xref: /freebsd/sys/contrib/openzfs/module/zfs/dmu_send.c (revision 3078531de10dcae44b253a35125c949ff4235284)
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 https://opensource.org/licenses/CDDL-1.0.
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, 2018 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) 2016 Actifio, Inc. All rights reserved.
29  * Copyright (c) 2019, Klara Inc.
30  * Copyright (c) 2019, Allan Jude
31  */
32 
33 #include <sys/dmu.h>
34 #include <sys/dmu_impl.h>
35 #include <sys/dmu_tx.h>
36 #include <sys/dbuf.h>
37 #include <sys/dnode.h>
38 #include <sys/zfs_context.h>
39 #include <sys/dmu_objset.h>
40 #include <sys/dmu_traverse.h>
41 #include <sys/dsl_dataset.h>
42 #include <sys/dsl_dir.h>
43 #include <sys/dsl_prop.h>
44 #include <sys/dsl_pool.h>
45 #include <sys/dsl_synctask.h>
46 #include <sys/spa_impl.h>
47 #include <sys/zfs_ioctl.h>
48 #include <sys/zap.h>
49 #include <sys/zio_checksum.h>
50 #include <sys/zfs_znode.h>
51 #include <zfs_fletcher.h>
52 #include <sys/avl.h>
53 #include <sys/ddt.h>
54 #include <sys/zfs_onexit.h>
55 #include <sys/dmu_send.h>
56 #include <sys/dmu_recv.h>
57 #include <sys/dsl_destroy.h>
58 #include <sys/blkptr.h>
59 #include <sys/dsl_bookmark.h>
60 #include <sys/zfeature.h>
61 #include <sys/bqueue.h>
62 #include <sys/zvol.h>
63 #include <sys/policy.h>
64 #include <sys/objlist.h>
65 #ifdef _KERNEL
66 #include <sys/zfs_vfsops.h>
67 #endif
68 
69 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
70 static int zfs_send_corrupt_data = B_FALSE;
71 /*
72  * This tunable controls the amount of data (measured in bytes) that will be
73  * prefetched by zfs send.  If the main thread is blocking on reads that haven't
74  * completed, this variable might need to be increased.  If instead the main
75  * thread is issuing new reads because the prefetches have fallen out of the
76  * cache, this may need to be decreased.
77  */
78 static int zfs_send_queue_length = SPA_MAXBLOCKSIZE;
79 /*
80  * This tunable controls the length of the queues that zfs send worker threads
81  * use to communicate.  If the send_main_thread is blocking on these queues,
82  * this variable may need to be increased.  If there is a significant slowdown
83  * at the start of a send as these threads consume all the available IO
84  * resources, this variable may need to be decreased.
85  */
86 static int zfs_send_no_prefetch_queue_length = 1024 * 1024;
87 /*
88  * These tunables control the fill fraction of the queues by zfs send.  The fill
89  * fraction controls the frequency with which threads have to be cv_signaled.
90  * If a lot of cpu time is being spent on cv_signal, then these should be tuned
91  * down.  If the queues empty before the signalled thread can catch up, then
92  * these should be tuned up.
93  */
94 static int zfs_send_queue_ff = 20;
95 static int zfs_send_no_prefetch_queue_ff = 20;
96 
97 /*
98  * Use this to override the recordsize calculation for fast zfs send estimates.
99  */
100 static int zfs_override_estimate_recordsize = 0;
101 
102 /* Set this tunable to FALSE to disable setting of DRR_FLAG_FREERECORDS */
103 static const boolean_t zfs_send_set_freerecords_bit = B_TRUE;
104 
105 /* Set this tunable to FALSE is disable sending unmodified spill blocks. */
106 static int zfs_send_unmodified_spill_blocks = B_TRUE;
107 
108 static inline boolean_t
109 overflow_multiply(uint64_t a, uint64_t b, uint64_t *c)
110 {
111 	uint64_t temp = a * b;
112 	if (b != 0 && temp / b != a)
113 		return (B_FALSE);
114 	*c = temp;
115 	return (B_TRUE);
116 }
117 
118 struct send_thread_arg {
119 	bqueue_t	q;
120 	objset_t	*os;		/* Objset to traverse */
121 	uint64_t	fromtxg;	/* Traverse from this txg */
122 	int		flags;		/* flags to pass to traverse_dataset */
123 	int		error_code;
124 	boolean_t	cancel;
125 	zbookmark_phys_t resume;
126 	uint64_t	*num_blocks_visited;
127 };
128 
129 struct redact_list_thread_arg {
130 	boolean_t		cancel;
131 	bqueue_t		q;
132 	zbookmark_phys_t	resume;
133 	redaction_list_t	*rl;
134 	boolean_t		mark_redact;
135 	int			error_code;
136 	uint64_t		*num_blocks_visited;
137 };
138 
139 struct send_merge_thread_arg {
140 	bqueue_t			q;
141 	objset_t			*os;
142 	struct redact_list_thread_arg	*from_arg;
143 	struct send_thread_arg		*to_arg;
144 	struct redact_list_thread_arg	*redact_arg;
145 	int				error;
146 	boolean_t			cancel;
147 };
148 
149 struct send_range {
150 	boolean_t		eos_marker; /* Marks the end of the stream */
151 	uint64_t		object;
152 	uint64_t		start_blkid;
153 	uint64_t		end_blkid;
154 	bqueue_node_t		ln;
155 	enum type {DATA, HOLE, OBJECT, OBJECT_RANGE, REDACT,
156 	    PREVIOUSLY_REDACTED} type;
157 	union {
158 		struct srd {
159 			dmu_object_type_t	obj_type;
160 			uint32_t		datablksz; // logical size
161 			uint32_t		datasz; // payload size
162 			blkptr_t		bp;
163 			arc_buf_t		*abuf;
164 			abd_t			*abd;
165 			kmutex_t		lock;
166 			kcondvar_t		cv;
167 			boolean_t		io_outstanding;
168 			boolean_t		io_compressed;
169 			int			io_err;
170 		} data;
171 		struct srh {
172 			uint32_t		datablksz;
173 		} hole;
174 		struct sro {
175 			/*
176 			 * This is a pointer because embedding it in the
177 			 * struct causes these structures to be massively larger
178 			 * for all range types; this makes the code much less
179 			 * memory efficient.
180 			 */
181 			dnode_phys_t		*dnp;
182 			blkptr_t		bp;
183 		} object;
184 		struct srr {
185 			uint32_t		datablksz;
186 		} redact;
187 		struct sror {
188 			blkptr_t		bp;
189 		} object_range;
190 	} sru;
191 };
192 
193 /*
194  * The list of data whose inclusion in a send stream can be pending from
195  * one call to backup_cb to another.  Multiple calls to dump_free(),
196  * dump_freeobjects(), and dump_redact() can be aggregated into a single
197  * DRR_FREE, DRR_FREEOBJECTS, or DRR_REDACT replay record.
198  */
199 typedef enum {
200 	PENDING_NONE,
201 	PENDING_FREE,
202 	PENDING_FREEOBJECTS,
203 	PENDING_REDACT
204 } dmu_pendop_t;
205 
206 typedef struct dmu_send_cookie {
207 	dmu_replay_record_t *dsc_drr;
208 	dmu_send_outparams_t *dsc_dso;
209 	offset_t *dsc_off;
210 	objset_t *dsc_os;
211 	zio_cksum_t dsc_zc;
212 	uint64_t dsc_toguid;
213 	uint64_t dsc_fromtxg;
214 	int dsc_err;
215 	dmu_pendop_t dsc_pending_op;
216 	uint64_t dsc_featureflags;
217 	uint64_t dsc_last_data_object;
218 	uint64_t dsc_last_data_offset;
219 	uint64_t dsc_resume_object;
220 	uint64_t dsc_resume_offset;
221 	boolean_t dsc_sent_begin;
222 	boolean_t dsc_sent_end;
223 } dmu_send_cookie_t;
224 
225 static int do_dump(dmu_send_cookie_t *dscp, struct send_range *range);
226 
227 static void
228 range_free(struct send_range *range)
229 {
230 	if (range->type == OBJECT) {
231 		size_t size = sizeof (dnode_phys_t) *
232 		    (range->sru.object.dnp->dn_extra_slots + 1);
233 		kmem_free(range->sru.object.dnp, size);
234 	} else if (range->type == DATA) {
235 		mutex_enter(&range->sru.data.lock);
236 		while (range->sru.data.io_outstanding)
237 			cv_wait(&range->sru.data.cv, &range->sru.data.lock);
238 		if (range->sru.data.abd != NULL)
239 			abd_free(range->sru.data.abd);
240 		if (range->sru.data.abuf != NULL) {
241 			arc_buf_destroy(range->sru.data.abuf,
242 			    &range->sru.data.abuf);
243 		}
244 		mutex_exit(&range->sru.data.lock);
245 
246 		cv_destroy(&range->sru.data.cv);
247 		mutex_destroy(&range->sru.data.lock);
248 	}
249 	kmem_free(range, sizeof (*range));
250 }
251 
252 /*
253  * For all record types except BEGIN, fill in the checksum (overlaid in
254  * drr_u.drr_checksum.drr_checksum).  The checksum verifies everything
255  * up to the start of the checksum itself.
256  */
257 static int
258 dump_record(dmu_send_cookie_t *dscp, void *payload, int payload_len)
259 {
260 	dmu_send_outparams_t *dso = dscp->dsc_dso;
261 	ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
262 	    ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
263 	(void) fletcher_4_incremental_native(dscp->dsc_drr,
264 	    offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
265 	    &dscp->dsc_zc);
266 	if (dscp->dsc_drr->drr_type == DRR_BEGIN) {
267 		dscp->dsc_sent_begin = B_TRUE;
268 	} else {
269 		ASSERT(ZIO_CHECKSUM_IS_ZERO(&dscp->dsc_drr->drr_u.
270 		    drr_checksum.drr_checksum));
271 		dscp->dsc_drr->drr_u.drr_checksum.drr_checksum = dscp->dsc_zc;
272 	}
273 	if (dscp->dsc_drr->drr_type == DRR_END) {
274 		dscp->dsc_sent_end = B_TRUE;
275 	}
276 	(void) fletcher_4_incremental_native(&dscp->dsc_drr->
277 	    drr_u.drr_checksum.drr_checksum,
278 	    sizeof (zio_cksum_t), &dscp->dsc_zc);
279 	*dscp->dsc_off += sizeof (dmu_replay_record_t);
280 	dscp->dsc_err = dso->dso_outfunc(dscp->dsc_os, dscp->dsc_drr,
281 	    sizeof (dmu_replay_record_t), dso->dso_arg);
282 	if (dscp->dsc_err != 0)
283 		return (SET_ERROR(EINTR));
284 	if (payload_len != 0) {
285 		*dscp->dsc_off += payload_len;
286 		/*
287 		 * payload is null when dso_dryrun == B_TRUE (i.e. when we're
288 		 * doing a send size calculation)
289 		 */
290 		if (payload != NULL) {
291 			(void) fletcher_4_incremental_native(
292 			    payload, payload_len, &dscp->dsc_zc);
293 		}
294 
295 		/*
296 		 * The code does not rely on this (len being a multiple of 8).
297 		 * We keep this assertion because of the corresponding assertion
298 		 * in receive_read().  Keeping this assertion ensures that we do
299 		 * not inadvertently break backwards compatibility (causing the
300 		 * assertion in receive_read() to trigger on old software).
301 		 *
302 		 * Raw sends cannot be received on old software, and so can
303 		 * bypass this assertion.
304 		 */
305 
306 		ASSERT((payload_len % 8 == 0) ||
307 		    (dscp->dsc_featureflags & DMU_BACKUP_FEATURE_RAW));
308 
309 		dscp->dsc_err = dso->dso_outfunc(dscp->dsc_os, payload,
310 		    payload_len, dso->dso_arg);
311 		if (dscp->dsc_err != 0)
312 			return (SET_ERROR(EINTR));
313 	}
314 	return (0);
315 }
316 
317 /*
318  * Fill in the drr_free struct, or perform aggregation if the previous record is
319  * also a free record, and the two are adjacent.
320  *
321  * Note that we send free records even for a full send, because we want to be
322  * able to receive a full send as a clone, which requires a list of all the free
323  * and freeobject records that were generated on the source.
324  */
325 static int
326 dump_free(dmu_send_cookie_t *dscp, uint64_t object, uint64_t offset,
327     uint64_t length)
328 {
329 	struct drr_free *drrf = &(dscp->dsc_drr->drr_u.drr_free);
330 
331 	/*
332 	 * When we receive a free record, dbuf_free_range() assumes
333 	 * that the receiving system doesn't have any dbufs in the range
334 	 * being freed.  This is always true because there is a one-record
335 	 * constraint: we only send one WRITE record for any given
336 	 * object,offset.  We know that the one-record constraint is
337 	 * true because we always send data in increasing order by
338 	 * object,offset.
339 	 *
340 	 * If the increasing-order constraint ever changes, we should find
341 	 * another way to assert that the one-record constraint is still
342 	 * satisfied.
343 	 */
344 	ASSERT(object > dscp->dsc_last_data_object ||
345 	    (object == dscp->dsc_last_data_object &&
346 	    offset > dscp->dsc_last_data_offset));
347 
348 	/*
349 	 * If there is a pending op, but it's not PENDING_FREE, push it out,
350 	 * since free block aggregation can only be done for blocks of the
351 	 * same type (i.e., DRR_FREE records can only be aggregated with
352 	 * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
353 	 * aggregated with other DRR_FREEOBJECTS records).
354 	 */
355 	if (dscp->dsc_pending_op != PENDING_NONE &&
356 	    dscp->dsc_pending_op != PENDING_FREE) {
357 		if (dump_record(dscp, NULL, 0) != 0)
358 			return (SET_ERROR(EINTR));
359 		dscp->dsc_pending_op = PENDING_NONE;
360 	}
361 
362 	if (dscp->dsc_pending_op == PENDING_FREE) {
363 		/*
364 		 * Check to see whether this free block can be aggregated
365 		 * with pending one.
366 		 */
367 		if (drrf->drr_object == object && drrf->drr_offset +
368 		    drrf->drr_length == offset) {
369 			if (offset + length < offset || length == UINT64_MAX)
370 				drrf->drr_length = UINT64_MAX;
371 			else
372 				drrf->drr_length += length;
373 			return (0);
374 		} else {
375 			/* not a continuation.  Push out pending record */
376 			if (dump_record(dscp, NULL, 0) != 0)
377 				return (SET_ERROR(EINTR));
378 			dscp->dsc_pending_op = PENDING_NONE;
379 		}
380 	}
381 	/* create a FREE record and make it pending */
382 	memset(dscp->dsc_drr, 0, sizeof (dmu_replay_record_t));
383 	dscp->dsc_drr->drr_type = DRR_FREE;
384 	drrf->drr_object = object;
385 	drrf->drr_offset = offset;
386 	if (offset + length < offset)
387 		drrf->drr_length = DMU_OBJECT_END;
388 	else
389 		drrf->drr_length = length;
390 	drrf->drr_toguid = dscp->dsc_toguid;
391 	if (length == DMU_OBJECT_END) {
392 		if (dump_record(dscp, NULL, 0) != 0)
393 			return (SET_ERROR(EINTR));
394 	} else {
395 		dscp->dsc_pending_op = PENDING_FREE;
396 	}
397 
398 	return (0);
399 }
400 
401 /*
402  * Fill in the drr_redact struct, or perform aggregation if the previous record
403  * is also a redaction record, and the two are adjacent.
404  */
405 static int
406 dump_redact(dmu_send_cookie_t *dscp, uint64_t object, uint64_t offset,
407     uint64_t length)
408 {
409 	struct drr_redact *drrr = &dscp->dsc_drr->drr_u.drr_redact;
410 
411 	/*
412 	 * If there is a pending op, but it's not PENDING_REDACT, push it out,
413 	 * since free block aggregation can only be done for blocks of the
414 	 * same type (i.e., DRR_REDACT records can only be aggregated with
415 	 * other DRR_REDACT records).
416 	 */
417 	if (dscp->dsc_pending_op != PENDING_NONE &&
418 	    dscp->dsc_pending_op != PENDING_REDACT) {
419 		if (dump_record(dscp, NULL, 0) != 0)
420 			return (SET_ERROR(EINTR));
421 		dscp->dsc_pending_op = PENDING_NONE;
422 	}
423 
424 	if (dscp->dsc_pending_op == PENDING_REDACT) {
425 		/*
426 		 * Check to see whether this redacted block can be aggregated
427 		 * with pending one.
428 		 */
429 		if (drrr->drr_object == object && drrr->drr_offset +
430 		    drrr->drr_length == offset) {
431 			drrr->drr_length += length;
432 			return (0);
433 		} else {
434 			/* not a continuation.  Push out pending record */
435 			if (dump_record(dscp, NULL, 0) != 0)
436 				return (SET_ERROR(EINTR));
437 			dscp->dsc_pending_op = PENDING_NONE;
438 		}
439 	}
440 	/* create a REDACT record and make it pending */
441 	memset(dscp->dsc_drr, 0, sizeof (dmu_replay_record_t));
442 	dscp->dsc_drr->drr_type = DRR_REDACT;
443 	drrr->drr_object = object;
444 	drrr->drr_offset = offset;
445 	drrr->drr_length = length;
446 	drrr->drr_toguid = dscp->dsc_toguid;
447 	dscp->dsc_pending_op = PENDING_REDACT;
448 
449 	return (0);
450 }
451 
452 static int
453 dmu_dump_write(dmu_send_cookie_t *dscp, dmu_object_type_t type, uint64_t object,
454     uint64_t offset, int lsize, int psize, const blkptr_t *bp,
455     boolean_t io_compressed, void *data)
456 {
457 	uint64_t payload_size;
458 	boolean_t raw = (dscp->dsc_featureflags & DMU_BACKUP_FEATURE_RAW);
459 	struct drr_write *drrw = &(dscp->dsc_drr->drr_u.drr_write);
460 
461 	/*
462 	 * We send data in increasing object, offset order.
463 	 * See comment in dump_free() for details.
464 	 */
465 	ASSERT(object > dscp->dsc_last_data_object ||
466 	    (object == dscp->dsc_last_data_object &&
467 	    offset > dscp->dsc_last_data_offset));
468 	dscp->dsc_last_data_object = object;
469 	dscp->dsc_last_data_offset = offset + lsize - 1;
470 
471 	/*
472 	 * If there is any kind of pending aggregation (currently either
473 	 * a grouping of free objects or free blocks), push it out to
474 	 * the stream, since aggregation can't be done across operations
475 	 * of different types.
476 	 */
477 	if (dscp->dsc_pending_op != PENDING_NONE) {
478 		if (dump_record(dscp, NULL, 0) != 0)
479 			return (SET_ERROR(EINTR));
480 		dscp->dsc_pending_op = PENDING_NONE;
481 	}
482 	/* write a WRITE record */
483 	memset(dscp->dsc_drr, 0, sizeof (dmu_replay_record_t));
484 	dscp->dsc_drr->drr_type = DRR_WRITE;
485 	drrw->drr_object = object;
486 	drrw->drr_type = type;
487 	drrw->drr_offset = offset;
488 	drrw->drr_toguid = dscp->dsc_toguid;
489 	drrw->drr_logical_size = lsize;
490 
491 	/* only set the compression fields if the buf is compressed or raw */
492 	boolean_t compressed =
493 	    (bp != NULL ? BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
494 	    io_compressed : lsize != psize);
495 	if (raw || compressed) {
496 		ASSERT(bp != NULL);
497 		ASSERT(raw || dscp->dsc_featureflags &
498 		    DMU_BACKUP_FEATURE_COMPRESSED);
499 		ASSERT(!BP_IS_EMBEDDED(bp));
500 		ASSERT3S(psize, >, 0);
501 
502 		if (raw) {
503 			ASSERT(BP_IS_PROTECTED(bp));
504 
505 			/*
506 			 * This is a raw protected block so we need to pass
507 			 * along everything the receiving side will need to
508 			 * interpret this block, including the byteswap, salt,
509 			 * IV, and MAC.
510 			 */
511 			if (BP_SHOULD_BYTESWAP(bp))
512 				drrw->drr_flags |= DRR_RAW_BYTESWAP;
513 			zio_crypt_decode_params_bp(bp, drrw->drr_salt,
514 			    drrw->drr_iv);
515 			zio_crypt_decode_mac_bp(bp, drrw->drr_mac);
516 		} else {
517 			/* this is a compressed block */
518 			ASSERT(dscp->dsc_featureflags &
519 			    DMU_BACKUP_FEATURE_COMPRESSED);
520 			ASSERT(!BP_SHOULD_BYTESWAP(bp));
521 			ASSERT(!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)));
522 			ASSERT3U(BP_GET_COMPRESS(bp), !=, ZIO_COMPRESS_OFF);
523 			ASSERT3S(lsize, >=, psize);
524 		}
525 
526 		/* set fields common to compressed and raw sends */
527 		drrw->drr_compressiontype = BP_GET_COMPRESS(bp);
528 		drrw->drr_compressed_size = psize;
529 		payload_size = drrw->drr_compressed_size;
530 	} else {
531 		payload_size = drrw->drr_logical_size;
532 	}
533 
534 	if (bp == NULL || BP_IS_EMBEDDED(bp) || (BP_IS_PROTECTED(bp) && !raw)) {
535 		/*
536 		 * There's no pre-computed checksum for partial-block writes,
537 		 * embedded BP's, or encrypted BP's that are being sent as
538 		 * plaintext, so (like fletcher4-checksummed blocks) userland
539 		 * will have to compute a dedup-capable checksum itself.
540 		 */
541 		drrw->drr_checksumtype = ZIO_CHECKSUM_OFF;
542 	} else {
543 		drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
544 		if (zio_checksum_table[drrw->drr_checksumtype].ci_flags &
545 		    ZCHECKSUM_FLAG_DEDUP)
546 			drrw->drr_flags |= DRR_CHECKSUM_DEDUP;
547 		DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
548 		DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
549 		DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
550 		DDK_SET_CRYPT(&drrw->drr_key, BP_IS_PROTECTED(bp));
551 		drrw->drr_key.ddk_cksum = bp->blk_cksum;
552 	}
553 
554 	if (dump_record(dscp, data, payload_size) != 0)
555 		return (SET_ERROR(EINTR));
556 	return (0);
557 }
558 
559 static int
560 dump_write_embedded(dmu_send_cookie_t *dscp, uint64_t object, uint64_t offset,
561     int blksz, const blkptr_t *bp)
562 {
563 	char buf[BPE_PAYLOAD_SIZE];
564 	struct drr_write_embedded *drrw =
565 	    &(dscp->dsc_drr->drr_u.drr_write_embedded);
566 
567 	if (dscp->dsc_pending_op != PENDING_NONE) {
568 		if (dump_record(dscp, NULL, 0) != 0)
569 			return (SET_ERROR(EINTR));
570 		dscp->dsc_pending_op = PENDING_NONE;
571 	}
572 
573 	ASSERT(BP_IS_EMBEDDED(bp));
574 
575 	memset(dscp->dsc_drr, 0, sizeof (dmu_replay_record_t));
576 	dscp->dsc_drr->drr_type = DRR_WRITE_EMBEDDED;
577 	drrw->drr_object = object;
578 	drrw->drr_offset = offset;
579 	drrw->drr_length = blksz;
580 	drrw->drr_toguid = dscp->dsc_toguid;
581 	drrw->drr_compression = BP_GET_COMPRESS(bp);
582 	drrw->drr_etype = BPE_GET_ETYPE(bp);
583 	drrw->drr_lsize = BPE_GET_LSIZE(bp);
584 	drrw->drr_psize = BPE_GET_PSIZE(bp);
585 
586 	decode_embedded_bp_compressed(bp, buf);
587 
588 	if (dump_record(dscp, buf, P2ROUNDUP(drrw->drr_psize, 8)) != 0)
589 		return (SET_ERROR(EINTR));
590 	return (0);
591 }
592 
593 static int
594 dump_spill(dmu_send_cookie_t *dscp, const blkptr_t *bp, uint64_t object,
595     void *data)
596 {
597 	struct drr_spill *drrs = &(dscp->dsc_drr->drr_u.drr_spill);
598 	uint64_t blksz = BP_GET_LSIZE(bp);
599 	uint64_t payload_size = blksz;
600 
601 	if (dscp->dsc_pending_op != PENDING_NONE) {
602 		if (dump_record(dscp, NULL, 0) != 0)
603 			return (SET_ERROR(EINTR));
604 		dscp->dsc_pending_op = PENDING_NONE;
605 	}
606 
607 	/* write a SPILL record */
608 	memset(dscp->dsc_drr, 0, sizeof (dmu_replay_record_t));
609 	dscp->dsc_drr->drr_type = DRR_SPILL;
610 	drrs->drr_object = object;
611 	drrs->drr_length = blksz;
612 	drrs->drr_toguid = dscp->dsc_toguid;
613 
614 	/* See comment in dump_dnode() for full details */
615 	if (zfs_send_unmodified_spill_blocks &&
616 	    (bp->blk_birth <= dscp->dsc_fromtxg)) {
617 		drrs->drr_flags |= DRR_SPILL_UNMODIFIED;
618 	}
619 
620 	/* handle raw send fields */
621 	if (dscp->dsc_featureflags & DMU_BACKUP_FEATURE_RAW) {
622 		ASSERT(BP_IS_PROTECTED(bp));
623 
624 		if (BP_SHOULD_BYTESWAP(bp))
625 			drrs->drr_flags |= DRR_RAW_BYTESWAP;
626 		drrs->drr_compressiontype = BP_GET_COMPRESS(bp);
627 		drrs->drr_compressed_size = BP_GET_PSIZE(bp);
628 		zio_crypt_decode_params_bp(bp, drrs->drr_salt, drrs->drr_iv);
629 		zio_crypt_decode_mac_bp(bp, drrs->drr_mac);
630 		payload_size = drrs->drr_compressed_size;
631 	}
632 
633 	if (dump_record(dscp, data, payload_size) != 0)
634 		return (SET_ERROR(EINTR));
635 	return (0);
636 }
637 
638 static int
639 dump_freeobjects(dmu_send_cookie_t *dscp, uint64_t firstobj, uint64_t numobjs)
640 {
641 	struct drr_freeobjects *drrfo = &(dscp->dsc_drr->drr_u.drr_freeobjects);
642 	uint64_t maxobj = DNODES_PER_BLOCK *
643 	    (DMU_META_DNODE(dscp->dsc_os)->dn_maxblkid + 1);
644 
645 	/*
646 	 * ZoL < 0.7 does not handle large FREEOBJECTS records correctly,
647 	 * leading to zfs recv never completing. to avoid this issue, don't
648 	 * send FREEOBJECTS records for object IDs which cannot exist on the
649 	 * receiving side.
650 	 */
651 	if (maxobj > 0) {
652 		if (maxobj <= firstobj)
653 			return (0);
654 
655 		if (maxobj < firstobj + numobjs)
656 			numobjs = maxobj - firstobj;
657 	}
658 
659 	/*
660 	 * If there is a pending op, but it's not PENDING_FREEOBJECTS,
661 	 * push it out, since free block aggregation can only be done for
662 	 * blocks of the same type (i.e., DRR_FREE records can only be
663 	 * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
664 	 * can only be aggregated with other DRR_FREEOBJECTS records).
665 	 */
666 	if (dscp->dsc_pending_op != PENDING_NONE &&
667 	    dscp->dsc_pending_op != PENDING_FREEOBJECTS) {
668 		if (dump_record(dscp, NULL, 0) != 0)
669 			return (SET_ERROR(EINTR));
670 		dscp->dsc_pending_op = PENDING_NONE;
671 	}
672 
673 	if (dscp->dsc_pending_op == PENDING_FREEOBJECTS) {
674 		/*
675 		 * See whether this free object array can be aggregated
676 		 * with pending one
677 		 */
678 		if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
679 			drrfo->drr_numobjs += numobjs;
680 			return (0);
681 		} else {
682 			/* can't be aggregated.  Push out pending record */
683 			if (dump_record(dscp, NULL, 0) != 0)
684 				return (SET_ERROR(EINTR));
685 			dscp->dsc_pending_op = PENDING_NONE;
686 		}
687 	}
688 
689 	/* write a FREEOBJECTS record */
690 	memset(dscp->dsc_drr, 0, sizeof (dmu_replay_record_t));
691 	dscp->dsc_drr->drr_type = DRR_FREEOBJECTS;
692 	drrfo->drr_firstobj = firstobj;
693 	drrfo->drr_numobjs = numobjs;
694 	drrfo->drr_toguid = dscp->dsc_toguid;
695 
696 	dscp->dsc_pending_op = PENDING_FREEOBJECTS;
697 
698 	return (0);
699 }
700 
701 static int
702 dump_dnode(dmu_send_cookie_t *dscp, const blkptr_t *bp, uint64_t object,
703     dnode_phys_t *dnp)
704 {
705 	struct drr_object *drro = &(dscp->dsc_drr->drr_u.drr_object);
706 	int bonuslen;
707 
708 	if (object < dscp->dsc_resume_object) {
709 		/*
710 		 * Note: when resuming, we will visit all the dnodes in
711 		 * the block of dnodes that we are resuming from.  In
712 		 * this case it's unnecessary to send the dnodes prior to
713 		 * the one we are resuming from.  We should be at most one
714 		 * block's worth of dnodes behind the resume point.
715 		 */
716 		ASSERT3U(dscp->dsc_resume_object - object, <,
717 		    1 << (DNODE_BLOCK_SHIFT - DNODE_SHIFT));
718 		return (0);
719 	}
720 
721 	if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
722 		return (dump_freeobjects(dscp, object, 1));
723 
724 	if (dscp->dsc_pending_op != PENDING_NONE) {
725 		if (dump_record(dscp, NULL, 0) != 0)
726 			return (SET_ERROR(EINTR));
727 		dscp->dsc_pending_op = PENDING_NONE;
728 	}
729 
730 	/* write an OBJECT record */
731 	memset(dscp->dsc_drr, 0, sizeof (dmu_replay_record_t));
732 	dscp->dsc_drr->drr_type = DRR_OBJECT;
733 	drro->drr_object = object;
734 	drro->drr_type = dnp->dn_type;
735 	drro->drr_bonustype = dnp->dn_bonustype;
736 	drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
737 	drro->drr_bonuslen = dnp->dn_bonuslen;
738 	drro->drr_dn_slots = dnp->dn_extra_slots + 1;
739 	drro->drr_checksumtype = dnp->dn_checksum;
740 	drro->drr_compress = dnp->dn_compress;
741 	drro->drr_toguid = dscp->dsc_toguid;
742 
743 	if (!(dscp->dsc_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
744 	    drro->drr_blksz > SPA_OLD_MAXBLOCKSIZE)
745 		drro->drr_blksz = SPA_OLD_MAXBLOCKSIZE;
746 
747 	bonuslen = P2ROUNDUP(dnp->dn_bonuslen, 8);
748 
749 	if ((dscp->dsc_featureflags & DMU_BACKUP_FEATURE_RAW)) {
750 		ASSERT(BP_IS_ENCRYPTED(bp));
751 
752 		if (BP_SHOULD_BYTESWAP(bp))
753 			drro->drr_flags |= DRR_RAW_BYTESWAP;
754 
755 		/* needed for reconstructing dnp on recv side */
756 		drro->drr_maxblkid = dnp->dn_maxblkid;
757 		drro->drr_indblkshift = dnp->dn_indblkshift;
758 		drro->drr_nlevels = dnp->dn_nlevels;
759 		drro->drr_nblkptr = dnp->dn_nblkptr;
760 
761 		/*
762 		 * Since we encrypt the entire bonus area, the (raw) part
763 		 * beyond the bonuslen is actually nonzero, so we need
764 		 * to send it.
765 		 */
766 		if (bonuslen != 0) {
767 			if (drro->drr_bonuslen > DN_MAX_BONUS_LEN(dnp))
768 				return (SET_ERROR(EINVAL));
769 			drro->drr_raw_bonuslen = DN_MAX_BONUS_LEN(dnp);
770 			bonuslen = drro->drr_raw_bonuslen;
771 		}
772 	}
773 
774 	/*
775 	 * DRR_OBJECT_SPILL is set for every dnode which references a
776 	 * spill block.	 This allows the receiving pool to definitively
777 	 * determine when a spill block should be kept or freed.
778 	 */
779 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
780 		drro->drr_flags |= DRR_OBJECT_SPILL;
781 
782 	if (dump_record(dscp, DN_BONUS(dnp), bonuslen) != 0)
783 		return (SET_ERROR(EINTR));
784 
785 	/* Free anything past the end of the file. */
786 	if (dump_free(dscp, object, (dnp->dn_maxblkid + 1) *
787 	    (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), DMU_OBJECT_END) != 0)
788 		return (SET_ERROR(EINTR));
789 
790 	/*
791 	 * Send DRR_SPILL records for unmodified spill blocks.	This is useful
792 	 * because changing certain attributes of the object (e.g. blocksize)
793 	 * can cause old versions of ZFS to incorrectly remove a spill block.
794 	 * Including these records in the stream forces an up to date version
795 	 * to always be written ensuring they're never lost.  Current versions
796 	 * of the code which understand the DRR_FLAG_SPILL_BLOCK feature can
797 	 * ignore these unmodified spill blocks.
798 	 */
799 	if (zfs_send_unmodified_spill_blocks &&
800 	    (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) &&
801 	    (DN_SPILL_BLKPTR(dnp)->blk_birth <= dscp->dsc_fromtxg)) {
802 		struct send_range record;
803 		blkptr_t *bp = DN_SPILL_BLKPTR(dnp);
804 
805 		memset(&record, 0, sizeof (struct send_range));
806 		record.type = DATA;
807 		record.object = object;
808 		record.eos_marker = B_FALSE;
809 		record.start_blkid = DMU_SPILL_BLKID;
810 		record.end_blkid = record.start_blkid + 1;
811 		record.sru.data.bp = *bp;
812 		record.sru.data.obj_type = dnp->dn_type;
813 		record.sru.data.datablksz = BP_GET_LSIZE(bp);
814 
815 		if (do_dump(dscp, &record) != 0)
816 			return (SET_ERROR(EINTR));
817 	}
818 
819 	if (dscp->dsc_err != 0)
820 		return (SET_ERROR(EINTR));
821 
822 	return (0);
823 }
824 
825 static int
826 dump_object_range(dmu_send_cookie_t *dscp, const blkptr_t *bp,
827     uint64_t firstobj, uint64_t numslots)
828 {
829 	struct drr_object_range *drror =
830 	    &(dscp->dsc_drr->drr_u.drr_object_range);
831 
832 	/* we only use this record type for raw sends */
833 	ASSERT(BP_IS_PROTECTED(bp));
834 	ASSERT(dscp->dsc_featureflags & DMU_BACKUP_FEATURE_RAW);
835 	ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
836 	ASSERT3U(BP_GET_TYPE(bp), ==, DMU_OT_DNODE);
837 	ASSERT0(BP_GET_LEVEL(bp));
838 
839 	if (dscp->dsc_pending_op != PENDING_NONE) {
840 		if (dump_record(dscp, NULL, 0) != 0)
841 			return (SET_ERROR(EINTR));
842 		dscp->dsc_pending_op = PENDING_NONE;
843 	}
844 
845 	memset(dscp->dsc_drr, 0, sizeof (dmu_replay_record_t));
846 	dscp->dsc_drr->drr_type = DRR_OBJECT_RANGE;
847 	drror->drr_firstobj = firstobj;
848 	drror->drr_numslots = numslots;
849 	drror->drr_toguid = dscp->dsc_toguid;
850 	if (BP_SHOULD_BYTESWAP(bp))
851 		drror->drr_flags |= DRR_RAW_BYTESWAP;
852 	zio_crypt_decode_params_bp(bp, drror->drr_salt, drror->drr_iv);
853 	zio_crypt_decode_mac_bp(bp, drror->drr_mac);
854 
855 	if (dump_record(dscp, NULL, 0) != 0)
856 		return (SET_ERROR(EINTR));
857 	return (0);
858 }
859 
860 static boolean_t
861 send_do_embed(const blkptr_t *bp, uint64_t featureflags)
862 {
863 	if (!BP_IS_EMBEDDED(bp))
864 		return (B_FALSE);
865 
866 	/*
867 	 * Compression function must be legacy, or explicitly enabled.
868 	 */
869 	if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS &&
870 	    !(featureflags & DMU_BACKUP_FEATURE_LZ4)))
871 		return (B_FALSE);
872 
873 	/*
874 	 * If we have not set the ZSTD feature flag, we can't send ZSTD
875 	 * compressed embedded blocks, as the receiver may not support them.
876 	 */
877 	if ((BP_GET_COMPRESS(bp) == ZIO_COMPRESS_ZSTD &&
878 	    !(featureflags & DMU_BACKUP_FEATURE_ZSTD)))
879 		return (B_FALSE);
880 
881 	/*
882 	 * Embed type must be explicitly enabled.
883 	 */
884 	switch (BPE_GET_ETYPE(bp)) {
885 	case BP_EMBEDDED_TYPE_DATA:
886 		if (featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)
887 			return (B_TRUE);
888 		break;
889 	default:
890 		return (B_FALSE);
891 	}
892 	return (B_FALSE);
893 }
894 
895 /*
896  * This function actually handles figuring out what kind of record needs to be
897  * dumped, and calling the appropriate helper function.  In most cases,
898  * the data has already been read by send_reader_thread().
899  */
900 static int
901 do_dump(dmu_send_cookie_t *dscp, struct send_range *range)
902 {
903 	int err = 0;
904 	switch (range->type) {
905 	case OBJECT:
906 		err = dump_dnode(dscp, &range->sru.object.bp, range->object,
907 		    range->sru.object.dnp);
908 		return (err);
909 	case OBJECT_RANGE: {
910 		ASSERT3U(range->start_blkid + 1, ==, range->end_blkid);
911 		if (!(dscp->dsc_featureflags & DMU_BACKUP_FEATURE_RAW)) {
912 			return (0);
913 		}
914 		uint64_t epb = BP_GET_LSIZE(&range->sru.object_range.bp) >>
915 		    DNODE_SHIFT;
916 		uint64_t firstobj = range->start_blkid * epb;
917 		err = dump_object_range(dscp, &range->sru.object_range.bp,
918 		    firstobj, epb);
919 		break;
920 	}
921 	case REDACT: {
922 		struct srr *srrp = &range->sru.redact;
923 		err = dump_redact(dscp, range->object, range->start_blkid *
924 		    srrp->datablksz, (range->end_blkid - range->start_blkid) *
925 		    srrp->datablksz);
926 		return (err);
927 	}
928 	case DATA: {
929 		struct srd *srdp = &range->sru.data;
930 		blkptr_t *bp = &srdp->bp;
931 		spa_t *spa =
932 		    dmu_objset_spa(dscp->dsc_os);
933 
934 		ASSERT3U(srdp->datablksz, ==, BP_GET_LSIZE(bp));
935 		ASSERT3U(range->start_blkid + 1, ==, range->end_blkid);
936 		if (BP_GET_TYPE(bp) == DMU_OT_SA) {
937 			arc_flags_t aflags = ARC_FLAG_WAIT;
938 			enum zio_flag zioflags = ZIO_FLAG_CANFAIL;
939 
940 			if (dscp->dsc_featureflags & DMU_BACKUP_FEATURE_RAW) {
941 				ASSERT(BP_IS_PROTECTED(bp));
942 				zioflags |= ZIO_FLAG_RAW;
943 			}
944 
945 			zbookmark_phys_t zb;
946 			ASSERT3U(range->start_blkid, ==, DMU_SPILL_BLKID);
947 			zb.zb_objset = dmu_objset_id(dscp->dsc_os);
948 			zb.zb_object = range->object;
949 			zb.zb_level = 0;
950 			zb.zb_blkid = range->start_blkid;
951 
952 			arc_buf_t *abuf = NULL;
953 			if (!dscp->dsc_dso->dso_dryrun && arc_read(NULL, spa,
954 			    bp, arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
955 			    zioflags, &aflags, &zb) != 0)
956 				return (SET_ERROR(EIO));
957 
958 			err = dump_spill(dscp, bp, zb.zb_object,
959 			    (abuf == NULL ? NULL : abuf->b_data));
960 			if (abuf != NULL)
961 				arc_buf_destroy(abuf, &abuf);
962 			return (err);
963 		}
964 		if (send_do_embed(bp, dscp->dsc_featureflags)) {
965 			err = dump_write_embedded(dscp, range->object,
966 			    range->start_blkid * srdp->datablksz,
967 			    srdp->datablksz, bp);
968 			return (err);
969 		}
970 		ASSERT(range->object > dscp->dsc_resume_object ||
971 		    (range->object == dscp->dsc_resume_object &&
972 		    range->start_blkid * srdp->datablksz >=
973 		    dscp->dsc_resume_offset));
974 		/* it's a level-0 block of a regular object */
975 
976 		mutex_enter(&srdp->lock);
977 		while (srdp->io_outstanding)
978 			cv_wait(&srdp->cv, &srdp->lock);
979 		err = srdp->io_err;
980 		mutex_exit(&srdp->lock);
981 
982 		if (err != 0) {
983 			if (zfs_send_corrupt_data &&
984 			    !dscp->dsc_dso->dso_dryrun) {
985 				/*
986 				 * Send a block filled with 0x"zfs badd bloc"
987 				 */
988 				srdp->abuf = arc_alloc_buf(spa, &srdp->abuf,
989 				    ARC_BUFC_DATA, srdp->datablksz);
990 				uint64_t *ptr;
991 				for (ptr = srdp->abuf->b_data;
992 				    (char *)ptr < (char *)srdp->abuf->b_data +
993 				    srdp->datablksz; ptr++)
994 					*ptr = 0x2f5baddb10cULL;
995 			} else {
996 				return (SET_ERROR(EIO));
997 			}
998 		}
999 
1000 		ASSERT(dscp->dsc_dso->dso_dryrun ||
1001 		    srdp->abuf != NULL || srdp->abd != NULL);
1002 
1003 		uint64_t offset = range->start_blkid * srdp->datablksz;
1004 
1005 		char *data = NULL;
1006 		if (srdp->abd != NULL) {
1007 			data = abd_to_buf(srdp->abd);
1008 			ASSERT3P(srdp->abuf, ==, NULL);
1009 		} else if (srdp->abuf != NULL) {
1010 			data = srdp->abuf->b_data;
1011 		}
1012 
1013 		/*
1014 		 * If we have large blocks stored on disk but the send flags
1015 		 * don't allow us to send large blocks, we split the data from
1016 		 * the arc buf into chunks.
1017 		 */
1018 		if (srdp->datablksz > SPA_OLD_MAXBLOCKSIZE &&
1019 		    !(dscp->dsc_featureflags &
1020 		    DMU_BACKUP_FEATURE_LARGE_BLOCKS)) {
1021 			if (dscp->dsc_featureflags & DMU_BACKUP_FEATURE_RAW)
1022 				return (SET_ERROR(ENOTSUP));
1023 
1024 			while (srdp->datablksz > 0 && err == 0) {
1025 				int n = MIN(srdp->datablksz,
1026 				    SPA_OLD_MAXBLOCKSIZE);
1027 				err = dmu_dump_write(dscp, srdp->obj_type,
1028 				    range->object, offset, n, n, NULL, B_FALSE,
1029 				    data);
1030 				offset += n;
1031 				/*
1032 				 * When doing dry run, data==NULL is used as a
1033 				 * sentinel value by
1034 				 * dmu_dump_write()->dump_record().
1035 				 */
1036 				if (data != NULL)
1037 					data += n;
1038 				srdp->datablksz -= n;
1039 			}
1040 		} else {
1041 			err = dmu_dump_write(dscp, srdp->obj_type,
1042 			    range->object, offset,
1043 			    srdp->datablksz, srdp->datasz, bp,
1044 			    srdp->io_compressed, data);
1045 		}
1046 		return (err);
1047 	}
1048 	case HOLE: {
1049 		struct srh *srhp = &range->sru.hole;
1050 		if (range->object == DMU_META_DNODE_OBJECT) {
1051 			uint32_t span = srhp->datablksz >> DNODE_SHIFT;
1052 			uint64_t first_obj = range->start_blkid * span;
1053 			uint64_t numobj = range->end_blkid * span - first_obj;
1054 			return (dump_freeobjects(dscp, first_obj, numobj));
1055 		}
1056 		uint64_t offset = 0;
1057 
1058 		/*
1059 		 * If this multiply overflows, we don't need to send this block.
1060 		 * Even if it has a birth time, it can never not be a hole, so
1061 		 * we don't need to send records for it.
1062 		 */
1063 		if (!overflow_multiply(range->start_blkid, srhp->datablksz,
1064 		    &offset)) {
1065 			return (0);
1066 		}
1067 		uint64_t len = 0;
1068 
1069 		if (!overflow_multiply(range->end_blkid, srhp->datablksz, &len))
1070 			len = UINT64_MAX;
1071 		len = len - offset;
1072 		return (dump_free(dscp, range->object, offset, len));
1073 	}
1074 	default:
1075 		panic("Invalid range type in do_dump: %d", range->type);
1076 	}
1077 	return (err);
1078 }
1079 
1080 static struct send_range *
1081 range_alloc(enum type type, uint64_t object, uint64_t start_blkid,
1082     uint64_t end_blkid, boolean_t eos)
1083 {
1084 	struct send_range *range = kmem_alloc(sizeof (*range), KM_SLEEP);
1085 	range->type = type;
1086 	range->object = object;
1087 	range->start_blkid = start_blkid;
1088 	range->end_blkid = end_blkid;
1089 	range->eos_marker = eos;
1090 	if (type == DATA) {
1091 		range->sru.data.abd = NULL;
1092 		range->sru.data.abuf = NULL;
1093 		mutex_init(&range->sru.data.lock, NULL, MUTEX_DEFAULT, NULL);
1094 		cv_init(&range->sru.data.cv, NULL, CV_DEFAULT, NULL);
1095 		range->sru.data.io_outstanding = 0;
1096 		range->sru.data.io_err = 0;
1097 		range->sru.data.io_compressed = B_FALSE;
1098 	}
1099 	return (range);
1100 }
1101 
1102 /*
1103  * This is the callback function to traverse_dataset that acts as a worker
1104  * thread for dmu_send_impl.
1105  */
1106 static int
1107 send_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
1108     const zbookmark_phys_t *zb, const struct dnode_phys *dnp, void *arg)
1109 {
1110 	(void) zilog;
1111 	struct send_thread_arg *sta = arg;
1112 	struct send_range *record;
1113 
1114 	ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
1115 	    zb->zb_object >= sta->resume.zb_object);
1116 
1117 	/*
1118 	 * All bps of an encrypted os should have the encryption bit set.
1119 	 * If this is not true it indicates tampering and we report an error.
1120 	 */
1121 	if (sta->os->os_encrypted &&
1122 	    !BP_IS_HOLE(bp) && !BP_USES_CRYPT(bp)) {
1123 		spa_log_error(spa, zb);
1124 		zfs_panic_recover("unencrypted block in encrypted "
1125 		    "object set %llu", dmu_objset_id(sta->os));
1126 		return (SET_ERROR(EIO));
1127 	}
1128 
1129 	if (sta->cancel)
1130 		return (SET_ERROR(EINTR));
1131 	if (zb->zb_object != DMU_META_DNODE_OBJECT &&
1132 	    DMU_OBJECT_IS_SPECIAL(zb->zb_object))
1133 		return (0);
1134 	atomic_inc_64(sta->num_blocks_visited);
1135 
1136 	if (zb->zb_level == ZB_DNODE_LEVEL) {
1137 		if (zb->zb_object == DMU_META_DNODE_OBJECT)
1138 			return (0);
1139 		record = range_alloc(OBJECT, zb->zb_object, 0, 0, B_FALSE);
1140 		record->sru.object.bp = *bp;
1141 		size_t size  = sizeof (*dnp) * (dnp->dn_extra_slots + 1);
1142 		record->sru.object.dnp = kmem_alloc(size, KM_SLEEP);
1143 		memcpy(record->sru.object.dnp, dnp, size);
1144 		bqueue_enqueue(&sta->q, record, sizeof (*record));
1145 		return (0);
1146 	}
1147 	if (zb->zb_level == 0 && zb->zb_object == DMU_META_DNODE_OBJECT &&
1148 	    !BP_IS_HOLE(bp)) {
1149 		record = range_alloc(OBJECT_RANGE, 0, zb->zb_blkid,
1150 		    zb->zb_blkid + 1, B_FALSE);
1151 		record->sru.object_range.bp = *bp;
1152 		bqueue_enqueue(&sta->q, record, sizeof (*record));
1153 		return (0);
1154 	}
1155 	if (zb->zb_level < 0 || (zb->zb_level > 0 && !BP_IS_HOLE(bp)))
1156 		return (0);
1157 	if (zb->zb_object == DMU_META_DNODE_OBJECT && !BP_IS_HOLE(bp))
1158 		return (0);
1159 
1160 	uint64_t span = bp_span_in_blocks(dnp->dn_indblkshift, zb->zb_level);
1161 	uint64_t start;
1162 
1163 	/*
1164 	 * If this multiply overflows, we don't need to send this block.
1165 	 * Even if it has a birth time, it can never not be a hole, so
1166 	 * we don't need to send records for it.
1167 	 */
1168 	if (!overflow_multiply(span, zb->zb_blkid, &start) || (!(zb->zb_blkid ==
1169 	    DMU_SPILL_BLKID || DMU_OT_IS_METADATA(dnp->dn_type)) &&
1170 	    span * zb->zb_blkid > dnp->dn_maxblkid)) {
1171 		ASSERT(BP_IS_HOLE(bp));
1172 		return (0);
1173 	}
1174 
1175 	if (zb->zb_blkid == DMU_SPILL_BLKID)
1176 		ASSERT3U(BP_GET_TYPE(bp), ==, DMU_OT_SA);
1177 
1178 	enum type record_type = DATA;
1179 	if (BP_IS_HOLE(bp))
1180 		record_type = HOLE;
1181 	else if (BP_IS_REDACTED(bp))
1182 		record_type = REDACT;
1183 	else
1184 		record_type = DATA;
1185 
1186 	record = range_alloc(record_type, zb->zb_object, start,
1187 	    (start + span < start ? 0 : start + span), B_FALSE);
1188 
1189 	uint64_t datablksz = (zb->zb_blkid == DMU_SPILL_BLKID ?
1190 	    BP_GET_LSIZE(bp) : dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
1191 
1192 	if (BP_IS_HOLE(bp)) {
1193 		record->sru.hole.datablksz = datablksz;
1194 	} else if (BP_IS_REDACTED(bp)) {
1195 		record->sru.redact.datablksz = datablksz;
1196 	} else {
1197 		record->sru.data.datablksz = datablksz;
1198 		record->sru.data.obj_type = dnp->dn_type;
1199 		record->sru.data.bp = *bp;
1200 	}
1201 
1202 	bqueue_enqueue(&sta->q, record, sizeof (*record));
1203 	return (0);
1204 }
1205 
1206 struct redact_list_cb_arg {
1207 	uint64_t *num_blocks_visited;
1208 	bqueue_t *q;
1209 	boolean_t *cancel;
1210 	boolean_t mark_redact;
1211 };
1212 
1213 static int
1214 redact_list_cb(redact_block_phys_t *rb, void *arg)
1215 {
1216 	struct redact_list_cb_arg *rlcap = arg;
1217 
1218 	atomic_inc_64(rlcap->num_blocks_visited);
1219 	if (*rlcap->cancel)
1220 		return (-1);
1221 
1222 	struct send_range *data = range_alloc(REDACT, rb->rbp_object,
1223 	    rb->rbp_blkid, rb->rbp_blkid + redact_block_get_count(rb), B_FALSE);
1224 	ASSERT3U(data->end_blkid, >, rb->rbp_blkid);
1225 	if (rlcap->mark_redact) {
1226 		data->type = REDACT;
1227 		data->sru.redact.datablksz = redact_block_get_size(rb);
1228 	} else {
1229 		data->type = PREVIOUSLY_REDACTED;
1230 	}
1231 	bqueue_enqueue(rlcap->q, data, sizeof (*data));
1232 
1233 	return (0);
1234 }
1235 
1236 /*
1237  * This function kicks off the traverse_dataset.  It also handles setting the
1238  * error code of the thread in case something goes wrong, and pushes the End of
1239  * Stream record when the traverse_dataset call has finished.
1240  */
1241 static __attribute__((noreturn)) void
1242 send_traverse_thread(void *arg)
1243 {
1244 	struct send_thread_arg *st_arg = arg;
1245 	int err = 0;
1246 	struct send_range *data;
1247 	fstrans_cookie_t cookie = spl_fstrans_mark();
1248 
1249 	err = traverse_dataset_resume(st_arg->os->os_dsl_dataset,
1250 	    st_arg->fromtxg, &st_arg->resume,
1251 	    st_arg->flags, send_cb, st_arg);
1252 
1253 	if (err != EINTR)
1254 		st_arg->error_code = err;
1255 	data = range_alloc(DATA, 0, 0, 0, B_TRUE);
1256 	bqueue_enqueue_flush(&st_arg->q, data, sizeof (*data));
1257 	spl_fstrans_unmark(cookie);
1258 	thread_exit();
1259 }
1260 
1261 /*
1262  * Utility function that causes End of Stream records to compare after of all
1263  * others, so that other threads' comparison logic can stay simple.
1264  */
1265 static int __attribute__((unused))
1266 send_range_after(const struct send_range *from, const struct send_range *to)
1267 {
1268 	if (from->eos_marker == B_TRUE)
1269 		return (1);
1270 	if (to->eos_marker == B_TRUE)
1271 		return (-1);
1272 
1273 	uint64_t from_obj = from->object;
1274 	uint64_t from_end_obj = from->object + 1;
1275 	uint64_t to_obj = to->object;
1276 	uint64_t to_end_obj = to->object + 1;
1277 	if (from_obj == 0) {
1278 		ASSERT(from->type == HOLE || from->type == OBJECT_RANGE);
1279 		from_obj = from->start_blkid << DNODES_PER_BLOCK_SHIFT;
1280 		from_end_obj = from->end_blkid << DNODES_PER_BLOCK_SHIFT;
1281 	}
1282 	if (to_obj == 0) {
1283 		ASSERT(to->type == HOLE || to->type == OBJECT_RANGE);
1284 		to_obj = to->start_blkid << DNODES_PER_BLOCK_SHIFT;
1285 		to_end_obj = to->end_blkid << DNODES_PER_BLOCK_SHIFT;
1286 	}
1287 
1288 	if (from_end_obj <= to_obj)
1289 		return (-1);
1290 	if (from_obj >= to_end_obj)
1291 		return (1);
1292 	int64_t cmp = TREE_CMP(to->type == OBJECT_RANGE, from->type ==
1293 	    OBJECT_RANGE);
1294 	if (unlikely(cmp))
1295 		return (cmp);
1296 	cmp = TREE_CMP(to->type == OBJECT, from->type == OBJECT);
1297 	if (unlikely(cmp))
1298 		return (cmp);
1299 	if (from->end_blkid <= to->start_blkid)
1300 		return (-1);
1301 	if (from->start_blkid >= to->end_blkid)
1302 		return (1);
1303 	return (0);
1304 }
1305 
1306 /*
1307  * Pop the new data off the queue, check that the records we receive are in
1308  * the right order, but do not free the old data.  This is used so that the
1309  * records can be sent on to the main thread without copying the data.
1310  */
1311 static struct send_range *
1312 get_next_range_nofree(bqueue_t *bq, struct send_range *prev)
1313 {
1314 	struct send_range *next = bqueue_dequeue(bq);
1315 	ASSERT3S(send_range_after(prev, next), ==, -1);
1316 	return (next);
1317 }
1318 
1319 /*
1320  * Pop the new data off the queue, check that the records we receive are in
1321  * the right order, and free the old data.
1322  */
1323 static struct send_range *
1324 get_next_range(bqueue_t *bq, struct send_range *prev)
1325 {
1326 	struct send_range *next = get_next_range_nofree(bq, prev);
1327 	range_free(prev);
1328 	return (next);
1329 }
1330 
1331 static __attribute__((noreturn)) void
1332 redact_list_thread(void *arg)
1333 {
1334 	struct redact_list_thread_arg *rlt_arg = arg;
1335 	struct send_range *record;
1336 	fstrans_cookie_t cookie = spl_fstrans_mark();
1337 	if (rlt_arg->rl != NULL) {
1338 		struct redact_list_cb_arg rlcba = {0};
1339 		rlcba.cancel = &rlt_arg->cancel;
1340 		rlcba.q = &rlt_arg->q;
1341 		rlcba.num_blocks_visited = rlt_arg->num_blocks_visited;
1342 		rlcba.mark_redact = rlt_arg->mark_redact;
1343 		int err = dsl_redaction_list_traverse(rlt_arg->rl,
1344 		    &rlt_arg->resume, redact_list_cb, &rlcba);
1345 		if (err != EINTR)
1346 			rlt_arg->error_code = err;
1347 	}
1348 	record = range_alloc(DATA, 0, 0, 0, B_TRUE);
1349 	bqueue_enqueue_flush(&rlt_arg->q, record, sizeof (*record));
1350 	spl_fstrans_unmark(cookie);
1351 
1352 	thread_exit();
1353 }
1354 
1355 /*
1356  * Compare the start point of the two provided ranges. End of stream ranges
1357  * compare last, objects compare before any data or hole inside that object and
1358  * multi-object holes that start at the same object.
1359  */
1360 static int
1361 send_range_start_compare(struct send_range *r1, struct send_range *r2)
1362 {
1363 	uint64_t r1_objequiv = r1->object;
1364 	uint64_t r1_l0equiv = r1->start_blkid;
1365 	uint64_t r2_objequiv = r2->object;
1366 	uint64_t r2_l0equiv = r2->start_blkid;
1367 	int64_t cmp = TREE_CMP(r1->eos_marker, r2->eos_marker);
1368 	if (unlikely(cmp))
1369 		return (cmp);
1370 	if (r1->object == 0) {
1371 		r1_objequiv = r1->start_blkid * DNODES_PER_BLOCK;
1372 		r1_l0equiv = 0;
1373 	}
1374 	if (r2->object == 0) {
1375 		r2_objequiv = r2->start_blkid * DNODES_PER_BLOCK;
1376 		r2_l0equiv = 0;
1377 	}
1378 
1379 	cmp = TREE_CMP(r1_objequiv, r2_objequiv);
1380 	if (likely(cmp))
1381 		return (cmp);
1382 	cmp = TREE_CMP(r2->type == OBJECT_RANGE, r1->type == OBJECT_RANGE);
1383 	if (unlikely(cmp))
1384 		return (cmp);
1385 	cmp = TREE_CMP(r2->type == OBJECT, r1->type == OBJECT);
1386 	if (unlikely(cmp))
1387 		return (cmp);
1388 
1389 	return (TREE_CMP(r1_l0equiv, r2_l0equiv));
1390 }
1391 
1392 enum q_idx {
1393 	REDACT_IDX = 0,
1394 	TO_IDX,
1395 	FROM_IDX,
1396 	NUM_THREADS
1397 };
1398 
1399 /*
1400  * This function returns the next range the send_merge_thread should operate on.
1401  * The inputs are two arrays; the first one stores the range at the front of the
1402  * queues stored in the second one.  The ranges are sorted in descending
1403  * priority order; the metadata from earlier ranges overrules metadata from
1404  * later ranges.  out_mask is used to return which threads the ranges came from;
1405  * bit i is set if ranges[i] started at the same place as the returned range.
1406  *
1407  * This code is not hardcoded to compare a specific number of threads; it could
1408  * be used with any number, just by changing the q_idx enum.
1409  *
1410  * The "next range" is the one with the earliest start; if two starts are equal,
1411  * the highest-priority range is the next to operate on.  If a higher-priority
1412  * range starts in the middle of the first range, then the first range will be
1413  * truncated to end where the higher-priority range starts, and we will operate
1414  * on that one next time.   In this way, we make sure that each block covered by
1415  * some range gets covered by a returned range, and each block covered is
1416  * returned using the metadata of the highest-priority range it appears in.
1417  *
1418  * For example, if the three ranges at the front of the queues were [2,4),
1419  * [3,5), and [1,3), then the ranges returned would be [1,2) with the metadata
1420  * from the third range, [2,4) with the metadata from the first range, and then
1421  * [4,5) with the metadata from the second.
1422  */
1423 static struct send_range *
1424 find_next_range(struct send_range **ranges, bqueue_t **qs, uint64_t *out_mask)
1425 {
1426 	int idx = 0; // index of the range with the earliest start
1427 	int i;
1428 	uint64_t bmask = 0;
1429 	for (i = 1; i < NUM_THREADS; i++) {
1430 		if (send_range_start_compare(ranges[i], ranges[idx]) < 0)
1431 			idx = i;
1432 	}
1433 	if (ranges[idx]->eos_marker) {
1434 		struct send_range *ret = range_alloc(DATA, 0, 0, 0, B_TRUE);
1435 		*out_mask = 0;
1436 		return (ret);
1437 	}
1438 	/*
1439 	 * Find all the ranges that start at that same point.
1440 	 */
1441 	for (i = 0; i < NUM_THREADS; i++) {
1442 		if (send_range_start_compare(ranges[i], ranges[idx]) == 0)
1443 			bmask |= 1 << i;
1444 	}
1445 	*out_mask = bmask;
1446 	/*
1447 	 * OBJECT_RANGE records only come from the TO thread, and should always
1448 	 * be treated as overlapping with nothing and sent on immediately.  They
1449 	 * are only used in raw sends, and are never redacted.
1450 	 */
1451 	if (ranges[idx]->type == OBJECT_RANGE) {
1452 		ASSERT3U(idx, ==, TO_IDX);
1453 		ASSERT3U(*out_mask, ==, 1 << TO_IDX);
1454 		struct send_range *ret = ranges[idx];
1455 		ranges[idx] = get_next_range_nofree(qs[idx], ranges[idx]);
1456 		return (ret);
1457 	}
1458 	/*
1459 	 * Find the first start or end point after the start of the first range.
1460 	 */
1461 	uint64_t first_change = ranges[idx]->end_blkid;
1462 	for (i = 0; i < NUM_THREADS; i++) {
1463 		if (i == idx || ranges[i]->eos_marker ||
1464 		    ranges[i]->object > ranges[idx]->object ||
1465 		    ranges[i]->object == DMU_META_DNODE_OBJECT)
1466 			continue;
1467 		ASSERT3U(ranges[i]->object, ==, ranges[idx]->object);
1468 		if (first_change > ranges[i]->start_blkid &&
1469 		    (bmask & (1 << i)) == 0)
1470 			first_change = ranges[i]->start_blkid;
1471 		else if (first_change > ranges[i]->end_blkid)
1472 			first_change = ranges[i]->end_blkid;
1473 	}
1474 	/*
1475 	 * Update all ranges to no longer overlap with the range we're
1476 	 * returning. All such ranges must start at the same place as the range
1477 	 * being returned, and end at or after first_change. Thus we update
1478 	 * their start to first_change. If that makes them size 0, then free
1479 	 * them and pull a new range from that thread.
1480 	 */
1481 	for (i = 0; i < NUM_THREADS; i++) {
1482 		if (i == idx || (bmask & (1 << i)) == 0)
1483 			continue;
1484 		ASSERT3U(first_change, >, ranges[i]->start_blkid);
1485 		ranges[i]->start_blkid = first_change;
1486 		ASSERT3U(ranges[i]->start_blkid, <=, ranges[i]->end_blkid);
1487 		if (ranges[i]->start_blkid == ranges[i]->end_blkid)
1488 			ranges[i] = get_next_range(qs[i], ranges[i]);
1489 	}
1490 	/*
1491 	 * Short-circuit the simple case; if the range doesn't overlap with
1492 	 * anything else, or it only overlaps with things that start at the same
1493 	 * place and are longer, send it on.
1494 	 */
1495 	if (first_change == ranges[idx]->end_blkid) {
1496 		struct send_range *ret = ranges[idx];
1497 		ranges[idx] = get_next_range_nofree(qs[idx], ranges[idx]);
1498 		return (ret);
1499 	}
1500 
1501 	/*
1502 	 * Otherwise, return a truncated copy of ranges[idx] and move the start
1503 	 * of ranges[idx] back to first_change.
1504 	 */
1505 	struct send_range *ret = kmem_alloc(sizeof (*ret), KM_SLEEP);
1506 	*ret = *ranges[idx];
1507 	ret->end_blkid = first_change;
1508 	ranges[idx]->start_blkid = first_change;
1509 	return (ret);
1510 }
1511 
1512 #define	FROM_AND_REDACT_BITS ((1 << REDACT_IDX) | (1 << FROM_IDX))
1513 
1514 /*
1515  * Merge the results from the from thread and the to thread, and then hand the
1516  * records off to send_prefetch_thread to prefetch them.  If this is not a
1517  * send from a redaction bookmark, the from thread will push an end of stream
1518  * record and stop, and we'll just send everything that was changed in the
1519  * to_ds since the ancestor's creation txg. If it is, then since
1520  * traverse_dataset has a canonical order, we can compare each change as
1521  * they're pulled off the queues.  That will give us a stream that is
1522  * appropriately sorted, and covers all records.  In addition, we pull the
1523  * data from the redact_list_thread and use that to determine which blocks
1524  * should be redacted.
1525  */
1526 static __attribute__((noreturn)) void
1527 send_merge_thread(void *arg)
1528 {
1529 	struct send_merge_thread_arg *smt_arg = arg;
1530 	struct send_range *front_ranges[NUM_THREADS];
1531 	bqueue_t *queues[NUM_THREADS];
1532 	int err = 0;
1533 	fstrans_cookie_t cookie = spl_fstrans_mark();
1534 
1535 	if (smt_arg->redact_arg == NULL) {
1536 		front_ranges[REDACT_IDX] =
1537 		    kmem_zalloc(sizeof (struct send_range), KM_SLEEP);
1538 		front_ranges[REDACT_IDX]->eos_marker = B_TRUE;
1539 		front_ranges[REDACT_IDX]->type = REDACT;
1540 		queues[REDACT_IDX] = NULL;
1541 	} else {
1542 		front_ranges[REDACT_IDX] =
1543 		    bqueue_dequeue(&smt_arg->redact_arg->q);
1544 		queues[REDACT_IDX] = &smt_arg->redact_arg->q;
1545 	}
1546 	front_ranges[TO_IDX] = bqueue_dequeue(&smt_arg->to_arg->q);
1547 	queues[TO_IDX] = &smt_arg->to_arg->q;
1548 	front_ranges[FROM_IDX] = bqueue_dequeue(&smt_arg->from_arg->q);
1549 	queues[FROM_IDX] = &smt_arg->from_arg->q;
1550 	uint64_t mask = 0;
1551 	struct send_range *range;
1552 	for (range = find_next_range(front_ranges, queues, &mask);
1553 	    !range->eos_marker && err == 0 && !smt_arg->cancel;
1554 	    range = find_next_range(front_ranges, queues, &mask)) {
1555 		/*
1556 		 * If the range in question was in both the from redact bookmark
1557 		 * and the bookmark we're using to redact, then don't send it.
1558 		 * It's already redacted on the receiving system, so a redaction
1559 		 * record would be redundant.
1560 		 */
1561 		if ((mask & FROM_AND_REDACT_BITS) == FROM_AND_REDACT_BITS) {
1562 			ASSERT3U(range->type, ==, REDACT);
1563 			range_free(range);
1564 			continue;
1565 		}
1566 		bqueue_enqueue(&smt_arg->q, range, sizeof (*range));
1567 
1568 		if (smt_arg->to_arg->error_code != 0) {
1569 			err = smt_arg->to_arg->error_code;
1570 		} else if (smt_arg->from_arg->error_code != 0) {
1571 			err = smt_arg->from_arg->error_code;
1572 		} else if (smt_arg->redact_arg != NULL &&
1573 		    smt_arg->redact_arg->error_code != 0) {
1574 			err = smt_arg->redact_arg->error_code;
1575 		}
1576 	}
1577 	if (smt_arg->cancel && err == 0)
1578 		err = SET_ERROR(EINTR);
1579 	smt_arg->error = err;
1580 	if (smt_arg->error != 0) {
1581 		smt_arg->to_arg->cancel = B_TRUE;
1582 		smt_arg->from_arg->cancel = B_TRUE;
1583 		if (smt_arg->redact_arg != NULL)
1584 			smt_arg->redact_arg->cancel = B_TRUE;
1585 	}
1586 	for (int i = 0; i < NUM_THREADS; i++) {
1587 		while (!front_ranges[i]->eos_marker) {
1588 			front_ranges[i] = get_next_range(queues[i],
1589 			    front_ranges[i]);
1590 		}
1591 		range_free(front_ranges[i]);
1592 	}
1593 	if (range == NULL)
1594 		range = kmem_zalloc(sizeof (*range), KM_SLEEP);
1595 	range->eos_marker = B_TRUE;
1596 	bqueue_enqueue_flush(&smt_arg->q, range, 1);
1597 	spl_fstrans_unmark(cookie);
1598 	thread_exit();
1599 }
1600 
1601 struct send_reader_thread_arg {
1602 	struct send_merge_thread_arg *smta;
1603 	bqueue_t q;
1604 	boolean_t cancel;
1605 	boolean_t issue_reads;
1606 	uint64_t featureflags;
1607 	int error;
1608 };
1609 
1610 static void
1611 dmu_send_read_done(zio_t *zio)
1612 {
1613 	struct send_range *range = zio->io_private;
1614 
1615 	mutex_enter(&range->sru.data.lock);
1616 	if (zio->io_error != 0) {
1617 		abd_free(range->sru.data.abd);
1618 		range->sru.data.abd = NULL;
1619 		range->sru.data.io_err = zio->io_error;
1620 	}
1621 
1622 	ASSERT(range->sru.data.io_outstanding);
1623 	range->sru.data.io_outstanding = B_FALSE;
1624 	cv_broadcast(&range->sru.data.cv);
1625 	mutex_exit(&range->sru.data.lock);
1626 }
1627 
1628 static void
1629 issue_data_read(struct send_reader_thread_arg *srta, struct send_range *range)
1630 {
1631 	struct srd *srdp = &range->sru.data;
1632 	blkptr_t *bp = &srdp->bp;
1633 	objset_t *os = srta->smta->os;
1634 
1635 	ASSERT3U(range->type, ==, DATA);
1636 	ASSERT3U(range->start_blkid + 1, ==, range->end_blkid);
1637 	/*
1638 	 * If we have large blocks stored on disk but
1639 	 * the send flags don't allow us to send large
1640 	 * blocks, we split the data from the arc buf
1641 	 * into chunks.
1642 	 */
1643 	boolean_t split_large_blocks =
1644 	    srdp->datablksz > SPA_OLD_MAXBLOCKSIZE &&
1645 	    !(srta->featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS);
1646 	/*
1647 	 * We should only request compressed data from the ARC if all
1648 	 * the following are true:
1649 	 *  - stream compression was requested
1650 	 *  - we aren't splitting large blocks into smaller chunks
1651 	 *  - the data won't need to be byteswapped before sending
1652 	 *  - this isn't an embedded block
1653 	 *  - this isn't metadata (if receiving on a different endian
1654 	 *    system it can be byteswapped more easily)
1655 	 */
1656 	boolean_t request_compressed =
1657 	    (srta->featureflags & DMU_BACKUP_FEATURE_COMPRESSED) &&
1658 	    !split_large_blocks && !BP_SHOULD_BYTESWAP(bp) &&
1659 	    !BP_IS_EMBEDDED(bp) && !DMU_OT_IS_METADATA(BP_GET_TYPE(bp));
1660 
1661 	enum zio_flag zioflags = ZIO_FLAG_CANFAIL;
1662 
1663 	if (srta->featureflags & DMU_BACKUP_FEATURE_RAW) {
1664 		zioflags |= ZIO_FLAG_RAW;
1665 		srdp->io_compressed = B_TRUE;
1666 	} else if (request_compressed) {
1667 		zioflags |= ZIO_FLAG_RAW_COMPRESS;
1668 		srdp->io_compressed = B_TRUE;
1669 	}
1670 
1671 	srdp->datasz = (zioflags & ZIO_FLAG_RAW_COMPRESS) ?
1672 	    BP_GET_PSIZE(bp) : BP_GET_LSIZE(bp);
1673 
1674 	if (!srta->issue_reads)
1675 		return;
1676 	if (BP_IS_REDACTED(bp))
1677 		return;
1678 	if (send_do_embed(bp, srta->featureflags))
1679 		return;
1680 
1681 	zbookmark_phys_t zb = {
1682 	    .zb_objset = dmu_objset_id(os),
1683 	    .zb_object = range->object,
1684 	    .zb_level = 0,
1685 	    .zb_blkid = range->start_blkid,
1686 	};
1687 
1688 	arc_flags_t aflags = ARC_FLAG_CACHED_ONLY;
1689 
1690 	int arc_err = arc_read(NULL, os->os_spa, bp,
1691 	    arc_getbuf_func, &srdp->abuf, ZIO_PRIORITY_ASYNC_READ,
1692 	    zioflags, &aflags, &zb);
1693 	/*
1694 	 * If the data is not already cached in the ARC, we read directly
1695 	 * from zio.  This avoids the performance overhead of adding a new
1696 	 * entry to the ARC, and we also avoid polluting the ARC cache with
1697 	 * data that is not likely to be used in the future.
1698 	 */
1699 	if (arc_err != 0) {
1700 		srdp->abd = abd_alloc_linear(srdp->datasz, B_FALSE);
1701 		srdp->io_outstanding = B_TRUE;
1702 		zio_nowait(zio_read(NULL, os->os_spa, bp, srdp->abd,
1703 		    srdp->datasz, dmu_send_read_done, range,
1704 		    ZIO_PRIORITY_ASYNC_READ, zioflags, &zb));
1705 	}
1706 }
1707 
1708 /*
1709  * Create a new record with the given values.
1710  */
1711 static void
1712 enqueue_range(struct send_reader_thread_arg *srta, bqueue_t *q, dnode_t *dn,
1713     uint64_t blkid, uint64_t count, const blkptr_t *bp, uint32_t datablksz)
1714 {
1715 	enum type range_type = (bp == NULL || BP_IS_HOLE(bp) ? HOLE :
1716 	    (BP_IS_REDACTED(bp) ? REDACT : DATA));
1717 
1718 	struct send_range *range = range_alloc(range_type, dn->dn_object,
1719 	    blkid, blkid + count, B_FALSE);
1720 
1721 	if (blkid == DMU_SPILL_BLKID)
1722 		ASSERT3U(BP_GET_TYPE(bp), ==, DMU_OT_SA);
1723 
1724 	switch (range_type) {
1725 	case HOLE:
1726 		range->sru.hole.datablksz = datablksz;
1727 		break;
1728 	case DATA:
1729 		ASSERT3U(count, ==, 1);
1730 		range->sru.data.datablksz = datablksz;
1731 		range->sru.data.obj_type = dn->dn_type;
1732 		range->sru.data.bp = *bp;
1733 		issue_data_read(srta, range);
1734 		break;
1735 	case REDACT:
1736 		range->sru.redact.datablksz = datablksz;
1737 		break;
1738 	default:
1739 		break;
1740 	}
1741 	bqueue_enqueue(q, range, datablksz);
1742 }
1743 
1744 /*
1745  * This thread is responsible for two things: First, it retrieves the correct
1746  * blkptr in the to ds if we need to send the data because of something from
1747  * the from thread.  As a result of this, we're the first ones to discover that
1748  * some indirect blocks can be discarded because they're not holes. Second,
1749  * it issues prefetches for the data we need to send.
1750  */
1751 static __attribute__((noreturn)) void
1752 send_reader_thread(void *arg)
1753 {
1754 	struct send_reader_thread_arg *srta = arg;
1755 	struct send_merge_thread_arg *smta = srta->smta;
1756 	bqueue_t *inq = &smta->q;
1757 	bqueue_t *outq = &srta->q;
1758 	objset_t *os = smta->os;
1759 	fstrans_cookie_t cookie = spl_fstrans_mark();
1760 	struct send_range *range = bqueue_dequeue(inq);
1761 	int err = 0;
1762 
1763 	/*
1764 	 * If the record we're analyzing is from a redaction bookmark from the
1765 	 * fromds, then we need to know whether or not it exists in the tods so
1766 	 * we know whether to create records for it or not. If it does, we need
1767 	 * the datablksz so we can generate an appropriate record for it.
1768 	 * Finally, if it isn't redacted, we need the blkptr so that we can send
1769 	 * a WRITE record containing the actual data.
1770 	 */
1771 	uint64_t last_obj = UINT64_MAX;
1772 	uint64_t last_obj_exists = B_TRUE;
1773 	while (!range->eos_marker && !srta->cancel && smta->error == 0 &&
1774 	    err == 0) {
1775 		switch (range->type) {
1776 		case DATA:
1777 			issue_data_read(srta, range);
1778 			bqueue_enqueue(outq, range, range->sru.data.datablksz);
1779 			range = get_next_range_nofree(inq, range);
1780 			break;
1781 		case HOLE:
1782 		case OBJECT:
1783 		case OBJECT_RANGE:
1784 		case REDACT: // Redacted blocks must exist
1785 			bqueue_enqueue(outq, range, sizeof (*range));
1786 			range = get_next_range_nofree(inq, range);
1787 			break;
1788 		case PREVIOUSLY_REDACTED: {
1789 			/*
1790 			 * This entry came from the "from bookmark" when
1791 			 * sending from a bookmark that has a redaction
1792 			 * list.  We need to check if this object/blkid
1793 			 * exists in the target ("to") dataset, and if
1794 			 * not then we drop this entry.  We also need
1795 			 * to fill in the block pointer so that we know
1796 			 * what to prefetch.
1797 			 *
1798 			 * To accomplish the above, we first cache whether or
1799 			 * not the last object we examined exists.  If it
1800 			 * doesn't, we can drop this record. If it does, we hold
1801 			 * the dnode and use it to call dbuf_dnode_findbp. We do
1802 			 * this instead of dbuf_bookmark_findbp because we will
1803 			 * often operate on large ranges, and holding the dnode
1804 			 * once is more efficient.
1805 			 */
1806 			boolean_t object_exists = B_TRUE;
1807 			/*
1808 			 * If the data is redacted, we only care if it exists,
1809 			 * so that we don't send records for objects that have
1810 			 * been deleted.
1811 			 */
1812 			dnode_t *dn;
1813 			if (range->object == last_obj && !last_obj_exists) {
1814 				/*
1815 				 * If we're still examining the same object as
1816 				 * previously, and it doesn't exist, we don't
1817 				 * need to call dbuf_bookmark_findbp.
1818 				 */
1819 				object_exists = B_FALSE;
1820 			} else {
1821 				err = dnode_hold(os, range->object, FTAG, &dn);
1822 				if (err == ENOENT) {
1823 					object_exists = B_FALSE;
1824 					err = 0;
1825 				}
1826 				last_obj = range->object;
1827 				last_obj_exists = object_exists;
1828 			}
1829 
1830 			if (err != 0) {
1831 				break;
1832 			} else if (!object_exists) {
1833 				/*
1834 				 * The block was modified, but doesn't
1835 				 * exist in the to dataset; if it was
1836 				 * deleted in the to dataset, then we'll
1837 				 * visit the hole bp for it at some point.
1838 				 */
1839 				range = get_next_range(inq, range);
1840 				continue;
1841 			}
1842 			uint64_t file_max =
1843 			    (dn->dn_maxblkid < range->end_blkid ?
1844 			    dn->dn_maxblkid : range->end_blkid);
1845 			/*
1846 			 * The object exists, so we need to try to find the
1847 			 * blkptr for each block in the range we're processing.
1848 			 */
1849 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
1850 			for (uint64_t blkid = range->start_blkid;
1851 			    blkid < file_max; blkid++) {
1852 				blkptr_t bp;
1853 				uint32_t datablksz =
1854 				    dn->dn_phys->dn_datablkszsec <<
1855 				    SPA_MINBLOCKSHIFT;
1856 				uint64_t offset = blkid * datablksz;
1857 				/*
1858 				 * This call finds the next non-hole block in
1859 				 * the object. This is to prevent a
1860 				 * performance problem where we're unredacting
1861 				 * a large hole. Using dnode_next_offset to
1862 				 * skip over the large hole avoids iterating
1863 				 * over every block in it.
1864 				 */
1865 				err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
1866 				    &offset, 1, 1, 0);
1867 				if (err == ESRCH) {
1868 					offset = UINT64_MAX;
1869 					err = 0;
1870 				} else if (err != 0) {
1871 					break;
1872 				}
1873 				if (offset != blkid * datablksz) {
1874 					/*
1875 					 * if there is a hole from here
1876 					 * (blkid) to offset
1877 					 */
1878 					offset = MIN(offset, file_max *
1879 					    datablksz);
1880 					uint64_t nblks = (offset / datablksz) -
1881 					    blkid;
1882 					enqueue_range(srta, outq, dn, blkid,
1883 					    nblks, NULL, datablksz);
1884 					blkid += nblks;
1885 				}
1886 				if (blkid >= file_max)
1887 					break;
1888 				err = dbuf_dnode_findbp(dn, 0, blkid, &bp,
1889 				    NULL, NULL);
1890 				if (err != 0)
1891 					break;
1892 				ASSERT(!BP_IS_HOLE(&bp));
1893 				enqueue_range(srta, outq, dn, blkid, 1, &bp,
1894 				    datablksz);
1895 			}
1896 			rw_exit(&dn->dn_struct_rwlock);
1897 			dnode_rele(dn, FTAG);
1898 			range = get_next_range(inq, range);
1899 		}
1900 		}
1901 	}
1902 	if (srta->cancel || err != 0) {
1903 		smta->cancel = B_TRUE;
1904 		srta->error = err;
1905 	} else if (smta->error != 0) {
1906 		srta->error = smta->error;
1907 	}
1908 	while (!range->eos_marker)
1909 		range = get_next_range(inq, range);
1910 
1911 	bqueue_enqueue_flush(outq, range, 1);
1912 	spl_fstrans_unmark(cookie);
1913 	thread_exit();
1914 }
1915 
1916 #define	NUM_SNAPS_NOT_REDACTED UINT64_MAX
1917 
1918 struct dmu_send_params {
1919 	/* Pool args */
1920 	const void *tag; // Tag dp was held with, will be used to release dp.
1921 	dsl_pool_t *dp;
1922 	/* To snapshot args */
1923 	const char *tosnap;
1924 	dsl_dataset_t *to_ds;
1925 	/* From snapshot args */
1926 	zfs_bookmark_phys_t ancestor_zb;
1927 	uint64_t *fromredactsnaps;
1928 	/* NUM_SNAPS_NOT_REDACTED if not sending from redaction bookmark */
1929 	uint64_t numfromredactsnaps;
1930 	/* Stream params */
1931 	boolean_t is_clone;
1932 	boolean_t embedok;
1933 	boolean_t large_block_ok;
1934 	boolean_t compressok;
1935 	boolean_t rawok;
1936 	boolean_t savedok;
1937 	uint64_t resumeobj;
1938 	uint64_t resumeoff;
1939 	uint64_t saved_guid;
1940 	zfs_bookmark_phys_t *redactbook;
1941 	/* Stream output params */
1942 	dmu_send_outparams_t *dso;
1943 
1944 	/* Stream progress params */
1945 	offset_t *off;
1946 	int outfd;
1947 	char saved_toname[MAXNAMELEN];
1948 };
1949 
1950 static int
1951 setup_featureflags(struct dmu_send_params *dspp, objset_t *os,
1952     uint64_t *featureflags)
1953 {
1954 	dsl_dataset_t *to_ds = dspp->to_ds;
1955 	dsl_pool_t *dp = dspp->dp;
1956 #ifdef _KERNEL
1957 	if (dmu_objset_type(os) == DMU_OST_ZFS) {
1958 		uint64_t version;
1959 		if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0)
1960 			return (SET_ERROR(EINVAL));
1961 
1962 		if (version >= ZPL_VERSION_SA)
1963 			*featureflags |= DMU_BACKUP_FEATURE_SA_SPILL;
1964 	}
1965 #endif
1966 
1967 	/* raw sends imply large_block_ok */
1968 	if ((dspp->rawok || dspp->large_block_ok) &&
1969 	    dsl_dataset_feature_is_active(to_ds, SPA_FEATURE_LARGE_BLOCKS)) {
1970 		*featureflags |= DMU_BACKUP_FEATURE_LARGE_BLOCKS;
1971 	}
1972 
1973 	/* encrypted datasets will not have embedded blocks */
1974 	if ((dspp->embedok || dspp->rawok) && !os->os_encrypted &&
1975 	    spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) {
1976 		*featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA;
1977 	}
1978 
1979 	/* raw send implies compressok */
1980 	if (dspp->compressok || dspp->rawok)
1981 		*featureflags |= DMU_BACKUP_FEATURE_COMPRESSED;
1982 
1983 	if (dspp->rawok && os->os_encrypted)
1984 		*featureflags |= DMU_BACKUP_FEATURE_RAW;
1985 
1986 	if ((*featureflags &
1987 	    (DMU_BACKUP_FEATURE_EMBED_DATA | DMU_BACKUP_FEATURE_COMPRESSED |
1988 	    DMU_BACKUP_FEATURE_RAW)) != 0 &&
1989 	    spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS)) {
1990 		*featureflags |= DMU_BACKUP_FEATURE_LZ4;
1991 	}
1992 
1993 	/*
1994 	 * We specifically do not include DMU_BACKUP_FEATURE_EMBED_DATA here to
1995 	 * allow sending ZSTD compressed datasets to a receiver that does not
1996 	 * support ZSTD
1997 	 */
1998 	if ((*featureflags &
1999 	    (DMU_BACKUP_FEATURE_COMPRESSED | DMU_BACKUP_FEATURE_RAW)) != 0 &&
2000 	    dsl_dataset_feature_is_active(to_ds, SPA_FEATURE_ZSTD_COMPRESS)) {
2001 		*featureflags |= DMU_BACKUP_FEATURE_ZSTD;
2002 	}
2003 
2004 	if (dspp->resumeobj != 0 || dspp->resumeoff != 0) {
2005 		*featureflags |= DMU_BACKUP_FEATURE_RESUMING;
2006 	}
2007 
2008 	if (dspp->redactbook != NULL) {
2009 		*featureflags |= DMU_BACKUP_FEATURE_REDACTED;
2010 	}
2011 
2012 	if (dsl_dataset_feature_is_active(to_ds, SPA_FEATURE_LARGE_DNODE)) {
2013 		*featureflags |= DMU_BACKUP_FEATURE_LARGE_DNODE;
2014 	}
2015 	return (0);
2016 }
2017 
2018 static dmu_replay_record_t *
2019 create_begin_record(struct dmu_send_params *dspp, objset_t *os,
2020     uint64_t featureflags)
2021 {
2022 	dmu_replay_record_t *drr = kmem_zalloc(sizeof (dmu_replay_record_t),
2023 	    KM_SLEEP);
2024 	drr->drr_type = DRR_BEGIN;
2025 
2026 	struct drr_begin *drrb = &drr->drr_u.drr_begin;
2027 	dsl_dataset_t *to_ds = dspp->to_ds;
2028 
2029 	drrb->drr_magic = DMU_BACKUP_MAGIC;
2030 	drrb->drr_creation_time = dsl_dataset_phys(to_ds)->ds_creation_time;
2031 	drrb->drr_type = dmu_objset_type(os);
2032 	drrb->drr_toguid = dsl_dataset_phys(to_ds)->ds_guid;
2033 	drrb->drr_fromguid = dspp->ancestor_zb.zbm_guid;
2034 
2035 	DMU_SET_STREAM_HDRTYPE(drrb->drr_versioninfo, DMU_SUBSTREAM);
2036 	DMU_SET_FEATUREFLAGS(drrb->drr_versioninfo, featureflags);
2037 
2038 	if (dspp->is_clone)
2039 		drrb->drr_flags |= DRR_FLAG_CLONE;
2040 	if (dsl_dataset_phys(dspp->to_ds)->ds_flags & DS_FLAG_CI_DATASET)
2041 		drrb->drr_flags |= DRR_FLAG_CI_DATA;
2042 	if (zfs_send_set_freerecords_bit)
2043 		drrb->drr_flags |= DRR_FLAG_FREERECORDS;
2044 	drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_SPILL_BLOCK;
2045 
2046 	if (dspp->savedok) {
2047 		drrb->drr_toguid = dspp->saved_guid;
2048 		strlcpy(drrb->drr_toname, dspp->saved_toname,
2049 		    sizeof (drrb->drr_toname));
2050 	} else {
2051 		dsl_dataset_name(to_ds, drrb->drr_toname);
2052 		if (!to_ds->ds_is_snapshot) {
2053 			(void) strlcat(drrb->drr_toname, "@--head--",
2054 			    sizeof (drrb->drr_toname));
2055 		}
2056 	}
2057 	return (drr);
2058 }
2059 
2060 static void
2061 setup_to_thread(struct send_thread_arg *to_arg, objset_t *to_os,
2062     dmu_sendstatus_t *dssp, uint64_t fromtxg, boolean_t rawok)
2063 {
2064 	VERIFY0(bqueue_init(&to_arg->q, zfs_send_no_prefetch_queue_ff,
2065 	    MAX(zfs_send_no_prefetch_queue_length, 2 * zfs_max_recordsize),
2066 	    offsetof(struct send_range, ln)));
2067 	to_arg->error_code = 0;
2068 	to_arg->cancel = B_FALSE;
2069 	to_arg->os = to_os;
2070 	to_arg->fromtxg = fromtxg;
2071 	to_arg->flags = TRAVERSE_PRE | TRAVERSE_PREFETCH_METADATA;
2072 	if (rawok)
2073 		to_arg->flags |= TRAVERSE_NO_DECRYPT;
2074 	if (zfs_send_corrupt_data)
2075 		to_arg->flags |= TRAVERSE_HARD;
2076 	to_arg->num_blocks_visited = &dssp->dss_blocks;
2077 	(void) thread_create(NULL, 0, send_traverse_thread, to_arg, 0,
2078 	    curproc, TS_RUN, minclsyspri);
2079 }
2080 
2081 static void
2082 setup_from_thread(struct redact_list_thread_arg *from_arg,
2083     redaction_list_t *from_rl, dmu_sendstatus_t *dssp)
2084 {
2085 	VERIFY0(bqueue_init(&from_arg->q, zfs_send_no_prefetch_queue_ff,
2086 	    MAX(zfs_send_no_prefetch_queue_length, 2 * zfs_max_recordsize),
2087 	    offsetof(struct send_range, ln)));
2088 	from_arg->error_code = 0;
2089 	from_arg->cancel = B_FALSE;
2090 	from_arg->rl = from_rl;
2091 	from_arg->mark_redact = B_FALSE;
2092 	from_arg->num_blocks_visited = &dssp->dss_blocks;
2093 	/*
2094 	 * If from_ds is null, send_traverse_thread just returns success and
2095 	 * enqueues an eos marker.
2096 	 */
2097 	(void) thread_create(NULL, 0, redact_list_thread, from_arg, 0,
2098 	    curproc, TS_RUN, minclsyspri);
2099 }
2100 
2101 static void
2102 setup_redact_list_thread(struct redact_list_thread_arg *rlt_arg,
2103     struct dmu_send_params *dspp, redaction_list_t *rl, dmu_sendstatus_t *dssp)
2104 {
2105 	if (dspp->redactbook == NULL)
2106 		return;
2107 
2108 	rlt_arg->cancel = B_FALSE;
2109 	VERIFY0(bqueue_init(&rlt_arg->q, zfs_send_no_prefetch_queue_ff,
2110 	    MAX(zfs_send_no_prefetch_queue_length, 2 * zfs_max_recordsize),
2111 	    offsetof(struct send_range, ln)));
2112 	rlt_arg->error_code = 0;
2113 	rlt_arg->mark_redact = B_TRUE;
2114 	rlt_arg->rl = rl;
2115 	rlt_arg->num_blocks_visited = &dssp->dss_blocks;
2116 
2117 	(void) thread_create(NULL, 0, redact_list_thread, rlt_arg, 0,
2118 	    curproc, TS_RUN, minclsyspri);
2119 }
2120 
2121 static void
2122 setup_merge_thread(struct send_merge_thread_arg *smt_arg,
2123     struct dmu_send_params *dspp, struct redact_list_thread_arg *from_arg,
2124     struct send_thread_arg *to_arg, struct redact_list_thread_arg *rlt_arg,
2125     objset_t *os)
2126 {
2127 	VERIFY0(bqueue_init(&smt_arg->q, zfs_send_no_prefetch_queue_ff,
2128 	    MAX(zfs_send_no_prefetch_queue_length, 2 * zfs_max_recordsize),
2129 	    offsetof(struct send_range, ln)));
2130 	smt_arg->cancel = B_FALSE;
2131 	smt_arg->error = 0;
2132 	smt_arg->from_arg = from_arg;
2133 	smt_arg->to_arg = to_arg;
2134 	if (dspp->redactbook != NULL)
2135 		smt_arg->redact_arg = rlt_arg;
2136 
2137 	smt_arg->os = os;
2138 	(void) thread_create(NULL, 0, send_merge_thread, smt_arg, 0, curproc,
2139 	    TS_RUN, minclsyspri);
2140 }
2141 
2142 static void
2143 setup_reader_thread(struct send_reader_thread_arg *srt_arg,
2144     struct dmu_send_params *dspp, struct send_merge_thread_arg *smt_arg,
2145     uint64_t featureflags)
2146 {
2147 	VERIFY0(bqueue_init(&srt_arg->q, zfs_send_queue_ff,
2148 	    MAX(zfs_send_queue_length, 2 * zfs_max_recordsize),
2149 	    offsetof(struct send_range, ln)));
2150 	srt_arg->smta = smt_arg;
2151 	srt_arg->issue_reads = !dspp->dso->dso_dryrun;
2152 	srt_arg->featureflags = featureflags;
2153 	(void) thread_create(NULL, 0, send_reader_thread, srt_arg, 0,
2154 	    curproc, TS_RUN, minclsyspri);
2155 }
2156 
2157 static int
2158 setup_resume_points(struct dmu_send_params *dspp,
2159     struct send_thread_arg *to_arg, struct redact_list_thread_arg *from_arg,
2160     struct redact_list_thread_arg *rlt_arg,
2161     struct send_merge_thread_arg *smt_arg, boolean_t resuming, objset_t *os,
2162     redaction_list_t *redact_rl, nvlist_t *nvl)
2163 {
2164 	(void) smt_arg;
2165 	dsl_dataset_t *to_ds = dspp->to_ds;
2166 	int err = 0;
2167 
2168 	uint64_t obj = 0;
2169 	uint64_t blkid = 0;
2170 	if (resuming) {
2171 		obj = dspp->resumeobj;
2172 		dmu_object_info_t to_doi;
2173 		err = dmu_object_info(os, obj, &to_doi);
2174 		if (err != 0)
2175 			return (err);
2176 
2177 		blkid = dspp->resumeoff / to_doi.doi_data_block_size;
2178 	}
2179 	/*
2180 	 * If we're resuming a redacted send, we can skip to the appropriate
2181 	 * point in the redaction bookmark by binary searching through it.
2182 	 */
2183 	if (redact_rl != NULL) {
2184 		SET_BOOKMARK(&rlt_arg->resume, to_ds->ds_object, obj, 0, blkid);
2185 	}
2186 
2187 	SET_BOOKMARK(&to_arg->resume, to_ds->ds_object, obj, 0, blkid);
2188 	if (nvlist_exists(nvl, BEGINNV_REDACT_FROM_SNAPS)) {
2189 		uint64_t objset = dspp->ancestor_zb.zbm_redaction_obj;
2190 		/*
2191 		 * Note: If the resume point is in an object whose
2192 		 * blocksize is different in the from vs to snapshots,
2193 		 * we will have divided by the "wrong" blocksize.
2194 		 * However, in this case fromsnap's send_cb() will
2195 		 * detect that the blocksize has changed and therefore
2196 		 * ignore this object.
2197 		 *
2198 		 * If we're resuming a send from a redaction bookmark,
2199 		 * we still cannot accidentally suggest blocks behind
2200 		 * the to_ds.  In addition, we know that any blocks in
2201 		 * the object in the to_ds will have to be sent, since
2202 		 * the size changed.  Therefore, we can't cause any harm
2203 		 * this way either.
2204 		 */
2205 		SET_BOOKMARK(&from_arg->resume, objset, obj, 0, blkid);
2206 	}
2207 	if (resuming) {
2208 		fnvlist_add_uint64(nvl, BEGINNV_RESUME_OBJECT, dspp->resumeobj);
2209 		fnvlist_add_uint64(nvl, BEGINNV_RESUME_OFFSET, dspp->resumeoff);
2210 	}
2211 	return (0);
2212 }
2213 
2214 static dmu_sendstatus_t *
2215 setup_send_progress(struct dmu_send_params *dspp)
2216 {
2217 	dmu_sendstatus_t *dssp = kmem_zalloc(sizeof (*dssp), KM_SLEEP);
2218 	dssp->dss_outfd = dspp->outfd;
2219 	dssp->dss_off = dspp->off;
2220 	dssp->dss_proc = curproc;
2221 	mutex_enter(&dspp->to_ds->ds_sendstream_lock);
2222 	list_insert_head(&dspp->to_ds->ds_sendstreams, dssp);
2223 	mutex_exit(&dspp->to_ds->ds_sendstream_lock);
2224 	return (dssp);
2225 }
2226 
2227 /*
2228  * Actually do the bulk of the work in a zfs send.
2229  *
2230  * The idea is that we want to do a send from ancestor_zb to to_ds.  We also
2231  * want to not send any data that has been modified by all the datasets in
2232  * redactsnaparr, and store the list of blocks that are redacted in this way in
2233  * a bookmark named redactbook, created on the to_ds.  We do this by creating
2234  * several worker threads, whose function is described below.
2235  *
2236  * There are three cases.
2237  * The first case is a redacted zfs send.  In this case there are 5 threads.
2238  * The first thread is the to_ds traversal thread: it calls dataset_traverse on
2239  * the to_ds and finds all the blocks that have changed since ancestor_zb (if
2240  * it's a full send, that's all blocks in the dataset).  It then sends those
2241  * blocks on to the send merge thread. The redact list thread takes the data
2242  * from the redaction bookmark and sends those blocks on to the send merge
2243  * thread.  The send merge thread takes the data from the to_ds traversal
2244  * thread, and combines it with the redaction records from the redact list
2245  * thread.  If a block appears in both the to_ds's data and the redaction data,
2246  * the send merge thread will mark it as redacted and send it on to the prefetch
2247  * thread.  Otherwise, the send merge thread will send the block on to the
2248  * prefetch thread unchanged. The prefetch thread will issue prefetch reads for
2249  * any data that isn't redacted, and then send the data on to the main thread.
2250  * The main thread behaves the same as in a normal send case, issuing demand
2251  * reads for data blocks and sending out records over the network
2252  *
2253  * The graphic below diagrams the flow of data in the case of a redacted zfs
2254  * send.  Each box represents a thread, and each line represents the flow of
2255  * data.
2256  *
2257  *             Records from the |
2258  *           redaction bookmark |
2259  * +--------------------+       |  +---------------------------+
2260  * |                    |       v  | Send Merge Thread         |
2261  * | Redact List Thread +----------> Apply redaction marks to  |
2262  * |                    |          | records as specified by   |
2263  * +--------------------+          | redaction ranges          |
2264  *                                 +----^---------------+------+
2265  *                                      |               | Merged data
2266  *                                      |               |
2267  *                                      |  +------------v--------+
2268  *                                      |  | Prefetch Thread     |
2269  * +--------------------+               |  | Issues prefetch     |
2270  * | to_ds Traversal    |               |  | reads of data blocks|
2271  * | Thread (finds      +---------------+  +------------+--------+
2272  * | candidate blocks)  |  Blocks modified              | Prefetched data
2273  * +--------------------+  by to_ds since               |
2274  *                         ancestor_zb     +------------v----+
2275  *                                         | Main Thread     |  File Descriptor
2276  *                                         | Sends data over +->(to zfs receive)
2277  *                                         | wire            |
2278  *                                         +-----------------+
2279  *
2280  * The second case is an incremental send from a redaction bookmark.  The to_ds
2281  * traversal thread and the main thread behave the same as in the redacted
2282  * send case.  The new thread is the from bookmark traversal thread.  It
2283  * iterates over the redaction list in the redaction bookmark, and enqueues
2284  * records for each block that was redacted in the original send.  The send
2285  * merge thread now has to merge the data from the two threads.  For details
2286  * about that process, see the header comment of send_merge_thread().  Any data
2287  * it decides to send on will be prefetched by the prefetch thread.  Note that
2288  * you can perform a redacted send from a redaction bookmark; in that case,
2289  * the data flow behaves very similarly to the flow in the redacted send case,
2290  * except with the addition of the bookmark traversal thread iterating over the
2291  * redaction bookmark.  The send_merge_thread also has to take on the
2292  * responsibility of merging the redact list thread's records, the bookmark
2293  * traversal thread's records, and the to_ds records.
2294  *
2295  * +---------------------+
2296  * |                     |
2297  * | Redact List Thread  +--------------+
2298  * |                     |              |
2299  * +---------------------+              |
2300  *        Blocks in redaction list      | Ranges modified by every secure snap
2301  *        of from bookmark              | (or EOS if not readcted)
2302  *                                      |
2303  * +---------------------+   |     +----v----------------------+
2304  * | bookmark Traversal  |   v     | Send Merge Thread         |
2305  * | Thread (finds       +---------> Merges bookmark, rlt, and |
2306  * | candidate blocks)   |         | to_ds send records        |
2307  * +---------------------+         +----^---------------+------+
2308  *                                      |               | Merged data
2309  *                                      |  +------------v--------+
2310  *                                      |  | Prefetch Thread     |
2311  * +--------------------+               |  | Issues prefetch     |
2312  * | to_ds Traversal    |               |  | reads of data blocks|
2313  * | Thread (finds      +---------------+  +------------+--------+
2314  * | candidate blocks)  |  Blocks modified              | Prefetched data
2315  * +--------------------+  by to_ds since  +------------v----+
2316  *                         ancestor_zb     | Main Thread     |  File Descriptor
2317  *                                         | Sends data over +->(to zfs receive)
2318  *                                         | wire            |
2319  *                                         +-----------------+
2320  *
2321  * The final case is a simple zfs full or incremental send.  The to_ds traversal
2322  * thread behaves the same as always. The redact list thread is never started.
2323  * The send merge thread takes all the blocks that the to_ds traversal thread
2324  * sends it, prefetches the data, and sends the blocks on to the main thread.
2325  * The main thread sends the data over the wire.
2326  *
2327  * To keep performance acceptable, we want to prefetch the data in the worker
2328  * threads.  While the to_ds thread could simply use the TRAVERSE_PREFETCH
2329  * feature built into traverse_dataset, the combining and deletion of records
2330  * due to redaction and sends from redaction bookmarks mean that we could
2331  * issue many unnecessary prefetches.  As a result, we only prefetch data
2332  * after we've determined that the record is not going to be redacted.  To
2333  * prevent the prefetching from getting too far ahead of the main thread, the
2334  * blocking queues that are used for communication are capped not by the
2335  * number of entries in the queue, but by the sum of the size of the
2336  * prefetches associated with them.  The limit on the amount of data that the
2337  * thread can prefetch beyond what the main thread has reached is controlled
2338  * by the global variable zfs_send_queue_length.  In addition, to prevent poor
2339  * performance in the beginning of a send, we also limit the distance ahead
2340  * that the traversal threads can be.  That distance is controlled by the
2341  * zfs_send_no_prefetch_queue_length tunable.
2342  *
2343  * Note: Releases dp using the specified tag.
2344  */
2345 static int
2346 dmu_send_impl(struct dmu_send_params *dspp)
2347 {
2348 	objset_t *os;
2349 	dmu_replay_record_t *drr;
2350 	dmu_sendstatus_t *dssp;
2351 	dmu_send_cookie_t dsc = {0};
2352 	int err;
2353 	uint64_t fromtxg = dspp->ancestor_zb.zbm_creation_txg;
2354 	uint64_t featureflags = 0;
2355 	struct redact_list_thread_arg *from_arg;
2356 	struct send_thread_arg *to_arg;
2357 	struct redact_list_thread_arg *rlt_arg;
2358 	struct send_merge_thread_arg *smt_arg;
2359 	struct send_reader_thread_arg *srt_arg;
2360 	struct send_range *range;
2361 	redaction_list_t *from_rl = NULL;
2362 	redaction_list_t *redact_rl = NULL;
2363 	boolean_t resuming = (dspp->resumeobj != 0 || dspp->resumeoff != 0);
2364 	boolean_t book_resuming = resuming;
2365 
2366 	dsl_dataset_t *to_ds = dspp->to_ds;
2367 	zfs_bookmark_phys_t *ancestor_zb = &dspp->ancestor_zb;
2368 	dsl_pool_t *dp = dspp->dp;
2369 	const void *tag = dspp->tag;
2370 
2371 	err = dmu_objset_from_ds(to_ds, &os);
2372 	if (err != 0) {
2373 		dsl_pool_rele(dp, tag);
2374 		return (err);
2375 	}
2376 
2377 	/*
2378 	 * If this is a non-raw send of an encrypted ds, we can ensure that
2379 	 * the objset_phys_t is authenticated. This is safe because this is
2380 	 * either a snapshot or we have owned the dataset, ensuring that
2381 	 * it can't be modified.
2382 	 */
2383 	if (!dspp->rawok && os->os_encrypted &&
2384 	    arc_is_unauthenticated(os->os_phys_buf)) {
2385 		zbookmark_phys_t zb;
2386 
2387 		SET_BOOKMARK(&zb, to_ds->ds_object, ZB_ROOT_OBJECT,
2388 		    ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
2389 		err = arc_untransform(os->os_phys_buf, os->os_spa,
2390 		    &zb, B_FALSE);
2391 		if (err != 0) {
2392 			dsl_pool_rele(dp, tag);
2393 			return (err);
2394 		}
2395 
2396 		ASSERT0(arc_is_unauthenticated(os->os_phys_buf));
2397 	}
2398 
2399 	if ((err = setup_featureflags(dspp, os, &featureflags)) != 0) {
2400 		dsl_pool_rele(dp, tag);
2401 		return (err);
2402 	}
2403 
2404 	/*
2405 	 * If we're doing a redacted send, hold the bookmark's redaction list.
2406 	 */
2407 	if (dspp->redactbook != NULL) {
2408 		err = dsl_redaction_list_hold_obj(dp,
2409 		    dspp->redactbook->zbm_redaction_obj, FTAG,
2410 		    &redact_rl);
2411 		if (err != 0) {
2412 			dsl_pool_rele(dp, tag);
2413 			return (SET_ERROR(EINVAL));
2414 		}
2415 		dsl_redaction_list_long_hold(dp, redact_rl, FTAG);
2416 	}
2417 
2418 	/*
2419 	 * If we're sending from a redaction bookmark, hold the redaction list
2420 	 * so that we can consider sending the redacted blocks.
2421 	 */
2422 	if (ancestor_zb->zbm_redaction_obj != 0) {
2423 		err = dsl_redaction_list_hold_obj(dp,
2424 		    ancestor_zb->zbm_redaction_obj, FTAG, &from_rl);
2425 		if (err != 0) {
2426 			if (redact_rl != NULL) {
2427 				dsl_redaction_list_long_rele(redact_rl, FTAG);
2428 				dsl_redaction_list_rele(redact_rl, FTAG);
2429 			}
2430 			dsl_pool_rele(dp, tag);
2431 			return (SET_ERROR(EINVAL));
2432 		}
2433 		dsl_redaction_list_long_hold(dp, from_rl, FTAG);
2434 	}
2435 
2436 	dsl_dataset_long_hold(to_ds, FTAG);
2437 
2438 	from_arg = kmem_zalloc(sizeof (*from_arg), KM_SLEEP);
2439 	to_arg = kmem_zalloc(sizeof (*to_arg), KM_SLEEP);
2440 	rlt_arg = kmem_zalloc(sizeof (*rlt_arg), KM_SLEEP);
2441 	smt_arg = kmem_zalloc(sizeof (*smt_arg), KM_SLEEP);
2442 	srt_arg = kmem_zalloc(sizeof (*srt_arg), KM_SLEEP);
2443 
2444 	drr = create_begin_record(dspp, os, featureflags);
2445 	dssp = setup_send_progress(dspp);
2446 
2447 	dsc.dsc_drr = drr;
2448 	dsc.dsc_dso = dspp->dso;
2449 	dsc.dsc_os = os;
2450 	dsc.dsc_off = dspp->off;
2451 	dsc.dsc_toguid = dsl_dataset_phys(to_ds)->ds_guid;
2452 	dsc.dsc_fromtxg = fromtxg;
2453 	dsc.dsc_pending_op = PENDING_NONE;
2454 	dsc.dsc_featureflags = featureflags;
2455 	dsc.dsc_resume_object = dspp->resumeobj;
2456 	dsc.dsc_resume_offset = dspp->resumeoff;
2457 
2458 	dsl_pool_rele(dp, tag);
2459 
2460 	void *payload = NULL;
2461 	size_t payload_len = 0;
2462 	nvlist_t *nvl = fnvlist_alloc();
2463 
2464 	/*
2465 	 * If we're doing a redacted send, we include the snapshots we're
2466 	 * redacted with respect to so that the target system knows what send
2467 	 * streams can be correctly received on top of this dataset. If we're
2468 	 * instead sending a redacted dataset, we include the snapshots that the
2469 	 * dataset was created with respect to.
2470 	 */
2471 	if (dspp->redactbook != NULL) {
2472 		fnvlist_add_uint64_array(nvl, BEGINNV_REDACT_SNAPS,
2473 		    redact_rl->rl_phys->rlp_snaps,
2474 		    redact_rl->rl_phys->rlp_num_snaps);
2475 	} else if (dsl_dataset_feature_is_active(to_ds,
2476 	    SPA_FEATURE_REDACTED_DATASETS)) {
2477 		uint64_t *tods_guids;
2478 		uint64_t length;
2479 		VERIFY(dsl_dataset_get_uint64_array_feature(to_ds,
2480 		    SPA_FEATURE_REDACTED_DATASETS, &length, &tods_guids));
2481 		fnvlist_add_uint64_array(nvl, BEGINNV_REDACT_SNAPS, tods_guids,
2482 		    length);
2483 	}
2484 
2485 	/*
2486 	 * If we're sending from a redaction bookmark, then we should retrieve
2487 	 * the guids of that bookmark so we can send them over the wire.
2488 	 */
2489 	if (from_rl != NULL) {
2490 		fnvlist_add_uint64_array(nvl, BEGINNV_REDACT_FROM_SNAPS,
2491 		    from_rl->rl_phys->rlp_snaps,
2492 		    from_rl->rl_phys->rlp_num_snaps);
2493 	}
2494 
2495 	/*
2496 	 * If the snapshot we're sending from is redacted, include the redaction
2497 	 * list in the stream.
2498 	 */
2499 	if (dspp->numfromredactsnaps != NUM_SNAPS_NOT_REDACTED) {
2500 		ASSERT3P(from_rl, ==, NULL);
2501 		fnvlist_add_uint64_array(nvl, BEGINNV_REDACT_FROM_SNAPS,
2502 		    dspp->fromredactsnaps, (uint_t)dspp->numfromredactsnaps);
2503 		if (dspp->numfromredactsnaps > 0) {
2504 			kmem_free(dspp->fromredactsnaps,
2505 			    dspp->numfromredactsnaps * sizeof (uint64_t));
2506 			dspp->fromredactsnaps = NULL;
2507 		}
2508 	}
2509 
2510 	if (resuming || book_resuming) {
2511 		err = setup_resume_points(dspp, to_arg, from_arg,
2512 		    rlt_arg, smt_arg, resuming, os, redact_rl, nvl);
2513 		if (err != 0)
2514 			goto out;
2515 	}
2516 
2517 	if (featureflags & DMU_BACKUP_FEATURE_RAW) {
2518 		uint64_t ivset_guid = (ancestor_zb != NULL) ?
2519 		    ancestor_zb->zbm_ivset_guid : 0;
2520 		nvlist_t *keynvl = NULL;
2521 		ASSERT(os->os_encrypted);
2522 
2523 		err = dsl_crypto_populate_key_nvlist(os, ivset_guid,
2524 		    &keynvl);
2525 		if (err != 0) {
2526 			fnvlist_free(nvl);
2527 			goto out;
2528 		}
2529 
2530 		fnvlist_add_nvlist(nvl, "crypt_keydata", keynvl);
2531 		fnvlist_free(keynvl);
2532 	}
2533 
2534 	if (!nvlist_empty(nvl)) {
2535 		payload = fnvlist_pack(nvl, &payload_len);
2536 		drr->drr_payloadlen = payload_len;
2537 	}
2538 
2539 	fnvlist_free(nvl);
2540 	err = dump_record(&dsc, payload, payload_len);
2541 	fnvlist_pack_free(payload, payload_len);
2542 	if (err != 0) {
2543 		err = dsc.dsc_err;
2544 		goto out;
2545 	}
2546 
2547 	setup_to_thread(to_arg, os, dssp, fromtxg, dspp->rawok);
2548 	setup_from_thread(from_arg, from_rl, dssp);
2549 	setup_redact_list_thread(rlt_arg, dspp, redact_rl, dssp);
2550 	setup_merge_thread(smt_arg, dspp, from_arg, to_arg, rlt_arg, os);
2551 	setup_reader_thread(srt_arg, dspp, smt_arg, featureflags);
2552 
2553 	range = bqueue_dequeue(&srt_arg->q);
2554 	while (err == 0 && !range->eos_marker) {
2555 		err = do_dump(&dsc, range);
2556 		range = get_next_range(&srt_arg->q, range);
2557 		if (issig(JUSTLOOKING) && issig(FORREAL))
2558 			err = SET_ERROR(EINTR);
2559 	}
2560 
2561 	/*
2562 	 * If we hit an error or are interrupted, cancel our worker threads and
2563 	 * clear the queue of any pending records.  The threads will pass the
2564 	 * cancel up the tree of worker threads, and each one will clean up any
2565 	 * pending records before exiting.
2566 	 */
2567 	if (err != 0) {
2568 		srt_arg->cancel = B_TRUE;
2569 		while (!range->eos_marker) {
2570 			range = get_next_range(&srt_arg->q, range);
2571 		}
2572 	}
2573 	range_free(range);
2574 
2575 	bqueue_destroy(&srt_arg->q);
2576 	bqueue_destroy(&smt_arg->q);
2577 	if (dspp->redactbook != NULL)
2578 		bqueue_destroy(&rlt_arg->q);
2579 	bqueue_destroy(&to_arg->q);
2580 	bqueue_destroy(&from_arg->q);
2581 
2582 	if (err == 0 && srt_arg->error != 0)
2583 		err = srt_arg->error;
2584 
2585 	if (err != 0)
2586 		goto out;
2587 
2588 	if (dsc.dsc_pending_op != PENDING_NONE)
2589 		if (dump_record(&dsc, NULL, 0) != 0)
2590 			err = SET_ERROR(EINTR);
2591 
2592 	if (err != 0) {
2593 		if (err == EINTR && dsc.dsc_err != 0)
2594 			err = dsc.dsc_err;
2595 		goto out;
2596 	}
2597 
2598 	/*
2599 	 * Send the DRR_END record if this is not a saved stream.
2600 	 * Otherwise, the omitted DRR_END record will signal to
2601 	 * the receive side that the stream is incomplete.
2602 	 */
2603 	if (!dspp->savedok) {
2604 		memset(drr, 0, sizeof (dmu_replay_record_t));
2605 		drr->drr_type = DRR_END;
2606 		drr->drr_u.drr_end.drr_checksum = dsc.dsc_zc;
2607 		drr->drr_u.drr_end.drr_toguid = dsc.dsc_toguid;
2608 
2609 		if (dump_record(&dsc, NULL, 0) != 0)
2610 			err = dsc.dsc_err;
2611 	}
2612 out:
2613 	mutex_enter(&to_ds->ds_sendstream_lock);
2614 	list_remove(&to_ds->ds_sendstreams, dssp);
2615 	mutex_exit(&to_ds->ds_sendstream_lock);
2616 
2617 	VERIFY(err != 0 || (dsc.dsc_sent_begin &&
2618 	    (dsc.dsc_sent_end || dspp->savedok)));
2619 
2620 	kmem_free(drr, sizeof (dmu_replay_record_t));
2621 	kmem_free(dssp, sizeof (dmu_sendstatus_t));
2622 	kmem_free(from_arg, sizeof (*from_arg));
2623 	kmem_free(to_arg, sizeof (*to_arg));
2624 	kmem_free(rlt_arg, sizeof (*rlt_arg));
2625 	kmem_free(smt_arg, sizeof (*smt_arg));
2626 	kmem_free(srt_arg, sizeof (*srt_arg));
2627 
2628 	dsl_dataset_long_rele(to_ds, FTAG);
2629 	if (from_rl != NULL) {
2630 		dsl_redaction_list_long_rele(from_rl, FTAG);
2631 		dsl_redaction_list_rele(from_rl, FTAG);
2632 	}
2633 	if (redact_rl != NULL) {
2634 		dsl_redaction_list_long_rele(redact_rl, FTAG);
2635 		dsl_redaction_list_rele(redact_rl, FTAG);
2636 	}
2637 
2638 	return (err);
2639 }
2640 
2641 int
2642 dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
2643     boolean_t embedok, boolean_t large_block_ok, boolean_t compressok,
2644     boolean_t rawok, boolean_t savedok, int outfd, offset_t *off,
2645     dmu_send_outparams_t *dsop)
2646 {
2647 	int err;
2648 	dsl_dataset_t *fromds;
2649 	ds_hold_flags_t dsflags;
2650 	struct dmu_send_params dspp = {0};
2651 	dspp.embedok = embedok;
2652 	dspp.large_block_ok = large_block_ok;
2653 	dspp.compressok = compressok;
2654 	dspp.outfd = outfd;
2655 	dspp.off = off;
2656 	dspp.dso = dsop;
2657 	dspp.tag = FTAG;
2658 	dspp.rawok = rawok;
2659 	dspp.savedok = savedok;
2660 
2661 	dsflags = (rawok) ? DS_HOLD_FLAG_NONE : DS_HOLD_FLAG_DECRYPT;
2662 	err = dsl_pool_hold(pool, FTAG, &dspp.dp);
2663 	if (err != 0)
2664 		return (err);
2665 
2666 	err = dsl_dataset_hold_obj_flags(dspp.dp, tosnap, dsflags, FTAG,
2667 	    &dspp.to_ds);
2668 	if (err != 0) {
2669 		dsl_pool_rele(dspp.dp, FTAG);
2670 		return (err);
2671 	}
2672 
2673 	if (fromsnap != 0) {
2674 		err = dsl_dataset_hold_obj_flags(dspp.dp, fromsnap, dsflags,
2675 		    FTAG, &fromds);
2676 		if (err != 0) {
2677 			dsl_dataset_rele_flags(dspp.to_ds, dsflags, FTAG);
2678 			dsl_pool_rele(dspp.dp, FTAG);
2679 			return (err);
2680 		}
2681 		dspp.ancestor_zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
2682 		dspp.ancestor_zb.zbm_creation_txg =
2683 		    dsl_dataset_phys(fromds)->ds_creation_txg;
2684 		dspp.ancestor_zb.zbm_creation_time =
2685 		    dsl_dataset_phys(fromds)->ds_creation_time;
2686 
2687 		if (dsl_dataset_is_zapified(fromds)) {
2688 			(void) zap_lookup(dspp.dp->dp_meta_objset,
2689 			    fromds->ds_object, DS_FIELD_IVSET_GUID, 8, 1,
2690 			    &dspp.ancestor_zb.zbm_ivset_guid);
2691 		}
2692 
2693 		/* See dmu_send for the reasons behind this. */
2694 		uint64_t *fromredact;
2695 
2696 		if (!dsl_dataset_get_uint64_array_feature(fromds,
2697 		    SPA_FEATURE_REDACTED_DATASETS,
2698 		    &dspp.numfromredactsnaps,
2699 		    &fromredact)) {
2700 			dspp.numfromredactsnaps = NUM_SNAPS_NOT_REDACTED;
2701 		} else if (dspp.numfromredactsnaps > 0) {
2702 			uint64_t size = dspp.numfromredactsnaps *
2703 			    sizeof (uint64_t);
2704 			dspp.fromredactsnaps = kmem_zalloc(size, KM_SLEEP);
2705 			memcpy(dspp.fromredactsnaps, fromredact, size);
2706 		}
2707 
2708 		boolean_t is_before =
2709 		    dsl_dataset_is_before(dspp.to_ds, fromds, 0);
2710 		dspp.is_clone = (dspp.to_ds->ds_dir !=
2711 		    fromds->ds_dir);
2712 		dsl_dataset_rele(fromds, FTAG);
2713 		if (!is_before) {
2714 			dsl_pool_rele(dspp.dp, FTAG);
2715 			err = SET_ERROR(EXDEV);
2716 		} else {
2717 			err = dmu_send_impl(&dspp);
2718 		}
2719 	} else {
2720 		dspp.numfromredactsnaps = NUM_SNAPS_NOT_REDACTED;
2721 		err = dmu_send_impl(&dspp);
2722 	}
2723 	dsl_dataset_rele(dspp.to_ds, FTAG);
2724 	return (err);
2725 }
2726 
2727 int
2728 dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok,
2729     boolean_t large_block_ok, boolean_t compressok, boolean_t rawok,
2730     boolean_t savedok, uint64_t resumeobj, uint64_t resumeoff,
2731     const char *redactbook, int outfd, offset_t *off,
2732     dmu_send_outparams_t *dsop)
2733 {
2734 	int err = 0;
2735 	ds_hold_flags_t dsflags;
2736 	boolean_t owned = B_FALSE;
2737 	dsl_dataset_t *fromds = NULL;
2738 	zfs_bookmark_phys_t book = {0};
2739 	struct dmu_send_params dspp = {0};
2740 
2741 	dsflags = (rawok) ? DS_HOLD_FLAG_NONE : DS_HOLD_FLAG_DECRYPT;
2742 	dspp.tosnap = tosnap;
2743 	dspp.embedok = embedok;
2744 	dspp.large_block_ok = large_block_ok;
2745 	dspp.compressok = compressok;
2746 	dspp.outfd = outfd;
2747 	dspp.off = off;
2748 	dspp.dso = dsop;
2749 	dspp.tag = FTAG;
2750 	dspp.resumeobj = resumeobj;
2751 	dspp.resumeoff = resumeoff;
2752 	dspp.rawok = rawok;
2753 	dspp.savedok = savedok;
2754 
2755 	if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL)
2756 		return (SET_ERROR(EINVAL));
2757 
2758 	err = dsl_pool_hold(tosnap, FTAG, &dspp.dp);
2759 	if (err != 0)
2760 		return (err);
2761 
2762 	if (strchr(tosnap, '@') == NULL && spa_writeable(dspp.dp->dp_spa)) {
2763 		/*
2764 		 * We are sending a filesystem or volume.  Ensure
2765 		 * that it doesn't change by owning the dataset.
2766 		 */
2767 
2768 		if (savedok) {
2769 			/*
2770 			 * We are looking for the dataset that represents the
2771 			 * partially received send stream. If this stream was
2772 			 * received as a new snapshot of an existing dataset,
2773 			 * this will be saved in a hidden clone named
2774 			 * "<pool>/<dataset>/%recv". Otherwise, the stream
2775 			 * will be saved in the live dataset itself. In
2776 			 * either case we need to use dsl_dataset_own_force()
2777 			 * because the stream is marked as inconsistent,
2778 			 * which would normally make it unavailable to be
2779 			 * owned.
2780 			 */
2781 			char *name = kmem_asprintf("%s/%s", tosnap,
2782 			    recv_clone_name);
2783 			err = dsl_dataset_own_force(dspp.dp, name, dsflags,
2784 			    FTAG, &dspp.to_ds);
2785 			if (err == ENOENT) {
2786 				err = dsl_dataset_own_force(dspp.dp, tosnap,
2787 				    dsflags, FTAG, &dspp.to_ds);
2788 			}
2789 
2790 			if (err == 0) {
2791 				err = zap_lookup(dspp.dp->dp_meta_objset,
2792 				    dspp.to_ds->ds_object,
2793 				    DS_FIELD_RESUME_TOGUID, 8, 1,
2794 				    &dspp.saved_guid);
2795 			}
2796 
2797 			if (err == 0) {
2798 				err = zap_lookup(dspp.dp->dp_meta_objset,
2799 				    dspp.to_ds->ds_object,
2800 				    DS_FIELD_RESUME_TONAME, 1,
2801 				    sizeof (dspp.saved_toname),
2802 				    dspp.saved_toname);
2803 			}
2804 			if (err != 0)
2805 				dsl_dataset_disown(dspp.to_ds, dsflags, FTAG);
2806 
2807 			kmem_strfree(name);
2808 		} else {
2809 			err = dsl_dataset_own(dspp.dp, tosnap, dsflags,
2810 			    FTAG, &dspp.to_ds);
2811 		}
2812 		owned = B_TRUE;
2813 	} else {
2814 		err = dsl_dataset_hold_flags(dspp.dp, tosnap, dsflags, FTAG,
2815 		    &dspp.to_ds);
2816 	}
2817 
2818 	if (err != 0) {
2819 		dsl_pool_rele(dspp.dp, FTAG);
2820 		return (err);
2821 	}
2822 
2823 	if (redactbook != NULL) {
2824 		char path[ZFS_MAX_DATASET_NAME_LEN];
2825 		(void) strlcpy(path, tosnap, sizeof (path));
2826 		char *at = strchr(path, '@');
2827 		if (at == NULL) {
2828 			err = EINVAL;
2829 		} else {
2830 			(void) snprintf(at, sizeof (path) - (at - path), "#%s",
2831 			    redactbook);
2832 			err = dsl_bookmark_lookup(dspp.dp, path,
2833 			    NULL, &book);
2834 			dspp.redactbook = &book;
2835 		}
2836 	}
2837 
2838 	if (err != 0) {
2839 		dsl_pool_rele(dspp.dp, FTAG);
2840 		if (owned)
2841 			dsl_dataset_disown(dspp.to_ds, dsflags, FTAG);
2842 		else
2843 			dsl_dataset_rele_flags(dspp.to_ds, dsflags, FTAG);
2844 		return (err);
2845 	}
2846 
2847 	if (fromsnap != NULL) {
2848 		zfs_bookmark_phys_t *zb = &dspp.ancestor_zb;
2849 		int fsnamelen;
2850 		if (strpbrk(tosnap, "@#") != NULL)
2851 			fsnamelen = strpbrk(tosnap, "@#") - tosnap;
2852 		else
2853 			fsnamelen = strlen(tosnap);
2854 
2855 		/*
2856 		 * If the fromsnap is in a different filesystem, then
2857 		 * mark the send stream as a clone.
2858 		 */
2859 		if (strncmp(tosnap, fromsnap, fsnamelen) != 0 ||
2860 		    (fromsnap[fsnamelen] != '@' &&
2861 		    fromsnap[fsnamelen] != '#')) {
2862 			dspp.is_clone = B_TRUE;
2863 		}
2864 
2865 		if (strchr(fromsnap, '@') != NULL) {
2866 			err = dsl_dataset_hold(dspp.dp, fromsnap, FTAG,
2867 			    &fromds);
2868 
2869 			if (err != 0) {
2870 				ASSERT3P(fromds, ==, NULL);
2871 			} else {
2872 				/*
2873 				 * We need to make a deep copy of the redact
2874 				 * snapshots of the from snapshot, because the
2875 				 * array will be freed when we evict from_ds.
2876 				 */
2877 				uint64_t *fromredact;
2878 				if (!dsl_dataset_get_uint64_array_feature(
2879 				    fromds, SPA_FEATURE_REDACTED_DATASETS,
2880 				    &dspp.numfromredactsnaps,
2881 				    &fromredact)) {
2882 					dspp.numfromredactsnaps =
2883 					    NUM_SNAPS_NOT_REDACTED;
2884 				} else if (dspp.numfromredactsnaps > 0) {
2885 					uint64_t size =
2886 					    dspp.numfromredactsnaps *
2887 					    sizeof (uint64_t);
2888 					dspp.fromredactsnaps = kmem_zalloc(size,
2889 					    KM_SLEEP);
2890 					memcpy(dspp.fromredactsnaps, fromredact,
2891 					    size);
2892 				}
2893 				if (!dsl_dataset_is_before(dspp.to_ds, fromds,
2894 				    0)) {
2895 					err = SET_ERROR(EXDEV);
2896 				} else {
2897 					zb->zbm_creation_txg =
2898 					    dsl_dataset_phys(fromds)->
2899 					    ds_creation_txg;
2900 					zb->zbm_creation_time =
2901 					    dsl_dataset_phys(fromds)->
2902 					    ds_creation_time;
2903 					zb->zbm_guid =
2904 					    dsl_dataset_phys(fromds)->ds_guid;
2905 					zb->zbm_redaction_obj = 0;
2906 
2907 					if (dsl_dataset_is_zapified(fromds)) {
2908 						(void) zap_lookup(
2909 						    dspp.dp->dp_meta_objset,
2910 						    fromds->ds_object,
2911 						    DS_FIELD_IVSET_GUID, 8, 1,
2912 						    &zb->zbm_ivset_guid);
2913 					}
2914 				}
2915 				dsl_dataset_rele(fromds, FTAG);
2916 			}
2917 		} else {
2918 			dspp.numfromredactsnaps = NUM_SNAPS_NOT_REDACTED;
2919 			err = dsl_bookmark_lookup(dspp.dp, fromsnap, dspp.to_ds,
2920 			    zb);
2921 			if (err == EXDEV && zb->zbm_redaction_obj != 0 &&
2922 			    zb->zbm_guid ==
2923 			    dsl_dataset_phys(dspp.to_ds)->ds_guid)
2924 				err = 0;
2925 		}
2926 
2927 		if (err == 0) {
2928 			/* dmu_send_impl will call dsl_pool_rele for us. */
2929 			err = dmu_send_impl(&dspp);
2930 		} else {
2931 			dsl_pool_rele(dspp.dp, FTAG);
2932 		}
2933 	} else {
2934 		dspp.numfromredactsnaps = NUM_SNAPS_NOT_REDACTED;
2935 		err = dmu_send_impl(&dspp);
2936 	}
2937 	if (owned)
2938 		dsl_dataset_disown(dspp.to_ds, dsflags, FTAG);
2939 	else
2940 		dsl_dataset_rele_flags(dspp.to_ds, dsflags, FTAG);
2941 	return (err);
2942 }
2943 
2944 static int
2945 dmu_adjust_send_estimate_for_indirects(dsl_dataset_t *ds, uint64_t uncompressed,
2946     uint64_t compressed, boolean_t stream_compressed, uint64_t *sizep)
2947 {
2948 	int err = 0;
2949 	uint64_t size;
2950 	/*
2951 	 * Assume that space (both on-disk and in-stream) is dominated by
2952 	 * data.  We will adjust for indirect blocks and the copies property,
2953 	 * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
2954 	 */
2955 
2956 	uint64_t recordsize;
2957 	uint64_t record_count;
2958 	objset_t *os;
2959 	VERIFY0(dmu_objset_from_ds(ds, &os));
2960 
2961 	/* Assume all (uncompressed) blocks are recordsize. */
2962 	if (zfs_override_estimate_recordsize != 0) {
2963 		recordsize = zfs_override_estimate_recordsize;
2964 	} else if (os->os_phys->os_type == DMU_OST_ZVOL) {
2965 		err = dsl_prop_get_int_ds(ds,
2966 		    zfs_prop_to_name(ZFS_PROP_VOLBLOCKSIZE), &recordsize);
2967 	} else {
2968 		err = dsl_prop_get_int_ds(ds,
2969 		    zfs_prop_to_name(ZFS_PROP_RECORDSIZE), &recordsize);
2970 	}
2971 	if (err != 0)
2972 		return (err);
2973 	record_count = uncompressed / recordsize;
2974 
2975 	/*
2976 	 * If we're estimating a send size for a compressed stream, use the
2977 	 * compressed data size to estimate the stream size. Otherwise, use the
2978 	 * uncompressed data size.
2979 	 */
2980 	size = stream_compressed ? compressed : uncompressed;
2981 
2982 	/*
2983 	 * Subtract out approximate space used by indirect blocks.
2984 	 * Assume most space is used by data blocks (non-indirect, non-dnode).
2985 	 * Assume no ditto blocks or internal fragmentation.
2986 	 *
2987 	 * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
2988 	 * block.
2989 	 */
2990 	size -= record_count * sizeof (blkptr_t);
2991 
2992 	/* Add in the space for the record associated with each block. */
2993 	size += record_count * sizeof (dmu_replay_record_t);
2994 
2995 	*sizep = size;
2996 
2997 	return (0);
2998 }
2999 
3000 int
3001 dmu_send_estimate_fast(dsl_dataset_t *origds, dsl_dataset_t *fromds,
3002     zfs_bookmark_phys_t *frombook, boolean_t stream_compressed,
3003     boolean_t saved, uint64_t *sizep)
3004 {
3005 	int err;
3006 	dsl_dataset_t *ds = origds;
3007 	uint64_t uncomp, comp;
3008 
3009 	ASSERT(dsl_pool_config_held(origds->ds_dir->dd_pool));
3010 	ASSERT(fromds == NULL || frombook == NULL);
3011 
3012 	/*
3013 	 * If this is a saved send we may actually be sending
3014 	 * from the %recv clone used for resuming.
3015 	 */
3016 	if (saved) {
3017 		objset_t *mos = origds->ds_dir->dd_pool->dp_meta_objset;
3018 		uint64_t guid;
3019 		char dsname[ZFS_MAX_DATASET_NAME_LEN + 6];
3020 
3021 		dsl_dataset_name(origds, dsname);
3022 		(void) strcat(dsname, "/");
3023 		(void) strcat(dsname, recv_clone_name);
3024 
3025 		err = dsl_dataset_hold(origds->ds_dir->dd_pool,
3026 		    dsname, FTAG, &ds);
3027 		if (err != ENOENT && err != 0) {
3028 			return (err);
3029 		} else if (err == ENOENT) {
3030 			ds = origds;
3031 		}
3032 
3033 		/* check that this dataset has partially received data */
3034 		err = zap_lookup(mos, ds->ds_object,
3035 		    DS_FIELD_RESUME_TOGUID, 8, 1, &guid);
3036 		if (err != 0) {
3037 			err = SET_ERROR(err == ENOENT ? EINVAL : err);
3038 			goto out;
3039 		}
3040 
3041 		err = zap_lookup(mos, ds->ds_object,
3042 		    DS_FIELD_RESUME_TONAME, 1, sizeof (dsname), dsname);
3043 		if (err != 0) {
3044 			err = SET_ERROR(err == ENOENT ? EINVAL : err);
3045 			goto out;
3046 		}
3047 	}
3048 
3049 	/* tosnap must be a snapshot or the target of a saved send */
3050 	if (!ds->ds_is_snapshot && ds == origds)
3051 		return (SET_ERROR(EINVAL));
3052 
3053 	if (fromds != NULL) {
3054 		uint64_t used;
3055 		if (!fromds->ds_is_snapshot) {
3056 			err = SET_ERROR(EINVAL);
3057 			goto out;
3058 		}
3059 
3060 		if (!dsl_dataset_is_before(ds, fromds, 0)) {
3061 			err = SET_ERROR(EXDEV);
3062 			goto out;
3063 		}
3064 
3065 		err = dsl_dataset_space_written(fromds, ds, &used, &comp,
3066 		    &uncomp);
3067 		if (err != 0)
3068 			goto out;
3069 	} else if (frombook != NULL) {
3070 		uint64_t used;
3071 		err = dsl_dataset_space_written_bookmark(frombook, ds, &used,
3072 		    &comp, &uncomp);
3073 		if (err != 0)
3074 			goto out;
3075 	} else {
3076 		uncomp = dsl_dataset_phys(ds)->ds_uncompressed_bytes;
3077 		comp = dsl_dataset_phys(ds)->ds_compressed_bytes;
3078 	}
3079 
3080 	err = dmu_adjust_send_estimate_for_indirects(ds, uncomp, comp,
3081 	    stream_compressed, sizep);
3082 	/*
3083 	 * Add the size of the BEGIN and END records to the estimate.
3084 	 */
3085 	*sizep += 2 * sizeof (dmu_replay_record_t);
3086 
3087 out:
3088 	if (ds != origds)
3089 		dsl_dataset_rele(ds, FTAG);
3090 	return (err);
3091 }
3092 
3093 ZFS_MODULE_PARAM(zfs_send, zfs_send_, corrupt_data, INT, ZMOD_RW,
3094 	"Allow sending corrupt data");
3095 
3096 ZFS_MODULE_PARAM(zfs_send, zfs_send_, queue_length, INT, ZMOD_RW,
3097 	"Maximum send queue length");
3098 
3099 ZFS_MODULE_PARAM(zfs_send, zfs_send_, unmodified_spill_blocks, INT, ZMOD_RW,
3100 	"Send unmodified spill blocks");
3101 
3102 ZFS_MODULE_PARAM(zfs_send, zfs_send_, no_prefetch_queue_length, INT, ZMOD_RW,
3103 	"Maximum send queue length for non-prefetch queues");
3104 
3105 ZFS_MODULE_PARAM(zfs_send, zfs_send_, queue_ff, INT, ZMOD_RW,
3106 	"Send queue fill fraction");
3107 
3108 ZFS_MODULE_PARAM(zfs_send, zfs_send_, no_prefetch_queue_ff, INT, ZMOD_RW,
3109 	"Send queue fill fraction for non-prefetch queues");
3110 
3111 ZFS_MODULE_PARAM(zfs_send, zfs_, override_estimate_recordsize, INT, ZMOD_RW,
3112 	"Override block size estimate with fixed size");
3113