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>
68 #include <sys/dirent.h>
71 #include <sys/filio.h>
72 #include <sys/resourcevar.h>
73 #include <sys/rwlock.h>
76 #include <sys/sleepqueue.h>
77 #include <sys/sysctl.h>
78 #include <sys/ttycom.h>
80 #include <sys/syslog.h>
81 #include <sys/unistd.h>
83 #include <sys/ktrace.h>
85 #include <security/audit/audit.h>
86 #include <security/mac/mac_framework.h>
89 #include <vm/vm_extern.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_page.h>
94 #include <vm/vm_pager.h>
97 #include <sys/pmckern.h>
100 static fo_rdwr_t vn_read;
101 static fo_rdwr_t vn_write;
102 static fo_rdwr_t vn_io_fault;
103 static fo_truncate_t vn_truncate;
104 static fo_ioctl_t vn_ioctl;
105 static fo_poll_t vn_poll;
106 static fo_kqfilter_t vn_kqfilter;
107 static fo_close_t vn_closefile;
108 static fo_mmap_t vn_mmap;
109 static fo_fallocate_t vn_fallocate;
110 static fo_fspacectl_t vn_fspacectl;
112 struct fileops vnops = {
113 .fo_read = vn_io_fault,
114 .fo_write = vn_io_fault,
115 .fo_truncate = vn_truncate,
116 .fo_ioctl = vn_ioctl,
118 .fo_kqfilter = vn_kqfilter,
119 .fo_stat = vn_statfile,
120 .fo_close = vn_closefile,
121 .fo_chmod = vn_chmod,
122 .fo_chown = vn_chown,
123 .fo_sendfile = vn_sendfile,
125 .fo_fill_kinfo = vn_fill_kinfo,
127 .fo_fallocate = vn_fallocate,
128 .fo_fspacectl = vn_fspacectl,
129 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
132 const u_int io_hold_cnt = 16;
133 static int vn_io_fault_enable = 1;
134 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
135 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
136 static int vn_io_fault_prefault = 0;
137 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
138 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
139 static int vn_io_pgcache_read_enable = 1;
140 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
141 &vn_io_pgcache_read_enable, 0,
142 "Enable copying from page cache for reads, avoiding fs");
143 static u_long vn_io_faults_cnt;
144 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
145 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
147 static int vfs_allow_read_dir = 0;
148 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
149 &vfs_allow_read_dir, 0,
150 "Enable read(2) of directory by root for filesystems that support it");
153 * Returns true if vn_io_fault mode of handling the i/o request should
157 do_vn_io_fault(struct vnode *vp, struct uio *uio)
161 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
162 (mp = vp->v_mount) != NULL &&
163 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
167 * Structure used to pass arguments to vn_io_fault1(), to do either
168 * file- or vnode-based I/O calls.
170 struct vn_io_fault_args {
178 struct fop_args_tag {
182 struct vop_args_tag {
188 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
189 struct vn_io_fault_args *args, struct thread *td);
192 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
194 struct thread *td = curthread;
196 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
200 open2nameif(int fmode, u_int vn_open_flags)
204 res = ISOPEN | LOCKLEAF;
205 if ((fmode & O_RESOLVE_BENEATH) != 0)
207 if ((fmode & O_EMPTY_PATH) != 0)
209 if ((fmode & FREAD) != 0)
211 if ((fmode & FWRITE) != 0)
213 if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
215 if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
217 if ((vn_open_flags & VN_OPEN_WANTIOCTLCAPS) != 0)
218 res |= WANTIOCTLCAPS;
223 * Common code for vnode open operations via a name lookup.
224 * Lookup the vnode and invoke VOP_CREATE if needed.
225 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
227 * Note that this does NOT free nameidata for the successful case,
228 * due to the NDINIT being done elsewhere.
231 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
232 struct ucred *cred, struct file *fp)
237 struct vattr *vap = &vat;
244 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
245 O_EXCL | O_DIRECTORY) ||
246 (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
248 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
249 ndp->ni_cnd.cn_nameiop = CREATE;
250 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
252 * Set NOCACHE to avoid flushing the cache when
253 * rolling in many files at once.
255 * Set NC_KEEPPOSENTRY to keep positive entries if they already
256 * exist despite NOCACHE.
258 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
259 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
260 ndp->ni_cnd.cn_flags |= FOLLOW;
261 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
263 if ((error = namei(ndp)) != 0)
265 if (ndp->ni_vp == NULL) {
268 vap->va_mode = cmode;
270 vap->va_vaflags |= VA_EXCLUSIVE;
271 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
274 if ((error = vn_start_write(NULL, &mp,
275 V_XSLEEP | V_PCATCH)) != 0)
280 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
281 ndp->ni_cnd.cn_flags |= MAKEENTRY;
283 error = mac_vnode_check_create(cred, ndp->ni_dvp,
287 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
290 if (error == 0 && (fmode & O_EXCL) != 0 &&
291 (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
293 vp->v_iflag |= VI_FOPENING;
297 VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
299 vn_finished_write(mp);
302 if (error == ERELOOKUP) {
310 if (ndp->ni_dvp == ndp->ni_vp)
316 if (fmode & O_EXCL) {
320 if (vp->v_type == VDIR) {
327 ndp->ni_cnd.cn_nameiop = LOOKUP;
328 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
329 ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
331 if ((fmode & FWRITE) == 0)
332 ndp->ni_cnd.cn_flags |= LOCKSHARED;
333 if ((error = namei(ndp)) != 0)
337 error = vn_open_vnode(vp, fmode, cred, curthread, fp);
340 vp->v_iflag &= ~VI_FOPENING;
357 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
360 int error, lock_flags, type;
362 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
363 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
365 KASSERT(fp != NULL, ("open with flock requires fp"));
366 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
369 lock_flags = VOP_ISLOCKED(vp);
372 lf.l_whence = SEEK_SET;
375 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
377 if ((fmode & FNONBLOCK) == 0)
379 if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
381 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
383 fp->f_flag |= FHASLOCK;
385 vn_lock(vp, lock_flags | LK_RETRY);
390 * Common code for vnode open operations once a vnode is located.
391 * Check permissions, and call the VOP_OPEN routine.
394 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
395 struct thread *td, struct file *fp)
400 if (vp->v_type == VLNK) {
401 if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
404 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
408 if ((fmode & O_PATH) == 0) {
409 if (vp->v_type == VSOCK)
411 if ((fmode & (FWRITE | O_TRUNC)) != 0) {
412 if (vp->v_type == VDIR)
416 if ((fmode & FREAD) != 0)
418 if ((fmode & O_APPEND) && (fmode & FWRITE))
421 if ((fmode & O_CREAT) != 0)
425 if ((fmode & FEXEC) != 0)
428 if ((fmode & O_VERIFY) != 0)
430 error = mac_vnode_check_open(cred, vp, accmode);
434 accmode &= ~(VCREAT | VVERIFY);
436 if ((fmode & O_CREAT) == 0 && accmode != 0) {
437 error = VOP_ACCESS(vp, accmode, cred, td);
441 if ((fmode & O_PATH) != 0) {
442 if (vp->v_type != VFIFO && vp->v_type != VSOCK &&
443 VOP_ACCESS(vp, VREAD, cred, td) == 0)
444 fp->f_flag |= FKQALLOWED;
448 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
449 vn_lock(vp, LK_UPGRADE | LK_RETRY);
450 error = VOP_OPEN(vp, fmode, cred, td, fp);
454 error = vn_open_vnode_advlock(vp, fmode, fp);
455 if (error == 0 && (fmode & FWRITE) != 0) {
456 error = VOP_ADD_WRITECOUNT(vp, 1);
458 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
459 __func__, vp, vp->v_writecount);
464 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
465 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
470 * Arrange the call by having fdrop() to use
471 * vn_closefile(). This is to satisfy
472 * filesystems like devfs or tmpfs, which
473 * override fo_close().
475 fp->f_flag |= FOPENFAILED;
477 if (fp->f_ops == &badfileops) {
478 fp->f_type = DTYPE_VNODE;
484 * If there is no fp, due to kernel-mode open,
485 * we can call VOP_CLOSE() now.
487 if ((vp->v_type == VFIFO ||
488 !MNT_EXTENDED_SHARED(vp->v_mount)) &&
489 VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
490 vn_lock(vp, LK_UPGRADE | LK_RETRY);
491 (void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
496 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
502 * Check for write permissions on the specified vnode.
503 * Prototype text segments cannot be written.
507 vn_writechk(struct vnode *vp)
510 ASSERT_VOP_LOCKED(vp, "vn_writechk");
512 * If there's shared text associated with
513 * the vnode, try to free it up once. If
514 * we fail, we can't allow writing.
526 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
527 struct thread *td, bool keep_ref)
530 int error, lock_flags;
532 lock_flags = vp->v_type != VFIFO && MNT_EXTENDED_SHARED(vp->v_mount) ?
533 LK_SHARED : LK_EXCLUSIVE;
535 vn_start_write(vp, &mp, V_WAIT);
536 vn_lock(vp, lock_flags | LK_RETRY);
537 AUDIT_ARG_VNODE1(vp);
538 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
539 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
540 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
541 __func__, vp, vp->v_writecount);
543 error = VOP_CLOSE(vp, flags, file_cred, td);
548 vn_finished_write(mp);
553 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
557 return (vn_close1(vp, flags, file_cred, td, false));
561 * Heuristic to detect sequential operation.
564 sequential_heuristic(struct uio *uio, struct file *fp)
568 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
571 if (fp->f_flag & FRDAHEAD)
572 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
575 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
576 * that the first I/O is normally considered to be slightly
577 * sequential. Seeking to offset 0 doesn't change sequentiality
578 * unless previous seeks have reduced f_seqcount to 0, in which
579 * case offset 0 is not special.
581 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
582 uio->uio_offset == fp->f_nextoff[rw]) {
584 * f_seqcount is in units of fixed-size blocks so that it
585 * depends mainly on the amount of sequential I/O and not
586 * much on the number of sequential I/O's. The fixed size
587 * of 16384 is hard-coded here since it is (not quite) just
588 * a magic size that works well here. This size is more
589 * closely related to the best I/O size for real disks than
590 * to any block size used by software.
592 if (uio->uio_resid >= IO_SEQMAX * 16384)
593 fp->f_seqcount[rw] = IO_SEQMAX;
595 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
596 if (fp->f_seqcount[rw] > IO_SEQMAX)
597 fp->f_seqcount[rw] = IO_SEQMAX;
599 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
602 /* Not sequential. Quickly draw-down sequentiality. */
603 if (fp->f_seqcount[rw] > 1)
604 fp->f_seqcount[rw] = 1;
606 fp->f_seqcount[rw] = 0;
611 * Package up an I/O request on a vnode into a uio and do it.
614 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
615 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
616 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
623 struct vn_io_fault_args args;
624 int error, lock_flags;
626 if (offset < 0 && vp->v_type != VCHR)
628 auio.uio_iov = &aiov;
630 aiov.iov_base = base;
632 auio.uio_resid = len;
633 auio.uio_offset = offset;
634 auio.uio_segflg = segflg;
639 if ((ioflg & IO_NODELOCKED) == 0) {
640 if ((ioflg & IO_RANGELOCKED) == 0) {
641 if (rw == UIO_READ) {
642 rl_cookie = vn_rangelock_rlock(vp, offset,
644 } else if ((ioflg & IO_APPEND) != 0) {
645 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
647 rl_cookie = vn_rangelock_wlock(vp, offset,
653 if (rw == UIO_WRITE) {
654 if (vp->v_type != VCHR &&
655 (error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH))
658 lock_flags = vn_lktype_write(mp, vp);
660 lock_flags = LK_SHARED;
661 vn_lock(vp, lock_flags | LK_RETRY);
665 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
667 if ((ioflg & IO_NOMACCHECK) == 0) {
669 error = mac_vnode_check_read(active_cred, file_cred,
672 error = mac_vnode_check_write(active_cred, file_cred,
677 if (file_cred != NULL)
681 if (do_vn_io_fault(vp, &auio)) {
682 args.kind = VN_IO_FAULT_VOP;
685 args.args.vop_args.vp = vp;
686 error = vn_io_fault1(vp, &auio, &args, td);
687 } else if (rw == UIO_READ) {
688 error = VOP_READ(vp, &auio, ioflg, cred);
689 } else /* if (rw == UIO_WRITE) */ {
690 error = VOP_WRITE(vp, &auio, ioflg, cred);
694 *aresid = auio.uio_resid;
696 if (auio.uio_resid && error == 0)
698 if ((ioflg & IO_NODELOCKED) == 0) {
701 vn_finished_write(mp);
704 if (rl_cookie != NULL)
705 vn_rangelock_unlock(vp, rl_cookie);
710 * Package up an I/O request on a vnode into a uio and do it. The I/O
711 * request is split up into smaller chunks and we try to avoid saturating
712 * the buffer cache while potentially holding a vnode locked, so we
713 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
714 * to give other processes a chance to lock the vnode (either other processes
715 * core'ing the same binary, or unrelated processes scanning the directory).
718 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
719 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
720 struct ucred *file_cred, size_t *aresid, struct thread *td)
729 * Force `offset' to a multiple of MAXBSIZE except possibly
730 * for the first chunk, so that filesystems only need to
731 * write full blocks except possibly for the first and last
734 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
738 if (rw != UIO_READ && vp->v_type == VREG)
741 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
742 ioflg, active_cred, file_cred, &iaresid, td);
743 len -= chunk; /* aresid calc already includes length */
747 base = (char *)base + chunk;
748 kern_yield(PRI_USER);
751 *aresid = len + iaresid;
755 #if OFF_MAX <= LONG_MAX
757 foffset_lock(struct file *fp, int flags)
759 volatile short *flagsp;
763 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
765 if ((flags & FOF_NOLOCK) != 0)
766 return (atomic_load_long(&fp->f_offset));
769 * According to McKusick the vn lock was protecting f_offset here.
770 * It is now protected by the FOFFSET_LOCKED flag.
772 flagsp = &fp->f_vnread_flags;
773 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
774 return (atomic_load_long(&fp->f_offset));
776 sleepq_lock(&fp->f_vnread_flags);
777 state = atomic_load_16(flagsp);
779 if ((state & FOFFSET_LOCKED) == 0) {
780 if (!atomic_fcmpset_acq_16(flagsp, &state,
785 if ((state & FOFFSET_LOCK_WAITING) == 0) {
786 if (!atomic_fcmpset_acq_16(flagsp, &state,
787 state | FOFFSET_LOCK_WAITING))
791 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
792 sleepq_wait(&fp->f_vnread_flags, PUSER -1);
794 sleepq_lock(&fp->f_vnread_flags);
795 state = atomic_load_16(flagsp);
797 res = atomic_load_long(&fp->f_offset);
798 sleepq_release(&fp->f_vnread_flags);
803 foffset_unlock(struct file *fp, off_t val, int flags)
805 volatile short *flagsp;
808 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
810 if ((flags & FOF_NOUPDATE) == 0)
811 atomic_store_long(&fp->f_offset, val);
812 if ((flags & FOF_NEXTOFF_R) != 0)
813 fp->f_nextoff[UIO_READ] = val;
814 if ((flags & FOF_NEXTOFF_W) != 0)
815 fp->f_nextoff[UIO_WRITE] = val;
817 if ((flags & FOF_NOLOCK) != 0)
820 flagsp = &fp->f_vnread_flags;
821 state = atomic_load_16(flagsp);
822 if ((state & FOFFSET_LOCK_WAITING) == 0 &&
823 atomic_cmpset_rel_16(flagsp, state, 0))
826 sleepq_lock(&fp->f_vnread_flags);
827 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
828 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
829 fp->f_vnread_flags = 0;
830 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
831 sleepq_release(&fp->f_vnread_flags);
835 foffset_lock(struct file *fp, int flags)
840 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
842 mtxp = mtx_pool_find(mtxpool_sleep, fp);
844 if ((flags & FOF_NOLOCK) == 0) {
845 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
846 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
847 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
850 fp->f_vnread_flags |= FOFFSET_LOCKED;
858 foffset_unlock(struct file *fp, off_t val, int flags)
862 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
864 mtxp = mtx_pool_find(mtxpool_sleep, fp);
866 if ((flags & FOF_NOUPDATE) == 0)
868 if ((flags & FOF_NEXTOFF_R) != 0)
869 fp->f_nextoff[UIO_READ] = val;
870 if ((flags & FOF_NEXTOFF_W) != 0)
871 fp->f_nextoff[UIO_WRITE] = val;
872 if ((flags & FOF_NOLOCK) == 0) {
873 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
874 ("Lost FOFFSET_LOCKED"));
875 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
876 wakeup(&fp->f_vnread_flags);
877 fp->f_vnread_flags = 0;
884 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
887 if ((flags & FOF_OFFSET) == 0)
888 uio->uio_offset = foffset_lock(fp, flags);
892 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
895 if ((flags & FOF_OFFSET) == 0)
896 foffset_unlock(fp, uio->uio_offset, flags);
900 get_advice(struct file *fp, struct uio *uio)
905 ret = POSIX_FADV_NORMAL;
906 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
909 mtxp = mtx_pool_find(mtxpool_sleep, fp);
911 if (fp->f_advice != NULL &&
912 uio->uio_offset >= fp->f_advice->fa_start &&
913 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
914 ret = fp->f_advice->fa_advice;
920 get_write_ioflag(struct file *fp)
928 mp = atomic_load_ptr(&vp->v_mount);
930 if ((fp->f_flag & O_DIRECT) != 0)
933 if ((fp->f_flag & O_FSYNC) != 0 ||
934 (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
938 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
939 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
940 * fall back to full O_SYNC behavior.
942 if ((fp->f_flag & O_DSYNC) != 0)
943 ioflag |= IO_SYNC | IO_DATASYNC;
949 vn_read_from_obj(struct vnode *vp, struct uio *uio)
952 vm_page_t ma[io_hold_cnt + 2];
957 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
958 obj = atomic_load_ptr(&vp->v_object);
960 return (EJUSTRETURN);
963 * Depends on type stability of vm_objects.
965 vm_object_pip_add(obj, 1);
966 if ((obj->flags & OBJ_DEAD) != 0) {
968 * Note that object might be already reused from the
969 * vnode, and the OBJ_DEAD flag cleared. This is fine,
970 * we recheck for DOOMED vnode state after all pages
971 * are busied, and retract then.
973 * But we check for OBJ_DEAD to ensure that we do not
974 * busy pages while vm_object_terminate_pages()
975 * processes the queue.
981 resid = uio->uio_resid;
982 off = uio->uio_offset;
983 for (i = 0; resid > 0; i++) {
984 MPASS(i < io_hold_cnt + 2);
985 ma[i] = vm_page_grab_unlocked(obj, atop(off),
986 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
992 * Skip invalid pages. Valid mask can be partial only
993 * at EOF, and we clip later.
995 if (vm_page_none_valid(ma[i])) {
996 vm_page_sunbusy(ma[i]);
1004 error = EJUSTRETURN;
1009 * Check VIRF_DOOMED after we busied our pages. Since
1010 * vgonel() terminates the vnode' vm_object, it cannot
1011 * process past pages busied by us.
1013 if (VN_IS_DOOMED(vp)) {
1014 error = EJUSTRETURN;
1018 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1019 if (resid > uio->uio_resid)
1020 resid = uio->uio_resid;
1023 * Unlocked read of vnp_size is safe because truncation cannot
1024 * pass busied page. But we load vnp_size into a local
1025 * variable so that possible concurrent extension does not
1026 * break calculation.
1028 #if defined(__powerpc__) && !defined(__powerpc64__)
1029 vsz = obj->un_pager.vnp.vnp_size;
1031 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1033 if (uio->uio_offset >= vsz) {
1034 error = EJUSTRETURN;
1037 if (uio->uio_offset + resid > vsz)
1038 resid = vsz - uio->uio_offset;
1040 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1043 for (j = 0; j < i; j++) {
1045 vm_page_reference(ma[j]);
1046 vm_page_sunbusy(ma[j]);
1049 vm_object_pip_wakeup(obj);
1052 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1056 * File table vnode read routine.
1059 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1067 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1069 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1072 if (fp->f_flag & FNONBLOCK)
1073 ioflag |= IO_NDELAY;
1074 if (fp->f_flag & O_DIRECT)
1075 ioflag |= IO_DIRECT;
1078 * Try to read from page cache. VIRF_DOOMED check is racy but
1079 * allows us to avoid unneeded work outright.
1081 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1082 (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1083 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1085 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1088 if (error != EJUSTRETURN)
1092 advice = get_advice(fp, uio);
1093 vn_lock(vp, LK_SHARED | LK_RETRY);
1096 case POSIX_FADV_NORMAL:
1097 case POSIX_FADV_SEQUENTIAL:
1098 case POSIX_FADV_NOREUSE:
1099 ioflag |= sequential_heuristic(uio, fp);
1101 case POSIX_FADV_RANDOM:
1102 /* Disable read-ahead for random I/O. */
1105 orig_offset = uio->uio_offset;
1108 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1111 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1112 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1114 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1115 orig_offset != uio->uio_offset)
1117 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1118 * for the backing file after a POSIX_FADV_NOREUSE
1121 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1122 POSIX_FADV_DONTNEED);
1127 * File table vnode write routine.
1130 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1138 bool need_finished_write;
1140 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1142 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1144 if (vp->v_type == VREG)
1147 if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1148 ioflag |= IO_APPEND;
1149 if ((fp->f_flag & FNONBLOCK) != 0)
1150 ioflag |= IO_NDELAY;
1151 ioflag |= get_write_ioflag(fp);
1154 need_finished_write = false;
1155 if (vp->v_type != VCHR) {
1156 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
1159 need_finished_write = true;
1162 advice = get_advice(fp, uio);
1164 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1166 case POSIX_FADV_NORMAL:
1167 case POSIX_FADV_SEQUENTIAL:
1168 case POSIX_FADV_NOREUSE:
1169 ioflag |= sequential_heuristic(uio, fp);
1171 case POSIX_FADV_RANDOM:
1172 /* XXX: Is this correct? */
1175 orig_offset = uio->uio_offset;
1178 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1181 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1182 fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1184 if (need_finished_write)
1185 vn_finished_write(mp);
1186 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1187 orig_offset != uio->uio_offset)
1189 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1190 * for the backing file after a POSIX_FADV_NOREUSE
1193 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1194 POSIX_FADV_DONTNEED);
1200 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1201 * prevent the following deadlock:
1203 * Assume that the thread A reads from the vnode vp1 into userspace
1204 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
1205 * currently not resident, then system ends up with the call chain
1206 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1207 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1208 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1209 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1210 * backed by the pages of vnode vp1, and some page in buf2 is not
1211 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1213 * To prevent the lock order reversal and deadlock, vn_io_fault() does
1214 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1215 * Instead, it first tries to do the whole range i/o with pagefaults
1216 * disabled. If all pages in the i/o buffer are resident and mapped,
1217 * VOP will succeed (ignoring the genuine filesystem errors).
1218 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1219 * i/o in chunks, with all pages in the chunk prefaulted and held
1220 * using vm_fault_quick_hold_pages().
1222 * Filesystems using this deadlock avoidance scheme should use the
1223 * array of the held pages from uio, saved in the curthread->td_ma,
1224 * instead of doing uiomove(). A helper function
1225 * vn_io_fault_uiomove() converts uiomove request into
1226 * uiomove_fromphys() over td_ma array.
1228 * Since vnode locks do not cover the whole i/o anymore, rangelocks
1229 * make the current i/o request atomic with respect to other i/os and
1234 * Decode vn_io_fault_args and perform the corresponding i/o.
1237 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1243 save = vm_fault_disable_pagefaults();
1244 switch (args->kind) {
1245 case VN_IO_FAULT_FOP:
1246 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1247 uio, args->cred, args->flags, td);
1249 case VN_IO_FAULT_VOP:
1250 if (uio->uio_rw == UIO_READ) {
1251 error = VOP_READ(args->args.vop_args.vp, uio,
1252 args->flags, args->cred);
1253 } else if (uio->uio_rw == UIO_WRITE) {
1254 error = VOP_WRITE(args->args.vop_args.vp, uio,
1255 args->flags, args->cred);
1259 panic("vn_io_fault_doio: unknown kind of io %d %d",
1260 args->kind, uio->uio_rw);
1262 vm_fault_enable_pagefaults(save);
1267 vn_io_fault_touch(char *base, const struct uio *uio)
1272 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1278 vn_io_fault_prefault_user(const struct uio *uio)
1281 const struct iovec *iov;
1286 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1287 ("vn_io_fault_prefault userspace"));
1291 resid = uio->uio_resid;
1292 base = iov->iov_base;
1295 error = vn_io_fault_touch(base, uio);
1298 if (len < PAGE_SIZE) {
1300 error = vn_io_fault_touch(base + len - 1, uio);
1305 if (++i >= uio->uio_iovcnt)
1307 iov = uio->uio_iov + i;
1308 base = iov->iov_base;
1320 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1321 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1322 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1323 * into args and call vn_io_fault1() to handle faults during the user
1324 * mode buffer accesses.
1327 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1330 vm_page_t ma[io_hold_cnt + 2];
1331 struct uio *uio_clone, short_uio;
1332 struct iovec short_iovec[1];
1333 vm_page_t *prev_td_ma;
1335 vm_offset_t addr, end;
1338 int error, cnt, saveheld, prev_td_ma_cnt;
1340 if (vn_io_fault_prefault) {
1341 error = vn_io_fault_prefault_user(uio);
1343 return (error); /* Or ignore ? */
1346 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1349 * The UFS follows IO_UNIT directive and replays back both
1350 * uio_offset and uio_resid if an error is encountered during the
1351 * operation. But, since the iovec may be already advanced,
1352 * uio is still in an inconsistent state.
1354 * Cache a copy of the original uio, which is advanced to the redo
1355 * point using UIO_NOCOPY below.
1357 uio_clone = cloneuio(uio);
1358 resid = uio->uio_resid;
1360 short_uio.uio_segflg = UIO_USERSPACE;
1361 short_uio.uio_rw = uio->uio_rw;
1362 short_uio.uio_td = uio->uio_td;
1364 error = vn_io_fault_doio(args, uio, td);
1365 if (error != EFAULT)
1368 atomic_add_long(&vn_io_faults_cnt, 1);
1369 uio_clone->uio_segflg = UIO_NOCOPY;
1370 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1371 uio_clone->uio_segflg = uio->uio_segflg;
1373 saveheld = curthread_pflags_set(TDP_UIOHELD);
1374 prev_td_ma = td->td_ma;
1375 prev_td_ma_cnt = td->td_ma_cnt;
1377 while (uio_clone->uio_resid != 0) {
1378 len = uio_clone->uio_iov->iov_len;
1380 KASSERT(uio_clone->uio_iovcnt >= 1,
1381 ("iovcnt underflow"));
1382 uio_clone->uio_iov++;
1383 uio_clone->uio_iovcnt--;
1386 if (len > ptoa(io_hold_cnt))
1387 len = ptoa(io_hold_cnt);
1388 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1389 end = round_page(addr + len);
1395 * A perfectly misaligned address and length could cause
1396 * both the start and the end of the chunk to use partial
1397 * page. +2 accounts for such a situation.
1399 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1400 addr, len, prot, ma, io_hold_cnt + 2);
1405 short_uio.uio_iov = &short_iovec[0];
1406 short_iovec[0].iov_base = (void *)addr;
1407 short_uio.uio_iovcnt = 1;
1408 short_uio.uio_resid = short_iovec[0].iov_len = len;
1409 short_uio.uio_offset = uio_clone->uio_offset;
1411 td->td_ma_cnt = cnt;
1413 error = vn_io_fault_doio(args, &short_uio, td);
1414 vm_page_unhold_pages(ma, cnt);
1415 adv = len - short_uio.uio_resid;
1417 uio_clone->uio_iov->iov_base =
1418 (char *)uio_clone->uio_iov->iov_base + adv;
1419 uio_clone->uio_iov->iov_len -= adv;
1420 uio_clone->uio_resid -= adv;
1421 uio_clone->uio_offset += adv;
1423 uio->uio_resid -= adv;
1424 uio->uio_offset += adv;
1426 if (error != 0 || adv == 0)
1429 td->td_ma = prev_td_ma;
1430 td->td_ma_cnt = prev_td_ma_cnt;
1431 curthread_pflags_restore(saveheld);
1433 free(uio_clone, M_IOV);
1438 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1439 int flags, struct thread *td)
1444 struct vn_io_fault_args args;
1447 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1451 * The ability to read(2) on a directory has historically been
1452 * allowed for all users, but this can and has been the source of
1453 * at least one security issue in the past. As such, it is now hidden
1454 * away behind a sysctl for those that actually need it to use it, and
1455 * restricted to root when it's turned on to make it relatively safe to
1456 * leave on for longer sessions of need.
1458 if (vp->v_type == VDIR) {
1459 KASSERT(uio->uio_rw == UIO_READ,
1460 ("illegal write attempted on a directory"));
1461 if (!vfs_allow_read_dir)
1463 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1467 foffset_lock_uio(fp, uio, flags);
1468 if (do_vn_io_fault(vp, uio)) {
1469 args.kind = VN_IO_FAULT_FOP;
1470 args.args.fop_args.fp = fp;
1471 args.args.fop_args.doio = doio;
1472 args.cred = active_cred;
1473 args.flags = flags | FOF_OFFSET;
1474 if (uio->uio_rw == UIO_READ) {
1475 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1476 uio->uio_offset + uio->uio_resid);
1477 } else if ((fp->f_flag & O_APPEND) != 0 ||
1478 (flags & FOF_OFFSET) == 0) {
1479 /* For appenders, punt and lock the whole range. */
1480 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1482 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1483 uio->uio_offset + uio->uio_resid);
1485 error = vn_io_fault1(vp, uio, &args, td);
1486 vn_rangelock_unlock(vp, rl_cookie);
1488 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1490 foffset_unlock_uio(fp, uio, flags);
1495 * Helper function to perform the requested uiomove operation using
1496 * the held pages for io->uio_iov[0].iov_base buffer instead of
1497 * copyin/copyout. Access to the pages with uiomove_fromphys()
1498 * instead of iov_base prevents page faults that could occur due to
1499 * pmap_collect() invalidating the mapping created by
1500 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1501 * object cleanup revoking the write access from page mappings.
1503 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1504 * instead of plain uiomove().
1507 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1509 struct uio transp_uio;
1510 struct iovec transp_iov[1];
1516 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1517 uio->uio_segflg != UIO_USERSPACE)
1518 return (uiomove(data, xfersize, uio));
1520 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1521 transp_iov[0].iov_base = data;
1522 transp_uio.uio_iov = &transp_iov[0];
1523 transp_uio.uio_iovcnt = 1;
1524 if (xfersize > uio->uio_resid)
1525 xfersize = uio->uio_resid;
1526 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1527 transp_uio.uio_offset = 0;
1528 transp_uio.uio_segflg = UIO_SYSSPACE;
1530 * Since transp_iov points to data, and td_ma page array
1531 * corresponds to original uio->uio_iov, we need to invert the
1532 * direction of the i/o operation as passed to
1533 * uiomove_fromphys().
1535 switch (uio->uio_rw) {
1537 transp_uio.uio_rw = UIO_READ;
1540 transp_uio.uio_rw = UIO_WRITE;
1543 transp_uio.uio_td = uio->uio_td;
1544 error = uiomove_fromphys(td->td_ma,
1545 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1546 xfersize, &transp_uio);
1547 adv = xfersize - transp_uio.uio_resid;
1549 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1550 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1552 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1554 td->td_ma_cnt -= pgadv;
1555 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1556 uio->uio_iov->iov_len -= adv;
1557 uio->uio_resid -= adv;
1558 uio->uio_offset += adv;
1563 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1567 vm_offset_t iov_base;
1571 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1572 uio->uio_segflg != UIO_USERSPACE)
1573 return (uiomove_fromphys(ma, offset, xfersize, uio));
1575 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1576 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1577 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1578 switch (uio->uio_rw) {
1580 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1584 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1588 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1590 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1592 td->td_ma_cnt -= pgadv;
1593 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1594 uio->uio_iov->iov_len -= cnt;
1595 uio->uio_resid -= cnt;
1596 uio->uio_offset += cnt;
1601 * File table truncate routine.
1604 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1616 * Lock the whole range for truncation. Otherwise split i/o
1617 * might happen partly before and partly after the truncation.
1619 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1620 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
1623 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1624 AUDIT_ARG_VNODE1(vp);
1625 if (vp->v_type == VDIR) {
1630 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1634 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1638 vn_finished_write(mp);
1640 vn_rangelock_unlock(vp, rl_cookie);
1641 if (error == ERELOOKUP)
1647 * Truncate a file that is already locked.
1650 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1656 error = VOP_ADD_WRITECOUNT(vp, 1);
1659 vattr.va_size = length;
1661 vattr.va_vaflags |= VA_SYNC;
1662 error = VOP_SETATTR(vp, &vattr, cred);
1663 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1669 * File table vnode stat routine.
1672 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred)
1674 struct vnode *vp = fp->f_vnode;
1677 vn_lock(vp, LK_SHARED | LK_RETRY);
1678 error = VOP_STAT(vp, sb, active_cred, fp->f_cred);
1685 * File table vnode ioctl routine.
1688 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1692 struct fiobmap2_arg *bmarg;
1697 switch (vp->v_type) {
1702 error = vn_getsize(vp, &size, active_cred);
1704 *(int *)data = size - fp->f_offset;
1707 bmarg = (struct fiobmap2_arg *)data;
1708 vn_lock(vp, LK_SHARED | LK_RETRY);
1710 error = mac_vnode_check_read(active_cred, fp->f_cred,
1714 error = VOP_BMAP(vp, bmarg->bn, NULL,
1715 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1722 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1727 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1735 * File table vnode poll routine.
1738 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1745 #if defined(MAC) || defined(AUDIT)
1746 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1747 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1748 AUDIT_ARG_VNODE1(vp);
1749 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1755 error = VOP_POLL(vp, events, fp->f_cred, td);
1760 * Acquire the requested lock and then check for validity. LK_RETRY
1761 * permits vn_lock to return doomed vnodes.
1763 static int __noinline
1764 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1768 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1769 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1772 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1774 if ((flags & LK_RETRY) == 0) {
1785 * Nothing to do if we got the lock.
1791 * Interlock was dropped by the call in _vn_lock.
1793 flags &= ~LK_INTERLOCK;
1795 error = VOP_LOCK1(vp, flags, file, line);
1796 } while (error != 0);
1801 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1805 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1806 ("vn_lock: no locktype (%d passed)", flags));
1807 VNPASS(vp->v_holdcnt > 0, vp);
1808 error = VOP_LOCK1(vp, flags, file, line);
1809 if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1810 return (_vn_lock_fallback(vp, flags, file, line, error));
1815 * File table vnode close routine.
1818 vn_closefile(struct file *fp, struct thread *td)
1826 fp->f_ops = &badfileops;
1827 ref = (fp->f_flag & FHASLOCK) != 0;
1829 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1831 if (__predict_false(ref)) {
1832 lf.l_whence = SEEK_SET;
1835 lf.l_type = F_UNLCK;
1836 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1843 * Preparing to start a filesystem write operation. If the operation is
1844 * permitted, then we bump the count of operations in progress and
1845 * proceed. If a suspend request is in progress, we wait until the
1846 * suspension is over, and then proceed.
1849 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1851 struct mount_pcpu *mpcpu;
1854 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1855 vfs_op_thread_enter(mp, mpcpu)) {
1856 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1857 vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1858 vfs_op_thread_exit(mp, mpcpu);
1863 mtx_assert(MNT_MTX(mp), MA_OWNED);
1870 * Check on status of suspension.
1872 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1873 mp->mnt_susp_owner != curthread) {
1875 if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) {
1876 if (flags & V_PCATCH)
1879 mflags |= (PUSER - 1);
1880 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1881 if ((flags & V_NOWAIT) != 0) {
1882 error = EWOULDBLOCK;
1885 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1891 if ((flags & V_XSLEEP) != 0)
1893 mp->mnt_writeopcount++;
1895 if (error != 0 || (flags & V_XSLEEP) != 0)
1902 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1907 KASSERT((flags & ~V_VALID_FLAGS) == 0,
1908 ("%s: invalid flags passed %d\n", __func__, flags));
1912 * If a vnode is provided, get and return the mount point that
1913 * to which it will write.
1916 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1918 if (error != EOPNOTSUPP)
1923 if ((mp = *mpp) == NULL)
1927 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1929 * As long as a vnode is not provided we need to acquire a
1930 * refcount for the provided mountpoint too, in order to
1931 * emulate a vfs_ref().
1936 error = vn_start_write_refed(mp, flags, false);
1937 if (error != 0 && (flags & V_NOWAIT) == 0)
1943 * Secondary suspension. Used by operations such as vop_inactive
1944 * routines that are needed by the higher level functions. These
1945 * are allowed to proceed until all the higher level functions have
1946 * completed (indicated by mnt_writeopcount dropping to zero). At that
1947 * time, these operations are halted until the suspension is over.
1950 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1955 KASSERT((flags & (~V_VALID_FLAGS | V_XSLEEP)) == 0,
1956 ("%s: invalid flags passed %d\n", __func__, flags));
1960 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1962 if (error != EOPNOTSUPP)
1968 * If we are not suspended or have not yet reached suspended
1969 * mode, then let the operation proceed.
1971 if ((mp = *mpp) == NULL)
1975 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1977 * As long as a vnode is not provided we need to acquire a
1978 * refcount for the provided mountpoint too, in order to
1979 * emulate a vfs_ref().
1984 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1985 mp->mnt_secondary_writes++;
1986 mp->mnt_secondary_accwrites++;
1990 if ((flags & V_NOWAIT) != 0) {
1994 return (EWOULDBLOCK);
1997 * Wait for the suspension to finish.
2000 if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) {
2001 if ((flags & V_PCATCH) != 0)
2004 mflags |= (PUSER - 1) | PDROP;
2005 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, "suspfs", 0);
2014 * Filesystem write operation has completed. If we are suspending and this
2015 * operation is the last one, notify the suspender that the suspension is
2019 vn_finished_write(struct mount *mp)
2021 struct mount_pcpu *mpcpu;
2027 if (vfs_op_thread_enter(mp, mpcpu)) {
2028 vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2029 vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2030 vfs_op_thread_exit(mp, mpcpu);
2035 vfs_assert_mount_counters(mp);
2037 c = --mp->mnt_writeopcount;
2038 if (mp->mnt_vfs_ops == 0) {
2039 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2044 vfs_dump_mount_counters(mp);
2045 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2046 wakeup(&mp->mnt_writeopcount);
2051 * Filesystem secondary write operation has completed. If we are
2052 * suspending and this operation is the last one, notify the suspender
2053 * that the suspension is now in effect.
2056 vn_finished_secondary_write(struct mount *mp)
2062 mp->mnt_secondary_writes--;
2063 if (mp->mnt_secondary_writes < 0)
2064 panic("vn_finished_secondary_write: neg cnt");
2065 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2066 mp->mnt_secondary_writes <= 0)
2067 wakeup(&mp->mnt_secondary_writes);
2072 * Request a filesystem to suspend write operations.
2075 vfs_write_suspend(struct mount *mp, int flags)
2082 vfs_assert_mount_counters(mp);
2083 if (mp->mnt_susp_owner == curthread) {
2084 vfs_op_exit_locked(mp);
2088 while (mp->mnt_kern_flag & MNTK_SUSPEND)
2089 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2092 * Unmount holds a write reference on the mount point. If we
2093 * own busy reference and drain for writers, we deadlock with
2094 * the reference draining in the unmount path. Callers of
2095 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2096 * vfs_busy() reference is owned and caller is not in the
2099 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2100 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2101 vfs_op_exit_locked(mp);
2106 mp->mnt_kern_flag |= MNTK_SUSPEND;
2107 mp->mnt_susp_owner = curthread;
2108 if (mp->mnt_writeopcount > 0)
2109 (void) msleep(&mp->mnt_writeopcount,
2110 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2113 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2114 vfs_write_resume(mp, 0);
2115 /* vfs_write_resume does vfs_op_exit() for us */
2121 * Request a filesystem to resume write operations.
2124 vfs_write_resume(struct mount *mp, int flags)
2128 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2129 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2130 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2132 mp->mnt_susp_owner = NULL;
2133 wakeup(&mp->mnt_writeopcount);
2134 wakeup(&mp->mnt_flag);
2135 curthread->td_pflags &= ~TDP_IGNSUSP;
2136 if ((flags & VR_START_WRITE) != 0) {
2138 mp->mnt_writeopcount++;
2141 if ((flags & VR_NO_SUSPCLR) == 0)
2144 } else if ((flags & VR_START_WRITE) != 0) {
2146 vn_start_write_refed(mp, 0, true);
2153 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2157 vfs_write_suspend_umnt(struct mount *mp)
2161 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2162 ("vfs_write_suspend_umnt: recursed"));
2164 /* dounmount() already called vn_start_write(). */
2166 vn_finished_write(mp);
2167 error = vfs_write_suspend(mp, 0);
2169 vn_start_write(NULL, &mp, V_WAIT);
2173 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2176 vn_start_write(NULL, &mp, V_WAIT);
2178 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2179 wakeup(&mp->mnt_flag);
2181 curthread->td_pflags |= TDP_IGNSUSP;
2186 * Implement kqueues for files by translating it to vnode operation.
2189 vn_kqfilter(struct file *fp, struct knote *kn)
2192 return (VOP_KQFILTER(fp->f_vnode, kn));
2196 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2198 if ((fp->f_flag & FKQALLOWED) == 0)
2200 return (vn_kqfilter(fp, kn));
2204 * Simplified in-kernel wrapper calls for extended attribute access.
2205 * Both calls pass in a NULL credential, authorizing as "kernel" access.
2206 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2209 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2210 const char *attrname, int *buflen, char *buf, struct thread *td)
2216 iov.iov_len = *buflen;
2219 auio.uio_iov = &iov;
2220 auio.uio_iovcnt = 1;
2221 auio.uio_rw = UIO_READ;
2222 auio.uio_segflg = UIO_SYSSPACE;
2224 auio.uio_offset = 0;
2225 auio.uio_resid = *buflen;
2227 if ((ioflg & IO_NODELOCKED) == 0)
2228 vn_lock(vp, LK_SHARED | LK_RETRY);
2230 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2232 /* authorize attribute retrieval as kernel */
2233 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2236 if ((ioflg & IO_NODELOCKED) == 0)
2240 *buflen = *buflen - auio.uio_resid;
2247 * XXX failure mode if partially written?
2250 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2251 const char *attrname, int buflen, char *buf, struct thread *td)
2258 iov.iov_len = buflen;
2261 auio.uio_iov = &iov;
2262 auio.uio_iovcnt = 1;
2263 auio.uio_rw = UIO_WRITE;
2264 auio.uio_segflg = UIO_SYSSPACE;
2266 auio.uio_offset = 0;
2267 auio.uio_resid = buflen;
2269 if ((ioflg & IO_NODELOCKED) == 0) {
2270 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2272 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2275 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2277 /* authorize attribute setting as kernel */
2278 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2280 if ((ioflg & IO_NODELOCKED) == 0) {
2281 vn_finished_write(mp);
2289 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2290 const char *attrname, struct thread *td)
2295 if ((ioflg & IO_NODELOCKED) == 0) {
2296 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2298 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2301 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2303 /* authorize attribute removal as kernel */
2304 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2305 if (error == EOPNOTSUPP)
2306 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2309 if ((ioflg & IO_NODELOCKED) == 0) {
2310 vn_finished_write(mp);
2318 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2322 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2326 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2329 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2334 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2335 int lkflags, struct vnode **rvp)
2340 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2342 ltype = VOP_ISLOCKED(vp);
2343 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2344 ("vn_vget_ino: vp not locked"));
2345 error = vfs_busy(mp, MBF_NOWAIT);
2349 error = vfs_busy(mp, 0);
2350 vn_lock(vp, ltype | LK_RETRY);
2354 if (VN_IS_DOOMED(vp)) {
2360 error = alloc(mp, alloc_arg, lkflags, rvp);
2362 if (error != 0 || *rvp != vp)
2363 vn_lock(vp, ltype | LK_RETRY);
2364 if (VN_IS_DOOMED(vp)) {
2377 vn_send_sigxfsz(struct proc *p)
2380 kern_psignal(p, SIGXFSZ);
2385 vn_rlimit_trunc(u_quad_t size, struct thread *td)
2387 if (size <= lim_cur(td, RLIMIT_FSIZE))
2389 vn_send_sigxfsz(td->td_proc);
2394 vn_rlimit_fsizex1(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2395 bool adj, struct thread *td)
2400 if (vp->v_type != VREG)
2404 * Handle file system maximum file size.
2406 if (maxfsz != 0 && uio->uio_offset + uio->uio_resid > maxfsz) {
2407 if (!adj || uio->uio_offset >= maxfsz)
2409 uio->uio_resid = maxfsz - uio->uio_offset;
2413 * This is kernel write (e.g. vnode_pager) or accounting
2414 * write, ignore limit.
2416 if (td == NULL || (td->td_pflags2 & TDP2_ACCT) != 0)
2420 * Calculate file size limit.
2422 ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2423 lim = __predict_false(ktr_write) ? td->td_ktr_io_lim :
2424 lim_cur(td, RLIMIT_FSIZE);
2427 * Is the limit reached?
2429 if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2433 * Prepared filesystems can handle writes truncated to the
2436 if (adj && (uoff_t)uio->uio_offset < lim) {
2437 uio->uio_resid = lim - (uoff_t)uio->uio_offset;
2441 if (!ktr_write || ktr_filesize_limit_signal)
2442 vn_send_sigxfsz(td->td_proc);
2447 * Helper for VOP_WRITE() implementations, the common code to
2448 * handle maximum supported file size on the filesystem, and
2449 * RLIMIT_FSIZE, except for special writes from accounting subsystem
2452 * For maximum file size (maxfsz argument):
2453 * - return EFBIG if uio_offset is beyond it
2454 * - otherwise, clamp uio_resid if write would extend file beyond maxfsz.
2457 * - return EFBIG and send SIGXFSZ if uio_offset is beyond the limit
2458 * - otherwise, clamp uio_resid if write would extend file beyond limit.
2460 * If clamping occured, the adjustment for uio_resid is stored in
2461 * *resid_adj, to be re-applied by vn_rlimit_fsizex_res() on return
2465 vn_rlimit_fsizex(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2466 ssize_t *resid_adj, struct thread *td)
2472 resid_orig = uio->uio_resid;
2473 adj = resid_adj != NULL;
2474 error = vn_rlimit_fsizex1(vp, uio, maxfsz, adj, td);
2476 *resid_adj = resid_orig - uio->uio_resid;
2481 vn_rlimit_fsizex_res(struct uio *uio, ssize_t resid_adj)
2483 uio->uio_resid += resid_adj;
2487 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2490 return (vn_rlimit_fsizex(vp, __DECONST(struct uio *, uio), 0, NULL,
2495 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2502 vn_lock(vp, LK_SHARED | LK_RETRY);
2503 AUDIT_ARG_VNODE1(vp);
2506 return (setfmode(td, active_cred, vp, mode));
2510 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2517 vn_lock(vp, LK_SHARED | LK_RETRY);
2518 AUDIT_ARG_VNODE1(vp);
2521 return (setfown(td, active_cred, vp, uid, gid));
2525 * Remove pages in the range ["start", "end") from the vnode's VM object. If
2526 * "end" is 0, then the range extends to the end of the object.
2529 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2533 if ((object = vp->v_object) == NULL)
2535 VM_OBJECT_WLOCK(object);
2536 vm_object_page_remove(object, start, end, 0);
2537 VM_OBJECT_WUNLOCK(object);
2541 * Like vn_pages_remove(), but skips invalid pages, which by definition are not
2542 * mapped into any process' address space. Filesystems may use this in
2543 * preference to vn_pages_remove() to avoid blocking on pages busied in
2544 * preparation for a VOP_GETPAGES.
2547 vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2551 if ((object = vp->v_object) == NULL)
2553 VM_OBJECT_WLOCK(object);
2554 vm_object_page_remove(object, start, end, OBJPR_VALIDONLY);
2555 VM_OBJECT_WUNLOCK(object);
2559 vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off,
2569 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2570 ("%s: Wrong command %lu", __func__, cmd));
2571 ASSERT_VOP_ELOCKED(vp, "vn_bmap_seekhole_locked");
2573 if (vp->v_type != VREG) {
2577 error = vn_getsize_locked(vp, &size, cred);
2581 if (noff < 0 || noff >= size) {
2586 /* See the comment in ufs_bmap_seekdata(). */
2589 VM_OBJECT_WLOCK(obj);
2590 vm_object_page_clean(obj, 0, 0, OBJPC_SYNC);
2591 VM_OBJECT_WUNLOCK(obj);
2594 bsize = vp->v_mount->mnt_stat.f_iosize;
2595 for (bn = noff / bsize; noff < size; bn++, noff += bsize -
2597 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2598 if (error == EOPNOTSUPP) {
2602 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2603 (bnp != -1 && cmd == FIOSEEKDATA)) {
2612 /* noff == size. There is an implicit hole at the end of file. */
2613 if (cmd == FIOSEEKDATA)
2622 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2626 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2627 ("%s: Wrong command %lu", __func__, cmd));
2629 if (vn_lock(vp, LK_EXCLUSIVE) != 0)
2631 error = vn_bmap_seekhole_locked(vp, cmd, off, cred);
2637 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2641 off_t foffset, fsize, size;
2644 cred = td->td_ucred;
2646 foffset = foffset_lock(fp, 0);
2647 noneg = (vp->v_type != VCHR);
2653 (offset > 0 && foffset > OFF_MAX - offset))) {
2660 error = vn_getsize(vp, &fsize, cred);
2665 * If the file references a disk device, then fetch
2666 * the media size and use that to determine the ending
2669 if (fsize == 0 && vp->v_type == VCHR &&
2670 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2672 if (noneg && offset > 0 && fsize > OFF_MAX - offset) {
2681 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2682 if (error == ENOTTY)
2686 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2687 if (error == ENOTTY)
2693 if (error == 0 && noneg && offset < 0)
2697 VFS_KNOTE_UNLOCKED(vp, 0);
2698 td->td_uretoff.tdu_off = offset;
2700 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2705 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2711 * Grant permission if the caller is the owner of the file, or
2712 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2713 * on the file. If the time pointer is null, then write
2714 * permission on the file is also sufficient.
2716 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2717 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2718 * will be allowed to set the times [..] to the current
2721 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2722 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2723 error = VOP_ACCESS(vp, VWRITE, cred, td);
2728 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2733 if (fp->f_type == DTYPE_FIFO)
2734 kif->kf_type = KF_TYPE_FIFO;
2736 kif->kf_type = KF_TYPE_VNODE;
2739 FILEDESC_SUNLOCK(fdp);
2740 error = vn_fill_kinfo_vnode(vp, kif);
2742 FILEDESC_SLOCK(fdp);
2747 vn_fill_junk(struct kinfo_file *kif)
2752 * Simulate vn_fullpath returning changing values for a given
2753 * vp during e.g. coredump.
2755 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2756 olen = strlen(kif->kf_path);
2758 strcpy(&kif->kf_path[len - 1], "$");
2760 for (; olen < len; olen++)
2761 strcpy(&kif->kf_path[olen], "A");
2765 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2768 char *fullpath, *freepath;
2771 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2774 error = vn_fullpath(vp, &fullpath, &freepath);
2776 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2778 if (freepath != NULL)
2779 free(freepath, M_TEMP);
2781 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2786 * Retrieve vnode attributes.
2788 va.va_fsid = VNOVAL;
2790 vn_lock(vp, LK_SHARED | LK_RETRY);
2791 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2795 if (va.va_fsid != VNOVAL)
2796 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2798 kif->kf_un.kf_file.kf_file_fsid =
2799 vp->v_mount->mnt_stat.f_fsid.val[0];
2800 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2801 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2802 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2803 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2804 kif->kf_un.kf_file.kf_file_size = va.va_size;
2805 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2806 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2807 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2808 kif->kf_un.kf_file.kf_file_nlink = va.va_nlink;
2813 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2814 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2818 struct pmckern_map_in pkm;
2824 boolean_t writecounted;
2827 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2828 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2830 * POSIX shared-memory objects are defined to have
2831 * kernel persistence, and are not defined to support
2832 * read(2)/write(2) -- or even open(2). Thus, we can
2833 * use MAP_ASYNC to trade on-disk coherence for speed.
2834 * The shm_open(3) library routine turns on the FPOSIXSHM
2835 * flag to request this behavior.
2837 if ((fp->f_flag & FPOSIXSHM) != 0)
2838 flags |= MAP_NOSYNC;
2843 * Ensure that file and memory protections are
2844 * compatible. Note that we only worry about
2845 * writability if mapping is shared; in this case,
2846 * current and max prot are dictated by the open file.
2847 * XXX use the vnode instead? Problem is: what
2848 * credentials do we use for determination? What if
2849 * proc does a setuid?
2852 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2853 maxprot = VM_PROT_NONE;
2854 if ((prot & VM_PROT_EXECUTE) != 0)
2857 maxprot = VM_PROT_EXECUTE;
2858 if ((fp->f_flag & FREAD) != 0)
2859 maxprot |= VM_PROT_READ;
2860 else if ((prot & VM_PROT_READ) != 0)
2864 * If we are sharing potential changes via MAP_SHARED and we
2865 * are trying to get write permission although we opened it
2866 * without asking for it, bail out.
2868 if ((flags & MAP_SHARED) != 0) {
2869 if ((fp->f_flag & FWRITE) != 0)
2870 maxprot |= VM_PROT_WRITE;
2871 else if ((prot & VM_PROT_WRITE) != 0)
2874 maxprot |= VM_PROT_WRITE;
2875 cap_maxprot |= VM_PROT_WRITE;
2877 maxprot &= cap_maxprot;
2880 * For regular files and shared memory, POSIX requires that
2881 * the value of foff be a legitimate offset within the data
2882 * object. In particular, negative offsets are invalid.
2883 * Blocking negative offsets and overflows here avoids
2884 * possible wraparound or user-level access into reserved
2885 * ranges of the data object later. In contrast, POSIX does
2886 * not dictate how offsets are used by device drivers, so in
2887 * the case of a device mapping a negative offset is passed
2894 foff > OFF_MAX - size)
2897 writecounted = FALSE;
2898 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2899 &foff, &object, &writecounted);
2902 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2903 foff, writecounted, td);
2906 * If this mapping was accounted for in the vnode's
2907 * writecount, then undo that now.
2910 vm_pager_release_writecount(object, 0, size);
2911 vm_object_deallocate(object);
2914 /* Inform hwpmc(4) if an executable is being mapped. */
2915 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2916 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2918 pkm.pm_address = (uintptr_t) *addr;
2919 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2927 vn_fsid(struct vnode *vp, struct vattr *va)
2931 f = &vp->v_mount->mnt_stat.f_fsid;
2932 va->va_fsid = (uint32_t)f->val[1];
2933 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2934 va->va_fsid += (uint32_t)f->val[0];
2938 vn_fsync_buf(struct vnode *vp, int waitfor)
2940 struct buf *bp, *nbp;
2943 int error, maxretry;
2946 maxretry = 10000; /* large, arbitrarily chosen */
2948 if (vp->v_type == VCHR) {
2950 mp = vp->v_rdev->si_mountpt;
2957 * MARK/SCAN initialization to avoid infinite loops.
2959 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2960 bp->b_vflags &= ~BV_SCANNED;
2965 * Flush all dirty buffers associated with a vnode.
2968 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2969 if ((bp->b_vflags & BV_SCANNED) != 0)
2971 bp->b_vflags |= BV_SCANNED;
2972 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2973 if (waitfor != MNT_WAIT)
2976 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2977 BO_LOCKPTR(bo)) != 0) {
2984 KASSERT(bp->b_bufobj == bo,
2985 ("bp %p wrong b_bufobj %p should be %p",
2986 bp, bp->b_bufobj, bo));
2987 if ((bp->b_flags & B_DELWRI) == 0)
2988 panic("fsync: not dirty");
2989 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2995 if (maxretry < 1000)
2996 pause("dirty", hz < 1000 ? 1 : hz / 1000);
3002 * If synchronous the caller expects us to completely resolve all
3003 * dirty buffers in the system. Wait for in-progress I/O to
3004 * complete (which could include background bitmap writes), then
3005 * retry if dirty blocks still exist.
3007 if (waitfor == MNT_WAIT) {
3008 bufobj_wwait(bo, 0, 0);
3009 if (bo->bo_dirty.bv_cnt > 0) {
3011 * If we are unable to write any of these buffers
3012 * then we fail now rather than trying endlessly
3013 * to write them out.
3015 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
3016 if ((error = bp->b_error) != 0)
3018 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
3019 (error == 0 && --maxretry >= 0))
3027 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
3033 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
3034 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
3035 * to do the actual copy.
3036 * vn_generic_copy_file_range() is factored out, so it can be called
3037 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
3038 * different file systems.
3041 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
3042 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
3043 struct ucred *outcred, struct thread *fsize_td)
3050 *lenp = 0; /* For error returns. */
3053 /* Do some sanity checks on the arguments. */
3054 if (invp->v_type == VDIR || outvp->v_type == VDIR)
3056 else if (*inoffp < 0 || *outoffp < 0 ||
3057 invp->v_type != VREG || outvp->v_type != VREG)
3062 /* Ensure offset + len does not wrap around. */
3065 if (uval > INT64_MAX)
3066 len = INT64_MAX - *inoffp;
3069 if (uval > INT64_MAX)
3070 len = INT64_MAX - *outoffp;
3075 * If the two vnode are for the same file system, call
3076 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
3077 * which can handle copies across multiple file systems.
3080 if (invp->v_mount == outvp->v_mount)
3081 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
3082 lenp, flags, incred, outcred, fsize_td);
3084 error = vn_generic_copy_file_range(invp, inoffp, outvp,
3085 outoffp, lenp, flags, incred, outcred, fsize_td);
3091 * Test len bytes of data starting at dat for all bytes == 0.
3092 * Return true if all bytes are zero, false otherwise.
3093 * Expects dat to be well aligned.
3096 mem_iszero(void *dat, int len)
3102 for (p = dat; len > 0; len -= sizeof(*p), p++) {
3103 if (len >= sizeof(*p)) {
3107 cp = (const char *)p;
3108 for (i = 0; i < len; i++, cp++)
3117 * Look for a hole in the output file and, if found, adjust *outoffp
3118 * and *xferp to skip past the hole.
3119 * *xferp is the entire hole length to be written and xfer2 is how many bytes
3120 * to be written as 0's upon return.
3123 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3124 off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3129 if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3130 *dataoffp = *outoffp;
3131 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3134 *holeoffp = *dataoffp;
3135 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3138 if (error != 0 || *holeoffp == *dataoffp) {
3140 * Since outvp is unlocked, it may be possible for
3141 * another thread to do a truncate(), lseek(), write()
3142 * creating a hole at startoff between the above
3143 * VOP_IOCTL() calls, if the other thread does not do
3145 * If that happens, *holeoffp == *dataoffp and finding
3146 * the hole has failed, so disable vn_skip_hole().
3148 *holeoffp = -1; /* Disable use of vn_skip_hole(). */
3151 KASSERT(*dataoffp >= *outoffp,
3152 ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3153 (intmax_t)*dataoffp, (intmax_t)*outoffp));
3154 KASSERT(*holeoffp > *dataoffp,
3155 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3156 (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3160 * If there is a hole before the data starts, advance *outoffp and
3161 * *xferp past the hole.
3163 if (*dataoffp > *outoffp) {
3164 delta = *dataoffp - *outoffp;
3165 if (delta >= *xferp) {
3166 /* Entire *xferp is a hole. */
3173 xfer2 = MIN(xfer2, *xferp);
3177 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3178 * that the write ends at the start of the hole.
3179 * *holeoffp should always be greater than *outoffp, but for the
3180 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3183 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3184 xfer2 = *holeoffp - *outoffp;
3189 * Write an xfer sized chunk to outvp in blksize blocks from dat.
3190 * dat is a maximum of blksize in length and can be written repeatedly in
3192 * If growfile == true, just grow the file via vn_truncate_locked() instead
3193 * of doing actual writes.
3194 * If checkhole == true, a hole is being punched, so skip over any hole
3195 * already in the output file.
3198 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3199 u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3202 off_t dataoff, holeoff, xfer2;
3206 * Loop around doing writes of blksize until write has been completed.
3207 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3208 * done for each iteration, since the xfer argument can be very
3209 * large if there is a large hole to punch in the output file.
3214 xfer2 = MIN(xfer, blksize);
3217 * Punching a hole. Skip writing if there is
3218 * already a hole in the output file.
3220 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3221 &dataoff, &holeoff, cred);
3226 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3231 error = vn_start_write(outvp, &mp, V_WAIT);
3235 error = vn_lock(outvp, LK_EXCLUSIVE);
3237 error = vn_truncate_locked(outvp, outoff + xfer,
3242 error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3244 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3245 outoff, UIO_SYSSPACE, IO_NODELOCKED,
3246 curthread->td_ucred, cred, NULL, curthread);
3253 vn_finished_write(mp);
3254 } while (!growfile && xfer > 0 && error == 0);
3259 * Copy a byte range of one file to another. This function can handle the
3260 * case where invp and outvp are on different file systems.
3261 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3262 * is no better file system specific way to do it.
3265 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3266 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3267 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3270 off_t startoff, endoff, xfer, xfer2;
3272 int error, interrupted;
3273 bool cantseek, readzeros, eof, lastblock, holetoeof;
3274 ssize_t aresid, r = 0;
3275 size_t copylen, len, savlen;
3276 off_t insize, outsize;
3278 long holein, holeout;
3279 struct timespec curts, endts;
3281 holein = holeout = 0;
3282 savlen = len = *lenp;
3287 error = vn_lock(invp, LK_SHARED);
3290 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3293 error = vn_getsize_locked(invp, &insize, incred);
3299 error = vn_start_write(outvp, &mp, V_WAIT);
3301 error = vn_lock(outvp, LK_EXCLUSIVE);
3304 * If fsize_td != NULL, do a vn_rlimit_fsizex() call,
3305 * now that outvp is locked.
3307 if (fsize_td != NULL) {
3310 io.uio_offset = *outoffp;
3312 error = vn_rlimit_fsizex(outvp, &io, 0, &r, fsize_td);
3313 len = savlen = io.uio_resid;
3315 * No need to call vn_rlimit_fsizex_res before return,
3316 * since the uio is local.
3319 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3322 * Holes that are past EOF do not need to be written as a block
3323 * of zero bytes. So, truncate the output file as far as
3324 * possible and then use size to decide if writing 0
3325 * bytes is necessary in the loop below.
3328 error = vn_getsize_locked(outvp, &outsize, outcred);
3329 if (error == 0 && outsize > *outoffp && outsize <= *outoffp + len) {
3331 error = mac_vnode_check_write(curthread->td_ucred,
3335 error = vn_truncate_locked(outvp, *outoffp,
3343 vn_finished_write(mp);
3347 if (holein == 0 && holeout > 0) {
3349 * For this special case, the input data will be scanned
3350 * for blocks of all 0 bytes. For these blocks, the
3351 * write can be skipped for the output file to create
3352 * an unallocated region.
3353 * Therefore, use the appropriate size for the output file.
3356 if (blksize <= 512) {
3358 * Use f_iosize, since ZFS reports a _PC_MIN_HOLE_SIZE
3359 * of 512, although it actually only creates
3360 * unallocated regions for blocks >= f_iosize.
3362 blksize = outvp->v_mount->mnt_stat.f_iosize;
3366 * Use the larger of the two f_iosize values. If they are
3367 * not the same size, one will normally be an exact multiple of
3368 * the other, since they are both likely to be a power of 2.
3370 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3371 outvp->v_mount->mnt_stat.f_iosize);
3374 /* Clip to sane limits. */
3377 else if (blksize > maxphys)
3379 dat = malloc(blksize, M_TEMP, M_WAITOK);
3382 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3383 * to find holes. Otherwise, just scan the read block for all 0s
3384 * in the inner loop where the data copying is done.
3385 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3386 * support holes on the server, but do not support FIOSEEKHOLE.
3387 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3388 * that this function should return after 1second with a partial
3391 if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3392 getnanouptime(&endts);
3395 timespecclear(&endts);
3396 holetoeof = eof = false;
3397 while (len > 0 && error == 0 && !eof && interrupted == 0) {
3398 endoff = 0; /* To shut up compilers. */
3404 * Find the next data area. If there is just a hole to EOF,
3405 * FIOSEEKDATA should fail with ENXIO.
3406 * (I do not know if any file system will report a hole to
3407 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3408 * will fail for those file systems.)
3410 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3411 * the code just falls through to the inner copy loop.
3415 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3417 if (error == ENXIO) {
3418 startoff = endoff = insize;
3419 eof = holetoeof = true;
3423 if (error == 0 && !holetoeof) {
3425 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3428 * Since invp is unlocked, it may be possible for
3429 * another thread to do a truncate(), lseek(), write()
3430 * creating a hole at startoff between the above
3431 * VOP_IOCTL() calls, if the other thread does not do
3433 * If that happens, startoff == endoff and finding
3434 * the hole has failed, so set an error.
3436 if (error == 0 && startoff == endoff)
3437 error = EINVAL; /* Any error. Reset to 0. */
3440 if (startoff > *inoffp) {
3441 /* Found hole before data block. */
3442 xfer = MIN(startoff - *inoffp, len);
3443 if (*outoffp < outsize) {
3444 /* Must write 0s to punch hole. */
3445 xfer2 = MIN(outsize - *outoffp,
3447 memset(dat, 0, MIN(xfer2, blksize));
3448 error = vn_write_outvp(outvp, dat,
3449 *outoffp, xfer2, blksize, false,
3450 holeout > 0, outcred);
3453 if (error == 0 && *outoffp + xfer >
3454 outsize && (xfer == len || holetoeof)) {
3455 /* Grow output file (hole at end). */
3456 error = vn_write_outvp(outvp, dat,
3457 *outoffp, xfer, blksize, true,
3465 interrupted = sig_intr();
3466 if (timespecisset(&endts) &&
3468 getnanouptime(&curts);
3469 if (timespeccmp(&curts,
3477 copylen = MIN(len, endoff - startoff);
3489 * Set first xfer to end at a block boundary, so that
3490 * holes are more likely detected in the loop below via
3491 * the for all bytes 0 method.
3493 xfer -= (*inoffp % blksize);
3495 /* Loop copying the data block. */
3496 while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3499 error = vn_lock(invp, LK_SHARED);
3502 error = vn_rdwr(UIO_READ, invp, dat, xfer,
3503 startoff, UIO_SYSSPACE, IO_NODELOCKED,
3504 curthread->td_ucred, incred, &aresid,
3508 if (error == 0 && aresid > 0) {
3509 /* Stop the copy at EOF on the input file. */
3516 * Skip the write for holes past the initial EOF
3517 * of the output file, unless this is the last
3518 * write of the output file at EOF.
3520 readzeros = cantseek ? mem_iszero(dat, xfer) :
3524 if (!cantseek || *outoffp < outsize ||
3525 lastblock || !readzeros)
3526 error = vn_write_outvp(outvp, dat,
3527 *outoffp, xfer, blksize,
3528 readzeros && lastblock &&
3529 *outoffp >= outsize, false,
3538 interrupted = sig_intr();
3539 if (timespecisset(&endts) &&
3541 getnanouptime(&curts);
3542 if (timespeccmp(&curts,
3554 *lenp = savlen - len;
3560 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3564 off_t olen, ooffset;
3567 int audited_vnode1 = 0;
3571 if (vp->v_type != VREG)
3574 /* Allocating blocks may take a long time, so iterate. */
3581 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
3584 error = vn_lock(vp, LK_EXCLUSIVE);
3586 vn_finished_write(mp);
3590 if (!audited_vnode1) {
3591 AUDIT_ARG_VNODE1(vp);
3596 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3599 error = VOP_ALLOCATE(vp, &offset, &len, 0,
3602 vn_finished_write(mp);
3604 if (olen + ooffset != offset + len) {
3605 panic("offset + len changed from %jx/%jx to %jx/%jx",
3606 ooffset, olen, offset, len);
3608 if (error != 0 || len == 0)
3610 KASSERT(olen > len, ("Iteration did not make progress?"));
3618 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags,
3619 int ioflag, struct ucred *cred, struct ucred *active_cred,
3620 struct ucred *file_cred)
3627 bool audited_vnode1 = false;
3636 if ((ioflag & (IO_NODELOCKED | IO_RANGELOCKED)) == 0)
3637 rl_cookie = vn_rangelock_wlock(vp, off, off + len);
3638 while (len > 0 && error == 0) {
3640 * Try to deallocate the longest range in one pass.
3641 * In case a pass takes too long to be executed, it returns
3642 * partial result. The residue will be proceeded in the next
3646 if ((ioflag & IO_NODELOCKED) == 0) {
3648 if ((error = vn_start_write(vp, &mp,
3649 V_WAIT | V_PCATCH)) != 0)
3651 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
3654 if (!audited_vnode1) {
3655 AUDIT_ARG_VNODE1(vp);
3656 audited_vnode1 = true;
3661 if ((ioflag & IO_NOMACCHECK) == 0)
3662 error = mac_vnode_check_write(active_cred, file_cred,
3666 error = VOP_DEALLOCATE(vp, &off, &len, flags, ioflag,
3669 if ((ioflag & IO_NODELOCKED) == 0) {
3672 vn_finished_write(mp);
3676 if (error == 0 && len != 0)
3680 if (rl_cookie != NULL)
3681 vn_rangelock_unlock(vp, rl_cookie);
3688 * This function is supposed to be used in the situations where the deallocation
3689 * is not triggered by a user request.
3692 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags,
3693 int ioflag, struct ucred *active_cred, struct ucred *file_cred)
3697 if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset ||
3700 if (vp->v_type != VREG)
3703 cred = file_cred != NOCRED ? file_cred : active_cred;
3704 return (vn_deallocate_impl(vp, offset, length, flags, ioflag, cred,
3705 active_cred, file_cred));
3709 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
3710 struct ucred *active_cred, struct thread *td)
3716 KASSERT(cmd == SPACECTL_DEALLOC, ("vn_fspacectl: Invalid cmd"));
3717 KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0,
3718 ("vn_fspacectl: non-zero flags"));
3719 KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset,
3720 ("vn_fspacectl: offset/length overflow or underflow"));
3723 if (vp->v_type != VREG)
3726 ioflag = get_write_ioflag(fp);
3729 case SPACECTL_DEALLOC:
3730 error = vn_deallocate_impl(vp, offset, length, flags, ioflag,
3731 active_cred, active_cred, fp->f_cred);
3734 panic("vn_fspacectl: unknown cmd %d", cmd);
3741 * Keep this assert as long as sizeof(struct dirent) is used as the maximum
3744 _Static_assert(_GENERIC_MAXDIRSIZ == sizeof(struct dirent),
3745 "'struct dirent' size must be a multiple of its alignment "
3746 "(see _GENERIC_DIRLEN())");
3749 * Returns successive directory entries through some caller's provided buffer.
3751 * This function automatically refills the provided buffer with calls to
3752 * VOP_READDIR() (after MAC permission checks).
3754 * 'td' is used for credentials and passed to uiomove(). 'dirbuf' is the
3755 * caller's buffer to fill and 'dirbuflen' its allocated size. 'dirbuf' must
3756 * be properly aligned to access 'struct dirent' structures and 'dirbuflen'
3757 * must be greater than GENERIC_MAXDIRSIZ to avoid VOP_READDIR() returning
3758 * EINVAL (the latter is not a strong guarantee (yet); but EINVAL will always
3759 * be returned if this requirement is not verified). '*dpp' points to the
3760 * current directory entry in the buffer and '*len' contains the remaining
3761 * valid bytes in 'dirbuf' after 'dpp' (including the pointed entry).
3763 * At first call (or when restarting the read), '*len' must have been set to 0,
3764 * '*off' to 0 (or any valid start offset) and '*eofflag' to 0. There are no
3765 * more entries as soon as '*len' is 0 after a call that returned 0. Calling
3766 * again this function after such a condition is considered an error and EINVAL
3767 * will be returned. Other possible error codes are those of VOP_READDIR(),
3768 * EINTEGRITY if the returned entries do not pass coherency tests, or EINVAL
3769 * (bad call). All errors are unrecoverable, i.e., the state ('*len', '*off'
3770 * and '*eofflag') must be re-initialized before a subsequent call. On error
3771 * or at end of directory, '*dpp' is reset to NULL.
3773 * '*len', '*off' and '*eofflag' are internal state the caller should not
3774 * tamper with except as explained above. '*off' is the next directory offset
3775 * to read from to refill the buffer. '*eofflag' is set to 0 or 1 by the last
3776 * internal call to VOP_READDIR() that returned without error, indicating
3777 * whether it reached the end of the directory, and to 2 by this function after
3778 * all entries have been read.
3781 vn_dir_next_dirent(struct vnode *vp, struct thread *td,
3782 char *dirbuf, size_t dirbuflen,
3783 struct dirent **dpp, size_t *len, off_t *off, int *eofflag)
3785 struct dirent *dp = NULL;
3791 ASSERT_VOP_LOCKED(vp, "vnode not locked");
3792 VNASSERT(vp->v_type == VDIR, vp, ("vnode is not a directory"));
3793 MPASS2((uintptr_t)dirbuf < (uintptr_t)dirbuf + dirbuflen,
3794 "Address space overflow");
3796 if (__predict_false(dirbuflen < GENERIC_MAXDIRSIZ)) {
3797 /* Don't take any chances in this case */
3806 * The caller continued to call us after an error (we set dp to
3807 * NULL in a previous iteration). Bail out right now.
3809 if (__predict_false(dp == NULL))
3812 MPASS(*len <= dirbuflen);
3813 MPASS2((uintptr_t)dirbuf <= (uintptr_t)dp &&
3814 (uintptr_t)dp + *len <= (uintptr_t)dirbuf + dirbuflen,
3815 "Filled range not inside buffer");
3817 reclen = dp->d_reclen;
3818 if (reclen >= *len) {
3819 /* End of buffer reached */
3822 dp = (struct dirent *)((char *)dp + reclen);
3830 /* Have to refill. */
3836 /* Nothing more to read. */
3837 *eofflag = 2; /* Remember the caller reached EOF. */
3841 /* The caller didn't test for EOF. */
3846 iov.iov_base = dirbuf;
3847 iov.iov_len = dirbuflen;
3851 uio.uio_offset = *off;
3852 uio.uio_resid = dirbuflen;
3853 uio.uio_segflg = UIO_SYSSPACE;
3854 uio.uio_rw = UIO_READ;
3858 error = mac_vnode_check_readdir(td->td_ucred, vp);
3861 error = VOP_READDIR(vp, &uio, td->td_ucred, eofflag,
3866 *len = dirbuflen - uio.uio_resid;
3867 *off = uio.uio_offset;
3870 /* Sanity check on INVARIANTS. */
3871 MPASS(*eofflag != 0);
3877 * Normalize the flag returned by VOP_READDIR(), since we use 2
3878 * as a sentinel value.
3883 dp = (struct dirent *)dirbuf;
3886 if (__predict_false(*len < GENERIC_MINDIRSIZ ||
3887 dp->d_reclen < GENERIC_MINDIRSIZ)) {
3901 * Checks whether a directory is empty or not.
3903 * If the directory is empty, returns 0, and if it is not, ENOTEMPTY. Other
3904 * values are genuine errors preventing the check.
3907 vn_dir_check_empty(struct vnode *vp)
3909 struct thread *const td = curthread;
3911 size_t dirbuflen, len;
3917 ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
3918 VNPASS(vp->v_type == VDIR, vp);
3920 error = VOP_GETATTR(vp, &va, td->td_ucred);
3924 dirbuflen = max(DEV_BSIZE, GENERIC_MAXDIRSIZ);
3925 if (dirbuflen < va.va_blocksize)
3926 dirbuflen = va.va_blocksize;
3927 dirbuf = malloc(dirbuflen, M_TEMP, M_WAITOK);
3934 error = vn_dir_next_dirent(vp, td, dirbuf, dirbuflen,
3935 &dp, &len, &off, &eofflag);
3946 * Skip whiteouts. Unionfs operates on filesystems only and
3947 * not on hierarchies, so these whiteouts would be shadowed on
3948 * the system hierarchy but not for a union using the
3949 * filesystem of their directories as the upper layer.
3950 * Additionally, unionfs currently transparently exposes
3951 * union-specific metadata of its upper layer, meaning that
3952 * whiteouts can be seen through the union view in empty
3953 * directories. Taking into account these whiteouts would then
3954 * prevent mounting another filesystem on such effectively
3955 * empty directories.
3957 if (dp->d_type == DT_WHT)
3961 * Any file in the directory which is not '.' or '..' indicates
3962 * the directory is not empty.
3964 switch (dp->d_namlen) {
3966 if (dp->d_name[1] != '.') {
3967 /* Can't be '..' (nor '.') */
3973 if (dp->d_name[0] != '.') {
3974 /* Can't be '..' nor '.' */
3987 free(dirbuf, M_TEMP);
3992 static u_long vn_lock_pair_pause_cnt;
3993 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3994 &vn_lock_pair_pause_cnt, 0,
3995 "Count of vn_lock_pair deadlocks");
3997 u_int vn_lock_pair_pause_max;
3998 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3999 &vn_lock_pair_pause_max, 0,
4000 "Max ticks for vn_lock_pair deadlock avoidance sleep");
4003 vn_lock_pair_pause(const char *wmesg)
4005 atomic_add_long(&vn_lock_pair_pause_cnt, 1);
4006 pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
4010 * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
4011 * vp1_locked indicates whether vp1 is locked; if not, vp1 must be
4012 * unlocked. Same for vp2 and vp2_locked. One of the vnodes can be
4015 * The function returns with both vnodes exclusively or shared locked,
4016 * according to corresponding lkflags, and guarantees that it does not
4017 * create lock order reversal with other threads during its execution.
4018 * Both vnodes could be unlocked temporary (and reclaimed).
4020 * If requesting shared locking, locked vnode lock must not be recursed.
4023 vn_lock_pair(struct vnode *vp1, bool vp1_locked, int lkflags1,
4024 struct vnode *vp2, bool vp2_locked, int lkflags2)
4028 MPASS(lkflags1 == LK_SHARED || lkflags1 == LK_EXCLUSIVE);
4029 MPASS(lkflags2 == LK_SHARED || lkflags2 == LK_EXCLUSIVE);
4031 if (vp1 == NULL && vp2 == NULL)
4035 if (lkflags1 == LK_SHARED &&
4036 (vp1->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
4037 lkflags1 = LK_EXCLUSIVE;
4038 if (vp1_locked && VOP_ISLOCKED(vp1) != LK_EXCLUSIVE) {
4039 ASSERT_VOP_LOCKED(vp1, "vp1");
4040 if (lkflags1 == LK_EXCLUSIVE) {
4042 ASSERT_VOP_UNLOCKED(vp1,
4043 "vp1 shared recursed");
4046 } else if (!vp1_locked)
4047 ASSERT_VOP_UNLOCKED(vp1, "vp1");
4053 if (lkflags2 == LK_SHARED &&
4054 (vp2->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
4055 lkflags2 = LK_EXCLUSIVE;
4056 if (vp2_locked && VOP_ISLOCKED(vp2) != LK_EXCLUSIVE) {
4057 ASSERT_VOP_LOCKED(vp2, "vp2");
4058 if (lkflags2 == LK_EXCLUSIVE) {
4060 ASSERT_VOP_UNLOCKED(vp2,
4061 "vp2 shared recursed");
4064 } else if (!vp2_locked)
4065 ASSERT_VOP_UNLOCKED(vp2, "vp2");
4070 if (!vp1_locked && !vp2_locked) {
4071 vn_lock(vp1, lkflags1 | LK_RETRY);
4075 while (!vp1_locked || !vp2_locked) {
4076 if (vp1_locked && vp2 != NULL) {
4078 error = VOP_LOCK1(vp2, lkflags2 | LK_NOWAIT,
4079 __FILE__, __LINE__);
4084 vn_lock_pair_pause("vlp1");
4086 vn_lock(vp2, lkflags2 | LK_RETRY);
4089 if (vp2_locked && vp1 != NULL) {
4091 error = VOP_LOCK1(vp1, lkflags1 | LK_NOWAIT,
4092 __FILE__, __LINE__);
4097 vn_lock_pair_pause("vlp2");
4099 vn_lock(vp1, lkflags1 | LK_RETRY);
4104 if (lkflags1 == LK_EXCLUSIVE)
4105 ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
4107 ASSERT_VOP_LOCKED(vp1, "vp1 ret");
4110 if (lkflags2 == LK_EXCLUSIVE)
4111 ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
4113 ASSERT_VOP_LOCKED(vp2, "vp2 ret");
4118 vn_lktype_write(struct mount *mp, struct vnode *vp)
4120 if (MNT_SHARED_WRITES(mp) ||
4121 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
4123 return (LK_EXCLUSIVE);