1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
25 * Copyright (c) 2014, Joyent, Inc. All rights reserved.
26 * Copyright (c) 2012, Martin Matuska <[email protected]>. All rights reserved.
27 * Copyright 2014 HybridCluster. All rights reserved.
28 * Copyright 2016 RackTop Systems.
29 * Copyright (c) 2014 Integros [integros.com]
30 */
31
32 #include <sys/dmu.h>
33 #include <sys/dmu_impl.h>
34 #include <sys/dmu_tx.h>
35 #include <sys/dbuf.h>
36 #include <sys/dnode.h>
37 #include <sys/zfs_context.h>
38 #include <sys/dmu_objset.h>
39 #include <sys/dmu_traverse.h>
40 #include <sys/dsl_dataset.h>
41 #include <sys/dsl_dir.h>
42 #include <sys/dsl_prop.h>
43 #include <sys/dsl_pool.h>
44 #include <sys/dsl_synctask.h>
45 #include <sys/zfs_ioctl.h>
46 #include <sys/zap.h>
47 #include <sys/zio_checksum.h>
48 #include <sys/zfs_znode.h>
49 #include <zfs_fletcher.h>
50 #include <sys/avl.h>
51 #include <sys/ddt.h>
52 #include <sys/zfs_onexit.h>
53 #include <sys/dmu_send.h>
54 #include <sys/dsl_destroy.h>
55 #include <sys/blkptr.h>
56 #include <sys/dsl_bookmark.h>
57 #include <sys/zfeature.h>
58 #include <sys/bqueue.h>
59
60 #ifdef __FreeBSD__
61 #undef dump_write
62 #define dump_write dmu_dump_write
63 #endif
64
65 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
66 int zfs_send_corrupt_data = B_FALSE;
67 int zfs_send_queue_length = 16 * 1024 * 1024;
68 int zfs_recv_queue_length = 16 * 1024 * 1024;
69 /* Set this tunable to FALSE to disable setting of DRR_FLAG_FREERECORDS */
70 int zfs_send_set_freerecords_bit = B_TRUE;
71
72 #ifdef _KERNEL
73 TUNABLE_INT("vfs.zfs.send_set_freerecords_bit", &zfs_send_set_freerecords_bit);
74 #endif
75
76 static char *dmu_recv_tag = "dmu_recv_tag";
77 const char *recv_clone_name = "%recv";
78
79 /*
80 * Use this to override the recordsize calculation for fast zfs send estimates.
81 */
82 uint64_t zfs_override_estimate_recordsize = 0;
83
84 #define BP_SPAN(datablkszsec, indblkshift, level) \
85 (((uint64_t)datablkszsec) << (SPA_MINBLOCKSHIFT + \
86 (level) * (indblkshift - SPA_BLKPTRSHIFT)))
87
88 static void byteswap_record(dmu_replay_record_t *drr);
89
90 struct send_thread_arg {
91 bqueue_t q;
92 dsl_dataset_t *ds; /* Dataset to traverse */
93 uint64_t fromtxg; /* Traverse from this txg */
94 int flags; /* flags to pass to traverse_dataset */
95 int error_code;
96 boolean_t cancel;
97 zbookmark_phys_t resume;
98 };
99
100 struct send_block_record {
101 boolean_t eos_marker; /* Marks the end of the stream */
102 blkptr_t bp;
103 zbookmark_phys_t zb;
104 uint8_t indblkshift;
105 uint16_t datablkszsec;
106 bqueue_node_t ln;
107 };
108
109 static int
dump_bytes(dmu_sendarg_t * dsp,void * buf,int len)110 dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
111 {
112 dsl_dataset_t *ds = dmu_objset_ds(dsp->dsa_os);
113 struct uio auio;
114 struct iovec aiov;
115
116 /*
117 * The code does not rely on this (len being a multiple of 8). We keep
118 * this assertion because of the corresponding assertion in
119 * receive_read(). Keeping this assertion ensures that we do not
120 * inadvertently break backwards compatibility (causing the assertion
121 * in receive_read() to trigger on old software).
122 *
123 * Removing the assertions could be rolled into a new feature that uses
124 * data that isn't 8-byte aligned; if the assertions were removed, a
125 * feature flag would have to be added.
126 */
127
128 ASSERT0(len % 8);
129
130 aiov.iov_base = buf;
131 aiov.iov_len = len;
132 auio.uio_iov = &aiov;
133 auio.uio_iovcnt = 1;
134 auio.uio_resid = len;
135 auio.uio_segflg = UIO_SYSSPACE;
136 auio.uio_rw = UIO_WRITE;
137 auio.uio_offset = (off_t)-1;
138 auio.uio_td = dsp->dsa_td;
139 #ifdef _KERNEL
140 if (dsp->dsa_fp->f_type == DTYPE_VNODE)
141 bwillwrite();
142 dsp->dsa_err = fo_write(dsp->dsa_fp, &auio, dsp->dsa_td->td_ucred, 0,
143 dsp->dsa_td);
144 #else
145 fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__);
146 dsp->dsa_err = EOPNOTSUPP;
147 #endif
148 mutex_enter(&ds->ds_sendstream_lock);
149 *dsp->dsa_off += len;
150 mutex_exit(&ds->ds_sendstream_lock);
151
152 return (dsp->dsa_err);
153 }
154
155 /*
156 * For all record types except BEGIN, fill in the checksum (overlaid in
157 * drr_u.drr_checksum.drr_checksum). The checksum verifies everything
158 * up to the start of the checksum itself.
159 */
160 static int
dump_record(dmu_sendarg_t * dsp,void * payload,int payload_len)161 dump_record(dmu_sendarg_t *dsp, void *payload, int payload_len)
162 {
163 ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
164 ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
165 (void) fletcher_4_incremental_native(dsp->dsa_drr,
166 offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
167 &dsp->dsa_zc);
168 if (dsp->dsa_drr->drr_type == DRR_BEGIN) {
169 dsp->dsa_sent_begin = B_TRUE;
170 } else {
171 ASSERT(ZIO_CHECKSUM_IS_ZERO(&dsp->dsa_drr->drr_u.
172 drr_checksum.drr_checksum));
173 dsp->dsa_drr->drr_u.drr_checksum.drr_checksum = dsp->dsa_zc;
174 }
175 if (dsp->dsa_drr->drr_type == DRR_END) {
176 dsp->dsa_sent_end = B_TRUE;
177 }
178 (void) fletcher_4_incremental_native(&dsp->dsa_drr->
179 drr_u.drr_checksum.drr_checksum,
180 sizeof (zio_cksum_t), &dsp->dsa_zc);
181 if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
182 return (SET_ERROR(EINTR));
183 if (payload_len != 0) {
184 (void) fletcher_4_incremental_native(payload, payload_len,
185 &dsp->dsa_zc);
186 if (dump_bytes(dsp, payload, payload_len) != 0)
187 return (SET_ERROR(EINTR));
188 }
189 return (0);
190 }
191
192 /*
193 * Fill in the drr_free struct, or perform aggregation if the previous record is
194 * also a free record, and the two are adjacent.
195 *
196 * Note that we send free records even for a full send, because we want to be
197 * able to receive a full send as a clone, which requires a list of all the free
198 * and freeobject records that were generated on the source.
199 */
200 static int
dump_free(dmu_sendarg_t * dsp,uint64_t object,uint64_t offset,uint64_t length)201 dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
202 uint64_t length)
203 {
204 struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free);
205
206 /*
207 * When we receive a free record, dbuf_free_range() assumes
208 * that the receiving system doesn't have any dbufs in the range
209 * being freed. This is always true because there is a one-record
210 * constraint: we only send one WRITE record for any given
211 * object,offset. We know that the one-record constraint is
212 * true because we always send data in increasing order by
213 * object,offset.
214 *
215 * If the increasing-order constraint ever changes, we should find
216 * another way to assert that the one-record constraint is still
217 * satisfied.
218 */
219 ASSERT(object > dsp->dsa_last_data_object ||
220 (object == dsp->dsa_last_data_object &&
221 offset > dsp->dsa_last_data_offset));
222
223 if (length != -1ULL && offset + length < offset)
224 length = -1ULL;
225
226 /*
227 * If there is a pending op, but it's not PENDING_FREE, push it out,
228 * since free block aggregation can only be done for blocks of the
229 * same type (i.e., DRR_FREE records can only be aggregated with
230 * other DRR_FREE records. DRR_FREEOBJECTS records can only be
231 * aggregated with other DRR_FREEOBJECTS records.
232 */
233 if (dsp->dsa_pending_op != PENDING_NONE &&
234 dsp->dsa_pending_op != PENDING_FREE) {
235 if (dump_record(dsp, NULL, 0) != 0)
236 return (SET_ERROR(EINTR));
237 dsp->dsa_pending_op = PENDING_NONE;
238 }
239
240 if (dsp->dsa_pending_op == PENDING_FREE) {
241 /*
242 * There should never be a PENDING_FREE if length is -1
243 * (because dump_dnode is the only place where this
244 * function is called with a -1, and only after flushing
245 * any pending record).
246 */
247 ASSERT(length != -1ULL);
248 /*
249 * Check to see whether this free block can be aggregated
250 * with pending one.
251 */
252 if (drrf->drr_object == object && drrf->drr_offset +
253 drrf->drr_length == offset) {
254 drrf->drr_length += length;
255 return (0);
256 } else {
257 /* not a continuation. Push out pending record */
258 if (dump_record(dsp, NULL, 0) != 0)
259 return (SET_ERROR(EINTR));
260 dsp->dsa_pending_op = PENDING_NONE;
261 }
262 }
263 /* create a FREE record and make it pending */
264 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
265 dsp->dsa_drr->drr_type = DRR_FREE;
266 drrf->drr_object = object;
267 drrf->drr_offset = offset;
268 drrf->drr_length = length;
269 drrf->drr_toguid = dsp->dsa_toguid;
270 if (length == -1ULL) {
271 if (dump_record(dsp, NULL, 0) != 0)
272 return (SET_ERROR(EINTR));
273 } else {
274 dsp->dsa_pending_op = PENDING_FREE;
275 }
276
277 return (0);
278 }
279
280 static int
dump_write(dmu_sendarg_t * dsp,dmu_object_type_t type,uint64_t object,uint64_t offset,int lsize,int psize,const blkptr_t * bp,void * data)281 dump_write(dmu_sendarg_t *dsp, dmu_object_type_t type,
282 uint64_t object, uint64_t offset, int lsize, int psize, const blkptr_t *bp,
283 void *data)
284 {
285 uint64_t payload_size;
286 struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write);
287
288 /*
289 * We send data in increasing object, offset order.
290 * See comment in dump_free() for details.
291 */
292 ASSERT(object > dsp->dsa_last_data_object ||
293 (object == dsp->dsa_last_data_object &&
294 offset > dsp->dsa_last_data_offset));
295 dsp->dsa_last_data_object = object;
296 dsp->dsa_last_data_offset = offset + lsize - 1;
297
298 /*
299 * If there is any kind of pending aggregation (currently either
300 * a grouping of free objects or free blocks), push it out to
301 * the stream, since aggregation can't be done across operations
302 * of different types.
303 */
304 if (dsp->dsa_pending_op != PENDING_NONE) {
305 if (dump_record(dsp, NULL, 0) != 0)
306 return (SET_ERROR(EINTR));
307 dsp->dsa_pending_op = PENDING_NONE;
308 }
309 /* write a WRITE record */
310 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
311 dsp->dsa_drr->drr_type = DRR_WRITE;
312 drrw->drr_object = object;
313 drrw->drr_type = type;
314 drrw->drr_offset = offset;
315 drrw->drr_toguid = dsp->dsa_toguid;
316 drrw->drr_logical_size = lsize;
317
318 /* only set the compression fields if the buf is compressed */
319 if (lsize != psize) {
320 ASSERT(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_COMPRESSED);
321 ASSERT(!BP_IS_EMBEDDED(bp));
322 ASSERT(!BP_SHOULD_BYTESWAP(bp));
323 ASSERT(!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)));
324 ASSERT3U(BP_GET_COMPRESS(bp), !=, ZIO_COMPRESS_OFF);
325 ASSERT3S(psize, >, 0);
326 ASSERT3S(lsize, >=, psize);
327
328 drrw->drr_compressiontype = BP_GET_COMPRESS(bp);
329 drrw->drr_compressed_size = psize;
330 payload_size = drrw->drr_compressed_size;
331 } else {
332 payload_size = drrw->drr_logical_size;
333 }
334
335 if (bp == NULL || BP_IS_EMBEDDED(bp)) {
336 /*
337 * There's no pre-computed checksum for partial-block
338 * writes or embedded BP's, so (like
339 * fletcher4-checkummed blocks) userland will have to
340 * compute a dedup-capable checksum itself.
341 */
342 drrw->drr_checksumtype = ZIO_CHECKSUM_OFF;
343 } else {
344 drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
345 if (zio_checksum_table[drrw->drr_checksumtype].ci_flags &
346 ZCHECKSUM_FLAG_DEDUP)
347 drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
348 DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
349 DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
350 DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
351 drrw->drr_key.ddk_cksum = bp->blk_cksum;
352 }
353
354 if (dump_record(dsp, data, payload_size) != 0)
355 return (SET_ERROR(EINTR));
356 return (0);
357 }
358
359 static int
dump_write_embedded(dmu_sendarg_t * dsp,uint64_t object,uint64_t offset,int blksz,const blkptr_t * bp)360 dump_write_embedded(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
361 int blksz, const blkptr_t *bp)
362 {
363 char buf[BPE_PAYLOAD_SIZE];
364 struct drr_write_embedded *drrw =
365 &(dsp->dsa_drr->drr_u.drr_write_embedded);
366
367 if (dsp->dsa_pending_op != PENDING_NONE) {
368 if (dump_record(dsp, NULL, 0) != 0)
369 return (EINTR);
370 dsp->dsa_pending_op = PENDING_NONE;
371 }
372
373 ASSERT(BP_IS_EMBEDDED(bp));
374
375 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
376 dsp->dsa_drr->drr_type = DRR_WRITE_EMBEDDED;
377 drrw->drr_object = object;
378 drrw->drr_offset = offset;
379 drrw->drr_length = blksz;
380 drrw->drr_toguid = dsp->dsa_toguid;
381 drrw->drr_compression = BP_GET_COMPRESS(bp);
382 drrw->drr_etype = BPE_GET_ETYPE(bp);
383 drrw->drr_lsize = BPE_GET_LSIZE(bp);
384 drrw->drr_psize = BPE_GET_PSIZE(bp);
385
386 decode_embedded_bp_compressed(bp, buf);
387
388 if (dump_record(dsp, buf, P2ROUNDUP(drrw->drr_psize, 8)) != 0)
389 return (EINTR);
390 return (0);
391 }
392
393 static int
dump_spill(dmu_sendarg_t * dsp,uint64_t object,int blksz,void * data)394 dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data)
395 {
396 struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill);
397
398 if (dsp->dsa_pending_op != PENDING_NONE) {
399 if (dump_record(dsp, NULL, 0) != 0)
400 return (SET_ERROR(EINTR));
401 dsp->dsa_pending_op = PENDING_NONE;
402 }
403
404 /* write a SPILL record */
405 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
406 dsp->dsa_drr->drr_type = DRR_SPILL;
407 drrs->drr_object = object;
408 drrs->drr_length = blksz;
409 drrs->drr_toguid = dsp->dsa_toguid;
410
411 if (dump_record(dsp, data, blksz) != 0)
412 return (SET_ERROR(EINTR));
413 return (0);
414 }
415
416 static int
dump_freeobjects(dmu_sendarg_t * dsp,uint64_t firstobj,uint64_t numobjs)417 dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
418 {
419 struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
420
421 /*
422 * If there is a pending op, but it's not PENDING_FREEOBJECTS,
423 * push it out, since free block aggregation can only be done for
424 * blocks of the same type (i.e., DRR_FREE records can only be
425 * aggregated with other DRR_FREE records. DRR_FREEOBJECTS records
426 * can only be aggregated with other DRR_FREEOBJECTS records.
427 */
428 if (dsp->dsa_pending_op != PENDING_NONE &&
429 dsp->dsa_pending_op != PENDING_FREEOBJECTS) {
430 if (dump_record(dsp, NULL, 0) != 0)
431 return (SET_ERROR(EINTR));
432 dsp->dsa_pending_op = PENDING_NONE;
433 }
434 if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) {
435 /*
436 * See whether this free object array can be aggregated
437 * with pending one
438 */
439 if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
440 drrfo->drr_numobjs += numobjs;
441 return (0);
442 } else {
443 /* can't be aggregated. Push out pending record */
444 if (dump_record(dsp, NULL, 0) != 0)
445 return (SET_ERROR(EINTR));
446 dsp->dsa_pending_op = PENDING_NONE;
447 }
448 }
449
450 /* write a FREEOBJECTS record */
451 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
452 dsp->dsa_drr->drr_type = DRR_FREEOBJECTS;
453 drrfo->drr_firstobj = firstobj;
454 drrfo->drr_numobjs = numobjs;
455 drrfo->drr_toguid = dsp->dsa_toguid;
456
457 dsp->dsa_pending_op = PENDING_FREEOBJECTS;
458
459 return (0);
460 }
461
462 static int
dump_dnode(dmu_sendarg_t * dsp,uint64_t object,dnode_phys_t * dnp)463 dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
464 {
465 struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object);
466
467 if (object < dsp->dsa_resume_object) {
468 /*
469 * Note: when resuming, we will visit all the dnodes in
470 * the block of dnodes that we are resuming from. In
471 * this case it's unnecessary to send the dnodes prior to
472 * the one we are resuming from. We should be at most one
473 * block's worth of dnodes behind the resume point.
474 */
475 ASSERT3U(dsp->dsa_resume_object - object, <,
476 1 << (DNODE_BLOCK_SHIFT - DNODE_SHIFT));
477 return (0);
478 }
479
480 if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
481 return (dump_freeobjects(dsp, object, 1));
482
483 if (dsp->dsa_pending_op != PENDING_NONE) {
484 if (dump_record(dsp, NULL, 0) != 0)
485 return (SET_ERROR(EINTR));
486 dsp->dsa_pending_op = PENDING_NONE;
487 }
488
489 /* write an OBJECT record */
490 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
491 dsp->dsa_drr->drr_type = DRR_OBJECT;
492 drro->drr_object = object;
493 drro->drr_type = dnp->dn_type;
494 drro->drr_bonustype = dnp->dn_bonustype;
495 drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
496 drro->drr_bonuslen = dnp->dn_bonuslen;
497 drro->drr_dn_slots = dnp->dn_extra_slots + 1;
498 drro->drr_checksumtype = dnp->dn_checksum;
499 drro->drr_compress = dnp->dn_compress;
500 drro->drr_toguid = dsp->dsa_toguid;
501
502 if (!(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
503 drro->drr_blksz > SPA_OLD_MAXBLOCKSIZE)
504 drro->drr_blksz = SPA_OLD_MAXBLOCKSIZE;
505
506 if (dump_record(dsp, DN_BONUS(dnp),
507 P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0) {
508 return (SET_ERROR(EINTR));
509 }
510
511 /* Free anything past the end of the file. */
512 if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) *
513 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0)
514 return (SET_ERROR(EINTR));
515 if (dsp->dsa_err != 0)
516 return (SET_ERROR(EINTR));
517 return (0);
518 }
519
520 static boolean_t
backup_do_embed(dmu_sendarg_t * dsp,const blkptr_t * bp)521 backup_do_embed(dmu_sendarg_t *dsp, const blkptr_t *bp)
522 {
523 if (!BP_IS_EMBEDDED(bp))
524 return (B_FALSE);
525
526 /*
527 * Compression function must be legacy, or explicitly enabled.
528 */
529 if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS &&
530 !(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LZ4)))
531 return (B_FALSE);
532
533 /*
534 * Embed type must be explicitly enabled.
535 */
536 switch (BPE_GET_ETYPE(bp)) {
537 case BP_EMBEDDED_TYPE_DATA:
538 if (dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)
539 return (B_TRUE);
540 break;
541 default:
542 return (B_FALSE);
543 }
544 return (B_FALSE);
545 }
546
547 /*
548 * This is the callback function to traverse_dataset that acts as the worker
549 * thread for dmu_send_impl.
550 */
551 /*ARGSUSED*/
552 static int
send_cb(spa_t * spa,zilog_t * zilog,const blkptr_t * bp,const zbookmark_phys_t * zb,const struct dnode_phys * dnp,void * arg)553 send_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
554 const zbookmark_phys_t *zb, const struct dnode_phys *dnp, void *arg)
555 {
556 struct send_thread_arg *sta = arg;
557 struct send_block_record *record;
558 uint64_t record_size;
559 int err = 0;
560
561 ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
562 zb->zb_object >= sta->resume.zb_object);
563
564 if (sta->cancel)
565 return (SET_ERROR(EINTR));
566
567 if (bp == NULL) {
568 ASSERT3U(zb->zb_level, ==, ZB_DNODE_LEVEL);
569 return (0);
570 } else if (zb->zb_level < 0) {
571 return (0);
572 }
573
574 record = kmem_zalloc(sizeof (struct send_block_record), KM_SLEEP);
575 record->eos_marker = B_FALSE;
576 record->bp = *bp;
577 record->zb = *zb;
578 record->indblkshift = dnp->dn_indblkshift;
579 record->datablkszsec = dnp->dn_datablkszsec;
580 record_size = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
581 bqueue_enqueue(&sta->q, record, record_size);
582
583 return (err);
584 }
585
586 /*
587 * This function kicks off the traverse_dataset. It also handles setting the
588 * error code of the thread in case something goes wrong, and pushes the End of
589 * Stream record when the traverse_dataset call has finished. If there is no
590 * dataset to traverse, the thread immediately pushes End of Stream marker.
591 */
592 static void
send_traverse_thread(void * arg)593 send_traverse_thread(void *arg)
594 {
595 struct send_thread_arg *st_arg = arg;
596 int err;
597 struct send_block_record *data;
598
599 if (st_arg->ds != NULL) {
600 err = traverse_dataset_resume(st_arg->ds,
601 st_arg->fromtxg, &st_arg->resume,
602 st_arg->flags, send_cb, st_arg);
603
604 if (err != EINTR)
605 st_arg->error_code = err;
606 }
607 data = kmem_zalloc(sizeof (*data), KM_SLEEP);
608 data->eos_marker = B_TRUE;
609 bqueue_enqueue(&st_arg->q, data, 1);
610 thread_exit();
611 }
612
613 /*
614 * This function actually handles figuring out what kind of record needs to be
615 * dumped, reading the data (which has hopefully been prefetched), and calling
616 * the appropriate helper function.
617 */
618 static int
do_dump(dmu_sendarg_t * dsa,struct send_block_record * data)619 do_dump(dmu_sendarg_t *dsa, struct send_block_record *data)
620 {
621 dsl_dataset_t *ds = dmu_objset_ds(dsa->dsa_os);
622 const blkptr_t *bp = &data->bp;
623 const zbookmark_phys_t *zb = &data->zb;
624 uint8_t indblkshift = data->indblkshift;
625 uint16_t dblkszsec = data->datablkszsec;
626 spa_t *spa = ds->ds_dir->dd_pool->dp_spa;
627 dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
628 int err = 0;
629
630 ASSERT3U(zb->zb_level, >=, 0);
631
632 ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
633 zb->zb_object >= dsa->dsa_resume_object);
634
635 if (zb->zb_object != DMU_META_DNODE_OBJECT &&
636 DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
637 return (0);
638 } else if (BP_IS_HOLE(bp) &&
639 zb->zb_object == DMU_META_DNODE_OBJECT) {
640 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
641 uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
642 err = dump_freeobjects(dsa, dnobj, span >> DNODE_SHIFT);
643 } else if (BP_IS_HOLE(bp)) {
644 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
645 uint64_t offset = zb->zb_blkid * span;
646 err = dump_free(dsa, zb->zb_object, offset, span);
647 } else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
648 return (0);
649 } else if (type == DMU_OT_DNODE) {
650 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
651 arc_flags_t aflags = ARC_FLAG_WAIT;
652 arc_buf_t *abuf;
653
654 ASSERT0(zb->zb_level);
655
656 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
657 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
658 &aflags, zb) != 0)
659 return (SET_ERROR(EIO));
660
661 dnode_phys_t *blk = abuf->b_data;
662 uint64_t dnobj = zb->zb_blkid * epb;
663 for (int i = 0; i < epb; i += blk[i].dn_extra_slots + 1) {
664 err = dump_dnode(dsa, dnobj + i, blk + i);
665 if (err != 0)
666 break;
667 }
668 arc_buf_destroy(abuf, &abuf);
669 } else if (type == DMU_OT_SA) {
670 arc_flags_t aflags = ARC_FLAG_WAIT;
671 arc_buf_t *abuf;
672 int blksz = BP_GET_LSIZE(bp);
673
674 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
675 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
676 &aflags, zb) != 0)
677 return (SET_ERROR(EIO));
678
679 err = dump_spill(dsa, zb->zb_object, blksz, abuf->b_data);
680 arc_buf_destroy(abuf, &abuf);
681 } else if (backup_do_embed(dsa, bp)) {
682 /* it's an embedded level-0 block of a regular object */
683 int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
684 ASSERT0(zb->zb_level);
685 err = dump_write_embedded(dsa, zb->zb_object,
686 zb->zb_blkid * blksz, blksz, bp);
687 } else {
688 /* it's a level-0 block of a regular object */
689 arc_flags_t aflags = ARC_FLAG_WAIT;
690 arc_buf_t *abuf;
691 int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
692 uint64_t offset;
693
694 /*
695 * If we have large blocks stored on disk but the send flags
696 * don't allow us to send large blocks, we split the data from
697 * the arc buf into chunks.
698 */
699 boolean_t split_large_blocks = blksz > SPA_OLD_MAXBLOCKSIZE &&
700 !(dsa->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS);
701 /*
702 * We should only request compressed data from the ARC if all
703 * the following are true:
704 * - stream compression was requested
705 * - we aren't splitting large blocks into smaller chunks
706 * - the data won't need to be byteswapped before sending
707 * - this isn't an embedded block
708 * - this isn't metadata (if receiving on a different endian
709 * system it can be byteswapped more easily)
710 */
711 boolean_t request_compressed =
712 (dsa->dsa_featureflags & DMU_BACKUP_FEATURE_COMPRESSED) &&
713 !split_large_blocks && !BP_SHOULD_BYTESWAP(bp) &&
714 !BP_IS_EMBEDDED(bp) && !DMU_OT_IS_METADATA(BP_GET_TYPE(bp));
715
716 ASSERT0(zb->zb_level);
717 ASSERT(zb->zb_object > dsa->dsa_resume_object ||
718 (zb->zb_object == dsa->dsa_resume_object &&
719 zb->zb_blkid * blksz >= dsa->dsa_resume_offset));
720
721 ASSERT0(zb->zb_level);
722 ASSERT(zb->zb_object > dsa->dsa_resume_object ||
723 (zb->zb_object == dsa->dsa_resume_object &&
724 zb->zb_blkid * blksz >= dsa->dsa_resume_offset));
725
726 ASSERT3U(blksz, ==, BP_GET_LSIZE(bp));
727
728 enum zio_flag zioflags = ZIO_FLAG_CANFAIL;
729 if (request_compressed)
730 zioflags |= ZIO_FLAG_RAW;
731 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
732 ZIO_PRIORITY_ASYNC_READ, zioflags, &aflags, zb) != 0) {
733 if (zfs_send_corrupt_data) {
734 /* Send a block filled with 0x"zfs badd bloc" */
735 abuf = arc_alloc_buf(spa, &abuf, ARC_BUFC_DATA,
736 blksz);
737 uint64_t *ptr;
738 for (ptr = abuf->b_data;
739 (char *)ptr < (char *)abuf->b_data + blksz;
740 ptr++)
741 *ptr = 0x2f5baddb10cULL;
742 } else {
743 return (SET_ERROR(EIO));
744 }
745 }
746
747 offset = zb->zb_blkid * blksz;
748
749 if (split_large_blocks) {
750 ASSERT3U(arc_get_compression(abuf), ==,
751 ZIO_COMPRESS_OFF);
752 char *buf = abuf->b_data;
753 while (blksz > 0 && err == 0) {
754 int n = MIN(blksz, SPA_OLD_MAXBLOCKSIZE);
755 err = dump_write(dsa, type, zb->zb_object,
756 offset, n, n, NULL, buf);
757 offset += n;
758 buf += n;
759 blksz -= n;
760 }
761 } else {
762 err = dump_write(dsa, type, zb->zb_object, offset,
763 blksz, arc_buf_size(abuf), bp, abuf->b_data);
764 }
765 arc_buf_destroy(abuf, &abuf);
766 }
767
768 ASSERT(err == 0 || err == EINTR);
769 return (err);
770 }
771
772 /*
773 * Pop the new data off the queue, and free the old data.
774 */
775 static struct send_block_record *
get_next_record(bqueue_t * bq,struct send_block_record * data)776 get_next_record(bqueue_t *bq, struct send_block_record *data)
777 {
778 struct send_block_record *tmp = bqueue_dequeue(bq);
779 kmem_free(data, sizeof (*data));
780 return (tmp);
781 }
782
783 /*
784 * Actually do the bulk of the work in a zfs send.
785 *
786 * Note: Releases dp using the specified tag.
787 */
788 static int
dmu_send_impl(void * tag,dsl_pool_t * dp,dsl_dataset_t * to_ds,zfs_bookmark_phys_t * ancestor_zb,boolean_t is_clone,boolean_t embedok,boolean_t large_block_ok,boolean_t compressok,int outfd,uint64_t resumeobj,uint64_t resumeoff,vnode_t * vp,offset_t * off)789 dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *to_ds,
790 zfs_bookmark_phys_t *ancestor_zb, boolean_t is_clone,
791 boolean_t embedok, boolean_t large_block_ok, boolean_t compressok,
792 int outfd, uint64_t resumeobj, uint64_t resumeoff,
793 #ifdef illumos
794 vnode_t *vp, offset_t *off)
795 #else
796 struct file *fp, offset_t *off)
797 #endif
798 {
799 objset_t *os;
800 dmu_replay_record_t *drr;
801 dmu_sendarg_t *dsp;
802 int err;
803 uint64_t fromtxg = 0;
804 uint64_t featureflags = 0;
805 struct send_thread_arg to_arg = { 0 };
806
807 err = dmu_objset_from_ds(to_ds, &os);
808 if (err != 0) {
809 dsl_pool_rele(dp, tag);
810 return (err);
811 }
812
813 drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
814 drr->drr_type = DRR_BEGIN;
815 drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
816 DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
817 DMU_SUBSTREAM);
818
819 #ifdef _KERNEL
820 if (dmu_objset_type(os) == DMU_OST_ZFS) {
821 uint64_t version;
822 if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0) {
823 kmem_free(drr, sizeof (dmu_replay_record_t));
824 dsl_pool_rele(dp, tag);
825 return (SET_ERROR(EINVAL));
826 }
827 if (version >= ZPL_VERSION_SA) {
828 featureflags |= DMU_BACKUP_FEATURE_SA_SPILL;
829 }
830 }
831 #endif
832
833 if (large_block_ok && to_ds->ds_feature_inuse[SPA_FEATURE_LARGE_BLOCKS])
834 featureflags |= DMU_BACKUP_FEATURE_LARGE_BLOCKS;
835 if (to_ds->ds_feature_inuse[SPA_FEATURE_LARGE_DNODE])
836 featureflags |= DMU_BACKUP_FEATURE_LARGE_DNODE;
837 if (embedok &&
838 spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) {
839 featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA;
840 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
841 featureflags |= DMU_BACKUP_FEATURE_LZ4;
842 }
843 if (compressok) {
844 featureflags |= DMU_BACKUP_FEATURE_COMPRESSED;
845 }
846 if ((featureflags &
847 (DMU_BACKUP_FEATURE_EMBED_DATA | DMU_BACKUP_FEATURE_COMPRESSED)) !=
848 0 && spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS)) {
849 featureflags |= DMU_BACKUP_FEATURE_LZ4;
850 }
851
852 if (resumeobj != 0 || resumeoff != 0) {
853 featureflags |= DMU_BACKUP_FEATURE_RESUMING;
854 }
855
856 DMU_SET_FEATUREFLAGS(drr->drr_u.drr_begin.drr_versioninfo,
857 featureflags);
858
859 drr->drr_u.drr_begin.drr_creation_time =
860 dsl_dataset_phys(to_ds)->ds_creation_time;
861 drr->drr_u.drr_begin.drr_type = dmu_objset_type(os);
862 if (is_clone)
863 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
864 drr->drr_u.drr_begin.drr_toguid = dsl_dataset_phys(to_ds)->ds_guid;
865 if (dsl_dataset_phys(to_ds)->ds_flags & DS_FLAG_CI_DATASET)
866 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
867 if (zfs_send_set_freerecords_bit)
868 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_FREERECORDS;
869
870 if (ancestor_zb != NULL) {
871 drr->drr_u.drr_begin.drr_fromguid =
872 ancestor_zb->zbm_guid;
873 fromtxg = ancestor_zb->zbm_creation_txg;
874 }
875 dsl_dataset_name(to_ds, drr->drr_u.drr_begin.drr_toname);
876 if (!to_ds->ds_is_snapshot) {
877 (void) strlcat(drr->drr_u.drr_begin.drr_toname, "@--head--",
878 sizeof (drr->drr_u.drr_begin.drr_toname));
879 }
880
881 dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP);
882
883 dsp->dsa_drr = drr;
884 dsp->dsa_outfd = outfd;
885 dsp->dsa_proc = curproc;
886 dsp->dsa_td = curthread;
887 dsp->dsa_fp = fp;
888 dsp->dsa_os = os;
889 dsp->dsa_off = off;
890 dsp->dsa_toguid = dsl_dataset_phys(to_ds)->ds_guid;
891 dsp->dsa_pending_op = PENDING_NONE;
892 dsp->dsa_featureflags = featureflags;
893 dsp->dsa_resume_object = resumeobj;
894 dsp->dsa_resume_offset = resumeoff;
895
896 mutex_enter(&to_ds->ds_sendstream_lock);
897 list_insert_head(&to_ds->ds_sendstreams, dsp);
898 mutex_exit(&to_ds->ds_sendstream_lock);
899
900 dsl_dataset_long_hold(to_ds, FTAG);
901 dsl_pool_rele(dp, tag);
902
903 void *payload = NULL;
904 size_t payload_len = 0;
905 if (resumeobj != 0 || resumeoff != 0) {
906 dmu_object_info_t to_doi;
907 err = dmu_object_info(os, resumeobj, &to_doi);
908 if (err != 0)
909 goto out;
910 SET_BOOKMARK(&to_arg.resume, to_ds->ds_object, resumeobj, 0,
911 resumeoff / to_doi.doi_data_block_size);
912
913 nvlist_t *nvl = fnvlist_alloc();
914 fnvlist_add_uint64(nvl, "resume_object", resumeobj);
915 fnvlist_add_uint64(nvl, "resume_offset", resumeoff);
916 payload = fnvlist_pack(nvl, &payload_len);
917 drr->drr_payloadlen = payload_len;
918 fnvlist_free(nvl);
919 }
920
921 err = dump_record(dsp, payload, payload_len);
922 fnvlist_pack_free(payload, payload_len);
923 if (err != 0) {
924 err = dsp->dsa_err;
925 goto out;
926 }
927
928 err = bqueue_init(&to_arg.q, zfs_send_queue_length,
929 offsetof(struct send_block_record, ln));
930 to_arg.error_code = 0;
931 to_arg.cancel = B_FALSE;
932 to_arg.ds = to_ds;
933 to_arg.fromtxg = fromtxg;
934 to_arg.flags = TRAVERSE_PRE | TRAVERSE_PREFETCH;
935 (void) thread_create(NULL, 0, send_traverse_thread, &to_arg, 0, &p0,
936 TS_RUN, minclsyspri);
937
938 struct send_block_record *to_data;
939 to_data = bqueue_dequeue(&to_arg.q);
940
941 while (!to_data->eos_marker && err == 0) {
942 err = do_dump(dsp, to_data);
943 to_data = get_next_record(&to_arg.q, to_data);
944 if (issig(JUSTLOOKING) && issig(FORREAL))
945 err = EINTR;
946 }
947
948 if (err != 0) {
949 to_arg.cancel = B_TRUE;
950 while (!to_data->eos_marker) {
951 to_data = get_next_record(&to_arg.q, to_data);
952 }
953 }
954 kmem_free(to_data, sizeof (*to_data));
955
956 bqueue_destroy(&to_arg.q);
957
958 if (err == 0 && to_arg.error_code != 0)
959 err = to_arg.error_code;
960
961 if (err != 0)
962 goto out;
963
964 if (dsp->dsa_pending_op != PENDING_NONE)
965 if (dump_record(dsp, NULL, 0) != 0)
966 err = SET_ERROR(EINTR);
967
968 if (err != 0) {
969 if (err == EINTR && dsp->dsa_err != 0)
970 err = dsp->dsa_err;
971 goto out;
972 }
973
974 bzero(drr, sizeof (dmu_replay_record_t));
975 drr->drr_type = DRR_END;
976 drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc;
977 drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid;
978
979 if (dump_record(dsp, NULL, 0) != 0)
980 err = dsp->dsa_err;
981
982 out:
983 mutex_enter(&to_ds->ds_sendstream_lock);
984 list_remove(&to_ds->ds_sendstreams, dsp);
985 mutex_exit(&to_ds->ds_sendstream_lock);
986
987 VERIFY(err != 0 || (dsp->dsa_sent_begin && dsp->dsa_sent_end));
988
989 kmem_free(drr, sizeof (dmu_replay_record_t));
990 kmem_free(dsp, sizeof (dmu_sendarg_t));
991
992 dsl_dataset_long_rele(to_ds, FTAG);
993
994 return (err);
995 }
996
997 int
dmu_send_obj(const char * pool,uint64_t tosnap,uint64_t fromsnap,boolean_t embedok,boolean_t large_block_ok,boolean_t compressok,int outfd,vnode_t * vp,offset_t * off)998 dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
999 boolean_t embedok, boolean_t large_block_ok, boolean_t compressok,
1000 #ifdef illumos
1001 int outfd, vnode_t *vp, offset_t *off)
1002 #else
1003 int outfd, struct file *fp, offset_t *off)
1004 #endif
1005 {
1006 dsl_pool_t *dp;
1007 dsl_dataset_t *ds;
1008 dsl_dataset_t *fromds = NULL;
1009 int err;
1010
1011 err = dsl_pool_hold(pool, FTAG, &dp);
1012 if (err != 0)
1013 return (err);
1014
1015 err = dsl_dataset_hold_obj(dp, tosnap, FTAG, &ds);
1016 if (err != 0) {
1017 dsl_pool_rele(dp, FTAG);
1018 return (err);
1019 }
1020
1021 if (fromsnap != 0) {
1022 zfs_bookmark_phys_t zb;
1023 boolean_t is_clone;
1024
1025 err = dsl_dataset_hold_obj(dp, fromsnap, FTAG, &fromds);
1026 if (err != 0) {
1027 dsl_dataset_rele(ds, FTAG);
1028 dsl_pool_rele(dp, FTAG);
1029 return (err);
1030 }
1031 if (!dsl_dataset_is_before(ds, fromds, 0))
1032 err = SET_ERROR(EXDEV);
1033 zb.zbm_creation_time =
1034 dsl_dataset_phys(fromds)->ds_creation_time;
1035 zb.zbm_creation_txg = dsl_dataset_phys(fromds)->ds_creation_txg;
1036 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
1037 is_clone = (fromds->ds_dir != ds->ds_dir);
1038 dsl_dataset_rele(fromds, FTAG);
1039 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
1040 embedok, large_block_ok, compressok, outfd, 0, 0, fp, off);
1041 } else {
1042 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
1043 embedok, large_block_ok, compressok, outfd, 0, 0, fp, off);
1044 }
1045 dsl_dataset_rele(ds, FTAG);
1046 return (err);
1047 }
1048
1049 int
dmu_send(const char * tosnap,const char * fromsnap,boolean_t embedok,boolean_t large_block_ok,boolean_t compressok,int outfd,uint64_t resumeobj,uint64_t resumeoff,vnode_t * vp,offset_t * off)1050 dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok,
1051 boolean_t large_block_ok, boolean_t compressok, int outfd,
1052 uint64_t resumeobj, uint64_t resumeoff,
1053 #ifdef illumos
1054 vnode_t *vp, offset_t *off)
1055 #else
1056 struct file *fp, offset_t *off)
1057 #endif
1058 {
1059 dsl_pool_t *dp;
1060 dsl_dataset_t *ds;
1061 int err;
1062 boolean_t owned = B_FALSE;
1063
1064 if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL)
1065 return (SET_ERROR(EINVAL));
1066
1067 err = dsl_pool_hold(tosnap, FTAG, &dp);
1068 if (err != 0)
1069 return (err);
1070
1071 if (strchr(tosnap, '@') == NULL && spa_writeable(dp->dp_spa)) {
1072 /*
1073 * We are sending a filesystem or volume. Ensure
1074 * that it doesn't change by owning the dataset.
1075 */
1076 err = dsl_dataset_own(dp, tosnap, FTAG, &ds);
1077 owned = B_TRUE;
1078 } else {
1079 err = dsl_dataset_hold(dp, tosnap, FTAG, &ds);
1080 }
1081 if (err != 0) {
1082 dsl_pool_rele(dp, FTAG);
1083 return (err);
1084 }
1085
1086 if (fromsnap != NULL) {
1087 zfs_bookmark_phys_t zb;
1088 boolean_t is_clone = B_FALSE;
1089 int fsnamelen = strchr(tosnap, '@') - tosnap;
1090
1091 /*
1092 * If the fromsnap is in a different filesystem, then
1093 * mark the send stream as a clone.
1094 */
1095 if (strncmp(tosnap, fromsnap, fsnamelen) != 0 ||
1096 (fromsnap[fsnamelen] != '@' &&
1097 fromsnap[fsnamelen] != '#')) {
1098 is_clone = B_TRUE;
1099 }
1100
1101 if (strchr(fromsnap, '@')) {
1102 dsl_dataset_t *fromds;
1103 err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds);
1104 if (err == 0) {
1105 if (!dsl_dataset_is_before(ds, fromds, 0))
1106 err = SET_ERROR(EXDEV);
1107 zb.zbm_creation_time =
1108 dsl_dataset_phys(fromds)->ds_creation_time;
1109 zb.zbm_creation_txg =
1110 dsl_dataset_phys(fromds)->ds_creation_txg;
1111 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
1112 is_clone = (ds->ds_dir != fromds->ds_dir);
1113 dsl_dataset_rele(fromds, FTAG);
1114 }
1115 } else {
1116 err = dsl_bookmark_lookup(dp, fromsnap, ds, &zb);
1117 }
1118 if (err != 0) {
1119 dsl_dataset_rele(ds, FTAG);
1120 dsl_pool_rele(dp, FTAG);
1121 return (err);
1122 }
1123 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
1124 embedok, large_block_ok, compressok,
1125 outfd, resumeobj, resumeoff, fp, off);
1126 } else {
1127 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
1128 embedok, large_block_ok, compressok,
1129 outfd, resumeobj, resumeoff, fp, off);
1130 }
1131 if (owned)
1132 dsl_dataset_disown(ds, FTAG);
1133 else
1134 dsl_dataset_rele(ds, FTAG);
1135 return (err);
1136 }
1137
1138 static int
dmu_adjust_send_estimate_for_indirects(dsl_dataset_t * ds,uint64_t uncompressed,uint64_t compressed,boolean_t stream_compressed,uint64_t * sizep)1139 dmu_adjust_send_estimate_for_indirects(dsl_dataset_t *ds, uint64_t uncompressed,
1140 uint64_t compressed, boolean_t stream_compressed, uint64_t *sizep)
1141 {
1142 int err = 0;
1143 uint64_t size;
1144 /*
1145 * Assume that space (both on-disk and in-stream) is dominated by
1146 * data. We will adjust for indirect blocks and the copies property,
1147 * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
1148 */
1149 uint64_t recordsize;
1150 uint64_t record_count;
1151 objset_t *os;
1152 VERIFY0(dmu_objset_from_ds(ds, &os));
1153
1154 /* Assume all (uncompressed) blocks are recordsize. */
1155 if (zfs_override_estimate_recordsize != 0) {
1156 recordsize = zfs_override_estimate_recordsize;
1157 } else if (os->os_phys->os_type == DMU_OST_ZVOL) {
1158 err = dsl_prop_get_int_ds(ds,
1159 zfs_prop_to_name(ZFS_PROP_VOLBLOCKSIZE), &recordsize);
1160 } else {
1161 err = dsl_prop_get_int_ds(ds,
1162 zfs_prop_to_name(ZFS_PROP_RECORDSIZE), &recordsize);
1163 }
1164 if (err != 0)
1165 return (err);
1166 record_count = uncompressed / recordsize;
1167
1168 /*
1169 * If we're estimating a send size for a compressed stream, use the
1170 * compressed data size to estimate the stream size. Otherwise, use the
1171 * uncompressed data size.
1172 */
1173 size = stream_compressed ? compressed : uncompressed;
1174
1175 /*
1176 * Subtract out approximate space used by indirect blocks.
1177 * Assume most space is used by data blocks (non-indirect, non-dnode).
1178 * Assume no ditto blocks or internal fragmentation.
1179 *
1180 * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
1181 * block.
1182 */
1183 size -= record_count * sizeof (blkptr_t);
1184
1185 /* Add in the space for the record associated with each block. */
1186 size += record_count * sizeof (dmu_replay_record_t);
1187
1188 *sizep = size;
1189
1190 return (0);
1191 }
1192
1193 int
dmu_send_estimate(dsl_dataset_t * ds,dsl_dataset_t * fromds,boolean_t stream_compressed,uint64_t * sizep)1194 dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds,
1195 boolean_t stream_compressed, uint64_t *sizep)
1196 {
1197 dsl_pool_t *dp = ds->ds_dir->dd_pool;
1198 int err;
1199 uint64_t uncomp, comp;
1200
1201 ASSERT(dsl_pool_config_held(dp));
1202
1203 /* tosnap must be a snapshot */
1204 if (!ds->ds_is_snapshot)
1205 return (SET_ERROR(EINVAL));
1206
1207 /* fromsnap, if provided, must be a snapshot */
1208 if (fromds != NULL && !fromds->ds_is_snapshot)
1209 return (SET_ERROR(EINVAL));
1210
1211 /*
1212 * fromsnap must be an earlier snapshot from the same fs as tosnap,
1213 * or the origin's fs.
1214 */
1215 if (fromds != NULL && !dsl_dataset_is_before(ds, fromds, 0))
1216 return (SET_ERROR(EXDEV));
1217
1218 /* Get compressed and uncompressed size estimates of changed data. */
1219 if (fromds == NULL) {
1220 uncomp = dsl_dataset_phys(ds)->ds_uncompressed_bytes;
1221 comp = dsl_dataset_phys(ds)->ds_compressed_bytes;
1222 } else {
1223 uint64_t used;
1224 err = dsl_dataset_space_written(fromds, ds,
1225 &used, &comp, &uncomp);
1226 if (err != 0)
1227 return (err);
1228 }
1229
1230 err = dmu_adjust_send_estimate_for_indirects(ds, uncomp, comp,
1231 stream_compressed, sizep);
1232 /*
1233 * Add the size of the BEGIN and END records to the estimate.
1234 */
1235 *sizep += 2 * sizeof (dmu_replay_record_t);
1236 return (err);
1237 }
1238
1239 struct calculate_send_arg {
1240 uint64_t uncompressed;
1241 uint64_t compressed;
1242 };
1243
1244 /*
1245 * Simple callback used to traverse the blocks of a snapshot and sum their
1246 * uncompressed and compressed sizes.
1247 */
1248 /* ARGSUSED */
1249 static int
dmu_calculate_send_traversal(spa_t * spa,zilog_t * zilog,const blkptr_t * bp,const zbookmark_phys_t * zb,const dnode_phys_t * dnp,void * arg)1250 dmu_calculate_send_traversal(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
1251 const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
1252 {
1253 struct calculate_send_arg *space = arg;
1254 if (bp != NULL && !BP_IS_HOLE(bp)) {
1255 space->uncompressed += BP_GET_UCSIZE(bp);
1256 space->compressed += BP_GET_PSIZE(bp);
1257 }
1258 return (0);
1259 }
1260
1261 /*
1262 * Given a desination snapshot and a TXG, calculate the approximate size of a
1263 * send stream sent from that TXG. from_txg may be zero, indicating that the
1264 * whole snapshot will be sent.
1265 */
1266 int
dmu_send_estimate_from_txg(dsl_dataset_t * ds,uint64_t from_txg,boolean_t stream_compressed,uint64_t * sizep)1267 dmu_send_estimate_from_txg(dsl_dataset_t *ds, uint64_t from_txg,
1268 boolean_t stream_compressed, uint64_t *sizep)
1269 {
1270 dsl_pool_t *dp = ds->ds_dir->dd_pool;
1271 int err;
1272 struct calculate_send_arg size = { 0 };
1273
1274 ASSERT(dsl_pool_config_held(dp));
1275
1276 /* tosnap must be a snapshot */
1277 if (!ds->ds_is_snapshot)
1278 return (SET_ERROR(EINVAL));
1279
1280 /* verify that from_txg is before the provided snapshot was taken */
1281 if (from_txg >= dsl_dataset_phys(ds)->ds_creation_txg) {
1282 return (SET_ERROR(EXDEV));
1283 }
1284
1285 /*
1286 * traverse the blocks of the snapshot with birth times after
1287 * from_txg, summing their uncompressed size
1288 */
1289 err = traverse_dataset(ds, from_txg, TRAVERSE_POST,
1290 dmu_calculate_send_traversal, &size);
1291 if (err)
1292 return (err);
1293
1294 err = dmu_adjust_send_estimate_for_indirects(ds, size.uncompressed,
1295 size.compressed, stream_compressed, sizep);
1296 return (err);
1297 }
1298
1299 typedef struct dmu_recv_begin_arg {
1300 const char *drba_origin;
1301 dmu_recv_cookie_t *drba_cookie;
1302 cred_t *drba_cred;
1303 uint64_t drba_snapobj;
1304 } dmu_recv_begin_arg_t;
1305
1306 static int
recv_begin_check_existing_impl(dmu_recv_begin_arg_t * drba,dsl_dataset_t * ds,uint64_t fromguid)1307 recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds,
1308 uint64_t fromguid)
1309 {
1310 uint64_t val;
1311 int error;
1312 dsl_pool_t *dp = ds->ds_dir->dd_pool;
1313
1314 /* Temporary clone name must not exist. */
1315 error = zap_lookup(dp->dp_meta_objset,
1316 dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, recv_clone_name,
1317 8, 1, &val);
1318 if (error != ENOENT)
1319 return (error == 0 ? SET_ERROR(EBUSY) : error);
1320
1321 /* Resume state must not be set. */
1322 if (dsl_dataset_has_resume_receive_state(ds))
1323 return (SET_ERROR(EBUSY));
1324
1325 /* New snapshot name must not exist. */
1326 error = zap_lookup(dp->dp_meta_objset,
1327 dsl_dataset_phys(ds)->ds_snapnames_zapobj,
1328 drba->drba_cookie->drc_tosnap, 8, 1, &val);
1329 if (error != ENOENT)
1330 return (error == 0 ? SET_ERROR(EEXIST) : error);
1331
1332 /*
1333 * Check snapshot limit before receiving. We'll recheck again at the
1334 * end, but might as well abort before receiving if we're already over
1335 * the limit.
1336 *
1337 * Note that we do not check the file system limit with
1338 * dsl_dir_fscount_check because the temporary %clones don't count
1339 * against that limit.
1340 */
1341 error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT,
1342 NULL, drba->drba_cred);
1343 if (error != 0)
1344 return (error);
1345
1346 if (fromguid != 0) {
1347 dsl_dataset_t *snap;
1348 uint64_t obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
1349
1350 /* Find snapshot in this dir that matches fromguid. */
1351 while (obj != 0) {
1352 error = dsl_dataset_hold_obj(dp, obj, FTAG,
1353 &snap);
1354 if (error != 0)
1355 return (SET_ERROR(ENODEV));
1356 if (snap->ds_dir != ds->ds_dir) {
1357 dsl_dataset_rele(snap, FTAG);
1358 return (SET_ERROR(ENODEV));
1359 }
1360 if (dsl_dataset_phys(snap)->ds_guid == fromguid)
1361 break;
1362 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
1363 dsl_dataset_rele(snap, FTAG);
1364 }
1365 if (obj == 0)
1366 return (SET_ERROR(ENODEV));
1367
1368 if (drba->drba_cookie->drc_force) {
1369 drba->drba_snapobj = obj;
1370 } else {
1371 /*
1372 * If we are not forcing, there must be no
1373 * changes since fromsnap.
1374 */
1375 if (dsl_dataset_modified_since_snap(ds, snap)) {
1376 dsl_dataset_rele(snap, FTAG);
1377 return (SET_ERROR(ETXTBSY));
1378 }
1379 drba->drba_snapobj = ds->ds_prev->ds_object;
1380 }
1381
1382 dsl_dataset_rele(snap, FTAG);
1383 } else {
1384 /* if full, then must be forced */
1385 if (!drba->drba_cookie->drc_force)
1386 return (SET_ERROR(EEXIST));
1387 /* start from $ORIGIN@$ORIGIN, if supported */
1388 drba->drba_snapobj = dp->dp_origin_snap != NULL ?
1389 dp->dp_origin_snap->ds_object : 0;
1390 }
1391
1392 return (0);
1393
1394 }
1395
1396 static int
dmu_recv_begin_check(void * arg,dmu_tx_t * tx)1397 dmu_recv_begin_check(void *arg, dmu_tx_t *tx)
1398 {
1399 dmu_recv_begin_arg_t *drba = arg;
1400 dsl_pool_t *dp = dmu_tx_pool(tx);
1401 struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1402 uint64_t fromguid = drrb->drr_fromguid;
1403 int flags = drrb->drr_flags;
1404 int error;
1405 uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1406 dsl_dataset_t *ds;
1407 const char *tofs = drba->drba_cookie->drc_tofs;
1408
1409 /* already checked */
1410 ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1411 ASSERT(!(featureflags & DMU_BACKUP_FEATURE_RESUMING));
1412
1413 if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1414 DMU_COMPOUNDSTREAM ||
1415 drrb->drr_type >= DMU_OST_NUMTYPES ||
1416 ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL))
1417 return (SET_ERROR(EINVAL));
1418
1419 /* Verify pool version supports SA if SA_SPILL feature set */
1420 if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1421 spa_version(dp->dp_spa) < SPA_VERSION_SA)
1422 return (SET_ERROR(ENOTSUP));
1423
1424 if (drba->drba_cookie->drc_resumable &&
1425 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EXTENSIBLE_DATASET))
1426 return (SET_ERROR(ENOTSUP));
1427
1428 /*
1429 * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1430 * record to a plain WRITE record, so the pool must have the
1431 * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1432 * records. Same with WRITE_EMBEDDED records that use LZ4 compression.
1433 */
1434 if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1435 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1436 return (SET_ERROR(ENOTSUP));
1437 if ((featureflags & DMU_BACKUP_FEATURE_LZ4) &&
1438 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1439 return (SET_ERROR(ENOTSUP));
1440
1441 /*
1442 * The receiving code doesn't know how to translate large blocks
1443 * to smaller ones, so the pool must have the LARGE_BLOCKS
1444 * feature enabled if the stream has LARGE_BLOCKS.
1445 */
1446 if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
1447 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS))
1448 return (SET_ERROR(ENOTSUP));
1449
1450 /*
1451 * The receiving code doesn't know how to translate large dnodes
1452 * to smaller ones, so the pool must have the LARGE_DNODE
1453 * feature enabled if the stream has LARGE_DNODE.
1454 */
1455 if ((featureflags & DMU_BACKUP_FEATURE_LARGE_DNODE) &&
1456 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_DNODE))
1457 return (SET_ERROR(ENOTSUP));
1458
1459 error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1460 if (error == 0) {
1461 /* target fs already exists; recv into temp clone */
1462
1463 /* Can't recv a clone into an existing fs */
1464 if (flags & DRR_FLAG_CLONE || drba->drba_origin) {
1465 dsl_dataset_rele(ds, FTAG);
1466 return (SET_ERROR(EINVAL));
1467 }
1468
1469 error = recv_begin_check_existing_impl(drba, ds, fromguid);
1470 dsl_dataset_rele(ds, FTAG);
1471 } else if (error == ENOENT) {
1472 /* target fs does not exist; must be a full backup or clone */
1473 char buf[ZFS_MAX_DATASET_NAME_LEN];
1474
1475 /*
1476 * If it's a non-clone incremental, we are missing the
1477 * target fs, so fail the recv.
1478 */
1479 if (fromguid != 0 && !(flags & DRR_FLAG_CLONE ||
1480 drba->drba_origin))
1481 return (SET_ERROR(ENOENT));
1482
1483 /*
1484 * If we're receiving a full send as a clone, and it doesn't
1485 * contain all the necessary free records and freeobject
1486 * records, reject it.
1487 */
1488 if (fromguid == 0 && drba->drba_origin &&
1489 !(flags & DRR_FLAG_FREERECORDS))
1490 return (SET_ERROR(EINVAL));
1491
1492 /* Open the parent of tofs */
1493 ASSERT3U(strlen(tofs), <, sizeof (buf));
1494 (void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1);
1495 error = dsl_dataset_hold(dp, buf, FTAG, &ds);
1496 if (error != 0)
1497 return (error);
1498
1499 /*
1500 * Check filesystem and snapshot limits before receiving. We'll
1501 * recheck snapshot limits again at the end (we create the
1502 * filesystems and increment those counts during begin_sync).
1503 */
1504 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1505 ZFS_PROP_FILESYSTEM_LIMIT, NULL, drba->drba_cred);
1506 if (error != 0) {
1507 dsl_dataset_rele(ds, FTAG);
1508 return (error);
1509 }
1510
1511 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1512 ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred);
1513 if (error != 0) {
1514 dsl_dataset_rele(ds, FTAG);
1515 return (error);
1516 }
1517
1518 if (drba->drba_origin != NULL) {
1519 dsl_dataset_t *origin;
1520 error = dsl_dataset_hold(dp, drba->drba_origin,
1521 FTAG, &origin);
1522 if (error != 0) {
1523 dsl_dataset_rele(ds, FTAG);
1524 return (error);
1525 }
1526 if (!origin->ds_is_snapshot) {
1527 dsl_dataset_rele(origin, FTAG);
1528 dsl_dataset_rele(ds, FTAG);
1529 return (SET_ERROR(EINVAL));
1530 }
1531 if (dsl_dataset_phys(origin)->ds_guid != fromguid &&
1532 fromguid != 0) {
1533 dsl_dataset_rele(origin, FTAG);
1534 dsl_dataset_rele(ds, FTAG);
1535 return (SET_ERROR(ENODEV));
1536 }
1537 dsl_dataset_rele(origin, FTAG);
1538 }
1539 dsl_dataset_rele(ds, FTAG);
1540 error = 0;
1541 }
1542 return (error);
1543 }
1544
1545 static void
dmu_recv_begin_sync(void * arg,dmu_tx_t * tx)1546 dmu_recv_begin_sync(void *arg, dmu_tx_t *tx)
1547 {
1548 dmu_recv_begin_arg_t *drba = arg;
1549 dsl_pool_t *dp = dmu_tx_pool(tx);
1550 objset_t *mos = dp->dp_meta_objset;
1551 struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1552 const char *tofs = drba->drba_cookie->drc_tofs;
1553 dsl_dataset_t *ds, *newds;
1554 uint64_t dsobj;
1555 int error;
1556 uint64_t crflags = 0;
1557
1558 if (drrb->drr_flags & DRR_FLAG_CI_DATA)
1559 crflags |= DS_FLAG_CI_DATASET;
1560
1561 error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1562 if (error == 0) {
1563 /* create temporary clone */
1564 dsl_dataset_t *snap = NULL;
1565 if (drba->drba_snapobj != 0) {
1566 VERIFY0(dsl_dataset_hold_obj(dp,
1567 drba->drba_snapobj, FTAG, &snap));
1568 }
1569 dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name,
1570 snap, crflags, drba->drba_cred, tx);
1571 if (drba->drba_snapobj != 0)
1572 dsl_dataset_rele(snap, FTAG);
1573 dsl_dataset_rele(ds, FTAG);
1574 } else {
1575 dsl_dir_t *dd;
1576 const char *tail;
1577 dsl_dataset_t *origin = NULL;
1578
1579 VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail));
1580
1581 if (drba->drba_origin != NULL) {
1582 VERIFY0(dsl_dataset_hold(dp, drba->drba_origin,
1583 FTAG, &origin));
1584 }
1585
1586 /* Create new dataset. */
1587 dsobj = dsl_dataset_create_sync(dd,
1588 strrchr(tofs, '/') + 1,
1589 origin, crflags, drba->drba_cred, tx);
1590 if (origin != NULL)
1591 dsl_dataset_rele(origin, FTAG);
1592 dsl_dir_rele(dd, FTAG);
1593 drba->drba_cookie->drc_newfs = B_TRUE;
1594 }
1595 VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds));
1596
1597 if (drba->drba_cookie->drc_resumable) {
1598 dsl_dataset_zapify(newds, tx);
1599 if (drrb->drr_fromguid != 0) {
1600 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_FROMGUID,
1601 8, 1, &drrb->drr_fromguid, tx));
1602 }
1603 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TOGUID,
1604 8, 1, &drrb->drr_toguid, tx));
1605 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TONAME,
1606 1, strlen(drrb->drr_toname) + 1, drrb->drr_toname, tx));
1607 uint64_t one = 1;
1608 uint64_t zero = 0;
1609 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OBJECT,
1610 8, 1, &one, tx));
1611 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OFFSET,
1612 8, 1, &zero, tx));
1613 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_BYTES,
1614 8, 1, &zero, tx));
1615 if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
1616 DMU_BACKUP_FEATURE_LARGE_BLOCKS) {
1617 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_LARGEBLOCK,
1618 8, 1, &one, tx));
1619 }
1620 if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
1621 DMU_BACKUP_FEATURE_EMBED_DATA) {
1622 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_EMBEDOK,
1623 8, 1, &one, tx));
1624 }
1625 if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
1626 DMU_BACKUP_FEATURE_COMPRESSED) {
1627 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_COMPRESSOK,
1628 8, 1, &one, tx));
1629 }
1630 }
1631
1632 dmu_buf_will_dirty(newds->ds_dbuf, tx);
1633 dsl_dataset_phys(newds)->ds_flags |= DS_FLAG_INCONSISTENT;
1634
1635 /*
1636 * If we actually created a non-clone, we need to create the
1637 * objset in our new dataset.
1638 */
1639 rrw_enter(&newds->ds_bp_rwlock, RW_READER, FTAG);
1640 if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) {
1641 (void) dmu_objset_create_impl(dp->dp_spa,
1642 newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx);
1643 }
1644 rrw_exit(&newds->ds_bp_rwlock, FTAG);
1645
1646 drba->drba_cookie->drc_ds = newds;
1647
1648 spa_history_log_internal_ds(newds, "receive", tx, "");
1649 }
1650
1651 static int
dmu_recv_resume_begin_check(void * arg,dmu_tx_t * tx)1652 dmu_recv_resume_begin_check(void *arg, dmu_tx_t *tx)
1653 {
1654 dmu_recv_begin_arg_t *drba = arg;
1655 dsl_pool_t *dp = dmu_tx_pool(tx);
1656 struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1657 int error;
1658 uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1659 dsl_dataset_t *ds;
1660 const char *tofs = drba->drba_cookie->drc_tofs;
1661
1662 /* already checked */
1663 ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1664 ASSERT(featureflags & DMU_BACKUP_FEATURE_RESUMING);
1665
1666 if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1667 DMU_COMPOUNDSTREAM ||
1668 drrb->drr_type >= DMU_OST_NUMTYPES)
1669 return (SET_ERROR(EINVAL));
1670
1671 /* Verify pool version supports SA if SA_SPILL feature set */
1672 if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1673 spa_version(dp->dp_spa) < SPA_VERSION_SA)
1674 return (SET_ERROR(ENOTSUP));
1675
1676 /*
1677 * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1678 * record to a plain WRITE record, so the pool must have the
1679 * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1680 * records. Same with WRITE_EMBEDDED records that use LZ4 compression.
1681 */
1682 if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1683 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1684 return (SET_ERROR(ENOTSUP));
1685 if ((featureflags & DMU_BACKUP_FEATURE_LZ4) &&
1686 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1687 return (SET_ERROR(ENOTSUP));
1688
1689 /* 6 extra bytes for /%recv */
1690 char recvname[ZFS_MAX_DATASET_NAME_LEN + 6];
1691
1692 (void) snprintf(recvname, sizeof (recvname), "%s/%s",
1693 tofs, recv_clone_name);
1694
1695 if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
1696 /* %recv does not exist; continue in tofs */
1697 error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1698 if (error != 0)
1699 return (error);
1700 }
1701
1702 /* check that ds is marked inconsistent */
1703 if (!DS_IS_INCONSISTENT(ds)) {
1704 dsl_dataset_rele(ds, FTAG);
1705 return (SET_ERROR(EINVAL));
1706 }
1707
1708 /* check that there is resuming data, and that the toguid matches */
1709 if (!dsl_dataset_is_zapified(ds)) {
1710 dsl_dataset_rele(ds, FTAG);
1711 return (SET_ERROR(EINVAL));
1712 }
1713 uint64_t val;
1714 error = zap_lookup(dp->dp_meta_objset, ds->ds_object,
1715 DS_FIELD_RESUME_TOGUID, sizeof (val), 1, &val);
1716 if (error != 0 || drrb->drr_toguid != val) {
1717 dsl_dataset_rele(ds, FTAG);
1718 return (SET_ERROR(EINVAL));
1719 }
1720
1721 /*
1722 * Check if the receive is still running. If so, it will be owned.
1723 * Note that nothing else can own the dataset (e.g. after the receive
1724 * fails) because it will be marked inconsistent.
1725 */
1726 if (dsl_dataset_has_owner(ds)) {
1727 dsl_dataset_rele(ds, FTAG);
1728 return (SET_ERROR(EBUSY));
1729 }
1730
1731 /* There should not be any snapshots of this fs yet. */
1732 if (ds->ds_prev != NULL && ds->ds_prev->ds_dir == ds->ds_dir) {
1733 dsl_dataset_rele(ds, FTAG);
1734 return (SET_ERROR(EINVAL));
1735 }
1736
1737 /*
1738 * Note: resume point will be checked when we process the first WRITE
1739 * record.
1740 */
1741
1742 /* check that the origin matches */
1743 val = 0;
1744 (void) zap_lookup(dp->dp_meta_objset, ds->ds_object,
1745 DS_FIELD_RESUME_FROMGUID, sizeof (val), 1, &val);
1746 if (drrb->drr_fromguid != val) {
1747 dsl_dataset_rele(ds, FTAG);
1748 return (SET_ERROR(EINVAL));
1749 }
1750
1751 dsl_dataset_rele(ds, FTAG);
1752 return (0);
1753 }
1754
1755 static void
dmu_recv_resume_begin_sync(void * arg,dmu_tx_t * tx)1756 dmu_recv_resume_begin_sync(void *arg, dmu_tx_t *tx)
1757 {
1758 dmu_recv_begin_arg_t *drba = arg;
1759 dsl_pool_t *dp = dmu_tx_pool(tx);
1760 const char *tofs = drba->drba_cookie->drc_tofs;
1761 dsl_dataset_t *ds;
1762 uint64_t dsobj;
1763 /* 6 extra bytes for /%recv */
1764 char recvname[ZFS_MAX_DATASET_NAME_LEN + 6];
1765
1766 (void) snprintf(recvname, sizeof (recvname), "%s/%s",
1767 tofs, recv_clone_name);
1768
1769 if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
1770 /* %recv does not exist; continue in tofs */
1771 VERIFY0(dsl_dataset_hold(dp, tofs, FTAG, &ds));
1772 drba->drba_cookie->drc_newfs = B_TRUE;
1773 }
1774
1775 /* clear the inconsistent flag so that we can own it */
1776 ASSERT(DS_IS_INCONSISTENT(ds));
1777 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1778 dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
1779 dsobj = ds->ds_object;
1780 dsl_dataset_rele(ds, FTAG);
1781
1782 VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &ds));
1783
1784 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1785 dsl_dataset_phys(ds)->ds_flags |= DS_FLAG_INCONSISTENT;
1786
1787 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
1788 ASSERT(!BP_IS_HOLE(dsl_dataset_get_blkptr(ds)));
1789 rrw_exit(&ds->ds_bp_rwlock, FTAG);
1790
1791 drba->drba_cookie->drc_ds = ds;
1792
1793 spa_history_log_internal_ds(ds, "resume receive", tx, "");
1794 }
1795
1796 /*
1797 * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
1798 * succeeds; otherwise we will leak the holds on the datasets.
1799 */
1800 int
dmu_recv_begin(char * tofs,char * tosnap,dmu_replay_record_t * drr_begin,boolean_t force,boolean_t resumable,char * origin,dmu_recv_cookie_t * drc)1801 dmu_recv_begin(char *tofs, char *tosnap, dmu_replay_record_t *drr_begin,
1802 boolean_t force, boolean_t resumable, char *origin, dmu_recv_cookie_t *drc)
1803 {
1804 dmu_recv_begin_arg_t drba = { 0 };
1805
1806 bzero(drc, sizeof (dmu_recv_cookie_t));
1807 drc->drc_drr_begin = drr_begin;
1808 drc->drc_drrb = &drr_begin->drr_u.drr_begin;
1809 drc->drc_tosnap = tosnap;
1810 drc->drc_tofs = tofs;
1811 drc->drc_force = force;
1812 drc->drc_resumable = resumable;
1813 drc->drc_cred = CRED();
1814 drc->drc_clone = (origin != NULL);
1815
1816 if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC)) {
1817 drc->drc_byteswap = B_TRUE;
1818 (void) fletcher_4_incremental_byteswap(drr_begin,
1819 sizeof (dmu_replay_record_t), &drc->drc_cksum);
1820 byteswap_record(drr_begin);
1821 } else if (drc->drc_drrb->drr_magic == DMU_BACKUP_MAGIC) {
1822 (void) fletcher_4_incremental_native(drr_begin,
1823 sizeof (dmu_replay_record_t), &drc->drc_cksum);
1824 } else {
1825 return (SET_ERROR(EINVAL));
1826 }
1827
1828 drba.drba_origin = origin;
1829 drba.drba_cookie = drc;
1830 drba.drba_cred = CRED();
1831
1832 if (DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo) &
1833 DMU_BACKUP_FEATURE_RESUMING) {
1834 return (dsl_sync_task(tofs,
1835 dmu_recv_resume_begin_check, dmu_recv_resume_begin_sync,
1836 &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1837 } else {
1838 return (dsl_sync_task(tofs,
1839 dmu_recv_begin_check, dmu_recv_begin_sync,
1840 &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1841 }
1842 }
1843
1844 struct receive_record_arg {
1845 dmu_replay_record_t header;
1846 void *payload; /* Pointer to a buffer containing the payload */
1847 /*
1848 * If the record is a write, pointer to the arc_buf_t containing the
1849 * payload.
1850 */
1851 arc_buf_t *write_buf;
1852 int payload_size;
1853 uint64_t bytes_read; /* bytes read from stream when record created */
1854 boolean_t eos_marker; /* Marks the end of the stream */
1855 bqueue_node_t node;
1856 };
1857
1858 struct receive_writer_arg {
1859 objset_t *os;
1860 boolean_t byteswap;
1861 bqueue_t q;
1862
1863 /*
1864 * These three args are used to signal to the main thread that we're
1865 * done.
1866 */
1867 kmutex_t mutex;
1868 kcondvar_t cv;
1869 boolean_t done;
1870
1871 int err;
1872 /* A map from guid to dataset to help handle dedup'd streams. */
1873 avl_tree_t *guid_to_ds_map;
1874 boolean_t resumable;
1875 uint64_t last_object;
1876 uint64_t last_offset;
1877 uint64_t max_object; /* highest object ID referenced in stream */
1878 uint64_t bytes_read; /* bytes read when current record created */
1879 };
1880
1881 struct objlist {
1882 list_t list; /* List of struct receive_objnode. */
1883 /*
1884 * Last object looked up. Used to assert that objects are being looked
1885 * up in ascending order.
1886 */
1887 uint64_t last_lookup;
1888 };
1889
1890 struct receive_objnode {
1891 list_node_t node;
1892 uint64_t object;
1893 };
1894
1895 struct receive_arg {
1896 objset_t *os;
1897 kthread_t *td;
1898 struct file *fp;
1899 uint64_t voff; /* The current offset in the stream */
1900 uint64_t bytes_read;
1901 /*
1902 * A record that has had its payload read in, but hasn't yet been handed
1903 * off to the worker thread.
1904 */
1905 struct receive_record_arg *rrd;
1906 /* A record that has had its header read in, but not its payload. */
1907 struct receive_record_arg *next_rrd;
1908 zio_cksum_t cksum;
1909 zio_cksum_t prev_cksum;
1910 int err;
1911 boolean_t byteswap;
1912 /* Sorted list of objects not to issue prefetches for. */
1913 struct objlist ignore_objlist;
1914 };
1915
1916 typedef struct guid_map_entry {
1917 uint64_t guid;
1918 dsl_dataset_t *gme_ds;
1919 avl_node_t avlnode;
1920 } guid_map_entry_t;
1921
1922 static int
guid_compare(const void * arg1,const void * arg2)1923 guid_compare(const void *arg1, const void *arg2)
1924 {
1925 const guid_map_entry_t *gmep1 = (const guid_map_entry_t *)arg1;
1926 const guid_map_entry_t *gmep2 = (const guid_map_entry_t *)arg2;
1927
1928 return (AVL_CMP(gmep1->guid, gmep2->guid));
1929 }
1930
1931 static void
free_guid_map_onexit(void * arg)1932 free_guid_map_onexit(void *arg)
1933 {
1934 avl_tree_t *ca = arg;
1935 void *cookie = NULL;
1936 guid_map_entry_t *gmep;
1937
1938 while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
1939 dsl_dataset_long_rele(gmep->gme_ds, gmep);
1940 dsl_dataset_rele(gmep->gme_ds, gmep);
1941 kmem_free(gmep, sizeof (guid_map_entry_t));
1942 }
1943 avl_destroy(ca);
1944 kmem_free(ca, sizeof (avl_tree_t));
1945 }
1946
1947 static int
restore_bytes(struct receive_arg * ra,void * buf,int len,off_t off,ssize_t * resid)1948 restore_bytes(struct receive_arg *ra, void *buf, int len, off_t off, ssize_t *resid)
1949 {
1950 struct uio auio;
1951 struct iovec aiov;
1952 int error;
1953
1954 aiov.iov_base = buf;
1955 aiov.iov_len = len;
1956 auio.uio_iov = &aiov;
1957 auio.uio_iovcnt = 1;
1958 auio.uio_resid = len;
1959 auio.uio_segflg = UIO_SYSSPACE;
1960 auio.uio_rw = UIO_READ;
1961 auio.uio_offset = off;
1962 auio.uio_td = ra->td;
1963 #ifdef _KERNEL
1964 error = fo_read(ra->fp, &auio, ra->td->td_ucred, FOF_OFFSET, ra->td);
1965 #else
1966 fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__);
1967 error = EOPNOTSUPP;
1968 #endif
1969 *resid = auio.uio_resid;
1970 return (error);
1971 }
1972
1973 static int
receive_read(struct receive_arg * ra,int len,void * buf)1974 receive_read(struct receive_arg *ra, int len, void *buf)
1975 {
1976 int done = 0;
1977
1978 /*
1979 * The code doesn't rely on this (lengths being multiples of 8). See
1980 * comment in dump_bytes.
1981 */
1982 ASSERT0(len % 8);
1983
1984 while (done < len) {
1985 ssize_t resid;
1986
1987 ra->err = restore_bytes(ra, buf + done,
1988 len - done, ra->voff, &resid);
1989
1990 if (resid == len - done) {
1991 /*
1992 * Note: ECKSUM indicates that the receive
1993 * was interrupted and can potentially be resumed.
1994 */
1995 ra->err = SET_ERROR(ECKSUM);
1996 }
1997 ra->voff += len - done - resid;
1998 done = len - resid;
1999 if (ra->err != 0)
2000 return (ra->err);
2001 }
2002
2003 ra->bytes_read += len;
2004
2005 ASSERT3U(done, ==, len);
2006 return (0);
2007 }
2008
2009 noinline static void
byteswap_record(dmu_replay_record_t * drr)2010 byteswap_record(dmu_replay_record_t *drr)
2011 {
2012 #define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
2013 #define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
2014 drr->drr_type = BSWAP_32(drr->drr_type);
2015 drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
2016
2017 switch (drr->drr_type) {
2018 case DRR_BEGIN:
2019 DO64(drr_begin.drr_magic);
2020 DO64(drr_begin.drr_versioninfo);
2021 DO64(drr_begin.drr_creation_time);
2022 DO32(drr_begin.drr_type);
2023 DO32(drr_begin.drr_flags);
2024 DO64(drr_begin.drr_toguid);
2025 DO64(drr_begin.drr_fromguid);
2026 break;
2027 case DRR_OBJECT:
2028 DO64(drr_object.drr_object);
2029 DO32(drr_object.drr_type);
2030 DO32(drr_object.drr_bonustype);
2031 DO32(drr_object.drr_blksz);
2032 DO32(drr_object.drr_bonuslen);
2033 DO64(drr_object.drr_toguid);
2034 break;
2035 case DRR_FREEOBJECTS:
2036 DO64(drr_freeobjects.drr_firstobj);
2037 DO64(drr_freeobjects.drr_numobjs);
2038 DO64(drr_freeobjects.drr_toguid);
2039 break;
2040 case DRR_WRITE:
2041 DO64(drr_write.drr_object);
2042 DO32(drr_write.drr_type);
2043 DO64(drr_write.drr_offset);
2044 DO64(drr_write.drr_logical_size);
2045 DO64(drr_write.drr_toguid);
2046 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write.drr_key.ddk_cksum);
2047 DO64(drr_write.drr_key.ddk_prop);
2048 DO64(drr_write.drr_compressed_size);
2049 break;
2050 case DRR_WRITE_BYREF:
2051 DO64(drr_write_byref.drr_object);
2052 DO64(drr_write_byref.drr_offset);
2053 DO64(drr_write_byref.drr_length);
2054 DO64(drr_write_byref.drr_toguid);
2055 DO64(drr_write_byref.drr_refguid);
2056 DO64(drr_write_byref.drr_refobject);
2057 DO64(drr_write_byref.drr_refoffset);
2058 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write_byref.
2059 drr_key.ddk_cksum);
2060 DO64(drr_write_byref.drr_key.ddk_prop);
2061 break;
2062 case DRR_WRITE_EMBEDDED:
2063 DO64(drr_write_embedded.drr_object);
2064 DO64(drr_write_embedded.drr_offset);
2065 DO64(drr_write_embedded.drr_length);
2066 DO64(drr_write_embedded.drr_toguid);
2067 DO32(drr_write_embedded.drr_lsize);
2068 DO32(drr_write_embedded.drr_psize);
2069 break;
2070 case DRR_FREE:
2071 DO64(drr_free.drr_object);
2072 DO64(drr_free.drr_offset);
2073 DO64(drr_free.drr_length);
2074 DO64(drr_free.drr_toguid);
2075 break;
2076 case DRR_SPILL:
2077 DO64(drr_spill.drr_object);
2078 DO64(drr_spill.drr_length);
2079 DO64(drr_spill.drr_toguid);
2080 break;
2081 case DRR_END:
2082 DO64(drr_end.drr_toguid);
2083 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_end.drr_checksum);
2084 break;
2085 }
2086
2087 if (drr->drr_type != DRR_BEGIN) {
2088 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_checksum.drr_checksum);
2089 }
2090
2091 #undef DO64
2092 #undef DO32
2093 }
2094
2095 static inline uint8_t
deduce_nblkptr(dmu_object_type_t bonus_type,uint64_t bonus_size)2096 deduce_nblkptr(dmu_object_type_t bonus_type, uint64_t bonus_size)
2097 {
2098 if (bonus_type == DMU_OT_SA) {
2099 return (1);
2100 } else {
2101 return (1 +
2102 ((DN_OLD_MAX_BONUSLEN -
2103 MIN(DN_OLD_MAX_BONUSLEN, bonus_size)) >> SPA_BLKPTRSHIFT));
2104 }
2105 }
2106
2107 static void
save_resume_state(struct receive_writer_arg * rwa,uint64_t object,uint64_t offset,dmu_tx_t * tx)2108 save_resume_state(struct receive_writer_arg *rwa,
2109 uint64_t object, uint64_t offset, dmu_tx_t *tx)
2110 {
2111 int txgoff = dmu_tx_get_txg(tx) & TXG_MASK;
2112
2113 if (!rwa->resumable)
2114 return;
2115
2116 /*
2117 * We use ds_resume_bytes[] != 0 to indicate that we need to
2118 * update this on disk, so it must not be 0.
2119 */
2120 ASSERT(rwa->bytes_read != 0);
2121
2122 /*
2123 * We only resume from write records, which have a valid
2124 * (non-meta-dnode) object number.
2125 */
2126 ASSERT(object != 0);
2127
2128 /*
2129 * For resuming to work correctly, we must receive records in order,
2130 * sorted by object,offset. This is checked by the callers, but
2131 * assert it here for good measure.
2132 */
2133 ASSERT3U(object, >=, rwa->os->os_dsl_dataset->ds_resume_object[txgoff]);
2134 ASSERT(object != rwa->os->os_dsl_dataset->ds_resume_object[txgoff] ||
2135 offset >= rwa->os->os_dsl_dataset->ds_resume_offset[txgoff]);
2136 ASSERT3U(rwa->bytes_read, >=,
2137 rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff]);
2138
2139 rwa->os->os_dsl_dataset->ds_resume_object[txgoff] = object;
2140 rwa->os->os_dsl_dataset->ds_resume_offset[txgoff] = offset;
2141 rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff] = rwa->bytes_read;
2142 }
2143
2144 noinline static int
receive_object(struct receive_writer_arg * rwa,struct drr_object * drro,void * data)2145 receive_object(struct receive_writer_arg *rwa, struct drr_object *drro,
2146 void *data)
2147 {
2148 dmu_object_info_t doi;
2149 dmu_tx_t *tx;
2150 uint64_t object;
2151 int err;
2152
2153 if (drro->drr_type == DMU_OT_NONE ||
2154 !DMU_OT_IS_VALID(drro->drr_type) ||
2155 !DMU_OT_IS_VALID(drro->drr_bonustype) ||
2156 drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
2157 drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
2158 P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
2159 drro->drr_blksz < SPA_MINBLOCKSIZE ||
2160 drro->drr_blksz > spa_maxblocksize(dmu_objset_spa(rwa->os)) ||
2161 drro->drr_bonuslen >
2162 DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(rwa->os)))) {
2163 return (SET_ERROR(EINVAL));
2164 }
2165
2166 err = dmu_object_info(rwa->os, drro->drr_object, &doi);
2167
2168 if (err != 0 && err != ENOENT)
2169 return (SET_ERROR(EINVAL));
2170 object = err == 0 ? drro->drr_object : DMU_NEW_OBJECT;
2171
2172 if (drro->drr_object > rwa->max_object)
2173 rwa->max_object = drro->drr_object;
2174
2175 /*
2176 * If we are losing blkptrs or changing the block size this must
2177 * be a new file instance. We must clear out the previous file
2178 * contents before we can change this type of metadata in the dnode.
2179 */
2180 if (err == 0) {
2181 int nblkptr;
2182
2183 nblkptr = deduce_nblkptr(drro->drr_bonustype,
2184 drro->drr_bonuslen);
2185
2186 if (drro->drr_blksz != doi.doi_data_block_size ||
2187 nblkptr < doi.doi_nblkptr) {
2188 err = dmu_free_long_range(rwa->os, drro->drr_object,
2189 0, DMU_OBJECT_END);
2190 if (err != 0)
2191 return (SET_ERROR(EINVAL));
2192 }
2193 }
2194
2195 tx = dmu_tx_create(rwa->os);
2196 dmu_tx_hold_bonus(tx, object);
2197 err = dmu_tx_assign(tx, TXG_WAIT);
2198 if (err != 0) {
2199 dmu_tx_abort(tx);
2200 return (err);
2201 }
2202
2203 if (object == DMU_NEW_OBJECT) {
2204 /* currently free, want to be allocated */
2205 err = dmu_object_claim_dnsize(rwa->os, drro->drr_object,
2206 drro->drr_type, drro->drr_blksz,
2207 drro->drr_bonustype, drro->drr_bonuslen,
2208 drro->drr_dn_slots << DNODE_SHIFT, tx);
2209 } else if (drro->drr_type != doi.doi_type ||
2210 drro->drr_blksz != doi.doi_data_block_size ||
2211 drro->drr_bonustype != doi.doi_bonus_type ||
2212 drro->drr_bonuslen != doi.doi_bonus_size) {
2213 /* currently allocated, but with different properties */
2214 err = dmu_object_reclaim(rwa->os, drro->drr_object,
2215 drro->drr_type, drro->drr_blksz,
2216 drro->drr_bonustype, drro->drr_bonuslen, tx);
2217 }
2218 if (err != 0) {
2219 dmu_tx_commit(tx);
2220 return (SET_ERROR(EINVAL));
2221 }
2222
2223 dmu_object_set_checksum(rwa->os, drro->drr_object,
2224 drro->drr_checksumtype, tx);
2225 dmu_object_set_compress(rwa->os, drro->drr_object,
2226 drro->drr_compress, tx);
2227
2228 if (data != NULL) {
2229 dmu_buf_t *db;
2230
2231 VERIFY0(dmu_bonus_hold(rwa->os, drro->drr_object, FTAG, &db));
2232 dmu_buf_will_dirty(db, tx);
2233
2234 ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
2235 bcopy(data, db->db_data, drro->drr_bonuslen);
2236 if (rwa->byteswap) {
2237 dmu_object_byteswap_t byteswap =
2238 DMU_OT_BYTESWAP(drro->drr_bonustype);
2239 dmu_ot_byteswap[byteswap].ob_func(db->db_data,
2240 drro->drr_bonuslen);
2241 }
2242 dmu_buf_rele(db, FTAG);
2243 }
2244 dmu_tx_commit(tx);
2245
2246 return (0);
2247 }
2248
2249 /* ARGSUSED */
2250 noinline static int
receive_freeobjects(struct receive_writer_arg * rwa,struct drr_freeobjects * drrfo)2251 receive_freeobjects(struct receive_writer_arg *rwa,
2252 struct drr_freeobjects *drrfo)
2253 {
2254 uint64_t obj;
2255 int next_err = 0;
2256
2257 if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
2258 return (SET_ERROR(EINVAL));
2259
2260 for (obj = drrfo->drr_firstobj == 0 ? 1 : drrfo->drr_firstobj;
2261 obj < drrfo->drr_firstobj + drrfo->drr_numobjs && next_err == 0;
2262 next_err = dmu_object_next(rwa->os, &obj, FALSE, 0)) {
2263 dmu_object_info_t doi;
2264 int err;
2265
2266 err = dmu_object_info(rwa->os, obj, &doi);
2267 if (err == ENOENT) {
2268 obj++;
2269 continue;
2270 } else if (err != 0) {
2271 return (err);
2272 }
2273
2274 err = dmu_free_long_object(rwa->os, obj);
2275 if (err != 0)
2276 return (err);
2277
2278 if (obj > rwa->max_object)
2279 rwa->max_object = obj;
2280 }
2281 if (next_err != ESRCH)
2282 return (next_err);
2283 return (0);
2284 }
2285
2286 noinline static int
receive_write(struct receive_writer_arg * rwa,struct drr_write * drrw,arc_buf_t * abuf)2287 receive_write(struct receive_writer_arg *rwa, struct drr_write *drrw,
2288 arc_buf_t *abuf)
2289 {
2290 dmu_tx_t *tx;
2291 int err;
2292
2293 if (drrw->drr_offset + drrw->drr_logical_size < drrw->drr_offset ||
2294 !DMU_OT_IS_VALID(drrw->drr_type))
2295 return (SET_ERROR(EINVAL));
2296
2297 /*
2298 * For resuming to work, records must be in increasing order
2299 * by (object, offset).
2300 */
2301 if (drrw->drr_object < rwa->last_object ||
2302 (drrw->drr_object == rwa->last_object &&
2303 drrw->drr_offset < rwa->last_offset)) {
2304 return (SET_ERROR(EINVAL));
2305 }
2306 rwa->last_object = drrw->drr_object;
2307 rwa->last_offset = drrw->drr_offset;
2308
2309 if (rwa->last_object > rwa->max_object)
2310 rwa->max_object = rwa->last_object;
2311
2312 if (dmu_object_info(rwa->os, drrw->drr_object, NULL) != 0)
2313 return (SET_ERROR(EINVAL));
2314
2315 tx = dmu_tx_create(rwa->os);
2316 dmu_tx_hold_write(tx, drrw->drr_object,
2317 drrw->drr_offset, drrw->drr_logical_size);
2318 err = dmu_tx_assign(tx, TXG_WAIT);
2319 if (err != 0) {
2320 dmu_tx_abort(tx);
2321 return (err);
2322 }
2323 if (rwa->byteswap) {
2324 dmu_object_byteswap_t byteswap =
2325 DMU_OT_BYTESWAP(drrw->drr_type);
2326 dmu_ot_byteswap[byteswap].ob_func(abuf->b_data,
2327 DRR_WRITE_PAYLOAD_SIZE(drrw));
2328 }
2329
2330 /* use the bonus buf to look up the dnode in dmu_assign_arcbuf */
2331 dmu_buf_t *bonus;
2332 if (dmu_bonus_hold(rwa->os, drrw->drr_object, FTAG, &bonus) != 0)
2333 return (SET_ERROR(EINVAL));
2334 dmu_assign_arcbuf(bonus, drrw->drr_offset, abuf, tx);
2335
2336 /*
2337 * Note: If the receive fails, we want the resume stream to start
2338 * with the same record that we last successfully received (as opposed
2339 * to the next record), so that we can verify that we are
2340 * resuming from the correct location.
2341 */
2342 save_resume_state(rwa, drrw->drr_object, drrw->drr_offset, tx);
2343 dmu_tx_commit(tx);
2344 dmu_buf_rele(bonus, FTAG);
2345
2346 return (0);
2347 }
2348
2349 /*
2350 * Handle a DRR_WRITE_BYREF record. This record is used in dedup'ed
2351 * streams to refer to a copy of the data that is already on the
2352 * system because it came in earlier in the stream. This function
2353 * finds the earlier copy of the data, and uses that copy instead of
2354 * data from the stream to fulfill this write.
2355 */
2356 static int
receive_write_byref(struct receive_writer_arg * rwa,struct drr_write_byref * drrwbr)2357 receive_write_byref(struct receive_writer_arg *rwa,
2358 struct drr_write_byref *drrwbr)
2359 {
2360 dmu_tx_t *tx;
2361 int err;
2362 guid_map_entry_t gmesrch;
2363 guid_map_entry_t *gmep;
2364 avl_index_t where;
2365 objset_t *ref_os = NULL;
2366 dmu_buf_t *dbp;
2367
2368 if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
2369 return (SET_ERROR(EINVAL));
2370
2371 /*
2372 * If the GUID of the referenced dataset is different from the
2373 * GUID of the target dataset, find the referenced dataset.
2374 */
2375 if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
2376 gmesrch.guid = drrwbr->drr_refguid;
2377 if ((gmep = avl_find(rwa->guid_to_ds_map, &gmesrch,
2378 &where)) == NULL) {
2379 return (SET_ERROR(EINVAL));
2380 }
2381 if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
2382 return (SET_ERROR(EINVAL));
2383 } else {
2384 ref_os = rwa->os;
2385 }
2386
2387 if (drrwbr->drr_object > rwa->max_object)
2388 rwa->max_object = drrwbr->drr_object;
2389
2390 err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
2391 drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH);
2392 if (err != 0)
2393 return (err);
2394
2395 tx = dmu_tx_create(rwa->os);
2396
2397 dmu_tx_hold_write(tx, drrwbr->drr_object,
2398 drrwbr->drr_offset, drrwbr->drr_length);
2399 err = dmu_tx_assign(tx, TXG_WAIT);
2400 if (err != 0) {
2401 dmu_tx_abort(tx);
2402 return (err);
2403 }
2404 dmu_write(rwa->os, drrwbr->drr_object,
2405 drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
2406 dmu_buf_rele(dbp, FTAG);
2407
2408 /* See comment in restore_write. */
2409 save_resume_state(rwa, drrwbr->drr_object, drrwbr->drr_offset, tx);
2410 dmu_tx_commit(tx);
2411 return (0);
2412 }
2413
2414 static int
receive_write_embedded(struct receive_writer_arg * rwa,struct drr_write_embedded * drrwe,void * data)2415 receive_write_embedded(struct receive_writer_arg *rwa,
2416 struct drr_write_embedded *drrwe, void *data)
2417 {
2418 dmu_tx_t *tx;
2419 int err;
2420
2421 if (drrwe->drr_offset + drrwe->drr_length < drrwe->drr_offset)
2422 return (EINVAL);
2423
2424 if (drrwe->drr_psize > BPE_PAYLOAD_SIZE)
2425 return (EINVAL);
2426
2427 if (drrwe->drr_etype >= NUM_BP_EMBEDDED_TYPES)
2428 return (EINVAL);
2429 if (drrwe->drr_compression >= ZIO_COMPRESS_FUNCTIONS)
2430 return (EINVAL);
2431
2432 if (drrwe->drr_object > rwa->max_object)
2433 rwa->max_object = drrwe->drr_object;
2434
2435 tx = dmu_tx_create(rwa->os);
2436
2437 dmu_tx_hold_write(tx, drrwe->drr_object,
2438 drrwe->drr_offset, drrwe->drr_length);
2439 err = dmu_tx_assign(tx, TXG_WAIT);
2440 if (err != 0) {
2441 dmu_tx_abort(tx);
2442 return (err);
2443 }
2444
2445 dmu_write_embedded(rwa->os, drrwe->drr_object,
2446 drrwe->drr_offset, data, drrwe->drr_etype,
2447 drrwe->drr_compression, drrwe->drr_lsize, drrwe->drr_psize,
2448 rwa->byteswap ^ ZFS_HOST_BYTEORDER, tx);
2449
2450 /* See comment in restore_write. */
2451 save_resume_state(rwa, drrwe->drr_object, drrwe->drr_offset, tx);
2452 dmu_tx_commit(tx);
2453 return (0);
2454 }
2455
2456 static int
receive_spill(struct receive_writer_arg * rwa,struct drr_spill * drrs,void * data)2457 receive_spill(struct receive_writer_arg *rwa, struct drr_spill *drrs,
2458 void *data)
2459 {
2460 dmu_tx_t *tx;
2461 dmu_buf_t *db, *db_spill;
2462 int err;
2463
2464 if (drrs->drr_length < SPA_MINBLOCKSIZE ||
2465 drrs->drr_length > spa_maxblocksize(dmu_objset_spa(rwa->os)))
2466 return (SET_ERROR(EINVAL));
2467
2468 if (dmu_object_info(rwa->os, drrs->drr_object, NULL) != 0)
2469 return (SET_ERROR(EINVAL));
2470
2471 if (drrs->drr_object > rwa->max_object)
2472 rwa->max_object = drrs->drr_object;
2473
2474 VERIFY0(dmu_bonus_hold(rwa->os, drrs->drr_object, FTAG, &db));
2475 if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
2476 dmu_buf_rele(db, FTAG);
2477 return (err);
2478 }
2479
2480 tx = dmu_tx_create(rwa->os);
2481
2482 dmu_tx_hold_spill(tx, db->db_object);
2483
2484 err = dmu_tx_assign(tx, TXG_WAIT);
2485 if (err != 0) {
2486 dmu_buf_rele(db, FTAG);
2487 dmu_buf_rele(db_spill, FTAG);
2488 dmu_tx_abort(tx);
2489 return (err);
2490 }
2491 dmu_buf_will_dirty(db_spill, tx);
2492
2493 if (db_spill->db_size < drrs->drr_length)
2494 VERIFY(0 == dbuf_spill_set_blksz(db_spill,
2495 drrs->drr_length, tx));
2496 bcopy(data, db_spill->db_data, drrs->drr_length);
2497
2498 dmu_buf_rele(db, FTAG);
2499 dmu_buf_rele(db_spill, FTAG);
2500
2501 dmu_tx_commit(tx);
2502 return (0);
2503 }
2504
2505 /* ARGSUSED */
2506 noinline static int
receive_free(struct receive_writer_arg * rwa,struct drr_free * drrf)2507 receive_free(struct receive_writer_arg *rwa, struct drr_free *drrf)
2508 {
2509 int err;
2510
2511 if (drrf->drr_length != -1ULL &&
2512 drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
2513 return (SET_ERROR(EINVAL));
2514
2515 if (dmu_object_info(rwa->os, drrf->drr_object, NULL) != 0)
2516 return (SET_ERROR(EINVAL));
2517
2518 if (drrf->drr_object > rwa->max_object)
2519 rwa->max_object = drrf->drr_object;
2520
2521 err = dmu_free_long_range(rwa->os, drrf->drr_object,
2522 drrf->drr_offset, drrf->drr_length);
2523
2524 return (err);
2525 }
2526
2527 /* used to destroy the drc_ds on error */
2528 static void
dmu_recv_cleanup_ds(dmu_recv_cookie_t * drc)2529 dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc)
2530 {
2531 if (drc->drc_resumable) {
2532 /* wait for our resume state to be written to disk */
2533 txg_wait_synced(drc->drc_ds->ds_dir->dd_pool, 0);
2534 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2535 } else {
2536 char name[ZFS_MAX_DATASET_NAME_LEN];
2537 dsl_dataset_name(drc->drc_ds, name);
2538 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2539 (void) dsl_destroy_head(name);
2540 }
2541 }
2542
2543 static void
receive_cksum(struct receive_arg * ra,int len,void * buf)2544 receive_cksum(struct receive_arg *ra, int len, void *buf)
2545 {
2546 if (ra->byteswap) {
2547 (void) fletcher_4_incremental_byteswap(buf, len, &ra->cksum);
2548 } else {
2549 (void) fletcher_4_incremental_native(buf, len, &ra->cksum);
2550 }
2551 }
2552
2553 /*
2554 * Read the payload into a buffer of size len, and update the current record's
2555 * payload field.
2556 * Allocate ra->next_rrd and read the next record's header into
2557 * ra->next_rrd->header.
2558 * Verify checksum of payload and next record.
2559 */
2560 static int
receive_read_payload_and_next_header(struct receive_arg * ra,int len,void * buf)2561 receive_read_payload_and_next_header(struct receive_arg *ra, int len, void *buf)
2562 {
2563 int err;
2564
2565 if (len != 0) {
2566 ASSERT3U(len, <=, SPA_MAXBLOCKSIZE);
2567 err = receive_read(ra, len, buf);
2568 if (err != 0)
2569 return (err);
2570 receive_cksum(ra, len, buf);
2571
2572 /* note: rrd is NULL when reading the begin record's payload */
2573 if (ra->rrd != NULL) {
2574 ra->rrd->payload = buf;
2575 ra->rrd->payload_size = len;
2576 ra->rrd->bytes_read = ra->bytes_read;
2577 }
2578 }
2579
2580 ra->prev_cksum = ra->cksum;
2581
2582 ra->next_rrd = kmem_zalloc(sizeof (*ra->next_rrd), KM_SLEEP);
2583 err = receive_read(ra, sizeof (ra->next_rrd->header),
2584 &ra->next_rrd->header);
2585 ra->next_rrd->bytes_read = ra->bytes_read;
2586 if (err != 0) {
2587 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2588 ra->next_rrd = NULL;
2589 return (err);
2590 }
2591 if (ra->next_rrd->header.drr_type == DRR_BEGIN) {
2592 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2593 ra->next_rrd = NULL;
2594 return (SET_ERROR(EINVAL));
2595 }
2596
2597 /*
2598 * Note: checksum is of everything up to but not including the
2599 * checksum itself.
2600 */
2601 ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
2602 ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
2603 receive_cksum(ra,
2604 offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
2605 &ra->next_rrd->header);
2606
2607 zio_cksum_t cksum_orig =
2608 ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
2609 zio_cksum_t *cksump =
2610 &ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
2611
2612 if (ra->byteswap)
2613 byteswap_record(&ra->next_rrd->header);
2614
2615 if ((!ZIO_CHECKSUM_IS_ZERO(cksump)) &&
2616 !ZIO_CHECKSUM_EQUAL(ra->cksum, *cksump)) {
2617 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2618 ra->next_rrd = NULL;
2619 return (SET_ERROR(ECKSUM));
2620 }
2621
2622 receive_cksum(ra, sizeof (cksum_orig), &cksum_orig);
2623
2624 return (0);
2625 }
2626
2627 static void
objlist_create(struct objlist * list)2628 objlist_create(struct objlist *list)
2629 {
2630 list_create(&list->list, sizeof (struct receive_objnode),
2631 offsetof(struct receive_objnode, node));
2632 list->last_lookup = 0;
2633 }
2634
2635 static void
objlist_destroy(struct objlist * list)2636 objlist_destroy(struct objlist *list)
2637 {
2638 for (struct receive_objnode *n = list_remove_head(&list->list);
2639 n != NULL; n = list_remove_head(&list->list)) {
2640 kmem_free(n, sizeof (*n));
2641 }
2642 list_destroy(&list->list);
2643 }
2644
2645 /*
2646 * This function looks through the objlist to see if the specified object number
2647 * is contained in the objlist. In the process, it will remove all object
2648 * numbers in the list that are smaller than the specified object number. Thus,
2649 * any lookup of an object number smaller than a previously looked up object
2650 * number will always return false; therefore, all lookups should be done in
2651 * ascending order.
2652 */
2653 static boolean_t
objlist_exists(struct objlist * list,uint64_t object)2654 objlist_exists(struct objlist *list, uint64_t object)
2655 {
2656 struct receive_objnode *node = list_head(&list->list);
2657 ASSERT3U(object, >=, list->last_lookup);
2658 list->last_lookup = object;
2659 while (node != NULL && node->object < object) {
2660 VERIFY3P(node, ==, list_remove_head(&list->list));
2661 kmem_free(node, sizeof (*node));
2662 node = list_head(&list->list);
2663 }
2664 return (node != NULL && node->object == object);
2665 }
2666
2667 /*
2668 * The objlist is a list of object numbers stored in ascending order. However,
2669 * the insertion of new object numbers does not seek out the correct location to
2670 * store a new object number; instead, it appends it to the list for simplicity.
2671 * Thus, any users must take care to only insert new object numbers in ascending
2672 * order.
2673 */
2674 static void
objlist_insert(struct objlist * list,uint64_t object)2675 objlist_insert(struct objlist *list, uint64_t object)
2676 {
2677 struct receive_objnode *node = kmem_zalloc(sizeof (*node), KM_SLEEP);
2678 node->object = object;
2679 #ifdef ZFS_DEBUG
2680 struct receive_objnode *last_object = list_tail(&list->list);
2681 uint64_t last_objnum = (last_object != NULL ? last_object->object : 0);
2682 ASSERT3U(node->object, >, last_objnum);
2683 #endif
2684 list_insert_tail(&list->list, node);
2685 }
2686
2687 /*
2688 * Issue the prefetch reads for any necessary indirect blocks.
2689 *
2690 * We use the object ignore list to tell us whether or not to issue prefetches
2691 * for a given object. We do this for both correctness (in case the blocksize
2692 * of an object has changed) and performance (if the object doesn't exist, don't
2693 * needlessly try to issue prefetches). We also trim the list as we go through
2694 * the stream to prevent it from growing to an unbounded size.
2695 *
2696 * The object numbers within will always be in sorted order, and any write
2697 * records we see will also be in sorted order, but they're not sorted with
2698 * respect to each other (i.e. we can get several object records before
2699 * receiving each object's write records). As a result, once we've reached a
2700 * given object number, we can safely remove any reference to lower object
2701 * numbers in the ignore list. In practice, we receive up to 32 object records
2702 * before receiving write records, so the list can have up to 32 nodes in it.
2703 */
2704 /* ARGSUSED */
2705 static void
receive_read_prefetch(struct receive_arg * ra,uint64_t object,uint64_t offset,uint64_t length)2706 receive_read_prefetch(struct receive_arg *ra,
2707 uint64_t object, uint64_t offset, uint64_t length)
2708 {
2709 if (!objlist_exists(&ra->ignore_objlist, object)) {
2710 dmu_prefetch(ra->os, object, 1, offset, length,
2711 ZIO_PRIORITY_SYNC_READ);
2712 }
2713 }
2714
2715 /*
2716 * Read records off the stream, issuing any necessary prefetches.
2717 */
2718 static int
receive_read_record(struct receive_arg * ra)2719 receive_read_record(struct receive_arg *ra)
2720 {
2721 int err;
2722
2723 switch (ra->rrd->header.drr_type) {
2724 case DRR_OBJECT:
2725 {
2726 struct drr_object *drro = &ra->rrd->header.drr_u.drr_object;
2727 uint32_t size = P2ROUNDUP(drro->drr_bonuslen, 8);
2728 void *buf = kmem_zalloc(size, KM_SLEEP);
2729 dmu_object_info_t doi;
2730 err = receive_read_payload_and_next_header(ra, size, buf);
2731 if (err != 0) {
2732 kmem_free(buf, size);
2733 return (err);
2734 }
2735 err = dmu_object_info(ra->os, drro->drr_object, &doi);
2736 /*
2737 * See receive_read_prefetch for an explanation why we're
2738 * storing this object in the ignore_obj_list.
2739 */
2740 if (err == ENOENT ||
2741 (err == 0 && doi.doi_data_block_size != drro->drr_blksz)) {
2742 objlist_insert(&ra->ignore_objlist, drro->drr_object);
2743 err = 0;
2744 }
2745 return (err);
2746 }
2747 case DRR_FREEOBJECTS:
2748 {
2749 err = receive_read_payload_and_next_header(ra, 0, NULL);
2750 return (err);
2751 }
2752 case DRR_WRITE:
2753 {
2754 struct drr_write *drrw = &ra->rrd->header.drr_u.drr_write;
2755 arc_buf_t *abuf;
2756 boolean_t is_meta = DMU_OT_IS_METADATA(drrw->drr_type);
2757 if (DRR_WRITE_COMPRESSED(drrw)) {
2758 ASSERT3U(drrw->drr_compressed_size, >, 0);
2759 ASSERT3U(drrw->drr_logical_size, >=,
2760 drrw->drr_compressed_size);
2761 ASSERT(!is_meta);
2762 abuf = arc_loan_compressed_buf(
2763 dmu_objset_spa(ra->os),
2764 drrw->drr_compressed_size, drrw->drr_logical_size,
2765 drrw->drr_compressiontype);
2766 } else {
2767 abuf = arc_loan_buf(dmu_objset_spa(ra->os),
2768 is_meta, drrw->drr_logical_size);
2769 }
2770
2771 err = receive_read_payload_and_next_header(ra,
2772 DRR_WRITE_PAYLOAD_SIZE(drrw), abuf->b_data);
2773 if (err != 0) {
2774 dmu_return_arcbuf(abuf);
2775 return (err);
2776 }
2777 ra->rrd->write_buf = abuf;
2778 receive_read_prefetch(ra, drrw->drr_object, drrw->drr_offset,
2779 drrw->drr_logical_size);
2780 return (err);
2781 }
2782 case DRR_WRITE_BYREF:
2783 {
2784 struct drr_write_byref *drrwb =
2785 &ra->rrd->header.drr_u.drr_write_byref;
2786 err = receive_read_payload_and_next_header(ra, 0, NULL);
2787 receive_read_prefetch(ra, drrwb->drr_object, drrwb->drr_offset,
2788 drrwb->drr_length);
2789 return (err);
2790 }
2791 case DRR_WRITE_EMBEDDED:
2792 {
2793 struct drr_write_embedded *drrwe =
2794 &ra->rrd->header.drr_u.drr_write_embedded;
2795 uint32_t size = P2ROUNDUP(drrwe->drr_psize, 8);
2796 void *buf = kmem_zalloc(size, KM_SLEEP);
2797
2798 err = receive_read_payload_and_next_header(ra, size, buf);
2799 if (err != 0) {
2800 kmem_free(buf, size);
2801 return (err);
2802 }
2803
2804 receive_read_prefetch(ra, drrwe->drr_object, drrwe->drr_offset,
2805 drrwe->drr_length);
2806 return (err);
2807 }
2808 case DRR_FREE:
2809 {
2810 /*
2811 * It might be beneficial to prefetch indirect blocks here, but
2812 * we don't really have the data to decide for sure.
2813 */
2814 err = receive_read_payload_and_next_header(ra, 0, NULL);
2815 return (err);
2816 }
2817 case DRR_END:
2818 {
2819 struct drr_end *drre = &ra->rrd->header.drr_u.drr_end;
2820 if (!ZIO_CHECKSUM_EQUAL(ra->prev_cksum, drre->drr_checksum))
2821 return (SET_ERROR(ECKSUM));
2822 return (0);
2823 }
2824 case DRR_SPILL:
2825 {
2826 struct drr_spill *drrs = &ra->rrd->header.drr_u.drr_spill;
2827 void *buf = kmem_zalloc(drrs->drr_length, KM_SLEEP);
2828 err = receive_read_payload_and_next_header(ra, drrs->drr_length,
2829 buf);
2830 if (err != 0)
2831 kmem_free(buf, drrs->drr_length);
2832 return (err);
2833 }
2834 default:
2835 return (SET_ERROR(EINVAL));
2836 }
2837 }
2838
2839 /*
2840 * Commit the records to the pool.
2841 */
2842 static int
receive_process_record(struct receive_writer_arg * rwa,struct receive_record_arg * rrd)2843 receive_process_record(struct receive_writer_arg *rwa,
2844 struct receive_record_arg *rrd)
2845 {
2846 int err;
2847
2848 /* Processing in order, therefore bytes_read should be increasing. */
2849 ASSERT3U(rrd->bytes_read, >=, rwa->bytes_read);
2850 rwa->bytes_read = rrd->bytes_read;
2851
2852 switch (rrd->header.drr_type) {
2853 case DRR_OBJECT:
2854 {
2855 struct drr_object *drro = &rrd->header.drr_u.drr_object;
2856 err = receive_object(rwa, drro, rrd->payload);
2857 kmem_free(rrd->payload, rrd->payload_size);
2858 rrd->payload = NULL;
2859 return (err);
2860 }
2861 case DRR_FREEOBJECTS:
2862 {
2863 struct drr_freeobjects *drrfo =
2864 &rrd->header.drr_u.drr_freeobjects;
2865 return (receive_freeobjects(rwa, drrfo));
2866 }
2867 case DRR_WRITE:
2868 {
2869 struct drr_write *drrw = &rrd->header.drr_u.drr_write;
2870 err = receive_write(rwa, drrw, rrd->write_buf);
2871 /* if receive_write() is successful, it consumes the arc_buf */
2872 if (err != 0)
2873 dmu_return_arcbuf(rrd->write_buf);
2874 rrd->write_buf = NULL;
2875 rrd->payload = NULL;
2876 return (err);
2877 }
2878 case DRR_WRITE_BYREF:
2879 {
2880 struct drr_write_byref *drrwbr =
2881 &rrd->header.drr_u.drr_write_byref;
2882 return (receive_write_byref(rwa, drrwbr));
2883 }
2884 case DRR_WRITE_EMBEDDED:
2885 {
2886 struct drr_write_embedded *drrwe =
2887 &rrd->header.drr_u.drr_write_embedded;
2888 err = receive_write_embedded(rwa, drrwe, rrd->payload);
2889 kmem_free(rrd->payload, rrd->payload_size);
2890 rrd->payload = NULL;
2891 return (err);
2892 }
2893 case DRR_FREE:
2894 {
2895 struct drr_free *drrf = &rrd->header.drr_u.drr_free;
2896 return (receive_free(rwa, drrf));
2897 }
2898 case DRR_SPILL:
2899 {
2900 struct drr_spill *drrs = &rrd->header.drr_u.drr_spill;
2901 err = receive_spill(rwa, drrs, rrd->payload);
2902 kmem_free(rrd->payload, rrd->payload_size);
2903 rrd->payload = NULL;
2904 return (err);
2905 }
2906 default:
2907 return (SET_ERROR(EINVAL));
2908 }
2909 }
2910
2911 /*
2912 * dmu_recv_stream's worker thread; pull records off the queue, and then call
2913 * receive_process_record When we're done, signal the main thread and exit.
2914 */
2915 static void
receive_writer_thread(void * arg)2916 receive_writer_thread(void *arg)
2917 {
2918 struct receive_writer_arg *rwa = arg;
2919 struct receive_record_arg *rrd;
2920 for (rrd = bqueue_dequeue(&rwa->q); !rrd->eos_marker;
2921 rrd = bqueue_dequeue(&rwa->q)) {
2922 /*
2923 * If there's an error, the main thread will stop putting things
2924 * on the queue, but we need to clear everything in it before we
2925 * can exit.
2926 */
2927 if (rwa->err == 0) {
2928 rwa->err = receive_process_record(rwa, rrd);
2929 } else if (rrd->write_buf != NULL) {
2930 dmu_return_arcbuf(rrd->write_buf);
2931 rrd->write_buf = NULL;
2932 rrd->payload = NULL;
2933 } else if (rrd->payload != NULL) {
2934 kmem_free(rrd->payload, rrd->payload_size);
2935 rrd->payload = NULL;
2936 }
2937 kmem_free(rrd, sizeof (*rrd));
2938 }
2939 kmem_free(rrd, sizeof (*rrd));
2940 mutex_enter(&rwa->mutex);
2941 rwa->done = B_TRUE;
2942 cv_signal(&rwa->cv);
2943 mutex_exit(&rwa->mutex);
2944 thread_exit();
2945 }
2946
2947 static int
resume_check(struct receive_arg * ra,nvlist_t * begin_nvl)2948 resume_check(struct receive_arg *ra, nvlist_t *begin_nvl)
2949 {
2950 uint64_t val;
2951 objset_t *mos = dmu_objset_pool(ra->os)->dp_meta_objset;
2952 uint64_t dsobj = dmu_objset_id(ra->os);
2953 uint64_t resume_obj, resume_off;
2954
2955 if (nvlist_lookup_uint64(begin_nvl,
2956 "resume_object", &resume_obj) != 0 ||
2957 nvlist_lookup_uint64(begin_nvl,
2958 "resume_offset", &resume_off) != 0) {
2959 return (SET_ERROR(EINVAL));
2960 }
2961 VERIFY0(zap_lookup(mos, dsobj,
2962 DS_FIELD_RESUME_OBJECT, sizeof (val), 1, &val));
2963 if (resume_obj != val)
2964 return (SET_ERROR(EINVAL));
2965 VERIFY0(zap_lookup(mos, dsobj,
2966 DS_FIELD_RESUME_OFFSET, sizeof (val), 1, &val));
2967 if (resume_off != val)
2968 return (SET_ERROR(EINVAL));
2969
2970 return (0);
2971 }
2972
2973 /*
2974 * Read in the stream's records, one by one, and apply them to the pool. There
2975 * are two threads involved; the thread that calls this function will spin up a
2976 * worker thread, read the records off the stream one by one, and issue
2977 * prefetches for any necessary indirect blocks. It will then push the records
2978 * onto an internal blocking queue. The worker thread will pull the records off
2979 * the queue, and actually write the data into the DMU. This way, the worker
2980 * thread doesn't have to wait for reads to complete, since everything it needs
2981 * (the indirect blocks) will be prefetched.
2982 *
2983 * NB: callers *must* call dmu_recv_end() if this succeeds.
2984 */
2985 int
dmu_recv_stream(dmu_recv_cookie_t * drc,struct file * fp,offset_t * voffp,int cleanup_fd,uint64_t * action_handlep)2986 dmu_recv_stream(dmu_recv_cookie_t *drc, struct file *fp, offset_t *voffp,
2987 int cleanup_fd, uint64_t *action_handlep)
2988 {
2989 int err = 0;
2990 struct receive_arg ra = { 0 };
2991 struct receive_writer_arg rwa = { 0 };
2992 int featureflags;
2993 nvlist_t *begin_nvl = NULL;
2994
2995 ra.byteswap = drc->drc_byteswap;
2996 ra.cksum = drc->drc_cksum;
2997 ra.td = curthread;
2998 ra.fp = fp;
2999 ra.voff = *voffp;
3000
3001 if (dsl_dataset_is_zapified(drc->drc_ds)) {
3002 (void) zap_lookup(drc->drc_ds->ds_dir->dd_pool->dp_meta_objset,
3003 drc->drc_ds->ds_object, DS_FIELD_RESUME_BYTES,
3004 sizeof (ra.bytes_read), 1, &ra.bytes_read);
3005 }
3006
3007 objlist_create(&ra.ignore_objlist);
3008
3009 /* these were verified in dmu_recv_begin */
3010 ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==,
3011 DMU_SUBSTREAM);
3012 ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES);
3013
3014 /*
3015 * Open the objset we are modifying.
3016 */
3017 VERIFY0(dmu_objset_from_ds(drc->drc_ds, &ra.os));
3018
3019 ASSERT(dsl_dataset_phys(drc->drc_ds)->ds_flags & DS_FLAG_INCONSISTENT);
3020
3021 featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
3022
3023 /* if this stream is dedup'ed, set up the avl tree for guid mapping */
3024 if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
3025 minor_t minor;
3026
3027 if (cleanup_fd == -1) {
3028 ra.err = SET_ERROR(EBADF);
3029 goto out;
3030 }
3031 ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
3032 if (ra.err != 0) {
3033 cleanup_fd = -1;
3034 goto out;
3035 }
3036
3037 if (*action_handlep == 0) {
3038 rwa.guid_to_ds_map =
3039 kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
3040 avl_create(rwa.guid_to_ds_map, guid_compare,
3041 sizeof (guid_map_entry_t),
3042 offsetof(guid_map_entry_t, avlnode));
3043 err = zfs_onexit_add_cb(minor,
3044 free_guid_map_onexit, rwa.guid_to_ds_map,
3045 action_handlep);
3046 if (ra.err != 0)
3047 goto out;
3048 } else {
3049 err = zfs_onexit_cb_data(minor, *action_handlep,
3050 (void **)&rwa.guid_to_ds_map);
3051 if (ra.err != 0)
3052 goto out;
3053 }
3054
3055 drc->drc_guid_to_ds_map = rwa.guid_to_ds_map;
3056 }
3057
3058 uint32_t payloadlen = drc->drc_drr_begin->drr_payloadlen;
3059 void *payload = NULL;
3060 if (payloadlen != 0)
3061 payload = kmem_alloc(payloadlen, KM_SLEEP);
3062
3063 err = receive_read_payload_and_next_header(&ra, payloadlen, payload);
3064 if (err != 0) {
3065 if (payloadlen != 0)
3066 kmem_free(payload, payloadlen);
3067 goto out;
3068 }
3069 if (payloadlen != 0) {
3070 err = nvlist_unpack(payload, payloadlen, &begin_nvl, KM_SLEEP);
3071 kmem_free(payload, payloadlen);
3072 if (err != 0)
3073 goto out;
3074 }
3075
3076 if (featureflags & DMU_BACKUP_FEATURE_RESUMING) {
3077 err = resume_check(&ra, begin_nvl);
3078 if (err != 0)
3079 goto out;
3080 }
3081
3082 (void) bqueue_init(&rwa.q, zfs_recv_queue_length,
3083 offsetof(struct receive_record_arg, node));
3084 cv_init(&rwa.cv, NULL, CV_DEFAULT, NULL);
3085 mutex_init(&rwa.mutex, NULL, MUTEX_DEFAULT, NULL);
3086 rwa.os = ra.os;
3087 rwa.byteswap = drc->drc_byteswap;
3088 rwa.resumable = drc->drc_resumable;
3089
3090 (void) thread_create(NULL, 0, receive_writer_thread, &rwa, 0, &p0,
3091 TS_RUN, minclsyspri);
3092 /*
3093 * We're reading rwa.err without locks, which is safe since we are the
3094 * only reader, and the worker thread is the only writer. It's ok if we
3095 * miss a write for an iteration or two of the loop, since the writer
3096 * thread will keep freeing records we send it until we send it an eos
3097 * marker.
3098 *
3099 * We can leave this loop in 3 ways: First, if rwa.err is
3100 * non-zero. In that case, the writer thread will free the rrd we just
3101 * pushed. Second, if we're interrupted; in that case, either it's the
3102 * first loop and ra.rrd was never allocated, or it's later, and ra.rrd
3103 * has been handed off to the writer thread who will free it. Finally,
3104 * if receive_read_record fails or we're at the end of the stream, then
3105 * we free ra.rrd and exit.
3106 */
3107 while (rwa.err == 0) {
3108 if (issig(JUSTLOOKING) && issig(FORREAL)) {
3109 err = SET_ERROR(EINTR);
3110 break;
3111 }
3112
3113 ASSERT3P(ra.rrd, ==, NULL);
3114 ra.rrd = ra.next_rrd;
3115 ra.next_rrd = NULL;
3116 /* Allocates and loads header into ra.next_rrd */
3117 err = receive_read_record(&ra);
3118
3119 if (ra.rrd->header.drr_type == DRR_END || err != 0) {
3120 kmem_free(ra.rrd, sizeof (*ra.rrd));
3121 ra.rrd = NULL;
3122 break;
3123 }
3124
3125 bqueue_enqueue(&rwa.q, ra.rrd,
3126 sizeof (struct receive_record_arg) + ra.rrd->payload_size);
3127 ra.rrd = NULL;
3128 }
3129 if (ra.next_rrd == NULL)
3130 ra.next_rrd = kmem_zalloc(sizeof (*ra.next_rrd), KM_SLEEP);
3131 ra.next_rrd->eos_marker = B_TRUE;
3132 bqueue_enqueue(&rwa.q, ra.next_rrd, 1);
3133
3134 mutex_enter(&rwa.mutex);
3135 while (!rwa.done) {
3136 cv_wait(&rwa.cv, &rwa.mutex);
3137 }
3138 mutex_exit(&rwa.mutex);
3139
3140 /*
3141 * If we are receiving a full stream as a clone, all object IDs which
3142 * are greater than the maximum ID referenced in the stream are
3143 * by definition unused and must be freed. Note that it's possible that
3144 * we've resumed this send and the first record we received was the END
3145 * record. In that case, max_object would be 0, but we shouldn't start
3146 * freeing all objects from there; instead we should start from the
3147 * resumeobj.
3148 */
3149 if (drc->drc_clone && drc->drc_drrb->drr_fromguid == 0) {
3150 uint64_t obj;
3151 if (nvlist_lookup_uint64(begin_nvl, "resume_object", &obj) != 0)
3152 obj = 0;
3153 if (rwa.max_object > obj)
3154 obj = rwa.max_object;
3155 obj++;
3156 int free_err = 0;
3157 int next_err = 0;
3158
3159 while (next_err == 0) {
3160 free_err = dmu_free_long_object(rwa.os, obj);
3161 if (free_err != 0 && free_err != ENOENT)
3162 break;
3163
3164 next_err = dmu_object_next(rwa.os, &obj, FALSE, 0);
3165 }
3166
3167 if (err == 0) {
3168 if (free_err != 0 && free_err != ENOENT)
3169 err = free_err;
3170 else if (next_err != ESRCH)
3171 err = next_err;
3172 }
3173 }
3174
3175 cv_destroy(&rwa.cv);
3176 mutex_destroy(&rwa.mutex);
3177 bqueue_destroy(&rwa.q);
3178 if (err == 0)
3179 err = rwa.err;
3180
3181 out:
3182 nvlist_free(begin_nvl);
3183 if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
3184 zfs_onexit_fd_rele(cleanup_fd);
3185
3186 if (err != 0) {
3187 /*
3188 * Clean up references. If receive is not resumable,
3189 * destroy what we created, so we don't leave it in
3190 * the inconsistent state.
3191 */
3192 dmu_recv_cleanup_ds(drc);
3193 }
3194
3195 *voffp = ra.voff;
3196 objlist_destroy(&ra.ignore_objlist);
3197 return (err);
3198 }
3199
3200 static int
dmu_recv_end_check(void * arg,dmu_tx_t * tx)3201 dmu_recv_end_check(void *arg, dmu_tx_t *tx)
3202 {
3203 dmu_recv_cookie_t *drc = arg;
3204 dsl_pool_t *dp = dmu_tx_pool(tx);
3205 int error;
3206
3207 ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag);
3208
3209 if (!drc->drc_newfs) {
3210 dsl_dataset_t *origin_head;
3211
3212 error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head);
3213 if (error != 0)
3214 return (error);
3215 if (drc->drc_force) {
3216 /*
3217 * We will destroy any snapshots in tofs (i.e. before
3218 * origin_head) that are after the origin (which is
3219 * the snap before drc_ds, because drc_ds can not
3220 * have any snaps of its own).
3221 */
3222 uint64_t obj;
3223
3224 obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3225 while (obj !=
3226 dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
3227 dsl_dataset_t *snap;
3228 error = dsl_dataset_hold_obj(dp, obj, FTAG,
3229 &snap);
3230 if (error != 0)
3231 break;
3232 if (snap->ds_dir != origin_head->ds_dir)
3233 error = SET_ERROR(EINVAL);
3234 if (error == 0) {
3235 error = dsl_destroy_snapshot_check_impl(
3236 snap, B_FALSE);
3237 }
3238 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
3239 dsl_dataset_rele(snap, FTAG);
3240 if (error != 0)
3241 break;
3242 }
3243 if (error != 0) {
3244 dsl_dataset_rele(origin_head, FTAG);
3245 return (error);
3246 }
3247 }
3248 error = dsl_dataset_clone_swap_check_impl(drc->drc_ds,
3249 origin_head, drc->drc_force, drc->drc_owner, tx);
3250 if (error != 0) {
3251 dsl_dataset_rele(origin_head, FTAG);
3252 return (error);
3253 }
3254 error = dsl_dataset_snapshot_check_impl(origin_head,
3255 drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
3256 dsl_dataset_rele(origin_head, FTAG);
3257 if (error != 0)
3258 return (error);
3259
3260 error = dsl_destroy_head_check_impl(drc->drc_ds, 1);
3261 } else {
3262 error = dsl_dataset_snapshot_check_impl(drc->drc_ds,
3263 drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
3264 }
3265 return (error);
3266 }
3267
3268 static void
dmu_recv_end_sync(void * arg,dmu_tx_t * tx)3269 dmu_recv_end_sync(void *arg, dmu_tx_t *tx)
3270 {
3271 dmu_recv_cookie_t *drc = arg;
3272 dsl_pool_t *dp = dmu_tx_pool(tx);
3273
3274 spa_history_log_internal_ds(drc->drc_ds, "finish receiving",
3275 tx, "snap=%s", drc->drc_tosnap);
3276
3277 if (!drc->drc_newfs) {
3278 dsl_dataset_t *origin_head;
3279
3280 VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG,
3281 &origin_head));
3282
3283 if (drc->drc_force) {
3284 /*
3285 * Destroy any snapshots of drc_tofs (origin_head)
3286 * after the origin (the snap before drc_ds).
3287 */
3288 uint64_t obj;
3289
3290 obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3291 while (obj !=
3292 dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
3293 dsl_dataset_t *snap;
3294 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG,
3295 &snap));
3296 ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir);
3297 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
3298 dsl_destroy_snapshot_sync_impl(snap,
3299 B_FALSE, tx);
3300 dsl_dataset_rele(snap, FTAG);
3301 }
3302 }
3303 VERIFY3P(drc->drc_ds->ds_prev, ==,
3304 origin_head->ds_prev);
3305
3306 dsl_dataset_clone_swap_sync_impl(drc->drc_ds,
3307 origin_head, tx);
3308 dsl_dataset_snapshot_sync_impl(origin_head,
3309 drc->drc_tosnap, tx);
3310
3311 /* set snapshot's creation time and guid */
3312 dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx);
3313 dsl_dataset_phys(origin_head->ds_prev)->ds_creation_time =
3314 drc->drc_drrb->drr_creation_time;
3315 dsl_dataset_phys(origin_head->ds_prev)->ds_guid =
3316 drc->drc_drrb->drr_toguid;
3317 dsl_dataset_phys(origin_head->ds_prev)->ds_flags &=
3318 ~DS_FLAG_INCONSISTENT;
3319
3320 dmu_buf_will_dirty(origin_head->ds_dbuf, tx);
3321 dsl_dataset_phys(origin_head)->ds_flags &=
3322 ~DS_FLAG_INCONSISTENT;
3323
3324 drc->drc_newsnapobj =
3325 dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3326
3327 dsl_dataset_rele(origin_head, FTAG);
3328 dsl_destroy_head_sync_impl(drc->drc_ds, tx);
3329
3330 if (drc->drc_owner != NULL)
3331 VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner);
3332 } else {
3333 dsl_dataset_t *ds = drc->drc_ds;
3334
3335 dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx);
3336
3337 /* set snapshot's creation time and guid */
3338 dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
3339 dsl_dataset_phys(ds->ds_prev)->ds_creation_time =
3340 drc->drc_drrb->drr_creation_time;
3341 dsl_dataset_phys(ds->ds_prev)->ds_guid =
3342 drc->drc_drrb->drr_toguid;
3343 dsl_dataset_phys(ds->ds_prev)->ds_flags &=
3344 ~DS_FLAG_INCONSISTENT;
3345
3346 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3347 dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
3348 if (dsl_dataset_has_resume_receive_state(ds)) {
3349 (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3350 DS_FIELD_RESUME_FROMGUID, tx);
3351 (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3352 DS_FIELD_RESUME_OBJECT, tx);
3353 (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3354 DS_FIELD_RESUME_OFFSET, tx);
3355 (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3356 DS_FIELD_RESUME_BYTES, tx);
3357 (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3358 DS_FIELD_RESUME_TOGUID, tx);
3359 (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3360 DS_FIELD_RESUME_TONAME, tx);
3361 }
3362 drc->drc_newsnapobj =
3363 dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj;
3364 }
3365 /*
3366 * Release the hold from dmu_recv_begin. This must be done before
3367 * we return to open context, so that when we free the dataset's dnode,
3368 * we can evict its bonus buffer.
3369 */
3370 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
3371 drc->drc_ds = NULL;
3372 }
3373
3374 static int
add_ds_to_guidmap(const char * name,avl_tree_t * guid_map,uint64_t snapobj)3375 add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj)
3376 {
3377 dsl_pool_t *dp;
3378 dsl_dataset_t *snapds;
3379 guid_map_entry_t *gmep;
3380 int err;
3381
3382 ASSERT(guid_map != NULL);
3383
3384 err = dsl_pool_hold(name, FTAG, &dp);
3385 if (err != 0)
3386 return (err);
3387 gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP);
3388 err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds);
3389 if (err == 0) {
3390 gmep->guid = dsl_dataset_phys(snapds)->ds_guid;
3391 gmep->gme_ds = snapds;
3392 avl_add(guid_map, gmep);
3393 dsl_dataset_long_hold(snapds, gmep);
3394 } else
3395 kmem_free(gmep, sizeof (*gmep));
3396
3397 dsl_pool_rele(dp, FTAG);
3398 return (err);
3399 }
3400
3401 static int dmu_recv_end_modified_blocks = 3;
3402
3403 static int
dmu_recv_existing_end(dmu_recv_cookie_t * drc)3404 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
3405 {
3406 #ifdef _KERNEL
3407 /*
3408 * We will be destroying the ds; make sure its origin is unmounted if
3409 * necessary.
3410 */
3411 char name[ZFS_MAX_DATASET_NAME_LEN];
3412 dsl_dataset_name(drc->drc_ds, name);
3413 zfs_destroy_unmount_origin(name);
3414 #endif
3415
3416 return (dsl_sync_task(drc->drc_tofs,
3417 dmu_recv_end_check, dmu_recv_end_sync, drc,
3418 dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL));
3419 }
3420
3421 static int
dmu_recv_new_end(dmu_recv_cookie_t * drc)3422 dmu_recv_new_end(dmu_recv_cookie_t *drc)
3423 {
3424 return (dsl_sync_task(drc->drc_tofs,
3425 dmu_recv_end_check, dmu_recv_end_sync, drc,
3426 dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL));
3427 }
3428
3429 int
dmu_recv_end(dmu_recv_cookie_t * drc,void * owner)3430 dmu_recv_end(dmu_recv_cookie_t *drc, void *owner)
3431 {
3432 int error;
3433
3434 drc->drc_owner = owner;
3435
3436 if (drc->drc_newfs)
3437 error = dmu_recv_new_end(drc);
3438 else
3439 error = dmu_recv_existing_end(drc);
3440
3441 if (error != 0) {
3442 dmu_recv_cleanup_ds(drc);
3443 } else if (drc->drc_guid_to_ds_map != NULL) {
3444 (void) add_ds_to_guidmap(drc->drc_tofs,
3445 drc->drc_guid_to_ds_map,
3446 drc->drc_newsnapobj);
3447 }
3448 return (error);
3449 }
3450
3451 /*
3452 * Return TRUE if this objset is currently being received into.
3453 */
3454 boolean_t
dmu_objset_is_receiving(objset_t * os)3455 dmu_objset_is_receiving(objset_t *os)
3456 {
3457 return (os->os_dsl_dataset != NULL &&
3458 os->os_dsl_dataset->ds_owner == dmu_recv_tag);
3459 }
3460