xref: /freebsd/sys/contrib/openzfs/module/zfs/dmu_send.c (revision 5ca8e32633c4ffbbcd6762e5888b6a4ba0708c6c)
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 uint_t 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 uint_t 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 uint_t zfs_send_queue_ff = 20;
95 static uint_t 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 uint_t 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 	uint32_t psize = drrw->drr_psize;
589 	uint32_t rsize = P2ROUNDUP(psize, 8);
590 
591 	if (psize != rsize)
592 		memset(buf + psize, 0, rsize - psize);
593 
594 	if (dump_record(dscp, buf, rsize) != 0)
595 		return (SET_ERROR(EINTR));
596 	return (0);
597 }
598 
599 static int
600 dump_spill(dmu_send_cookie_t *dscp, const blkptr_t *bp, uint64_t object,
601     void *data)
602 {
603 	struct drr_spill *drrs = &(dscp->dsc_drr->drr_u.drr_spill);
604 	uint64_t blksz = BP_GET_LSIZE(bp);
605 	uint64_t payload_size = blksz;
606 
607 	if (dscp->dsc_pending_op != PENDING_NONE) {
608 		if (dump_record(dscp, NULL, 0) != 0)
609 			return (SET_ERROR(EINTR));
610 		dscp->dsc_pending_op = PENDING_NONE;
611 	}
612 
613 	/* write a SPILL record */
614 	memset(dscp->dsc_drr, 0, sizeof (dmu_replay_record_t));
615 	dscp->dsc_drr->drr_type = DRR_SPILL;
616 	drrs->drr_object = object;
617 	drrs->drr_length = blksz;
618 	drrs->drr_toguid = dscp->dsc_toguid;
619 
620 	/* See comment in dump_dnode() for full details */
621 	if (zfs_send_unmodified_spill_blocks &&
622 	    (bp->blk_birth <= dscp->dsc_fromtxg)) {
623 		drrs->drr_flags |= DRR_SPILL_UNMODIFIED;
624 	}
625 
626 	/* handle raw send fields */
627 	if (dscp->dsc_featureflags & DMU_BACKUP_FEATURE_RAW) {
628 		ASSERT(BP_IS_PROTECTED(bp));
629 
630 		if (BP_SHOULD_BYTESWAP(bp))
631 			drrs->drr_flags |= DRR_RAW_BYTESWAP;
632 		drrs->drr_compressiontype = BP_GET_COMPRESS(bp);
633 		drrs->drr_compressed_size = BP_GET_PSIZE(bp);
634 		zio_crypt_decode_params_bp(bp, drrs->drr_salt, drrs->drr_iv);
635 		zio_crypt_decode_mac_bp(bp, drrs->drr_mac);
636 		payload_size = drrs->drr_compressed_size;
637 	}
638 
639 	if (dump_record(dscp, data, payload_size) != 0)
640 		return (SET_ERROR(EINTR));
641 	return (0);
642 }
643 
644 static int
645 dump_freeobjects(dmu_send_cookie_t *dscp, uint64_t firstobj, uint64_t numobjs)
646 {
647 	struct drr_freeobjects *drrfo = &(dscp->dsc_drr->drr_u.drr_freeobjects);
648 	uint64_t maxobj = DNODES_PER_BLOCK *
649 	    (DMU_META_DNODE(dscp->dsc_os)->dn_maxblkid + 1);
650 
651 	/*
652 	 * ZoL < 0.7 does not handle large FREEOBJECTS records correctly,
653 	 * leading to zfs recv never completing. to avoid this issue, don't
654 	 * send FREEOBJECTS records for object IDs which cannot exist on the
655 	 * receiving side.
656 	 */
657 	if (maxobj > 0) {
658 		if (maxobj <= firstobj)
659 			return (0);
660 
661 		if (maxobj < firstobj + numobjs)
662 			numobjs = maxobj - firstobj;
663 	}
664 
665 	/*
666 	 * If there is a pending op, but it's not PENDING_FREEOBJECTS,
667 	 * push it out, since free block aggregation can only be done for
668 	 * blocks of the same type (i.e., DRR_FREE records can only be
669 	 * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
670 	 * can only be aggregated with other DRR_FREEOBJECTS records).
671 	 */
672 	if (dscp->dsc_pending_op != PENDING_NONE &&
673 	    dscp->dsc_pending_op != PENDING_FREEOBJECTS) {
674 		if (dump_record(dscp, NULL, 0) != 0)
675 			return (SET_ERROR(EINTR));
676 		dscp->dsc_pending_op = PENDING_NONE;
677 	}
678 
679 	if (dscp->dsc_pending_op == PENDING_FREEOBJECTS) {
680 		/*
681 		 * See whether this free object array can be aggregated
682 		 * with pending one
683 		 */
684 		if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
685 			drrfo->drr_numobjs += numobjs;
686 			return (0);
687 		} else {
688 			/* can't be aggregated.  Push out pending record */
689 			if (dump_record(dscp, NULL, 0) != 0)
690 				return (SET_ERROR(EINTR));
691 			dscp->dsc_pending_op = PENDING_NONE;
692 		}
693 	}
694 
695 	/* write a FREEOBJECTS record */
696 	memset(dscp->dsc_drr, 0, sizeof (dmu_replay_record_t));
697 	dscp->dsc_drr->drr_type = DRR_FREEOBJECTS;
698 	drrfo->drr_firstobj = firstobj;
699 	drrfo->drr_numobjs = numobjs;
700 	drrfo->drr_toguid = dscp->dsc_toguid;
701 
702 	dscp->dsc_pending_op = PENDING_FREEOBJECTS;
703 
704 	return (0);
705 }
706 
707 static int
708 dump_dnode(dmu_send_cookie_t *dscp, const blkptr_t *bp, uint64_t object,
709     dnode_phys_t *dnp)
710 {
711 	struct drr_object *drro = &(dscp->dsc_drr->drr_u.drr_object);
712 	int bonuslen;
713 
714 	if (object < dscp->dsc_resume_object) {
715 		/*
716 		 * Note: when resuming, we will visit all the dnodes in
717 		 * the block of dnodes that we are resuming from.  In
718 		 * this case it's unnecessary to send the dnodes prior to
719 		 * the one we are resuming from.  We should be at most one
720 		 * block's worth of dnodes behind the resume point.
721 		 */
722 		ASSERT3U(dscp->dsc_resume_object - object, <,
723 		    1 << (DNODE_BLOCK_SHIFT - DNODE_SHIFT));
724 		return (0);
725 	}
726 
727 	if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
728 		return (dump_freeobjects(dscp, object, 1));
729 
730 	if (dscp->dsc_pending_op != PENDING_NONE) {
731 		if (dump_record(dscp, NULL, 0) != 0)
732 			return (SET_ERROR(EINTR));
733 		dscp->dsc_pending_op = PENDING_NONE;
734 	}
735 
736 	/* write an OBJECT record */
737 	memset(dscp->dsc_drr, 0, sizeof (dmu_replay_record_t));
738 	dscp->dsc_drr->drr_type = DRR_OBJECT;
739 	drro->drr_object = object;
740 	drro->drr_type = dnp->dn_type;
741 	drro->drr_bonustype = dnp->dn_bonustype;
742 	drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
743 	drro->drr_bonuslen = dnp->dn_bonuslen;
744 	drro->drr_dn_slots = dnp->dn_extra_slots + 1;
745 	drro->drr_checksumtype = dnp->dn_checksum;
746 	drro->drr_compress = dnp->dn_compress;
747 	drro->drr_toguid = dscp->dsc_toguid;
748 
749 	if (!(dscp->dsc_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
750 	    drro->drr_blksz > SPA_OLD_MAXBLOCKSIZE)
751 		drro->drr_blksz = SPA_OLD_MAXBLOCKSIZE;
752 
753 	bonuslen = P2ROUNDUP(dnp->dn_bonuslen, 8);
754 
755 	if ((dscp->dsc_featureflags & DMU_BACKUP_FEATURE_RAW)) {
756 		ASSERT(BP_IS_ENCRYPTED(bp));
757 
758 		if (BP_SHOULD_BYTESWAP(bp))
759 			drro->drr_flags |= DRR_RAW_BYTESWAP;
760 
761 		/* needed for reconstructing dnp on recv side */
762 		drro->drr_maxblkid = dnp->dn_maxblkid;
763 		drro->drr_indblkshift = dnp->dn_indblkshift;
764 		drro->drr_nlevels = dnp->dn_nlevels;
765 		drro->drr_nblkptr = dnp->dn_nblkptr;
766 
767 		/*
768 		 * Since we encrypt the entire bonus area, the (raw) part
769 		 * beyond the bonuslen is actually nonzero, so we need
770 		 * to send it.
771 		 */
772 		if (bonuslen != 0) {
773 			if (drro->drr_bonuslen > DN_MAX_BONUS_LEN(dnp))
774 				return (SET_ERROR(EINVAL));
775 			drro->drr_raw_bonuslen = DN_MAX_BONUS_LEN(dnp);
776 			bonuslen = drro->drr_raw_bonuslen;
777 		}
778 	}
779 
780 	/*
781 	 * DRR_OBJECT_SPILL is set for every dnode which references a
782 	 * spill block.	 This allows the receiving pool to definitively
783 	 * determine when a spill block should be kept or freed.
784 	 */
785 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
786 		drro->drr_flags |= DRR_OBJECT_SPILL;
787 
788 	if (dump_record(dscp, DN_BONUS(dnp), bonuslen) != 0)
789 		return (SET_ERROR(EINTR));
790 
791 	/* Free anything past the end of the file. */
792 	if (dump_free(dscp, object, (dnp->dn_maxblkid + 1) *
793 	    (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), DMU_OBJECT_END) != 0)
794 		return (SET_ERROR(EINTR));
795 
796 	/*
797 	 * Send DRR_SPILL records for unmodified spill blocks.	This is useful
798 	 * because changing certain attributes of the object (e.g. blocksize)
799 	 * can cause old versions of ZFS to incorrectly remove a spill block.
800 	 * Including these records in the stream forces an up to date version
801 	 * to always be written ensuring they're never lost.  Current versions
802 	 * of the code which understand the DRR_FLAG_SPILL_BLOCK feature can
803 	 * ignore these unmodified spill blocks.
804 	 */
805 	if (zfs_send_unmodified_spill_blocks &&
806 	    (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) &&
807 	    (DN_SPILL_BLKPTR(dnp)->blk_birth <= dscp->dsc_fromtxg)) {
808 		struct send_range record;
809 		blkptr_t *bp = DN_SPILL_BLKPTR(dnp);
810 
811 		memset(&record, 0, sizeof (struct send_range));
812 		record.type = DATA;
813 		record.object = object;
814 		record.eos_marker = B_FALSE;
815 		record.start_blkid = DMU_SPILL_BLKID;
816 		record.end_blkid = record.start_blkid + 1;
817 		record.sru.data.bp = *bp;
818 		record.sru.data.obj_type = dnp->dn_type;
819 		record.sru.data.datablksz = BP_GET_LSIZE(bp);
820 
821 		if (do_dump(dscp, &record) != 0)
822 			return (SET_ERROR(EINTR));
823 	}
824 
825 	if (dscp->dsc_err != 0)
826 		return (SET_ERROR(EINTR));
827 
828 	return (0);
829 }
830 
831 static int
832 dump_object_range(dmu_send_cookie_t *dscp, const blkptr_t *bp,
833     uint64_t firstobj, uint64_t numslots)
834 {
835 	struct drr_object_range *drror =
836 	    &(dscp->dsc_drr->drr_u.drr_object_range);
837 
838 	/* we only use this record type for raw sends */
839 	ASSERT(BP_IS_PROTECTED(bp));
840 	ASSERT(dscp->dsc_featureflags & DMU_BACKUP_FEATURE_RAW);
841 	ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
842 	ASSERT3U(BP_GET_TYPE(bp), ==, DMU_OT_DNODE);
843 	ASSERT0(BP_GET_LEVEL(bp));
844 
845 	if (dscp->dsc_pending_op != PENDING_NONE) {
846 		if (dump_record(dscp, NULL, 0) != 0)
847 			return (SET_ERROR(EINTR));
848 		dscp->dsc_pending_op = PENDING_NONE;
849 	}
850 
851 	memset(dscp->dsc_drr, 0, sizeof (dmu_replay_record_t));
852 	dscp->dsc_drr->drr_type = DRR_OBJECT_RANGE;
853 	drror->drr_firstobj = firstobj;
854 	drror->drr_numslots = numslots;
855 	drror->drr_toguid = dscp->dsc_toguid;
856 	if (BP_SHOULD_BYTESWAP(bp))
857 		drror->drr_flags |= DRR_RAW_BYTESWAP;
858 	zio_crypt_decode_params_bp(bp, drror->drr_salt, drror->drr_iv);
859 	zio_crypt_decode_mac_bp(bp, drror->drr_mac);
860 
861 	if (dump_record(dscp, NULL, 0) != 0)
862 		return (SET_ERROR(EINTR));
863 	return (0);
864 }
865 
866 static boolean_t
867 send_do_embed(const blkptr_t *bp, uint64_t featureflags)
868 {
869 	if (!BP_IS_EMBEDDED(bp))
870 		return (B_FALSE);
871 
872 	/*
873 	 * Compression function must be legacy, or explicitly enabled.
874 	 */
875 	if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS &&
876 	    !(featureflags & DMU_BACKUP_FEATURE_LZ4)))
877 		return (B_FALSE);
878 
879 	/*
880 	 * If we have not set the ZSTD feature flag, we can't send ZSTD
881 	 * compressed embedded blocks, as the receiver may not support them.
882 	 */
883 	if ((BP_GET_COMPRESS(bp) == ZIO_COMPRESS_ZSTD &&
884 	    !(featureflags & DMU_BACKUP_FEATURE_ZSTD)))
885 		return (B_FALSE);
886 
887 	/*
888 	 * Embed type must be explicitly enabled.
889 	 */
890 	switch (BPE_GET_ETYPE(bp)) {
891 	case BP_EMBEDDED_TYPE_DATA:
892 		if (featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)
893 			return (B_TRUE);
894 		break;
895 	default:
896 		return (B_FALSE);
897 	}
898 	return (B_FALSE);
899 }
900 
901 /*
902  * This function actually handles figuring out what kind of record needs to be
903  * dumped, and calling the appropriate helper function.  In most cases,
904  * the data has already been read by send_reader_thread().
905  */
906 static int
907 do_dump(dmu_send_cookie_t *dscp, struct send_range *range)
908 {
909 	int err = 0;
910 	switch (range->type) {
911 	case OBJECT:
912 		err = dump_dnode(dscp, &range->sru.object.bp, range->object,
913 		    range->sru.object.dnp);
914 		return (err);
915 	case OBJECT_RANGE: {
916 		ASSERT3U(range->start_blkid + 1, ==, range->end_blkid);
917 		if (!(dscp->dsc_featureflags & DMU_BACKUP_FEATURE_RAW)) {
918 			return (0);
919 		}
920 		uint64_t epb = BP_GET_LSIZE(&range->sru.object_range.bp) >>
921 		    DNODE_SHIFT;
922 		uint64_t firstobj = range->start_blkid * epb;
923 		err = dump_object_range(dscp, &range->sru.object_range.bp,
924 		    firstobj, epb);
925 		break;
926 	}
927 	case REDACT: {
928 		struct srr *srrp = &range->sru.redact;
929 		err = dump_redact(dscp, range->object, range->start_blkid *
930 		    srrp->datablksz, (range->end_blkid - range->start_blkid) *
931 		    srrp->datablksz);
932 		return (err);
933 	}
934 	case DATA: {
935 		struct srd *srdp = &range->sru.data;
936 		blkptr_t *bp = &srdp->bp;
937 		spa_t *spa =
938 		    dmu_objset_spa(dscp->dsc_os);
939 
940 		ASSERT3U(srdp->datablksz, ==, BP_GET_LSIZE(bp));
941 		ASSERT3U(range->start_blkid + 1, ==, range->end_blkid);
942 		if (BP_GET_TYPE(bp) == DMU_OT_SA) {
943 			arc_flags_t aflags = ARC_FLAG_WAIT;
944 			zio_flag_t zioflags = ZIO_FLAG_CANFAIL;
945 
946 			if (dscp->dsc_featureflags & DMU_BACKUP_FEATURE_RAW) {
947 				ASSERT(BP_IS_PROTECTED(bp));
948 				zioflags |= ZIO_FLAG_RAW;
949 			}
950 
951 			zbookmark_phys_t zb;
952 			ASSERT3U(range->start_blkid, ==, DMU_SPILL_BLKID);
953 			zb.zb_objset = dmu_objset_id(dscp->dsc_os);
954 			zb.zb_object = range->object;
955 			zb.zb_level = 0;
956 			zb.zb_blkid = range->start_blkid;
957 
958 			arc_buf_t *abuf = NULL;
959 			if (!dscp->dsc_dso->dso_dryrun && arc_read(NULL, spa,
960 			    bp, arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
961 			    zioflags, &aflags, &zb) != 0)
962 				return (SET_ERROR(EIO));
963 
964 			err = dump_spill(dscp, bp, zb.zb_object,
965 			    (abuf == NULL ? NULL : abuf->b_data));
966 			if (abuf != NULL)
967 				arc_buf_destroy(abuf, &abuf);
968 			return (err);
969 		}
970 		if (send_do_embed(bp, dscp->dsc_featureflags)) {
971 			err = dump_write_embedded(dscp, range->object,
972 			    range->start_blkid * srdp->datablksz,
973 			    srdp->datablksz, bp);
974 			return (err);
975 		}
976 		ASSERT(range->object > dscp->dsc_resume_object ||
977 		    (range->object == dscp->dsc_resume_object &&
978 		    range->start_blkid * srdp->datablksz >=
979 		    dscp->dsc_resume_offset));
980 		/* it's a level-0 block of a regular object */
981 
982 		mutex_enter(&srdp->lock);
983 		while (srdp->io_outstanding)
984 			cv_wait(&srdp->cv, &srdp->lock);
985 		err = srdp->io_err;
986 		mutex_exit(&srdp->lock);
987 
988 		if (err != 0) {
989 			if (zfs_send_corrupt_data &&
990 			    !dscp->dsc_dso->dso_dryrun) {
991 				/*
992 				 * Send a block filled with 0x"zfs badd bloc"
993 				 */
994 				srdp->abuf = arc_alloc_buf(spa, &srdp->abuf,
995 				    ARC_BUFC_DATA, srdp->datablksz);
996 				uint64_t *ptr;
997 				for (ptr = srdp->abuf->b_data;
998 				    (char *)ptr < (char *)srdp->abuf->b_data +
999 				    srdp->datablksz; ptr++)
1000 					*ptr = 0x2f5baddb10cULL;
1001 			} else {
1002 				return (SET_ERROR(EIO));
1003 			}
1004 		}
1005 
1006 		ASSERT(dscp->dsc_dso->dso_dryrun ||
1007 		    srdp->abuf != NULL || srdp->abd != NULL);
1008 
1009 		uint64_t offset = range->start_blkid * srdp->datablksz;
1010 
1011 		char *data = NULL;
1012 		if (srdp->abd != NULL) {
1013 			data = abd_to_buf(srdp->abd);
1014 			ASSERT3P(srdp->abuf, ==, NULL);
1015 		} else if (srdp->abuf != NULL) {
1016 			data = srdp->abuf->b_data;
1017 		}
1018 
1019 		/*
1020 		 * If we have large blocks stored on disk but the send flags
1021 		 * don't allow us to send large blocks, we split the data from
1022 		 * the arc buf into chunks.
1023 		 */
1024 		if (srdp->datablksz > SPA_OLD_MAXBLOCKSIZE &&
1025 		    !(dscp->dsc_featureflags &
1026 		    DMU_BACKUP_FEATURE_LARGE_BLOCKS)) {
1027 			while (srdp->datablksz > 0 && err == 0) {
1028 				int n = MIN(srdp->datablksz,
1029 				    SPA_OLD_MAXBLOCKSIZE);
1030 				err = dmu_dump_write(dscp, srdp->obj_type,
1031 				    range->object, offset, n, n, NULL, B_FALSE,
1032 				    data);
1033 				offset += n;
1034 				/*
1035 				 * When doing dry run, data==NULL is used as a
1036 				 * sentinel value by
1037 				 * dmu_dump_write()->dump_record().
1038 				 */
1039 				if (data != NULL)
1040 					data += n;
1041 				srdp->datablksz -= n;
1042 			}
1043 		} else {
1044 			err = dmu_dump_write(dscp, srdp->obj_type,
1045 			    range->object, offset,
1046 			    srdp->datablksz, srdp->datasz, bp,
1047 			    srdp->io_compressed, data);
1048 		}
1049 		return (err);
1050 	}
1051 	case HOLE: {
1052 		struct srh *srhp = &range->sru.hole;
1053 		if (range->object == DMU_META_DNODE_OBJECT) {
1054 			uint32_t span = srhp->datablksz >> DNODE_SHIFT;
1055 			uint64_t first_obj = range->start_blkid * span;
1056 			uint64_t numobj = range->end_blkid * span - first_obj;
1057 			return (dump_freeobjects(dscp, first_obj, numobj));
1058 		}
1059 		uint64_t offset = 0;
1060 
1061 		/*
1062 		 * If this multiply overflows, we don't need to send this block.
1063 		 * Even if it has a birth time, it can never not be a hole, so
1064 		 * we don't need to send records for it.
1065 		 */
1066 		if (!overflow_multiply(range->start_blkid, srhp->datablksz,
1067 		    &offset)) {
1068 			return (0);
1069 		}
1070 		uint64_t len = 0;
1071 
1072 		if (!overflow_multiply(range->end_blkid, srhp->datablksz, &len))
1073 			len = UINT64_MAX;
1074 		len = len - offset;
1075 		return (dump_free(dscp, range->object, offset, len));
1076 	}
1077 	default:
1078 		panic("Invalid range type in do_dump: %d", range->type);
1079 	}
1080 	return (err);
1081 }
1082 
1083 static struct send_range *
1084 range_alloc(enum type type, uint64_t object, uint64_t start_blkid,
1085     uint64_t end_blkid, boolean_t eos)
1086 {
1087 	struct send_range *range = kmem_alloc(sizeof (*range), KM_SLEEP);
1088 	range->type = type;
1089 	range->object = object;
1090 	range->start_blkid = start_blkid;
1091 	range->end_blkid = end_blkid;
1092 	range->eos_marker = eos;
1093 	if (type == DATA) {
1094 		range->sru.data.abd = NULL;
1095 		range->sru.data.abuf = NULL;
1096 		mutex_init(&range->sru.data.lock, NULL, MUTEX_DEFAULT, NULL);
1097 		cv_init(&range->sru.data.cv, NULL, CV_DEFAULT, NULL);
1098 		range->sru.data.io_outstanding = 0;
1099 		range->sru.data.io_err = 0;
1100 		range->sru.data.io_compressed = B_FALSE;
1101 	}
1102 	return (range);
1103 }
1104 
1105 /*
1106  * This is the callback function to traverse_dataset that acts as a worker
1107  * thread for dmu_send_impl.
1108  */
1109 static int
1110 send_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
1111     const zbookmark_phys_t *zb, const struct dnode_phys *dnp, void *arg)
1112 {
1113 	(void) zilog;
1114 	struct send_thread_arg *sta = arg;
1115 	struct send_range *record;
1116 
1117 	ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
1118 	    zb->zb_object >= sta->resume.zb_object);
1119 
1120 	/*
1121 	 * All bps of an encrypted os should have the encryption bit set.
1122 	 * If this is not true it indicates tampering and we report an error.
1123 	 */
1124 	if (sta->os->os_encrypted &&
1125 	    !BP_IS_HOLE(bp) && !BP_USES_CRYPT(bp)) {
1126 		spa_log_error(spa, zb, &bp->blk_birth);
1127 		return (SET_ERROR(EIO));
1128 	}
1129 
1130 	if (sta->cancel)
1131 		return (SET_ERROR(EINTR));
1132 	if (zb->zb_object != DMU_META_DNODE_OBJECT &&
1133 	    DMU_OBJECT_IS_SPECIAL(zb->zb_object))
1134 		return (0);
1135 	atomic_inc_64(sta->num_blocks_visited);
1136 
1137 	if (zb->zb_level == ZB_DNODE_LEVEL) {
1138 		if (zb->zb_object == DMU_META_DNODE_OBJECT)
1139 			return (0);
1140 		record = range_alloc(OBJECT, zb->zb_object, 0, 0, B_FALSE);
1141 		record->sru.object.bp = *bp;
1142 		size_t size  = sizeof (*dnp) * (dnp->dn_extra_slots + 1);
1143 		record->sru.object.dnp = kmem_alloc(size, KM_SLEEP);
1144 		memcpy(record->sru.object.dnp, dnp, size);
1145 		bqueue_enqueue(&sta->q, record, sizeof (*record));
1146 		return (0);
1147 	}
1148 	if (zb->zb_level == 0 && zb->zb_object == DMU_META_DNODE_OBJECT &&
1149 	    !BP_IS_HOLE(bp)) {
1150 		record = range_alloc(OBJECT_RANGE, 0, zb->zb_blkid,
1151 		    zb->zb_blkid + 1, B_FALSE);
1152 		record->sru.object_range.bp = *bp;
1153 		bqueue_enqueue(&sta->q, record, sizeof (*record));
1154 		return (0);
1155 	}
1156 	if (zb->zb_level < 0 || (zb->zb_level > 0 && !BP_IS_HOLE(bp)))
1157 		return (0);
1158 	if (zb->zb_object == DMU_META_DNODE_OBJECT && !BP_IS_HOLE(bp))
1159 		return (0);
1160 
1161 	uint64_t span = bp_span_in_blocks(dnp->dn_indblkshift, zb->zb_level);
1162 	uint64_t start;
1163 
1164 	/*
1165 	 * If this multiply overflows, we don't need to send this block.
1166 	 * Even if it has a birth time, it can never not be a hole, so
1167 	 * we don't need to send records for it.
1168 	 */
1169 	if (!overflow_multiply(span, zb->zb_blkid, &start) || (!(zb->zb_blkid ==
1170 	    DMU_SPILL_BLKID || DMU_OT_IS_METADATA(dnp->dn_type)) &&
1171 	    span * zb->zb_blkid > dnp->dn_maxblkid)) {
1172 		ASSERT(BP_IS_HOLE(bp));
1173 		return (0);
1174 	}
1175 
1176 	if (zb->zb_blkid == DMU_SPILL_BLKID)
1177 		ASSERT3U(BP_GET_TYPE(bp), ==, DMU_OT_SA);
1178 
1179 	enum type record_type = DATA;
1180 	if (BP_IS_HOLE(bp))
1181 		record_type = HOLE;
1182 	else if (BP_IS_REDACTED(bp))
1183 		record_type = REDACT;
1184 	else
1185 		record_type = DATA;
1186 
1187 	record = range_alloc(record_type, zb->zb_object, start,
1188 	    (start + span < start ? 0 : start + span), B_FALSE);
1189 
1190 	uint64_t datablksz = (zb->zb_blkid == DMU_SPILL_BLKID ?
1191 	    BP_GET_LSIZE(bp) : dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
1192 
1193 	if (BP_IS_HOLE(bp)) {
1194 		record->sru.hole.datablksz = datablksz;
1195 	} else if (BP_IS_REDACTED(bp)) {
1196 		record->sru.redact.datablksz = datablksz;
1197 	} else {
1198 		record->sru.data.datablksz = datablksz;
1199 		record->sru.data.obj_type = dnp->dn_type;
1200 		record->sru.data.bp = *bp;
1201 	}
1202 
1203 	bqueue_enqueue(&sta->q, record, sizeof (*record));
1204 	return (0);
1205 }
1206 
1207 struct redact_list_cb_arg {
1208 	uint64_t *num_blocks_visited;
1209 	bqueue_t *q;
1210 	boolean_t *cancel;
1211 	boolean_t mark_redact;
1212 };
1213 
1214 static int
1215 redact_list_cb(redact_block_phys_t *rb, void *arg)
1216 {
1217 	struct redact_list_cb_arg *rlcap = arg;
1218 
1219 	atomic_inc_64(rlcap->num_blocks_visited);
1220 	if (*rlcap->cancel)
1221 		return (-1);
1222 
1223 	struct send_range *data = range_alloc(REDACT, rb->rbp_object,
1224 	    rb->rbp_blkid, rb->rbp_blkid + redact_block_get_count(rb), B_FALSE);
1225 	ASSERT3U(data->end_blkid, >, rb->rbp_blkid);
1226 	if (rlcap->mark_redact) {
1227 		data->type = REDACT;
1228 		data->sru.redact.datablksz = redact_block_get_size(rb);
1229 	} else {
1230 		data->type = PREVIOUSLY_REDACTED;
1231 	}
1232 	bqueue_enqueue(rlcap->q, data, sizeof (*data));
1233 
1234 	return (0);
1235 }
1236 
1237 /*
1238  * This function kicks off the traverse_dataset.  It also handles setting the
1239  * error code of the thread in case something goes wrong, and pushes the End of
1240  * Stream record when the traverse_dataset call has finished.
1241  */
1242 static __attribute__((noreturn)) void
1243 send_traverse_thread(void *arg)
1244 {
1245 	struct send_thread_arg *st_arg = arg;
1246 	int err = 0;
1247 	struct send_range *data;
1248 	fstrans_cookie_t cookie = spl_fstrans_mark();
1249 
1250 	err = traverse_dataset_resume(st_arg->os->os_dsl_dataset,
1251 	    st_arg->fromtxg, &st_arg->resume,
1252 	    st_arg->flags, send_cb, st_arg);
1253 
1254 	if (err != EINTR)
1255 		st_arg->error_code = err;
1256 	data = range_alloc(DATA, 0, 0, 0, B_TRUE);
1257 	bqueue_enqueue_flush(&st_arg->q, data, sizeof (*data));
1258 	spl_fstrans_unmark(cookie);
1259 	thread_exit();
1260 }
1261 
1262 /*
1263  * Utility function that causes End of Stream records to compare after of all
1264  * others, so that other threads' comparison logic can stay simple.
1265  */
1266 static int __attribute__((unused))
1267 send_range_after(const struct send_range *from, const struct send_range *to)
1268 {
1269 	if (from->eos_marker == B_TRUE)
1270 		return (1);
1271 	if (to->eos_marker == B_TRUE)
1272 		return (-1);
1273 
1274 	uint64_t from_obj = from->object;
1275 	uint64_t from_end_obj = from->object + 1;
1276 	uint64_t to_obj = to->object;
1277 	uint64_t to_end_obj = to->object + 1;
1278 	if (from_obj == 0) {
1279 		ASSERT(from->type == HOLE || from->type == OBJECT_RANGE);
1280 		from_obj = from->start_blkid << DNODES_PER_BLOCK_SHIFT;
1281 		from_end_obj = from->end_blkid << DNODES_PER_BLOCK_SHIFT;
1282 	}
1283 	if (to_obj == 0) {
1284 		ASSERT(to->type == HOLE || to->type == OBJECT_RANGE);
1285 		to_obj = to->start_blkid << DNODES_PER_BLOCK_SHIFT;
1286 		to_end_obj = to->end_blkid << DNODES_PER_BLOCK_SHIFT;
1287 	}
1288 
1289 	if (from_end_obj <= to_obj)
1290 		return (-1);
1291 	if (from_obj >= to_end_obj)
1292 		return (1);
1293 	int64_t cmp = TREE_CMP(to->type == OBJECT_RANGE, from->type ==
1294 	    OBJECT_RANGE);
1295 	if (unlikely(cmp))
1296 		return (cmp);
1297 	cmp = TREE_CMP(to->type == OBJECT, from->type == OBJECT);
1298 	if (unlikely(cmp))
1299 		return (cmp);
1300 	if (from->end_blkid <= to->start_blkid)
1301 		return (-1);
1302 	if (from->start_blkid >= to->end_blkid)
1303 		return (1);
1304 	return (0);
1305 }
1306 
1307 /*
1308  * Pop the new data off the queue, check that the records we receive are in
1309  * the right order, but do not free the old data.  This is used so that the
1310  * records can be sent on to the main thread without copying the data.
1311  */
1312 static struct send_range *
1313 get_next_range_nofree(bqueue_t *bq, struct send_range *prev)
1314 {
1315 	struct send_range *next = bqueue_dequeue(bq);
1316 	ASSERT3S(send_range_after(prev, next), ==, -1);
1317 	return (next);
1318 }
1319 
1320 /*
1321  * Pop the new data off the queue, check that the records we receive are in
1322  * the right order, and free the old data.
1323  */
1324 static struct send_range *
1325 get_next_range(bqueue_t *bq, struct send_range *prev)
1326 {
1327 	struct send_range *next = get_next_range_nofree(bq, prev);
1328 	range_free(prev);
1329 	return (next);
1330 }
1331 
1332 static __attribute__((noreturn)) void
1333 redact_list_thread(void *arg)
1334 {
1335 	struct redact_list_thread_arg *rlt_arg = arg;
1336 	struct send_range *record;
1337 	fstrans_cookie_t cookie = spl_fstrans_mark();
1338 	if (rlt_arg->rl != NULL) {
1339 		struct redact_list_cb_arg rlcba = {0};
1340 		rlcba.cancel = &rlt_arg->cancel;
1341 		rlcba.q = &rlt_arg->q;
1342 		rlcba.num_blocks_visited = rlt_arg->num_blocks_visited;
1343 		rlcba.mark_redact = rlt_arg->mark_redact;
1344 		int err = dsl_redaction_list_traverse(rlt_arg->rl,
1345 		    &rlt_arg->resume, redact_list_cb, &rlcba);
1346 		if (err != EINTR)
1347 			rlt_arg->error_code = err;
1348 	}
1349 	record = range_alloc(DATA, 0, 0, 0, B_TRUE);
1350 	bqueue_enqueue_flush(&rlt_arg->q, record, sizeof (*record));
1351 	spl_fstrans_unmark(cookie);
1352 
1353 	thread_exit();
1354 }
1355 
1356 /*
1357  * Compare the start point of the two provided ranges. End of stream ranges
1358  * compare last, objects compare before any data or hole inside that object and
1359  * multi-object holes that start at the same object.
1360  */
1361 static int
1362 send_range_start_compare(struct send_range *r1, struct send_range *r2)
1363 {
1364 	uint64_t r1_objequiv = r1->object;
1365 	uint64_t r1_l0equiv = r1->start_blkid;
1366 	uint64_t r2_objequiv = r2->object;
1367 	uint64_t r2_l0equiv = r2->start_blkid;
1368 	int64_t cmp = TREE_CMP(r1->eos_marker, r2->eos_marker);
1369 	if (unlikely(cmp))
1370 		return (cmp);
1371 	if (r1->object == 0) {
1372 		r1_objequiv = r1->start_blkid * DNODES_PER_BLOCK;
1373 		r1_l0equiv = 0;
1374 	}
1375 	if (r2->object == 0) {
1376 		r2_objequiv = r2->start_blkid * DNODES_PER_BLOCK;
1377 		r2_l0equiv = 0;
1378 	}
1379 
1380 	cmp = TREE_CMP(r1_objequiv, r2_objequiv);
1381 	if (likely(cmp))
1382 		return (cmp);
1383 	cmp = TREE_CMP(r2->type == OBJECT_RANGE, r1->type == OBJECT_RANGE);
1384 	if (unlikely(cmp))
1385 		return (cmp);
1386 	cmp = TREE_CMP(r2->type == OBJECT, r1->type == OBJECT);
1387 	if (unlikely(cmp))
1388 		return (cmp);
1389 
1390 	return (TREE_CMP(r1_l0equiv, r2_l0equiv));
1391 }
1392 
1393 enum q_idx {
1394 	REDACT_IDX = 0,
1395 	TO_IDX,
1396 	FROM_IDX,
1397 	NUM_THREADS
1398 };
1399 
1400 /*
1401  * This function returns the next range the send_merge_thread should operate on.
1402  * The inputs are two arrays; the first one stores the range at the front of the
1403  * queues stored in the second one.  The ranges are sorted in descending
1404  * priority order; the metadata from earlier ranges overrules metadata from
1405  * later ranges.  out_mask is used to return which threads the ranges came from;
1406  * bit i is set if ranges[i] started at the same place as the returned range.
1407  *
1408  * This code is not hardcoded to compare a specific number of threads; it could
1409  * be used with any number, just by changing the q_idx enum.
1410  *
1411  * The "next range" is the one with the earliest start; if two starts are equal,
1412  * the highest-priority range is the next to operate on.  If a higher-priority
1413  * range starts in the middle of the first range, then the first range will be
1414  * truncated to end where the higher-priority range starts, and we will operate
1415  * on that one next time.   In this way, we make sure that each block covered by
1416  * some range gets covered by a returned range, and each block covered is
1417  * returned using the metadata of the highest-priority range it appears in.
1418  *
1419  * For example, if the three ranges at the front of the queues were [2,4),
1420  * [3,5), and [1,3), then the ranges returned would be [1,2) with the metadata
1421  * from the third range, [2,4) with the metadata from the first range, and then
1422  * [4,5) with the metadata from the second.
1423  */
1424 static struct send_range *
1425 find_next_range(struct send_range **ranges, bqueue_t **qs, uint64_t *out_mask)
1426 {
1427 	int idx = 0; // index of the range with the earliest start
1428 	int i;
1429 	uint64_t bmask = 0;
1430 	for (i = 1; i < NUM_THREADS; i++) {
1431 		if (send_range_start_compare(ranges[i], ranges[idx]) < 0)
1432 			idx = i;
1433 	}
1434 	if (ranges[idx]->eos_marker) {
1435 		struct send_range *ret = range_alloc(DATA, 0, 0, 0, B_TRUE);
1436 		*out_mask = 0;
1437 		return (ret);
1438 	}
1439 	/*
1440 	 * Find all the ranges that start at that same point.
1441 	 */
1442 	for (i = 0; i < NUM_THREADS; i++) {
1443 		if (send_range_start_compare(ranges[i], ranges[idx]) == 0)
1444 			bmask |= 1 << i;
1445 	}
1446 	*out_mask = bmask;
1447 	/*
1448 	 * OBJECT_RANGE records only come from the TO thread, and should always
1449 	 * be treated as overlapping with nothing and sent on immediately.  They
1450 	 * are only used in raw sends, and are never redacted.
1451 	 */
1452 	if (ranges[idx]->type == OBJECT_RANGE) {
1453 		ASSERT3U(idx, ==, TO_IDX);
1454 		ASSERT3U(*out_mask, ==, 1 << TO_IDX);
1455 		struct send_range *ret = ranges[idx];
1456 		ranges[idx] = get_next_range_nofree(qs[idx], ranges[idx]);
1457 		return (ret);
1458 	}
1459 	/*
1460 	 * Find the first start or end point after the start of the first range.
1461 	 */
1462 	uint64_t first_change = ranges[idx]->end_blkid;
1463 	for (i = 0; i < NUM_THREADS; i++) {
1464 		if (i == idx || ranges[i]->eos_marker ||
1465 		    ranges[i]->object > ranges[idx]->object ||
1466 		    ranges[i]->object == DMU_META_DNODE_OBJECT)
1467 			continue;
1468 		ASSERT3U(ranges[i]->object, ==, ranges[idx]->object);
1469 		if (first_change > ranges[i]->start_blkid &&
1470 		    (bmask & (1 << i)) == 0)
1471 			first_change = ranges[i]->start_blkid;
1472 		else if (first_change > ranges[i]->end_blkid)
1473 			first_change = ranges[i]->end_blkid;
1474 	}
1475 	/*
1476 	 * Update all ranges to no longer overlap with the range we're
1477 	 * returning. All such ranges must start at the same place as the range
1478 	 * being returned, and end at or after first_change. Thus we update
1479 	 * their start to first_change. If that makes them size 0, then free
1480 	 * them and pull a new range from that thread.
1481 	 */
1482 	for (i = 0; i < NUM_THREADS; i++) {
1483 		if (i == idx || (bmask & (1 << i)) == 0)
1484 			continue;
1485 		ASSERT3U(first_change, >, ranges[i]->start_blkid);
1486 		ranges[i]->start_blkid = first_change;
1487 		ASSERT3U(ranges[i]->start_blkid, <=, ranges[i]->end_blkid);
1488 		if (ranges[i]->start_blkid == ranges[i]->end_blkid)
1489 			ranges[i] = get_next_range(qs[i], ranges[i]);
1490 	}
1491 	/*
1492 	 * Short-circuit the simple case; if the range doesn't overlap with
1493 	 * anything else, or it only overlaps with things that start at the same
1494 	 * place and are longer, send it on.
1495 	 */
1496 	if (first_change == ranges[idx]->end_blkid) {
1497 		struct send_range *ret = ranges[idx];
1498 		ranges[idx] = get_next_range_nofree(qs[idx], ranges[idx]);
1499 		return (ret);
1500 	}
1501 
1502 	/*
1503 	 * Otherwise, return a truncated copy of ranges[idx] and move the start
1504 	 * of ranges[idx] back to first_change.
1505 	 */
1506 	struct send_range *ret = kmem_alloc(sizeof (*ret), KM_SLEEP);
1507 	*ret = *ranges[idx];
1508 	ret->end_blkid = first_change;
1509 	ranges[idx]->start_blkid = first_change;
1510 	return (ret);
1511 }
1512 
1513 #define	FROM_AND_REDACT_BITS ((1 << REDACT_IDX) | (1 << FROM_IDX))
1514 
1515 /*
1516  * Merge the results from the from thread and the to thread, and then hand the
1517  * records off to send_prefetch_thread to prefetch them.  If this is not a
1518  * send from a redaction bookmark, the from thread will push an end of stream
1519  * record and stop, and we'll just send everything that was changed in the
1520  * to_ds since the ancestor's creation txg. If it is, then since
1521  * traverse_dataset has a canonical order, we can compare each change as
1522  * they're pulled off the queues.  That will give us a stream that is
1523  * appropriately sorted, and covers all records.  In addition, we pull the
1524  * data from the redact_list_thread and use that to determine which blocks
1525  * should be redacted.
1526  */
1527 static __attribute__((noreturn)) void
1528 send_merge_thread(void *arg)
1529 {
1530 	struct send_merge_thread_arg *smt_arg = arg;
1531 	struct send_range *front_ranges[NUM_THREADS];
1532 	bqueue_t *queues[NUM_THREADS];
1533 	int err = 0;
1534 	fstrans_cookie_t cookie = spl_fstrans_mark();
1535 
1536 	if (smt_arg->redact_arg == NULL) {
1537 		front_ranges[REDACT_IDX] =
1538 		    kmem_zalloc(sizeof (struct send_range), KM_SLEEP);
1539 		front_ranges[REDACT_IDX]->eos_marker = B_TRUE;
1540 		front_ranges[REDACT_IDX]->type = REDACT;
1541 		queues[REDACT_IDX] = NULL;
1542 	} else {
1543 		front_ranges[REDACT_IDX] =
1544 		    bqueue_dequeue(&smt_arg->redact_arg->q);
1545 		queues[REDACT_IDX] = &smt_arg->redact_arg->q;
1546 	}
1547 	front_ranges[TO_IDX] = bqueue_dequeue(&smt_arg->to_arg->q);
1548 	queues[TO_IDX] = &smt_arg->to_arg->q;
1549 	front_ranges[FROM_IDX] = bqueue_dequeue(&smt_arg->from_arg->q);
1550 	queues[FROM_IDX] = &smt_arg->from_arg->q;
1551 	uint64_t mask = 0;
1552 	struct send_range *range;
1553 	for (range = find_next_range(front_ranges, queues, &mask);
1554 	    !range->eos_marker && err == 0 && !smt_arg->cancel;
1555 	    range = find_next_range(front_ranges, queues, &mask)) {
1556 		/*
1557 		 * If the range in question was in both the from redact bookmark
1558 		 * and the bookmark we're using to redact, then don't send it.
1559 		 * It's already redacted on the receiving system, so a redaction
1560 		 * record would be redundant.
1561 		 */
1562 		if ((mask & FROM_AND_REDACT_BITS) == FROM_AND_REDACT_BITS) {
1563 			ASSERT3U(range->type, ==, REDACT);
1564 			range_free(range);
1565 			continue;
1566 		}
1567 		bqueue_enqueue(&smt_arg->q, range, sizeof (*range));
1568 
1569 		if (smt_arg->to_arg->error_code != 0) {
1570 			err = smt_arg->to_arg->error_code;
1571 		} else if (smt_arg->from_arg->error_code != 0) {
1572 			err = smt_arg->from_arg->error_code;
1573 		} else if (smt_arg->redact_arg != NULL &&
1574 		    smt_arg->redact_arg->error_code != 0) {
1575 			err = smt_arg->redact_arg->error_code;
1576 		}
1577 	}
1578 	if (smt_arg->cancel && err == 0)
1579 		err = SET_ERROR(EINTR);
1580 	smt_arg->error = err;
1581 	if (smt_arg->error != 0) {
1582 		smt_arg->to_arg->cancel = B_TRUE;
1583 		smt_arg->from_arg->cancel = B_TRUE;
1584 		if (smt_arg->redact_arg != NULL)
1585 			smt_arg->redact_arg->cancel = B_TRUE;
1586 	}
1587 	for (int i = 0; i < NUM_THREADS; i++) {
1588 		while (!front_ranges[i]->eos_marker) {
1589 			front_ranges[i] = get_next_range(queues[i],
1590 			    front_ranges[i]);
1591 		}
1592 		range_free(front_ranges[i]);
1593 	}
1594 	range->eos_marker = B_TRUE;
1595 	bqueue_enqueue_flush(&smt_arg->q, range, 1);
1596 	spl_fstrans_unmark(cookie);
1597 	thread_exit();
1598 }
1599 
1600 struct send_reader_thread_arg {
1601 	struct send_merge_thread_arg *smta;
1602 	bqueue_t q;
1603 	boolean_t cancel;
1604 	boolean_t issue_reads;
1605 	uint64_t featureflags;
1606 	int error;
1607 };
1608 
1609 static void
1610 dmu_send_read_done(zio_t *zio)
1611 {
1612 	struct send_range *range = zio->io_private;
1613 
1614 	mutex_enter(&range->sru.data.lock);
1615 	if (zio->io_error != 0) {
1616 		abd_free(range->sru.data.abd);
1617 		range->sru.data.abd = NULL;
1618 		range->sru.data.io_err = zio->io_error;
1619 	}
1620 
1621 	ASSERT(range->sru.data.io_outstanding);
1622 	range->sru.data.io_outstanding = B_FALSE;
1623 	cv_broadcast(&range->sru.data.cv);
1624 	mutex_exit(&range->sru.data.lock);
1625 }
1626 
1627 static void
1628 issue_data_read(struct send_reader_thread_arg *srta, struct send_range *range)
1629 {
1630 	struct srd *srdp = &range->sru.data;
1631 	blkptr_t *bp = &srdp->bp;
1632 	objset_t *os = srta->smta->os;
1633 
1634 	ASSERT3U(range->type, ==, DATA);
1635 	ASSERT3U(range->start_blkid + 1, ==, range->end_blkid);
1636 	/*
1637 	 * If we have large blocks stored on disk but
1638 	 * the send flags don't allow us to send large
1639 	 * blocks, we split the data from the arc buf
1640 	 * into chunks.
1641 	 */
1642 	boolean_t split_large_blocks =
1643 	    srdp->datablksz > SPA_OLD_MAXBLOCKSIZE &&
1644 	    !(srta->featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS);
1645 	/*
1646 	 * We should only request compressed data from the ARC if all
1647 	 * the following are true:
1648 	 *  - stream compression was requested
1649 	 *  - we aren't splitting large blocks into smaller chunks
1650 	 *  - the data won't need to be byteswapped before sending
1651 	 *  - this isn't an embedded block
1652 	 *  - this isn't metadata (if receiving on a different endian
1653 	 *    system it can be byteswapped more easily)
1654 	 */
1655 	boolean_t request_compressed =
1656 	    (srta->featureflags & DMU_BACKUP_FEATURE_COMPRESSED) &&
1657 	    !split_large_blocks && !BP_SHOULD_BYTESWAP(bp) &&
1658 	    !BP_IS_EMBEDDED(bp) && !DMU_OT_IS_METADATA(BP_GET_TYPE(bp));
1659 
1660 	zio_flag_t zioflags = ZIO_FLAG_CANFAIL;
1661 
1662 	if (srta->featureflags & DMU_BACKUP_FEATURE_RAW) {
1663 		zioflags |= ZIO_FLAG_RAW;
1664 		srdp->io_compressed = B_TRUE;
1665 	} else if (request_compressed) {
1666 		zioflags |= ZIO_FLAG_RAW_COMPRESS;
1667 		srdp->io_compressed = B_TRUE;
1668 	}
1669 
1670 	srdp->datasz = (zioflags & ZIO_FLAG_RAW_COMPRESS) ?
1671 	    BP_GET_PSIZE(bp) : BP_GET_LSIZE(bp);
1672 
1673 	if (!srta->issue_reads)
1674 		return;
1675 	if (BP_IS_REDACTED(bp))
1676 		return;
1677 	if (send_do_embed(bp, srta->featureflags))
1678 		return;
1679 
1680 	zbookmark_phys_t zb = {
1681 	    .zb_objset = dmu_objset_id(os),
1682 	    .zb_object = range->object,
1683 	    .zb_level = 0,
1684 	    .zb_blkid = range->start_blkid,
1685 	};
1686 
1687 	arc_flags_t aflags = ARC_FLAG_CACHED_ONLY;
1688 
1689 	int arc_err = arc_read(NULL, os->os_spa, bp,
1690 	    arc_getbuf_func, &srdp->abuf, ZIO_PRIORITY_ASYNC_READ,
1691 	    zioflags, &aflags, &zb);
1692 	/*
1693 	 * If the data is not already cached in the ARC, we read directly
1694 	 * from zio.  This avoids the performance overhead of adding a new
1695 	 * entry to the ARC, and we also avoid polluting the ARC cache with
1696 	 * data that is not likely to be used in the future.
1697 	 */
1698 	if (arc_err != 0) {
1699 		srdp->abd = abd_alloc_linear(srdp->datasz, B_FALSE);
1700 		srdp->io_outstanding = B_TRUE;
1701 		zio_nowait(zio_read(NULL, os->os_spa, bp, srdp->abd,
1702 		    srdp->datasz, dmu_send_read_done, range,
1703 		    ZIO_PRIORITY_ASYNC_READ, zioflags, &zb));
1704 	}
1705 }
1706 
1707 /*
1708  * Create a new record with the given values.
1709  */
1710 static void
1711 enqueue_range(struct send_reader_thread_arg *srta, bqueue_t *q, dnode_t *dn,
1712     uint64_t blkid, uint64_t count, const blkptr_t *bp, uint32_t datablksz)
1713 {
1714 	enum type range_type = (bp == NULL || BP_IS_HOLE(bp) ? HOLE :
1715 	    (BP_IS_REDACTED(bp) ? REDACT : DATA));
1716 
1717 	struct send_range *range = range_alloc(range_type, dn->dn_object,
1718 	    blkid, blkid + count, B_FALSE);
1719 
1720 	if (blkid == DMU_SPILL_BLKID) {
1721 		ASSERT3P(bp, !=, NULL);
1722 		ASSERT3U(BP_GET_TYPE(bp), ==, DMU_OT_SA);
1723 	}
1724 
1725 	switch (range_type) {
1726 	case HOLE:
1727 		range->sru.hole.datablksz = datablksz;
1728 		break;
1729 	case DATA:
1730 		ASSERT3U(count, ==, 1);
1731 		range->sru.data.datablksz = datablksz;
1732 		range->sru.data.obj_type = dn->dn_type;
1733 		range->sru.data.bp = *bp;
1734 		issue_data_read(srta, range);
1735 		break;
1736 	case REDACT:
1737 		range->sru.redact.datablksz = datablksz;
1738 		break;
1739 	default:
1740 		break;
1741 	}
1742 	bqueue_enqueue(q, range, datablksz);
1743 }
1744 
1745 /*
1746  * This thread is responsible for two things: First, it retrieves the correct
1747  * blkptr in the to ds if we need to send the data because of something from
1748  * the from thread.  As a result of this, we're the first ones to discover that
1749  * some indirect blocks can be discarded because they're not holes. Second,
1750  * it issues prefetches for the data we need to send.
1751  */
1752 static __attribute__((noreturn)) void
1753 send_reader_thread(void *arg)
1754 {
1755 	struct send_reader_thread_arg *srta = arg;
1756 	struct send_merge_thread_arg *smta = srta->smta;
1757 	bqueue_t *inq = &smta->q;
1758 	bqueue_t *outq = &srta->q;
1759 	objset_t *os = smta->os;
1760 	fstrans_cookie_t cookie = spl_fstrans_mark();
1761 	struct send_range *range = bqueue_dequeue(inq);
1762 	int err = 0;
1763 
1764 	/*
1765 	 * If the record we're analyzing is from a redaction bookmark from the
1766 	 * fromds, then we need to know whether or not it exists in the tods so
1767 	 * we know whether to create records for it or not. If it does, we need
1768 	 * the datablksz so we can generate an appropriate record for it.
1769 	 * Finally, if it isn't redacted, we need the blkptr so that we can send
1770 	 * a WRITE record containing the actual data.
1771 	 */
1772 	uint64_t last_obj = UINT64_MAX;
1773 	uint64_t last_obj_exists = B_TRUE;
1774 	while (!range->eos_marker && !srta->cancel && smta->error == 0 &&
1775 	    err == 0) {
1776 		switch (range->type) {
1777 		case DATA:
1778 			issue_data_read(srta, range);
1779 			bqueue_enqueue(outq, range, range->sru.data.datablksz);
1780 			range = get_next_range_nofree(inq, range);
1781 			break;
1782 		case HOLE:
1783 		case OBJECT:
1784 		case OBJECT_RANGE:
1785 		case REDACT: // Redacted blocks must exist
1786 			bqueue_enqueue(outq, range, sizeof (*range));
1787 			range = get_next_range_nofree(inq, range);
1788 			break;
1789 		case PREVIOUSLY_REDACTED: {
1790 			/*
1791 			 * This entry came from the "from bookmark" when
1792 			 * sending from a bookmark that has a redaction
1793 			 * list.  We need to check if this object/blkid
1794 			 * exists in the target ("to") dataset, and if
1795 			 * not then we drop this entry.  We also need
1796 			 * to fill in the block pointer so that we know
1797 			 * what to prefetch.
1798 			 *
1799 			 * To accomplish the above, we first cache whether or
1800 			 * not the last object we examined exists.  If it
1801 			 * doesn't, we can drop this record. If it does, we hold
1802 			 * the dnode and use it to call dbuf_dnode_findbp. We do
1803 			 * this instead of dbuf_bookmark_findbp because we will
1804 			 * often operate on large ranges, and holding the dnode
1805 			 * once is more efficient.
1806 			 */
1807 			boolean_t object_exists = B_TRUE;
1808 			/*
1809 			 * If the data is redacted, we only care if it exists,
1810 			 * so that we don't send records for objects that have
1811 			 * been deleted.
1812 			 */
1813 			dnode_t *dn;
1814 			if (range->object == last_obj && !last_obj_exists) {
1815 				/*
1816 				 * If we're still examining the same object as
1817 				 * previously, and it doesn't exist, we don't
1818 				 * need to call dbuf_bookmark_findbp.
1819 				 */
1820 				object_exists = B_FALSE;
1821 			} else {
1822 				err = dnode_hold(os, range->object, FTAG, &dn);
1823 				if (err == ENOENT) {
1824 					object_exists = B_FALSE;
1825 					err = 0;
1826 				}
1827 				last_obj = range->object;
1828 				last_obj_exists = object_exists;
1829 			}
1830 
1831 			if (err != 0) {
1832 				break;
1833 			} else if (!object_exists) {
1834 				/*
1835 				 * The block was modified, but doesn't
1836 				 * exist in the to dataset; if it was
1837 				 * deleted in the to dataset, then we'll
1838 				 * visit the hole bp for it at some point.
1839 				 */
1840 				range = get_next_range(inq, range);
1841 				continue;
1842 			}
1843 			uint64_t file_max =
1844 			    MIN(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 
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 
1966 	/* raw sends imply large_block_ok */
1967 	if ((dspp->rawok || dspp->large_block_ok) &&
1968 	    dsl_dataset_feature_is_active(to_ds, SPA_FEATURE_LARGE_BLOCKS)) {
1969 		*featureflags |= DMU_BACKUP_FEATURE_LARGE_BLOCKS;
1970 	}
1971 
1972 	/* encrypted datasets will not have embedded blocks */
1973 	if ((dspp->embedok || dspp->rawok) && !os->os_encrypted &&
1974 	    spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) {
1975 		*featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA;
1976 	}
1977 
1978 	/* raw send implies compressok */
1979 	if (dspp->compressok || dspp->rawok)
1980 		*featureflags |= DMU_BACKUP_FEATURE_COMPRESSED;
1981 
1982 	if (dspp->rawok && os->os_encrypted)
1983 		*featureflags |= DMU_BACKUP_FEATURE_RAW;
1984 
1985 	if ((*featureflags &
1986 	    (DMU_BACKUP_FEATURE_EMBED_DATA | DMU_BACKUP_FEATURE_COMPRESSED |
1987 	    DMU_BACKUP_FEATURE_RAW)) != 0 &&
1988 	    spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS)) {
1989 		*featureflags |= DMU_BACKUP_FEATURE_LZ4;
1990 	}
1991 
1992 	/*
1993 	 * We specifically do not include DMU_BACKUP_FEATURE_EMBED_DATA here to
1994 	 * allow sending ZSTD compressed datasets to a receiver that does not
1995 	 * support ZSTD
1996 	 */
1997 	if ((*featureflags &
1998 	    (DMU_BACKUP_FEATURE_COMPRESSED | DMU_BACKUP_FEATURE_RAW)) != 0 &&
1999 	    dsl_dataset_feature_is_active(to_ds, SPA_FEATURE_ZSTD_COMPRESS)) {
2000 		*featureflags |= DMU_BACKUP_FEATURE_ZSTD;
2001 	}
2002 
2003 	if (dspp->resumeobj != 0 || dspp->resumeoff != 0) {
2004 		*featureflags |= DMU_BACKUP_FEATURE_RESUMING;
2005 	}
2006 
2007 	if (dspp->redactbook != NULL) {
2008 		*featureflags |= DMU_BACKUP_FEATURE_REDACTED;
2009 	}
2010 
2011 	if (dsl_dataset_feature_is_active(to_ds, SPA_FEATURE_LARGE_DNODE)) {
2012 		*featureflags |= DMU_BACKUP_FEATURE_LARGE_DNODE;
2013 	}
2014 	return (0);
2015 }
2016 
2017 static dmu_replay_record_t *
2018 create_begin_record(struct dmu_send_params *dspp, objset_t *os,
2019     uint64_t featureflags)
2020 {
2021 	dmu_replay_record_t *drr = kmem_zalloc(sizeof (dmu_replay_record_t),
2022 	    KM_SLEEP);
2023 	drr->drr_type = DRR_BEGIN;
2024 
2025 	struct drr_begin *drrb = &drr->drr_u.drr_begin;
2026 	dsl_dataset_t *to_ds = dspp->to_ds;
2027 
2028 	drrb->drr_magic = DMU_BACKUP_MAGIC;
2029 	drrb->drr_creation_time = dsl_dataset_phys(to_ds)->ds_creation_time;
2030 	drrb->drr_type = dmu_objset_type(os);
2031 	drrb->drr_toguid = dsl_dataset_phys(to_ds)->ds_guid;
2032 	drrb->drr_fromguid = dspp->ancestor_zb.zbm_guid;
2033 
2034 	DMU_SET_STREAM_HDRTYPE(drrb->drr_versioninfo, DMU_SUBSTREAM);
2035 	DMU_SET_FEATUREFLAGS(drrb->drr_versioninfo, featureflags);
2036 
2037 	if (dspp->is_clone)
2038 		drrb->drr_flags |= DRR_FLAG_CLONE;
2039 	if (dsl_dataset_phys(dspp->to_ds)->ds_flags & DS_FLAG_CI_DATASET)
2040 		drrb->drr_flags |= DRR_FLAG_CI_DATA;
2041 	if (zfs_send_set_freerecords_bit)
2042 		drrb->drr_flags |= DRR_FLAG_FREERECORDS;
2043 	drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_SPILL_BLOCK;
2044 
2045 	if (dspp->savedok) {
2046 		drrb->drr_toguid = dspp->saved_guid;
2047 		strlcpy(drrb->drr_toname, dspp->saved_toname,
2048 		    sizeof (drrb->drr_toname));
2049 	} else {
2050 		dsl_dataset_name(to_ds, drrb->drr_toname);
2051 		if (!to_ds->ds_is_snapshot) {
2052 			(void) strlcat(drrb->drr_toname, "@--head--",
2053 			    sizeof (drrb->drr_toname));
2054 		}
2055 	}
2056 	return (drr);
2057 }
2058 
2059 static void
2060 setup_to_thread(struct send_thread_arg *to_arg, objset_t *to_os,
2061     dmu_sendstatus_t *dssp, uint64_t fromtxg, boolean_t rawok)
2062 {
2063 	VERIFY0(bqueue_init(&to_arg->q, zfs_send_no_prefetch_queue_ff,
2064 	    MAX(zfs_send_no_prefetch_queue_length, 2 * zfs_max_recordsize),
2065 	    offsetof(struct send_range, ln)));
2066 	to_arg->error_code = 0;
2067 	to_arg->cancel = B_FALSE;
2068 	to_arg->os = to_os;
2069 	to_arg->fromtxg = fromtxg;
2070 	to_arg->flags = TRAVERSE_PRE | TRAVERSE_PREFETCH_METADATA;
2071 	if (rawok)
2072 		to_arg->flags |= TRAVERSE_NO_DECRYPT;
2073 	if (zfs_send_corrupt_data)
2074 		to_arg->flags |= TRAVERSE_HARD;
2075 	to_arg->num_blocks_visited = &dssp->dss_blocks;
2076 	(void) thread_create(NULL, 0, send_traverse_thread, to_arg, 0,
2077 	    curproc, TS_RUN, minclsyspri);
2078 }
2079 
2080 static void
2081 setup_from_thread(struct redact_list_thread_arg *from_arg,
2082     redaction_list_t *from_rl, dmu_sendstatus_t *dssp)
2083 {
2084 	VERIFY0(bqueue_init(&from_arg->q, zfs_send_no_prefetch_queue_ff,
2085 	    MAX(zfs_send_no_prefetch_queue_length, 2 * zfs_max_recordsize),
2086 	    offsetof(struct send_range, ln)));
2087 	from_arg->error_code = 0;
2088 	from_arg->cancel = B_FALSE;
2089 	from_arg->rl = from_rl;
2090 	from_arg->mark_redact = B_FALSE;
2091 	from_arg->num_blocks_visited = &dssp->dss_blocks;
2092 	/*
2093 	 * If from_ds is null, send_traverse_thread just returns success and
2094 	 * enqueues an eos marker.
2095 	 */
2096 	(void) thread_create(NULL, 0, redact_list_thread, from_arg, 0,
2097 	    curproc, TS_RUN, minclsyspri);
2098 }
2099 
2100 static void
2101 setup_redact_list_thread(struct redact_list_thread_arg *rlt_arg,
2102     struct dmu_send_params *dspp, redaction_list_t *rl, dmu_sendstatus_t *dssp)
2103 {
2104 	if (dspp->redactbook == NULL)
2105 		return;
2106 
2107 	rlt_arg->cancel = B_FALSE;
2108 	VERIFY0(bqueue_init(&rlt_arg->q, zfs_send_no_prefetch_queue_ff,
2109 	    MAX(zfs_send_no_prefetch_queue_length, 2 * zfs_max_recordsize),
2110 	    offsetof(struct send_range, ln)));
2111 	rlt_arg->error_code = 0;
2112 	rlt_arg->mark_redact = B_TRUE;
2113 	rlt_arg->rl = rl;
2114 	rlt_arg->num_blocks_visited = &dssp->dss_blocks;
2115 
2116 	(void) thread_create(NULL, 0, redact_list_thread, rlt_arg, 0,
2117 	    curproc, TS_RUN, minclsyspri);
2118 }
2119 
2120 static void
2121 setup_merge_thread(struct send_merge_thread_arg *smt_arg,
2122     struct dmu_send_params *dspp, struct redact_list_thread_arg *from_arg,
2123     struct send_thread_arg *to_arg, struct redact_list_thread_arg *rlt_arg,
2124     objset_t *os)
2125 {
2126 	VERIFY0(bqueue_init(&smt_arg->q, zfs_send_no_prefetch_queue_ff,
2127 	    MAX(zfs_send_no_prefetch_queue_length, 2 * zfs_max_recordsize),
2128 	    offsetof(struct send_range, ln)));
2129 	smt_arg->cancel = B_FALSE;
2130 	smt_arg->error = 0;
2131 	smt_arg->from_arg = from_arg;
2132 	smt_arg->to_arg = to_arg;
2133 	if (dspp->redactbook != NULL)
2134 		smt_arg->redact_arg = rlt_arg;
2135 
2136 	smt_arg->os = os;
2137 	(void) thread_create(NULL, 0, send_merge_thread, smt_arg, 0, curproc,
2138 	    TS_RUN, minclsyspri);
2139 }
2140 
2141 static void
2142 setup_reader_thread(struct send_reader_thread_arg *srt_arg,
2143     struct dmu_send_params *dspp, struct send_merge_thread_arg *smt_arg,
2144     uint64_t featureflags)
2145 {
2146 	VERIFY0(bqueue_init(&srt_arg->q, zfs_send_queue_ff,
2147 	    MAX(zfs_send_queue_length, 2 * zfs_max_recordsize),
2148 	    offsetof(struct send_range, ln)));
2149 	srt_arg->smta = smt_arg;
2150 	srt_arg->issue_reads = !dspp->dso->dso_dryrun;
2151 	srt_arg->featureflags = featureflags;
2152 	(void) thread_create(NULL, 0, send_reader_thread, srt_arg, 0,
2153 	    curproc, TS_RUN, minclsyspri);
2154 }
2155 
2156 static int
2157 setup_resume_points(struct dmu_send_params *dspp,
2158     struct send_thread_arg *to_arg, struct redact_list_thread_arg *from_arg,
2159     struct redact_list_thread_arg *rlt_arg,
2160     struct send_merge_thread_arg *smt_arg, boolean_t resuming, objset_t *os,
2161     redaction_list_t *redact_rl, nvlist_t *nvl)
2162 {
2163 	(void) smt_arg;
2164 	dsl_dataset_t *to_ds = dspp->to_ds;
2165 	int err = 0;
2166 
2167 	uint64_t obj = 0;
2168 	uint64_t blkid = 0;
2169 	if (resuming) {
2170 		obj = dspp->resumeobj;
2171 		dmu_object_info_t to_doi;
2172 		err = dmu_object_info(os, obj, &to_doi);
2173 		if (err != 0)
2174 			return (err);
2175 
2176 		blkid = dspp->resumeoff / to_doi.doi_data_block_size;
2177 	}
2178 	/*
2179 	 * If we're resuming a redacted send, we can skip to the appropriate
2180 	 * point in the redaction bookmark by binary searching through it.
2181 	 */
2182 	if (redact_rl != NULL) {
2183 		SET_BOOKMARK(&rlt_arg->resume, to_ds->ds_object, obj, 0, blkid);
2184 	}
2185 
2186 	SET_BOOKMARK(&to_arg->resume, to_ds->ds_object, obj, 0, blkid);
2187 	if (nvlist_exists(nvl, BEGINNV_REDACT_FROM_SNAPS)) {
2188 		uint64_t objset = dspp->ancestor_zb.zbm_redaction_obj;
2189 		/*
2190 		 * Note: If the resume point is in an object whose
2191 		 * blocksize is different in the from vs to snapshots,
2192 		 * we will have divided by the "wrong" blocksize.
2193 		 * However, in this case fromsnap's send_cb() will
2194 		 * detect that the blocksize has changed and therefore
2195 		 * ignore this object.
2196 		 *
2197 		 * If we're resuming a send from a redaction bookmark,
2198 		 * we still cannot accidentally suggest blocks behind
2199 		 * the to_ds.  In addition, we know that any blocks in
2200 		 * the object in the to_ds will have to be sent, since
2201 		 * the size changed.  Therefore, we can't cause any harm
2202 		 * this way either.
2203 		 */
2204 		SET_BOOKMARK(&from_arg->resume, objset, obj, 0, blkid);
2205 	}
2206 	if (resuming) {
2207 		fnvlist_add_uint64(nvl, BEGINNV_RESUME_OBJECT, dspp->resumeobj);
2208 		fnvlist_add_uint64(nvl, BEGINNV_RESUME_OFFSET, dspp->resumeoff);
2209 	}
2210 	return (0);
2211 }
2212 
2213 static dmu_sendstatus_t *
2214 setup_send_progress(struct dmu_send_params *dspp)
2215 {
2216 	dmu_sendstatus_t *dssp = kmem_zalloc(sizeof (*dssp), KM_SLEEP);
2217 	dssp->dss_outfd = dspp->outfd;
2218 	dssp->dss_off = dspp->off;
2219 	dssp->dss_proc = curproc;
2220 	mutex_enter(&dspp->to_ds->ds_sendstream_lock);
2221 	list_insert_head(&dspp->to_ds->ds_sendstreams, dssp);
2222 	mutex_exit(&dspp->to_ds->ds_sendstream_lock);
2223 	return (dssp);
2224 }
2225 
2226 /*
2227  * Actually do the bulk of the work in a zfs send.
2228  *
2229  * The idea is that we want to do a send from ancestor_zb to to_ds.  We also
2230  * want to not send any data that has been modified by all the datasets in
2231  * redactsnaparr, and store the list of blocks that are redacted in this way in
2232  * a bookmark named redactbook, created on the to_ds.  We do this by creating
2233  * several worker threads, whose function is described below.
2234  *
2235  * There are three cases.
2236  * The first case is a redacted zfs send.  In this case there are 5 threads.
2237  * The first thread is the to_ds traversal thread: it calls dataset_traverse on
2238  * the to_ds and finds all the blocks that have changed since ancestor_zb (if
2239  * it's a full send, that's all blocks in the dataset).  It then sends those
2240  * blocks on to the send merge thread. The redact list thread takes the data
2241  * from the redaction bookmark and sends those blocks on to the send merge
2242  * thread.  The send merge thread takes the data from the to_ds traversal
2243  * thread, and combines it with the redaction records from the redact list
2244  * thread.  If a block appears in both the to_ds's data and the redaction data,
2245  * the send merge thread will mark it as redacted and send it on to the prefetch
2246  * thread.  Otherwise, the send merge thread will send the block on to the
2247  * prefetch thread unchanged. The prefetch thread will issue prefetch reads for
2248  * any data that isn't redacted, and then send the data on to the main thread.
2249  * The main thread behaves the same as in a normal send case, issuing demand
2250  * reads for data blocks and sending out records over the network
2251  *
2252  * The graphic below diagrams the flow of data in the case of a redacted zfs
2253  * send.  Each box represents a thread, and each line represents the flow of
2254  * data.
2255  *
2256  *             Records from the |
2257  *           redaction bookmark |
2258  * +--------------------+       |  +---------------------------+
2259  * |                    |       v  | Send Merge Thread         |
2260  * | Redact List Thread +----------> Apply redaction marks to  |
2261  * |                    |          | records as specified by   |
2262  * +--------------------+          | redaction ranges          |
2263  *                                 +----^---------------+------+
2264  *                                      |               | Merged data
2265  *                                      |               |
2266  *                                      |  +------------v--------+
2267  *                                      |  | Prefetch Thread     |
2268  * +--------------------+               |  | Issues prefetch     |
2269  * | to_ds Traversal    |               |  | reads of data blocks|
2270  * | Thread (finds      +---------------+  +------------+--------+
2271  * | candidate blocks)  |  Blocks modified              | Prefetched data
2272  * +--------------------+  by to_ds since               |
2273  *                         ancestor_zb     +------------v----+
2274  *                                         | Main Thread     |  File Descriptor
2275  *                                         | Sends data over +->(to zfs receive)
2276  *                                         | wire            |
2277  *                                         +-----------------+
2278  *
2279  * The second case is an incremental send from a redaction bookmark.  The to_ds
2280  * traversal thread and the main thread behave the same as in the redacted
2281  * send case.  The new thread is the from bookmark traversal thread.  It
2282  * iterates over the redaction list in the redaction bookmark, and enqueues
2283  * records for each block that was redacted in the original send.  The send
2284  * merge thread now has to merge the data from the two threads.  For details
2285  * about that process, see the header comment of send_merge_thread().  Any data
2286  * it decides to send on will be prefetched by the prefetch thread.  Note that
2287  * you can perform a redacted send from a redaction bookmark; in that case,
2288  * the data flow behaves very similarly to the flow in the redacted send case,
2289  * except with the addition of the bookmark traversal thread iterating over the
2290  * redaction bookmark.  The send_merge_thread also has to take on the
2291  * responsibility of merging the redact list thread's records, the bookmark
2292  * traversal thread's records, and the to_ds records.
2293  *
2294  * +---------------------+
2295  * |                     |
2296  * | Redact List Thread  +--------------+
2297  * |                     |              |
2298  * +---------------------+              |
2299  *        Blocks in redaction list      | Ranges modified by every secure snap
2300  *        of from bookmark              | (or EOS if not readcted)
2301  *                                      |
2302  * +---------------------+   |     +----v----------------------+
2303  * | bookmark Traversal  |   v     | Send Merge Thread         |
2304  * | Thread (finds       +---------> Merges bookmark, rlt, and |
2305  * | candidate blocks)   |         | to_ds send records        |
2306  * +---------------------+         +----^---------------+------+
2307  *                                      |               | Merged data
2308  *                                      |  +------------v--------+
2309  *                                      |  | Prefetch Thread     |
2310  * +--------------------+               |  | Issues prefetch     |
2311  * | to_ds Traversal    |               |  | reads of data blocks|
2312  * | Thread (finds      +---------------+  +------------+--------+
2313  * | candidate blocks)  |  Blocks modified              | Prefetched data
2314  * +--------------------+  by to_ds since  +------------v----+
2315  *                         ancestor_zb     | Main Thread     |  File Descriptor
2316  *                                         | Sends data over +->(to zfs receive)
2317  *                                         | wire            |
2318  *                                         +-----------------+
2319  *
2320  * The final case is a simple zfs full or incremental send.  The to_ds traversal
2321  * thread behaves the same as always. The redact list thread is never started.
2322  * The send merge thread takes all the blocks that the to_ds traversal thread
2323  * sends it, prefetches the data, and sends the blocks on to the main thread.
2324  * The main thread sends the data over the wire.
2325  *
2326  * To keep performance acceptable, we want to prefetch the data in the worker
2327  * threads.  While the to_ds thread could simply use the TRAVERSE_PREFETCH
2328  * feature built into traverse_dataset, the combining and deletion of records
2329  * due to redaction and sends from redaction bookmarks mean that we could
2330  * issue many unnecessary prefetches.  As a result, we only prefetch data
2331  * after we've determined that the record is not going to be redacted.  To
2332  * prevent the prefetching from getting too far ahead of the main thread, the
2333  * blocking queues that are used for communication are capped not by the
2334  * number of entries in the queue, but by the sum of the size of the
2335  * prefetches associated with them.  The limit on the amount of data that the
2336  * thread can prefetch beyond what the main thread has reached is controlled
2337  * by the global variable zfs_send_queue_length.  In addition, to prevent poor
2338  * performance in the beginning of a send, we also limit the distance ahead
2339  * that the traversal threads can be.  That distance is controlled by the
2340  * zfs_send_no_prefetch_queue_length tunable.
2341  *
2342  * Note: Releases dp using the specified tag.
2343  */
2344 static int
2345 dmu_send_impl(struct dmu_send_params *dspp)
2346 {
2347 	objset_t *os;
2348 	dmu_replay_record_t *drr;
2349 	dmu_sendstatus_t *dssp;
2350 	dmu_send_cookie_t dsc = {0};
2351 	int err;
2352 	uint64_t fromtxg = dspp->ancestor_zb.zbm_creation_txg;
2353 	uint64_t featureflags = 0;
2354 	struct redact_list_thread_arg *from_arg;
2355 	struct send_thread_arg *to_arg;
2356 	struct redact_list_thread_arg *rlt_arg;
2357 	struct send_merge_thread_arg *smt_arg;
2358 	struct send_reader_thread_arg *srt_arg;
2359 	struct send_range *range;
2360 	redaction_list_t *from_rl = NULL;
2361 	redaction_list_t *redact_rl = NULL;
2362 	boolean_t resuming = (dspp->resumeobj != 0 || dspp->resumeoff != 0);
2363 	boolean_t book_resuming = resuming;
2364 
2365 	dsl_dataset_t *to_ds = dspp->to_ds;
2366 	zfs_bookmark_phys_t *ancestor_zb = &dspp->ancestor_zb;
2367 	dsl_pool_t *dp = dspp->dp;
2368 	const void *tag = dspp->tag;
2369 
2370 	err = dmu_objset_from_ds(to_ds, &os);
2371 	if (err != 0) {
2372 		dsl_pool_rele(dp, tag);
2373 		return (err);
2374 	}
2375 
2376 	/*
2377 	 * If this is a non-raw send of an encrypted ds, we can ensure that
2378 	 * the objset_phys_t is authenticated. This is safe because this is
2379 	 * either a snapshot or we have owned the dataset, ensuring that
2380 	 * it can't be modified.
2381 	 */
2382 	if (!dspp->rawok && os->os_encrypted &&
2383 	    arc_is_unauthenticated(os->os_phys_buf)) {
2384 		zbookmark_phys_t zb;
2385 
2386 		SET_BOOKMARK(&zb, to_ds->ds_object, ZB_ROOT_OBJECT,
2387 		    ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
2388 		err = arc_untransform(os->os_phys_buf, os->os_spa,
2389 		    &zb, B_FALSE);
2390 		if (err != 0) {
2391 			dsl_pool_rele(dp, tag);
2392 			return (err);
2393 		}
2394 
2395 		ASSERT0(arc_is_unauthenticated(os->os_phys_buf));
2396 	}
2397 
2398 	if ((err = setup_featureflags(dspp, os, &featureflags)) != 0) {
2399 		dsl_pool_rele(dp, tag);
2400 		return (err);
2401 	}
2402 
2403 	/*
2404 	 * If we're doing a redacted send, hold the bookmark's redaction list.
2405 	 */
2406 	if (dspp->redactbook != NULL) {
2407 		err = dsl_redaction_list_hold_obj(dp,
2408 		    dspp->redactbook->zbm_redaction_obj, FTAG,
2409 		    &redact_rl);
2410 		if (err != 0) {
2411 			dsl_pool_rele(dp, tag);
2412 			return (SET_ERROR(EINVAL));
2413 		}
2414 		dsl_redaction_list_long_hold(dp, redact_rl, FTAG);
2415 	}
2416 
2417 	/*
2418 	 * If we're sending from a redaction bookmark, hold the redaction list
2419 	 * so that we can consider sending the redacted blocks.
2420 	 */
2421 	if (ancestor_zb->zbm_redaction_obj != 0) {
2422 		err = dsl_redaction_list_hold_obj(dp,
2423 		    ancestor_zb->zbm_redaction_obj, FTAG, &from_rl);
2424 		if (err != 0) {
2425 			if (redact_rl != NULL) {
2426 				dsl_redaction_list_long_rele(redact_rl, FTAG);
2427 				dsl_redaction_list_rele(redact_rl, FTAG);
2428 			}
2429 			dsl_pool_rele(dp, tag);
2430 			return (SET_ERROR(EINVAL));
2431 		}
2432 		dsl_redaction_list_long_hold(dp, from_rl, FTAG);
2433 	}
2434 
2435 	dsl_dataset_long_hold(to_ds, FTAG);
2436 
2437 	from_arg = kmem_zalloc(sizeof (*from_arg), KM_SLEEP);
2438 	to_arg = kmem_zalloc(sizeof (*to_arg), KM_SLEEP);
2439 	rlt_arg = kmem_zalloc(sizeof (*rlt_arg), KM_SLEEP);
2440 	smt_arg = kmem_zalloc(sizeof (*smt_arg), KM_SLEEP);
2441 	srt_arg = kmem_zalloc(sizeof (*srt_arg), KM_SLEEP);
2442 
2443 	drr = create_begin_record(dspp, os, featureflags);
2444 	dssp = setup_send_progress(dspp);
2445 
2446 	dsc.dsc_drr = drr;
2447 	dsc.dsc_dso = dspp->dso;
2448 	dsc.dsc_os = os;
2449 	dsc.dsc_off = dspp->off;
2450 	dsc.dsc_toguid = dsl_dataset_phys(to_ds)->ds_guid;
2451 	dsc.dsc_fromtxg = fromtxg;
2452 	dsc.dsc_pending_op = PENDING_NONE;
2453 	dsc.dsc_featureflags = featureflags;
2454 	dsc.dsc_resume_object = dspp->resumeobj;
2455 	dsc.dsc_resume_offset = dspp->resumeoff;
2456 
2457 	dsl_pool_rele(dp, tag);
2458 
2459 	void *payload = NULL;
2460 	size_t payload_len = 0;
2461 	nvlist_t *nvl = fnvlist_alloc();
2462 
2463 	/*
2464 	 * If we're doing a redacted send, we include the snapshots we're
2465 	 * redacted with respect to so that the target system knows what send
2466 	 * streams can be correctly received on top of this dataset. If we're
2467 	 * instead sending a redacted dataset, we include the snapshots that the
2468 	 * dataset was created with respect to.
2469 	 */
2470 	if (dspp->redactbook != NULL) {
2471 		fnvlist_add_uint64_array(nvl, BEGINNV_REDACT_SNAPS,
2472 		    redact_rl->rl_phys->rlp_snaps,
2473 		    redact_rl->rl_phys->rlp_num_snaps);
2474 	} else if (dsl_dataset_feature_is_active(to_ds,
2475 	    SPA_FEATURE_REDACTED_DATASETS)) {
2476 		uint64_t *tods_guids;
2477 		uint64_t length;
2478 		VERIFY(dsl_dataset_get_uint64_array_feature(to_ds,
2479 		    SPA_FEATURE_REDACTED_DATASETS, &length, &tods_guids));
2480 		fnvlist_add_uint64_array(nvl, BEGINNV_REDACT_SNAPS, tods_guids,
2481 		    length);
2482 	}
2483 
2484 	/*
2485 	 * If we're sending from a redaction bookmark, then we should retrieve
2486 	 * the guids of that bookmark so we can send them over the wire.
2487 	 */
2488 	if (from_rl != NULL) {
2489 		fnvlist_add_uint64_array(nvl, BEGINNV_REDACT_FROM_SNAPS,
2490 		    from_rl->rl_phys->rlp_snaps,
2491 		    from_rl->rl_phys->rlp_num_snaps);
2492 	}
2493 
2494 	/*
2495 	 * If the snapshot we're sending from is redacted, include the redaction
2496 	 * list in the stream.
2497 	 */
2498 	if (dspp->numfromredactsnaps != NUM_SNAPS_NOT_REDACTED) {
2499 		ASSERT3P(from_rl, ==, NULL);
2500 		fnvlist_add_uint64_array(nvl, BEGINNV_REDACT_FROM_SNAPS,
2501 		    dspp->fromredactsnaps, (uint_t)dspp->numfromredactsnaps);
2502 		if (dspp->numfromredactsnaps > 0) {
2503 			kmem_free(dspp->fromredactsnaps,
2504 			    dspp->numfromredactsnaps * sizeof (uint64_t));
2505 			dspp->fromredactsnaps = NULL;
2506 		}
2507 	}
2508 
2509 	if (resuming || book_resuming) {
2510 		err = setup_resume_points(dspp, to_arg, from_arg,
2511 		    rlt_arg, smt_arg, resuming, os, redact_rl, nvl);
2512 		if (err != 0)
2513 			goto out;
2514 	}
2515 
2516 	if (featureflags & DMU_BACKUP_FEATURE_RAW) {
2517 		uint64_t ivset_guid = ancestor_zb->zbm_ivset_guid;
2518 		nvlist_t *keynvl = NULL;
2519 		ASSERT(os->os_encrypted);
2520 
2521 		err = dsl_crypto_populate_key_nvlist(os, ivset_guid,
2522 		    &keynvl);
2523 		if (err != 0) {
2524 			fnvlist_free(nvl);
2525 			goto out;
2526 		}
2527 
2528 		fnvlist_add_nvlist(nvl, "crypt_keydata", keynvl);
2529 		fnvlist_free(keynvl);
2530 	}
2531 
2532 	if (!nvlist_empty(nvl)) {
2533 		payload = fnvlist_pack(nvl, &payload_len);
2534 		drr->drr_payloadlen = payload_len;
2535 	}
2536 
2537 	fnvlist_free(nvl);
2538 	err = dump_record(&dsc, payload, payload_len);
2539 	fnvlist_pack_free(payload, payload_len);
2540 	if (err != 0) {
2541 		err = dsc.dsc_err;
2542 		goto out;
2543 	}
2544 
2545 	setup_to_thread(to_arg, os, dssp, fromtxg, dspp->rawok);
2546 	setup_from_thread(from_arg, from_rl, dssp);
2547 	setup_redact_list_thread(rlt_arg, dspp, redact_rl, dssp);
2548 	setup_merge_thread(smt_arg, dspp, from_arg, to_arg, rlt_arg, os);
2549 	setup_reader_thread(srt_arg, dspp, smt_arg, featureflags);
2550 
2551 	range = bqueue_dequeue(&srt_arg->q);
2552 	while (err == 0 && !range->eos_marker) {
2553 		err = do_dump(&dsc, range);
2554 		range = get_next_range(&srt_arg->q, range);
2555 		if (issig(JUSTLOOKING) && issig(FORREAL))
2556 			err = SET_ERROR(EINTR);
2557 	}
2558 
2559 	/*
2560 	 * If we hit an error or are interrupted, cancel our worker threads and
2561 	 * clear the queue of any pending records.  The threads will pass the
2562 	 * cancel up the tree of worker threads, and each one will clean up any
2563 	 * pending records before exiting.
2564 	 */
2565 	if (err != 0) {
2566 		srt_arg->cancel = B_TRUE;
2567 		while (!range->eos_marker) {
2568 			range = get_next_range(&srt_arg->q, range);
2569 		}
2570 	}
2571 	range_free(range);
2572 
2573 	bqueue_destroy(&srt_arg->q);
2574 	bqueue_destroy(&smt_arg->q);
2575 	if (dspp->redactbook != NULL)
2576 		bqueue_destroy(&rlt_arg->q);
2577 	bqueue_destroy(&to_arg->q);
2578 	bqueue_destroy(&from_arg->q);
2579 
2580 	if (err == 0 && srt_arg->error != 0)
2581 		err = srt_arg->error;
2582 
2583 	if (err != 0)
2584 		goto out;
2585 
2586 	if (dsc.dsc_pending_op != PENDING_NONE)
2587 		if (dump_record(&dsc, NULL, 0) != 0)
2588 			err = SET_ERROR(EINTR);
2589 
2590 	if (err != 0) {
2591 		if (err == EINTR && dsc.dsc_err != 0)
2592 			err = dsc.dsc_err;
2593 		goto out;
2594 	}
2595 
2596 	/*
2597 	 * Send the DRR_END record if this is not a saved stream.
2598 	 * Otherwise, the omitted DRR_END record will signal to
2599 	 * the receive side that the stream is incomplete.
2600 	 */
2601 	if (!dspp->savedok) {
2602 		memset(drr, 0, sizeof (dmu_replay_record_t));
2603 		drr->drr_type = DRR_END;
2604 		drr->drr_u.drr_end.drr_checksum = dsc.dsc_zc;
2605 		drr->drr_u.drr_end.drr_toguid = dsc.dsc_toguid;
2606 
2607 		if (dump_record(&dsc, NULL, 0) != 0)
2608 			err = dsc.dsc_err;
2609 	}
2610 out:
2611 	mutex_enter(&to_ds->ds_sendstream_lock);
2612 	list_remove(&to_ds->ds_sendstreams, dssp);
2613 	mutex_exit(&to_ds->ds_sendstream_lock);
2614 
2615 	VERIFY(err != 0 || (dsc.dsc_sent_begin &&
2616 	    (dsc.dsc_sent_end || dspp->savedok)));
2617 
2618 	kmem_free(drr, sizeof (dmu_replay_record_t));
2619 	kmem_free(dssp, sizeof (dmu_sendstatus_t));
2620 	kmem_free(from_arg, sizeof (*from_arg));
2621 	kmem_free(to_arg, sizeof (*to_arg));
2622 	kmem_free(rlt_arg, sizeof (*rlt_arg));
2623 	kmem_free(smt_arg, sizeof (*smt_arg));
2624 	kmem_free(srt_arg, sizeof (*srt_arg));
2625 
2626 	dsl_dataset_long_rele(to_ds, FTAG);
2627 	if (from_rl != NULL) {
2628 		dsl_redaction_list_long_rele(from_rl, FTAG);
2629 		dsl_redaction_list_rele(from_rl, FTAG);
2630 	}
2631 	if (redact_rl != NULL) {
2632 		dsl_redaction_list_long_rele(redact_rl, FTAG);
2633 		dsl_redaction_list_rele(redact_rl, FTAG);
2634 	}
2635 
2636 	return (err);
2637 }
2638 
2639 int
2640 dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
2641     boolean_t embedok, boolean_t large_block_ok, boolean_t compressok,
2642     boolean_t rawok, boolean_t savedok, int outfd, offset_t *off,
2643     dmu_send_outparams_t *dsop)
2644 {
2645 	int err;
2646 	dsl_dataset_t *fromds;
2647 	ds_hold_flags_t dsflags;
2648 	struct dmu_send_params dspp = {0};
2649 	dspp.embedok = embedok;
2650 	dspp.large_block_ok = large_block_ok;
2651 	dspp.compressok = compressok;
2652 	dspp.outfd = outfd;
2653 	dspp.off = off;
2654 	dspp.dso = dsop;
2655 	dspp.tag = FTAG;
2656 	dspp.rawok = rawok;
2657 	dspp.savedok = savedok;
2658 
2659 	dsflags = (rawok) ? DS_HOLD_FLAG_NONE : DS_HOLD_FLAG_DECRYPT;
2660 	err = dsl_pool_hold(pool, FTAG, &dspp.dp);
2661 	if (err != 0)
2662 		return (err);
2663 
2664 	err = dsl_dataset_hold_obj_flags(dspp.dp, tosnap, dsflags, FTAG,
2665 	    &dspp.to_ds);
2666 	if (err != 0) {
2667 		dsl_pool_rele(dspp.dp, FTAG);
2668 		return (err);
2669 	}
2670 
2671 	if (fromsnap != 0) {
2672 		err = dsl_dataset_hold_obj_flags(dspp.dp, fromsnap, dsflags,
2673 		    FTAG, &fromds);
2674 		if (err != 0) {
2675 			dsl_dataset_rele_flags(dspp.to_ds, dsflags, FTAG);
2676 			dsl_pool_rele(dspp.dp, FTAG);
2677 			return (err);
2678 		}
2679 		dspp.ancestor_zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
2680 		dspp.ancestor_zb.zbm_creation_txg =
2681 		    dsl_dataset_phys(fromds)->ds_creation_txg;
2682 		dspp.ancestor_zb.zbm_creation_time =
2683 		    dsl_dataset_phys(fromds)->ds_creation_time;
2684 
2685 		if (dsl_dataset_is_zapified(fromds)) {
2686 			(void) zap_lookup(dspp.dp->dp_meta_objset,
2687 			    fromds->ds_object, DS_FIELD_IVSET_GUID, 8, 1,
2688 			    &dspp.ancestor_zb.zbm_ivset_guid);
2689 		}
2690 
2691 		/* See dmu_send for the reasons behind this. */
2692 		uint64_t *fromredact;
2693 
2694 		if (!dsl_dataset_get_uint64_array_feature(fromds,
2695 		    SPA_FEATURE_REDACTED_DATASETS,
2696 		    &dspp.numfromredactsnaps,
2697 		    &fromredact)) {
2698 			dspp.numfromredactsnaps = NUM_SNAPS_NOT_REDACTED;
2699 		} else if (dspp.numfromredactsnaps > 0) {
2700 			uint64_t size = dspp.numfromredactsnaps *
2701 			    sizeof (uint64_t);
2702 			dspp.fromredactsnaps = kmem_zalloc(size, KM_SLEEP);
2703 			memcpy(dspp.fromredactsnaps, fromredact, size);
2704 		}
2705 
2706 		boolean_t is_before =
2707 		    dsl_dataset_is_before(dspp.to_ds, fromds, 0);
2708 		dspp.is_clone = (dspp.to_ds->ds_dir !=
2709 		    fromds->ds_dir);
2710 		dsl_dataset_rele(fromds, FTAG);
2711 		if (!is_before) {
2712 			dsl_pool_rele(dspp.dp, FTAG);
2713 			err = SET_ERROR(EXDEV);
2714 		} else {
2715 			err = dmu_send_impl(&dspp);
2716 		}
2717 	} else {
2718 		dspp.numfromredactsnaps = NUM_SNAPS_NOT_REDACTED;
2719 		err = dmu_send_impl(&dspp);
2720 	}
2721 	if (dspp.fromredactsnaps)
2722 		kmem_free(dspp.fromredactsnaps,
2723 		    dspp.numfromredactsnaps * sizeof (uint64_t));
2724 
2725 	dsl_dataset_rele(dspp.to_ds, FTAG);
2726 	return (err);
2727 }
2728 
2729 int
2730 dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok,
2731     boolean_t large_block_ok, boolean_t compressok, boolean_t rawok,
2732     boolean_t savedok, uint64_t resumeobj, uint64_t resumeoff,
2733     const char *redactbook, int outfd, offset_t *off,
2734     dmu_send_outparams_t *dsop)
2735 {
2736 	int err = 0;
2737 	ds_hold_flags_t dsflags;
2738 	boolean_t owned = B_FALSE;
2739 	dsl_dataset_t *fromds = NULL;
2740 	zfs_bookmark_phys_t book = {0};
2741 	struct dmu_send_params dspp = {0};
2742 
2743 	dsflags = (rawok) ? DS_HOLD_FLAG_NONE : DS_HOLD_FLAG_DECRYPT;
2744 	dspp.tosnap = tosnap;
2745 	dspp.embedok = embedok;
2746 	dspp.large_block_ok = large_block_ok;
2747 	dspp.compressok = compressok;
2748 	dspp.outfd = outfd;
2749 	dspp.off = off;
2750 	dspp.dso = dsop;
2751 	dspp.tag = FTAG;
2752 	dspp.resumeobj = resumeobj;
2753 	dspp.resumeoff = resumeoff;
2754 	dspp.rawok = rawok;
2755 	dspp.savedok = savedok;
2756 
2757 	if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL)
2758 		return (SET_ERROR(EINVAL));
2759 
2760 	err = dsl_pool_hold(tosnap, FTAG, &dspp.dp);
2761 	if (err != 0)
2762 		return (err);
2763 
2764 	if (strchr(tosnap, '@') == NULL && spa_writeable(dspp.dp->dp_spa)) {
2765 		/*
2766 		 * We are sending a filesystem or volume.  Ensure
2767 		 * that it doesn't change by owning the dataset.
2768 		 */
2769 
2770 		if (savedok) {
2771 			/*
2772 			 * We are looking for the dataset that represents the
2773 			 * partially received send stream. If this stream was
2774 			 * received as a new snapshot of an existing dataset,
2775 			 * this will be saved in a hidden clone named
2776 			 * "<pool>/<dataset>/%recv". Otherwise, the stream
2777 			 * will be saved in the live dataset itself. In
2778 			 * either case we need to use dsl_dataset_own_force()
2779 			 * because the stream is marked as inconsistent,
2780 			 * which would normally make it unavailable to be
2781 			 * owned.
2782 			 */
2783 			char *name = kmem_asprintf("%s/%s", tosnap,
2784 			    recv_clone_name);
2785 			err = dsl_dataset_own_force(dspp.dp, name, dsflags,
2786 			    FTAG, &dspp.to_ds);
2787 			if (err == ENOENT) {
2788 				err = dsl_dataset_own_force(dspp.dp, tosnap,
2789 				    dsflags, FTAG, &dspp.to_ds);
2790 			}
2791 
2792 			if (err == 0) {
2793 				owned = B_TRUE;
2794 				err = zap_lookup(dspp.dp->dp_meta_objset,
2795 				    dspp.to_ds->ds_object,
2796 				    DS_FIELD_RESUME_TOGUID, 8, 1,
2797 				    &dspp.saved_guid);
2798 			}
2799 
2800 			if (err == 0) {
2801 				err = zap_lookup(dspp.dp->dp_meta_objset,
2802 				    dspp.to_ds->ds_object,
2803 				    DS_FIELD_RESUME_TONAME, 1,
2804 				    sizeof (dspp.saved_toname),
2805 				    dspp.saved_toname);
2806 			}
2807 			/* Only disown if there was an error in the lookups */
2808 			if (owned && (err != 0))
2809 				dsl_dataset_disown(dspp.to_ds, dsflags, FTAG);
2810 
2811 			kmem_strfree(name);
2812 		} else {
2813 			err = dsl_dataset_own(dspp.dp, tosnap, dsflags,
2814 			    FTAG, &dspp.to_ds);
2815 			if (err == 0)
2816 				owned = B_TRUE;
2817 		}
2818 	} else {
2819 		err = dsl_dataset_hold_flags(dspp.dp, tosnap, dsflags, FTAG,
2820 		    &dspp.to_ds);
2821 	}
2822 
2823 	if (err != 0) {
2824 		/* Note: dsl dataset is not owned at this point */
2825 		dsl_pool_rele(dspp.dp, FTAG);
2826 		return (err);
2827 	}
2828 
2829 	if (redactbook != NULL) {
2830 		char path[ZFS_MAX_DATASET_NAME_LEN];
2831 		(void) strlcpy(path, tosnap, sizeof (path));
2832 		char *at = strchr(path, '@');
2833 		if (at == NULL) {
2834 			err = EINVAL;
2835 		} else {
2836 			(void) snprintf(at, sizeof (path) - (at - path), "#%s",
2837 			    redactbook);
2838 			err = dsl_bookmark_lookup(dspp.dp, path,
2839 			    NULL, &book);
2840 			dspp.redactbook = &book;
2841 		}
2842 	}
2843 
2844 	if (err != 0) {
2845 		dsl_pool_rele(dspp.dp, FTAG);
2846 		if (owned)
2847 			dsl_dataset_disown(dspp.to_ds, dsflags, FTAG);
2848 		else
2849 			dsl_dataset_rele_flags(dspp.to_ds, dsflags, FTAG);
2850 		return (err);
2851 	}
2852 
2853 	if (fromsnap != NULL) {
2854 		zfs_bookmark_phys_t *zb = &dspp.ancestor_zb;
2855 		int fsnamelen;
2856 		if (strpbrk(tosnap, "@#") != NULL)
2857 			fsnamelen = strpbrk(tosnap, "@#") - tosnap;
2858 		else
2859 			fsnamelen = strlen(tosnap);
2860 
2861 		/*
2862 		 * If the fromsnap is in a different filesystem, then
2863 		 * mark the send stream as a clone.
2864 		 */
2865 		if (strncmp(tosnap, fromsnap, fsnamelen) != 0 ||
2866 		    (fromsnap[fsnamelen] != '@' &&
2867 		    fromsnap[fsnamelen] != '#')) {
2868 			dspp.is_clone = B_TRUE;
2869 		}
2870 
2871 		if (strchr(fromsnap, '@') != NULL) {
2872 			err = dsl_dataset_hold(dspp.dp, fromsnap, FTAG,
2873 			    &fromds);
2874 
2875 			if (err != 0) {
2876 				ASSERT3P(fromds, ==, NULL);
2877 			} else {
2878 				/*
2879 				 * We need to make a deep copy of the redact
2880 				 * snapshots of the from snapshot, because the
2881 				 * array will be freed when we evict from_ds.
2882 				 */
2883 				uint64_t *fromredact;
2884 				if (!dsl_dataset_get_uint64_array_feature(
2885 				    fromds, SPA_FEATURE_REDACTED_DATASETS,
2886 				    &dspp.numfromredactsnaps,
2887 				    &fromredact)) {
2888 					dspp.numfromredactsnaps =
2889 					    NUM_SNAPS_NOT_REDACTED;
2890 				} else if (dspp.numfromredactsnaps > 0) {
2891 					uint64_t size =
2892 					    dspp.numfromredactsnaps *
2893 					    sizeof (uint64_t);
2894 					dspp.fromredactsnaps = kmem_zalloc(size,
2895 					    KM_SLEEP);
2896 					memcpy(dspp.fromredactsnaps, fromredact,
2897 					    size);
2898 				}
2899 				if (!dsl_dataset_is_before(dspp.to_ds, fromds,
2900 				    0)) {
2901 					err = SET_ERROR(EXDEV);
2902 				} else {
2903 					zb->zbm_creation_txg =
2904 					    dsl_dataset_phys(fromds)->
2905 					    ds_creation_txg;
2906 					zb->zbm_creation_time =
2907 					    dsl_dataset_phys(fromds)->
2908 					    ds_creation_time;
2909 					zb->zbm_guid =
2910 					    dsl_dataset_phys(fromds)->ds_guid;
2911 					zb->zbm_redaction_obj = 0;
2912 
2913 					if (dsl_dataset_is_zapified(fromds)) {
2914 						(void) zap_lookup(
2915 						    dspp.dp->dp_meta_objset,
2916 						    fromds->ds_object,
2917 						    DS_FIELD_IVSET_GUID, 8, 1,
2918 						    &zb->zbm_ivset_guid);
2919 					}
2920 				}
2921 				dsl_dataset_rele(fromds, FTAG);
2922 			}
2923 		} else {
2924 			dspp.numfromredactsnaps = NUM_SNAPS_NOT_REDACTED;
2925 			err = dsl_bookmark_lookup(dspp.dp, fromsnap, dspp.to_ds,
2926 			    zb);
2927 			if (err == EXDEV && zb->zbm_redaction_obj != 0 &&
2928 			    zb->zbm_guid ==
2929 			    dsl_dataset_phys(dspp.to_ds)->ds_guid)
2930 				err = 0;
2931 		}
2932 
2933 		if (err == 0) {
2934 			/* dmu_send_impl will call dsl_pool_rele for us. */
2935 			err = dmu_send_impl(&dspp);
2936 		} else {
2937 			if (dspp.fromredactsnaps)
2938 				kmem_free(dspp.fromredactsnaps,
2939 				    dspp.numfromredactsnaps *
2940 				    sizeof (uint64_t));
2941 			dsl_pool_rele(dspp.dp, FTAG);
2942 		}
2943 	} else {
2944 		dspp.numfromredactsnaps = NUM_SNAPS_NOT_REDACTED;
2945 		err = dmu_send_impl(&dspp);
2946 	}
2947 	if (owned)
2948 		dsl_dataset_disown(dspp.to_ds, dsflags, FTAG);
2949 	else
2950 		dsl_dataset_rele_flags(dspp.to_ds, dsflags, FTAG);
2951 	return (err);
2952 }
2953 
2954 static int
2955 dmu_adjust_send_estimate_for_indirects(dsl_dataset_t *ds, uint64_t uncompressed,
2956     uint64_t compressed, boolean_t stream_compressed, uint64_t *sizep)
2957 {
2958 	int err = 0;
2959 	uint64_t size;
2960 	/*
2961 	 * Assume that space (both on-disk and in-stream) is dominated by
2962 	 * data.  We will adjust for indirect blocks and the copies property,
2963 	 * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
2964 	 */
2965 
2966 	uint64_t recordsize;
2967 	uint64_t record_count;
2968 	objset_t *os;
2969 	VERIFY0(dmu_objset_from_ds(ds, &os));
2970 
2971 	/* Assume all (uncompressed) blocks are recordsize. */
2972 	if (zfs_override_estimate_recordsize != 0) {
2973 		recordsize = zfs_override_estimate_recordsize;
2974 	} else if (os->os_phys->os_type == DMU_OST_ZVOL) {
2975 		err = dsl_prop_get_int_ds(ds,
2976 		    zfs_prop_to_name(ZFS_PROP_VOLBLOCKSIZE), &recordsize);
2977 	} else {
2978 		err = dsl_prop_get_int_ds(ds,
2979 		    zfs_prop_to_name(ZFS_PROP_RECORDSIZE), &recordsize);
2980 	}
2981 	if (err != 0)
2982 		return (err);
2983 	record_count = uncompressed / recordsize;
2984 
2985 	/*
2986 	 * If we're estimating a send size for a compressed stream, use the
2987 	 * compressed data size to estimate the stream size. Otherwise, use the
2988 	 * uncompressed data size.
2989 	 */
2990 	size = stream_compressed ? compressed : uncompressed;
2991 
2992 	/*
2993 	 * Subtract out approximate space used by indirect blocks.
2994 	 * Assume most space is used by data blocks (non-indirect, non-dnode).
2995 	 * Assume no ditto blocks or internal fragmentation.
2996 	 *
2997 	 * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
2998 	 * block.
2999 	 */
3000 	size -= record_count * sizeof (blkptr_t);
3001 
3002 	/* Add in the space for the record associated with each block. */
3003 	size += record_count * sizeof (dmu_replay_record_t);
3004 
3005 	*sizep = size;
3006 
3007 	return (0);
3008 }
3009 
3010 int
3011 dmu_send_estimate_fast(dsl_dataset_t *origds, dsl_dataset_t *fromds,
3012     zfs_bookmark_phys_t *frombook, boolean_t stream_compressed,
3013     boolean_t saved, uint64_t *sizep)
3014 {
3015 	int err;
3016 	dsl_dataset_t *ds = origds;
3017 	uint64_t uncomp, comp;
3018 
3019 	ASSERT(dsl_pool_config_held(origds->ds_dir->dd_pool));
3020 	ASSERT(fromds == NULL || frombook == NULL);
3021 
3022 	/*
3023 	 * If this is a saved send we may actually be sending
3024 	 * from the %recv clone used for resuming.
3025 	 */
3026 	if (saved) {
3027 		objset_t *mos = origds->ds_dir->dd_pool->dp_meta_objset;
3028 		uint64_t guid;
3029 		char dsname[ZFS_MAX_DATASET_NAME_LEN + 6];
3030 
3031 		dsl_dataset_name(origds, dsname);
3032 		(void) strcat(dsname, "/");
3033 		(void) strlcat(dsname, recv_clone_name, sizeof (dsname));
3034 
3035 		err = dsl_dataset_hold(origds->ds_dir->dd_pool,
3036 		    dsname, FTAG, &ds);
3037 		if (err != ENOENT && err != 0) {
3038 			return (err);
3039 		} else if (err == ENOENT) {
3040 			ds = origds;
3041 		}
3042 
3043 		/* check that this dataset has partially received data */
3044 		err = zap_lookup(mos, ds->ds_object,
3045 		    DS_FIELD_RESUME_TOGUID, 8, 1, &guid);
3046 		if (err != 0) {
3047 			err = SET_ERROR(err == ENOENT ? EINVAL : err);
3048 			goto out;
3049 		}
3050 
3051 		err = zap_lookup(mos, ds->ds_object,
3052 		    DS_FIELD_RESUME_TONAME, 1, sizeof (dsname), dsname);
3053 		if (err != 0) {
3054 			err = SET_ERROR(err == ENOENT ? EINVAL : err);
3055 			goto out;
3056 		}
3057 	}
3058 
3059 	/* tosnap must be a snapshot or the target of a saved send */
3060 	if (!ds->ds_is_snapshot && ds == origds)
3061 		return (SET_ERROR(EINVAL));
3062 
3063 	if (fromds != NULL) {
3064 		uint64_t used;
3065 		if (!fromds->ds_is_snapshot) {
3066 			err = SET_ERROR(EINVAL);
3067 			goto out;
3068 		}
3069 
3070 		if (!dsl_dataset_is_before(ds, fromds, 0)) {
3071 			err = SET_ERROR(EXDEV);
3072 			goto out;
3073 		}
3074 
3075 		err = dsl_dataset_space_written(fromds, ds, &used, &comp,
3076 		    &uncomp);
3077 		if (err != 0)
3078 			goto out;
3079 	} else if (frombook != NULL) {
3080 		uint64_t used;
3081 		err = dsl_dataset_space_written_bookmark(frombook, ds, &used,
3082 		    &comp, &uncomp);
3083 		if (err != 0)
3084 			goto out;
3085 	} else {
3086 		uncomp = dsl_dataset_phys(ds)->ds_uncompressed_bytes;
3087 		comp = dsl_dataset_phys(ds)->ds_compressed_bytes;
3088 	}
3089 
3090 	err = dmu_adjust_send_estimate_for_indirects(ds, uncomp, comp,
3091 	    stream_compressed, sizep);
3092 	/*
3093 	 * Add the size of the BEGIN and END records to the estimate.
3094 	 */
3095 	*sizep += 2 * sizeof (dmu_replay_record_t);
3096 
3097 out:
3098 	if (ds != origds)
3099 		dsl_dataset_rele(ds, FTAG);
3100 	return (err);
3101 }
3102 
3103 ZFS_MODULE_PARAM(zfs_send, zfs_send_, corrupt_data, INT, ZMOD_RW,
3104 	"Allow sending corrupt data");
3105 
3106 ZFS_MODULE_PARAM(zfs_send, zfs_send_, queue_length, UINT, ZMOD_RW,
3107 	"Maximum send queue length");
3108 
3109 ZFS_MODULE_PARAM(zfs_send, zfs_send_, unmodified_spill_blocks, INT, ZMOD_RW,
3110 	"Send unmodified spill blocks");
3111 
3112 ZFS_MODULE_PARAM(zfs_send, zfs_send_, no_prefetch_queue_length, UINT, ZMOD_RW,
3113 	"Maximum send queue length for non-prefetch queues");
3114 
3115 ZFS_MODULE_PARAM(zfs_send, zfs_send_, queue_ff, UINT, ZMOD_RW,
3116 	"Send queue fill fraction");
3117 
3118 ZFS_MODULE_PARAM(zfs_send, zfs_send_, no_prefetch_queue_ff, UINT, ZMOD_RW,
3119 	"Send queue fill fraction for non-prefetch queues");
3120 
3121 ZFS_MODULE_PARAM(zfs_send, zfs_, override_estimate_recordsize, UINT, ZMOD_RW,
3122 	"Override block size estimate with fixed size");
3123