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_RESOLVE_BENEATH) != 0)
195 if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
197 if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
203 * Common code for vnode open operations via a name lookup.
204 * Lookup the vnode and invoke VOP_CREATE if needed.
205 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
207 * Note that this does NOT free nameidata for the successful case,
208 * due to the NDINIT being done elsewhere.
211 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
212 struct ucred *cred, struct file *fp)
216 struct thread *td = ndp->ni_cnd.cn_thread;
218 struct vattr *vap = &vat;
223 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
224 O_EXCL | O_DIRECTORY))
226 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
227 ndp->ni_cnd.cn_nameiop = CREATE;
228 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
230 * Set NOCACHE to avoid flushing the cache when
231 * rolling in many files at once.
233 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE;
234 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
235 ndp->ni_cnd.cn_flags |= FOLLOW;
236 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
238 if ((error = namei(ndp)) != 0)
240 if (ndp->ni_vp == NULL) {
243 vap->va_mode = cmode;
245 vap->va_vaflags |= VA_EXCLUSIVE;
246 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
247 NDFREE(ndp, NDF_ONLY_PNBUF);
249 if ((error = vn_start_write(NULL, &mp,
250 V_XSLEEP | PCATCH)) != 0)
254 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
255 ndp->ni_cnd.cn_flags |= MAKEENTRY;
257 error = mac_vnode_check_create(cred, ndp->ni_dvp,
261 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
264 vn_finished_write(mp);
266 NDFREE(ndp, NDF_ONLY_PNBUF);
272 if (ndp->ni_dvp == ndp->ni_vp)
278 if (fmode & O_EXCL) {
282 if (vp->v_type == VDIR) {
289 ndp->ni_cnd.cn_nameiop = LOOKUP;
290 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
291 ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
293 if ((fmode & FWRITE) == 0)
294 ndp->ni_cnd.cn_flags |= LOCKSHARED;
295 if ((error = namei(ndp)) != 0)
299 error = vn_open_vnode(vp, fmode, cred, td, fp);
305 NDFREE(ndp, NDF_ONLY_PNBUF);
313 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
316 int error, lock_flags, type;
318 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
319 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
321 KASSERT(fp != NULL, ("open with flock requires fp"));
322 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
325 lock_flags = VOP_ISLOCKED(vp);
328 lf.l_whence = SEEK_SET;
331 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
333 if ((fmode & FNONBLOCK) == 0)
335 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
337 fp->f_flag |= FHASLOCK;
339 vn_lock(vp, lock_flags | LK_RETRY);
340 if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0)
346 * Common code for vnode open operations once a vnode is located.
347 * Check permissions, and call the VOP_OPEN routine.
350 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
351 struct thread *td, struct file *fp)
356 if (vp->v_type == VLNK)
358 if (vp->v_type == VSOCK)
360 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
363 if (fmode & (FWRITE | O_TRUNC)) {
364 if (vp->v_type == VDIR)
372 if ((fmode & O_APPEND) && (fmode & FWRITE))
377 if (fmode & O_VERIFY)
379 error = mac_vnode_check_open(cred, vp, accmode);
383 accmode &= ~(VCREAT | VVERIFY);
385 if ((fmode & O_CREAT) == 0 && accmode != 0) {
386 error = VOP_ACCESS(vp, accmode, cred, td);
390 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
391 vn_lock(vp, LK_UPGRADE | LK_RETRY);
392 error = VOP_OPEN(vp, fmode, cred, td, fp);
396 error = vn_open_vnode_advlock(vp, fmode, fp);
397 if (error == 0 && (fmode & FWRITE) != 0) {
398 error = VOP_ADD_WRITECOUNT(vp, 1);
400 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
401 __func__, vp, vp->v_writecount);
406 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
407 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
412 * Arrange the call by having fdrop() to use
413 * vn_closefile(). This is to satisfy
414 * filesystems like devfs or tmpfs, which
415 * override fo_close().
417 fp->f_flag |= FOPENFAILED;
419 if (fp->f_ops == &badfileops) {
420 fp->f_type = DTYPE_VNODE;
426 * If there is no fp, due to kernel-mode open,
427 * we can call VOP_CLOSE() now.
429 if (vp->v_type != VFIFO && (fmode & FWRITE) != 0 &&
430 !MNT_EXTENDED_SHARED(vp->v_mount) &&
431 VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
432 vn_lock(vp, LK_UPGRADE | LK_RETRY);
433 (void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
438 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
444 * Check for write permissions on the specified vnode.
445 * Prototype text segments cannot be written.
449 vn_writechk(struct vnode *vp)
452 ASSERT_VOP_LOCKED(vp, "vn_writechk");
454 * If there's shared text associated with
455 * the vnode, try to free it up once. If
456 * we fail, we can't allow writing.
468 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
469 struct thread *td, bool keep_ref)
472 int error, lock_flags;
474 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
475 MNT_EXTENDED_SHARED(vp->v_mount))
476 lock_flags = LK_SHARED;
478 lock_flags = LK_EXCLUSIVE;
480 vn_start_write(vp, &mp, V_WAIT);
481 vn_lock(vp, lock_flags | LK_RETRY);
482 AUDIT_ARG_VNODE1(vp);
483 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
484 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
485 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
486 __func__, vp, vp->v_writecount);
488 error = VOP_CLOSE(vp, flags, file_cred, td);
493 vn_finished_write(mp);
498 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
502 return (vn_close1(vp, flags, file_cred, td, false));
506 * Heuristic to detect sequential operation.
509 sequential_heuristic(struct uio *uio, struct file *fp)
512 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
513 if (fp->f_flag & FRDAHEAD)
514 return (fp->f_seqcount << IO_SEQSHIFT);
517 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
518 * that the first I/O is normally considered to be slightly
519 * sequential. Seeking to offset 0 doesn't change sequentiality
520 * unless previous seeks have reduced f_seqcount to 0, in which
521 * case offset 0 is not special.
523 if ((uio->uio_offset == 0 && fp->f_seqcount > 0) ||
524 uio->uio_offset == fp->f_nextoff) {
526 * f_seqcount is in units of fixed-size blocks so that it
527 * depends mainly on the amount of sequential I/O and not
528 * much on the number of sequential I/O's. The fixed size
529 * of 16384 is hard-coded here since it is (not quite) just
530 * a magic size that works well here. This size is more
531 * closely related to the best I/O size for real disks than
532 * to any block size used by software.
534 if (uio->uio_resid >= IO_SEQMAX * 16384)
535 fp->f_seqcount = IO_SEQMAX;
537 fp->f_seqcount += howmany(uio->uio_resid, 16384);
538 if (fp->f_seqcount > IO_SEQMAX)
539 fp->f_seqcount = IO_SEQMAX;
541 return (fp->f_seqcount << IO_SEQSHIFT);
544 /* Not sequential. Quickly draw-down sequentiality. */
545 if (fp->f_seqcount > 1)
553 * Package up an I/O request on a vnode into a uio and do it.
556 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
557 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
558 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
565 struct vn_io_fault_args args;
566 int error, lock_flags;
568 if (offset < 0 && vp->v_type != VCHR)
570 auio.uio_iov = &aiov;
572 aiov.iov_base = base;
574 auio.uio_resid = len;
575 auio.uio_offset = offset;
576 auio.uio_segflg = segflg;
581 if ((ioflg & IO_NODELOCKED) == 0) {
582 if ((ioflg & IO_RANGELOCKED) == 0) {
583 if (rw == UIO_READ) {
584 rl_cookie = vn_rangelock_rlock(vp, offset,
587 rl_cookie = vn_rangelock_wlock(vp, offset,
593 if (rw == UIO_WRITE) {
594 if (vp->v_type != VCHR &&
595 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
598 if (MNT_SHARED_WRITES(mp) ||
599 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
600 lock_flags = LK_SHARED;
602 lock_flags = LK_EXCLUSIVE;
604 lock_flags = LK_SHARED;
605 vn_lock(vp, lock_flags | LK_RETRY);
609 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
611 if ((ioflg & IO_NOMACCHECK) == 0) {
613 error = mac_vnode_check_read(active_cred, file_cred,
616 error = mac_vnode_check_write(active_cred, file_cred,
621 if (file_cred != NULL)
625 if (do_vn_io_fault(vp, &auio)) {
626 args.kind = VN_IO_FAULT_VOP;
629 args.args.vop_args.vp = vp;
630 error = vn_io_fault1(vp, &auio, &args, td);
631 } else if (rw == UIO_READ) {
632 error = VOP_READ(vp, &auio, ioflg, cred);
633 } else /* if (rw == UIO_WRITE) */ {
634 error = VOP_WRITE(vp, &auio, ioflg, cred);
638 *aresid = auio.uio_resid;
640 if (auio.uio_resid && error == 0)
642 if ((ioflg & IO_NODELOCKED) == 0) {
645 vn_finished_write(mp);
648 if (rl_cookie != NULL)
649 vn_rangelock_unlock(vp, rl_cookie);
654 * Package up an I/O request on a vnode into a uio and do it. The I/O
655 * request is split up into smaller chunks and we try to avoid saturating
656 * the buffer cache while potentially holding a vnode locked, so we
657 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
658 * to give other processes a chance to lock the vnode (either other processes
659 * core'ing the same binary, or unrelated processes scanning the directory).
662 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
663 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
664 struct ucred *file_cred, size_t *aresid, struct thread *td)
673 * Force `offset' to a multiple of MAXBSIZE except possibly
674 * for the first chunk, so that filesystems only need to
675 * write full blocks except possibly for the first and last
678 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
682 if (rw != UIO_READ && vp->v_type == VREG)
685 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
686 ioflg, active_cred, file_cred, &iaresid, td);
687 len -= chunk; /* aresid calc already includes length */
691 base = (char *)base + chunk;
692 kern_yield(PRI_USER);
695 *aresid = len + iaresid;
700 foffset_lock(struct file *fp, int flags)
705 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
707 #if OFF_MAX <= LONG_MAX
709 * Caller only wants the current f_offset value. Assume that
710 * the long and shorter integer types reads are atomic.
712 if ((flags & FOF_NOLOCK) != 0)
713 return (fp->f_offset);
717 * According to McKusick the vn lock was protecting f_offset here.
718 * It is now protected by the FOFFSET_LOCKED flag.
720 mtxp = mtx_pool_find(mtxpool_sleep, fp);
722 if ((flags & FOF_NOLOCK) == 0) {
723 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
724 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
725 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
728 fp->f_vnread_flags |= FOFFSET_LOCKED;
736 foffset_unlock(struct file *fp, off_t val, int flags)
740 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
742 #if OFF_MAX <= LONG_MAX
743 if ((flags & FOF_NOLOCK) != 0) {
744 if ((flags & FOF_NOUPDATE) == 0)
746 if ((flags & FOF_NEXTOFF) != 0)
752 mtxp = mtx_pool_find(mtxpool_sleep, fp);
754 if ((flags & FOF_NOUPDATE) == 0)
756 if ((flags & FOF_NEXTOFF) != 0)
758 if ((flags & FOF_NOLOCK) == 0) {
759 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
760 ("Lost FOFFSET_LOCKED"));
761 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
762 wakeup(&fp->f_vnread_flags);
763 fp->f_vnread_flags = 0;
769 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
772 if ((flags & FOF_OFFSET) == 0)
773 uio->uio_offset = foffset_lock(fp, flags);
777 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
780 if ((flags & FOF_OFFSET) == 0)
781 foffset_unlock(fp, uio->uio_offset, flags);
785 get_advice(struct file *fp, struct uio *uio)
790 ret = POSIX_FADV_NORMAL;
791 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
794 mtxp = mtx_pool_find(mtxpool_sleep, fp);
796 if (fp->f_advice != NULL &&
797 uio->uio_offset >= fp->f_advice->fa_start &&
798 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
799 ret = fp->f_advice->fa_advice;
805 * File table vnode read routine.
808 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
816 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
818 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
821 if (fp->f_flag & FNONBLOCK)
823 if (fp->f_flag & O_DIRECT)
825 advice = get_advice(fp, uio);
826 vn_lock(vp, LK_SHARED | LK_RETRY);
829 case POSIX_FADV_NORMAL:
830 case POSIX_FADV_SEQUENTIAL:
831 case POSIX_FADV_NOREUSE:
832 ioflag |= sequential_heuristic(uio, fp);
834 case POSIX_FADV_RANDOM:
835 /* Disable read-ahead for random I/O. */
838 orig_offset = uio->uio_offset;
841 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
844 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
845 fp->f_nextoff = uio->uio_offset;
847 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
848 orig_offset != uio->uio_offset)
850 * Use POSIX_FADV_DONTNEED to flush pages and buffers
851 * for the backing file after a POSIX_FADV_NOREUSE
854 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
855 POSIX_FADV_DONTNEED);
860 * File table vnode write routine.
863 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
869 int error, ioflag, lock_flags;
872 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
874 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
876 if (vp->v_type == VREG)
879 if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
881 if (fp->f_flag & FNONBLOCK)
883 if (fp->f_flag & O_DIRECT)
885 if ((fp->f_flag & O_FSYNC) ||
886 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
889 if (vp->v_type != VCHR &&
890 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
893 advice = get_advice(fp, uio);
895 if (MNT_SHARED_WRITES(mp) ||
896 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
897 lock_flags = LK_SHARED;
899 lock_flags = LK_EXCLUSIVE;
902 vn_lock(vp, lock_flags | LK_RETRY);
904 case POSIX_FADV_NORMAL:
905 case POSIX_FADV_SEQUENTIAL:
906 case POSIX_FADV_NOREUSE:
907 ioflag |= sequential_heuristic(uio, fp);
909 case POSIX_FADV_RANDOM:
910 /* XXX: Is this correct? */
913 orig_offset = uio->uio_offset;
916 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
919 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
920 fp->f_nextoff = uio->uio_offset;
922 if (vp->v_type != VCHR)
923 vn_finished_write(mp);
924 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
925 orig_offset != uio->uio_offset)
927 * Use POSIX_FADV_DONTNEED to flush pages and buffers
928 * for the backing file after a POSIX_FADV_NOREUSE
931 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
932 POSIX_FADV_DONTNEED);
938 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
939 * prevent the following deadlock:
941 * Assume that the thread A reads from the vnode vp1 into userspace
942 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
943 * currently not resident, then system ends up with the call chain
944 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
945 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
946 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
947 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
948 * backed by the pages of vnode vp1, and some page in buf2 is not
949 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
951 * To prevent the lock order reversal and deadlock, vn_io_fault() does
952 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
953 * Instead, it first tries to do the whole range i/o with pagefaults
954 * disabled. If all pages in the i/o buffer are resident and mapped,
955 * VOP will succeed (ignoring the genuine filesystem errors).
956 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
957 * i/o in chunks, with all pages in the chunk prefaulted and held
958 * using vm_fault_quick_hold_pages().
960 * Filesystems using this deadlock avoidance scheme should use the
961 * array of the held pages from uio, saved in the curthread->td_ma,
962 * instead of doing uiomove(). A helper function
963 * vn_io_fault_uiomove() converts uiomove request into
964 * uiomove_fromphys() over td_ma array.
966 * Since vnode locks do not cover the whole i/o anymore, rangelocks
967 * make the current i/o request atomic with respect to other i/os and
972 * Decode vn_io_fault_args and perform the corresponding i/o.
975 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
981 save = vm_fault_disable_pagefaults();
982 switch (args->kind) {
983 case VN_IO_FAULT_FOP:
984 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
985 uio, args->cred, args->flags, td);
987 case VN_IO_FAULT_VOP:
988 if (uio->uio_rw == UIO_READ) {
989 error = VOP_READ(args->args.vop_args.vp, uio,
990 args->flags, args->cred);
991 } else if (uio->uio_rw == UIO_WRITE) {
992 error = VOP_WRITE(args->args.vop_args.vp, uio,
993 args->flags, args->cred);
997 panic("vn_io_fault_doio: unknown kind of io %d %d",
998 args->kind, uio->uio_rw);
1000 vm_fault_enable_pagefaults(save);
1005 vn_io_fault_touch(char *base, const struct uio *uio)
1010 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1016 vn_io_fault_prefault_user(const struct uio *uio)
1019 const struct iovec *iov;
1024 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1025 ("vn_io_fault_prefault userspace"));
1029 resid = uio->uio_resid;
1030 base = iov->iov_base;
1033 error = vn_io_fault_touch(base, uio);
1036 if (len < PAGE_SIZE) {
1038 error = vn_io_fault_touch(base + len - 1, uio);
1043 if (++i >= uio->uio_iovcnt)
1045 iov = uio->uio_iov + i;
1046 base = iov->iov_base;
1058 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1059 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1060 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1061 * into args and call vn_io_fault1() to handle faults during the user
1062 * mode buffer accesses.
1065 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1068 vm_page_t ma[io_hold_cnt + 2];
1069 struct uio *uio_clone, short_uio;
1070 struct iovec short_iovec[1];
1071 vm_page_t *prev_td_ma;
1073 vm_offset_t addr, end;
1076 int error, cnt, saveheld, prev_td_ma_cnt;
1078 if (vn_io_fault_prefault) {
1079 error = vn_io_fault_prefault_user(uio);
1081 return (error); /* Or ignore ? */
1084 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1087 * The UFS follows IO_UNIT directive and replays back both
1088 * uio_offset and uio_resid if an error is encountered during the
1089 * operation. But, since the iovec may be already advanced,
1090 * uio is still in an inconsistent state.
1092 * Cache a copy of the original uio, which is advanced to the redo
1093 * point using UIO_NOCOPY below.
1095 uio_clone = cloneuio(uio);
1096 resid = uio->uio_resid;
1098 short_uio.uio_segflg = UIO_USERSPACE;
1099 short_uio.uio_rw = uio->uio_rw;
1100 short_uio.uio_td = uio->uio_td;
1102 error = vn_io_fault_doio(args, uio, td);
1103 if (error != EFAULT)
1106 atomic_add_long(&vn_io_faults_cnt, 1);
1107 uio_clone->uio_segflg = UIO_NOCOPY;
1108 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1109 uio_clone->uio_segflg = uio->uio_segflg;
1111 saveheld = curthread_pflags_set(TDP_UIOHELD);
1112 prev_td_ma = td->td_ma;
1113 prev_td_ma_cnt = td->td_ma_cnt;
1115 while (uio_clone->uio_resid != 0) {
1116 len = uio_clone->uio_iov->iov_len;
1118 KASSERT(uio_clone->uio_iovcnt >= 1,
1119 ("iovcnt underflow"));
1120 uio_clone->uio_iov++;
1121 uio_clone->uio_iovcnt--;
1124 if (len > io_hold_cnt * PAGE_SIZE)
1125 len = io_hold_cnt * PAGE_SIZE;
1126 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1127 end = round_page(addr + len);
1132 cnt = atop(end - trunc_page(addr));
1134 * A perfectly misaligned address and length could cause
1135 * both the start and the end of the chunk to use partial
1136 * page. +2 accounts for such a situation.
1138 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1139 addr, len, prot, ma, io_hold_cnt + 2);
1144 short_uio.uio_iov = &short_iovec[0];
1145 short_iovec[0].iov_base = (void *)addr;
1146 short_uio.uio_iovcnt = 1;
1147 short_uio.uio_resid = short_iovec[0].iov_len = len;
1148 short_uio.uio_offset = uio_clone->uio_offset;
1150 td->td_ma_cnt = cnt;
1152 error = vn_io_fault_doio(args, &short_uio, td);
1153 vm_page_unhold_pages(ma, cnt);
1154 adv = len - short_uio.uio_resid;
1156 uio_clone->uio_iov->iov_base =
1157 (char *)uio_clone->uio_iov->iov_base + adv;
1158 uio_clone->uio_iov->iov_len -= adv;
1159 uio_clone->uio_resid -= adv;
1160 uio_clone->uio_offset += adv;
1162 uio->uio_resid -= adv;
1163 uio->uio_offset += adv;
1165 if (error != 0 || adv == 0)
1168 td->td_ma = prev_td_ma;
1169 td->td_ma_cnt = prev_td_ma_cnt;
1170 curthread_pflags_restore(saveheld);
1172 free(uio_clone, M_IOV);
1177 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1178 int flags, struct thread *td)
1183 struct vn_io_fault_args args;
1186 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1190 * The ability to read(2) on a directory has historically been
1191 * allowed for all users, but this can and has been the source of
1192 * at least one security issue in the past. As such, it is now hidden
1193 * away behind a sysctl for those that actually need it to use it.
1195 if (vp->v_type == VDIR) {
1196 KASSERT(uio->uio_rw == UIO_READ,
1197 ("illegal write attempted on a directory"));
1198 if (!vfs_allow_read_dir)
1202 foffset_lock_uio(fp, uio, flags);
1203 if (do_vn_io_fault(vp, uio)) {
1204 args.kind = VN_IO_FAULT_FOP;
1205 args.args.fop_args.fp = fp;
1206 args.args.fop_args.doio = doio;
1207 args.cred = active_cred;
1208 args.flags = flags | FOF_OFFSET;
1209 if (uio->uio_rw == UIO_READ) {
1210 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1211 uio->uio_offset + uio->uio_resid);
1212 } else if ((fp->f_flag & O_APPEND) != 0 ||
1213 (flags & FOF_OFFSET) == 0) {
1214 /* For appenders, punt and lock the whole range. */
1215 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1217 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1218 uio->uio_offset + uio->uio_resid);
1220 error = vn_io_fault1(vp, uio, &args, td);
1221 vn_rangelock_unlock(vp, rl_cookie);
1223 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1225 foffset_unlock_uio(fp, uio, flags);
1230 * Helper function to perform the requested uiomove operation using
1231 * the held pages for io->uio_iov[0].iov_base buffer instead of
1232 * copyin/copyout. Access to the pages with uiomove_fromphys()
1233 * instead of iov_base prevents page faults that could occur due to
1234 * pmap_collect() invalidating the mapping created by
1235 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1236 * object cleanup revoking the write access from page mappings.
1238 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1239 * instead of plain uiomove().
1242 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1244 struct uio transp_uio;
1245 struct iovec transp_iov[1];
1251 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1252 uio->uio_segflg != UIO_USERSPACE)
1253 return (uiomove(data, xfersize, uio));
1255 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1256 transp_iov[0].iov_base = data;
1257 transp_uio.uio_iov = &transp_iov[0];
1258 transp_uio.uio_iovcnt = 1;
1259 if (xfersize > uio->uio_resid)
1260 xfersize = uio->uio_resid;
1261 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1262 transp_uio.uio_offset = 0;
1263 transp_uio.uio_segflg = UIO_SYSSPACE;
1265 * Since transp_iov points to data, and td_ma page array
1266 * corresponds to original uio->uio_iov, we need to invert the
1267 * direction of the i/o operation as passed to
1268 * uiomove_fromphys().
1270 switch (uio->uio_rw) {
1272 transp_uio.uio_rw = UIO_READ;
1275 transp_uio.uio_rw = UIO_WRITE;
1278 transp_uio.uio_td = uio->uio_td;
1279 error = uiomove_fromphys(td->td_ma,
1280 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1281 xfersize, &transp_uio);
1282 adv = xfersize - transp_uio.uio_resid;
1284 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1285 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1287 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1289 td->td_ma_cnt -= pgadv;
1290 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1291 uio->uio_iov->iov_len -= adv;
1292 uio->uio_resid -= adv;
1293 uio->uio_offset += adv;
1298 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1302 vm_offset_t iov_base;
1306 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1307 uio->uio_segflg != UIO_USERSPACE)
1308 return (uiomove_fromphys(ma, offset, xfersize, uio));
1310 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1311 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1312 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1313 switch (uio->uio_rw) {
1315 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1319 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1323 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1325 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1327 td->td_ma_cnt -= pgadv;
1328 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1329 uio->uio_iov->iov_len -= cnt;
1330 uio->uio_resid -= cnt;
1331 uio->uio_offset += cnt;
1337 * File table truncate routine.
1340 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1352 * Lock the whole range for truncation. Otherwise split i/o
1353 * might happen partly before and partly after the truncation.
1355 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1356 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1359 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1360 AUDIT_ARG_VNODE1(vp);
1361 if (vp->v_type == VDIR) {
1366 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1370 error = VOP_ADD_WRITECOUNT(vp, 1);
1373 vattr.va_size = length;
1374 if ((fp->f_flag & O_FSYNC) != 0)
1375 vattr.va_vaflags |= VA_SYNC;
1376 error = VOP_SETATTR(vp, &vattr, fp->f_cred);
1377 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1381 vn_finished_write(mp);
1383 vn_rangelock_unlock(vp, rl_cookie);
1388 * File table vnode stat routine.
1391 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1394 struct vnode *vp = fp->f_vnode;
1397 vn_lock(vp, LK_SHARED | LK_RETRY);
1398 error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1405 * Stat a vnode; implementation for the stat syscall
1408 vn_stat(struct vnode *vp, struct stat *sb, struct ucred *active_cred,
1409 struct ucred *file_cred, struct thread *td)
1416 AUDIT_ARG_VNODE1(vp);
1418 error = mac_vnode_check_stat(active_cred, file_cred, vp);
1426 * Initialize defaults for new and unusual fields, so that file
1427 * systems which don't support these fields don't need to know
1430 vap->va_birthtime.tv_sec = -1;
1431 vap->va_birthtime.tv_nsec = 0;
1432 vap->va_fsid = VNOVAL;
1433 vap->va_rdev = NODEV;
1435 error = VOP_GETATTR(vp, vap, active_cred);
1440 * Zero the spare stat fields
1442 bzero(sb, sizeof *sb);
1445 * Copy from vattr table
1447 if (vap->va_fsid != VNOVAL)
1448 sb->st_dev = vap->va_fsid;
1450 sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1451 sb->st_ino = vap->va_fileid;
1452 mode = vap->va_mode;
1453 switch (vap->va_type) {
1479 sb->st_nlink = vap->va_nlink;
1480 sb->st_uid = vap->va_uid;
1481 sb->st_gid = vap->va_gid;
1482 sb->st_rdev = vap->va_rdev;
1483 if (vap->va_size > OFF_MAX)
1485 sb->st_size = vap->va_size;
1486 sb->st_atim = vap->va_atime;
1487 sb->st_mtim = vap->va_mtime;
1488 sb->st_ctim = vap->va_ctime;
1489 sb->st_birthtim = vap->va_birthtime;
1492 * According to www.opengroup.org, the meaning of st_blksize is
1493 * "a filesystem-specific preferred I/O block size for this
1494 * object. In some filesystem types, this may vary from file
1496 * Use minimum/default of PAGE_SIZE (e.g. for VCHR).
1499 sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1501 sb->st_flags = vap->va_flags;
1502 if (priv_check(td, PRIV_VFS_GENERATION))
1505 sb->st_gen = vap->va_gen;
1507 sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1512 * File table vnode ioctl routine.
1515 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1520 struct fiobmap2_arg *bmarg;
1524 switch (vp->v_type) {
1529 vn_lock(vp, LK_SHARED | LK_RETRY);
1530 error = VOP_GETATTR(vp, &vattr, active_cred);
1533 *(int *)data = vattr.va_size - fp->f_offset;
1536 bmarg = (struct fiobmap2_arg *)data;
1537 vn_lock(vp, LK_SHARED | LK_RETRY);
1539 error = mac_vnode_check_read(active_cred, fp->f_cred,
1543 error = VOP_BMAP(vp, bmarg->bn, NULL,
1544 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1551 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1556 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1564 * File table vnode poll routine.
1567 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1575 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1576 AUDIT_ARG_VNODE1(vp);
1577 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1582 error = VOP_POLL(vp, events, fp->f_cred, td);
1587 * Acquire the requested lock and then check for validity. LK_RETRY
1588 * permits vn_lock to return doomed vnodes.
1591 _vn_lock(struct vnode *vp, int flags, char *file, int line)
1595 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1596 ("vn_lock: no locktype"));
1597 VNASSERT(vp->v_holdcnt != 0, vp, ("vn_lock: zero hold count"));
1599 error = VOP_LOCK1(vp, flags, file, line);
1600 flags &= ~LK_INTERLOCK; /* Interlock is always dropped. */
1601 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1602 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1604 if ((flags & LK_RETRY) == 0) {
1605 if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0) {
1609 } else if (error != 0)
1615 * File table vnode close routine.
1618 vn_closefile(struct file *fp, struct thread *td)
1626 fp->f_ops = &badfileops;
1627 ref = (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1629 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1631 if (__predict_false(ref)) {
1632 lf.l_whence = SEEK_SET;
1635 lf.l_type = F_UNLCK;
1636 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1643 vn_suspendable(struct mount *mp)
1646 return (mp->mnt_op->vfs_susp_clean != NULL);
1650 * Preparing to start a filesystem write operation. If the operation is
1651 * permitted, then we bump the count of operations in progress and
1652 * proceed. If a suspend request is in progress, we wait until the
1653 * suspension is over, and then proceed.
1656 vn_start_write_locked(struct mount *mp, int flags)
1660 mtx_assert(MNT_MTX(mp), MA_OWNED);
1664 * Check on status of suspension.
1666 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1667 mp->mnt_susp_owner != curthread) {
1668 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1669 (flags & PCATCH) : 0) | (PUSER - 1);
1670 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1671 if (flags & V_NOWAIT) {
1672 error = EWOULDBLOCK;
1675 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1681 if (flags & V_XSLEEP)
1683 mp->mnt_writeopcount++;
1685 if (error != 0 || (flags & V_XSLEEP) != 0)
1692 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1697 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1698 ("V_MNTREF requires mp"));
1702 * If a vnode is provided, get and return the mount point that
1703 * to which it will write.
1706 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1708 if (error != EOPNOTSUPP)
1713 if ((mp = *mpp) == NULL)
1716 if (!vn_suspendable(mp)) {
1717 if (vp != NULL || (flags & V_MNTREF) != 0)
1723 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1725 * As long as a vnode is not provided we need to acquire a
1726 * refcount for the provided mountpoint too, in order to
1727 * emulate a vfs_ref().
1730 if (vp == NULL && (flags & V_MNTREF) == 0)
1733 return (vn_start_write_locked(mp, flags));
1737 * Secondary suspension. Used by operations such as vop_inactive
1738 * routines that are needed by the higher level functions. These
1739 * are allowed to proceed until all the higher level functions have
1740 * completed (indicated by mnt_writeopcount dropping to zero). At that
1741 * time, these operations are halted until the suspension is over.
1744 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1749 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1750 ("V_MNTREF requires mp"));
1754 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1756 if (error != EOPNOTSUPP)
1762 * If we are not suspended or have not yet reached suspended
1763 * mode, then let the operation proceed.
1765 if ((mp = *mpp) == NULL)
1768 if (!vn_suspendable(mp)) {
1769 if (vp != NULL || (flags & V_MNTREF) != 0)
1775 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1777 * As long as a vnode is not provided we need to acquire a
1778 * refcount for the provided mountpoint too, in order to
1779 * emulate a vfs_ref().
1782 if (vp == NULL && (flags & V_MNTREF) == 0)
1784 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1785 mp->mnt_secondary_writes++;
1786 mp->mnt_secondary_accwrites++;
1790 if (flags & V_NOWAIT) {
1793 return (EWOULDBLOCK);
1796 * Wait for the suspension to finish.
1798 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1799 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1808 * Filesystem write operation has completed. If we are suspending and this
1809 * operation is the last one, notify the suspender that the suspension is
1813 vn_finished_write(struct mount *mp)
1815 if (mp == NULL || !vn_suspendable(mp))
1819 mp->mnt_writeopcount--;
1820 if (mp->mnt_writeopcount < 0)
1821 panic("vn_finished_write: neg cnt");
1822 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1823 mp->mnt_writeopcount <= 0)
1824 wakeup(&mp->mnt_writeopcount);
1830 * Filesystem secondary write operation has completed. If we are
1831 * suspending and this operation is the last one, notify the suspender
1832 * that the suspension is now in effect.
1835 vn_finished_secondary_write(struct mount *mp)
1837 if (mp == NULL || !vn_suspendable(mp))
1841 mp->mnt_secondary_writes--;
1842 if (mp->mnt_secondary_writes < 0)
1843 panic("vn_finished_secondary_write: neg cnt");
1844 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1845 mp->mnt_secondary_writes <= 0)
1846 wakeup(&mp->mnt_secondary_writes);
1853 * Request a filesystem to suspend write operations.
1856 vfs_write_suspend(struct mount *mp, int flags)
1860 MPASS(vn_suspendable(mp));
1863 if (mp->mnt_susp_owner == curthread) {
1867 while (mp->mnt_kern_flag & MNTK_SUSPEND)
1868 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1871 * Unmount holds a write reference on the mount point. If we
1872 * own busy reference and drain for writers, we deadlock with
1873 * the reference draining in the unmount path. Callers of
1874 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1875 * vfs_busy() reference is owned and caller is not in the
1878 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1879 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1884 mp->mnt_kern_flag |= MNTK_SUSPEND;
1885 mp->mnt_susp_owner = curthread;
1886 if (mp->mnt_writeopcount > 0)
1887 (void) msleep(&mp->mnt_writeopcount,
1888 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1891 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0)
1892 vfs_write_resume(mp, 0);
1897 * Request a filesystem to resume write operations.
1900 vfs_write_resume(struct mount *mp, int flags)
1903 MPASS(vn_suspendable(mp));
1906 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1907 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
1908 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
1910 mp->mnt_susp_owner = NULL;
1911 wakeup(&mp->mnt_writeopcount);
1912 wakeup(&mp->mnt_flag);
1913 curthread->td_pflags &= ~TDP_IGNSUSP;
1914 if ((flags & VR_START_WRITE) != 0) {
1916 mp->mnt_writeopcount++;
1919 if ((flags & VR_NO_SUSPCLR) == 0)
1921 } else if ((flags & VR_START_WRITE) != 0) {
1923 vn_start_write_locked(mp, 0);
1930 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
1934 vfs_write_suspend_umnt(struct mount *mp)
1938 MPASS(vn_suspendable(mp));
1939 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
1940 ("vfs_write_suspend_umnt: recursed"));
1942 /* dounmount() already called vn_start_write(). */
1944 vn_finished_write(mp);
1945 error = vfs_write_suspend(mp, 0);
1947 vn_start_write(NULL, &mp, V_WAIT);
1951 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
1954 vn_start_write(NULL, &mp, V_WAIT);
1956 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
1957 wakeup(&mp->mnt_flag);
1959 curthread->td_pflags |= TDP_IGNSUSP;
1964 * Implement kqueues for files by translating it to vnode operation.
1967 vn_kqfilter(struct file *fp, struct knote *kn)
1970 return (VOP_KQFILTER(fp->f_vnode, kn));
1974 * Simplified in-kernel wrapper calls for extended attribute access.
1975 * Both calls pass in a NULL credential, authorizing as "kernel" access.
1976 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
1979 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
1980 const char *attrname, int *buflen, char *buf, struct thread *td)
1986 iov.iov_len = *buflen;
1989 auio.uio_iov = &iov;
1990 auio.uio_iovcnt = 1;
1991 auio.uio_rw = UIO_READ;
1992 auio.uio_segflg = UIO_SYSSPACE;
1994 auio.uio_offset = 0;
1995 auio.uio_resid = *buflen;
1997 if ((ioflg & IO_NODELOCKED) == 0)
1998 vn_lock(vp, LK_SHARED | LK_RETRY);
2000 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2002 /* authorize attribute retrieval as kernel */
2003 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2006 if ((ioflg & IO_NODELOCKED) == 0)
2010 *buflen = *buflen - auio.uio_resid;
2017 * XXX failure mode if partially written?
2020 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2021 const char *attrname, int buflen, char *buf, struct thread *td)
2028 iov.iov_len = buflen;
2031 auio.uio_iov = &iov;
2032 auio.uio_iovcnt = 1;
2033 auio.uio_rw = UIO_WRITE;
2034 auio.uio_segflg = UIO_SYSSPACE;
2036 auio.uio_offset = 0;
2037 auio.uio_resid = buflen;
2039 if ((ioflg & IO_NODELOCKED) == 0) {
2040 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2042 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2045 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2047 /* authorize attribute setting as kernel */
2048 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2050 if ((ioflg & IO_NODELOCKED) == 0) {
2051 vn_finished_write(mp);
2059 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2060 const char *attrname, struct thread *td)
2065 if ((ioflg & IO_NODELOCKED) == 0) {
2066 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2068 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2071 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2073 /* authorize attribute removal as kernel */
2074 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2075 if (error == EOPNOTSUPP)
2076 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2079 if ((ioflg & IO_NODELOCKED) == 0) {
2080 vn_finished_write(mp);
2088 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2092 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2096 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2099 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2104 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2105 int lkflags, struct vnode **rvp)
2110 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2112 ltype = VOP_ISLOCKED(vp);
2113 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2114 ("vn_vget_ino: vp not locked"));
2115 error = vfs_busy(mp, MBF_NOWAIT);
2119 error = vfs_busy(mp, 0);
2120 vn_lock(vp, ltype | LK_RETRY);
2124 if (vp->v_iflag & VI_DOOMED) {
2130 error = alloc(mp, alloc_arg, lkflags, rvp);
2132 if (error != 0 || *rvp != vp)
2133 vn_lock(vp, ltype | LK_RETRY);
2134 if (vp->v_iflag & VI_DOOMED) {
2147 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2151 if (vp->v_type != VREG || td == NULL)
2153 if ((uoff_t)uio->uio_offset + uio->uio_resid >
2154 lim_cur(td, RLIMIT_FSIZE)) {
2155 PROC_LOCK(td->td_proc);
2156 kern_psignal(td->td_proc, SIGXFSZ);
2157 PROC_UNLOCK(td->td_proc);
2164 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2171 vn_lock(vp, LK_SHARED | LK_RETRY);
2172 AUDIT_ARG_VNODE1(vp);
2175 return (setfmode(td, active_cred, vp, mode));
2179 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2186 vn_lock(vp, LK_SHARED | LK_RETRY);
2187 AUDIT_ARG_VNODE1(vp);
2190 return (setfown(td, active_cred, vp, uid, gid));
2194 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2198 if ((object = vp->v_object) == NULL)
2200 VM_OBJECT_WLOCK(object);
2201 vm_object_page_remove(object, start, end, 0);
2202 VM_OBJECT_WUNLOCK(object);
2206 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2214 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2215 ("Wrong command %lu", cmd));
2217 if (vn_lock(vp, LK_SHARED) != 0)
2219 if (vp->v_type != VREG) {
2223 error = VOP_GETATTR(vp, &va, cred);
2227 if (noff >= va.va_size) {
2231 bsize = vp->v_mount->mnt_stat.f_iosize;
2232 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2234 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2235 if (error == EOPNOTSUPP) {
2239 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2240 (bnp != -1 && cmd == FIOSEEKDATA)) {
2247 if (noff > va.va_size)
2249 /* noff == va.va_size. There is an implicit hole at the end of file. */
2250 if (cmd == FIOSEEKDATA)
2260 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2265 off_t foffset, size;
2268 cred = td->td_ucred;
2270 foffset = foffset_lock(fp, 0);
2271 noneg = (vp->v_type != VCHR);
2277 (offset > 0 && foffset > OFF_MAX - offset))) {
2284 vn_lock(vp, LK_SHARED | LK_RETRY);
2285 error = VOP_GETATTR(vp, &vattr, cred);
2291 * If the file references a disk device, then fetch
2292 * the media size and use that to determine the ending
2295 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2296 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2297 vattr.va_size = size;
2299 (vattr.va_size > OFF_MAX ||
2300 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2304 offset += vattr.va_size;
2309 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2312 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2317 if (error == 0 && noneg && offset < 0)
2321 VFS_KNOTE_UNLOCKED(vp, 0);
2322 td->td_uretoff.tdu_off = offset;
2324 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2329 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2335 * Grant permission if the caller is the owner of the file, or
2336 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2337 * on the file. If the time pointer is null, then write
2338 * permission on the file is also sufficient.
2340 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2341 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2342 * will be allowed to set the times [..] to the current
2345 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2346 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2347 error = VOP_ACCESS(vp, VWRITE, cred, td);
2352 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2357 if (fp->f_type == DTYPE_FIFO)
2358 kif->kf_type = KF_TYPE_FIFO;
2360 kif->kf_type = KF_TYPE_VNODE;
2363 FILEDESC_SUNLOCK(fdp);
2364 error = vn_fill_kinfo_vnode(vp, kif);
2366 FILEDESC_SLOCK(fdp);
2371 vn_fill_junk(struct kinfo_file *kif)
2376 * Simulate vn_fullpath returning changing values for a given
2377 * vp during e.g. coredump.
2379 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2380 olen = strlen(kif->kf_path);
2382 strcpy(&kif->kf_path[len - 1], "$");
2384 for (; olen < len; olen++)
2385 strcpy(&kif->kf_path[olen], "A");
2389 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2392 char *fullpath, *freepath;
2395 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2398 error = vn_fullpath(curthread, vp, &fullpath, &freepath);
2400 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2402 if (freepath != NULL)
2403 free(freepath, M_TEMP);
2405 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2410 * Retrieve vnode attributes.
2412 va.va_fsid = VNOVAL;
2414 vn_lock(vp, LK_SHARED | LK_RETRY);
2415 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2419 if (va.va_fsid != VNOVAL)
2420 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2422 kif->kf_un.kf_file.kf_file_fsid =
2423 vp->v_mount->mnt_stat.f_fsid.val[0];
2424 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2425 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2426 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2427 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2428 kif->kf_un.kf_file.kf_file_size = va.va_size;
2429 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2430 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2431 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2436 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2437 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2441 struct pmckern_map_in pkm;
2447 boolean_t writecounted;
2450 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2451 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2453 * POSIX shared-memory objects are defined to have
2454 * kernel persistence, and are not defined to support
2455 * read(2)/write(2) -- or even open(2). Thus, we can
2456 * use MAP_ASYNC to trade on-disk coherence for speed.
2457 * The shm_open(3) library routine turns on the FPOSIXSHM
2458 * flag to request this behavior.
2460 if ((fp->f_flag & FPOSIXSHM) != 0)
2461 flags |= MAP_NOSYNC;
2466 * Ensure that file and memory protections are
2467 * compatible. Note that we only worry about
2468 * writability if mapping is shared; in this case,
2469 * current and max prot are dictated by the open file.
2470 * XXX use the vnode instead? Problem is: what
2471 * credentials do we use for determination? What if
2472 * proc does a setuid?
2475 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2476 maxprot = VM_PROT_NONE;
2477 if ((prot & VM_PROT_EXECUTE) != 0)
2480 maxprot = VM_PROT_EXECUTE;
2481 if ((fp->f_flag & FREAD) != 0)
2482 maxprot |= VM_PROT_READ;
2483 else if ((prot & VM_PROT_READ) != 0)
2487 * If we are sharing potential changes via MAP_SHARED and we
2488 * are trying to get write permission although we opened it
2489 * without asking for it, bail out.
2491 if ((flags & MAP_SHARED) != 0) {
2492 if ((fp->f_flag & FWRITE) != 0)
2493 maxprot |= VM_PROT_WRITE;
2494 else if ((prot & VM_PROT_WRITE) != 0)
2497 maxprot |= VM_PROT_WRITE;
2498 cap_maxprot |= VM_PROT_WRITE;
2500 maxprot &= cap_maxprot;
2503 * For regular files and shared memory, POSIX requires that
2504 * the value of foff be a legitimate offset within the data
2505 * object. In particular, negative offsets are invalid.
2506 * Blocking negative offsets and overflows here avoids
2507 * possible wraparound or user-level access into reserved
2508 * ranges of the data object later. In contrast, POSIX does
2509 * not dictate how offsets are used by device drivers, so in
2510 * the case of a device mapping a negative offset is passed
2517 foff < 0 || foff > OFF_MAX - size)
2520 writecounted = FALSE;
2521 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2522 &foff, &object, &writecounted);
2525 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2526 foff, writecounted, td);
2529 * If this mapping was accounted for in the vnode's
2530 * writecount, then undo that now.
2533 vm_pager_release_writecount(object, 0, size);
2534 vm_object_deallocate(object);
2537 /* Inform hwpmc(4) if an executable is being mapped. */
2538 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2539 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2541 pkm.pm_address = (uintptr_t) *addr;
2542 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2550 vn_fsid(struct vnode *vp, struct vattr *va)
2554 f = &vp->v_mount->mnt_stat.f_fsid;
2555 va->va_fsid = (uint32_t)f->val[1];
2556 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2557 va->va_fsid += (uint32_t)f->val[0];
2561 vn_fsync_buf(struct vnode *vp, int waitfor)
2563 struct buf *bp, *nbp;
2566 int error, maxretry;
2569 maxretry = 10000; /* large, arbitrarily chosen */
2571 if (vp->v_type == VCHR) {
2573 mp = vp->v_rdev->si_mountpt;
2580 * MARK/SCAN initialization to avoid infinite loops.
2582 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2583 bp->b_vflags &= ~BV_SCANNED;
2588 * Flush all dirty buffers associated with a vnode.
2591 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2592 if ((bp->b_vflags & BV_SCANNED) != 0)
2594 bp->b_vflags |= BV_SCANNED;
2595 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2596 if (waitfor != MNT_WAIT)
2599 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2600 BO_LOCKPTR(bo)) != 0) {
2607 KASSERT(bp->b_bufobj == bo,
2608 ("bp %p wrong b_bufobj %p should be %p",
2609 bp, bp->b_bufobj, bo));
2610 if ((bp->b_flags & B_DELWRI) == 0)
2611 panic("fsync: not dirty");
2612 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2618 if (maxretry < 1000)
2619 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2625 * If synchronous the caller expects us to completely resolve all
2626 * dirty buffers in the system. Wait for in-progress I/O to
2627 * complete (which could include background bitmap writes), then
2628 * retry if dirty blocks still exist.
2630 if (waitfor == MNT_WAIT) {
2631 bufobj_wwait(bo, 0, 0);
2632 if (bo->bo_dirty.bv_cnt > 0) {
2634 * If we are unable to write any of these buffers
2635 * then we fail now rather than trying endlessly
2636 * to write them out.
2638 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2639 if ((error = bp->b_error) != 0)
2641 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2642 (error == 0 && --maxretry >= 0))
2650 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);