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 * Common code for vnode open operations via a name lookup.
189 * Lookup the vnode and invoke VOP_CREATE if needed.
190 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
192 * Note that this does NOT free nameidata for the successful case,
193 * due to the NDINIT being done elsewhere.
196 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
197 struct ucred *cred, struct file *fp)
201 struct thread *td = ndp->ni_cnd.cn_thread;
203 struct vattr *vap = &vat;
208 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
209 O_EXCL | O_DIRECTORY))
211 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
212 ndp->ni_cnd.cn_nameiop = CREATE;
214 * Set NOCACHE to avoid flushing the cache when
215 * rolling in many files at once.
217 ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE;
218 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
219 ndp->ni_cnd.cn_flags |= FOLLOW;
220 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
221 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
222 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
223 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
224 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
226 if ((error = namei(ndp)) != 0)
228 if (ndp->ni_vp == NULL) {
231 vap->va_mode = cmode;
233 vap->va_vaflags |= VA_EXCLUSIVE;
234 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
235 NDFREE(ndp, NDF_ONLY_PNBUF);
237 if ((error = vn_start_write(NULL, &mp,
238 V_XSLEEP | PCATCH)) != 0)
242 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
243 ndp->ni_cnd.cn_flags |= MAKEENTRY;
245 error = mac_vnode_check_create(cred, ndp->ni_dvp,
249 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
252 vn_finished_write(mp);
254 NDFREE(ndp, NDF_ONLY_PNBUF);
260 if (ndp->ni_dvp == ndp->ni_vp)
266 if (fmode & O_EXCL) {
270 if (vp->v_type == VDIR) {
277 ndp->ni_cnd.cn_nameiop = LOOKUP;
278 ndp->ni_cnd.cn_flags = ISOPEN |
279 ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
280 if (!(fmode & FWRITE))
281 ndp->ni_cnd.cn_flags |= LOCKSHARED;
282 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
283 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
284 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
285 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
286 if ((error = namei(ndp)) != 0)
290 error = vn_open_vnode(vp, fmode, cred, td, fp);
296 NDFREE(ndp, NDF_ONLY_PNBUF);
304 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
307 int error, lock_flags, type;
309 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
310 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
312 KASSERT(fp != NULL, ("open with flock requires fp"));
313 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
316 lock_flags = VOP_ISLOCKED(vp);
319 lf.l_whence = SEEK_SET;
322 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
324 if ((fmode & FNONBLOCK) == 0)
326 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
328 fp->f_flag |= FHASLOCK;
330 vn_lock(vp, lock_flags | LK_RETRY);
331 if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0)
337 * Common code for vnode open operations once a vnode is located.
338 * Check permissions, and call the VOP_OPEN routine.
341 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
342 struct thread *td, struct file *fp)
347 if (vp->v_type == VLNK)
349 if (vp->v_type == VSOCK)
351 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
354 if (fmode & (FWRITE | O_TRUNC)) {
355 if (vp->v_type == VDIR)
363 if ((fmode & O_APPEND) && (fmode & FWRITE))
368 if (fmode & O_VERIFY)
370 error = mac_vnode_check_open(cred, vp, accmode);
374 accmode &= ~(VCREAT | VVERIFY);
376 if ((fmode & O_CREAT) == 0 && accmode != 0) {
377 error = VOP_ACCESS(vp, accmode, cred, td);
381 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
382 vn_lock(vp, LK_UPGRADE | LK_RETRY);
383 error = VOP_OPEN(vp, fmode, cred, td, fp);
387 error = vn_open_vnode_advlock(vp, fmode, fp);
388 if (error == 0 && (fmode & FWRITE) != 0) {
389 error = VOP_ADD_WRITECOUNT(vp, 1);
391 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
392 __func__, vp, vp->v_writecount);
397 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
398 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
399 * Arrange for that by having fdrop() to use vn_closefile().
402 fp->f_flag |= FOPENFAILED;
404 if (fp->f_ops == &badfileops) {
405 fp->f_type = DTYPE_VNODE;
411 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
417 * Check for write permissions on the specified vnode.
418 * Prototype text segments cannot be written.
422 vn_writechk(struct vnode *vp)
425 ASSERT_VOP_LOCKED(vp, "vn_writechk");
427 * If there's shared text associated with
428 * the vnode, try to free it up once. If
429 * we fail, we can't allow writing.
441 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
442 struct thread *td, bool keep_ref)
445 int error, lock_flags;
447 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
448 MNT_EXTENDED_SHARED(vp->v_mount))
449 lock_flags = LK_SHARED;
451 lock_flags = LK_EXCLUSIVE;
453 vn_start_write(vp, &mp, V_WAIT);
454 vn_lock(vp, lock_flags | LK_RETRY);
455 AUDIT_ARG_VNODE1(vp);
456 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
457 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
458 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
459 __func__, vp, vp->v_writecount);
461 error = VOP_CLOSE(vp, flags, file_cred, td);
466 vn_finished_write(mp);
471 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
475 return (vn_close1(vp, flags, file_cred, td, false));
479 * Heuristic to detect sequential operation.
482 sequential_heuristic(struct uio *uio, struct file *fp)
485 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
486 if (fp->f_flag & FRDAHEAD)
487 return (fp->f_seqcount << IO_SEQSHIFT);
490 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
491 * that the first I/O is normally considered to be slightly
492 * sequential. Seeking to offset 0 doesn't change sequentiality
493 * unless previous seeks have reduced f_seqcount to 0, in which
494 * case offset 0 is not special.
496 if ((uio->uio_offset == 0 && fp->f_seqcount > 0) ||
497 uio->uio_offset == fp->f_nextoff) {
499 * f_seqcount is in units of fixed-size blocks so that it
500 * depends mainly on the amount of sequential I/O and not
501 * much on the number of sequential I/O's. The fixed size
502 * of 16384 is hard-coded here since it is (not quite) just
503 * a magic size that works well here. This size is more
504 * closely related to the best I/O size for real disks than
505 * to any block size used by software.
507 if (uio->uio_resid >= IO_SEQMAX * 16384)
508 fp->f_seqcount = IO_SEQMAX;
510 fp->f_seqcount += howmany(uio->uio_resid, 16384);
511 if (fp->f_seqcount > IO_SEQMAX)
512 fp->f_seqcount = IO_SEQMAX;
514 return (fp->f_seqcount << IO_SEQSHIFT);
517 /* Not sequential. Quickly draw-down sequentiality. */
518 if (fp->f_seqcount > 1)
526 * Package up an I/O request on a vnode into a uio and do it.
529 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
530 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
531 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
538 struct vn_io_fault_args args;
539 int error, lock_flags;
541 if (offset < 0 && vp->v_type != VCHR)
543 auio.uio_iov = &aiov;
545 aiov.iov_base = base;
547 auio.uio_resid = len;
548 auio.uio_offset = offset;
549 auio.uio_segflg = segflg;
554 if ((ioflg & IO_NODELOCKED) == 0) {
555 if ((ioflg & IO_RANGELOCKED) == 0) {
556 if (rw == UIO_READ) {
557 rl_cookie = vn_rangelock_rlock(vp, offset,
560 rl_cookie = vn_rangelock_wlock(vp, offset,
566 if (rw == UIO_WRITE) {
567 if (vp->v_type != VCHR &&
568 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
571 if (MNT_SHARED_WRITES(mp) ||
572 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
573 lock_flags = LK_SHARED;
575 lock_flags = LK_EXCLUSIVE;
577 lock_flags = LK_SHARED;
578 vn_lock(vp, lock_flags | LK_RETRY);
582 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
584 if ((ioflg & IO_NOMACCHECK) == 0) {
586 error = mac_vnode_check_read(active_cred, file_cred,
589 error = mac_vnode_check_write(active_cred, file_cred,
594 if (file_cred != NULL)
598 if (do_vn_io_fault(vp, &auio)) {
599 args.kind = VN_IO_FAULT_VOP;
602 args.args.vop_args.vp = vp;
603 error = vn_io_fault1(vp, &auio, &args, td);
604 } else if (rw == UIO_READ) {
605 error = VOP_READ(vp, &auio, ioflg, cred);
606 } else /* if (rw == UIO_WRITE) */ {
607 error = VOP_WRITE(vp, &auio, ioflg, cred);
611 *aresid = auio.uio_resid;
613 if (auio.uio_resid && error == 0)
615 if ((ioflg & IO_NODELOCKED) == 0) {
618 vn_finished_write(mp);
621 if (rl_cookie != NULL)
622 vn_rangelock_unlock(vp, rl_cookie);
627 * Package up an I/O request on a vnode into a uio and do it. The I/O
628 * request is split up into smaller chunks and we try to avoid saturating
629 * the buffer cache while potentially holding a vnode locked, so we
630 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
631 * to give other processes a chance to lock the vnode (either other processes
632 * core'ing the same binary, or unrelated processes scanning the directory).
635 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
636 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
637 struct ucred *file_cred, size_t *aresid, struct thread *td)
646 * Force `offset' to a multiple of MAXBSIZE except possibly
647 * for the first chunk, so that filesystems only need to
648 * write full blocks except possibly for the first and last
651 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
655 if (rw != UIO_READ && vp->v_type == VREG)
658 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
659 ioflg, active_cred, file_cred, &iaresid, td);
660 len -= chunk; /* aresid calc already includes length */
664 base = (char *)base + chunk;
665 kern_yield(PRI_USER);
668 *aresid = len + iaresid;
673 foffset_lock(struct file *fp, int flags)
678 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
680 #if OFF_MAX <= LONG_MAX
682 * Caller only wants the current f_offset value. Assume that
683 * the long and shorter integer types reads are atomic.
685 if ((flags & FOF_NOLOCK) != 0)
686 return (fp->f_offset);
690 * According to McKusick the vn lock was protecting f_offset here.
691 * It is now protected by the FOFFSET_LOCKED flag.
693 mtxp = mtx_pool_find(mtxpool_sleep, fp);
695 if ((flags & FOF_NOLOCK) == 0) {
696 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
697 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
698 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
701 fp->f_vnread_flags |= FOFFSET_LOCKED;
709 foffset_unlock(struct file *fp, off_t val, int flags)
713 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
715 #if OFF_MAX <= LONG_MAX
716 if ((flags & FOF_NOLOCK) != 0) {
717 if ((flags & FOF_NOUPDATE) == 0)
719 if ((flags & FOF_NEXTOFF) != 0)
725 mtxp = mtx_pool_find(mtxpool_sleep, fp);
727 if ((flags & FOF_NOUPDATE) == 0)
729 if ((flags & FOF_NEXTOFF) != 0)
731 if ((flags & FOF_NOLOCK) == 0) {
732 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
733 ("Lost FOFFSET_LOCKED"));
734 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
735 wakeup(&fp->f_vnread_flags);
736 fp->f_vnread_flags = 0;
742 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
745 if ((flags & FOF_OFFSET) == 0)
746 uio->uio_offset = foffset_lock(fp, flags);
750 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
753 if ((flags & FOF_OFFSET) == 0)
754 foffset_unlock(fp, uio->uio_offset, flags);
758 get_advice(struct file *fp, struct uio *uio)
763 ret = POSIX_FADV_NORMAL;
764 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
767 mtxp = mtx_pool_find(mtxpool_sleep, fp);
769 if (fp->f_advice != NULL &&
770 uio->uio_offset >= fp->f_advice->fa_start &&
771 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
772 ret = fp->f_advice->fa_advice;
778 * File table vnode read routine.
781 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
789 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
791 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
794 if (fp->f_flag & FNONBLOCK)
796 if (fp->f_flag & O_DIRECT)
798 advice = get_advice(fp, uio);
799 vn_lock(vp, LK_SHARED | LK_RETRY);
802 case POSIX_FADV_NORMAL:
803 case POSIX_FADV_SEQUENTIAL:
804 case POSIX_FADV_NOREUSE:
805 ioflag |= sequential_heuristic(uio, fp);
807 case POSIX_FADV_RANDOM:
808 /* Disable read-ahead for random I/O. */
811 orig_offset = uio->uio_offset;
814 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
817 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
818 fp->f_nextoff = uio->uio_offset;
820 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
821 orig_offset != uio->uio_offset)
823 * Use POSIX_FADV_DONTNEED to flush pages and buffers
824 * for the backing file after a POSIX_FADV_NOREUSE
827 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
828 POSIX_FADV_DONTNEED);
833 * File table vnode write routine.
836 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
842 int error, ioflag, lock_flags;
845 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
847 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
849 if (vp->v_type == VREG)
852 if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
854 if (fp->f_flag & FNONBLOCK)
856 if (fp->f_flag & O_DIRECT)
858 if ((fp->f_flag & O_FSYNC) ||
859 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
862 if (vp->v_type != VCHR &&
863 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
866 advice = get_advice(fp, uio);
868 if (MNT_SHARED_WRITES(mp) ||
869 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
870 lock_flags = LK_SHARED;
872 lock_flags = LK_EXCLUSIVE;
875 vn_lock(vp, lock_flags | LK_RETRY);
877 case POSIX_FADV_NORMAL:
878 case POSIX_FADV_SEQUENTIAL:
879 case POSIX_FADV_NOREUSE:
880 ioflag |= sequential_heuristic(uio, fp);
882 case POSIX_FADV_RANDOM:
883 /* XXX: Is this correct? */
886 orig_offset = uio->uio_offset;
889 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
892 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
893 fp->f_nextoff = uio->uio_offset;
895 if (vp->v_type != VCHR)
896 vn_finished_write(mp);
897 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
898 orig_offset != uio->uio_offset)
900 * Use POSIX_FADV_DONTNEED to flush pages and buffers
901 * for the backing file after a POSIX_FADV_NOREUSE
904 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
905 POSIX_FADV_DONTNEED);
911 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
912 * prevent the following deadlock:
914 * Assume that the thread A reads from the vnode vp1 into userspace
915 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
916 * currently not resident, then system ends up with the call chain
917 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
918 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
919 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
920 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
921 * backed by the pages of vnode vp1, and some page in buf2 is not
922 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
924 * To prevent the lock order reversal and deadlock, vn_io_fault() does
925 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
926 * Instead, it first tries to do the whole range i/o with pagefaults
927 * disabled. If all pages in the i/o buffer are resident and mapped,
928 * VOP will succeed (ignoring the genuine filesystem errors).
929 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
930 * i/o in chunks, with all pages in the chunk prefaulted and held
931 * using vm_fault_quick_hold_pages().
933 * Filesystems using this deadlock avoidance scheme should use the
934 * array of the held pages from uio, saved in the curthread->td_ma,
935 * instead of doing uiomove(). A helper function
936 * vn_io_fault_uiomove() converts uiomove request into
937 * uiomove_fromphys() over td_ma array.
939 * Since vnode locks do not cover the whole i/o anymore, rangelocks
940 * make the current i/o request atomic with respect to other i/os and
945 * Decode vn_io_fault_args and perform the corresponding i/o.
948 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
954 save = vm_fault_disable_pagefaults();
955 switch (args->kind) {
956 case VN_IO_FAULT_FOP:
957 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
958 uio, args->cred, args->flags, td);
960 case VN_IO_FAULT_VOP:
961 if (uio->uio_rw == UIO_READ) {
962 error = VOP_READ(args->args.vop_args.vp, uio,
963 args->flags, args->cred);
964 } else if (uio->uio_rw == UIO_WRITE) {
965 error = VOP_WRITE(args->args.vop_args.vp, uio,
966 args->flags, args->cred);
970 panic("vn_io_fault_doio: unknown kind of io %d %d",
971 args->kind, uio->uio_rw);
973 vm_fault_enable_pagefaults(save);
978 vn_io_fault_touch(char *base, const struct uio *uio)
983 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
989 vn_io_fault_prefault_user(const struct uio *uio)
992 const struct iovec *iov;
997 KASSERT(uio->uio_segflg == UIO_USERSPACE,
998 ("vn_io_fault_prefault userspace"));
1002 resid = uio->uio_resid;
1003 base = iov->iov_base;
1006 error = vn_io_fault_touch(base, uio);
1009 if (len < PAGE_SIZE) {
1011 error = vn_io_fault_touch(base + len - 1, uio);
1016 if (++i >= uio->uio_iovcnt)
1018 iov = uio->uio_iov + i;
1019 base = iov->iov_base;
1031 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1032 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1033 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1034 * into args and call vn_io_fault1() to handle faults during the user
1035 * mode buffer accesses.
1038 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1041 vm_page_t ma[io_hold_cnt + 2];
1042 struct uio *uio_clone, short_uio;
1043 struct iovec short_iovec[1];
1044 vm_page_t *prev_td_ma;
1046 vm_offset_t addr, end;
1049 int error, cnt, saveheld, prev_td_ma_cnt;
1051 if (vn_io_fault_prefault) {
1052 error = vn_io_fault_prefault_user(uio);
1054 return (error); /* Or ignore ? */
1057 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1060 * The UFS follows IO_UNIT directive and replays back both
1061 * uio_offset and uio_resid if an error is encountered during the
1062 * operation. But, since the iovec may be already advanced,
1063 * uio is still in an inconsistent state.
1065 * Cache a copy of the original uio, which is advanced to the redo
1066 * point using UIO_NOCOPY below.
1068 uio_clone = cloneuio(uio);
1069 resid = uio->uio_resid;
1071 short_uio.uio_segflg = UIO_USERSPACE;
1072 short_uio.uio_rw = uio->uio_rw;
1073 short_uio.uio_td = uio->uio_td;
1075 error = vn_io_fault_doio(args, uio, td);
1076 if (error != EFAULT)
1079 atomic_add_long(&vn_io_faults_cnt, 1);
1080 uio_clone->uio_segflg = UIO_NOCOPY;
1081 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1082 uio_clone->uio_segflg = uio->uio_segflg;
1084 saveheld = curthread_pflags_set(TDP_UIOHELD);
1085 prev_td_ma = td->td_ma;
1086 prev_td_ma_cnt = td->td_ma_cnt;
1088 while (uio_clone->uio_resid != 0) {
1089 len = uio_clone->uio_iov->iov_len;
1091 KASSERT(uio_clone->uio_iovcnt >= 1,
1092 ("iovcnt underflow"));
1093 uio_clone->uio_iov++;
1094 uio_clone->uio_iovcnt--;
1097 if (len > io_hold_cnt * PAGE_SIZE)
1098 len = io_hold_cnt * PAGE_SIZE;
1099 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1100 end = round_page(addr + len);
1105 cnt = atop(end - trunc_page(addr));
1107 * A perfectly misaligned address and length could cause
1108 * both the start and the end of the chunk to use partial
1109 * page. +2 accounts for such a situation.
1111 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1112 addr, len, prot, ma, io_hold_cnt + 2);
1117 short_uio.uio_iov = &short_iovec[0];
1118 short_iovec[0].iov_base = (void *)addr;
1119 short_uio.uio_iovcnt = 1;
1120 short_uio.uio_resid = short_iovec[0].iov_len = len;
1121 short_uio.uio_offset = uio_clone->uio_offset;
1123 td->td_ma_cnt = cnt;
1125 error = vn_io_fault_doio(args, &short_uio, td);
1126 vm_page_unhold_pages(ma, cnt);
1127 adv = len - short_uio.uio_resid;
1129 uio_clone->uio_iov->iov_base =
1130 (char *)uio_clone->uio_iov->iov_base + adv;
1131 uio_clone->uio_iov->iov_len -= adv;
1132 uio_clone->uio_resid -= adv;
1133 uio_clone->uio_offset += adv;
1135 uio->uio_resid -= adv;
1136 uio->uio_offset += adv;
1138 if (error != 0 || adv == 0)
1141 td->td_ma = prev_td_ma;
1142 td->td_ma_cnt = prev_td_ma_cnt;
1143 curthread_pflags_restore(saveheld);
1145 free(uio_clone, M_IOV);
1150 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1151 int flags, struct thread *td)
1156 struct vn_io_fault_args args;
1159 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1163 * The ability to read(2) on a directory has historically been
1164 * allowed for all users, but this can and has been the source of
1165 * at least one security issue in the past. As such, it is now hidden
1166 * away behind a sysctl for those that actually need it to use it.
1168 if (vp->v_type == VDIR) {
1169 KASSERT(uio->uio_rw == UIO_READ,
1170 ("illegal write attempted on a directory"));
1171 if (!vfs_allow_read_dir)
1175 foffset_lock_uio(fp, uio, flags);
1176 if (do_vn_io_fault(vp, uio)) {
1177 args.kind = VN_IO_FAULT_FOP;
1178 args.args.fop_args.fp = fp;
1179 args.args.fop_args.doio = doio;
1180 args.cred = active_cred;
1181 args.flags = flags | FOF_OFFSET;
1182 if (uio->uio_rw == UIO_READ) {
1183 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1184 uio->uio_offset + uio->uio_resid);
1185 } else if ((fp->f_flag & O_APPEND) != 0 ||
1186 (flags & FOF_OFFSET) == 0) {
1187 /* For appenders, punt and lock the whole range. */
1188 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1190 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1191 uio->uio_offset + uio->uio_resid);
1193 error = vn_io_fault1(vp, uio, &args, td);
1194 vn_rangelock_unlock(vp, rl_cookie);
1196 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1198 foffset_unlock_uio(fp, uio, flags);
1203 * Helper function to perform the requested uiomove operation using
1204 * the held pages for io->uio_iov[0].iov_base buffer instead of
1205 * copyin/copyout. Access to the pages with uiomove_fromphys()
1206 * instead of iov_base prevents page faults that could occur due to
1207 * pmap_collect() invalidating the mapping created by
1208 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1209 * object cleanup revoking the write access from page mappings.
1211 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1212 * instead of plain uiomove().
1215 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1217 struct uio transp_uio;
1218 struct iovec transp_iov[1];
1224 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1225 uio->uio_segflg != UIO_USERSPACE)
1226 return (uiomove(data, xfersize, uio));
1228 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1229 transp_iov[0].iov_base = data;
1230 transp_uio.uio_iov = &transp_iov[0];
1231 transp_uio.uio_iovcnt = 1;
1232 if (xfersize > uio->uio_resid)
1233 xfersize = uio->uio_resid;
1234 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1235 transp_uio.uio_offset = 0;
1236 transp_uio.uio_segflg = UIO_SYSSPACE;
1238 * Since transp_iov points to data, and td_ma page array
1239 * corresponds to original uio->uio_iov, we need to invert the
1240 * direction of the i/o operation as passed to
1241 * uiomove_fromphys().
1243 switch (uio->uio_rw) {
1245 transp_uio.uio_rw = UIO_READ;
1248 transp_uio.uio_rw = UIO_WRITE;
1251 transp_uio.uio_td = uio->uio_td;
1252 error = uiomove_fromphys(td->td_ma,
1253 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1254 xfersize, &transp_uio);
1255 adv = xfersize - transp_uio.uio_resid;
1257 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1258 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1260 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1262 td->td_ma_cnt -= pgadv;
1263 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1264 uio->uio_iov->iov_len -= adv;
1265 uio->uio_resid -= adv;
1266 uio->uio_offset += adv;
1271 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1275 vm_offset_t iov_base;
1279 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1280 uio->uio_segflg != UIO_USERSPACE)
1281 return (uiomove_fromphys(ma, offset, xfersize, uio));
1283 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1284 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1285 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1286 switch (uio->uio_rw) {
1288 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1292 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1296 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1298 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1300 td->td_ma_cnt -= pgadv;
1301 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1302 uio->uio_iov->iov_len -= cnt;
1303 uio->uio_resid -= cnt;
1304 uio->uio_offset += cnt;
1310 * File table truncate routine.
1313 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1325 * Lock the whole range for truncation. Otherwise split i/o
1326 * might happen partly before and partly after the truncation.
1328 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1329 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1332 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1333 AUDIT_ARG_VNODE1(vp);
1334 if (vp->v_type == VDIR) {
1339 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1343 error = VOP_ADD_WRITECOUNT(vp, 1);
1346 vattr.va_size = length;
1347 if ((fp->f_flag & O_FSYNC) != 0)
1348 vattr.va_vaflags |= VA_SYNC;
1349 error = VOP_SETATTR(vp, &vattr, fp->f_cred);
1350 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1354 vn_finished_write(mp);
1356 vn_rangelock_unlock(vp, rl_cookie);
1361 * File table vnode stat routine.
1364 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1367 struct vnode *vp = fp->f_vnode;
1370 vn_lock(vp, LK_SHARED | LK_RETRY);
1371 error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1378 * Stat a vnode; implementation for the stat syscall
1381 vn_stat(struct vnode *vp, struct stat *sb, struct ucred *active_cred,
1382 struct ucred *file_cred, struct thread *td)
1389 AUDIT_ARG_VNODE1(vp);
1391 error = mac_vnode_check_stat(active_cred, file_cred, vp);
1399 * Initialize defaults for new and unusual fields, so that file
1400 * systems which don't support these fields don't need to know
1403 vap->va_birthtime.tv_sec = -1;
1404 vap->va_birthtime.tv_nsec = 0;
1405 vap->va_fsid = VNOVAL;
1406 vap->va_rdev = NODEV;
1408 error = VOP_GETATTR(vp, vap, active_cred);
1413 * Zero the spare stat fields
1415 bzero(sb, sizeof *sb);
1418 * Copy from vattr table
1420 if (vap->va_fsid != VNOVAL)
1421 sb->st_dev = vap->va_fsid;
1423 sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1424 sb->st_ino = vap->va_fileid;
1425 mode = vap->va_mode;
1426 switch (vap->va_type) {
1452 sb->st_nlink = vap->va_nlink;
1453 sb->st_uid = vap->va_uid;
1454 sb->st_gid = vap->va_gid;
1455 sb->st_rdev = vap->va_rdev;
1456 if (vap->va_size > OFF_MAX)
1458 sb->st_size = vap->va_size;
1459 sb->st_atim = vap->va_atime;
1460 sb->st_mtim = vap->va_mtime;
1461 sb->st_ctim = vap->va_ctime;
1462 sb->st_birthtim = vap->va_birthtime;
1465 * According to www.opengroup.org, the meaning of st_blksize is
1466 * "a filesystem-specific preferred I/O block size for this
1467 * object. In some filesystem types, this may vary from file
1469 * Use minimum/default of PAGE_SIZE (e.g. for VCHR).
1472 sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1474 sb->st_flags = vap->va_flags;
1475 if (priv_check(td, PRIV_VFS_GENERATION))
1478 sb->st_gen = vap->va_gen;
1480 sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1485 * File table vnode ioctl routine.
1488 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1493 struct fiobmap2_arg *bmarg;
1497 switch (vp->v_type) {
1502 vn_lock(vp, LK_SHARED | LK_RETRY);
1503 error = VOP_GETATTR(vp, &vattr, active_cred);
1506 *(int *)data = vattr.va_size - fp->f_offset;
1509 bmarg = (struct fiobmap2_arg *)data;
1510 vn_lock(vp, LK_SHARED | LK_RETRY);
1512 error = mac_vnode_check_read(active_cred, fp->f_cred,
1516 error = VOP_BMAP(vp, bmarg->bn, NULL,
1517 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1524 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1529 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1537 * File table vnode poll routine.
1540 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1548 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1549 AUDIT_ARG_VNODE1(vp);
1550 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1555 error = VOP_POLL(vp, events, fp->f_cred, td);
1560 * Acquire the requested lock and then check for validity. LK_RETRY
1561 * permits vn_lock to return doomed vnodes.
1564 _vn_lock(struct vnode *vp, int flags, char *file, int line)
1568 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1569 ("vn_lock: no locktype"));
1570 VNASSERT(vp->v_holdcnt != 0, vp, ("vn_lock: zero hold count"));
1572 error = VOP_LOCK1(vp, flags, file, line);
1573 flags &= ~LK_INTERLOCK; /* Interlock is always dropped. */
1574 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1575 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1577 if ((flags & LK_RETRY) == 0) {
1578 if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0) {
1582 } else if (error != 0)
1588 * File table vnode close routine.
1591 vn_closefile(struct file *fp, struct thread *td)
1599 fp->f_ops = &badfileops;
1600 ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1602 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1604 if (__predict_false(ref)) {
1605 lf.l_whence = SEEK_SET;
1608 lf.l_type = F_UNLCK;
1609 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1616 vn_suspendable(struct mount *mp)
1619 return (mp->mnt_op->vfs_susp_clean != NULL);
1623 * Preparing to start a filesystem write operation. If the operation is
1624 * permitted, then we bump the count of operations in progress and
1625 * proceed. If a suspend request is in progress, we wait until the
1626 * suspension is over, and then proceed.
1629 vn_start_write_locked(struct mount *mp, int flags)
1633 mtx_assert(MNT_MTX(mp), MA_OWNED);
1637 * Check on status of suspension.
1639 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1640 mp->mnt_susp_owner != curthread) {
1641 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1642 (flags & PCATCH) : 0) | (PUSER - 1);
1643 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1644 if (flags & V_NOWAIT) {
1645 error = EWOULDBLOCK;
1648 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1654 if (flags & V_XSLEEP)
1656 mp->mnt_writeopcount++;
1658 if (error != 0 || (flags & V_XSLEEP) != 0)
1665 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1670 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1671 ("V_MNTREF requires mp"));
1675 * If a vnode is provided, get and return the mount point that
1676 * to which it will write.
1679 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1681 if (error != EOPNOTSUPP)
1686 if ((mp = *mpp) == NULL)
1689 if (!vn_suspendable(mp)) {
1690 if (vp != NULL || (flags & V_MNTREF) != 0)
1696 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1698 * As long as a vnode is not provided we need to acquire a
1699 * refcount for the provided mountpoint too, in order to
1700 * emulate a vfs_ref().
1703 if (vp == NULL && (flags & V_MNTREF) == 0)
1706 return (vn_start_write_locked(mp, flags));
1710 * Secondary suspension. Used by operations such as vop_inactive
1711 * routines that are needed by the higher level functions. These
1712 * are allowed to proceed until all the higher level functions have
1713 * completed (indicated by mnt_writeopcount dropping to zero). At that
1714 * time, these operations are halted until the suspension is over.
1717 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1722 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1723 ("V_MNTREF requires mp"));
1727 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1729 if (error != EOPNOTSUPP)
1735 * If we are not suspended or have not yet reached suspended
1736 * mode, then let the operation proceed.
1738 if ((mp = *mpp) == NULL)
1741 if (!vn_suspendable(mp)) {
1742 if (vp != NULL || (flags & V_MNTREF) != 0)
1748 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1750 * As long as a vnode is not provided we need to acquire a
1751 * refcount for the provided mountpoint too, in order to
1752 * emulate a vfs_ref().
1755 if (vp == NULL && (flags & V_MNTREF) == 0)
1757 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1758 mp->mnt_secondary_writes++;
1759 mp->mnt_secondary_accwrites++;
1763 if (flags & V_NOWAIT) {
1766 return (EWOULDBLOCK);
1769 * Wait for the suspension to finish.
1771 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1772 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1781 * Filesystem write operation has completed. If we are suspending and this
1782 * operation is the last one, notify the suspender that the suspension is
1786 vn_finished_write(struct mount *mp)
1788 if (mp == NULL || !vn_suspendable(mp))
1792 mp->mnt_writeopcount--;
1793 if (mp->mnt_writeopcount < 0)
1794 panic("vn_finished_write: neg cnt");
1795 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1796 mp->mnt_writeopcount <= 0)
1797 wakeup(&mp->mnt_writeopcount);
1803 * Filesystem secondary write operation has completed. If we are
1804 * suspending and this operation is the last one, notify the suspender
1805 * that the suspension is now in effect.
1808 vn_finished_secondary_write(struct mount *mp)
1810 if (mp == NULL || !vn_suspendable(mp))
1814 mp->mnt_secondary_writes--;
1815 if (mp->mnt_secondary_writes < 0)
1816 panic("vn_finished_secondary_write: neg cnt");
1817 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1818 mp->mnt_secondary_writes <= 0)
1819 wakeup(&mp->mnt_secondary_writes);
1826 * Request a filesystem to suspend write operations.
1829 vfs_write_suspend(struct mount *mp, int flags)
1833 MPASS(vn_suspendable(mp));
1836 if (mp->mnt_susp_owner == curthread) {
1840 while (mp->mnt_kern_flag & MNTK_SUSPEND)
1841 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1844 * Unmount holds a write reference on the mount point. If we
1845 * own busy reference and drain for writers, we deadlock with
1846 * the reference draining in the unmount path. Callers of
1847 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1848 * vfs_busy() reference is owned and caller is not in the
1851 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1852 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1857 mp->mnt_kern_flag |= MNTK_SUSPEND;
1858 mp->mnt_susp_owner = curthread;
1859 if (mp->mnt_writeopcount > 0)
1860 (void) msleep(&mp->mnt_writeopcount,
1861 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1864 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0)
1865 vfs_write_resume(mp, 0);
1870 * Request a filesystem to resume write operations.
1873 vfs_write_resume(struct mount *mp, int flags)
1876 MPASS(vn_suspendable(mp));
1879 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1880 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
1881 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
1883 mp->mnt_susp_owner = NULL;
1884 wakeup(&mp->mnt_writeopcount);
1885 wakeup(&mp->mnt_flag);
1886 curthread->td_pflags &= ~TDP_IGNSUSP;
1887 if ((flags & VR_START_WRITE) != 0) {
1889 mp->mnt_writeopcount++;
1892 if ((flags & VR_NO_SUSPCLR) == 0)
1894 } else if ((flags & VR_START_WRITE) != 0) {
1896 vn_start_write_locked(mp, 0);
1903 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
1907 vfs_write_suspend_umnt(struct mount *mp)
1911 MPASS(vn_suspendable(mp));
1912 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
1913 ("vfs_write_suspend_umnt: recursed"));
1915 /* dounmount() already called vn_start_write(). */
1917 vn_finished_write(mp);
1918 error = vfs_write_suspend(mp, 0);
1920 vn_start_write(NULL, &mp, V_WAIT);
1924 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
1927 vn_start_write(NULL, &mp, V_WAIT);
1929 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
1930 wakeup(&mp->mnt_flag);
1932 curthread->td_pflags |= TDP_IGNSUSP;
1937 * Implement kqueues for files by translating it to vnode operation.
1940 vn_kqfilter(struct file *fp, struct knote *kn)
1943 return (VOP_KQFILTER(fp->f_vnode, kn));
1947 * Simplified in-kernel wrapper calls for extended attribute access.
1948 * Both calls pass in a NULL credential, authorizing as "kernel" access.
1949 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
1952 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
1953 const char *attrname, int *buflen, char *buf, struct thread *td)
1959 iov.iov_len = *buflen;
1962 auio.uio_iov = &iov;
1963 auio.uio_iovcnt = 1;
1964 auio.uio_rw = UIO_READ;
1965 auio.uio_segflg = UIO_SYSSPACE;
1967 auio.uio_offset = 0;
1968 auio.uio_resid = *buflen;
1970 if ((ioflg & IO_NODELOCKED) == 0)
1971 vn_lock(vp, LK_SHARED | LK_RETRY);
1973 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
1975 /* authorize attribute retrieval as kernel */
1976 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
1979 if ((ioflg & IO_NODELOCKED) == 0)
1983 *buflen = *buflen - auio.uio_resid;
1990 * XXX failure mode if partially written?
1993 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
1994 const char *attrname, int buflen, char *buf, struct thread *td)
2001 iov.iov_len = buflen;
2004 auio.uio_iov = &iov;
2005 auio.uio_iovcnt = 1;
2006 auio.uio_rw = UIO_WRITE;
2007 auio.uio_segflg = UIO_SYSSPACE;
2009 auio.uio_offset = 0;
2010 auio.uio_resid = buflen;
2012 if ((ioflg & IO_NODELOCKED) == 0) {
2013 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2015 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2018 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2020 /* authorize attribute setting as kernel */
2021 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2023 if ((ioflg & IO_NODELOCKED) == 0) {
2024 vn_finished_write(mp);
2032 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2033 const char *attrname, struct thread *td)
2038 if ((ioflg & IO_NODELOCKED) == 0) {
2039 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2041 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2044 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2046 /* authorize attribute removal as kernel */
2047 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2048 if (error == EOPNOTSUPP)
2049 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2052 if ((ioflg & IO_NODELOCKED) == 0) {
2053 vn_finished_write(mp);
2061 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2065 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2069 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2072 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2077 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2078 int lkflags, struct vnode **rvp)
2083 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2085 ltype = VOP_ISLOCKED(vp);
2086 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2087 ("vn_vget_ino: vp not locked"));
2088 error = vfs_busy(mp, MBF_NOWAIT);
2092 error = vfs_busy(mp, 0);
2093 vn_lock(vp, ltype | LK_RETRY);
2097 if (vp->v_iflag & VI_DOOMED) {
2103 error = alloc(mp, alloc_arg, lkflags, rvp);
2105 if (error != 0 || *rvp != vp)
2106 vn_lock(vp, ltype | LK_RETRY);
2107 if (vp->v_iflag & VI_DOOMED) {
2120 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2124 if (vp->v_type != VREG || td == NULL)
2126 if ((uoff_t)uio->uio_offset + uio->uio_resid >
2127 lim_cur(td, RLIMIT_FSIZE)) {
2128 PROC_LOCK(td->td_proc);
2129 kern_psignal(td->td_proc, SIGXFSZ);
2130 PROC_UNLOCK(td->td_proc);
2137 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2144 vn_lock(vp, LK_SHARED | LK_RETRY);
2145 AUDIT_ARG_VNODE1(vp);
2148 return (setfmode(td, active_cred, vp, mode));
2152 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2159 vn_lock(vp, LK_SHARED | LK_RETRY);
2160 AUDIT_ARG_VNODE1(vp);
2163 return (setfown(td, active_cred, vp, uid, gid));
2167 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2171 if ((object = vp->v_object) == NULL)
2173 VM_OBJECT_WLOCK(object);
2174 vm_object_page_remove(object, start, end, 0);
2175 VM_OBJECT_WUNLOCK(object);
2179 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2187 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2188 ("Wrong command %lu", cmd));
2190 if (vn_lock(vp, LK_SHARED) != 0)
2192 if (vp->v_type != VREG) {
2196 error = VOP_GETATTR(vp, &va, cred);
2200 if (noff >= va.va_size) {
2204 bsize = vp->v_mount->mnt_stat.f_iosize;
2205 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2207 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2208 if (error == EOPNOTSUPP) {
2212 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2213 (bnp != -1 && cmd == FIOSEEKDATA)) {
2220 if (noff > va.va_size)
2222 /* noff == va.va_size. There is an implicit hole at the end of file. */
2223 if (cmd == FIOSEEKDATA)
2233 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2238 off_t foffset, size;
2241 cred = td->td_ucred;
2243 foffset = foffset_lock(fp, 0);
2244 noneg = (vp->v_type != VCHR);
2250 (offset > 0 && foffset > OFF_MAX - offset))) {
2257 vn_lock(vp, LK_SHARED | LK_RETRY);
2258 error = VOP_GETATTR(vp, &vattr, cred);
2264 * If the file references a disk device, then fetch
2265 * the media size and use that to determine the ending
2268 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2269 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2270 vattr.va_size = size;
2272 (vattr.va_size > OFF_MAX ||
2273 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2277 offset += vattr.va_size;
2282 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2285 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2290 if (error == 0 && noneg && offset < 0)
2294 VFS_KNOTE_UNLOCKED(vp, 0);
2295 td->td_uretoff.tdu_off = offset;
2297 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2302 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2308 * Grant permission if the caller is the owner of the file, or
2309 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2310 * on the file. If the time pointer is null, then write
2311 * permission on the file is also sufficient.
2313 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2314 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2315 * will be allowed to set the times [..] to the current
2318 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2319 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2320 error = VOP_ACCESS(vp, VWRITE, cred, td);
2325 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2330 if (fp->f_type == DTYPE_FIFO)
2331 kif->kf_type = KF_TYPE_FIFO;
2333 kif->kf_type = KF_TYPE_VNODE;
2336 FILEDESC_SUNLOCK(fdp);
2337 error = vn_fill_kinfo_vnode(vp, kif);
2339 FILEDESC_SLOCK(fdp);
2344 vn_fill_junk(struct kinfo_file *kif)
2349 * Simulate vn_fullpath returning changing values for a given
2350 * vp during e.g. coredump.
2352 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2353 olen = strlen(kif->kf_path);
2355 strcpy(&kif->kf_path[len - 1], "$");
2357 for (; olen < len; olen++)
2358 strcpy(&kif->kf_path[olen], "A");
2362 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2365 char *fullpath, *freepath;
2368 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2371 error = vn_fullpath(curthread, vp, &fullpath, &freepath);
2373 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2375 if (freepath != NULL)
2376 free(freepath, M_TEMP);
2378 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2383 * Retrieve vnode attributes.
2385 va.va_fsid = VNOVAL;
2387 vn_lock(vp, LK_SHARED | LK_RETRY);
2388 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2392 if (va.va_fsid != VNOVAL)
2393 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2395 kif->kf_un.kf_file.kf_file_fsid =
2396 vp->v_mount->mnt_stat.f_fsid.val[0];
2397 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2398 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2399 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2400 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2401 kif->kf_un.kf_file.kf_file_size = va.va_size;
2402 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2403 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2404 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2409 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2410 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2414 struct pmckern_map_in pkm;
2420 boolean_t writecounted;
2423 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2424 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2426 * POSIX shared-memory objects are defined to have
2427 * kernel persistence, and are not defined to support
2428 * read(2)/write(2) -- or even open(2). Thus, we can
2429 * use MAP_ASYNC to trade on-disk coherence for speed.
2430 * The shm_open(3) library routine turns on the FPOSIXSHM
2431 * flag to request this behavior.
2433 if ((fp->f_flag & FPOSIXSHM) != 0)
2434 flags |= MAP_NOSYNC;
2439 * Ensure that file and memory protections are
2440 * compatible. Note that we only worry about
2441 * writability if mapping is shared; in this case,
2442 * current and max prot are dictated by the open file.
2443 * XXX use the vnode instead? Problem is: what
2444 * credentials do we use for determination? What if
2445 * proc does a setuid?
2448 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2449 maxprot = VM_PROT_NONE;
2450 if ((prot & VM_PROT_EXECUTE) != 0)
2453 maxprot = VM_PROT_EXECUTE;
2454 if ((fp->f_flag & FREAD) != 0)
2455 maxprot |= VM_PROT_READ;
2456 else if ((prot & VM_PROT_READ) != 0)
2460 * If we are sharing potential changes via MAP_SHARED and we
2461 * are trying to get write permission although we opened it
2462 * without asking for it, bail out.
2464 if ((flags & MAP_SHARED) != 0) {
2465 if ((fp->f_flag & FWRITE) != 0)
2466 maxprot |= VM_PROT_WRITE;
2467 else if ((prot & VM_PROT_WRITE) != 0)
2470 maxprot |= VM_PROT_WRITE;
2471 cap_maxprot |= VM_PROT_WRITE;
2473 maxprot &= cap_maxprot;
2476 * For regular files and shared memory, POSIX requires that
2477 * the value of foff be a legitimate offset within the data
2478 * object. In particular, negative offsets are invalid.
2479 * Blocking negative offsets and overflows here avoids
2480 * possible wraparound or user-level access into reserved
2481 * ranges of the data object later. In contrast, POSIX does
2482 * not dictate how offsets are used by device drivers, so in
2483 * the case of a device mapping a negative offset is passed
2490 foff < 0 || foff > OFF_MAX - size)
2493 writecounted = FALSE;
2494 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2495 &foff, &object, &writecounted);
2498 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2499 foff, writecounted, td);
2502 * If this mapping was accounted for in the vnode's
2503 * writecount, then undo that now.
2506 vm_pager_release_writecount(object, 0, size);
2507 vm_object_deallocate(object);
2510 /* Inform hwpmc(4) if an executable is being mapped. */
2511 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2512 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2514 pkm.pm_address = (uintptr_t) *addr;
2515 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2523 vn_fsid(struct vnode *vp, struct vattr *va)
2527 f = &vp->v_mount->mnt_stat.f_fsid;
2528 va->va_fsid = (uint32_t)f->val[1];
2529 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2530 va->va_fsid += (uint32_t)f->val[0];
2534 vn_fsync_buf(struct vnode *vp, int waitfor)
2536 struct buf *bp, *nbp;
2539 int error, maxretry;
2542 maxretry = 10000; /* large, arbitrarily chosen */
2544 if (vp->v_type == VCHR) {
2546 mp = vp->v_rdev->si_mountpt;
2553 * MARK/SCAN initialization to avoid infinite loops.
2555 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2556 bp->b_vflags &= ~BV_SCANNED;
2561 * Flush all dirty buffers associated with a vnode.
2564 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2565 if ((bp->b_vflags & BV_SCANNED) != 0)
2567 bp->b_vflags |= BV_SCANNED;
2568 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2569 if (waitfor != MNT_WAIT)
2572 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2573 BO_LOCKPTR(bo)) != 0) {
2580 KASSERT(bp->b_bufobj == bo,
2581 ("bp %p wrong b_bufobj %p should be %p",
2582 bp, bp->b_bufobj, bo));
2583 if ((bp->b_flags & B_DELWRI) == 0)
2584 panic("fsync: not dirty");
2585 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2591 if (maxretry < 1000)
2592 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2598 * If synchronous the caller expects us to completely resolve all
2599 * dirty buffers in the system. Wait for in-progress I/O to
2600 * complete (which could include background bitmap writes), then
2601 * retry if dirty blocks still exist.
2603 if (waitfor == MNT_WAIT) {
2604 bufobj_wwait(bo, 0, 0);
2605 if (bo->bo_dirty.bv_cnt > 0) {
2607 * If we are unable to write any of these buffers
2608 * then we fail now rather than trying endlessly
2609 * to write them out.
2611 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2612 if ((error = bp->b_error) != 0)
2614 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2615 (error == 0 && --maxretry >= 0))
2623 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);