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>
50 #include <sys/fcntl.h>
56 #include <sys/limits.h>
59 #include <sys/mount.h>
60 #include <sys/mutex.h>
61 #include <sys/namei.h>
62 #include <sys/vnode.h>
65 #include <sys/filio.h>
66 #include <sys/resourcevar.h>
67 #include <sys/rwlock.h>
69 #include <sys/sysctl.h>
70 #include <sys/ttycom.h>
72 #include <sys/syslog.h>
73 #include <sys/unistd.h>
76 #include <security/audit/audit.h>
77 #include <security/mac/mac_framework.h>
80 #include <vm/vm_extern.h>
82 #include <vm/vm_map.h>
83 #include <vm/vm_object.h>
84 #include <vm/vm_page.h>
85 #include <vm/vnode_pager.h>
87 static fo_rdwr_t vn_read;
88 static fo_rdwr_t vn_write;
89 static fo_rdwr_t vn_io_fault;
90 static fo_truncate_t vn_truncate;
91 static fo_ioctl_t vn_ioctl;
92 static fo_poll_t vn_poll;
93 static fo_kqfilter_t vn_kqfilter;
94 static fo_stat_t vn_statfile;
95 static fo_close_t vn_closefile;
96 static fo_mmap_t vn_mmap;
98 struct fileops vnops = {
99 .fo_read = vn_io_fault,
100 .fo_write = vn_io_fault,
101 .fo_truncate = vn_truncate,
102 .fo_ioctl = vn_ioctl,
104 .fo_kqfilter = vn_kqfilter,
105 .fo_stat = vn_statfile,
106 .fo_close = vn_closefile,
107 .fo_chmod = vn_chmod,
108 .fo_chown = vn_chown,
109 .fo_sendfile = vn_sendfile,
111 .fo_fill_kinfo = vn_fill_kinfo,
113 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
116 static const int io_hold_cnt = 16;
117 static int vn_io_fault_enable = 1;
118 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW,
119 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
120 static int vn_io_fault_prefault = 0;
121 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RW,
122 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
123 static u_long vn_io_faults_cnt;
124 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
125 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
128 * Returns true if vn_io_fault mode of handling the i/o request should
132 do_vn_io_fault(struct vnode *vp, struct uio *uio)
136 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
137 (mp = vp->v_mount) != NULL &&
138 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
142 * Structure used to pass arguments to vn_io_fault1(), to do either
143 * file- or vnode-based I/O calls.
145 struct vn_io_fault_args {
153 struct fop_args_tag {
157 struct vop_args_tag {
163 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
164 struct vn_io_fault_args *args, struct thread *td);
167 vn_open(ndp, flagp, cmode, fp)
168 struct nameidata *ndp;
172 struct thread *td = ndp->ni_cnd.cn_thread;
174 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
178 * Common code for vnode open operations via a name lookup.
179 * Lookup the vnode and invoke VOP_CREATE if needed.
180 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
182 * Note that this does NOT free nameidata for the successful case,
183 * due to the NDINIT being done elsewhere.
186 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
187 struct ucred *cred, struct file *fp)
191 struct thread *td = ndp->ni_cnd.cn_thread;
193 struct vattr *vap = &vat;
198 if (fmode & O_CREAT) {
199 ndp->ni_cnd.cn_nameiop = CREATE;
201 * Set NOCACHE to avoid flushing the cache when
202 * rolling in many files at once.
204 ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE;
205 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
206 ndp->ni_cnd.cn_flags |= FOLLOW;
207 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
208 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
209 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
210 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
212 if ((error = namei(ndp)) != 0)
214 if (ndp->ni_vp == NULL) {
217 vap->va_mode = cmode;
219 vap->va_vaflags |= VA_EXCLUSIVE;
220 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
221 NDFREE(ndp, NDF_ONLY_PNBUF);
223 if ((error = vn_start_write(NULL, &mp,
224 V_XSLEEP | PCATCH)) != 0)
228 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
229 ndp->ni_cnd.cn_flags |= MAKEENTRY;
231 error = mac_vnode_check_create(cred, ndp->ni_dvp,
235 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
238 vn_finished_write(mp);
240 NDFREE(ndp, NDF_ONLY_PNBUF);
246 if (ndp->ni_dvp == ndp->ni_vp)
252 if (fmode & O_EXCL) {
259 ndp->ni_cnd.cn_nameiop = LOOKUP;
260 ndp->ni_cnd.cn_flags = ISOPEN |
261 ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
262 if (!(fmode & FWRITE))
263 ndp->ni_cnd.cn_flags |= LOCKSHARED;
264 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
265 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
266 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
267 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
268 if ((error = namei(ndp)) != 0)
272 error = vn_open_vnode(vp, fmode, cred, td, fp);
278 NDFREE(ndp, NDF_ONLY_PNBUF);
286 * Common code for vnode open operations once a vnode is located.
287 * Check permissions, and call the VOP_OPEN routine.
290 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
291 struct thread *td, struct file *fp)
296 int error, have_flock, lock_flags, type;
298 if (vp->v_type == VLNK)
300 if (vp->v_type == VSOCK)
302 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
305 if (fmode & (FWRITE | O_TRUNC)) {
306 if (vp->v_type == VDIR)
314 if ((fmode & O_APPEND) && (fmode & FWRITE))
319 if (fmode & O_VERIFY)
321 error = mac_vnode_check_open(cred, vp, accmode);
325 accmode &= ~(VCREAT | VVERIFY);
327 if ((fmode & O_CREAT) == 0) {
328 if (accmode & VWRITE) {
329 error = vn_writechk(vp);
334 error = VOP_ACCESS(vp, accmode, cred, td);
339 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
340 vn_lock(vp, LK_UPGRADE | LK_RETRY);
341 if ((error = VOP_OPEN(vp, fmode, cred, td, fp)) != 0)
344 if (fmode & (O_EXLOCK | O_SHLOCK)) {
345 KASSERT(fp != NULL, ("open with flock requires fp"));
346 lock_flags = VOP_ISLOCKED(vp);
348 lf.l_whence = SEEK_SET;
351 if (fmode & O_EXLOCK)
356 if ((fmode & FNONBLOCK) == 0)
358 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
359 have_flock = (error == 0);
360 vn_lock(vp, lock_flags | LK_RETRY);
361 if (error == 0 && vp->v_iflag & VI_DOOMED)
364 * Another thread might have used this vnode as an
365 * executable while the vnode lock was dropped.
366 * Ensure the vnode is still able to be opened for
367 * writing after the lock has been obtained.
369 if (error == 0 && accmode & VWRITE)
370 error = vn_writechk(vp);
374 lf.l_whence = SEEK_SET;
378 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf,
381 vn_start_write(vp, &mp, V_WAIT);
382 vn_lock(vp, lock_flags | LK_RETRY);
383 (void)VOP_CLOSE(vp, fmode, cred, td);
384 vn_finished_write(mp);
385 /* Prevent second close from fdrop()->vn_close(). */
387 fp->f_ops= &badfileops;
390 fp->f_flag |= FHASLOCK;
392 if (fmode & FWRITE) {
393 VOP_ADD_WRITECOUNT(vp, 1);
394 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
395 __func__, vp, vp->v_writecount);
397 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
402 * Check for write permissions on the specified vnode.
403 * Prototype text segments cannot be written.
407 register struct vnode *vp;
410 ASSERT_VOP_LOCKED(vp, "vn_writechk");
412 * If there's shared text associated with
413 * the vnode, try to free it up once. If
414 * we fail, we can't allow writing.
426 vn_close(vp, flags, file_cred, td)
427 register struct vnode *vp;
429 struct ucred *file_cred;
433 int error, lock_flags;
435 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
436 MNT_EXTENDED_SHARED(vp->v_mount))
437 lock_flags = LK_SHARED;
439 lock_flags = LK_EXCLUSIVE;
441 vn_start_write(vp, &mp, V_WAIT);
442 vn_lock(vp, lock_flags | LK_RETRY);
443 if (flags & FWRITE) {
444 VNASSERT(vp->v_writecount > 0, vp,
445 ("vn_close: negative writecount"));
446 VOP_ADD_WRITECOUNT(vp, -1);
447 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
448 __func__, vp, vp->v_writecount);
450 error = VOP_CLOSE(vp, flags, file_cred, td);
452 vn_finished_write(mp);
457 * Heuristic to detect sequential operation.
460 sequential_heuristic(struct uio *uio, struct file *fp)
463 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
464 if (fp->f_flag & FRDAHEAD)
465 return (fp->f_seqcount << IO_SEQSHIFT);
468 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
469 * that the first I/O is normally considered to be slightly
470 * sequential. Seeking to offset 0 doesn't change sequentiality
471 * unless previous seeks have reduced f_seqcount to 0, in which
472 * case offset 0 is not special.
474 if ((uio->uio_offset == 0 && fp->f_seqcount > 0) ||
475 uio->uio_offset == fp->f_nextoff) {
477 * f_seqcount is in units of fixed-size blocks so that it
478 * depends mainly on the amount of sequential I/O and not
479 * much on the number of sequential I/O's. The fixed size
480 * of 16384 is hard-coded here since it is (not quite) just
481 * a magic size that works well here. This size is more
482 * closely related to the best I/O size for real disks than
483 * to any block size used by software.
485 fp->f_seqcount += howmany(uio->uio_resid, 16384);
486 if (fp->f_seqcount > IO_SEQMAX)
487 fp->f_seqcount = IO_SEQMAX;
488 return (fp->f_seqcount << IO_SEQSHIFT);
491 /* Not sequential. Quickly draw-down sequentiality. */
492 if (fp->f_seqcount > 1)
500 * Package up an I/O request on a vnode into a uio and do it.
503 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
504 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
505 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
512 struct vn_io_fault_args args;
513 int error, lock_flags;
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)
749 mtxp = mtx_pool_find(mtxpool_sleep, fp);
751 if (uio->uio_offset >= fp->f_advice->fa_start &&
752 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
753 ret = fp->f_advice->fa_advice;
759 * File table vnode read routine.
762 vn_read(fp, uio, active_cred, flags, td)
765 struct ucred *active_cred;
773 off_t offset, start, end;
775 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
777 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
780 if (fp->f_flag & FNONBLOCK)
782 if (fp->f_flag & O_DIRECT)
784 advice = get_advice(fp, uio);
785 vn_lock(vp, LK_SHARED | LK_RETRY);
788 case POSIX_FADV_NORMAL:
789 case POSIX_FADV_SEQUENTIAL:
790 case POSIX_FADV_NOREUSE:
791 ioflag |= sequential_heuristic(uio, fp);
793 case POSIX_FADV_RANDOM:
794 /* Disable read-ahead for random I/O. */
797 offset = uio->uio_offset;
800 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
803 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
804 fp->f_nextoff = uio->uio_offset;
806 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
807 offset != uio->uio_offset) {
809 * Use POSIX_FADV_DONTNEED to flush clean pages and
810 * buffers for the backing file after a
811 * POSIX_FADV_NOREUSE read(2). To optimize the common
812 * case of using POSIX_FADV_NOREUSE with sequential
813 * access, track the previous implicit DONTNEED
814 * request and grow this request to include the
815 * current read(2) in addition to the previous
816 * DONTNEED. With purely sequential access this will
817 * cause the DONTNEED requests to continously grow to
818 * cover all of the previously read regions of the
819 * file. This allows filesystem blocks that are
820 * accessed by multiple calls to read(2) to be flushed
821 * once the last read(2) finishes.
824 end = uio->uio_offset - 1;
825 mtxp = mtx_pool_find(mtxpool_sleep, fp);
827 if (fp->f_advice != NULL &&
828 fp->f_advice->fa_advice == POSIX_FADV_NOREUSE) {
829 if (start != 0 && fp->f_advice->fa_prevend + 1 == start)
830 start = fp->f_advice->fa_prevstart;
831 else if (fp->f_advice->fa_prevstart != 0 &&
832 fp->f_advice->fa_prevstart == end + 1)
833 end = fp->f_advice->fa_prevend;
834 fp->f_advice->fa_prevstart = start;
835 fp->f_advice->fa_prevend = end;
838 error = VOP_ADVISE(vp, start, end, POSIX_FADV_DONTNEED);
844 * File table vnode write routine.
847 vn_write(fp, uio, active_cred, flags, td)
850 struct ucred *active_cred;
857 int error, ioflag, lock_flags;
859 off_t offset, start, end;
861 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
863 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
865 if (vp->v_type == VREG)
868 if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
870 if (fp->f_flag & FNONBLOCK)
872 if (fp->f_flag & O_DIRECT)
874 if ((fp->f_flag & O_FSYNC) ||
875 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
878 if (vp->v_type != VCHR &&
879 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
882 advice = get_advice(fp, uio);
884 if (MNT_SHARED_WRITES(mp) ||
885 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
886 lock_flags = LK_SHARED;
888 lock_flags = LK_EXCLUSIVE;
891 vn_lock(vp, lock_flags | LK_RETRY);
893 case POSIX_FADV_NORMAL:
894 case POSIX_FADV_SEQUENTIAL:
895 case POSIX_FADV_NOREUSE:
896 ioflag |= sequential_heuristic(uio, fp);
898 case POSIX_FADV_RANDOM:
899 /* XXX: Is this correct? */
902 offset = uio->uio_offset;
905 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
908 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
909 fp->f_nextoff = uio->uio_offset;
911 if (vp->v_type != VCHR)
912 vn_finished_write(mp);
913 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
914 offset != uio->uio_offset) {
916 * Use POSIX_FADV_DONTNEED to flush clean pages and
917 * buffers for the backing file after a
918 * POSIX_FADV_NOREUSE write(2). To optimize the
919 * common case of using POSIX_FADV_NOREUSE with
920 * sequential access, track the previous implicit
921 * DONTNEED request and grow this request to include
922 * the current write(2) in addition to the previous
923 * DONTNEED. With purely sequential access this will
924 * cause the DONTNEED requests to continously grow to
925 * cover all of the previously written regions of the
928 * Note that the blocks just written are almost
929 * certainly still dirty, so this only works when
930 * VOP_ADVISE() calls from subsequent writes push out
931 * the data written by this write(2) once the backing
932 * buffers are clean. However, as compared to forcing
933 * IO_DIRECT, this gives much saner behavior. Write
934 * clustering is still allowed, and clean pages are
935 * merely moved to the cache page queue rather than
936 * outright thrown away. This means a subsequent
937 * read(2) can still avoid hitting the disk if the
938 * pages have not been reclaimed.
940 * This does make POSIX_FADV_NOREUSE largely useless
941 * with non-sequential access. However, sequential
942 * access is the more common use case and the flag is
946 end = uio->uio_offset - 1;
947 mtxp = mtx_pool_find(mtxpool_sleep, fp);
949 if (fp->f_advice != NULL &&
950 fp->f_advice->fa_advice == POSIX_FADV_NOREUSE) {
951 if (start != 0 && fp->f_advice->fa_prevend + 1 == start)
952 start = fp->f_advice->fa_prevstart;
953 else if (fp->f_advice->fa_prevstart != 0 &&
954 fp->f_advice->fa_prevstart == end + 1)
955 end = fp->f_advice->fa_prevend;
956 fp->f_advice->fa_prevstart = start;
957 fp->f_advice->fa_prevend = end;
960 error = VOP_ADVISE(vp, start, end, POSIX_FADV_DONTNEED);
968 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
969 * prevent the following deadlock:
971 * Assume that the thread A reads from the vnode vp1 into userspace
972 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
973 * currently not resident, then system ends up with the call chain
974 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
975 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
976 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
977 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
978 * backed by the pages of vnode vp1, and some page in buf2 is not
979 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
981 * To prevent the lock order reversal and deadlock, vn_io_fault() does
982 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
983 * Instead, it first tries to do the whole range i/o with pagefaults
984 * disabled. If all pages in the i/o buffer are resident and mapped,
985 * VOP will succeed (ignoring the genuine filesystem errors).
986 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
987 * i/o in chunks, with all pages in the chunk prefaulted and held
988 * using vm_fault_quick_hold_pages().
990 * Filesystems using this deadlock avoidance scheme should use the
991 * array of the held pages from uio, saved in the curthread->td_ma,
992 * instead of doing uiomove(). A helper function
993 * vn_io_fault_uiomove() converts uiomove request into
994 * uiomove_fromphys() over td_ma array.
996 * Since vnode locks do not cover the whole i/o anymore, rangelocks
997 * make the current i/o request atomic with respect to other i/os and
1002 * Decode vn_io_fault_args and perform the corresponding i/o.
1005 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1009 switch (args->kind) {
1010 case VN_IO_FAULT_FOP:
1011 return ((args->args.fop_args.doio)(args->args.fop_args.fp,
1012 uio, args->cred, args->flags, td));
1013 case VN_IO_FAULT_VOP:
1014 if (uio->uio_rw == UIO_READ) {
1015 return (VOP_READ(args->args.vop_args.vp, uio,
1016 args->flags, args->cred));
1017 } else if (uio->uio_rw == UIO_WRITE) {
1018 return (VOP_WRITE(args->args.vop_args.vp, uio,
1019 args->flags, args->cred));
1023 panic("vn_io_fault_doio: unknown kind of io %d %d", args->kind,
1028 vn_io_fault_touch(char *base, const struct uio *uio)
1033 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1039 vn_io_fault_prefault_user(const struct uio *uio)
1042 const struct iovec *iov;
1047 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1048 ("vn_io_fault_prefault userspace"));
1052 resid = uio->uio_resid;
1053 base = iov->iov_base;
1056 error = vn_io_fault_touch(base, uio);
1059 if (len < PAGE_SIZE) {
1061 error = vn_io_fault_touch(base + len - 1, uio);
1066 if (++i >= uio->uio_iovcnt)
1068 iov = uio->uio_iov + i;
1069 base = iov->iov_base;
1081 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1082 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1083 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1084 * into args and call vn_io_fault1() to handle faults during the user
1085 * mode buffer accesses.
1088 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1091 vm_page_t ma[io_hold_cnt + 2];
1092 struct uio *uio_clone, short_uio;
1093 struct iovec short_iovec[1];
1094 vm_page_t *prev_td_ma;
1096 vm_offset_t addr, end;
1099 int error, cnt, save, saveheld, prev_td_ma_cnt;
1101 if (vn_io_fault_prefault) {
1102 error = vn_io_fault_prefault_user(uio);
1104 return (error); /* Or ignore ? */
1107 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1110 * The UFS follows IO_UNIT directive and replays back both
1111 * uio_offset and uio_resid if an error is encountered during the
1112 * operation. But, since the iovec may be already advanced,
1113 * uio is still in an inconsistent state.
1115 * Cache a copy of the original uio, which is advanced to the redo
1116 * point using UIO_NOCOPY below.
1118 uio_clone = cloneuio(uio);
1119 resid = uio->uio_resid;
1121 short_uio.uio_segflg = UIO_USERSPACE;
1122 short_uio.uio_rw = uio->uio_rw;
1123 short_uio.uio_td = uio->uio_td;
1125 save = vm_fault_disable_pagefaults();
1126 error = vn_io_fault_doio(args, uio, td);
1127 if (error != EFAULT)
1130 atomic_add_long(&vn_io_faults_cnt, 1);
1131 uio_clone->uio_segflg = UIO_NOCOPY;
1132 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1133 uio_clone->uio_segflg = uio->uio_segflg;
1135 saveheld = curthread_pflags_set(TDP_UIOHELD);
1136 prev_td_ma = td->td_ma;
1137 prev_td_ma_cnt = td->td_ma_cnt;
1139 while (uio_clone->uio_resid != 0) {
1140 len = uio_clone->uio_iov->iov_len;
1142 KASSERT(uio_clone->uio_iovcnt >= 1,
1143 ("iovcnt underflow"));
1144 uio_clone->uio_iov++;
1145 uio_clone->uio_iovcnt--;
1148 if (len > io_hold_cnt * PAGE_SIZE)
1149 len = io_hold_cnt * PAGE_SIZE;
1150 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1151 end = round_page(addr + len);
1156 cnt = atop(end - trunc_page(addr));
1158 * A perfectly misaligned address and length could cause
1159 * both the start and the end of the chunk to use partial
1160 * page. +2 accounts for such a situation.
1162 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1163 addr, len, prot, ma, io_hold_cnt + 2);
1168 short_uio.uio_iov = &short_iovec[0];
1169 short_iovec[0].iov_base = (void *)addr;
1170 short_uio.uio_iovcnt = 1;
1171 short_uio.uio_resid = short_iovec[0].iov_len = len;
1172 short_uio.uio_offset = uio_clone->uio_offset;
1174 td->td_ma_cnt = cnt;
1176 error = vn_io_fault_doio(args, &short_uio, td);
1177 vm_page_unhold_pages(ma, cnt);
1178 adv = len - short_uio.uio_resid;
1180 uio_clone->uio_iov->iov_base =
1181 (char *)uio_clone->uio_iov->iov_base + adv;
1182 uio_clone->uio_iov->iov_len -= adv;
1183 uio_clone->uio_resid -= adv;
1184 uio_clone->uio_offset += adv;
1186 uio->uio_resid -= adv;
1187 uio->uio_offset += adv;
1189 if (error != 0 || adv == 0)
1192 td->td_ma = prev_td_ma;
1193 td->td_ma_cnt = prev_td_ma_cnt;
1194 curthread_pflags_restore(saveheld);
1196 vm_fault_enable_pagefaults(save);
1197 free(uio_clone, M_IOV);
1202 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1203 int flags, struct thread *td)
1208 struct vn_io_fault_args args;
1211 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1213 foffset_lock_uio(fp, uio, flags);
1214 if (do_vn_io_fault(vp, uio)) {
1215 args.kind = VN_IO_FAULT_FOP;
1216 args.args.fop_args.fp = fp;
1217 args.args.fop_args.doio = doio;
1218 args.cred = active_cred;
1219 args.flags = flags | FOF_OFFSET;
1220 if (uio->uio_rw == UIO_READ) {
1221 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1222 uio->uio_offset + uio->uio_resid);
1223 } else if ((fp->f_flag & O_APPEND) != 0 ||
1224 (flags & FOF_OFFSET) == 0) {
1225 /* For appenders, punt and lock the whole range. */
1226 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1228 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1229 uio->uio_offset + uio->uio_resid);
1231 error = vn_io_fault1(vp, uio, &args, td);
1232 vn_rangelock_unlock(vp, rl_cookie);
1234 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1236 foffset_unlock_uio(fp, uio, flags);
1241 * Helper function to perform the requested uiomove operation using
1242 * the held pages for io->uio_iov[0].iov_base buffer instead of
1243 * copyin/copyout. Access to the pages with uiomove_fromphys()
1244 * instead of iov_base prevents page faults that could occur due to
1245 * pmap_collect() invalidating the mapping created by
1246 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1247 * object cleanup revoking the write access from page mappings.
1249 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1250 * instead of plain uiomove().
1253 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1255 struct uio transp_uio;
1256 struct iovec transp_iov[1];
1262 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1263 uio->uio_segflg != UIO_USERSPACE)
1264 return (uiomove(data, xfersize, uio));
1266 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1267 transp_iov[0].iov_base = data;
1268 transp_uio.uio_iov = &transp_iov[0];
1269 transp_uio.uio_iovcnt = 1;
1270 if (xfersize > uio->uio_resid)
1271 xfersize = uio->uio_resid;
1272 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1273 transp_uio.uio_offset = 0;
1274 transp_uio.uio_segflg = UIO_SYSSPACE;
1276 * Since transp_iov points to data, and td_ma page array
1277 * corresponds to original uio->uio_iov, we need to invert the
1278 * direction of the i/o operation as passed to
1279 * uiomove_fromphys().
1281 switch (uio->uio_rw) {
1283 transp_uio.uio_rw = UIO_READ;
1286 transp_uio.uio_rw = UIO_WRITE;
1289 transp_uio.uio_td = uio->uio_td;
1290 error = uiomove_fromphys(td->td_ma,
1291 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1292 xfersize, &transp_uio);
1293 adv = xfersize - transp_uio.uio_resid;
1295 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1296 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1298 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1300 td->td_ma_cnt -= pgadv;
1301 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1302 uio->uio_iov->iov_len -= adv;
1303 uio->uio_resid -= adv;
1304 uio->uio_offset += adv;
1309 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1313 vm_offset_t iov_base;
1317 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1318 uio->uio_segflg != UIO_USERSPACE)
1319 return (uiomove_fromphys(ma, offset, xfersize, uio));
1321 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1322 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1323 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1324 switch (uio->uio_rw) {
1326 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1330 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1334 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1336 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1338 td->td_ma_cnt -= pgadv;
1339 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1340 uio->uio_iov->iov_len -= cnt;
1341 uio->uio_resid -= cnt;
1342 uio->uio_offset += cnt;
1348 * File table truncate routine.
1351 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1363 * Lock the whole range for truncation. Otherwise split i/o
1364 * might happen partly before and partly after the truncation.
1366 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1367 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1370 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1371 if (vp->v_type == VDIR) {
1376 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1380 error = vn_writechk(vp);
1383 vattr.va_size = length;
1384 error = VOP_SETATTR(vp, &vattr, fp->f_cred);
1388 vn_finished_write(mp);
1390 vn_rangelock_unlock(vp, rl_cookie);
1395 * File table vnode stat routine.
1398 vn_statfile(fp, sb, active_cred, td)
1401 struct ucred *active_cred;
1404 struct vnode *vp = fp->f_vnode;
1407 vn_lock(vp, LK_SHARED | LK_RETRY);
1408 error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1415 * Stat a vnode; implementation for the stat syscall
1418 vn_stat(vp, sb, active_cred, file_cred, td)
1420 register struct stat *sb;
1421 struct ucred *active_cred;
1422 struct ucred *file_cred;
1426 register struct vattr *vap;
1431 error = mac_vnode_check_stat(active_cred, file_cred, vp);
1439 * Initialize defaults for new and unusual fields, so that file
1440 * systems which don't support these fields don't need to know
1443 vap->va_birthtime.tv_sec = -1;
1444 vap->va_birthtime.tv_nsec = 0;
1445 vap->va_fsid = VNOVAL;
1446 vap->va_rdev = NODEV;
1448 error = VOP_GETATTR(vp, vap, active_cred);
1453 * Zero the spare stat fields
1455 bzero(sb, sizeof *sb);
1458 * Copy from vattr table
1460 if (vap->va_fsid != VNOVAL)
1461 sb->st_dev = vap->va_fsid;
1463 sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1464 sb->st_ino = vap->va_fileid;
1465 mode = vap->va_mode;
1466 switch (vap->va_type) {
1492 sb->st_nlink = vap->va_nlink;
1493 sb->st_uid = vap->va_uid;
1494 sb->st_gid = vap->va_gid;
1495 sb->st_rdev = vap->va_rdev;
1496 if (vap->va_size > OFF_MAX)
1498 sb->st_size = vap->va_size;
1499 sb->st_atim = vap->va_atime;
1500 sb->st_mtim = vap->va_mtime;
1501 sb->st_ctim = vap->va_ctime;
1502 sb->st_birthtim = vap->va_birthtime;
1505 * According to www.opengroup.org, the meaning of st_blksize is
1506 * "a filesystem-specific preferred I/O block size for this
1507 * object. In some filesystem types, this may vary from file
1509 * Use miminum/default of PAGE_SIZE (e.g. for VCHR).
1512 sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1514 sb->st_flags = vap->va_flags;
1515 if (priv_check(td, PRIV_VFS_GENERATION))
1518 sb->st_gen = vap->va_gen;
1520 sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1525 * File table vnode ioctl routine.
1528 vn_ioctl(fp, com, data, active_cred, td)
1532 struct ucred *active_cred;
1540 switch (vp->v_type) {
1545 vn_lock(vp, LK_SHARED | LK_RETRY);
1546 error = VOP_GETATTR(vp, &vattr, active_cred);
1549 *(int *)data = vattr.va_size - fp->f_offset;
1555 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1564 * File table vnode poll routine.
1567 vn_poll(fp, events, active_cred, td)
1570 struct ucred *active_cred;
1578 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1579 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1584 error = VOP_POLL(vp, events, fp->f_cred, td);
1589 * Acquire the requested lock and then check for validity. LK_RETRY
1590 * permits vn_lock to return doomed vnodes.
1593 _vn_lock(struct vnode *vp, int flags, char *file, int line)
1597 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1598 ("vn_lock called with no locktype."));
1600 #ifdef DEBUG_VFS_LOCKS
1601 KASSERT(vp->v_holdcnt != 0,
1602 ("vn_lock %p: zero hold count", vp));
1604 error = VOP_LOCK1(vp, flags, file, line);
1605 flags &= ~LK_INTERLOCK; /* Interlock is always dropped. */
1606 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1607 ("LK_RETRY set with incompatible flags (0x%x) or an error occured (%d)",
1610 * Callers specify LK_RETRY if they wish to get dead vnodes.
1611 * If RETRY is not set, we return ENOENT instead.
1613 if (error == 0 && vp->v_iflag & VI_DOOMED &&
1614 (flags & LK_RETRY) == 0) {
1619 } while (flags & LK_RETRY && error != 0);
1624 * File table vnode close routine.
1627 vn_closefile(fp, td)
1636 fp->f_ops = &badfileops;
1638 if (fp->f_type == DTYPE_VNODE && fp->f_flag & FHASLOCK)
1641 error = vn_close(vp, fp->f_flag, fp->f_cred, td);
1643 if (fp->f_type == DTYPE_VNODE && fp->f_flag & FHASLOCK) {
1644 lf.l_whence = SEEK_SET;
1647 lf.l_type = F_UNLCK;
1648 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1655 vn_suspendable(struct mount *mp)
1658 return (mp->mnt_op->vfs_susp_clean != NULL);
1662 * Preparing to start a filesystem write operation. If the operation is
1663 * permitted, then we bump the count of operations in progress and
1664 * proceed. If a suspend request is in progress, we wait until the
1665 * suspension is over, and then proceed.
1668 vn_start_write_locked(struct mount *mp, int flags)
1672 mtx_assert(MNT_MTX(mp), MA_OWNED);
1676 * Check on status of suspension.
1678 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1679 mp->mnt_susp_owner != curthread) {
1680 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1681 (flags & PCATCH) : 0) | (PUSER - 1);
1682 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1683 if (flags & V_NOWAIT) {
1684 error = EWOULDBLOCK;
1687 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1693 if (flags & V_XSLEEP)
1695 mp->mnt_writeopcount++;
1697 if (error != 0 || (flags & V_XSLEEP) != 0)
1704 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1709 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1710 ("V_MNTREF requires mp"));
1714 * If a vnode is provided, get and return the mount point that
1715 * to which it will write.
1718 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1720 if (error != EOPNOTSUPP)
1725 if ((mp = *mpp) == NULL)
1728 if (!vn_suspendable(mp)) {
1729 if (vp != NULL || (flags & V_MNTREF) != 0)
1735 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1737 * As long as a vnode is not provided we need to acquire a
1738 * refcount for the provided mountpoint too, in order to
1739 * emulate a vfs_ref().
1742 if (vp == NULL && (flags & V_MNTREF) == 0)
1745 return (vn_start_write_locked(mp, flags));
1749 * Secondary suspension. Used by operations such as vop_inactive
1750 * routines that are needed by the higher level functions. These
1751 * are allowed to proceed until all the higher level functions have
1752 * completed (indicated by mnt_writeopcount dropping to zero). At that
1753 * time, these operations are halted until the suspension is over.
1756 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1761 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1762 ("V_MNTREF requires mp"));
1766 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1768 if (error != EOPNOTSUPP)
1774 * If we are not suspended or have not yet reached suspended
1775 * mode, then let the operation proceed.
1777 if ((mp = *mpp) == NULL)
1780 if (!vn_suspendable(mp)) {
1781 if (vp != NULL || (flags & V_MNTREF) != 0)
1787 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1789 * As long as a vnode is not provided we need to acquire a
1790 * refcount for the provided mountpoint too, in order to
1791 * emulate a vfs_ref().
1794 if (vp == NULL && (flags & V_MNTREF) == 0)
1796 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1797 mp->mnt_secondary_writes++;
1798 mp->mnt_secondary_accwrites++;
1802 if (flags & V_NOWAIT) {
1805 return (EWOULDBLOCK);
1808 * Wait for the suspension to finish.
1810 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1811 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1820 * Filesystem write operation has completed. If we are suspending and this
1821 * operation is the last one, notify the suspender that the suspension is
1825 vn_finished_write(mp)
1828 if (mp == NULL || !vn_suspendable(mp))
1832 mp->mnt_writeopcount--;
1833 if (mp->mnt_writeopcount < 0)
1834 panic("vn_finished_write: neg cnt");
1835 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1836 mp->mnt_writeopcount <= 0)
1837 wakeup(&mp->mnt_writeopcount);
1843 * Filesystem secondary write operation has completed. If we are
1844 * suspending and this operation is the last one, notify the suspender
1845 * that the suspension is now in effect.
1848 vn_finished_secondary_write(mp)
1851 if (mp == NULL || !vn_suspendable(mp))
1855 mp->mnt_secondary_writes--;
1856 if (mp->mnt_secondary_writes < 0)
1857 panic("vn_finished_secondary_write: neg cnt");
1858 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1859 mp->mnt_secondary_writes <= 0)
1860 wakeup(&mp->mnt_secondary_writes);
1867 * Request a filesystem to suspend write operations.
1870 vfs_write_suspend(struct mount *mp, int flags)
1874 MPASS(vn_suspendable(mp));
1877 if (mp->mnt_susp_owner == curthread) {
1881 while (mp->mnt_kern_flag & MNTK_SUSPEND)
1882 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1885 * Unmount holds a write reference on the mount point. If we
1886 * own busy reference and drain for writers, we deadlock with
1887 * the reference draining in the unmount path. Callers of
1888 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1889 * vfs_busy() reference is owned and caller is not in the
1892 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1893 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1898 mp->mnt_kern_flag |= MNTK_SUSPEND;
1899 mp->mnt_susp_owner = curthread;
1900 if (mp->mnt_writeopcount > 0)
1901 (void) msleep(&mp->mnt_writeopcount,
1902 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1905 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0)
1906 vfs_write_resume(mp, 0);
1911 * Request a filesystem to resume write operations.
1914 vfs_write_resume(struct mount *mp, int flags)
1917 MPASS(vn_suspendable(mp));
1920 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1921 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
1922 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
1924 mp->mnt_susp_owner = NULL;
1925 wakeup(&mp->mnt_writeopcount);
1926 wakeup(&mp->mnt_flag);
1927 curthread->td_pflags &= ~TDP_IGNSUSP;
1928 if ((flags & VR_START_WRITE) != 0) {
1930 mp->mnt_writeopcount++;
1933 if ((flags & VR_NO_SUSPCLR) == 0)
1935 } else if ((flags & VR_START_WRITE) != 0) {
1937 vn_start_write_locked(mp, 0);
1944 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
1948 vfs_write_suspend_umnt(struct mount *mp)
1952 MPASS(vn_suspendable(mp));
1953 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
1954 ("vfs_write_suspend_umnt: recursed"));
1956 /* dounmount() already called vn_start_write(). */
1958 vn_finished_write(mp);
1959 error = vfs_write_suspend(mp, 0);
1961 vn_start_write(NULL, &mp, V_WAIT);
1965 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
1968 vn_start_write(NULL, &mp, V_WAIT);
1970 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
1971 wakeup(&mp->mnt_flag);
1973 curthread->td_pflags |= TDP_IGNSUSP;
1978 * Implement kqueues for files by translating it to vnode operation.
1981 vn_kqfilter(struct file *fp, struct knote *kn)
1984 return (VOP_KQFILTER(fp->f_vnode, kn));
1988 * Simplified in-kernel wrapper calls for extended attribute access.
1989 * Both calls pass in a NULL credential, authorizing as "kernel" access.
1990 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
1993 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
1994 const char *attrname, int *buflen, char *buf, struct thread *td)
2000 iov.iov_len = *buflen;
2003 auio.uio_iov = &iov;
2004 auio.uio_iovcnt = 1;
2005 auio.uio_rw = UIO_READ;
2006 auio.uio_segflg = UIO_SYSSPACE;
2008 auio.uio_offset = 0;
2009 auio.uio_resid = *buflen;
2011 if ((ioflg & IO_NODELOCKED) == 0)
2012 vn_lock(vp, LK_SHARED | LK_RETRY);
2014 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2016 /* authorize attribute retrieval as kernel */
2017 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2020 if ((ioflg & IO_NODELOCKED) == 0)
2024 *buflen = *buflen - auio.uio_resid;
2031 * XXX failure mode if partially written?
2034 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2035 const char *attrname, int buflen, char *buf, struct thread *td)
2042 iov.iov_len = buflen;
2045 auio.uio_iov = &iov;
2046 auio.uio_iovcnt = 1;
2047 auio.uio_rw = UIO_WRITE;
2048 auio.uio_segflg = UIO_SYSSPACE;
2050 auio.uio_offset = 0;
2051 auio.uio_resid = buflen;
2053 if ((ioflg & IO_NODELOCKED) == 0) {
2054 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2056 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2059 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2061 /* authorize attribute setting as kernel */
2062 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2064 if ((ioflg & IO_NODELOCKED) == 0) {
2065 vn_finished_write(mp);
2073 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2074 const char *attrname, struct thread *td)
2079 if ((ioflg & IO_NODELOCKED) == 0) {
2080 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2082 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2085 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2087 /* authorize attribute removal as kernel */
2088 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2089 if (error == EOPNOTSUPP)
2090 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2093 if ((ioflg & IO_NODELOCKED) == 0) {
2094 vn_finished_write(mp);
2102 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2106 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2110 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2113 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2118 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2119 int lkflags, struct vnode **rvp)
2124 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2126 ltype = VOP_ISLOCKED(vp);
2127 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2128 ("vn_vget_ino: vp not locked"));
2129 error = vfs_busy(mp, MBF_NOWAIT);
2133 error = vfs_busy(mp, 0);
2134 vn_lock(vp, ltype | LK_RETRY);
2138 if (vp->v_iflag & VI_DOOMED) {
2144 error = alloc(mp, alloc_arg, lkflags, rvp);
2147 vn_lock(vp, ltype | LK_RETRY);
2148 if (vp->v_iflag & VI_DOOMED) {
2161 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2165 if (vp->v_type != VREG || td == NULL)
2167 if ((uoff_t)uio->uio_offset + uio->uio_resid >
2168 lim_cur(td, RLIMIT_FSIZE)) {
2169 PROC_LOCK(td->td_proc);
2170 kern_psignal(td->td_proc, SIGXFSZ);
2171 PROC_UNLOCK(td->td_proc);
2178 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2185 vn_lock(vp, LK_SHARED | LK_RETRY);
2186 AUDIT_ARG_VNODE1(vp);
2189 return (setfmode(td, active_cred, vp, mode));
2193 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2200 vn_lock(vp, LK_SHARED | LK_RETRY);
2201 AUDIT_ARG_VNODE1(vp);
2204 return (setfown(td, active_cred, vp, uid, gid));
2208 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2212 if ((object = vp->v_object) == NULL)
2214 VM_OBJECT_WLOCK(object);
2215 vm_object_page_remove(object, start, end, 0);
2216 VM_OBJECT_WUNLOCK(object);
2220 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2228 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2229 ("Wrong command %lu", cmd));
2231 if (vn_lock(vp, LK_SHARED) != 0)
2233 if (vp->v_type != VREG) {
2237 error = VOP_GETATTR(vp, &va, cred);
2241 if (noff >= va.va_size) {
2245 bsize = vp->v_mount->mnt_stat.f_iosize;
2246 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize) {
2247 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2248 if (error == EOPNOTSUPP) {
2252 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2253 (bnp != -1 && cmd == FIOSEEKDATA)) {
2260 if (noff > va.va_size)
2262 /* noff == va.va_size. There is an implicit hole at the end of file. */
2263 if (cmd == FIOSEEKDATA)
2273 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2278 off_t foffset, size;
2281 cred = td->td_ucred;
2283 foffset = foffset_lock(fp, 0);
2284 noneg = (vp->v_type != VCHR);
2290 (offset > 0 && foffset > OFF_MAX - offset))) {
2297 vn_lock(vp, LK_SHARED | LK_RETRY);
2298 error = VOP_GETATTR(vp, &vattr, cred);
2304 * If the file references a disk device, then fetch
2305 * the media size and use that to determine the ending
2308 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2309 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2310 vattr.va_size = size;
2312 (vattr.va_size > OFF_MAX ||
2313 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2317 offset += vattr.va_size;
2322 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2325 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2330 if (error == 0 && noneg && offset < 0)
2334 VFS_KNOTE_UNLOCKED(vp, 0);
2335 td->td_uretoff.tdu_off = offset;
2337 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2342 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2348 * Grant permission if the caller is the owner of the file, or
2349 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2350 * on the file. If the time pointer is null, then write
2351 * permission on the file is also sufficient.
2353 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2354 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2355 * will be allowed to set the times [..] to the current
2358 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2359 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2360 error = VOP_ACCESS(vp, VWRITE, cred, td);
2365 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2370 if (fp->f_type == DTYPE_FIFO)
2371 kif->kf_type = KF_TYPE_FIFO;
2373 kif->kf_type = KF_TYPE_VNODE;
2376 FILEDESC_SUNLOCK(fdp);
2377 error = vn_fill_kinfo_vnode(vp, kif);
2379 FILEDESC_SLOCK(fdp);
2384 vn_fill_junk(struct kinfo_file *kif)
2389 * Simulate vn_fullpath returning changing values for a given
2390 * vp during e.g. coredump.
2392 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2393 olen = strlen(kif->kf_path);
2395 strcpy(&kif->kf_path[len - 1], "$");
2397 for (; olen < len; olen++)
2398 strcpy(&kif->kf_path[olen], "A");
2402 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2405 char *fullpath, *freepath;
2408 kif->kf_vnode_type = vntype_to_kinfo(vp->v_type);
2411 error = vn_fullpath(curthread, vp, &fullpath, &freepath);
2413 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2415 if (freepath != NULL)
2416 free(freepath, M_TEMP);
2418 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2423 * Retrieve vnode attributes.
2425 va.va_fsid = VNOVAL;
2427 vn_lock(vp, LK_SHARED | LK_RETRY);
2428 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2432 if (va.va_fsid != VNOVAL)
2433 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2435 kif->kf_un.kf_file.kf_file_fsid =
2436 vp->v_mount->mnt_stat.f_fsid.val[0];
2437 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2438 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2439 kif->kf_un.kf_file.kf_file_size = va.va_size;
2440 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2445 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2446 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2450 struct pmckern_map_in pkm;
2456 boolean_t writecounted;
2459 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2460 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2462 * POSIX shared-memory objects are defined to have
2463 * kernel persistence, and are not defined to support
2464 * read(2)/write(2) -- or even open(2). Thus, we can
2465 * use MAP_ASYNC to trade on-disk coherence for speed.
2466 * The shm_open(3) library routine turns on the FPOSIXSHM
2467 * flag to request this behavior.
2469 if ((fp->f_flag & FPOSIXSHM) != 0)
2470 flags |= MAP_NOSYNC;
2475 * Ensure that file and memory protections are
2476 * compatible. Note that we only worry about
2477 * writability if mapping is shared; in this case,
2478 * current and max prot are dictated by the open file.
2479 * XXX use the vnode instead? Problem is: what
2480 * credentials do we use for determination? What if
2481 * proc does a setuid?
2484 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0)
2485 maxprot = VM_PROT_NONE;
2487 maxprot = VM_PROT_EXECUTE;
2488 if ((fp->f_flag & FREAD) != 0)
2489 maxprot |= VM_PROT_READ;
2490 else if ((prot & VM_PROT_READ) != 0)
2494 * If we are sharing potential changes via MAP_SHARED and we
2495 * are trying to get write permission although we opened it
2496 * without asking for it, bail out.
2498 if ((flags & MAP_SHARED) != 0) {
2499 if ((fp->f_flag & FWRITE) != 0)
2500 maxprot |= VM_PROT_WRITE;
2501 else if ((prot & VM_PROT_WRITE) != 0)
2504 maxprot |= VM_PROT_WRITE;
2505 cap_maxprot |= VM_PROT_WRITE;
2507 maxprot &= cap_maxprot;
2509 writecounted = FALSE;
2510 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2511 &foff, &object, &writecounted);
2514 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2515 foff, writecounted, td);
2518 * If this mapping was accounted for in the vnode's
2519 * writecount, then undo that now.
2522 vnode_pager_release_writecount(object, 0, size);
2523 vm_object_deallocate(object);
2526 /* Inform hwpmc(4) if an executable is being mapped. */
2527 if (error == 0 && (prot & VM_PROT_EXECUTE) != 0) {
2529 pkm.pm_address = (uintptr_t) addr;
2530 PMC_CALL_HOOK(td, PMC_FN_MMAP, (void *) &pkm);