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>
75 #include <sys/sleepqueue.h>
76 #include <sys/sysctl.h>
77 #include <sys/ttycom.h>
79 #include <sys/syslog.h>
80 #include <sys/unistd.h>
83 #include <security/audit/audit.h>
84 #include <security/mac/mac_framework.h>
87 #include <vm/vm_extern.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_object.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_pager.h>
95 #include <sys/pmckern.h>
98 static fo_rdwr_t vn_read;
99 static fo_rdwr_t vn_write;
100 static fo_rdwr_t vn_io_fault;
101 static fo_truncate_t vn_truncate;
102 static fo_ioctl_t vn_ioctl;
103 static fo_poll_t vn_poll;
104 static fo_kqfilter_t vn_kqfilter;
105 static fo_close_t vn_closefile;
106 static fo_mmap_t vn_mmap;
107 static fo_fallocate_t vn_fallocate;
109 struct fileops vnops = {
110 .fo_read = vn_io_fault,
111 .fo_write = vn_io_fault,
112 .fo_truncate = vn_truncate,
113 .fo_ioctl = vn_ioctl,
115 .fo_kqfilter = vn_kqfilter,
116 .fo_stat = vn_statfile,
117 .fo_close = vn_closefile,
118 .fo_chmod = vn_chmod,
119 .fo_chown = vn_chown,
120 .fo_sendfile = vn_sendfile,
122 .fo_fill_kinfo = vn_fill_kinfo,
124 .fo_fallocate = vn_fallocate,
125 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
128 const u_int io_hold_cnt = 16;
129 static int vn_io_fault_enable = 1;
130 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
131 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
132 static int vn_io_fault_prefault = 0;
133 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
134 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
135 static int vn_io_pgcache_read_enable = 1;
136 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
137 &vn_io_pgcache_read_enable, 0,
138 "Enable copying from page cache for reads, avoiding fs");
139 static u_long vn_io_faults_cnt;
140 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
141 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
143 static int vfs_allow_read_dir = 0;
144 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
145 &vfs_allow_read_dir, 0,
146 "Enable read(2) of directory by root for filesystems that support it");
149 * Returns true if vn_io_fault mode of handling the i/o request should
153 do_vn_io_fault(struct vnode *vp, struct uio *uio)
157 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
158 (mp = vp->v_mount) != NULL &&
159 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
163 * Structure used to pass arguments to vn_io_fault1(), to do either
164 * file- or vnode-based I/O calls.
166 struct vn_io_fault_args {
174 struct fop_args_tag {
178 struct vop_args_tag {
184 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
185 struct vn_io_fault_args *args, struct thread *td);
188 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
190 struct thread *td = ndp->ni_cnd.cn_thread;
192 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
196 open2nameif(int fmode, u_int vn_open_flags)
200 res = ISOPEN | LOCKLEAF;
201 if ((fmode & O_RESOLVE_BENEATH) != 0)
203 if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
205 if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
211 * Common code for vnode open operations via a name lookup.
212 * Lookup the vnode and invoke VOP_CREATE if needed.
213 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
215 * Note that this does NOT free nameidata for the successful case,
216 * due to the NDINIT being done elsewhere.
219 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
220 struct ucred *cred, struct file *fp)
224 struct thread *td = ndp->ni_cnd.cn_thread;
226 struct vattr *vap = &vat;
233 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
234 O_EXCL | O_DIRECTORY))
236 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
237 ndp->ni_cnd.cn_nameiop = CREATE;
238 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
240 * Set NOCACHE to avoid flushing the cache when
241 * rolling in many files at once.
243 * Set NC_KEEPPOSENTRY to keep positive entries if they already
244 * exist despite NOCACHE.
246 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
247 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
248 ndp->ni_cnd.cn_flags |= FOLLOW;
249 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
251 if ((error = namei(ndp)) != 0)
253 if (ndp->ni_vp == NULL) {
256 vap->va_mode = cmode;
258 vap->va_vaflags |= VA_EXCLUSIVE;
259 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
260 NDFREE(ndp, NDF_ONLY_PNBUF);
262 if ((error = vn_start_write(NULL, &mp,
263 V_XSLEEP | PCATCH)) != 0)
268 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
269 ndp->ni_cnd.cn_flags |= MAKEENTRY;
271 error = mac_vnode_check_create(cred, ndp->ni_dvp,
275 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
278 if (error == 0 && (fmode & O_EXCL) != 0 &&
279 (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
281 vp->v_iflag |= VI_FOPENING;
285 VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
287 vn_finished_write(mp);
289 NDFREE(ndp, NDF_ONLY_PNBUF);
290 if (error == ERELOOKUP) {
298 if (ndp->ni_dvp == ndp->ni_vp)
304 if (fmode & O_EXCL) {
308 if (vp->v_type == VDIR) {
315 ndp->ni_cnd.cn_nameiop = LOOKUP;
316 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
317 ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
319 if ((fmode & FWRITE) == 0)
320 ndp->ni_cnd.cn_flags |= LOCKSHARED;
321 if ((error = namei(ndp)) != 0)
325 error = vn_open_vnode(vp, fmode, cred, td, fp);
328 vp->v_iflag &= ~VI_FOPENING;
337 NDFREE(ndp, NDF_ONLY_PNBUF);
345 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
348 int error, lock_flags, type;
350 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
351 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
353 KASSERT(fp != NULL, ("open with flock requires fp"));
354 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
357 lock_flags = VOP_ISLOCKED(vp);
360 lf.l_whence = SEEK_SET;
363 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
365 if ((fmode & FNONBLOCK) == 0)
367 if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
369 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
371 fp->f_flag |= FHASLOCK;
373 vn_lock(vp, lock_flags | LK_RETRY);
378 * Common code for vnode open operations once a vnode is located.
379 * Check permissions, and call the VOP_OPEN routine.
382 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
383 struct thread *td, struct file *fp)
388 if (vp->v_type == VLNK) {
389 if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
392 if (vp->v_type == VSOCK)
394 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
398 if ((fmode & O_PATH) == 0) {
399 if ((fmode & (FWRITE | O_TRUNC)) != 0) {
400 if (vp->v_type == VDIR)
404 if ((fmode & FREAD) != 0)
406 if ((fmode & O_APPEND) && (fmode & FWRITE))
409 if ((fmode & O_CREAT) != 0)
413 if ((fmode & FEXEC) != 0)
416 if ((fmode & O_VERIFY) != 0)
418 error = mac_vnode_check_open(cred, vp, accmode);
422 accmode &= ~(VCREAT | VVERIFY);
424 if ((fmode & O_CREAT) == 0 && accmode != 0) {
425 error = VOP_ACCESS(vp, accmode, cred, td);
429 if ((fmode & O_PATH) != 0) {
430 if (vp->v_type == VFIFO)
433 error = VOP_ACCESS(vp, VREAD, cred, td);
435 fp->f_flag |= FKQALLOWED;
439 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
440 vn_lock(vp, LK_UPGRADE | LK_RETRY);
441 error = VOP_OPEN(vp, fmode, cred, td, fp);
445 error = vn_open_vnode_advlock(vp, fmode, fp);
446 if (error == 0 && (fmode & FWRITE) != 0) {
447 error = VOP_ADD_WRITECOUNT(vp, 1);
449 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
450 __func__, vp, vp->v_writecount);
455 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
456 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
461 * Arrange the call by having fdrop() to use
462 * vn_closefile(). This is to satisfy
463 * filesystems like devfs or tmpfs, which
464 * override fo_close().
466 fp->f_flag |= FOPENFAILED;
468 if (fp->f_ops == &badfileops) {
469 fp->f_type = DTYPE_VNODE;
475 * If there is no fp, due to kernel-mode open,
476 * we can call VOP_CLOSE() now.
478 if (vp->v_type != VFIFO && (fmode & FWRITE) != 0 &&
479 !MNT_EXTENDED_SHARED(vp->v_mount) &&
480 VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
481 vn_lock(vp, LK_UPGRADE | LK_RETRY);
482 (void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
487 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
493 * Check for write permissions on the specified vnode.
494 * Prototype text segments cannot be written.
498 vn_writechk(struct vnode *vp)
501 ASSERT_VOP_LOCKED(vp, "vn_writechk");
503 * If there's shared text associated with
504 * the vnode, try to free it up once. If
505 * we fail, we can't allow writing.
517 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
518 struct thread *td, bool keep_ref)
521 int error, lock_flags;
523 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
524 MNT_EXTENDED_SHARED(vp->v_mount))
525 lock_flags = LK_SHARED;
527 lock_flags = LK_EXCLUSIVE;
529 vn_start_write(vp, &mp, V_WAIT);
530 vn_lock(vp, lock_flags | LK_RETRY);
531 AUDIT_ARG_VNODE1(vp);
532 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
533 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
534 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
535 __func__, vp, vp->v_writecount);
537 error = VOP_CLOSE(vp, flags, file_cred, td);
542 vn_finished_write(mp);
547 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
551 return (vn_close1(vp, flags, file_cred, td, false));
555 * Heuristic to detect sequential operation.
558 sequential_heuristic(struct uio *uio, struct file *fp)
562 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
565 if (fp->f_flag & FRDAHEAD)
566 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
569 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
570 * that the first I/O is normally considered to be slightly
571 * sequential. Seeking to offset 0 doesn't change sequentiality
572 * unless previous seeks have reduced f_seqcount to 0, in which
573 * case offset 0 is not special.
575 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
576 uio->uio_offset == fp->f_nextoff[rw]) {
578 * f_seqcount is in units of fixed-size blocks so that it
579 * depends mainly on the amount of sequential I/O and not
580 * much on the number of sequential I/O's. The fixed size
581 * of 16384 is hard-coded here since it is (not quite) just
582 * a magic size that works well here. This size is more
583 * closely related to the best I/O size for real disks than
584 * to any block size used by software.
586 if (uio->uio_resid >= IO_SEQMAX * 16384)
587 fp->f_seqcount[rw] = IO_SEQMAX;
589 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
590 if (fp->f_seqcount[rw] > IO_SEQMAX)
591 fp->f_seqcount[rw] = IO_SEQMAX;
593 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
596 /* Not sequential. Quickly draw-down sequentiality. */
597 if (fp->f_seqcount[rw] > 1)
598 fp->f_seqcount[rw] = 1;
600 fp->f_seqcount[rw] = 0;
605 * Package up an I/O request on a vnode into a uio and do it.
608 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
609 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
610 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
617 struct vn_io_fault_args args;
618 int error, lock_flags;
620 if (offset < 0 && vp->v_type != VCHR)
622 auio.uio_iov = &aiov;
624 aiov.iov_base = base;
626 auio.uio_resid = len;
627 auio.uio_offset = offset;
628 auio.uio_segflg = segflg;
633 if ((ioflg & IO_NODELOCKED) == 0) {
634 if ((ioflg & IO_RANGELOCKED) == 0) {
635 if (rw == UIO_READ) {
636 rl_cookie = vn_rangelock_rlock(vp, offset,
638 } else if ((ioflg & IO_APPEND) != 0) {
639 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
641 rl_cookie = vn_rangelock_wlock(vp, offset,
647 if (rw == UIO_WRITE) {
648 if (vp->v_type != VCHR &&
649 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
652 if (MNT_SHARED_WRITES(mp) ||
653 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
654 lock_flags = LK_SHARED;
656 lock_flags = LK_EXCLUSIVE;
658 lock_flags = LK_SHARED;
659 vn_lock(vp, lock_flags | LK_RETRY);
663 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
665 if ((ioflg & IO_NOMACCHECK) == 0) {
667 error = mac_vnode_check_read(active_cred, file_cred,
670 error = mac_vnode_check_write(active_cred, file_cred,
675 if (file_cred != NULL)
679 if (do_vn_io_fault(vp, &auio)) {
680 args.kind = VN_IO_FAULT_VOP;
683 args.args.vop_args.vp = vp;
684 error = vn_io_fault1(vp, &auio, &args, td);
685 } else if (rw == UIO_READ) {
686 error = VOP_READ(vp, &auio, ioflg, cred);
687 } else /* if (rw == UIO_WRITE) */ {
688 error = VOP_WRITE(vp, &auio, ioflg, cred);
692 *aresid = auio.uio_resid;
694 if (auio.uio_resid && error == 0)
696 if ((ioflg & IO_NODELOCKED) == 0) {
699 vn_finished_write(mp);
702 if (rl_cookie != NULL)
703 vn_rangelock_unlock(vp, rl_cookie);
708 * Package up an I/O request on a vnode into a uio and do it. The I/O
709 * request is split up into smaller chunks and we try to avoid saturating
710 * the buffer cache while potentially holding a vnode locked, so we
711 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
712 * to give other processes a chance to lock the vnode (either other processes
713 * core'ing the same binary, or unrelated processes scanning the directory).
716 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
717 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
718 struct ucred *file_cred, size_t *aresid, struct thread *td)
727 * Force `offset' to a multiple of MAXBSIZE except possibly
728 * for the first chunk, so that filesystems only need to
729 * write full blocks except possibly for the first and last
732 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
736 if (rw != UIO_READ && vp->v_type == VREG)
739 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
740 ioflg, active_cred, file_cred, &iaresid, td);
741 len -= chunk; /* aresid calc already includes length */
745 base = (char *)base + chunk;
746 kern_yield(PRI_USER);
749 *aresid = len + iaresid;
753 #if OFF_MAX <= LONG_MAX
755 foffset_lock(struct file *fp, int flags)
757 volatile short *flagsp;
761 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
763 if ((flags & FOF_NOLOCK) != 0)
764 return (atomic_load_long(&fp->f_offset));
767 * According to McKusick the vn lock was protecting f_offset here.
768 * It is now protected by the FOFFSET_LOCKED flag.
770 flagsp = &fp->f_vnread_flags;
771 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
772 return (atomic_load_long(&fp->f_offset));
774 sleepq_lock(&fp->f_vnread_flags);
775 state = atomic_load_16(flagsp);
777 if ((state & FOFFSET_LOCKED) == 0) {
778 if (!atomic_fcmpset_acq_16(flagsp, &state,
783 if ((state & FOFFSET_LOCK_WAITING) == 0) {
784 if (!atomic_fcmpset_acq_16(flagsp, &state,
785 state | FOFFSET_LOCK_WAITING))
789 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
790 sleepq_wait(&fp->f_vnread_flags, PUSER -1);
792 sleepq_lock(&fp->f_vnread_flags);
793 state = atomic_load_16(flagsp);
795 res = atomic_load_long(&fp->f_offset);
796 sleepq_release(&fp->f_vnread_flags);
801 foffset_unlock(struct file *fp, off_t val, int flags)
803 volatile short *flagsp;
806 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
808 if ((flags & FOF_NOUPDATE) == 0)
809 atomic_store_long(&fp->f_offset, val);
810 if ((flags & FOF_NEXTOFF_R) != 0)
811 fp->f_nextoff[UIO_READ] = val;
812 if ((flags & FOF_NEXTOFF_W) != 0)
813 fp->f_nextoff[UIO_WRITE] = val;
815 if ((flags & FOF_NOLOCK) != 0)
818 flagsp = &fp->f_vnread_flags;
819 state = atomic_load_16(flagsp);
820 if ((state & FOFFSET_LOCK_WAITING) == 0 &&
821 atomic_cmpset_rel_16(flagsp, state, 0))
824 sleepq_lock(&fp->f_vnread_flags);
825 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
826 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
827 fp->f_vnread_flags = 0;
828 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
829 sleepq_release(&fp->f_vnread_flags);
833 foffset_lock(struct file *fp, int flags)
838 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
840 mtxp = mtx_pool_find(mtxpool_sleep, fp);
842 if ((flags & FOF_NOLOCK) == 0) {
843 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
844 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
845 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
848 fp->f_vnread_flags |= FOFFSET_LOCKED;
856 foffset_unlock(struct file *fp, off_t val, int flags)
860 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
862 mtxp = mtx_pool_find(mtxpool_sleep, fp);
864 if ((flags & FOF_NOUPDATE) == 0)
866 if ((flags & FOF_NEXTOFF_R) != 0)
867 fp->f_nextoff[UIO_READ] = val;
868 if ((flags & FOF_NEXTOFF_W) != 0)
869 fp->f_nextoff[UIO_WRITE] = val;
870 if ((flags & FOF_NOLOCK) == 0) {
871 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
872 ("Lost FOFFSET_LOCKED"));
873 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
874 wakeup(&fp->f_vnread_flags);
875 fp->f_vnread_flags = 0;
882 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
885 if ((flags & FOF_OFFSET) == 0)
886 uio->uio_offset = foffset_lock(fp, flags);
890 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
893 if ((flags & FOF_OFFSET) == 0)
894 foffset_unlock(fp, uio->uio_offset, flags);
898 get_advice(struct file *fp, struct uio *uio)
903 ret = POSIX_FADV_NORMAL;
904 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
907 mtxp = mtx_pool_find(mtxpool_sleep, fp);
909 if (fp->f_advice != NULL &&
910 uio->uio_offset >= fp->f_advice->fa_start &&
911 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
912 ret = fp->f_advice->fa_advice;
918 vn_read_from_obj(struct vnode *vp, struct uio *uio)
921 vm_page_t ma[io_hold_cnt + 2];
926 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
927 obj = atomic_load_ptr(&vp->v_object);
929 return (EJUSTRETURN);
932 * Depends on type stability of vm_objects.
934 vm_object_pip_add(obj, 1);
935 if ((obj->flags & OBJ_DEAD) != 0) {
937 * Note that object might be already reused from the
938 * vnode, and the OBJ_DEAD flag cleared. This is fine,
939 * we recheck for DOOMED vnode state after all pages
940 * are busied, and retract then.
942 * But we check for OBJ_DEAD to ensure that we do not
943 * busy pages while vm_object_terminate_pages()
944 * processes the queue.
950 resid = uio->uio_resid;
951 off = uio->uio_offset;
952 for (i = 0; resid > 0; i++) {
953 MPASS(i < io_hold_cnt + 2);
954 ma[i] = vm_page_grab_unlocked(obj, atop(off),
955 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
961 * Skip invalid pages. Valid mask can be partial only
962 * at EOF, and we clip later.
964 if (vm_page_none_valid(ma[i])) {
965 vm_page_sunbusy(ma[i]);
978 * Check VIRF_DOOMED after we busied our pages. Since
979 * vgonel() terminates the vnode' vm_object, it cannot
980 * process past pages busied by us.
982 if (VN_IS_DOOMED(vp)) {
987 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
988 if (resid > uio->uio_resid)
989 resid = uio->uio_resid;
992 * Unlocked read of vnp_size is safe because truncation cannot
993 * pass busied page. But we load vnp_size into a local
994 * variable so that possible concurrent extension does not
997 #if defined(__powerpc__) && !defined(__powerpc64__)
998 vsz = obj->un_pager.vnp.vnp_size;
1000 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1002 if (uio->uio_offset >= vsz) {
1003 error = EJUSTRETURN;
1006 if (uio->uio_offset + resid > vsz)
1007 resid = vsz - uio->uio_offset;
1009 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1012 for (j = 0; j < i; j++) {
1014 vm_page_reference(ma[j]);
1015 vm_page_sunbusy(ma[j]);
1018 vm_object_pip_wakeup(obj);
1021 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1025 * File table vnode read routine.
1028 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1036 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1038 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1041 if (fp->f_flag & FNONBLOCK)
1042 ioflag |= IO_NDELAY;
1043 if (fp->f_flag & O_DIRECT)
1044 ioflag |= IO_DIRECT;
1047 * Try to read from page cache. VIRF_DOOMED check is racy but
1048 * allows us to avoid unneeded work outright.
1050 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1051 (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1052 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1054 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1057 if (error != EJUSTRETURN)
1061 advice = get_advice(fp, uio);
1062 vn_lock(vp, LK_SHARED | LK_RETRY);
1065 case POSIX_FADV_NORMAL:
1066 case POSIX_FADV_SEQUENTIAL:
1067 case POSIX_FADV_NOREUSE:
1068 ioflag |= sequential_heuristic(uio, fp);
1070 case POSIX_FADV_RANDOM:
1071 /* Disable read-ahead for random I/O. */
1074 orig_offset = uio->uio_offset;
1077 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1080 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1081 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1083 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1084 orig_offset != uio->uio_offset)
1086 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1087 * for the backing file after a POSIX_FADV_NOREUSE
1090 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1091 POSIX_FADV_DONTNEED);
1096 * File table vnode write routine.
1099 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1105 int error, ioflag, lock_flags;
1108 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1110 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1112 if (vp->v_type == VREG)
1115 if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
1116 ioflag |= IO_APPEND;
1117 if (fp->f_flag & FNONBLOCK)
1118 ioflag |= IO_NDELAY;
1119 if (fp->f_flag & O_DIRECT)
1120 ioflag |= IO_DIRECT;
1121 if ((fp->f_flag & O_FSYNC) ||
1122 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
1125 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
1126 * implementations that don't understand IO_DATASYNC fall back to full
1129 if (fp->f_flag & O_DSYNC)
1130 ioflag |= IO_SYNC | IO_DATASYNC;
1132 if (vp->v_type != VCHR &&
1133 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
1136 advice = get_advice(fp, uio);
1138 if (MNT_SHARED_WRITES(mp) ||
1139 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
1140 lock_flags = LK_SHARED;
1142 lock_flags = LK_EXCLUSIVE;
1145 vn_lock(vp, lock_flags | LK_RETRY);
1147 case POSIX_FADV_NORMAL:
1148 case POSIX_FADV_SEQUENTIAL:
1149 case POSIX_FADV_NOREUSE:
1150 ioflag |= sequential_heuristic(uio, fp);
1152 case POSIX_FADV_RANDOM:
1153 /* XXX: Is this correct? */
1156 orig_offset = uio->uio_offset;
1159 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1162 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1163 fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1165 if (vp->v_type != VCHR)
1166 vn_finished_write(mp);
1167 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1168 orig_offset != uio->uio_offset)
1170 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1171 * for the backing file after a POSIX_FADV_NOREUSE
1174 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1175 POSIX_FADV_DONTNEED);
1181 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1182 * prevent the following deadlock:
1184 * Assume that the thread A reads from the vnode vp1 into userspace
1185 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
1186 * currently not resident, then system ends up with the call chain
1187 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1188 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1189 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1190 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1191 * backed by the pages of vnode vp1, and some page in buf2 is not
1192 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1194 * To prevent the lock order reversal and deadlock, vn_io_fault() does
1195 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1196 * Instead, it first tries to do the whole range i/o with pagefaults
1197 * disabled. If all pages in the i/o buffer are resident and mapped,
1198 * VOP will succeed (ignoring the genuine filesystem errors).
1199 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1200 * i/o in chunks, with all pages in the chunk prefaulted and held
1201 * using vm_fault_quick_hold_pages().
1203 * Filesystems using this deadlock avoidance scheme should use the
1204 * array of the held pages from uio, saved in the curthread->td_ma,
1205 * instead of doing uiomove(). A helper function
1206 * vn_io_fault_uiomove() converts uiomove request into
1207 * uiomove_fromphys() over td_ma array.
1209 * Since vnode locks do not cover the whole i/o anymore, rangelocks
1210 * make the current i/o request atomic with respect to other i/os and
1215 * Decode vn_io_fault_args and perform the corresponding i/o.
1218 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1224 save = vm_fault_disable_pagefaults();
1225 switch (args->kind) {
1226 case VN_IO_FAULT_FOP:
1227 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1228 uio, args->cred, args->flags, td);
1230 case VN_IO_FAULT_VOP:
1231 if (uio->uio_rw == UIO_READ) {
1232 error = VOP_READ(args->args.vop_args.vp, uio,
1233 args->flags, args->cred);
1234 } else if (uio->uio_rw == UIO_WRITE) {
1235 error = VOP_WRITE(args->args.vop_args.vp, uio,
1236 args->flags, args->cred);
1240 panic("vn_io_fault_doio: unknown kind of io %d %d",
1241 args->kind, uio->uio_rw);
1243 vm_fault_enable_pagefaults(save);
1248 vn_io_fault_touch(char *base, const struct uio *uio)
1253 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1259 vn_io_fault_prefault_user(const struct uio *uio)
1262 const struct iovec *iov;
1267 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1268 ("vn_io_fault_prefault userspace"));
1272 resid = uio->uio_resid;
1273 base = iov->iov_base;
1276 error = vn_io_fault_touch(base, uio);
1279 if (len < PAGE_SIZE) {
1281 error = vn_io_fault_touch(base + len - 1, uio);
1286 if (++i >= uio->uio_iovcnt)
1288 iov = uio->uio_iov + i;
1289 base = iov->iov_base;
1301 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1302 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1303 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1304 * into args and call vn_io_fault1() to handle faults during the user
1305 * mode buffer accesses.
1308 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1311 vm_page_t ma[io_hold_cnt + 2];
1312 struct uio *uio_clone, short_uio;
1313 struct iovec short_iovec[1];
1314 vm_page_t *prev_td_ma;
1316 vm_offset_t addr, end;
1319 int error, cnt, saveheld, prev_td_ma_cnt;
1321 if (vn_io_fault_prefault) {
1322 error = vn_io_fault_prefault_user(uio);
1324 return (error); /* Or ignore ? */
1327 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1330 * The UFS follows IO_UNIT directive and replays back both
1331 * uio_offset and uio_resid if an error is encountered during the
1332 * operation. But, since the iovec may be already advanced,
1333 * uio is still in an inconsistent state.
1335 * Cache a copy of the original uio, which is advanced to the redo
1336 * point using UIO_NOCOPY below.
1338 uio_clone = cloneuio(uio);
1339 resid = uio->uio_resid;
1341 short_uio.uio_segflg = UIO_USERSPACE;
1342 short_uio.uio_rw = uio->uio_rw;
1343 short_uio.uio_td = uio->uio_td;
1345 error = vn_io_fault_doio(args, uio, td);
1346 if (error != EFAULT)
1349 atomic_add_long(&vn_io_faults_cnt, 1);
1350 uio_clone->uio_segflg = UIO_NOCOPY;
1351 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1352 uio_clone->uio_segflg = uio->uio_segflg;
1354 saveheld = curthread_pflags_set(TDP_UIOHELD);
1355 prev_td_ma = td->td_ma;
1356 prev_td_ma_cnt = td->td_ma_cnt;
1358 while (uio_clone->uio_resid != 0) {
1359 len = uio_clone->uio_iov->iov_len;
1361 KASSERT(uio_clone->uio_iovcnt >= 1,
1362 ("iovcnt underflow"));
1363 uio_clone->uio_iov++;
1364 uio_clone->uio_iovcnt--;
1367 if (len > ptoa(io_hold_cnt))
1368 len = ptoa(io_hold_cnt);
1369 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1370 end = round_page(addr + len);
1375 cnt = atop(end - trunc_page(addr));
1377 * A perfectly misaligned address and length could cause
1378 * both the start and the end of the chunk to use partial
1379 * page. +2 accounts for such a situation.
1381 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1382 addr, len, prot, ma, io_hold_cnt + 2);
1387 short_uio.uio_iov = &short_iovec[0];
1388 short_iovec[0].iov_base = (void *)addr;
1389 short_uio.uio_iovcnt = 1;
1390 short_uio.uio_resid = short_iovec[0].iov_len = len;
1391 short_uio.uio_offset = uio_clone->uio_offset;
1393 td->td_ma_cnt = cnt;
1395 error = vn_io_fault_doio(args, &short_uio, td);
1396 vm_page_unhold_pages(ma, cnt);
1397 adv = len - short_uio.uio_resid;
1399 uio_clone->uio_iov->iov_base =
1400 (char *)uio_clone->uio_iov->iov_base + adv;
1401 uio_clone->uio_iov->iov_len -= adv;
1402 uio_clone->uio_resid -= adv;
1403 uio_clone->uio_offset += adv;
1405 uio->uio_resid -= adv;
1406 uio->uio_offset += adv;
1408 if (error != 0 || adv == 0)
1411 td->td_ma = prev_td_ma;
1412 td->td_ma_cnt = prev_td_ma_cnt;
1413 curthread_pflags_restore(saveheld);
1415 free(uio_clone, M_IOV);
1420 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1421 int flags, struct thread *td)
1426 struct vn_io_fault_args args;
1429 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1433 * The ability to read(2) on a directory has historically been
1434 * allowed for all users, but this can and has been the source of
1435 * at least one security issue in the past. As such, it is now hidden
1436 * away behind a sysctl for those that actually need it to use it, and
1437 * restricted to root when it's turned on to make it relatively safe to
1438 * leave on for longer sessions of need.
1440 if (vp->v_type == VDIR) {
1441 KASSERT(uio->uio_rw == UIO_READ,
1442 ("illegal write attempted on a directory"));
1443 if (!vfs_allow_read_dir)
1445 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1449 foffset_lock_uio(fp, uio, flags);
1450 if (do_vn_io_fault(vp, uio)) {
1451 args.kind = VN_IO_FAULT_FOP;
1452 args.args.fop_args.fp = fp;
1453 args.args.fop_args.doio = doio;
1454 args.cred = active_cred;
1455 args.flags = flags | FOF_OFFSET;
1456 if (uio->uio_rw == UIO_READ) {
1457 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1458 uio->uio_offset + uio->uio_resid);
1459 } else if ((fp->f_flag & O_APPEND) != 0 ||
1460 (flags & FOF_OFFSET) == 0) {
1461 /* For appenders, punt and lock the whole range. */
1462 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1464 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1465 uio->uio_offset + uio->uio_resid);
1467 error = vn_io_fault1(vp, uio, &args, td);
1468 vn_rangelock_unlock(vp, rl_cookie);
1470 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1472 foffset_unlock_uio(fp, uio, flags);
1477 * Helper function to perform the requested uiomove operation using
1478 * the held pages for io->uio_iov[0].iov_base buffer instead of
1479 * copyin/copyout. Access to the pages with uiomove_fromphys()
1480 * instead of iov_base prevents page faults that could occur due to
1481 * pmap_collect() invalidating the mapping created by
1482 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1483 * object cleanup revoking the write access from page mappings.
1485 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1486 * instead of plain uiomove().
1489 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1491 struct uio transp_uio;
1492 struct iovec transp_iov[1];
1498 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1499 uio->uio_segflg != UIO_USERSPACE)
1500 return (uiomove(data, xfersize, uio));
1502 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1503 transp_iov[0].iov_base = data;
1504 transp_uio.uio_iov = &transp_iov[0];
1505 transp_uio.uio_iovcnt = 1;
1506 if (xfersize > uio->uio_resid)
1507 xfersize = uio->uio_resid;
1508 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1509 transp_uio.uio_offset = 0;
1510 transp_uio.uio_segflg = UIO_SYSSPACE;
1512 * Since transp_iov points to data, and td_ma page array
1513 * corresponds to original uio->uio_iov, we need to invert the
1514 * direction of the i/o operation as passed to
1515 * uiomove_fromphys().
1517 switch (uio->uio_rw) {
1519 transp_uio.uio_rw = UIO_READ;
1522 transp_uio.uio_rw = UIO_WRITE;
1525 transp_uio.uio_td = uio->uio_td;
1526 error = uiomove_fromphys(td->td_ma,
1527 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1528 xfersize, &transp_uio);
1529 adv = xfersize - transp_uio.uio_resid;
1531 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1532 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1534 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1536 td->td_ma_cnt -= pgadv;
1537 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1538 uio->uio_iov->iov_len -= adv;
1539 uio->uio_resid -= adv;
1540 uio->uio_offset += adv;
1545 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1549 vm_offset_t iov_base;
1553 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1554 uio->uio_segflg != UIO_USERSPACE)
1555 return (uiomove_fromphys(ma, offset, xfersize, uio));
1557 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1558 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1559 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1560 switch (uio->uio_rw) {
1562 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1566 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1570 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1572 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1574 td->td_ma_cnt -= pgadv;
1575 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1576 uio->uio_iov->iov_len -= cnt;
1577 uio->uio_resid -= cnt;
1578 uio->uio_offset += cnt;
1583 * File table truncate routine.
1586 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1598 * Lock the whole range for truncation. Otherwise split i/o
1599 * might happen partly before and partly after the truncation.
1601 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1602 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1605 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1606 AUDIT_ARG_VNODE1(vp);
1607 if (vp->v_type == VDIR) {
1612 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1616 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1620 vn_finished_write(mp);
1622 vn_rangelock_unlock(vp, rl_cookie);
1623 if (error == ERELOOKUP)
1629 * Truncate a file that is already locked.
1632 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1638 error = VOP_ADD_WRITECOUNT(vp, 1);
1641 vattr.va_size = length;
1643 vattr.va_vaflags |= VA_SYNC;
1644 error = VOP_SETATTR(vp, &vattr, cred);
1645 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1651 * File table vnode stat routine.
1654 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1657 struct vnode *vp = fp->f_vnode;
1660 vn_lock(vp, LK_SHARED | LK_RETRY);
1661 error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1668 * File table vnode ioctl routine.
1671 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1676 struct fiobmap2_arg *bmarg;
1680 switch (vp->v_type) {
1685 vn_lock(vp, LK_SHARED | LK_RETRY);
1686 error = VOP_GETATTR(vp, &vattr, active_cred);
1689 *(int *)data = vattr.va_size - fp->f_offset;
1692 bmarg = (struct fiobmap2_arg *)data;
1693 vn_lock(vp, LK_SHARED | LK_RETRY);
1695 error = mac_vnode_check_read(active_cred, fp->f_cred,
1699 error = VOP_BMAP(vp, bmarg->bn, NULL,
1700 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1707 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1712 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1720 * File table vnode poll routine.
1723 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1730 #if defined(MAC) || defined(AUDIT)
1731 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1732 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1733 AUDIT_ARG_VNODE1(vp);
1734 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1740 error = VOP_POLL(vp, events, fp->f_cred, td);
1745 * Acquire the requested lock and then check for validity. LK_RETRY
1746 * permits vn_lock to return doomed vnodes.
1748 static int __noinline
1749 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1753 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1754 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1757 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1759 if ((flags & LK_RETRY) == 0) {
1770 * Nothing to do if we got the lock.
1776 * Interlock was dropped by the call in _vn_lock.
1778 flags &= ~LK_INTERLOCK;
1780 error = VOP_LOCK1(vp, flags, file, line);
1781 } while (error != 0);
1786 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1790 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1791 ("vn_lock: no locktype (%d passed)", flags));
1792 VNPASS(vp->v_holdcnt > 0, vp);
1793 error = VOP_LOCK1(vp, flags, file, line);
1794 if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1795 return (_vn_lock_fallback(vp, flags, file, line, error));
1800 * File table vnode close routine.
1803 vn_closefile(struct file *fp, struct thread *td)
1811 fp->f_ops = &badfileops;
1812 ref = (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1814 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1816 if (__predict_false(ref)) {
1817 lf.l_whence = SEEK_SET;
1820 lf.l_type = F_UNLCK;
1821 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1828 * Preparing to start a filesystem write operation. If the operation is
1829 * permitted, then we bump the count of operations in progress and
1830 * proceed. If a suspend request is in progress, we wait until the
1831 * suspension is over, and then proceed.
1834 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1836 struct mount_pcpu *mpcpu;
1839 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1840 vfs_op_thread_enter(mp, mpcpu)) {
1841 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1842 vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1843 vfs_op_thread_exit(mp, mpcpu);
1848 mtx_assert(MNT_MTX(mp), MA_OWNED);
1855 * Check on status of suspension.
1857 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1858 mp->mnt_susp_owner != curthread) {
1859 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1860 (flags & PCATCH) : 0) | (PUSER - 1);
1861 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1862 if (flags & V_NOWAIT) {
1863 error = EWOULDBLOCK;
1866 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1872 if (flags & V_XSLEEP)
1874 mp->mnt_writeopcount++;
1876 if (error != 0 || (flags & V_XSLEEP) != 0)
1883 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1888 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1889 ("V_MNTREF requires mp"));
1893 * If a vnode is provided, get and return the mount point that
1894 * to which it will write.
1897 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1899 if (error != EOPNOTSUPP)
1904 if ((mp = *mpp) == NULL)
1908 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1910 * As long as a vnode is not provided we need to acquire a
1911 * refcount for the provided mountpoint too, in order to
1912 * emulate a vfs_ref().
1914 if (vp == NULL && (flags & V_MNTREF) == 0)
1917 return (vn_start_write_refed(mp, flags, false));
1921 * Secondary suspension. Used by operations such as vop_inactive
1922 * routines that are needed by the higher level functions. These
1923 * are allowed to proceed until all the higher level functions have
1924 * completed (indicated by mnt_writeopcount dropping to zero). At that
1925 * time, these operations are halted until the suspension is over.
1928 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1933 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1934 ("V_MNTREF requires mp"));
1938 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1940 if (error != EOPNOTSUPP)
1946 * If we are not suspended or have not yet reached suspended
1947 * mode, then let the operation proceed.
1949 if ((mp = *mpp) == NULL)
1953 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1955 * As long as a vnode is not provided we need to acquire a
1956 * refcount for the provided mountpoint too, in order to
1957 * emulate a vfs_ref().
1960 if (vp == NULL && (flags & V_MNTREF) == 0)
1962 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1963 mp->mnt_secondary_writes++;
1964 mp->mnt_secondary_accwrites++;
1968 if (flags & V_NOWAIT) {
1971 return (EWOULDBLOCK);
1974 * Wait for the suspension to finish.
1976 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1977 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1986 * Filesystem write operation has completed. If we are suspending and this
1987 * operation is the last one, notify the suspender that the suspension is
1991 vn_finished_write(struct mount *mp)
1993 struct mount_pcpu *mpcpu;
1999 if (vfs_op_thread_enter(mp, mpcpu)) {
2000 vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2001 vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2002 vfs_op_thread_exit(mp, mpcpu);
2007 vfs_assert_mount_counters(mp);
2009 c = --mp->mnt_writeopcount;
2010 if (mp->mnt_vfs_ops == 0) {
2011 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2016 vfs_dump_mount_counters(mp);
2017 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2018 wakeup(&mp->mnt_writeopcount);
2023 * Filesystem secondary write operation has completed. If we are
2024 * suspending and this operation is the last one, notify the suspender
2025 * that the suspension is now in effect.
2028 vn_finished_secondary_write(struct mount *mp)
2034 mp->mnt_secondary_writes--;
2035 if (mp->mnt_secondary_writes < 0)
2036 panic("vn_finished_secondary_write: neg cnt");
2037 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2038 mp->mnt_secondary_writes <= 0)
2039 wakeup(&mp->mnt_secondary_writes);
2044 * Request a filesystem to suspend write operations.
2047 vfs_write_suspend(struct mount *mp, int flags)
2054 vfs_assert_mount_counters(mp);
2055 if (mp->mnt_susp_owner == curthread) {
2056 vfs_op_exit_locked(mp);
2060 while (mp->mnt_kern_flag & MNTK_SUSPEND)
2061 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2064 * Unmount holds a write reference on the mount point. If we
2065 * own busy reference and drain for writers, we deadlock with
2066 * the reference draining in the unmount path. Callers of
2067 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2068 * vfs_busy() reference is owned and caller is not in the
2071 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2072 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2073 vfs_op_exit_locked(mp);
2078 mp->mnt_kern_flag |= MNTK_SUSPEND;
2079 mp->mnt_susp_owner = curthread;
2080 if (mp->mnt_writeopcount > 0)
2081 (void) msleep(&mp->mnt_writeopcount,
2082 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2085 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2086 vfs_write_resume(mp, 0);
2087 /* vfs_write_resume does vfs_op_exit() for us */
2093 * Request a filesystem to resume write operations.
2096 vfs_write_resume(struct mount *mp, int flags)
2100 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2101 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2102 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2104 mp->mnt_susp_owner = NULL;
2105 wakeup(&mp->mnt_writeopcount);
2106 wakeup(&mp->mnt_flag);
2107 curthread->td_pflags &= ~TDP_IGNSUSP;
2108 if ((flags & VR_START_WRITE) != 0) {
2110 mp->mnt_writeopcount++;
2113 if ((flags & VR_NO_SUSPCLR) == 0)
2116 } else if ((flags & VR_START_WRITE) != 0) {
2118 vn_start_write_refed(mp, 0, true);
2125 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2129 vfs_write_suspend_umnt(struct mount *mp)
2133 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2134 ("vfs_write_suspend_umnt: recursed"));
2136 /* dounmount() already called vn_start_write(). */
2138 vn_finished_write(mp);
2139 error = vfs_write_suspend(mp, 0);
2141 vn_start_write(NULL, &mp, V_WAIT);
2145 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2148 vn_start_write(NULL, &mp, V_WAIT);
2150 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2151 wakeup(&mp->mnt_flag);
2153 curthread->td_pflags |= TDP_IGNSUSP;
2158 * Implement kqueues for files by translating it to vnode operation.
2161 vn_kqfilter(struct file *fp, struct knote *kn)
2164 return (VOP_KQFILTER(fp->f_vnode, kn));
2168 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2170 if ((fp->f_flag & FKQALLOWED) == 0)
2172 return (vn_kqfilter(fp, kn));
2176 * Simplified in-kernel wrapper calls for extended attribute access.
2177 * Both calls pass in a NULL credential, authorizing as "kernel" access.
2178 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2181 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2182 const char *attrname, int *buflen, char *buf, struct thread *td)
2188 iov.iov_len = *buflen;
2191 auio.uio_iov = &iov;
2192 auio.uio_iovcnt = 1;
2193 auio.uio_rw = UIO_READ;
2194 auio.uio_segflg = UIO_SYSSPACE;
2196 auio.uio_offset = 0;
2197 auio.uio_resid = *buflen;
2199 if ((ioflg & IO_NODELOCKED) == 0)
2200 vn_lock(vp, LK_SHARED | LK_RETRY);
2202 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2204 /* authorize attribute retrieval as kernel */
2205 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2208 if ((ioflg & IO_NODELOCKED) == 0)
2212 *buflen = *buflen - auio.uio_resid;
2219 * XXX failure mode if partially written?
2222 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2223 const char *attrname, int buflen, char *buf, struct thread *td)
2230 iov.iov_len = buflen;
2233 auio.uio_iov = &iov;
2234 auio.uio_iovcnt = 1;
2235 auio.uio_rw = UIO_WRITE;
2236 auio.uio_segflg = UIO_SYSSPACE;
2238 auio.uio_offset = 0;
2239 auio.uio_resid = buflen;
2241 if ((ioflg & IO_NODELOCKED) == 0) {
2242 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2244 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2247 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2249 /* authorize attribute setting as kernel */
2250 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2252 if ((ioflg & IO_NODELOCKED) == 0) {
2253 vn_finished_write(mp);
2261 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2262 const char *attrname, struct thread *td)
2267 if ((ioflg & IO_NODELOCKED) == 0) {
2268 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2270 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2273 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2275 /* authorize attribute removal as kernel */
2276 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2277 if (error == EOPNOTSUPP)
2278 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2281 if ((ioflg & IO_NODELOCKED) == 0) {
2282 vn_finished_write(mp);
2290 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2294 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2298 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2301 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2306 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2307 int lkflags, struct vnode **rvp)
2312 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2314 ltype = VOP_ISLOCKED(vp);
2315 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2316 ("vn_vget_ino: vp not locked"));
2317 error = vfs_busy(mp, MBF_NOWAIT);
2321 error = vfs_busy(mp, 0);
2322 vn_lock(vp, ltype | LK_RETRY);
2326 if (VN_IS_DOOMED(vp)) {
2332 error = alloc(mp, alloc_arg, lkflags, rvp);
2334 if (error != 0 || *rvp != vp)
2335 vn_lock(vp, ltype | LK_RETRY);
2336 if (VN_IS_DOOMED(vp)) {
2349 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2353 if (vp->v_type != VREG || td == NULL)
2355 if ((uoff_t)uio->uio_offset + uio->uio_resid >
2356 lim_cur(td, RLIMIT_FSIZE)) {
2357 PROC_LOCK(td->td_proc);
2358 kern_psignal(td->td_proc, SIGXFSZ);
2359 PROC_UNLOCK(td->td_proc);
2366 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2373 vn_lock(vp, LK_SHARED | LK_RETRY);
2374 AUDIT_ARG_VNODE1(vp);
2377 return (setfmode(td, active_cred, vp, mode));
2381 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2388 vn_lock(vp, LK_SHARED | LK_RETRY);
2389 AUDIT_ARG_VNODE1(vp);
2392 return (setfown(td, active_cred, vp, uid, gid));
2396 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2400 if ((object = vp->v_object) == NULL)
2402 VM_OBJECT_WLOCK(object);
2403 vm_object_page_remove(object, start, end, 0);
2404 VM_OBJECT_WUNLOCK(object);
2408 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2416 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2417 ("Wrong command %lu", cmd));
2419 if (vn_lock(vp, LK_SHARED) != 0)
2421 if (vp->v_type != VREG) {
2425 error = VOP_GETATTR(vp, &va, cred);
2429 if (noff >= va.va_size) {
2433 bsize = vp->v_mount->mnt_stat.f_iosize;
2434 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2436 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2437 if (error == EOPNOTSUPP) {
2441 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2442 (bnp != -1 && cmd == FIOSEEKDATA)) {
2449 if (noff > va.va_size)
2451 /* noff == va.va_size. There is an implicit hole at the end of file. */
2452 if (cmd == FIOSEEKDATA)
2462 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2467 off_t foffset, size;
2470 cred = td->td_ucred;
2472 foffset = foffset_lock(fp, 0);
2473 noneg = (vp->v_type != VCHR);
2479 (offset > 0 && foffset > OFF_MAX - offset))) {
2486 vn_lock(vp, LK_SHARED | LK_RETRY);
2487 error = VOP_GETATTR(vp, &vattr, cred);
2493 * If the file references a disk device, then fetch
2494 * the media size and use that to determine the ending
2497 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2498 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2499 vattr.va_size = size;
2501 (vattr.va_size > OFF_MAX ||
2502 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2506 offset += vattr.va_size;
2511 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2512 if (error == ENOTTY)
2516 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2517 if (error == ENOTTY)
2523 if (error == 0 && noneg && offset < 0)
2527 VFS_KNOTE_UNLOCKED(vp, 0);
2528 td->td_uretoff.tdu_off = offset;
2530 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2535 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2541 * Grant permission if the caller is the owner of the file, or
2542 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2543 * on the file. If the time pointer is null, then write
2544 * permission on the file is also sufficient.
2546 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2547 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2548 * will be allowed to set the times [..] to the current
2551 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2552 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2553 error = VOP_ACCESS(vp, VWRITE, cred, td);
2558 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2563 if (fp->f_type == DTYPE_FIFO)
2564 kif->kf_type = KF_TYPE_FIFO;
2566 kif->kf_type = KF_TYPE_VNODE;
2569 FILEDESC_SUNLOCK(fdp);
2570 error = vn_fill_kinfo_vnode(vp, kif);
2572 FILEDESC_SLOCK(fdp);
2577 vn_fill_junk(struct kinfo_file *kif)
2582 * Simulate vn_fullpath returning changing values for a given
2583 * vp during e.g. coredump.
2585 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2586 olen = strlen(kif->kf_path);
2588 strcpy(&kif->kf_path[len - 1], "$");
2590 for (; olen < len; olen++)
2591 strcpy(&kif->kf_path[olen], "A");
2595 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2598 char *fullpath, *freepath;
2601 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2604 error = vn_fullpath(vp, &fullpath, &freepath);
2606 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2608 if (freepath != NULL)
2609 free(freepath, M_TEMP);
2611 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2616 * Retrieve vnode attributes.
2618 va.va_fsid = VNOVAL;
2620 vn_lock(vp, LK_SHARED | LK_RETRY);
2621 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2625 if (va.va_fsid != VNOVAL)
2626 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2628 kif->kf_un.kf_file.kf_file_fsid =
2629 vp->v_mount->mnt_stat.f_fsid.val[0];
2630 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2631 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2632 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2633 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2634 kif->kf_un.kf_file.kf_file_size = va.va_size;
2635 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2636 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2637 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2642 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2643 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2647 struct pmckern_map_in pkm;
2653 boolean_t writecounted;
2656 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2657 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2659 * POSIX shared-memory objects are defined to have
2660 * kernel persistence, and are not defined to support
2661 * read(2)/write(2) -- or even open(2). Thus, we can
2662 * use MAP_ASYNC to trade on-disk coherence for speed.
2663 * The shm_open(3) library routine turns on the FPOSIXSHM
2664 * flag to request this behavior.
2666 if ((fp->f_flag & FPOSIXSHM) != 0)
2667 flags |= MAP_NOSYNC;
2672 * Ensure that file and memory protections are
2673 * compatible. Note that we only worry about
2674 * writability if mapping is shared; in this case,
2675 * current and max prot are dictated by the open file.
2676 * XXX use the vnode instead? Problem is: what
2677 * credentials do we use for determination? What if
2678 * proc does a setuid?
2681 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2682 maxprot = VM_PROT_NONE;
2683 if ((prot & VM_PROT_EXECUTE) != 0)
2686 maxprot = VM_PROT_EXECUTE;
2687 if ((fp->f_flag & FREAD) != 0)
2688 maxprot |= VM_PROT_READ;
2689 else if ((prot & VM_PROT_READ) != 0)
2693 * If we are sharing potential changes via MAP_SHARED and we
2694 * are trying to get write permission although we opened it
2695 * without asking for it, bail out.
2697 if ((flags & MAP_SHARED) != 0) {
2698 if ((fp->f_flag & FWRITE) != 0)
2699 maxprot |= VM_PROT_WRITE;
2700 else if ((prot & VM_PROT_WRITE) != 0)
2703 maxprot |= VM_PROT_WRITE;
2704 cap_maxprot |= VM_PROT_WRITE;
2706 maxprot &= cap_maxprot;
2709 * For regular files and shared memory, POSIX requires that
2710 * the value of foff be a legitimate offset within the data
2711 * object. In particular, negative offsets are invalid.
2712 * Blocking negative offsets and overflows here avoids
2713 * possible wraparound or user-level access into reserved
2714 * ranges of the data object later. In contrast, POSIX does
2715 * not dictate how offsets are used by device drivers, so in
2716 * the case of a device mapping a negative offset is passed
2723 foff > OFF_MAX - size)
2726 writecounted = FALSE;
2727 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2728 &foff, &object, &writecounted);
2731 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2732 foff, writecounted, td);
2735 * If this mapping was accounted for in the vnode's
2736 * writecount, then undo that now.
2739 vm_pager_release_writecount(object, 0, size);
2740 vm_object_deallocate(object);
2743 /* Inform hwpmc(4) if an executable is being mapped. */
2744 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2745 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2747 pkm.pm_address = (uintptr_t) *addr;
2748 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2756 vn_fsid(struct vnode *vp, struct vattr *va)
2760 f = &vp->v_mount->mnt_stat.f_fsid;
2761 va->va_fsid = (uint32_t)f->val[1];
2762 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2763 va->va_fsid += (uint32_t)f->val[0];
2767 vn_fsync_buf(struct vnode *vp, int waitfor)
2769 struct buf *bp, *nbp;
2772 int error, maxretry;
2775 maxretry = 10000; /* large, arbitrarily chosen */
2777 if (vp->v_type == VCHR) {
2779 mp = vp->v_rdev->si_mountpt;
2786 * MARK/SCAN initialization to avoid infinite loops.
2788 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2789 bp->b_vflags &= ~BV_SCANNED;
2794 * Flush all dirty buffers associated with a vnode.
2797 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2798 if ((bp->b_vflags & BV_SCANNED) != 0)
2800 bp->b_vflags |= BV_SCANNED;
2801 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2802 if (waitfor != MNT_WAIT)
2805 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2806 BO_LOCKPTR(bo)) != 0) {
2813 KASSERT(bp->b_bufobj == bo,
2814 ("bp %p wrong b_bufobj %p should be %p",
2815 bp, bp->b_bufobj, bo));
2816 if ((bp->b_flags & B_DELWRI) == 0)
2817 panic("fsync: not dirty");
2818 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2824 if (maxretry < 1000)
2825 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2831 * If synchronous the caller expects us to completely resolve all
2832 * dirty buffers in the system. Wait for in-progress I/O to
2833 * complete (which could include background bitmap writes), then
2834 * retry if dirty blocks still exist.
2836 if (waitfor == MNT_WAIT) {
2837 bufobj_wwait(bo, 0, 0);
2838 if (bo->bo_dirty.bv_cnt > 0) {
2840 * If we are unable to write any of these buffers
2841 * then we fail now rather than trying endlessly
2842 * to write them out.
2844 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2845 if ((error = bp->b_error) != 0)
2847 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2848 (error == 0 && --maxretry >= 0))
2856 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2862 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
2863 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2864 * to do the actual copy.
2865 * vn_generic_copy_file_range() is factored out, so it can be called
2866 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2867 * different file systems.
2870 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2871 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2872 struct ucred *outcred, struct thread *fsize_td)
2879 *lenp = 0; /* For error returns. */
2882 /* Do some sanity checks on the arguments. */
2883 if (invp->v_type == VDIR || outvp->v_type == VDIR)
2885 else if (*inoffp < 0 || *outoffp < 0 ||
2886 invp->v_type != VREG || outvp->v_type != VREG)
2891 /* Ensure offset + len does not wrap around. */
2894 if (uval > INT64_MAX)
2895 len = INT64_MAX - *inoffp;
2898 if (uval > INT64_MAX)
2899 len = INT64_MAX - *outoffp;
2904 * If the two vnode are for the same file system, call
2905 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2906 * which can handle copies across multiple file systems.
2909 if (invp->v_mount == outvp->v_mount)
2910 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2911 lenp, flags, incred, outcred, fsize_td);
2913 error = vn_generic_copy_file_range(invp, inoffp, outvp,
2914 outoffp, lenp, flags, incred, outcred, fsize_td);
2920 * Test len bytes of data starting at dat for all bytes == 0.
2921 * Return true if all bytes are zero, false otherwise.
2922 * Expects dat to be well aligned.
2925 mem_iszero(void *dat, int len)
2931 for (p = dat; len > 0; len -= sizeof(*p), p++) {
2932 if (len >= sizeof(*p)) {
2936 cp = (const char *)p;
2937 for (i = 0; i < len; i++, cp++)
2946 * Look for a hole in the output file and, if found, adjust *outoffp
2947 * and *xferp to skip past the hole.
2948 * *xferp is the entire hole length to be written and xfer2 is how many bytes
2949 * to be written as 0's upon return.
2952 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2953 off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2958 if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2959 *dataoffp = *outoffp;
2960 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2963 *holeoffp = *dataoffp;
2964 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2967 if (error != 0 || *holeoffp == *dataoffp) {
2969 * Since outvp is unlocked, it may be possible for
2970 * another thread to do a truncate(), lseek(), write()
2971 * creating a hole at startoff between the above
2972 * VOP_IOCTL() calls, if the other thread does not do
2974 * If that happens, *holeoffp == *dataoffp and finding
2975 * the hole has failed, so disable vn_skip_hole().
2977 *holeoffp = -1; /* Disable use of vn_skip_hole(). */
2980 KASSERT(*dataoffp >= *outoffp,
2981 ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2982 (intmax_t)*dataoffp, (intmax_t)*outoffp));
2983 KASSERT(*holeoffp > *dataoffp,
2984 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2985 (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2989 * If there is a hole before the data starts, advance *outoffp and
2990 * *xferp past the hole.
2992 if (*dataoffp > *outoffp) {
2993 delta = *dataoffp - *outoffp;
2994 if (delta >= *xferp) {
2995 /* Entire *xferp is a hole. */
3002 xfer2 = MIN(xfer2, *xferp);
3006 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3007 * that the write ends at the start of the hole.
3008 * *holeoffp should always be greater than *outoffp, but for the
3009 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3012 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3013 xfer2 = *holeoffp - *outoffp;
3018 * Write an xfer sized chunk to outvp in blksize blocks from dat.
3019 * dat is a maximum of blksize in length and can be written repeatedly in
3021 * If growfile == true, just grow the file via vn_truncate_locked() instead
3022 * of doing actual writes.
3023 * If checkhole == true, a hole is being punched, so skip over any hole
3024 * already in the output file.
3027 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3028 u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3031 off_t dataoff, holeoff, xfer2;
3035 * Loop around doing writes of blksize until write has been completed.
3036 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3037 * done for each iteration, since the xfer argument can be very
3038 * large if there is a large hole to punch in the output file.
3043 xfer2 = MIN(xfer, blksize);
3046 * Punching a hole. Skip writing if there is
3047 * already a hole in the output file.
3049 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3050 &dataoff, &holeoff, cred);
3055 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3060 error = vn_start_write(outvp, &mp, V_WAIT);
3064 error = vn_lock(outvp, LK_EXCLUSIVE);
3066 error = vn_truncate_locked(outvp, outoff + xfer,
3071 if (MNT_SHARED_WRITES(mp))
3074 lckf = LK_EXCLUSIVE;
3075 error = vn_lock(outvp, lckf);
3077 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3078 outoff, UIO_SYSSPACE, IO_NODELOCKED,
3079 curthread->td_ucred, cred, NULL, curthread);
3086 vn_finished_write(mp);
3087 } while (!growfile && xfer > 0 && error == 0);
3092 * Copy a byte range of one file to another. This function can handle the
3093 * case where invp and outvp are on different file systems.
3094 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3095 * is no better file system specific way to do it.
3098 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3099 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3100 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3105 off_t startoff, endoff, xfer, xfer2;
3107 int error, interrupted;
3108 bool cantseek, readzeros, eof, lastblock;
3110 size_t copylen, len, rem, savlen;
3112 long holein, holeout;
3114 holein = holeout = 0;
3115 savlen = len = *lenp;
3120 error = vn_lock(invp, LK_SHARED);
3123 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3128 error = vn_start_write(outvp, &mp, V_WAIT);
3130 error = vn_lock(outvp, LK_EXCLUSIVE);
3133 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3134 * now that outvp is locked.
3136 if (fsize_td != NULL) {
3137 io.uio_offset = *outoffp;
3139 error = vn_rlimit_fsize(outvp, &io, fsize_td);
3143 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3146 * Holes that are past EOF do not need to be written as a block
3147 * of zero bytes. So, truncate the output file as far as
3148 * possible and then use va.va_size to decide if writing 0
3149 * bytes is necessary in the loop below.
3152 error = VOP_GETATTR(outvp, &va, outcred);
3153 if (error == 0 && va.va_size > *outoffp && va.va_size <=
3156 error = mac_vnode_check_write(curthread->td_ucred,
3160 error = vn_truncate_locked(outvp, *outoffp,
3163 va.va_size = *outoffp;
3168 vn_finished_write(mp);
3173 * Set the blksize to the larger of the hole sizes for invp and outvp.
3174 * If hole sizes aren't available, set the blksize to the larger
3175 * f_iosize of invp and outvp.
3176 * This code expects the hole sizes and f_iosizes to be powers of 2.
3177 * This value is clipped at 4Kbytes and 1Mbyte.
3179 blksize = MAX(holein, holeout);
3181 /* Clip len to end at an exact multiple of hole size. */
3183 rem = *inoffp % blksize;
3185 rem = blksize - rem;
3186 if (len > rem && len - rem > blksize)
3187 len = savlen = rounddown(len - rem, blksize) + rem;
3191 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3192 outvp->v_mount->mnt_stat.f_iosize);
3195 else if (blksize > 1024 * 1024)
3196 blksize = 1024 * 1024;
3197 dat = malloc(blksize, M_TEMP, M_WAITOK);
3200 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3201 * to find holes. Otherwise, just scan the read block for all 0s
3202 * in the inner loop where the data copying is done.
3203 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3204 * support holes on the server, but do not support FIOSEEKHOLE.
3207 while (len > 0 && error == 0 && !eof && interrupted == 0) {
3208 endoff = 0; /* To shut up compilers. */
3214 * Find the next data area. If there is just a hole to EOF,
3215 * FIOSEEKDATA should fail and then we drop down into the
3216 * inner loop and create the hole on the outvp file.
3217 * (I do not know if any file system will report a hole to
3218 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3219 * will fail for those file systems.)
3221 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3222 * the code just falls through to the inner copy loop.
3226 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3230 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3233 * Since invp is unlocked, it may be possible for
3234 * another thread to do a truncate(), lseek(), write()
3235 * creating a hole at startoff between the above
3236 * VOP_IOCTL() calls, if the other thread does not do
3238 * If that happens, startoff == endoff and finding
3239 * the hole has failed, so set an error.
3241 if (error == 0 && startoff == endoff)
3242 error = EINVAL; /* Any error. Reset to 0. */
3245 if (startoff > *inoffp) {
3246 /* Found hole before data block. */
3247 xfer = MIN(startoff - *inoffp, len);
3248 if (*outoffp < va.va_size) {
3249 /* Must write 0s to punch hole. */
3250 xfer2 = MIN(va.va_size - *outoffp,
3252 memset(dat, 0, MIN(xfer2, blksize));
3253 error = vn_write_outvp(outvp, dat,
3254 *outoffp, xfer2, blksize, false,
3255 holeout > 0, outcred);
3258 if (error == 0 && *outoffp + xfer >
3259 va.va_size && xfer == len)
3260 /* Grow last block. */
3261 error = vn_write_outvp(outvp, dat,
3262 *outoffp, xfer, blksize, true,
3269 interrupted = sig_intr();
3272 copylen = MIN(len, endoff - startoff);
3284 * Set first xfer to end at a block boundary, so that
3285 * holes are more likely detected in the loop below via
3286 * the for all bytes 0 method.
3288 xfer -= (*inoffp % blksize);
3290 /* Loop copying the data block. */
3291 while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3294 error = vn_lock(invp, LK_SHARED);
3297 error = vn_rdwr(UIO_READ, invp, dat, xfer,
3298 startoff, UIO_SYSSPACE, IO_NODELOCKED,
3299 curthread->td_ucred, incred, &aresid,
3303 if (error == 0 && aresid > 0) {
3304 /* Stop the copy at EOF on the input file. */
3311 * Skip the write for holes past the initial EOF
3312 * of the output file, unless this is the last
3313 * write of the output file at EOF.
3315 readzeros = cantseek ? mem_iszero(dat, xfer) :
3319 if (!cantseek || *outoffp < va.va_size ||
3320 lastblock || !readzeros)
3321 error = vn_write_outvp(outvp, dat,
3322 *outoffp, xfer, blksize,
3323 readzeros && lastblock &&
3324 *outoffp >= va.va_size, false,
3333 interrupted = sig_intr();
3340 *lenp = savlen - len;
3346 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3350 off_t olen, ooffset;
3353 int audited_vnode1 = 0;
3357 if (vp->v_type != VREG)
3360 /* Allocating blocks may take a long time, so iterate. */
3367 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3370 error = vn_lock(vp, LK_EXCLUSIVE);
3372 vn_finished_write(mp);
3376 if (!audited_vnode1) {
3377 AUDIT_ARG_VNODE1(vp);
3382 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3385 error = VOP_ALLOCATE(vp, &offset, &len);
3387 vn_finished_write(mp);
3389 if (olen + ooffset != offset + len) {
3390 panic("offset + len changed from %jx/%jx to %jx/%jx",
3391 ooffset, olen, offset, len);
3393 if (error != 0 || len == 0)
3395 KASSERT(olen > len, ("Iteration did not make progress?"));
3402 static u_long vn_lock_pair_pause_cnt;
3403 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3404 &vn_lock_pair_pause_cnt, 0,
3405 "Count of vn_lock_pair deadlocks");
3407 u_int vn_lock_pair_pause_max;
3408 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3409 &vn_lock_pair_pause_max, 0,
3410 "Max ticks for vn_lock_pair deadlock avoidance sleep");
3413 vn_lock_pair_pause(const char *wmesg)
3415 atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3416 pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3420 * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3421 * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3422 * must be unlocked. Same for vp2 and vp2_locked. One of the vnodes
3425 * The function returns with both vnodes exclusively locked, and
3426 * guarantees that it does not create lock order reversal with other
3427 * threads during its execution. Both vnodes could be unlocked
3428 * temporary (and reclaimed).
3431 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3436 if (vp1 == NULL && vp2 == NULL)
3440 ASSERT_VOP_ELOCKED(vp1, "vp1");
3442 ASSERT_VOP_UNLOCKED(vp1, "vp1");
3448 ASSERT_VOP_ELOCKED(vp2, "vp2");
3450 ASSERT_VOP_UNLOCKED(vp2, "vp2");
3454 if (!vp1_locked && !vp2_locked) {
3455 vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3460 if (vp1_locked && vp2_locked)
3462 if (vp1_locked && vp2 != NULL) {
3464 error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3465 __FILE__, __LINE__);
3470 vn_lock_pair_pause("vlp1");
3472 vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3475 if (vp2_locked && vp1 != NULL) {
3477 error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3478 __FILE__, __LINE__);
3483 vn_lock_pair_pause("vlp2");
3485 vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3490 ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3492 ASSERT_VOP_ELOCKED(vp2, "vp2 ret");