2 * Copyright (c) 1982, 1986, 1989, 1993
3 * The Regents of the University of California. All rights reserved.
4 * (c) UNIX System Laboratories, Inc.
5 * All or some portions of this file are derived from material licensed
6 * to the University of California by American Telephone and Telegraph
7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8 * the permission of UNIX System Laboratories, Inc.
10 * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
11 * Copyright (c) 2013, 2014 The FreeBSD Foundation
13 * Portions of this software were developed by Konstantin Belousov
14 * under sponsorship from the FreeBSD Foundation.
16 * Redistribution and use in source and binary forms, with or without
17 * modification, are permitted provided that the following conditions
19 * 1. Redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer.
21 * 2. Redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution.
24 * 4. Neither the name of the University nor the names of its contributors
25 * may be used to endorse or promote products derived from this software
26 * without specific prior written permission.
28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40 * @(#)vfs_vnops.c 8.2 (Berkeley) 1/21/94
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
46 #include <sys/param.h>
47 #include <sys/systm.h>
49 #include <sys/fcntl.h>
55 #include <sys/limits.h>
57 #include <sys/mount.h>
58 #include <sys/mutex.h>
59 #include <sys/namei.h>
60 #include <sys/vnode.h>
63 #include <sys/filio.h>
64 #include <sys/resourcevar.h>
65 #include <sys/rwlock.h>
67 #include <sys/sysctl.h>
68 #include <sys/ttycom.h>
70 #include <sys/syslog.h>
71 #include <sys/unistd.h>
73 #include <security/audit/audit.h>
74 #include <security/mac/mac_framework.h>
77 #include <vm/vm_extern.h>
79 #include <vm/vm_map.h>
80 #include <vm/vm_object.h>
81 #include <vm/vm_page.h>
83 static fo_rdwr_t vn_read;
84 static fo_rdwr_t vn_write;
85 static fo_rdwr_t vn_io_fault;
86 static fo_truncate_t vn_truncate;
87 static fo_ioctl_t vn_ioctl;
88 static fo_poll_t vn_poll;
89 static fo_kqfilter_t vn_kqfilter;
90 static fo_stat_t vn_statfile;
91 static fo_close_t vn_closefile;
93 struct fileops vnops = {
94 .fo_read = vn_io_fault,
95 .fo_write = vn_io_fault,
96 .fo_truncate = vn_truncate,
99 .fo_kqfilter = vn_kqfilter,
100 .fo_stat = vn_statfile,
101 .fo_close = vn_closefile,
102 .fo_chmod = vn_chmod,
103 .fo_chown = vn_chown,
104 .fo_sendfile = vn_sendfile,
106 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
109 static const int io_hold_cnt = 16;
110 static int vn_io_fault_enable = 1;
111 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW,
112 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
113 static u_long vn_io_faults_cnt;
114 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
115 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
118 * Returns true if vn_io_fault mode of handling the i/o request should
122 do_vn_io_fault(struct vnode *vp, struct uio *uio)
126 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
127 (mp = vp->v_mount) != NULL &&
128 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
132 * Structure used to pass arguments to vn_io_fault1(), to do either
133 * file- or vnode-based I/O calls.
135 struct vn_io_fault_args {
143 struct fop_args_tag {
147 struct vop_args_tag {
153 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
154 struct vn_io_fault_args *args, struct thread *td);
157 vn_open(ndp, flagp, cmode, fp)
158 struct nameidata *ndp;
162 struct thread *td = ndp->ni_cnd.cn_thread;
164 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
168 * Common code for vnode open operations via a name lookup.
169 * Lookup the vnode and invoke VOP_CREATE if needed.
170 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
172 * Note that this does NOT free nameidata for the successful case,
173 * due to the NDINIT being done elsewhere.
176 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
177 struct ucred *cred, struct file *fp)
181 struct thread *td = ndp->ni_cnd.cn_thread;
183 struct vattr *vap = &vat;
188 if (fmode & O_CREAT) {
189 ndp->ni_cnd.cn_nameiop = CREATE;
191 * Set NOCACHE to avoid flushing the cache when
192 * rolling in many files at once.
194 ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE;
195 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
196 ndp->ni_cnd.cn_flags |= FOLLOW;
197 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
198 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
199 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
200 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
202 if ((error = namei(ndp)) != 0)
204 if (ndp->ni_vp == NULL) {
207 vap->va_mode = cmode;
209 vap->va_vaflags |= VA_EXCLUSIVE;
210 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
211 NDFREE(ndp, NDF_ONLY_PNBUF);
213 if ((error = vn_start_write(NULL, &mp,
214 V_XSLEEP | PCATCH)) != 0)
218 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
219 ndp->ni_cnd.cn_flags |= MAKEENTRY;
221 error = mac_vnode_check_create(cred, ndp->ni_dvp,
225 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
228 vn_finished_write(mp);
230 NDFREE(ndp, NDF_ONLY_PNBUF);
236 if (ndp->ni_dvp == ndp->ni_vp)
242 if (fmode & O_EXCL) {
249 ndp->ni_cnd.cn_nameiop = LOOKUP;
250 ndp->ni_cnd.cn_flags = ISOPEN |
251 ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
252 if (!(fmode & FWRITE))
253 ndp->ni_cnd.cn_flags |= LOCKSHARED;
254 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
255 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
256 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
257 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
258 if ((error = namei(ndp)) != 0)
262 error = vn_open_vnode(vp, fmode, cred, td, fp);
268 NDFREE(ndp, NDF_ONLY_PNBUF);
276 * Common code for vnode open operations once a vnode is located.
277 * Check permissions, and call the VOP_OPEN routine.
280 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
281 struct thread *td, struct file *fp)
286 int error, have_flock, lock_flags, type;
288 if (vp->v_type == VLNK)
290 if (vp->v_type == VSOCK)
292 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
295 if (fmode & (FWRITE | O_TRUNC)) {
296 if (vp->v_type == VDIR)
304 if ((fmode & O_APPEND) && (fmode & FWRITE))
307 error = mac_vnode_check_open(cred, vp, accmode);
311 if ((fmode & O_CREAT) == 0) {
312 if (accmode & VWRITE) {
313 error = vn_writechk(vp);
318 error = VOP_ACCESS(vp, accmode, cred, td);
323 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
324 vn_lock(vp, LK_UPGRADE | LK_RETRY);
325 if ((error = VOP_OPEN(vp, fmode, cred, td, fp)) != 0)
328 if (fmode & (O_EXLOCK | O_SHLOCK)) {
329 KASSERT(fp != NULL, ("open with flock requires fp"));
330 lock_flags = VOP_ISLOCKED(vp);
332 lf.l_whence = SEEK_SET;
335 if (fmode & O_EXLOCK)
340 if ((fmode & FNONBLOCK) == 0)
342 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
343 have_flock = (error == 0);
344 vn_lock(vp, lock_flags | LK_RETRY);
345 if (error == 0 && vp->v_iflag & VI_DOOMED)
348 * Another thread might have used this vnode as an
349 * executable while the vnode lock was dropped.
350 * Ensure the vnode is still able to be opened for
351 * writing after the lock has been obtained.
353 if (error == 0 && accmode & VWRITE)
354 error = vn_writechk(vp);
358 lf.l_whence = SEEK_SET;
362 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf,
365 vn_start_write(vp, &mp, V_WAIT);
366 vn_lock(vp, lock_flags | LK_RETRY);
367 (void)VOP_CLOSE(vp, fmode, cred, td);
368 vn_finished_write(mp);
369 /* Prevent second close from fdrop()->vn_close(). */
371 fp->f_ops= &badfileops;
374 fp->f_flag |= FHASLOCK;
376 if (fmode & FWRITE) {
377 VOP_ADD_WRITECOUNT(vp, 1);
378 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
379 __func__, vp, vp->v_writecount);
381 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
386 * Check for write permissions on the specified vnode.
387 * Prototype text segments cannot be written.
391 register struct vnode *vp;
394 ASSERT_VOP_LOCKED(vp, "vn_writechk");
396 * If there's shared text associated with
397 * the vnode, try to free it up once. If
398 * we fail, we can't allow writing.
410 vn_close(vp, flags, file_cred, td)
411 register struct vnode *vp;
413 struct ucred *file_cred;
417 int error, lock_flags;
419 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
420 MNT_EXTENDED_SHARED(vp->v_mount))
421 lock_flags = LK_SHARED;
423 lock_flags = LK_EXCLUSIVE;
425 vn_start_write(vp, &mp, V_WAIT);
426 vn_lock(vp, lock_flags | LK_RETRY);
427 if (flags & FWRITE) {
428 VNASSERT(vp->v_writecount > 0, vp,
429 ("vn_close: negative writecount"));
430 VOP_ADD_WRITECOUNT(vp, -1);
431 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
432 __func__, vp, vp->v_writecount);
434 error = VOP_CLOSE(vp, flags, file_cred, td);
436 vn_finished_write(mp);
441 * Heuristic to detect sequential operation.
444 sequential_heuristic(struct uio *uio, struct file *fp)
447 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
448 if (fp->f_flag & FRDAHEAD)
449 return (fp->f_seqcount << IO_SEQSHIFT);
452 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
453 * that the first I/O is normally considered to be slightly
454 * sequential. Seeking to offset 0 doesn't change sequentiality
455 * unless previous seeks have reduced f_seqcount to 0, in which
456 * case offset 0 is not special.
458 if ((uio->uio_offset == 0 && fp->f_seqcount > 0) ||
459 uio->uio_offset == fp->f_nextoff) {
461 * f_seqcount is in units of fixed-size blocks so that it
462 * depends mainly on the amount of sequential I/O and not
463 * much on the number of sequential I/O's. The fixed size
464 * of 16384 is hard-coded here since it is (not quite) just
465 * a magic size that works well here. This size is more
466 * closely related to the best I/O size for real disks than
467 * to any block size used by software.
469 fp->f_seqcount += howmany(uio->uio_resid, 16384);
470 if (fp->f_seqcount > IO_SEQMAX)
471 fp->f_seqcount = IO_SEQMAX;
472 return (fp->f_seqcount << IO_SEQSHIFT);
475 /* Not sequential. Quickly draw-down sequentiality. */
476 if (fp->f_seqcount > 1)
484 * Package up an I/O request on a vnode into a uio and do it.
487 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
488 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
489 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
496 struct vn_io_fault_args args;
497 int error, lock_flags;
499 auio.uio_iov = &aiov;
501 aiov.iov_base = base;
503 auio.uio_resid = len;
504 auio.uio_offset = offset;
505 auio.uio_segflg = segflg;
510 if ((ioflg & IO_NODELOCKED) == 0) {
511 if ((ioflg & IO_RANGELOCKED) == 0) {
512 if (rw == UIO_READ) {
513 rl_cookie = vn_rangelock_rlock(vp, offset,
516 rl_cookie = vn_rangelock_wlock(vp, offset,
522 if (rw == UIO_WRITE) {
523 if (vp->v_type != VCHR &&
524 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
527 if (MNT_SHARED_WRITES(mp) ||
528 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
529 lock_flags = LK_SHARED;
531 lock_flags = LK_EXCLUSIVE;
533 lock_flags = LK_SHARED;
534 vn_lock(vp, lock_flags | LK_RETRY);
538 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
540 if ((ioflg & IO_NOMACCHECK) == 0) {
542 error = mac_vnode_check_read(active_cred, file_cred,
545 error = mac_vnode_check_write(active_cred, file_cred,
550 if (file_cred != NULL)
554 if (do_vn_io_fault(vp, &auio)) {
555 args.kind = VN_IO_FAULT_VOP;
558 args.args.vop_args.vp = vp;
559 error = vn_io_fault1(vp, &auio, &args, td);
560 } else if (rw == UIO_READ) {
561 error = VOP_READ(vp, &auio, ioflg, cred);
562 } else /* if (rw == UIO_WRITE) */ {
563 error = VOP_WRITE(vp, &auio, ioflg, cred);
567 *aresid = auio.uio_resid;
569 if (auio.uio_resid && error == 0)
571 if ((ioflg & IO_NODELOCKED) == 0) {
574 vn_finished_write(mp);
577 if (rl_cookie != NULL)
578 vn_rangelock_unlock(vp, rl_cookie);
583 * Package up an I/O request on a vnode into a uio and do it. The I/O
584 * request is split up into smaller chunks and we try to avoid saturating
585 * the buffer cache while potentially holding a vnode locked, so we
586 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
587 * to give other processes a chance to lock the vnode (either other processes
588 * core'ing the same binary, or unrelated processes scanning the directory).
591 vn_rdwr_inchunks(rw, vp, base, len, offset, segflg, ioflg, active_cred,
592 file_cred, aresid, td)
600 struct ucred *active_cred;
601 struct ucred *file_cred;
612 * Force `offset' to a multiple of MAXBSIZE except possibly
613 * for the first chunk, so that filesystems only need to
614 * write full blocks except possibly for the first and last
617 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
621 if (rw != UIO_READ && vp->v_type == VREG)
624 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
625 ioflg, active_cred, file_cred, &iaresid, td);
626 len -= chunk; /* aresid calc already includes length */
630 base = (char *)base + chunk;
631 kern_yield(PRI_USER);
634 *aresid = len + iaresid;
639 foffset_lock(struct file *fp, int flags)
644 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
646 #if OFF_MAX <= LONG_MAX
648 * Caller only wants the current f_offset value. Assume that
649 * the long and shorter integer types reads are atomic.
651 if ((flags & FOF_NOLOCK) != 0)
652 return (fp->f_offset);
656 * According to McKusick the vn lock was protecting f_offset here.
657 * It is now protected by the FOFFSET_LOCKED flag.
659 mtxp = mtx_pool_find(mtxpool_sleep, fp);
661 if ((flags & FOF_NOLOCK) == 0) {
662 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
663 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
664 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
667 fp->f_vnread_flags |= FOFFSET_LOCKED;
675 foffset_unlock(struct file *fp, off_t val, int flags)
679 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
681 #if OFF_MAX <= LONG_MAX
682 if ((flags & FOF_NOLOCK) != 0) {
683 if ((flags & FOF_NOUPDATE) == 0)
685 if ((flags & FOF_NEXTOFF) != 0)
691 mtxp = mtx_pool_find(mtxpool_sleep, fp);
693 if ((flags & FOF_NOUPDATE) == 0)
695 if ((flags & FOF_NEXTOFF) != 0)
697 if ((flags & FOF_NOLOCK) == 0) {
698 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
699 ("Lost FOFFSET_LOCKED"));
700 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
701 wakeup(&fp->f_vnread_flags);
702 fp->f_vnread_flags = 0;
708 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
711 if ((flags & FOF_OFFSET) == 0)
712 uio->uio_offset = foffset_lock(fp, flags);
716 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
719 if ((flags & FOF_OFFSET) == 0)
720 foffset_unlock(fp, uio->uio_offset, flags);
724 get_advice(struct file *fp, struct uio *uio)
729 ret = POSIX_FADV_NORMAL;
730 if (fp->f_advice == NULL)
733 mtxp = mtx_pool_find(mtxpool_sleep, fp);
735 if (uio->uio_offset >= fp->f_advice->fa_start &&
736 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
737 ret = fp->f_advice->fa_advice;
743 * File table vnode read routine.
746 vn_read(fp, uio, active_cred, flags, td)
749 struct ucred *active_cred;
757 off_t offset, start, end;
759 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
761 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
764 if (fp->f_flag & FNONBLOCK)
766 if (fp->f_flag & O_DIRECT)
768 advice = get_advice(fp, uio);
769 vn_lock(vp, LK_SHARED | LK_RETRY);
772 case POSIX_FADV_NORMAL:
773 case POSIX_FADV_SEQUENTIAL:
774 case POSIX_FADV_NOREUSE:
775 ioflag |= sequential_heuristic(uio, fp);
777 case POSIX_FADV_RANDOM:
778 /* Disable read-ahead for random I/O. */
781 offset = uio->uio_offset;
784 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
787 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
788 fp->f_nextoff = uio->uio_offset;
790 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
791 offset != uio->uio_offset) {
793 * Use POSIX_FADV_DONTNEED to flush clean pages and
794 * buffers for the backing file after a
795 * POSIX_FADV_NOREUSE read(2). To optimize the common
796 * case of using POSIX_FADV_NOREUSE with sequential
797 * access, track the previous implicit DONTNEED
798 * request and grow this request to include the
799 * current read(2) in addition to the previous
800 * DONTNEED. With purely sequential access this will
801 * cause the DONTNEED requests to continously grow to
802 * cover all of the previously read regions of the
803 * file. This allows filesystem blocks that are
804 * accessed by multiple calls to read(2) to be flushed
805 * once the last read(2) finishes.
808 end = uio->uio_offset - 1;
809 mtxp = mtx_pool_find(mtxpool_sleep, fp);
811 if (fp->f_advice != NULL &&
812 fp->f_advice->fa_advice == POSIX_FADV_NOREUSE) {
813 if (start != 0 && fp->f_advice->fa_prevend + 1 == start)
814 start = fp->f_advice->fa_prevstart;
815 else if (fp->f_advice->fa_prevstart != 0 &&
816 fp->f_advice->fa_prevstart == end + 1)
817 end = fp->f_advice->fa_prevend;
818 fp->f_advice->fa_prevstart = start;
819 fp->f_advice->fa_prevend = end;
822 error = VOP_ADVISE(vp, start, end, POSIX_FADV_DONTNEED);
828 * File table vnode write routine.
831 vn_write(fp, uio, active_cred, flags, td)
834 struct ucred *active_cred;
841 int error, ioflag, lock_flags;
843 off_t offset, start, end;
845 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
847 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
849 if (vp->v_type == VREG)
852 if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
854 if (fp->f_flag & FNONBLOCK)
856 if (fp->f_flag & O_DIRECT)
858 if ((fp->f_flag & O_FSYNC) ||
859 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
862 if (vp->v_type != VCHR &&
863 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
866 advice = get_advice(fp, uio);
868 if (MNT_SHARED_WRITES(mp) ||
869 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
870 lock_flags = LK_SHARED;
872 lock_flags = LK_EXCLUSIVE;
875 vn_lock(vp, lock_flags | LK_RETRY);
877 case POSIX_FADV_NORMAL:
878 case POSIX_FADV_SEQUENTIAL:
879 case POSIX_FADV_NOREUSE:
880 ioflag |= sequential_heuristic(uio, fp);
882 case POSIX_FADV_RANDOM:
883 /* XXX: Is this correct? */
886 offset = uio->uio_offset;
889 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
892 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
893 fp->f_nextoff = uio->uio_offset;
895 if (vp->v_type != VCHR)
896 vn_finished_write(mp);
897 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
898 offset != uio->uio_offset) {
900 * Use POSIX_FADV_DONTNEED to flush clean pages and
901 * buffers for the backing file after a
902 * POSIX_FADV_NOREUSE write(2). To optimize the
903 * common case of using POSIX_FADV_NOREUSE with
904 * sequential access, track the previous implicit
905 * DONTNEED request and grow this request to include
906 * the current write(2) in addition to the previous
907 * DONTNEED. With purely sequential access this will
908 * cause the DONTNEED requests to continously grow to
909 * cover all of the previously written regions of the
912 * Note that the blocks just written are almost
913 * certainly still dirty, so this only works when
914 * VOP_ADVISE() calls from subsequent writes push out
915 * the data written by this write(2) once the backing
916 * buffers are clean. However, as compared to forcing
917 * IO_DIRECT, this gives much saner behavior. Write
918 * clustering is still allowed, and clean pages are
919 * merely moved to the cache page queue rather than
920 * outright thrown away. This means a subsequent
921 * read(2) can still avoid hitting the disk if the
922 * pages have not been reclaimed.
924 * This does make POSIX_FADV_NOREUSE largely useless
925 * with non-sequential access. However, sequential
926 * access is the more common use case and the flag is
930 end = uio->uio_offset - 1;
931 mtxp = mtx_pool_find(mtxpool_sleep, fp);
933 if (fp->f_advice != NULL &&
934 fp->f_advice->fa_advice == POSIX_FADV_NOREUSE) {
935 if (start != 0 && fp->f_advice->fa_prevend + 1 == start)
936 start = fp->f_advice->fa_prevstart;
937 else if (fp->f_advice->fa_prevstart != 0 &&
938 fp->f_advice->fa_prevstart == end + 1)
939 end = fp->f_advice->fa_prevend;
940 fp->f_advice->fa_prevstart = start;
941 fp->f_advice->fa_prevend = end;
944 error = VOP_ADVISE(vp, start, end, POSIX_FADV_DONTNEED);
952 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
953 * prevent the following deadlock:
955 * Assume that the thread A reads from the vnode vp1 into userspace
956 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
957 * currently not resident, then system ends up with the call chain
958 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
959 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
960 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
961 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
962 * backed by the pages of vnode vp1, and some page in buf2 is not
963 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
965 * To prevent the lock order reversal and deadlock, vn_io_fault() does
966 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
967 * Instead, it first tries to do the whole range i/o with pagefaults
968 * disabled. If all pages in the i/o buffer are resident and mapped,
969 * VOP will succeed (ignoring the genuine filesystem errors).
970 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
971 * i/o in chunks, with all pages in the chunk prefaulted and held
972 * using vm_fault_quick_hold_pages().
974 * Filesystems using this deadlock avoidance scheme should use the
975 * array of the held pages from uio, saved in the curthread->td_ma,
976 * instead of doing uiomove(). A helper function
977 * vn_io_fault_uiomove() converts uiomove request into
978 * uiomove_fromphys() over td_ma array.
980 * Since vnode locks do not cover the whole i/o anymore, rangelocks
981 * make the current i/o request atomic with respect to other i/os and
986 * Decode vn_io_fault_args and perform the corresponding i/o.
989 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
993 switch (args->kind) {
994 case VN_IO_FAULT_FOP:
995 return ((args->args.fop_args.doio)(args->args.fop_args.fp,
996 uio, args->cred, args->flags, td));
997 case VN_IO_FAULT_VOP:
998 if (uio->uio_rw == UIO_READ) {
999 return (VOP_READ(args->args.vop_args.vp, uio,
1000 args->flags, args->cred));
1001 } else if (uio->uio_rw == UIO_WRITE) {
1002 return (VOP_WRITE(args->args.vop_args.vp, uio,
1003 args->flags, args->cred));
1007 panic("vn_io_fault_doio: unknown kind of io %d %d", args->kind,
1012 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1013 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1014 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1015 * into args and call vn_io_fault1() to handle faults during the user
1016 * mode buffer accesses.
1019 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1022 vm_page_t ma[io_hold_cnt + 2];
1023 struct uio *uio_clone, short_uio;
1024 struct iovec short_iovec[1];
1025 vm_page_t *prev_td_ma;
1027 vm_offset_t addr, end;
1030 int error, cnt, save, saveheld, prev_td_ma_cnt;
1032 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1035 * The UFS follows IO_UNIT directive and replays back both
1036 * uio_offset and uio_resid if an error is encountered during the
1037 * operation. But, since the iovec may be already advanced,
1038 * uio is still in an inconsistent state.
1040 * Cache a copy of the original uio, which is advanced to the redo
1041 * point using UIO_NOCOPY below.
1043 uio_clone = cloneuio(uio);
1044 resid = uio->uio_resid;
1046 short_uio.uio_segflg = UIO_USERSPACE;
1047 short_uio.uio_rw = uio->uio_rw;
1048 short_uio.uio_td = uio->uio_td;
1050 save = vm_fault_disable_pagefaults();
1051 error = vn_io_fault_doio(args, uio, td);
1052 if (error != EFAULT)
1055 atomic_add_long(&vn_io_faults_cnt, 1);
1056 uio_clone->uio_segflg = UIO_NOCOPY;
1057 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1058 uio_clone->uio_segflg = uio->uio_segflg;
1060 saveheld = curthread_pflags_set(TDP_UIOHELD);
1061 prev_td_ma = td->td_ma;
1062 prev_td_ma_cnt = td->td_ma_cnt;
1064 while (uio_clone->uio_resid != 0) {
1065 len = uio_clone->uio_iov->iov_len;
1067 KASSERT(uio_clone->uio_iovcnt >= 1,
1068 ("iovcnt underflow"));
1069 uio_clone->uio_iov++;
1070 uio_clone->uio_iovcnt--;
1073 if (len > io_hold_cnt * PAGE_SIZE)
1074 len = io_hold_cnt * PAGE_SIZE;
1075 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1076 end = round_page(addr + len);
1081 cnt = atop(end - trunc_page(addr));
1083 * A perfectly misaligned address and length could cause
1084 * both the start and the end of the chunk to use partial
1085 * page. +2 accounts for such a situation.
1087 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1088 addr, len, prot, ma, io_hold_cnt + 2);
1093 short_uio.uio_iov = &short_iovec[0];
1094 short_iovec[0].iov_base = (void *)addr;
1095 short_uio.uio_iovcnt = 1;
1096 short_uio.uio_resid = short_iovec[0].iov_len = len;
1097 short_uio.uio_offset = uio_clone->uio_offset;
1099 td->td_ma_cnt = cnt;
1101 error = vn_io_fault_doio(args, &short_uio, td);
1102 vm_page_unhold_pages(ma, cnt);
1103 adv = len - short_uio.uio_resid;
1105 uio_clone->uio_iov->iov_base =
1106 (char *)uio_clone->uio_iov->iov_base + adv;
1107 uio_clone->uio_iov->iov_len -= adv;
1108 uio_clone->uio_resid -= adv;
1109 uio_clone->uio_offset += adv;
1111 uio->uio_resid -= adv;
1112 uio->uio_offset += adv;
1114 if (error != 0 || adv == 0)
1117 td->td_ma = prev_td_ma;
1118 td->td_ma_cnt = prev_td_ma_cnt;
1119 curthread_pflags_restore(saveheld);
1121 vm_fault_enable_pagefaults(save);
1122 free(uio_clone, M_IOV);
1127 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1128 int flags, struct thread *td)
1133 struct vn_io_fault_args args;
1136 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1138 foffset_lock_uio(fp, uio, flags);
1139 if (do_vn_io_fault(vp, uio)) {
1140 args.kind = VN_IO_FAULT_FOP;
1141 args.args.fop_args.fp = fp;
1142 args.args.fop_args.doio = doio;
1143 args.cred = active_cred;
1144 args.flags = flags | FOF_OFFSET;
1145 if (uio->uio_rw == UIO_READ) {
1146 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1147 uio->uio_offset + uio->uio_resid);
1148 } else if ((fp->f_flag & O_APPEND) != 0 ||
1149 (flags & FOF_OFFSET) == 0) {
1150 /* For appenders, punt and lock the whole range. */
1151 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1153 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1154 uio->uio_offset + uio->uio_resid);
1156 error = vn_io_fault1(vp, uio, &args, td);
1157 vn_rangelock_unlock(vp, rl_cookie);
1159 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1161 foffset_unlock_uio(fp, uio, flags);
1166 * Helper function to perform the requested uiomove operation using
1167 * the held pages for io->uio_iov[0].iov_base buffer instead of
1168 * copyin/copyout. Access to the pages with uiomove_fromphys()
1169 * instead of iov_base prevents page faults that could occur due to
1170 * pmap_collect() invalidating the mapping created by
1171 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1172 * object cleanup revoking the write access from page mappings.
1174 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1175 * instead of plain uiomove().
1178 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1180 struct uio transp_uio;
1181 struct iovec transp_iov[1];
1187 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1188 uio->uio_segflg != UIO_USERSPACE)
1189 return (uiomove(data, xfersize, uio));
1191 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1192 transp_iov[0].iov_base = data;
1193 transp_uio.uio_iov = &transp_iov[0];
1194 transp_uio.uio_iovcnt = 1;
1195 if (xfersize > uio->uio_resid)
1196 xfersize = uio->uio_resid;
1197 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1198 transp_uio.uio_offset = 0;
1199 transp_uio.uio_segflg = UIO_SYSSPACE;
1201 * Since transp_iov points to data, and td_ma page array
1202 * corresponds to original uio->uio_iov, we need to invert the
1203 * direction of the i/o operation as passed to
1204 * uiomove_fromphys().
1206 switch (uio->uio_rw) {
1208 transp_uio.uio_rw = UIO_READ;
1211 transp_uio.uio_rw = UIO_WRITE;
1214 transp_uio.uio_td = uio->uio_td;
1215 error = uiomove_fromphys(td->td_ma,
1216 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1217 xfersize, &transp_uio);
1218 adv = xfersize - transp_uio.uio_resid;
1220 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1221 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1223 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1225 td->td_ma_cnt -= pgadv;
1226 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1227 uio->uio_iov->iov_len -= adv;
1228 uio->uio_resid -= adv;
1229 uio->uio_offset += adv;
1234 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1238 vm_offset_t iov_base;
1242 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1243 uio->uio_segflg != UIO_USERSPACE)
1244 return (uiomove_fromphys(ma, offset, xfersize, uio));
1246 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1247 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1248 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1249 switch (uio->uio_rw) {
1251 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1255 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1259 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1261 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1263 td->td_ma_cnt -= pgadv;
1264 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1265 uio->uio_iov->iov_len -= cnt;
1266 uio->uio_resid -= cnt;
1267 uio->uio_offset += cnt;
1273 * File table truncate routine.
1276 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1288 * Lock the whole range for truncation. Otherwise split i/o
1289 * might happen partly before and partly after the truncation.
1291 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1292 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1295 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1296 if (vp->v_type == VDIR) {
1301 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1305 error = vn_writechk(vp);
1308 vattr.va_size = length;
1309 error = VOP_SETATTR(vp, &vattr, fp->f_cred);
1313 vn_finished_write(mp);
1315 vn_rangelock_unlock(vp, rl_cookie);
1320 * File table vnode stat routine.
1323 vn_statfile(fp, sb, active_cred, td)
1326 struct ucred *active_cred;
1329 struct vnode *vp = fp->f_vnode;
1332 vn_lock(vp, LK_SHARED | LK_RETRY);
1333 error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1340 * Stat a vnode; implementation for the stat syscall
1343 vn_stat(vp, sb, active_cred, file_cred, td)
1345 register struct stat *sb;
1346 struct ucred *active_cred;
1347 struct ucred *file_cred;
1351 register struct vattr *vap;
1356 error = mac_vnode_check_stat(active_cred, file_cred, vp);
1364 * Initialize defaults for new and unusual fields, so that file
1365 * systems which don't support these fields don't need to know
1368 vap->va_birthtime.tv_sec = -1;
1369 vap->va_birthtime.tv_nsec = 0;
1370 vap->va_fsid = VNOVAL;
1371 vap->va_rdev = NODEV;
1373 error = VOP_GETATTR(vp, vap, active_cred);
1378 * Zero the spare stat fields
1380 bzero(sb, sizeof *sb);
1383 * Copy from vattr table
1385 if (vap->va_fsid != VNOVAL)
1386 sb->st_dev = vap->va_fsid;
1388 sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1389 sb->st_ino = vap->va_fileid;
1390 mode = vap->va_mode;
1391 switch (vap->va_type) {
1417 sb->st_nlink = vap->va_nlink;
1418 sb->st_uid = vap->va_uid;
1419 sb->st_gid = vap->va_gid;
1420 sb->st_rdev = vap->va_rdev;
1421 if (vap->va_size > OFF_MAX)
1423 sb->st_size = vap->va_size;
1424 sb->st_atim = vap->va_atime;
1425 sb->st_mtim = vap->va_mtime;
1426 sb->st_ctim = vap->va_ctime;
1427 sb->st_birthtim = vap->va_birthtime;
1430 * According to www.opengroup.org, the meaning of st_blksize is
1431 * "a filesystem-specific preferred I/O block size for this
1432 * object. In some filesystem types, this may vary from file
1434 * Use miminum/default of PAGE_SIZE (e.g. for VCHR).
1437 sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1439 sb->st_flags = vap->va_flags;
1440 if (priv_check(td, PRIV_VFS_GENERATION))
1443 sb->st_gen = vap->va_gen;
1445 sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1450 * File table vnode ioctl routine.
1453 vn_ioctl(fp, com, data, active_cred, td)
1457 struct ucred *active_cred;
1465 switch (vp->v_type) {
1470 vn_lock(vp, LK_SHARED | LK_RETRY);
1471 error = VOP_GETATTR(vp, &vattr, active_cred);
1474 *(int *)data = vattr.va_size - fp->f_offset;
1480 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1489 * File table vnode poll routine.
1492 vn_poll(fp, events, active_cred, td)
1495 struct ucred *active_cred;
1503 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1504 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1509 error = VOP_POLL(vp, events, fp->f_cred, td);
1514 * Acquire the requested lock and then check for validity. LK_RETRY
1515 * permits vn_lock to return doomed vnodes.
1518 _vn_lock(struct vnode *vp, int flags, char *file, int line)
1522 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1523 ("vn_lock called with no locktype."));
1525 #ifdef DEBUG_VFS_LOCKS
1526 KASSERT(vp->v_holdcnt != 0,
1527 ("vn_lock %p: zero hold count", vp));
1529 error = VOP_LOCK1(vp, flags, file, line);
1530 flags &= ~LK_INTERLOCK; /* Interlock is always dropped. */
1531 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1532 ("LK_RETRY set with incompatible flags (0x%x) or an error occured (%d)",
1535 * Callers specify LK_RETRY if they wish to get dead vnodes.
1536 * If RETRY is not set, we return ENOENT instead.
1538 if (error == 0 && vp->v_iflag & VI_DOOMED &&
1539 (flags & LK_RETRY) == 0) {
1544 } while (flags & LK_RETRY && error != 0);
1549 * File table vnode close routine.
1552 vn_closefile(fp, td)
1561 fp->f_ops = &badfileops;
1563 if (fp->f_type == DTYPE_VNODE && fp->f_flag & FHASLOCK)
1566 error = vn_close(vp, fp->f_flag, fp->f_cred, td);
1568 if (fp->f_type == DTYPE_VNODE && fp->f_flag & FHASLOCK) {
1569 lf.l_whence = SEEK_SET;
1572 lf.l_type = F_UNLCK;
1573 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1580 * Preparing to start a filesystem write operation. If the operation is
1581 * permitted, then we bump the count of operations in progress and
1582 * proceed. If a suspend request is in progress, we wait until the
1583 * suspension is over, and then proceed.
1586 vn_start_write_locked(struct mount *mp, int flags)
1590 mtx_assert(MNT_MTX(mp), MA_OWNED);
1594 * Check on status of suspension.
1596 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1597 mp->mnt_susp_owner != curthread) {
1598 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1599 (flags & PCATCH) : 0) | (PUSER - 1);
1600 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1601 if (flags & V_NOWAIT) {
1602 error = EWOULDBLOCK;
1605 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1611 if (flags & V_XSLEEP)
1613 mp->mnt_writeopcount++;
1615 if (error != 0 || (flags & V_XSLEEP) != 0)
1622 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1627 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1628 ("V_MNTREF requires mp"));
1632 * If a vnode is provided, get and return the mount point that
1633 * to which it will write.
1636 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1638 if (error != EOPNOTSUPP)
1643 if ((mp = *mpp) == NULL)
1647 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1649 * As long as a vnode is not provided we need to acquire a
1650 * refcount for the provided mountpoint too, in order to
1651 * emulate a vfs_ref().
1654 if (vp == NULL && (flags & V_MNTREF) == 0)
1657 return (vn_start_write_locked(mp, flags));
1661 * Secondary suspension. Used by operations such as vop_inactive
1662 * routines that are needed by the higher level functions. These
1663 * are allowed to proceed until all the higher level functions have
1664 * completed (indicated by mnt_writeopcount dropping to zero). At that
1665 * time, these operations are halted until the suspension is over.
1668 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1673 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1674 ("V_MNTREF requires mp"));
1678 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1680 if (error != EOPNOTSUPP)
1686 * If we are not suspended or have not yet reached suspended
1687 * mode, then let the operation proceed.
1689 if ((mp = *mpp) == NULL)
1693 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1695 * As long as a vnode is not provided we need to acquire a
1696 * refcount for the provided mountpoint too, in order to
1697 * emulate a vfs_ref().
1700 if (vp == NULL && (flags & V_MNTREF) == 0)
1702 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1703 mp->mnt_secondary_writes++;
1704 mp->mnt_secondary_accwrites++;
1708 if (flags & V_NOWAIT) {
1711 return (EWOULDBLOCK);
1714 * Wait for the suspension to finish.
1716 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1717 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1726 * Filesystem write operation has completed. If we are suspending and this
1727 * operation is the last one, notify the suspender that the suspension is
1731 vn_finished_write(mp)
1738 mp->mnt_writeopcount--;
1739 if (mp->mnt_writeopcount < 0)
1740 panic("vn_finished_write: neg cnt");
1741 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1742 mp->mnt_writeopcount <= 0)
1743 wakeup(&mp->mnt_writeopcount);
1749 * Filesystem secondary write operation has completed. If we are
1750 * suspending and this operation is the last one, notify the suspender
1751 * that the suspension is now in effect.
1754 vn_finished_secondary_write(mp)
1761 mp->mnt_secondary_writes--;
1762 if (mp->mnt_secondary_writes < 0)
1763 panic("vn_finished_secondary_write: neg cnt");
1764 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1765 mp->mnt_secondary_writes <= 0)
1766 wakeup(&mp->mnt_secondary_writes);
1773 * Request a filesystem to suspend write operations.
1776 vfs_write_suspend(struct mount *mp, int flags)
1781 if (mp->mnt_susp_owner == curthread) {
1785 while (mp->mnt_kern_flag & MNTK_SUSPEND)
1786 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1789 * Unmount holds a write reference on the mount point. If we
1790 * own busy reference and drain for writers, we deadlock with
1791 * the reference draining in the unmount path. Callers of
1792 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1793 * vfs_busy() reference is owned and caller is not in the
1796 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1797 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1802 mp->mnt_kern_flag |= MNTK_SUSPEND;
1803 mp->mnt_susp_owner = curthread;
1804 if (mp->mnt_writeopcount > 0)
1805 (void) msleep(&mp->mnt_writeopcount,
1806 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1809 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0)
1810 vfs_write_resume(mp, 0);
1815 * Request a filesystem to resume write operations.
1818 vfs_write_resume(struct mount *mp, int flags)
1822 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1823 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
1824 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
1826 mp->mnt_susp_owner = NULL;
1827 wakeup(&mp->mnt_writeopcount);
1828 wakeup(&mp->mnt_flag);
1829 curthread->td_pflags &= ~TDP_IGNSUSP;
1830 if ((flags & VR_START_WRITE) != 0) {
1832 mp->mnt_writeopcount++;
1835 if ((flags & VR_NO_SUSPCLR) == 0)
1837 } else if ((flags & VR_START_WRITE) != 0) {
1839 vn_start_write_locked(mp, 0);
1846 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
1850 vfs_write_suspend_umnt(struct mount *mp)
1854 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
1855 ("vfs_write_suspend_umnt: recursed"));
1857 /* dounmount() already called vn_start_write(). */
1859 vn_finished_write(mp);
1860 error = vfs_write_suspend(mp, 0);
1862 vn_start_write(NULL, &mp, V_WAIT);
1866 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
1869 vn_start_write(NULL, &mp, V_WAIT);
1871 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
1872 wakeup(&mp->mnt_flag);
1874 curthread->td_pflags |= TDP_IGNSUSP;
1879 * Implement kqueues for files by translating it to vnode operation.
1882 vn_kqfilter(struct file *fp, struct knote *kn)
1885 return (VOP_KQFILTER(fp->f_vnode, kn));
1889 * Simplified in-kernel wrapper calls for extended attribute access.
1890 * Both calls pass in a NULL credential, authorizing as "kernel" access.
1891 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
1894 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
1895 const char *attrname, int *buflen, char *buf, struct thread *td)
1901 iov.iov_len = *buflen;
1904 auio.uio_iov = &iov;
1905 auio.uio_iovcnt = 1;
1906 auio.uio_rw = UIO_READ;
1907 auio.uio_segflg = UIO_SYSSPACE;
1909 auio.uio_offset = 0;
1910 auio.uio_resid = *buflen;
1912 if ((ioflg & IO_NODELOCKED) == 0)
1913 vn_lock(vp, LK_SHARED | LK_RETRY);
1915 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
1917 /* authorize attribute retrieval as kernel */
1918 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
1921 if ((ioflg & IO_NODELOCKED) == 0)
1925 *buflen = *buflen - auio.uio_resid;
1932 * XXX failure mode if partially written?
1935 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
1936 const char *attrname, int buflen, char *buf, struct thread *td)
1943 iov.iov_len = buflen;
1946 auio.uio_iov = &iov;
1947 auio.uio_iovcnt = 1;
1948 auio.uio_rw = UIO_WRITE;
1949 auio.uio_segflg = UIO_SYSSPACE;
1951 auio.uio_offset = 0;
1952 auio.uio_resid = buflen;
1954 if ((ioflg & IO_NODELOCKED) == 0) {
1955 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
1957 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1960 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
1962 /* authorize attribute setting as kernel */
1963 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
1965 if ((ioflg & IO_NODELOCKED) == 0) {
1966 vn_finished_write(mp);
1974 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
1975 const char *attrname, struct thread *td)
1980 if ((ioflg & IO_NODELOCKED) == 0) {
1981 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
1983 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1986 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
1988 /* authorize attribute removal as kernel */
1989 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
1990 if (error == EOPNOTSUPP)
1991 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
1994 if ((ioflg & IO_NODELOCKED) == 0) {
1995 vn_finished_write(mp);
2003 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2007 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2011 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2014 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2019 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2020 int lkflags, struct vnode **rvp)
2025 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2027 ltype = VOP_ISLOCKED(vp);
2028 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2029 ("vn_vget_ino: vp not locked"));
2030 error = vfs_busy(mp, MBF_NOWAIT);
2034 error = vfs_busy(mp, 0);
2035 vn_lock(vp, ltype | LK_RETRY);
2039 if (vp->v_iflag & VI_DOOMED) {
2045 error = alloc(mp, alloc_arg, lkflags, rvp);
2048 vn_lock(vp, ltype | LK_RETRY);
2049 if (vp->v_iflag & VI_DOOMED) {
2062 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2063 const struct thread *td)
2066 if (vp->v_type != VREG || td == NULL)
2068 PROC_LOCK(td->td_proc);
2069 if ((uoff_t)uio->uio_offset + uio->uio_resid >
2070 lim_cur(td->td_proc, RLIMIT_FSIZE)) {
2071 kern_psignal(td->td_proc, SIGXFSZ);
2072 PROC_UNLOCK(td->td_proc);
2075 PROC_UNLOCK(td->td_proc);
2080 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2087 vn_lock(vp, LK_SHARED | LK_RETRY);
2088 AUDIT_ARG_VNODE1(vp);
2091 return (setfmode(td, active_cred, vp, mode));
2095 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2102 vn_lock(vp, LK_SHARED | LK_RETRY);
2103 AUDIT_ARG_VNODE1(vp);
2106 return (setfown(td, active_cred, vp, uid, gid));
2110 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2114 if ((object = vp->v_object) == NULL)
2116 VM_OBJECT_WLOCK(object);
2117 vm_object_page_remove(object, start, end, 0);
2118 VM_OBJECT_WUNLOCK(object);
2122 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2130 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2131 ("Wrong command %lu", cmd));
2133 if (vn_lock(vp, LK_SHARED) != 0)
2135 if (vp->v_type != VREG) {
2139 error = VOP_GETATTR(vp, &va, cred);
2143 if (noff >= va.va_size) {
2147 bsize = vp->v_mount->mnt_stat.f_iosize;
2148 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize) {
2149 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2150 if (error == EOPNOTSUPP) {
2154 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2155 (bnp != -1 && cmd == FIOSEEKDATA)) {
2162 if (noff > va.va_size)
2164 /* noff == va.va_size. There is an implicit hole at the end of file. */
2165 if (cmd == FIOSEEKDATA)
2175 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2180 off_t foffset, size;
2183 cred = td->td_ucred;
2185 foffset = foffset_lock(fp, 0);
2186 noneg = (vp->v_type != VCHR);
2192 (offset > 0 && foffset > OFF_MAX - offset))) {
2199 vn_lock(vp, LK_SHARED | LK_RETRY);
2200 error = VOP_GETATTR(vp, &vattr, cred);
2206 * If the file references a disk device, then fetch
2207 * the media size and use that to determine the ending
2210 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2211 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2212 vattr.va_size = size;
2214 (vattr.va_size > OFF_MAX ||
2215 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2219 offset += vattr.va_size;
2224 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2227 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2232 if (error == 0 && noneg && offset < 0)
2236 VFS_KNOTE_UNLOCKED(vp, 0);
2237 *(off_t *)(td->td_retval) = offset;
2239 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2244 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2250 * Grant permission if the caller is the owner of the file, or
2251 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2252 * on the file. If the time pointer is null, then write
2253 * permission on the file is also sufficient.
2255 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2256 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2257 * will be allowed to set the times [..] to the current
2260 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2261 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2262 error = VOP_ACCESS(vp, VWRITE, cred, td);