2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1982, 1986, 1989, 1993
5 * The Regents of the University of California. All rights reserved.
6 * (c) UNIX System Laboratories, Inc.
7 * All or some portions of this file are derived from material licensed
8 * to the University of California by American Telephone and Telegraph
9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10 * the permission of UNIX System Laboratories, Inc.
12 * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13 * Copyright (c) 2013, 2014 The FreeBSD Foundation
15 * Portions of this software were developed by Konstantin Belousov
16 * under sponsorship from the FreeBSD Foundation.
18 * Redistribution and use in source and binary forms, with or without
19 * modification, are permitted provided that the following conditions
21 * 1. Redistributions of source code must retain the above copyright
22 * notice, this list of conditions and the following disclaimer.
23 * 2. Redistributions in binary form must reproduce the above copyright
24 * notice, this list of conditions and the following disclaimer in the
25 * documentation and/or other materials provided with the distribution.
26 * 3. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42 * @(#)vfs_vnops.c 8.2 (Berkeley) 1/21/94
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
48 #include "opt_hwpmc_hooks.h"
50 #include <sys/param.h>
51 #include <sys/systm.h>
54 #include <sys/fcntl.h>
61 #include <sys/limits.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
74 #include <sys/sysctl.h>
75 #include <sys/ttycom.h>
77 #include <sys/syslog.h>
78 #include <sys/unistd.h>
81 #include <security/audit/audit.h>
82 #include <security/mac/mac_framework.h>
85 #include <vm/vm_extern.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vnode_pager.h>
93 #include <sys/pmckern.h>
96 static fo_rdwr_t vn_read;
97 static fo_rdwr_t vn_write;
98 static fo_rdwr_t vn_io_fault;
99 static fo_truncate_t vn_truncate;
100 static fo_ioctl_t vn_ioctl;
101 static fo_poll_t vn_poll;
102 static fo_kqfilter_t vn_kqfilter;
103 static fo_stat_t vn_statfile;
104 static fo_close_t vn_closefile;
105 static fo_mmap_t vn_mmap;
107 struct fileops vnops = {
108 .fo_read = vn_io_fault,
109 .fo_write = vn_io_fault,
110 .fo_truncate = vn_truncate,
111 .fo_ioctl = vn_ioctl,
113 .fo_kqfilter = vn_kqfilter,
114 .fo_stat = vn_statfile,
115 .fo_close = vn_closefile,
116 .fo_chmod = vn_chmod,
117 .fo_chown = vn_chown,
118 .fo_sendfile = vn_sendfile,
120 .fo_fill_kinfo = vn_fill_kinfo,
122 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
125 static const int io_hold_cnt = 16;
126 static int vn_io_fault_enable = 1;
127 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW,
128 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
129 static int vn_io_fault_prefault = 0;
130 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RW,
131 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
132 static u_long vn_io_faults_cnt;
133 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
134 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
137 * Returns true if vn_io_fault mode of handling the i/o request should
141 do_vn_io_fault(struct vnode *vp, struct uio *uio)
145 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
146 (mp = vp->v_mount) != NULL &&
147 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
151 * Structure used to pass arguments to vn_io_fault1(), to do either
152 * file- or vnode-based I/O calls.
154 struct vn_io_fault_args {
162 struct fop_args_tag {
166 struct vop_args_tag {
172 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
173 struct vn_io_fault_args *args, struct thread *td);
176 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
178 struct thread *td = ndp->ni_cnd.cn_thread;
180 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
184 * Common code for vnode open operations via a name lookup.
185 * Lookup the vnode and invoke VOP_CREATE if needed.
186 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
188 * Note that this does NOT free nameidata for the successful case,
189 * due to the NDINIT being done elsewhere.
192 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
193 struct ucred *cred, struct file *fp)
197 struct thread *td = ndp->ni_cnd.cn_thread;
199 struct vattr *vap = &vat;
204 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
205 O_EXCL | O_DIRECTORY))
207 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
208 ndp->ni_cnd.cn_nameiop = CREATE;
210 * Set NOCACHE to avoid flushing the cache when
211 * rolling in many files at once.
213 ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE;
214 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
215 ndp->ni_cnd.cn_flags |= FOLLOW;
216 if ((fmode & O_BENEATH) != 0)
217 ndp->ni_cnd.cn_flags |= BENEATH;
218 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
219 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
220 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
221 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
223 if ((error = namei(ndp)) != 0)
225 if (ndp->ni_vp == NULL) {
228 vap->va_mode = cmode;
230 vap->va_vaflags |= VA_EXCLUSIVE;
231 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
232 NDFREE(ndp, NDF_ONLY_PNBUF);
234 if ((error = vn_start_write(NULL, &mp,
235 V_XSLEEP | PCATCH)) != 0)
239 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
240 ndp->ni_cnd.cn_flags |= MAKEENTRY;
242 error = mac_vnode_check_create(cred, ndp->ni_dvp,
246 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
249 vn_finished_write(mp);
251 NDFREE(ndp, NDF_ONLY_PNBUF);
257 if (ndp->ni_dvp == ndp->ni_vp)
263 if (fmode & O_EXCL) {
270 ndp->ni_cnd.cn_nameiop = LOOKUP;
271 ndp->ni_cnd.cn_flags = ISOPEN |
272 ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
273 if (!(fmode & FWRITE))
274 ndp->ni_cnd.cn_flags |= LOCKSHARED;
275 if ((fmode & O_BENEATH) != 0)
276 ndp->ni_cnd.cn_flags |= BENEATH;
277 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
278 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
279 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
280 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
281 if ((error = namei(ndp)) != 0)
285 error = vn_open_vnode(vp, fmode, cred, td, fp);
291 NDFREE(ndp, NDF_ONLY_PNBUF);
299 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
302 int error, lock_flags, type;
304 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
305 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
307 KASSERT(fp != NULL, ("open with flock requires fp"));
308 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
311 lock_flags = VOP_ISLOCKED(vp);
314 lf.l_whence = SEEK_SET;
317 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
319 if ((fmode & FNONBLOCK) == 0)
321 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
323 fp->f_flag |= FHASLOCK;
325 vn_lock(vp, lock_flags | LK_RETRY);
326 if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0)
332 * Common code for vnode open operations once a vnode is located.
333 * Check permissions, and call the VOP_OPEN routine.
336 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
337 struct thread *td, struct file *fp)
342 if (vp->v_type == VLNK)
344 if (vp->v_type == VSOCK)
346 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
349 if (fmode & (FWRITE | O_TRUNC)) {
350 if (vp->v_type == VDIR)
358 if ((fmode & O_APPEND) && (fmode & FWRITE))
363 if (fmode & O_VERIFY)
365 error = mac_vnode_check_open(cred, vp, accmode);
369 accmode &= ~(VCREAT | VVERIFY);
371 if ((fmode & O_CREAT) == 0 && accmode != 0) {
372 error = VOP_ACCESS(vp, accmode, cred, td);
376 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
377 vn_lock(vp, LK_UPGRADE | LK_RETRY);
378 error = VOP_OPEN(vp, fmode, cred, td, fp);
382 error = vn_open_vnode_advlock(vp, fmode, fp);
383 if (error == 0 && (fmode & FWRITE) != 0) {
384 error = VOP_ADD_WRITECOUNT(vp, 1);
386 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
387 __func__, vp, vp->v_writecount);
392 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
393 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
394 * Arrange for that by having fdrop() to use vn_closefile().
397 fp->f_flag |= FOPENFAILED;
399 if (fp->f_ops == &badfileops) {
400 fp->f_type = DTYPE_VNODE;
406 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
412 * Check for write permissions on the specified vnode.
413 * Prototype text segments cannot be written.
417 vn_writechk(struct vnode *vp)
420 ASSERT_VOP_LOCKED(vp, "vn_writechk");
422 * If there's shared text associated with
423 * the vnode, try to free it up once. If
424 * we fail, we can't allow writing.
436 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
437 struct thread *td, bool keep_ref)
440 int error, lock_flags;
442 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
443 MNT_EXTENDED_SHARED(vp->v_mount))
444 lock_flags = LK_SHARED;
446 lock_flags = LK_EXCLUSIVE;
448 vn_start_write(vp, &mp, V_WAIT);
449 vn_lock(vp, lock_flags | LK_RETRY);
450 AUDIT_ARG_VNODE1(vp);
451 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
452 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
453 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
454 __func__, vp, vp->v_writecount);
456 error = VOP_CLOSE(vp, flags, file_cred, td);
461 vn_finished_write(mp);
466 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
470 return (vn_close1(vp, flags, file_cred, td, false));
474 * Heuristic to detect sequential operation.
477 sequential_heuristic(struct uio *uio, struct file *fp)
480 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
481 if (fp->f_flag & FRDAHEAD)
482 return (fp->f_seqcount << IO_SEQSHIFT);
485 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
486 * that the first I/O is normally considered to be slightly
487 * sequential. Seeking to offset 0 doesn't change sequentiality
488 * unless previous seeks have reduced f_seqcount to 0, in which
489 * case offset 0 is not special.
491 if ((uio->uio_offset == 0 && fp->f_seqcount > 0) ||
492 uio->uio_offset == fp->f_nextoff) {
494 * f_seqcount is in units of fixed-size blocks so that it
495 * depends mainly on the amount of sequential I/O and not
496 * much on the number of sequential I/O's. The fixed size
497 * of 16384 is hard-coded here since it is (not quite) just
498 * a magic size that works well here. This size is more
499 * closely related to the best I/O size for real disks than
500 * to any block size used by software.
502 if (uio->uio_resid >= IO_SEQMAX * 16384)
503 fp->f_seqcount = IO_SEQMAX;
505 fp->f_seqcount += howmany(uio->uio_resid, 16384);
506 if (fp->f_seqcount > IO_SEQMAX)
507 fp->f_seqcount = IO_SEQMAX;
509 return (fp->f_seqcount << IO_SEQSHIFT);
512 /* Not sequential. Quickly draw-down sequentiality. */
513 if (fp->f_seqcount > 1)
521 * Package up an I/O request on a vnode into a uio and do it.
524 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
525 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
526 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
533 struct vn_io_fault_args args;
534 int error, lock_flags;
536 if (offset < 0 && vp->v_type != VCHR)
538 auio.uio_iov = &aiov;
540 aiov.iov_base = base;
542 auio.uio_resid = len;
543 auio.uio_offset = offset;
544 auio.uio_segflg = segflg;
549 if ((ioflg & IO_NODELOCKED) == 0) {
550 if ((ioflg & IO_RANGELOCKED) == 0) {
551 if (rw == UIO_READ) {
552 rl_cookie = vn_rangelock_rlock(vp, offset,
555 rl_cookie = vn_rangelock_wlock(vp, offset,
561 if (rw == UIO_WRITE) {
562 if (vp->v_type != VCHR &&
563 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
566 if (MNT_SHARED_WRITES(mp) ||
567 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
568 lock_flags = LK_SHARED;
570 lock_flags = LK_EXCLUSIVE;
572 lock_flags = LK_SHARED;
573 vn_lock(vp, lock_flags | LK_RETRY);
577 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
579 if ((ioflg & IO_NOMACCHECK) == 0) {
581 error = mac_vnode_check_read(active_cred, file_cred,
584 error = mac_vnode_check_write(active_cred, file_cred,
589 if (file_cred != NULL)
593 if (do_vn_io_fault(vp, &auio)) {
594 args.kind = VN_IO_FAULT_VOP;
597 args.args.vop_args.vp = vp;
598 error = vn_io_fault1(vp, &auio, &args, td);
599 } else if (rw == UIO_READ) {
600 error = VOP_READ(vp, &auio, ioflg, cred);
601 } else /* if (rw == UIO_WRITE) */ {
602 error = VOP_WRITE(vp, &auio, ioflg, cred);
606 *aresid = auio.uio_resid;
608 if (auio.uio_resid && error == 0)
610 if ((ioflg & IO_NODELOCKED) == 0) {
613 vn_finished_write(mp);
616 if (rl_cookie != NULL)
617 vn_rangelock_unlock(vp, rl_cookie);
622 * Package up an I/O request on a vnode into a uio and do it. The I/O
623 * request is split up into smaller chunks and we try to avoid saturating
624 * the buffer cache while potentially holding a vnode locked, so we
625 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
626 * to give other processes a chance to lock the vnode (either other processes
627 * core'ing the same binary, or unrelated processes scanning the directory).
630 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
631 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
632 struct ucred *file_cred, size_t *aresid, struct thread *td)
641 * Force `offset' to a multiple of MAXBSIZE except possibly
642 * for the first chunk, so that filesystems only need to
643 * write full blocks except possibly for the first and last
646 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
650 if (rw != UIO_READ && vp->v_type == VREG)
653 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
654 ioflg, active_cred, file_cred, &iaresid, td);
655 len -= chunk; /* aresid calc already includes length */
659 base = (char *)base + chunk;
660 kern_yield(PRI_USER);
663 *aresid = len + iaresid;
668 foffset_lock(struct file *fp, int flags)
673 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
675 #if OFF_MAX <= LONG_MAX
677 * Caller only wants the current f_offset value. Assume that
678 * the long and shorter integer types reads are atomic.
680 if ((flags & FOF_NOLOCK) != 0)
681 return (fp->f_offset);
685 * According to McKusick the vn lock was protecting f_offset here.
686 * It is now protected by the FOFFSET_LOCKED flag.
688 mtxp = mtx_pool_find(mtxpool_sleep, fp);
690 if ((flags & FOF_NOLOCK) == 0) {
691 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
692 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
693 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
696 fp->f_vnread_flags |= FOFFSET_LOCKED;
704 foffset_unlock(struct file *fp, off_t val, int flags)
708 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
710 #if OFF_MAX <= LONG_MAX
711 if ((flags & FOF_NOLOCK) != 0) {
712 if ((flags & FOF_NOUPDATE) == 0)
714 if ((flags & FOF_NEXTOFF) != 0)
720 mtxp = mtx_pool_find(mtxpool_sleep, fp);
722 if ((flags & FOF_NOUPDATE) == 0)
724 if ((flags & FOF_NEXTOFF) != 0)
726 if ((flags & FOF_NOLOCK) == 0) {
727 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
728 ("Lost FOFFSET_LOCKED"));
729 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
730 wakeup(&fp->f_vnread_flags);
731 fp->f_vnread_flags = 0;
737 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
740 if ((flags & FOF_OFFSET) == 0)
741 uio->uio_offset = foffset_lock(fp, flags);
745 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
748 if ((flags & FOF_OFFSET) == 0)
749 foffset_unlock(fp, uio->uio_offset, flags);
753 get_advice(struct file *fp, struct uio *uio)
758 ret = POSIX_FADV_NORMAL;
759 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
762 mtxp = mtx_pool_find(mtxpool_sleep, fp);
764 if (fp->f_advice != NULL &&
765 uio->uio_offset >= fp->f_advice->fa_start &&
766 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
767 ret = fp->f_advice->fa_advice;
773 * File table vnode read routine.
776 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
784 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
786 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
789 if (fp->f_flag & FNONBLOCK)
791 if (fp->f_flag & O_DIRECT)
793 advice = get_advice(fp, uio);
794 vn_lock(vp, LK_SHARED | LK_RETRY);
797 case POSIX_FADV_NORMAL:
798 case POSIX_FADV_SEQUENTIAL:
799 case POSIX_FADV_NOREUSE:
800 ioflag |= sequential_heuristic(uio, fp);
802 case POSIX_FADV_RANDOM:
803 /* Disable read-ahead for random I/O. */
806 orig_offset = uio->uio_offset;
809 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
812 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
813 fp->f_nextoff = uio->uio_offset;
815 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
816 orig_offset != uio->uio_offset)
818 * Use POSIX_FADV_DONTNEED to flush pages and buffers
819 * for the backing file after a POSIX_FADV_NOREUSE
822 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
823 POSIX_FADV_DONTNEED);
828 * File table vnode write routine.
831 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
837 int error, ioflag, lock_flags;
840 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
842 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
844 if (vp->v_type == VREG)
847 if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
849 if (fp->f_flag & FNONBLOCK)
851 if (fp->f_flag & O_DIRECT)
853 if ((fp->f_flag & O_FSYNC) ||
854 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
857 if (vp->v_type != VCHR &&
858 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
861 advice = get_advice(fp, uio);
863 if (MNT_SHARED_WRITES(mp) ||
864 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
865 lock_flags = LK_SHARED;
867 lock_flags = LK_EXCLUSIVE;
870 vn_lock(vp, lock_flags | LK_RETRY);
872 case POSIX_FADV_NORMAL:
873 case POSIX_FADV_SEQUENTIAL:
874 case POSIX_FADV_NOREUSE:
875 ioflag |= sequential_heuristic(uio, fp);
877 case POSIX_FADV_RANDOM:
878 /* XXX: Is this correct? */
881 orig_offset = uio->uio_offset;
884 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
887 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
888 fp->f_nextoff = uio->uio_offset;
890 if (vp->v_type != VCHR)
891 vn_finished_write(mp);
892 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
893 orig_offset != uio->uio_offset)
895 * Use POSIX_FADV_DONTNEED to flush pages and buffers
896 * for the backing file after a POSIX_FADV_NOREUSE
899 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
900 POSIX_FADV_DONTNEED);
906 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
907 * prevent the following deadlock:
909 * Assume that the thread A reads from the vnode vp1 into userspace
910 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
911 * currently not resident, then system ends up with the call chain
912 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
913 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
914 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
915 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
916 * backed by the pages of vnode vp1, and some page in buf2 is not
917 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
919 * To prevent the lock order reversal and deadlock, vn_io_fault() does
920 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
921 * Instead, it first tries to do the whole range i/o with pagefaults
922 * disabled. If all pages in the i/o buffer are resident and mapped,
923 * VOP will succeed (ignoring the genuine filesystem errors).
924 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
925 * i/o in chunks, with all pages in the chunk prefaulted and held
926 * using vm_fault_quick_hold_pages().
928 * Filesystems using this deadlock avoidance scheme should use the
929 * array of the held pages from uio, saved in the curthread->td_ma,
930 * instead of doing uiomove(). A helper function
931 * vn_io_fault_uiomove() converts uiomove request into
932 * uiomove_fromphys() over td_ma array.
934 * Since vnode locks do not cover the whole i/o anymore, rangelocks
935 * make the current i/o request atomic with respect to other i/os and
940 * Decode vn_io_fault_args and perform the corresponding i/o.
943 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
949 save = vm_fault_disable_pagefaults();
950 switch (args->kind) {
951 case VN_IO_FAULT_FOP:
952 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
953 uio, args->cred, args->flags, td);
955 case VN_IO_FAULT_VOP:
956 if (uio->uio_rw == UIO_READ) {
957 error = VOP_READ(args->args.vop_args.vp, uio,
958 args->flags, args->cred);
959 } else if (uio->uio_rw == UIO_WRITE) {
960 error = VOP_WRITE(args->args.vop_args.vp, uio,
961 args->flags, args->cred);
965 panic("vn_io_fault_doio: unknown kind of io %d %d",
966 args->kind, uio->uio_rw);
968 vm_fault_enable_pagefaults(save);
973 vn_io_fault_touch(char *base, const struct uio *uio)
978 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
984 vn_io_fault_prefault_user(const struct uio *uio)
987 const struct iovec *iov;
992 KASSERT(uio->uio_segflg == UIO_USERSPACE,
993 ("vn_io_fault_prefault userspace"));
997 resid = uio->uio_resid;
998 base = iov->iov_base;
1001 error = vn_io_fault_touch(base, uio);
1004 if (len < PAGE_SIZE) {
1006 error = vn_io_fault_touch(base + len - 1, uio);
1011 if (++i >= uio->uio_iovcnt)
1013 iov = uio->uio_iov + i;
1014 base = iov->iov_base;
1026 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1027 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1028 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1029 * into args and call vn_io_fault1() to handle faults during the user
1030 * mode buffer accesses.
1033 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1036 vm_page_t ma[io_hold_cnt + 2];
1037 struct uio *uio_clone, short_uio;
1038 struct iovec short_iovec[1];
1039 vm_page_t *prev_td_ma;
1041 vm_offset_t addr, end;
1044 int error, cnt, saveheld, prev_td_ma_cnt;
1046 if (vn_io_fault_prefault) {
1047 error = vn_io_fault_prefault_user(uio);
1049 return (error); /* Or ignore ? */
1052 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1055 * The UFS follows IO_UNIT directive and replays back both
1056 * uio_offset and uio_resid if an error is encountered during the
1057 * operation. But, since the iovec may be already advanced,
1058 * uio is still in an inconsistent state.
1060 * Cache a copy of the original uio, which is advanced to the redo
1061 * point using UIO_NOCOPY below.
1063 uio_clone = cloneuio(uio);
1064 resid = uio->uio_resid;
1066 short_uio.uio_segflg = UIO_USERSPACE;
1067 short_uio.uio_rw = uio->uio_rw;
1068 short_uio.uio_td = uio->uio_td;
1070 error = vn_io_fault_doio(args, uio, td);
1071 if (error != EFAULT)
1074 atomic_add_long(&vn_io_faults_cnt, 1);
1075 uio_clone->uio_segflg = UIO_NOCOPY;
1076 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1077 uio_clone->uio_segflg = uio->uio_segflg;
1079 saveheld = curthread_pflags_set(TDP_UIOHELD);
1080 prev_td_ma = td->td_ma;
1081 prev_td_ma_cnt = td->td_ma_cnt;
1083 while (uio_clone->uio_resid != 0) {
1084 len = uio_clone->uio_iov->iov_len;
1086 KASSERT(uio_clone->uio_iovcnt >= 1,
1087 ("iovcnt underflow"));
1088 uio_clone->uio_iov++;
1089 uio_clone->uio_iovcnt--;
1092 if (len > io_hold_cnt * PAGE_SIZE)
1093 len = io_hold_cnt * PAGE_SIZE;
1094 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1095 end = round_page(addr + len);
1100 cnt = atop(end - trunc_page(addr));
1102 * A perfectly misaligned address and length could cause
1103 * both the start and the end of the chunk to use partial
1104 * page. +2 accounts for such a situation.
1106 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1107 addr, len, prot, ma, io_hold_cnt + 2);
1112 short_uio.uio_iov = &short_iovec[0];
1113 short_iovec[0].iov_base = (void *)addr;
1114 short_uio.uio_iovcnt = 1;
1115 short_uio.uio_resid = short_iovec[0].iov_len = len;
1116 short_uio.uio_offset = uio_clone->uio_offset;
1118 td->td_ma_cnt = cnt;
1120 error = vn_io_fault_doio(args, &short_uio, td);
1121 vm_page_unhold_pages(ma, cnt);
1122 adv = len - short_uio.uio_resid;
1124 uio_clone->uio_iov->iov_base =
1125 (char *)uio_clone->uio_iov->iov_base + adv;
1126 uio_clone->uio_iov->iov_len -= adv;
1127 uio_clone->uio_resid -= adv;
1128 uio_clone->uio_offset += adv;
1130 uio->uio_resid -= adv;
1131 uio->uio_offset += adv;
1133 if (error != 0 || adv == 0)
1136 td->td_ma = prev_td_ma;
1137 td->td_ma_cnt = prev_td_ma_cnt;
1138 curthread_pflags_restore(saveheld);
1140 free(uio_clone, M_IOV);
1145 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1146 int flags, struct thread *td)
1151 struct vn_io_fault_args args;
1154 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1156 foffset_lock_uio(fp, uio, flags);
1157 if (do_vn_io_fault(vp, uio)) {
1158 args.kind = VN_IO_FAULT_FOP;
1159 args.args.fop_args.fp = fp;
1160 args.args.fop_args.doio = doio;
1161 args.cred = active_cred;
1162 args.flags = flags | FOF_OFFSET;
1163 if (uio->uio_rw == UIO_READ) {
1164 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1165 uio->uio_offset + uio->uio_resid);
1166 } else if ((fp->f_flag & O_APPEND) != 0 ||
1167 (flags & FOF_OFFSET) == 0) {
1168 /* For appenders, punt and lock the whole range. */
1169 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1171 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1172 uio->uio_offset + uio->uio_resid);
1174 error = vn_io_fault1(vp, uio, &args, td);
1175 vn_rangelock_unlock(vp, rl_cookie);
1177 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1179 foffset_unlock_uio(fp, uio, flags);
1184 * Helper function to perform the requested uiomove operation using
1185 * the held pages for io->uio_iov[0].iov_base buffer instead of
1186 * copyin/copyout. Access to the pages with uiomove_fromphys()
1187 * instead of iov_base prevents page faults that could occur due to
1188 * pmap_collect() invalidating the mapping created by
1189 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1190 * object cleanup revoking the write access from page mappings.
1192 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1193 * instead of plain uiomove().
1196 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1198 struct uio transp_uio;
1199 struct iovec transp_iov[1];
1205 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1206 uio->uio_segflg != UIO_USERSPACE)
1207 return (uiomove(data, xfersize, uio));
1209 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1210 transp_iov[0].iov_base = data;
1211 transp_uio.uio_iov = &transp_iov[0];
1212 transp_uio.uio_iovcnt = 1;
1213 if (xfersize > uio->uio_resid)
1214 xfersize = uio->uio_resid;
1215 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1216 transp_uio.uio_offset = 0;
1217 transp_uio.uio_segflg = UIO_SYSSPACE;
1219 * Since transp_iov points to data, and td_ma page array
1220 * corresponds to original uio->uio_iov, we need to invert the
1221 * direction of the i/o operation as passed to
1222 * uiomove_fromphys().
1224 switch (uio->uio_rw) {
1226 transp_uio.uio_rw = UIO_READ;
1229 transp_uio.uio_rw = UIO_WRITE;
1232 transp_uio.uio_td = uio->uio_td;
1233 error = uiomove_fromphys(td->td_ma,
1234 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1235 xfersize, &transp_uio);
1236 adv = xfersize - transp_uio.uio_resid;
1238 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1239 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1241 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1243 td->td_ma_cnt -= pgadv;
1244 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1245 uio->uio_iov->iov_len -= adv;
1246 uio->uio_resid -= adv;
1247 uio->uio_offset += adv;
1252 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1256 vm_offset_t iov_base;
1260 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1261 uio->uio_segflg != UIO_USERSPACE)
1262 return (uiomove_fromphys(ma, offset, xfersize, uio));
1264 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1265 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1266 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1267 switch (uio->uio_rw) {
1269 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1273 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1277 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1279 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1281 td->td_ma_cnt -= pgadv;
1282 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1283 uio->uio_iov->iov_len -= cnt;
1284 uio->uio_resid -= cnt;
1285 uio->uio_offset += cnt;
1291 * File table truncate routine.
1294 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1305 * Lock the whole range for truncation. Otherwise split i/o
1306 * might happen partly before and partly after the truncation.
1308 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1309 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1312 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1313 AUDIT_ARG_VNODE1(vp);
1314 if (vp->v_type == VDIR) {
1319 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1323 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1327 vn_finished_write(mp);
1329 vn_rangelock_unlock(vp, rl_cookie);
1334 * Truncate a file that is already locked.
1337 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1343 error = VOP_ADD_WRITECOUNT(vp, 1);
1346 vattr.va_size = length;
1348 vattr.va_vaflags |= VA_SYNC;
1349 error = VOP_SETATTR(vp, &vattr, cred);
1350 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1356 * File table vnode stat routine.
1359 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1362 struct vnode *vp = fp->f_vnode;
1365 vn_lock(vp, LK_SHARED | LK_RETRY);
1366 error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1373 * Stat a vnode; implementation for the stat syscall
1376 vn_stat(struct vnode *vp, struct stat *sb, struct ucred *active_cred,
1377 struct ucred *file_cred, struct thread *td)
1384 AUDIT_ARG_VNODE1(vp);
1386 error = mac_vnode_check_stat(active_cred, file_cred, vp);
1394 * Initialize defaults for new and unusual fields, so that file
1395 * systems which don't support these fields don't need to know
1398 vap->va_birthtime.tv_sec = -1;
1399 vap->va_birthtime.tv_nsec = 0;
1400 vap->va_fsid = VNOVAL;
1401 vap->va_rdev = NODEV;
1403 error = VOP_GETATTR(vp, vap, active_cred);
1408 * Zero the spare stat fields
1410 bzero(sb, sizeof *sb);
1413 * Copy from vattr table
1415 if (vap->va_fsid != VNOVAL)
1416 sb->st_dev = vap->va_fsid;
1418 sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1419 sb->st_ino = vap->va_fileid;
1420 mode = vap->va_mode;
1421 switch (vap->va_type) {
1447 sb->st_nlink = vap->va_nlink;
1448 sb->st_uid = vap->va_uid;
1449 sb->st_gid = vap->va_gid;
1450 sb->st_rdev = vap->va_rdev;
1451 if (vap->va_size > OFF_MAX)
1453 sb->st_size = vap->va_size;
1454 sb->st_atim = vap->va_atime;
1455 sb->st_mtim = vap->va_mtime;
1456 sb->st_ctim = vap->va_ctime;
1457 sb->st_birthtim = vap->va_birthtime;
1460 * According to www.opengroup.org, the meaning of st_blksize is
1461 * "a filesystem-specific preferred I/O block size for this
1462 * object. In some filesystem types, this may vary from file
1464 * Use miminum/default of PAGE_SIZE (e.g. for VCHR).
1467 sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1469 sb->st_flags = vap->va_flags;
1470 if (priv_check(td, PRIV_VFS_GENERATION))
1473 sb->st_gen = vap->va_gen;
1475 sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1480 * File table vnode ioctl routine.
1483 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1488 struct fiobmap2_arg *bmarg;
1492 switch (vp->v_type) {
1497 vn_lock(vp, LK_SHARED | LK_RETRY);
1498 error = VOP_GETATTR(vp, &vattr, active_cred);
1501 *(int *)data = vattr.va_size - fp->f_offset;
1504 bmarg = (struct fiobmap2_arg *)data;
1505 vn_lock(vp, LK_SHARED | LK_RETRY);
1507 error = mac_vnode_check_read(active_cred, fp->f_cred,
1511 error = VOP_BMAP(vp, bmarg->bn, NULL,
1512 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1519 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1524 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1532 * File table vnode poll routine.
1535 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1543 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1544 AUDIT_ARG_VNODE1(vp);
1545 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1550 error = VOP_POLL(vp, events, fp->f_cred, td);
1555 * Acquire the requested lock and then check for validity. LK_RETRY
1556 * permits vn_lock to return doomed vnodes.
1559 _vn_lock(struct vnode *vp, int flags, char *file, int line)
1563 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1564 ("vn_lock: no locktype"));
1565 VNASSERT(vp->v_holdcnt != 0, vp, ("vn_lock: zero hold count"));
1567 error = VOP_LOCK1(vp, flags, file, line);
1568 flags &= ~LK_INTERLOCK; /* Interlock is always dropped. */
1569 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1570 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1572 if ((flags & LK_RETRY) == 0) {
1573 if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0) {
1577 } else if (error != 0)
1583 * File table vnode close routine.
1586 vn_closefile(struct file *fp, struct thread *td)
1594 fp->f_ops = &badfileops;
1595 ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1597 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1599 if (__predict_false(ref)) {
1600 lf.l_whence = SEEK_SET;
1603 lf.l_type = F_UNLCK;
1604 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1611 vn_suspendable(struct mount *mp)
1614 return (mp->mnt_op->vfs_susp_clean != NULL);
1618 * Preparing to start a filesystem write operation. If the operation is
1619 * permitted, then we bump the count of operations in progress and
1620 * proceed. If a suspend request is in progress, we wait until the
1621 * suspension is over, and then proceed.
1624 vn_start_write_locked(struct mount *mp, int flags)
1628 mtx_assert(MNT_MTX(mp), MA_OWNED);
1632 * Check on status of suspension.
1634 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1635 mp->mnt_susp_owner != curthread) {
1636 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1637 (flags & PCATCH) : 0) | (PUSER - 1);
1638 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1639 if (flags & V_NOWAIT) {
1640 error = EWOULDBLOCK;
1643 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1649 if (flags & V_XSLEEP)
1651 mp->mnt_writeopcount++;
1653 if (error != 0 || (flags & V_XSLEEP) != 0)
1660 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1665 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1666 ("V_MNTREF requires mp"));
1670 * If a vnode is provided, get and return the mount point that
1671 * to which it will write.
1674 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1676 if (error != EOPNOTSUPP)
1681 if ((mp = *mpp) == NULL)
1684 if (!vn_suspendable(mp)) {
1685 if (vp != NULL || (flags & V_MNTREF) != 0)
1691 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1693 * As long as a vnode is not provided we need to acquire a
1694 * refcount for the provided mountpoint too, in order to
1695 * emulate a vfs_ref().
1698 if (vp == NULL && (flags & V_MNTREF) == 0)
1701 return (vn_start_write_locked(mp, flags));
1705 * Secondary suspension. Used by operations such as vop_inactive
1706 * routines that are needed by the higher level functions. These
1707 * are allowed to proceed until all the higher level functions have
1708 * completed (indicated by mnt_writeopcount dropping to zero). At that
1709 * time, these operations are halted until the suspension is over.
1712 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1717 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1718 ("V_MNTREF requires mp"));
1722 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1724 if (error != EOPNOTSUPP)
1730 * If we are not suspended or have not yet reached suspended
1731 * mode, then let the operation proceed.
1733 if ((mp = *mpp) == NULL)
1736 if (!vn_suspendable(mp)) {
1737 if (vp != NULL || (flags & V_MNTREF) != 0)
1743 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1745 * As long as a vnode is not provided we need to acquire a
1746 * refcount for the provided mountpoint too, in order to
1747 * emulate a vfs_ref().
1750 if (vp == NULL && (flags & V_MNTREF) == 0)
1752 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1753 mp->mnt_secondary_writes++;
1754 mp->mnt_secondary_accwrites++;
1758 if (flags & V_NOWAIT) {
1761 return (EWOULDBLOCK);
1764 * Wait for the suspension to finish.
1766 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1767 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1776 * Filesystem write operation has completed. If we are suspending and this
1777 * operation is the last one, notify the suspender that the suspension is
1781 vn_finished_write(struct mount *mp)
1783 if (mp == NULL || !vn_suspendable(mp))
1787 mp->mnt_writeopcount--;
1788 if (mp->mnt_writeopcount < 0)
1789 panic("vn_finished_write: neg cnt");
1790 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1791 mp->mnt_writeopcount <= 0)
1792 wakeup(&mp->mnt_writeopcount);
1798 * Filesystem secondary write operation has completed. If we are
1799 * suspending and this operation is the last one, notify the suspender
1800 * that the suspension is now in effect.
1803 vn_finished_secondary_write(struct mount *mp)
1805 if (mp == NULL || !vn_suspendable(mp))
1809 mp->mnt_secondary_writes--;
1810 if (mp->mnt_secondary_writes < 0)
1811 panic("vn_finished_secondary_write: neg cnt");
1812 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1813 mp->mnt_secondary_writes <= 0)
1814 wakeup(&mp->mnt_secondary_writes);
1821 * Request a filesystem to suspend write operations.
1824 vfs_write_suspend(struct mount *mp, int flags)
1828 MPASS(vn_suspendable(mp));
1831 if (mp->mnt_susp_owner == curthread) {
1835 while (mp->mnt_kern_flag & MNTK_SUSPEND)
1836 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1839 * Unmount holds a write reference on the mount point. If we
1840 * own busy reference and drain for writers, we deadlock with
1841 * the reference draining in the unmount path. Callers of
1842 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1843 * vfs_busy() reference is owned and caller is not in the
1846 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1847 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1852 mp->mnt_kern_flag |= MNTK_SUSPEND;
1853 mp->mnt_susp_owner = curthread;
1854 if (mp->mnt_writeopcount > 0)
1855 (void) msleep(&mp->mnt_writeopcount,
1856 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1859 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0)
1860 vfs_write_resume(mp, 0);
1865 * Request a filesystem to resume write operations.
1868 vfs_write_resume(struct mount *mp, int flags)
1871 MPASS(vn_suspendable(mp));
1874 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1875 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
1876 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
1878 mp->mnt_susp_owner = NULL;
1879 wakeup(&mp->mnt_writeopcount);
1880 wakeup(&mp->mnt_flag);
1881 curthread->td_pflags &= ~TDP_IGNSUSP;
1882 if ((flags & VR_START_WRITE) != 0) {
1884 mp->mnt_writeopcount++;
1887 if ((flags & VR_NO_SUSPCLR) == 0)
1889 } else if ((flags & VR_START_WRITE) != 0) {
1891 vn_start_write_locked(mp, 0);
1898 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
1902 vfs_write_suspend_umnt(struct mount *mp)
1906 MPASS(vn_suspendable(mp));
1907 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
1908 ("vfs_write_suspend_umnt: recursed"));
1910 /* dounmount() already called vn_start_write(). */
1912 vn_finished_write(mp);
1913 error = vfs_write_suspend(mp, 0);
1915 vn_start_write(NULL, &mp, V_WAIT);
1919 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
1922 vn_start_write(NULL, &mp, V_WAIT);
1924 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
1925 wakeup(&mp->mnt_flag);
1927 curthread->td_pflags |= TDP_IGNSUSP;
1932 * Implement kqueues for files by translating it to vnode operation.
1935 vn_kqfilter(struct file *fp, struct knote *kn)
1938 return (VOP_KQFILTER(fp->f_vnode, kn));
1942 * Simplified in-kernel wrapper calls for extended attribute access.
1943 * Both calls pass in a NULL credential, authorizing as "kernel" access.
1944 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
1947 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
1948 const char *attrname, int *buflen, char *buf, struct thread *td)
1954 iov.iov_len = *buflen;
1957 auio.uio_iov = &iov;
1958 auio.uio_iovcnt = 1;
1959 auio.uio_rw = UIO_READ;
1960 auio.uio_segflg = UIO_SYSSPACE;
1962 auio.uio_offset = 0;
1963 auio.uio_resid = *buflen;
1965 if ((ioflg & IO_NODELOCKED) == 0)
1966 vn_lock(vp, LK_SHARED | LK_RETRY);
1968 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
1970 /* authorize attribute retrieval as kernel */
1971 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
1974 if ((ioflg & IO_NODELOCKED) == 0)
1978 *buflen = *buflen - auio.uio_resid;
1985 * XXX failure mode if partially written?
1988 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
1989 const char *attrname, int buflen, char *buf, struct thread *td)
1996 iov.iov_len = buflen;
1999 auio.uio_iov = &iov;
2000 auio.uio_iovcnt = 1;
2001 auio.uio_rw = UIO_WRITE;
2002 auio.uio_segflg = UIO_SYSSPACE;
2004 auio.uio_offset = 0;
2005 auio.uio_resid = buflen;
2007 if ((ioflg & IO_NODELOCKED) == 0) {
2008 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2010 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2013 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2015 /* authorize attribute setting as kernel */
2016 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2018 if ((ioflg & IO_NODELOCKED) == 0) {
2019 vn_finished_write(mp);
2027 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2028 const char *attrname, struct thread *td)
2033 if ((ioflg & IO_NODELOCKED) == 0) {
2034 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2036 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2039 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2041 /* authorize attribute removal as kernel */
2042 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2043 if (error == EOPNOTSUPP)
2044 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2047 if ((ioflg & IO_NODELOCKED) == 0) {
2048 vn_finished_write(mp);
2056 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2060 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2064 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2067 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2072 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2073 int lkflags, struct vnode **rvp)
2078 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2080 ltype = VOP_ISLOCKED(vp);
2081 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2082 ("vn_vget_ino: vp not locked"));
2083 error = vfs_busy(mp, MBF_NOWAIT);
2087 error = vfs_busy(mp, 0);
2088 vn_lock(vp, ltype | LK_RETRY);
2092 if (vp->v_iflag & VI_DOOMED) {
2098 error = alloc(mp, alloc_arg, lkflags, rvp);
2101 vn_lock(vp, ltype | LK_RETRY);
2102 if (vp->v_iflag & VI_DOOMED) {
2115 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2119 if (vp->v_type != VREG || td == NULL)
2121 if ((uoff_t)uio->uio_offset + uio->uio_resid >
2122 lim_cur(td, RLIMIT_FSIZE)) {
2123 PROC_LOCK(td->td_proc);
2124 kern_psignal(td->td_proc, SIGXFSZ);
2125 PROC_UNLOCK(td->td_proc);
2132 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2139 vn_lock(vp, LK_SHARED | LK_RETRY);
2140 AUDIT_ARG_VNODE1(vp);
2143 return (setfmode(td, active_cred, vp, mode));
2147 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2154 vn_lock(vp, LK_SHARED | LK_RETRY);
2155 AUDIT_ARG_VNODE1(vp);
2158 return (setfown(td, active_cred, vp, uid, gid));
2162 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2166 if ((object = vp->v_object) == NULL)
2168 VM_OBJECT_WLOCK(object);
2169 vm_object_page_remove(object, start, end, 0);
2170 VM_OBJECT_WUNLOCK(object);
2174 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2182 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2183 ("Wrong command %lu", cmd));
2185 if (vn_lock(vp, LK_SHARED) != 0)
2187 if (vp->v_type != VREG) {
2191 error = VOP_GETATTR(vp, &va, cred);
2195 if (noff >= va.va_size) {
2199 bsize = vp->v_mount->mnt_stat.f_iosize;
2200 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2202 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2203 if (error == EOPNOTSUPP) {
2207 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2208 (bnp != -1 && cmd == FIOSEEKDATA)) {
2215 if (noff > va.va_size)
2217 /* noff == va.va_size. There is an implicit hole at the end of file. */
2218 if (cmd == FIOSEEKDATA)
2228 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2233 off_t foffset, size;
2236 cred = td->td_ucred;
2238 foffset = foffset_lock(fp, 0);
2239 noneg = (vp->v_type != VCHR);
2245 (offset > 0 && foffset > OFF_MAX - offset))) {
2252 vn_lock(vp, LK_SHARED | LK_RETRY);
2253 error = VOP_GETATTR(vp, &vattr, cred);
2259 * If the file references a disk device, then fetch
2260 * the media size and use that to determine the ending
2263 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2264 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2265 vattr.va_size = size;
2267 (vattr.va_size > OFF_MAX ||
2268 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2272 offset += vattr.va_size;
2277 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2280 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2285 if (error == 0 && noneg && offset < 0)
2289 VFS_KNOTE_UNLOCKED(vp, 0);
2290 td->td_uretoff.tdu_off = offset;
2292 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2297 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2303 * Grant permission if the caller is the owner of the file, or
2304 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2305 * on the file. If the time pointer is null, then write
2306 * permission on the file is also sufficient.
2308 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2309 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2310 * will be allowed to set the times [..] to the current
2313 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2314 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2315 error = VOP_ACCESS(vp, VWRITE, cred, td);
2320 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2325 if (fp->f_type == DTYPE_FIFO)
2326 kif->kf_type = KF_TYPE_FIFO;
2328 kif->kf_type = KF_TYPE_VNODE;
2331 FILEDESC_SUNLOCK(fdp);
2332 error = vn_fill_kinfo_vnode(vp, kif);
2334 FILEDESC_SLOCK(fdp);
2339 vn_fill_junk(struct kinfo_file *kif)
2344 * Simulate vn_fullpath returning changing values for a given
2345 * vp during e.g. coredump.
2347 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2348 olen = strlen(kif->kf_path);
2350 strcpy(&kif->kf_path[len - 1], "$");
2352 for (; olen < len; olen++)
2353 strcpy(&kif->kf_path[olen], "A");
2357 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2360 char *fullpath, *freepath;
2363 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2366 error = vn_fullpath(curthread, vp, &fullpath, &freepath);
2368 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2370 if (freepath != NULL)
2371 free(freepath, M_TEMP);
2373 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2378 * Retrieve vnode attributes.
2380 va.va_fsid = VNOVAL;
2382 vn_lock(vp, LK_SHARED | LK_RETRY);
2383 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2387 if (va.va_fsid != VNOVAL)
2388 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2390 kif->kf_un.kf_file.kf_file_fsid =
2391 vp->v_mount->mnt_stat.f_fsid.val[0];
2392 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2393 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2394 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2395 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2396 kif->kf_un.kf_file.kf_file_size = va.va_size;
2397 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2398 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2399 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2404 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2405 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2409 struct pmckern_map_in pkm;
2415 boolean_t writecounted;
2418 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2419 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2421 * POSIX shared-memory objects are defined to have
2422 * kernel persistence, and are not defined to support
2423 * read(2)/write(2) -- or even open(2). Thus, we can
2424 * use MAP_ASYNC to trade on-disk coherence for speed.
2425 * The shm_open(3) library routine turns on the FPOSIXSHM
2426 * flag to request this behavior.
2428 if ((fp->f_flag & FPOSIXSHM) != 0)
2429 flags |= MAP_NOSYNC;
2434 * Ensure that file and memory protections are
2435 * compatible. Note that we only worry about
2436 * writability if mapping is shared; in this case,
2437 * current and max prot are dictated by the open file.
2438 * XXX use the vnode instead? Problem is: what
2439 * credentials do we use for determination? What if
2440 * proc does a setuid?
2443 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2444 maxprot = VM_PROT_NONE;
2445 if ((prot & VM_PROT_EXECUTE) != 0)
2448 maxprot = VM_PROT_EXECUTE;
2449 if ((fp->f_flag & FREAD) != 0)
2450 maxprot |= VM_PROT_READ;
2451 else if ((prot & VM_PROT_READ) != 0)
2455 * If we are sharing potential changes via MAP_SHARED and we
2456 * are trying to get write permission although we opened it
2457 * without asking for it, bail out.
2459 if ((flags & MAP_SHARED) != 0) {
2460 if ((fp->f_flag & FWRITE) != 0)
2461 maxprot |= VM_PROT_WRITE;
2462 else if ((prot & VM_PROT_WRITE) != 0)
2465 maxprot |= VM_PROT_WRITE;
2466 cap_maxprot |= VM_PROT_WRITE;
2468 maxprot &= cap_maxprot;
2471 * For regular files and shared memory, POSIX requires that
2472 * the value of foff be a legitimate offset within the data
2473 * object. In particular, negative offsets are invalid.
2474 * Blocking negative offsets and overflows here avoids
2475 * possible wraparound or user-level access into reserved
2476 * ranges of the data object later. In contrast, POSIX does
2477 * not dictate how offsets are used by device drivers, so in
2478 * the case of a device mapping a negative offset is passed
2485 foff < 0 || foff > OFF_MAX - size)
2488 writecounted = FALSE;
2489 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2490 &foff, &object, &writecounted);
2493 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2494 foff, writecounted, td);
2497 * If this mapping was accounted for in the vnode's
2498 * writecount, then undo that now.
2501 vnode_pager_release_writecount(object, 0, size);
2502 vm_object_deallocate(object);
2505 /* Inform hwpmc(4) if an executable is being mapped. */
2506 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2507 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2509 pkm.pm_address = (uintptr_t) *addr;
2510 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2518 vn_fsid(struct vnode *vp, struct vattr *va)
2522 f = &vp->v_mount->mnt_stat.f_fsid;
2523 va->va_fsid = (uint32_t)f->val[1];
2524 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2525 va->va_fsid += (uint32_t)f->val[0];
2529 vn_fsync_buf(struct vnode *vp, int waitfor)
2531 struct buf *bp, *nbp;
2534 int error, maxretry;
2537 maxretry = 10000; /* large, arbitrarily chosen */
2539 if (vp->v_type == VCHR) {
2541 mp = vp->v_rdev->si_mountpt;
2548 * MARK/SCAN initialization to avoid infinite loops.
2550 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2551 bp->b_vflags &= ~BV_SCANNED;
2556 * Flush all dirty buffers associated with a vnode.
2559 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2560 if ((bp->b_vflags & BV_SCANNED) != 0)
2562 bp->b_vflags |= BV_SCANNED;
2563 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2564 if (waitfor != MNT_WAIT)
2567 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2568 BO_LOCKPTR(bo)) != 0) {
2575 KASSERT(bp->b_bufobj == bo,
2576 ("bp %p wrong b_bufobj %p should be %p",
2577 bp, bp->b_bufobj, bo));
2578 if ((bp->b_flags & B_DELWRI) == 0)
2579 panic("fsync: not dirty");
2580 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2586 if (maxretry < 1000)
2587 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2593 * If synchronous the caller expects us to completely resolve all
2594 * dirty buffers in the system. Wait for in-progress I/O to
2595 * complete (which could include background bitmap writes), then
2596 * retry if dirty blocks still exist.
2598 if (waitfor == MNT_WAIT) {
2599 bufobj_wwait(bo, 0, 0);
2600 if (bo->bo_dirty.bv_cnt > 0) {
2602 * If we are unable to write any of these buffers
2603 * then we fail now rather than trying endlessly
2604 * to write them out.
2606 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2607 if ((error = bp->b_error) != 0)
2609 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2610 (error == 0 && --maxretry >= 0))
2618 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2624 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
2625 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2626 * to do the actual copy.
2627 * vn_generic_copy_file_range() is factored out, so it can be called
2628 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2629 * different file systems.
2632 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2633 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2634 struct ucred *outcred, struct thread *fsize_td)
2639 uint64_t uvalin, uvalout;
2642 *lenp = 0; /* For error returns. */
2645 /* Do some sanity checks on the arguments. */
2650 if (invp->v_type == VDIR || outvp->v_type == VDIR)
2652 else if (*inoffp < 0 || uvalin > INT64_MAX || uvalin <
2653 (uint64_t)*inoffp || *outoffp < 0 || uvalout > INT64_MAX ||
2654 uvalout < (uint64_t)*outoffp || invp->v_type != VREG ||
2655 outvp->v_type != VREG)
2657 else if (invp == outvp)
2662 error = vn_lock(invp, LK_SHARED);
2665 /* Check that the offset + len does not go past EOF of invp. */
2666 error = VOP_GETATTR(invp, &va, incred);
2667 if (error == 0 && va.va_size < *inoffp + len)
2669 VOP_UNLOCK(invp, 0);
2674 * If the two vnode are for the same file system, call
2675 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2676 * which can handle copies across multiple file systems.
2679 if (invp->v_mount == outvp->v_mount)
2680 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2681 lenp, flags, incred, outcred, fsize_td);
2683 error = vn_generic_copy_file_range(invp, inoffp, outvp,
2684 outoffp, lenp, flags, incred, outcred, fsize_td);
2690 * Test len bytes of data starting at dat for all bytes == 0.
2691 * Return true if all bytes are zero, false otherwise.
2692 * Expects dat to be well aligned.
2695 mem_iszero(void *dat, int len)
2701 for (p = dat; len > 0; len -= sizeof(*p), p++) {
2702 if (len >= sizeof(*p)) {
2706 cp = (const char *)p;
2707 for (i = 0; i < len; i++, cp++)
2716 * Write an xfer sized chunk to outvp in blksize blocks from dat.
2717 * dat is a maximum of blksize in length and can be written repeatedly in
2719 * If growfile == true, just grow the file via vn_truncate_locked() instead
2720 * of doing actual writes.
2723 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
2724 u_long blksize, bool growfile, struct ucred *cred)
2731 * Loop around doing writes of blksize until write has been completed.
2732 * Lock/unlock on each loop iteration so that a bwillwrite() can be
2733 * done for each iteration, since the xfer argument can be very
2734 * large if there is a large hole to punch in the output file.
2739 error = vn_start_write(outvp, &mp, V_WAIT);
2741 if (MNT_SHARED_WRITES(mp))
2744 lckf = LK_EXCLUSIVE;
2745 error = vn_lock(outvp, lckf);
2749 error = vn_truncate_locked(outvp, outoff + xfer,
2752 xfer2 = MIN(xfer, blksize);
2753 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
2754 outoff, UIO_SYSSPACE, IO_NODELOCKED,
2755 curthread->td_ucred, cred, NULL, curthread);
2759 VOP_UNLOCK(outvp, 0);
2762 vn_finished_write(mp);
2763 } while (!growfile && xfer > 0 && error == 0);
2768 * Copy a byte range of one file to another. This function can handle the
2769 * case where invp and outvp are on different file systems.
2770 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
2771 * is no better file system specific way to do it.
2774 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
2775 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
2776 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
2781 off_t startoff, endoff, xfer, xfer2;
2784 bool cantseek, readzeros;
2786 size_t copylen, len, savlen;
2788 long holein, holeout;
2790 holein = holeout = 0;
2791 savlen = len = *lenp;
2795 error = vn_lock(invp, LK_SHARED);
2798 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
2800 VOP_UNLOCK(invp, 0);
2803 error = vn_start_write(outvp, &mp, V_WAIT);
2805 error = vn_lock(outvp, LK_EXCLUSIVE);
2808 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
2809 * now that outvp is locked.
2811 if (fsize_td != NULL) {
2812 io.uio_offset = *outoffp;
2814 error = vn_rlimit_fsize(outvp, &io, fsize_td);
2818 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
2821 * Holes that are past EOF do not need to be written as a block
2822 * of zero bytes. So, truncate the output file as far as
2823 * possible and then use va.va_size to decide if writing 0
2824 * bytes is necessary in the loop below.
2827 error = VOP_GETATTR(outvp, &va, outcred);
2828 if (error == 0 && va.va_size > *outoffp && va.va_size <=
2831 error = mac_vnode_check_write(curthread->td_ucred,
2835 error = vn_truncate_locked(outvp, *outoffp,
2838 va.va_size = *outoffp;
2840 VOP_UNLOCK(outvp, 0);
2843 vn_finished_write(mp);
2848 * Set the blksize to the larger of the hole sizes for invp and outvp.
2849 * If hole sizes aren't available, set the blksize to the larger
2850 * f_iosize of invp and outvp.
2851 * This code expects the hole sizes and f_iosizes to be powers of 2.
2852 * This value is clipped at 4Kbytes and 1Mbyte.
2854 blksize = MAX(holein, holeout);
2856 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
2857 outvp->v_mount->mnt_stat.f_iosize);
2860 else if (blksize > 1024 * 1024)
2861 blksize = 1024 * 1024;
2862 dat = malloc(blksize, M_TEMP, M_WAITOK);
2865 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
2866 * to find holes. Otherwise, just scan the read block for all 0s
2867 * in the inner loop where the data copying is done.
2868 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
2869 * support holes on the server, but do not support FIOSEEKHOLE.
2871 while (len > 0 && error == 0) {
2872 endoff = 0; /* To shut up compilers. */
2878 * Find the next data area. If there is just a hole to EOF,
2879 * FIOSEEKDATA should fail and then we drop down into the
2880 * inner loop and create the hole on the outvp file.
2881 * (I do not know if any file system will report a hole to
2882 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
2883 * will fail for those file systems.)
2885 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
2886 * the code just falls through to the inner copy loop.
2890 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
2894 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
2897 * Since invp is unlocked, it may be possible for
2898 * another thread to do a truncate(), lseek(), write()
2899 * creating a hole at startoff between the above
2900 * VOP_IOCTL() calls, if the other thread does not do
2902 * If that happens, startoff == endoff and finding
2903 * the hole has failed, so set an error.
2905 if (error == 0 && startoff == endoff)
2906 error = EINVAL; /* Any error. Reset to 0. */
2909 if (startoff > *inoffp) {
2910 /* Found hole before data block. */
2911 xfer = MIN(startoff - *inoffp, len);
2912 if (*outoffp < va.va_size) {
2913 /* Must write 0s to punch hole. */
2914 xfer2 = MIN(va.va_size - *outoffp,
2916 memset(dat, 0, MIN(xfer2, blksize));
2917 error = vn_write_outvp(outvp, dat,
2918 *outoffp, xfer2, blksize, false,
2922 if (error == 0 && *outoffp + xfer >
2923 va.va_size && xfer == len)
2924 /* Grow last block. */
2925 error = vn_write_outvp(outvp, dat,
2926 *outoffp, xfer, blksize, true,
2934 copylen = MIN(len, endoff - startoff);
2946 * Set first xfer to end at a block boundary, so that
2947 * holes are more likely detected in the loop below via
2948 * the for all bytes 0 method.
2950 xfer -= (*inoffp % blksize);
2952 /* Loop copying the data block. */
2953 while (copylen > 0 && error == 0) {
2956 error = vn_lock(invp, LK_SHARED);
2959 error = vn_rdwr(UIO_READ, invp, dat, xfer,
2960 startoff, UIO_SYSSPACE, IO_NODELOCKED,
2961 curthread->td_ucred, incred, &aresid,
2963 VOP_UNLOCK(invp, 0);
2965 * Linux considers a range that exceeds EOF to
2966 * be an error, so we will too.
2968 if (error == 0 && aresid > 0)
2972 * Skip the write for holes past the initial EOF
2973 * of the output file, unless this is the last
2974 * write of the output file at EOF.
2976 readzeros = cantseek ? mem_iszero(dat, xfer) :
2978 if (!cantseek || *outoffp < va.va_size ||
2979 xfer == len || !readzeros)
2980 error = vn_write_outvp(outvp, dat,
2981 *outoffp, xfer, blksize,
2982 readzeros && xfer == len &&
2983 *outoffp >= va.va_size, outcred);
2996 *lenp = savlen - len;