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.
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.
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]
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
30 /* Portions Copyright 2007 Jeremy Teo */
31 /* Portions Copyright 2010 Robert Milkowski */
33 #include <sys/types.h>
34 #include <sys/param.h>
36 #include <sys/sysmacros.h>
38 #include <sys/uio_impl.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>
49 #include <sys/dmu_objset.h>
53 #include <sys/policy.h>
54 #include <sys/zfeature.h>
55 #include <sys/zfs_vnops.h>
56 #include <sys/zfs_quota.h>
57 #include <sys/zfs_vfsops.h>
58 #include <sys/zfs_znode.h>
61 static ulong_t zfs_fsync_sync_cnt
= 4;
64 zfs_fsync(znode_t
*zp
, int syncflag
, cred_t
*cr
)
67 zfsvfs_t
*zfsvfs
= ZTOZSB(zp
);
69 (void) tsd_set(zfs_fsyncer_key
, (void *)(uintptr_t)zfs_fsync_sync_cnt
);
71 if (zfsvfs
->z_os
->os_sync
!= ZFS_SYNC_DISABLED
) {
72 if ((error
= zfs_enter_verify_zp(zfsvfs
, zp
, FTAG
)) != 0)
74 atomic_inc_32(&zp
->z_sync_writes_cnt
);
75 zil_commit(zfsvfs
->z_log
, zp
->z_id
);
76 atomic_dec_32(&zp
->z_sync_writes_cnt
);
77 zfs_exit(zfsvfs
, FTAG
);
80 tsd_set(zfs_fsyncer_key
, NULL
);
86 #if defined(SEEK_HOLE) && defined(SEEK_DATA)
88 * Lseek support for finding holes (cmd == SEEK_HOLE) and
89 * data (cmd == SEEK_DATA). "off" is an in/out parameter.
92 zfs_holey_common(znode_t
*zp
, ulong_t cmd
, loff_t
*off
)
94 zfs_locked_range_t
*lr
;
95 uint64_t noff
= (uint64_t)*off
; /* new offset */
100 file_sz
= zp
->z_size
;
101 if (noff
>= file_sz
) {
102 return (SET_ERROR(ENXIO
));
105 if (cmd
== F_SEEK_HOLE
)
110 /* Flush any mmap()'d data to disk */
111 if (zn_has_cached_data(zp
, 0, file_sz
- 1))
112 zn_flush_cached_data(zp
, B_FALSE
);
114 lr
= zfs_rangelock_enter(&zp
->z_rangelock
, 0, UINT64_MAX
, RL_READER
);
115 error
= dmu_offset_next(ZTOZSB(zp
)->z_os
, zp
->z_id
, hole
, &noff
);
116 zfs_rangelock_exit(lr
);
119 return (SET_ERROR(ENXIO
));
121 /* File was dirty, so fall back to using generic logic */
122 if (error
== EBUSY
) {
130 * We could find a hole that begins after the logical end-of-file,
131 * because dmu_offset_next() only works on whole blocks. If the
132 * EOF falls mid-block, then indicate that the "virtual hole"
133 * at the end of the file begins at the logical EOF, rather than
134 * at the end of the last block.
136 if (noff
> file_sz
) {
148 zfs_holey(znode_t
*zp
, ulong_t cmd
, loff_t
*off
)
150 zfsvfs_t
*zfsvfs
= ZTOZSB(zp
);
153 if ((error
= zfs_enter_verify_zp(zfsvfs
, zp
, FTAG
)) != 0)
156 error
= zfs_holey_common(zp
, cmd
, off
);
158 zfs_exit(zfsvfs
, FTAG
);
161 #endif /* SEEK_HOLE && SEEK_DATA */
164 zfs_access(znode_t
*zp
, int mode
, int flag
, cred_t
*cr
)
166 zfsvfs_t
*zfsvfs
= ZTOZSB(zp
);
169 if ((error
= zfs_enter_verify_zp(zfsvfs
, zp
, FTAG
)) != 0)
172 if (flag
& V_ACE_MASK
)
173 #if defined(__linux__)
174 error
= zfs_zaccess(zp
, mode
, flag
, B_FALSE
, cr
,
177 error
= zfs_zaccess(zp
, mode
, flag
, B_FALSE
, cr
,
181 #if defined(__linux__)
182 error
= zfs_zaccess_rwx(zp
, mode
, flag
, cr
, zfs_init_idmap
);
184 error
= zfs_zaccess_rwx(zp
, mode
, flag
, cr
, NULL
);
187 zfs_exit(zfsvfs
, FTAG
);
191 static uint64_t zfs_vnops_read_chunk_size
= 1024 * 1024; /* Tunable */
194 * Read bytes from specified file into supplied buffer.
196 * IN: zp - inode of file to be read from.
197 * uio - structure supplying read location, range info,
199 * ioflag - O_SYNC flags; used to provide FRSYNC semantics.
200 * O_DIRECT flag; used to bypass page cache.
201 * cr - credentials of caller.
203 * OUT: uio - updated offset and range, buffer filled.
205 * RETURN: 0 on success, error code on failure.
208 * inode - atime updated if byte count > 0
211 zfs_read(struct znode
*zp
, zfs_uio_t
*uio
, int ioflag
, cred_t
*cr
)
215 boolean_t frsync
= B_FALSE
;
217 zfsvfs_t
*zfsvfs
= ZTOZSB(zp
);
218 if ((error
= zfs_enter_verify_zp(zfsvfs
, zp
, FTAG
)) != 0)
221 if (zp
->z_pflags
& ZFS_AV_QUARANTINED
) {
222 zfs_exit(zfsvfs
, FTAG
);
223 return (SET_ERROR(EACCES
));
226 /* We don't copy out anything useful for directories. */
227 if (Z_ISDIR(ZTOTYPE(zp
))) {
228 zfs_exit(zfsvfs
, FTAG
);
229 return (SET_ERROR(EISDIR
));
233 * Validate file offset
235 if (zfs_uio_offset(uio
) < (offset_t
)0) {
236 zfs_exit(zfsvfs
, FTAG
);
237 return (SET_ERROR(EINVAL
));
241 * Fasttrack empty reads
243 if (zfs_uio_resid(uio
) == 0) {
244 zfs_exit(zfsvfs
, FTAG
);
250 * If we're in FRSYNC mode, sync out this znode before reading it.
251 * Only do this for non-snapshots.
253 * Some platforms do not support FRSYNC and instead map it
254 * to O_SYNC, which results in unnecessary calls to zil_commit. We
255 * only honor FRSYNC requests on platforms which support it.
257 frsync
= !!(ioflag
& FRSYNC
);
260 (frsync
|| zfsvfs
->z_os
->os_sync
== ZFS_SYNC_ALWAYS
))
261 zil_commit(zfsvfs
->z_log
, zp
->z_id
);
264 * Lock the range against changes.
266 zfs_locked_range_t
*lr
= zfs_rangelock_enter(&zp
->z_rangelock
,
267 zfs_uio_offset(uio
), zfs_uio_resid(uio
), RL_READER
);
270 * If we are reading past end-of-file we can skip
271 * to the end; but we might still need to set atime.
273 if (zfs_uio_offset(uio
) >= zp
->z_size
) {
278 ASSERT(zfs_uio_offset(uio
) < zp
->z_size
);
279 #if defined(__linux__)
280 ssize_t start_offset
= zfs_uio_offset(uio
);
282 ssize_t n
= MIN(zfs_uio_resid(uio
), zp
->z_size
- zfs_uio_offset(uio
));
283 ssize_t start_resid
= n
;
286 ssize_t nbytes
= MIN(n
, zfs_vnops_read_chunk_size
-
287 P2PHASE(zfs_uio_offset(uio
), zfs_vnops_read_chunk_size
));
289 if (zfs_uio_segflg(uio
) == UIO_NOCOPY
)
290 error
= mappedread_sf(zp
, nbytes
, uio
);
293 if (zn_has_cached_data(zp
, zfs_uio_offset(uio
),
294 zfs_uio_offset(uio
) + nbytes
- 1) && !(ioflag
& O_DIRECT
)) {
295 error
= mappedread(zp
, nbytes
, uio
);
297 error
= dmu_read_uio_dbuf(sa_get_db(zp
->z_sa_hdl
),
302 /* convert checksum errors into IO errors */
304 error
= SET_ERROR(EIO
);
306 #if defined(__linux__)
308 * if we actually read some bytes, bubbling EFAULT
309 * up to become EAGAIN isn't what we want here...
311 * ...on Linux, at least. On FBSD, doing this breaks.
313 if (error
== EFAULT
&&
314 (zfs_uio_offset(uio
) - start_offset
) != 0)
323 int64_t nread
= start_resid
- n
;
324 dataset_kstats_update_read_kstats(&zfsvfs
->z_kstat
, nread
);
325 task_io_account_read(nread
);
327 zfs_rangelock_exit(lr
);
329 ZFS_ACCESSTIME_STAMP(zfsvfs
, zp
);
330 zfs_exit(zfsvfs
, FTAG
);
335 zfs_clear_setid_bits_if_necessary(zfsvfs_t
*zfsvfs
, znode_t
*zp
, cred_t
*cr
,
336 uint64_t *clear_setid_bits_txgp
, dmu_tx_t
*tx
)
338 zilog_t
*zilog
= zfsvfs
->z_log
;
339 const uint64_t uid
= KUID_TO_SUID(ZTOUID(zp
));
341 ASSERT(clear_setid_bits_txgp
!= NULL
);
345 * Clear Set-UID/Set-GID bits on successful write if not
346 * privileged and at least one of the execute bits is set.
348 * It would be nice to do this after all writes have
349 * been done, but that would still expose the ISUID/ISGID
350 * to another app after the partial write is committed.
352 * Note: we don't call zfs_fuid_map_id() here because
353 * user 0 is not an ephemeral uid.
355 mutex_enter(&zp
->z_acl_lock
);
356 if ((zp
->z_mode
& (S_IXUSR
| (S_IXUSR
>> 3) | (S_IXUSR
>> 6))) != 0 &&
357 (zp
->z_mode
& (S_ISUID
| S_ISGID
)) != 0 &&
358 secpolicy_vnode_setid_retain(zp
, cr
,
359 ((zp
->z_mode
& S_ISUID
) != 0 && uid
== 0)) != 0) {
362 zp
->z_mode
&= ~(S_ISUID
| S_ISGID
);
363 newmode
= zp
->z_mode
;
364 (void) sa_update(zp
->z_sa_hdl
, SA_ZPL_MODE(zfsvfs
),
365 (void *)&newmode
, sizeof (uint64_t), tx
);
367 mutex_exit(&zp
->z_acl_lock
);
370 * Make sure SUID/SGID bits will be removed when we replay the
371 * log. If the setid bits are keep coming back, don't log more
372 * than one TX_SETATTR per transaction group.
374 if (*clear_setid_bits_txgp
!= dmu_tx_get_txg(tx
)) {
377 va
.va_mask
= ATTR_MODE
;
378 va
.va_nodeid
= zp
->z_id
;
379 va
.va_mode
= newmode
;
380 zfs_log_setattr(zilog
, tx
, TX_SETATTR
, zp
, &va
,
382 *clear_setid_bits_txgp
= dmu_tx_get_txg(tx
);
385 mutex_exit(&zp
->z_acl_lock
);
390 * Write the bytes to a file.
392 * IN: zp - znode of file to be written to.
393 * uio - structure supplying write location, range info,
395 * ioflag - O_APPEND flag set if in append mode.
396 * O_DIRECT flag; used to bypass page cache.
397 * cr - credentials of caller.
399 * OUT: uio - updated offset and range.
401 * RETURN: 0 if success
402 * error code if failure
405 * ip - ctime|mtime updated if byte count > 0
408 zfs_write(znode_t
*zp
, zfs_uio_t
*uio
, int ioflag
, cred_t
*cr
)
410 int error
= 0, error1
;
411 ssize_t start_resid
= zfs_uio_resid(uio
);
412 uint64_t clear_setid_bits_txg
= 0;
415 * Fasttrack empty write
417 ssize_t n
= start_resid
;
421 zfsvfs_t
*zfsvfs
= ZTOZSB(zp
);
422 if ((error
= zfs_enter_verify_zp(zfsvfs
, zp
, FTAG
)) != 0)
425 sa_bulk_attr_t bulk
[4];
427 uint64_t mtime
[2], ctime
[2];
428 SA_ADD_BULK_ATTR(bulk
, count
, SA_ZPL_MTIME(zfsvfs
), NULL
, &mtime
, 16);
429 SA_ADD_BULK_ATTR(bulk
, count
, SA_ZPL_CTIME(zfsvfs
), NULL
, &ctime
, 16);
430 SA_ADD_BULK_ATTR(bulk
, count
, SA_ZPL_SIZE(zfsvfs
), NULL
,
432 SA_ADD_BULK_ATTR(bulk
, count
, SA_ZPL_FLAGS(zfsvfs
), NULL
,
436 * Callers might not be able to detect properly that we are read-only,
437 * so check it explicitly here.
439 if (zfs_is_readonly(zfsvfs
)) {
440 zfs_exit(zfsvfs
, FTAG
);
441 return (SET_ERROR(EROFS
));
445 * If immutable or not appending then return EPERM.
446 * Intentionally allow ZFS_READONLY through here.
447 * See zfs_zaccess_common()
449 if ((zp
->z_pflags
& ZFS_IMMUTABLE
) ||
450 ((zp
->z_pflags
& ZFS_APPENDONLY
) && !(ioflag
& O_APPEND
) &&
451 (zfs_uio_offset(uio
) < zp
->z_size
))) {
452 zfs_exit(zfsvfs
, FTAG
);
453 return (SET_ERROR(EPERM
));
457 * Validate file offset
459 offset_t woff
= ioflag
& O_APPEND
? zp
->z_size
: zfs_uio_offset(uio
);
461 zfs_exit(zfsvfs
, FTAG
);
462 return (SET_ERROR(EINVAL
));
465 const uint64_t max_blksz
= zfsvfs
->z_max_blksz
;
468 * Pre-fault the pages to ensure slow (eg NFS) pages
470 * Skip this if uio contains loaned arc_buf.
472 if (zfs_uio_prefaultpages(MIN(n
, max_blksz
), uio
)) {
473 zfs_exit(zfsvfs
, FTAG
);
474 return (SET_ERROR(EFAULT
));
478 * If in append mode, set the io offset pointer to eof.
480 zfs_locked_range_t
*lr
;
481 if (ioflag
& O_APPEND
) {
483 * Obtain an appending range lock to guarantee file append
484 * semantics. We reset the write offset once we have the lock.
486 lr
= zfs_rangelock_enter(&zp
->z_rangelock
, 0, n
, RL_APPEND
);
487 woff
= lr
->lr_offset
;
488 if (lr
->lr_length
== UINT64_MAX
) {
490 * We overlocked the file because this write will cause
491 * the file block size to increase.
492 * Note that zp_size cannot change with this lock held.
496 zfs_uio_setoffset(uio
, woff
);
499 * Note that if the file block size will change as a result of
500 * this write, then this range lock will lock the entire file
501 * so that we can re-write the block safely.
503 lr
= zfs_rangelock_enter(&zp
->z_rangelock
, woff
, n
, RL_WRITER
);
506 if (zn_rlimit_fsize_uio(zp
, uio
)) {
507 zfs_rangelock_exit(lr
);
508 zfs_exit(zfsvfs
, FTAG
);
509 return (SET_ERROR(EFBIG
));
512 const rlim64_t limit
= MAXOFFSET_T
;
515 zfs_rangelock_exit(lr
);
516 zfs_exit(zfsvfs
, FTAG
);
517 return (SET_ERROR(EFBIG
));
520 if (n
> limit
- woff
)
523 uint64_t end_size
= MAX(zp
->z_size
, woff
+ n
);
524 zilog_t
*zilog
= zfsvfs
->z_log
;
526 const uint64_t uid
= KUID_TO_SUID(ZTOUID(zp
));
527 const uint64_t gid
= KGID_TO_SGID(ZTOGID(zp
));
528 const uint64_t projid
= zp
->z_projid
;
531 * Write the file in reasonable size chunks. Each chunk is written
532 * in a separate transaction; this keeps the intent log records small
533 * and allows us to do more fine-grained space accounting.
536 woff
= zfs_uio_offset(uio
);
538 if (zfs_id_overblockquota(zfsvfs
, DMU_USERUSED_OBJECT
, uid
) ||
539 zfs_id_overblockquota(zfsvfs
, DMU_GROUPUSED_OBJECT
, gid
) ||
540 (projid
!= ZFS_DEFAULT_PROJID
&&
541 zfs_id_overblockquota(zfsvfs
, DMU_PROJECTUSED_OBJECT
,
543 error
= SET_ERROR(EDQUOT
);
547 arc_buf_t
*abuf
= NULL
;
548 if (n
>= max_blksz
&& woff
>= zp
->z_size
&&
549 P2PHASE(woff
, max_blksz
) == 0 &&
550 zp
->z_blksz
== max_blksz
) {
552 * This write covers a full block. "Borrow" a buffer
553 * from the dmu so that we can fill it before we enter
554 * a transaction. This avoids the possibility of
555 * holding up the transaction if the data copy hangs
556 * up on a pagefault (e.g., from an NFS server mapping).
560 abuf
= dmu_request_arcbuf(sa_get_db(zp
->z_sa_hdl
),
562 ASSERT(abuf
!= NULL
);
563 ASSERT(arc_buf_size(abuf
) == max_blksz
);
564 if ((error
= zfs_uiocopy(abuf
->b_data
, max_blksz
,
565 UIO_WRITE
, uio
, &cbytes
))) {
566 dmu_return_arcbuf(abuf
);
569 ASSERT3S(cbytes
, ==, max_blksz
);
573 * Start a transaction.
575 dmu_tx_t
*tx
= dmu_tx_create(zfsvfs
->z_os
);
576 dmu_tx_hold_sa(tx
, zp
->z_sa_hdl
, B_FALSE
);
577 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)sa_get_db(zp
->z_sa_hdl
);
579 dmu_tx_hold_write_by_dnode(tx
, DB_DNODE(db
), woff
,
582 zfs_sa_upgrade_txholds(tx
, zp
);
583 error
= dmu_tx_assign(tx
, TXG_WAIT
);
587 dmu_return_arcbuf(abuf
);
592 * NB: We must call zfs_clear_setid_bits_if_necessary before
593 * committing the transaction!
597 * If rangelock_enter() over-locked we grow the blocksize
598 * and then reduce the lock range. This will only happen
599 * on the first iteration since rangelock_reduce() will
600 * shrink down lr_length to the appropriate size.
602 if (lr
->lr_length
== UINT64_MAX
) {
605 if (zp
->z_blksz
> max_blksz
) {
607 * File's blocksize is already larger than the
608 * "recordsize" property. Only let it grow to
609 * the next power of 2.
611 ASSERT(!ISP2(zp
->z_blksz
));
612 new_blksz
= MIN(end_size
,
613 1 << highbit64(zp
->z_blksz
));
615 new_blksz
= MIN(end_size
, max_blksz
);
617 zfs_grow_blocksize(zp
, new_blksz
, tx
);
618 zfs_rangelock_reduce(lr
, woff
, n
);
622 * XXX - should we really limit each write to z_max_blksz?
623 * Perhaps we should use SPA_MAXBLOCKSIZE chunks?
625 const ssize_t nbytes
=
626 MIN(n
, max_blksz
- P2PHASE(woff
, max_blksz
));
630 tx_bytes
= zfs_uio_resid(uio
);
631 zfs_uio_fault_disable(uio
, B_TRUE
);
632 error
= dmu_write_uio_dbuf(sa_get_db(zp
->z_sa_hdl
),
634 zfs_uio_fault_disable(uio
, B_FALSE
);
636 if (error
== EFAULT
) {
637 zfs_clear_setid_bits_if_necessary(zfsvfs
, zp
,
638 cr
, &clear_setid_bits_txg
, tx
);
641 * Account for partial writes before
642 * continuing the loop.
643 * Update needs to occur before the next
644 * zfs_uio_prefaultpages, or prefaultpages may
645 * error, and we may break the loop early.
647 if (tx_bytes
!= zfs_uio_resid(uio
))
648 n
-= tx_bytes
- zfs_uio_resid(uio
);
649 if (zfs_uio_prefaultpages(MIN(n
, max_blksz
),
657 * On FreeBSD, EFAULT should be propagated back to the
658 * VFS, which will handle faulting and will retry.
660 if (error
!= 0 && error
!= EFAULT
) {
661 zfs_clear_setid_bits_if_necessary(zfsvfs
, zp
,
662 cr
, &clear_setid_bits_txg
, tx
);
666 tx_bytes
-= zfs_uio_resid(uio
);
668 /* Implied by abuf != NULL: */
669 ASSERT3S(n
, >=, max_blksz
);
670 ASSERT0(P2PHASE(woff
, max_blksz
));
672 * We can simplify nbytes to MIN(n, max_blksz) since
673 * P2PHASE(woff, max_blksz) is 0, and knowing
674 * n >= max_blksz lets us simplify further:
676 ASSERT3S(nbytes
, ==, max_blksz
);
678 * Thus, we're writing a full block at a block-aligned
679 * offset and extending the file past EOF.
681 * dmu_assign_arcbuf_by_dbuf() will directly assign the
682 * arc buffer to a dbuf.
684 error
= dmu_assign_arcbuf_by_dbuf(
685 sa_get_db(zp
->z_sa_hdl
), woff
, abuf
, tx
);
688 * XXX This might not be necessary if
689 * dmu_assign_arcbuf_by_dbuf is guaranteed
692 zfs_clear_setid_bits_if_necessary(zfsvfs
, zp
,
693 cr
, &clear_setid_bits_txg
, tx
);
694 dmu_return_arcbuf(abuf
);
698 ASSERT3S(nbytes
, <=, zfs_uio_resid(uio
));
699 zfs_uioskip(uio
, nbytes
);
703 zn_has_cached_data(zp
, woff
, woff
+ tx_bytes
- 1) &&
704 !(ioflag
& O_DIRECT
)) {
705 update_pages(zp
, woff
, tx_bytes
, zfsvfs
->z_os
);
709 * If we made no progress, we're done. If we made even
710 * partial progress, update the znode and ZIL accordingly.
713 (void) sa_update(zp
->z_sa_hdl
, SA_ZPL_SIZE(zfsvfs
),
714 (void *)&zp
->z_size
, sizeof (uint64_t), tx
);
720 zfs_clear_setid_bits_if_necessary(zfsvfs
, zp
, cr
,
721 &clear_setid_bits_txg
, tx
);
723 zfs_tstamp_update_setup(zp
, CONTENT_MODIFIED
, mtime
, ctime
);
726 * Update the file size (zp_size) if it has changed;
727 * account for possible concurrent updates.
729 while ((end_size
= zp
->z_size
) < zfs_uio_offset(uio
)) {
730 (void) atomic_cas_64(&zp
->z_size
, end_size
,
731 zfs_uio_offset(uio
));
732 ASSERT(error
== 0 || error
== EFAULT
);
735 * If we are replaying and eof is non zero then force
736 * the file size to the specified eof. Note, there's no
737 * concurrency during replay.
739 if (zfsvfs
->z_replay
&& zfsvfs
->z_replay_eof
!= 0)
740 zp
->z_size
= zfsvfs
->z_replay_eof
;
742 error1
= sa_bulk_update(zp
->z_sa_hdl
, bulk
, count
, tx
);
744 /* Avoid clobbering EFAULT. */
748 * NB: During replay, the TX_SETATTR record logged by
749 * zfs_clear_setid_bits_if_necessary must precede any of
750 * the TX_WRITE records logged here.
752 zfs_log_write(zilog
, tx
, TX_WRITE
, zp
, woff
, tx_bytes
, ioflag
,
759 ASSERT3S(tx_bytes
, ==, nbytes
);
763 if (zfs_uio_prefaultpages(MIN(n
, max_blksz
), uio
)) {
764 error
= SET_ERROR(EFAULT
);
770 zfs_znode_update_vfs(zp
);
771 zfs_rangelock_exit(lr
);
774 * If we're in replay mode, or we made no progress, or the
775 * uio data is inaccessible return an error. Otherwise, it's
776 * at least a partial write, so it's successful.
778 if (zfsvfs
->z_replay
|| zfs_uio_resid(uio
) == start_resid
||
780 zfs_exit(zfsvfs
, FTAG
);
784 if (ioflag
& (O_SYNC
| O_DSYNC
) ||
785 zfsvfs
->z_os
->os_sync
== ZFS_SYNC_ALWAYS
)
786 zil_commit(zilog
, zp
->z_id
);
788 const int64_t nwritten
= start_resid
- zfs_uio_resid(uio
);
789 dataset_kstats_update_write_kstats(&zfsvfs
->z_kstat
, nwritten
);
790 task_io_account_write(nwritten
);
792 zfs_exit(zfsvfs
, FTAG
);
797 zfs_getsecattr(znode_t
*zp
, vsecattr_t
*vsecp
, int flag
, cred_t
*cr
)
799 zfsvfs_t
*zfsvfs
= ZTOZSB(zp
);
801 boolean_t skipaclchk
= (flag
& ATTR_NOACLCHECK
) ? B_TRUE
: B_FALSE
;
803 if ((error
= zfs_enter_verify_zp(zfsvfs
, zp
, FTAG
)) != 0)
805 error
= zfs_getacl(zp
, vsecp
, skipaclchk
, cr
);
806 zfs_exit(zfsvfs
, FTAG
);
812 zfs_setsecattr(znode_t
*zp
, vsecattr_t
*vsecp
, int flag
, cred_t
*cr
)
814 zfsvfs_t
*zfsvfs
= ZTOZSB(zp
);
816 boolean_t skipaclchk
= (flag
& ATTR_NOACLCHECK
) ? B_TRUE
: B_FALSE
;
817 zilog_t
*zilog
= zfsvfs
->z_log
;
819 if ((error
= zfs_enter_verify_zp(zfsvfs
, zp
, FTAG
)) != 0)
822 error
= zfs_setacl(zp
, vsecp
, skipaclchk
, cr
);
824 if (zfsvfs
->z_os
->os_sync
== ZFS_SYNC_ALWAYS
)
825 zil_commit(zilog
, 0);
827 zfs_exit(zfsvfs
, FTAG
);
832 static int zil_fault_io
= 0;
835 static void zfs_get_done(zgd_t
*zgd
, int error
);
838 * Get data to generate a TX_WRITE intent log record.
841 zfs_get_data(void *arg
, uint64_t gen
, lr_write_t
*lr
, char *buf
,
842 struct lwb
*lwb
, zio_t
*zio
)
844 zfsvfs_t
*zfsvfs
= arg
;
845 objset_t
*os
= zfsvfs
->z_os
;
847 uint64_t object
= lr
->lr_foid
;
848 uint64_t offset
= lr
->lr_offset
;
849 uint64_t size
= lr
->lr_length
;
855 ASSERT3P(lwb
, !=, NULL
);
856 ASSERT3P(zio
, !=, NULL
);
857 ASSERT3U(size
, !=, 0);
860 * Nothing to do if the file has been removed
862 if (zfs_zget(zfsvfs
, object
, &zp
) != 0)
863 return (SET_ERROR(ENOENT
));
864 if (zp
->z_unlinked
) {
866 * Release the vnode asynchronously as we currently have the
867 * txg stopped from syncing.
870 return (SET_ERROR(ENOENT
));
872 /* check if generation number matches */
873 if (sa_lookup(zp
->z_sa_hdl
, SA_ZPL_GEN(zfsvfs
), &zp_gen
,
874 sizeof (zp_gen
)) != 0) {
876 return (SET_ERROR(EIO
));
880 return (SET_ERROR(ENOENT
));
883 zgd
= kmem_zalloc(sizeof (zgd_t
), KM_SLEEP
);
885 zgd
->zgd_private
= zp
;
888 * Write records come in two flavors: immediate and indirect.
889 * For small writes it's cheaper to store the data with the
890 * log record (immediate); for large writes it's cheaper to
891 * sync the data and get a pointer to it (indirect) so that
892 * we don't have to write the data twice.
894 if (buf
!= NULL
) { /* immediate write */
895 zgd
->zgd_lr
= zfs_rangelock_enter(&zp
->z_rangelock
,
896 offset
, size
, RL_READER
);
897 /* test for truncation needs to be done while range locked */
898 if (offset
>= zp
->z_size
) {
899 error
= SET_ERROR(ENOENT
);
901 error
= dmu_read(os
, object
, offset
, size
, buf
,
902 DMU_READ_NO_PREFETCH
);
904 ASSERT(error
== 0 || error
== ENOENT
);
905 } else { /* indirect write */
907 * Have to lock the whole block to ensure when it's
908 * written out and its checksum is being calculated
909 * that no one can change the data. We need to re-check
910 * blocksize after we get the lock in case it's changed!
915 blkoff
= ISP2(size
) ? P2PHASE(offset
, size
) : offset
;
917 zgd
->zgd_lr
= zfs_rangelock_enter(&zp
->z_rangelock
,
918 offset
, size
, RL_READER
);
919 if (zp
->z_blksz
== size
)
922 zfs_rangelock_exit(zgd
->zgd_lr
);
924 /* test for truncation needs to be done while range locked */
925 if (lr
->lr_offset
>= zp
->z_size
)
926 error
= SET_ERROR(ENOENT
);
929 error
= SET_ERROR(EIO
);
934 error
= dmu_buf_hold(os
, object
, offset
, zgd
, &db
,
935 DMU_READ_NO_PREFETCH
);
938 blkptr_t
*bp
= &lr
->lr_blkptr
;
943 ASSERT(db
->db_offset
== offset
);
944 ASSERT(db
->db_size
== size
);
946 error
= dmu_sync(zio
, lr
->lr_common
.lrc_txg
,
948 ASSERT(error
|| lr
->lr_length
<= size
);
951 * On success, we need to wait for the write I/O
952 * initiated by dmu_sync() to complete before we can
953 * release this dbuf. We will finish everything up
954 * in the zfs_get_done() callback.
959 if (error
== EALREADY
) {
960 lr
->lr_common
.lrc_txtype
= TX_WRITE2
;
962 * TX_WRITE2 relies on the data previously
963 * written by the TX_WRITE that caused
964 * EALREADY. We zero out the BP because
965 * it is the old, currently-on-disk BP.
974 zfs_get_done(zgd
, error
);
981 zfs_get_done(zgd_t
*zgd
, int error
)
984 znode_t
*zp
= zgd
->zgd_private
;
987 dmu_buf_rele(zgd
->zgd_db
, zgd
);
989 zfs_rangelock_exit(zgd
->zgd_lr
);
992 * Release the vnode asynchronously as we currently have the
993 * txg stopped from syncing.
997 kmem_free(zgd
, sizeof (zgd_t
));
1001 zfs_enter_two(zfsvfs_t
*zfsvfs1
, zfsvfs_t
*zfsvfs2
, const char *tag
)
1005 /* Swap. Not sure if the order of zfs_enter()s is important. */
1006 if (zfsvfs1
> zfsvfs2
) {
1007 zfsvfs_t
*tmpzfsvfs
;
1009 tmpzfsvfs
= zfsvfs2
;
1011 zfsvfs1
= tmpzfsvfs
;
1014 error
= zfs_enter(zfsvfs1
, tag
);
1017 if (zfsvfs1
!= zfsvfs2
) {
1018 error
= zfs_enter(zfsvfs2
, tag
);
1020 zfs_exit(zfsvfs1
, tag
);
1029 zfs_exit_two(zfsvfs_t
*zfsvfs1
, zfsvfs_t
*zfsvfs2
, const char *tag
)
1032 zfs_exit(zfsvfs1
, tag
);
1033 if (zfsvfs1
!= zfsvfs2
)
1034 zfs_exit(zfsvfs2
, tag
);
1038 * We split each clone request in chunks that can fit into a single ZIL
1039 * log entry. Each ZIL log entry can fit 130816 bytes for a block cloning
1040 * operation (see zil_max_log_data() and zfs_log_clone_range()). This gives
1041 * us room for storing 1022 block pointers.
1043 * On success, the function return the number of bytes copied in *lenp.
1044 * Note, it doesn't return how much bytes are left to be copied.
1047 zfs_clone_range(znode_t
*inzp
, uint64_t *inoffp
, znode_t
*outzp
,
1048 uint64_t *outoffp
, uint64_t *lenp
, cred_t
*cr
)
1050 zfsvfs_t
*inzfsvfs
, *outzfsvfs
;
1051 objset_t
*inos
, *outos
;
1052 zfs_locked_range_t
*inlr
, *outlr
;
1056 uint64_t inoff
, outoff
, len
, done
;
1057 uint64_t outsize
, size
;
1060 sa_bulk_attr_t bulk
[3];
1061 uint64_t mtime
[2], ctime
[2];
1062 uint64_t uid
, gid
, projid
;
1064 size_t maxblocks
, nbps
;
1066 uint64_t clear_setid_bits_txg
= 0;
1073 inzfsvfs
= ZTOZSB(inzp
);
1074 outzfsvfs
= ZTOZSB(outzp
);
1075 inos
= inzfsvfs
->z_os
;
1076 outos
= outzfsvfs
->z_os
;
1079 * Both source and destination have to belong to the same storage pool.
1081 if (dmu_objset_spa(inos
) != dmu_objset_spa(outos
)) {
1082 zfs_exit_two(inzfsvfs
, outzfsvfs
, FTAG
);
1083 return (SET_ERROR(EXDEV
));
1087 * We need to call zfs_enter() potentially on two different datasets,
1088 * so we need a dedicated function for that.
1090 error
= zfs_enter_two(inzfsvfs
, outzfsvfs
, FTAG
);
1094 ASSERT(!outzfsvfs
->z_replay
);
1096 error
= zfs_verify_zp(inzp
);
1098 error
= zfs_verify_zp(outzp
);
1100 zfs_exit_two(inzfsvfs
, outzfsvfs
, FTAG
);
1104 if (!spa_feature_is_enabled(dmu_objset_spa(outos
),
1105 SPA_FEATURE_BLOCK_CLONING
)) {
1106 zfs_exit_two(inzfsvfs
, outzfsvfs
, FTAG
);
1107 return (SET_ERROR(EXDEV
));
1111 * We don't copy source file's flags that's why we don't allow to clone
1112 * files that are in quarantine.
1114 if (inzp
->z_pflags
& ZFS_AV_QUARANTINED
) {
1115 zfs_exit_two(inzfsvfs
, outzfsvfs
, FTAG
);
1116 return (SET_ERROR(EACCES
));
1119 if (inoff
>= inzp
->z_size
) {
1121 zfs_exit_two(inzfsvfs
, outzfsvfs
, FTAG
);
1124 if (len
> inzp
->z_size
- inoff
) {
1125 len
= inzp
->z_size
- inoff
;
1129 zfs_exit_two(inzfsvfs
, outzfsvfs
, FTAG
);
1134 * Callers might not be able to detect properly that we are read-only,
1135 * so check it explicitly here.
1137 if (zfs_is_readonly(outzfsvfs
)) {
1138 zfs_exit_two(inzfsvfs
, outzfsvfs
, FTAG
);
1139 return (SET_ERROR(EROFS
));
1143 * If immutable or not appending then return EPERM.
1144 * Intentionally allow ZFS_READONLY through here.
1145 * See zfs_zaccess_common()
1147 if ((outzp
->z_pflags
& ZFS_IMMUTABLE
) != 0) {
1148 zfs_exit_two(inzfsvfs
, outzfsvfs
, FTAG
);
1149 return (SET_ERROR(EPERM
));
1153 * No overlapping if we are cloning within the same file.
1155 if (inzp
== outzp
) {
1156 if (inoff
< outoff
+ len
&& outoff
< inoff
+ len
) {
1157 zfs_exit_two(inzfsvfs
, outzfsvfs
, FTAG
);
1158 return (SET_ERROR(EINVAL
));
1163 * Maintain predictable lock order.
1165 if (inzp
< outzp
|| (inzp
== outzp
&& inoff
< outoff
)) {
1166 inlr
= zfs_rangelock_enter(&inzp
->z_rangelock
, inoff
, len
,
1168 outlr
= zfs_rangelock_enter(&outzp
->z_rangelock
, outoff
, len
,
1171 outlr
= zfs_rangelock_enter(&outzp
->z_rangelock
, outoff
, len
,
1173 inlr
= zfs_rangelock_enter(&inzp
->z_rangelock
, inoff
, len
,
1177 inblksz
= inzp
->z_blksz
;
1180 * We cannot clone into files with different block size.
1182 if (inblksz
!= outzp
->z_blksz
&& outzp
->z_size
> inblksz
) {
1183 error
= SET_ERROR(EXDEV
);
1188 * Offsets and len must be at block boundries.
1190 if ((inoff
% inblksz
) != 0 || (outoff
% inblksz
) != 0) {
1191 error
= SET_ERROR(EXDEV
);
1195 * Length must be multipe of blksz, except for the end of the file.
1197 if ((len
% inblksz
) != 0 &&
1198 (len
< inzp
->z_size
- inoff
|| len
< outzp
->z_size
- outoff
)) {
1199 error
= SET_ERROR(EXDEV
);
1203 error
= zn_rlimit_fsize(outoff
+ len
);
1208 if (inoff
>= MAXOFFSET_T
|| outoff
>= MAXOFFSET_T
) {
1209 error
= SET_ERROR(EFBIG
);
1213 SA_ADD_BULK_ATTR(bulk
, count
, SA_ZPL_MTIME(outzfsvfs
), NULL
,
1215 SA_ADD_BULK_ATTR(bulk
, count
, SA_ZPL_CTIME(outzfsvfs
), NULL
,
1217 SA_ADD_BULK_ATTR(bulk
, count
, SA_ZPL_SIZE(outzfsvfs
), NULL
,
1220 zilog
= outzfsvfs
->z_log
;
1221 maxblocks
= zil_max_log_data(zilog
, sizeof (lr_clone_range_t
)) /
1224 uid
= KUID_TO_SUID(ZTOUID(outzp
));
1225 gid
= KGID_TO_SGID(ZTOGID(outzp
));
1226 projid
= outzp
->z_projid
;
1228 bps
= kmem_alloc(sizeof (bps
[0]) * maxblocks
, KM_SLEEP
);
1231 * Clone the file in reasonable size chunks. Each chunk is cloned
1232 * in a separate transaction; this keeps the intent log records small
1233 * and allows us to do more fine-grained space accounting.
1236 size
= MIN(inblksz
* maxblocks
, len
);
1238 if (zfs_id_overblockquota(outzfsvfs
, DMU_USERUSED_OBJECT
,
1240 zfs_id_overblockquota(outzfsvfs
, DMU_GROUPUSED_OBJECT
,
1242 (projid
!= ZFS_DEFAULT_PROJID
&&
1243 zfs_id_overblockquota(outzfsvfs
, DMU_PROJECTUSED_OBJECT
,
1245 error
= SET_ERROR(EDQUOT
);
1250 * Start a transaction.
1252 tx
= dmu_tx_create(outos
);
1255 error
= dmu_read_l0_bps(inos
, inzp
->z_id
, inoff
, size
, tx
, bps
,
1260 * If we are tyring to clone a block that was created
1261 * in the current transaction group. Return an error,
1262 * so the caller can fallback to just copying the data.
1264 if (error
== EAGAIN
) {
1265 error
= SET_ERROR(EXDEV
);
1270 * Encrypted data is fine as long as it comes from the same
1272 * TODO: We want to extend it in the future to allow cloning to
1273 * datasets with the same keys, like clones or to be able to
1274 * clone a file from a snapshot of an encrypted dataset into the
1277 if (BP_IS_PROTECTED(&bps
[0])) {
1278 if (inzfsvfs
!= outzfsvfs
) {
1280 error
= SET_ERROR(EXDEV
);
1285 dmu_tx_hold_sa(tx
, outzp
->z_sa_hdl
, B_FALSE
);
1286 db
= (dmu_buf_impl_t
*)sa_get_db(outzp
->z_sa_hdl
);
1288 dmu_tx_hold_clone_by_dnode(tx
, DB_DNODE(db
), outoff
, size
);
1290 zfs_sa_upgrade_txholds(tx
, outzp
);
1291 error
= dmu_tx_assign(tx
, TXG_WAIT
);
1298 * Copy source znode's block size. This only happens on the
1299 * first iteration since zfs_rangelock_reduce() will shrink down
1300 * lr_len to the appropriate size.
1302 if (outlr
->lr_length
== UINT64_MAX
) {
1303 zfs_grow_blocksize(outzp
, inblksz
, tx
);
1305 * Round range lock up to the block boundary, so we
1306 * prevent appends until we are done.
1308 zfs_rangelock_reduce(outlr
, outoff
,
1309 ((len
- 1) / inblksz
+ 1) * inblksz
);
1312 error
= dmu_brt_clone(outos
, outzp
->z_id
, outoff
, size
, tx
,
1313 bps
, nbps
, B_FALSE
);
1319 zfs_clear_setid_bits_if_necessary(outzfsvfs
, outzp
, cr
,
1320 &clear_setid_bits_txg
, tx
);
1322 zfs_tstamp_update_setup(outzp
, CONTENT_MODIFIED
, mtime
, ctime
);
1325 * Update the file size (zp_size) if it has changed;
1326 * account for possible concurrent updates.
1328 while ((outsize
= outzp
->z_size
) < outoff
+ size
) {
1329 (void) atomic_cas_64(&outzp
->z_size
, outsize
,
1333 error
= sa_bulk_update(outzp
->z_sa_hdl
, bulk
, count
, tx
);
1335 zfs_log_clone_range(zilog
, tx
, TX_CLONE_RANGE
, outzp
, outoff
,
1336 size
, inblksz
, bps
, nbps
);
1349 kmem_free(bps
, sizeof (bps
[0]) * maxblocks
);
1350 zfs_znode_update_vfs(outzp
);
1353 zfs_rangelock_exit(outlr
);
1354 zfs_rangelock_exit(inlr
);
1358 * If we have made at least partial progress, reset the error.
1362 ZFS_ACCESSTIME_STAMP(inzfsvfs
, inzp
);
1364 if (outos
->os_sync
== ZFS_SYNC_ALWAYS
) {
1365 zil_commit(zilog
, outzp
->z_id
);
1373 zfs_exit_two(inzfsvfs
, outzfsvfs
, FTAG
);
1379 * Usual pattern would be to call zfs_clone_range() from zfs_replay_clone(),
1380 * but we cannot do that, because when replaying we don't have source znode
1381 * available. This is why we need a dedicated replay function.
1384 zfs_clone_range_replay(znode_t
*zp
, uint64_t off
, uint64_t len
, uint64_t blksz
,
1385 const blkptr_t
*bps
, size_t nbps
)
1392 sa_bulk_attr_t bulk
[3];
1393 uint64_t mtime
[2], ctime
[2];
1395 ASSERT3U(off
, <, MAXOFFSET_T
);
1396 ASSERT3U(len
, >, 0);
1397 ASSERT3U(nbps
, >, 0);
1399 zfsvfs
= ZTOZSB(zp
);
1401 ASSERT(spa_feature_is_enabled(dmu_objset_spa(zfsvfs
->z_os
),
1402 SPA_FEATURE_BLOCK_CLONING
));
1404 if ((error
= zfs_enter_verify_zp(zfsvfs
, zp
, FTAG
)) != 0)
1407 ASSERT(zfsvfs
->z_replay
);
1408 ASSERT(!zfs_is_readonly(zfsvfs
));
1410 if ((off
% blksz
) != 0) {
1411 zfs_exit(zfsvfs
, FTAG
);
1412 return (SET_ERROR(EINVAL
));
1415 SA_ADD_BULK_ATTR(bulk
, count
, SA_ZPL_MTIME(zfsvfs
), NULL
, &mtime
, 16);
1416 SA_ADD_BULK_ATTR(bulk
, count
, SA_ZPL_CTIME(zfsvfs
), NULL
, &ctime
, 16);
1417 SA_ADD_BULK_ATTR(bulk
, count
, SA_ZPL_SIZE(zfsvfs
), NULL
,
1421 * Start a transaction.
1423 tx
= dmu_tx_create(zfsvfs
->z_os
);
1425 dmu_tx_hold_sa(tx
, zp
->z_sa_hdl
, B_FALSE
);
1426 db
= (dmu_buf_impl_t
*)sa_get_db(zp
->z_sa_hdl
);
1428 dmu_tx_hold_clone_by_dnode(tx
, DB_DNODE(db
), off
, len
);
1430 zfs_sa_upgrade_txholds(tx
, zp
);
1431 error
= dmu_tx_assign(tx
, TXG_WAIT
);
1434 zfs_exit(zfsvfs
, FTAG
);
1438 if (zp
->z_blksz
< blksz
)
1439 zfs_grow_blocksize(zp
, blksz
, tx
);
1441 dmu_brt_clone(zfsvfs
->z_os
, zp
->z_id
, off
, len
, tx
, bps
, nbps
, B_TRUE
);
1443 zfs_tstamp_update_setup(zp
, CONTENT_MODIFIED
, mtime
, ctime
);
1445 if (zp
->z_size
< off
+ len
)
1446 zp
->z_size
= off
+ len
;
1448 error
= sa_bulk_update(zp
->z_sa_hdl
, bulk
, count
, tx
);
1451 * zil_replaying() not only check if we are replaying ZIL, but also
1452 * updates the ZIL header to record replay progress.
1454 VERIFY(zil_replaying(zfsvfs
->z_log
, tx
));
1458 zfs_znode_update_vfs(zp
);
1460 zfs_exit(zfsvfs
, FTAG
);
1465 EXPORT_SYMBOL(zfs_access
);
1466 EXPORT_SYMBOL(zfs_fsync
);
1467 EXPORT_SYMBOL(zfs_holey
);
1468 EXPORT_SYMBOL(zfs_read
);
1469 EXPORT_SYMBOL(zfs_write
);
1470 EXPORT_SYMBOL(zfs_getsecattr
);
1471 EXPORT_SYMBOL(zfs_setsecattr
);
1472 EXPORT_SYMBOL(zfs_clone_range
);
1473 EXPORT_SYMBOL(zfs_clone_range_replay
);
1475 ZFS_MODULE_PARAM(zfs_vnops
, zfs_vnops_
, read_chunk_size
, U64
, ZMOD_RW
,
1476 "Bytes to read per chunk");