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 int vn_io_fault_prefault = 0;
114 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RW,
115 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
116 static u_long vn_io_faults_cnt;
117 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
118 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
121 * Returns true if vn_io_fault mode of handling the i/o request should
125 do_vn_io_fault(struct vnode *vp, struct uio *uio)
129 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
130 (mp = vp->v_mount) != NULL &&
131 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
135 * Structure used to pass arguments to vn_io_fault1(), to do either
136 * file- or vnode-based I/O calls.
138 struct vn_io_fault_args {
146 struct fop_args_tag {
150 struct vop_args_tag {
156 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
157 struct vn_io_fault_args *args, struct thread *td);
160 vn_open(ndp, flagp, cmode, fp)
161 struct nameidata *ndp;
165 struct thread *td = ndp->ni_cnd.cn_thread;
167 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
171 * Common code for vnode open operations via a name lookup.
172 * Lookup the vnode and invoke VOP_CREATE if needed.
173 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
175 * Note that this does NOT free nameidata for the successful case,
176 * due to the NDINIT being done elsewhere.
179 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
180 struct ucred *cred, struct file *fp)
184 struct thread *td = ndp->ni_cnd.cn_thread;
186 struct vattr *vap = &vat;
191 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
192 O_EXCL | O_DIRECTORY))
194 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
195 ndp->ni_cnd.cn_nameiop = CREATE;
197 * Set NOCACHE to avoid flushing the cache when
198 * rolling in many files at once.
200 ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE;
201 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
202 ndp->ni_cnd.cn_flags |= FOLLOW;
203 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
204 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
205 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
206 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
208 if ((error = namei(ndp)) != 0)
210 if (ndp->ni_vp == NULL) {
213 vap->va_mode = cmode;
215 vap->va_vaflags |= VA_EXCLUSIVE;
216 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
217 NDFREE(ndp, NDF_ONLY_PNBUF);
219 if ((error = vn_start_write(NULL, &mp,
220 V_XSLEEP | PCATCH)) != 0)
224 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
225 ndp->ni_cnd.cn_flags |= MAKEENTRY;
227 error = mac_vnode_check_create(cred, ndp->ni_dvp,
231 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
234 vn_finished_write(mp);
236 NDFREE(ndp, NDF_ONLY_PNBUF);
242 if (ndp->ni_dvp == ndp->ni_vp)
248 if (fmode & O_EXCL) {
255 ndp->ni_cnd.cn_nameiop = LOOKUP;
256 ndp->ni_cnd.cn_flags = ISOPEN |
257 ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
258 if (!(fmode & FWRITE))
259 ndp->ni_cnd.cn_flags |= LOCKSHARED;
260 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
261 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
262 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
263 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
264 if ((error = namei(ndp)) != 0)
268 error = vn_open_vnode(vp, fmode, cred, td, fp);
274 NDFREE(ndp, NDF_ONLY_PNBUF);
282 * Common code for vnode open operations once a vnode is located.
283 * Check permissions, and call the VOP_OPEN routine.
286 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
287 struct thread *td, struct file *fp)
292 int error, have_flock, lock_flags, type;
294 if (vp->v_type == VLNK)
296 if (vp->v_type == VSOCK)
298 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
301 if (fmode & (FWRITE | O_TRUNC)) {
302 if (vp->v_type == VDIR)
310 if ((fmode & O_APPEND) && (fmode & FWRITE))
313 error = mac_vnode_check_open(cred, vp, accmode);
317 if ((fmode & O_CREAT) == 0) {
318 if (accmode & VWRITE) {
319 error = vn_writechk(vp);
324 error = VOP_ACCESS(vp, accmode, cred, td);
329 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
330 vn_lock(vp, LK_UPGRADE | LK_RETRY);
331 if ((error = VOP_OPEN(vp, fmode, cred, td, fp)) != 0)
334 if (fmode & (O_EXLOCK | O_SHLOCK)) {
335 KASSERT(fp != NULL, ("open with flock requires fp"));
336 lock_flags = VOP_ISLOCKED(vp);
338 lf.l_whence = SEEK_SET;
341 if (fmode & O_EXLOCK)
346 if ((fmode & FNONBLOCK) == 0)
348 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
349 have_flock = (error == 0);
350 vn_lock(vp, lock_flags | LK_RETRY);
351 if (error == 0 && vp->v_iflag & VI_DOOMED)
354 * Another thread might have used this vnode as an
355 * executable while the vnode lock was dropped.
356 * Ensure the vnode is still able to be opened for
357 * writing after the lock has been obtained.
359 if (error == 0 && accmode & VWRITE)
360 error = vn_writechk(vp);
364 lf.l_whence = SEEK_SET;
368 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf,
371 vn_start_write(vp, &mp, V_WAIT);
372 vn_lock(vp, lock_flags | LK_RETRY);
373 (void)VOP_CLOSE(vp, fmode, cred, td);
374 vn_finished_write(mp);
375 /* Prevent second close from fdrop()->vn_close(). */
377 fp->f_ops= &badfileops;
380 fp->f_flag |= FHASLOCK;
382 if (fmode & FWRITE) {
383 VOP_ADD_WRITECOUNT(vp, 1);
384 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
385 __func__, vp, vp->v_writecount);
387 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
392 * Check for write permissions on the specified vnode.
393 * Prototype text segments cannot be written.
397 register struct vnode *vp;
400 ASSERT_VOP_LOCKED(vp, "vn_writechk");
402 * If there's shared text associated with
403 * the vnode, try to free it up once. If
404 * we fail, we can't allow writing.
416 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
417 struct thread *td, bool keep_ref)
420 int error, lock_flags;
422 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
423 MNT_EXTENDED_SHARED(vp->v_mount))
424 lock_flags = LK_SHARED;
426 lock_flags = LK_EXCLUSIVE;
428 vn_start_write(vp, &mp, V_WAIT);
429 vn_lock(vp, lock_flags | LK_RETRY);
430 if (flags & FWRITE) {
431 VNASSERT(vp->v_writecount > 0, vp,
432 ("vn_close: negative writecount"));
433 VOP_ADD_WRITECOUNT(vp, -1);
434 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
435 __func__, vp, vp->v_writecount);
437 error = VOP_CLOSE(vp, flags, file_cred, td);
442 vn_finished_write(mp);
447 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
451 return (vn_close1(vp, flags, file_cred, td, false));
455 * Heuristic to detect sequential operation.
458 sequential_heuristic(struct uio *uio, struct file *fp)
461 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
462 if (fp->f_flag & FRDAHEAD)
463 return (fp->f_seqcount << IO_SEQSHIFT);
466 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
467 * that the first I/O is normally considered to be slightly
468 * sequential. Seeking to offset 0 doesn't change sequentiality
469 * unless previous seeks have reduced f_seqcount to 0, in which
470 * case offset 0 is not special.
472 if ((uio->uio_offset == 0 && fp->f_seqcount > 0) ||
473 uio->uio_offset == fp->f_nextoff) {
475 * f_seqcount is in units of fixed-size blocks so that it
476 * depends mainly on the amount of sequential I/O and not
477 * much on the number of sequential I/O's. The fixed size
478 * of 16384 is hard-coded here since it is (not quite) just
479 * a magic size that works well here. This size is more
480 * closely related to the best I/O size for real disks than
481 * to any block size used by software.
483 fp->f_seqcount += howmany(uio->uio_resid, 16384);
484 if (fp->f_seqcount > IO_SEQMAX)
485 fp->f_seqcount = IO_SEQMAX;
486 return (fp->f_seqcount << IO_SEQSHIFT);
489 /* Not sequential. Quickly draw-down sequentiality. */
490 if (fp->f_seqcount > 1)
498 * Package up an I/O request on a vnode into a uio and do it.
501 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
502 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
503 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
510 struct vn_io_fault_args args;
511 int error, lock_flags;
513 if (offset < 0 && vp->v_type != VCHR)
515 auio.uio_iov = &aiov;
517 aiov.iov_base = base;
519 auio.uio_resid = len;
520 auio.uio_offset = offset;
521 auio.uio_segflg = segflg;
526 if ((ioflg & IO_NODELOCKED) == 0) {
527 if ((ioflg & IO_RANGELOCKED) == 0) {
528 if (rw == UIO_READ) {
529 rl_cookie = vn_rangelock_rlock(vp, offset,
532 rl_cookie = vn_rangelock_wlock(vp, offset,
538 if (rw == UIO_WRITE) {
539 if (vp->v_type != VCHR &&
540 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
543 if (MNT_SHARED_WRITES(mp) ||
544 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
545 lock_flags = LK_SHARED;
547 lock_flags = LK_EXCLUSIVE;
549 lock_flags = LK_SHARED;
550 vn_lock(vp, lock_flags | LK_RETRY);
554 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
556 if ((ioflg & IO_NOMACCHECK) == 0) {
558 error = mac_vnode_check_read(active_cred, file_cred,
561 error = mac_vnode_check_write(active_cred, file_cred,
566 if (file_cred != NULL)
570 if (do_vn_io_fault(vp, &auio)) {
571 args.kind = VN_IO_FAULT_VOP;
574 args.args.vop_args.vp = vp;
575 error = vn_io_fault1(vp, &auio, &args, td);
576 } else if (rw == UIO_READ) {
577 error = VOP_READ(vp, &auio, ioflg, cred);
578 } else /* if (rw == UIO_WRITE) */ {
579 error = VOP_WRITE(vp, &auio, ioflg, cred);
583 *aresid = auio.uio_resid;
585 if (auio.uio_resid && error == 0)
587 if ((ioflg & IO_NODELOCKED) == 0) {
590 vn_finished_write(mp);
593 if (rl_cookie != NULL)
594 vn_rangelock_unlock(vp, rl_cookie);
599 * Package up an I/O request on a vnode into a uio and do it. The I/O
600 * request is split up into smaller chunks and we try to avoid saturating
601 * the buffer cache while potentially holding a vnode locked, so we
602 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
603 * to give other processes a chance to lock the vnode (either other processes
604 * core'ing the same binary, or unrelated processes scanning the directory).
607 vn_rdwr_inchunks(rw, vp, base, len, offset, segflg, ioflg, active_cred,
608 file_cred, aresid, td)
616 struct ucred *active_cred;
617 struct ucred *file_cred;
628 * Force `offset' to a multiple of MAXBSIZE except possibly
629 * for the first chunk, so that filesystems only need to
630 * write full blocks except possibly for the first and last
633 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
637 if (rw != UIO_READ && vp->v_type == VREG)
640 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
641 ioflg, active_cred, file_cred, &iaresid, td);
642 len -= chunk; /* aresid calc already includes length */
646 base = (char *)base + chunk;
647 kern_yield(PRI_USER);
650 *aresid = len + iaresid;
655 foffset_lock(struct file *fp, int flags)
660 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
662 #if OFF_MAX <= LONG_MAX
664 * Caller only wants the current f_offset value. Assume that
665 * the long and shorter integer types reads are atomic.
667 if ((flags & FOF_NOLOCK) != 0)
668 return (fp->f_offset);
672 * According to McKusick the vn lock was protecting f_offset here.
673 * It is now protected by the FOFFSET_LOCKED flag.
675 mtxp = mtx_pool_find(mtxpool_sleep, fp);
677 if ((flags & FOF_NOLOCK) == 0) {
678 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
679 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
680 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
683 fp->f_vnread_flags |= FOFFSET_LOCKED;
691 foffset_unlock(struct file *fp, off_t val, int flags)
695 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
697 #if OFF_MAX <= LONG_MAX
698 if ((flags & FOF_NOLOCK) != 0) {
699 if ((flags & FOF_NOUPDATE) == 0)
701 if ((flags & FOF_NEXTOFF) != 0)
707 mtxp = mtx_pool_find(mtxpool_sleep, fp);
709 if ((flags & FOF_NOUPDATE) == 0)
711 if ((flags & FOF_NEXTOFF) != 0)
713 if ((flags & FOF_NOLOCK) == 0) {
714 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
715 ("Lost FOFFSET_LOCKED"));
716 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
717 wakeup(&fp->f_vnread_flags);
718 fp->f_vnread_flags = 0;
724 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
727 if ((flags & FOF_OFFSET) == 0)
728 uio->uio_offset = foffset_lock(fp, flags);
732 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
735 if ((flags & FOF_OFFSET) == 0)
736 foffset_unlock(fp, uio->uio_offset, flags);
740 get_advice(struct file *fp, struct uio *uio)
745 ret = POSIX_FADV_NORMAL;
746 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
749 mtxp = mtx_pool_find(mtxpool_sleep, fp);
751 if (fp->f_advice != NULL &&
752 uio->uio_offset >= fp->f_advice->fa_start &&
753 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
754 ret = fp->f_advice->fa_advice;
760 * File table vnode read routine.
763 vn_read(fp, uio, active_cred, flags, td)
766 struct ucred *active_cred;
774 off_t offset, start, end;
776 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
778 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
781 if (fp->f_flag & FNONBLOCK)
783 if (fp->f_flag & O_DIRECT)
785 advice = get_advice(fp, uio);
786 vn_lock(vp, LK_SHARED | LK_RETRY);
789 case POSIX_FADV_NORMAL:
790 case POSIX_FADV_SEQUENTIAL:
791 case POSIX_FADV_NOREUSE:
792 ioflag |= sequential_heuristic(uio, fp);
794 case POSIX_FADV_RANDOM:
795 /* Disable read-ahead for random I/O. */
798 offset = uio->uio_offset;
801 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
804 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
805 fp->f_nextoff = uio->uio_offset;
807 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
808 offset != uio->uio_offset) {
810 * Use POSIX_FADV_DONTNEED to flush clean pages and
811 * buffers for the backing file after a
812 * POSIX_FADV_NOREUSE read(2). To optimize the common
813 * case of using POSIX_FADV_NOREUSE with sequential
814 * access, track the previous implicit DONTNEED
815 * request and grow this request to include the
816 * current read(2) in addition to the previous
817 * DONTNEED. With purely sequential access this will
818 * cause the DONTNEED requests to continously grow to
819 * cover all of the previously read regions of the
820 * file. This allows filesystem blocks that are
821 * accessed by multiple calls to read(2) to be flushed
822 * once the last read(2) finishes.
825 end = uio->uio_offset - 1;
826 mtxp = mtx_pool_find(mtxpool_sleep, fp);
828 if (fp->f_advice != NULL &&
829 fp->f_advice->fa_advice == POSIX_FADV_NOREUSE) {
830 if (start != 0 && fp->f_advice->fa_prevend + 1 == start)
831 start = fp->f_advice->fa_prevstart;
832 else if (fp->f_advice->fa_prevstart != 0 &&
833 fp->f_advice->fa_prevstart == end + 1)
834 end = fp->f_advice->fa_prevend;
835 fp->f_advice->fa_prevstart = start;
836 fp->f_advice->fa_prevend = end;
839 error = VOP_ADVISE(vp, start, end, POSIX_FADV_DONTNEED);
845 * File table vnode write routine.
848 vn_write(fp, uio, active_cred, flags, td)
851 struct ucred *active_cred;
858 int error, ioflag, lock_flags;
860 off_t offset, start, end;
862 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
864 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
866 if (vp->v_type == VREG)
869 if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
871 if (fp->f_flag & FNONBLOCK)
873 if (fp->f_flag & O_DIRECT)
875 if ((fp->f_flag & O_FSYNC) ||
876 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
879 if (vp->v_type != VCHR &&
880 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
883 advice = get_advice(fp, uio);
885 if (MNT_SHARED_WRITES(mp) ||
886 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
887 lock_flags = LK_SHARED;
889 lock_flags = LK_EXCLUSIVE;
892 vn_lock(vp, lock_flags | LK_RETRY);
894 case POSIX_FADV_NORMAL:
895 case POSIX_FADV_SEQUENTIAL:
896 case POSIX_FADV_NOREUSE:
897 ioflag |= sequential_heuristic(uio, fp);
899 case POSIX_FADV_RANDOM:
900 /* XXX: Is this correct? */
903 offset = uio->uio_offset;
906 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
909 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
910 fp->f_nextoff = uio->uio_offset;
912 if (vp->v_type != VCHR)
913 vn_finished_write(mp);
914 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
915 offset != uio->uio_offset) {
917 * Use POSIX_FADV_DONTNEED to flush clean pages and
918 * buffers for the backing file after a
919 * POSIX_FADV_NOREUSE write(2). To optimize the
920 * common case of using POSIX_FADV_NOREUSE with
921 * sequential access, track the previous implicit
922 * DONTNEED request and grow this request to include
923 * the current write(2) in addition to the previous
924 * DONTNEED. With purely sequential access this will
925 * cause the DONTNEED requests to continously grow to
926 * cover all of the previously written regions of the
929 * Note that the blocks just written are almost
930 * certainly still dirty, so this only works when
931 * VOP_ADVISE() calls from subsequent writes push out
932 * the data written by this write(2) once the backing
933 * buffers are clean. However, as compared to forcing
934 * IO_DIRECT, this gives much saner behavior. Write
935 * clustering is still allowed, and clean pages are
936 * merely moved to the cache page queue rather than
937 * outright thrown away. This means a subsequent
938 * read(2) can still avoid hitting the disk if the
939 * pages have not been reclaimed.
941 * This does make POSIX_FADV_NOREUSE largely useless
942 * with non-sequential access. However, sequential
943 * access is the more common use case and the flag is
947 end = uio->uio_offset - 1;
948 mtxp = mtx_pool_find(mtxpool_sleep, fp);
950 if (fp->f_advice != NULL &&
951 fp->f_advice->fa_advice == POSIX_FADV_NOREUSE) {
952 if (start != 0 && fp->f_advice->fa_prevend + 1 == start)
953 start = fp->f_advice->fa_prevstart;
954 else if (fp->f_advice->fa_prevstart != 0 &&
955 fp->f_advice->fa_prevstart == end + 1)
956 end = fp->f_advice->fa_prevend;
957 fp->f_advice->fa_prevstart = start;
958 fp->f_advice->fa_prevend = end;
961 error = VOP_ADVISE(vp, start, end, POSIX_FADV_DONTNEED);
969 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
970 * prevent the following deadlock:
972 * Assume that the thread A reads from the vnode vp1 into userspace
973 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
974 * currently not resident, then system ends up with the call chain
975 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
976 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
977 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
978 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
979 * backed by the pages of vnode vp1, and some page in buf2 is not
980 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
982 * To prevent the lock order reversal and deadlock, vn_io_fault() does
983 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
984 * Instead, it first tries to do the whole range i/o with pagefaults
985 * disabled. If all pages in the i/o buffer are resident and mapped,
986 * VOP will succeed (ignoring the genuine filesystem errors).
987 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
988 * i/o in chunks, with all pages in the chunk prefaulted and held
989 * using vm_fault_quick_hold_pages().
991 * Filesystems using this deadlock avoidance scheme should use the
992 * array of the held pages from uio, saved in the curthread->td_ma,
993 * instead of doing uiomove(). A helper function
994 * vn_io_fault_uiomove() converts uiomove request into
995 * uiomove_fromphys() over td_ma array.
997 * Since vnode locks do not cover the whole i/o anymore, rangelocks
998 * make the current i/o request atomic with respect to other i/os and
1003 * Decode vn_io_fault_args and perform the corresponding i/o.
1006 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1010 switch (args->kind) {
1011 case VN_IO_FAULT_FOP:
1012 return ((args->args.fop_args.doio)(args->args.fop_args.fp,
1013 uio, args->cred, args->flags, td));
1014 case VN_IO_FAULT_VOP:
1015 if (uio->uio_rw == UIO_READ) {
1016 return (VOP_READ(args->args.vop_args.vp, uio,
1017 args->flags, args->cred));
1018 } else if (uio->uio_rw == UIO_WRITE) {
1019 return (VOP_WRITE(args->args.vop_args.vp, uio,
1020 args->flags, args->cred));
1024 panic("vn_io_fault_doio: unknown kind of io %d %d", args->kind,
1029 vn_io_fault_touch(char *base, const struct uio *uio)
1034 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1040 vn_io_fault_prefault_user(const struct uio *uio)
1043 const struct iovec *iov;
1048 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1049 ("vn_io_fault_prefault userspace"));
1053 resid = uio->uio_resid;
1054 base = iov->iov_base;
1057 error = vn_io_fault_touch(base, uio);
1060 if (len < PAGE_SIZE) {
1062 error = vn_io_fault_touch(base + len - 1, uio);
1067 if (++i >= uio->uio_iovcnt)
1069 iov = uio->uio_iov + i;
1070 base = iov->iov_base;
1082 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1083 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1084 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1085 * into args and call vn_io_fault1() to handle faults during the user
1086 * mode buffer accesses.
1089 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1092 vm_page_t ma[io_hold_cnt + 2];
1093 struct uio *uio_clone, short_uio;
1094 struct iovec short_iovec[1];
1095 vm_page_t *prev_td_ma;
1097 vm_offset_t addr, end;
1100 int error, cnt, save, saveheld, prev_td_ma_cnt;
1102 if (vn_io_fault_prefault) {
1103 error = vn_io_fault_prefault_user(uio);
1105 return (error); /* Or ignore ? */
1108 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1111 * The UFS follows IO_UNIT directive and replays back both
1112 * uio_offset and uio_resid if an error is encountered during the
1113 * operation. But, since the iovec may be already advanced,
1114 * uio is still in an inconsistent state.
1116 * Cache a copy of the original uio, which is advanced to the redo
1117 * point using UIO_NOCOPY below.
1119 uio_clone = cloneuio(uio);
1120 resid = uio->uio_resid;
1122 short_uio.uio_segflg = UIO_USERSPACE;
1123 short_uio.uio_rw = uio->uio_rw;
1124 short_uio.uio_td = uio->uio_td;
1126 save = vm_fault_disable_pagefaults();
1127 error = vn_io_fault_doio(args, uio, td);
1128 if (error != EFAULT)
1131 atomic_add_long(&vn_io_faults_cnt, 1);
1132 uio_clone->uio_segflg = UIO_NOCOPY;
1133 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1134 uio_clone->uio_segflg = uio->uio_segflg;
1136 saveheld = curthread_pflags_set(TDP_UIOHELD);
1137 prev_td_ma = td->td_ma;
1138 prev_td_ma_cnt = td->td_ma_cnt;
1140 while (uio_clone->uio_resid != 0) {
1141 len = uio_clone->uio_iov->iov_len;
1143 KASSERT(uio_clone->uio_iovcnt >= 1,
1144 ("iovcnt underflow"));
1145 uio_clone->uio_iov++;
1146 uio_clone->uio_iovcnt--;
1149 if (len > io_hold_cnt * PAGE_SIZE)
1150 len = io_hold_cnt * PAGE_SIZE;
1151 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1152 end = round_page(addr + len);
1157 cnt = atop(end - trunc_page(addr));
1159 * A perfectly misaligned address and length could cause
1160 * both the start and the end of the chunk to use partial
1161 * page. +2 accounts for such a situation.
1163 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1164 addr, len, prot, ma, io_hold_cnt + 2);
1169 short_uio.uio_iov = &short_iovec[0];
1170 short_iovec[0].iov_base = (void *)addr;
1171 short_uio.uio_iovcnt = 1;
1172 short_uio.uio_resid = short_iovec[0].iov_len = len;
1173 short_uio.uio_offset = uio_clone->uio_offset;
1175 td->td_ma_cnt = cnt;
1177 error = vn_io_fault_doio(args, &short_uio, td);
1178 vm_page_unhold_pages(ma, cnt);
1179 adv = len - short_uio.uio_resid;
1181 uio_clone->uio_iov->iov_base =
1182 (char *)uio_clone->uio_iov->iov_base + adv;
1183 uio_clone->uio_iov->iov_len -= adv;
1184 uio_clone->uio_resid -= adv;
1185 uio_clone->uio_offset += adv;
1187 uio->uio_resid -= adv;
1188 uio->uio_offset += adv;
1190 if (error != 0 || adv == 0)
1193 td->td_ma = prev_td_ma;
1194 td->td_ma_cnt = prev_td_ma_cnt;
1195 curthread_pflags_restore(saveheld);
1197 vm_fault_enable_pagefaults(save);
1198 free(uio_clone, M_IOV);
1203 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1204 int flags, struct thread *td)
1209 struct vn_io_fault_args args;
1212 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1214 foffset_lock_uio(fp, uio, flags);
1215 if (do_vn_io_fault(vp, uio)) {
1216 args.kind = VN_IO_FAULT_FOP;
1217 args.args.fop_args.fp = fp;
1218 args.args.fop_args.doio = doio;
1219 args.cred = active_cred;
1220 args.flags = flags | FOF_OFFSET;
1221 if (uio->uio_rw == UIO_READ) {
1222 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1223 uio->uio_offset + uio->uio_resid);
1224 } else if ((fp->f_flag & O_APPEND) != 0 ||
1225 (flags & FOF_OFFSET) == 0) {
1226 /* For appenders, punt and lock the whole range. */
1227 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1229 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1230 uio->uio_offset + uio->uio_resid);
1232 error = vn_io_fault1(vp, uio, &args, td);
1233 vn_rangelock_unlock(vp, rl_cookie);
1235 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1237 foffset_unlock_uio(fp, uio, flags);
1242 * Helper function to perform the requested uiomove operation using
1243 * the held pages for io->uio_iov[0].iov_base buffer instead of
1244 * copyin/copyout. Access to the pages with uiomove_fromphys()
1245 * instead of iov_base prevents page faults that could occur due to
1246 * pmap_collect() invalidating the mapping created by
1247 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1248 * object cleanup revoking the write access from page mappings.
1250 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1251 * instead of plain uiomove().
1254 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1256 struct uio transp_uio;
1257 struct iovec transp_iov[1];
1263 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1264 uio->uio_segflg != UIO_USERSPACE)
1265 return (uiomove(data, xfersize, uio));
1267 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1268 transp_iov[0].iov_base = data;
1269 transp_uio.uio_iov = &transp_iov[0];
1270 transp_uio.uio_iovcnt = 1;
1271 if (xfersize > uio->uio_resid)
1272 xfersize = uio->uio_resid;
1273 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1274 transp_uio.uio_offset = 0;
1275 transp_uio.uio_segflg = UIO_SYSSPACE;
1277 * Since transp_iov points to data, and td_ma page array
1278 * corresponds to original uio->uio_iov, we need to invert the
1279 * direction of the i/o operation as passed to
1280 * uiomove_fromphys().
1282 switch (uio->uio_rw) {
1284 transp_uio.uio_rw = UIO_READ;
1287 transp_uio.uio_rw = UIO_WRITE;
1290 transp_uio.uio_td = uio->uio_td;
1291 error = uiomove_fromphys(td->td_ma,
1292 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1293 xfersize, &transp_uio);
1294 adv = xfersize - transp_uio.uio_resid;
1296 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1297 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1299 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1301 td->td_ma_cnt -= pgadv;
1302 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1303 uio->uio_iov->iov_len -= adv;
1304 uio->uio_resid -= adv;
1305 uio->uio_offset += adv;
1310 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1314 vm_offset_t iov_base;
1318 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1319 uio->uio_segflg != UIO_USERSPACE)
1320 return (uiomove_fromphys(ma, offset, xfersize, uio));
1322 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1323 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1324 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1325 switch (uio->uio_rw) {
1327 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1331 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1335 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1337 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1339 td->td_ma_cnt -= pgadv;
1340 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1341 uio->uio_iov->iov_len -= cnt;
1342 uio->uio_resid -= cnt;
1343 uio->uio_offset += cnt;
1349 * File table truncate routine.
1352 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1364 * Lock the whole range for truncation. Otherwise split i/o
1365 * might happen partly before and partly after the truncation.
1367 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1368 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1371 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1372 if (vp->v_type == VDIR) {
1377 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1381 error = vn_writechk(vp);
1384 vattr.va_size = length;
1385 if ((fp->f_flag & O_FSYNC) != 0)
1386 vattr.va_vaflags |= VA_SYNC;
1387 error = VOP_SETATTR(vp, &vattr, fp->f_cred);
1391 vn_finished_write(mp);
1393 vn_rangelock_unlock(vp, rl_cookie);
1398 * File table vnode stat routine.
1401 vn_statfile(fp, sb, active_cred, td)
1404 struct ucred *active_cred;
1407 struct vnode *vp = fp->f_vnode;
1410 vn_lock(vp, LK_SHARED | LK_RETRY);
1411 error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1418 * Stat a vnode; implementation for the stat syscall
1421 vn_stat(vp, sb, active_cred, file_cred, td)
1423 register struct stat *sb;
1424 struct ucred *active_cred;
1425 struct ucred *file_cred;
1429 register struct vattr *vap;
1434 error = mac_vnode_check_stat(active_cred, file_cred, vp);
1442 * Initialize defaults for new and unusual fields, so that file
1443 * systems which don't support these fields don't need to know
1446 vap->va_birthtime.tv_sec = -1;
1447 vap->va_birthtime.tv_nsec = 0;
1448 vap->va_fsid = VNOVAL;
1449 vap->va_rdev = NODEV;
1451 error = VOP_GETATTR(vp, vap, active_cred);
1456 * Zero the spare stat fields
1458 bzero(sb, sizeof *sb);
1461 * Copy from vattr table
1463 if (vap->va_fsid != VNOVAL)
1464 sb->st_dev = vap->va_fsid;
1466 sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1467 sb->st_ino = vap->va_fileid;
1468 mode = vap->va_mode;
1469 switch (vap->va_type) {
1495 sb->st_nlink = vap->va_nlink;
1496 sb->st_uid = vap->va_uid;
1497 sb->st_gid = vap->va_gid;
1498 sb->st_rdev = vap->va_rdev;
1499 if (vap->va_size > OFF_MAX)
1501 sb->st_size = vap->va_size;
1502 sb->st_atim = vap->va_atime;
1503 sb->st_mtim = vap->va_mtime;
1504 sb->st_ctim = vap->va_ctime;
1505 sb->st_birthtim = vap->va_birthtime;
1508 * According to www.opengroup.org, the meaning of st_blksize is
1509 * "a filesystem-specific preferred I/O block size for this
1510 * object. In some filesystem types, this may vary from file
1512 * Use miminum/default of PAGE_SIZE (e.g. for VCHR).
1515 sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1517 sb->st_flags = vap->va_flags;
1518 if (priv_check(td, PRIV_VFS_GENERATION))
1521 sb->st_gen = vap->va_gen;
1523 sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1528 * File table vnode ioctl routine.
1531 vn_ioctl(fp, com, data, active_cred, td)
1535 struct ucred *active_cred;
1543 switch (vp->v_type) {
1548 vn_lock(vp, LK_SHARED | LK_RETRY);
1549 error = VOP_GETATTR(vp, &vattr, active_cred);
1552 *(int *)data = vattr.va_size - fp->f_offset;
1558 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1567 * File table vnode poll routine.
1570 vn_poll(fp, events, active_cred, td)
1573 struct ucred *active_cred;
1581 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1582 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1587 error = VOP_POLL(vp, events, fp->f_cred, td);
1592 * Acquire the requested lock and then check for validity. LK_RETRY
1593 * permits vn_lock to return doomed vnodes.
1596 _vn_lock(struct vnode *vp, int flags, char *file, int line)
1600 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1601 ("vn_lock called with no locktype."));
1603 #ifdef DEBUG_VFS_LOCKS
1604 KASSERT(vp->v_holdcnt != 0,
1605 ("vn_lock %p: zero hold count", vp));
1607 error = VOP_LOCK1(vp, flags, file, line);
1608 flags &= ~LK_INTERLOCK; /* Interlock is always dropped. */
1609 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1610 ("LK_RETRY set with incompatible flags (0x%x) or an error occurred (%d)",
1613 * Callers specify LK_RETRY if they wish to get dead vnodes.
1614 * If RETRY is not set, we return ENOENT instead.
1616 if (error == 0 && vp->v_iflag & VI_DOOMED &&
1617 (flags & LK_RETRY) == 0) {
1622 } while (flags & LK_RETRY && error != 0);
1627 * File table vnode close routine.
1630 vn_closefile(fp, td)
1640 fp->f_ops = &badfileops;
1641 ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1643 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1645 if (__predict_false(ref)) {
1646 lf.l_whence = SEEK_SET;
1649 lf.l_type = F_UNLCK;
1650 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1657 * Preparing to start a filesystem write operation. If the operation is
1658 * permitted, then we bump the count of operations in progress and
1659 * proceed. If a suspend request is in progress, we wait until the
1660 * suspension is over, and then proceed.
1663 vn_start_write_locked(struct mount *mp, int flags)
1667 mtx_assert(MNT_MTX(mp), MA_OWNED);
1671 * Check on status of suspension.
1673 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1674 mp->mnt_susp_owner != curthread) {
1675 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1676 (flags & PCATCH) : 0) | (PUSER - 1);
1677 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1678 if (flags & V_NOWAIT) {
1679 error = EWOULDBLOCK;
1682 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1688 if (flags & V_XSLEEP)
1690 mp->mnt_writeopcount++;
1692 if (error != 0 || (flags & V_XSLEEP) != 0)
1699 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1704 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1705 ("V_MNTREF requires mp"));
1709 * If a vnode is provided, get and return the mount point that
1710 * to which it will write.
1713 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1715 if (error != EOPNOTSUPP)
1720 if ((mp = *mpp) == NULL)
1724 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1726 * As long as a vnode is not provided we need to acquire a
1727 * refcount for the provided mountpoint too, in order to
1728 * emulate a vfs_ref().
1731 if (vp == NULL && (flags & V_MNTREF) == 0)
1734 return (vn_start_write_locked(mp, flags));
1738 * Secondary suspension. Used by operations such as vop_inactive
1739 * routines that are needed by the higher level functions. These
1740 * are allowed to proceed until all the higher level functions have
1741 * completed (indicated by mnt_writeopcount dropping to zero). At that
1742 * time, these operations are halted until the suspension is over.
1745 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1750 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1751 ("V_MNTREF requires mp"));
1755 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1757 if (error != EOPNOTSUPP)
1763 * If we are not suspended or have not yet reached suspended
1764 * mode, then let the operation proceed.
1766 if ((mp = *mpp) == NULL)
1770 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1772 * As long as a vnode is not provided we need to acquire a
1773 * refcount for the provided mountpoint too, in order to
1774 * emulate a vfs_ref().
1777 if (vp == NULL && (flags & V_MNTREF) == 0)
1779 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1780 mp->mnt_secondary_writes++;
1781 mp->mnt_secondary_accwrites++;
1785 if (flags & V_NOWAIT) {
1788 return (EWOULDBLOCK);
1791 * Wait for the suspension to finish.
1793 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1794 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1803 * Filesystem write operation has completed. If we are suspending and this
1804 * operation is the last one, notify the suspender that the suspension is
1808 vn_finished_write(mp)
1815 mp->mnt_writeopcount--;
1816 if (mp->mnt_writeopcount < 0)
1817 panic("vn_finished_write: neg cnt");
1818 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1819 mp->mnt_writeopcount <= 0)
1820 wakeup(&mp->mnt_writeopcount);
1826 * Filesystem secondary write operation has completed. If we are
1827 * suspending and this operation is the last one, notify the suspender
1828 * that the suspension is now in effect.
1831 vn_finished_secondary_write(mp)
1838 mp->mnt_secondary_writes--;
1839 if (mp->mnt_secondary_writes < 0)
1840 panic("vn_finished_secondary_write: neg cnt");
1841 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1842 mp->mnt_secondary_writes <= 0)
1843 wakeup(&mp->mnt_secondary_writes);
1850 * Request a filesystem to suspend write operations.
1853 vfs_write_suspend(struct mount *mp, int flags)
1858 if (mp->mnt_susp_owner == curthread) {
1862 while (mp->mnt_kern_flag & MNTK_SUSPEND)
1863 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1866 * Unmount holds a write reference on the mount point. If we
1867 * own busy reference and drain for writers, we deadlock with
1868 * the reference draining in the unmount path. Callers of
1869 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1870 * vfs_busy() reference is owned and caller is not in the
1873 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1874 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1879 mp->mnt_kern_flag |= MNTK_SUSPEND;
1880 mp->mnt_susp_owner = curthread;
1881 if (mp->mnt_writeopcount > 0)
1882 (void) msleep(&mp->mnt_writeopcount,
1883 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1886 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0)
1887 vfs_write_resume(mp, 0);
1892 * Request a filesystem to resume write operations.
1895 vfs_write_resume(struct mount *mp, int flags)
1899 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1900 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
1901 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
1903 mp->mnt_susp_owner = NULL;
1904 wakeup(&mp->mnt_writeopcount);
1905 wakeup(&mp->mnt_flag);
1906 curthread->td_pflags &= ~TDP_IGNSUSP;
1907 if ((flags & VR_START_WRITE) != 0) {
1909 mp->mnt_writeopcount++;
1912 if ((flags & VR_NO_SUSPCLR) == 0)
1914 } else if ((flags & VR_START_WRITE) != 0) {
1916 vn_start_write_locked(mp, 0);
1923 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
1927 vfs_write_suspend_umnt(struct mount *mp)
1931 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
1932 ("vfs_write_suspend_umnt: recursed"));
1934 /* dounmount() already called vn_start_write(). */
1936 vn_finished_write(mp);
1937 error = vfs_write_suspend(mp, 0);
1939 vn_start_write(NULL, &mp, V_WAIT);
1943 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
1946 vn_start_write(NULL, &mp, V_WAIT);
1948 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
1949 wakeup(&mp->mnt_flag);
1951 curthread->td_pflags |= TDP_IGNSUSP;
1956 * Implement kqueues for files by translating it to vnode operation.
1959 vn_kqfilter(struct file *fp, struct knote *kn)
1962 return (VOP_KQFILTER(fp->f_vnode, kn));
1966 * Simplified in-kernel wrapper calls for extended attribute access.
1967 * Both calls pass in a NULL credential, authorizing as "kernel" access.
1968 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
1971 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
1972 const char *attrname, int *buflen, char *buf, struct thread *td)
1978 iov.iov_len = *buflen;
1981 auio.uio_iov = &iov;
1982 auio.uio_iovcnt = 1;
1983 auio.uio_rw = UIO_READ;
1984 auio.uio_segflg = UIO_SYSSPACE;
1986 auio.uio_offset = 0;
1987 auio.uio_resid = *buflen;
1989 if ((ioflg & IO_NODELOCKED) == 0)
1990 vn_lock(vp, LK_SHARED | LK_RETRY);
1992 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
1994 /* authorize attribute retrieval as kernel */
1995 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
1998 if ((ioflg & IO_NODELOCKED) == 0)
2002 *buflen = *buflen - auio.uio_resid;
2009 * XXX failure mode if partially written?
2012 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2013 const char *attrname, int buflen, char *buf, struct thread *td)
2020 iov.iov_len = buflen;
2023 auio.uio_iov = &iov;
2024 auio.uio_iovcnt = 1;
2025 auio.uio_rw = UIO_WRITE;
2026 auio.uio_segflg = UIO_SYSSPACE;
2028 auio.uio_offset = 0;
2029 auio.uio_resid = buflen;
2031 if ((ioflg & IO_NODELOCKED) == 0) {
2032 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2034 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2037 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2039 /* authorize attribute setting as kernel */
2040 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2042 if ((ioflg & IO_NODELOCKED) == 0) {
2043 vn_finished_write(mp);
2051 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2052 const char *attrname, struct thread *td)
2057 if ((ioflg & IO_NODELOCKED) == 0) {
2058 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2060 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2063 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2065 /* authorize attribute removal as kernel */
2066 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2067 if (error == EOPNOTSUPP)
2068 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2071 if ((ioflg & IO_NODELOCKED) == 0) {
2072 vn_finished_write(mp);
2080 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2084 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2088 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2091 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2096 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2097 int lkflags, struct vnode **rvp)
2102 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2104 ltype = VOP_ISLOCKED(vp);
2105 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2106 ("vn_vget_ino: vp not locked"));
2107 error = vfs_busy(mp, MBF_NOWAIT);
2111 error = vfs_busy(mp, 0);
2112 vn_lock(vp, ltype | LK_RETRY);
2116 if (vp->v_iflag & VI_DOOMED) {
2122 error = alloc(mp, alloc_arg, lkflags, rvp);
2125 vn_lock(vp, ltype | LK_RETRY);
2126 if (vp->v_iflag & VI_DOOMED) {
2139 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2140 const struct thread *td)
2143 if (vp->v_type != VREG || td == NULL)
2145 PROC_LOCK(td->td_proc);
2146 if ((uoff_t)uio->uio_offset + uio->uio_resid >
2147 lim_cur(td->td_proc, RLIMIT_FSIZE)) {
2148 kern_psignal(td->td_proc, SIGXFSZ);
2149 PROC_UNLOCK(td->td_proc);
2152 PROC_UNLOCK(td->td_proc);
2157 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2164 vn_lock(vp, LK_SHARED | LK_RETRY);
2165 AUDIT_ARG_VNODE1(vp);
2168 return (setfmode(td, active_cred, vp, mode));
2172 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2179 vn_lock(vp, LK_SHARED | LK_RETRY);
2180 AUDIT_ARG_VNODE1(vp);
2183 return (setfown(td, active_cred, vp, uid, gid));
2187 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2191 if ((object = vp->v_object) == NULL)
2193 VM_OBJECT_WLOCK(object);
2194 vm_object_page_remove(object, start, end, 0);
2195 VM_OBJECT_WUNLOCK(object);
2199 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2207 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2208 ("Wrong command %lu", cmd));
2210 if (vn_lock(vp, LK_SHARED) != 0)
2212 if (vp->v_type != VREG) {
2216 error = VOP_GETATTR(vp, &va, cred);
2220 if (noff >= va.va_size) {
2224 bsize = vp->v_mount->mnt_stat.f_iosize;
2225 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize) {
2226 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2227 if (error == EOPNOTSUPP) {
2231 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2232 (bnp != -1 && cmd == FIOSEEKDATA)) {
2239 if (noff > va.va_size)
2241 /* noff == va.va_size. There is an implicit hole at the end of file. */
2242 if (cmd == FIOSEEKDATA)
2252 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2257 off_t foffset, size;
2260 cred = td->td_ucred;
2262 foffset = foffset_lock(fp, 0);
2263 noneg = (vp->v_type != VCHR);
2269 (offset > 0 && foffset > OFF_MAX - offset))) {
2276 vn_lock(vp, LK_SHARED | LK_RETRY);
2277 error = VOP_GETATTR(vp, &vattr, cred);
2283 * If the file references a disk device, then fetch
2284 * the media size and use that to determine the ending
2287 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2288 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2289 vattr.va_size = size;
2291 (vattr.va_size > OFF_MAX ||
2292 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2296 offset += vattr.va_size;
2301 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2304 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2309 if (error == 0 && noneg && offset < 0)
2313 VFS_KNOTE_UNLOCKED(vp, 0);
2314 *(off_t *)(td->td_retval) = offset;
2316 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2321 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2327 * Grant permission if the caller is the owner of the file, or
2328 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2329 * on the file. If the time pointer is null, then write
2330 * permission on the file is also sufficient.
2332 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2333 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2334 * will be allowed to set the times [..] to the current
2337 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2338 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2339 error = VOP_ACCESS(vp, VWRITE, cred, td);