2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1982, 1986, 1989, 1993
5 * The Regents of the University of California. All rights reserved.
6 * (c) UNIX System Laboratories, Inc.
7 * All or some portions of this file are derived from material licensed
8 * to the University of California by American Telephone and Telegraph
9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10 * the permission of UNIX System Laboratories, Inc.
12 * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13 * Copyright (c) 2013, 2014 The FreeBSD Foundation
15 * Portions of this software were developed by Konstantin Belousov
16 * under sponsorship from the FreeBSD Foundation.
18 * Redistribution and use in source and binary forms, with or without
19 * modification, are permitted provided that the following conditions
21 * 1. Redistributions of source code must retain the above copyright
22 * notice, this list of conditions and the following disclaimer.
23 * 2. Redistributions in binary form must reproduce the above copyright
24 * notice, this list of conditions and the following disclaimer in the
25 * documentation and/or other materials provided with the distribution.
26 * 3. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42 * @(#)vfs_vnops.c 8.2 (Berkeley) 1/21/94
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
48 #include "opt_hwpmc_hooks.h"
50 #include <sys/param.h>
51 #include <sys/systm.h>
54 #include <sys/fcntl.h>
61 #include <sys/limits.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
74 #include <sys/sleepqueue.h>
75 #include <sys/sysctl.h>
76 #include <sys/ttycom.h>
78 #include <sys/syslog.h>
79 #include <sys/unistd.h>
82 #include <security/audit/audit.h>
83 #include <security/mac/mac_framework.h>
86 #include <vm/vm_extern.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pager.h>
94 #include <sys/pmckern.h>
97 static fo_rdwr_t vn_read;
98 static fo_rdwr_t vn_write;
99 static fo_rdwr_t vn_io_fault;
100 static fo_truncate_t vn_truncate;
101 static fo_ioctl_t vn_ioctl;
102 static fo_poll_t vn_poll;
103 static fo_kqfilter_t vn_kqfilter;
104 static fo_stat_t vn_statfile;
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 static const int io_hold_cnt = 16;
129 static int vn_io_fault_enable = 1;
130 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW,
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_RW,
134 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
135 static u_long vn_io_faults_cnt;
136 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
137 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
139 static int vfs_allow_read_dir = 0;
140 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
141 &vfs_allow_read_dir, 0,
142 "Enable read(2) of directory by root for filesystems that support it");
145 * Returns true if vn_io_fault mode of handling the i/o request should
149 do_vn_io_fault(struct vnode *vp, struct uio *uio)
153 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
154 (mp = vp->v_mount) != NULL &&
155 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
159 * Structure used to pass arguments to vn_io_fault1(), to do either
160 * file- or vnode-based I/O calls.
162 struct vn_io_fault_args {
170 struct fop_args_tag {
174 struct vop_args_tag {
180 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
181 struct vn_io_fault_args *args, struct thread *td);
184 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
186 struct thread *td = ndp->ni_cnd.cn_thread;
188 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
192 * Common code for vnode open operations via a name lookup.
193 * Lookup the vnode and invoke VOP_CREATE if needed.
194 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
196 * Note that this does NOT free nameidata for the successful case,
197 * due to the NDINIT being done elsewhere.
200 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
201 struct ucred *cred, struct file *fp)
205 struct thread *td = ndp->ni_cnd.cn_thread;
207 struct vattr *vap = &vat;
212 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
213 O_EXCL | O_DIRECTORY))
215 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
216 ndp->ni_cnd.cn_nameiop = CREATE;
218 * Set NOCACHE to avoid flushing the cache when
219 * rolling in many files at once.
221 ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE;
222 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
223 ndp->ni_cnd.cn_flags |= FOLLOW;
224 if ((fmode & O_BENEATH) != 0)
225 ndp->ni_cnd.cn_flags |= BENEATH;
226 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
227 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
228 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
229 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
230 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
232 if ((error = namei(ndp)) != 0)
234 if (ndp->ni_vp == NULL) {
237 vap->va_mode = cmode;
239 vap->va_vaflags |= VA_EXCLUSIVE;
240 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
241 NDFREE(ndp, NDF_ONLY_PNBUF);
243 if ((error = vn_start_write(NULL, &mp,
244 V_XSLEEP | PCATCH)) != 0)
248 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
249 ndp->ni_cnd.cn_flags |= MAKEENTRY;
251 error = mac_vnode_check_create(cred, ndp->ni_dvp,
255 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
258 vn_finished_write(mp);
260 NDFREE(ndp, NDF_ONLY_PNBUF);
266 if (ndp->ni_dvp == ndp->ni_vp)
272 if (fmode & O_EXCL) {
276 if (vp->v_type == VDIR) {
283 ndp->ni_cnd.cn_nameiop = LOOKUP;
284 ndp->ni_cnd.cn_flags = ISOPEN |
285 ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
286 if (!(fmode & FWRITE))
287 ndp->ni_cnd.cn_flags |= LOCKSHARED;
288 if ((fmode & O_BENEATH) != 0)
289 ndp->ni_cnd.cn_flags |= BENEATH;
290 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
291 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
292 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
293 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
294 if ((error = namei(ndp)) != 0)
298 error = vn_open_vnode(vp, fmode, cred, td, fp);
304 NDFREE(ndp, NDF_ONLY_PNBUF);
312 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
315 int error, lock_flags, type;
317 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
318 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
320 KASSERT(fp != NULL, ("open with flock requires fp"));
321 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
324 lock_flags = VOP_ISLOCKED(vp);
327 lf.l_whence = SEEK_SET;
330 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
332 if ((fmode & FNONBLOCK) == 0)
334 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
336 fp->f_flag |= FHASLOCK;
338 vn_lock(vp, lock_flags | LK_RETRY);
339 if (error == 0 && VN_IS_DOOMED(vp))
345 * Common code for vnode open operations once a vnode is located.
346 * Check permissions, and call the VOP_OPEN routine.
349 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
350 struct thread *td, struct file *fp)
355 if (vp->v_type == VLNK)
357 if (vp->v_type == VSOCK)
359 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
362 if (fmode & (FWRITE | O_TRUNC)) {
363 if (vp->v_type == VDIR)
371 if ((fmode & O_APPEND) && (fmode & FWRITE))
376 if (fmode & O_VERIFY)
378 error = mac_vnode_check_open(cred, vp, accmode);
382 accmode &= ~(VCREAT | VVERIFY);
384 if ((fmode & O_CREAT) == 0 && accmode != 0) {
385 error = VOP_ACCESS(vp, accmode, cred, td);
389 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
390 vn_lock(vp, LK_UPGRADE | LK_RETRY);
391 error = VOP_OPEN(vp, fmode, cred, td, fp);
395 error = vn_open_vnode_advlock(vp, fmode, fp);
396 if (error == 0 && (fmode & FWRITE) != 0) {
397 error = VOP_ADD_WRITECOUNT(vp, 1);
399 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
400 __func__, vp, vp->v_writecount);
405 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
406 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
407 * Arrange for that by having fdrop() to use vn_closefile().
410 fp->f_flag |= FOPENFAILED;
412 if (fp->f_ops == &badfileops) {
413 fp->f_type = DTYPE_VNODE;
419 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
425 * Check for write permissions on the specified vnode.
426 * Prototype text segments cannot be written.
430 vn_writechk(struct vnode *vp)
433 ASSERT_VOP_LOCKED(vp, "vn_writechk");
435 * If there's shared text associated with
436 * the vnode, try to free it up once. If
437 * we fail, we can't allow writing.
449 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
450 struct thread *td, bool keep_ref)
453 int error, lock_flags;
455 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
456 MNT_EXTENDED_SHARED(vp->v_mount))
457 lock_flags = LK_SHARED;
459 lock_flags = LK_EXCLUSIVE;
461 vn_start_write(vp, &mp, V_WAIT);
462 vn_lock(vp, lock_flags | LK_RETRY);
463 AUDIT_ARG_VNODE1(vp);
464 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
465 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
466 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
467 __func__, vp, vp->v_writecount);
469 error = VOP_CLOSE(vp, flags, file_cred, td);
474 vn_finished_write(mp);
479 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
483 return (vn_close1(vp, flags, file_cred, td, false));
487 * Heuristic to detect sequential operation.
490 sequential_heuristic(struct uio *uio, struct file *fp)
493 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
494 if (fp->f_flag & FRDAHEAD)
495 return (fp->f_seqcount << IO_SEQSHIFT);
498 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
499 * that the first I/O is normally considered to be slightly
500 * sequential. Seeking to offset 0 doesn't change sequentiality
501 * unless previous seeks have reduced f_seqcount to 0, in which
502 * case offset 0 is not special.
504 if ((uio->uio_offset == 0 && fp->f_seqcount > 0) ||
505 uio->uio_offset == fp->f_nextoff) {
507 * f_seqcount is in units of fixed-size blocks so that it
508 * depends mainly on the amount of sequential I/O and not
509 * much on the number of sequential I/O's. The fixed size
510 * of 16384 is hard-coded here since it is (not quite) just
511 * a magic size that works well here. This size is more
512 * closely related to the best I/O size for real disks than
513 * to any block size used by software.
515 if (uio->uio_resid >= IO_SEQMAX * 16384)
516 fp->f_seqcount = IO_SEQMAX;
518 fp->f_seqcount += howmany(uio->uio_resid, 16384);
519 if (fp->f_seqcount > IO_SEQMAX)
520 fp->f_seqcount = IO_SEQMAX;
522 return (fp->f_seqcount << IO_SEQSHIFT);
525 /* Not sequential. Quickly draw-down sequentiality. */
526 if (fp->f_seqcount > 1)
534 * Package up an I/O request on a vnode into a uio and do it.
537 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
538 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
539 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
546 struct vn_io_fault_args args;
547 int error, lock_flags;
549 if (offset < 0 && vp->v_type != VCHR)
551 auio.uio_iov = &aiov;
553 aiov.iov_base = base;
555 auio.uio_resid = len;
556 auio.uio_offset = offset;
557 auio.uio_segflg = segflg;
562 if ((ioflg & IO_NODELOCKED) == 0) {
563 if ((ioflg & IO_RANGELOCKED) == 0) {
564 if (rw == UIO_READ) {
565 rl_cookie = vn_rangelock_rlock(vp, offset,
568 rl_cookie = vn_rangelock_wlock(vp, offset,
574 if (rw == UIO_WRITE) {
575 if (vp->v_type != VCHR &&
576 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
579 if (MNT_SHARED_WRITES(mp) ||
580 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
581 lock_flags = LK_SHARED;
583 lock_flags = LK_EXCLUSIVE;
585 lock_flags = LK_SHARED;
586 vn_lock(vp, lock_flags | LK_RETRY);
590 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
592 if ((ioflg & IO_NOMACCHECK) == 0) {
594 error = mac_vnode_check_read(active_cred, file_cred,
597 error = mac_vnode_check_write(active_cred, file_cred,
602 if (file_cred != NULL)
606 if (do_vn_io_fault(vp, &auio)) {
607 args.kind = VN_IO_FAULT_VOP;
610 args.args.vop_args.vp = vp;
611 error = vn_io_fault1(vp, &auio, &args, td);
612 } else if (rw == UIO_READ) {
613 error = VOP_READ(vp, &auio, ioflg, cred);
614 } else /* if (rw == UIO_WRITE) */ {
615 error = VOP_WRITE(vp, &auio, ioflg, cred);
619 *aresid = auio.uio_resid;
621 if (auio.uio_resid && error == 0)
623 if ((ioflg & IO_NODELOCKED) == 0) {
626 vn_finished_write(mp);
629 if (rl_cookie != NULL)
630 vn_rangelock_unlock(vp, rl_cookie);
635 * Package up an I/O request on a vnode into a uio and do it. The I/O
636 * request is split up into smaller chunks and we try to avoid saturating
637 * the buffer cache while potentially holding a vnode locked, so we
638 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
639 * to give other processes a chance to lock the vnode (either other processes
640 * core'ing the same binary, or unrelated processes scanning the directory).
643 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
644 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
645 struct ucred *file_cred, size_t *aresid, struct thread *td)
654 * Force `offset' to a multiple of MAXBSIZE except possibly
655 * for the first chunk, so that filesystems only need to
656 * write full blocks except possibly for the first and last
659 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
663 if (rw != UIO_READ && vp->v_type == VREG)
666 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
667 ioflg, active_cred, file_cred, &iaresid, td);
668 len -= chunk; /* aresid calc already includes length */
672 base = (char *)base + chunk;
673 kern_yield(PRI_USER);
676 *aresid = len + iaresid;
680 #if OFF_MAX <= LONG_MAX
682 foffset_lock(struct file *fp, int flags)
684 volatile short *flagsp;
688 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
690 if ((flags & FOF_NOLOCK) != 0)
691 return (atomic_load_long(&fp->f_offset));
694 * According to McKusick the vn lock was protecting f_offset here.
695 * It is now protected by the FOFFSET_LOCKED flag.
697 flagsp = &fp->f_vnread_flags;
698 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
699 return (atomic_load_long(&fp->f_offset));
701 sleepq_lock(&fp->f_vnread_flags);
702 state = atomic_load_16(flagsp);
704 if ((state & FOFFSET_LOCKED) == 0) {
705 if (!atomic_fcmpset_acq_16(flagsp, &state,
710 if ((state & FOFFSET_LOCK_WAITING) == 0) {
711 if (!atomic_fcmpset_acq_16(flagsp, &state,
712 state | FOFFSET_LOCK_WAITING))
716 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
717 sleepq_wait(&fp->f_vnread_flags, PUSER -1);
719 sleepq_lock(&fp->f_vnread_flags);
720 state = atomic_load_16(flagsp);
722 res = atomic_load_long(&fp->f_offset);
723 sleepq_release(&fp->f_vnread_flags);
728 foffset_unlock(struct file *fp, off_t val, int flags)
730 volatile short *flagsp;
733 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
735 if ((flags & FOF_NOUPDATE) == 0)
736 atomic_store_long(&fp->f_offset, val);
737 if ((flags & FOF_NEXTOFF) != 0)
740 if ((flags & FOF_NOLOCK) != 0)
743 flagsp = &fp->f_vnread_flags;
744 state = atomic_load_16(flagsp);
745 if ((state & FOFFSET_LOCK_WAITING) == 0 &&
746 atomic_cmpset_rel_16(flagsp, state, 0))
749 sleepq_lock(&fp->f_vnread_flags);
750 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
751 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
752 fp->f_vnread_flags = 0;
753 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
754 sleepq_release(&fp->f_vnread_flags);
758 foffset_lock(struct file *fp, int flags)
763 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
765 mtxp = mtx_pool_find(mtxpool_sleep, fp);
767 if ((flags & FOF_NOLOCK) == 0) {
768 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
769 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
770 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
773 fp->f_vnread_flags |= FOFFSET_LOCKED;
781 foffset_unlock(struct file *fp, off_t val, int flags)
785 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
787 mtxp = mtx_pool_find(mtxpool_sleep, fp);
789 if ((flags & FOF_NOUPDATE) == 0)
791 if ((flags & FOF_NEXTOFF) != 0)
793 if ((flags & FOF_NOLOCK) == 0) {
794 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
795 ("Lost FOFFSET_LOCKED"));
796 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
797 wakeup(&fp->f_vnread_flags);
798 fp->f_vnread_flags = 0;
805 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
808 if ((flags & FOF_OFFSET) == 0)
809 uio->uio_offset = foffset_lock(fp, flags);
813 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
816 if ((flags & FOF_OFFSET) == 0)
817 foffset_unlock(fp, uio->uio_offset, flags);
821 get_advice(struct file *fp, struct uio *uio)
826 ret = POSIX_FADV_NORMAL;
827 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
830 mtxp = mtx_pool_find(mtxpool_sleep, fp);
832 if (fp->f_advice != NULL &&
833 uio->uio_offset >= fp->f_advice->fa_start &&
834 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
835 ret = fp->f_advice->fa_advice;
841 * File table vnode read routine.
844 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
852 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
854 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
857 if (fp->f_flag & FNONBLOCK)
859 if (fp->f_flag & O_DIRECT)
861 advice = get_advice(fp, uio);
862 vn_lock(vp, LK_SHARED | LK_RETRY);
865 case POSIX_FADV_NORMAL:
866 case POSIX_FADV_SEQUENTIAL:
867 case POSIX_FADV_NOREUSE:
868 ioflag |= sequential_heuristic(uio, fp);
870 case POSIX_FADV_RANDOM:
871 /* Disable read-ahead for random I/O. */
874 orig_offset = uio->uio_offset;
877 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
880 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
881 fp->f_nextoff = uio->uio_offset;
883 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
884 orig_offset != uio->uio_offset)
886 * Use POSIX_FADV_DONTNEED to flush pages and buffers
887 * for the backing file after a POSIX_FADV_NOREUSE
890 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
891 POSIX_FADV_DONTNEED);
896 * File table vnode write routine.
899 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
905 int error, ioflag, lock_flags;
908 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
910 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
912 if (vp->v_type == VREG)
915 if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
917 if (fp->f_flag & FNONBLOCK)
919 if (fp->f_flag & O_DIRECT)
921 if ((fp->f_flag & O_FSYNC) ||
922 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
925 if (vp->v_type != VCHR &&
926 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
929 advice = get_advice(fp, uio);
931 if (MNT_SHARED_WRITES(mp) ||
932 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
933 lock_flags = LK_SHARED;
935 lock_flags = LK_EXCLUSIVE;
938 vn_lock(vp, lock_flags | LK_RETRY);
940 case POSIX_FADV_NORMAL:
941 case POSIX_FADV_SEQUENTIAL:
942 case POSIX_FADV_NOREUSE:
943 ioflag |= sequential_heuristic(uio, fp);
945 case POSIX_FADV_RANDOM:
946 /* XXX: Is this correct? */
949 orig_offset = uio->uio_offset;
952 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
955 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
956 fp->f_nextoff = uio->uio_offset;
958 if (vp->v_type != VCHR)
959 vn_finished_write(mp);
960 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
961 orig_offset != uio->uio_offset)
963 * Use POSIX_FADV_DONTNEED to flush pages and buffers
964 * for the backing file after a POSIX_FADV_NOREUSE
967 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
968 POSIX_FADV_DONTNEED);
974 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
975 * prevent the following deadlock:
977 * Assume that the thread A reads from the vnode vp1 into userspace
978 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
979 * currently not resident, then system ends up with the call chain
980 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
981 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
982 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
983 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
984 * backed by the pages of vnode vp1, and some page in buf2 is not
985 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
987 * To prevent the lock order reversal and deadlock, vn_io_fault() does
988 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
989 * Instead, it first tries to do the whole range i/o with pagefaults
990 * disabled. If all pages in the i/o buffer are resident and mapped,
991 * VOP will succeed (ignoring the genuine filesystem errors).
992 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
993 * i/o in chunks, with all pages in the chunk prefaulted and held
994 * using vm_fault_quick_hold_pages().
996 * Filesystems using this deadlock avoidance scheme should use the
997 * array of the held pages from uio, saved in the curthread->td_ma,
998 * instead of doing uiomove(). A helper function
999 * vn_io_fault_uiomove() converts uiomove request into
1000 * uiomove_fromphys() over td_ma array.
1002 * Since vnode locks do not cover the whole i/o anymore, rangelocks
1003 * make the current i/o request atomic with respect to other i/os and
1008 * Decode vn_io_fault_args and perform the corresponding i/o.
1011 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1017 save = vm_fault_disable_pagefaults();
1018 switch (args->kind) {
1019 case VN_IO_FAULT_FOP:
1020 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1021 uio, args->cred, args->flags, td);
1023 case VN_IO_FAULT_VOP:
1024 if (uio->uio_rw == UIO_READ) {
1025 error = VOP_READ(args->args.vop_args.vp, uio,
1026 args->flags, args->cred);
1027 } else if (uio->uio_rw == UIO_WRITE) {
1028 error = VOP_WRITE(args->args.vop_args.vp, uio,
1029 args->flags, args->cred);
1033 panic("vn_io_fault_doio: unknown kind of io %d %d",
1034 args->kind, uio->uio_rw);
1036 vm_fault_enable_pagefaults(save);
1041 vn_io_fault_touch(char *base, const struct uio *uio)
1046 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1052 vn_io_fault_prefault_user(const struct uio *uio)
1055 const struct iovec *iov;
1060 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1061 ("vn_io_fault_prefault userspace"));
1065 resid = uio->uio_resid;
1066 base = iov->iov_base;
1069 error = vn_io_fault_touch(base, uio);
1072 if (len < PAGE_SIZE) {
1074 error = vn_io_fault_touch(base + len - 1, uio);
1079 if (++i >= uio->uio_iovcnt)
1081 iov = uio->uio_iov + i;
1082 base = iov->iov_base;
1094 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1095 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1096 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1097 * into args and call vn_io_fault1() to handle faults during the user
1098 * mode buffer accesses.
1101 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1104 vm_page_t ma[io_hold_cnt + 2];
1105 struct uio *uio_clone, short_uio;
1106 struct iovec short_iovec[1];
1107 vm_page_t *prev_td_ma;
1109 vm_offset_t addr, end;
1112 int error, cnt, saveheld, prev_td_ma_cnt;
1114 if (vn_io_fault_prefault) {
1115 error = vn_io_fault_prefault_user(uio);
1117 return (error); /* Or ignore ? */
1120 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1123 * The UFS follows IO_UNIT directive and replays back both
1124 * uio_offset and uio_resid if an error is encountered during the
1125 * operation. But, since the iovec may be already advanced,
1126 * uio is still in an inconsistent state.
1128 * Cache a copy of the original uio, which is advanced to the redo
1129 * point using UIO_NOCOPY below.
1131 uio_clone = cloneuio(uio);
1132 resid = uio->uio_resid;
1134 short_uio.uio_segflg = UIO_USERSPACE;
1135 short_uio.uio_rw = uio->uio_rw;
1136 short_uio.uio_td = uio->uio_td;
1138 error = vn_io_fault_doio(args, uio, td);
1139 if (error != EFAULT)
1142 atomic_add_long(&vn_io_faults_cnt, 1);
1143 uio_clone->uio_segflg = UIO_NOCOPY;
1144 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1145 uio_clone->uio_segflg = uio->uio_segflg;
1147 saveheld = curthread_pflags_set(TDP_UIOHELD);
1148 prev_td_ma = td->td_ma;
1149 prev_td_ma_cnt = td->td_ma_cnt;
1151 while (uio_clone->uio_resid != 0) {
1152 len = uio_clone->uio_iov->iov_len;
1154 KASSERT(uio_clone->uio_iovcnt >= 1,
1155 ("iovcnt underflow"));
1156 uio_clone->uio_iov++;
1157 uio_clone->uio_iovcnt--;
1160 if (len > io_hold_cnt * PAGE_SIZE)
1161 len = io_hold_cnt * PAGE_SIZE;
1162 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1163 end = round_page(addr + len);
1168 cnt = atop(end - trunc_page(addr));
1170 * A perfectly misaligned address and length could cause
1171 * both the start and the end of the chunk to use partial
1172 * page. +2 accounts for such a situation.
1174 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1175 addr, len, prot, ma, io_hold_cnt + 2);
1180 short_uio.uio_iov = &short_iovec[0];
1181 short_iovec[0].iov_base = (void *)addr;
1182 short_uio.uio_iovcnt = 1;
1183 short_uio.uio_resid = short_iovec[0].iov_len = len;
1184 short_uio.uio_offset = uio_clone->uio_offset;
1186 td->td_ma_cnt = cnt;
1188 error = vn_io_fault_doio(args, &short_uio, td);
1189 vm_page_unhold_pages(ma, cnt);
1190 adv = len - short_uio.uio_resid;
1192 uio_clone->uio_iov->iov_base =
1193 (char *)uio_clone->uio_iov->iov_base + adv;
1194 uio_clone->uio_iov->iov_len -= adv;
1195 uio_clone->uio_resid -= adv;
1196 uio_clone->uio_offset += adv;
1198 uio->uio_resid -= adv;
1199 uio->uio_offset += adv;
1201 if (error != 0 || adv == 0)
1204 td->td_ma = prev_td_ma;
1205 td->td_ma_cnt = prev_td_ma_cnt;
1206 curthread_pflags_restore(saveheld);
1208 free(uio_clone, M_IOV);
1213 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1214 int flags, struct thread *td)
1219 struct vn_io_fault_args args;
1222 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1226 * The ability to read(2) on a directory has historically been
1227 * allowed for all users, but this can and has been the source of
1228 * at least one security issue in the past. As such, it is now hidden
1229 * away behind a sysctl for those that actually need it to use it, and
1230 * restricted to root when it's turned on to make it relatively safe to
1231 * leave on for longer sessions of need.
1233 if (vp->v_type == VDIR) {
1234 KASSERT(uio->uio_rw == UIO_READ,
1235 ("illegal write attempted on a directory"));
1236 if (!vfs_allow_read_dir)
1238 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1242 foffset_lock_uio(fp, uio, flags);
1243 if (do_vn_io_fault(vp, uio)) {
1244 args.kind = VN_IO_FAULT_FOP;
1245 args.args.fop_args.fp = fp;
1246 args.args.fop_args.doio = doio;
1247 args.cred = active_cred;
1248 args.flags = flags | FOF_OFFSET;
1249 if (uio->uio_rw == UIO_READ) {
1250 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1251 uio->uio_offset + uio->uio_resid);
1252 } else if ((fp->f_flag & O_APPEND) != 0 ||
1253 (flags & FOF_OFFSET) == 0) {
1254 /* For appenders, punt and lock the whole range. */
1255 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1257 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1258 uio->uio_offset + uio->uio_resid);
1260 error = vn_io_fault1(vp, uio, &args, td);
1261 vn_rangelock_unlock(vp, rl_cookie);
1263 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1265 foffset_unlock_uio(fp, uio, flags);
1270 * Helper function to perform the requested uiomove operation using
1271 * the held pages for io->uio_iov[0].iov_base buffer instead of
1272 * copyin/copyout. Access to the pages with uiomove_fromphys()
1273 * instead of iov_base prevents page faults that could occur due to
1274 * pmap_collect() invalidating the mapping created by
1275 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1276 * object cleanup revoking the write access from page mappings.
1278 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1279 * instead of plain uiomove().
1282 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1284 struct uio transp_uio;
1285 struct iovec transp_iov[1];
1291 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1292 uio->uio_segflg != UIO_USERSPACE)
1293 return (uiomove(data, xfersize, uio));
1295 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1296 transp_iov[0].iov_base = data;
1297 transp_uio.uio_iov = &transp_iov[0];
1298 transp_uio.uio_iovcnt = 1;
1299 if (xfersize > uio->uio_resid)
1300 xfersize = uio->uio_resid;
1301 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1302 transp_uio.uio_offset = 0;
1303 transp_uio.uio_segflg = UIO_SYSSPACE;
1305 * Since transp_iov points to data, and td_ma page array
1306 * corresponds to original uio->uio_iov, we need to invert the
1307 * direction of the i/o operation as passed to
1308 * uiomove_fromphys().
1310 switch (uio->uio_rw) {
1312 transp_uio.uio_rw = UIO_READ;
1315 transp_uio.uio_rw = UIO_WRITE;
1318 transp_uio.uio_td = uio->uio_td;
1319 error = uiomove_fromphys(td->td_ma,
1320 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1321 xfersize, &transp_uio);
1322 adv = xfersize - transp_uio.uio_resid;
1324 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1325 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1327 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1329 td->td_ma_cnt -= pgadv;
1330 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1331 uio->uio_iov->iov_len -= adv;
1332 uio->uio_resid -= adv;
1333 uio->uio_offset += adv;
1338 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1342 vm_offset_t iov_base;
1346 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1347 uio->uio_segflg != UIO_USERSPACE)
1348 return (uiomove_fromphys(ma, offset, xfersize, uio));
1350 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1351 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1352 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1353 switch (uio->uio_rw) {
1355 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1359 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1363 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1365 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1367 td->td_ma_cnt -= pgadv;
1368 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1369 uio->uio_iov->iov_len -= cnt;
1370 uio->uio_resid -= cnt;
1371 uio->uio_offset += cnt;
1376 * File table truncate routine.
1379 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1390 * Lock the whole range for truncation. Otherwise split i/o
1391 * might happen partly before and partly after the truncation.
1393 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1394 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1397 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1398 AUDIT_ARG_VNODE1(vp);
1399 if (vp->v_type == VDIR) {
1404 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1408 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1412 vn_finished_write(mp);
1414 vn_rangelock_unlock(vp, rl_cookie);
1419 * Truncate a file that is already locked.
1422 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1428 error = VOP_ADD_WRITECOUNT(vp, 1);
1431 vattr.va_size = length;
1433 vattr.va_vaflags |= VA_SYNC;
1434 error = VOP_SETATTR(vp, &vattr, cred);
1435 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1441 * File table vnode stat routine.
1444 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1447 struct vnode *vp = fp->f_vnode;
1450 vn_lock(vp, LK_SHARED | LK_RETRY);
1451 error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1458 * Stat a vnode; implementation for the stat syscall
1461 vn_stat(struct vnode *vp, struct stat *sb, struct ucred *active_cred,
1462 struct ucred *file_cred, struct thread *td)
1469 AUDIT_ARG_VNODE1(vp);
1471 error = mac_vnode_check_stat(active_cred, file_cred, vp);
1479 * Initialize defaults for new and unusual fields, so that file
1480 * systems which don't support these fields don't need to know
1483 vap->va_birthtime.tv_sec = -1;
1484 vap->va_birthtime.tv_nsec = 0;
1485 vap->va_fsid = VNOVAL;
1486 vap->va_rdev = NODEV;
1488 error = VOP_GETATTR(vp, vap, active_cred);
1493 * Zero the spare stat fields
1495 bzero(sb, sizeof *sb);
1498 * Copy from vattr table
1500 if (vap->va_fsid != VNOVAL)
1501 sb->st_dev = vap->va_fsid;
1503 sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1504 sb->st_ino = vap->va_fileid;
1505 mode = vap->va_mode;
1506 switch (vap->va_type) {
1532 sb->st_nlink = vap->va_nlink;
1533 sb->st_uid = vap->va_uid;
1534 sb->st_gid = vap->va_gid;
1535 sb->st_rdev = vap->va_rdev;
1536 if (vap->va_size > OFF_MAX)
1538 sb->st_size = vap->va_size;
1539 sb->st_atim.tv_sec = vap->va_atime.tv_sec;
1540 sb->st_atim.tv_nsec = vap->va_atime.tv_nsec;
1541 sb->st_mtim.tv_sec = vap->va_mtime.tv_sec;
1542 sb->st_mtim.tv_nsec = vap->va_mtime.tv_nsec;
1543 sb->st_ctim.tv_sec = vap->va_ctime.tv_sec;
1544 sb->st_ctim.tv_nsec = vap->va_ctime.tv_nsec;
1545 sb->st_birthtim.tv_sec = vap->va_birthtime.tv_sec;
1546 sb->st_birthtim.tv_nsec = vap->va_birthtime.tv_nsec;
1549 * According to www.opengroup.org, the meaning of st_blksize is
1550 * "a filesystem-specific preferred I/O block size for this
1551 * object. In some filesystem types, this may vary from file
1553 * Use miminum/default of PAGE_SIZE (e.g. for VCHR).
1556 sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1558 sb->st_flags = vap->va_flags;
1559 if (priv_check_cred_vfs_generation(td->td_ucred))
1562 sb->st_gen = vap->va_gen;
1564 sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1569 * File table vnode ioctl routine.
1572 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1577 struct fiobmap2_arg *bmarg;
1581 switch (vp->v_type) {
1586 vn_lock(vp, LK_SHARED | LK_RETRY);
1587 error = VOP_GETATTR(vp, &vattr, active_cred);
1590 *(int *)data = vattr.va_size - fp->f_offset;
1593 bmarg = (struct fiobmap2_arg *)data;
1594 vn_lock(vp, LK_SHARED | LK_RETRY);
1596 error = mac_vnode_check_read(active_cred, fp->f_cred,
1600 error = VOP_BMAP(vp, bmarg->bn, NULL,
1601 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1608 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1613 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1621 * File table vnode poll routine.
1624 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1632 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1633 AUDIT_ARG_VNODE1(vp);
1634 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1639 error = VOP_POLL(vp, events, fp->f_cred, td);
1644 * Acquire the requested lock and then check for validity. LK_RETRY
1645 * permits vn_lock to return doomed vnodes.
1647 static int __noinline
1648 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1652 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1653 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1656 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1658 if ((flags & LK_RETRY) == 0) {
1669 * Nothing to do if we got the lock.
1675 * Interlock was dropped by the call in _vn_lock.
1677 flags &= ~LK_INTERLOCK;
1679 error = VOP_LOCK1(vp, flags, file, line);
1680 } while (error != 0);
1685 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1689 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1690 ("vn_lock: no locktype (%d passed)", flags));
1691 VNPASS(vp->v_holdcnt > 0, vp);
1692 error = VOP_LOCK1(vp, flags, file, line);
1693 if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1694 return (_vn_lock_fallback(vp, flags, file, line, error));
1699 * File table vnode close routine.
1702 vn_closefile(struct file *fp, struct thread *td)
1710 fp->f_ops = &badfileops;
1711 ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1713 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1715 if (__predict_false(ref)) {
1716 lf.l_whence = SEEK_SET;
1719 lf.l_type = F_UNLCK;
1720 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1727 * Preparing to start a filesystem write operation. If the operation is
1728 * permitted, then we bump the count of operations in progress and
1729 * proceed. If a suspend request is in progress, we wait until the
1730 * suspension is over, and then proceed.
1733 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1737 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1738 vfs_op_thread_enter(mp)) {
1739 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1740 vfs_mp_count_add_pcpu(mp, writeopcount, 1);
1741 vfs_op_thread_exit(mp);
1746 mtx_assert(MNT_MTX(mp), MA_OWNED);
1753 * Check on status of suspension.
1755 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1756 mp->mnt_susp_owner != curthread) {
1757 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1758 (flags & PCATCH) : 0) | (PUSER - 1);
1759 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1760 if (flags & V_NOWAIT) {
1761 error = EWOULDBLOCK;
1764 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1770 if (flags & V_XSLEEP)
1772 mp->mnt_writeopcount++;
1774 if (error != 0 || (flags & V_XSLEEP) != 0)
1781 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1786 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1787 ("V_MNTREF requires mp"));
1791 * If a vnode is provided, get and return the mount point that
1792 * to which it will write.
1795 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1797 if (error != EOPNOTSUPP)
1802 if ((mp = *mpp) == NULL)
1806 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1808 * As long as a vnode is not provided we need to acquire a
1809 * refcount for the provided mountpoint too, in order to
1810 * emulate a vfs_ref().
1812 if (vp == NULL && (flags & V_MNTREF) == 0)
1815 return (vn_start_write_refed(mp, flags, false));
1819 * Secondary suspension. Used by operations such as vop_inactive
1820 * routines that are needed by the higher level functions. These
1821 * are allowed to proceed until all the higher level functions have
1822 * completed (indicated by mnt_writeopcount dropping to zero). At that
1823 * time, these operations are halted until the suspension is over.
1826 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1831 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1832 ("V_MNTREF requires mp"));
1836 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1838 if (error != EOPNOTSUPP)
1844 * If we are not suspended or have not yet reached suspended
1845 * mode, then let the operation proceed.
1847 if ((mp = *mpp) == NULL)
1851 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1853 * As long as a vnode is not provided we need to acquire a
1854 * refcount for the provided mountpoint too, in order to
1855 * emulate a vfs_ref().
1858 if (vp == NULL && (flags & V_MNTREF) == 0)
1860 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1861 mp->mnt_secondary_writes++;
1862 mp->mnt_secondary_accwrites++;
1866 if (flags & V_NOWAIT) {
1869 return (EWOULDBLOCK);
1872 * Wait for the suspension to finish.
1874 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1875 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1884 * Filesystem write operation has completed. If we are suspending and this
1885 * operation is the last one, notify the suspender that the suspension is
1889 vn_finished_write(struct mount *mp)
1896 if (vfs_op_thread_enter(mp)) {
1897 vfs_mp_count_sub_pcpu(mp, writeopcount, 1);
1898 vfs_mp_count_sub_pcpu(mp, ref, 1);
1899 vfs_op_thread_exit(mp);
1904 vfs_assert_mount_counters(mp);
1906 c = --mp->mnt_writeopcount;
1907 if (mp->mnt_vfs_ops == 0) {
1908 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1913 vfs_dump_mount_counters(mp);
1914 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
1915 wakeup(&mp->mnt_writeopcount);
1920 * Filesystem secondary write operation has completed. If we are
1921 * suspending and this operation is the last one, notify the suspender
1922 * that the suspension is now in effect.
1925 vn_finished_secondary_write(struct mount *mp)
1931 mp->mnt_secondary_writes--;
1932 if (mp->mnt_secondary_writes < 0)
1933 panic("vn_finished_secondary_write: neg cnt");
1934 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1935 mp->mnt_secondary_writes <= 0)
1936 wakeup(&mp->mnt_secondary_writes);
1941 * Request a filesystem to suspend write operations.
1944 vfs_write_suspend(struct mount *mp, int flags)
1951 vfs_assert_mount_counters(mp);
1952 if (mp->mnt_susp_owner == curthread) {
1953 vfs_op_exit_locked(mp);
1957 while (mp->mnt_kern_flag & MNTK_SUSPEND)
1958 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1961 * Unmount holds a write reference on the mount point. If we
1962 * own busy reference and drain for writers, we deadlock with
1963 * the reference draining in the unmount path. Callers of
1964 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1965 * vfs_busy() reference is owned and caller is not in the
1968 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1969 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1970 vfs_op_exit_locked(mp);
1975 mp->mnt_kern_flag |= MNTK_SUSPEND;
1976 mp->mnt_susp_owner = curthread;
1977 if (mp->mnt_writeopcount > 0)
1978 (void) msleep(&mp->mnt_writeopcount,
1979 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1982 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
1983 vfs_write_resume(mp, 0);
1984 /* vfs_write_resume does vfs_op_exit() for us */
1990 * Request a filesystem to resume write operations.
1993 vfs_write_resume(struct mount *mp, int flags)
1997 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1998 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
1999 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2001 mp->mnt_susp_owner = NULL;
2002 wakeup(&mp->mnt_writeopcount);
2003 wakeup(&mp->mnt_flag);
2004 curthread->td_pflags &= ~TDP_IGNSUSP;
2005 if ((flags & VR_START_WRITE) != 0) {
2007 mp->mnt_writeopcount++;
2010 if ((flags & VR_NO_SUSPCLR) == 0)
2013 } else if ((flags & VR_START_WRITE) != 0) {
2015 vn_start_write_refed(mp, 0, true);
2022 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2026 vfs_write_suspend_umnt(struct mount *mp)
2030 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2031 ("vfs_write_suspend_umnt: recursed"));
2033 /* dounmount() already called vn_start_write(). */
2035 vn_finished_write(mp);
2036 error = vfs_write_suspend(mp, 0);
2038 vn_start_write(NULL, &mp, V_WAIT);
2042 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2045 vn_start_write(NULL, &mp, V_WAIT);
2047 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2048 wakeup(&mp->mnt_flag);
2050 curthread->td_pflags |= TDP_IGNSUSP;
2055 * Implement kqueues for files by translating it to vnode operation.
2058 vn_kqfilter(struct file *fp, struct knote *kn)
2061 return (VOP_KQFILTER(fp->f_vnode, kn));
2065 * Simplified in-kernel wrapper calls for extended attribute access.
2066 * Both calls pass in a NULL credential, authorizing as "kernel" access.
2067 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2070 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2071 const char *attrname, int *buflen, char *buf, struct thread *td)
2077 iov.iov_len = *buflen;
2080 auio.uio_iov = &iov;
2081 auio.uio_iovcnt = 1;
2082 auio.uio_rw = UIO_READ;
2083 auio.uio_segflg = UIO_SYSSPACE;
2085 auio.uio_offset = 0;
2086 auio.uio_resid = *buflen;
2088 if ((ioflg & IO_NODELOCKED) == 0)
2089 vn_lock(vp, LK_SHARED | LK_RETRY);
2091 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2093 /* authorize attribute retrieval as kernel */
2094 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2097 if ((ioflg & IO_NODELOCKED) == 0)
2101 *buflen = *buflen - auio.uio_resid;
2108 * XXX failure mode if partially written?
2111 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2112 const char *attrname, int buflen, char *buf, struct thread *td)
2119 iov.iov_len = buflen;
2122 auio.uio_iov = &iov;
2123 auio.uio_iovcnt = 1;
2124 auio.uio_rw = UIO_WRITE;
2125 auio.uio_segflg = UIO_SYSSPACE;
2127 auio.uio_offset = 0;
2128 auio.uio_resid = buflen;
2130 if ((ioflg & IO_NODELOCKED) == 0) {
2131 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2133 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2136 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2138 /* authorize attribute setting as kernel */
2139 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2141 if ((ioflg & IO_NODELOCKED) == 0) {
2142 vn_finished_write(mp);
2150 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2151 const char *attrname, struct thread *td)
2156 if ((ioflg & IO_NODELOCKED) == 0) {
2157 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2159 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2162 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2164 /* authorize attribute removal as kernel */
2165 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2166 if (error == EOPNOTSUPP)
2167 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2170 if ((ioflg & IO_NODELOCKED) == 0) {
2171 vn_finished_write(mp);
2179 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2183 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2187 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2190 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2195 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2196 int lkflags, struct vnode **rvp)
2201 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2203 ltype = VOP_ISLOCKED(vp);
2204 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2205 ("vn_vget_ino: vp not locked"));
2206 error = vfs_busy(mp, MBF_NOWAIT);
2210 error = vfs_busy(mp, 0);
2211 vn_lock(vp, ltype | LK_RETRY);
2215 if (VN_IS_DOOMED(vp)) {
2221 error = alloc(mp, alloc_arg, lkflags, rvp);
2223 if (error != 0 || *rvp != vp)
2224 vn_lock(vp, ltype | LK_RETRY);
2225 if (VN_IS_DOOMED(vp)) {
2238 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2242 if (vp->v_type != VREG || td == NULL)
2244 if ((uoff_t)uio->uio_offset + uio->uio_resid >
2245 lim_cur(td, RLIMIT_FSIZE)) {
2246 PROC_LOCK(td->td_proc);
2247 kern_psignal(td->td_proc, SIGXFSZ);
2248 PROC_UNLOCK(td->td_proc);
2255 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2262 vn_lock(vp, LK_SHARED | LK_RETRY);
2263 AUDIT_ARG_VNODE1(vp);
2266 return (setfmode(td, active_cred, vp, mode));
2270 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2277 vn_lock(vp, LK_SHARED | LK_RETRY);
2278 AUDIT_ARG_VNODE1(vp);
2281 return (setfown(td, active_cred, vp, uid, gid));
2285 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2289 if ((object = vp->v_object) == NULL)
2291 VM_OBJECT_WLOCK(object);
2292 vm_object_page_remove(object, start, end, 0);
2293 VM_OBJECT_WUNLOCK(object);
2297 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2305 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2306 ("Wrong command %lu", cmd));
2308 if (vn_lock(vp, LK_SHARED) != 0)
2310 if (vp->v_type != VREG) {
2314 error = VOP_GETATTR(vp, &va, cred);
2318 if (noff >= va.va_size) {
2322 bsize = vp->v_mount->mnt_stat.f_iosize;
2323 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2325 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2326 if (error == EOPNOTSUPP) {
2330 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2331 (bnp != -1 && cmd == FIOSEEKDATA)) {
2338 if (noff > va.va_size)
2340 /* noff == va.va_size. There is an implicit hole at the end of file. */
2341 if (cmd == FIOSEEKDATA)
2351 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2356 off_t foffset, size;
2359 cred = td->td_ucred;
2361 foffset = foffset_lock(fp, 0);
2362 noneg = (vp->v_type != VCHR);
2368 (offset > 0 && foffset > OFF_MAX - offset))) {
2375 vn_lock(vp, LK_SHARED | LK_RETRY);
2376 error = VOP_GETATTR(vp, &vattr, cred);
2382 * If the file references a disk device, then fetch
2383 * the media size and use that to determine the ending
2386 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2387 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2388 vattr.va_size = size;
2390 (vattr.va_size > OFF_MAX ||
2391 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2395 offset += vattr.va_size;
2400 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2401 if (error == ENOTTY)
2405 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2406 if (error == ENOTTY)
2412 if (error == 0 && noneg && offset < 0)
2416 VFS_KNOTE_UNLOCKED(vp, 0);
2417 td->td_uretoff.tdu_off = offset;
2419 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2424 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2430 * Grant permission if the caller is the owner of the file, or
2431 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2432 * on the file. If the time pointer is null, then write
2433 * permission on the file is also sufficient.
2435 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2436 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2437 * will be allowed to set the times [..] to the current
2440 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2441 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2442 error = VOP_ACCESS(vp, VWRITE, cred, td);
2447 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2452 if (fp->f_type == DTYPE_FIFO)
2453 kif->kf_type = KF_TYPE_FIFO;
2455 kif->kf_type = KF_TYPE_VNODE;
2458 FILEDESC_SUNLOCK(fdp);
2459 error = vn_fill_kinfo_vnode(vp, kif);
2461 FILEDESC_SLOCK(fdp);
2466 vn_fill_junk(struct kinfo_file *kif)
2471 * Simulate vn_fullpath returning changing values for a given
2472 * vp during e.g. coredump.
2474 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2475 olen = strlen(kif->kf_path);
2477 strcpy(&kif->kf_path[len - 1], "$");
2479 for (; olen < len; olen++)
2480 strcpy(&kif->kf_path[olen], "A");
2484 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2487 char *fullpath, *freepath;
2490 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2493 error = vn_fullpath(curthread, vp, &fullpath, &freepath);
2495 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2497 if (freepath != NULL)
2498 free(freepath, M_TEMP);
2500 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2505 * Retrieve vnode attributes.
2507 va.va_fsid = VNOVAL;
2509 vn_lock(vp, LK_SHARED | LK_RETRY);
2510 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2514 if (va.va_fsid != VNOVAL)
2515 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2517 kif->kf_un.kf_file.kf_file_fsid =
2518 vp->v_mount->mnt_stat.f_fsid.val[0];
2519 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2520 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2521 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2522 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2523 kif->kf_un.kf_file.kf_file_size = va.va_size;
2524 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2525 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2526 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2531 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2532 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2536 struct pmckern_map_in pkm;
2542 boolean_t writecounted;
2545 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2546 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2548 * POSIX shared-memory objects are defined to have
2549 * kernel persistence, and are not defined to support
2550 * read(2)/write(2) -- or even open(2). Thus, we can
2551 * use MAP_ASYNC to trade on-disk coherence for speed.
2552 * The shm_open(3) library routine turns on the FPOSIXSHM
2553 * flag to request this behavior.
2555 if ((fp->f_flag & FPOSIXSHM) != 0)
2556 flags |= MAP_NOSYNC;
2561 * Ensure that file and memory protections are
2562 * compatible. Note that we only worry about
2563 * writability if mapping is shared; in this case,
2564 * current and max prot are dictated by the open file.
2565 * XXX use the vnode instead? Problem is: what
2566 * credentials do we use for determination? What if
2567 * proc does a setuid?
2570 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2571 maxprot = VM_PROT_NONE;
2572 if ((prot & VM_PROT_EXECUTE) != 0)
2575 maxprot = VM_PROT_EXECUTE;
2576 if ((fp->f_flag & FREAD) != 0)
2577 maxprot |= VM_PROT_READ;
2578 else if ((prot & VM_PROT_READ) != 0)
2582 * If we are sharing potential changes via MAP_SHARED and we
2583 * are trying to get write permission although we opened it
2584 * without asking for it, bail out.
2586 if ((flags & MAP_SHARED) != 0) {
2587 if ((fp->f_flag & FWRITE) != 0)
2588 maxprot |= VM_PROT_WRITE;
2589 else if ((prot & VM_PROT_WRITE) != 0)
2592 maxprot |= VM_PROT_WRITE;
2593 cap_maxprot |= VM_PROT_WRITE;
2595 maxprot &= cap_maxprot;
2598 * For regular files and shared memory, POSIX requires that
2599 * the value of foff be a legitimate offset within the data
2600 * object. In particular, negative offsets are invalid.
2601 * Blocking negative offsets and overflows here avoids
2602 * possible wraparound or user-level access into reserved
2603 * ranges of the data object later. In contrast, POSIX does
2604 * not dictate how offsets are used by device drivers, so in
2605 * the case of a device mapping a negative offset is passed
2612 foff < 0 || foff > OFF_MAX - size)
2615 writecounted = FALSE;
2616 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2617 &foff, &object, &writecounted);
2620 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2621 foff, writecounted, td);
2624 * If this mapping was accounted for in the vnode's
2625 * writecount, then undo that now.
2628 vm_pager_release_writecount(object, 0, size);
2629 vm_object_deallocate(object);
2632 /* Inform hwpmc(4) if an executable is being mapped. */
2633 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2634 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2636 pkm.pm_address = (uintptr_t) *addr;
2637 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2645 vn_fsid(struct vnode *vp, struct vattr *va)
2649 f = &vp->v_mount->mnt_stat.f_fsid;
2650 va->va_fsid = (uint32_t)f->val[1];
2651 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2652 va->va_fsid += (uint32_t)f->val[0];
2656 vn_fsync_buf(struct vnode *vp, int waitfor)
2658 struct buf *bp, *nbp;
2661 int error, maxretry;
2664 maxretry = 10000; /* large, arbitrarily chosen */
2666 if (vp->v_type == VCHR) {
2668 mp = vp->v_rdev->si_mountpt;
2675 * MARK/SCAN initialization to avoid infinite loops.
2677 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2678 bp->b_vflags &= ~BV_SCANNED;
2683 * Flush all dirty buffers associated with a vnode.
2686 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2687 if ((bp->b_vflags & BV_SCANNED) != 0)
2689 bp->b_vflags |= BV_SCANNED;
2690 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2691 if (waitfor != MNT_WAIT)
2694 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2695 BO_LOCKPTR(bo)) != 0) {
2702 KASSERT(bp->b_bufobj == bo,
2703 ("bp %p wrong b_bufobj %p should be %p",
2704 bp, bp->b_bufobj, bo));
2705 if ((bp->b_flags & B_DELWRI) == 0)
2706 panic("fsync: not dirty");
2707 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2713 if (maxretry < 1000)
2714 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2720 * If synchronous the caller expects us to completely resolve all
2721 * dirty buffers in the system. Wait for in-progress I/O to
2722 * complete (which could include background bitmap writes), then
2723 * retry if dirty blocks still exist.
2725 if (waitfor == MNT_WAIT) {
2726 bufobj_wwait(bo, 0, 0);
2727 if (bo->bo_dirty.bv_cnt > 0) {
2729 * If we are unable to write any of these buffers
2730 * then we fail now rather than trying endlessly
2731 * to write them out.
2733 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2734 if ((error = bp->b_error) != 0)
2736 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2737 (error == 0 && --maxretry >= 0))
2745 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2751 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
2752 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2753 * to do the actual copy.
2754 * vn_generic_copy_file_range() is factored out, so it can be called
2755 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2756 * different file systems.
2759 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2760 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2761 struct ucred *outcred, struct thread *fsize_td)
2765 uint64_t uvalin, uvalout;
2768 *lenp = 0; /* For error returns. */
2771 /* Do some sanity checks on the arguments. */
2776 if (invp->v_type == VDIR || outvp->v_type == VDIR)
2778 else if (*inoffp < 0 || uvalin > INT64_MAX || uvalin <
2779 (uint64_t)*inoffp || *outoffp < 0 || uvalout > INT64_MAX ||
2780 uvalout < (uint64_t)*outoffp || invp->v_type != VREG ||
2781 outvp->v_type != VREG)
2787 * If the two vnode are for the same file system, call
2788 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2789 * which can handle copies across multiple file systems.
2792 if (invp->v_mount == outvp->v_mount)
2793 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2794 lenp, flags, incred, outcred, fsize_td);
2796 error = vn_generic_copy_file_range(invp, inoffp, outvp,
2797 outoffp, lenp, flags, incred, outcred, fsize_td);
2803 * Test len bytes of data starting at dat for all bytes == 0.
2804 * Return true if all bytes are zero, false otherwise.
2805 * Expects dat to be well aligned.
2808 mem_iszero(void *dat, int len)
2814 for (p = dat; len > 0; len -= sizeof(*p), p++) {
2815 if (len >= sizeof(*p)) {
2819 cp = (const char *)p;
2820 for (i = 0; i < len; i++, cp++)
2829 * Look for a hole in the output file and, if found, adjust *outoffp
2830 * and *xferp to skip past the hole.
2831 * *xferp is the entire hole length to be written and xfer2 is how many bytes
2832 * to be written as 0's upon return.
2835 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2836 off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2841 if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2842 *dataoffp = *outoffp;
2843 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2846 *holeoffp = *dataoffp;
2847 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2850 if (error != 0 || *holeoffp == *dataoffp) {
2852 * Since outvp is unlocked, it may be possible for
2853 * another thread to do a truncate(), lseek(), write()
2854 * creating a hole at startoff between the above
2855 * VOP_IOCTL() calls, if the other thread does not do
2857 * If that happens, *holeoffp == *dataoffp and finding
2858 * the hole has failed, so disable vn_skip_hole().
2860 *holeoffp = -1; /* Disable use of vn_skip_hole(). */
2863 KASSERT(*dataoffp >= *outoffp,
2864 ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2865 (intmax_t)*dataoffp, (intmax_t)*outoffp));
2866 KASSERT(*holeoffp > *dataoffp,
2867 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2868 (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2872 * If there is a hole before the data starts, advance *outoffp and
2873 * *xferp past the hole.
2875 if (*dataoffp > *outoffp) {
2876 delta = *dataoffp - *outoffp;
2877 if (delta >= *xferp) {
2878 /* Entire *xferp is a hole. */
2885 xfer2 = MIN(xfer2, *xferp);
2889 * If a hole starts before the end of this xfer2, reduce this xfer2 so
2890 * that the write ends at the start of the hole.
2891 * *holeoffp should always be greater than *outoffp, but for the
2892 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
2895 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
2896 xfer2 = *holeoffp - *outoffp;
2901 * Write an xfer sized chunk to outvp in blksize blocks from dat.
2902 * dat is a maximum of blksize in length and can be written repeatedly in
2904 * If growfile == true, just grow the file via vn_truncate_locked() instead
2905 * of doing actual writes.
2906 * If checkhole == true, a hole is being punched, so skip over any hole
2907 * already in the output file.
2910 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
2911 u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
2914 off_t dataoff, holeoff, xfer2;
2918 * Loop around doing writes of blksize until write has been completed.
2919 * Lock/unlock on each loop iteration so that a bwillwrite() can be
2920 * done for each iteration, since the xfer argument can be very
2921 * large if there is a large hole to punch in the output file.
2926 xfer2 = MIN(xfer, blksize);
2929 * Punching a hole. Skip writing if there is
2930 * already a hole in the output file.
2932 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
2933 &dataoff, &holeoff, cred);
2938 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
2943 error = vn_start_write(outvp, &mp, V_WAIT);
2945 if (MNT_SHARED_WRITES(mp))
2948 lckf = LK_EXCLUSIVE;
2949 error = vn_lock(outvp, lckf);
2953 error = vn_truncate_locked(outvp, outoff + xfer,
2956 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
2957 outoff, UIO_SYSSPACE, IO_NODELOCKED,
2958 curthread->td_ucred, cred, NULL, curthread);
2965 vn_finished_write(mp);
2966 } while (!growfile && xfer > 0 && error == 0);
2971 * Copy a byte range of one file to another. This function can handle the
2972 * case where invp and outvp are on different file systems.
2973 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
2974 * is no better file system specific way to do it.
2977 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
2978 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
2979 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
2984 off_t startoff, endoff, xfer, xfer2;
2987 bool cantseek, readzeros, eof, lastblock;
2989 size_t copylen, len, savlen;
2991 long holein, holeout;
2993 holein = holeout = 0;
2994 savlen = len = *lenp;
2998 error = vn_lock(invp, LK_SHARED);
3001 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3006 error = vn_start_write(outvp, &mp, V_WAIT);
3008 error = vn_lock(outvp, LK_EXCLUSIVE);
3011 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3012 * now that outvp is locked.
3014 if (fsize_td != NULL) {
3015 io.uio_offset = *outoffp;
3017 error = vn_rlimit_fsize(outvp, &io, fsize_td);
3021 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3024 * Holes that are past EOF do not need to be written as a block
3025 * of zero bytes. So, truncate the output file as far as
3026 * possible and then use va.va_size to decide if writing 0
3027 * bytes is necessary in the loop below.
3030 error = VOP_GETATTR(outvp, &va, outcred);
3031 if (error == 0 && va.va_size > *outoffp && va.va_size <=
3034 error = mac_vnode_check_write(curthread->td_ucred,
3038 error = vn_truncate_locked(outvp, *outoffp,
3041 va.va_size = *outoffp;
3046 vn_finished_write(mp);
3051 * Set the blksize to the larger of the hole sizes for invp and outvp.
3052 * If hole sizes aren't available, set the blksize to the larger
3053 * f_iosize of invp and outvp.
3054 * This code expects the hole sizes and f_iosizes to be powers of 2.
3055 * This value is clipped at 4Kbytes and 1Mbyte.
3057 blksize = MAX(holein, holeout);
3059 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3060 outvp->v_mount->mnt_stat.f_iosize);
3063 else if (blksize > 1024 * 1024)
3064 blksize = 1024 * 1024;
3065 dat = malloc(blksize, M_TEMP, M_WAITOK);
3068 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3069 * to find holes. Otherwise, just scan the read block for all 0s
3070 * in the inner loop where the data copying is done.
3071 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3072 * support holes on the server, but do not support FIOSEEKHOLE.
3075 while (len > 0 && error == 0 && !eof) {
3076 endoff = 0; /* To shut up compilers. */
3082 * Find the next data area. If there is just a hole to EOF,
3083 * FIOSEEKDATA should fail and then we drop down into the
3084 * inner loop and create the hole on the outvp file.
3085 * (I do not know if any file system will report a hole to
3086 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3087 * will fail for those file systems.)
3089 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3090 * the code just falls through to the inner copy loop.
3094 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3098 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3101 * Since invp is unlocked, it may be possible for
3102 * another thread to do a truncate(), lseek(), write()
3103 * creating a hole at startoff between the above
3104 * VOP_IOCTL() calls, if the other thread does not do
3106 * If that happens, startoff == endoff and finding
3107 * the hole has failed, so set an error.
3109 if (error == 0 && startoff == endoff)
3110 error = EINVAL; /* Any error. Reset to 0. */
3113 if (startoff > *inoffp) {
3114 /* Found hole before data block. */
3115 xfer = MIN(startoff - *inoffp, len);
3116 if (*outoffp < va.va_size) {
3117 /* Must write 0s to punch hole. */
3118 xfer2 = MIN(va.va_size - *outoffp,
3120 memset(dat, 0, MIN(xfer2, blksize));
3121 error = vn_write_outvp(outvp, dat,
3122 *outoffp, xfer2, blksize, false,
3123 holeout > 0, outcred);
3126 if (error == 0 && *outoffp + xfer >
3127 va.va_size && xfer == len)
3128 /* Grow last block. */
3129 error = vn_write_outvp(outvp, dat,
3130 *outoffp, xfer, blksize, true,
3138 copylen = MIN(len, endoff - startoff);
3150 * Set first xfer to end at a block boundary, so that
3151 * holes are more likely detected in the loop below via
3152 * the for all bytes 0 method.
3154 xfer -= (*inoffp % blksize);
3156 /* Loop copying the data block. */
3157 while (copylen > 0 && error == 0 && !eof) {
3160 error = vn_lock(invp, LK_SHARED);
3163 error = vn_rdwr(UIO_READ, invp, dat, xfer,
3164 startoff, UIO_SYSSPACE, IO_NODELOCKED,
3165 curthread->td_ucred, incred, &aresid,
3169 if (error == 0 && aresid > 0) {
3170 /* Stop the copy at EOF on the input file. */
3177 * Skip the write for holes past the initial EOF
3178 * of the output file, unless this is the last
3179 * write of the output file at EOF.
3181 readzeros = cantseek ? mem_iszero(dat, xfer) :
3185 if (!cantseek || *outoffp < va.va_size ||
3186 lastblock || !readzeros)
3187 error = vn_write_outvp(outvp, dat,
3188 *outoffp, xfer, blksize,
3189 readzeros && lastblock &&
3190 *outoffp >= va.va_size, false,
3204 *lenp = savlen - len;
3210 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3214 off_t olen, ooffset;
3217 int audited_vnode1 = 0;
3221 if (vp->v_type != VREG)
3224 /* Allocating blocks may take a long time, so iterate. */
3231 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3234 error = vn_lock(vp, LK_EXCLUSIVE);
3236 vn_finished_write(mp);
3240 if (!audited_vnode1) {
3241 AUDIT_ARG_VNODE1(vp);
3246 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3249 error = VOP_ALLOCATE(vp, &offset, &len);
3251 vn_finished_write(mp);
3253 if (olen + ooffset != offset + len) {
3254 panic("offset + len changed from %jx/%jx to %jx/%jx",
3255 ooffset, olen, offset, len);
3257 if (error != 0 || len == 0)
3259 KASSERT(olen > len, ("Iteration did not make progress?"));