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>
60 #include <sys/limits.h>
63 #include <sys/mount.h>
64 #include <sys/mutex.h>
65 #include <sys/namei.h>
66 #include <sys/vnode.h>
69 #include <sys/filio.h>
70 #include <sys/resourcevar.h>
71 #include <sys/rwlock.h>
73 #include <sys/sysctl.h>
74 #include <sys/ttycom.h>
76 #include <sys/syslog.h>
77 #include <sys/unistd.h>
80 #include <security/audit/audit.h>
81 #include <security/mac/mac_framework.h>
84 #include <vm/vm_extern.h>
86 #include <vm/vm_map.h>
87 #include <vm/vm_object.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_pager.h>
92 #include <sys/pmckern.h>
95 static fo_rdwr_t vn_read;
96 static fo_rdwr_t vn_write;
97 static fo_rdwr_t vn_io_fault;
98 static fo_truncate_t vn_truncate;
99 static fo_ioctl_t vn_ioctl;
100 static fo_poll_t vn_poll;
101 static fo_kqfilter_t vn_kqfilter;
102 static fo_stat_t vn_statfile;
103 static fo_close_t vn_closefile;
104 static fo_mmap_t vn_mmap;
106 struct fileops vnops = {
107 .fo_read = vn_io_fault,
108 .fo_write = vn_io_fault,
109 .fo_truncate = vn_truncate,
110 .fo_ioctl = vn_ioctl,
112 .fo_kqfilter = vn_kqfilter,
113 .fo_stat = vn_statfile,
114 .fo_close = vn_closefile,
115 .fo_chmod = vn_chmod,
116 .fo_chown = vn_chown,
117 .fo_sendfile = vn_sendfile,
119 .fo_fill_kinfo = vn_fill_kinfo,
121 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
124 static const int io_hold_cnt = 16;
125 static int vn_io_fault_enable = 1;
126 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW,
127 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
128 static int vn_io_fault_prefault = 0;
129 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RW,
130 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
131 static u_long vn_io_faults_cnt;
132 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
133 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
135 static int vfs_allow_read_dir = 0;
136 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
137 &vfs_allow_read_dir, 0,
138 "Enable read(2) of directory for filesystems that support it");
141 * Returns true if vn_io_fault mode of handling the i/o request should
145 do_vn_io_fault(struct vnode *vp, struct uio *uio)
149 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
150 (mp = vp->v_mount) != NULL &&
151 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
155 * Structure used to pass arguments to vn_io_fault1(), to do either
156 * file- or vnode-based I/O calls.
158 struct vn_io_fault_args {
166 struct fop_args_tag {
170 struct vop_args_tag {
176 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
177 struct vn_io_fault_args *args, struct thread *td);
180 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
182 struct thread *td = ndp->ni_cnd.cn_thread;
184 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
188 open2nameif(int fmode, u_int vn_open_flags)
192 res = ISOPEN | LOCKLEAF;
193 if ((fmode & O_BENEATH) != 0)
195 if ((fmode & O_RESOLVE_BENEATH) != 0)
197 if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
199 if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
205 * Common code for vnode open operations via a name lookup.
206 * Lookup the vnode and invoke VOP_CREATE if needed.
207 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
209 * Note that this does NOT free nameidata for the successful case,
210 * due to the NDINIT being done elsewhere.
213 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
214 struct ucred *cred, struct file *fp)
218 struct thread *td = ndp->ni_cnd.cn_thread;
220 struct vattr *vap = &vat;
225 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
226 O_EXCL | O_DIRECTORY))
228 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
229 ndp->ni_cnd.cn_nameiop = CREATE;
230 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
232 * Set NOCACHE to avoid flushing the cache when
233 * rolling in many files at once.
235 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE;
236 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
237 ndp->ni_cnd.cn_flags |= FOLLOW;
238 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
240 if ((error = namei(ndp)) != 0)
242 if (ndp->ni_vp == NULL) {
245 vap->va_mode = cmode;
247 vap->va_vaflags |= VA_EXCLUSIVE;
248 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
249 NDFREE(ndp, NDF_ONLY_PNBUF);
251 if ((error = vn_start_write(NULL, &mp,
252 V_XSLEEP | PCATCH)) != 0)
256 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
257 ndp->ni_cnd.cn_flags |= MAKEENTRY;
259 error = mac_vnode_check_create(cred, ndp->ni_dvp,
263 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
266 vn_finished_write(mp);
268 NDFREE(ndp, NDF_ONLY_PNBUF);
274 if (ndp->ni_dvp == ndp->ni_vp)
280 if (fmode & O_EXCL) {
284 if (vp->v_type == VDIR) {
291 ndp->ni_cnd.cn_nameiop = LOOKUP;
292 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
293 ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
295 if ((fmode & FWRITE) == 0)
296 ndp->ni_cnd.cn_flags |= LOCKSHARED;
297 if ((error = namei(ndp)) != 0)
301 error = vn_open_vnode(vp, fmode, cred, td, fp);
307 NDFREE(ndp, NDF_ONLY_PNBUF);
315 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
318 int error, lock_flags, type;
320 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
321 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
323 KASSERT(fp != NULL, ("open with flock requires fp"));
324 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
327 lock_flags = VOP_ISLOCKED(vp);
330 lf.l_whence = SEEK_SET;
333 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
335 if ((fmode & FNONBLOCK) == 0)
337 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
339 fp->f_flag |= FHASLOCK;
341 vn_lock(vp, lock_flags | LK_RETRY);
342 if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0)
348 * Common code for vnode open operations once a vnode is located.
349 * Check permissions, and call the VOP_OPEN routine.
352 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
353 struct thread *td, struct file *fp)
358 if (vp->v_type == VLNK)
360 if (vp->v_type == VSOCK)
362 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
365 if (fmode & (FWRITE | O_TRUNC)) {
366 if (vp->v_type == VDIR)
374 if ((fmode & O_APPEND) && (fmode & FWRITE))
379 if (fmode & O_VERIFY)
381 error = mac_vnode_check_open(cred, vp, accmode);
385 accmode &= ~(VCREAT | VVERIFY);
387 if ((fmode & O_CREAT) == 0 && accmode != 0) {
388 error = VOP_ACCESS(vp, accmode, cred, td);
392 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
393 vn_lock(vp, LK_UPGRADE | LK_RETRY);
394 error = VOP_OPEN(vp, fmode, cred, td, fp);
398 error = vn_open_vnode_advlock(vp, fmode, fp);
399 if (error == 0 && (fmode & FWRITE) != 0) {
400 error = VOP_ADD_WRITECOUNT(vp, 1);
402 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
403 __func__, vp, vp->v_writecount);
408 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
409 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
410 * Arrange for that by having fdrop() to use vn_closefile().
413 fp->f_flag |= FOPENFAILED;
415 if (fp->f_ops == &badfileops) {
416 fp->f_type = DTYPE_VNODE;
422 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
428 * Check for write permissions on the specified vnode.
429 * Prototype text segments cannot be written.
433 vn_writechk(struct vnode *vp)
436 ASSERT_VOP_LOCKED(vp, "vn_writechk");
438 * If there's shared text associated with
439 * the vnode, try to free it up once. If
440 * we fail, we can't allow writing.
452 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
453 struct thread *td, bool keep_ref)
456 int error, lock_flags;
458 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
459 MNT_EXTENDED_SHARED(vp->v_mount))
460 lock_flags = LK_SHARED;
462 lock_flags = LK_EXCLUSIVE;
464 vn_start_write(vp, &mp, V_WAIT);
465 vn_lock(vp, lock_flags | LK_RETRY);
466 AUDIT_ARG_VNODE1(vp);
467 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
468 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
469 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
470 __func__, vp, vp->v_writecount);
472 error = VOP_CLOSE(vp, flags, file_cred, td);
477 vn_finished_write(mp);
482 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
486 return (vn_close1(vp, flags, file_cred, td, false));
490 * Heuristic to detect sequential operation.
493 sequential_heuristic(struct uio *uio, struct file *fp)
496 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
497 if (fp->f_flag & FRDAHEAD)
498 return (fp->f_seqcount << IO_SEQSHIFT);
501 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
502 * that the first I/O is normally considered to be slightly
503 * sequential. Seeking to offset 0 doesn't change sequentiality
504 * unless previous seeks have reduced f_seqcount to 0, in which
505 * case offset 0 is not special.
507 if ((uio->uio_offset == 0 && fp->f_seqcount > 0) ||
508 uio->uio_offset == fp->f_nextoff) {
510 * f_seqcount is in units of fixed-size blocks so that it
511 * depends mainly on the amount of sequential I/O and not
512 * much on the number of sequential I/O's. The fixed size
513 * of 16384 is hard-coded here since it is (not quite) just
514 * a magic size that works well here. This size is more
515 * closely related to the best I/O size for real disks than
516 * to any block size used by software.
518 if (uio->uio_resid >= IO_SEQMAX * 16384)
519 fp->f_seqcount = IO_SEQMAX;
521 fp->f_seqcount += howmany(uio->uio_resid, 16384);
522 if (fp->f_seqcount > IO_SEQMAX)
523 fp->f_seqcount = IO_SEQMAX;
525 return (fp->f_seqcount << IO_SEQSHIFT);
528 /* Not sequential. Quickly draw-down sequentiality. */
529 if (fp->f_seqcount > 1)
537 * Package up an I/O request on a vnode into a uio and do it.
540 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
541 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
542 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
549 struct vn_io_fault_args args;
550 int error, lock_flags;
552 if (offset < 0 && vp->v_type != VCHR)
554 auio.uio_iov = &aiov;
556 aiov.iov_base = base;
558 auio.uio_resid = len;
559 auio.uio_offset = offset;
560 auio.uio_segflg = segflg;
565 if ((ioflg & IO_NODELOCKED) == 0) {
566 if ((ioflg & IO_RANGELOCKED) == 0) {
567 if (rw == UIO_READ) {
568 rl_cookie = vn_rangelock_rlock(vp, offset,
571 rl_cookie = vn_rangelock_wlock(vp, offset,
577 if (rw == UIO_WRITE) {
578 if (vp->v_type != VCHR &&
579 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
582 if (MNT_SHARED_WRITES(mp) ||
583 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
584 lock_flags = LK_SHARED;
586 lock_flags = LK_EXCLUSIVE;
588 lock_flags = LK_SHARED;
589 vn_lock(vp, lock_flags | LK_RETRY);
593 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
595 if ((ioflg & IO_NOMACCHECK) == 0) {
597 error = mac_vnode_check_read(active_cred, file_cred,
600 error = mac_vnode_check_write(active_cred, file_cred,
605 if (file_cred != NULL)
609 if (do_vn_io_fault(vp, &auio)) {
610 args.kind = VN_IO_FAULT_VOP;
613 args.args.vop_args.vp = vp;
614 error = vn_io_fault1(vp, &auio, &args, td);
615 } else if (rw == UIO_READ) {
616 error = VOP_READ(vp, &auio, ioflg, cred);
617 } else /* if (rw == UIO_WRITE) */ {
618 error = VOP_WRITE(vp, &auio, ioflg, cred);
622 *aresid = auio.uio_resid;
624 if (auio.uio_resid && error == 0)
626 if ((ioflg & IO_NODELOCKED) == 0) {
629 vn_finished_write(mp);
632 if (rl_cookie != NULL)
633 vn_rangelock_unlock(vp, rl_cookie);
638 * Package up an I/O request on a vnode into a uio and do it. The I/O
639 * request is split up into smaller chunks and we try to avoid saturating
640 * the buffer cache while potentially holding a vnode locked, so we
641 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
642 * to give other processes a chance to lock the vnode (either other processes
643 * core'ing the same binary, or unrelated processes scanning the directory).
646 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
647 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
648 struct ucred *file_cred, size_t *aresid, struct thread *td)
657 * Force `offset' to a multiple of MAXBSIZE except possibly
658 * for the first chunk, so that filesystems only need to
659 * write full blocks except possibly for the first and last
662 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
666 if (rw != UIO_READ && vp->v_type == VREG)
669 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
670 ioflg, active_cred, file_cred, &iaresid, td);
671 len -= chunk; /* aresid calc already includes length */
675 base = (char *)base + chunk;
676 kern_yield(PRI_USER);
679 *aresid = len + iaresid;
684 foffset_lock(struct file *fp, int flags)
689 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
691 #if OFF_MAX <= LONG_MAX
693 * Caller only wants the current f_offset value. Assume that
694 * the long and shorter integer types reads are atomic.
696 if ((flags & FOF_NOLOCK) != 0)
697 return (fp->f_offset);
701 * According to McKusick the vn lock was protecting f_offset here.
702 * It is now protected by the FOFFSET_LOCKED flag.
704 mtxp = mtx_pool_find(mtxpool_sleep, fp);
706 if ((flags & FOF_NOLOCK) == 0) {
707 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
708 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
709 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
712 fp->f_vnread_flags |= FOFFSET_LOCKED;
720 foffset_unlock(struct file *fp, off_t val, int flags)
724 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
726 #if OFF_MAX <= LONG_MAX
727 if ((flags & FOF_NOLOCK) != 0) {
728 if ((flags & FOF_NOUPDATE) == 0)
730 if ((flags & FOF_NEXTOFF) != 0)
736 mtxp = mtx_pool_find(mtxpool_sleep, fp);
738 if ((flags & FOF_NOUPDATE) == 0)
740 if ((flags & FOF_NEXTOFF) != 0)
742 if ((flags & FOF_NOLOCK) == 0) {
743 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
744 ("Lost FOFFSET_LOCKED"));
745 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
746 wakeup(&fp->f_vnread_flags);
747 fp->f_vnread_flags = 0;
753 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
756 if ((flags & FOF_OFFSET) == 0)
757 uio->uio_offset = foffset_lock(fp, flags);
761 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
764 if ((flags & FOF_OFFSET) == 0)
765 foffset_unlock(fp, uio->uio_offset, flags);
769 get_advice(struct file *fp, struct uio *uio)
774 ret = POSIX_FADV_NORMAL;
775 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
778 mtxp = mtx_pool_find(mtxpool_sleep, fp);
780 if (fp->f_advice != NULL &&
781 uio->uio_offset >= fp->f_advice->fa_start &&
782 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
783 ret = fp->f_advice->fa_advice;
789 * File table vnode read routine.
792 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
800 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
802 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
805 if (fp->f_flag & FNONBLOCK)
807 if (fp->f_flag & O_DIRECT)
809 advice = get_advice(fp, uio);
810 vn_lock(vp, LK_SHARED | LK_RETRY);
813 case POSIX_FADV_NORMAL:
814 case POSIX_FADV_SEQUENTIAL:
815 case POSIX_FADV_NOREUSE:
816 ioflag |= sequential_heuristic(uio, fp);
818 case POSIX_FADV_RANDOM:
819 /* Disable read-ahead for random I/O. */
822 orig_offset = uio->uio_offset;
825 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
828 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
829 fp->f_nextoff = uio->uio_offset;
831 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
832 orig_offset != uio->uio_offset)
834 * Use POSIX_FADV_DONTNEED to flush pages and buffers
835 * for the backing file after a POSIX_FADV_NOREUSE
838 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
839 POSIX_FADV_DONTNEED);
844 * File table vnode write routine.
847 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
853 int error, ioflag, lock_flags;
856 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
858 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
860 if (vp->v_type == VREG)
863 if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
865 if (fp->f_flag & FNONBLOCK)
867 if (fp->f_flag & O_DIRECT)
869 if ((fp->f_flag & O_FSYNC) ||
870 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
873 if (vp->v_type != VCHR &&
874 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
877 advice = get_advice(fp, uio);
879 if (MNT_SHARED_WRITES(mp) ||
880 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
881 lock_flags = LK_SHARED;
883 lock_flags = LK_EXCLUSIVE;
886 vn_lock(vp, lock_flags | LK_RETRY);
888 case POSIX_FADV_NORMAL:
889 case POSIX_FADV_SEQUENTIAL:
890 case POSIX_FADV_NOREUSE:
891 ioflag |= sequential_heuristic(uio, fp);
893 case POSIX_FADV_RANDOM:
894 /* XXX: Is this correct? */
897 orig_offset = uio->uio_offset;
900 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
903 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
904 fp->f_nextoff = uio->uio_offset;
906 if (vp->v_type != VCHR)
907 vn_finished_write(mp);
908 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
909 orig_offset != uio->uio_offset)
911 * Use POSIX_FADV_DONTNEED to flush pages and buffers
912 * for the backing file after a POSIX_FADV_NOREUSE
915 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
916 POSIX_FADV_DONTNEED);
922 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
923 * prevent the following deadlock:
925 * Assume that the thread A reads from the vnode vp1 into userspace
926 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
927 * currently not resident, then system ends up with the call chain
928 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
929 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
930 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
931 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
932 * backed by the pages of vnode vp1, and some page in buf2 is not
933 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
935 * To prevent the lock order reversal and deadlock, vn_io_fault() does
936 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
937 * Instead, it first tries to do the whole range i/o with pagefaults
938 * disabled. If all pages in the i/o buffer are resident and mapped,
939 * VOP will succeed (ignoring the genuine filesystem errors).
940 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
941 * i/o in chunks, with all pages in the chunk prefaulted and held
942 * using vm_fault_quick_hold_pages().
944 * Filesystems using this deadlock avoidance scheme should use the
945 * array of the held pages from uio, saved in the curthread->td_ma,
946 * instead of doing uiomove(). A helper function
947 * vn_io_fault_uiomove() converts uiomove request into
948 * uiomove_fromphys() over td_ma array.
950 * Since vnode locks do not cover the whole i/o anymore, rangelocks
951 * make the current i/o request atomic with respect to other i/os and
956 * Decode vn_io_fault_args and perform the corresponding i/o.
959 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
965 save = vm_fault_disable_pagefaults();
966 switch (args->kind) {
967 case VN_IO_FAULT_FOP:
968 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
969 uio, args->cred, args->flags, td);
971 case VN_IO_FAULT_VOP:
972 if (uio->uio_rw == UIO_READ) {
973 error = VOP_READ(args->args.vop_args.vp, uio,
974 args->flags, args->cred);
975 } else if (uio->uio_rw == UIO_WRITE) {
976 error = VOP_WRITE(args->args.vop_args.vp, uio,
977 args->flags, args->cred);
981 panic("vn_io_fault_doio: unknown kind of io %d %d",
982 args->kind, uio->uio_rw);
984 vm_fault_enable_pagefaults(save);
989 vn_io_fault_touch(char *base, const struct uio *uio)
994 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1000 vn_io_fault_prefault_user(const struct uio *uio)
1003 const struct iovec *iov;
1008 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1009 ("vn_io_fault_prefault userspace"));
1013 resid = uio->uio_resid;
1014 base = iov->iov_base;
1017 error = vn_io_fault_touch(base, uio);
1020 if (len < PAGE_SIZE) {
1022 error = vn_io_fault_touch(base + len - 1, uio);
1027 if (++i >= uio->uio_iovcnt)
1029 iov = uio->uio_iov + i;
1030 base = iov->iov_base;
1042 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1043 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1044 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1045 * into args and call vn_io_fault1() to handle faults during the user
1046 * mode buffer accesses.
1049 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1052 vm_page_t ma[io_hold_cnt + 2];
1053 struct uio *uio_clone, short_uio;
1054 struct iovec short_iovec[1];
1055 vm_page_t *prev_td_ma;
1057 vm_offset_t addr, end;
1060 int error, cnt, saveheld, prev_td_ma_cnt;
1062 if (vn_io_fault_prefault) {
1063 error = vn_io_fault_prefault_user(uio);
1065 return (error); /* Or ignore ? */
1068 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1071 * The UFS follows IO_UNIT directive and replays back both
1072 * uio_offset and uio_resid if an error is encountered during the
1073 * operation. But, since the iovec may be already advanced,
1074 * uio is still in an inconsistent state.
1076 * Cache a copy of the original uio, which is advanced to the redo
1077 * point using UIO_NOCOPY below.
1079 uio_clone = cloneuio(uio);
1080 resid = uio->uio_resid;
1082 short_uio.uio_segflg = UIO_USERSPACE;
1083 short_uio.uio_rw = uio->uio_rw;
1084 short_uio.uio_td = uio->uio_td;
1086 error = vn_io_fault_doio(args, uio, td);
1087 if (error != EFAULT)
1090 atomic_add_long(&vn_io_faults_cnt, 1);
1091 uio_clone->uio_segflg = UIO_NOCOPY;
1092 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1093 uio_clone->uio_segflg = uio->uio_segflg;
1095 saveheld = curthread_pflags_set(TDP_UIOHELD);
1096 prev_td_ma = td->td_ma;
1097 prev_td_ma_cnt = td->td_ma_cnt;
1099 while (uio_clone->uio_resid != 0) {
1100 len = uio_clone->uio_iov->iov_len;
1102 KASSERT(uio_clone->uio_iovcnt >= 1,
1103 ("iovcnt underflow"));
1104 uio_clone->uio_iov++;
1105 uio_clone->uio_iovcnt--;
1108 if (len > io_hold_cnt * PAGE_SIZE)
1109 len = io_hold_cnt * PAGE_SIZE;
1110 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1111 end = round_page(addr + len);
1116 cnt = atop(end - trunc_page(addr));
1118 * A perfectly misaligned address and length could cause
1119 * both the start and the end of the chunk to use partial
1120 * page. +2 accounts for such a situation.
1122 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1123 addr, len, prot, ma, io_hold_cnt + 2);
1128 short_uio.uio_iov = &short_iovec[0];
1129 short_iovec[0].iov_base = (void *)addr;
1130 short_uio.uio_iovcnt = 1;
1131 short_uio.uio_resid = short_iovec[0].iov_len = len;
1132 short_uio.uio_offset = uio_clone->uio_offset;
1134 td->td_ma_cnt = cnt;
1136 error = vn_io_fault_doio(args, &short_uio, td);
1137 vm_page_unhold_pages(ma, cnt);
1138 adv = len - short_uio.uio_resid;
1140 uio_clone->uio_iov->iov_base =
1141 (char *)uio_clone->uio_iov->iov_base + adv;
1142 uio_clone->uio_iov->iov_len -= adv;
1143 uio_clone->uio_resid -= adv;
1144 uio_clone->uio_offset += adv;
1146 uio->uio_resid -= adv;
1147 uio->uio_offset += adv;
1149 if (error != 0 || adv == 0)
1152 td->td_ma = prev_td_ma;
1153 td->td_ma_cnt = prev_td_ma_cnt;
1154 curthread_pflags_restore(saveheld);
1156 free(uio_clone, M_IOV);
1161 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1162 int flags, struct thread *td)
1167 struct vn_io_fault_args args;
1170 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1174 * The ability to read(2) on a directory has historically been
1175 * allowed for all users, but this can and has been the source of
1176 * at least one security issue in the past. As such, it is now hidden
1177 * away behind a sysctl for those that actually need it to use it.
1179 if (vp->v_type == VDIR) {
1180 KASSERT(uio->uio_rw == UIO_READ,
1181 ("illegal write attempted on a directory"));
1182 if (!vfs_allow_read_dir)
1186 foffset_lock_uio(fp, uio, flags);
1187 if (do_vn_io_fault(vp, uio)) {
1188 args.kind = VN_IO_FAULT_FOP;
1189 args.args.fop_args.fp = fp;
1190 args.args.fop_args.doio = doio;
1191 args.cred = active_cred;
1192 args.flags = flags | FOF_OFFSET;
1193 if (uio->uio_rw == UIO_READ) {
1194 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1195 uio->uio_offset + uio->uio_resid);
1196 } else if ((fp->f_flag & O_APPEND) != 0 ||
1197 (flags & FOF_OFFSET) == 0) {
1198 /* For appenders, punt and lock the whole range. */
1199 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1201 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1202 uio->uio_offset + uio->uio_resid);
1204 error = vn_io_fault1(vp, uio, &args, td);
1205 vn_rangelock_unlock(vp, rl_cookie);
1207 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1209 foffset_unlock_uio(fp, uio, flags);
1214 * Helper function to perform the requested uiomove operation using
1215 * the held pages for io->uio_iov[0].iov_base buffer instead of
1216 * copyin/copyout. Access to the pages with uiomove_fromphys()
1217 * instead of iov_base prevents page faults that could occur due to
1218 * pmap_collect() invalidating the mapping created by
1219 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1220 * object cleanup revoking the write access from page mappings.
1222 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1223 * instead of plain uiomove().
1226 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1228 struct uio transp_uio;
1229 struct iovec transp_iov[1];
1235 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1236 uio->uio_segflg != UIO_USERSPACE)
1237 return (uiomove(data, xfersize, uio));
1239 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1240 transp_iov[0].iov_base = data;
1241 transp_uio.uio_iov = &transp_iov[0];
1242 transp_uio.uio_iovcnt = 1;
1243 if (xfersize > uio->uio_resid)
1244 xfersize = uio->uio_resid;
1245 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1246 transp_uio.uio_offset = 0;
1247 transp_uio.uio_segflg = UIO_SYSSPACE;
1249 * Since transp_iov points to data, and td_ma page array
1250 * corresponds to original uio->uio_iov, we need to invert the
1251 * direction of the i/o operation as passed to
1252 * uiomove_fromphys().
1254 switch (uio->uio_rw) {
1256 transp_uio.uio_rw = UIO_READ;
1259 transp_uio.uio_rw = UIO_WRITE;
1262 transp_uio.uio_td = uio->uio_td;
1263 error = uiomove_fromphys(td->td_ma,
1264 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1265 xfersize, &transp_uio);
1266 adv = xfersize - transp_uio.uio_resid;
1268 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1269 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1271 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1273 td->td_ma_cnt -= pgadv;
1274 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1275 uio->uio_iov->iov_len -= adv;
1276 uio->uio_resid -= adv;
1277 uio->uio_offset += adv;
1282 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1286 vm_offset_t iov_base;
1290 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1291 uio->uio_segflg != UIO_USERSPACE)
1292 return (uiomove_fromphys(ma, offset, xfersize, uio));
1294 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1295 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1296 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1297 switch (uio->uio_rw) {
1299 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1303 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1307 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1309 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1311 td->td_ma_cnt -= pgadv;
1312 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1313 uio->uio_iov->iov_len -= cnt;
1314 uio->uio_resid -= cnt;
1315 uio->uio_offset += cnt;
1321 * File table truncate routine.
1324 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1336 * Lock the whole range for truncation. Otherwise split i/o
1337 * might happen partly before and partly after the truncation.
1339 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1340 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1343 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1344 AUDIT_ARG_VNODE1(vp);
1345 if (vp->v_type == VDIR) {
1350 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1354 error = VOP_ADD_WRITECOUNT(vp, 1);
1357 vattr.va_size = length;
1358 if ((fp->f_flag & O_FSYNC) != 0)
1359 vattr.va_vaflags |= VA_SYNC;
1360 error = VOP_SETATTR(vp, &vattr, fp->f_cred);
1361 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1365 vn_finished_write(mp);
1367 vn_rangelock_unlock(vp, rl_cookie);
1372 * File table vnode stat routine.
1375 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1378 struct vnode *vp = fp->f_vnode;
1381 vn_lock(vp, LK_SHARED | LK_RETRY);
1382 error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1389 * Stat a vnode; implementation for the stat syscall
1392 vn_stat(struct vnode *vp, struct stat *sb, struct ucred *active_cred,
1393 struct ucred *file_cred, struct thread *td)
1400 AUDIT_ARG_VNODE1(vp);
1402 error = mac_vnode_check_stat(active_cred, file_cred, vp);
1410 * Initialize defaults for new and unusual fields, so that file
1411 * systems which don't support these fields don't need to know
1414 vap->va_birthtime.tv_sec = -1;
1415 vap->va_birthtime.tv_nsec = 0;
1416 vap->va_fsid = VNOVAL;
1417 vap->va_rdev = NODEV;
1419 error = VOP_GETATTR(vp, vap, active_cred);
1424 * Zero the spare stat fields
1426 bzero(sb, sizeof *sb);
1429 * Copy from vattr table
1431 if (vap->va_fsid != VNOVAL)
1432 sb->st_dev = vap->va_fsid;
1434 sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1435 sb->st_ino = vap->va_fileid;
1436 mode = vap->va_mode;
1437 switch (vap->va_type) {
1463 sb->st_nlink = vap->va_nlink;
1464 sb->st_uid = vap->va_uid;
1465 sb->st_gid = vap->va_gid;
1466 sb->st_rdev = vap->va_rdev;
1467 if (vap->va_size > OFF_MAX)
1469 sb->st_size = vap->va_size;
1470 sb->st_atim = vap->va_atime;
1471 sb->st_mtim = vap->va_mtime;
1472 sb->st_ctim = vap->va_ctime;
1473 sb->st_birthtim = vap->va_birthtime;
1476 * According to www.opengroup.org, the meaning of st_blksize is
1477 * "a filesystem-specific preferred I/O block size for this
1478 * object. In some filesystem types, this may vary from file
1480 * Use minimum/default of PAGE_SIZE (e.g. for VCHR).
1483 sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1485 sb->st_flags = vap->va_flags;
1486 if (priv_check(td, PRIV_VFS_GENERATION))
1489 sb->st_gen = vap->va_gen;
1491 sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1496 * File table vnode ioctl routine.
1499 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1504 struct fiobmap2_arg *bmarg;
1508 switch (vp->v_type) {
1513 vn_lock(vp, LK_SHARED | LK_RETRY);
1514 error = VOP_GETATTR(vp, &vattr, active_cred);
1517 *(int *)data = vattr.va_size - fp->f_offset;
1520 bmarg = (struct fiobmap2_arg *)data;
1521 vn_lock(vp, LK_SHARED | LK_RETRY);
1523 error = mac_vnode_check_read(active_cred, fp->f_cred,
1527 error = VOP_BMAP(vp, bmarg->bn, NULL,
1528 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1535 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1540 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1548 * File table vnode poll routine.
1551 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1559 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1560 AUDIT_ARG_VNODE1(vp);
1561 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1566 error = VOP_POLL(vp, events, fp->f_cred, td);
1571 * Acquire the requested lock and then check for validity. LK_RETRY
1572 * permits vn_lock to return doomed vnodes.
1575 _vn_lock(struct vnode *vp, int flags, char *file, int line)
1579 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1580 ("vn_lock: no locktype"));
1581 VNASSERT(vp->v_holdcnt != 0, vp, ("vn_lock: zero hold count"));
1583 error = VOP_LOCK1(vp, flags, file, line);
1584 flags &= ~LK_INTERLOCK; /* Interlock is always dropped. */
1585 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1586 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1588 if ((flags & LK_RETRY) == 0) {
1589 if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0) {
1593 } else if (error != 0)
1599 * File table vnode close routine.
1602 vn_closefile(struct file *fp, struct thread *td)
1610 fp->f_ops = &badfileops;
1611 ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1613 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1615 if (__predict_false(ref)) {
1616 lf.l_whence = SEEK_SET;
1619 lf.l_type = F_UNLCK;
1620 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1627 vn_suspendable(struct mount *mp)
1630 return (mp->mnt_op->vfs_susp_clean != NULL);
1634 * Preparing to start a filesystem write operation. If the operation is
1635 * permitted, then we bump the count of operations in progress and
1636 * proceed. If a suspend request is in progress, we wait until the
1637 * suspension is over, and then proceed.
1640 vn_start_write_locked(struct mount *mp, int flags)
1644 mtx_assert(MNT_MTX(mp), MA_OWNED);
1648 * Check on status of suspension.
1650 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1651 mp->mnt_susp_owner != curthread) {
1652 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1653 (flags & PCATCH) : 0) | (PUSER - 1);
1654 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1655 if (flags & V_NOWAIT) {
1656 error = EWOULDBLOCK;
1659 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1665 if (flags & V_XSLEEP)
1667 mp->mnt_writeopcount++;
1669 if (error != 0 || (flags & V_XSLEEP) != 0)
1676 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1681 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1682 ("V_MNTREF requires mp"));
1686 * If a vnode is provided, get and return the mount point that
1687 * to which it will write.
1690 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1692 if (error != EOPNOTSUPP)
1697 if ((mp = *mpp) == NULL)
1700 if (!vn_suspendable(mp)) {
1701 if (vp != NULL || (flags & V_MNTREF) != 0)
1707 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1709 * As long as a vnode is not provided we need to acquire a
1710 * refcount for the provided mountpoint too, in order to
1711 * emulate a vfs_ref().
1714 if (vp == NULL && (flags & V_MNTREF) == 0)
1717 return (vn_start_write_locked(mp, flags));
1721 * Secondary suspension. Used by operations such as vop_inactive
1722 * routines that are needed by the higher level functions. These
1723 * are allowed to proceed until all the higher level functions have
1724 * completed (indicated by mnt_writeopcount dropping to zero). At that
1725 * time, these operations are halted until the suspension is over.
1728 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1733 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1734 ("V_MNTREF requires mp"));
1738 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1740 if (error != EOPNOTSUPP)
1746 * If we are not suspended or have not yet reached suspended
1747 * mode, then let the operation proceed.
1749 if ((mp = *mpp) == NULL)
1752 if (!vn_suspendable(mp)) {
1753 if (vp != NULL || (flags & V_MNTREF) != 0)
1759 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1761 * As long as a vnode is not provided we need to acquire a
1762 * refcount for the provided mountpoint too, in order to
1763 * emulate a vfs_ref().
1766 if (vp == NULL && (flags & V_MNTREF) == 0)
1768 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1769 mp->mnt_secondary_writes++;
1770 mp->mnt_secondary_accwrites++;
1774 if (flags & V_NOWAIT) {
1777 return (EWOULDBLOCK);
1780 * Wait for the suspension to finish.
1782 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1783 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1792 * Filesystem write operation has completed. If we are suspending and this
1793 * operation is the last one, notify the suspender that the suspension is
1797 vn_finished_write(struct mount *mp)
1799 if (mp == NULL || !vn_suspendable(mp))
1803 mp->mnt_writeopcount--;
1804 if (mp->mnt_writeopcount < 0)
1805 panic("vn_finished_write: neg cnt");
1806 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1807 mp->mnt_writeopcount <= 0)
1808 wakeup(&mp->mnt_writeopcount);
1814 * Filesystem secondary write operation has completed. If we are
1815 * suspending and this operation is the last one, notify the suspender
1816 * that the suspension is now in effect.
1819 vn_finished_secondary_write(struct mount *mp)
1821 if (mp == NULL || !vn_suspendable(mp))
1825 mp->mnt_secondary_writes--;
1826 if (mp->mnt_secondary_writes < 0)
1827 panic("vn_finished_secondary_write: neg cnt");
1828 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1829 mp->mnt_secondary_writes <= 0)
1830 wakeup(&mp->mnt_secondary_writes);
1837 * Request a filesystem to suspend write operations.
1840 vfs_write_suspend(struct mount *mp, int flags)
1844 MPASS(vn_suspendable(mp));
1847 if (mp->mnt_susp_owner == curthread) {
1851 while (mp->mnt_kern_flag & MNTK_SUSPEND)
1852 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1855 * Unmount holds a write reference on the mount point. If we
1856 * own busy reference and drain for writers, we deadlock with
1857 * the reference draining in the unmount path. Callers of
1858 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1859 * vfs_busy() reference is owned and caller is not in the
1862 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1863 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1868 mp->mnt_kern_flag |= MNTK_SUSPEND;
1869 mp->mnt_susp_owner = curthread;
1870 if (mp->mnt_writeopcount > 0)
1871 (void) msleep(&mp->mnt_writeopcount,
1872 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1875 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0)
1876 vfs_write_resume(mp, 0);
1881 * Request a filesystem to resume write operations.
1884 vfs_write_resume(struct mount *mp, int flags)
1887 MPASS(vn_suspendable(mp));
1890 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1891 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
1892 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
1894 mp->mnt_susp_owner = NULL;
1895 wakeup(&mp->mnt_writeopcount);
1896 wakeup(&mp->mnt_flag);
1897 curthread->td_pflags &= ~TDP_IGNSUSP;
1898 if ((flags & VR_START_WRITE) != 0) {
1900 mp->mnt_writeopcount++;
1903 if ((flags & VR_NO_SUSPCLR) == 0)
1905 } else if ((flags & VR_START_WRITE) != 0) {
1907 vn_start_write_locked(mp, 0);
1914 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
1918 vfs_write_suspend_umnt(struct mount *mp)
1922 MPASS(vn_suspendable(mp));
1923 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
1924 ("vfs_write_suspend_umnt: recursed"));
1926 /* dounmount() already called vn_start_write(). */
1928 vn_finished_write(mp);
1929 error = vfs_write_suspend(mp, 0);
1931 vn_start_write(NULL, &mp, V_WAIT);
1935 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
1938 vn_start_write(NULL, &mp, V_WAIT);
1940 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
1941 wakeup(&mp->mnt_flag);
1943 curthread->td_pflags |= TDP_IGNSUSP;
1948 * Implement kqueues for files by translating it to vnode operation.
1951 vn_kqfilter(struct file *fp, struct knote *kn)
1954 return (VOP_KQFILTER(fp->f_vnode, kn));
1958 * Simplified in-kernel wrapper calls for extended attribute access.
1959 * Both calls pass in a NULL credential, authorizing as "kernel" access.
1960 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
1963 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
1964 const char *attrname, int *buflen, char *buf, struct thread *td)
1970 iov.iov_len = *buflen;
1973 auio.uio_iov = &iov;
1974 auio.uio_iovcnt = 1;
1975 auio.uio_rw = UIO_READ;
1976 auio.uio_segflg = UIO_SYSSPACE;
1978 auio.uio_offset = 0;
1979 auio.uio_resid = *buflen;
1981 if ((ioflg & IO_NODELOCKED) == 0)
1982 vn_lock(vp, LK_SHARED | LK_RETRY);
1984 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
1986 /* authorize attribute retrieval as kernel */
1987 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
1990 if ((ioflg & IO_NODELOCKED) == 0)
1994 *buflen = *buflen - auio.uio_resid;
2001 * XXX failure mode if partially written?
2004 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2005 const char *attrname, int buflen, char *buf, struct thread *td)
2012 iov.iov_len = buflen;
2015 auio.uio_iov = &iov;
2016 auio.uio_iovcnt = 1;
2017 auio.uio_rw = UIO_WRITE;
2018 auio.uio_segflg = UIO_SYSSPACE;
2020 auio.uio_offset = 0;
2021 auio.uio_resid = buflen;
2023 if ((ioflg & IO_NODELOCKED) == 0) {
2024 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2026 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2029 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2031 /* authorize attribute setting as kernel */
2032 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2034 if ((ioflg & IO_NODELOCKED) == 0) {
2035 vn_finished_write(mp);
2043 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2044 const char *attrname, struct thread *td)
2049 if ((ioflg & IO_NODELOCKED) == 0) {
2050 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2052 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2055 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2057 /* authorize attribute removal as kernel */
2058 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2059 if (error == EOPNOTSUPP)
2060 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2063 if ((ioflg & IO_NODELOCKED) == 0) {
2064 vn_finished_write(mp);
2072 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2076 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2080 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2083 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2088 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2089 int lkflags, struct vnode **rvp)
2094 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2096 ltype = VOP_ISLOCKED(vp);
2097 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2098 ("vn_vget_ino: vp not locked"));
2099 error = vfs_busy(mp, MBF_NOWAIT);
2103 error = vfs_busy(mp, 0);
2104 vn_lock(vp, ltype | LK_RETRY);
2108 if (vp->v_iflag & VI_DOOMED) {
2114 error = alloc(mp, alloc_arg, lkflags, rvp);
2116 if (error != 0 || *rvp != vp)
2117 vn_lock(vp, ltype | LK_RETRY);
2118 if (vp->v_iflag & VI_DOOMED) {
2131 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2135 if (vp->v_type != VREG || td == NULL)
2137 if ((uoff_t)uio->uio_offset + uio->uio_resid >
2138 lim_cur(td, RLIMIT_FSIZE)) {
2139 PROC_LOCK(td->td_proc);
2140 kern_psignal(td->td_proc, SIGXFSZ);
2141 PROC_UNLOCK(td->td_proc);
2148 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2155 vn_lock(vp, LK_SHARED | LK_RETRY);
2156 AUDIT_ARG_VNODE1(vp);
2159 return (setfmode(td, active_cred, vp, mode));
2163 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2170 vn_lock(vp, LK_SHARED | LK_RETRY);
2171 AUDIT_ARG_VNODE1(vp);
2174 return (setfown(td, active_cred, vp, uid, gid));
2178 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2182 if ((object = vp->v_object) == NULL)
2184 VM_OBJECT_WLOCK(object);
2185 vm_object_page_remove(object, start, end, 0);
2186 VM_OBJECT_WUNLOCK(object);
2190 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2198 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2199 ("Wrong command %lu", cmd));
2201 if (vn_lock(vp, LK_SHARED) != 0)
2203 if (vp->v_type != VREG) {
2207 error = VOP_GETATTR(vp, &va, cred);
2211 if (noff >= va.va_size) {
2215 bsize = vp->v_mount->mnt_stat.f_iosize;
2216 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2218 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2219 if (error == EOPNOTSUPP) {
2223 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2224 (bnp != -1 && cmd == FIOSEEKDATA)) {
2231 if (noff > va.va_size)
2233 /* noff == va.va_size. There is an implicit hole at the end of file. */
2234 if (cmd == FIOSEEKDATA)
2244 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2249 off_t foffset, size;
2252 cred = td->td_ucred;
2254 foffset = foffset_lock(fp, 0);
2255 noneg = (vp->v_type != VCHR);
2261 (offset > 0 && foffset > OFF_MAX - offset))) {
2268 vn_lock(vp, LK_SHARED | LK_RETRY);
2269 error = VOP_GETATTR(vp, &vattr, cred);
2275 * If the file references a disk device, then fetch
2276 * the media size and use that to determine the ending
2279 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2280 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2281 vattr.va_size = size;
2283 (vattr.va_size > OFF_MAX ||
2284 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2288 offset += vattr.va_size;
2293 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2296 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2301 if (error == 0 && noneg && offset < 0)
2305 VFS_KNOTE_UNLOCKED(vp, 0);
2306 td->td_uretoff.tdu_off = offset;
2308 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2313 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2319 * Grant permission if the caller is the owner of the file, or
2320 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2321 * on the file. If the time pointer is null, then write
2322 * permission on the file is also sufficient.
2324 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2325 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2326 * will be allowed to set the times [..] to the current
2329 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2330 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2331 error = VOP_ACCESS(vp, VWRITE, cred, td);
2336 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2341 if (fp->f_type == DTYPE_FIFO)
2342 kif->kf_type = KF_TYPE_FIFO;
2344 kif->kf_type = KF_TYPE_VNODE;
2347 FILEDESC_SUNLOCK(fdp);
2348 error = vn_fill_kinfo_vnode(vp, kif);
2350 FILEDESC_SLOCK(fdp);
2355 vn_fill_junk(struct kinfo_file *kif)
2360 * Simulate vn_fullpath returning changing values for a given
2361 * vp during e.g. coredump.
2363 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2364 olen = strlen(kif->kf_path);
2366 strcpy(&kif->kf_path[len - 1], "$");
2368 for (; olen < len; olen++)
2369 strcpy(&kif->kf_path[olen], "A");
2373 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2376 char *fullpath, *freepath;
2379 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2382 error = vn_fullpath(curthread, vp, &fullpath, &freepath);
2384 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2386 if (freepath != NULL)
2387 free(freepath, M_TEMP);
2389 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2394 * Retrieve vnode attributes.
2396 va.va_fsid = VNOVAL;
2398 vn_lock(vp, LK_SHARED | LK_RETRY);
2399 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2403 if (va.va_fsid != VNOVAL)
2404 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2406 kif->kf_un.kf_file.kf_file_fsid =
2407 vp->v_mount->mnt_stat.f_fsid.val[0];
2408 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2409 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2410 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2411 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2412 kif->kf_un.kf_file.kf_file_size = va.va_size;
2413 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2414 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2415 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2420 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2421 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2425 struct pmckern_map_in pkm;
2431 boolean_t writecounted;
2434 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2435 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2437 * POSIX shared-memory objects are defined to have
2438 * kernel persistence, and are not defined to support
2439 * read(2)/write(2) -- or even open(2). Thus, we can
2440 * use MAP_ASYNC to trade on-disk coherence for speed.
2441 * The shm_open(3) library routine turns on the FPOSIXSHM
2442 * flag to request this behavior.
2444 if ((fp->f_flag & FPOSIXSHM) != 0)
2445 flags |= MAP_NOSYNC;
2450 * Ensure that file and memory protections are
2451 * compatible. Note that we only worry about
2452 * writability if mapping is shared; in this case,
2453 * current and max prot are dictated by the open file.
2454 * XXX use the vnode instead? Problem is: what
2455 * credentials do we use for determination? What if
2456 * proc does a setuid?
2459 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2460 maxprot = VM_PROT_NONE;
2461 if ((prot & VM_PROT_EXECUTE) != 0)
2464 maxprot = VM_PROT_EXECUTE;
2465 if ((fp->f_flag & FREAD) != 0)
2466 maxprot |= VM_PROT_READ;
2467 else if ((prot & VM_PROT_READ) != 0)
2471 * If we are sharing potential changes via MAP_SHARED and we
2472 * are trying to get write permission although we opened it
2473 * without asking for it, bail out.
2475 if ((flags & MAP_SHARED) != 0) {
2476 if ((fp->f_flag & FWRITE) != 0)
2477 maxprot |= VM_PROT_WRITE;
2478 else if ((prot & VM_PROT_WRITE) != 0)
2481 maxprot |= VM_PROT_WRITE;
2482 cap_maxprot |= VM_PROT_WRITE;
2484 maxprot &= cap_maxprot;
2487 * For regular files and shared memory, POSIX requires that
2488 * the value of foff be a legitimate offset within the data
2489 * object. In particular, negative offsets are invalid.
2490 * Blocking negative offsets and overflows here avoids
2491 * possible wraparound or user-level access into reserved
2492 * ranges of the data object later. In contrast, POSIX does
2493 * not dictate how offsets are used by device drivers, so in
2494 * the case of a device mapping a negative offset is passed
2501 foff < 0 || foff > OFF_MAX - size)
2504 writecounted = FALSE;
2505 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2506 &foff, &object, &writecounted);
2509 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2510 foff, writecounted, td);
2513 * If this mapping was accounted for in the vnode's
2514 * writecount, then undo that now.
2517 vm_pager_release_writecount(object, 0, size);
2518 vm_object_deallocate(object);
2521 /* Inform hwpmc(4) if an executable is being mapped. */
2522 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2523 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2525 pkm.pm_address = (uintptr_t) *addr;
2526 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2534 vn_fsid(struct vnode *vp, struct vattr *va)
2538 f = &vp->v_mount->mnt_stat.f_fsid;
2539 va->va_fsid = (uint32_t)f->val[1];
2540 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2541 va->va_fsid += (uint32_t)f->val[0];
2545 vn_fsync_buf(struct vnode *vp, int waitfor)
2547 struct buf *bp, *nbp;
2550 int error, maxretry;
2553 maxretry = 10000; /* large, arbitrarily chosen */
2555 if (vp->v_type == VCHR) {
2557 mp = vp->v_rdev->si_mountpt;
2564 * MARK/SCAN initialization to avoid infinite loops.
2566 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2567 bp->b_vflags &= ~BV_SCANNED;
2572 * Flush all dirty buffers associated with a vnode.
2575 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2576 if ((bp->b_vflags & BV_SCANNED) != 0)
2578 bp->b_vflags |= BV_SCANNED;
2579 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2580 if (waitfor != MNT_WAIT)
2583 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2584 BO_LOCKPTR(bo)) != 0) {
2591 KASSERT(bp->b_bufobj == bo,
2592 ("bp %p wrong b_bufobj %p should be %p",
2593 bp, bp->b_bufobj, bo));
2594 if ((bp->b_flags & B_DELWRI) == 0)
2595 panic("fsync: not dirty");
2596 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2602 if (maxretry < 1000)
2603 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2609 * If synchronous the caller expects us to completely resolve all
2610 * dirty buffers in the system. Wait for in-progress I/O to
2611 * complete (which could include background bitmap writes), then
2612 * retry if dirty blocks still exist.
2614 if (waitfor == MNT_WAIT) {
2615 bufobj_wwait(bo, 0, 0);
2616 if (bo->bo_dirty.bv_cnt > 0) {
2618 * If we are unable to write any of these buffers
2619 * then we fail now rather than trying endlessly
2620 * to write them out.
2622 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2623 if ((error = bp->b_error) != 0)
2625 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2626 (error == 0 && --maxretry >= 0))
2634 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);