr_offset, tx);
2122 	dmu_tx_commit(tx);
2123 	dmu_buf_rele(bonus, FTAG);
2124 
2125 	return (0);
2126 }
2127 
2128 /*
2129  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
2130  * streams to refer to a copy of the data that is already on the
2131  * system because it came in earlier in the stream.  This function
2132  * finds the earlier copy of the data, and uses that copy instead of
2133  * data from the stream to fulfill this write.
2134  */
2135 static int
2136 receive_write_byref(struct receive_writer_arg *rwa,
2137     struct drr_write_byref *drrwbr)
2138 {
2139 	dmu_tx_t *tx;
2140 	int err;
2141 	guid_map_entry_t gmesrch;
2142 	guid_map_entry_t *gmep;
2143 	avl_index_t where;
2144 	objset_t *ref_os = NULL;
2145 	dmu_buf_t *dbp;
2146 
2147 	if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
2148 		return (SET_ERROR(EINVAL));
2149 
2150 	/*
2151 	 * If the GUID of the referenced dataset is different from the
2152 	 * GUID of the target dataset, find the referenced dataset.
2153 	 */
2154 	if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
2155 		gmesrch.guid = drrwbr->drr_refguid;
2156 		if ((gmep = avl_find(rwa->guid_to_ds_map, &gmesrch,
2157 		    &where)) == NULL) {
2158 			return (SET_ERROR(EINVAL));
2159 		}
2160 		if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
2161 			return (SET_ERROR(EINVAL));
2162 	} else {
2163 		ref_os = rwa->os;
2164 	}
2165 
2166 	err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
2167 	    drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH);
2168 	if (err != 0)
2169 		return (err);
2170 
2171 	tx = dmu_tx_create(rwa->os);
2172 
2173 	dmu_tx_hold_write(tx, drrwbr->drr_object,
2174 	    drrwbr->drr_offset, drrwbr->drr_length);
2175 	err = dmu_tx_assign(tx, TXG_WAIT);
2176 	if (err != 0) {
2177 		dmu_tx_abort(tx);
2178 		return (err);
2179 	}
2180 	dmu_write(rwa->os, drrwbr->drr_object,
2181 	    drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
2182 	dmu_buf_rele(dbp, FTAG);
2183 
2184 	/* See comment in restore_write. */
2185 	save_resume_state(rwa, drrwbr->drr_object, drrwbr->drr_offset, tx);
2186 	dmu_tx_commit(tx);
2187 	return (0);
2188 }
2189 
2190 static int
2191 receive_write_embedded(struct receive_writer_arg *rwa,
2192     struct drr_write_embedded *drrwe, void *data)
2193 {
2194 	dmu_tx_t *tx;
2195 	int err;
2196 
2197 	if (drrwe->drr_offset + drrwe->drr_length < drrwe->drr_offset)
2198 		return (EINVAL);
2199 
2200 	if (drrwe->drr_psize > BPE_PAYLOAD_SIZE)
2201 		return (EINVAL);
2202 
2203 	if (drrwe->drr_etype >= NUM_BP_EMBEDDED_TYPES)
2204 		return (EINVAL);
2205 	if (drrwe->drr_compression >= ZIO_COMPRESS_FUNCTIONS)
2206 		return (EINVAL);
2207 
2208 	tx = dmu_tx_create(rwa->os);
2209 
2210 	dmu_tx_hold_write(tx, drrwe->drr_object,
2211 	    drrwe->drr_offset, drrwe->drr_length);
2212 	err = dmu_tx_assign(tx, TXG_WAIT);
2213 	if (err != 0) {
2214 		dmu_tx_abort(tx);
2215 		return (err);
2216 	}
2217 
2218 	dmu_write_embedded(rwa->os, drrwe->drr_object,
2219 	    drrwe->drr_offset, data, drrwe->drr_etype,
2220 	    drrwe->drr_compression, drrwe->drr_lsize, drrwe->drr_psize,
2221 	    rwa->byteswap ^ ZFS_HOST_BYTEORDER, tx);
2222 
2223 	/* See comment in restore_write. */
2224 	save_resume_state(rwa, drrwe->drr_object, drrwe->drr_offset, tx);
2225 	dmu_tx_commit(tx);
2226 	return (0);
2227 }
2228 
2229 static int
2230 receive_spill(struct receive_writer_arg *rwa, struct drr_spill *drrs,
2231     void *data)
2232 {
2233 	dmu_tx_t *tx;
2234 	dmu_buf_t *db, *db_spill;
2235 	int err;
2236 
2237 	if (drrs->drr_length < SPA_MINBLOCKSIZE ||
2238 	    drrs->drr_length > spa_maxblocksize(dmu_objset_spa(rwa->os)))
2239 		return (SET_ERROR(EINVAL));
2240 
2241 	if (dmu_object_info(rwa->os, drrs->drr_object, NULL) != 0)
2242 		return (SET_ERROR(EINVAL));
2243 
2244 	VERIFY0(dmu_bonus_hold(rwa->os, drrs->drr_object, FTAG, &db));
2245 	if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
2246 		dmu_buf_rele(db, FTAG);
2247 		return (err);
2248 	}
2249 
2250 	tx = dmu_tx_create(rwa->os);
2251 
2252 	dmu_tx_hold_spill(tx, db->db_object);
2253 
2254 	err = dmu_tx_assign(tx, TXG_WAIT);
2255 	if (err != 0) {
2256 		dmu_buf_rele(db, FTAG);
2257 		dmu_buf_rele(db_spill, FTAG);
2258 		dmu_tx_abort(tx);
2259 		return (err);
2260 	}
2261 	dmu_buf_will_dirty(db_spill, tx);
2262 
2263 	if (db_spill->db_size < drrs->drr_length)
2264 		VERIFY(0 == dbuf_spill_set_blksz(db_spill,
2265 		    drrs->drr_length, tx));
2266 	bcopy(data, db_spill->db_data, drrs->drr_length);
2267 
2268 	dmu_buf_rele(db, FTAG);
2269 	dmu_buf_rele(db_spill, FTAG);
2270 
2271 	dmu_tx_commit(tx);
2272 	return (0);
2273 }
2274 
2275 /* ARGSUSED */
2276 static int
2277 receive_free(struct receive_writer_arg *rwa, struct drr_free *drrf)
2278 {
2279 	int err;
2280 
2281 	if (drrf->drr_length != -1ULL &&
2282 	    drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
2283 		return (SET_ERROR(EINVAL));
2284 
2285 	if (dmu_object_info(rwa->os, drrf->drr_object, NULL) != 0)
2286 		return (SET_ERROR(EINVAL));
2287 
2288 	err = dmu_free_long_range(rwa->os, drrf->drr_object,
2289 	    drrf->drr_offset, drrf->drr_length);
2290 
2291 	return (err);
2292 }
2293 
2294 /* used to destroy the drc_ds on error */
2295 static void
2296 dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc)
2297 {
2298 	if (drc->drc_resumable) {
2299 		/* wait for our resume state to be written to disk */
2300 		txg_wait_synced(drc->drc_ds->ds_dir->dd_pool, 0);
2301 		dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2302 	} else {
2303 		char name[MAXNAMELEN];
2304 		dsl_dataset_name(drc->drc_ds, name);
2305 		dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2306 		(void) dsl_destroy_head(name);
2307 	}
2308 }
2309 
2310 static void
2311 receive_cksum(struct receive_arg *ra, int len, void *buf)
2312 {
2313 	if (ra->byteswap) {
2314 		fletcher_4_incremental_byteswap(buf, len, &ra->cksum);
2315 	} else {
2316 		fletcher_4_incremental_native(buf, len, &ra->cksum);
2317 	}
2318 }
2319 
2320 /*
2321  * Read the payload into a buffer of size len, and update the current record's
2322  * payload field.
2323  * Allocate ra->next_rrd and read the next record's header into
2324  * ra->next_rrd->header.
2325  * Verify checksum of payload and next record.
2326  */
2327 static int
2328 receive_read_payload_and_next_header(struct receive_arg *ra, int len, void *buf)
2329 {
2330 	int err;
2331 
2332 	if (len != 0) {
2333 		ASSERT3U(len, <=, SPA_MAXBLOCKSIZE);
2334 		err = receive_read(ra, len, buf);
2335 		if (err != 0)
2336 			return (err);
2337 		receive_cksum(ra, len, buf);
2338 
2339 		/* note: rrd is NULL when reading the begin record's payload */
2340 		if (ra->rrd != NULL) {
2341 			ra->rrd->payload = buf;
2342 			ra->rrd->payload_size = len;
2343 			ra->rrd->bytes_read = ra->bytes_read;
2344 		}
2345 	}
2346 
2347 	ra->prev_cksum = ra->cksum;
2348 
2349 	ra->next_rrd = kmem_zalloc(sizeof (*ra->next_rrd), KM_SLEEP);
2350 	err = receive_read(ra, sizeof (ra->next_rrd->header),
2351 	    &ra->next_rrd->header);
2352 	ra->next_rrd->bytes_read = ra->bytes_read;
2353 	if (err != 0) {
2354 		kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2355 		ra->next_rrd = NULL;
2356 		return (err);
2357 	}
2358 	if (ra->next_rrd->header.drr_type == DRR_BEGIN) {
2359 		kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2360 		ra->next_rrd = NULL;
2361 		return (SET_ERROR(EINVAL));
2362 	}
2363 
2364 	/*
2365 	 * Note: checksum is of everything up to but not including the
2366 	 * checksum itself.
2367 	 */
2368 	ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
2369 	    ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
2370 	receive_cksum(ra,
2371 	    offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
2372 	    &ra->next_rrd->header);
2373 
2374 	zio_cksum_t cksum_orig =
2375 	    ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
2376 	zio_cksum_t *cksump =
2377 	    &ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
2378 
2379 	if (ra->byteswap)
2380 		byteswap_record(&ra->next_rrd->header);
2381 
2382 	if ((!ZIO_CHECKSUM_IS_ZERO(cksump)) &&
2383 	    !ZIO_CHECKSUM_EQUAL(ra->cksum, *cksump)) {
2384 		kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2385 		ra->next_rrd = NULL;
2386 		return (SET_ERROR(ECKSUM));
2387 	}
2388 
2389 	receive_cksum(ra, sizeof (cksum_orig), &cksum_orig);
2390 
2391 	return (0);
2392 }
2393 
2394 static void
2395 objlist_create(struct objlist *list)
2396 {
2397 	list_create(&list->list, sizeof (struct receive_objnode),
2398 	    offsetof(struct receive_objnode, node));
2399 	list->last_lookup = 0;
2400 }
2401 
2402 static void
2403 objlist_destroy(struct objlist *list)
2404 {
2405 	for (struct receive_objnode *n = list_remove_head(&list->list);
2406 	    n != NULL; n = list_remove_head(&list->list)) {
2407 		kmem_free(n, sizeof (*n));
2408 	}
2409 	list_destroy(&list->list);
2410 }
2411 
2412 /*
2413  * This function looks through the objlist to see if the specified object number
2414  * is contained in the objlist.  In the process, it will remove all object
2415  * numbers in the list that are smaller than the specified object number.  Thus,
2416  * any lookup of an object number smaller than a previously looked up object
2417  * number will always return false; therefore, all lookups should be done in
2418  * ascending order.
2419  */
2420 static boolean_t
2421 objlist_exists(struct objlist *list, uint64_t object)
2422 {
2423 	struct receive_objnode *node = list_head(&list->list);
2424 	ASSERT3U(object, >=, list->last_lookup);
2425 	list->last_lookup = object;
2426 	while (node != NULL && node->object < object) {
2427 		VERIFY3P(node, ==, list_remove_head(&list->list));
2428 		kmem_free(node, sizeof (*node));
2429 		node = list_head(&list->list);
2430 	}
2431 	return (node != NULL && node->object == object);
2432 }
2433 
2434 /*
2435  * The objlist is a list of object numbers stored in ascending order.  However,
2436  * the insertion of new object numbers does not seek out the correct location to
2437  * store a new object number; instead, it appends it to the list for simplicity.
2438  * Thus, any users must take care to only insert new object numbers in ascending
2439  * order.
2440  */
2441 static void
2442 objlist_insert(struct objlist *list, uint64_t object)
2443 {
2444 	struct receive_objnode *node = kmem_zalloc(sizeof (*node), KM_SLEEP);
2445 	node->object = object;
2446 #ifdef ZFS_DEBUG
2447 	struct receive_objnode *last_object = list_tail(&list->list);
2448 	uint64_t last_objnum = (last_object != NULL ? last_object->object : 0);
2449 	ASSERT3U(node->object, >, last_objnum);
2450 #endif
2451 	list_insert_tail(&list->list, node);
2452 }
2453 
2454 /*
2455  * Issue the prefetch reads for any necessary indirect blocks.
2456  *
2457  * We use the object ignore list to tell us whether or not to issue prefetches
2458  * for a given object.  We do this for both correctness (in case the blocksize
2459  * of an object has changed) and performance (if the object doesn't exist, don't
2460  * needlessly try to issue prefetches).  We also trim the list as we go through
2461  * the stream to prevent it from growing to an unbounded size.
2462  *
2463  * The object numbers within will always be in sorted order, and any write
2464  * records we see will also be in sorted order, but they're not sorted with
2465  * respect to each other (i.e. we can get several object records before
2466  * receiving each object's write records).  As a result, once we've reached a
2467  * given object number, we can safely remove any reference to lower object
2468  * numbers in the ignore list. In practice, we receive up to 32 object records
2469  * before receiving write records, so the list can have up to 32 nodes in it.
2470  */
2471 /* ARGSUSED */
2472 static void
2473 receive_read_prefetch(struct receive_arg *ra,
2474     uint64_t object, uint64_t offset, uint64_t length)
2475 {
2476 	if (!objlist_exists(&ra->ignore_objlist, object)) {
2477 		dmu_prefetch(ra->os, object, 1, offset, length,
2478 		    ZIO_PRIORITY_SYNC_READ);
2479 	}
2480 }
2481 
2482 /*
2483  * Read records off the stream, issuing any necessary prefetches.
2484  */
2485 static int
2486 receive_read_record(struct receive_arg *ra)
2487 {
2488 	int err;
2489 
2490 	switch (ra->rrd->header.drr_type) {
2491 	case DRR_OBJECT:
2492 	{
2493 		struct drr_object *drro = &ra->rrd->header.drr_u.drr_object;
2494 		uint32_t size = P2ROUNDUP(drro->drr_bonuslen, 8);
2495 		void *buf = kmem_zalloc(size, KM_SLEEP);
2496 		dmu_object_info_t doi;
2497 		err = receive_read_payload_and_next_header(ra, size, buf);
2498 		if (err != 0) {
2499 			kmem_free(buf, size);
2500 			return (err);
2501 		}
2502 		err = dmu_object_info(ra->os, drro->drr_object, &doi);
2503 		/*
2504 		 * See receive_read_prefetch for an explanation why we're
2505 		 * storing this object in the ignore_obj_list.
2506 		 */
2507 		if (err == ENOENT ||
2508 		    (err == 0 && doi.doi_data_block_size != drro->drr_blksz)) {
2509 			objlist_insert(&ra->ignore_objlist, drro->drr_object);
2510 			err = 0;
2511 		}
2512 		return (err);
2513 	}
2514 	case DRR_FREEOBJECTS:
2515 	{
2516 		err = receive_read_payload_and_next_header(ra, 0, NULL);
2517 		return (err);
2518 	}
2519 	case DRR_WRITE:
2520 	{
2521 		struct drr_write *drrw = &ra->rrd->header.drr_u.drr_write;
2522 		arc_buf_t *abuf = arc_loan_buf(dmu_objset_spa(ra->os),
2523 		    drrw->drr_length);
2524 
2525 		err = receive_read_payload_and_next_header(ra,
2526 		    drrw->drr_length, abuf->b_data);
2527 		if (err != 0) {
2528 			dmu_return_arcbuf(abuf);
2529 			return (err);
2530 		}
2531 		ra->rrd->write_buf = abuf;
2532 		receive_read_prefetch(ra, drrw->drr_object, drrw->drr_offset,
2533 		    drrw->drr_length);
2534 		return (err);
2535 	}
2536 	case DRR_WRITE_BYREF:
2537 	{
2538 		struct drr_write_byref *drrwb =
2539 		    &ra->rrd->header.drr_u.drr_write_byref;
2540 		err = receive_read_payload_and_next_header(ra, 0, NULL);
2541 		receive_read_prefetch(ra, drrwb->drr_object, drrwb->drr_offset,
2542 		    drrwb->drr_length);
2543 		return (err);
2544 	}
2545 	case DRR_WRITE_EMBEDDED:
2546 	{
2547 		struct drr_write_embedded *drrwe =
2548 		    &ra->rrd->header.drr_u.drr_write_embedded;
2549 		uint32_t size = P2ROUNDUP(drrwe->drr_psize, 8);
2550 		void *buf = kmem_zalloc(size, KM_SLEEP);
2551 
2552 		err = receive_read_payload_and_next_header(ra, size, buf);
2553 		if (err != 0) {
2554 			kmem_free(buf, size);
2555 			return (err);
2556 		}
2557 
2558 		receive_read_prefetch(ra, drrwe->drr_object, drrwe->drr_offset,
2559 		    drrwe->drr_length);
2560 		return (err);
2561 	}
2562 	case DRR_FREE:
2563 	{
2564 		/*
2565 		 * It might be beneficial to prefetch indirect blocks here, but
2566 		 * we don't really have the data to decide for sure.
2567 		 */
2568 		err = receive_read_payload_and_next_header(ra, 0, NULL);
2569 		return (err);
2570 	}
2571 	case DRR_END:
2572 	{
2573 		struct drr_end *drre = &ra->rrd->header.drr_u.drr_end;
2574 		if (!ZIO_CHECKSUM_EQUAL(ra->prev_cksum, drre->drr_checksum))
2575 			return (SET_ERROR(ECKSUM));
2576 		return (0);
2577 	}
2578 	case DRR_SPILL:
2579 	{
2580 		struct drr_spill *drrs = &ra->rrd->header.drr_u.drr_spill;
2581 		void *buf = kmem_zalloc(drrs->drr_length, KM_SLEEP);
2582 		err = receive_read_payload_and_next_header(ra, drrs->drr_length,
2583 		    buf);
2584 		if (err != 0)
2585 			kmem_free(buf, drrs->drr_length);
2586 		return (err);
2587 	}
2588 	default:
2589 		return (SET_ERROR(EINVAL));
2590 	}
2591 }
2592 
2593 /*
2594  * Commit the records to the pool.
2595  */
2596 static int
2597 receive_process_record(struct receive_writer_arg *rwa,
2598     struct receive_record_arg *rrd)
2599 {
2600 	int err;
2601 
2602 	/* Processing in order, therefore bytes_read should be increasing. */
2603 	ASSERT3U(rrd->bytes_read, >=, rwa->bytes_read);
2604 	rwa->bytes_read = rrd->bytes_read;
2605 
2606 	switch (rrd->header.drr_type) {
2607 	case DRR_OBJECT:
2608 	{
2609 		struct drr_object *drro = &rrd->header.drr_u.drr_object;
2610 		err = receive_object(rwa, drro, rrd->payload);
2611 		kmem_free(rrd->payload, rrd->payload_size);
2612 		rrd->payload = NULL;
2613 		return (err);
2614 	}
2615 	case DRR_FREEOBJECTS:
2616 	{
2617 		struct drr_freeobjects *drrfo =
2618 		    &rrd->header.drr_u.drr_freeobjects;
2619 		return (receive_freeobjects(rwa, drrfo));
2620 	}
2621 	case DRR_WRITE:
2622 	{
2623 		struct drr_write *drrw = &rrd->header.drr_u.drr_write;
2624 		err = receive_write(rwa, drrw, rrd->write_buf);
2625 		/* if receive_write() is successful, it consumes the arc_buf */
2626 		if (err != 0)
2627 			dmu_return_arcbuf(rrd->write_buf);
2628 		rrd->write_buf = NULL;
2629 		rrd->payload = NULL;
2630 		return (err);
2631 	}
2632 	case DRR_WRITE_BYREF:
2633 	{
2634 		struct drr_write_byref *drrwbr =
2635 		    &rrd->header.drr_u.drr_write_byref;
2636 		return (receive_write_byref(rwa, drrwbr));
2637 	}
2638 	case DRR_WRITE_EMBEDDED:
2639 	{
2640 		struct drr_write_embedded *drrwe =
2641 		    &rrd->header.drr_u.drr_write_embedded;
2642 		err = receive_write_embedded(rwa, drrwe, rrd->payload);
2643 		kmem_free(rrd->payload, rrd->payload_size);
2644 		rrd->payload = NULL;
2645 		return (err);
2646 	}
2647 	case DRR_FREE:
2648 	{
2649 		struct drr_free *drrf = &rrd->header.drr_u.drr_free;
2650 		return (receive_free(rwa, drrf));
2651 	}
2652 	case DRR_SPILL:
2653 	{
2654 		struct drr_spill *drrs = &rrd->header.drr_u.drr_spill;
2655 		err = receive_spill(rwa, drrs, rrd->payload);
2656 		kmem_free(rrd->payload, rrd->payload_size);
2657 		rrd->payload = NULL;
2658 		return (err);
2659 	}
2660 	default:
2661 		return (SET_ERROR(EINVAL));
2662 	}
2663 }
2664 
2665 /*
2666  * dmu_recv_stream's worker thread; pull records off the queue, and then call
2667  * receive_process_record  When we're done, signal the main thread and exit.
2668  */
2669 static void
2670 receive_writer_thread(void *arg)
2671 {
2672 	struct receive_writer_arg *rwa = arg;
2673 	struct receive_record_arg *rrd;
2674 	for (rrd = bqueue_dequeue(&rwa->q); !rrd->eos_marker;
2675 	    rrd = bqueue_dequeue(&rwa->q)) {
2676 		/*
2677 		 * If there's an error, the main thread will stop putting things
2678 		 * on the queue, but we need to clear everything in it before we
2679 		 * can exit.
2680 		 */
2681 		if (rwa->err == 0) {
2682 			rwa->err = receive_process_record(rwa, rrd);
2683 		} else if (rrd->write_buf != NULL) {
2684 			dmu_return_arcbuf(rrd->write_buf);
2685 			rrd->write_buf = NULL;
2686 			rrd->payload = NULL;
2687 		} else if (rrd->payload != NULL) {
2688 			kmem_free(rrd->payload, rrd->payload_size);
2689 			rrd->payload = NULL;
2690 		}
2691 		kmem_free(rrd, sizeof (*rrd));
2692 	}
2693 	kmem_free(rrd, sizeof (*rrd));
2694 	mutex_enter(&rwa->mutex);
2695 	rwa->done = B_TRUE;
2696 	cv_signal(&rwa->cv);
2697 	mutex_exit(&rwa->mutex);
2698 }
2699 
2700 static int
2701 resume_check(struct receive_arg *ra, nvlist_t *begin_nvl)
2702 {
2703 	uint64_t val;
2704 	objset_t *mos = dmu_objset_pool(ra->os)->dp_meta_objset;
2705 	uint64_t dsobj = dmu_objset_id(ra->os);
2706 	uint64_t resume_obj, resume_off;
2707 
2708 	if (nvlist_lookup_uint64(begin_nvl,
2709 	    "resume_object", &resume_obj) != 0 ||
2710 	    nvlist_lookup_uint64(begin_nvl,
2711 	    "resume_offset", &resume_off) != 0) {
2712 		return (SET_ERROR(EINVAL));
2713 	}
2714 	VERIFY0(zap_lookup(mos, dsobj,
2715 	    DS_FIELD_RESUME_OBJECT, sizeof (val), 1, &val));
2716 	if (resume_obj != val)
2717 		return (SET_ERROR(EINVAL));
2718 	VERIFY0(zap_lookup(mos, dsobj,
2719 	    DS_FIELD_RESUME_OFFSET, sizeof (val), 1, &val));
2720 	if (resume_off != val)
2721 		return (SET_ERROR(EINVAL));
2722 
2723 	return (0);
2724 }
2725 
2726 /*
2727  * Read in the stream's records, one by one, and apply them to the pool.  There
2728  * are two threads involved; the thread that calls this function will spin up a
2729  * worker thread, read the records off the stream one by one, and issue
2730  * prefetches for any necessary indirect blocks.  It will then push the records
2731  * onto an internal blocking queue.  The worker thread will pull the records off
2732  * the queue, and actually write the data into the DMU.  This way, the worker
2733  * thread doesn't have to wait for reads to complete, since everything it needs
2734  * (the indirect blocks) will be prefetched.
2735  *
2736  * NB: callers *must* call dmu_recv_end() if this succeeds.
2737  */
2738 int
2739 dmu_recv_stream(dmu_recv_cookie_t *drc, vnode_t *vp, offset_t *voffp,
2740     int cleanup_fd, uint64_t *action_handlep)
2741 {
2742 	int err = 0;
2743 	struct receive_arg ra = { 0 };
2744 	struct receive_writer_arg rwa = { 0 };
2745 	int featureflags;
2746 	nvlist_t *begin_nvl = NULL;
2747 
2748 	ra.byteswap = drc->drc_byteswap;
2749 	ra.cksum = drc->drc_cksum;
2750 	ra.vp = vp;
2751 	ra.voff = *voffp;
2752 
2753 	if (dsl_dataset_is_zapified(drc->drc_ds)) {
2754 		(void) zap_lookup(drc->drc_ds->ds_dir->dd_pool->dp_meta_objset,
2755 		    drc->drc_ds->ds_object, DS_FIELD_RESUME_BYTES,
2756 		    sizeof (ra.bytes_read), 1, &ra.bytes_read);
2757 	}
2758 
2759 	objlist_create(&ra.ignore_objlist);
2760 
2761 	/* these were verified in dmu_recv_begin */
2762 	ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==,
2763 	    DMU_SUBSTREAM);
2764 	ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES);
2765 
2766 	/*
2767 	 * Open the objset we are modifying.
2768 	 */
2769 	VERIFY0(dmu_objset_from_ds(drc->drc_ds, &ra.os));
2770 
2771 	ASSERT(dsl_dataset_phys(drc->drc_ds)->ds_flags & DS_FLAG_INCONSISTENT);
2772 
2773 	featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
2774 
2775 	/* if this stream is dedup'ed, set up the avl tree for guid mapping */
2776 	if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
2777 		minor_t minor;
2778 
2779 		if (cleanup_fd == -1) {
2780 			ra.err = SET_ERROR(EBADF);
2781 			goto out;
2782 		}
2783 		ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
2784 		if (ra.err != 0) {
2785 			cleanup_fd = -1;
2786 			goto out;
2787 		}
2788 
2789 		if (*action_handlep == 0) {
2790 			rwa.guid_to_ds_map =
2791 			    kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
2792 			avl_create(rwa.guid_to_ds_map, guid_compare,
2793 			    sizeof (guid_map_entry_t),
2794 			    offsetof(guid_map_entry_t, avlnode));
2795 			err = zfs_onexit_add_cb(minor,
2796 			    free_guid_map_onexit, rwa.guid_to_ds_map,
2797 			    action_handlep);
2798 			if (ra.err != 0)
2799 				goto out;
2800 		} else {
2801 			err = zfs_onexit_cb_data(minor, *action_handlep,
2802 			    (void **)&rwa.guid_to_ds_map);
2803 			if (ra.err != 0)
2804 				goto out;
2805 		}
2806 
2807 		drc->drc_guid_to_ds_map = rwa.guid_to_ds_map;
2808 	}
2809 
2810 	uint32_t payloadlen = drc->drc_drr_begin->drr_payloadlen;
2811 	void *payload = NULL;
2812 	if (payloadlen != 0)
2813 		payload = kmem_alloc(payloadlen, KM_SLEEP);
2814 
2815 	err = receive_read_payload_and_next_header(&ra, payloadlen, payload);
2816 	if (err != 0) {
2817 		if (payloadlen != 0)
2818 			kmem_free(payload, payloadlen);
2819 		goto out;
2820 	}
2821 	if (payloadlen != 0) {
2822 		err = nvlist_unpack(payload, payloadlen, &begin_nvl, KM_SLEEP);
2823 		kmem_free(payload, payloadlen);
2824 		if (err != 0)
2825 			goto out;
2826 	}
2827 
2828 	if (featureflags & DMU_BACKUP_FEATURE_RESUMING) {
2829 		err = resume_check(&ra, begin_nvl);
2830 		if (err != 0)
2831 			goto out;
2832 	}
2833 
2834 	(void) bqueue_init(&rwa.q, zfs_recv_queue_length,
2835 	    offsetof(struct receive_record_arg, node));
2836 	cv_init(&rwa.cv, NULL, CV_DEFAULT, NULL);
2837 	mutex_init(&rwa.mutex, NULL, MUTEX_DEFAULT, NULL);
2838 	rwa.os = ra.os;
2839 	rwa.byteswap = drc->drc_byteswap;
2840 	rwa.resumable = drc->drc_resumable;
2841 
2842 	(void) thread_create(NULL, 0, receive_writer_thread, &rwa, 0, curproc,
2843 	    TS_RUN, minclsyspri);
2844 	/*
2845 	 * We're reading rwa.err without locks, which is safe since we are the
2846 	 * only reader, and the worker thread is the only writer.  It's ok if we
2847 	 * miss a write for an iteration or two of the loop, since the writer
2848 	 * thread will keep freeing records we send it until we send it an eos
2849 	 * marker.
2850 	 *
2851 	 * We can leave this loop in 3 ways:  First, if rwa.err is
2852 	 * non-zero.  In that case, the writer thread will free the rrd we just
2853 	 * pushed.  Second, if  we're interrupted; in that case, either it's the
2854 	 * first loop and ra.rrd was never allocated, or it's later, and ra.rrd
2855 	 * has been handed off to the writer thread who will free it.  Finally,
2856 	 * if receive_read_record fails or we're at the end of the stream, then
2857 	 * we free ra.rrd and exit.
2858 	 */
2859 	while (rwa.err == 0) {
2860 		if (issig(JUSTLOOKING) && issig(FORREAL)) {
2861 			err = SET_ERROR(EINTR);
2862 			break;
2863 		}
2864 
2865 		ASSERT3P(ra.rrd, ==, NULL);
2866 		ra.rrd = ra.next_rrd;
2867 		ra.next_rrd = NULL;
2868 		/* Allocates and loads header into ra.next_rrd */
2869 		err = receive_read_record(&ra);
2870 
2871 		if (ra.rrd->header.drr_type == DRR_END || err != 0) {
2872 			kmem_free(ra.rrd, sizeof (*ra.rrd));
2873 			ra.rrd = NULL;
2874 			break;
2875 		}
2876 
2877 		bqueue_enqueue(&rwa.q, ra.rrd,
2878 		    sizeof (struct receive_record_arg) + ra.rrd->payload_size);
2879 		ra.rrd = NULL;
2880 	}
2881 	if (ra.next_rrd == NULL)
2882 		ra.next_rrd = kmem_zalloc(sizeof (*ra.next_rrd), KM_SLEEP);
2883 	ra.next_rrd->eos_marker = B_TRUE;
2884 	bqueue_enqueue(&rwa.q, ra.next_rrd, 1);
2885 
2886 	mutex_enter(&rwa.mutex);
2887 	while (!rwa.done) {
2888 		cv_wait(&rwa.cv, &rwa.mutex);
2889 	}
2890 	mutex_exit(&rwa.mutex);
2891 
2892 	cv_destroy(&rwa.cv);
2893 	mutex_destroy(&rwa.mutex);
2894 	bqueue_destroy(&rwa.q);
2895 	if (err == 0)
2896 		err = rwa.err;
2897 
2898 out:
2899 	nvlist_free(begin_nvl);
2900 	if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
2901 		zfs_onexit_fd_rele(cleanup_fd);
2902 
2903 	if (err != 0) {
2904 		/*
2905 		 * Clean up references. If receive is not resumable,
2906 		 * destroy what we created, so we don't leave it in
2907 		 * the inconsistent state.
2908 		 */
2909 		dmu_recv_cleanup_ds(drc);
2910 	}
2911 
2912 	*voffp = ra.voff;
2913 	objlist_destroy(&ra.ignore_objlist);
2914 	return (err);
2915 }
2916 
2917 static int
2918 dmu_recv_end_check(void *arg, dmu_tx_t *tx)
2919 {
2920 	dmu_recv_cookie_t *drc = arg;
2921 	dsl_pool_t *dp = dmu_tx_pool(tx);
2922 	int error;
2923 
2924 	ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag);
2925 
2926 	if (!drc->drc_newfs) {
2927 		dsl_dataset_t *origin_head;
2928 
2929 		error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head);
2930 		if (error != 0)
2931 			return (error);
2932 		if (drc->drc_force) {
2933 			/*
2934 			 * We will destroy any snapshots in tofs (i.e. before
2935 			 * origin_head) that are after the origin (which is
2936 			 * the snap before drc_ds, because drc_ds can not
2937 			 * have any snaps of its own).
2938 			 */
2939 			uint64_t obj;
2940 
2941 			obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
2942 			while (obj !=
2943 			    dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
2944 				dsl_dataset_t *snap;
2945 				error = dsl_dataset_hold_obj(dp, obj, FTAG,
2946 				    &snap);
2947 				if (error != 0)
2948 					break;
2949 				if (snap->ds_dir != origin_head->ds_dir)
2950 					error = SET_ERROR(EINVAL);
2951 				if (error == 0)  {
2952 					error = dsl_destroy_snapshot_check_impl(
2953 					    snap, B_FALSE);
2954 				}
2955 				obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
2956 				dsl_dataset_rele(snap, FTAG);
2957 				if (error != 0)
2958 					break;
2959 			}
2960 			if (error != 0) {
2961 				dsl_dataset_rele(origin_head, FTAG);
2962 				return (error);
2963 			}
2964 		}
2965 		error = dsl_dataset_clone_swap_check_impl(drc->drc_ds,
2966 		    origin_head, drc->drc_force, drc->drc_owner, tx);
2967 		if (error != 0) {
2968 			dsl_dataset_rele(origin_head, FTAG);
2969 			return (error);
2970 		}
2971 		error = dsl_dataset_snapshot_check_impl(origin_head,
2972 		    drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
2973 		dsl_dataset_rele(origin_head, FTAG);
2974 		if (error != 0)
2975 			return (error);
2976 
2977 		error = dsl_destroy_head_check_impl(drc->drc_ds, 1);
2978 	} else {
2979 		error = dsl_dataset_snapshot_check_impl(drc->drc_ds,
2980 		    drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
2981 	}
2982 	return (error);
2983 }
2984 
2985 static void
2986 dmu_recv_end_sync(void *arg, dmu_tx_t *tx)
2987 {
2988 	dmu_recv_cookie_t *drc = arg;
2989 	dsl_pool_t *dp = dmu_tx_pool(tx);
2990 
2991 	spa_history_log_internal_ds(drc->drc_ds, "finish receiving",
2992 	    tx, "snap=%s", drc->drc_tosnap);
2993 
2994 	if (!drc->drc_newfs) {
2995 		dsl_dataset_t *origin_head;
2996 
2997 		VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG,
2998 		    &origin_head));
2999 
3000 		if (drc->drc_force) {
3001 			/*
3002 			 * Destroy any snapshots of drc_tofs (origin_head)
3003 			 * after the origin (the snap before drc_ds).
3004 			 */
3005 			uint64_t obj;
3006 
3007 			obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3008 			while (obj !=
3009 			    dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
3010 				dsl_dataset_t *snap;
3011 				VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG,
3012 				    &snap));
3013 				ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir);
3014 				obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
3015 				dsl_destroy_snapshot_sync_impl(snap,
3016 				    B_FALSE, tx);
3017 				dsl_dataset_rele(snap, FTAG);
3018 			}
3019 		}
3020 		VERIFY3P(drc->drc_ds->ds_prev, ==,
3021 		    origin_head->ds_prev);
3022 
3023 		dsl_dataset_clone_swap_sync_impl(drc->drc_ds,
3024 		    origin_head, tx);
3025 		dsl_dataset_snapshot_sync_impl(origin_head,
3026 		    drc->drc_tosnap, tx);
3027 
3028 		/* set snapshot's creation time and guid */
3029 		dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx);
3030 		dsl_dataset_phys(origin_head->ds_prev)->ds_creation_time =
3031 		    drc->drc_drrb->drr_creation_time;
3032 		dsl_dataset_phys(origin_head->ds_prev)->ds_guid =
3033 		    drc->drc_drrb->drr_toguid;
3034 		dsl_dataset_phys(origin_head->ds_prev)->ds_flags &=
3035 		    ~DS_FLAG_INCONSISTENT;
3036 
3037 		dmu_buf_will_dirty(origin_head->ds_dbuf, tx);
3038 		dsl_dataset_phys(origin_head)->ds_flags &=
3039 		    ~DS_FLAG_INCONSISTENT;
3040 
3041 		dsl_dataset_rele(origin_head, FTAG);
3042 		dsl_destroy_head_sync_impl(drc->drc_ds, tx);
3043 
3044 		if (drc->drc_owner != NULL)
3045 			VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner);
3046 	} else {
3047 		dsl_dataset_t *ds = drc->drc_ds;
3048 
3049 		dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx);
3050 
3051 		/* set snapshot's creation time and guid */
3052 		dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
3053 		dsl_dataset_phys(ds->ds_prev)->ds_creation_time =
3054 		    drc->drc_drrb->drr_creation_time;
3055 		dsl_dataset_phys(ds->ds_prev)->ds_guid =
3056 		    drc->drc_drrb->drr_toguid;
3057 		dsl_dataset_phys(ds->ds_prev)->ds_flags &=
3058 		    ~DS_FLAG_INCONSISTENT;
3059 
3060 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
3061 		dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
3062 		if (dsl_dataset_has_resume_receive_state(ds)) {
3063 			(void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3064 			    DS_FIELD_RESUME_FROMGUID, tx);
3065 			(void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3066 			    DS_FIELD_RESUME_OBJECT, tx);
3067 			(void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3068 			    DS_FIELD_RESUME_OFFSET, tx);
3069 			(void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3070 			    DS_FIELD_RESUME_BYTES, tx);
3071 			(void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3072 			    DS_FIELD_RESUME_TOGUID, tx);
3073 			(void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3074 			    DS_FIELD_RESUME_TONAME, tx);
3075 		}
3076 	}
3077 	drc->drc_newsnapobj = dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj;
3078 	/*
3079 	 * Release the hold from dmu_recv_begin.  This must be done before
3080 	 * we return to open context, so that when we free the dataset's dnode,
3081 	 * we can evict its bonus buffer.
3082 	 */
3083 	dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
3084 	drc->drc_ds = NULL;
3085 }
3086 
3087 static int
3088 add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj)
3089 {
3090 	dsl_pool_t *dp;
3091 	dsl_dataset_t *snapds;
3092 	guid_map_entry_t *gmep;
3093 	int err;
3094 
3095 	ASSERT(guid_map != NULL);
3096 
3097 	err = dsl_pool_hold(name, FTAG, &dp);
3098 	if (err != 0)
3099 		return (err);
3100 	gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP);
3101 	err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds);
3102 	if (err == 0) {
3103 		gmep->guid = dsl_dataset_phys(snapds)->ds_guid;
3104 		gmep->gme_ds = snapds;
3105 		avl_add(guid_map, gmep);
3106 		dsl_dataset_long_hold(snapds, gmep);
3107 	} else {
3108 		kmem_free(gmep, sizeof (*gmep));
3109 	}
3110 
3111 	dsl_pool_rele(dp, FTAG);
3112 	return (err);
3113 }
3114 
3115 static int dmu_recv_end_modified_blocks = 3;
3116 
3117 static int
3118 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
3119 {
3120 	int error;
3121 	char name[MAXNAMELEN];
3122 
3123 #ifdef _KERNEL
3124 	/*
3125 	 * We will be destroying the ds; make sure its origin is unmounted if
3126 	 * necessary.
3127 	 */
3128 	dsl_dataset_name(drc->drc_ds, name);
3129 	zfs_destroy_unmount_origin(name);
3130 #endif
3131 
3132 	error = dsl_sync_task(drc->drc_tofs,
3133 	    dmu_recv_end_check, dmu_recv_end_sync, drc,
3134 	    dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL);
3135 
3136 	if (error != 0)
3137 		dmu_recv_cleanup_ds(drc);
3138 	return (error);
3139 }
3140 
3141 static int
3142 dmu_recv_new_end(dmu_recv_cookie_t *drc)
3143 {
3144 	int error;
3145 
3146 	error = dsl_sync_task(drc->drc_tofs,
3147 	    dmu_recv_end_check, dmu_recv_end_sync, drc,
3148 	    dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL);
3149 
3150 	if (error != 0) {
3151 		dmu_recv_cleanup_ds(drc);
3152 	} else if (drc->drc_guid_to_ds_map != NULL) {
3153 		(void) add_ds_to_guidmap(drc->drc_tofs,
3154 		    drc->drc_guid_to_ds_map,
3155 		    drc->drc_newsnapobj);
3156 	}
3157 	return (error);
3158 }
3159 
3160 int
3161 dmu_recv_end(dmu_recv_cookie_t *drc, void *owner)
3162 {
3163 	drc->drc_owner = owner;
3164 
3165 	if (drc->drc_newfs)
3166 		return (dmu_recv_new_end(drc));
3167 	else
3168 		return (dmu_recv_existing_end(drc));
3169 }
3170 
3171 /*
3172  * Return TRUE if this objset is currently being received into.
3173  */
3174 boolean_t
3175 dmu_objset_is_receiving(objset_t *os)
3176 {
3177 	return (os->os_dsl_dataset != NULL &&
3178 	    os->os_dsl_dataset->ds_owner == dmu_recv_tag);
3179 }
3180