2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1982, 1986, 1989, 1993
5 * The Regents of the University of California. All rights reserved.
6 * (c) UNIX System Laboratories, Inc.
7 * All or some portions of this file are derived from material licensed
8 * to the University of California by American Telephone and Telegraph
9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10 * the permission of UNIX System Laboratories, Inc.
12 * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13 * Copyright (c) 2013, 2014 The FreeBSD Foundation
15 * Portions of this software were developed by Konstantin Belousov
16 * under sponsorship from the FreeBSD Foundation.
18 * Redistribution and use in source and binary forms, with or without
19 * modification, are permitted provided that the following conditions
21 * 1. Redistributions of source code must retain the above copyright
22 * notice, this list of conditions and the following disclaimer.
23 * 2. Redistributions in binary form must reproduce the above copyright
24 * notice, this list of conditions and the following disclaimer in the
25 * documentation and/or other materials provided with the distribution.
26 * 3. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42 * @(#)vfs_vnops.c 8.2 (Berkeley) 1/21/94
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
48 #include "opt_hwpmc_hooks.h"
50 #include <sys/param.h>
51 #include <sys/systm.h>
54 #include <sys/fcntl.h>
61 #include <sys/limits.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
74 #include <sys/sysctl.h>
75 #include <sys/ttycom.h>
77 #include <sys/syslog.h>
78 #include <sys/unistd.h>
81 #include <security/audit/audit.h>
82 #include <security/mac/mac_framework.h>
85 #include <vm/vm_extern.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pager.h>
93 #include <sys/pmckern.h>
96 static fo_rdwr_t vn_read;
97 static fo_rdwr_t vn_write;
98 static fo_rdwr_t vn_io_fault;
99 static fo_truncate_t vn_truncate;
100 static fo_ioctl_t vn_ioctl;
101 static fo_poll_t vn_poll;
102 static fo_kqfilter_t vn_kqfilter;
103 static fo_stat_t vn_statfile;
104 static fo_close_t vn_closefile;
105 static fo_mmap_t vn_mmap;
106 static fo_fallocate_t vn_fallocate;
108 struct fileops vnops = {
109 .fo_read = vn_io_fault,
110 .fo_write = vn_io_fault,
111 .fo_truncate = vn_truncate,
112 .fo_ioctl = vn_ioctl,
114 .fo_kqfilter = vn_kqfilter,
115 .fo_stat = vn_statfile,
116 .fo_close = vn_closefile,
117 .fo_chmod = vn_chmod,
118 .fo_chown = vn_chown,
119 .fo_sendfile = vn_sendfile,
121 .fo_fill_kinfo = vn_fill_kinfo,
123 .fo_fallocate = vn_fallocate,
124 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
127 static const int io_hold_cnt = 16;
128 static int vn_io_fault_enable = 1;
129 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW,
130 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
131 static int vn_io_fault_prefault = 0;
132 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RW,
133 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
134 static u_long vn_io_faults_cnt;
135 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
136 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
139 * Returns true if vn_io_fault mode of handling the i/o request should
143 do_vn_io_fault(struct vnode *vp, struct uio *uio)
147 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
148 (mp = vp->v_mount) != NULL &&
149 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
153 * Structure used to pass arguments to vn_io_fault1(), to do either
154 * file- or vnode-based I/O calls.
156 struct vn_io_fault_args {
164 struct fop_args_tag {
168 struct vop_args_tag {
174 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
175 struct vn_io_fault_args *args, struct thread *td);
178 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
180 struct thread *td = ndp->ni_cnd.cn_thread;
182 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
186 * Common code for vnode open operations via a name lookup.
187 * Lookup the vnode and invoke VOP_CREATE if needed.
188 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
190 * Note that this does NOT free nameidata for the successful case,
191 * due to the NDINIT being done elsewhere.
194 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
195 struct ucred *cred, struct file *fp)
199 struct thread *td = ndp->ni_cnd.cn_thread;
201 struct vattr *vap = &vat;
206 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
207 O_EXCL | O_DIRECTORY))
209 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
210 ndp->ni_cnd.cn_nameiop = CREATE;
212 * Set NOCACHE to avoid flushing the cache when
213 * rolling in many files at once.
215 ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE;
216 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
217 ndp->ni_cnd.cn_flags |= FOLLOW;
218 if ((fmode & O_BENEATH) != 0)
219 ndp->ni_cnd.cn_flags |= BENEATH;
220 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
221 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
222 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
223 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
224 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
226 if ((error = namei(ndp)) != 0)
228 if (ndp->ni_vp == NULL) {
231 vap->va_mode = cmode;
233 vap->va_vaflags |= VA_EXCLUSIVE;
234 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
235 NDFREE(ndp, NDF_ONLY_PNBUF);
237 if ((error = vn_start_write(NULL, &mp,
238 V_XSLEEP | PCATCH)) != 0)
242 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
243 ndp->ni_cnd.cn_flags |= MAKEENTRY;
245 error = mac_vnode_check_create(cred, ndp->ni_dvp,
249 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
252 vn_finished_write(mp);
254 NDFREE(ndp, NDF_ONLY_PNBUF);
260 if (ndp->ni_dvp == ndp->ni_vp)
266 if (fmode & O_EXCL) {
270 if (vp->v_type == VDIR) {
277 ndp->ni_cnd.cn_nameiop = LOOKUP;
278 ndp->ni_cnd.cn_flags = ISOPEN |
279 ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
280 if (!(fmode & FWRITE))
281 ndp->ni_cnd.cn_flags |= LOCKSHARED;
282 if ((fmode & O_BENEATH) != 0)
283 ndp->ni_cnd.cn_flags |= BENEATH;
284 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
285 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
286 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
287 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
288 if ((error = namei(ndp)) != 0)
292 error = vn_open_vnode(vp, fmode, cred, td, fp);
298 NDFREE(ndp, NDF_ONLY_PNBUF);
306 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
309 int error, lock_flags, type;
311 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
312 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
314 KASSERT(fp != NULL, ("open with flock requires fp"));
315 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
318 lock_flags = VOP_ISLOCKED(vp);
321 lf.l_whence = SEEK_SET;
324 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
326 if ((fmode & FNONBLOCK) == 0)
328 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
330 fp->f_flag |= FHASLOCK;
332 vn_lock(vp, lock_flags | LK_RETRY);
333 if (error == 0 && VN_IS_DOOMED(vp))
339 * Common code for vnode open operations once a vnode is located.
340 * Check permissions, and call the VOP_OPEN routine.
343 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
344 struct thread *td, struct file *fp)
349 if (vp->v_type == VLNK)
351 if (vp->v_type == VSOCK)
353 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
356 if (fmode & (FWRITE | O_TRUNC)) {
357 if (vp->v_type == VDIR)
365 if ((fmode & O_APPEND) && (fmode & FWRITE))
370 if (fmode & O_VERIFY)
372 error = mac_vnode_check_open(cred, vp, accmode);
376 accmode &= ~(VCREAT | VVERIFY);
378 if ((fmode & O_CREAT) == 0 && accmode != 0) {
379 error = VOP_ACCESS(vp, accmode, cred, td);
383 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
384 vn_lock(vp, LK_UPGRADE | LK_RETRY);
385 error = VOP_OPEN(vp, fmode, cred, td, fp);
389 error = vn_open_vnode_advlock(vp, fmode, fp);
390 if (error == 0 && (fmode & FWRITE) != 0) {
391 error = VOP_ADD_WRITECOUNT(vp, 1);
393 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
394 __func__, vp, vp->v_writecount);
399 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
400 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
401 * Arrange for that by having fdrop() to use vn_closefile().
404 fp->f_flag |= FOPENFAILED;
406 if (fp->f_ops == &badfileops) {
407 fp->f_type = DTYPE_VNODE;
413 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
419 * Check for write permissions on the specified vnode.
420 * Prototype text segments cannot be written.
424 vn_writechk(struct vnode *vp)
427 ASSERT_VOP_LOCKED(vp, "vn_writechk");
429 * If there's shared text associated with
430 * the vnode, try to free it up once. If
431 * we fail, we can't allow writing.
443 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
444 struct thread *td, bool keep_ref)
447 int error, lock_flags;
449 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
450 MNT_EXTENDED_SHARED(vp->v_mount))
451 lock_flags = LK_SHARED;
453 lock_flags = LK_EXCLUSIVE;
455 vn_start_write(vp, &mp, V_WAIT);
456 vn_lock(vp, lock_flags | LK_RETRY);
457 AUDIT_ARG_VNODE1(vp);
458 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
459 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
460 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
461 __func__, vp, vp->v_writecount);
463 error = VOP_CLOSE(vp, flags, file_cred, td);
468 vn_finished_write(mp);
473 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
477 return (vn_close1(vp, flags, file_cred, td, false));
481 * Heuristic to detect sequential operation.
484 sequential_heuristic(struct uio *uio, struct file *fp)
487 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
488 if (fp->f_flag & FRDAHEAD)
489 return (fp->f_seqcount << IO_SEQSHIFT);
492 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
493 * that the first I/O is normally considered to be slightly
494 * sequential. Seeking to offset 0 doesn't change sequentiality
495 * unless previous seeks have reduced f_seqcount to 0, in which
496 * case offset 0 is not special.
498 if ((uio->uio_offset == 0 && fp->f_seqcount > 0) ||
499 uio->uio_offset == fp->f_nextoff) {
501 * f_seqcount is in units of fixed-size blocks so that it
502 * depends mainly on the amount of sequential I/O and not
503 * much on the number of sequential I/O's. The fixed size
504 * of 16384 is hard-coded here since it is (not quite) just
505 * a magic size that works well here. This size is more
506 * closely related to the best I/O size for real disks than
507 * to any block size used by software.
509 if (uio->uio_resid >= IO_SEQMAX * 16384)
510 fp->f_seqcount = IO_SEQMAX;
512 fp->f_seqcount += howmany(uio->uio_resid, 16384);
513 if (fp->f_seqcount > IO_SEQMAX)
514 fp->f_seqcount = IO_SEQMAX;
516 return (fp->f_seqcount << IO_SEQSHIFT);
519 /* Not sequential. Quickly draw-down sequentiality. */
520 if (fp->f_seqcount > 1)
528 * Package up an I/O request on a vnode into a uio and do it.
531 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
532 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
533 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
540 struct vn_io_fault_args args;
541 int error, lock_flags;
543 if (offset < 0 && vp->v_type != VCHR)
545 auio.uio_iov = &aiov;
547 aiov.iov_base = base;
549 auio.uio_resid = len;
550 auio.uio_offset = offset;
551 auio.uio_segflg = segflg;
556 if ((ioflg & IO_NODELOCKED) == 0) {
557 if ((ioflg & IO_RANGELOCKED) == 0) {
558 if (rw == UIO_READ) {
559 rl_cookie = vn_rangelock_rlock(vp, offset,
562 rl_cookie = vn_rangelock_wlock(vp, offset,
568 if (rw == UIO_WRITE) {
569 if (vp->v_type != VCHR &&
570 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
573 if (MNT_SHARED_WRITES(mp) ||
574 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
575 lock_flags = LK_SHARED;
577 lock_flags = LK_EXCLUSIVE;
579 lock_flags = LK_SHARED;
580 vn_lock(vp, lock_flags | LK_RETRY);
584 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
586 if ((ioflg & IO_NOMACCHECK) == 0) {
588 error = mac_vnode_check_read(active_cred, file_cred,
591 error = mac_vnode_check_write(active_cred, file_cred,
596 if (file_cred != NULL)
600 if (do_vn_io_fault(vp, &auio)) {
601 args.kind = VN_IO_FAULT_VOP;
604 args.args.vop_args.vp = vp;
605 error = vn_io_fault1(vp, &auio, &args, td);
606 } else if (rw == UIO_READ) {
607 error = VOP_READ(vp, &auio, ioflg, cred);
608 } else /* if (rw == UIO_WRITE) */ {
609 error = VOP_WRITE(vp, &auio, ioflg, cred);
613 *aresid = auio.uio_resid;
615 if (auio.uio_resid && error == 0)
617 if ((ioflg & IO_NODELOCKED) == 0) {
620 vn_finished_write(mp);
623 if (rl_cookie != NULL)
624 vn_rangelock_unlock(vp, rl_cookie);
629 * Package up an I/O request on a vnode into a uio and do it. The I/O
630 * request is split up into smaller chunks and we try to avoid saturating
631 * the buffer cache while potentially holding a vnode locked, so we
632 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
633 * to give other processes a chance to lock the vnode (either other processes
634 * core'ing the same binary, or unrelated processes scanning the directory).
637 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
638 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
639 struct ucred *file_cred, size_t *aresid, struct thread *td)
648 * Force `offset' to a multiple of MAXBSIZE except possibly
649 * for the first chunk, so that filesystems only need to
650 * write full blocks except possibly for the first and last
653 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
657 if (rw != UIO_READ && vp->v_type == VREG)
660 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
661 ioflg, active_cred, file_cred, &iaresid, td);
662 len -= chunk; /* aresid calc already includes length */
666 base = (char *)base + chunk;
667 kern_yield(PRI_USER);
670 *aresid = len + iaresid;
675 foffset_lock(struct file *fp, int flags)
680 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
682 #if OFF_MAX <= LONG_MAX
684 * Caller only wants the current f_offset value. Assume that
685 * the long and shorter integer types reads are atomic.
687 if ((flags & FOF_NOLOCK) != 0)
688 return (fp->f_offset);
692 * According to McKusick the vn lock was protecting f_offset here.
693 * It is now protected by the FOFFSET_LOCKED flag.
695 mtxp = mtx_pool_find(mtxpool_sleep, fp);
697 if ((flags & FOF_NOLOCK) == 0) {
698 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
699 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
700 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
703 fp->f_vnread_flags |= FOFFSET_LOCKED;
711 foffset_unlock(struct file *fp, off_t val, int flags)
715 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
717 #if OFF_MAX <= LONG_MAX
718 if ((flags & FOF_NOLOCK) != 0) {
719 if ((flags & FOF_NOUPDATE) == 0)
721 if ((flags & FOF_NEXTOFF) != 0)
727 mtxp = mtx_pool_find(mtxpool_sleep, fp);
729 if ((flags & FOF_NOUPDATE) == 0)
731 if ((flags & FOF_NEXTOFF) != 0)
733 if ((flags & FOF_NOLOCK) == 0) {
734 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
735 ("Lost FOFFSET_LOCKED"));
736 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
737 wakeup(&fp->f_vnread_flags);
738 fp->f_vnread_flags = 0;
744 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
747 if ((flags & FOF_OFFSET) == 0)
748 uio->uio_offset = foffset_lock(fp, flags);
752 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
755 if ((flags & FOF_OFFSET) == 0)
756 foffset_unlock(fp, uio->uio_offset, flags);
760 get_advice(struct file *fp, struct uio *uio)
765 ret = POSIX_FADV_NORMAL;
766 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
769 mtxp = mtx_pool_find(mtxpool_sleep, fp);
771 if (fp->f_advice != NULL &&
772 uio->uio_offset >= fp->f_advice->fa_start &&
773 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
774 ret = fp->f_advice->fa_advice;
780 * File table vnode read routine.
783 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
791 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
793 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
796 if (fp->f_flag & FNONBLOCK)
798 if (fp->f_flag & O_DIRECT)
800 advice = get_advice(fp, uio);
801 vn_lock(vp, LK_SHARED | LK_RETRY);
804 case POSIX_FADV_NORMAL:
805 case POSIX_FADV_SEQUENTIAL:
806 case POSIX_FADV_NOREUSE:
807 ioflag |= sequential_heuristic(uio, fp);
809 case POSIX_FADV_RANDOM:
810 /* Disable read-ahead for random I/O. */
813 orig_offset = uio->uio_offset;
816 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
819 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
820 fp->f_nextoff = uio->uio_offset;
822 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
823 orig_offset != uio->uio_offset)
825 * Use POSIX_FADV_DONTNEED to flush pages and buffers
826 * for the backing file after a POSIX_FADV_NOREUSE
829 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
830 POSIX_FADV_DONTNEED);
835 * File table vnode write routine.
838 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
844 int error, ioflag, lock_flags;
847 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
849 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
851 if (vp->v_type == VREG)
854 if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
856 if (fp->f_flag & FNONBLOCK)
858 if (fp->f_flag & O_DIRECT)
860 if ((fp->f_flag & O_FSYNC) ||
861 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
864 if (vp->v_type != VCHR &&
865 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
868 advice = get_advice(fp, uio);
870 if (MNT_SHARED_WRITES(mp) ||
871 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
872 lock_flags = LK_SHARED;
874 lock_flags = LK_EXCLUSIVE;
877 vn_lock(vp, lock_flags | LK_RETRY);
879 case POSIX_FADV_NORMAL:
880 case POSIX_FADV_SEQUENTIAL:
881 case POSIX_FADV_NOREUSE:
882 ioflag |= sequential_heuristic(uio, fp);
884 case POSIX_FADV_RANDOM:
885 /* XXX: Is this correct? */
888 orig_offset = uio->uio_offset;
891 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
894 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
895 fp->f_nextoff = uio->uio_offset;
897 if (vp->v_type != VCHR)
898 vn_finished_write(mp);
899 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
900 orig_offset != uio->uio_offset)
902 * Use POSIX_FADV_DONTNEED to flush pages and buffers
903 * for the backing file after a POSIX_FADV_NOREUSE
906 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
907 POSIX_FADV_DONTNEED);
913 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
914 * prevent the following deadlock:
916 * Assume that the thread A reads from the vnode vp1 into userspace
917 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
918 * currently not resident, then system ends up with the call chain
919 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
920 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
921 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
922 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
923 * backed by the pages of vnode vp1, and some page in buf2 is not
924 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
926 * To prevent the lock order reversal and deadlock, vn_io_fault() does
927 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
928 * Instead, it first tries to do the whole range i/o with pagefaults
929 * disabled. If all pages in the i/o buffer are resident and mapped,
930 * VOP will succeed (ignoring the genuine filesystem errors).
931 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
932 * i/o in chunks, with all pages in the chunk prefaulted and held
933 * using vm_fault_quick_hold_pages().
935 * Filesystems using this deadlock avoidance scheme should use the
936 * array of the held pages from uio, saved in the curthread->td_ma,
937 * instead of doing uiomove(). A helper function
938 * vn_io_fault_uiomove() converts uiomove request into
939 * uiomove_fromphys() over td_ma array.
941 * Since vnode locks do not cover the whole i/o anymore, rangelocks
942 * make the current i/o request atomic with respect to other i/os and
947 * Decode vn_io_fault_args and perform the corresponding i/o.
950 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
956 save = vm_fault_disable_pagefaults();
957 switch (args->kind) {
958 case VN_IO_FAULT_FOP:
959 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
960 uio, args->cred, args->flags, td);
962 case VN_IO_FAULT_VOP:
963 if (uio->uio_rw == UIO_READ) {
964 error = VOP_READ(args->args.vop_args.vp, uio,
965 args->flags, args->cred);
966 } else if (uio->uio_rw == UIO_WRITE) {
967 error = VOP_WRITE(args->args.vop_args.vp, uio,
968 args->flags, args->cred);
972 panic("vn_io_fault_doio: unknown kind of io %d %d",
973 args->kind, uio->uio_rw);
975 vm_fault_enable_pagefaults(save);
980 vn_io_fault_touch(char *base, const struct uio *uio)
985 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
991 vn_io_fault_prefault_user(const struct uio *uio)
994 const struct iovec *iov;
999 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1000 ("vn_io_fault_prefault userspace"));
1004 resid = uio->uio_resid;
1005 base = iov->iov_base;
1008 error = vn_io_fault_touch(base, uio);
1011 if (len < PAGE_SIZE) {
1013 error = vn_io_fault_touch(base + len - 1, uio);
1018 if (++i >= uio->uio_iovcnt)
1020 iov = uio->uio_iov + i;
1021 base = iov->iov_base;
1033 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1034 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1035 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1036 * into args and call vn_io_fault1() to handle faults during the user
1037 * mode buffer accesses.
1040 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1043 vm_page_t ma[io_hold_cnt + 2];
1044 struct uio *uio_clone, short_uio;
1045 struct iovec short_iovec[1];
1046 vm_page_t *prev_td_ma;
1048 vm_offset_t addr, end;
1051 int error, cnt, saveheld, prev_td_ma_cnt;
1053 if (vn_io_fault_prefault) {
1054 error = vn_io_fault_prefault_user(uio);
1056 return (error); /* Or ignore ? */
1059 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1062 * The UFS follows IO_UNIT directive and replays back both
1063 * uio_offset and uio_resid if an error is encountered during the
1064 * operation. But, since the iovec may be already advanced,
1065 * uio is still in an inconsistent state.
1067 * Cache a copy of the original uio, which is advanced to the redo
1068 * point using UIO_NOCOPY below.
1070 uio_clone = cloneuio(uio);
1071 resid = uio->uio_resid;
1073 short_uio.uio_segflg = UIO_USERSPACE;
1074 short_uio.uio_rw = uio->uio_rw;
1075 short_uio.uio_td = uio->uio_td;
1077 error = vn_io_fault_doio(args, uio, td);
1078 if (error != EFAULT)
1081 atomic_add_long(&vn_io_faults_cnt, 1);
1082 uio_clone->uio_segflg = UIO_NOCOPY;
1083 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1084 uio_clone->uio_segflg = uio->uio_segflg;
1086 saveheld = curthread_pflags_set(TDP_UIOHELD);
1087 prev_td_ma = td->td_ma;
1088 prev_td_ma_cnt = td->td_ma_cnt;
1090 while (uio_clone->uio_resid != 0) {
1091 len = uio_clone->uio_iov->iov_len;
1093 KASSERT(uio_clone->uio_iovcnt >= 1,
1094 ("iovcnt underflow"));
1095 uio_clone->uio_iov++;
1096 uio_clone->uio_iovcnt--;
1099 if (len > io_hold_cnt * PAGE_SIZE)
1100 len = io_hold_cnt * PAGE_SIZE;
1101 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1102 end = round_page(addr + len);
1107 cnt = atop(end - trunc_page(addr));
1109 * A perfectly misaligned address and length could cause
1110 * both the start and the end of the chunk to use partial
1111 * page. +2 accounts for such a situation.
1113 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1114 addr, len, prot, ma, io_hold_cnt + 2);
1119 short_uio.uio_iov = &short_iovec[0];
1120 short_iovec[0].iov_base = (void *)addr;
1121 short_uio.uio_iovcnt = 1;
1122 short_uio.uio_resid = short_iovec[0].iov_len = len;
1123 short_uio.uio_offset = uio_clone->uio_offset;
1125 td->td_ma_cnt = cnt;
1127 error = vn_io_fault_doio(args, &short_uio, td);
1128 vm_page_unhold_pages(ma, cnt);
1129 adv = len - short_uio.uio_resid;
1131 uio_clone->uio_iov->iov_base =
1132 (char *)uio_clone->uio_iov->iov_base + adv;
1133 uio_clone->uio_iov->iov_len -= adv;
1134 uio_clone->uio_resid -= adv;
1135 uio_clone->uio_offset += adv;
1137 uio->uio_resid -= adv;
1138 uio->uio_offset += adv;
1140 if (error != 0 || adv == 0)
1143 td->td_ma = prev_td_ma;
1144 td->td_ma_cnt = prev_td_ma_cnt;
1145 curthread_pflags_restore(saveheld);
1147 free(uio_clone, M_IOV);
1152 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1153 int flags, struct thread *td)
1158 struct vn_io_fault_args args;
1161 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1163 foffset_lock_uio(fp, uio, flags);
1164 if (do_vn_io_fault(vp, uio)) {
1165 args.kind = VN_IO_FAULT_FOP;
1166 args.args.fop_args.fp = fp;
1167 args.args.fop_args.doio = doio;
1168 args.cred = active_cred;
1169 args.flags = flags | FOF_OFFSET;
1170 if (uio->uio_rw == UIO_READ) {
1171 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1172 uio->uio_offset + uio->uio_resid);
1173 } else if ((fp->f_flag & O_APPEND) != 0 ||
1174 (flags & FOF_OFFSET) == 0) {
1175 /* For appenders, punt and lock the whole range. */
1176 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1178 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1179 uio->uio_offset + uio->uio_resid);
1181 error = vn_io_fault1(vp, uio, &args, td);
1182 vn_rangelock_unlock(vp, rl_cookie);
1184 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1186 foffset_unlock_uio(fp, uio, flags);
1191 * Helper function to perform the requested uiomove operation using
1192 * the held pages for io->uio_iov[0].iov_base buffer instead of
1193 * copyin/copyout. Access to the pages with uiomove_fromphys()
1194 * instead of iov_base prevents page faults that could occur due to
1195 * pmap_collect() invalidating the mapping created by
1196 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1197 * object cleanup revoking the write access from page mappings.
1199 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1200 * instead of plain uiomove().
1203 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1205 struct uio transp_uio;
1206 struct iovec transp_iov[1];
1212 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1213 uio->uio_segflg != UIO_USERSPACE)
1214 return (uiomove(data, xfersize, uio));
1216 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1217 transp_iov[0].iov_base = data;
1218 transp_uio.uio_iov = &transp_iov[0];
1219 transp_uio.uio_iovcnt = 1;
1220 if (xfersize > uio->uio_resid)
1221 xfersize = uio->uio_resid;
1222 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1223 transp_uio.uio_offset = 0;
1224 transp_uio.uio_segflg = UIO_SYSSPACE;
1226 * Since transp_iov points to data, and td_ma page array
1227 * corresponds to original uio->uio_iov, we need to invert the
1228 * direction of the i/o operation as passed to
1229 * uiomove_fromphys().
1231 switch (uio->uio_rw) {
1233 transp_uio.uio_rw = UIO_READ;
1236 transp_uio.uio_rw = UIO_WRITE;
1239 transp_uio.uio_td = uio->uio_td;
1240 error = uiomove_fromphys(td->td_ma,
1241 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1242 xfersize, &transp_uio);
1243 adv = xfersize - transp_uio.uio_resid;
1245 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1246 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1248 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1250 td->td_ma_cnt -= pgadv;
1251 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1252 uio->uio_iov->iov_len -= adv;
1253 uio->uio_resid -= adv;
1254 uio->uio_offset += adv;
1259 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1263 vm_offset_t iov_base;
1267 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1268 uio->uio_segflg != UIO_USERSPACE)
1269 return (uiomove_fromphys(ma, offset, xfersize, uio));
1271 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1272 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1273 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1274 switch (uio->uio_rw) {
1276 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1280 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1284 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1286 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1288 td->td_ma_cnt -= pgadv;
1289 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1290 uio->uio_iov->iov_len -= cnt;
1291 uio->uio_resid -= cnt;
1292 uio->uio_offset += cnt;
1298 * File table truncate routine.
1301 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1312 * Lock the whole range for truncation. Otherwise split i/o
1313 * might happen partly before and partly after the truncation.
1315 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1316 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1319 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1320 AUDIT_ARG_VNODE1(vp);
1321 if (vp->v_type == VDIR) {
1326 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1330 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1334 vn_finished_write(mp);
1336 vn_rangelock_unlock(vp, rl_cookie);
1341 * Truncate a file that is already locked.
1344 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1350 error = VOP_ADD_WRITECOUNT(vp, 1);
1353 vattr.va_size = length;
1355 vattr.va_vaflags |= VA_SYNC;
1356 error = VOP_SETATTR(vp, &vattr, cred);
1357 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1363 * File table vnode stat routine.
1366 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1369 struct vnode *vp = fp->f_vnode;
1372 vn_lock(vp, LK_SHARED | LK_RETRY);
1373 error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1380 * Stat a vnode; implementation for the stat syscall
1383 vn_stat(struct vnode *vp, struct stat *sb, struct ucred *active_cred,
1384 struct ucred *file_cred, struct thread *td)
1391 AUDIT_ARG_VNODE1(vp);
1393 error = mac_vnode_check_stat(active_cred, file_cred, vp);
1401 * Initialize defaults for new and unusual fields, so that file
1402 * systems which don't support these fields don't need to know
1405 vap->va_birthtime.tv_sec = -1;
1406 vap->va_birthtime.tv_nsec = 0;
1407 vap->va_fsid = VNOVAL;
1408 vap->va_rdev = NODEV;
1410 error = VOP_GETATTR(vp, vap, active_cred);
1415 * Zero the spare stat fields
1417 bzero(sb, sizeof *sb);
1420 * Copy from vattr table
1422 if (vap->va_fsid != VNOVAL)
1423 sb->st_dev = vap->va_fsid;
1425 sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1426 sb->st_ino = vap->va_fileid;
1427 mode = vap->va_mode;
1428 switch (vap->va_type) {
1454 sb->st_nlink = vap->va_nlink;
1455 sb->st_uid = vap->va_uid;
1456 sb->st_gid = vap->va_gid;
1457 sb->st_rdev = vap->va_rdev;
1458 if (vap->va_size > OFF_MAX)
1460 sb->st_size = vap->va_size;
1461 sb->st_atim.tv_sec = vap->va_atime.tv_sec;
1462 sb->st_atim.tv_nsec = vap->va_atime.tv_nsec;
1463 sb->st_mtim.tv_sec = vap->va_mtime.tv_sec;
1464 sb->st_mtim.tv_nsec = vap->va_mtime.tv_nsec;
1465 sb->st_ctim.tv_sec = vap->va_ctime.tv_sec;
1466 sb->st_ctim.tv_nsec = vap->va_ctime.tv_nsec;
1467 sb->st_birthtim.tv_sec = vap->va_birthtime.tv_sec;
1468 sb->st_birthtim.tv_nsec = vap->va_birthtime.tv_nsec;
1471 * According to www.opengroup.org, the meaning of st_blksize is
1472 * "a filesystem-specific preferred I/O block size for this
1473 * object. In some filesystem types, this may vary from file
1475 * Use miminum/default of PAGE_SIZE (e.g. for VCHR).
1478 sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1480 sb->st_flags = vap->va_flags;
1481 if (priv_check(td, PRIV_VFS_GENERATION))
1484 sb->st_gen = vap->va_gen;
1486 sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1491 * File table vnode ioctl routine.
1494 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1499 struct fiobmap2_arg *bmarg;
1503 switch (vp->v_type) {
1508 vn_lock(vp, LK_SHARED | LK_RETRY);
1509 error = VOP_GETATTR(vp, &vattr, active_cred);
1512 *(int *)data = vattr.va_size - fp->f_offset;
1515 bmarg = (struct fiobmap2_arg *)data;
1516 vn_lock(vp, LK_SHARED | LK_RETRY);
1518 error = mac_vnode_check_read(active_cred, fp->f_cred,
1522 error = VOP_BMAP(vp, bmarg->bn, NULL,
1523 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1530 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1535 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1543 * File table vnode poll routine.
1546 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1554 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1555 AUDIT_ARG_VNODE1(vp);
1556 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1561 error = VOP_POLL(vp, events, fp->f_cred, td);
1566 * Acquire the requested lock and then check for validity. LK_RETRY
1567 * permits vn_lock to return doomed vnodes.
1569 static int __noinline
1570 _vn_lock_fallback(struct vnode *vp, int flags, char *file, int line, int error)
1573 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1574 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1577 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1579 if ((flags & LK_RETRY) == 0) {
1590 * Nothing to do if we got the lock.
1596 * Interlock was dropped by the call in _vn_lock.
1598 flags &= ~LK_INTERLOCK;
1600 error = VOP_LOCK1(vp, flags, file, line);
1601 } while (error != 0);
1606 _vn_lock(struct vnode *vp, int flags, char *file, int line)
1610 VNASSERT((flags & LK_TYPE_MASK) != 0, vp, ("vn_lock: no locktype"));
1611 VNASSERT(vp->v_holdcnt != 0, vp, ("vn_lock: zero hold count"));
1612 error = VOP_LOCK1(vp, flags, file, line);
1613 if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1614 return (_vn_lock_fallback(vp, flags, file, line, error));
1619 * File table vnode close routine.
1622 vn_closefile(struct file *fp, struct thread *td)
1630 fp->f_ops = &badfileops;
1631 ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1633 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1635 if (__predict_false(ref)) {
1636 lf.l_whence = SEEK_SET;
1639 lf.l_type = F_UNLCK;
1640 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1647 vn_suspendable(struct mount *mp)
1650 return (mp->mnt_op->vfs_susp_clean != NULL);
1654 * Preparing to start a filesystem write operation. If the operation is
1655 * permitted, then we bump the count of operations in progress and
1656 * proceed. If a suspend request is in progress, we wait until the
1657 * suspension is over, and then proceed.
1660 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1664 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1665 vfs_op_thread_enter(mp)) {
1666 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1667 vfs_mp_count_add_pcpu(mp, writeopcount, 1);
1668 vfs_op_thread_exit(mp);
1673 mtx_assert(MNT_MTX(mp), MA_OWNED);
1680 * Check on status of suspension.
1682 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1683 mp->mnt_susp_owner != curthread) {
1684 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1685 (flags & PCATCH) : 0) | (PUSER - 1);
1686 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1687 if (flags & V_NOWAIT) {
1688 error = EWOULDBLOCK;
1691 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1697 if (flags & V_XSLEEP)
1699 mp->mnt_writeopcount++;
1701 if (error != 0 || (flags & V_XSLEEP) != 0)
1708 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1713 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1714 ("V_MNTREF requires mp"));
1718 * If a vnode is provided, get and return the mount point that
1719 * to which it will write.
1722 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1724 if (error != EOPNOTSUPP)
1729 if ((mp = *mpp) == NULL)
1732 if (!vn_suspendable(mp)) {
1733 if (vp != NULL || (flags & V_MNTREF) != 0)
1739 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1741 * As long as a vnode is not provided we need to acquire a
1742 * refcount for the provided mountpoint too, in order to
1743 * emulate a vfs_ref().
1745 if (vp == NULL && (flags & V_MNTREF) == 0)
1748 return (vn_start_write_refed(mp, flags, false));
1752 * Secondary suspension. Used by operations such as vop_inactive
1753 * routines that are needed by the higher level functions. These
1754 * are allowed to proceed until all the higher level functions have
1755 * completed (indicated by mnt_writeopcount dropping to zero). At that
1756 * time, these operations are halted until the suspension is over.
1759 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1764 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1765 ("V_MNTREF requires mp"));
1769 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1771 if (error != EOPNOTSUPP)
1777 * If we are not suspended or have not yet reached suspended
1778 * mode, then let the operation proceed.
1780 if ((mp = *mpp) == NULL)
1783 if (!vn_suspendable(mp)) {
1784 if (vp != NULL || (flags & V_MNTREF) != 0)
1790 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1792 * As long as a vnode is not provided we need to acquire a
1793 * refcount for the provided mountpoint too, in order to
1794 * emulate a vfs_ref().
1797 if (vp == NULL && (flags & V_MNTREF) == 0)
1799 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1800 mp->mnt_secondary_writes++;
1801 mp->mnt_secondary_accwrites++;
1805 if (flags & V_NOWAIT) {
1808 return (EWOULDBLOCK);
1811 * Wait for the suspension to finish.
1813 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1814 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1823 * Filesystem write operation has completed. If we are suspending and this
1824 * operation is the last one, notify the suspender that the suspension is
1828 vn_finished_write(struct mount *mp)
1832 if (mp == NULL || !vn_suspendable(mp))
1835 if (vfs_op_thread_enter(mp)) {
1836 vfs_mp_count_sub_pcpu(mp, writeopcount, 1);
1837 vfs_mp_count_sub_pcpu(mp, ref, 1);
1838 vfs_op_thread_exit(mp);
1843 vfs_assert_mount_counters(mp);
1845 c = --mp->mnt_writeopcount;
1846 if (mp->mnt_vfs_ops == 0) {
1847 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1852 vfs_dump_mount_counters(mp);
1853 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
1854 wakeup(&mp->mnt_writeopcount);
1860 * Filesystem secondary write operation has completed. If we are
1861 * suspending and this operation is the last one, notify the suspender
1862 * that the suspension is now in effect.
1865 vn_finished_secondary_write(struct mount *mp)
1867 if (mp == NULL || !vn_suspendable(mp))
1871 mp->mnt_secondary_writes--;
1872 if (mp->mnt_secondary_writes < 0)
1873 panic("vn_finished_secondary_write: neg cnt");
1874 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1875 mp->mnt_secondary_writes <= 0)
1876 wakeup(&mp->mnt_secondary_writes);
1883 * Request a filesystem to suspend write operations.
1886 vfs_write_suspend(struct mount *mp, int flags)
1890 MPASS(vn_suspendable(mp));
1895 vfs_assert_mount_counters(mp);
1896 if (mp->mnt_susp_owner == curthread) {
1897 vfs_op_exit_locked(mp);
1901 while (mp->mnt_kern_flag & MNTK_SUSPEND)
1902 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1905 * Unmount holds a write reference on the mount point. If we
1906 * own busy reference and drain for writers, we deadlock with
1907 * the reference draining in the unmount path. Callers of
1908 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1909 * vfs_busy() reference is owned and caller is not in the
1912 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1913 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1914 vfs_op_exit_locked(mp);
1919 mp->mnt_kern_flag |= MNTK_SUSPEND;
1920 mp->mnt_susp_owner = curthread;
1921 if (mp->mnt_writeopcount > 0)
1922 (void) msleep(&mp->mnt_writeopcount,
1923 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1926 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
1927 vfs_write_resume(mp, 0);
1934 * Request a filesystem to resume write operations.
1937 vfs_write_resume(struct mount *mp, int flags)
1940 MPASS(vn_suspendable(mp));
1943 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1944 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
1945 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
1947 mp->mnt_susp_owner = NULL;
1948 wakeup(&mp->mnt_writeopcount);
1949 wakeup(&mp->mnt_flag);
1950 curthread->td_pflags &= ~TDP_IGNSUSP;
1951 if ((flags & VR_START_WRITE) != 0) {
1953 mp->mnt_writeopcount++;
1956 if ((flags & VR_NO_SUSPCLR) == 0)
1959 } else if ((flags & VR_START_WRITE) != 0) {
1961 vn_start_write_refed(mp, 0, true);
1968 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
1972 vfs_write_suspend_umnt(struct mount *mp)
1976 MPASS(vn_suspendable(mp));
1977 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
1978 ("vfs_write_suspend_umnt: recursed"));
1980 /* dounmount() already called vn_start_write(). */
1982 vn_finished_write(mp);
1983 error = vfs_write_suspend(mp, 0);
1985 vn_start_write(NULL, &mp, V_WAIT);
1989 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
1992 vn_start_write(NULL, &mp, V_WAIT);
1994 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
1995 wakeup(&mp->mnt_flag);
1997 curthread->td_pflags |= TDP_IGNSUSP;
2002 * Implement kqueues for files by translating it to vnode operation.
2005 vn_kqfilter(struct file *fp, struct knote *kn)
2008 return (VOP_KQFILTER(fp->f_vnode, kn));
2012 * Simplified in-kernel wrapper calls for extended attribute access.
2013 * Both calls pass in a NULL credential, authorizing as "kernel" access.
2014 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2017 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2018 const char *attrname, int *buflen, char *buf, struct thread *td)
2024 iov.iov_len = *buflen;
2027 auio.uio_iov = &iov;
2028 auio.uio_iovcnt = 1;
2029 auio.uio_rw = UIO_READ;
2030 auio.uio_segflg = UIO_SYSSPACE;
2032 auio.uio_offset = 0;
2033 auio.uio_resid = *buflen;
2035 if ((ioflg & IO_NODELOCKED) == 0)
2036 vn_lock(vp, LK_SHARED | LK_RETRY);
2038 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2040 /* authorize attribute retrieval as kernel */
2041 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2044 if ((ioflg & IO_NODELOCKED) == 0)
2048 *buflen = *buflen - auio.uio_resid;
2055 * XXX failure mode if partially written?
2058 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2059 const char *attrname, int buflen, char *buf, struct thread *td)
2066 iov.iov_len = buflen;
2069 auio.uio_iov = &iov;
2070 auio.uio_iovcnt = 1;
2071 auio.uio_rw = UIO_WRITE;
2072 auio.uio_segflg = UIO_SYSSPACE;
2074 auio.uio_offset = 0;
2075 auio.uio_resid = buflen;
2077 if ((ioflg & IO_NODELOCKED) == 0) {
2078 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2080 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2083 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2085 /* authorize attribute setting as kernel */
2086 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2088 if ((ioflg & IO_NODELOCKED) == 0) {
2089 vn_finished_write(mp);
2097 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2098 const char *attrname, struct thread *td)
2103 if ((ioflg & IO_NODELOCKED) == 0) {
2104 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2106 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2109 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2111 /* authorize attribute removal as kernel */
2112 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2113 if (error == EOPNOTSUPP)
2114 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2117 if ((ioflg & IO_NODELOCKED) == 0) {
2118 vn_finished_write(mp);
2126 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2130 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2134 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2137 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2142 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2143 int lkflags, struct vnode **rvp)
2148 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2150 ltype = VOP_ISLOCKED(vp);
2151 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2152 ("vn_vget_ino: vp not locked"));
2153 error = vfs_busy(mp, MBF_NOWAIT);
2157 error = vfs_busy(mp, 0);
2158 vn_lock(vp, ltype | LK_RETRY);
2162 if (VN_IS_DOOMED(vp)) {
2168 error = alloc(mp, alloc_arg, lkflags, rvp);
2170 if (error != 0 || *rvp != vp)
2171 vn_lock(vp, ltype | LK_RETRY);
2172 if (VN_IS_DOOMED(vp)) {
2185 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2189 if (vp->v_type != VREG || td == NULL)
2191 if ((uoff_t)uio->uio_offset + uio->uio_resid >
2192 lim_cur(td, RLIMIT_FSIZE)) {
2193 PROC_LOCK(td->td_proc);
2194 kern_psignal(td->td_proc, SIGXFSZ);
2195 PROC_UNLOCK(td->td_proc);
2202 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2209 vn_lock(vp, LK_SHARED | LK_RETRY);
2210 AUDIT_ARG_VNODE1(vp);
2213 return (setfmode(td, active_cred, vp, mode));
2217 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2224 vn_lock(vp, LK_SHARED | LK_RETRY);
2225 AUDIT_ARG_VNODE1(vp);
2228 return (setfown(td, active_cred, vp, uid, gid));
2232 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2236 if ((object = vp->v_object) == NULL)
2238 VM_OBJECT_WLOCK(object);
2239 vm_object_page_remove(object, start, end, 0);
2240 VM_OBJECT_WUNLOCK(object);
2244 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2252 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2253 ("Wrong command %lu", cmd));
2255 if (vn_lock(vp, LK_SHARED) != 0)
2257 if (vp->v_type != VREG) {
2261 error = VOP_GETATTR(vp, &va, cred);
2265 if (noff >= va.va_size) {
2269 bsize = vp->v_mount->mnt_stat.f_iosize;
2270 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2272 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2273 if (error == EOPNOTSUPP) {
2277 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2278 (bnp != -1 && cmd == FIOSEEKDATA)) {
2285 if (noff > va.va_size)
2287 /* noff == va.va_size. There is an implicit hole at the end of file. */
2288 if (cmd == FIOSEEKDATA)
2298 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2303 off_t foffset, size;
2306 cred = td->td_ucred;
2308 foffset = foffset_lock(fp, 0);
2309 noneg = (vp->v_type != VCHR);
2315 (offset > 0 && foffset > OFF_MAX - offset))) {
2322 vn_lock(vp, LK_SHARED | LK_RETRY);
2323 error = VOP_GETATTR(vp, &vattr, cred);
2329 * If the file references a disk device, then fetch
2330 * the media size and use that to determine the ending
2333 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2334 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2335 vattr.va_size = size;
2337 (vattr.va_size > OFF_MAX ||
2338 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2342 offset += vattr.va_size;
2347 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2348 if (error == ENOTTY)
2352 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2353 if (error == ENOTTY)
2359 if (error == 0 && noneg && offset < 0)
2363 VFS_KNOTE_UNLOCKED(vp, 0);
2364 td->td_uretoff.tdu_off = offset;
2366 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2371 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2377 * Grant permission if the caller is the owner of the file, or
2378 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2379 * on the file. If the time pointer is null, then write
2380 * permission on the file is also sufficient.
2382 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2383 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2384 * will be allowed to set the times [..] to the current
2387 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2388 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2389 error = VOP_ACCESS(vp, VWRITE, cred, td);
2394 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2399 if (fp->f_type == DTYPE_FIFO)
2400 kif->kf_type = KF_TYPE_FIFO;
2402 kif->kf_type = KF_TYPE_VNODE;
2405 FILEDESC_SUNLOCK(fdp);
2406 error = vn_fill_kinfo_vnode(vp, kif);
2408 FILEDESC_SLOCK(fdp);
2413 vn_fill_junk(struct kinfo_file *kif)
2418 * Simulate vn_fullpath returning changing values for a given
2419 * vp during e.g. coredump.
2421 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2422 olen = strlen(kif->kf_path);
2424 strcpy(&kif->kf_path[len - 1], "$");
2426 for (; olen < len; olen++)
2427 strcpy(&kif->kf_path[olen], "A");
2431 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2434 char *fullpath, *freepath;
2437 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2440 error = vn_fullpath(curthread, vp, &fullpath, &freepath);
2442 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2444 if (freepath != NULL)
2445 free(freepath, M_TEMP);
2447 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2452 * Retrieve vnode attributes.
2454 va.va_fsid = VNOVAL;
2456 vn_lock(vp, LK_SHARED | LK_RETRY);
2457 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2461 if (va.va_fsid != VNOVAL)
2462 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2464 kif->kf_un.kf_file.kf_file_fsid =
2465 vp->v_mount->mnt_stat.f_fsid.val[0];
2466 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2467 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2468 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2469 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2470 kif->kf_un.kf_file.kf_file_size = va.va_size;
2471 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2472 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2473 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2478 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2479 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2483 struct pmckern_map_in pkm;
2489 boolean_t writecounted;
2492 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2493 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2495 * POSIX shared-memory objects are defined to have
2496 * kernel persistence, and are not defined to support
2497 * read(2)/write(2) -- or even open(2). Thus, we can
2498 * use MAP_ASYNC to trade on-disk coherence for speed.
2499 * The shm_open(3) library routine turns on the FPOSIXSHM
2500 * flag to request this behavior.
2502 if ((fp->f_flag & FPOSIXSHM) != 0)
2503 flags |= MAP_NOSYNC;
2508 * Ensure that file and memory protections are
2509 * compatible. Note that we only worry about
2510 * writability if mapping is shared; in this case,
2511 * current and max prot are dictated by the open file.
2512 * XXX use the vnode instead? Problem is: what
2513 * credentials do we use for determination? What if
2514 * proc does a setuid?
2517 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2518 maxprot = VM_PROT_NONE;
2519 if ((prot & VM_PROT_EXECUTE) != 0)
2522 maxprot = VM_PROT_EXECUTE;
2523 if ((fp->f_flag & FREAD) != 0)
2524 maxprot |= VM_PROT_READ;
2525 else if ((prot & VM_PROT_READ) != 0)
2529 * If we are sharing potential changes via MAP_SHARED and we
2530 * are trying to get write permission although we opened it
2531 * without asking for it, bail out.
2533 if ((flags & MAP_SHARED) != 0) {
2534 if ((fp->f_flag & FWRITE) != 0)
2535 maxprot |= VM_PROT_WRITE;
2536 else if ((prot & VM_PROT_WRITE) != 0)
2539 maxprot |= VM_PROT_WRITE;
2540 cap_maxprot |= VM_PROT_WRITE;
2542 maxprot &= cap_maxprot;
2545 * For regular files and shared memory, POSIX requires that
2546 * the value of foff be a legitimate offset within the data
2547 * object. In particular, negative offsets are invalid.
2548 * Blocking negative offsets and overflows here avoids
2549 * possible wraparound or user-level access into reserved
2550 * ranges of the data object later. In contrast, POSIX does
2551 * not dictate how offsets are used by device drivers, so in
2552 * the case of a device mapping a negative offset is passed
2559 foff < 0 || foff > OFF_MAX - size)
2562 writecounted = FALSE;
2563 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2564 &foff, &object, &writecounted);
2567 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2568 foff, writecounted, td);
2571 * If this mapping was accounted for in the vnode's
2572 * writecount, then undo that now.
2575 vm_pager_release_writecount(object, 0, size);
2576 vm_object_deallocate(object);
2579 /* Inform hwpmc(4) if an executable is being mapped. */
2580 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2581 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2583 pkm.pm_address = (uintptr_t) *addr;
2584 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2592 vn_fsid(struct vnode *vp, struct vattr *va)
2596 f = &vp->v_mount->mnt_stat.f_fsid;
2597 va->va_fsid = (uint32_t)f->val[1];
2598 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2599 va->va_fsid += (uint32_t)f->val[0];
2603 vn_fsync_buf(struct vnode *vp, int waitfor)
2605 struct buf *bp, *nbp;
2608 int error, maxretry;
2611 maxretry = 10000; /* large, arbitrarily chosen */
2613 if (vp->v_type == VCHR) {
2615 mp = vp->v_rdev->si_mountpt;
2622 * MARK/SCAN initialization to avoid infinite loops.
2624 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2625 bp->b_vflags &= ~BV_SCANNED;
2630 * Flush all dirty buffers associated with a vnode.
2633 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2634 if ((bp->b_vflags & BV_SCANNED) != 0)
2636 bp->b_vflags |= BV_SCANNED;
2637 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2638 if (waitfor != MNT_WAIT)
2641 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2642 BO_LOCKPTR(bo)) != 0) {
2649 KASSERT(bp->b_bufobj == bo,
2650 ("bp %p wrong b_bufobj %p should be %p",
2651 bp, bp->b_bufobj, bo));
2652 if ((bp->b_flags & B_DELWRI) == 0)
2653 panic("fsync: not dirty");
2654 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2660 if (maxretry < 1000)
2661 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2667 * If synchronous the caller expects us to completely resolve all
2668 * dirty buffers in the system. Wait for in-progress I/O to
2669 * complete (which could include background bitmap writes), then
2670 * retry if dirty blocks still exist.
2672 if (waitfor == MNT_WAIT) {
2673 bufobj_wwait(bo, 0, 0);
2674 if (bo->bo_dirty.bv_cnt > 0) {
2676 * If we are unable to write any of these buffers
2677 * then we fail now rather than trying endlessly
2678 * to write them out.
2680 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2681 if ((error = bp->b_error) != 0)
2683 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2684 (error == 0 && --maxretry >= 0))
2692 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2698 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
2699 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2700 * to do the actual copy.
2701 * vn_generic_copy_file_range() is factored out, so it can be called
2702 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2703 * different file systems.
2706 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2707 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2708 struct ucred *outcred, struct thread *fsize_td)
2712 uint64_t uvalin, uvalout;
2715 *lenp = 0; /* For error returns. */
2718 /* Do some sanity checks on the arguments. */
2723 if (invp->v_type == VDIR || outvp->v_type == VDIR)
2725 else if (*inoffp < 0 || uvalin > INT64_MAX || uvalin <
2726 (uint64_t)*inoffp || *outoffp < 0 || uvalout > INT64_MAX ||
2727 uvalout < (uint64_t)*outoffp || invp->v_type != VREG ||
2728 outvp->v_type != VREG)
2734 * If the two vnode are for the same file system, call
2735 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2736 * which can handle copies across multiple file systems.
2739 if (invp->v_mount == outvp->v_mount)
2740 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2741 lenp, flags, incred, outcred, fsize_td);
2743 error = vn_generic_copy_file_range(invp, inoffp, outvp,
2744 outoffp, lenp, flags, incred, outcred, fsize_td);
2750 * Test len bytes of data starting at dat for all bytes == 0.
2751 * Return true if all bytes are zero, false otherwise.
2752 * Expects dat to be well aligned.
2755 mem_iszero(void *dat, int len)
2761 for (p = dat; len > 0; len -= sizeof(*p), p++) {
2762 if (len >= sizeof(*p)) {
2766 cp = (const char *)p;
2767 for (i = 0; i < len; i++, cp++)
2776 * Look for a hole in the output file and, if found, adjust *outoffp
2777 * and *xferp to skip past the hole.
2778 * *xferp is the entire hole length to be written and xfer2 is how many bytes
2779 * to be written as 0's upon return.
2782 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2783 off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2788 if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2789 *dataoffp = *outoffp;
2790 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2793 *holeoffp = *dataoffp;
2794 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2797 if (error != 0 || *holeoffp == *dataoffp) {
2799 * Since outvp is unlocked, it may be possible for
2800 * another thread to do a truncate(), lseek(), write()
2801 * creating a hole at startoff between the above
2802 * VOP_IOCTL() calls, if the other thread does not do
2804 * If that happens, *holeoffp == *dataoffp and finding
2805 * the hole has failed, so disable vn_skip_hole().
2807 *holeoffp = -1; /* Disable use of vn_skip_hole(). */
2810 KASSERT(*dataoffp >= *outoffp,
2811 ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2812 (intmax_t)*dataoffp, (intmax_t)*outoffp));
2813 KASSERT(*holeoffp > *dataoffp,
2814 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2815 (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2819 * If there is a hole before the data starts, advance *outoffp and
2820 * *xferp past the hole.
2822 if (*dataoffp > *outoffp) {
2823 delta = *dataoffp - *outoffp;
2824 if (delta >= *xferp) {
2825 /* Entire *xferp is a hole. */
2832 xfer2 = MIN(xfer2, *xferp);
2836 * If a hole starts before the end of this xfer2, reduce this xfer2 so
2837 * that the write ends at the start of the hole.
2838 * *holeoffp should always be greater than *outoffp, but for the
2839 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
2842 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
2843 xfer2 = *holeoffp - *outoffp;
2848 * Write an xfer sized chunk to outvp in blksize blocks from dat.
2849 * dat is a maximum of blksize in length and can be written repeatedly in
2851 * If growfile == true, just grow the file via vn_truncate_locked() instead
2852 * of doing actual writes.
2853 * If checkhole == true, a hole is being punched, so skip over any hole
2854 * already in the output file.
2857 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
2858 u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
2861 off_t dataoff, holeoff, xfer2;
2865 * Loop around doing writes of blksize until write has been completed.
2866 * Lock/unlock on each loop iteration so that a bwillwrite() can be
2867 * done for each iteration, since the xfer argument can be very
2868 * large if there is a large hole to punch in the output file.
2873 xfer2 = MIN(xfer, blksize);
2876 * Punching a hole. Skip writing if there is
2877 * already a hole in the output file.
2879 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
2880 &dataoff, &holeoff, cred);
2885 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
2890 error = vn_start_write(outvp, &mp, V_WAIT);
2892 if (MNT_SHARED_WRITES(mp))
2895 lckf = LK_EXCLUSIVE;
2896 error = vn_lock(outvp, lckf);
2900 error = vn_truncate_locked(outvp, outoff + xfer,
2903 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
2904 outoff, UIO_SYSSPACE, IO_NODELOCKED,
2905 curthread->td_ucred, cred, NULL, curthread);
2912 vn_finished_write(mp);
2913 } while (!growfile && xfer > 0 && error == 0);
2918 * Copy a byte range of one file to another. This function can handle the
2919 * case where invp and outvp are on different file systems.
2920 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
2921 * is no better file system specific way to do it.
2924 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
2925 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
2926 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
2931 off_t startoff, endoff, xfer, xfer2;
2934 bool cantseek, readzeros, eof, lastblock;
2936 size_t copylen, len, savlen;
2938 long holein, holeout;
2940 holein = holeout = 0;
2941 savlen = len = *lenp;
2945 error = vn_lock(invp, LK_SHARED);
2948 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
2953 error = vn_start_write(outvp, &mp, V_WAIT);
2955 error = vn_lock(outvp, LK_EXCLUSIVE);
2958 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
2959 * now that outvp is locked.
2961 if (fsize_td != NULL) {
2962 io.uio_offset = *outoffp;
2964 error = vn_rlimit_fsize(outvp, &io, fsize_td);
2968 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
2971 * Holes that are past EOF do not need to be written as a block
2972 * of zero bytes. So, truncate the output file as far as
2973 * possible and then use va.va_size to decide if writing 0
2974 * bytes is necessary in the loop below.
2977 error = VOP_GETATTR(outvp, &va, outcred);
2978 if (error == 0 && va.va_size > *outoffp && va.va_size <=
2981 error = mac_vnode_check_write(curthread->td_ucred,
2985 error = vn_truncate_locked(outvp, *outoffp,
2988 va.va_size = *outoffp;
2993 vn_finished_write(mp);
2998 * Set the blksize to the larger of the hole sizes for invp and outvp.
2999 * If hole sizes aren't available, set the blksize to the larger
3000 * f_iosize of invp and outvp.
3001 * This code expects the hole sizes and f_iosizes to be powers of 2.
3002 * This value is clipped at 4Kbytes and 1Mbyte.
3004 blksize = MAX(holein, holeout);
3006 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3007 outvp->v_mount->mnt_stat.f_iosize);
3010 else if (blksize > 1024 * 1024)
3011 blksize = 1024 * 1024;
3012 dat = malloc(blksize, M_TEMP, M_WAITOK);
3015 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3016 * to find holes. Otherwise, just scan the read block for all 0s
3017 * in the inner loop where the data copying is done.
3018 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3019 * support holes on the server, but do not support FIOSEEKHOLE.
3022 while (len > 0 && error == 0 && !eof) {
3023 endoff = 0; /* To shut up compilers. */
3029 * Find the next data area. If there is just a hole to EOF,
3030 * FIOSEEKDATA should fail and then we drop down into the
3031 * inner loop and create the hole on the outvp file.
3032 * (I do not know if any file system will report a hole to
3033 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3034 * will fail for those file systems.)
3036 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3037 * the code just falls through to the inner copy loop.
3041 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3045 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3048 * Since invp is unlocked, it may be possible for
3049 * another thread to do a truncate(), lseek(), write()
3050 * creating a hole at startoff between the above
3051 * VOP_IOCTL() calls, if the other thread does not do
3053 * If that happens, startoff == endoff and finding
3054 * the hole has failed, so set an error.
3056 if (error == 0 && startoff == endoff)
3057 error = EINVAL; /* Any error. Reset to 0. */
3060 if (startoff > *inoffp) {
3061 /* Found hole before data block. */
3062 xfer = MIN(startoff - *inoffp, len);
3063 if (*outoffp < va.va_size) {
3064 /* Must write 0s to punch hole. */
3065 xfer2 = MIN(va.va_size - *outoffp,
3067 memset(dat, 0, MIN(xfer2, blksize));
3068 error = vn_write_outvp(outvp, dat,
3069 *outoffp, xfer2, blksize, false,
3070 holeout > 0, outcred);
3073 if (error == 0 && *outoffp + xfer >
3074 va.va_size && xfer == len)
3075 /* Grow last block. */
3076 error = vn_write_outvp(outvp, dat,
3077 *outoffp, xfer, blksize, true,
3085 copylen = MIN(len, endoff - startoff);
3097 * Set first xfer to end at a block boundary, so that
3098 * holes are more likely detected in the loop below via
3099 * the for all bytes 0 method.
3101 xfer -= (*inoffp % blksize);
3103 /* Loop copying the data block. */
3104 while (copylen > 0 && error == 0 && !eof) {
3107 error = vn_lock(invp, LK_SHARED);
3110 error = vn_rdwr(UIO_READ, invp, dat, xfer,
3111 startoff, UIO_SYSSPACE, IO_NODELOCKED,
3112 curthread->td_ucred, incred, &aresid,
3116 if (error == 0 && aresid > 0) {
3117 /* Stop the copy at EOF on the input file. */
3124 * Skip the write for holes past the initial EOF
3125 * of the output file, unless this is the last
3126 * write of the output file at EOF.
3128 readzeros = cantseek ? mem_iszero(dat, xfer) :
3132 if (!cantseek || *outoffp < va.va_size ||
3133 lastblock || !readzeros)
3134 error = vn_write_outvp(outvp, dat,
3135 *outoffp, xfer, blksize,
3136 readzeros && lastblock &&
3137 *outoffp >= va.va_size, false,
3151 *lenp = savlen - len;
3157 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3161 off_t olen, ooffset;
3164 int audited_vnode1 = 0;
3168 if (vp->v_type != VREG)
3171 /* Allocating blocks may take a long time, so iterate. */
3178 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3181 error = vn_lock(vp, LK_EXCLUSIVE);
3183 vn_finished_write(mp);
3187 if (!audited_vnode1) {
3188 AUDIT_ARG_VNODE1(vp);
3193 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3196 error = VOP_ALLOCATE(vp, &offset, &len);
3198 vn_finished_write(mp);
3200 if (olen + ooffset != offset + len) {
3201 panic("offset + len changed from %jx/%jx to %jx/%jx",
3202 ooffset, olen, offset, len);
3204 if (error != 0 || len == 0)
3206 KASSERT(olen > len, ("Iteration did not make progress?"));