2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1982, 1986, 1989, 1993
5 * The Regents of the University of California. All rights reserved.
6 * (c) UNIX System Laboratories, Inc.
7 * All or some portions of this file are derived from material licensed
8 * to the University of California by American Telephone and Telegraph
9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10 * the permission of UNIX System Laboratories, Inc.
12 * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13 * Copyright (c) 2013, 2014 The FreeBSD Foundation
15 * Portions of this software were developed by Konstantin Belousov
16 * under sponsorship from the FreeBSD Foundation.
18 * Redistribution and use in source and binary forms, with or without
19 * modification, are permitted provided that the following conditions
21 * 1. Redistributions of source code must retain the above copyright
22 * notice, this list of conditions and the following disclaimer.
23 * 2. Redistributions in binary form must reproduce the above copyright
24 * notice, this list of conditions and the following disclaimer in the
25 * documentation and/or other materials provided with the distribution.
26 * 3. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42 * @(#)vfs_vnops.c 8.2 (Berkeley) 1/21/94
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
48 #include "opt_hwpmc_hooks.h"
50 #include <sys/param.h>
51 #include <sys/systm.h>
54 #include <sys/fcntl.h>
61 #include <sys/limits.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
75 #include <sys/sleepqueue.h>
76 #include <sys/sysctl.h>
77 #include <sys/ttycom.h>
79 #include <sys/syslog.h>
80 #include <sys/unistd.h>
83 #include <security/audit/audit.h>
84 #include <security/mac/mac_framework.h>
87 #include <vm/vm_extern.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_object.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_pager.h>
95 #include <sys/pmckern.h>
98 static fo_rdwr_t vn_read;
99 static fo_rdwr_t vn_write;
100 static fo_rdwr_t vn_io_fault;
101 static fo_truncate_t vn_truncate;
102 static fo_ioctl_t vn_ioctl;
103 static fo_poll_t vn_poll;
104 static fo_kqfilter_t vn_kqfilter;
105 static fo_close_t vn_closefile;
106 static fo_mmap_t vn_mmap;
107 static fo_fallocate_t vn_fallocate;
109 struct fileops vnops = {
110 .fo_read = vn_io_fault,
111 .fo_write = vn_io_fault,
112 .fo_truncate = vn_truncate,
113 .fo_ioctl = vn_ioctl,
115 .fo_kqfilter = vn_kqfilter,
116 .fo_stat = vn_statfile,
117 .fo_close = vn_closefile,
118 .fo_chmod = vn_chmod,
119 .fo_chown = vn_chown,
120 .fo_sendfile = vn_sendfile,
122 .fo_fill_kinfo = vn_fill_kinfo,
124 .fo_fallocate = vn_fallocate,
125 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
128 const u_int io_hold_cnt = 16;
129 static int vn_io_fault_enable = 1;
130 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
131 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
132 static int vn_io_fault_prefault = 0;
133 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
134 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
135 static int vn_io_pgcache_read_enable = 1;
136 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
137 &vn_io_pgcache_read_enable, 0,
138 "Enable copying from page cache for reads, avoiding fs");
139 static u_long vn_io_faults_cnt;
140 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
141 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
143 static int vfs_allow_read_dir = 0;
144 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
145 &vfs_allow_read_dir, 0,
146 "Enable read(2) of directory by root for filesystems that support it");
149 * Returns true if vn_io_fault mode of handling the i/o request should
153 do_vn_io_fault(struct vnode *vp, struct uio *uio)
157 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
158 (mp = vp->v_mount) != NULL &&
159 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
163 * Structure used to pass arguments to vn_io_fault1(), to do either
164 * file- or vnode-based I/O calls.
166 struct vn_io_fault_args {
174 struct fop_args_tag {
178 struct vop_args_tag {
184 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
185 struct vn_io_fault_args *args, struct thread *td);
188 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
190 struct thread *td = ndp->ni_cnd.cn_thread;
192 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
196 open2nameif(int fmode, u_int vn_open_flags)
200 res = ISOPEN | LOCKLEAF;
201 if ((fmode & O_RESOLVE_BENEATH) != 0)
203 if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
205 if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
211 * Common code for vnode open operations via a name lookup.
212 * Lookup the vnode and invoke VOP_CREATE if needed.
213 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
215 * Note that this does NOT free nameidata for the successful case,
216 * due to the NDINIT being done elsewhere.
219 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
220 struct ucred *cred, struct file *fp)
224 struct thread *td = ndp->ni_cnd.cn_thread;
226 struct vattr *vap = &vat;
233 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
234 O_EXCL | O_DIRECTORY))
236 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
237 ndp->ni_cnd.cn_nameiop = CREATE;
238 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
240 * Set NOCACHE to avoid flushing the cache when
241 * rolling in many files at once.
243 * Set NC_KEEPPOSENTRY to keep positive entries if they already
244 * exist despite NOCACHE.
246 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
247 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
248 ndp->ni_cnd.cn_flags |= FOLLOW;
249 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
251 if ((error = namei(ndp)) != 0)
253 if (ndp->ni_vp == NULL) {
256 vap->va_mode = cmode;
258 vap->va_vaflags |= VA_EXCLUSIVE;
259 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
260 NDFREE(ndp, NDF_ONLY_PNBUF);
262 if ((error = vn_start_write(NULL, &mp,
263 V_XSLEEP | PCATCH)) != 0)
268 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
269 ndp->ni_cnd.cn_flags |= MAKEENTRY;
271 error = mac_vnode_check_create(cred, ndp->ni_dvp,
275 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
278 if (error == 0 && (fmode & O_EXCL) != 0 &&
279 (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
281 vp->v_iflag |= VI_FOPENING;
285 VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
287 vn_finished_write(mp);
289 NDFREE(ndp, NDF_ONLY_PNBUF);
290 if (error == ERELOOKUP) {
298 if (ndp->ni_dvp == ndp->ni_vp)
304 if (fmode & O_EXCL) {
308 if (vp->v_type == VDIR) {
315 ndp->ni_cnd.cn_nameiop = LOOKUP;
316 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
317 ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
319 if ((fmode & FWRITE) == 0)
320 ndp->ni_cnd.cn_flags |= LOCKSHARED;
321 if ((error = namei(ndp)) != 0)
325 error = vn_open_vnode(vp, fmode, cred, td, fp);
328 vp->v_iflag &= ~VI_FOPENING;
337 NDFREE(ndp, NDF_ONLY_PNBUF);
345 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
348 int error, lock_flags, type;
350 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
351 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
353 KASSERT(fp != NULL, ("open with flock requires fp"));
354 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
357 lock_flags = VOP_ISLOCKED(vp);
360 lf.l_whence = SEEK_SET;
363 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
365 if ((fmode & FNONBLOCK) == 0)
367 if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
369 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
371 fp->f_flag |= FHASLOCK;
373 vn_lock(vp, lock_flags | LK_RETRY);
378 * Common code for vnode open operations once a vnode is located.
379 * Check permissions, and call the VOP_OPEN routine.
382 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
383 struct thread *td, struct file *fp)
388 if (vp->v_type == VLNK) {
389 if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
392 if (vp->v_type == VSOCK)
394 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
398 if ((fmode & O_PATH) == 0) {
399 if ((fmode & (FWRITE | O_TRUNC)) != 0) {
400 if (vp->v_type == VDIR)
404 if ((fmode & FREAD) != 0)
406 if ((fmode & O_APPEND) && (fmode & FWRITE))
409 if ((fmode & O_CREAT) != 0)
413 if ((fmode & FEXEC) != 0)
416 if ((fmode & O_VERIFY) != 0)
418 error = mac_vnode_check_open(cred, vp, accmode);
422 accmode &= ~(VCREAT | VVERIFY);
424 if ((fmode & O_CREAT) == 0 && accmode != 0) {
425 error = VOP_ACCESS(vp, accmode, cred, td);
429 if ((fmode & O_PATH) != 0)
432 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
433 vn_lock(vp, LK_UPGRADE | LK_RETRY);
434 error = VOP_OPEN(vp, fmode, cred, td, fp);
438 error = vn_open_vnode_advlock(vp, fmode, fp);
439 if (error == 0 && (fmode & FWRITE) != 0) {
440 error = VOP_ADD_WRITECOUNT(vp, 1);
442 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
443 __func__, vp, vp->v_writecount);
448 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
449 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
450 * Arrange for that by having fdrop() to use vn_closefile().
453 fp->f_flag |= FOPENFAILED;
455 if (fp->f_ops == &badfileops) {
456 fp->f_type = DTYPE_VNODE;
462 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
468 * Check for write permissions on the specified vnode.
469 * Prototype text segments cannot be written.
473 vn_writechk(struct vnode *vp)
476 ASSERT_VOP_LOCKED(vp, "vn_writechk");
478 * If there's shared text associated with
479 * the vnode, try to free it up once. If
480 * we fail, we can't allow writing.
492 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
493 struct thread *td, bool keep_ref)
496 int error, lock_flags;
498 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
499 MNT_EXTENDED_SHARED(vp->v_mount))
500 lock_flags = LK_SHARED;
502 lock_flags = LK_EXCLUSIVE;
504 vn_start_write(vp, &mp, V_WAIT);
505 vn_lock(vp, lock_flags | LK_RETRY);
506 AUDIT_ARG_VNODE1(vp);
507 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
508 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
509 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
510 __func__, vp, vp->v_writecount);
512 error = VOP_CLOSE(vp, flags, file_cred, td);
517 vn_finished_write(mp);
522 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
526 return (vn_close1(vp, flags, file_cred, td, false));
530 * Heuristic to detect sequential operation.
533 sequential_heuristic(struct uio *uio, struct file *fp)
537 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
540 if (fp->f_flag & FRDAHEAD)
541 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
544 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
545 * that the first I/O is normally considered to be slightly
546 * sequential. Seeking to offset 0 doesn't change sequentiality
547 * unless previous seeks have reduced f_seqcount to 0, in which
548 * case offset 0 is not special.
550 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
551 uio->uio_offset == fp->f_nextoff[rw]) {
553 * f_seqcount is in units of fixed-size blocks so that it
554 * depends mainly on the amount of sequential I/O and not
555 * much on the number of sequential I/O's. The fixed size
556 * of 16384 is hard-coded here since it is (not quite) just
557 * a magic size that works well here. This size is more
558 * closely related to the best I/O size for real disks than
559 * to any block size used by software.
561 if (uio->uio_resid >= IO_SEQMAX * 16384)
562 fp->f_seqcount[rw] = IO_SEQMAX;
564 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
565 if (fp->f_seqcount[rw] > IO_SEQMAX)
566 fp->f_seqcount[rw] = IO_SEQMAX;
568 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
571 /* Not sequential. Quickly draw-down sequentiality. */
572 if (fp->f_seqcount[rw] > 1)
573 fp->f_seqcount[rw] = 1;
575 fp->f_seqcount[rw] = 0;
580 * Package up an I/O request on a vnode into a uio and do it.
583 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
584 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
585 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
592 struct vn_io_fault_args args;
593 int error, lock_flags;
595 if (offset < 0 && vp->v_type != VCHR)
597 auio.uio_iov = &aiov;
599 aiov.iov_base = base;
601 auio.uio_resid = len;
602 auio.uio_offset = offset;
603 auio.uio_segflg = segflg;
608 if ((ioflg & IO_NODELOCKED) == 0) {
609 if ((ioflg & IO_RANGELOCKED) == 0) {
610 if (rw == UIO_READ) {
611 rl_cookie = vn_rangelock_rlock(vp, offset,
613 } else if ((ioflg & IO_APPEND) != 0) {
614 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
616 rl_cookie = vn_rangelock_wlock(vp, offset,
622 if (rw == UIO_WRITE) {
623 if (vp->v_type != VCHR &&
624 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
627 if (MNT_SHARED_WRITES(mp) ||
628 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
629 lock_flags = LK_SHARED;
631 lock_flags = LK_EXCLUSIVE;
633 lock_flags = LK_SHARED;
634 vn_lock(vp, lock_flags | LK_RETRY);
638 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
640 if ((ioflg & IO_NOMACCHECK) == 0) {
642 error = mac_vnode_check_read(active_cred, file_cred,
645 error = mac_vnode_check_write(active_cred, file_cred,
650 if (file_cred != NULL)
654 if (do_vn_io_fault(vp, &auio)) {
655 args.kind = VN_IO_FAULT_VOP;
658 args.args.vop_args.vp = vp;
659 error = vn_io_fault1(vp, &auio, &args, td);
660 } else if (rw == UIO_READ) {
661 error = VOP_READ(vp, &auio, ioflg, cred);
662 } else /* if (rw == UIO_WRITE) */ {
663 error = VOP_WRITE(vp, &auio, ioflg, cred);
667 *aresid = auio.uio_resid;
669 if (auio.uio_resid && error == 0)
671 if ((ioflg & IO_NODELOCKED) == 0) {
674 vn_finished_write(mp);
677 if (rl_cookie != NULL)
678 vn_rangelock_unlock(vp, rl_cookie);
683 * Package up an I/O request on a vnode into a uio and do it. The I/O
684 * request is split up into smaller chunks and we try to avoid saturating
685 * the buffer cache while potentially holding a vnode locked, so we
686 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
687 * to give other processes a chance to lock the vnode (either other processes
688 * core'ing the same binary, or unrelated processes scanning the directory).
691 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
692 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
693 struct ucred *file_cred, size_t *aresid, struct thread *td)
702 * Force `offset' to a multiple of MAXBSIZE except possibly
703 * for the first chunk, so that filesystems only need to
704 * write full blocks except possibly for the first and last
707 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
711 if (rw != UIO_READ && vp->v_type == VREG)
714 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
715 ioflg, active_cred, file_cred, &iaresid, td);
716 len -= chunk; /* aresid calc already includes length */
720 base = (char *)base + chunk;
721 kern_yield(PRI_USER);
724 *aresid = len + iaresid;
728 #if OFF_MAX <= LONG_MAX
730 foffset_lock(struct file *fp, int flags)
732 volatile short *flagsp;
736 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
738 if ((flags & FOF_NOLOCK) != 0)
739 return (atomic_load_long(&fp->f_offset));
742 * According to McKusick the vn lock was protecting f_offset here.
743 * It is now protected by the FOFFSET_LOCKED flag.
745 flagsp = &fp->f_vnread_flags;
746 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
747 return (atomic_load_long(&fp->f_offset));
749 sleepq_lock(&fp->f_vnread_flags);
750 state = atomic_load_16(flagsp);
752 if ((state & FOFFSET_LOCKED) == 0) {
753 if (!atomic_fcmpset_acq_16(flagsp, &state,
758 if ((state & FOFFSET_LOCK_WAITING) == 0) {
759 if (!atomic_fcmpset_acq_16(flagsp, &state,
760 state | FOFFSET_LOCK_WAITING))
764 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
765 sleepq_wait(&fp->f_vnread_flags, PUSER -1);
767 sleepq_lock(&fp->f_vnread_flags);
768 state = atomic_load_16(flagsp);
770 res = atomic_load_long(&fp->f_offset);
771 sleepq_release(&fp->f_vnread_flags);
776 foffset_unlock(struct file *fp, off_t val, int flags)
778 volatile short *flagsp;
781 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
783 if ((flags & FOF_NOUPDATE) == 0)
784 atomic_store_long(&fp->f_offset, val);
785 if ((flags & FOF_NEXTOFF_R) != 0)
786 fp->f_nextoff[UIO_READ] = val;
787 if ((flags & FOF_NEXTOFF_W) != 0)
788 fp->f_nextoff[UIO_WRITE] = val;
790 if ((flags & FOF_NOLOCK) != 0)
793 flagsp = &fp->f_vnread_flags;
794 state = atomic_load_16(flagsp);
795 if ((state & FOFFSET_LOCK_WAITING) == 0 &&
796 atomic_cmpset_rel_16(flagsp, state, 0))
799 sleepq_lock(&fp->f_vnread_flags);
800 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
801 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
802 fp->f_vnread_flags = 0;
803 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
804 sleepq_release(&fp->f_vnread_flags);
808 foffset_lock(struct file *fp, int flags)
813 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
815 mtxp = mtx_pool_find(mtxpool_sleep, fp);
817 if ((flags & FOF_NOLOCK) == 0) {
818 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
819 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
820 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
823 fp->f_vnread_flags |= FOFFSET_LOCKED;
831 foffset_unlock(struct file *fp, off_t val, int flags)
835 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
837 mtxp = mtx_pool_find(mtxpool_sleep, fp);
839 if ((flags & FOF_NOUPDATE) == 0)
841 if ((flags & FOF_NEXTOFF_R) != 0)
842 fp->f_nextoff[UIO_READ] = val;
843 if ((flags & FOF_NEXTOFF_W) != 0)
844 fp->f_nextoff[UIO_WRITE] = val;
845 if ((flags & FOF_NOLOCK) == 0) {
846 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
847 ("Lost FOFFSET_LOCKED"));
848 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
849 wakeup(&fp->f_vnread_flags);
850 fp->f_vnread_flags = 0;
857 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
860 if ((flags & FOF_OFFSET) == 0)
861 uio->uio_offset = foffset_lock(fp, flags);
865 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
868 if ((flags & FOF_OFFSET) == 0)
869 foffset_unlock(fp, uio->uio_offset, flags);
873 get_advice(struct file *fp, struct uio *uio)
878 ret = POSIX_FADV_NORMAL;
879 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
882 mtxp = mtx_pool_find(mtxpool_sleep, fp);
884 if (fp->f_advice != NULL &&
885 uio->uio_offset >= fp->f_advice->fa_start &&
886 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
887 ret = fp->f_advice->fa_advice;
893 vn_read_from_obj(struct vnode *vp, struct uio *uio)
896 vm_page_t ma[io_hold_cnt + 2];
901 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
902 obj = atomic_load_ptr(&vp->v_object);
904 return (EJUSTRETURN);
907 * Depends on type stability of vm_objects.
909 vm_object_pip_add(obj, 1);
910 if ((obj->flags & OBJ_DEAD) != 0) {
912 * Note that object might be already reused from the
913 * vnode, and the OBJ_DEAD flag cleared. This is fine,
914 * we recheck for DOOMED vnode state after all pages
915 * are busied, and retract then.
917 * But we check for OBJ_DEAD to ensure that we do not
918 * busy pages while vm_object_terminate_pages()
919 * processes the queue.
925 resid = uio->uio_resid;
926 off = uio->uio_offset;
927 for (i = 0; resid > 0; i++) {
928 MPASS(i < io_hold_cnt + 2);
929 ma[i] = vm_page_grab_unlocked(obj, atop(off),
930 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
936 * Skip invalid pages. Valid mask can be partial only
937 * at EOF, and we clip later.
939 if (vm_page_none_valid(ma[i])) {
940 vm_page_sunbusy(ma[i]);
953 * Check VIRF_DOOMED after we busied our pages. Since
954 * vgonel() terminates the vnode' vm_object, it cannot
955 * process past pages busied by us.
957 if (VN_IS_DOOMED(vp)) {
962 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
963 if (resid > uio->uio_resid)
964 resid = uio->uio_resid;
967 * Unlocked read of vnp_size is safe because truncation cannot
968 * pass busied page. But we load vnp_size into a local
969 * variable so that possible concurrent extension does not
972 #if defined(__powerpc__) && !defined(__powerpc64__)
973 vsz = obj->un_pager.vnp.vnp_size;
975 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
977 if (uio->uio_offset >= vsz) {
981 if (uio->uio_offset + resid > vsz)
982 resid = vsz - uio->uio_offset;
984 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
987 for (j = 0; j < i; j++) {
989 vm_page_reference(ma[j]);
990 vm_page_sunbusy(ma[j]);
993 vm_object_pip_wakeup(obj);
996 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1000 * File table vnode read routine.
1003 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1011 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1013 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1016 if (fp->f_flag & FNONBLOCK)
1017 ioflag |= IO_NDELAY;
1018 if (fp->f_flag & O_DIRECT)
1019 ioflag |= IO_DIRECT;
1022 * Try to read from page cache. VIRF_DOOMED check is racy but
1023 * allows us to avoid unneeded work outright.
1025 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1026 (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1027 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1029 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1032 if (error != EJUSTRETURN)
1036 advice = get_advice(fp, uio);
1037 vn_lock(vp, LK_SHARED | LK_RETRY);
1040 case POSIX_FADV_NORMAL:
1041 case POSIX_FADV_SEQUENTIAL:
1042 case POSIX_FADV_NOREUSE:
1043 ioflag |= sequential_heuristic(uio, fp);
1045 case POSIX_FADV_RANDOM:
1046 /* Disable read-ahead for random I/O. */
1049 orig_offset = uio->uio_offset;
1052 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1055 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1056 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1058 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1059 orig_offset != uio->uio_offset)
1061 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1062 * for the backing file after a POSIX_FADV_NOREUSE
1065 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1066 POSIX_FADV_DONTNEED);
1071 * File table vnode write routine.
1074 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1080 int error, ioflag, lock_flags;
1083 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1085 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1087 if (vp->v_type == VREG)
1090 if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
1091 ioflag |= IO_APPEND;
1092 if (fp->f_flag & FNONBLOCK)
1093 ioflag |= IO_NDELAY;
1094 if (fp->f_flag & O_DIRECT)
1095 ioflag |= IO_DIRECT;
1096 if ((fp->f_flag & O_FSYNC) ||
1097 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
1100 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
1101 * implementations that don't understand IO_DATASYNC fall back to full
1104 if (fp->f_flag & O_DSYNC)
1105 ioflag |= IO_SYNC | IO_DATASYNC;
1107 if (vp->v_type != VCHR &&
1108 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
1111 advice = get_advice(fp, uio);
1113 if (MNT_SHARED_WRITES(mp) ||
1114 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
1115 lock_flags = LK_SHARED;
1117 lock_flags = LK_EXCLUSIVE;
1120 vn_lock(vp, lock_flags | LK_RETRY);
1122 case POSIX_FADV_NORMAL:
1123 case POSIX_FADV_SEQUENTIAL:
1124 case POSIX_FADV_NOREUSE:
1125 ioflag |= sequential_heuristic(uio, fp);
1127 case POSIX_FADV_RANDOM:
1128 /* XXX: Is this correct? */
1131 orig_offset = uio->uio_offset;
1134 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1137 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1138 fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1140 if (vp->v_type != VCHR)
1141 vn_finished_write(mp);
1142 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1143 orig_offset != uio->uio_offset)
1145 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1146 * for the backing file after a POSIX_FADV_NOREUSE
1149 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1150 POSIX_FADV_DONTNEED);
1156 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1157 * prevent the following deadlock:
1159 * Assume that the thread A reads from the vnode vp1 into userspace
1160 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
1161 * currently not resident, then system ends up with the call chain
1162 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1163 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1164 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1165 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1166 * backed by the pages of vnode vp1, and some page in buf2 is not
1167 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1169 * To prevent the lock order reversal and deadlock, vn_io_fault() does
1170 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1171 * Instead, it first tries to do the whole range i/o with pagefaults
1172 * disabled. If all pages in the i/o buffer are resident and mapped,
1173 * VOP will succeed (ignoring the genuine filesystem errors).
1174 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1175 * i/o in chunks, with all pages in the chunk prefaulted and held
1176 * using vm_fault_quick_hold_pages().
1178 * Filesystems using this deadlock avoidance scheme should use the
1179 * array of the held pages from uio, saved in the curthread->td_ma,
1180 * instead of doing uiomove(). A helper function
1181 * vn_io_fault_uiomove() converts uiomove request into
1182 * uiomove_fromphys() over td_ma array.
1184 * Since vnode locks do not cover the whole i/o anymore, rangelocks
1185 * make the current i/o request atomic with respect to other i/os and
1190 * Decode vn_io_fault_args and perform the corresponding i/o.
1193 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1199 save = vm_fault_disable_pagefaults();
1200 switch (args->kind) {
1201 case VN_IO_FAULT_FOP:
1202 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1203 uio, args->cred, args->flags, td);
1205 case VN_IO_FAULT_VOP:
1206 if (uio->uio_rw == UIO_READ) {
1207 error = VOP_READ(args->args.vop_args.vp, uio,
1208 args->flags, args->cred);
1209 } else if (uio->uio_rw == UIO_WRITE) {
1210 error = VOP_WRITE(args->args.vop_args.vp, uio,
1211 args->flags, args->cred);
1215 panic("vn_io_fault_doio: unknown kind of io %d %d",
1216 args->kind, uio->uio_rw);
1218 vm_fault_enable_pagefaults(save);
1223 vn_io_fault_touch(char *base, const struct uio *uio)
1228 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1234 vn_io_fault_prefault_user(const struct uio *uio)
1237 const struct iovec *iov;
1242 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1243 ("vn_io_fault_prefault userspace"));
1247 resid = uio->uio_resid;
1248 base = iov->iov_base;
1251 error = vn_io_fault_touch(base, uio);
1254 if (len < PAGE_SIZE) {
1256 error = vn_io_fault_touch(base + len - 1, uio);
1261 if (++i >= uio->uio_iovcnt)
1263 iov = uio->uio_iov + i;
1264 base = iov->iov_base;
1276 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1277 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1278 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1279 * into args and call vn_io_fault1() to handle faults during the user
1280 * mode buffer accesses.
1283 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1286 vm_page_t ma[io_hold_cnt + 2];
1287 struct uio *uio_clone, short_uio;
1288 struct iovec short_iovec[1];
1289 vm_page_t *prev_td_ma;
1291 vm_offset_t addr, end;
1294 int error, cnt, saveheld, prev_td_ma_cnt;
1296 if (vn_io_fault_prefault) {
1297 error = vn_io_fault_prefault_user(uio);
1299 return (error); /* Or ignore ? */
1302 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1305 * The UFS follows IO_UNIT directive and replays back both
1306 * uio_offset and uio_resid if an error is encountered during the
1307 * operation. But, since the iovec may be already advanced,
1308 * uio is still in an inconsistent state.
1310 * Cache a copy of the original uio, which is advanced to the redo
1311 * point using UIO_NOCOPY below.
1313 uio_clone = cloneuio(uio);
1314 resid = uio->uio_resid;
1316 short_uio.uio_segflg = UIO_USERSPACE;
1317 short_uio.uio_rw = uio->uio_rw;
1318 short_uio.uio_td = uio->uio_td;
1320 error = vn_io_fault_doio(args, uio, td);
1321 if (error != EFAULT)
1324 atomic_add_long(&vn_io_faults_cnt, 1);
1325 uio_clone->uio_segflg = UIO_NOCOPY;
1326 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1327 uio_clone->uio_segflg = uio->uio_segflg;
1329 saveheld = curthread_pflags_set(TDP_UIOHELD);
1330 prev_td_ma = td->td_ma;
1331 prev_td_ma_cnt = td->td_ma_cnt;
1333 while (uio_clone->uio_resid != 0) {
1334 len = uio_clone->uio_iov->iov_len;
1336 KASSERT(uio_clone->uio_iovcnt >= 1,
1337 ("iovcnt underflow"));
1338 uio_clone->uio_iov++;
1339 uio_clone->uio_iovcnt--;
1342 if (len > ptoa(io_hold_cnt))
1343 len = ptoa(io_hold_cnt);
1344 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1345 end = round_page(addr + len);
1350 cnt = atop(end - trunc_page(addr));
1352 * A perfectly misaligned address and length could cause
1353 * both the start and the end of the chunk to use partial
1354 * page. +2 accounts for such a situation.
1356 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1357 addr, len, prot, ma, io_hold_cnt + 2);
1362 short_uio.uio_iov = &short_iovec[0];
1363 short_iovec[0].iov_base = (void *)addr;
1364 short_uio.uio_iovcnt = 1;
1365 short_uio.uio_resid = short_iovec[0].iov_len = len;
1366 short_uio.uio_offset = uio_clone->uio_offset;
1368 td->td_ma_cnt = cnt;
1370 error = vn_io_fault_doio(args, &short_uio, td);
1371 vm_page_unhold_pages(ma, cnt);
1372 adv = len - short_uio.uio_resid;
1374 uio_clone->uio_iov->iov_base =
1375 (char *)uio_clone->uio_iov->iov_base + adv;
1376 uio_clone->uio_iov->iov_len -= adv;
1377 uio_clone->uio_resid -= adv;
1378 uio_clone->uio_offset += adv;
1380 uio->uio_resid -= adv;
1381 uio->uio_offset += adv;
1383 if (error != 0 || adv == 0)
1386 td->td_ma = prev_td_ma;
1387 td->td_ma_cnt = prev_td_ma_cnt;
1388 curthread_pflags_restore(saveheld);
1390 free(uio_clone, M_IOV);
1395 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1396 int flags, struct thread *td)
1401 struct vn_io_fault_args args;
1404 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1408 * The ability to read(2) on a directory has historically been
1409 * allowed for all users, but this can and has been the source of
1410 * at least one security issue in the past. As such, it is now hidden
1411 * away behind a sysctl for those that actually need it to use it, and
1412 * restricted to root when it's turned on to make it relatively safe to
1413 * leave on for longer sessions of need.
1415 if (vp->v_type == VDIR) {
1416 KASSERT(uio->uio_rw == UIO_READ,
1417 ("illegal write attempted on a directory"));
1418 if (!vfs_allow_read_dir)
1420 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1424 foffset_lock_uio(fp, uio, flags);
1425 if (do_vn_io_fault(vp, uio)) {
1426 args.kind = VN_IO_FAULT_FOP;
1427 args.args.fop_args.fp = fp;
1428 args.args.fop_args.doio = doio;
1429 args.cred = active_cred;
1430 args.flags = flags | FOF_OFFSET;
1431 if (uio->uio_rw == UIO_READ) {
1432 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1433 uio->uio_offset + uio->uio_resid);
1434 } else if ((fp->f_flag & O_APPEND) != 0 ||
1435 (flags & FOF_OFFSET) == 0) {
1436 /* For appenders, punt and lock the whole range. */
1437 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1439 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1440 uio->uio_offset + uio->uio_resid);
1442 error = vn_io_fault1(vp, uio, &args, td);
1443 vn_rangelock_unlock(vp, rl_cookie);
1445 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1447 foffset_unlock_uio(fp, uio, flags);
1452 * Helper function to perform the requested uiomove operation using
1453 * the held pages for io->uio_iov[0].iov_base buffer instead of
1454 * copyin/copyout. Access to the pages with uiomove_fromphys()
1455 * instead of iov_base prevents page faults that could occur due to
1456 * pmap_collect() invalidating the mapping created by
1457 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1458 * object cleanup revoking the write access from page mappings.
1460 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1461 * instead of plain uiomove().
1464 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1466 struct uio transp_uio;
1467 struct iovec transp_iov[1];
1473 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1474 uio->uio_segflg != UIO_USERSPACE)
1475 return (uiomove(data, xfersize, uio));
1477 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1478 transp_iov[0].iov_base = data;
1479 transp_uio.uio_iov = &transp_iov[0];
1480 transp_uio.uio_iovcnt = 1;
1481 if (xfersize > uio->uio_resid)
1482 xfersize = uio->uio_resid;
1483 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1484 transp_uio.uio_offset = 0;
1485 transp_uio.uio_segflg = UIO_SYSSPACE;
1487 * Since transp_iov points to data, and td_ma page array
1488 * corresponds to original uio->uio_iov, we need to invert the
1489 * direction of the i/o operation as passed to
1490 * uiomove_fromphys().
1492 switch (uio->uio_rw) {
1494 transp_uio.uio_rw = UIO_READ;
1497 transp_uio.uio_rw = UIO_WRITE;
1500 transp_uio.uio_td = uio->uio_td;
1501 error = uiomove_fromphys(td->td_ma,
1502 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1503 xfersize, &transp_uio);
1504 adv = xfersize - transp_uio.uio_resid;
1506 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1507 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1509 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1511 td->td_ma_cnt -= pgadv;
1512 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1513 uio->uio_iov->iov_len -= adv;
1514 uio->uio_resid -= adv;
1515 uio->uio_offset += adv;
1520 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1524 vm_offset_t iov_base;
1528 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1529 uio->uio_segflg != UIO_USERSPACE)
1530 return (uiomove_fromphys(ma, offset, xfersize, uio));
1532 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1533 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1534 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1535 switch (uio->uio_rw) {
1537 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1541 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1545 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1547 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1549 td->td_ma_cnt -= pgadv;
1550 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1551 uio->uio_iov->iov_len -= cnt;
1552 uio->uio_resid -= cnt;
1553 uio->uio_offset += cnt;
1558 * File table truncate routine.
1561 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1573 * Lock the whole range for truncation. Otherwise split i/o
1574 * might happen partly before and partly after the truncation.
1576 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1577 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1580 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1581 AUDIT_ARG_VNODE1(vp);
1582 if (vp->v_type == VDIR) {
1587 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1591 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1595 vn_finished_write(mp);
1597 vn_rangelock_unlock(vp, rl_cookie);
1598 if (error == ERELOOKUP)
1604 * Truncate a file that is already locked.
1607 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1613 error = VOP_ADD_WRITECOUNT(vp, 1);
1616 vattr.va_size = length;
1618 vattr.va_vaflags |= VA_SYNC;
1619 error = VOP_SETATTR(vp, &vattr, cred);
1620 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1626 * File table vnode stat routine.
1629 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1632 struct vnode *vp = fp->f_vnode;
1635 vn_lock(vp, LK_SHARED | LK_RETRY);
1636 error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1643 * File table vnode ioctl routine.
1646 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1651 struct fiobmap2_arg *bmarg;
1655 switch (vp->v_type) {
1660 vn_lock(vp, LK_SHARED | LK_RETRY);
1661 error = VOP_GETATTR(vp, &vattr, active_cred);
1664 *(int *)data = vattr.va_size - fp->f_offset;
1667 bmarg = (struct fiobmap2_arg *)data;
1668 vn_lock(vp, LK_SHARED | LK_RETRY);
1670 error = mac_vnode_check_read(active_cred, fp->f_cred,
1674 error = VOP_BMAP(vp, bmarg->bn, NULL,
1675 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1682 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1687 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1695 * File table vnode poll routine.
1698 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1705 #if defined(MAC) || defined(AUDIT)
1706 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1707 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1708 AUDIT_ARG_VNODE1(vp);
1709 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1715 error = VOP_POLL(vp, events, fp->f_cred, td);
1720 * Acquire the requested lock and then check for validity. LK_RETRY
1721 * permits vn_lock to return doomed vnodes.
1723 static int __noinline
1724 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1728 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1729 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1732 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1734 if ((flags & LK_RETRY) == 0) {
1745 * Nothing to do if we got the lock.
1751 * Interlock was dropped by the call in _vn_lock.
1753 flags &= ~LK_INTERLOCK;
1755 error = VOP_LOCK1(vp, flags, file, line);
1756 } while (error != 0);
1761 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1765 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1766 ("vn_lock: no locktype (%d passed)", flags));
1767 VNPASS(vp->v_holdcnt > 0, vp);
1768 error = VOP_LOCK1(vp, flags, file, line);
1769 if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1770 return (_vn_lock_fallback(vp, flags, file, line, error));
1775 * File table vnode close routine.
1778 vn_closefile(struct file *fp, struct thread *td)
1786 fp->f_ops = &badfileops;
1787 ref = (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1789 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1791 if (__predict_false(ref)) {
1792 lf.l_whence = SEEK_SET;
1795 lf.l_type = F_UNLCK;
1796 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1803 * Preparing to start a filesystem write operation. If the operation is
1804 * permitted, then we bump the count of operations in progress and
1805 * proceed. If a suspend request is in progress, we wait until the
1806 * suspension is over, and then proceed.
1809 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1811 struct mount_pcpu *mpcpu;
1814 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1815 vfs_op_thread_enter(mp, mpcpu)) {
1816 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1817 vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1818 vfs_op_thread_exit(mp, mpcpu);
1823 mtx_assert(MNT_MTX(mp), MA_OWNED);
1830 * Check on status of suspension.
1832 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1833 mp->mnt_susp_owner != curthread) {
1834 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1835 (flags & PCATCH) : 0) | (PUSER - 1);
1836 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1837 if (flags & V_NOWAIT) {
1838 error = EWOULDBLOCK;
1841 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1847 if (flags & V_XSLEEP)
1849 mp->mnt_writeopcount++;
1851 if (error != 0 || (flags & V_XSLEEP) != 0)
1858 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1863 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1864 ("V_MNTREF requires mp"));
1868 * If a vnode is provided, get and return the mount point that
1869 * to which it will write.
1872 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1874 if (error != EOPNOTSUPP)
1879 if ((mp = *mpp) == NULL)
1883 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1885 * As long as a vnode is not provided we need to acquire a
1886 * refcount for the provided mountpoint too, in order to
1887 * emulate a vfs_ref().
1889 if (vp == NULL && (flags & V_MNTREF) == 0)
1892 return (vn_start_write_refed(mp, flags, false));
1896 * Secondary suspension. Used by operations such as vop_inactive
1897 * routines that are needed by the higher level functions. These
1898 * are allowed to proceed until all the higher level functions have
1899 * completed (indicated by mnt_writeopcount dropping to zero). At that
1900 * time, these operations are halted until the suspension is over.
1903 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1908 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1909 ("V_MNTREF requires mp"));
1913 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1915 if (error != EOPNOTSUPP)
1921 * If we are not suspended or have not yet reached suspended
1922 * mode, then let the operation proceed.
1924 if ((mp = *mpp) == NULL)
1928 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1930 * As long as a vnode is not provided we need to acquire a
1931 * refcount for the provided mountpoint too, in order to
1932 * emulate a vfs_ref().
1935 if (vp == NULL && (flags & V_MNTREF) == 0)
1937 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1938 mp->mnt_secondary_writes++;
1939 mp->mnt_secondary_accwrites++;
1943 if (flags & V_NOWAIT) {
1946 return (EWOULDBLOCK);
1949 * Wait for the suspension to finish.
1951 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1952 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1961 * Filesystem write operation has completed. If we are suspending and this
1962 * operation is the last one, notify the suspender that the suspension is
1966 vn_finished_write(struct mount *mp)
1968 struct mount_pcpu *mpcpu;
1974 if (vfs_op_thread_enter(mp, mpcpu)) {
1975 vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
1976 vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
1977 vfs_op_thread_exit(mp, mpcpu);
1982 vfs_assert_mount_counters(mp);
1984 c = --mp->mnt_writeopcount;
1985 if (mp->mnt_vfs_ops == 0) {
1986 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1991 vfs_dump_mount_counters(mp);
1992 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
1993 wakeup(&mp->mnt_writeopcount);
1998 * Filesystem secondary write operation has completed. If we are
1999 * suspending and this operation is the last one, notify the suspender
2000 * that the suspension is now in effect.
2003 vn_finished_secondary_write(struct mount *mp)
2009 mp->mnt_secondary_writes--;
2010 if (mp->mnt_secondary_writes < 0)
2011 panic("vn_finished_secondary_write: neg cnt");
2012 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2013 mp->mnt_secondary_writes <= 0)
2014 wakeup(&mp->mnt_secondary_writes);
2019 * Request a filesystem to suspend write operations.
2022 vfs_write_suspend(struct mount *mp, int flags)
2029 vfs_assert_mount_counters(mp);
2030 if (mp->mnt_susp_owner == curthread) {
2031 vfs_op_exit_locked(mp);
2035 while (mp->mnt_kern_flag & MNTK_SUSPEND)
2036 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2039 * Unmount holds a write reference on the mount point. If we
2040 * own busy reference and drain for writers, we deadlock with
2041 * the reference draining in the unmount path. Callers of
2042 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2043 * vfs_busy() reference is owned and caller is not in the
2046 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2047 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2048 vfs_op_exit_locked(mp);
2053 mp->mnt_kern_flag |= MNTK_SUSPEND;
2054 mp->mnt_susp_owner = curthread;
2055 if (mp->mnt_writeopcount > 0)
2056 (void) msleep(&mp->mnt_writeopcount,
2057 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2060 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2061 vfs_write_resume(mp, 0);
2062 /* vfs_write_resume does vfs_op_exit() for us */
2068 * Request a filesystem to resume write operations.
2071 vfs_write_resume(struct mount *mp, int flags)
2075 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2076 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2077 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2079 mp->mnt_susp_owner = NULL;
2080 wakeup(&mp->mnt_writeopcount);
2081 wakeup(&mp->mnt_flag);
2082 curthread->td_pflags &= ~TDP_IGNSUSP;
2083 if ((flags & VR_START_WRITE) != 0) {
2085 mp->mnt_writeopcount++;
2088 if ((flags & VR_NO_SUSPCLR) == 0)
2091 } else if ((flags & VR_START_WRITE) != 0) {
2093 vn_start_write_refed(mp, 0, true);
2100 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2104 vfs_write_suspend_umnt(struct mount *mp)
2108 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2109 ("vfs_write_suspend_umnt: recursed"));
2111 /* dounmount() already called vn_start_write(). */
2113 vn_finished_write(mp);
2114 error = vfs_write_suspend(mp, 0);
2116 vn_start_write(NULL, &mp, V_WAIT);
2120 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2123 vn_start_write(NULL, &mp, V_WAIT);
2125 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2126 wakeup(&mp->mnt_flag);
2128 curthread->td_pflags |= TDP_IGNSUSP;
2133 * Implement kqueues for files by translating it to vnode operation.
2136 vn_kqfilter(struct file *fp, struct knote *kn)
2139 return (VOP_KQFILTER(fp->f_vnode, kn));
2143 * Simplified in-kernel wrapper calls for extended attribute access.
2144 * Both calls pass in a NULL credential, authorizing as "kernel" access.
2145 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2148 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2149 const char *attrname, int *buflen, char *buf, struct thread *td)
2155 iov.iov_len = *buflen;
2158 auio.uio_iov = &iov;
2159 auio.uio_iovcnt = 1;
2160 auio.uio_rw = UIO_READ;
2161 auio.uio_segflg = UIO_SYSSPACE;
2163 auio.uio_offset = 0;
2164 auio.uio_resid = *buflen;
2166 if ((ioflg & IO_NODELOCKED) == 0)
2167 vn_lock(vp, LK_SHARED | LK_RETRY);
2169 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2171 /* authorize attribute retrieval as kernel */
2172 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2175 if ((ioflg & IO_NODELOCKED) == 0)
2179 *buflen = *buflen - auio.uio_resid;
2186 * XXX failure mode if partially written?
2189 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2190 const char *attrname, int buflen, char *buf, struct thread *td)
2197 iov.iov_len = buflen;
2200 auio.uio_iov = &iov;
2201 auio.uio_iovcnt = 1;
2202 auio.uio_rw = UIO_WRITE;
2203 auio.uio_segflg = UIO_SYSSPACE;
2205 auio.uio_offset = 0;
2206 auio.uio_resid = buflen;
2208 if ((ioflg & IO_NODELOCKED) == 0) {
2209 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2211 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2214 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2216 /* authorize attribute setting as kernel */
2217 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2219 if ((ioflg & IO_NODELOCKED) == 0) {
2220 vn_finished_write(mp);
2228 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2229 const char *attrname, struct thread *td)
2234 if ((ioflg & IO_NODELOCKED) == 0) {
2235 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2237 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2240 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2242 /* authorize attribute removal as kernel */
2243 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2244 if (error == EOPNOTSUPP)
2245 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2248 if ((ioflg & IO_NODELOCKED) == 0) {
2249 vn_finished_write(mp);
2257 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2261 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2265 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2268 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2273 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2274 int lkflags, struct vnode **rvp)
2279 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2281 ltype = VOP_ISLOCKED(vp);
2282 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2283 ("vn_vget_ino: vp not locked"));
2284 error = vfs_busy(mp, MBF_NOWAIT);
2288 error = vfs_busy(mp, 0);
2289 vn_lock(vp, ltype | LK_RETRY);
2293 if (VN_IS_DOOMED(vp)) {
2299 error = alloc(mp, alloc_arg, lkflags, rvp);
2301 if (error != 0 || *rvp != vp)
2302 vn_lock(vp, ltype | LK_RETRY);
2303 if (VN_IS_DOOMED(vp)) {
2316 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2320 if (vp->v_type != VREG || td == NULL)
2322 if ((uoff_t)uio->uio_offset + uio->uio_resid >
2323 lim_cur(td, RLIMIT_FSIZE)) {
2324 PROC_LOCK(td->td_proc);
2325 kern_psignal(td->td_proc, SIGXFSZ);
2326 PROC_UNLOCK(td->td_proc);
2333 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2340 vn_lock(vp, LK_SHARED | LK_RETRY);
2341 AUDIT_ARG_VNODE1(vp);
2344 return (setfmode(td, active_cred, vp, mode));
2348 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2355 vn_lock(vp, LK_SHARED | LK_RETRY);
2356 AUDIT_ARG_VNODE1(vp);
2359 return (setfown(td, active_cred, vp, uid, gid));
2363 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2367 if ((object = vp->v_object) == NULL)
2369 VM_OBJECT_WLOCK(object);
2370 vm_object_page_remove(object, start, end, 0);
2371 VM_OBJECT_WUNLOCK(object);
2375 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2383 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2384 ("Wrong command %lu", cmd));
2386 if (vn_lock(vp, LK_SHARED) != 0)
2388 if (vp->v_type != VREG) {
2392 error = VOP_GETATTR(vp, &va, cred);
2396 if (noff >= va.va_size) {
2400 bsize = vp->v_mount->mnt_stat.f_iosize;
2401 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2403 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2404 if (error == EOPNOTSUPP) {
2408 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2409 (bnp != -1 && cmd == FIOSEEKDATA)) {
2416 if (noff > va.va_size)
2418 /* noff == va.va_size. There is an implicit hole at the end of file. */
2419 if (cmd == FIOSEEKDATA)
2429 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2434 off_t foffset, size;
2437 cred = td->td_ucred;
2439 foffset = foffset_lock(fp, 0);
2440 noneg = (vp->v_type != VCHR);
2446 (offset > 0 && foffset > OFF_MAX - offset))) {
2453 vn_lock(vp, LK_SHARED | LK_RETRY);
2454 error = VOP_GETATTR(vp, &vattr, cred);
2460 * If the file references a disk device, then fetch
2461 * the media size and use that to determine the ending
2464 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2465 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2466 vattr.va_size = size;
2468 (vattr.va_size > OFF_MAX ||
2469 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2473 offset += vattr.va_size;
2478 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2479 if (error == ENOTTY)
2483 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2484 if (error == ENOTTY)
2490 if (error == 0 && noneg && offset < 0)
2494 VFS_KNOTE_UNLOCKED(vp, 0);
2495 td->td_uretoff.tdu_off = offset;
2497 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2502 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2508 * Grant permission if the caller is the owner of the file, or
2509 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2510 * on the file. If the time pointer is null, then write
2511 * permission on the file is also sufficient.
2513 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2514 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2515 * will be allowed to set the times [..] to the current
2518 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2519 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2520 error = VOP_ACCESS(vp, VWRITE, cred, td);
2525 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2530 if (fp->f_type == DTYPE_FIFO)
2531 kif->kf_type = KF_TYPE_FIFO;
2533 kif->kf_type = KF_TYPE_VNODE;
2536 FILEDESC_SUNLOCK(fdp);
2537 error = vn_fill_kinfo_vnode(vp, kif);
2539 FILEDESC_SLOCK(fdp);
2544 vn_fill_junk(struct kinfo_file *kif)
2549 * Simulate vn_fullpath returning changing values for a given
2550 * vp during e.g. coredump.
2552 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2553 olen = strlen(kif->kf_path);
2555 strcpy(&kif->kf_path[len - 1], "$");
2557 for (; olen < len; olen++)
2558 strcpy(&kif->kf_path[olen], "A");
2562 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2565 char *fullpath, *freepath;
2568 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2571 error = vn_fullpath(vp, &fullpath, &freepath);
2573 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2575 if (freepath != NULL)
2576 free(freepath, M_TEMP);
2578 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2583 * Retrieve vnode attributes.
2585 va.va_fsid = VNOVAL;
2587 vn_lock(vp, LK_SHARED | LK_RETRY);
2588 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2592 if (va.va_fsid != VNOVAL)
2593 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2595 kif->kf_un.kf_file.kf_file_fsid =
2596 vp->v_mount->mnt_stat.f_fsid.val[0];
2597 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2598 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2599 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2600 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2601 kif->kf_un.kf_file.kf_file_size = va.va_size;
2602 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2603 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2604 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2609 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2610 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2614 struct pmckern_map_in pkm;
2620 boolean_t writecounted;
2623 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2624 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2626 * POSIX shared-memory objects are defined to have
2627 * kernel persistence, and are not defined to support
2628 * read(2)/write(2) -- or even open(2). Thus, we can
2629 * use MAP_ASYNC to trade on-disk coherence for speed.
2630 * The shm_open(3) library routine turns on the FPOSIXSHM
2631 * flag to request this behavior.
2633 if ((fp->f_flag & FPOSIXSHM) != 0)
2634 flags |= MAP_NOSYNC;
2639 * Ensure that file and memory protections are
2640 * compatible. Note that we only worry about
2641 * writability if mapping is shared; in this case,
2642 * current and max prot are dictated by the open file.
2643 * XXX use the vnode instead? Problem is: what
2644 * credentials do we use for determination? What if
2645 * proc does a setuid?
2648 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2649 maxprot = VM_PROT_NONE;
2650 if ((prot & VM_PROT_EXECUTE) != 0)
2653 maxprot = VM_PROT_EXECUTE;
2654 if ((fp->f_flag & FREAD) != 0)
2655 maxprot |= VM_PROT_READ;
2656 else if ((prot & VM_PROT_READ) != 0)
2660 * If we are sharing potential changes via MAP_SHARED and we
2661 * are trying to get write permission although we opened it
2662 * without asking for it, bail out.
2664 if ((flags & MAP_SHARED) != 0) {
2665 if ((fp->f_flag & FWRITE) != 0)
2666 maxprot |= VM_PROT_WRITE;
2667 else if ((prot & VM_PROT_WRITE) != 0)
2670 maxprot |= VM_PROT_WRITE;
2671 cap_maxprot |= VM_PROT_WRITE;
2673 maxprot &= cap_maxprot;
2676 * For regular files and shared memory, POSIX requires that
2677 * the value of foff be a legitimate offset within the data
2678 * object. In particular, negative offsets are invalid.
2679 * Blocking negative offsets and overflows here avoids
2680 * possible wraparound or user-level access into reserved
2681 * ranges of the data object later. In contrast, POSIX does
2682 * not dictate how offsets are used by device drivers, so in
2683 * the case of a device mapping a negative offset is passed
2690 foff > OFF_MAX - size)
2693 writecounted = FALSE;
2694 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2695 &foff, &object, &writecounted);
2698 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2699 foff, writecounted, td);
2702 * If this mapping was accounted for in the vnode's
2703 * writecount, then undo that now.
2706 vm_pager_release_writecount(object, 0, size);
2707 vm_object_deallocate(object);
2710 /* Inform hwpmc(4) if an executable is being mapped. */
2711 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2712 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2714 pkm.pm_address = (uintptr_t) *addr;
2715 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2723 vn_fsid(struct vnode *vp, struct vattr *va)
2727 f = &vp->v_mount->mnt_stat.f_fsid;
2728 va->va_fsid = (uint32_t)f->val[1];
2729 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2730 va->va_fsid += (uint32_t)f->val[0];
2734 vn_fsync_buf(struct vnode *vp, int waitfor)
2736 struct buf *bp, *nbp;
2739 int error, maxretry;
2742 maxretry = 10000; /* large, arbitrarily chosen */
2744 if (vp->v_type == VCHR) {
2746 mp = vp->v_rdev->si_mountpt;
2753 * MARK/SCAN initialization to avoid infinite loops.
2755 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2756 bp->b_vflags &= ~BV_SCANNED;
2761 * Flush all dirty buffers associated with a vnode.
2764 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2765 if ((bp->b_vflags & BV_SCANNED) != 0)
2767 bp->b_vflags |= BV_SCANNED;
2768 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2769 if (waitfor != MNT_WAIT)
2772 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2773 BO_LOCKPTR(bo)) != 0) {
2780 KASSERT(bp->b_bufobj == bo,
2781 ("bp %p wrong b_bufobj %p should be %p",
2782 bp, bp->b_bufobj, bo));
2783 if ((bp->b_flags & B_DELWRI) == 0)
2784 panic("fsync: not dirty");
2785 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2791 if (maxretry < 1000)
2792 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2798 * If synchronous the caller expects us to completely resolve all
2799 * dirty buffers in the system. Wait for in-progress I/O to
2800 * complete (which could include background bitmap writes), then
2801 * retry if dirty blocks still exist.
2803 if (waitfor == MNT_WAIT) {
2804 bufobj_wwait(bo, 0, 0);
2805 if (bo->bo_dirty.bv_cnt > 0) {
2807 * If we are unable to write any of these buffers
2808 * then we fail now rather than trying endlessly
2809 * to write them out.
2811 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2812 if ((error = bp->b_error) != 0)
2814 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2815 (error == 0 && --maxretry >= 0))
2823 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2829 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
2830 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2831 * to do the actual copy.
2832 * vn_generic_copy_file_range() is factored out, so it can be called
2833 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2834 * different file systems.
2837 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2838 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2839 struct ucred *outcred, struct thread *fsize_td)
2846 *lenp = 0; /* For error returns. */
2849 /* Do some sanity checks on the arguments. */
2850 if (invp->v_type == VDIR || outvp->v_type == VDIR)
2852 else if (*inoffp < 0 || *outoffp < 0 ||
2853 invp->v_type != VREG || outvp->v_type != VREG)
2858 /* Ensure offset + len does not wrap around. */
2861 if (uval > INT64_MAX)
2862 len = INT64_MAX - *inoffp;
2865 if (uval > INT64_MAX)
2866 len = INT64_MAX - *outoffp;
2871 * If the two vnode are for the same file system, call
2872 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2873 * which can handle copies across multiple file systems.
2876 if (invp->v_mount == outvp->v_mount)
2877 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2878 lenp, flags, incred, outcred, fsize_td);
2880 error = vn_generic_copy_file_range(invp, inoffp, outvp,
2881 outoffp, lenp, flags, incred, outcred, fsize_td);
2887 * Test len bytes of data starting at dat for all bytes == 0.
2888 * Return true if all bytes are zero, false otherwise.
2889 * Expects dat to be well aligned.
2892 mem_iszero(void *dat, int len)
2898 for (p = dat; len > 0; len -= sizeof(*p), p++) {
2899 if (len >= sizeof(*p)) {
2903 cp = (const char *)p;
2904 for (i = 0; i < len; i++, cp++)
2913 * Look for a hole in the output file and, if found, adjust *outoffp
2914 * and *xferp to skip past the hole.
2915 * *xferp is the entire hole length to be written and xfer2 is how many bytes
2916 * to be written as 0's upon return.
2919 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2920 off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2925 if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2926 *dataoffp = *outoffp;
2927 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2930 *holeoffp = *dataoffp;
2931 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2934 if (error != 0 || *holeoffp == *dataoffp) {
2936 * Since outvp is unlocked, it may be possible for
2937 * another thread to do a truncate(), lseek(), write()
2938 * creating a hole at startoff between the above
2939 * VOP_IOCTL() calls, if the other thread does not do
2941 * If that happens, *holeoffp == *dataoffp and finding
2942 * the hole has failed, so disable vn_skip_hole().
2944 *holeoffp = -1; /* Disable use of vn_skip_hole(). */
2947 KASSERT(*dataoffp >= *outoffp,
2948 ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2949 (intmax_t)*dataoffp, (intmax_t)*outoffp));
2950 KASSERT(*holeoffp > *dataoffp,
2951 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2952 (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2956 * If there is a hole before the data starts, advance *outoffp and
2957 * *xferp past the hole.
2959 if (*dataoffp > *outoffp) {
2960 delta = *dataoffp - *outoffp;
2961 if (delta >= *xferp) {
2962 /* Entire *xferp is a hole. */
2969 xfer2 = MIN(xfer2, *xferp);
2973 * If a hole starts before the end of this xfer2, reduce this xfer2 so
2974 * that the write ends at the start of the hole.
2975 * *holeoffp should always be greater than *outoffp, but for the
2976 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
2979 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
2980 xfer2 = *holeoffp - *outoffp;
2985 * Write an xfer sized chunk to outvp in blksize blocks from dat.
2986 * dat is a maximum of blksize in length and can be written repeatedly in
2988 * If growfile == true, just grow the file via vn_truncate_locked() instead
2989 * of doing actual writes.
2990 * If checkhole == true, a hole is being punched, so skip over any hole
2991 * already in the output file.
2994 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
2995 u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
2998 off_t dataoff, holeoff, xfer2;
3002 * Loop around doing writes of blksize until write has been completed.
3003 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3004 * done for each iteration, since the xfer argument can be very
3005 * large if there is a large hole to punch in the output file.
3010 xfer2 = MIN(xfer, blksize);
3013 * Punching a hole. Skip writing if there is
3014 * already a hole in the output file.
3016 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3017 &dataoff, &holeoff, cred);
3022 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3027 error = vn_start_write(outvp, &mp, V_WAIT);
3031 error = vn_lock(outvp, LK_EXCLUSIVE);
3033 error = vn_truncate_locked(outvp, outoff + xfer,
3038 if (MNT_SHARED_WRITES(mp))
3041 lckf = LK_EXCLUSIVE;
3042 error = vn_lock(outvp, lckf);
3044 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3045 outoff, UIO_SYSSPACE, IO_NODELOCKED,
3046 curthread->td_ucred, cred, NULL, curthread);
3053 vn_finished_write(mp);
3054 } while (!growfile && xfer > 0 && error == 0);
3059 * Copy a byte range of one file to another. This function can handle the
3060 * case where invp and outvp are on different file systems.
3061 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3062 * is no better file system specific way to do it.
3065 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3066 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3067 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3072 off_t startoff, endoff, xfer, xfer2;
3074 int error, interrupted;
3075 bool cantseek, readzeros, eof, lastblock;
3077 size_t copylen, len, rem, savlen;
3079 long holein, holeout;
3081 holein = holeout = 0;
3082 savlen = len = *lenp;
3087 error = vn_lock(invp, LK_SHARED);
3090 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3095 error = vn_start_write(outvp, &mp, V_WAIT);
3097 error = vn_lock(outvp, LK_EXCLUSIVE);
3100 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3101 * now that outvp is locked.
3103 if (fsize_td != NULL) {
3104 io.uio_offset = *outoffp;
3106 error = vn_rlimit_fsize(outvp, &io, fsize_td);
3110 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3113 * Holes that are past EOF do not need to be written as a block
3114 * of zero bytes. So, truncate the output file as far as
3115 * possible and then use va.va_size to decide if writing 0
3116 * bytes is necessary in the loop below.
3119 error = VOP_GETATTR(outvp, &va, outcred);
3120 if (error == 0 && va.va_size > *outoffp && va.va_size <=
3123 error = mac_vnode_check_write(curthread->td_ucred,
3127 error = vn_truncate_locked(outvp, *outoffp,
3130 va.va_size = *outoffp;
3135 vn_finished_write(mp);
3140 * Set the blksize to the larger of the hole sizes for invp and outvp.
3141 * If hole sizes aren't available, set the blksize to the larger
3142 * f_iosize of invp and outvp.
3143 * This code expects the hole sizes and f_iosizes to be powers of 2.
3144 * This value is clipped at 4Kbytes and 1Mbyte.
3146 blksize = MAX(holein, holeout);
3148 /* Clip len to end at an exact multiple of hole size. */
3150 rem = *inoffp % blksize;
3152 rem = blksize - rem;
3153 if (len > rem && len - rem > blksize)
3154 len = savlen = rounddown(len - rem, blksize) + rem;
3158 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3159 outvp->v_mount->mnt_stat.f_iosize);
3162 else if (blksize > 1024 * 1024)
3163 blksize = 1024 * 1024;
3164 dat = malloc(blksize, M_TEMP, M_WAITOK);
3167 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3168 * to find holes. Otherwise, just scan the read block for all 0s
3169 * in the inner loop where the data copying is done.
3170 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3171 * support holes on the server, but do not support FIOSEEKHOLE.
3174 while (len > 0 && error == 0 && !eof && interrupted == 0) {
3175 endoff = 0; /* To shut up compilers. */
3181 * Find the next data area. If there is just a hole to EOF,
3182 * FIOSEEKDATA should fail and then we drop down into the
3183 * inner loop and create the hole on the outvp file.
3184 * (I do not know if any file system will report a hole to
3185 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3186 * will fail for those file systems.)
3188 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3189 * the code just falls through to the inner copy loop.
3193 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3197 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3200 * Since invp is unlocked, it may be possible for
3201 * another thread to do a truncate(), lseek(), write()
3202 * creating a hole at startoff between the above
3203 * VOP_IOCTL() calls, if the other thread does not do
3205 * If that happens, startoff == endoff and finding
3206 * the hole has failed, so set an error.
3208 if (error == 0 && startoff == endoff)
3209 error = EINVAL; /* Any error. Reset to 0. */
3212 if (startoff > *inoffp) {
3213 /* Found hole before data block. */
3214 xfer = MIN(startoff - *inoffp, len);
3215 if (*outoffp < va.va_size) {
3216 /* Must write 0s to punch hole. */
3217 xfer2 = MIN(va.va_size - *outoffp,
3219 memset(dat, 0, MIN(xfer2, blksize));
3220 error = vn_write_outvp(outvp, dat,
3221 *outoffp, xfer2, blksize, false,
3222 holeout > 0, outcred);
3225 if (error == 0 && *outoffp + xfer >
3226 va.va_size && xfer == len)
3227 /* Grow last block. */
3228 error = vn_write_outvp(outvp, dat,
3229 *outoffp, xfer, blksize, true,
3236 interrupted = sig_intr();
3239 copylen = MIN(len, endoff - startoff);
3251 * Set first xfer to end at a block boundary, so that
3252 * holes are more likely detected in the loop below via
3253 * the for all bytes 0 method.
3255 xfer -= (*inoffp % blksize);
3257 /* Loop copying the data block. */
3258 while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3261 error = vn_lock(invp, LK_SHARED);
3264 error = vn_rdwr(UIO_READ, invp, dat, xfer,
3265 startoff, UIO_SYSSPACE, IO_NODELOCKED,
3266 curthread->td_ucred, incred, &aresid,
3270 if (error == 0 && aresid > 0) {
3271 /* Stop the copy at EOF on the input file. */
3278 * Skip the write for holes past the initial EOF
3279 * of the output file, unless this is the last
3280 * write of the output file at EOF.
3282 readzeros = cantseek ? mem_iszero(dat, xfer) :
3286 if (!cantseek || *outoffp < va.va_size ||
3287 lastblock || !readzeros)
3288 error = vn_write_outvp(outvp, dat,
3289 *outoffp, xfer, blksize,
3290 readzeros && lastblock &&
3291 *outoffp >= va.va_size, false,
3300 interrupted = sig_intr();
3307 *lenp = savlen - len;
3313 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3317 off_t olen, ooffset;
3320 int audited_vnode1 = 0;
3324 if (vp->v_type != VREG)
3327 /* Allocating blocks may take a long time, so iterate. */
3334 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3337 error = vn_lock(vp, LK_EXCLUSIVE);
3339 vn_finished_write(mp);
3343 if (!audited_vnode1) {
3344 AUDIT_ARG_VNODE1(vp);
3349 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3352 error = VOP_ALLOCATE(vp, &offset, &len);
3354 vn_finished_write(mp);
3356 if (olen + ooffset != offset + len) {
3357 panic("offset + len changed from %jx/%jx to %jx/%jx",
3358 ooffset, olen, offset, len);
3360 if (error != 0 || len == 0)
3362 KASSERT(olen > len, ("Iteration did not make progress?"));
3369 static u_long vn_lock_pair_pause_cnt;
3370 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3371 &vn_lock_pair_pause_cnt, 0,
3372 "Count of vn_lock_pair deadlocks");
3374 u_int vn_lock_pair_pause_max;
3375 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3376 &vn_lock_pair_pause_max, 0,
3377 "Max ticks for vn_lock_pair deadlock avoidance sleep");
3380 vn_lock_pair_pause(const char *wmesg)
3382 atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3383 pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3387 * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3388 * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3389 * must be unlocked. Same for vp2 and vp2_locked. One of the vnodes
3392 * The function returns with both vnodes exclusively locked, and
3393 * guarantees that it does not create lock order reversal with other
3394 * threads during its execution. Both vnodes could be unlocked
3395 * temporary (and reclaimed).
3398 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3403 if (vp1 == NULL && vp2 == NULL)
3407 ASSERT_VOP_ELOCKED(vp1, "vp1");
3409 ASSERT_VOP_UNLOCKED(vp1, "vp1");
3415 ASSERT_VOP_ELOCKED(vp2, "vp2");
3417 ASSERT_VOP_UNLOCKED(vp2, "vp2");
3421 if (!vp1_locked && !vp2_locked) {
3422 vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3427 if (vp1_locked && vp2_locked)
3429 if (vp1_locked && vp2 != NULL) {
3431 error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3432 __FILE__, __LINE__);
3437 vn_lock_pair_pause("vlp1");
3439 vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3442 if (vp2_locked && vp1 != NULL) {
3444 error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3445 __FILE__, __LINE__);
3450 vn_lock_pair_pause("vlp2");
3452 vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3457 ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3459 ASSERT_VOP_ELOCKED(vp2, "vp2 ret");