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 <mm@FreeBSD.org>. 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 #ifdef __NetBSD__
66 #ifdef _KERNEL
67 #define FOF_OFFSET FOF_UPDATE_OFFSET
68 #define td_ucred l_cred
69 #define bwillwrite() /* nothing */
70
71 static int
fo_write(struct file * fp,struct uio * uio,cred_t * cred,int flags,kthread_t * thr)72 fo_write(struct file *fp, struct uio *uio, cred_t *cred, int flags, kthread_t *thr)
73 {
74
75 if (fp->f_type == DTYPE_VNODE)
76 flags |= FOF_UPDATE_OFFSET;
77 return (*fp->f_ops->fo_write)(fp, &fp->f_offset, uio, cred, flags);
78 }
79
80 static int
fo_read(struct file * fp,struct uio * uio,cred_t * cred,int flags,kthread_t * thr)81 fo_read(struct file *fp, struct uio *uio, cred_t *cred, int flags, kthread_t *thr)
82 {
83
84 if (fp->f_type == DTYPE_VNODE)
85 flags |= FOF_UPDATE_OFFSET;
86 return (*fp->f_ops->fo_read)(fp, &fp->f_offset, uio, cred, flags);
87 }
88 #endif
89 #endif
90
91 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
92 int zfs_send_corrupt_data = B_FALSE;
93 int zfs_send_queue_length = 16 * 1024 * 1024;
94 int zfs_recv_queue_length = 16 * 1024 * 1024;
95 /* Set this tunable to FALSE to disable setting of DRR_FLAG_FREERECORDS */
96 int zfs_send_set_freerecords_bit = B_TRUE;
97
98 #ifdef _KERNEL
99 TUNABLE_INT("vfs.zfs.send_set_freerecords_bit", &zfs_send_set_freerecords_bit);
100 #endif
101
102 static char *dmu_recv_tag = "dmu_recv_tag";
103 const char *recv_clone_name = "%recv";
104
105 #define BP_SPAN(datablkszsec, indblkshift, level) \
106 (((uint64_t)datablkszsec) << (SPA_MINBLOCKSHIFT + \
107 (level) * (indblkshift - SPA_BLKPTRSHIFT)))
108
109 static void byteswap_record(dmu_replay_record_t *drr);
110
111 struct send_thread_arg {
112 bqueue_t q;
113 dsl_dataset_t *ds; /* Dataset to traverse */
114 uint64_t fromtxg; /* Traverse from this txg */
115 int flags; /* flags to pass to traverse_dataset */
116 int error_code;
117 boolean_t cancel;
118 zbookmark_phys_t resume;
119 };
120
121 struct send_block_record {
122 boolean_t eos_marker; /* Marks the end of the stream */
123 blkptr_t bp;
124 zbookmark_phys_t zb;
125 uint8_t indblkshift;
126 uint16_t datablkszsec;
127 bqueue_node_t ln;
128 };
129
130 static int
dump_bytes(dmu_sendarg_t * dsp,void * buf,int len)131 dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
132 {
133 dsl_dataset_t *ds = dmu_objset_ds(dsp->dsa_os);
134 struct uio auio;
135 struct iovec aiov;
136
137 /*
138 * The code does not rely on this (len being a multiple of 8). We keep
139 * this assertion because of the corresponding assertion in
140 * receive_read(). Keeping this assertion ensures that we do not
141 * inadvertently break backwards compatibility (causing the assertion
142 * in receive_read() to trigger on old software).
143 *
144 * Removing the assertions could be rolled into a new feature that uses
145 * data that isn't 8-byte aligned; if the assertions were removed, a
146 * feature flag would have to be added.
147 */
148
149 ASSERT0(len % 8);
150
151 aiov.iov_base = buf;
152 aiov.iov_len = len;
153 auio.uio_iov = &aiov;
154 auio.uio_iovcnt = 1;
155 auio.uio_resid = len;
156 #ifdef __NetBSD__
157 #ifdef _KERNEL
158 auio.uio_vmspace = vmspace_kernel();
159 #endif
160 #else
161 auio.uio_segflg = UIO_SYSSPACE;
162 #endif
163 auio.uio_rw = UIO_WRITE;
164 auio.uio_offset = (off_t)-1;
165 #ifdef __FreeBSD__
166 auio.uio_td = dsp->dsa_td;
167 #endif
168 #ifdef _KERNEL
169 if (dsp->dsa_fp->f_type == DTYPE_VNODE)
170 bwillwrite();
171 dsp->dsa_err = fo_write(dsp->dsa_fp, &auio, dsp->dsa_td->td_ucred, 0,
172 dsp->dsa_td);
173 #else
174 fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__);
175 dsp->dsa_err = EOPNOTSUPP;
176 #endif
177 mutex_enter(&ds->ds_sendstream_lock);
178 *dsp->dsa_off += len;
179 mutex_exit(&ds->ds_sendstream_lock);
180
181 return (dsp->dsa_err);
182 }
183
184 /*
185 * For all record types except BEGIN, fill in the checksum (overlaid in
186 * drr_u.drr_checksum.drr_checksum). The checksum verifies everything
187 * up to the start of the checksum itself.
188 */
189 static int
dump_record(dmu_sendarg_t * dsp,void * payload,int payload_len)190 dump_record(dmu_sendarg_t *dsp, void *payload, int payload_len)
191 {
192 ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
193 ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
194 fletcher_4_incremental_native(dsp->dsa_drr,
195 offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
196 &dsp->dsa_zc);
197 if (dsp->dsa_drr->drr_type == DRR_BEGIN) {
198 dsp->dsa_sent_begin = B_TRUE;
199 } else {
200 ASSERT(ZIO_CHECKSUM_IS_ZERO(&dsp->dsa_drr->drr_u.
201 drr_checksum.drr_checksum));
202 dsp->dsa_drr->drr_u.drr_checksum.drr_checksum = dsp->dsa_zc;
203 }
204 if (dsp->dsa_drr->drr_type == DRR_END) {
205 dsp->dsa_sent_end = B_TRUE;
206 }
207 fletcher_4_incremental_native(&dsp->dsa_drr->
208 drr_u.drr_checksum.drr_checksum,
209 sizeof (zio_cksum_t), &dsp->dsa_zc);
210 if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
211 return (SET_ERROR(EINTR));
212 if (payload_len != 0) {
213 fletcher_4_incremental_native(payload, payload_len,
214 &dsp->dsa_zc);
215 if (dump_bytes(dsp, payload, payload_len) != 0)
216 return (SET_ERROR(EINTR));
217 }
218 return (0);
219 }
220
221 /*
222 * Fill in the drr_free struct, or perform aggregation if the previous record is
223 * also a free record, and the two are adjacent.
224 *
225 * Note that we send free records even for a full send, because we want to be
226 * able to receive a full send as a clone, which requires a list of all the free
227 * and freeobject records that were generated on the source.
228 */
229 static int
dump_free(dmu_sendarg_t * dsp,uint64_t object,uint64_t offset,uint64_t length)230 dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
231 uint64_t length)
232 {
233 struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free);
234
235 /*
236 * When we receive a free record, dbuf_free_range() assumes
237 * that the receiving system doesn't have any dbufs in the range
238 * being freed. This is always true because there is a one-record
239 * constraint: we only send one WRITE record for any given
240 * object,offset. We know that the one-record constraint is
241 * true because we always send data in increasing order by
242 * object,offset.
243 *
244 * If the increasing-order constraint ever changes, we should find
245 * another way to assert that the one-record constraint is still
246 * satisfied.
247 */
248 ASSERT(object > dsp->dsa_last_data_object ||
249 (object == dsp->dsa_last_data_object &&
250 offset > dsp->dsa_last_data_offset));
251
252 if (length != -1ULL && offset + length < offset)
253 length = -1ULL;
254
255 /*
256 * If there is a pending op, but it's not PENDING_FREE, push it out,
257 * since free block aggregation can only be done for blocks of the
258 * same type (i.e., DRR_FREE records can only be aggregated with
259 * other DRR_FREE records. DRR_FREEOBJECTS records can only be
260 * aggregated with other DRR_FREEOBJECTS records.
261 */
262 if (dsp->dsa_pending_op != PENDING_NONE &&
263 dsp->dsa_pending_op != PENDING_FREE) {
264 if (dump_record(dsp, NULL, 0) != 0)
265 return (SET_ERROR(EINTR));
266 dsp->dsa_pending_op = PENDING_NONE;
267 }
268
269 if (dsp->dsa_pending_op == PENDING_FREE) {
270 /*
271 * There should never be a PENDING_FREE if length is -1
272 * (because dump_dnode is the only place where this
273 * function is called with a -1, and only after flushing
274 * any pending record).
275 */
276 ASSERT(length != -1ULL);
277 /*
278 * Check to see whether this free block can be aggregated
279 * with pending one.
280 */
281 if (drrf->drr_object == object && drrf->drr_offset +
282 drrf->drr_length == offset) {
283 drrf->drr_length += length;
284 return (0);
285 } else {
286 /* not a continuation. Push out pending record */
287 if (dump_record(dsp, NULL, 0) != 0)
288 return (SET_ERROR(EINTR));
289 dsp->dsa_pending_op = PENDING_NONE;
290 }
291 }
292 /* create a FREE record and make it pending */
293 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
294 dsp->dsa_drr->drr_type = DRR_FREE;
295 drrf->drr_object = object;
296 drrf->drr_offset = offset;
297 drrf->drr_length = length;
298 drrf->drr_toguid = dsp->dsa_toguid;
299 if (length == -1ULL) {
300 if (dump_record(dsp, NULL, 0) != 0)
301 return (SET_ERROR(EINTR));
302 } else {
303 dsp->dsa_pending_op = PENDING_FREE;
304 }
305
306 return (0);
307 }
308
309 static int
dump_write(dmu_sendarg_t * dsp,dmu_object_type_t type,uint64_t object,uint64_t offset,int blksz,const blkptr_t * bp,void * data)310 dump_write(dmu_sendarg_t *dsp, dmu_object_type_t type,
311 uint64_t object, uint64_t offset, int blksz, const blkptr_t *bp, void *data)
312 {
313 struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write);
314
315 /*
316 * We send data in increasing object, offset order.
317 * See comment in dump_free() for details.
318 */
319 ASSERT(object > dsp->dsa_last_data_object ||
320 (object == dsp->dsa_last_data_object &&
321 offset > dsp->dsa_last_data_offset));
322 dsp->dsa_last_data_object = object;
323 dsp->dsa_last_data_offset = offset + blksz - 1;
324
325 /*
326 * If there is any kind of pending aggregation (currently either
327 * a grouping of free objects or free blocks), push it out to
328 * the stream, since aggregation can't be done across operations
329 * of different types.
330 */
331 if (dsp->dsa_pending_op != PENDING_NONE) {
332 if (dump_record(dsp, NULL, 0) != 0)
333 return (SET_ERROR(EINTR));
334 dsp->dsa_pending_op = PENDING_NONE;
335 }
336 /* write a WRITE record */
337 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
338 dsp->dsa_drr->drr_type = DRR_WRITE;
339 drrw->drr_object = object;
340 drrw->drr_type = type;
341 drrw->drr_offset = offset;
342 drrw->drr_length = blksz;
343 drrw->drr_toguid = dsp->dsa_toguid;
344 if (bp == NULL || BP_IS_EMBEDDED(bp)) {
345 /*
346 * There's no pre-computed checksum for partial-block
347 * writes or embedded BP's, so (like
348 * fletcher4-checkummed blocks) userland will have to
349 * compute a dedup-capable checksum itself.
350 */
351 drrw->drr_checksumtype = ZIO_CHECKSUM_OFF;
352 } else {
353 drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
354 if (zio_checksum_table[drrw->drr_checksumtype].ci_flags &
355 ZCHECKSUM_FLAG_DEDUP)
356 drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
357 DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
358 DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
359 DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
360 drrw->drr_key.ddk_cksum = bp->blk_cksum;
361 }
362
363 if (dump_record(dsp, data, blksz) != 0)
364 return (SET_ERROR(EINTR));
365 return (0);
366 }
367
368 static int
dump_write_embedded(dmu_sendarg_t * dsp,uint64_t object,uint64_t offset,int blksz,const blkptr_t * bp)369 dump_write_embedded(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
370 int blksz, const blkptr_t *bp)
371 {
372 char buf[BPE_PAYLOAD_SIZE];
373 struct drr_write_embedded *drrw =
374 &(dsp->dsa_drr->drr_u.drr_write_embedded);
375
376 if (dsp->dsa_pending_op != PENDING_NONE) {
377 if (dump_record(dsp, NULL, 0) != 0)
378 return (EINTR);
379 dsp->dsa_pending_op = PENDING_NONE;
380 }
381
382 ASSERT(BP_IS_EMBEDDED(bp));
383
384 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
385 dsp->dsa_drr->drr_type = DRR_WRITE_EMBEDDED;
386 drrw->drr_object = object;
387 drrw->drr_offset = offset;
388 drrw->drr_length = blksz;
389 drrw->drr_toguid = dsp->dsa_toguid;
390 drrw->drr_compression = BP_GET_COMPRESS(bp);
391 drrw->drr_etype = BPE_GET_ETYPE(bp);
392 drrw->drr_lsize = BPE_GET_LSIZE(bp);
393 drrw->drr_psize = BPE_GET_PSIZE(bp);
394
395 decode_embedded_bp_compressed(bp, buf);
396
397 if (dump_record(dsp, buf, P2ROUNDUP(drrw->drr_psize, 8)) != 0)
398 return (EINTR);
399 return (0);
400 }
401
402 static int
dump_spill(dmu_sendarg_t * dsp,uint64_t object,int blksz,void * data)403 dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data)
404 {
405 struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill);
406
407 if (dsp->dsa_pending_op != PENDING_NONE) {
408 if (dump_record(dsp, NULL, 0) != 0)
409 return (SET_ERROR(EINTR));
410 dsp->dsa_pending_op = PENDING_NONE;
411 }
412
413 /* write a SPILL record */
414 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
415 dsp->dsa_drr->drr_type = DRR_SPILL;
416 drrs->drr_object = object;
417 drrs->drr_length = blksz;
418 drrs->drr_toguid = dsp->dsa_toguid;
419
420 if (dump_record(dsp, data, blksz) != 0)
421 return (SET_ERROR(EINTR));
422 return (0);
423 }
424
425 static int
dump_freeobjects(dmu_sendarg_t * dsp,uint64_t firstobj,uint64_t numobjs)426 dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
427 {
428 struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
429
430 /*
431 * If there is a pending op, but it's not PENDING_FREEOBJECTS,
432 * push it out, since free block aggregation can only be done for
433 * blocks of the same type (i.e., DRR_FREE records can only be
434 * aggregated with other DRR_FREE records. DRR_FREEOBJECTS records
435 * can only be aggregated with other DRR_FREEOBJECTS records.
436 */
437 if (dsp->dsa_pending_op != PENDING_NONE &&
438 dsp->dsa_pending_op != PENDING_FREEOBJECTS) {
439 if (dump_record(dsp, NULL, 0) != 0)
440 return (SET_ERROR(EINTR));
441 dsp->dsa_pending_op = PENDING_NONE;
442 }
443 if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) {
444 /*
445 * See whether this free object array can be aggregated
446 * with pending one
447 */
448 if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
449 drrfo->drr_numobjs += numobjs;
450 return (0);
451 } else {
452 /* can't be aggregated. Push out pending record */
453 if (dump_record(dsp, NULL, 0) != 0)
454 return (SET_ERROR(EINTR));
455 dsp->dsa_pending_op = PENDING_NONE;
456 }
457 }
458
459 /* write a FREEOBJECTS record */
460 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
461 dsp->dsa_drr->drr_type = DRR_FREEOBJECTS;
462 drrfo->drr_firstobj = firstobj;
463 drrfo->drr_numobjs = numobjs;
464 drrfo->drr_toguid = dsp->dsa_toguid;
465
466 dsp->dsa_pending_op = PENDING_FREEOBJECTS;
467
468 return (0);
469 }
470
471 static int
dump_dnode(dmu_sendarg_t * dsp,uint64_t object,dnode_phys_t * dnp)472 dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
473 {
474 struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object);
475
476 if (object < dsp->dsa_resume_object) {
477 /*
478 * Note: when resuming, we will visit all the dnodes in
479 * the block of dnodes that we are resuming from. In
480 * this case it's unnecessary to send the dnodes prior to
481 * the one we are resuming from. We should be at most one
482 * block's worth of dnodes behind the resume point.
483 */
484 ASSERT3U(dsp->dsa_resume_object - object, <,
485 1 << (DNODE_BLOCK_SHIFT - DNODE_SHIFT));
486 return (0);
487 }
488
489 if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
490 return (dump_freeobjects(dsp, object, 1));
491
492 if (dsp->dsa_pending_op != PENDING_NONE) {
493 if (dump_record(dsp, NULL, 0) != 0)
494 return (SET_ERROR(EINTR));
495 dsp->dsa_pending_op = PENDING_NONE;
496 }
497
498 /* write an OBJECT record */
499 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
500 dsp->dsa_drr->drr_type = DRR_OBJECT;
501 drro->drr_object = object;
502 drro->drr_type = dnp->dn_type;
503 drro->drr_bonustype = dnp->dn_bonustype;
504 drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
505 drro->drr_bonuslen = dnp->dn_bonuslen;
506 drro->drr_checksumtype = dnp->dn_checksum;
507 drro->drr_compress = dnp->dn_compress;
508 drro->drr_toguid = dsp->dsa_toguid;
509
510 if (!(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
511 drro->drr_blksz > SPA_OLD_MAXBLOCKSIZE)
512 drro->drr_blksz = SPA_OLD_MAXBLOCKSIZE;
513
514 if (dump_record(dsp, DN_BONUS(dnp),
515 P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0) {
516 return (SET_ERROR(EINTR));
517 }
518
519 /* Free anything past the end of the file. */
520 if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) *
521 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0)
522 return (SET_ERROR(EINTR));
523 if (dsp->dsa_err != 0)
524 return (SET_ERROR(EINTR));
525 return (0);
526 }
527
528 static boolean_t
backup_do_embed(dmu_sendarg_t * dsp,const blkptr_t * bp)529 backup_do_embed(dmu_sendarg_t *dsp, const blkptr_t *bp)
530 {
531 if (!BP_IS_EMBEDDED(bp))
532 return (B_FALSE);
533
534 /*
535 * Compression function must be legacy, or explicitly enabled.
536 */
537 if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS &&
538 !(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4)))
539 return (B_FALSE);
540
541 /*
542 * Embed type must be explicitly enabled.
543 */
544 switch (BPE_GET_ETYPE(bp)) {
545 case BP_EMBEDDED_TYPE_DATA:
546 if (dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)
547 return (B_TRUE);
548 break;
549 default:
550 return (B_FALSE);
551 }
552 return (B_FALSE);
553 }
554
555 /*
556 * This is the callback function to traverse_dataset that acts as the worker
557 * thread for dmu_send_impl.
558 */
559 /*ARGSUSED*/
560 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)561 send_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
562 const zbookmark_phys_t *zb, const struct dnode_phys *dnp, void *arg)
563 {
564 struct send_thread_arg *sta = arg;
565 struct send_block_record *record;
566 uint64_t record_size;
567 int err = 0;
568
569 ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
570 zb->zb_object >= sta->resume.zb_object);
571
572 if (sta->cancel)
573 return (SET_ERROR(EINTR));
574
575 if (bp == NULL) {
576 ASSERT3U(zb->zb_level, ==, ZB_DNODE_LEVEL);
577 return (0);
578 } else if (zb->zb_level < 0) {
579 return (0);
580 }
581
582 record = kmem_zalloc(sizeof (struct send_block_record), KM_SLEEP);
583 record->eos_marker = B_FALSE;
584 record->bp = *bp;
585 record->zb = *zb;
586 record->indblkshift = dnp->dn_indblkshift;
587 record->datablkszsec = dnp->dn_datablkszsec;
588 record_size = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
589 bqueue_enqueue(&sta->q, record, record_size);
590
591 return (err);
592 }
593
594 /*
595 * This function kicks off the traverse_dataset. It also handles setting the
596 * error code of the thread in case something goes wrong, and pushes the End of
597 * Stream record when the traverse_dataset call has finished. If there is no
598 * dataset to traverse, the thread immediately pushes End of Stream marker.
599 */
600 static void
send_traverse_thread(void * arg)601 send_traverse_thread(void *arg)
602 {
603 struct send_thread_arg *st_arg = arg;
604 int err;
605 struct send_block_record *data;
606
607 if (st_arg->ds != NULL) {
608 err = traverse_dataset_resume(st_arg->ds,
609 st_arg->fromtxg, &st_arg->resume,
610 st_arg->flags, send_cb, st_arg);
611
612 if (err != EINTR)
613 st_arg->error_code = err;
614 }
615 data = kmem_zalloc(sizeof (*data), KM_SLEEP);
616 data->eos_marker = B_TRUE;
617 bqueue_enqueue(&st_arg->q, data, 1);
618 thread_exit();
619 }
620
621 /*
622 * This function actually handles figuring out what kind of record needs to be
623 * dumped, reading the data (which has hopefully been prefetched), and calling
624 * the appropriate helper function.
625 */
626 static int
do_dump(dmu_sendarg_t * dsa,struct send_block_record * data)627 do_dump(dmu_sendarg_t *dsa, struct send_block_record *data)
628 {
629 dsl_dataset_t *ds = dmu_objset_ds(dsa->dsa_os);
630 const blkptr_t *bp = &data->bp;
631 const zbookmark_phys_t *zb = &data->zb;
632 uint8_t indblkshift = data->indblkshift;
633 uint16_t dblkszsec = data->datablkszsec;
634 spa_t *spa = ds->ds_dir->dd_pool->dp_spa;
635 dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
636 int err = 0;
637
638 ASSERT3U(zb->zb_level, >=, 0);
639
640 ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
641 zb->zb_object >= dsa->dsa_resume_object);
642
643 if (zb->zb_object != DMU_META_DNODE_OBJECT &&
644 DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
645 return (0);
646 } else if (BP_IS_HOLE(bp) &&
647 zb->zb_object == DMU_META_DNODE_OBJECT) {
648 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
649 uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
650 err = dump_freeobjects(dsa, dnobj, span >> DNODE_SHIFT);
651 } else if (BP_IS_HOLE(bp)) {
652 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
653 uint64_t offset = zb->zb_blkid * span;
654 err = dump_free(dsa, zb->zb_object, offset, span);
655 } else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
656 return (0);
657 } else if (type == DMU_OT_DNODE) {
658 int blksz = BP_GET_LSIZE(bp);
659 arc_flags_t aflags = ARC_FLAG_WAIT;
660 arc_buf_t *abuf;
661
662 ASSERT0(zb->zb_level);
663
664 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
665 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
666 &aflags, zb) != 0)
667 return (SET_ERROR(EIO));
668
669 dnode_phys_t *blk = abuf->b_data;
670 uint64_t dnobj = zb->zb_blkid * (blksz >> DNODE_SHIFT);
671 for (int i = 0; i < blksz >> DNODE_SHIFT; i++) {
672 err = dump_dnode(dsa, dnobj + i, blk + i);
673 if (err != 0)
674 break;
675 }
676 arc_buf_destroy(abuf, &abuf);
677 } else if (type == DMU_OT_SA) {
678 arc_flags_t aflags = ARC_FLAG_WAIT;
679 arc_buf_t *abuf;
680 int blksz = BP_GET_LSIZE(bp);
681
682 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
683 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
684 &aflags, zb) != 0)
685 return (SET_ERROR(EIO));
686
687 err = dump_spill(dsa, zb->zb_object, blksz, abuf->b_data);
688 arc_buf_destroy(abuf, &abuf);
689 } else if (backup_do_embed(dsa, bp)) {
690 /* it's an embedded level-0 block of a regular object */
691 int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
692 ASSERT0(zb->zb_level);
693 err = dump_write_embedded(dsa, zb->zb_object,
694 zb->zb_blkid * blksz, blksz, bp);
695 } else {
696 /* it's a level-0 block of a regular object */
697 arc_flags_t aflags = ARC_FLAG_WAIT;
698 arc_buf_t *abuf;
699 int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
700 uint64_t offset;
701
702 ASSERT0(zb->zb_level);
703 ASSERT(zb->zb_object > dsa->dsa_resume_object ||
704 (zb->zb_object == dsa->dsa_resume_object &&
705 zb->zb_blkid * blksz >= dsa->dsa_resume_offset));
706
707 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
708 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
709 &aflags, zb) != 0) {
710 if (zfs_send_corrupt_data) {
711 /* Send a block filled with 0x"zfs badd bloc" */
712 abuf = arc_alloc_buf(spa, blksz, &abuf,
713 ARC_BUFC_DATA);
714 uint64_t *ptr;
715 for (ptr = abuf->b_data;
716 (char *)ptr < (char *)abuf->b_data + blksz;
717 ptr++)
718 *ptr = 0x2f5baddb10cULL;
719 } else {
720 return (SET_ERROR(EIO));
721 }
722 }
723
724 offset = zb->zb_blkid * blksz;
725
726 if (!(dsa->dsa_featureflags &
727 DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
728 blksz > SPA_OLD_MAXBLOCKSIZE) {
729 char *buf = abuf->b_data;
730 while (blksz > 0 && err == 0) {
731 int n = MIN(blksz, SPA_OLD_MAXBLOCKSIZE);
732 err = dump_write(dsa, type, zb->zb_object,
733 offset, n, NULL, buf);
734 offset += n;
735 buf += n;
736 blksz -= n;
737 }
738 } else {
739 err = dump_write(dsa, type, zb->zb_object,
740 offset, blksz, bp, abuf->b_data);
741 }
742 arc_buf_destroy(abuf, &abuf);
743 }
744
745 ASSERT(err == 0 || err == EINTR);
746 return (err);
747 }
748
749 /*
750 * Pop the new data off the queue, and free the old data.
751 */
752 static struct send_block_record *
get_next_record(bqueue_t * bq,struct send_block_record * data)753 get_next_record(bqueue_t *bq, struct send_block_record *data)
754 {
755 struct send_block_record *tmp = bqueue_dequeue(bq);
756 kmem_free(data, sizeof (*data));
757 return (tmp);
758 }
759
760 /*
761 * Actually do the bulk of the work in a zfs send.
762 *
763 * Note: Releases dp using the specified tag.
764 */
765 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,int outfd,uint64_t resumeobj,uint64_t resumeoff,vnode_t * vp,offset_t * off)766 dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *to_ds,
767 zfs_bookmark_phys_t *ancestor_zb,
768 boolean_t is_clone, boolean_t embedok, boolean_t large_block_ok, int outfd,
769 uint64_t resumeobj, uint64_t resumeoff,
770 #ifdef illumos
771 vnode_t *vp, offset_t *off)
772 #else
773 struct file *fp, offset_t *off)
774 #endif
775 {
776 objset_t *os;
777 dmu_replay_record_t *drr;
778 dmu_sendarg_t *dsp;
779 int err;
780 uint64_t fromtxg = 0;
781 uint64_t featureflags = 0;
782 struct send_thread_arg to_arg = { 0 };
783
784 err = dmu_objset_from_ds(to_ds, &os);
785 if (err != 0) {
786 dsl_pool_rele(dp, tag);
787 return (err);
788 }
789
790 drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
791 drr->drr_type = DRR_BEGIN;
792 drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
793 DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
794 DMU_SUBSTREAM);
795
796 #ifdef _KERNEL
797 if (dmu_objset_type(os) == DMU_OST_ZFS) {
798 uint64_t version;
799 if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0) {
800 kmem_free(drr, sizeof (dmu_replay_record_t));
801 dsl_pool_rele(dp, tag);
802 return (SET_ERROR(EINVAL));
803 }
804 if (version >= ZPL_VERSION_SA) {
805 featureflags |= DMU_BACKUP_FEATURE_SA_SPILL;
806 }
807 }
808 #endif
809
810 if (large_block_ok && to_ds->ds_feature_inuse[SPA_FEATURE_LARGE_BLOCKS])
811 featureflags |= DMU_BACKUP_FEATURE_LARGE_BLOCKS;
812 if (embedok &&
813 spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) {
814 featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA;
815 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
816 featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA_LZ4;
817 }
818
819 if (resumeobj != 0 || resumeoff != 0) {
820 featureflags |= DMU_BACKUP_FEATURE_RESUMING;
821 }
822
823 DMU_SET_FEATUREFLAGS(drr->drr_u.drr_begin.drr_versioninfo,
824 featureflags);
825
826 drr->drr_u.drr_begin.drr_creation_time =
827 dsl_dataset_phys(to_ds)->ds_creation_time;
828 drr->drr_u.drr_begin.drr_type = dmu_objset_type(os);
829 if (is_clone)
830 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
831 drr->drr_u.drr_begin.drr_toguid = dsl_dataset_phys(to_ds)->ds_guid;
832 if (dsl_dataset_phys(to_ds)->ds_flags & DS_FLAG_CI_DATASET)
833 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
834 if (zfs_send_set_freerecords_bit)
835 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_FREERECORDS;
836
837 if (ancestor_zb != NULL) {
838 drr->drr_u.drr_begin.drr_fromguid =
839 ancestor_zb->zbm_guid;
840 fromtxg = ancestor_zb->zbm_creation_txg;
841 }
842 dsl_dataset_name(to_ds, drr->drr_u.drr_begin.drr_toname);
843 if (!to_ds->ds_is_snapshot) {
844 (void) strlcat(drr->drr_u.drr_begin.drr_toname, "@--head--",
845 sizeof (drr->drr_u.drr_begin.drr_toname));
846 }
847
848 dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP);
849
850 dsp->dsa_drr = drr;
851 dsp->dsa_outfd = outfd;
852 dsp->dsa_proc = curproc;
853 dsp->dsa_td = curthread;
854 dsp->dsa_fp = fp;
855 dsp->dsa_os = os;
856 dsp->dsa_off = off;
857 dsp->dsa_toguid = dsl_dataset_phys(to_ds)->ds_guid;
858 dsp->dsa_pending_op = PENDING_NONE;
859 dsp->dsa_featureflags = featureflags;
860 dsp->dsa_resume_object = resumeobj;
861 dsp->dsa_resume_offset = resumeoff;
862
863 mutex_enter(&to_ds->ds_sendstream_lock);
864 list_insert_head(&to_ds->ds_sendstreams, dsp);
865 mutex_exit(&to_ds->ds_sendstream_lock);
866
867 dsl_dataset_long_hold(to_ds, FTAG);
868 dsl_pool_rele(dp, tag);
869
870 void *payload = NULL;
871 size_t payload_len = 0;
872 if (resumeobj != 0 || resumeoff != 0) {
873 dmu_object_info_t to_doi;
874 err = dmu_object_info(os, resumeobj, &to_doi);
875 if (err != 0)
876 goto out;
877 SET_BOOKMARK(&to_arg.resume, to_ds->ds_object, resumeobj, 0,
878 resumeoff / to_doi.doi_data_block_size);
879
880 nvlist_t *nvl = fnvlist_alloc();
881 fnvlist_add_uint64(nvl, "resume_object", resumeobj);
882 fnvlist_add_uint64(nvl, "resume_offset", resumeoff);
883 payload = fnvlist_pack(nvl, &payload_len);
884 drr->drr_payloadlen = payload_len;
885 fnvlist_free(nvl);
886 }
887
888 err = dump_record(dsp, payload, payload_len);
889 fnvlist_pack_free(payload, payload_len);
890 if (err != 0) {
891 err = dsp->dsa_err;
892 goto out;
893 }
894
895 err = bqueue_init(&to_arg.q, zfs_send_queue_length,
896 offsetof(struct send_block_record, ln));
897 to_arg.error_code = 0;
898 to_arg.cancel = B_FALSE;
899 to_arg.ds = to_ds;
900 to_arg.fromtxg = fromtxg;
901 to_arg.flags = TRAVERSE_PRE | TRAVERSE_PREFETCH;
902 (void) thread_create(NULL, 0, send_traverse_thread, &to_arg, 0, &p0,
903 TS_RUN, minclsyspri);
904
905 struct send_block_record *to_data;
906 to_data = bqueue_dequeue(&to_arg.q);
907
908 while (!to_data->eos_marker && err == 0) {
909 err = do_dump(dsp, to_data);
910 to_data = get_next_record(&to_arg.q, to_data);
911 if (issig(JUSTLOOKING) && issig(FORREAL))
912 err = EINTR;
913 }
914
915 if (err != 0) {
916 to_arg.cancel = B_TRUE;
917 while (!to_data->eos_marker) {
918 to_data = get_next_record(&to_arg.q, to_data);
919 }
920 }
921 kmem_free(to_data, sizeof (*to_data));
922
923 bqueue_destroy(&to_arg.q);
924
925 if (err == 0 && to_arg.error_code != 0)
926 err = to_arg.error_code;
927
928 if (err != 0)
929 goto out;
930
931 if (dsp->dsa_pending_op != PENDING_NONE)
932 if (dump_record(dsp, NULL, 0) != 0)
933 err = SET_ERROR(EINTR);
934
935 if (err != 0) {
936 if (err == EINTR && dsp->dsa_err != 0)
937 err = dsp->dsa_err;
938 goto out;
939 }
940
941 bzero(drr, sizeof (dmu_replay_record_t));
942 drr->drr_type = DRR_END;
943 drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc;
944 drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid;
945
946 if (dump_record(dsp, NULL, 0) != 0)
947 err = dsp->dsa_err;
948
949 out:
950 mutex_enter(&to_ds->ds_sendstream_lock);
951 list_remove(&to_ds->ds_sendstreams, dsp);
952 mutex_exit(&to_ds->ds_sendstream_lock);
953
954 VERIFY(err != 0 || (dsp->dsa_sent_begin && dsp->dsa_sent_end));
955
956 kmem_free(drr, sizeof (dmu_replay_record_t));
957 kmem_free(dsp, sizeof (dmu_sendarg_t));
958
959 dsl_dataset_long_rele(to_ds, FTAG);
960
961 return (err);
962 }
963
964 int
dmu_send_obj(const char * pool,uint64_t tosnap,uint64_t fromsnap,boolean_t embedok,boolean_t large_block_ok,int outfd,vnode_t * vp,offset_t * off)965 dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
966 boolean_t embedok, boolean_t large_block_ok,
967 #ifdef illumos
968 int outfd, vnode_t *vp, offset_t *off)
969 #else
970 int outfd, struct file *fp, offset_t *off)
971 #endif
972 {
973 dsl_pool_t *dp;
974 dsl_dataset_t *ds;
975 dsl_dataset_t *fromds = NULL;
976 int err;
977
978 err = dsl_pool_hold(pool, FTAG, &dp);
979 if (err != 0)
980 return (err);
981
982 err = dsl_dataset_hold_obj(dp, tosnap, FTAG, &ds);
983 if (err != 0) {
984 dsl_pool_rele(dp, FTAG);
985 return (err);
986 }
987
988 if (fromsnap != 0) {
989 zfs_bookmark_phys_t zb;
990 boolean_t is_clone;
991
992 err = dsl_dataset_hold_obj(dp, fromsnap, FTAG, &fromds);
993 if (err != 0) {
994 dsl_dataset_rele(ds, FTAG);
995 dsl_pool_rele(dp, FTAG);
996 return (err);
997 }
998 if (!dsl_dataset_is_before(ds, fromds, 0))
999 err = SET_ERROR(EXDEV);
1000 zb.zbm_creation_time =
1001 dsl_dataset_phys(fromds)->ds_creation_time;
1002 zb.zbm_creation_txg = dsl_dataset_phys(fromds)->ds_creation_txg;
1003 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
1004 is_clone = (fromds->ds_dir != ds->ds_dir);
1005 dsl_dataset_rele(fromds, FTAG);
1006 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
1007 embedok, large_block_ok, outfd, 0, 0, fp, off);
1008 } else {
1009 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
1010 embedok, large_block_ok, outfd, 0, 0, fp, off);
1011 }
1012 dsl_dataset_rele(ds, FTAG);
1013 return (err);
1014 }
1015
1016 int
dmu_send(const char * tosnap,const char * fromsnap,boolean_t embedok,boolean_t large_block_ok,int outfd,uint64_t resumeobj,uint64_t resumeoff,vnode_t * vp,offset_t * off)1017 dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok,
1018 boolean_t large_block_ok, int outfd, uint64_t resumeobj, uint64_t resumeoff,
1019 #ifdef illumos
1020 vnode_t *vp, offset_t *off)
1021 #else
1022 struct file *fp, offset_t *off)
1023 #endif
1024 {
1025 dsl_pool_t *dp;
1026 dsl_dataset_t *ds;
1027 int err;
1028 boolean_t owned = B_FALSE;
1029
1030 if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL)
1031 return (SET_ERROR(EINVAL));
1032
1033 err = dsl_pool_hold(tosnap, FTAG, &dp);
1034 if (err != 0)
1035 return (err);
1036
1037 if (strchr(tosnap, '@') == NULL && spa_writeable(dp->dp_spa)) {
1038 /*
1039 * We are sending a filesystem or volume. Ensure
1040 * that it doesn't change by owning the dataset.
1041 */
1042 err = dsl_dataset_own(dp, tosnap, FTAG, &ds);
1043 owned = B_TRUE;
1044 } else {
1045 err = dsl_dataset_hold(dp, tosnap, FTAG, &ds);
1046 }
1047 if (err != 0) {
1048 dsl_pool_rele(dp, FTAG);
1049 return (err);
1050 }
1051
1052 if (fromsnap != NULL) {
1053 zfs_bookmark_phys_t zb;
1054 boolean_t is_clone = B_FALSE;
1055 int fsnamelen = strchr(tosnap, '@') - tosnap;
1056
1057 /*
1058 * If the fromsnap is in a different filesystem, then
1059 * mark the send stream as a clone.
1060 */
1061 if (strncmp(tosnap, fromsnap, fsnamelen) != 0 ||
1062 (fromsnap[fsnamelen] != '@' &&
1063 fromsnap[fsnamelen] != '#')) {
1064 is_clone = B_TRUE;
1065 }
1066
1067 if (strchr(fromsnap, '@')) {
1068 dsl_dataset_t *fromds;
1069 err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds);
1070 if (err == 0) {
1071 if (!dsl_dataset_is_before(ds, fromds, 0))
1072 err = SET_ERROR(EXDEV);
1073 zb.zbm_creation_time =
1074 dsl_dataset_phys(fromds)->ds_creation_time;
1075 zb.zbm_creation_txg =
1076 dsl_dataset_phys(fromds)->ds_creation_txg;
1077 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
1078 is_clone = (ds->ds_dir != fromds->ds_dir);
1079 dsl_dataset_rele(fromds, FTAG);
1080 }
1081 } else {
1082 err = dsl_bookmark_lookup(dp, fromsnap, ds, &zb);
1083 }
1084 if (err != 0) {
1085 dsl_dataset_rele(ds, FTAG);
1086 dsl_pool_rele(dp, FTAG);
1087 return (err);
1088 }
1089 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
1090 embedok, large_block_ok,
1091 outfd, resumeobj, resumeoff, fp, off);
1092 } else {
1093 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
1094 embedok, large_block_ok,
1095 outfd, resumeobj, resumeoff, fp, off);
1096 }
1097 if (owned)
1098 dsl_dataset_disown(ds, FTAG);
1099 else
1100 dsl_dataset_rele(ds, FTAG);
1101 return (err);
1102 }
1103
1104 static int
dmu_adjust_send_estimate_for_indirects(dsl_dataset_t * ds,uint64_t size,uint64_t * sizep)1105 dmu_adjust_send_estimate_for_indirects(dsl_dataset_t *ds, uint64_t size,
1106 uint64_t *sizep)
1107 {
1108 int err;
1109 /*
1110 * Assume that space (both on-disk and in-stream) is dominated by
1111 * data. We will adjust for indirect blocks and the copies property,
1112 * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
1113 */
1114
1115 /*
1116 * Subtract out approximate space used by indirect blocks.
1117 * Assume most space is used by data blocks (non-indirect, non-dnode).
1118 * Assume all blocks are recordsize. Assume ditto blocks and
1119 * internal fragmentation counter out compression.
1120 *
1121 * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
1122 * block, which we observe in practice.
1123 */
1124 uint64_t recordsize;
1125 err = dsl_prop_get_int_ds(ds, "recordsize", &recordsize);
1126 if (err != 0)
1127 return (err);
1128 size -= size / recordsize * sizeof (blkptr_t);
1129
1130 /* Add in the space for the record associated with each block. */
1131 size += size / recordsize * sizeof (dmu_replay_record_t);
1132
1133 *sizep = size;
1134
1135 return (0);
1136 }
1137
1138 int
dmu_send_estimate(dsl_dataset_t * ds,dsl_dataset_t * fromds,uint64_t * sizep)1139 dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds, uint64_t *sizep)
1140 {
1141 dsl_pool_t *dp = ds->ds_dir->dd_pool;
1142 int err;
1143 uint64_t size;
1144
1145 ASSERT(dsl_pool_config_held(dp));
1146
1147 /* tosnap must be a snapshot */
1148 if (!ds->ds_is_snapshot)
1149 return (SET_ERROR(EINVAL));
1150
1151 /* fromsnap, if provided, must be a snapshot */
1152 if (fromds != NULL && !fromds->ds_is_snapshot)
1153 return (SET_ERROR(EINVAL));
1154
1155 /*
1156 * fromsnap must be an earlier snapshot from the same fs as tosnap,
1157 * or the origin's fs.
1158 */
1159 if (fromds != NULL && !dsl_dataset_is_before(ds, fromds, 0))
1160 return (SET_ERROR(EXDEV));
1161
1162 /* Get uncompressed size estimate of changed data. */
1163 if (fromds == NULL) {
1164 size = dsl_dataset_phys(ds)->ds_uncompressed_bytes;
1165 } else {
1166 uint64_t used, comp;
1167 err = dsl_dataset_space_written(fromds, ds,
1168 &used, &comp, &size);
1169 if (err != 0)
1170 return (err);
1171 }
1172
1173 err = dmu_adjust_send_estimate_for_indirects(ds, size, sizep);
1174 return (err);
1175 }
1176
1177 /*
1178 * Simple callback used to traverse the blocks of a snapshot and sum their
1179 * uncompressed size
1180 */
1181 /* ARGSUSED */
1182 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)1183 dmu_calculate_send_traversal(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
1184 const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
1185 {
1186 uint64_t *spaceptr = arg;
1187 if (bp != NULL && !BP_IS_HOLE(bp)) {
1188 *spaceptr += BP_GET_UCSIZE(bp);
1189 }
1190 return (0);
1191 }
1192
1193 /*
1194 * Given a desination snapshot and a TXG, calculate the approximate size of a
1195 * send stream sent from that TXG. from_txg may be zero, indicating that the
1196 * whole snapshot will be sent.
1197 */
1198 int
dmu_send_estimate_from_txg(dsl_dataset_t * ds,uint64_t from_txg,uint64_t * sizep)1199 dmu_send_estimate_from_txg(dsl_dataset_t *ds, uint64_t from_txg,
1200 uint64_t *sizep)
1201 {
1202 dsl_pool_t *dp = ds->ds_dir->dd_pool;
1203 int err;
1204 uint64_t size = 0;
1205
1206 ASSERT(dsl_pool_config_held(dp));
1207
1208 /* tosnap must be a snapshot */
1209 if (!dsl_dataset_is_snapshot(ds))
1210 return (SET_ERROR(EINVAL));
1211
1212 /* verify that from_txg is before the provided snapshot was taken */
1213 if (from_txg >= dsl_dataset_phys(ds)->ds_creation_txg) {
1214 return (SET_ERROR(EXDEV));
1215 }
1216
1217 /*
1218 * traverse the blocks of the snapshot with birth times after
1219 * from_txg, summing their uncompressed size
1220 */
1221 err = traverse_dataset(ds, from_txg, TRAVERSE_POST,
1222 dmu_calculate_send_traversal, &size);
1223 if (err)
1224 return (err);
1225
1226 err = dmu_adjust_send_estimate_for_indirects(ds, size, sizep);
1227 return (err);
1228 }
1229
1230 typedef struct dmu_recv_begin_arg {
1231 const char *drba_origin;
1232 dmu_recv_cookie_t *drba_cookie;
1233 cred_t *drba_cred;
1234 uint64_t drba_snapobj;
1235 } dmu_recv_begin_arg_t;
1236
1237 static int
recv_begin_check_existing_impl(dmu_recv_begin_arg_t * drba,dsl_dataset_t * ds,uint64_t fromguid)1238 recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds,
1239 uint64_t fromguid)
1240 {
1241 uint64_t val;
1242 int error;
1243 dsl_pool_t *dp = ds->ds_dir->dd_pool;
1244
1245 /* temporary clone name must not exist */
1246 error = zap_lookup(dp->dp_meta_objset,
1247 dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, recv_clone_name,
1248 8, 1, &val);
1249 if (error != ENOENT)
1250 return (error == 0 ? EBUSY : error);
1251
1252 /* new snapshot name must not exist */
1253 error = zap_lookup(dp->dp_meta_objset,
1254 dsl_dataset_phys(ds)->ds_snapnames_zapobj,
1255 drba->drba_cookie->drc_tosnap, 8, 1, &val);
1256 if (error != ENOENT)
1257 return (error == 0 ? EEXIST : error);
1258
1259 /*
1260 * Check snapshot limit before receiving. We'll recheck again at the
1261 * end, but might as well abort before receiving if we're already over
1262 * the limit.
1263 *
1264 * Note that we do not check the file system limit with
1265 * dsl_dir_fscount_check because the temporary %clones don't count
1266 * against that limit.
1267 */
1268 error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT,
1269 NULL, drba->drba_cred);
1270 if (error != 0)
1271 return (error);
1272
1273 if (fromguid != 0) {
1274 dsl_dataset_t *snap;
1275 uint64_t obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
1276
1277 /* Find snapshot in this dir that matches fromguid. */
1278 while (obj != 0) {
1279 error = dsl_dataset_hold_obj(dp, obj, FTAG,
1280 &snap);
1281 if (error != 0)
1282 return (SET_ERROR(ENODEV));
1283 if (snap->ds_dir != ds->ds_dir) {
1284 dsl_dataset_rele(snap, FTAG);
1285 return (SET_ERROR(ENODEV));
1286 }
1287 if (dsl_dataset_phys(snap)->ds_guid == fromguid)
1288 break;
1289 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
1290 dsl_dataset_rele(snap, FTAG);
1291 }
1292 if (obj == 0)
1293 return (SET_ERROR(ENODEV));
1294
1295 if (drba->drba_cookie->drc_force) {
1296 drba->drba_snapobj = obj;
1297 } else {
1298 /*
1299 * If we are not forcing, there must be no
1300 * changes since fromsnap.
1301 */
1302 if (dsl_dataset_modified_since_snap(ds, snap)) {
1303 dsl_dataset_rele(snap, FTAG);
1304 return (SET_ERROR(ETXTBSY));
1305 }
1306 drba->drba_snapobj = ds->ds_prev->ds_object;
1307 }
1308
1309 dsl_dataset_rele(snap, FTAG);
1310 } else {
1311 /* if full, then must be forced */
1312 if (!drba->drba_cookie->drc_force)
1313 return (SET_ERROR(EEXIST));
1314 /* start from $ORIGIN@$ORIGIN, if supported */
1315 drba->drba_snapobj = dp->dp_origin_snap != NULL ?
1316 dp->dp_origin_snap->ds_object : 0;
1317 }
1318
1319 return (0);
1320
1321 }
1322
1323 static int
dmu_recv_begin_check(void * arg,dmu_tx_t * tx)1324 dmu_recv_begin_check(void *arg, dmu_tx_t *tx)
1325 {
1326 dmu_recv_begin_arg_t *drba = arg;
1327 dsl_pool_t *dp = dmu_tx_pool(tx);
1328 struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1329 uint64_t fromguid = drrb->drr_fromguid;
1330 int flags = drrb->drr_flags;
1331 int error;
1332 uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1333 dsl_dataset_t *ds;
1334 const char *tofs = drba->drba_cookie->drc_tofs;
1335
1336 /* already checked */
1337 ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1338 ASSERT(!(featureflags & DMU_BACKUP_FEATURE_RESUMING));
1339
1340 if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1341 DMU_COMPOUNDSTREAM ||
1342 drrb->drr_type >= DMU_OST_NUMTYPES ||
1343 ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL))
1344 return (SET_ERROR(EINVAL));
1345
1346 /* Verify pool version supports SA if SA_SPILL feature set */
1347 if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1348 spa_version(dp->dp_spa) < SPA_VERSION_SA)
1349 return (SET_ERROR(ENOTSUP));
1350
1351 if (drba->drba_cookie->drc_resumable &&
1352 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EXTENSIBLE_DATASET))
1353 return (SET_ERROR(ENOTSUP));
1354
1355 /*
1356 * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1357 * record to a plan WRITE record, so the pool must have the
1358 * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1359 * records. Same with WRITE_EMBEDDED records that use LZ4 compression.
1360 */
1361 if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1362 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1363 return (SET_ERROR(ENOTSUP));
1364 if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4) &&
1365 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1366 return (SET_ERROR(ENOTSUP));
1367
1368 /*
1369 * The receiving code doesn't know how to translate large blocks
1370 * to smaller ones, so the pool must have the LARGE_BLOCKS
1371 * feature enabled if the stream has LARGE_BLOCKS.
1372 */
1373 if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
1374 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS))
1375 return (SET_ERROR(ENOTSUP));
1376
1377 error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1378 if (error == 0) {
1379 /* target fs already exists; recv into temp clone */
1380
1381 /* Can't recv a clone into an existing fs */
1382 if (flags & DRR_FLAG_CLONE || drba->drba_origin) {
1383 dsl_dataset_rele(ds, FTAG);
1384 return (SET_ERROR(EINVAL));
1385 }
1386
1387 error = recv_begin_check_existing_impl(drba, ds, fromguid);
1388 dsl_dataset_rele(ds, FTAG);
1389 } else if (error == ENOENT) {
1390 /* target fs does not exist; must be a full backup or clone */
1391 char buf[ZFS_MAX_DATASET_NAME_LEN];
1392
1393 /*
1394 * If it's a non-clone incremental, we are missing the
1395 * target fs, so fail the recv.
1396 */
1397 if (fromguid != 0 && !(flags & DRR_FLAG_CLONE ||
1398 drba->drba_origin))
1399 return (SET_ERROR(ENOENT));
1400
1401 /*
1402 * If we're receiving a full send as a clone, and it doesn't
1403 * contain all the necessary free records and freeobject
1404 * records, reject it.
1405 */
1406 if (fromguid == 0 && drba->drba_origin &&
1407 !(flags & DRR_FLAG_FREERECORDS))
1408 return (SET_ERROR(EINVAL));
1409
1410 /* Open the parent of tofs */
1411 ASSERT3U(strlen(tofs), <, sizeof (buf));
1412 (void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1);
1413 error = dsl_dataset_hold(dp, buf, FTAG, &ds);
1414 if (error != 0)
1415 return (error);
1416
1417 /*
1418 * Check filesystem and snapshot limits before receiving. We'll
1419 * recheck snapshot limits again at the end (we create the
1420 * filesystems and increment those counts during begin_sync).
1421 */
1422 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1423 ZFS_PROP_FILESYSTEM_LIMIT, NULL, drba->drba_cred);
1424 if (error != 0) {
1425 dsl_dataset_rele(ds, FTAG);
1426 return (error);
1427 }
1428
1429 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1430 ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred);
1431 if (error != 0) {
1432 dsl_dataset_rele(ds, FTAG);
1433 return (error);
1434 }
1435
1436 if (drba->drba_origin != NULL) {
1437 dsl_dataset_t *origin;
1438 error = dsl_dataset_hold(dp, drba->drba_origin,
1439 FTAG, &origin);
1440 if (error != 0) {
1441 dsl_dataset_rele(ds, FTAG);
1442 return (error);
1443 }
1444 if (!origin->ds_is_snapshot) {
1445 dsl_dataset_rele(origin, FTAG);
1446 dsl_dataset_rele(ds, FTAG);
1447 return (SET_ERROR(EINVAL));
1448 }
1449 if (dsl_dataset_phys(origin)->ds_guid != fromguid &&
1450 fromguid != 0) {
1451 dsl_dataset_rele(origin, FTAG);
1452 dsl_dataset_rele(ds, FTAG);
1453 return (SET_ERROR(ENODEV));
1454 }
1455 dsl_dataset_rele(origin, FTAG);
1456 }
1457 dsl_dataset_rele(ds, FTAG);
1458 error = 0;
1459 }
1460 return (error);
1461 }
1462
1463 static void
dmu_recv_begin_sync(void * arg,dmu_tx_t * tx)1464 dmu_recv_begin_sync(void *arg, dmu_tx_t *tx)
1465 {
1466 dmu_recv_begin_arg_t *drba = arg;
1467 dsl_pool_t *dp = dmu_tx_pool(tx);
1468 objset_t *mos = dp->dp_meta_objset;
1469 struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1470 const char *tofs = drba->drba_cookie->drc_tofs;
1471 dsl_dataset_t *ds, *newds;
1472 uint64_t dsobj;
1473 int error;
1474 uint64_t crflags = 0;
1475
1476 if (drrb->drr_flags & DRR_FLAG_CI_DATA)
1477 crflags |= DS_FLAG_CI_DATASET;
1478
1479 error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1480 if (error == 0) {
1481 /* create temporary clone */
1482 dsl_dataset_t *snap = NULL;
1483 if (drba->drba_snapobj != 0) {
1484 VERIFY0(dsl_dataset_hold_obj(dp,
1485 drba->drba_snapobj, FTAG, &snap));
1486 }
1487 dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name,
1488 snap, crflags, drba->drba_cred, tx);
1489 if (drba->drba_snapobj != 0)
1490 dsl_dataset_rele(snap, FTAG);
1491 dsl_dataset_rele(ds, FTAG);
1492 } else {
1493 dsl_dir_t *dd;
1494 const char *tail;
1495 dsl_dataset_t *origin = NULL;
1496
1497 VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail));
1498
1499 if (drba->drba_origin != NULL) {
1500 VERIFY0(dsl_dataset_hold(dp, drba->drba_origin,
1501 FTAG, &origin));
1502 }
1503
1504 /* Create new dataset. */
1505 dsobj = dsl_dataset_create_sync(dd,
1506 strrchr(tofs, '/') + 1,
1507 origin, crflags, drba->drba_cred, tx);
1508 if (origin != NULL)
1509 dsl_dataset_rele(origin, FTAG);
1510 dsl_dir_rele(dd, FTAG);
1511 drba->drba_cookie->drc_newfs = B_TRUE;
1512 }
1513 VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds));
1514
1515 if (drba->drba_cookie->drc_resumable) {
1516 dsl_dataset_zapify(newds, tx);
1517 if (drrb->drr_fromguid != 0) {
1518 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_FROMGUID,
1519 8, 1, &drrb->drr_fromguid, tx));
1520 }
1521 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TOGUID,
1522 8, 1, &drrb->drr_toguid, tx));
1523 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TONAME,
1524 1, strlen(drrb->drr_toname) + 1, drrb->drr_toname, tx));
1525 uint64_t one = 1;
1526 uint64_t zero = 0;
1527 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OBJECT,
1528 8, 1, &one, tx));
1529 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OFFSET,
1530 8, 1, &zero, tx));
1531 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_BYTES,
1532 8, 1, &zero, tx));
1533 if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
1534 DMU_BACKUP_FEATURE_EMBED_DATA) {
1535 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_EMBEDOK,
1536 8, 1, &one, tx));
1537 }
1538 }
1539
1540 dmu_buf_will_dirty(newds->ds_dbuf, tx);
1541 dsl_dataset_phys(newds)->ds_flags |= DS_FLAG_INCONSISTENT;
1542
1543 /*
1544 * If we actually created a non-clone, we need to create the
1545 * objset in our new dataset.
1546 */
1547 rrw_enter(&newds->ds_bp_rwlock, RW_READER, FTAG);
1548 if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) {
1549 (void) dmu_objset_create_impl(dp->dp_spa,
1550 newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx);
1551 }
1552 rrw_exit(&newds->ds_bp_rwlock, FTAG);
1553
1554 drba->drba_cookie->drc_ds = newds;
1555
1556 spa_history_log_internal_ds(newds, "receive", tx, "");
1557 }
1558
1559 static int
dmu_recv_resume_begin_check(void * arg,dmu_tx_t * tx)1560 dmu_recv_resume_begin_check(void *arg, dmu_tx_t *tx)
1561 {
1562 dmu_recv_begin_arg_t *drba = arg;
1563 dsl_pool_t *dp = dmu_tx_pool(tx);
1564 struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1565 int error;
1566 uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1567 dsl_dataset_t *ds;
1568 const char *tofs = drba->drba_cookie->drc_tofs;
1569
1570 /* already checked */
1571 ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1572 ASSERT(featureflags & DMU_BACKUP_FEATURE_RESUMING);
1573
1574 if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1575 DMU_COMPOUNDSTREAM ||
1576 drrb->drr_type >= DMU_OST_NUMTYPES)
1577 return (SET_ERROR(EINVAL));
1578
1579 /* Verify pool version supports SA if SA_SPILL feature set */
1580 if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1581 spa_version(dp->dp_spa) < SPA_VERSION_SA)
1582 return (SET_ERROR(ENOTSUP));
1583
1584 /*
1585 * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1586 * record to a plain WRITE record, so the pool must have the
1587 * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1588 * records. Same with WRITE_EMBEDDED records that use LZ4 compression.
1589 */
1590 if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1591 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1592 return (SET_ERROR(ENOTSUP));
1593 if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4) &&
1594 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1595 return (SET_ERROR(ENOTSUP));
1596
1597 /* 6 extra bytes for /%recv */
1598 char recvname[ZFS_MAX_DATASET_NAME_LEN + 6];
1599
1600 (void) snprintf(recvname, sizeof (recvname), "%s/%s",
1601 tofs, recv_clone_name);
1602
1603 if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
1604 /* %recv does not exist; continue in tofs */
1605 error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1606 if (error != 0)
1607 return (error);
1608 }
1609
1610 /* check that ds is marked inconsistent */
1611 if (!DS_IS_INCONSISTENT(ds)) {
1612 dsl_dataset_rele(ds, FTAG);
1613 return (SET_ERROR(EINVAL));
1614 }
1615
1616 /* check that there is resuming data, and that the toguid matches */
1617 if (!dsl_dataset_is_zapified(ds)) {
1618 dsl_dataset_rele(ds, FTAG);
1619 return (SET_ERROR(EINVAL));
1620 }
1621 uint64_t val;
1622 error = zap_lookup(dp->dp_meta_objset, ds->ds_object,
1623 DS_FIELD_RESUME_TOGUID, sizeof (val), 1, &val);
1624 if (error != 0 || drrb->drr_toguid != val) {
1625 dsl_dataset_rele(ds, FTAG);
1626 return (SET_ERROR(EINVAL));
1627 }
1628
1629 /*
1630 * Check if the receive is still running. If so, it will be owned.
1631 * Note that nothing else can own the dataset (e.g. after the receive
1632 * fails) because it will be marked inconsistent.
1633 */
1634 if (dsl_dataset_has_owner(ds)) {
1635 dsl_dataset_rele(ds, FTAG);
1636 return (SET_ERROR(EBUSY));
1637 }
1638
1639 /* There should not be any snapshots of this fs yet. */
1640 if (ds->ds_prev != NULL && ds->ds_prev->ds_dir == ds->ds_dir) {
1641 dsl_dataset_rele(ds, FTAG);
1642 return (SET_ERROR(EINVAL));
1643 }
1644
1645 /*
1646 * Note: resume point will be checked when we process the first WRITE
1647 * record.
1648 */
1649
1650 /* check that the origin matches */
1651 val = 0;
1652 (void) zap_lookup(dp->dp_meta_objset, ds->ds_object,
1653 DS_FIELD_RESUME_FROMGUID, sizeof (val), 1, &val);
1654 if (drrb->drr_fromguid != val) {
1655 dsl_dataset_rele(ds, FTAG);
1656 return (SET_ERROR(EINVAL));
1657 }
1658
1659 dsl_dataset_rele(ds, FTAG);
1660 return (0);
1661 }
1662
1663 static void
dmu_recv_resume_begin_sync(void * arg,dmu_tx_t * tx)1664 dmu_recv_resume_begin_sync(void *arg, dmu_tx_t *tx)
1665 {
1666 dmu_recv_begin_arg_t *drba = arg;
1667 dsl_pool_t *dp = dmu_tx_pool(tx);
1668 const char *tofs = drba->drba_cookie->drc_tofs;
1669 dsl_dataset_t *ds;
1670 uint64_t dsobj;
1671 /* 6 extra bytes for /%recv */
1672 char recvname[ZFS_MAX_DATASET_NAME_LEN + 6];
1673
1674 (void) snprintf(recvname, sizeof (recvname), "%s/%s",
1675 tofs, recv_clone_name);
1676
1677 if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
1678 /* %recv does not exist; continue in tofs */
1679 VERIFY0(dsl_dataset_hold(dp, tofs, FTAG, &ds));
1680 drba->drba_cookie->drc_newfs = B_TRUE;
1681 }
1682
1683 /* clear the inconsistent flag so that we can own it */
1684 ASSERT(DS_IS_INCONSISTENT(ds));
1685 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1686 dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
1687 dsobj = ds->ds_object;
1688 dsl_dataset_rele(ds, FTAG);
1689
1690 VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &ds));
1691
1692 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1693 dsl_dataset_phys(ds)->ds_flags |= DS_FLAG_INCONSISTENT;
1694
1695 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
1696 ASSERT(!BP_IS_HOLE(dsl_dataset_get_blkptr(ds)));
1697 rrw_exit(&ds->ds_bp_rwlock, FTAG);
1698
1699 drba->drba_cookie->drc_ds = ds;
1700
1701 spa_history_log_internal_ds(ds, "resume receive", tx, "");
1702 }
1703
1704 /*
1705 * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
1706 * succeeds; otherwise we will leak the holds on the datasets.
1707 */
1708 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)1709 dmu_recv_begin(char *tofs, char *tosnap, dmu_replay_record_t *drr_begin,
1710 boolean_t force, boolean_t resumable, char *origin, dmu_recv_cookie_t *drc)
1711 {
1712 dmu_recv_begin_arg_t drba = { 0 };
1713
1714 bzero(drc, sizeof (dmu_recv_cookie_t));
1715 drc->drc_drr_begin = drr_begin;
1716 drc->drc_drrb = &drr_begin->drr_u.drr_begin;
1717 drc->drc_tosnap = tosnap;
1718 drc->drc_tofs = tofs;
1719 drc->drc_force = force;
1720 drc->drc_resumable = resumable;
1721 drc->drc_cred = CRED();
1722
1723 if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC)) {
1724 drc->drc_byteswap = B_TRUE;
1725 fletcher_4_incremental_byteswap(drr_begin,
1726 sizeof (dmu_replay_record_t), &drc->drc_cksum);
1727 byteswap_record(drr_begin);
1728 } else if (drc->drc_drrb->drr_magic == DMU_BACKUP_MAGIC) {
1729 fletcher_4_incremental_native(drr_begin,
1730 sizeof (dmu_replay_record_t), &drc->drc_cksum);
1731 } else {
1732 return (SET_ERROR(EINVAL));
1733 }
1734
1735 drba.drba_origin = origin;
1736 drba.drba_cookie = drc;
1737 drba.drba_cred = CRED();
1738
1739 if (DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo) &
1740 DMU_BACKUP_FEATURE_RESUMING) {
1741 return (dsl_sync_task(tofs,
1742 dmu_recv_resume_begin_check, dmu_recv_resume_begin_sync,
1743 &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1744 } else {
1745 return (dsl_sync_task(tofs,
1746 dmu_recv_begin_check, dmu_recv_begin_sync,
1747 &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1748 }
1749 }
1750
1751 struct receive_record_arg {
1752 dmu_replay_record_t header;
1753 void *payload; /* Pointer to a buffer containing the payload */
1754 /*
1755 * If the record is a write, pointer to the arc_buf_t containing the
1756 * payload.
1757 */
1758 arc_buf_t *write_buf;
1759 int payload_size;
1760 uint64_t bytes_read; /* bytes read from stream when record created */
1761 boolean_t eos_marker; /* Marks the end of the stream */
1762 bqueue_node_t node;
1763 };
1764
1765 struct receive_writer_arg {
1766 objset_t *os;
1767 boolean_t byteswap;
1768 bqueue_t q;
1769
1770 /*
1771 * These three args are used to signal to the main thread that we're
1772 * done.
1773 */
1774 kmutex_t mutex;
1775 kcondvar_t cv;
1776 boolean_t done;
1777
1778 int err;
1779 /* A map from guid to dataset to help handle dedup'd streams. */
1780 avl_tree_t *guid_to_ds_map;
1781 boolean_t resumable;
1782 uint64_t last_object, last_offset;
1783 uint64_t bytes_read; /* bytes read when current record created */
1784 };
1785
1786 struct objlist {
1787 list_t list; /* List of struct receive_objnode. */
1788 /*
1789 * Last object looked up. Used to assert that objects are being looked
1790 * up in ascending order.
1791 */
1792 uint64_t last_lookup;
1793 };
1794
1795 struct receive_objnode {
1796 list_node_t node;
1797 uint64_t object;
1798 };
1799
1800 struct receive_arg {
1801 objset_t *os;
1802 kthread_t *td;
1803 struct file *fp;
1804 uint64_t voff; /* The current offset in the stream */
1805 uint64_t bytes_read;
1806 /*
1807 * A record that has had its payload read in, but hasn't yet been handed
1808 * off to the worker thread.
1809 */
1810 struct receive_record_arg *rrd;
1811 /* A record that has had its header read in, but not its payload. */
1812 struct receive_record_arg *next_rrd;
1813 zio_cksum_t cksum;
1814 zio_cksum_t prev_cksum;
1815 int err;
1816 boolean_t byteswap;
1817 /* Sorted list of objects not to issue prefetches for. */
1818 struct objlist ignore_objlist;
1819 };
1820
1821 typedef struct guid_map_entry {
1822 uint64_t guid;
1823 dsl_dataset_t *gme_ds;
1824 avl_node_t avlnode;
1825 } guid_map_entry_t;
1826
1827 static int
guid_compare(const void * arg1,const void * arg2)1828 guid_compare(const void *arg1, const void *arg2)
1829 {
1830 const guid_map_entry_t *gmep1 = arg1;
1831 const guid_map_entry_t *gmep2 = arg2;
1832
1833 if (gmep1->guid < gmep2->guid)
1834 return (-1);
1835 else if (gmep1->guid > gmep2->guid)
1836 return (1);
1837 return (0);
1838 }
1839
1840 static void
free_guid_map_onexit(void * arg)1841 free_guid_map_onexit(void *arg)
1842 {
1843 avl_tree_t *ca = arg;
1844 void *cookie = NULL;
1845 guid_map_entry_t *gmep;
1846
1847 while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
1848 dsl_dataset_long_rele(gmep->gme_ds, gmep);
1849 dsl_dataset_rele(gmep->gme_ds, gmep);
1850 kmem_free(gmep, sizeof (guid_map_entry_t));
1851 }
1852 avl_destroy(ca);
1853 kmem_free(ca, sizeof (avl_tree_t));
1854 }
1855
1856 static int
restore_bytes(struct receive_arg * ra,void * buf,int len,off_t off,ssize_t * resid)1857 restore_bytes(struct receive_arg *ra, void *buf, int len, off_t off, ssize_t *resid)
1858 {
1859 struct uio auio;
1860 struct iovec aiov;
1861 int error;
1862
1863 aiov.iov_base = buf;
1864 aiov.iov_len = len;
1865 auio.uio_iov = &aiov;
1866 auio.uio_iovcnt = 1;
1867 auio.uio_resid = len;
1868 #ifdef __NetBSD__
1869 #ifdef _KERNEL
1870 auio.uio_vmspace = vmspace_kernel();
1871 #endif
1872 #else
1873 auio.uio_segflg = UIO_SYSSPACE;
1874 #endif
1875 auio.uio_rw = UIO_READ;
1876 auio.uio_offset = off;
1877 #ifdef __FreeBSD__
1878 auio.uio_td = ra->td;
1879 #endif
1880 #ifdef _KERNEL
1881 error = fo_read(ra->fp, &auio, ra->td->td_ucred, FOF_OFFSET, ra->td);
1882 #else
1883 fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__);
1884 error = EOPNOTSUPP;
1885 #endif
1886 *resid = auio.uio_resid;
1887 return (error);
1888 }
1889
1890 static int
receive_read(struct receive_arg * ra,int len,void * buf)1891 receive_read(struct receive_arg *ra, int len, void *buf)
1892 {
1893 int done = 0;
1894
1895 /*
1896 * The code doesn't rely on this (lengths being multiples of 8). See
1897 * comment in dump_bytes.
1898 */
1899 ASSERT0(len % 8);
1900
1901 while (done < len) {
1902 ssize_t resid;
1903
1904 ra->err = restore_bytes(ra, buf + done,
1905 len - done, ra->voff, &resid);
1906
1907 if (resid == len - done) {
1908 /*
1909 * Note: ECKSUM indicates that the receive
1910 * was interrupted and can potentially be resumed.
1911 */
1912 ra->err = SET_ERROR(ECKSUM);
1913 }
1914 ra->voff += len - done - resid;
1915 done = len - resid;
1916 if (ra->err != 0)
1917 return (ra->err);
1918 }
1919
1920 ra->bytes_read += len;
1921
1922 ASSERT3U(done, ==, len);
1923 return (0);
1924 }
1925
1926 static void
byteswap_record(dmu_replay_record_t * drr)1927 byteswap_record(dmu_replay_record_t *drr)
1928 {
1929 #define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
1930 #define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
1931 drr->drr_type = BSWAP_32(drr->drr_type);
1932 drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
1933
1934 switch (drr->drr_type) {
1935 case DRR_BEGIN:
1936 DO64(drr_begin.drr_magic);
1937 DO64(drr_begin.drr_versioninfo);
1938 DO64(drr_begin.drr_creation_time);
1939 DO32(drr_begin.drr_type);
1940 DO32(drr_begin.drr_flags);
1941 DO64(drr_begin.drr_toguid);
1942 DO64(drr_begin.drr_fromguid);
1943 break;
1944 case DRR_OBJECT:
1945 DO64(drr_object.drr_object);
1946 DO32(drr_object.drr_type);
1947 DO32(drr_object.drr_bonustype);
1948 DO32(drr_object.drr_blksz);
1949 DO32(drr_object.drr_bonuslen);
1950 DO64(drr_object.drr_toguid);
1951 break;
1952 case DRR_FREEOBJECTS:
1953 DO64(drr_freeobjects.drr_firstobj);
1954 DO64(drr_freeobjects.drr_numobjs);
1955 DO64(drr_freeobjects.drr_toguid);
1956 break;
1957 case DRR_WRITE:
1958 DO64(drr_write.drr_object);
1959 DO32(drr_write.drr_type);
1960 DO64(drr_write.drr_offset);
1961 DO64(drr_write.drr_length);
1962 DO64(drr_write.drr_toguid);
1963 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write.drr_key.ddk_cksum);
1964 DO64(drr_write.drr_key.ddk_prop);
1965 break;
1966 case DRR_WRITE_BYREF:
1967 DO64(drr_write_byref.drr_object);
1968 DO64(drr_write_byref.drr_offset);
1969 DO64(drr_write_byref.drr_length);
1970 DO64(drr_write_byref.drr_toguid);
1971 DO64(drr_write_byref.drr_refguid);
1972 DO64(drr_write_byref.drr_refobject);
1973 DO64(drr_write_byref.drr_refoffset);
1974 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write_byref.
1975 drr_key.ddk_cksum);
1976 DO64(drr_write_byref.drr_key.ddk_prop);
1977 break;
1978 case DRR_WRITE_EMBEDDED:
1979 DO64(drr_write_embedded.drr_object);
1980 DO64(drr_write_embedded.drr_offset);
1981 DO64(drr_write_embedded.drr_length);
1982 DO64(drr_write_embedded.drr_toguid);
1983 DO32(drr_write_embedded.drr_lsize);
1984 DO32(drr_write_embedded.drr_psize);
1985 break;
1986 case DRR_FREE:
1987 DO64(drr_free.drr_object);
1988 DO64(drr_free.drr_offset);
1989 DO64(drr_free.drr_length);
1990 DO64(drr_free.drr_toguid);
1991 break;
1992 case DRR_SPILL:
1993 DO64(drr_spill.drr_object);
1994 DO64(drr_spill.drr_length);
1995 DO64(drr_spill.drr_toguid);
1996 break;
1997 case DRR_END:
1998 DO64(drr_end.drr_toguid);
1999 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_end.drr_checksum);
2000 break;
2001 }
2002
2003 if (drr->drr_type != DRR_BEGIN) {
2004 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_checksum.drr_checksum);
2005 }
2006
2007 #undef DO64
2008 #undef DO32
2009 }
2010
2011 static inline uint8_t
deduce_nblkptr(dmu_object_type_t bonus_type,uint64_t bonus_size)2012 deduce_nblkptr(dmu_object_type_t bonus_type, uint64_t bonus_size)
2013 {
2014 if (bonus_type == DMU_OT_SA) {
2015 return (1);
2016 } else {
2017 return (1 +
2018 ((DN_MAX_BONUSLEN - bonus_size) >> SPA_BLKPTRSHIFT));
2019 }
2020 }
2021
2022 static void
save_resume_state(struct receive_writer_arg * rwa,uint64_t object,uint64_t offset,dmu_tx_t * tx)2023 save_resume_state(struct receive_writer_arg *rwa,
2024 uint64_t object, uint64_t offset, dmu_tx_t *tx)
2025 {
2026 int txgoff = dmu_tx_get_txg(tx) & TXG_MASK;
2027
2028 if (!rwa->resumable)
2029 return;
2030
2031 /*
2032 * We use ds_resume_bytes[] != 0 to indicate that we need to
2033 * update this on disk, so it must not be 0.
2034 */
2035 ASSERT(rwa->bytes_read != 0);
2036
2037 /*
2038 * We only resume from write records, which have a valid
2039 * (non-meta-dnode) object number.
2040 */
2041 ASSERT(object != 0);
2042
2043 /*
2044 * For resuming to work correctly, we must receive records in order,
2045 * sorted by object,offset. This is checked by the callers, but
2046 * assert it here for good measure.
2047 */
2048 ASSERT3U(object, >=, rwa->os->os_dsl_dataset->ds_resume_object[txgoff]);
2049 ASSERT(object != rwa->os->os_dsl_dataset->ds_resume_object[txgoff] ||
2050 offset >= rwa->os->os_dsl_dataset->ds_resume_offset[txgoff]);
2051 ASSERT3U(rwa->bytes_read, >=,
2052 rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff]);
2053
2054 rwa->os->os_dsl_dataset->ds_resume_object[txgoff] = object;
2055 rwa->os->os_dsl_dataset->ds_resume_offset[txgoff] = offset;
2056 rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff] = rwa->bytes_read;
2057 }
2058
2059 static int
receive_object(struct receive_writer_arg * rwa,struct drr_object * drro,void * data)2060 receive_object(struct receive_writer_arg *rwa, struct drr_object *drro,
2061 void *data)
2062 {
2063 dmu_object_info_t doi;
2064 dmu_tx_t *tx;
2065 uint64_t object;
2066 int err;
2067
2068 if (drro->drr_type == DMU_OT_NONE ||
2069 !DMU_OT_IS_VALID(drro->drr_type) ||
2070 !DMU_OT_IS_VALID(drro->drr_bonustype) ||
2071 drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
2072 drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
2073 P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
2074 drro->drr_blksz < SPA_MINBLOCKSIZE ||
2075 drro->drr_blksz > spa_maxblocksize(dmu_objset_spa(rwa->os)) ||
2076 drro->drr_bonuslen > DN_MAX_BONUSLEN) {
2077 return (SET_ERROR(EINVAL));
2078 }
2079
2080 err = dmu_object_info(rwa->os, drro->drr_object, &doi);
2081
2082 if (err != 0 && err != ENOENT)
2083 return (SET_ERROR(EINVAL));
2084 object = err == 0 ? drro->drr_object : DMU_NEW_OBJECT;
2085
2086 /*
2087 * If we are losing blkptrs or changing the block size this must
2088 * be a new file instance. We must clear out the previous file
2089 * contents before we can change this type of metadata in the dnode.
2090 */
2091 if (err == 0) {
2092 int nblkptr;
2093
2094 nblkptr = deduce_nblkptr(drro->drr_bonustype,
2095 drro->drr_bonuslen);
2096
2097 if (drro->drr_blksz != doi.doi_data_block_size ||
2098 nblkptr < doi.doi_nblkptr) {
2099 err = dmu_free_long_range(rwa->os, drro->drr_object,
2100 0, DMU_OBJECT_END);
2101 if (err != 0)
2102 return (SET_ERROR(EINVAL));
2103 }
2104 }
2105
2106 tx = dmu_tx_create(rwa->os);
2107 dmu_tx_hold_bonus(tx, object);
2108 err = dmu_tx_assign(tx, TXG_WAIT);
2109 if (err != 0) {
2110 dmu_tx_abort(tx);
2111 return (err);
2112 }
2113
2114 if (object == DMU_NEW_OBJECT) {
2115 /* currently free, want to be allocated */
2116 err = dmu_object_claim(rwa->os, drro->drr_object,
2117 drro->drr_type, drro->drr_blksz,
2118 drro->drr_bonustype, drro->drr_bonuslen, tx);
2119 } else if (drro->drr_type != doi.doi_type ||
2120 drro->drr_blksz != doi.doi_data_block_size ||
2121 drro->drr_bonustype != doi.doi_bonus_type ||
2122 drro->drr_bonuslen != doi.doi_bonus_size) {
2123 /* currently allocated, but with different properties */
2124 err = dmu_object_reclaim(rwa->os, drro->drr_object,
2125 drro->drr_type, drro->drr_blksz,
2126 drro->drr_bonustype, drro->drr_bonuslen, tx);
2127 }
2128 if (err != 0) {
2129 dmu_tx_commit(tx);
2130 return (SET_ERROR(EINVAL));
2131 }
2132
2133 dmu_object_set_checksum(rwa->os, drro->drr_object,
2134 drro->drr_checksumtype, tx);
2135 dmu_object_set_compress(rwa->os, drro->drr_object,
2136 drro->drr_compress, tx);
2137
2138 if (data != NULL) {
2139 dmu_buf_t *db;
2140
2141 VERIFY0(dmu_bonus_hold(rwa->os, drro->drr_object, FTAG, &db));
2142 dmu_buf_will_dirty(db, tx);
2143
2144 ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
2145 bcopy(data, db->db_data, drro->drr_bonuslen);
2146 if (rwa->byteswap) {
2147 dmu_object_byteswap_t byteswap =
2148 DMU_OT_BYTESWAP(drro->drr_bonustype);
2149 dmu_ot_byteswap[byteswap].ob_func(db->db_data,
2150 drro->drr_bonuslen);
2151 }
2152 dmu_buf_rele(db, FTAG);
2153 }
2154 dmu_tx_commit(tx);
2155
2156 return (0);
2157 }
2158
2159 /* ARGSUSED */
2160 static int
receive_freeobjects(struct receive_writer_arg * rwa,struct drr_freeobjects * drrfo)2161 receive_freeobjects(struct receive_writer_arg *rwa,
2162 struct drr_freeobjects *drrfo)
2163 {
2164 uint64_t obj;
2165 int next_err = 0;
2166
2167 if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
2168 return (SET_ERROR(EINVAL));
2169
2170 for (obj = drrfo->drr_firstobj;
2171 obj < drrfo->drr_firstobj + drrfo->drr_numobjs && next_err == 0;
2172 next_err = dmu_object_next(rwa->os, &obj, FALSE, 0)) {
2173 int err;
2174
2175 if (dmu_object_info(rwa->os, obj, NULL) != 0)
2176 continue;
2177
2178 err = dmu_free_long_object(rwa->os, obj);
2179 if (err != 0)
2180 return (err);
2181 }
2182 if (next_err != ESRCH)
2183 return (next_err);
2184 return (0);
2185 }
2186
2187 static int
receive_write(struct receive_writer_arg * rwa,struct drr_write * drrw,arc_buf_t * abuf)2188 receive_write(struct receive_writer_arg *rwa, struct drr_write *drrw,
2189 arc_buf_t *abuf)
2190 {
2191 dmu_tx_t *tx;
2192 int err;
2193
2194 if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
2195 !DMU_OT_IS_VALID(drrw->drr_type))
2196 return (SET_ERROR(EINVAL));
2197
2198 /*
2199 * For resuming to work, records must be in increasing order
2200 * by (object, offset).
2201 */
2202 if (drrw->drr_object < rwa->last_object ||
2203 (drrw->drr_object == rwa->last_object &&
2204 drrw->drr_offset < rwa->last_offset)) {
2205 return (SET_ERROR(EINVAL));
2206 }
2207 rwa->last_object = drrw->drr_object;
2208 rwa->last_offset = drrw->drr_offset;
2209
2210 if (dmu_object_info(rwa->os, drrw->drr_object, NULL) != 0)
2211 return (SET_ERROR(EINVAL));
2212
2213 tx = dmu_tx_create(rwa->os);
2214
2215 dmu_tx_hold_write(tx, drrw->drr_object,
2216 drrw->drr_offset, drrw->drr_length);
2217 err = dmu_tx_assign(tx, TXG_WAIT);
2218 if (err != 0) {
2219 dmu_tx_abort(tx);
2220 return (err);
2221 }
2222 if (rwa->byteswap) {
2223 dmu_object_byteswap_t byteswap =
2224 DMU_OT_BYTESWAP(drrw->drr_type);
2225 dmu_ot_byteswap[byteswap].ob_func(abuf->b_data,
2226 drrw->drr_length);
2227 }
2228
2229 dmu_buf_t *bonus;
2230 if (dmu_bonus_hold(rwa->os, drrw->drr_object, FTAG, &bonus) != 0)
2231 return (SET_ERROR(EINVAL));
2232 dmu_assign_arcbuf(bonus, drrw->drr_offset, abuf, tx);
2233
2234 /*
2235 * Note: If the receive fails, we want the resume stream to start
2236 * with the same record that we last successfully received (as opposed
2237 * to the next record), so that we can verify that we are
2238 * resuming from the correct location.
2239 */
2240 save_resume_state(rwa, drrw->drr_object, drrw->drr_offset, tx);
2241 dmu_tx_commit(tx);
2242 dmu_buf_rele(bonus, FTAG);
2243
2244 return (0);
2245 }
2246
2247 /*
2248 * Handle a DRR_WRITE_BYREF record. This record is used in dedup'ed
2249 * streams to refer to a copy of the data that is already on the
2250 * system because it came in earlier in the stream. This function
2251 * finds the earlier copy of the data, and uses that copy instead of
2252 * data from the stream to fulfill this write.
2253 */
2254 static int
receive_write_byref(struct receive_writer_arg * rwa,struct drr_write_byref * drrwbr)2255 receive_write_byref(struct receive_writer_arg *rwa,
2256 struct drr_write_byref *drrwbr)
2257 {
2258 dmu_tx_t *tx;
2259 int err;
2260 guid_map_entry_t gmesrch;
2261 guid_map_entry_t *gmep;
2262 avl_index_t where;
2263 objset_t *ref_os = NULL;
2264 dmu_buf_t *dbp;
2265
2266 if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
2267 return (SET_ERROR(EINVAL));
2268
2269 /*
2270 * If the GUID of the referenced dataset is different from the
2271 * GUID of the target dataset, find the referenced dataset.
2272 */
2273 if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
2274 gmesrch.guid = drrwbr->drr_refguid;
2275 if ((gmep = avl_find(rwa->guid_to_ds_map, &gmesrch,
2276 &where)) == NULL) {
2277 return (SET_ERROR(EINVAL));
2278 }
2279 if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
2280 return (SET_ERROR(EINVAL));
2281 } else {
2282 ref_os = rwa->os;
2283 }
2284
2285 err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
2286 drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH);
2287 if (err != 0)
2288 return (err);
2289
2290 tx = dmu_tx_create(rwa->os);
2291
2292 dmu_tx_hold_write(tx, drrwbr->drr_object,
2293 drrwbr->drr_offset, drrwbr->drr_length);
2294 err = dmu_tx_assign(tx, TXG_WAIT);
2295 if (err != 0) {
2296 dmu_tx_abort(tx);
2297 return (err);
2298 }
2299 dmu_write(rwa->os, drrwbr->drr_object,
2300 drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
2301 dmu_buf_rele(dbp, FTAG);
2302
2303 /* See comment in restore_write. */
2304 save_resume_state(rwa, drrwbr->drr_object, drrwbr->drr_offset, tx);
2305 dmu_tx_commit(tx);
2306 return (0);
2307 }
2308
2309 static int
receive_write_embedded(struct receive_writer_arg * rwa,struct drr_write_embedded * drrwe,void * data)2310 receive_write_embedded(struct receive_writer_arg *rwa,
2311 struct drr_write_embedded *drrwe, void *data)
2312 {
2313 dmu_tx_t *tx;
2314 int err;
2315
2316 if (drrwe->drr_offset + drrwe->drr_length < drrwe->drr_offset)
2317 return (EINVAL);
2318
2319 if (drrwe->drr_psize > BPE_PAYLOAD_SIZE)
2320 return (EINVAL);
2321
2322 if (drrwe->drr_etype >= NUM_BP_EMBEDDED_TYPES)
2323 return (EINVAL);
2324 if (drrwe->drr_compression >= ZIO_COMPRESS_FUNCTIONS)
2325 return (EINVAL);
2326
2327 tx = dmu_tx_create(rwa->os);
2328
2329 dmu_tx_hold_write(tx, drrwe->drr_object,
2330 drrwe->drr_offset, drrwe->drr_length);
2331 err = dmu_tx_assign(tx, TXG_WAIT);
2332 if (err != 0) {
2333 dmu_tx_abort(tx);
2334 return (err);
2335 }
2336
2337 dmu_write_embedded(rwa->os, drrwe->drr_object,
2338 drrwe->drr_offset, data, drrwe->drr_etype,
2339 drrwe->drr_compression, drrwe->drr_lsize, drrwe->drr_psize,
2340 rwa->byteswap ^ ZFS_HOST_BYTEORDER, tx);
2341
2342 /* See comment in restore_write. */
2343 save_resume_state(rwa, drrwe->drr_object, drrwe->drr_offset, tx);
2344 dmu_tx_commit(tx);
2345 return (0);
2346 }
2347
2348 static int
receive_spill(struct receive_writer_arg * rwa,struct drr_spill * drrs,void * data)2349 receive_spill(struct receive_writer_arg *rwa, struct drr_spill *drrs,
2350 void *data)
2351 {
2352 dmu_tx_t *tx;
2353 dmu_buf_t *db, *db_spill;
2354 int err;
2355
2356 if (drrs->drr_length < SPA_MINBLOCKSIZE ||
2357 drrs->drr_length > spa_maxblocksize(dmu_objset_spa(rwa->os)))
2358 return (SET_ERROR(EINVAL));
2359
2360 if (dmu_object_info(rwa->os, drrs->drr_object, NULL) != 0)
2361 return (SET_ERROR(EINVAL));
2362
2363 VERIFY0(dmu_bonus_hold(rwa->os, drrs->drr_object, FTAG, &db));
2364 if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
2365 dmu_buf_rele(db, FTAG);
2366 return (err);
2367 }
2368
2369 tx = dmu_tx_create(rwa->os);
2370
2371 dmu_tx_hold_spill(tx, db->db_object);
2372
2373 err = dmu_tx_assign(tx, TXG_WAIT);
2374 if (err != 0) {
2375 dmu_buf_rele(db, FTAG);
2376 dmu_buf_rele(db_spill, FTAG);
2377 dmu_tx_abort(tx);
2378 return (err);
2379 }
2380 dmu_buf_will_dirty(db_spill, tx);
2381
2382 if (db_spill->db_size < drrs->drr_length)
2383 VERIFY(0 == dbuf_spill_set_blksz(db_spill,
2384 drrs->drr_length, tx));
2385 bcopy(data, db_spill->db_data, drrs->drr_length);
2386
2387 dmu_buf_rele(db, FTAG);
2388 dmu_buf_rele(db_spill, FTAG);
2389
2390 dmu_tx_commit(tx);
2391 return (0);
2392 }
2393
2394 /* ARGSUSED */
2395 static int
receive_free(struct receive_writer_arg * rwa,struct drr_free * drrf)2396 receive_free(struct receive_writer_arg *rwa, struct drr_free *drrf)
2397 {
2398 int err;
2399
2400 if (drrf->drr_length != -1ULL &&
2401 drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
2402 return (SET_ERROR(EINVAL));
2403
2404 if (dmu_object_info(rwa->os, drrf->drr_object, NULL) != 0)
2405 return (SET_ERROR(EINVAL));
2406
2407 err = dmu_free_long_range(rwa->os, drrf->drr_object,
2408 drrf->drr_offset, drrf->drr_length);
2409
2410 return (err);
2411 }
2412
2413 /* used to destroy the drc_ds on error */
2414 static void
dmu_recv_cleanup_ds(dmu_recv_cookie_t * drc)2415 dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc)
2416 {
2417 if (drc->drc_resumable) {
2418 /* wait for our resume state to be written to disk */
2419 txg_wait_synced(drc->drc_ds->ds_dir->dd_pool, 0);
2420 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2421 } else {
2422 char name[ZFS_MAX_DATASET_NAME_LEN];
2423 dsl_dataset_name(drc->drc_ds, name);
2424 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2425 (void) dsl_destroy_head(name);
2426 }
2427 }
2428
2429 static void
receive_cksum(struct receive_arg * ra,int len,void * buf)2430 receive_cksum(struct receive_arg *ra, int len, void *buf)
2431 {
2432 if (ra->byteswap) {
2433 fletcher_4_incremental_byteswap(buf, len, &ra->cksum);
2434 } else {
2435 fletcher_4_incremental_native(buf, len, &ra->cksum);
2436 }
2437 }
2438
2439 /*
2440 * Read the payload into a buffer of size len, and update the current record's
2441 * payload field.
2442 * Allocate ra->next_rrd and read the next record's header into
2443 * ra->next_rrd->header.
2444 * Verify checksum of payload and next record.
2445 */
2446 static int
receive_read_payload_and_next_header(struct receive_arg * ra,int len,void * buf)2447 receive_read_payload_and_next_header(struct receive_arg *ra, int len, void *buf)
2448 {
2449 int err;
2450
2451 if (len != 0) {
2452 ASSERT3U(len, <=, SPA_MAXBLOCKSIZE);
2453 err = receive_read(ra, len, buf);
2454 if (err != 0)
2455 return (err);
2456 receive_cksum(ra, len, buf);
2457
2458 /* note: rrd is NULL when reading the begin record's payload */
2459 if (ra->rrd != NULL) {
2460 ra->rrd->payload = buf;
2461 ra->rrd->payload_size = len;
2462 ra->rrd->bytes_read = ra->bytes_read;
2463 }
2464 }
2465
2466 ra->prev_cksum = ra->cksum;
2467
2468 ra->next_rrd = kmem_zalloc(sizeof (*ra->next_rrd), KM_SLEEP);
2469 err = receive_read(ra, sizeof (ra->next_rrd->header),
2470 &ra->next_rrd->header);
2471 ra->next_rrd->bytes_read = ra->bytes_read;
2472 if (err != 0) {
2473 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2474 ra->next_rrd = NULL;
2475 return (err);
2476 }
2477 if (ra->next_rrd->header.drr_type == DRR_BEGIN) {
2478 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2479 ra->next_rrd = NULL;
2480 return (SET_ERROR(EINVAL));
2481 }
2482
2483 /*
2484 * Note: checksum is of everything up to but not including the
2485 * checksum itself.
2486 */
2487 ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
2488 ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
2489 receive_cksum(ra,
2490 offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
2491 &ra->next_rrd->header);
2492
2493 zio_cksum_t cksum_orig =
2494 ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
2495 zio_cksum_t *cksump =
2496 &ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
2497
2498 if (ra->byteswap)
2499 byteswap_record(&ra->next_rrd->header);
2500
2501 if ((!ZIO_CHECKSUM_IS_ZERO(cksump)) &&
2502 !ZIO_CHECKSUM_EQUAL(ra->cksum, *cksump)) {
2503 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2504 ra->next_rrd = NULL;
2505 return (SET_ERROR(ECKSUM));
2506 }
2507
2508 receive_cksum(ra, sizeof (cksum_orig), &cksum_orig);
2509
2510 return (0);
2511 }
2512
2513 static void
objlist_create(struct objlist * list)2514 objlist_create(struct objlist *list)
2515 {
2516 list_create(&list->list, sizeof (struct receive_objnode),
2517 offsetof(struct receive_objnode, node));
2518 list->last_lookup = 0;
2519 }
2520
2521 static void
objlist_destroy(struct objlist * list)2522 objlist_destroy(struct objlist *list)
2523 {
2524 for (struct receive_objnode *n = list_remove_head(&list->list);
2525 n != NULL; n = list_remove_head(&list->list)) {
2526 kmem_free(n, sizeof (*n));
2527 }
2528 list_destroy(&list->list);
2529 }
2530
2531 /*
2532 * This function looks through the objlist to see if the specified object number
2533 * is contained in the objlist. In the process, it will remove all object
2534 * numbers in the list that are smaller than the specified object number. Thus,
2535 * any lookup of an object number smaller than a previously looked up object
2536 * number will always return false; therefore, all lookups should be done in
2537 * ascending order.
2538 */
2539 static boolean_t
objlist_exists(struct objlist * list,uint64_t object)2540 objlist_exists(struct objlist *list, uint64_t object)
2541 {
2542 struct receive_objnode *node = list_head(&list->list);
2543 ASSERT3U(object, >=, list->last_lookup);
2544 list->last_lookup = object;
2545 while (node != NULL && node->object < object) {
2546 VERIFY3P(node, ==, list_remove_head(&list->list));
2547 kmem_free(node, sizeof (*node));
2548 node = list_head(&list->list);
2549 }
2550 return (node != NULL && node->object == object);
2551 }
2552
2553 /*
2554 * The objlist is a list of object numbers stored in ascending order. However,
2555 * the insertion of new object numbers does not seek out the correct location to
2556 * store a new object number; instead, it appends it to the list for simplicity.
2557 * Thus, any users must take care to only insert new object numbers in ascending
2558 * order.
2559 */
2560 static void
objlist_insert(struct objlist * list,uint64_t object)2561 objlist_insert(struct objlist *list, uint64_t object)
2562 {
2563 struct receive_objnode *node = kmem_zalloc(sizeof (*node), KM_SLEEP);
2564 node->object = object;
2565 #ifdef ZFS_DEBUG
2566 struct receive_objnode *last_object = list_tail(&list->list);
2567 uint64_t last_objnum = (last_object != NULL ? last_object->object : 0);
2568 ASSERT3U(node->object, >, last_objnum);
2569 #endif
2570 list_insert_tail(&list->list, node);
2571 }
2572
2573 /*
2574 * Issue the prefetch reads for any necessary indirect blocks.
2575 *
2576 * We use the object ignore list to tell us whether or not to issue prefetches
2577 * for a given object. We do this for both correctness (in case the blocksize
2578 * of an object has changed) and performance (if the object doesn't exist, don't
2579 * needlessly try to issue prefetches). We also trim the list as we go through
2580 * the stream to prevent it from growing to an unbounded size.
2581 *
2582 * The object numbers within will always be in sorted order, and any write
2583 * records we see will also be in sorted order, but they're not sorted with
2584 * respect to each other (i.e. we can get several object records before
2585 * receiving each object's write records). As a result, once we've reached a
2586 * given object number, we can safely remove any reference to lower object
2587 * numbers in the ignore list. In practice, we receive up to 32 object records
2588 * before receiving write records, so the list can have up to 32 nodes in it.
2589 */
2590 /* ARGSUSED */
2591 static void
receive_read_prefetch(struct receive_arg * ra,uint64_t object,uint64_t offset,uint64_t length)2592 receive_read_prefetch(struct receive_arg *ra,
2593 uint64_t object, uint64_t offset, uint64_t length)
2594 {
2595 if (!objlist_exists(&ra->ignore_objlist, object)) {
2596 dmu_prefetch(ra->os, object, 1, offset, length,
2597 ZIO_PRIORITY_SYNC_READ);
2598 }
2599 }
2600
2601 /*
2602 * Read records off the stream, issuing any necessary prefetches.
2603 */
2604 static int
receive_read_record(struct receive_arg * ra)2605 receive_read_record(struct receive_arg *ra)
2606 {
2607 int err;
2608
2609 switch (ra->rrd->header.drr_type) {
2610 case DRR_OBJECT:
2611 {
2612 struct drr_object *drro = &ra->rrd->header.drr_u.drr_object;
2613 uint32_t size = P2ROUNDUP(drro->drr_bonuslen, 8);
2614 void *buf = kmem_zalloc(size, KM_SLEEP);
2615 dmu_object_info_t doi;
2616 err = receive_read_payload_and_next_header(ra, size, buf);
2617 if (err != 0) {
2618 kmem_free(buf, size);
2619 return (err);
2620 }
2621 err = dmu_object_info(ra->os, drro->drr_object, &doi);
2622 /*
2623 * See receive_read_prefetch for an explanation why we're
2624 * storing this object in the ignore_obj_list.
2625 */
2626 if (err == ENOENT ||
2627 (err == 0 && doi.doi_data_block_size != drro->drr_blksz)) {
2628 objlist_insert(&ra->ignore_objlist, drro->drr_object);
2629 err = 0;
2630 }
2631 return (err);
2632 }
2633 case DRR_FREEOBJECTS:
2634 {
2635 err = receive_read_payload_and_next_header(ra, 0, NULL);
2636 return (err);
2637 }
2638 case DRR_WRITE:
2639 {
2640 struct drr_write *drrw = &ra->rrd->header.drr_u.drr_write;
2641 arc_buf_t *abuf = arc_loan_buf(dmu_objset_spa(ra->os),
2642 drrw->drr_length);
2643
2644 err = receive_read_payload_and_next_header(ra,
2645 drrw->drr_length, abuf->b_data);
2646 if (err != 0) {
2647 dmu_return_arcbuf(abuf);
2648 return (err);
2649 }
2650 ra->rrd->write_buf = abuf;
2651 receive_read_prefetch(ra, drrw->drr_object, drrw->drr_offset,
2652 drrw->drr_length);
2653 return (err);
2654 }
2655 case DRR_WRITE_BYREF:
2656 {
2657 struct drr_write_byref *drrwb =
2658 &ra->rrd->header.drr_u.drr_write_byref;
2659 err = receive_read_payload_and_next_header(ra, 0, NULL);
2660 receive_read_prefetch(ra, drrwb->drr_object, drrwb->drr_offset,
2661 drrwb->drr_length);
2662 return (err);
2663 }
2664 case DRR_WRITE_EMBEDDED:
2665 {
2666 struct drr_write_embedded *drrwe =
2667 &ra->rrd->header.drr_u.drr_write_embedded;
2668 uint32_t size = P2ROUNDUP(drrwe->drr_psize, 8);
2669 void *buf = kmem_zalloc(size, KM_SLEEP);
2670
2671 err = receive_read_payload_and_next_header(ra, size, buf);
2672 if (err != 0) {
2673 kmem_free(buf, size);
2674 return (err);
2675 }
2676
2677 receive_read_prefetch(ra, drrwe->drr_object, drrwe->drr_offset,
2678 drrwe->drr_length);
2679 return (err);
2680 }
2681 case DRR_FREE:
2682 {
2683 /*
2684 * It might be beneficial to prefetch indirect blocks here, but
2685 * we don't really have the data to decide for sure.
2686 */
2687 err = receive_read_payload_and_next_header(ra, 0, NULL);
2688 return (err);
2689 }
2690 case DRR_END:
2691 {
2692 struct drr_end *drre = &ra->rrd->header.drr_u.drr_end;
2693 if (!ZIO_CHECKSUM_EQUAL(ra->prev_cksum, drre->drr_checksum))
2694 return (SET_ERROR(ECKSUM));
2695 return (0);
2696 }
2697 case DRR_SPILL:
2698 {
2699 struct drr_spill *drrs = &ra->rrd->header.drr_u.drr_spill;
2700 void *buf = kmem_zalloc(drrs->drr_length, KM_SLEEP);
2701 err = receive_read_payload_and_next_header(ra, drrs->drr_length,
2702 buf);
2703 if (err != 0)
2704 kmem_free(buf, drrs->drr_length);
2705 return (err);
2706 }
2707 default:
2708 return (SET_ERROR(EINVAL));
2709 }
2710 }
2711
2712 /*
2713 * Commit the records to the pool.
2714 */
2715 static int
receive_process_record(struct receive_writer_arg * rwa,struct receive_record_arg * rrd)2716 receive_process_record(struct receive_writer_arg *rwa,
2717 struct receive_record_arg *rrd)
2718 {
2719 int err;
2720
2721 /* Processing in order, therefore bytes_read should be increasing. */
2722 ASSERT3U(rrd->bytes_read, >=, rwa->bytes_read);
2723 rwa->bytes_read = rrd->bytes_read;
2724
2725 switch (rrd->header.drr_type) {
2726 case DRR_OBJECT:
2727 {
2728 struct drr_object *drro = &rrd->header.drr_u.drr_object;
2729 err = receive_object(rwa, drro, rrd->payload);
2730 kmem_free(rrd->payload, rrd->payload_size);
2731 rrd->payload = NULL;
2732 return (err);
2733 }
2734 case DRR_FREEOBJECTS:
2735 {
2736 struct drr_freeobjects *drrfo =
2737 &rrd->header.drr_u.drr_freeobjects;
2738 return (receive_freeobjects(rwa, drrfo));
2739 }
2740 case DRR_WRITE:
2741 {
2742 struct drr_write *drrw = &rrd->header.drr_u.drr_write;
2743 err = receive_write(rwa, drrw, rrd->write_buf);
2744 /* if receive_write() is successful, it consumes the arc_buf */
2745 if (err != 0)
2746 dmu_return_arcbuf(rrd->write_buf);
2747 rrd->write_buf = NULL;
2748 rrd->payload = NULL;
2749 return (err);
2750 }
2751 case DRR_WRITE_BYREF:
2752 {
2753 struct drr_write_byref *drrwbr =
2754 &rrd->header.drr_u.drr_write_byref;
2755 return (receive_write_byref(rwa, drrwbr));
2756 }
2757 case DRR_WRITE_EMBEDDED:
2758 {
2759 struct drr_write_embedded *drrwe =
2760 &rrd->header.drr_u.drr_write_embedded;
2761 err = receive_write_embedded(rwa, drrwe, rrd->payload);
2762 kmem_free(rrd->payload, rrd->payload_size);
2763 rrd->payload = NULL;
2764 return (err);
2765 }
2766 case DRR_FREE:
2767 {
2768 struct drr_free *drrf = &rrd->header.drr_u.drr_free;
2769 return (receive_free(rwa, drrf));
2770 }
2771 case DRR_SPILL:
2772 {
2773 struct drr_spill *drrs = &rrd->header.drr_u.drr_spill;
2774 err = receive_spill(rwa, drrs, rrd->payload);
2775 kmem_free(rrd->payload, rrd->payload_size);
2776 rrd->payload = NULL;
2777 return (err);
2778 }
2779 default:
2780 return (SET_ERROR(EINVAL));
2781 }
2782 }
2783
2784 /*
2785 * dmu_recv_stream's worker thread; pull records off the queue, and then call
2786 * receive_process_record When we're done, signal the main thread and exit.
2787 */
2788 static void
receive_writer_thread(void * arg)2789 receive_writer_thread(void *arg)
2790 {
2791 struct receive_writer_arg *rwa = arg;
2792 struct receive_record_arg *rrd;
2793 for (rrd = bqueue_dequeue(&rwa->q); !rrd->eos_marker;
2794 rrd = bqueue_dequeue(&rwa->q)) {
2795 /*
2796 * If there's an error, the main thread will stop putting things
2797 * on the queue, but we need to clear everything in it before we
2798 * can exit.
2799 */
2800 if (rwa->err == 0) {
2801 rwa->err = receive_process_record(rwa, rrd);
2802 } else if (rrd->write_buf != NULL) {
2803 dmu_return_arcbuf(rrd->write_buf);
2804 rrd->write_buf = NULL;
2805 rrd->payload = NULL;
2806 } else if (rrd->payload != NULL) {
2807 kmem_free(rrd->payload, rrd->payload_size);
2808 rrd->payload = NULL;
2809 }
2810 kmem_free(rrd, sizeof (*rrd));
2811 }
2812 kmem_free(rrd, sizeof (*rrd));
2813 mutex_enter(&rwa->mutex);
2814 rwa->done = B_TRUE;
2815 cv_signal(&rwa->cv);
2816 mutex_exit(&rwa->mutex);
2817 thread_exit();
2818 }
2819
2820 static int
resume_check(struct receive_arg * ra,nvlist_t * begin_nvl)2821 resume_check(struct receive_arg *ra, nvlist_t *begin_nvl)
2822 {
2823 uint64_t val;
2824 objset_t *mos = dmu_objset_pool(ra->os)->dp_meta_objset;
2825 uint64_t dsobj = dmu_objset_id(ra->os);
2826 uint64_t resume_obj, resume_off;
2827
2828 if (nvlist_lookup_uint64(begin_nvl,
2829 "resume_object", &resume_obj) != 0 ||
2830 nvlist_lookup_uint64(begin_nvl,
2831 "resume_offset", &resume_off) != 0) {
2832 return (SET_ERROR(EINVAL));
2833 }
2834 VERIFY0(zap_lookup(mos, dsobj,
2835 DS_FIELD_RESUME_OBJECT, sizeof (val), 1, &val));
2836 if (resume_obj != val)
2837 return (SET_ERROR(EINVAL));
2838 VERIFY0(zap_lookup(mos, dsobj,
2839 DS_FIELD_RESUME_OFFSET, sizeof (val), 1, &val));
2840 if (resume_off != val)
2841 return (SET_ERROR(EINVAL));
2842
2843 return (0);
2844 }
2845
2846 /*
2847 * Read in the stream's records, one by one, and apply them to the pool. There
2848 * are two threads involved; the thread that calls this function will spin up a
2849 * worker thread, read the records off the stream one by one, and issue
2850 * prefetches for any necessary indirect blocks. It will then push the records
2851 * onto an internal blocking queue. The worker thread will pull the records off
2852 * the queue, and actually write the data into the DMU. This way, the worker
2853 * thread doesn't have to wait for reads to complete, since everything it needs
2854 * (the indirect blocks) will be prefetched.
2855 *
2856 * NB: callers *must* call dmu_recv_end() if this succeeds.
2857 */
2858 int
dmu_recv_stream(dmu_recv_cookie_t * drc,struct file * fp,offset_t * voffp,int cleanup_fd,uint64_t * action_handlep)2859 dmu_recv_stream(dmu_recv_cookie_t *drc, struct file *fp, offset_t *voffp,
2860 int cleanup_fd, uint64_t *action_handlep)
2861 {
2862 int err = 0;
2863 struct receive_arg ra = { 0 };
2864 struct receive_writer_arg rwa = { 0 };
2865 int featureflags;
2866 nvlist_t *begin_nvl = NULL;
2867
2868 ra.byteswap = drc->drc_byteswap;
2869 ra.cksum = drc->drc_cksum;
2870 ra.td = curthread;
2871 ra.fp = fp;
2872 ra.voff = *voffp;
2873
2874 if (dsl_dataset_is_zapified(drc->drc_ds)) {
2875 (void) zap_lookup(drc->drc_ds->ds_dir->dd_pool->dp_meta_objset,
2876 drc->drc_ds->ds_object, DS_FIELD_RESUME_BYTES,
2877 sizeof (ra.bytes_read), 1, &ra.bytes_read);
2878 }
2879
2880 objlist_create(&ra.ignore_objlist);
2881
2882 /* these were verified in dmu_recv_begin */
2883 ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==,
2884 DMU_SUBSTREAM);
2885 ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES);
2886
2887 /*
2888 * Open the objset we are modifying.
2889 */
2890 VERIFY0(dmu_objset_from_ds(drc->drc_ds, &ra.os));
2891
2892 ASSERT(dsl_dataset_phys(drc->drc_ds)->ds_flags & DS_FLAG_INCONSISTENT);
2893
2894 featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
2895
2896 /* if this stream is dedup'ed, set up the avl tree for guid mapping */
2897 if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
2898 minor_t minor;
2899
2900 if (cleanup_fd == -1) {
2901 ra.err = SET_ERROR(EBADF);
2902 goto out;
2903 }
2904 ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
2905 if (ra.err != 0) {
2906 cleanup_fd = -1;
2907 goto out;
2908 }
2909
2910 if (*action_handlep == 0) {
2911 rwa.guid_to_ds_map =
2912 kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
2913 avl_create(rwa.guid_to_ds_map, guid_compare,
2914 sizeof (guid_map_entry_t),
2915 offsetof(guid_map_entry_t, avlnode));
2916 err = zfs_onexit_add_cb(minor,
2917 free_guid_map_onexit, rwa.guid_to_ds_map,
2918 action_handlep);
2919 if (ra.err != 0)
2920 goto out;
2921 } else {
2922 err = zfs_onexit_cb_data(minor, *action_handlep,
2923 (void **)&rwa.guid_to_ds_map);
2924 if (ra.err != 0)
2925 goto out;
2926 }
2927
2928 drc->drc_guid_to_ds_map = rwa.guid_to_ds_map;
2929 }
2930
2931 uint32_t payloadlen = drc->drc_drr_begin->drr_payloadlen;
2932 void *payload = NULL;
2933 if (payloadlen != 0)
2934 payload = kmem_alloc(payloadlen, KM_SLEEP);
2935
2936 err = receive_read_payload_and_next_header(&ra, payloadlen, payload);
2937 if (err != 0) {
2938 if (payloadlen != 0)
2939 kmem_free(payload, payloadlen);
2940 goto out;
2941 }
2942 if (payloadlen != 0) {
2943 err = nvlist_unpack(payload, payloadlen, &begin_nvl, KM_SLEEP);
2944 kmem_free(payload, payloadlen);
2945 if (err != 0)
2946 goto out;
2947 }
2948
2949 if (featureflags & DMU_BACKUP_FEATURE_RESUMING) {
2950 err = resume_check(&ra, begin_nvl);
2951 if (err != 0)
2952 goto out;
2953 }
2954
2955 (void) bqueue_init(&rwa.q, zfs_recv_queue_length,
2956 offsetof(struct receive_record_arg, node));
2957 cv_init(&rwa.cv, NULL, CV_DEFAULT, NULL);
2958 mutex_init(&rwa.mutex, NULL, MUTEX_DEFAULT, NULL);
2959 rwa.os = ra.os;
2960 rwa.byteswap = drc->drc_byteswap;
2961 rwa.resumable = drc->drc_resumable;
2962
2963 (void) thread_create(NULL, 0, receive_writer_thread, &rwa, 0, &p0,
2964 TS_RUN, minclsyspri);
2965 /*
2966 * We're reading rwa.err without locks, which is safe since we are the
2967 * only reader, and the worker thread is the only writer. It's ok if we
2968 * miss a write for an iteration or two of the loop, since the writer
2969 * thread will keep freeing records we send it until we send it an eos
2970 * marker.
2971 *
2972 * We can leave this loop in 3 ways: First, if rwa.err is
2973 * non-zero. In that case, the writer thread will free the rrd we just
2974 * pushed. Second, if we're interrupted; in that case, either it's the
2975 * first loop and ra.rrd was never allocated, or it's later, and ra.rrd
2976 * has been handed off to the writer thread who will free it. Finally,
2977 * if receive_read_record fails or we're at the end of the stream, then
2978 * we free ra.rrd and exit.
2979 */
2980 while (rwa.err == 0) {
2981 if (issig(JUSTLOOKING) && issig(FORREAL)) {
2982 err = SET_ERROR(EINTR);
2983 break;
2984 }
2985
2986 ASSERT3P(ra.rrd, ==, NULL);
2987 ra.rrd = ra.next_rrd;
2988 ra.next_rrd = NULL;
2989 /* Allocates and loads header into ra.next_rrd */
2990 err = receive_read_record(&ra);
2991
2992 if (ra.rrd->header.drr_type == DRR_END || err != 0) {
2993 kmem_free(ra.rrd, sizeof (*ra.rrd));
2994 ra.rrd = NULL;
2995 break;
2996 }
2997
2998 bqueue_enqueue(&rwa.q, ra.rrd,
2999 sizeof (struct receive_record_arg) + ra.rrd->payload_size);
3000 ra.rrd = NULL;
3001 }
3002 if (ra.next_rrd == NULL)
3003 ra.next_rrd = kmem_zalloc(sizeof (*ra.next_rrd), KM_SLEEP);
3004 ra.next_rrd->eos_marker = B_TRUE;
3005 bqueue_enqueue(&rwa.q, ra.next_rrd, 1);
3006
3007 mutex_enter(&rwa.mutex);
3008 while (!rwa.done) {
3009 cv_wait(&rwa.cv, &rwa.mutex);
3010 }
3011 mutex_exit(&rwa.mutex);
3012
3013 cv_destroy(&rwa.cv);
3014 mutex_destroy(&rwa.mutex);
3015 bqueue_destroy(&rwa.q);
3016 if (err == 0)
3017 err = rwa.err;
3018
3019 out:
3020 nvlist_free(begin_nvl);
3021 if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
3022 zfs_onexit_fd_rele(cleanup_fd);
3023
3024 if (err != 0) {
3025 /*
3026 * Clean up references. If receive is not resumable,
3027 * destroy what we created, so we don't leave it in
3028 * the inconsistent state.
3029 */
3030 dmu_recv_cleanup_ds(drc);
3031 }
3032
3033 *voffp = ra.voff;
3034 objlist_destroy(&ra.ignore_objlist);
3035 return (err);
3036 }
3037
3038 static int
dmu_recv_end_check(void * arg,dmu_tx_t * tx)3039 dmu_recv_end_check(void *arg, dmu_tx_t *tx)
3040 {
3041 dmu_recv_cookie_t *drc = arg;
3042 dsl_pool_t *dp = dmu_tx_pool(tx);
3043 int error;
3044
3045 ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag);
3046
3047 if (!drc->drc_newfs) {
3048 dsl_dataset_t *origin_head;
3049
3050 error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head);
3051 if (error != 0)
3052 return (error);
3053 if (drc->drc_force) {
3054 /*
3055 * We will destroy any snapshots in tofs (i.e. before
3056 * origin_head) that are after the origin (which is
3057 * the snap before drc_ds, because drc_ds can not
3058 * have any snaps of its own).
3059 */
3060 uint64_t obj;
3061
3062 obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3063 while (obj !=
3064 dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
3065 dsl_dataset_t *snap;
3066 error = dsl_dataset_hold_obj(dp, obj, FTAG,
3067 &snap);
3068 if (error != 0)
3069 break;
3070 if (snap->ds_dir != origin_head->ds_dir)
3071 error = SET_ERROR(EINVAL);
3072 if (error == 0) {
3073 error = dsl_destroy_snapshot_check_impl(
3074 snap, B_FALSE);
3075 }
3076 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
3077 dsl_dataset_rele(snap, FTAG);
3078 if (error != 0)
3079 break;
3080 }
3081 if (error != 0) {
3082 dsl_dataset_rele(origin_head, FTAG);
3083 return (error);
3084 }
3085 }
3086 error = dsl_dataset_clone_swap_check_impl(drc->drc_ds,
3087 origin_head, drc->drc_force, drc->drc_owner, tx);
3088 if (error != 0) {
3089 dsl_dataset_rele(origin_head, FTAG);
3090 return (error);
3091 }
3092 error = dsl_dataset_snapshot_check_impl(origin_head,
3093 drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
3094 dsl_dataset_rele(origin_head, FTAG);
3095 if (error != 0)
3096 return (error);
3097
3098 error = dsl_destroy_head_check_impl(drc->drc_ds, 1);
3099 } else {
3100 error = dsl_dataset_snapshot_check_impl(drc->drc_ds,
3101 drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
3102 }
3103 return (error);
3104 }
3105
3106 static void
dmu_recv_end_sync(void * arg,dmu_tx_t * tx)3107 dmu_recv_end_sync(void *arg, dmu_tx_t *tx)
3108 {
3109 dmu_recv_cookie_t *drc = arg;
3110 dsl_pool_t *dp = dmu_tx_pool(tx);
3111
3112 spa_history_log_internal_ds(drc->drc_ds, "finish receiving",
3113 tx, "snap=%s", drc->drc_tosnap);
3114
3115 if (!drc->drc_newfs) {
3116 dsl_dataset_t *origin_head;
3117
3118 VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG,
3119 &origin_head));
3120
3121 if (drc->drc_force) {
3122 /*
3123 * Destroy any snapshots of drc_tofs (origin_head)
3124 * after the origin (the snap before drc_ds).
3125 */
3126 uint64_t obj;
3127
3128 obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3129 while (obj !=
3130 dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
3131 dsl_dataset_t *snap;
3132 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG,
3133 &snap));
3134 ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir);
3135 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
3136 dsl_destroy_snapshot_sync_impl(snap,
3137 B_FALSE, tx);
3138 dsl_dataset_rele(snap, FTAG);
3139 }
3140 }
3141 VERIFY3P(drc->drc_ds->ds_prev, ==,
3142 origin_head->ds_prev);
3143
3144 dsl_dataset_clone_swap_sync_impl(drc->drc_ds,
3145 origin_head, tx);
3146 dsl_dataset_snapshot_sync_impl(origin_head,
3147 drc->drc_tosnap, tx);
3148
3149 /* set snapshot's creation time and guid */
3150 dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx);
3151 dsl_dataset_phys(origin_head->ds_prev)->ds_creation_time =
3152 drc->drc_drrb->drr_creation_time;
3153 dsl_dataset_phys(origin_head->ds_prev)->ds_guid =
3154 drc->drc_drrb->drr_toguid;
3155 dsl_dataset_phys(origin_head->ds_prev)->ds_flags &=
3156 ~DS_FLAG_INCONSISTENT;
3157
3158 dmu_buf_will_dirty(origin_head->ds_dbuf, tx);
3159 dsl_dataset_phys(origin_head)->ds_flags &=
3160 ~DS_FLAG_INCONSISTENT;
3161
3162 drc->drc_newsnapobj =
3163 dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3164
3165 dsl_dataset_rele(origin_head, FTAG);
3166 dsl_destroy_head_sync_impl(drc->drc_ds, tx);
3167
3168 if (drc->drc_owner != NULL)
3169 VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner);
3170 } else {
3171 dsl_dataset_t *ds = drc->drc_ds;
3172
3173 dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx);
3174
3175 /* set snapshot's creation time and guid */
3176 dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
3177 dsl_dataset_phys(ds->ds_prev)->ds_creation_time =
3178 drc->drc_drrb->drr_creation_time;
3179 dsl_dataset_phys(ds->ds_prev)->ds_guid =
3180 drc->drc_drrb->drr_toguid;
3181 dsl_dataset_phys(ds->ds_prev)->ds_flags &=
3182 ~DS_FLAG_INCONSISTENT;
3183
3184 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3185 dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
3186 if (dsl_dataset_has_resume_receive_state(ds)) {
3187 (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3188 DS_FIELD_RESUME_FROMGUID, tx);
3189 (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3190 DS_FIELD_RESUME_OBJECT, tx);
3191 (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3192 DS_FIELD_RESUME_OFFSET, tx);
3193 (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3194 DS_FIELD_RESUME_BYTES, tx);
3195 (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3196 DS_FIELD_RESUME_TOGUID, tx);
3197 (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3198 DS_FIELD_RESUME_TONAME, tx);
3199 }
3200 drc->drc_newsnapobj =
3201 dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj;
3202 }
3203 /*
3204 * Release the hold from dmu_recv_begin. This must be done before
3205 * we return to open context, so that when we free the dataset's dnode,
3206 * we can evict its bonus buffer.
3207 */
3208 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
3209 drc->drc_ds = NULL;
3210 }
3211
3212 static int
add_ds_to_guidmap(const char * name,avl_tree_t * guid_map,uint64_t snapobj)3213 add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj)
3214 {
3215 dsl_pool_t *dp;
3216 dsl_dataset_t *snapds;
3217 guid_map_entry_t *gmep;
3218 int err;
3219
3220 ASSERT(guid_map != NULL);
3221
3222 err = dsl_pool_hold(name, FTAG, &dp);
3223 if (err != 0)
3224 return (err);
3225 gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP);
3226 err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds);
3227 if (err == 0) {
3228 gmep->guid = dsl_dataset_phys(snapds)->ds_guid;
3229 gmep->gme_ds = snapds;
3230 avl_add(guid_map, gmep);
3231 dsl_dataset_long_hold(snapds, gmep);
3232 } else
3233 kmem_free(gmep, sizeof (*gmep));
3234
3235 dsl_pool_rele(dp, FTAG);
3236 return (err);
3237 }
3238
3239 static int dmu_recv_end_modified_blocks = 3;
3240
3241 static int
dmu_recv_existing_end(dmu_recv_cookie_t * drc)3242 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
3243 {
3244 #ifdef _KERNEL
3245 /*
3246 * We will be destroying the ds; make sure its origin is unmounted if
3247 * necessary.
3248 */
3249 char name[ZFS_MAX_DATASET_NAME_LEN];
3250 dsl_dataset_name(drc->drc_ds, name);
3251 zfs_destroy_unmount_origin(name);
3252 #endif
3253
3254 return (dsl_sync_task(drc->drc_tofs,
3255 dmu_recv_end_check, dmu_recv_end_sync, drc,
3256 dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL));
3257 }
3258
3259 static int
dmu_recv_new_end(dmu_recv_cookie_t * drc)3260 dmu_recv_new_end(dmu_recv_cookie_t *drc)
3261 {
3262 return (dsl_sync_task(drc->drc_tofs,
3263 dmu_recv_end_check, dmu_recv_end_sync, drc,
3264 dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL));
3265 }
3266
3267 int
dmu_recv_end(dmu_recv_cookie_t * drc,void * owner)3268 dmu_recv_end(dmu_recv_cookie_t *drc, void *owner)
3269 {
3270 int error;
3271
3272 drc->drc_owner = owner;
3273
3274 if (drc->drc_newfs)
3275 error = dmu_recv_new_end(drc);
3276 else
3277 error = dmu_recv_existing_end(drc);
3278
3279 if (error != 0) {
3280 dmu_recv_cleanup_ds(drc);
3281 } else if (drc->drc_guid_to_ds_map != NULL) {
3282 (void) add_ds_to_guidmap(drc->drc_tofs,
3283 drc->drc_guid_to_ds_map,
3284 drc->drc_newsnapobj);
3285 }
3286 return (error);
3287 }
3288
3289 /*
3290 * Return TRUE if this objset is currently being received into.
3291 */
3292 boolean_t
dmu_objset_is_receiving(objset_t * os)3293 dmu_objset_is_receiving(objset_t *os)
3294 {
3295 return (os->os_dsl_dataset != NULL &&
3296 os->os_dsl_dataset->ds_owner == dmu_recv_tag);
3297 }
3298