xref: /freebsd/sys/contrib/openzfs/module/zfs/zfs_vnops.c (revision d0b3ecdc274930e190ea233b6b69ff03782eaf8d)
1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3  * CDDL HEADER START
4  *
5  * The contents of this file are subject to the terms of the
6  * Common Development and Distribution License (the "License").
7  * You may not use this file except in compliance with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or https://opensource.org/licenses/CDDL-1.0.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 
23 /*
24  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
25  * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
26  * Copyright (c) 2015 by Chunwei Chen. All rights reserved.
27  * Copyright 2017 Nexenta Systems, Inc.
28  * Copyright (c) 2021, 2022 by Pawel Jakub Dawidek
29  * Copyright (c) 2025, Rob Norris <robn@despairlabs.com>
30  * Copyright (c) 2025, Klara, Inc.
31  */
32 
33 /* Portions Copyright 2007 Jeremy Teo */
34 /* Portions Copyright 2010 Robert Milkowski */
35 
36 #include <sys/types.h>
37 #include <sys/param.h>
38 #include <sys/time.h>
39 #include <sys/sysmacros.h>
40 #include <sys/vfs.h>
41 #include <sys/file.h>
42 #include <sys/stat.h>
43 #include <sys/kmem.h>
44 #include <sys/cmn_err.h>
45 #include <sys/errno.h>
46 #include <sys/zfs_dir.h>
47 #include <sys/zfs_acl.h>
48 #include <sys/zfs_ioctl.h>
49 #include <sys/fs/zfs.h>
50 #include <sys/dmu.h>
51 #include <sys/dmu_objset.h>
52 #include <sys/dsl_crypt.h>
53 #include <sys/dsl_dataset.h>
54 #include <sys/spa.h>
55 #include <sys/txg.h>
56 #include <sys/brt.h>
57 #include <sys/dbuf.h>
58 #include <sys/policy.h>
59 #include <sys/zfeature.h>
60 #include <sys/zfs_vnops.h>
61 #include <sys/zfs_quota.h>
62 #include <sys/zfs_vfsops.h>
63 #include <sys/zfs_znode.h>
64 
65 /*
66  * Enables access to the block cloning feature. If this setting is 0, then even
67  * if feature@block_cloning is enabled, using functions and system calls that
68  * attempt to clone blocks will act as though the feature is disabled.
69  */
70 int zfs_bclone_enabled = 1;
71 
72 /*
73  * Restricts block cloning between datasets with different properties
74  * (checksum, compression, copies, dedup, or special_small_blocks).
75  */
76 int zfs_bclone_strict_properties = 1;
77 
78 /*
79  * When set to 1 the FICLONE and FICLONERANGE ioctls will wait for any dirty
80  * data to be written to disk before proceeding. This ensures that the clone
81  * operation reliably succeeds, even if a file is modified and then immediately
82  * cloned. Note that for small files this may be slower than simply copying
83  * the file. When set to 0 the clone operation will immediately fail if it
84  * encounters any dirty blocks. By default waiting is enabled.
85  */
86 int zfs_bclone_wait_dirty = 1;
87 
88 /*
89  * Enable Direct I/O. If this setting is 0, then all I/O requests will be
90  * directed through the ARC acting as though the dataset property direct was
91  * set to disabled.
92  *
93  * Disabled by default on FreeBSD until a potential range locking issue in
94  * zfs_getpages() can be resolved.
95  */
96 #ifdef __FreeBSD__
97 static int zfs_dio_enabled = 0;
98 #else
99 static int zfs_dio_enabled = 1;
100 #endif
101 
102 /*
103  * Strictly enforce alignment for Direct I/O requests, returning EINVAL
104  * if not page-aligned instead of silently falling back to uncached I/O.
105  */
106 static int zfs_dio_strict = 0;
107 
108 
109 /*
110  * Maximum bytes to read per chunk in zfs_read().
111  */
112 #ifdef _ILP32
113 static uint64_t zfs_vnops_read_chunk_size = 1024 * 1024;
114 #else
115 static uint64_t zfs_vnops_read_chunk_size = DMU_MAX_ACCESS / 2;
116 #endif
117 
118 int
zfs_fsync(znode_t * zp,int syncflag,cred_t * cr)119 zfs_fsync(znode_t *zp, int syncflag, cred_t *cr)
120 {
121 	int error = 0;
122 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
123 
124 	if (zfsvfs->z_os->os_sync != ZFS_SYNC_DISABLED) {
125 		if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0)
126 			return (error);
127 		error = zil_commit(zfsvfs->z_log, zp->z_id);
128 		zfs_exit(zfsvfs, FTAG);
129 	}
130 	return (error);
131 }
132 
133 
134 #if defined(SEEK_HOLE) && defined(SEEK_DATA)
135 /*
136  * Lseek support for finding holes (cmd == SEEK_HOLE) and
137  * data (cmd == SEEK_DATA). "off" is an in/out parameter.
138  */
139 static int
zfs_holey_common(znode_t * zp,ulong_t cmd,loff_t * off)140 zfs_holey_common(znode_t *zp, ulong_t cmd, loff_t *off)
141 {
142 	zfs_locked_range_t *lr;
143 	uint64_t noff = (uint64_t)*off; /* new offset */
144 	uint64_t file_sz;
145 	int error;
146 	boolean_t hole;
147 
148 	file_sz = zp->z_size;
149 	if (noff >= file_sz)  {
150 		return (SET_ERROR(ENXIO));
151 	}
152 
153 	if (cmd == F_SEEK_HOLE)
154 		hole = B_TRUE;
155 	else
156 		hole = B_FALSE;
157 
158 	/* Flush any mmap()'d data to disk */
159 	if (zn_has_cached_data(zp, 0, file_sz - 1))
160 		zn_flush_cached_data(zp, B_TRUE);
161 
162 	lr = zfs_rangelock_enter(&zp->z_rangelock, 0, UINT64_MAX, RL_READER);
163 	error = dmu_offset_next(ZTOZSB(zp)->z_os, zp->z_id, hole, &noff);
164 	zfs_rangelock_exit(lr);
165 
166 	if (error == ESRCH)
167 		return (SET_ERROR(ENXIO));
168 
169 	/* File was dirty, so fall back to using generic logic */
170 	if (error == EBUSY) {
171 		if (hole)
172 			*off = file_sz;
173 
174 		return (0);
175 	}
176 
177 	/*
178 	 * We could find a hole that begins after the logical end-of-file,
179 	 * because dmu_offset_next() only works on whole blocks.  If the
180 	 * EOF falls mid-block, then indicate that the "virtual hole"
181 	 * at the end of the file begins at the logical EOF, rather than
182 	 * at the end of the last block.
183 	 */
184 	if (noff > file_sz) {
185 		ASSERT(hole);
186 		noff = file_sz;
187 	}
188 
189 	if (noff < *off)
190 		return (error);
191 	*off = noff;
192 	return (error);
193 }
194 
195 int
zfs_holey(znode_t * zp,ulong_t cmd,loff_t * off)196 zfs_holey(znode_t *zp, ulong_t cmd, loff_t *off)
197 {
198 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
199 	int error;
200 
201 	if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0)
202 		return (error);
203 
204 	error = zfs_holey_common(zp, cmd, off);
205 
206 	zfs_exit(zfsvfs, FTAG);
207 	return (error);
208 }
209 #endif /* SEEK_HOLE && SEEK_DATA */
210 
211 int
zfs_access(znode_t * zp,int mode,int flag,cred_t * cr)212 zfs_access(znode_t *zp, int mode, int flag, cred_t *cr)
213 {
214 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
215 	int error;
216 
217 	if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0)
218 		return (error);
219 
220 	if (flag & V_ACE_MASK)
221 #if defined(__linux__)
222 		error = zfs_zaccess(zp, mode, flag, B_FALSE, cr,
223 		    zfs_init_idmap);
224 #else
225 		error = zfs_zaccess(zp, mode, flag, B_FALSE, cr,
226 		    NULL);
227 #endif
228 	else
229 #if defined(__linux__)
230 		error = zfs_zaccess_rwx(zp, mode, flag, cr, zfs_init_idmap);
231 #else
232 		error = zfs_zaccess_rwx(zp, mode, flag, cr, NULL);
233 #endif
234 
235 	zfs_exit(zfsvfs, FTAG);
236 	return (error);
237 }
238 
239 /*
240  * Determine if Direct I/O has been requested (either via the O_DIRECT flag or
241  * the "direct" dataset property). When inherited by the property only apply
242  * the O_DIRECT flag to correctly aligned IO requests. The rational for this
243  * is it allows the property to be safely set on a dataset without forcing
244  * all of the applications to be aware of the alignment restrictions. When
245  * O_DIRECT is explicitly requested by an application return EINVAL if the
246  * request is unaligned.  In all cases, if the range for this request has
247  * been mmap'ed then we will perform buffered I/O to keep the mapped region
248  * synhronized with the ARC.
249  *
250  * It is possible that a file's pages could be mmap'ed after it is checked
251  * here. If so, that is handled coorarding in zfs_write(). See comments in the
252  * following area for how this is handled:
253  * zfs_write() -> update_pages()
254  */
255 static int
zfs_setup_direct(struct znode * zp,zfs_uio_t * uio,zfs_uio_rw_t rw,int * ioflagp)256 zfs_setup_direct(struct znode *zp, zfs_uio_t *uio, zfs_uio_rw_t rw,
257     int *ioflagp)
258 {
259 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
260 	objset_t *os = zfsvfs->z_os;
261 	int ioflag = *ioflagp;
262 	int error = 0;
263 
264 	if (os->os_direct == ZFS_DIRECT_ALWAYS) {
265 		/* Force either direct or uncached I/O. */
266 		ioflag |= O_DIRECT;
267 	}
268 
269 	if ((ioflag & O_DIRECT) == 0)
270 		goto out;
271 
272 	if (!zfs_dio_enabled || os->os_direct == ZFS_DIRECT_DISABLED) {
273 		/*
274 		 * Direct I/O is disabled.  The I/O request will be directed
275 		 * through the ARC as uncached I/O.
276 		 */
277 		goto out;
278 	}
279 
280 	if (!zfs_uio_page_aligned(uio) ||
281 	    !zfs_uio_aligned(uio, PAGE_SIZE)) {
282 		/*
283 		 * Misaligned requests can be executed through the ARC as
284 		 * uncached I/O.  But if O_DIRECT was set by user and we
285 		 * were set to be strict, then it is a failure.
286 		 */
287 		if ((*ioflagp & O_DIRECT) && zfs_dio_strict)
288 			error = SET_ERROR(EINVAL);
289 		goto out;
290 	}
291 
292 	if (zn_has_cached_data(zp, zfs_uio_offset(uio),
293 	    zfs_uio_offset(uio) + zfs_uio_resid(uio) - 1)) {
294 		/*
295 		 * The region is mmap'ed.  The I/O request will be directed
296 		 * through the ARC as uncached I/O.
297 		 */
298 		goto out;
299 	}
300 
301 	/*
302 	 * For short writes the page mapping of Direct I/O makes no sense.
303 	 * Direct them through the ARC as uncached I/O.
304 	 */
305 	if (rw == UIO_WRITE && zfs_uio_resid(uio) < zp->z_blksz)
306 		goto out;
307 
308 	error = zfs_uio_get_dio_pages_alloc(uio, rw);
309 	if (error)
310 		goto out;
311 	ASSERT(uio->uio_extflg & UIO_DIRECT);
312 
313 out:
314 	*ioflagp = ioflag;
315 	return (error);
316 }
317 
318 /*
319  * Read bytes from specified file into supplied buffer.
320  *
321  *	IN:	zp	- inode of file to be read from.
322  *		uio	- structure supplying read location, range info,
323  *			  and return buffer.
324  *		ioflag	- O_SYNC flags; used to provide FRSYNC semantics.
325  *			  O_DIRECT flag; used to bypass page cache.
326  *		cr	- credentials of caller.
327  *
328  *	OUT:	uio	- updated offset and range, buffer filled.
329  *
330  *	RETURN:	0 on success, error code on failure.
331  *
332  * Side Effects:
333  *	inode - atime updated if byte count > 0
334  */
335 int
zfs_read(struct znode * zp,zfs_uio_t * uio,int ioflag,cred_t * cr)336 zfs_read(struct znode *zp, zfs_uio_t *uio, int ioflag, cred_t *cr)
337 {
338 	(void) cr;
339 	int error = 0;
340 	boolean_t frsync = B_FALSE;
341 	boolean_t dio_checksum_failure = B_FALSE;
342 
343 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
344 	if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0)
345 		return (error);
346 
347 	if (zp->z_pflags & ZFS_AV_QUARANTINED) {
348 		zfs_exit(zfsvfs, FTAG);
349 		return (SET_ERROR(EACCES));
350 	}
351 
352 	/* We don't copy out anything useful for directories. */
353 	if (Z_ISDIR(ZTOTYPE(zp))) {
354 		zfs_exit(zfsvfs, FTAG);
355 		return (SET_ERROR(EISDIR));
356 	}
357 
358 	/*
359 	 * Validate file offset
360 	 */
361 	if (zfs_uio_offset(uio) < (offset_t)0) {
362 		zfs_exit(zfsvfs, FTAG);
363 		return (SET_ERROR(EINVAL));
364 	}
365 
366 	/*
367 	 * Fasttrack empty reads
368 	 */
369 	if (zfs_uio_resid(uio) == 0) {
370 		zfs_exit(zfsvfs, FTAG);
371 		return (0);
372 	}
373 
374 #ifdef FRSYNC
375 	/*
376 	 * If we're in FRSYNC mode, sync out this znode before reading it.
377 	 * Only do this for non-snapshots.
378 	 *
379 	 * Some platforms do not support FRSYNC and instead map it
380 	 * to O_SYNC, which results in unnecessary calls to zil_commit. We
381 	 * only honor FRSYNC requests on platforms which support it.
382 	 */
383 	frsync = !!(ioflag & FRSYNC);
384 #endif
385 	if (zfsvfs->z_log &&
386 	    (frsync || zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS)) {
387 		error = zil_commit(zfsvfs->z_log, zp->z_id);
388 		if (error != 0) {
389 			zfs_exit(zfsvfs, FTAG);
390 			return (error);
391 		}
392 	}
393 
394 	/*
395 	 * Lock the range against changes.
396 	 */
397 	zfs_locked_range_t *lr = zfs_rangelock_enter(&zp->z_rangelock,
398 	    zfs_uio_offset(uio), zfs_uio_resid(uio), RL_READER);
399 
400 	/*
401 	 * If we are reading past end-of-file we can skip
402 	 * to the end; but we might still need to set atime.
403 	 */
404 	if (zfs_uio_offset(uio) >= zp->z_size) {
405 		error = 0;
406 		goto out;
407 	}
408 	ASSERT(zfs_uio_offset(uio) < zp->z_size);
409 
410 	/*
411 	 * Setting up Direct I/O if requested.
412 	 */
413 	error = zfs_setup_direct(zp, uio, UIO_READ, &ioflag);
414 	if (error) {
415 		goto out;
416 	}
417 
418 #if defined(__linux__)
419 	ssize_t start_offset = zfs_uio_offset(uio);
420 #endif
421 	uint_t blksz = zp->z_blksz;
422 	ssize_t chunk_size;
423 	ssize_t n = MIN(zfs_uio_resid(uio), zp->z_size - zfs_uio_offset(uio));
424 	ssize_t start_resid = n;
425 	ssize_t dio_remaining_resid = 0;
426 
427 	dmu_flags_t dflags = DMU_READ_PREFETCH;
428 	if (ioflag & O_DIRECT)
429 		dflags |= DMU_UNCACHEDIO;
430 	if (uio->uio_extflg & UIO_DIRECT) {
431 		/*
432 		 * All pages for an O_DIRECT request ahve already been mapped
433 		 * so there's no compelling reason to handle this uio in
434 		 * smaller chunks.
435 		 */
436 		chunk_size = DMU_MAX_ACCESS;
437 
438 		/*
439 		 * In the event that the O_DIRECT request is reading the entire
440 		 * file, it is possible file's length is not page sized
441 		 * aligned. However, lower layers expect that the Direct I/O
442 		 * request is page-aligned. In this case, as much of the file
443 		 * that can be read using Direct I/O happens and the remaining
444 		 * amount will be read through the ARC.
445 		 *
446 		 * This is still consistent with the semantics of Direct I/O in
447 		 * ZFS as at a minimum the I/O request must be page-aligned.
448 		 */
449 		dio_remaining_resid = n - P2ALIGN_TYPED(n, PAGE_SIZE, ssize_t);
450 		if (dio_remaining_resid != 0)
451 			n -= dio_remaining_resid;
452 		dflags |= DMU_DIRECTIO;
453 	} else {
454 		chunk_size = MIN(MAX(zfs_vnops_read_chunk_size, blksz),
455 		    DMU_MAX_ACCESS / 2);
456 	}
457 
458 	while (n > 0) {
459 		ssize_t nbytes = MIN(n, chunk_size -
460 		    P2PHASE(zfs_uio_offset(uio), blksz));
461 #ifdef UIO_NOCOPY
462 		if (zfs_uio_segflg(uio) == UIO_NOCOPY)
463 			error = mappedread_sf(zp, nbytes, uio);
464 		else
465 #endif
466 		if (zn_has_cached_data(zp, zfs_uio_offset(uio),
467 		    zfs_uio_offset(uio) + nbytes - 1)) {
468 			error = mappedread(zp, nbytes, uio);
469 		} else {
470 			error = dmu_read_uio_dbuf(sa_get_db(zp->z_sa_hdl),
471 			    uio, nbytes, dflags);
472 		}
473 
474 		if (error) {
475 			/* convert checksum errors into IO errors */
476 			if (error == ECKSUM) {
477 				/*
478 				 * If a Direct I/O read returned a checksum
479 				 * verify error, then it must be treated as
480 				 * suspicious. The contents of the buffer could
481 				 * have beeen manipulated while the I/O was in
482 				 * flight. In this case, the remainder of I/O
483 				 * request will just be reissued through the
484 				 * ARC.
485 				 */
486 				if (uio->uio_extflg & UIO_DIRECT) {
487 					dio_checksum_failure = B_TRUE;
488 					uio->uio_extflg &= ~UIO_DIRECT;
489 					n += dio_remaining_resid;
490 					dio_remaining_resid = 0;
491 					continue;
492 				} else {
493 					error = SET_ERROR(EIO);
494 				}
495 			}
496 
497 #if defined(__linux__)
498 			/*
499 			 * if we actually read some bytes, bubbling EFAULT
500 			 * up to become EAGAIN isn't what we want here...
501 			 *
502 			 * ...on Linux, at least. On FBSD, doing this breaks.
503 			 */
504 			if (error == EFAULT &&
505 			    (zfs_uio_offset(uio) - start_offset) != 0)
506 				error = 0;
507 #endif
508 			break;
509 		}
510 
511 		n -= nbytes;
512 	}
513 
514 	if (error == 0 && (uio->uio_extflg & UIO_DIRECT) &&
515 	    dio_remaining_resid != 0) {
516 		/*
517 		 * Temporarily remove the UIO_DIRECT flag from the UIO so the
518 		 * remainder of the file can be read using the ARC.
519 		 */
520 		uio->uio_extflg &= ~UIO_DIRECT;
521 		dflags &= ~DMU_DIRECTIO;
522 
523 		if (zn_has_cached_data(zp, zfs_uio_offset(uio),
524 		    zfs_uio_offset(uio) + dio_remaining_resid - 1)) {
525 			error = mappedread(zp, dio_remaining_resid, uio);
526 		} else {
527 			error = dmu_read_uio_dbuf(sa_get_db(zp->z_sa_hdl), uio,
528 			    dio_remaining_resid, dflags);
529 		}
530 		uio->uio_extflg |= UIO_DIRECT;
531 		dflags |= DMU_DIRECTIO;
532 
533 		if (error != 0)
534 			n += dio_remaining_resid;
535 	} else if (error && (uio->uio_extflg & UIO_DIRECT)) {
536 		n += dio_remaining_resid;
537 	}
538 	int64_t nread = start_resid - n;
539 
540 	dataset_kstats_update_read_kstats(&zfsvfs->z_kstat, nread);
541 out:
542 	zfs_rangelock_exit(lr);
543 
544 	if (dio_checksum_failure == B_TRUE)
545 		uio->uio_extflg |= UIO_DIRECT;
546 
547 	/*
548 	 * Cleanup for Direct I/O if requested.
549 	 */
550 	if (uio->uio_extflg & UIO_DIRECT)
551 		zfs_uio_free_dio_pages(uio, UIO_READ);
552 
553 	ZFS_ACCESSTIME_STAMP(zfsvfs, zp);
554 	zfs_exit(zfsvfs, FTAG);
555 	return (error);
556 }
557 
558 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)559 zfs_clear_setid_bits_if_necessary(zfsvfs_t *zfsvfs, znode_t *zp, cred_t *cr,
560     uint64_t *clear_setid_bits_txgp, dmu_tx_t *tx)
561 {
562 	zilog_t *zilog = zfsvfs->z_log;
563 	const uint64_t uid = KUID_TO_SUID(ZTOUID(zp));
564 
565 	ASSERT(clear_setid_bits_txgp != NULL);
566 	ASSERT(tx != NULL);
567 
568 	/*
569 	 * Clear Set-UID/Set-GID bits on successful write if not
570 	 * privileged and at least one of the execute bits is set.
571 	 *
572 	 * It would be nice to do this after all writes have
573 	 * been done, but that would still expose the ISUID/ISGID
574 	 * to another app after the partial write is committed.
575 	 *
576 	 * Note: we don't call zfs_fuid_map_id() here because
577 	 * user 0 is not an ephemeral uid.
578 	 */
579 	mutex_enter(&zp->z_acl_lock);
580 	if ((zp->z_mode & (S_IXUSR | (S_IXUSR >> 3) | (S_IXUSR >> 6))) != 0 &&
581 	    (zp->z_mode & (S_ISUID | S_ISGID)) != 0 &&
582 	    secpolicy_vnode_setid_retain(zp, cr,
583 	    ((zp->z_mode & S_ISUID) != 0 && uid == 0)) != 0) {
584 		uint64_t newmode;
585 
586 		zp->z_mode &= ~(S_ISUID | S_ISGID);
587 		newmode = zp->z_mode;
588 		(void) sa_update(zp->z_sa_hdl, SA_ZPL_MODE(zfsvfs),
589 		    (void *)&newmode, sizeof (uint64_t), tx);
590 
591 		mutex_exit(&zp->z_acl_lock);
592 
593 		/*
594 		 * Make sure SUID/SGID bits will be removed when we replay the
595 		 * log. If the setid bits are keep coming back, don't log more
596 		 * than one TX_SETATTR per transaction group.
597 		 */
598 		if (*clear_setid_bits_txgp != dmu_tx_get_txg(tx)) {
599 			vattr_t va = {0};
600 
601 			va.va_mask = ATTR_MODE;
602 			va.va_nodeid = zp->z_id;
603 			va.va_mode = newmode;
604 			zfs_log_setattr(zilog, tx, TX_SETATTR, zp, &va,
605 			    ATTR_MODE, NULL);
606 			*clear_setid_bits_txgp = dmu_tx_get_txg(tx);
607 		}
608 	} else {
609 		mutex_exit(&zp->z_acl_lock);
610 	}
611 }
612 
613 /*
614  * Write the bytes to a file.
615  *
616  *	IN:	zp	- znode of file to be written to.
617  *		uio	- structure supplying write location, range info,
618  *			  and data buffer.
619  *		ioflag	- O_APPEND flag set if in append mode.
620  *			  O_DIRECT flag; used to bypass page cache.
621  *		cr	- credentials of caller.
622  *
623  *	OUT:	uio	- updated offset and range.
624  *
625  *	RETURN:	0 if success
626  *		error code if failure
627  *
628  * Timestamps:
629  *	ip - ctime|mtime updated if byte count > 0
630  */
631 int
zfs_write(znode_t * zp,zfs_uio_t * uio,int ioflag,cred_t * cr)632 zfs_write(znode_t *zp, zfs_uio_t *uio, int ioflag, cred_t *cr)
633 {
634 	int error = 0, error1;
635 	ssize_t start_resid = zfs_uio_resid(uio);
636 	uint64_t clear_setid_bits_txg = 0;
637 	boolean_t o_direct_defer = B_FALSE;
638 
639 	/*
640 	 * Fasttrack empty write
641 	 */
642 	ssize_t n = start_resid;
643 	if (n == 0)
644 		return (0);
645 
646 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
647 	if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0)
648 		return (error);
649 
650 	sa_bulk_attr_t bulk[5];
651 	int count = 0;
652 	uint64_t mtime[2], ctime[2];
653 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(zfsvfs), NULL, &mtime, 16);
654 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, 16);
655 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_SIZE(zfsvfs), NULL,
656 	    &zp->z_size, 8);
657 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_FLAGS(zfsvfs), NULL,
658 	    &zp->z_pflags, 8);
659 	if (zp->z_is_sa)
660 		SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_SEQ(zfsvfs), NULL,
661 		    &zp->z_seq, 8);
662 
663 	/*
664 	 * Callers might not be able to detect properly that we are read-only,
665 	 * so check it explicitly here.
666 	 */
667 	if (zfs_is_readonly(zfsvfs)) {
668 		zfs_exit(zfsvfs, FTAG);
669 		return (SET_ERROR(EROFS));
670 	}
671 
672 	/*
673 	 * If immutable or not appending then return EPERM.
674 	 * Intentionally allow ZFS_READONLY through here.
675 	 * See zfs_zaccess_common()
676 	 */
677 	if ((zp->z_pflags & ZFS_IMMUTABLE) ||
678 	    ((zp->z_pflags & ZFS_APPENDONLY) && !(ioflag & O_APPEND) &&
679 	    (zfs_uio_offset(uio) < zp->z_size))) {
680 		zfs_exit(zfsvfs, FTAG);
681 		return (SET_ERROR(EPERM));
682 	}
683 
684 	/*
685 	 * Validate file offset
686 	 */
687 	offset_t woff = ioflag & O_APPEND ? zp->z_size : zfs_uio_offset(uio);
688 	if (woff < 0) {
689 		zfs_exit(zfsvfs, FTAG);
690 		return (SET_ERROR(EINVAL));
691 	}
692 
693 	/*
694 	 * Setting up Direct I/O if requested.
695 	 */
696 	error = zfs_setup_direct(zp, uio, UIO_WRITE, &ioflag);
697 	if (error) {
698 		zfs_exit(zfsvfs, FTAG);
699 		return (SET_ERROR(error));
700 	}
701 
702 	/*
703 	 * Pre-fault the pages to ensure slow (eg NFS) pages
704 	 * don't hold up txg.
705 	 */
706 	ssize_t pfbytes = MIN(n, DMU_MAX_ACCESS >> 1);
707 	if (zfs_uio_prefaultpages(pfbytes, uio)) {
708 		zfs_exit(zfsvfs, FTAG);
709 		return (SET_ERROR(EFAULT));
710 	}
711 
712 	/*
713 	 * If in append mode, set the io offset pointer to eof.
714 	 */
715 	zfs_locked_range_t *lr;
716 	if (ioflag & O_APPEND) {
717 		/*
718 		 * Obtain an appending range lock to guarantee file append
719 		 * semantics.  We reset the write offset once we have the lock.
720 		 */
721 		lr = zfs_rangelock_enter(&zp->z_rangelock, 0, n, RL_APPEND);
722 		woff = lr->lr_offset;
723 		if (lr->lr_length == UINT64_MAX) {
724 			/*
725 			 * We overlocked the file because this write will cause
726 			 * the file block size to increase.
727 			 * Note that zp_size cannot change with this lock held.
728 			 */
729 			woff = zp->z_size;
730 		}
731 		zfs_uio_setoffset(uio, woff);
732 		/*
733 		 * We need to update the starting offset as well because it is
734 		 * set previously in the ZPL (Linux) and VNOPS (FreeBSD)
735 		 * layers.
736 		 */
737 		zfs_uio_setsoffset(uio, woff);
738 	} else {
739 		/*
740 		 * Note that if the file block size will change as a result of
741 		 * this write, then this range lock will lock the entire file
742 		 * so that we can re-write the block safely.
743 		 */
744 		lr = zfs_rangelock_enter(&zp->z_rangelock, woff, n, RL_WRITER);
745 	}
746 
747 	if (zn_rlimit_fsize_uio(zp, uio)) {
748 		zfs_rangelock_exit(lr);
749 		zfs_exit(zfsvfs, FTAG);
750 		return (SET_ERROR(EFBIG));
751 	}
752 
753 	const rlim64_t limit = MAXOFFSET_T;
754 
755 	if (woff >= limit) {
756 		zfs_rangelock_exit(lr);
757 		zfs_exit(zfsvfs, FTAG);
758 		return (SET_ERROR(EFBIG));
759 	}
760 
761 	if (n > limit - woff)
762 		n = limit - woff;
763 
764 	uint64_t end_size = MAX(zp->z_size, woff + n);
765 	zilog_t *zilog = zfsvfs->z_log;
766 	boolean_t commit = (ioflag & (O_SYNC | O_DSYNC)) ||
767 	    (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS);
768 
769 	const uint64_t uid = KUID_TO_SUID(ZTOUID(zp));
770 	const uint64_t gid = KGID_TO_SGID(ZTOGID(zp));
771 	const uint64_t projid = zp->z_projid;
772 
773 	/*
774 	 * In the event we are increasing the file block size
775 	 * (lr_length == UINT64_MAX), we will direct the write to the ARC.
776 	 * Because zfs_grow_blocksize() will read from the ARC in order to
777 	 * grow the dbuf, we avoid doing Direct I/O here as that would cause
778 	 * data written to disk to be overwritten by data in the ARC during
779 	 * the sync phase. Besides writing data twice to disk, we also
780 	 * want to avoid consistency concerns between data in the the ARC and
781 	 * on disk while growing the file's blocksize.
782 	 *
783 	 * We will only temporarily remove Direct I/O and put it back after
784 	 * we have grown the blocksize. We do this in the event a request
785 	 * is larger than max_blksz, so further requests to
786 	 * dmu_write_uio_dbuf() will still issue the requests using Direct
787 	 * IO.
788 	 *
789 	 * As an example:
790 	 * The first block to file is being written as a 4k request with
791 	 * a recorsize of 1K. The first 1K issued in the loop below will go
792 	 * through the ARC; however, the following 3 1K requests will
793 	 * use Direct I/O.
794 	 */
795 	if (uio->uio_extflg & UIO_DIRECT && lr->lr_length == UINT64_MAX) {
796 		uio->uio_extflg &= ~UIO_DIRECT;
797 		o_direct_defer = B_TRUE;
798 	}
799 
800 	/*
801 	 * Write the file in reasonable size chunks.  Each chunk is written
802 	 * in a separate transaction; this keeps the intent log records small
803 	 * and allows us to do more fine-grained space accounting.
804 	 */
805 	while (n > 0) {
806 		woff = zfs_uio_offset(uio);
807 
808 		if (zfs_id_overblockquota(zfsvfs, DMU_USERUSED_OBJECT, uid) ||
809 		    zfs_id_overblockquota(zfsvfs, DMU_GROUPUSED_OBJECT, gid) ||
810 		    (projid != ZFS_DEFAULT_PROJID &&
811 		    zfs_id_overblockquota(zfsvfs, DMU_PROJECTUSED_OBJECT,
812 		    projid))) {
813 			error = SET_ERROR(EDQUOT);
814 			break;
815 		}
816 
817 		uint64_t blksz;
818 		if (lr->lr_length == UINT64_MAX && zp->z_size <= zp->z_blksz) {
819 			if (zp->z_blksz > zfsvfs->z_max_blksz &&
820 			    !ISP2(zp->z_blksz)) {
821 				/*
822 				 * File's blocksize is already larger than the
823 				 * "recordsize" property.  Only let it grow to
824 				 * the next power of 2.
825 				 */
826 				blksz = 1 << highbit64(zp->z_blksz);
827 			} else {
828 				blksz = zfsvfs->z_max_blksz;
829 			}
830 			blksz = MIN(blksz, P2ROUNDUP(end_size,
831 			    SPA_MINBLOCKSIZE));
832 			blksz = MAX(blksz, zp->z_blksz);
833 		} else {
834 			blksz = zp->z_blksz;
835 		}
836 
837 		arc_buf_t *abuf = NULL;
838 		ssize_t nbytes = n;
839 		if (n >= blksz && woff >= zp->z_size &&
840 		    P2PHASE(woff, blksz) == 0 &&
841 		    !(uio->uio_extflg & UIO_DIRECT) &&
842 		    (blksz >= SPA_OLD_MAXBLOCKSIZE || n < 4 * blksz)) {
843 			/*
844 			 * This write covers a full block.  "Borrow" a buffer
845 			 * from the dmu so that we can fill it before we enter
846 			 * a transaction.  This avoids the possibility of
847 			 * holding up the transaction if the data copy hangs
848 			 * up on a pagefault (e.g., from an NFS server mapping).
849 			 */
850 			abuf = dmu_request_arcbuf(sa_get_db(zp->z_sa_hdl),
851 			    blksz);
852 			ASSERT(abuf != NULL);
853 			ASSERT(arc_buf_size(abuf) == blksz);
854 			if ((error = zfs_uiocopy(abuf->b_data, blksz,
855 			    UIO_WRITE, uio, &nbytes))) {
856 				dmu_return_arcbuf(abuf);
857 				break;
858 			}
859 			ASSERT3S(nbytes, ==, blksz);
860 		} else {
861 			nbytes = MIN(n, (DMU_MAX_ACCESS >> 1) -
862 			    P2PHASE(woff, blksz));
863 			if (pfbytes < nbytes) {
864 				if (zfs_uio_prefaultpages(nbytes, uio)) {
865 					error = SET_ERROR(EFAULT);
866 					break;
867 				}
868 				pfbytes = nbytes;
869 			}
870 		}
871 
872 		/*
873 		 * Start a transaction.
874 		 */
875 		dmu_tx_t *tx = dmu_tx_create(zfsvfs->z_os);
876 		dmu_tx_hold_sa(tx, zp->z_sa_hdl, ZFS_SEQ_MAY_GROW(zp));
877 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)sa_get_db(zp->z_sa_hdl);
878 		DB_DNODE_ENTER(db);
879 		dmu_tx_hold_write_by_dnode(tx, DB_DNODE(db), woff, nbytes);
880 		DB_DNODE_EXIT(db);
881 		zfs_sa_upgrade_txholds(tx, zp);
882 		error = dmu_tx_assign(tx, DMU_TX_WAIT);
883 		if (error) {
884 			dmu_tx_abort(tx);
885 			if (abuf != NULL)
886 				dmu_return_arcbuf(abuf);
887 			break;
888 		}
889 
890 		/*
891 		 * NB: We must call zfs_clear_setid_bits_if_necessary before
892 		 * committing the transaction!
893 		 */
894 
895 		/*
896 		 * If rangelock_enter() over-locked we grow the blocksize
897 		 * and then reduce the lock range.  This will only happen
898 		 * on the first iteration since rangelock_reduce() will
899 		 * shrink down lr_length to the appropriate size.
900 		 */
901 		if (lr->lr_length == UINT64_MAX) {
902 			zfs_grow_blocksize(zp, blksz, tx);
903 			zfs_rangelock_reduce(lr, woff, n);
904 		}
905 
906 		dmu_flags_t dflags = DMU_READ_PREFETCH;
907 		if (ioflag & O_DIRECT)
908 			dflags |= DMU_UNCACHEDIO;
909 		if (uio->uio_extflg & UIO_DIRECT)
910 			dflags |= DMU_DIRECTIO;
911 
912 		ssize_t tx_bytes;
913 		if (abuf == NULL) {
914 			tx_bytes = zfs_uio_resid(uio);
915 			zfs_uio_fault_disable(uio, B_TRUE);
916 			error = dmu_write_uio_dbuf(sa_get_db(zp->z_sa_hdl),
917 			    uio, nbytes, tx, dflags);
918 			zfs_uio_fault_disable(uio, B_FALSE);
919 #ifdef __linux__
920 			if (error == EFAULT) {
921 				zfs_clear_setid_bits_if_necessary(zfsvfs, zp,
922 				    cr, &clear_setid_bits_txg, tx);
923 				dmu_tx_commit(tx);
924 				/*
925 				 * Account for partial writes before
926 				 * continuing the loop.
927 				 * Update needs to occur before the next
928 				 * zfs_uio_prefaultpages, or prefaultpages may
929 				 * error, and we may break the loop early.
930 				 */
931 				n -= tx_bytes - zfs_uio_resid(uio);
932 				pfbytes -= tx_bytes - zfs_uio_resid(uio);
933 				continue;
934 			}
935 #endif
936 			/*
937 			 * On FreeBSD, EFAULT should be propagated back to the
938 			 * VFS, which will handle faulting and will retry.
939 			 */
940 			if (error != 0 && error != EFAULT) {
941 				zfs_clear_setid_bits_if_necessary(zfsvfs, zp,
942 				    cr, &clear_setid_bits_txg, tx);
943 				dmu_tx_commit(tx);
944 				break;
945 			}
946 			tx_bytes -= zfs_uio_resid(uio);
947 		} else {
948 			/*
949 			 * Thus, we're writing a full block at a block-aligned
950 			 * offset and extending the file past EOF.
951 			 *
952 			 * dmu_assign_arcbuf_by_dbuf() will directly assign the
953 			 * arc buffer to a dbuf.
954 			 */
955 			error = dmu_assign_arcbuf_by_dbuf(
956 			    sa_get_db(zp->z_sa_hdl), woff, abuf, tx, dflags);
957 			if (error != 0) {
958 				/*
959 				 * XXX This might not be necessary if
960 				 * dmu_assign_arcbuf_by_dbuf is guaranteed
961 				 * to be atomic.
962 				 */
963 				zfs_clear_setid_bits_if_necessary(zfsvfs, zp,
964 				    cr, &clear_setid_bits_txg, tx);
965 				dmu_return_arcbuf(abuf);
966 				dmu_tx_commit(tx);
967 				break;
968 			}
969 			ASSERT3S(nbytes, <=, zfs_uio_resid(uio));
970 			zfs_uioskip(uio, nbytes);
971 			tx_bytes = nbytes;
972 		}
973 		/*
974 		 * There is a window where a file's pages can be mmap'ed after
975 		 * zfs_setup_direct() is called. This is due to the fact that
976 		 * the rangelock in this function is acquired after calling
977 		 * zfs_setup_direct(). This is done so that
978 		 * zfs_uio_prefaultpages() does not attempt to fault in pages
979 		 * on Linux for Direct I/O requests. This is not necessary as
980 		 * the pages are pinned in memory and can not be faulted out.
981 		 * Ideally, the rangelock would be held before calling
982 		 * zfs_setup_direct() and zfs_uio_prefaultpages(); however,
983 		 * this can lead to a deadlock as zfs_getpage() also acquires
984 		 * the rangelock as a RL_WRITER and prefaulting the pages can
985 		 * lead to zfs_getpage() being called.
986 		 *
987 		 * In the case of the pages being mapped after
988 		 * zfs_setup_direct() is called, the call to update_pages()
989 		 * will still be made to make sure there is consistency between
990 		 * the ARC and the Linux page cache. This is an ufortunate
991 		 * situation as the data will be read back into the ARC after
992 		 * the Direct I/O write has completed, but this is the penality
993 		 * for writing to a mmap'ed region of a file using Direct I/O.
994 		 */
995 		if (tx_bytes &&
996 		    zn_has_cached_data(zp, woff, woff + tx_bytes - 1)) {
997 			update_pages(zp, woff, tx_bytes, zfsvfs->z_os);
998 		}
999 
1000 		/*
1001 		 * If we made no progress, we're done.  If we made even
1002 		 * partial progress, update the znode and ZIL accordingly.
1003 		 */
1004 		if (tx_bytes == 0) {
1005 			(void) sa_update(zp->z_sa_hdl, SA_ZPL_SIZE(zfsvfs),
1006 			    (void *)&zp->z_size, sizeof (uint64_t), tx);
1007 			dmu_tx_commit(tx);
1008 			ASSERT(error != 0);
1009 			break;
1010 		}
1011 
1012 		zfs_clear_setid_bits_if_necessary(zfsvfs, zp, cr,
1013 		    &clear_setid_bits_txg, tx);
1014 
1015 		zfs_tstamp_update_setup(zp, CONTENT_MODIFIED, mtime, ctime);
1016 		if (zp->z_is_sa)
1017 			zp->z_has_seq = B_TRUE;
1018 
1019 		/*
1020 		 * Update the file size (zp_size) if it has changed;
1021 		 * account for possible concurrent updates.
1022 		 */
1023 		while ((end_size = zp->z_size) < zfs_uio_offset(uio)) {
1024 			(void) atomic_cas_64(&zp->z_size, end_size,
1025 			    zfs_uio_offset(uio));
1026 			ASSERT(error == 0 || error == EFAULT);
1027 		}
1028 		/*
1029 		 * If we are replaying and eof is non zero then force
1030 		 * the file size to the specified eof. Note, there's no
1031 		 * concurrency during replay.
1032 		 */
1033 		if (zfsvfs->z_replay && zfsvfs->z_replay_eof != 0)
1034 			zp->z_size = zfsvfs->z_replay_eof;
1035 
1036 		ASSERT3S(count, <=, ARRAY_SIZE(bulk));
1037 		error1 = sa_bulk_update(zp->z_sa_hdl, bulk, count, tx);
1038 		if (error1 != 0)
1039 			/* Avoid clobbering EFAULT. */
1040 			error = error1;
1041 
1042 		/*
1043 		 * NB: During replay, the TX_SETATTR record logged by
1044 		 * zfs_clear_setid_bits_if_necessary must precede any of
1045 		 * the TX_WRITE records logged here.
1046 		 */
1047 		zfs_log_write(zilog, tx, TX_WRITE, zp, woff, tx_bytes, commit,
1048 		    uio->uio_extflg & UIO_DIRECT ? B_TRUE : B_FALSE, NULL,
1049 		    NULL);
1050 
1051 		dmu_tx_commit(tx);
1052 
1053 		/*
1054 		 * Direct I/O was deferred in order to grow the first block.
1055 		 * At this point it can be re-enabled for subsequent writes.
1056 		 */
1057 		if (o_direct_defer) {
1058 			ASSERT(ioflag & O_DIRECT);
1059 			uio->uio_extflg |= UIO_DIRECT;
1060 			o_direct_defer = B_FALSE;
1061 		}
1062 
1063 		if (error != 0)
1064 			break;
1065 		ASSERT3S(tx_bytes, ==, nbytes);
1066 		n -= nbytes;
1067 		pfbytes -= nbytes;
1068 	}
1069 
1070 	if (o_direct_defer) {
1071 		ASSERT(ioflag & O_DIRECT);
1072 		uio->uio_extflg |= UIO_DIRECT;
1073 		o_direct_defer = B_FALSE;
1074 	}
1075 
1076 	zfs_znode_update_vfs(zp);
1077 	zfs_rangelock_exit(lr);
1078 
1079 	/*
1080 	 * Cleanup for Direct I/O if requested.
1081 	 */
1082 	if (uio->uio_extflg & UIO_DIRECT)
1083 		zfs_uio_free_dio_pages(uio, UIO_WRITE);
1084 
1085 	/*
1086 	 * If we're in replay mode, or we made no progress, or the
1087 	 * uio data is inaccessible return an error.  Otherwise, it's
1088 	 * at least a partial write, so it's successful.
1089 	 */
1090 	if (zfsvfs->z_replay || zfs_uio_resid(uio) == start_resid ||
1091 	    error == EFAULT) {
1092 		zfs_exit(zfsvfs, FTAG);
1093 		return (error);
1094 	}
1095 
1096 	if (commit) {
1097 		error = zil_commit(zilog, zp->z_id);
1098 		if (error != 0) {
1099 			zfs_exit(zfsvfs, FTAG);
1100 			return (error);
1101 		}
1102 	}
1103 
1104 	int64_t nwritten = start_resid - zfs_uio_resid(uio);
1105 	dataset_kstats_update_write_kstats(&zfsvfs->z_kstat, nwritten);
1106 
1107 	zfs_exit(zfsvfs, FTAG);
1108 	return (0);
1109 }
1110 
1111 /*
1112  * Check if a block should be skipped during rewrite.
1113  * Returns B_TRUE if block should be skipped.
1114  */
1115 static boolean_t
zfs_rewrite_skip(dmu_buf_t * db,objset_t * os,uint64_t flags)1116 zfs_rewrite_skip(dmu_buf_t *db, objset_t *os, uint64_t flags)
1117 {
1118 	/*
1119 	 * This may be slightly stale and racy, but should be OK for
1120 	 * the advisory use.
1121 	 */
1122 	blkptr_t *bp = dmu_buf_get_blkptr(db);
1123 	if (bp == NULL)
1124 		return (B_TRUE);
1125 
1126 	if (flags & ZFS_REWRITE_SKIP_SNAPSHOT) {
1127 		if (dmu_objset_block_is_shared(os, bp))
1128 			return (B_TRUE);
1129 	}
1130 
1131 	if (flags & ZFS_REWRITE_SKIP_BRT) {
1132 		if (brt_maybe_exists(os->os_spa, bp))
1133 			return (B_TRUE);
1134 	}
1135 
1136 	return (B_FALSE);
1137 }
1138 
1139 /*
1140  * Rewrite a range of file as-is without modification.
1141  *
1142  *	IN:	zp	- znode of file to be rewritten.
1143  *		off	- Offset of the range to rewrite.
1144  *		len	- Length of the range to rewrite.
1145  *		flags	- Random rewrite parameters.
1146  *		arg	- flags-specific argument.
1147  *
1148  *	RETURN:	0 if success
1149  *		error code if failure
1150  */
1151 int
zfs_rewrite(znode_t * zp,uint64_t off,uint64_t len,uint64_t flags,uint64_t arg)1152 zfs_rewrite(znode_t *zp, uint64_t off, uint64_t len, uint64_t flags,
1153     uint64_t arg)
1154 {
1155 	int error;
1156 
1157 #define	ZFS_REWRITE_VALID_FLAGS \
1158 	(ZFS_REWRITE_PHYSICAL | ZFS_REWRITE_SKIP_SNAPSHOT | \
1159 	ZFS_REWRITE_SKIP_BRT)
1160 
1161 	if ((flags & ~ZFS_REWRITE_VALID_FLAGS) != 0 || arg != 0)
1162 		return (SET_ERROR(EINVAL));
1163 
1164 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
1165 	if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0)
1166 		return (error);
1167 
1168 	/* Check if physical rewrite is allowed */
1169 	spa_t *spa = zfsvfs->z_os->os_spa;
1170 	if ((flags & ZFS_REWRITE_PHYSICAL) &&
1171 	    !spa_feature_is_enabled(spa, SPA_FEATURE_PHYSICAL_REWRITE)) {
1172 		zfs_exit(zfsvfs, FTAG);
1173 		return (SET_ERROR(ENOTSUP));
1174 	}
1175 
1176 	if (zfs_is_readonly(zfsvfs)) {
1177 		zfs_exit(zfsvfs, FTAG);
1178 		return (SET_ERROR(EROFS));
1179 	}
1180 
1181 	if (off >= zp->z_size) {
1182 		zfs_exit(zfsvfs, FTAG);
1183 		return (0);
1184 	}
1185 	if (len == 0 || len > zp->z_size - off)
1186 		len = zp->z_size - off;
1187 
1188 	/* Flush any mmap()'d data to disk */
1189 	if (zn_has_cached_data(zp, off, off + len - 1))
1190 		zn_flush_cached_data(zp, B_TRUE);
1191 
1192 	zfs_locked_range_t *lr;
1193 	lr = zfs_rangelock_enter(&zp->z_rangelock, off, len, RL_WRITER);
1194 
1195 	const uint64_t uid = KUID_TO_SUID(ZTOUID(zp));
1196 	const uint64_t gid = KGID_TO_SGID(ZTOGID(zp));
1197 	const uint64_t projid = zp->z_projid;
1198 
1199 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)sa_get_db(zp->z_sa_hdl);
1200 	DB_DNODE_ENTER(db);
1201 	dnode_t *dn = DB_DNODE(db);
1202 
1203 	uint64_t n, noff = off, nr = 0, nw = 0;
1204 	while (len > 0) {
1205 		/*
1206 		 * Rewrite only actual data, skipping any holes.  This might
1207 		 * be inaccurate for dirty files, but we don't really care.
1208 		 */
1209 		if (noff == off) {
1210 			/* Find next data in the file. */
1211 			error = dnode_next_offset(dn, 0, &noff, 1, 1, 0);
1212 			if (error || noff >= off + len) {
1213 				if (error == ESRCH)	/* No more data. */
1214 					error = 0;
1215 				break;
1216 			}
1217 			ASSERT3U(noff, >=, off);
1218 			len -= noff - off;
1219 			off = noff;
1220 
1221 			/* Find where the data end. */
1222 			error = dnode_next_offset(dn, DNODE_FIND_HOLE, &noff,
1223 			    1, 1, 0);
1224 			if (error != 0)
1225 				noff = off + len;
1226 		}
1227 		ASSERT3U(noff, >, off);
1228 
1229 		if (zfs_id_overblockquota(zfsvfs, DMU_USERUSED_OBJECT, uid) ||
1230 		    zfs_id_overblockquota(zfsvfs, DMU_GROUPUSED_OBJECT, gid) ||
1231 		    (projid != ZFS_DEFAULT_PROJID &&
1232 		    zfs_id_overblockquota(zfsvfs, DMU_PROJECTUSED_OBJECT,
1233 		    projid))) {
1234 			error = SET_ERROR(EDQUOT);
1235 			break;
1236 		}
1237 
1238 		n = MIN(MIN(len, noff - off),
1239 		    DMU_MAX_ACCESS / 2 - P2PHASE(off, zp->z_blksz));
1240 
1241 		dmu_tx_t *tx = dmu_tx_create(zfsvfs->z_os);
1242 		dmu_tx_hold_write_by_dnode(tx, dn, off, n);
1243 		error = dmu_tx_assign(tx, DMU_TX_WAIT);
1244 		if (error) {
1245 			dmu_tx_abort(tx);
1246 			break;
1247 		}
1248 
1249 		/* Mark all dbufs within range as dirty to trigger rewrite. */
1250 		dmu_buf_t **dbp;
1251 		int numbufs;
1252 		error = dmu_buf_hold_array_by_dnode(dn, off, n, TRUE, FTAG,
1253 		    &numbufs, &dbp, DMU_READ_PREFETCH | DMU_UNCACHEDIO);
1254 		if (error) {
1255 			dmu_tx_commit(tx);
1256 			break;
1257 		}
1258 		for (int i = 0; i < numbufs; i++) {
1259 			nr += dbp[i]->db_size;
1260 			if (dmu_buf_is_dirty(dbp[i], tx))
1261 				continue;
1262 
1263 			if (zfs_rewrite_skip(dbp[i], zfsvfs->z_os, flags))
1264 				continue;
1265 
1266 			nw += dbp[i]->db_size;
1267 			if (flags & ZFS_REWRITE_PHYSICAL)
1268 				dmu_buf_will_rewrite(dbp[i], tx);
1269 			else
1270 				dmu_buf_will_dirty(dbp[i], tx);
1271 		}
1272 		dmu_buf_rele_array(dbp, numbufs, FTAG);
1273 
1274 		dmu_tx_commit(tx);
1275 
1276 		len -= n;
1277 		off += n;
1278 
1279 		if (issig()) {
1280 			error = SET_ERROR(EINTR);
1281 			break;
1282 		}
1283 	}
1284 
1285 	DB_DNODE_EXIT(db);
1286 
1287 	dataset_kstats_update_read_kstats(&zfsvfs->z_kstat, nr);
1288 	dataset_kstats_update_write_kstats(&zfsvfs->z_kstat, nw);
1289 
1290 	zfs_rangelock_exit(lr);
1291 	zfs_exit(zfsvfs, FTAG);
1292 	return (error);
1293 }
1294 
1295 int
zfs_getsecattr(znode_t * zp,vsecattr_t * vsecp,int flag,cred_t * cr)1296 zfs_getsecattr(znode_t *zp, vsecattr_t *vsecp, int flag, cred_t *cr)
1297 {
1298 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
1299 	int error;
1300 	boolean_t skipaclchk = (flag & ATTR_NOACLCHECK) ? B_TRUE : B_FALSE;
1301 
1302 	if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0)
1303 		return (error);
1304 	error = zfs_getacl(zp, vsecp, skipaclchk, cr);
1305 	zfs_exit(zfsvfs, FTAG);
1306 
1307 	return (error);
1308 }
1309 
1310 int
zfs_setsecattr(znode_t * zp,vsecattr_t * vsecp,int flag,cred_t * cr)1311 zfs_setsecattr(znode_t *zp, vsecattr_t *vsecp, int flag, cred_t *cr)
1312 {
1313 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
1314 	int error;
1315 	boolean_t skipaclchk = (flag & ATTR_NOACLCHECK) ? B_TRUE : B_FALSE;
1316 	zilog_t	*zilog;
1317 
1318 	if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0)
1319 		return (error);
1320 	zilog = zfsvfs->z_log;
1321 	error = zfs_setacl(zp, vsecp, skipaclchk, cr);
1322 
1323 	if (error == 0 && zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS)
1324 		error = zil_commit(zilog, 0);
1325 
1326 	zfs_exit(zfsvfs, FTAG);
1327 	return (error);
1328 }
1329 
1330 /*
1331  * Get the optimal alignment to ensure direct IO can be performed without
1332  * incurring any RMW penalty on write. If direct IO is not enabled for this
1333  * file, returns an error.
1334  */
1335 int
zfs_get_direct_alignment(znode_t * zp,uint64_t * alignp)1336 zfs_get_direct_alignment(znode_t *zp, uint64_t *alignp)
1337 {
1338 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
1339 
1340 	if (!zfs_dio_enabled || zfsvfs->z_os->os_direct == ZFS_DIRECT_DISABLED)
1341 		return (SET_ERROR(EOPNOTSUPP));
1342 
1343 	/*
1344 	 * If the file has multiple blocks, then its block size is fixed
1345 	 * forever, and so is the ideal alignment.
1346 	 *
1347 	 * If however it only has a single block, then we want to return the
1348 	 * max block size it could possibly grown to (ie, the dataset
1349 	 * recordsize). We do this so that a program querying alignment
1350 	 * immediately after the file is created gets a value that won't change
1351 	 * once the file has grown into the second block and beyond.
1352 	 *
1353 	 * Because we don't have a count of blocks easily available here, we
1354 	 * check if the apparent file size is smaller than its current block
1355 	 * size (meaning, the file hasn't yet grown into the current block
1356 	 * size) and then, check if the block size is smaller than the dataset
1357 	 * maximum (meaning, if the file grew past the current block size, the
1358 	 * block size could would be increased).
1359 	 */
1360 	if (zp->z_size <= zp->z_blksz && zp->z_blksz < zfsvfs->z_max_blksz)
1361 		*alignp = MAX(zfsvfs->z_max_blksz, PAGE_SIZE);
1362 	else
1363 		*alignp = MAX(zp->z_blksz, PAGE_SIZE);
1364 
1365 	return (0);
1366 }
1367 
1368 #ifdef ZFS_DEBUG
1369 static int zil_fault_io = 0;
1370 #endif
1371 
1372 static void zfs_get_done(zgd_t *zgd, int error);
1373 
1374 /*
1375  * Get data to generate a TX_WRITE intent log record.
1376  */
1377 int
zfs_get_data(void * arg,uint64_t gen,lr_write_t * lr,char * buf,struct lwb * lwb,zio_t * zio)1378 zfs_get_data(void *arg, uint64_t gen, lr_write_t *lr, char *buf,
1379     struct lwb *lwb, zio_t *zio)
1380 {
1381 	zfsvfs_t *zfsvfs = arg;
1382 	objset_t *os = zfsvfs->z_os;
1383 	znode_t *zp;
1384 	uint64_t object = lr->lr_foid;
1385 	uint64_t offset = lr->lr_offset;
1386 	uint64_t size = lr->lr_length;
1387 	zgd_t *zgd;
1388 	int error = 0;
1389 	uint64_t zp_gen;
1390 
1391 	ASSERT3P(lwb, !=, NULL);
1392 	ASSERT3U(size, !=, 0);
1393 
1394 	/*
1395 	 * Nothing to do if the file has been removed
1396 	 */
1397 	if (zfs_zget(zfsvfs, object, &zp) != 0)
1398 		return (SET_ERROR(ENOENT));
1399 	if (zp->z_unlinked) {
1400 		/*
1401 		 * Release the vnode asynchronously as we currently have the
1402 		 * txg stopped from syncing.
1403 		 */
1404 		zfs_zrele_async(zp);
1405 		return (SET_ERROR(ENOENT));
1406 	}
1407 	/* check if generation number matches */
1408 	if (sa_lookup(zp->z_sa_hdl, SA_ZPL_GEN(zfsvfs), &zp_gen,
1409 	    sizeof (zp_gen)) != 0) {
1410 		zfs_zrele_async(zp);
1411 		return (SET_ERROR(EIO));
1412 	}
1413 	if (zp_gen != gen) {
1414 		zfs_zrele_async(zp);
1415 		return (SET_ERROR(ENOENT));
1416 	}
1417 
1418 	zgd = kmem_zalloc(sizeof (zgd_t), KM_SLEEP);
1419 	zgd->zgd_lwb = lwb;
1420 	zgd->zgd_private = zp;
1421 
1422 	/*
1423 	 * Write records come in two flavors: immediate and indirect.
1424 	 * For small writes it's cheaper to store the data with the
1425 	 * log record (immediate); for large writes it's cheaper to
1426 	 * sync the data and get a pointer to it (indirect) so that
1427 	 * we don't have to write the data twice.
1428 	 */
1429 	if (buf != NULL) { /* immediate write */
1430 		zgd->zgd_lr = zfs_rangelock_enter(&zp->z_rangelock, offset,
1431 		    size, RL_READER);
1432 		/* test for truncation needs to be done while range locked */
1433 		if (offset >= zp->z_size) {
1434 			error = SET_ERROR(ENOENT);
1435 		} else {
1436 			error = dmu_read(os, object, offset, size, buf,
1437 			    DMU_READ_NO_PREFETCH | DMU_KEEP_CACHING);
1438 		}
1439 		ASSERT(error == 0 || error == ENOENT);
1440 	} else { /* indirect write */
1441 		ASSERT3P(zio, !=, NULL);
1442 		/*
1443 		 * Have to lock the whole block to ensure when it's
1444 		 * written out and its checksum is being calculated
1445 		 * that no one can change the data. We need to re-check
1446 		 * blocksize after we get the lock in case it's changed!
1447 		 */
1448 		for (;;) {
1449 			uint64_t blkoff;
1450 			size = zp->z_blksz;
1451 			blkoff = ISP2(size) ? P2PHASE(offset, size) : offset;
1452 			offset -= blkoff;
1453 			zgd->zgd_lr = zfs_rangelock_enter(&zp->z_rangelock,
1454 			    offset, size, RL_READER);
1455 			if (zp->z_blksz == size)
1456 				break;
1457 			offset += blkoff;
1458 			zfs_rangelock_exit(zgd->zgd_lr);
1459 		}
1460 		/* test for truncation needs to be done while range locked */
1461 		if (lr->lr_offset >= zp->z_size)
1462 			error = SET_ERROR(ENOENT);
1463 #ifdef ZFS_DEBUG
1464 		if (zil_fault_io) {
1465 			error = SET_ERROR(EIO);
1466 			zil_fault_io = 0;
1467 		}
1468 #endif
1469 
1470 		dmu_buf_t *dbp;
1471 		if (error == 0)
1472 			error = dmu_buf_hold_noread(os, object, offset, zgd,
1473 			    &dbp);
1474 
1475 		if (error == 0) {
1476 			zgd->zgd_db = dbp;
1477 			dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp;
1478 			boolean_t direct_write = B_FALSE;
1479 			mutex_enter(&db->db_mtx);
1480 			dbuf_dirty_record_t *dr =
1481 			    dbuf_find_dirty_eq(db, lr->lr_common.lrc_txg);
1482 			if (dr != NULL && dr->dt.dl.dr_diowrite)
1483 				direct_write = B_TRUE;
1484 			mutex_exit(&db->db_mtx);
1485 
1486 			/*
1487 			 * All Direct I/O writes will have already completed and
1488 			 * the block pointer can be immediately stored in the
1489 			 * log record.
1490 			 */
1491 			if (direct_write) {
1492 				/*
1493 				 * A Direct I/O write always covers an entire
1494 				 * block.
1495 				 */
1496 				ASSERT3U(dbp->db_size, ==, zp->z_blksz);
1497 				lr->lr_blkptr = dr->dt.dl.dr_overridden_by;
1498 				zfs_get_done(zgd, 0);
1499 				return (0);
1500 			}
1501 
1502 			blkptr_t *bp = &lr->lr_blkptr;
1503 			zgd->zgd_bp = bp;
1504 
1505 			ASSERT3U(dbp->db_offset, ==, offset);
1506 			ASSERT3U(dbp->db_size, ==, size);
1507 
1508 			error = dmu_sync(zio, lr->lr_common.lrc_txg,
1509 			    zfs_get_done, zgd);
1510 			ASSERT(error || lr->lr_length <= size);
1511 
1512 			/*
1513 			 * On success, we need to wait for the write I/O
1514 			 * initiated by dmu_sync() to complete before we can
1515 			 * release this dbuf.  We will finish everything up
1516 			 * in the zfs_get_done() callback.
1517 			 */
1518 			if (error == 0)
1519 				return (0);
1520 
1521 			if (error == EALREADY) {
1522 				lr->lr_common.lrc_txtype = TX_WRITE2;
1523 				/*
1524 				 * TX_WRITE2 relies on the data previously
1525 				 * written by the TX_WRITE that caused
1526 				 * EALREADY.  We zero out the BP because
1527 				 * it is the old, currently-on-disk BP.
1528 				 */
1529 				zgd->zgd_bp = NULL;
1530 				BP_ZERO(bp);
1531 				error = 0;
1532 			}
1533 		}
1534 	}
1535 
1536 	zfs_get_done(zgd, error);
1537 
1538 	return (error);
1539 }
1540 
1541 static void
zfs_get_done(zgd_t * zgd,int error)1542 zfs_get_done(zgd_t *zgd, int error)
1543 {
1544 	(void) error;
1545 	znode_t *zp = zgd->zgd_private;
1546 
1547 	if (zgd->zgd_db)
1548 		dmu_buf_rele(zgd->zgd_db, zgd);
1549 
1550 	zfs_rangelock_exit(zgd->zgd_lr);
1551 
1552 	/*
1553 	 * Release the vnode asynchronously as we currently have the
1554 	 * txg stopped from syncing.
1555 	 */
1556 	zfs_zrele_async(zp);
1557 
1558 	kmem_free(zgd, sizeof (zgd_t));
1559 }
1560 
1561 static int
zfs_enter_two(zfsvfs_t * zfsvfs1,zfsvfs_t * zfsvfs2,const char * tag)1562 zfs_enter_two(zfsvfs_t *zfsvfs1, zfsvfs_t *zfsvfs2, const char *tag)
1563 {
1564 	int error;
1565 
1566 	/* Swap. Not sure if the order of zfs_enter()s is important. */
1567 	if (zfsvfs1 > zfsvfs2) {
1568 		zfsvfs_t *tmpzfsvfs;
1569 
1570 		tmpzfsvfs = zfsvfs2;
1571 		zfsvfs2 = zfsvfs1;
1572 		zfsvfs1 = tmpzfsvfs;
1573 	}
1574 
1575 	error = zfs_enter(zfsvfs1, tag);
1576 	if (error != 0)
1577 		return (error);
1578 	if (zfsvfs1 != zfsvfs2) {
1579 		error = zfs_enter(zfsvfs2, tag);
1580 		if (error != 0) {
1581 			zfs_exit(zfsvfs1, tag);
1582 			return (error);
1583 		}
1584 	}
1585 
1586 	return (0);
1587 }
1588 
1589 static void
zfs_exit_two(zfsvfs_t * zfsvfs1,zfsvfs_t * zfsvfs2,const char * tag)1590 zfs_exit_two(zfsvfs_t *zfsvfs1, zfsvfs_t *zfsvfs2, const char *tag)
1591 {
1592 
1593 	zfs_exit(zfsvfs1, tag);
1594 	if (zfsvfs1 != zfsvfs2)
1595 		zfs_exit(zfsvfs2, tag);
1596 }
1597 
1598 /*
1599  * We split each clone request in chunks that can fit into a single ZIL
1600  * log entry. Each ZIL log entry can fit 130816 bytes for a block cloning
1601  * operation (see zil_max_log_data() and zfs_log_clone_range()). This gives
1602  * us room for storing 1022 block pointers.
1603  *
1604  * On success, the function return the number of bytes copied in *lenp.
1605  * Note, it doesn't return how much bytes are left to be copied.
1606  * On errors which are caused by any file system limitations or
1607  * brt limitations `EINVAL` is returned. In the most cases a user
1608  * requested bad parameters, it could be possible to clone the file but
1609  * some parameters don't match the requirements.
1610  */
1611 int
zfs_clone_range(znode_t * inzp,uint64_t * inoffp,znode_t * outzp,uint64_t * outoffp,uint64_t * lenp,cred_t * cr)1612 zfs_clone_range(znode_t *inzp, uint64_t *inoffp, znode_t *outzp,
1613     uint64_t *outoffp, uint64_t *lenp, cred_t *cr)
1614 {
1615 	zfsvfs_t	*inzfsvfs, *outzfsvfs;
1616 	objset_t	*inos, *outos;
1617 	zfs_locked_range_t *inlr, *outlr;
1618 	dmu_buf_impl_t	*db;
1619 	dmu_tx_t	*tx;
1620 	zilog_t		*zilog;
1621 	uint64_t	inoff, outoff, len, done;
1622 	uint64_t	outsize, size;
1623 	int		error;
1624 	int		count = 0;
1625 	sa_bulk_attr_t	bulk[5];
1626 	uint64_t	mtime[2], ctime[2];
1627 	uint64_t	uid, gid, projid;
1628 	blkptr_t	*bps;
1629 	size_t		maxblocks, nbps;
1630 	uint_t		inblksz;
1631 	uint64_t	clear_setid_bits_txg = 0;
1632 	uint64_t	last_synced_txg = 0;
1633 
1634 	inoff = *inoffp;
1635 	outoff = *outoffp;
1636 	len = *lenp;
1637 	done = 0;
1638 
1639 	inzfsvfs = ZTOZSB(inzp);
1640 	outzfsvfs = ZTOZSB(outzp);
1641 
1642 	/*
1643 	 * We need to call zfs_enter() potentially on two different datasets,
1644 	 * so we need a dedicated function for that.
1645 	 */
1646 	error = zfs_enter_two(inzfsvfs, outzfsvfs, FTAG);
1647 	if (error != 0)
1648 		return (error);
1649 
1650 	inos = inzfsvfs->z_os;
1651 	outos = outzfsvfs->z_os;
1652 
1653 	/*
1654 	 * Both source and destination have to belong to the same storage pool.
1655 	 */
1656 	if (dmu_objset_spa(inos) != dmu_objset_spa(outos)) {
1657 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1658 		return (SET_ERROR(EXDEV));
1659 	}
1660 
1661 	/*
1662 	 * outos and inos belongs to the same storage pool.
1663 	 * see a few lines above, only one check.
1664 	 */
1665 	if (!spa_feature_is_enabled(dmu_objset_spa(outos),
1666 	    SPA_FEATURE_BLOCK_CLONING)) {
1667 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1668 		return (SET_ERROR(EOPNOTSUPP));
1669 	}
1670 
1671 	ASSERT(!outzfsvfs->z_replay);
1672 
1673 	/*
1674 	 * Block cloning from an unencrypted dataset into an encrypted
1675 	 * dataset and vice versa is not supported.
1676 	 */
1677 	if (inos->os_encrypted != outos->os_encrypted) {
1678 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1679 		return (SET_ERROR(EXDEV));
1680 	}
1681 
1682 	/*
1683 	 * Cloning across encrypted datasets is possible only if they
1684 	 * share the same master key.
1685 	 */
1686 	if (inos != outos && inos->os_encrypted &&
1687 	    !dmu_objset_crypto_key_equal(inos, outos)) {
1688 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1689 		return (SET_ERROR(EXDEV));
1690 	}
1691 
1692 	/*
1693 	 * Cloning between datasets with different properties is possible,
1694 	 * but it may cause confusions when copying data between them and
1695 	 * expecting new properties to apply.
1696 	 */
1697 	if (zfs_bclone_strict_properties && inos != outos &&
1698 	    !inzfsvfs->z_issnap &&
1699 	    (inos->os_checksum != outos->os_checksum ||
1700 	    inos->os_compress != outos->os_compress ||
1701 	    inos->os_copies != outos->os_copies ||
1702 	    inos->os_dedup_checksum != outos->os_dedup_checksum)) {
1703 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1704 		return (SET_ERROR(EXDEV));
1705 	}
1706 
1707 	error = zfs_verify_zp(inzp);
1708 	if (error == 0)
1709 		error = zfs_verify_zp(outzp);
1710 	if (error != 0) {
1711 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1712 		return (error);
1713 	}
1714 
1715 	/*
1716 	 * We don't copy source file's flags that's why we don't allow to clone
1717 	 * files that are in quarantine.
1718 	 */
1719 	if (inzp->z_pflags & ZFS_AV_QUARANTINED) {
1720 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1721 		return (SET_ERROR(EACCES));
1722 	}
1723 
1724 	if (inoff >= inzp->z_size) {
1725 		*lenp = 0;
1726 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1727 		return (0);
1728 	}
1729 	if (len > inzp->z_size - inoff) {
1730 		len = inzp->z_size - inoff;
1731 	}
1732 	if (len == 0) {
1733 		*lenp = 0;
1734 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1735 		return (0);
1736 	}
1737 
1738 	/*
1739 	 * Callers might not be able to detect properly that we are read-only,
1740 	 * so check it explicitly here.
1741 	 */
1742 	if (zfs_is_readonly(outzfsvfs)) {
1743 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1744 		return (SET_ERROR(EROFS));
1745 	}
1746 
1747 	/*
1748 	 * If immutable or not appending then return EPERM.
1749 	 * Intentionally allow ZFS_READONLY through here.
1750 	 * See zfs_zaccess_common()
1751 	 */
1752 	if ((outzp->z_pflags & ZFS_IMMUTABLE) != 0) {
1753 		zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1754 		return (SET_ERROR(EPERM));
1755 	}
1756 
1757 	/*
1758 	 * No overlapping if we are cloning within the same file.
1759 	 */
1760 	if (inzp == outzp) {
1761 		if (inoff < outoff + len && outoff < inoff + len) {
1762 			zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
1763 			return (SET_ERROR(EINVAL));
1764 		}
1765 	}
1766 
1767 	/* Flush any mmap()'d data to disk */
1768 	if (zn_has_cached_data(inzp, inoff, inoff + len - 1))
1769 		zn_flush_cached_data(inzp, B_TRUE);
1770 
1771 	/*
1772 	 * Maintain predictable lock order.
1773 	 */
1774 	if (inzp < outzp || (inzp == outzp && inoff < outoff)) {
1775 		inlr = zfs_rangelock_enter(&inzp->z_rangelock, inoff, len,
1776 		    RL_READER);
1777 		outlr = zfs_rangelock_enter(&outzp->z_rangelock, outoff, len,
1778 		    RL_WRITER);
1779 	} else {
1780 		outlr = zfs_rangelock_enter(&outzp->z_rangelock, outoff, len,
1781 		    RL_WRITER);
1782 		inlr = zfs_rangelock_enter(&inzp->z_rangelock, inoff, len,
1783 		    RL_READER);
1784 	}
1785 
1786 	inblksz = inzp->z_blksz;
1787 
1788 	/*
1789 	 * Cloning between datasets with different special_small_blocks would
1790 	 * bypass storage tier migration that would occur with a regular copy.
1791 	 */
1792 	if (zfs_bclone_strict_properties && inos != outos &&
1793 	    !inzfsvfs->z_issnap && spa_has_special(dmu_objset_spa(inos))) {
1794 		uint64_t in_smallblk = inos->os_zpl_special_smallblock;
1795 		uint64_t out_smallblk = outos->os_zpl_special_smallblock;
1796 		if (in_smallblk != out_smallblk) {
1797 			uint64_t min_smallblk = MIN(in_smallblk, out_smallblk);
1798 			uint64_t max_smallblk = MAX(in_smallblk, out_smallblk);
1799 			if (min_smallblk < inblksz &&
1800 			    (inos->os_compress != ZIO_COMPRESS_OFF ||
1801 			    max_smallblk >= inblksz)) {
1802 				error = SET_ERROR(EXDEV);
1803 				goto unlock;
1804 			}
1805 		}
1806 	}
1807 
1808 	/*
1809 	 * We cannot clone into a file with different block size if we can't
1810 	 * grow it (block size is already bigger, has more than one block, or
1811 	 * not locked for growth).  There are other possible reasons for the
1812 	 * grow to fail, but we cover what we can before opening transaction
1813 	 * and the rest detect after we try to do it.
1814 	 */
1815 	if (inblksz < outzp->z_blksz) {
1816 		error = SET_ERROR(EINVAL);
1817 		goto unlock;
1818 	}
1819 	if (inblksz != outzp->z_blksz && (outzp->z_size > outzp->z_blksz ||
1820 	    outlr->lr_length != UINT64_MAX)) {
1821 		error = SET_ERROR(EINVAL);
1822 		goto unlock;
1823 	}
1824 
1825 	/*
1826 	 * Block size must be power-of-2 if destination offset != 0.
1827 	 * There can be no multiple blocks of non-power-of-2 size.
1828 	 */
1829 	if (outoff != 0 && !ISP2(inblksz)) {
1830 		error = SET_ERROR(EINVAL);
1831 		goto unlock;
1832 	}
1833 
1834 	/*
1835 	 * Offsets and len must be at block boundries.
1836 	 */
1837 	if ((inoff % inblksz) != 0 || (outoff % inblksz) != 0) {
1838 		error = SET_ERROR(EINVAL);
1839 		goto unlock;
1840 	}
1841 	/*
1842 	 * Length must be multipe of blksz, except for the end of the file.
1843 	 */
1844 	if ((len % inblksz) != 0 &&
1845 	    (len < inzp->z_size - inoff || len < outzp->z_size - outoff)) {
1846 		error = SET_ERROR(EINVAL);
1847 		goto unlock;
1848 	}
1849 
1850 	/*
1851 	 * If we are copying only one block and it is smaller than recordsize
1852 	 * property, do not allow destination to grow beyond one block if it
1853 	 * is not there yet.  Otherwise the destination will get stuck with
1854 	 * that block size forever, that can be as small as 512 bytes, no
1855 	 * matter how big the destination grow later.
1856 	 */
1857 	if (len <= inblksz && inblksz < outzfsvfs->z_max_blksz &&
1858 	    outzp->z_size <= inblksz && outoff + len > inblksz) {
1859 		error = SET_ERROR(EINVAL);
1860 		goto unlock;
1861 	}
1862 
1863 	error = zn_rlimit_fsize(outoff + len);
1864 	if (error != 0) {
1865 		goto unlock;
1866 	}
1867 
1868 	if (inoff >= MAXOFFSET_T || outoff >= MAXOFFSET_T) {
1869 		error = SET_ERROR(EFBIG);
1870 		goto unlock;
1871 	}
1872 
1873 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(outzfsvfs), NULL,
1874 	    &mtime, 16);
1875 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(outzfsvfs), NULL,
1876 	    &ctime, 16);
1877 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_SIZE(outzfsvfs), NULL,
1878 	    &outzp->z_size, 8);
1879 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_FLAGS(outzfsvfs), NULL,
1880 	    &outzp->z_pflags, 8);
1881 	if (outzp->z_is_sa)
1882 		SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_SEQ(outzfsvfs), NULL,
1883 		    &outzp->z_seq, 8);
1884 
1885 	zilog = outzfsvfs->z_log;
1886 	maxblocks = zil_max_log_data(zilog, sizeof (lr_clone_range_t)) /
1887 	    sizeof (bps[0]);
1888 
1889 	uid = KUID_TO_SUID(ZTOUID(outzp));
1890 	gid = KGID_TO_SGID(ZTOGID(outzp));
1891 	projid = outzp->z_projid;
1892 
1893 	bps = vmem_alloc(sizeof (bps[0]) * maxblocks, KM_SLEEP);
1894 
1895 	/*
1896 	 * Clone the file in reasonable size chunks.  Each chunk is cloned
1897 	 * in a separate transaction; this keeps the intent log records small
1898 	 * and allows us to do more fine-grained space accounting.
1899 	 */
1900 	while (len > 0) {
1901 		size = MIN(inblksz * maxblocks, len);
1902 
1903 		if (zfs_id_overblockquota(outzfsvfs, DMU_USERUSED_OBJECT,
1904 		    uid) ||
1905 		    zfs_id_overblockquota(outzfsvfs, DMU_GROUPUSED_OBJECT,
1906 		    gid) ||
1907 		    (projid != ZFS_DEFAULT_PROJID &&
1908 		    zfs_id_overblockquota(outzfsvfs, DMU_PROJECTUSED_OBJECT,
1909 		    projid))) {
1910 			error = SET_ERROR(EDQUOT);
1911 			break;
1912 		}
1913 
1914 		nbps = maxblocks;
1915 		last_synced_txg = spa_last_synced_txg(dmu_objset_spa(inos));
1916 		error = dmu_read_l0_bps(inos, inzp->z_id, inoff, size, bps,
1917 		    &nbps);
1918 		if (error != 0) {
1919 			/*
1920 			 * If we are trying to clone a block that was created
1921 			 * in the current transaction group, the error will be
1922 			 * EAGAIN here.  Based on zfs_bclone_wait_dirty either
1923 			 * return a shortened range to the caller so it can
1924 			 * fallback, or wait for the next TXG and check again.
1925 			 */
1926 			if (error == EAGAIN && zfs_bclone_wait_dirty) {
1927 				txg_wait_flag_t wait_flags =
1928 				    spa_get_failmode(dmu_objset_spa(inos)) ==
1929 				    ZIO_FAILURE_MODE_CONTINUE ?
1930 				    TXG_WAIT_SUSPEND : 0;
1931 				error = txg_wait_synced_flags(
1932 				    dmu_objset_pool(inos), last_synced_txg + 1,
1933 				    wait_flags);
1934 				if (error == 0)
1935 					continue;
1936 				ASSERT3U(error, ==, ESHUTDOWN);
1937 				error = SET_ERROR(EIO);
1938 			}
1939 
1940 			break;
1941 		}
1942 
1943 		/*
1944 		 * Start a transaction.
1945 		 */
1946 		tx = dmu_tx_create(outos);
1947 		dmu_tx_hold_sa(tx, outzp->z_sa_hdl, ZFS_SEQ_MAY_GROW(outzp));
1948 		db = (dmu_buf_impl_t *)sa_get_db(outzp->z_sa_hdl);
1949 		DB_DNODE_ENTER(db);
1950 		dmu_tx_hold_clone_by_dnode(tx, DB_DNODE(db), outoff, size,
1951 		    inblksz);
1952 		DB_DNODE_EXIT(db);
1953 		zfs_sa_upgrade_txholds(tx, outzp);
1954 		error = dmu_tx_assign(tx, DMU_TX_WAIT);
1955 		if (error != 0) {
1956 			dmu_tx_abort(tx);
1957 			break;
1958 		}
1959 
1960 		/*
1961 		 * Copy source znode's block size. This is done only if the
1962 		 * whole znode is locked (see zfs_rangelock_cb()) and only
1963 		 * on the first iteration since zfs_rangelock_reduce() will
1964 		 * shrink down lr_length to the appropriate size.
1965 		 */
1966 		if (outlr->lr_length == UINT64_MAX) {
1967 			zfs_grow_blocksize(outzp, inblksz, tx);
1968 
1969 			/*
1970 			 * Block growth may fail for many reasons we can not
1971 			 * predict here.  If it happen the cloning is doomed.
1972 			 */
1973 			if (inblksz != outzp->z_blksz) {
1974 				error = SET_ERROR(EINVAL);
1975 				dmu_tx_commit(tx);
1976 				break;
1977 			}
1978 
1979 			/*
1980 			 * Round range lock up to the block boundary, so we
1981 			 * prevent appends until we are done.
1982 			 */
1983 			zfs_rangelock_reduce(outlr, outoff,
1984 			    ((len - 1) / inblksz + 1) * inblksz);
1985 		}
1986 
1987 		error = dmu_brt_clone(outos, outzp->z_id, outoff, size, tx,
1988 		    bps, nbps);
1989 		if (error != 0) {
1990 			dmu_tx_commit(tx);
1991 			break;
1992 		}
1993 
1994 		if (zn_has_cached_data(outzp, outoff, outoff + size - 1)) {
1995 			update_pages(outzp, outoff, size, outos);
1996 		}
1997 
1998 		zfs_clear_setid_bits_if_necessary(outzfsvfs, outzp, cr,
1999 		    &clear_setid_bits_txg, tx);
2000 
2001 		zfs_tstamp_update_setup(outzp, CONTENT_MODIFIED, mtime, ctime);
2002 		if (outzp->z_is_sa)
2003 			outzp->z_has_seq = B_TRUE;
2004 
2005 		/*
2006 		 * Update the file size (zp_size) if it has changed;
2007 		 * account for possible concurrent updates.
2008 		 */
2009 		while ((outsize = outzp->z_size) < outoff + size) {
2010 			(void) atomic_cas_64(&outzp->z_size, outsize,
2011 			    outoff + size);
2012 		}
2013 
2014 		ASSERT3S(count, <=, ARRAY_SIZE(bulk));
2015 		error = sa_bulk_update(outzp->z_sa_hdl, bulk, count, tx);
2016 
2017 		zfs_log_clone_range(zilog, tx, TX_CLONE_RANGE, outzp, outoff,
2018 		    size, inblksz, bps, nbps);
2019 
2020 		dmu_tx_commit(tx);
2021 
2022 		if (error != 0)
2023 			break;
2024 
2025 		inoff += size;
2026 		outoff += size;
2027 		len -= size;
2028 		done += size;
2029 
2030 		if (issig()) {
2031 			error = SET_ERROR(EINTR);
2032 			break;
2033 		}
2034 	}
2035 
2036 	vmem_free(bps, sizeof (bps[0]) * maxblocks);
2037 	zfs_znode_update_vfs(outzp);
2038 
2039 unlock:
2040 	zfs_rangelock_exit(outlr);
2041 	zfs_rangelock_exit(inlr);
2042 
2043 	if (done > 0) {
2044 		/*
2045 		 * If we have made at least partial progress, reset the error.
2046 		 */
2047 		error = 0;
2048 
2049 		ZFS_ACCESSTIME_STAMP(inzfsvfs, inzp);
2050 
2051 		if (outos->os_sync == ZFS_SYNC_ALWAYS) {
2052 			error = zil_commit(zilog, outzp->z_id);
2053 		}
2054 
2055 		*inoffp += done;
2056 		*outoffp += done;
2057 		*lenp = done;
2058 	} else {
2059 		/*
2060 		 * If we made no progress, there must be a good reason.
2061 		 * EOF is handled explicitly above, before the loop.
2062 		 */
2063 		ASSERT3S(error, !=, 0);
2064 	}
2065 
2066 	zfs_exit_two(inzfsvfs, outzfsvfs, FTAG);
2067 
2068 	return (error);
2069 }
2070 
2071 /*
2072  * Usual pattern would be to call zfs_clone_range() from zfs_replay_clone(),
2073  * but we cannot do that, because when replaying we don't have source znode
2074  * available. This is why we need a dedicated replay function.
2075  */
2076 int
zfs_clone_range_replay(znode_t * zp,uint64_t off,uint64_t len,uint64_t blksz,const blkptr_t * bps,size_t nbps)2077 zfs_clone_range_replay(znode_t *zp, uint64_t off, uint64_t len, uint64_t blksz,
2078     const blkptr_t *bps, size_t nbps)
2079 {
2080 	zfsvfs_t	*zfsvfs;
2081 	dmu_buf_impl_t	*db;
2082 	dmu_tx_t	*tx;
2083 	int		error;
2084 	int		count = 0;
2085 	sa_bulk_attr_t	bulk[5];
2086 	uint64_t	mtime[2], ctime[2];
2087 
2088 	ASSERT3U(off, <, MAXOFFSET_T);
2089 	ASSERT3U(len, >, 0);
2090 	ASSERT3U(nbps, >, 0);
2091 
2092 	zfsvfs = ZTOZSB(zp);
2093 
2094 	ASSERT(spa_feature_is_enabled(dmu_objset_spa(zfsvfs->z_os),
2095 	    SPA_FEATURE_BLOCK_CLONING));
2096 
2097 	if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0)
2098 		return (error);
2099 
2100 	ASSERT(zfsvfs->z_replay);
2101 	ASSERT(!zfs_is_readonly(zfsvfs));
2102 
2103 	if ((off % blksz) != 0) {
2104 		zfs_exit(zfsvfs, FTAG);
2105 		return (SET_ERROR(EINVAL));
2106 	}
2107 
2108 	/*
2109 	 * Start a transaction.
2110 	 */
2111 	tx = dmu_tx_create(zfsvfs->z_os);
2112 
2113 	dmu_tx_hold_sa(tx, zp->z_sa_hdl, ZFS_SEQ_MAY_GROW(zp));
2114 	db = (dmu_buf_impl_t *)sa_get_db(zp->z_sa_hdl);
2115 	DB_DNODE_ENTER(db);
2116 	dmu_tx_hold_clone_by_dnode(tx, DB_DNODE(db), off, len, blksz);
2117 	DB_DNODE_EXIT(db);
2118 	zfs_sa_upgrade_txholds(tx, zp);
2119 	error = dmu_tx_assign(tx, DMU_TX_WAIT);
2120 	if (error != 0) {
2121 		dmu_tx_abort(tx);
2122 		zfs_exit(zfsvfs, FTAG);
2123 		return (error);
2124 	}
2125 
2126 	if (zp->z_blksz < blksz)
2127 		zfs_grow_blocksize(zp, blksz, tx);
2128 
2129 	dmu_brt_clone(zfsvfs->z_os, zp->z_id, off, len, tx, bps, nbps);
2130 
2131 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(zfsvfs), NULL, &mtime, 16);
2132 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, 16);
2133 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_SIZE(zfsvfs), NULL,
2134 	    &zp->z_size, 8);
2135 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_FLAGS(zfsvfs), NULL,
2136 	    &zp->z_pflags, 8);
2137 	zfs_tstamp_update_setup(zp, CONTENT_MODIFIED, mtime, ctime);
2138 	ZFS_PERSIST_SEQ(zp, bulk, count);
2139 
2140 	if (zp->z_size < off + len)
2141 		zp->z_size = off + len;
2142 
2143 	ASSERT3S(count, <=, ARRAY_SIZE(bulk));
2144 	error = sa_bulk_update(zp->z_sa_hdl, bulk, count, tx);
2145 
2146 	/*
2147 	 * zil_replaying() not only check if we are replaying ZIL, but also
2148 	 * updates the ZIL header to record replay progress.
2149 	 */
2150 	VERIFY(zil_replaying(zfsvfs->z_log, tx));
2151 
2152 	dmu_tx_commit(tx);
2153 
2154 	zfs_znode_update_vfs(zp);
2155 
2156 	zfs_exit(zfsvfs, FTAG);
2157 
2158 	return (error);
2159 }
2160 
2161 EXPORT_SYMBOL(zfs_access);
2162 EXPORT_SYMBOL(zfs_fsync);
2163 EXPORT_SYMBOL(zfs_holey);
2164 EXPORT_SYMBOL(zfs_read);
2165 EXPORT_SYMBOL(zfs_write);
2166 EXPORT_SYMBOL(zfs_getsecattr);
2167 EXPORT_SYMBOL(zfs_setsecattr);
2168 EXPORT_SYMBOL(zfs_clone_range);
2169 EXPORT_SYMBOL(zfs_clone_range_replay);
2170 
2171 ZFS_MODULE_PARAM(zfs_vnops, zfs_vnops_, read_chunk_size, U64, ZMOD_RW,
2172 	"Bytes to read per chunk");
2173 
2174 ZFS_MODULE_PARAM(zfs, zfs_, bclone_enabled, INT, ZMOD_RW,
2175 	"Enable block cloning");
2176 
2177 ZFS_MODULE_PARAM(zfs, zfs_, bclone_strict_properties, INT, ZMOD_RW,
2178 	"Restrict cross-dataset cloning with different properties");
2179 
2180 ZFS_MODULE_PARAM(zfs, zfs_, bclone_wait_dirty, INT, ZMOD_RW,
2181 	"Wait for dirty blocks when cloning");
2182 
2183 ZFS_MODULE_PARAM(zfs, zfs_, dio_enabled, INT, ZMOD_RW,
2184 	"Enable Direct I/O");
2185 
2186 ZFS_MODULE_PARAM(zfs, zfs_, dio_strict, INT, ZMOD_RW,
2187 	"Return errors on misaligned Direct I/O");
2188