1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or https://opensource.org/licenses/CDDL-1.0.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
25  * Copyright (c) 2015 by Chunwei Chen. All rights reserved.
26  * Copyright 2017 Nexenta Systems, Inc.
27  * Copyright (c) 2021, 2022 by Pawel Jakub Dawidek
28  */
29 
30 /* Portions Copyright 2007 Jeremy Teo */
31 /* Portions Copyright 2010 Robert Milkowski */
32 
33 #include <sys/types.h>
34 #include <sys/param.h>
35 #include <sys/time.h>
36 #include <sys/sysmacros.h>
37 #include <sys/vfs.h>
38 #include <sys/uio_impl.h>
39 #include <sys/file.h>
40 #include <sys/stat.h>
41 #include <sys/kmem.h>
42 #include <sys/cmn_err.h>
43 #include <sys/errno.h>
44 #include <sys/zfs_dir.h>
45 #include <sys/zfs_acl.h>
46 #include <sys/zfs_ioctl.h>
47 #include <sys/fs/zfs.h>
48 #include <sys/dmu.h>
49 #include <sys/dmu_objset.h>
50 #include <sys/dsl_crypt.h>
51 #include <sys/spa.h>
52 #include <sys/txg.h>
53 #include <sys/dbuf.h>
54 #include <sys/policy.h>
55 #include <sys/zfeature.h>
56 #include <sys/zfs_vnops.h>
57 #include <sys/zfs_quota.h>
58 #include <sys/zfs_vfsops.h>
59 #include <sys/zfs_znode.h>
60 
61 /*
62  * Enable the experimental block cloning feature.  If this setting is 0, then
63  * even if feature@block_cloning is enabled, attempts to clone blocks will act
64  * as though the feature is disabled.
65  */
66 int zfs_bclone_enabled = 0;
67 
68 /*
69  * When set zfs_clone_range() waits for dirty data to be written to disk.
70  * This allows the clone operation to reliably succeed when a file is modified
71  * and then immediately cloned. For small files this may be slower than making
72  * a copy of the file and is therefore not the default.  However, in certain
73  * scenarios this behavior may be desirable so a tunable is provided.
74  */
75 static int zfs_bclone_wait_dirty = 0;
76 
77 /*
78  * Maximum bytes to read per chunk in zfs_read().
79  */
80 static uint64_t zfs_vnops_read_chunk_size = 1024 * 1024;
81 
82 static ulong_t zfs_fsync_sync_cnt = 4;
83 
84 int
zfs_fsync(znode_t * zp,int syncflag,cred_t * cr)85 zfs_fsync(znode_t *zp, int syncflag, cred_t *cr)
86 {
87 	int error = 0;
88 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
89 
90 	(void) tsd_set(zfs_fsyncer_key, (void *)(uintptr_t)zfs_fsync_sync_cnt);
91 
92 	if (zfsvfs->z_os->os_sync != ZFS_SYNC_DISABLED) {
93 		if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0)
94 			goto out;
95 		atomic_inc_32(&zp->z_sync_writes_cnt);
96 		zil_commit(zfsvfs->z_log, zp->z_id);
97 		atomic_dec_32(&zp->z_sync_writes_cnt);
98 		zfs_exit(zfsvfs, FTAG);
99 	}
100 out:
101 	tsd_set(zfs_fsyncer_key, NULL);
102 
103 	return (error);
104 }
105 
106 
107 #if defined(SEEK_HOLE) && defined(SEEK_DATA)
108 /*
109  * Lseek support for finding holes (cmd == SEEK_HOLE) and
110  * data (cmd == SEEK_DATA). "off" is an in/out parameter.
111  */
112 static int
zfs_holey_common(znode_t * zp,ulong_t cmd,loff_t * off)113 zfs_holey_common(znode_t *zp, ulong_t cmd, loff_t *off)
114 {
115 	zfs_locked_range_t *lr;
116 	uint64_t noff = (uint64_t)*off; /* new offset */
117 	uint64_t file_sz;
118 	int error;
119 	boolean_t hole;
120 
121 	file_sz = zp->z_size;
122 	if (noff >= file_sz)  {
123 		return (SET_ERROR(ENXIO));
124 	}
125 
126 	if (cmd == F_SEEK_HOLE)
127 		hole = B_TRUE;
128 	else
129 		hole = B_FALSE;
130 
131 	/* Flush any mmap()'d data to disk */
132 	if (zn_has_cached_data(zp, 0, file_sz - 1))
133 		zn_flush_cached_data(zp, B_TRUE);
134 
135 	lr = zfs_rangelock_enter(&zp->z_rangelock, 0, UINT64_MAX, RL_READER);
136 	error = dmu_offset_next(ZTOZSB(zp)->z_os, zp->z_id, hole, &noff);
137 	zfs_rangelock_exit(lr);
138 
139 	if (error == ESRCH)
140 		return (SET_ERROR(ENXIO));
141 
142 	/* File was dirty, so fall back to using generic logic */
143 	if (error == EBUSY) {
144 		if (hole)
145 			*off = file_sz;
146 
147 		return (0);
148 	}
149 
150 	/*
151 	 * We could find a hole that begins after the logical end-of-file,
152 	 * because dmu_offset_next() only works on whole blocks.  If the
153 	 * EOF falls mid-block, then indicate that the "virtual hole"
154 	 * at the end of the file begins at the logical EOF, rather than
155 	 * at the end of the last block.
156 	 */
157 	if (noff > file_sz) {
158 		ASSERT(hole);
159 		noff = file_sz;
160 	}
161 
162 	if (noff < *off)
163 		return (error);
164 	*off = noff;
165 	return (error);
166 }
167 
168 int
zfs_holey(znode_t * zp,ulong_t cmd,loff_t * off)169 zfs_holey(znode_t *zp, ulong_t cmd, loff_t *off)
170 {
171 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
172 	int error;
173 
174 	if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0)
175 		return (error);
176 
177 	error = zfs_holey_common(zp, cmd, off);
178 
179 	zfs_exit(zfsvfs, FTAG);
180 	return (error);
181 }
182 #endif /* SEEK_HOLE && SEEK_DATA */
183 
184 int
zfs_access(znode_t * zp,int mode,int flag,cred_t * cr)185 zfs_access(znode_t *zp, int mode, int flag, cred_t *cr)
186 {
187 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
188 	int error;
189 
190 	if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0)
191 		return (error);
192 
193 	if (flag & V_ACE_MASK)
194 #if defined(__linux__)
195 		error = zfs_zaccess(zp, mode, flag, B_FALSE, cr,
196 		    zfs_init_idmap);
197 #else
198 		error = zfs_zaccess(zp, mode, flag, B_FALSE, cr,
199 		    NULL);
200 #endif
201 	else
202 #if defined(__linux__)
203 		error = zfs_zaccess_rwx(zp, mode, flag, cr, zfs_init_idmap);
204 #else
205 		error = zfs_zaccess_rwx(zp, mode, flag, cr, NULL);
206 #endif
207 
208 	zfs_exit(zfsvfs, FTAG);
209 	return (error);
210 }
211 
212 /*
213  * Read bytes from specified file into supplied buffer.
214  *
215  *	IN:	zp	- inode of file to be read from.
216  *		uio	- structure supplying read location, range info,
217  *			  and return buffer.
218  *		ioflag	- O_SYNC flags; used to provide FRSYNC semantics.
219  *			  O_DIRECT flag; used to bypass page cache.
220  *		cr	- credentials of caller.
221  *
222  *	OUT:	uio	- updated offset and range, buffer filled.
223  *
224  *	RETURN:	0 on success, error code on failure.
225  *
226  * Side Effects:
227  *	inode - atime updated if byte count > 0
228  */
229 int
zfs_read(struct znode * zp,zfs_uio_t * uio,int ioflag,cred_t * cr)230 zfs_read(struct znode *zp, zfs_uio_t *uio, int ioflag, cred_t *cr)
231 {
232 	(void) cr;
233 	int error = 0;
234 	boolean_t frsync = B_FALSE;
235 
236 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
237 	if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0)
238 		return (error);
239 
240 	if (zp->z_pflags & ZFS_AV_QUARANTINED) {
241 		zfs_exit(zfsvfs, FTAG);
242 		return (SET_ERROR(EACCES));
243 	}
244 
245 	/* We don't copy out anything useful for directories. */
246 	if (Z_ISDIR(ZTOTYPE(zp))) {
247 		zfs_exit(zfsvfs, FTAG);
248 		return (SET_ERROR(EISDIR));
249 	}
250 
251 	/*
252 	 * Validate file offset
253 	 */
254 	if (zfs_uio_offset(uio) < (offset_t)0) {
255 		zfs_exit(zfsvfs, FTAG);
256 		return (SET_ERROR(EINVAL));
257 	}
258 
259 	/*
260 	 * Fasttrack empty reads
261 	 */
262 	if (zfs_uio_resid(uio) == 0) {
263 		zfs_exit(zfsvfs, FTAG);
264 		return (0);
265 	}
266 
267 #ifdef FRSYNC
268 	/*
269 	 * If we're in FRSYNC mode, sync out this znode before reading it.
270 	 * Only do this for non-snapshots.
271 	 *
272 	 * Some platforms do not support FRSYNC and instead map it
273 	 * to O_SYNC, which results in unnecessary calls to zil_commit. We
274 	 * only honor FRSYNC requests on platforms which support it.
275 	 */
276 	frsync = !!(ioflag & FRSYNC);
277 #endif
278 	if (zfsvfs->z_log &&
279 	    (frsync || zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS))
280 		zil_commit(zfsvfs->z_log, zp->z_id);
281 
282 	/*
283 	 * Lock the range against changes.
284 	 */
285 	zfs_locked_range_t *lr = zfs_rangelock_enter(&zp->z_rangelock,
286 	    zfs_uio_offset(uio), zfs_uio_resid(uio), RL_READER);
287 
288 	/*
289 	 * If we are reading past end-of-file we can skip
290 	 * to the end; but we might still need to set atime.
291 	 */
292 	if (zfs_uio_offset(uio) >= zp->z_size) {
293 		error = 0;
294 		goto out;
295 	}
296 
297 	ASSERT(zfs_uio_offset(uio) < zp->z_size);
298 #if defined(__linux__)
299 	ssize_t start_offset = zfs_uio_offset(uio);
300 #endif
301 	ssize_t n = MIN(zfs_uio_resid(uio), zp->z_size - zfs_uio_offset(uio));
302 	ssize_t start_resid = n;
303 
304 	while (n > 0) {
305 		ssize_t nbytes = MIN(n, zfs_vnops_read_chunk_size -
306 		    P2PHASE(zfs_uio_offset(uio), zfs_vnops_read_chunk_size));
307 #ifdef UIO_NOCOPY
308 		if (zfs_uio_segflg(uio) == UIO_NOCOPY)
309 			error = mappedread_sf(zp, nbytes, uio);
310 		else
311 #endif
312 		if (zn_has_cached_data(zp, zfs_uio_offset(uio),
313 		    zfs_uio_offset(uio) + nbytes - 1) && !(ioflag & O_DIRECT)) {
314 			error = mappedread(zp, nbytes, uio);
315 		} else {
316 			error = dmu_read_uio_dbuf(sa_get_db(zp->z_sa_hdl),
317 			    uio, nbytes);
318 		}
319 
320 		if (error) {
321 			/* convert checksum errors into IO errors */
322 			if (error == ECKSUM)
323 				error = SET_ERROR(EIO);
324 
325 #if defined(__linux__)
326 			/*
327 			 * if we actually read some bytes, bubbling EFAULT
328 			 * up to become EAGAIN isn't what we want here...
329 			 *
330 			 * ...on Linux, at least. On FBSD, doing this breaks.
331 			 */
332 			if (error == EFAULT &&
333 			    (zfs_uio_offset(uio) - start_offset) != 0)
334 				error = 0;
335 #endif
336 			break;
337 		}
338 
339 		n -= nbytes;
340 	}
341 
342 	int64_t nread = start_resid - n;
343 	dataset_kstats_update_read_kstats(&zfsvfs->z_kstat, nread);
344 	task_io_account_read(nread);
345 out:
346 	zfs_rangelock_exit(lr);
347 
348 	ZFS_ACCESSTIME_STAMP(zfsvfs, zp);
349 	zfs_exit(zfsvfs, FTAG);
350 	return (error);
351 }
352 
353 static void
zfs_clear_setid_bits_if_necessary(zfsvfs_t * zfsvfs,znode_t * zp,cred_t * cr,uint64_t * clear_setid_bits_txgp,dmu_tx_t * tx)354 zfs_clear_setid_bits_if_necessary(zfsvfs_t *zfsvfs, znode_t *zp, cred_t *cr,
355     uint64_t *clear_setid_bits_txgp, dmu_tx_t *tx)
356 {
357 	zilog_t *zilog = zfsvfs->z_log;
358 	const uint64_t uid = KUID_TO_SUID(ZTOUID(zp));
359 
360 	ASSERT(clear_setid_bits_txgp != NULL);
361 	ASSERT(tx != NULL);
362 
363 	/*
364 	 * Clear Set-UID/Set-GID bits on successful write if not
365 	 * privileged and at least one of the execute bits is set.
366 	 *
367 	 * It would be nice to do this after all writes have
368 	 * been done, but that would still expose the ISUID/ISGID
369 	 * to another app after the partial write is committed.
370 	 *
371 	 * Note: we don't call zfs_fuid_map_id() here because
372 	 * user 0 is not an ephemeral uid.
373 	 */
374 	mutex_enter(&zp->z_acl_lock);
375 	if ((zp->z_mode & (S_IXUSR | (S_IXUSR >> 3) | (S_IXUSR >> 6))) != 0 &&
376 	    (zp->z_mode & (S_ISUID | S_ISGID)) != 0 &&
377 	    secpolicy_vnode_setid_retain(zp, cr,
378 	    ((zp->z_mode & S_ISUID) != 0 && uid == 0)) != 0) {
379 		uint64_t newmode;
380 
381 		zp->z_mode &= ~(S_ISUID | S_ISGID);
382 		newmode = zp->z_mode;
383 		(void) sa_update(zp->z_sa_hdl, SA_ZPL_MODE(zfsvfs),
384 		    (void *)&newmode, sizeof (uint64_t), tx);
385 
386 		mutex_exit(&zp->z_acl_lock);
387 
388 		/*
389 		 * Make sure SUID/SGID bits will be removed when we replay the
390 		 * log. If the setid bits are keep coming back, don't log more
391 		 * than one TX_SETATTR per transaction group.
392 		 */
393 		if (*clear_setid_bits_txgp != dmu_tx_get_txg(tx)) {
394 			vattr_t va = {0};
395 
396 			va.va_mask = ATTR_MODE;
397 			va.va_nodeid = zp->z_id;
398 			va.va_mode = newmode;
399 			zfs_log_setattr(zilog, tx, TX_SETATTR, zp, &va,
400 			    ATTR_MODE, NULL);
401 			*clear_setid_bits_txgp = dmu_tx_get_txg(tx);
402 		}
403 	} else {
404 		mutex_exit(&zp->z_acl_lock);
405 	}
406 }
407 
408 /*
409  * Write the bytes to a file.
410  *
411  *	IN:	zp	- znode of file to be written to.
412  *		uio	- structure supplying write location, range info,
413  *			  and data buffer.
414  *		ioflag	- O_APPEND flag set if in append mode.
415  *			  O_DIRECT flag; used to bypass page cache.
416  *		cr	- credentials of caller.
417  *
418  *	OUT:	uio	- updated offset and range.
419  *
420  *	RETURN:	0 if success
421  *		error code if failure
422  *
423  * Timestamps:
424  *	ip - ctime|mtime updated if byte count > 0
425  */
426 int
zfs_write(znode_t * zp,zfs_uio_t * uio,int ioflag,cred_t * cr)427 zfs_write(znode_t *zp, zfs_uio_t *uio, int ioflag, cred_t *cr)
428 {
429 	int error = 0, error1;
430 	ssize_t start_resid = zfs_uio_resid(uio);
431 	uint64_t clear_setid_bits_txg = 0;
432 
433 	/*
434 	 * Fasttrack empty write
435 	 */
436 	ssize_t n = start_resid;
437 	if (n == 0)
438 		return (0);
439 
440 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
441 	if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0)
442 		return (error);
443 
444 	sa_bulk_attr_t bulk[4];
445 	int count = 0;
446 	uint64_t mtime[2], ctime[2];
447 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(zfsvfs), NULL, &mtime, 16);
448 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, 16);
449 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_SIZE(zfsvfs), NULL,
450 	    &zp->z_size, 8);
451 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_FLAGS(zfsvfs), NULL,
452 	    &zp->z_pflags, 8);
453 
454 	/*
455 	 * Callers might not be able to detect properly that we are read-only,
456 	 * so check it explicitly here.
457 	 */
458 	if (zfs_is_readonly(zfsvfs)) {
459 		zfs_exit(zfsvfs, FTAG);
460 		return (SET_ERROR(EROFS));
461 	}
462 
463 	/*
464 	 * If immutable or not appending then return EPERM.
465 	 * Intentionally allow ZFS_READONLY through here.
466 	 * See zfs_zaccess_common()
467 	 */
468 	if ((zp->z_pflags & ZFS_IMMUTABLE) ||
469 	    ((zp->z_pflags & ZFS_APPENDONLY) && !(ioflag & O_APPEND) &&
470 	    (zfs_uio_offset(uio) < zp->z_size))) {
471 		zfs_exit(zfsvfs, FTAG);
472 		return (SET_ERROR(EPERM));
473 	}
474 
475 	/*
476 	 * Validate file offset
477 	 */
478 	offset_t woff = ioflag & O_APPEND ? zp->z_size : zfs_uio_offset(uio);
479 	if (woff < 0) {
480 		zfs_exit(zfsvfs, FTAG);
481 		return (SET_ERROR(EINVAL));
482 	}
483 
484 	/*
485 	 * Pre-fault the pages to ensure slow (eg NFS) pages
486 	 * don't hold up txg.
487 	 */
488 	ssize_t pfbytes = MIN(n, DMU_MAX_ACCESS >> 1);
489 	if (zfs_uio_prefaultpages(pfbytes, uio)) {
490 		zfs_exit(zfsvfs, FTAG);
491 		return (SET_ERROR(EFAULT));
492 	}
493 
494 	/*
495 	 * If in append mode, set the io offset pointer to eof.
496 	 */
497 	zfs_locked_range_t *lr;
498 	if (ioflag & O_APPEND) {
499 		/*
500 		 * Obtain an appending range lock to guarantee file append
501 		 * semantics.  We reset the write offset once we have the lock.
502 		 */
503 		lr = zfs_rangelock_enter(&zp->z_rangelock, 0, n, RL_APPEND);
504 		woff = lr->lr_offset;
505 		if (lr->lr_length == UINT64_MAX) {
506 			/*
507 			 * We overlocked the file because this write will cause
508 			 * the file block size to increase.
509 			 * Note that zp_size cannot change with this lock held.
510 			 */
511 			woff = zp->z_size;
512 		}
513 		zfs_uio_setoffset(uio, woff);
514 	} else {
515 		/*
516 		 * Note that if the file block size will change as a result of
517 		 * this write, then this range lock will lock the entire file
518 		 * so that we can re-write the block safely.
519 		 */
520 		lr = zfs_rangelock_enter(&zp->z_rangelock, woff, n, RL_WRITER);
521 	}
522 
523 	if (zn_rlimit_fsize_uio(zp, uio)) {
524 		zfs_rangelock_exit(lr);
525 		zfs_exit(zfsvfs, FTAG);
526 		return (SET_ERROR(EFBIG));
527 	}
528 
529 	const rlim64_t limit = MAXOFFSET_T;
530 
531 	if (woff >= limit) {
532 		zfs_rangelock_exit(lr);
533 		zfs_exit(zfsvfs, FTAG);
534 		return (SET_ERROR(EFBIG));
535 	}
536 
537 	if (n > limit - woff)
538 		n = limit - woff;
539 
540 	uint64_t end_size = MAX(zp->z_size, woff + n);
541 	zilog_t *zilog = zfsvfs->z_log;
542 
543 	const uint64_t uid = KUID_TO_SUID(ZTOUID(zp));
544 	const uint64_t gid = KGID_TO_SGID(ZTOGID(zp));
545 	const uint64_t projid = zp->z_projid;
546 
547 	/*
548 	 * Write the file in reasonable size chunks.  Each chunk is written
549 	 * in a separate transaction; this keeps the intent log records small
550 	 * and allows us to do more fine-grained space accounting.
551 	 */
552 	while (n > 0) {
553 		woff = zfs_uio_offset(uio);
554 
555 		if (zfs_id_overblockquota(zfsvfs, DMU_USERUSED_OBJECT, uid) ||
556 		    zfs_id_overblockquota(zfsvfs, DMU_GROUPUSED_OBJECT, gid) ||
557 		    (projid != ZFS_DEFAULT_PROJID &&
558 		    zfs_id_overblockquota(zfsvfs, DMU_PROJECTUSED_OBJECT,
559 		    projid))) {
560 			error = SET_ERROR(EDQUOT);
561 			break;
562 		}
563 
564 		uint64_t blksz;
565 		if (lr->lr_length == UINT64_MAX && zp->z_size <= zp->z_blksz) {
566 			if (zp->z_blksz > zfsvfs->z_max_blksz &&
567 			    !ISP2(zp->z_blksz)) {
568 				/*
569 				 * File's blocksize is already larger than the
570 				 * "recordsize" property.  Only let it grow to
571 				 * the next power of 2.
572 				 */
573 				blksz = 1 << highbit64(zp->z_blksz);
574 			} else {
575 				blksz = zfsvfs->z_max_blksz;
576 			}
577 			blksz = MIN(blksz, P2ROUNDUP(end_size,
578 			    SPA_MINBLOCKSIZE));
579 			blksz = MAX(blksz, zp->z_blksz);
580 		} else {
581 			blksz = zp->z_blksz;
582 		}
583 
584 		arc_buf_t *abuf = NULL;
585 		ssize_t nbytes = n;
586 		if (n >= blksz && woff >= zp->z_size &&
587 		    P2PHASE(woff, blksz) == 0 &&
588 		    (blksz >= SPA_OLD_MAXBLOCKSIZE || n < 4 * blksz)) {
589 			/*
590 			 * This write covers a full block.  "Borrow" a buffer
591 			 * from the dmu so that we can fill it before we enter
592 			 * a transaction.  This avoids the possibility of
593 			 * holding up the transaction if the data copy hangs
594 			 * up on a pagefault (e.g., from an NFS server mapping).
595 			 */
596 			abuf = dmu_request_arcbuf(sa_get_db(zp->z_sa_hdl),
597 			    blksz);
598 			ASSERT(abuf != NULL);
599 			ASSERT(arc_buf_size(abuf) == blksz);
600 			if ((error = zfs_uiocopy(abuf->b_data, blksz,
601 			    UIO_WRITE, uio, &nbytes))) {
602 				dmu_return_arcbuf(abuf);
603 				break;
604 			}
605 			ASSERT3S(nbytes, ==, blksz);
606 		} else {
607 			nbytes = MIN(n, (DMU_MAX_ACCESS >> 1) -
608 			    P2PHASE(woff, blksz));
609 			if (pfbytes < nbytes) {
610 				if (zfs_uio_prefaultpages(nbytes, uio)) {
611 					error = SET_ERROR(EFAULT);
612 					break;
613 				}
614 				pfbytes = nbytes;
615 			}
616 		}
617 
618 		/*
619 		 * Start a transaction.
620 		 */
621 		dmu_tx_t *tx = dmu_tx_create(zfsvfs->z_os);
622 		dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_FALSE);
623 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)sa_get_db(zp->z_sa_hdl);
624 		DB_DNODE_ENTER(db);
625 		dmu_tx_hold_write_by_dnode(tx, DB_DNODE(db), woff, nbytes);
626 		DB_DNODE_EXIT(db);
627 		zfs_sa_upgrade_txholds(tx, zp);
628 		error = dmu_tx_assign(tx, TXG_WAIT);
629 		if (error) {
630 			dmu_tx_abort(tx);
631 			if (abuf != NULL)
632 				dmu_return_arcbuf(abuf);
633 			break;
634 		}
635 
636 		/*
637 		 * NB: We must call zfs_clear_setid_bits_if_necessary before
638 		 * committing the transaction!
639 		 */
640 
641 		/*
642 		 * If rangelock_enter() over-locked we grow the blocksize
643 		 * and then reduce the lock range.  This will only happen
644 		 * on the first iteration since rangelock_reduce() will
645 		 * shrink down lr_length to the appropriate size.
646 		 */
647 		if (lr->lr_length == UINT64_MAX) {
648 			zfs_grow_blocksize(zp, blksz, tx);
649 			zfs_rangelock_reduce(lr, woff, n);
650 		}
651 
652 		ssize_t tx_bytes;
653 		if (abuf == NULL) {
654 			tx_bytes = zfs_uio_resid(uio);
655 			zfs_uio_fault_disable(uio, B_TRUE);
656 			error = dmu_write_uio_dbuf(sa_get_db(zp->z_sa_hdl),
657 			    uio, nbytes, tx);
658 			zfs_uio_fault_disable(uio, B_FALSE);
659 #ifdef __linux__
660 			if (error == EFAULT) {
661 				zfs_clear_setid_bits_if_necessary(zfsvfs, zp,
662 				    cr, &clear_setid_bits_txg, tx);
663 				dmu_tx_commit(tx);
664 				/*
665 				 * Account for partial writes before
666 				 * continuing the loop.
667 				 * Update needs to occur before the next
668 				 * zfs_uio_prefaultpages, or prefaultpages may
669 				 * error, and we may break the loop early.
670 				 */
671 				n -= tx_bytes - zfs_uio_resid(uio);
672 				pfbytes -= tx_bytes - zfs_uio_resid(uio);
673 				continue;
674 			}
675 #endif
676 			/*
677 			 * On FreeBSD, EFAULT should be propagated back to the
678 			 * VFS, which will handle faulting and will retry.
679 			 */
680 			if (error != 0 && error != EFAULT) {
681 				zfs_clear_setid_bits_if_necessary(zfsvfs, zp,
682 				    cr, &clear_setid_bits_txg, tx);
683 				dmu_tx_commit(tx);
684 				break;
685 			}
686 			tx_bytes -= zfs_uio_resid(uio);
687 		} else {
688 			/*
689 			 * Thus, we're writing a full block at a block-aligned
690 			 * offset and extending the file past EOF.
691 			 *
692 			 * dmu_assign_arcbuf_by_dbuf() will directly assign the
693 			 * arc buffer to a dbuf.
694 			 */
695 			error = dmu_assign_arcbuf_by_dbuf(
696 			    sa_get_db(zp->z_sa_hdl), woff, abuf, tx);
697 			if (error != 0) {
698 				/*
699 				 * XXX This might not be necessary if
700 				 * dmu_assign_arcbuf_by_dbuf is guaranteed
701 				 * to be atomic.
702 				 */
703 				zfs_clear_setid_bits_if_necessary(zfsvfs, zp,
704 				    cr, &clear_setid_bits_txg, tx);
705 				dmu_return_arcbuf(abuf);
706 				dmu_tx_commit(tx);
707 				break;
708 			}
709 			ASSERT3S(nbytes, <=, zfs_uio_resid(uio));
710 			zfs_uioskip(uio, nbytes);
711 			tx_bytes = nbytes;
712 		}
713 		if (tx_bytes &&
714 		    zn_has_cached_data(zp, woff, woff + tx_bytes - 1) &&
715 		    !(ioflag & O_DIRECT)) {
716 			update_pages(zp, woff, tx_bytes, zfsvfs->z_os);
717 		}
718 
719 		/*
720 		 * If we made no progress, we're done.  If we made even
721 		 * partial progress, update the znode and ZIL accordingly.
722 		 */
723 		if (tx_bytes == 0) {
724 			(void) sa_update(zp->z_sa_hdl, SA_ZPL_SIZE(zfsvfs),
725 			    (void *)&zp->z_size, sizeof (uint64_t), tx);
726 			dmu_tx_commit(tx);
727 			ASSERT(error != 0);
728 			break;
729 		}
730 
731 		zfs_clear_setid_bits_if_necessary(zfsvfs, zp, cr,
732 		    &clear_setid_bits_txg, tx);
733 
734 		zfs_tstamp_update_setup(zp, CONTENT_MODIFIED, mtime, ctime);
735 
736 		/*
737 		 * Update the file size (zp_size) if it has changed;
738 		 * account for possible concurrent updates.
739 		 */
740 		while ((end_size = zp->z_size) < zfs_uio_offset(uio)) {
741 			(void) atomic_cas_64(&zp->z_size, end_size,
742 			    zfs_uio_offset(uio));
743 			ASSERT(error == 0 || error == EFAULT);
744 		}
745 		/*
746 		 * If we are replaying and eof is non zero then force
747 		 * the file size to the specified eof. Note, there's no
748 		 * concurrency during replay.
749 		 */
750 		if (zfsvfs->z_replay && zfsvfs->z_replay_eof != 0)
751 			zp->z_size = zfsvfs->z_replay_eof;
752 
753 		error1 = sa_bulk_update(zp->z_sa_hdl, bulk, count, tx);
754 		if (error1 != 0)
755 			/* Avoid clobbering EFAULT. */
756 			error = error1;
757 
758 		/*
759 		 * NB: During replay, the TX_SETATTR record logged by
760 		 * zfs_clear_setid_bits_if_necessary must precede any of
761 		 * the TX_WRITE records logged here.
762 		 */
763 		zfs_log_write(zilog, tx, TX_WRITE, zp, woff, tx_bytes, ioflag,
764 		    NULL, NULL);
765 
766 		dmu_tx_commit(tx);
767 
768 		if (error != 0)
769 			break;
770 		ASSERT3S(tx_bytes, ==, nbytes);
771 		n -= nbytes;
772 		pfbytes -= nbytes;
773 	}
774 
775 	zfs_znode_update_vfs(zp);
776 	zfs_rangelock_exit(lr);
777 
778 	/*
779 	 * If we're in replay mode, or we made no progress, or the
780 	 * uio data is inaccessible return an error.  Otherwise, it's
781 	 * at least a partial write, so it's successful.
782 	 */
783 	if (zfsvfs->z_replay || zfs_uio_resid(uio) == start_resid ||
784 	    error == EFAULT) {
785 		zfs_exit(zfsvfs, FTAG);
786 		return (error);
787 	}
788 
789 	if (ioflag & (O_SYNC | O_DSYNC) ||
790 	    zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS)
791 		zil_commit(zilog, zp->z_id);
792 
793 	const int64_t nwritten = start_resid - zfs_uio_resid(uio);
794 	dataset_kstats_update_write_kstats(&zfsvfs->z_kstat, nwritten);
795 	task_io_account_write(nwritten);
796 
797 	zfs_exit(zfsvfs, FTAG);
798 	return (0);
799 }
800 
801 int
zfs_getsecattr(znode_t * zp,vsecattr_t * vsecp,int flag,cred_t * cr)802 zfs_getsecattr(znode_t *zp, vsecattr_t *vsecp, int flag, cred_t *cr)
803 {
804 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
805 	int error;
806 	boolean_t skipaclchk = (flag & ATTR_NOACLCHECK) ? B_TRUE : B_FALSE;
807 
808 	if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0)
809 		return (error);
810 	error = zfs_getacl(zp, vsecp, skipaclchk, cr);
811 	zfs_exit(zfsvfs, FTAG);
812 
813 	return (error);
814 }
815 
816 int
zfs_setsecattr(znode_t * zp,vsecattr_t * vsecp,int flag,cred_t * cr)817 zfs_setsecattr(znode_t *zp, vsecattr_t *vsecp, int flag, cred_t *cr)
818 {
819 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
820 	int error;
821 	boolean_t skipaclchk = (flag & ATTR_NOACLCHECK) ? B_TRUE : B_FALSE;
822 	zilog_t	*zilog;
823 
824 	if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0)
825 		return (error);
826 	zilog = zfsvfs->z_log;
827 	error = zfs_setacl(zp, vsecp, skipaclchk, cr);
828 
829 	if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS)
830 		zil_commit(zilog, 0);
831 
832 	zfs_exit(zfsvfs, FTAG);
833 	return (error);
834 }
835 
836 #ifdef ZFS_DEBUG
837 static int zil_fault_io = 0;
838 #endif
839 
840 static void zfs_get_done(zgd_t *zgd, int error);
841 
842 /*
843  * Get data to generate a TX_WRITE intent log record.
844  */
845 int
zfs_get_data(void * arg,uint64_t gen,lr_write_t * lr,char * buf,struct lwb * lwb,zio_t * zio)846 zfs_get_data(void *arg, uint64_t gen, lr_write_t *lr, char *buf,
847     struct lwb *lwb, zio_t *zio)
848 {
849 	zfsvfs_t *zfsvfs = arg;
850 	objset_t *os = zfsvfs->z_os;
851 	znode_t *zp;
852 	uint64_t object = lr->lr_foid;
853 	uint64_t offset = lr->lr_offset;
854 	uint64_t size = lr->lr_length;
855 	dmu_buf_t *db;
856 	zgd_t *zgd;
857 	int error = 0;
858 	uint64_t zp_gen;
859 
860 	ASSERT3P(lwb, !=, NULL);
861 	ASSERT3U(size, !=, 0);
862 
863 	/*
864 	 * Nothing to do if the file has been removed
865 	 */
866 	if (zfs_zget(zfsvfs, object, &zp) != 0)
867 		return (SET_ERROR(ENOENT));
868 	if (zp->z_unlinked) {
869 		/*
870 		 * Release the vnode asynchronously as we currently have the
871 		 * txg stopped from syncing.
872 		 */
873 		zfs_zrele_async(zp);
874 		return (SET_ERROR(ENOENT));
875 	}
876 	/* check if generation number matches */
877 	if (sa_lookup(zp->z_sa_hdl, SA_ZPL_GEN(zfsvfs), &zp_gen,
878 	    sizeof (zp_gen)) != 0) {
879 		zfs_zrele_async(zp);
880 		return (SET_ERROR(EIO));
881 	}
882 	if (zp_gen != gen) {
883 		zfs_zrele_async(zp);
884 		return (SET_ERROR(ENOENT));
885 	}
886 
887 	zgd = kmem_zalloc(sizeof (zgd_t), KM_SLEEP);
888 	zgd->zgd_lwb = lwb;
889 	zgd->zgd_private = zp;
890 
891 	/*
892 	 * Write records come in two flavors: immediate and indirect.
893 	 * For small writes it's cheaper to store the data with the
894 	 * log record (immediate); for large writes it's cheaper to
895 	 * sync the data and get a pointer to it (indirect) so that
896 	 * we don't have to write the data twice.
897 	 */
898 	if (buf != NULL) { /* immediate write */
899 		zgd->zgd_lr = zfs_rangelock_enter(&zp->z_rangelock,
900 		    offset, size, RL_READER);
901 		/* test for truncation needs to be done while range locked */
902 		if (offset >= zp->z_size) {
903 			error = SET_ERROR(ENOENT);
904 		} else {
905 			error = dmu_read(os, object, offset, size, buf,
906 			    DMU_READ_NO_PREFETCH);
907 		}
908 		ASSERT(error == 0 || error == ENOENT);
909 	} else { /* indirect write */
910 		ASSERT3P(zio, !=, NULL);
911 		/*
912 		 * Have to lock the whole block to ensure when it's
913 		 * written out and its checksum is being calculated
914 		 * that no one can change the data. We need to re-check
915 		 * blocksize after we get the lock in case it's changed!
916 		 */
917 		for (;;) {
918 			uint64_t blkoff;
919 			size = zp->z_blksz;
920 			blkoff = ISP2(size) ? P2PHASE(offset, size) : offset;
921 			offset -= blkoff;
922 			zgd->zgd_lr = zfs_rangelock_enter(&zp->z_rangelock,
923 			    offset, size, RL_READER);
924 			if (zp->z_blksz == size)
925 				break;
926 			offset += blkoff;
927 			zfs_rangelock_exit(zgd->zgd_lr);
928 		}
929 		/* test for truncation needs to be done while range locked */
930 		if (lr->lr_offset >= zp->z_size)
931 			error = SET_ERROR(ENOENT);
932 #ifdef ZFS_DEBUG
933 		if (zil_fault_io) {
934 			error = SET_ERROR(EIO);
935 			zil_fault_io = 0;
936 		}
937 #endif
938 		if (error == 0)
939 			error = dmu_buf_hold_noread(os, object, offset, zgd,
940 			    &db);
941 
942 		if (error == 0) {
943 			blkptr_t *bp = &lr->lr_blkptr;
944 
945 			zgd->zgd_db = db;
946 			zgd->zgd_bp = bp;
947 
948 			ASSERT(db->db_offset == offset);
949 			ASSERT(db->db_size == size);
950 
951 			error = dmu_sync(zio, lr->lr_common.lrc_txg,
952 			    zfs_get_done, zgd);
953 			ASSERT(error || lr->lr_length <= size);
954 
955 			/*
956 			 * On success, we need to wait for the write I/O
957 			 * initiated by dmu_sync() to complete before we can
958 			 * release this dbuf.  We will finish everything up
959 			 * in the zfs_get_done() callback.
960 			 */
961 			if (error == 0)
962 				return (0);
963 
964 			if (error == EALREADY) {
965 				lr->lr_common.lrc_txtype = TX_WRITE2;
966 				/*
967 				 * TX_WRITE2 relies on the data previously
968 				 * written by the TX_WRITE that caused
969 				 * EALREADY.  We zero out the BP because
970 				 * it is the old, currently-on-disk BP.
971 				 */
972 				zgd->zgd_bp = NULL;
973 				BP_ZERO(bp);
974 				error = 0;
975 			}
976 		}
977 	}
978 
979 	zfs_get_done(zgd, error);
980 
981 	return (error);
982 }
983 
984 
985 static void
zfs_get_done(zgd_t * zgd,int error)986 zfs_get_done(zgd_t *zgd, int error)
987 {
988 	(void) error;
989 	znode_t *zp = zgd->zgd_private;
990 
991 	if (zgd->zgd_db)
992 		dmu_buf_rele(zgd->zgd_db, zgd);
993 
994 	zfs_rangelock_exit(zgd->zgd_lr);
995 
996 	/*
997 	 * Release the vnode asynchronously as we currently have the
998 	 * txg stopped from syncing.
999 	 */
1000 	zfs_zrele_async(zp);
1001 
1002 	kmem_free(zgd, sizeof (zgd_t));
1003 }
1004 
1005 static int
zfs_enter_two(zfsvfs_t * zfsvfs1,zfsvfs_t * zfsvfs2,const char * tag)1006 zfs_enter_two(zfsvfs_t *zfsvfs1, zfsvfs_t *zfsvfs2, const char *tag)
1007 {
1008 	int error;
1009 
1010 	/* Swap. Not sure if the order of zfs_enter()s is important. */
1011 	if (zfsvfs1 > zfsvfs2) {
1012 		zfsvfs_t *tmpzfsvfs;
1013 
1014 		tmpzfsvfs = zfsvfs2;
1015 		zfsvfs2 = zfsvfs1;
1016 		zfsvfs1 = tmpzfsvfs;
1017 	}
1018 
1019 	error = zfs_enter(zfsvfs1, tag);
1020 	if (error != 0)
1021 		return (error);
1022 	if (zfsvfs1 != zfsvfs2) {
1023 		error = zfs_enter(zfsvfs2, tag);
1024 		if (error != 0) {
1025 			zfs_exit(zfsvfs1, tag);
1026 			return (error);
1027 		}
1028 	}
1029 
1030 	return (0);
1031 }
1032 
1033 static void
zfs_exit_two(zfsvfs_t * zfsvfs1,zfsvfs_t * zfsvfs2,const char * tag)1034 zfs_exit_two(zfsvfs_t *zfsvfs1, zfsvfs_t *zfsvfs2, const char *tag)
1035 {
1036 
1037 	zfs_exit(zfsvfs1, tag);
1038 	if (zfsvfs1 != zfsvfs2)
1039 		zfs_exit(zfsvfs2, tag);
1040 }
1041 
1042 /*
1043  * We split each clone request in chunks that can fit into a single ZIL
1044  * log entry. Each ZIL log entry can fit 130816 bytes for a block cloning
1045  * operation (see zil_max_log_data() and zfs_log_clone_range()). This gives
1046  * us room for storing 1022 block pointers.
1047  *
1048  * On success, the function return the number of bytes copied in *lenp.
1049  * Note, it doesn't return how much bytes are left to be copied.
1050  * On errors which are caused by any file system limitations or
1051  * brt limitations `EINVAL` is returned. In the most cases a user
1052  * requested bad parameters, it could be possible to clone the file but
1053  * some parameters don't match the requirements.
1054  */
1055 int
zfs_clone_range(znode_t * inzp,uint64_t * inoffp,znode_t * outzp,uint64_t * outoffp,uint64_t * lenp,cred_t * cr)1056 zfs_clone_range(znode_t *inzp, uint64_t *inoffp, znode_t *outzp,
1057     uint64_t *outoffp, uint64_t *lenp, cred_t *cr)
1058 {
1059 	zfsvfs_t	*inzfsvfs, *outzfsvfs;
1060 	objset_t	*inos, *outos;
1061 	zfs_locked_range_t *inlr, *outlr;
1062 	dmu_buf_impl_t	*db;
1063 	dmu_tx_t	*tx;
1064 	zilog_t		*zilog;
1065 	uint64_t	inoff, outoff, len, done;
1066 	uint64_t	outsize, size;
1067 	int		error;
1068 	int		count = 0;
1069 	sa_bulk_attr_t	bulk[3];
1070 	uint64_t	mtime[2], ctime[2];
1071 	uint64_t	uid, gid, projid;
1072 	blkptr_t	*bps;
1073 	size_t		maxblocks, nbps;
1074 	uint_t		inblksz;
1075 	uint64_t	clear_setid_bits_txg = 0;
1076 	uint64_t	last_synced_txg = 0;
1077 
1078 	inoff = *inoffp;
1079 	outoff = *outoffp;
1080 	len = *lenp;
1081 	done = 0;
1082 
1083 	inzfsvfs = ZTOZSB(inzp);
1084 	outzfsvfs = ZTOZSB(outzp);
1085 
1086 	/*
1087 	 * We need to call zfs_enter() potentially on two different datasets,
1088 	 * so we need a dedicated function for that.
1089 	 */
1090 	error = zfs_enter_two(inzfsvfs, outzfsvfs, FTAG);
1091 	if (error != 0)
1092 		return (error);
1093 
1094 	inos = inzfsvfs->z_os;
1095 	outos = outzfsvfs->z_os;
1096 
1097 	/*
1098 	 * Both source and destination have to belong to the same storage pool.
1099 	 */
1100 	if (dmu_objset_spa(inos) != dmu_objset_spa(outos)) {
1101 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1102 		return (SET_ERROR(EXDEV));
1103 	}
1104 
1105 	/*
1106 	 * outos and inos belongs to the same storage pool.
1107 	 * see a few lines above, only one check.
1108 	 */
1109 	if (!spa_feature_is_enabled(dmu_objset_spa(outos),
1110 	    SPA_FEATURE_BLOCK_CLONING)) {
1111 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1112 		return (SET_ERROR(EOPNOTSUPP));
1113 	}
1114 
1115 	ASSERT(!outzfsvfs->z_replay);
1116 
1117 	/*
1118 	 * Block cloning from an unencrypted dataset into an encrypted
1119 	 * dataset and vice versa is not supported.
1120 	 */
1121 	if (inos->os_encrypted != outos->os_encrypted) {
1122 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1123 		return (SET_ERROR(EXDEV));
1124 	}
1125 
1126 	/*
1127 	 * Cloning across encrypted datasets is possible only if they
1128 	 * share the same master key.
1129 	 */
1130 	if (inos != outos && inos->os_encrypted &&
1131 	    !dmu_objset_crypto_key_equal(inos, outos)) {
1132 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1133 		return (SET_ERROR(EXDEV));
1134 	}
1135 
1136 	error = zfs_verify_zp(inzp);
1137 	if (error == 0)
1138 		error = zfs_verify_zp(outzp);
1139 	if (error != 0) {
1140 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1141 		return (error);
1142 	}
1143 
1144 	/*
1145 	 * We don't copy source file's flags that's why we don't allow to clone
1146 	 * files that are in quarantine.
1147 	 */
1148 	if (inzp->z_pflags & ZFS_AV_QUARANTINED) {
1149 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1150 		return (SET_ERROR(EACCES));
1151 	}
1152 
1153 	if (inoff >= inzp->z_size) {
1154 		*lenp = 0;
1155 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1156 		return (0);
1157 	}
1158 	if (len > inzp->z_size - inoff) {
1159 		len = inzp->z_size - inoff;
1160 	}
1161 	if (len == 0) {
1162 		*lenp = 0;
1163 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1164 		return (0);
1165 	}
1166 
1167 	/*
1168 	 * Callers might not be able to detect properly that we are read-only,
1169 	 * so check it explicitly here.
1170 	 */
1171 	if (zfs_is_readonly(outzfsvfs)) {
1172 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1173 		return (SET_ERROR(EROFS));
1174 	}
1175 
1176 	/*
1177 	 * If immutable or not appending then return EPERM.
1178 	 * Intentionally allow ZFS_READONLY through here.
1179 	 * See zfs_zaccess_common()
1180 	 */
1181 	if ((outzp->z_pflags & ZFS_IMMUTABLE) != 0) {
1182 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1183 		return (SET_ERROR(EPERM));
1184 	}
1185 
1186 	/*
1187 	 * No overlapping if we are cloning within the same file.
1188 	 */
1189 	if (inzp == outzp) {
1190 		if (inoff < outoff + len && outoff < inoff + len) {
1191 			zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1192 			return (SET_ERROR(EINVAL));
1193 		}
1194 	}
1195 
1196 	/* Flush any mmap()'d data to disk */
1197 	if (zn_has_cached_data(inzp, inoff, inoff + len - 1))
1198 		zn_flush_cached_data(inzp, B_TRUE);
1199 
1200 	/*
1201 	 * Maintain predictable lock order.
1202 	 */
1203 	if (inzp < outzp || (inzp == outzp && inoff < outoff)) {
1204 		inlr = zfs_rangelock_enter(&inzp->z_rangelock, inoff, len,
1205 		    RL_READER);
1206 		outlr = zfs_rangelock_enter(&outzp->z_rangelock, outoff, len,
1207 		    RL_WRITER);
1208 	} else {
1209 		outlr = zfs_rangelock_enter(&outzp->z_rangelock, outoff, len,
1210 		    RL_WRITER);
1211 		inlr = zfs_rangelock_enter(&inzp->z_rangelock, inoff, len,
1212 		    RL_READER);
1213 	}
1214 
1215 	inblksz = inzp->z_blksz;
1216 
1217 	/*
1218 	 * We cannot clone into a file with different block size if we can't
1219 	 * grow it (block size is already bigger, has more than one block, or
1220 	 * not locked for growth).  There are other possible reasons for the
1221 	 * grow to fail, but we cover what we can before opening transaction
1222 	 * and the rest detect after we try to do it.
1223 	 */
1224 	if (inblksz < outzp->z_blksz) {
1225 		error = SET_ERROR(EINVAL);
1226 		goto unlock;
1227 	}
1228 	if (inblksz != outzp->z_blksz && (outzp->z_size > outzp->z_blksz ||
1229 	    outlr->lr_length != UINT64_MAX)) {
1230 		error = SET_ERROR(EINVAL);
1231 		goto unlock;
1232 	}
1233 
1234 	/*
1235 	 * Block size must be power-of-2 if destination offset != 0.
1236 	 * There can be no multiple blocks of non-power-of-2 size.
1237 	 */
1238 	if (outoff != 0 && !ISP2(inblksz)) {
1239 		error = SET_ERROR(EINVAL);
1240 		goto unlock;
1241 	}
1242 
1243 	/*
1244 	 * Offsets and len must be at block boundries.
1245 	 */
1246 	if ((inoff % inblksz) != 0 || (outoff % inblksz) != 0) {
1247 		error = SET_ERROR(EINVAL);
1248 		goto unlock;
1249 	}
1250 	/*
1251 	 * Length must be multipe of blksz, except for the end of the file.
1252 	 */
1253 	if ((len % inblksz) != 0 &&
1254 	    (len < inzp->z_size - inoff || len < outzp->z_size - outoff)) {
1255 		error = SET_ERROR(EINVAL);
1256 		goto unlock;
1257 	}
1258 
1259 	/*
1260 	 * If we are copying only one block and it is smaller than recordsize
1261 	 * property, do not allow destination to grow beyond one block if it
1262 	 * is not there yet.  Otherwise the destination will get stuck with
1263 	 * that block size forever, that can be as small as 512 bytes, no
1264 	 * matter how big the destination grow later.
1265 	 */
1266 	if (len <= inblksz && inblksz < outzfsvfs->z_max_blksz &&
1267 	    outzp->z_size <= inblksz && outoff + len > inblksz) {
1268 		error = SET_ERROR(EINVAL);
1269 		goto unlock;
1270 	}
1271 
1272 	error = zn_rlimit_fsize(outoff + len);
1273 	if (error != 0) {
1274 		goto unlock;
1275 	}
1276 
1277 	if (inoff >= MAXOFFSET_T || outoff >= MAXOFFSET_T) {
1278 		error = SET_ERROR(EFBIG);
1279 		goto unlock;
1280 	}
1281 
1282 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(outzfsvfs), NULL,
1283 	    &mtime, 16);
1284 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(outzfsvfs), NULL,
1285 	    &ctime, 16);
1286 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_SIZE(outzfsvfs), NULL,
1287 	    &outzp->z_size, 8);
1288 
1289 	zilog = outzfsvfs->z_log;
1290 	maxblocks = zil_max_log_data(zilog, sizeof (lr_clone_range_t)) /
1291 	    sizeof (bps[0]);
1292 
1293 	uid = KUID_TO_SUID(ZTOUID(outzp));
1294 	gid = KGID_TO_SGID(ZTOGID(outzp));
1295 	projid = outzp->z_projid;
1296 
1297 	bps = vmem_alloc(sizeof (bps[0]) * maxblocks, KM_SLEEP);
1298 
1299 	/*
1300 	 * Clone the file in reasonable size chunks.  Each chunk is cloned
1301 	 * in a separate transaction; this keeps the intent log records small
1302 	 * and allows us to do more fine-grained space accounting.
1303 	 */
1304 	while (len > 0) {
1305 		size = MIN(inblksz * maxblocks, len);
1306 
1307 		if (zfs_id_overblockquota(outzfsvfs, DMU_USERUSED_OBJECT,
1308 		    uid) ||
1309 		    zfs_id_overblockquota(outzfsvfs, DMU_GROUPUSED_OBJECT,
1310 		    gid) ||
1311 		    (projid != ZFS_DEFAULT_PROJID &&
1312 		    zfs_id_overblockquota(outzfsvfs, DMU_PROJECTUSED_OBJECT,
1313 		    projid))) {
1314 			error = SET_ERROR(EDQUOT);
1315 			break;
1316 		}
1317 
1318 		nbps = maxblocks;
1319 		last_synced_txg = spa_last_synced_txg(dmu_objset_spa(inos));
1320 		error = dmu_read_l0_bps(inos, inzp->z_id, inoff, size, bps,
1321 		    &nbps);
1322 		if (error != 0) {
1323 			/*
1324 			 * If we are trying to clone a block that was created
1325 			 * in the current transaction group, the error will be
1326 			 * EAGAIN here.  Based on zfs_bclone_wait_dirty either
1327 			 * return a shortened range to the caller so it can
1328 			 * fallback, or wait for the next TXG and check again.
1329 			 */
1330 			if (error == EAGAIN && zfs_bclone_wait_dirty) {
1331 				txg_wait_synced(dmu_objset_pool(inos),
1332 				    last_synced_txg + 1);
1333 				continue;
1334 			}
1335 
1336 			break;
1337 		}
1338 
1339 		/*
1340 		 * Start a transaction.
1341 		 */
1342 		tx = dmu_tx_create(outos);
1343 		dmu_tx_hold_sa(tx, outzp->z_sa_hdl, B_FALSE);
1344 		db = (dmu_buf_impl_t *)sa_get_db(outzp->z_sa_hdl);
1345 		DB_DNODE_ENTER(db);
1346 		dmu_tx_hold_clone_by_dnode(tx, DB_DNODE(db), outoff, size);
1347 		DB_DNODE_EXIT(db);
1348 		zfs_sa_upgrade_txholds(tx, outzp);
1349 		error = dmu_tx_assign(tx, TXG_WAIT);
1350 		if (error != 0) {
1351 			dmu_tx_abort(tx);
1352 			break;
1353 		}
1354 
1355 		/*
1356 		 * Copy source znode's block size. This is done only if the
1357 		 * whole znode is locked (see zfs_rangelock_cb()) and only
1358 		 * on the first iteration since zfs_rangelock_reduce() will
1359 		 * shrink down lr_length to the appropriate size.
1360 		 */
1361 		if (outlr->lr_length == UINT64_MAX) {
1362 			zfs_grow_blocksize(outzp, inblksz, tx);
1363 
1364 			/*
1365 			 * Block growth may fail for many reasons we can not
1366 			 * predict here.  If it happen the cloning is doomed.
1367 			 */
1368 			if (inblksz != outzp->z_blksz) {
1369 				error = SET_ERROR(EINVAL);
1370 				dmu_tx_abort(tx);
1371 				break;
1372 			}
1373 
1374 			/*
1375 			 * Round range lock up to the block boundary, so we
1376 			 * prevent appends until we are done.
1377 			 */
1378 			zfs_rangelock_reduce(outlr, outoff,
1379 			    ((len - 1) / inblksz + 1) * inblksz);
1380 		}
1381 
1382 		error = dmu_brt_clone(outos, outzp->z_id, outoff, size, tx,
1383 		    bps, nbps);
1384 		if (error != 0) {
1385 			dmu_tx_commit(tx);
1386 			break;
1387 		}
1388 
1389 		if (zn_has_cached_data(outzp, outoff, outoff + size - 1)) {
1390 			update_pages(outzp, outoff, size, outos);
1391 		}
1392 
1393 		zfs_clear_setid_bits_if_necessary(outzfsvfs, outzp, cr,
1394 		    &clear_setid_bits_txg, tx);
1395 
1396 		zfs_tstamp_update_setup(outzp, CONTENT_MODIFIED, mtime, ctime);
1397 
1398 		/*
1399 		 * Update the file size (zp_size) if it has changed;
1400 		 * account for possible concurrent updates.
1401 		 */
1402 		while ((outsize = outzp->z_size) < outoff + size) {
1403 			(void) atomic_cas_64(&outzp->z_size, outsize,
1404 			    outoff + size);
1405 		}
1406 
1407 		error = sa_bulk_update(outzp->z_sa_hdl, bulk, count, tx);
1408 
1409 		zfs_log_clone_range(zilog, tx, TX_CLONE_RANGE, outzp, outoff,
1410 		    size, inblksz, bps, nbps);
1411 
1412 		dmu_tx_commit(tx);
1413 
1414 		if (error != 0)
1415 			break;
1416 
1417 		inoff += size;
1418 		outoff += size;
1419 		len -= size;
1420 		done += size;
1421 	}
1422 
1423 	vmem_free(bps, sizeof (bps[0]) * maxblocks);
1424 	zfs_znode_update_vfs(outzp);
1425 
1426 unlock:
1427 	zfs_rangelock_exit(outlr);
1428 	zfs_rangelock_exit(inlr);
1429 
1430 	if (done > 0) {
1431 		/*
1432 		 * If we have made at least partial progress, reset the error.
1433 		 */
1434 		error = 0;
1435 
1436 		ZFS_ACCESSTIME_STAMP(inzfsvfs, inzp);
1437 
1438 		if (outos->os_sync == ZFS_SYNC_ALWAYS) {
1439 			zil_commit(zilog, outzp->z_id);
1440 		}
1441 
1442 		*inoffp += done;
1443 		*outoffp += done;
1444 		*lenp = done;
1445 	} else {
1446 		/*
1447 		 * If we made no progress, there must be a good reason.
1448 		 * EOF is handled explicitly above, before the loop.
1449 		 */
1450 		ASSERT3S(error, !=, 0);
1451 	}
1452 
1453 	zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1454 
1455 	return (error);
1456 }
1457 
1458 /*
1459  * Usual pattern would be to call zfs_clone_range() from zfs_replay_clone(),
1460  * but we cannot do that, because when replaying we don't have source znode
1461  * available. This is why we need a dedicated replay function.
1462  */
1463 int
zfs_clone_range_replay(znode_t * zp,uint64_t off,uint64_t len,uint64_t blksz,const blkptr_t * bps,size_t nbps)1464 zfs_clone_range_replay(znode_t *zp, uint64_t off, uint64_t len, uint64_t blksz,
1465     const blkptr_t *bps, size_t nbps)
1466 {
1467 	zfsvfs_t	*zfsvfs;
1468 	dmu_buf_impl_t	*db;
1469 	dmu_tx_t	*tx;
1470 	int		error;
1471 	int		count = 0;
1472 	sa_bulk_attr_t	bulk[3];
1473 	uint64_t	mtime[2], ctime[2];
1474 
1475 	ASSERT3U(off, <, MAXOFFSET_T);
1476 	ASSERT3U(len, >, 0);
1477 	ASSERT3U(nbps, >, 0);
1478 
1479 	zfsvfs = ZTOZSB(zp);
1480 
1481 	ASSERT(spa_feature_is_enabled(dmu_objset_spa(zfsvfs->z_os),
1482 	    SPA_FEATURE_BLOCK_CLONING));
1483 
1484 	if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0)
1485 		return (error);
1486 
1487 	ASSERT(zfsvfs->z_replay);
1488 	ASSERT(!zfs_is_readonly(zfsvfs));
1489 
1490 	if ((off % blksz) != 0) {
1491 		zfs_exit(zfsvfs, FTAG);
1492 		return (SET_ERROR(EINVAL));
1493 	}
1494 
1495 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(zfsvfs), NULL, &mtime, 16);
1496 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, 16);
1497 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_SIZE(zfsvfs), NULL,
1498 	    &zp->z_size, 8);
1499 
1500 	/*
1501 	 * Start a transaction.
1502 	 */
1503 	tx = dmu_tx_create(zfsvfs->z_os);
1504 
1505 	dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_FALSE);
1506 	db = (dmu_buf_impl_t *)sa_get_db(zp->z_sa_hdl);
1507 	DB_DNODE_ENTER(db);
1508 	dmu_tx_hold_clone_by_dnode(tx, DB_DNODE(db), off, len);
1509 	DB_DNODE_EXIT(db);
1510 	zfs_sa_upgrade_txholds(tx, zp);
1511 	error = dmu_tx_assign(tx, TXG_WAIT);
1512 	if (error != 0) {
1513 		dmu_tx_abort(tx);
1514 		zfs_exit(zfsvfs, FTAG);
1515 		return (error);
1516 	}
1517 
1518 	if (zp->z_blksz < blksz)
1519 		zfs_grow_blocksize(zp, blksz, tx);
1520 
1521 	dmu_brt_clone(zfsvfs->z_os, zp->z_id, off, len, tx, bps, nbps);
1522 
1523 	zfs_tstamp_update_setup(zp, CONTENT_MODIFIED, mtime, ctime);
1524 
1525 	if (zp->z_size < off + len)
1526 		zp->z_size = off + len;
1527 
1528 	error = sa_bulk_update(zp->z_sa_hdl, bulk, count, tx);
1529 
1530 	/*
1531 	 * zil_replaying() not only check if we are replaying ZIL, but also
1532 	 * updates the ZIL header to record replay progress.
1533 	 */
1534 	VERIFY(zil_replaying(zfsvfs->z_log, tx));
1535 
1536 	dmu_tx_commit(tx);
1537 
1538 	zfs_znode_update_vfs(zp);
1539 
1540 	zfs_exit(zfsvfs, FTAG);
1541 
1542 	return (error);
1543 }
1544 
1545 EXPORT_SYMBOL(zfs_access);
1546 EXPORT_SYMBOL(zfs_fsync);
1547 EXPORT_SYMBOL(zfs_holey);
1548 EXPORT_SYMBOL(zfs_read);
1549 EXPORT_SYMBOL(zfs_write);
1550 EXPORT_SYMBOL(zfs_getsecattr);
1551 EXPORT_SYMBOL(zfs_setsecattr);
1552 EXPORT_SYMBOL(zfs_clone_range);
1553 EXPORT_SYMBOL(zfs_clone_range_replay);
1554 
1555 ZFS_MODULE_PARAM(zfs_vnops, zfs_vnops_, read_chunk_size, U64, ZMOD_RW,
1556 	"Bytes to read per chunk");
1557 
1558 ZFS_MODULE_PARAM(zfs, zfs_, bclone_enabled, INT, ZMOD_RW,
1559 	"Enable block cloning");
1560 
1561 ZFS_MODULE_PARAM(zfs, zfs_, bclone_wait_dirty, INT, ZMOD_RW,
1562 	"Wait for dirty blocks when cloning");
1563