2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1982, 1986, 1989, 1993
5 * The Regents of the University of California. All rights reserved.
6 * (c) UNIX System Laboratories, Inc.
7 * All or some portions of this file are derived from material licensed
8 * to the University of California by American Telephone and Telegraph
9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10 * the permission of UNIX System Laboratories, Inc.
12 * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13 * Copyright (c) 2013, 2014 The FreeBSD Foundation
15 * Portions of this software were developed by Konstantin Belousov
16 * under sponsorship from the FreeBSD Foundation.
18 * Redistribution and use in source and binary forms, with or without
19 * modification, are permitted provided that the following conditions
21 * 1. Redistributions of source code must retain the above copyright
22 * notice, this list of conditions and the following disclaimer.
23 * 2. Redistributions in binary form must reproduce the above copyright
24 * notice, this list of conditions and the following disclaimer in the
25 * documentation and/or other materials provided with the distribution.
26 * 3. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42 * @(#)vfs_vnops.c 8.2 (Berkeley) 1/21/94
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
48 #include "opt_hwpmc_hooks.h"
50 #include <sys/param.h>
51 #include <sys/systm.h>
54 #include <sys/fcntl.h>
61 #include <sys/limits.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
74 #include <sys/sleepqueue.h>
75 #include <sys/sysctl.h>
76 #include <sys/ttycom.h>
78 #include <sys/syslog.h>
79 #include <sys/unistd.h>
82 #include <security/audit/audit.h>
83 #include <security/mac/mac_framework.h>
86 #include <vm/vm_extern.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pager.h>
94 #include <sys/pmckern.h>
97 static fo_rdwr_t vn_read;
98 static fo_rdwr_t vn_write;
99 static fo_rdwr_t vn_io_fault;
100 static fo_truncate_t vn_truncate;
101 static fo_ioctl_t vn_ioctl;
102 static fo_poll_t vn_poll;
103 static fo_kqfilter_t vn_kqfilter;
104 static fo_stat_t vn_statfile;
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_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;
231 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
232 O_EXCL | O_DIRECTORY))
234 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
235 ndp->ni_cnd.cn_nameiop = CREATE;
236 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
238 * Set NOCACHE to avoid flushing the cache when
239 * rolling in many files at once.
241 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE;
242 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
243 ndp->ni_cnd.cn_flags |= FOLLOW;
244 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
246 if ((error = namei(ndp)) != 0)
248 if (ndp->ni_vp == NULL) {
251 vap->va_mode = cmode;
253 vap->va_vaflags |= VA_EXCLUSIVE;
254 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
255 NDFREE(ndp, NDF_ONLY_PNBUF);
257 if ((error = vn_start_write(NULL, &mp,
258 V_XSLEEP | PCATCH)) != 0)
262 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
263 ndp->ni_cnd.cn_flags |= MAKEENTRY;
265 error = mac_vnode_check_create(cred, ndp->ni_dvp,
269 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
272 vn_finished_write(mp);
274 NDFREE(ndp, NDF_ONLY_PNBUF);
280 if (ndp->ni_dvp == ndp->ni_vp)
286 if (fmode & O_EXCL) {
290 if (vp->v_type == VDIR) {
297 ndp->ni_cnd.cn_nameiop = LOOKUP;
298 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
299 ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
301 if ((fmode & FWRITE) == 0)
302 ndp->ni_cnd.cn_flags |= LOCKSHARED;
303 if ((error = namei(ndp)) != 0)
307 error = vn_open_vnode(vp, fmode, cred, td, fp);
313 NDFREE(ndp, NDF_ONLY_PNBUF);
321 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
324 int error, lock_flags, type;
326 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
327 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
329 KASSERT(fp != NULL, ("open with flock requires fp"));
330 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
333 lock_flags = VOP_ISLOCKED(vp);
336 lf.l_whence = SEEK_SET;
339 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
341 if ((fmode & FNONBLOCK) == 0)
343 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
345 fp->f_flag |= FHASLOCK;
347 vn_lock(vp, lock_flags | LK_RETRY);
348 if (error == 0 && VN_IS_DOOMED(vp))
354 * Common code for vnode open operations once a vnode is located.
355 * Check permissions, and call the VOP_OPEN routine.
358 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
359 struct thread *td, struct file *fp)
364 if (vp->v_type == VLNK)
366 if (vp->v_type == VSOCK)
368 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
371 if (fmode & (FWRITE | O_TRUNC)) {
372 if (vp->v_type == VDIR)
380 if ((fmode & O_APPEND) && (fmode & FWRITE))
385 if (fmode & O_VERIFY)
387 error = mac_vnode_check_open(cred, vp, accmode);
391 accmode &= ~(VCREAT | VVERIFY);
393 if ((fmode & O_CREAT) == 0 && accmode != 0) {
394 error = VOP_ACCESS(vp, accmode, cred, td);
398 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
399 vn_lock(vp, LK_UPGRADE | LK_RETRY);
400 error = VOP_OPEN(vp, fmode, cred, td, fp);
404 error = vn_open_vnode_advlock(vp, fmode, fp);
405 if (error == 0 && (fmode & FWRITE) != 0) {
406 error = VOP_ADD_WRITECOUNT(vp, 1);
408 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
409 __func__, vp, vp->v_writecount);
414 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
415 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
416 * Arrange for that by having fdrop() to use vn_closefile().
419 fp->f_flag |= FOPENFAILED;
421 if (fp->f_ops == &badfileops) {
422 fp->f_type = DTYPE_VNODE;
428 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
434 * Check for write permissions on the specified vnode.
435 * Prototype text segments cannot be written.
439 vn_writechk(struct vnode *vp)
442 ASSERT_VOP_LOCKED(vp, "vn_writechk");
444 * If there's shared text associated with
445 * the vnode, try to free it up once. If
446 * we fail, we can't allow writing.
458 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
459 struct thread *td, bool keep_ref)
462 int error, lock_flags;
464 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
465 MNT_EXTENDED_SHARED(vp->v_mount))
466 lock_flags = LK_SHARED;
468 lock_flags = LK_EXCLUSIVE;
470 vn_start_write(vp, &mp, V_WAIT);
471 vn_lock(vp, lock_flags | LK_RETRY);
472 AUDIT_ARG_VNODE1(vp);
473 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
474 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
475 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
476 __func__, vp, vp->v_writecount);
478 error = VOP_CLOSE(vp, flags, file_cred, td);
483 vn_finished_write(mp);
488 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
492 return (vn_close1(vp, flags, file_cred, td, false));
496 * Heuristic to detect sequential operation.
499 sequential_heuristic(struct uio *uio, struct file *fp)
503 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
506 if (fp->f_flag & FRDAHEAD)
507 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
510 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
511 * that the first I/O is normally considered to be slightly
512 * sequential. Seeking to offset 0 doesn't change sequentiality
513 * unless previous seeks have reduced f_seqcount to 0, in which
514 * case offset 0 is not special.
516 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
517 uio->uio_offset == fp->f_nextoff[rw]) {
519 * f_seqcount is in units of fixed-size blocks so that it
520 * depends mainly on the amount of sequential I/O and not
521 * much on the number of sequential I/O's. The fixed size
522 * of 16384 is hard-coded here since it is (not quite) just
523 * a magic size that works well here. This size is more
524 * closely related to the best I/O size for real disks than
525 * to any block size used by software.
527 if (uio->uio_resid >= IO_SEQMAX * 16384)
528 fp->f_seqcount[rw] = IO_SEQMAX;
530 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
531 if (fp->f_seqcount[rw] > IO_SEQMAX)
532 fp->f_seqcount[rw] = IO_SEQMAX;
534 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
537 /* Not sequential. Quickly draw-down sequentiality. */
538 if (fp->f_seqcount[rw] > 1)
539 fp->f_seqcount[rw] = 1;
541 fp->f_seqcount[rw] = 0;
546 * Package up an I/O request on a vnode into a uio and do it.
549 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
550 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
551 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
558 struct vn_io_fault_args args;
559 int error, lock_flags;
561 if (offset < 0 && vp->v_type != VCHR)
563 auio.uio_iov = &aiov;
565 aiov.iov_base = base;
567 auio.uio_resid = len;
568 auio.uio_offset = offset;
569 auio.uio_segflg = segflg;
574 if ((ioflg & IO_NODELOCKED) == 0) {
575 if ((ioflg & IO_RANGELOCKED) == 0) {
576 if (rw == UIO_READ) {
577 rl_cookie = vn_rangelock_rlock(vp, offset,
580 rl_cookie = vn_rangelock_wlock(vp, offset,
586 if (rw == UIO_WRITE) {
587 if (vp->v_type != VCHR &&
588 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
591 if (MNT_SHARED_WRITES(mp) ||
592 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
593 lock_flags = LK_SHARED;
595 lock_flags = LK_EXCLUSIVE;
597 lock_flags = LK_SHARED;
598 vn_lock(vp, lock_flags | LK_RETRY);
602 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
604 if ((ioflg & IO_NOMACCHECK) == 0) {
606 error = mac_vnode_check_read(active_cred, file_cred,
609 error = mac_vnode_check_write(active_cred, file_cred,
614 if (file_cred != NULL)
618 if (do_vn_io_fault(vp, &auio)) {
619 args.kind = VN_IO_FAULT_VOP;
622 args.args.vop_args.vp = vp;
623 error = vn_io_fault1(vp, &auio, &args, td);
624 } else if (rw == UIO_READ) {
625 error = VOP_READ(vp, &auio, ioflg, cred);
626 } else /* if (rw == UIO_WRITE) */ {
627 error = VOP_WRITE(vp, &auio, ioflg, cred);
631 *aresid = auio.uio_resid;
633 if (auio.uio_resid && error == 0)
635 if ((ioflg & IO_NODELOCKED) == 0) {
638 vn_finished_write(mp);
641 if (rl_cookie != NULL)
642 vn_rangelock_unlock(vp, rl_cookie);
647 * Package up an I/O request on a vnode into a uio and do it. The I/O
648 * request is split up into smaller chunks and we try to avoid saturating
649 * the buffer cache while potentially holding a vnode locked, so we
650 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
651 * to give other processes a chance to lock the vnode (either other processes
652 * core'ing the same binary, or unrelated processes scanning the directory).
655 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
656 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
657 struct ucred *file_cred, size_t *aresid, struct thread *td)
666 * Force `offset' to a multiple of MAXBSIZE except possibly
667 * for the first chunk, so that filesystems only need to
668 * write full blocks except possibly for the first and last
671 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
675 if (rw != UIO_READ && vp->v_type == VREG)
678 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
679 ioflg, active_cred, file_cred, &iaresid, td);
680 len -= chunk; /* aresid calc already includes length */
684 base = (char *)base + chunk;
685 kern_yield(PRI_USER);
688 *aresid = len + iaresid;
692 #if OFF_MAX <= LONG_MAX
694 foffset_lock(struct file *fp, int flags)
696 volatile short *flagsp;
700 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
702 if ((flags & FOF_NOLOCK) != 0)
703 return (atomic_load_long(&fp->f_offset));
706 * According to McKusick the vn lock was protecting f_offset here.
707 * It is now protected by the FOFFSET_LOCKED flag.
709 flagsp = &fp->f_vnread_flags;
710 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
711 return (atomic_load_long(&fp->f_offset));
713 sleepq_lock(&fp->f_vnread_flags);
714 state = atomic_load_16(flagsp);
716 if ((state & FOFFSET_LOCKED) == 0) {
717 if (!atomic_fcmpset_acq_16(flagsp, &state,
722 if ((state & FOFFSET_LOCK_WAITING) == 0) {
723 if (!atomic_fcmpset_acq_16(flagsp, &state,
724 state | FOFFSET_LOCK_WAITING))
728 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
729 sleepq_wait(&fp->f_vnread_flags, PUSER -1);
731 sleepq_lock(&fp->f_vnread_flags);
732 state = atomic_load_16(flagsp);
734 res = atomic_load_long(&fp->f_offset);
735 sleepq_release(&fp->f_vnread_flags);
740 foffset_unlock(struct file *fp, off_t val, int flags)
742 volatile short *flagsp;
745 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
747 if ((flags & FOF_NOUPDATE) == 0)
748 atomic_store_long(&fp->f_offset, val);
749 if ((flags & FOF_NEXTOFF_R) != 0)
750 fp->f_nextoff[UIO_READ] = val;
751 if ((flags & FOF_NEXTOFF_W) != 0)
752 fp->f_nextoff[UIO_WRITE] = val;
754 if ((flags & FOF_NOLOCK) != 0)
757 flagsp = &fp->f_vnread_flags;
758 state = atomic_load_16(flagsp);
759 if ((state & FOFFSET_LOCK_WAITING) == 0 &&
760 atomic_cmpset_rel_16(flagsp, state, 0))
763 sleepq_lock(&fp->f_vnread_flags);
764 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
765 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
766 fp->f_vnread_flags = 0;
767 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
768 sleepq_release(&fp->f_vnread_flags);
772 foffset_lock(struct file *fp, int flags)
777 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
779 mtxp = mtx_pool_find(mtxpool_sleep, fp);
781 if ((flags & FOF_NOLOCK) == 0) {
782 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
783 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
784 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
787 fp->f_vnread_flags |= FOFFSET_LOCKED;
795 foffset_unlock(struct file *fp, off_t val, int flags)
799 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
801 mtxp = mtx_pool_find(mtxpool_sleep, fp);
803 if ((flags & FOF_NOUPDATE) == 0)
805 if ((flags & FOF_NEXTOFF_R) != 0)
806 fp->f_nextoff[UIO_READ] = val;
807 if ((flags & FOF_NEXTOFF_W) != 0)
808 fp->f_nextoff[UIO_WRITE] = val;
809 if ((flags & FOF_NOLOCK) == 0) {
810 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
811 ("Lost FOFFSET_LOCKED"));
812 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
813 wakeup(&fp->f_vnread_flags);
814 fp->f_vnread_flags = 0;
821 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
824 if ((flags & FOF_OFFSET) == 0)
825 uio->uio_offset = foffset_lock(fp, flags);
829 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
832 if ((flags & FOF_OFFSET) == 0)
833 foffset_unlock(fp, uio->uio_offset, flags);
837 get_advice(struct file *fp, struct uio *uio)
842 ret = POSIX_FADV_NORMAL;
843 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
846 mtxp = mtx_pool_find(mtxpool_sleep, fp);
848 if (fp->f_advice != NULL &&
849 uio->uio_offset >= fp->f_advice->fa_start &&
850 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
851 ret = fp->f_advice->fa_advice;
857 vn_read_from_obj(struct vnode *vp, struct uio *uio)
860 vm_page_t ma[io_hold_cnt + 2];
866 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
868 MPASS(obj->type == OBJT_VNODE);
871 * Depends on type stability of vm_objects.
873 vm_object_pip_add(obj, 1);
874 if ((obj->flags & OBJ_DEAD) != 0) {
876 * Note that object might be already reused from the
877 * vnode, and the OBJ_DEAD flag cleared. This is fine,
878 * we recheck for DOOMED vnode state after all pages
879 * are busied, and retract then.
881 * But we check for OBJ_DEAD to ensure that we do not
882 * busy pages while vm_object_terminate_pages()
883 * processes the queue.
889 resid = uio->uio_resid;
890 off = uio->uio_offset;
891 for (i = 0; resid > 0; i++) {
892 MPASS(i < io_hold_cnt + 2);
893 ma[i] = vm_page_grab_unlocked(obj, atop(off),
894 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
900 * Skip invalid pages. Valid mask can be partial only
901 * at EOF, and we clip later.
903 if (vm_page_none_valid(ma[i])) {
904 vm_page_sunbusy(ma[i]);
917 * Check VIRF_DOOMED after we busied our pages. Since
918 * vgonel() terminates the vnode' vm_object, it cannot
919 * process past pages busied by us.
921 if (VN_IS_DOOMED(vp)) {
926 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
927 if (resid > uio->uio_resid)
928 resid = uio->uio_resid;
931 * Unlocked read of vnp_size is safe because truncation cannot
932 * pass busied page. But we load vnp_size into a local
933 * variable so that possible concurrent extension does not
936 #if defined(__powerpc__) && !defined(__powerpc64__)
937 vsz = obj->un_pager.vnp.vnp_size;
939 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
941 if (uio->uio_offset + resid > vsz)
942 resid = vsz - uio->uio_offset;
944 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
947 for (j = 0; j < i; j++) {
949 vm_page_reference(ma[j]);
950 vm_page_sunbusy(ma[j]);
953 vm_object_pip_wakeup(obj);
956 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
960 * File table vnode read routine.
963 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
971 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
973 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
976 if (fp->f_flag & FNONBLOCK)
978 if (fp->f_flag & O_DIRECT)
982 * Try to read from page cache. VIRF_DOOMED check is racy but
983 * allows us to avoid unneeded work outright.
985 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
986 (vp->v_irflag & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
987 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
989 fp->f_nextoff[UIO_READ] = uio->uio_offset;
992 if (error != EJUSTRETURN)
996 advice = get_advice(fp, uio);
997 vn_lock(vp, LK_SHARED | LK_RETRY);
1000 case POSIX_FADV_NORMAL:
1001 case POSIX_FADV_SEQUENTIAL:
1002 case POSIX_FADV_NOREUSE:
1003 ioflag |= sequential_heuristic(uio, fp);
1005 case POSIX_FADV_RANDOM:
1006 /* Disable read-ahead for random I/O. */
1009 orig_offset = uio->uio_offset;
1012 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1015 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1016 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1018 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1019 orig_offset != uio->uio_offset)
1021 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1022 * for the backing file after a POSIX_FADV_NOREUSE
1025 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1026 POSIX_FADV_DONTNEED);
1031 * File table vnode write routine.
1034 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1040 int error, ioflag, lock_flags;
1043 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1045 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1047 if (vp->v_type == VREG)
1050 if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
1051 ioflag |= IO_APPEND;
1052 if (fp->f_flag & FNONBLOCK)
1053 ioflag |= IO_NDELAY;
1054 if (fp->f_flag & O_DIRECT)
1055 ioflag |= IO_DIRECT;
1056 if ((fp->f_flag & O_FSYNC) ||
1057 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
1060 if (vp->v_type != VCHR &&
1061 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
1064 advice = get_advice(fp, uio);
1066 if (MNT_SHARED_WRITES(mp) ||
1067 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
1068 lock_flags = LK_SHARED;
1070 lock_flags = LK_EXCLUSIVE;
1073 vn_lock(vp, lock_flags | LK_RETRY);
1075 case POSIX_FADV_NORMAL:
1076 case POSIX_FADV_SEQUENTIAL:
1077 case POSIX_FADV_NOREUSE:
1078 ioflag |= sequential_heuristic(uio, fp);
1080 case POSIX_FADV_RANDOM:
1081 /* XXX: Is this correct? */
1084 orig_offset = uio->uio_offset;
1087 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1090 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1091 fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1093 if (vp->v_type != VCHR)
1094 vn_finished_write(mp);
1095 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1096 orig_offset != uio->uio_offset)
1098 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1099 * for the backing file after a POSIX_FADV_NOREUSE
1102 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1103 POSIX_FADV_DONTNEED);
1109 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1110 * prevent the following deadlock:
1112 * Assume that the thread A reads from the vnode vp1 into userspace
1113 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
1114 * currently not resident, then system ends up with the call chain
1115 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1116 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1117 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1118 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1119 * backed by the pages of vnode vp1, and some page in buf2 is not
1120 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1122 * To prevent the lock order reversal and deadlock, vn_io_fault() does
1123 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1124 * Instead, it first tries to do the whole range i/o with pagefaults
1125 * disabled. If all pages in the i/o buffer are resident and mapped,
1126 * VOP will succeed (ignoring the genuine filesystem errors).
1127 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1128 * i/o in chunks, with all pages in the chunk prefaulted and held
1129 * using vm_fault_quick_hold_pages().
1131 * Filesystems using this deadlock avoidance scheme should use the
1132 * array of the held pages from uio, saved in the curthread->td_ma,
1133 * instead of doing uiomove(). A helper function
1134 * vn_io_fault_uiomove() converts uiomove request into
1135 * uiomove_fromphys() over td_ma array.
1137 * Since vnode locks do not cover the whole i/o anymore, rangelocks
1138 * make the current i/o request atomic with respect to other i/os and
1143 * Decode vn_io_fault_args and perform the corresponding i/o.
1146 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1152 save = vm_fault_disable_pagefaults();
1153 switch (args->kind) {
1154 case VN_IO_FAULT_FOP:
1155 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1156 uio, args->cred, args->flags, td);
1158 case VN_IO_FAULT_VOP:
1159 if (uio->uio_rw == UIO_READ) {
1160 error = VOP_READ(args->args.vop_args.vp, uio,
1161 args->flags, args->cred);
1162 } else if (uio->uio_rw == UIO_WRITE) {
1163 error = VOP_WRITE(args->args.vop_args.vp, uio,
1164 args->flags, args->cred);
1168 panic("vn_io_fault_doio: unknown kind of io %d %d",
1169 args->kind, uio->uio_rw);
1171 vm_fault_enable_pagefaults(save);
1176 vn_io_fault_touch(char *base, const struct uio *uio)
1181 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1187 vn_io_fault_prefault_user(const struct uio *uio)
1190 const struct iovec *iov;
1195 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1196 ("vn_io_fault_prefault userspace"));
1200 resid = uio->uio_resid;
1201 base = iov->iov_base;
1204 error = vn_io_fault_touch(base, uio);
1207 if (len < PAGE_SIZE) {
1209 error = vn_io_fault_touch(base + len - 1, uio);
1214 if (++i >= uio->uio_iovcnt)
1216 iov = uio->uio_iov + i;
1217 base = iov->iov_base;
1229 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1230 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1231 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1232 * into args and call vn_io_fault1() to handle faults during the user
1233 * mode buffer accesses.
1236 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1239 vm_page_t ma[io_hold_cnt + 2];
1240 struct uio *uio_clone, short_uio;
1241 struct iovec short_iovec[1];
1242 vm_page_t *prev_td_ma;
1244 vm_offset_t addr, end;
1247 int error, cnt, saveheld, prev_td_ma_cnt;
1249 if (vn_io_fault_prefault) {
1250 error = vn_io_fault_prefault_user(uio);
1252 return (error); /* Or ignore ? */
1255 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1258 * The UFS follows IO_UNIT directive and replays back both
1259 * uio_offset and uio_resid if an error is encountered during the
1260 * operation. But, since the iovec may be already advanced,
1261 * uio is still in an inconsistent state.
1263 * Cache a copy of the original uio, which is advanced to the redo
1264 * point using UIO_NOCOPY below.
1266 uio_clone = cloneuio(uio);
1267 resid = uio->uio_resid;
1269 short_uio.uio_segflg = UIO_USERSPACE;
1270 short_uio.uio_rw = uio->uio_rw;
1271 short_uio.uio_td = uio->uio_td;
1273 error = vn_io_fault_doio(args, uio, td);
1274 if (error != EFAULT)
1277 atomic_add_long(&vn_io_faults_cnt, 1);
1278 uio_clone->uio_segflg = UIO_NOCOPY;
1279 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1280 uio_clone->uio_segflg = uio->uio_segflg;
1282 saveheld = curthread_pflags_set(TDP_UIOHELD);
1283 prev_td_ma = td->td_ma;
1284 prev_td_ma_cnt = td->td_ma_cnt;
1286 while (uio_clone->uio_resid != 0) {
1287 len = uio_clone->uio_iov->iov_len;
1289 KASSERT(uio_clone->uio_iovcnt >= 1,
1290 ("iovcnt underflow"));
1291 uio_clone->uio_iov++;
1292 uio_clone->uio_iovcnt--;
1295 if (len > ptoa(io_hold_cnt))
1296 len = ptoa(io_hold_cnt);
1297 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1298 end = round_page(addr + len);
1303 cnt = atop(end - trunc_page(addr));
1305 * A perfectly misaligned address and length could cause
1306 * both the start and the end of the chunk to use partial
1307 * page. +2 accounts for such a situation.
1309 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1310 addr, len, prot, ma, io_hold_cnt + 2);
1315 short_uio.uio_iov = &short_iovec[0];
1316 short_iovec[0].iov_base = (void *)addr;
1317 short_uio.uio_iovcnt = 1;
1318 short_uio.uio_resid = short_iovec[0].iov_len = len;
1319 short_uio.uio_offset = uio_clone->uio_offset;
1321 td->td_ma_cnt = cnt;
1323 error = vn_io_fault_doio(args, &short_uio, td);
1324 vm_page_unhold_pages(ma, cnt);
1325 adv = len - short_uio.uio_resid;
1327 uio_clone->uio_iov->iov_base =
1328 (char *)uio_clone->uio_iov->iov_base + adv;
1329 uio_clone->uio_iov->iov_len -= adv;
1330 uio_clone->uio_resid -= adv;
1331 uio_clone->uio_offset += adv;
1333 uio->uio_resid -= adv;
1334 uio->uio_offset += adv;
1336 if (error != 0 || adv == 0)
1339 td->td_ma = prev_td_ma;
1340 td->td_ma_cnt = prev_td_ma_cnt;
1341 curthread_pflags_restore(saveheld);
1343 free(uio_clone, M_IOV);
1348 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1349 int flags, struct thread *td)
1354 struct vn_io_fault_args args;
1357 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1361 * The ability to read(2) on a directory has historically been
1362 * allowed for all users, but this can and has been the source of
1363 * at least one security issue in the past. As such, it is now hidden
1364 * away behind a sysctl for those that actually need it to use it, and
1365 * restricted to root when it's turned on to make it relatively safe to
1366 * leave on for longer sessions of need.
1368 if (vp->v_type == VDIR) {
1369 KASSERT(uio->uio_rw == UIO_READ,
1370 ("illegal write attempted on a directory"));
1371 if (!vfs_allow_read_dir)
1373 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1377 foffset_lock_uio(fp, uio, flags);
1378 if (do_vn_io_fault(vp, uio)) {
1379 args.kind = VN_IO_FAULT_FOP;
1380 args.args.fop_args.fp = fp;
1381 args.args.fop_args.doio = doio;
1382 args.cred = active_cred;
1383 args.flags = flags | FOF_OFFSET;
1384 if (uio->uio_rw == UIO_READ) {
1385 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1386 uio->uio_offset + uio->uio_resid);
1387 } else if ((fp->f_flag & O_APPEND) != 0 ||
1388 (flags & FOF_OFFSET) == 0) {
1389 /* For appenders, punt and lock the whole range. */
1390 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1392 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1393 uio->uio_offset + uio->uio_resid);
1395 error = vn_io_fault1(vp, uio, &args, td);
1396 vn_rangelock_unlock(vp, rl_cookie);
1398 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1400 foffset_unlock_uio(fp, uio, flags);
1405 * Helper function to perform the requested uiomove operation using
1406 * the held pages for io->uio_iov[0].iov_base buffer instead of
1407 * copyin/copyout. Access to the pages with uiomove_fromphys()
1408 * instead of iov_base prevents page faults that could occur due to
1409 * pmap_collect() invalidating the mapping created by
1410 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1411 * object cleanup revoking the write access from page mappings.
1413 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1414 * instead of plain uiomove().
1417 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1419 struct uio transp_uio;
1420 struct iovec transp_iov[1];
1426 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1427 uio->uio_segflg != UIO_USERSPACE)
1428 return (uiomove(data, xfersize, uio));
1430 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1431 transp_iov[0].iov_base = data;
1432 transp_uio.uio_iov = &transp_iov[0];
1433 transp_uio.uio_iovcnt = 1;
1434 if (xfersize > uio->uio_resid)
1435 xfersize = uio->uio_resid;
1436 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1437 transp_uio.uio_offset = 0;
1438 transp_uio.uio_segflg = UIO_SYSSPACE;
1440 * Since transp_iov points to data, and td_ma page array
1441 * corresponds to original uio->uio_iov, we need to invert the
1442 * direction of the i/o operation as passed to
1443 * uiomove_fromphys().
1445 switch (uio->uio_rw) {
1447 transp_uio.uio_rw = UIO_READ;
1450 transp_uio.uio_rw = UIO_WRITE;
1453 transp_uio.uio_td = uio->uio_td;
1454 error = uiomove_fromphys(td->td_ma,
1455 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1456 xfersize, &transp_uio);
1457 adv = xfersize - transp_uio.uio_resid;
1459 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1460 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1462 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1464 td->td_ma_cnt -= pgadv;
1465 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1466 uio->uio_iov->iov_len -= adv;
1467 uio->uio_resid -= adv;
1468 uio->uio_offset += adv;
1473 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1477 vm_offset_t iov_base;
1481 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1482 uio->uio_segflg != UIO_USERSPACE)
1483 return (uiomove_fromphys(ma, offset, xfersize, uio));
1485 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1486 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1487 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1488 switch (uio->uio_rw) {
1490 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1494 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1498 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1500 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1502 td->td_ma_cnt -= pgadv;
1503 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1504 uio->uio_iov->iov_len -= cnt;
1505 uio->uio_resid -= cnt;
1506 uio->uio_offset += cnt;
1511 * File table truncate routine.
1514 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1525 * Lock the whole range for truncation. Otherwise split i/o
1526 * might happen partly before and partly after the truncation.
1528 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1529 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1532 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1533 AUDIT_ARG_VNODE1(vp);
1534 if (vp->v_type == VDIR) {
1539 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1543 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1547 vn_finished_write(mp);
1549 vn_rangelock_unlock(vp, rl_cookie);
1554 * Truncate a file that is already locked.
1557 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1563 error = VOP_ADD_WRITECOUNT(vp, 1);
1566 vattr.va_size = length;
1568 vattr.va_vaflags |= VA_SYNC;
1569 error = VOP_SETATTR(vp, &vattr, cred);
1570 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1576 * File table vnode stat routine.
1579 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1582 struct vnode *vp = fp->f_vnode;
1585 vn_lock(vp, LK_SHARED | LK_RETRY);
1586 error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1593 * File table vnode ioctl routine.
1596 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1601 struct fiobmap2_arg *bmarg;
1605 switch (vp->v_type) {
1610 vn_lock(vp, LK_SHARED | LK_RETRY);
1611 error = VOP_GETATTR(vp, &vattr, active_cred);
1614 *(int *)data = vattr.va_size - fp->f_offset;
1617 bmarg = (struct fiobmap2_arg *)data;
1618 vn_lock(vp, LK_SHARED | LK_RETRY);
1620 error = mac_vnode_check_read(active_cred, fp->f_cred,
1624 error = VOP_BMAP(vp, bmarg->bn, NULL,
1625 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1632 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1637 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1645 * File table vnode poll routine.
1648 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1655 #if defined(MAC) || defined(AUDIT)
1656 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1657 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1658 AUDIT_ARG_VNODE1(vp);
1659 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1665 error = VOP_POLL(vp, events, fp->f_cred, td);
1670 * Acquire the requested lock and then check for validity. LK_RETRY
1671 * permits vn_lock to return doomed vnodes.
1673 static int __noinline
1674 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1678 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1679 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1682 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1684 if ((flags & LK_RETRY) == 0) {
1695 * Nothing to do if we got the lock.
1701 * Interlock was dropped by the call in _vn_lock.
1703 flags &= ~LK_INTERLOCK;
1705 error = VOP_LOCK1(vp, flags, file, line);
1706 } while (error != 0);
1711 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1715 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1716 ("vn_lock: no locktype (%d passed)", flags));
1717 VNPASS(vp->v_holdcnt > 0, vp);
1718 error = VOP_LOCK1(vp, flags, file, line);
1719 if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1720 return (_vn_lock_fallback(vp, flags, file, line, error));
1725 * File table vnode close routine.
1728 vn_closefile(struct file *fp, struct thread *td)
1736 fp->f_ops = &badfileops;
1737 ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1739 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1741 if (__predict_false(ref)) {
1742 lf.l_whence = SEEK_SET;
1745 lf.l_type = F_UNLCK;
1746 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1753 * Preparing to start a filesystem write operation. If the operation is
1754 * permitted, then we bump the count of operations in progress and
1755 * proceed. If a suspend request is in progress, we wait until the
1756 * suspension is over, and then proceed.
1759 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1763 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1764 vfs_op_thread_enter(mp)) {
1765 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1766 vfs_mp_count_add_pcpu(mp, writeopcount, 1);
1767 vfs_op_thread_exit(mp);
1772 mtx_assert(MNT_MTX(mp), MA_OWNED);
1779 * Check on status of suspension.
1781 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1782 mp->mnt_susp_owner != curthread) {
1783 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1784 (flags & PCATCH) : 0) | (PUSER - 1);
1785 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1786 if (flags & V_NOWAIT) {
1787 error = EWOULDBLOCK;
1790 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1796 if (flags & V_XSLEEP)
1798 mp->mnt_writeopcount++;
1800 if (error != 0 || (flags & V_XSLEEP) != 0)
1807 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1812 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1813 ("V_MNTREF requires mp"));
1817 * If a vnode is provided, get and return the mount point that
1818 * to which it will write.
1821 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1823 if (error != EOPNOTSUPP)
1828 if ((mp = *mpp) == NULL)
1832 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1834 * As long as a vnode is not provided we need to acquire a
1835 * refcount for the provided mountpoint too, in order to
1836 * emulate a vfs_ref().
1838 if (vp == NULL && (flags & V_MNTREF) == 0)
1841 return (vn_start_write_refed(mp, flags, false));
1845 * Secondary suspension. Used by operations such as vop_inactive
1846 * routines that are needed by the higher level functions. These
1847 * are allowed to proceed until all the higher level functions have
1848 * completed (indicated by mnt_writeopcount dropping to zero). At that
1849 * time, these operations are halted until the suspension is over.
1852 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1857 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1858 ("V_MNTREF requires mp"));
1862 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1864 if (error != EOPNOTSUPP)
1870 * If we are not suspended or have not yet reached suspended
1871 * mode, then let the operation proceed.
1873 if ((mp = *mpp) == NULL)
1877 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1879 * As long as a vnode is not provided we need to acquire a
1880 * refcount for the provided mountpoint too, in order to
1881 * emulate a vfs_ref().
1884 if (vp == NULL && (flags & V_MNTREF) == 0)
1886 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1887 mp->mnt_secondary_writes++;
1888 mp->mnt_secondary_accwrites++;
1892 if (flags & V_NOWAIT) {
1895 return (EWOULDBLOCK);
1898 * Wait for the suspension to finish.
1900 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1901 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1910 * Filesystem write operation has completed. If we are suspending and this
1911 * operation is the last one, notify the suspender that the suspension is
1915 vn_finished_write(struct mount *mp)
1922 if (vfs_op_thread_enter(mp)) {
1923 vfs_mp_count_sub_pcpu(mp, writeopcount, 1);
1924 vfs_mp_count_sub_pcpu(mp, ref, 1);
1925 vfs_op_thread_exit(mp);
1930 vfs_assert_mount_counters(mp);
1932 c = --mp->mnt_writeopcount;
1933 if (mp->mnt_vfs_ops == 0) {
1934 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1939 vfs_dump_mount_counters(mp);
1940 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
1941 wakeup(&mp->mnt_writeopcount);
1946 * Filesystem secondary write operation has completed. If we are
1947 * suspending and this operation is the last one, notify the suspender
1948 * that the suspension is now in effect.
1951 vn_finished_secondary_write(struct mount *mp)
1957 mp->mnt_secondary_writes--;
1958 if (mp->mnt_secondary_writes < 0)
1959 panic("vn_finished_secondary_write: neg cnt");
1960 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1961 mp->mnt_secondary_writes <= 0)
1962 wakeup(&mp->mnt_secondary_writes);
1967 * Request a filesystem to suspend write operations.
1970 vfs_write_suspend(struct mount *mp, int flags)
1977 vfs_assert_mount_counters(mp);
1978 if (mp->mnt_susp_owner == curthread) {
1979 vfs_op_exit_locked(mp);
1983 while (mp->mnt_kern_flag & MNTK_SUSPEND)
1984 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1987 * Unmount holds a write reference on the mount point. If we
1988 * own busy reference and drain for writers, we deadlock with
1989 * the reference draining in the unmount path. Callers of
1990 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1991 * vfs_busy() reference is owned and caller is not in the
1994 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1995 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1996 vfs_op_exit_locked(mp);
2001 mp->mnt_kern_flag |= MNTK_SUSPEND;
2002 mp->mnt_susp_owner = curthread;
2003 if (mp->mnt_writeopcount > 0)
2004 (void) msleep(&mp->mnt_writeopcount,
2005 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2008 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2009 vfs_write_resume(mp, 0);
2010 /* vfs_write_resume does vfs_op_exit() for us */
2016 * Request a filesystem to resume write operations.
2019 vfs_write_resume(struct mount *mp, int flags)
2023 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2024 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2025 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2027 mp->mnt_susp_owner = NULL;
2028 wakeup(&mp->mnt_writeopcount);
2029 wakeup(&mp->mnt_flag);
2030 curthread->td_pflags &= ~TDP_IGNSUSP;
2031 if ((flags & VR_START_WRITE) != 0) {
2033 mp->mnt_writeopcount++;
2036 if ((flags & VR_NO_SUSPCLR) == 0)
2039 } else if ((flags & VR_START_WRITE) != 0) {
2041 vn_start_write_refed(mp, 0, true);
2048 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2052 vfs_write_suspend_umnt(struct mount *mp)
2056 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2057 ("vfs_write_suspend_umnt: recursed"));
2059 /* dounmount() already called vn_start_write(). */
2061 vn_finished_write(mp);
2062 error = vfs_write_suspend(mp, 0);
2064 vn_start_write(NULL, &mp, V_WAIT);
2068 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2071 vn_start_write(NULL, &mp, V_WAIT);
2073 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2074 wakeup(&mp->mnt_flag);
2076 curthread->td_pflags |= TDP_IGNSUSP;
2081 * Implement kqueues for files by translating it to vnode operation.
2084 vn_kqfilter(struct file *fp, struct knote *kn)
2087 return (VOP_KQFILTER(fp->f_vnode, kn));
2091 * Simplified in-kernel wrapper calls for extended attribute access.
2092 * Both calls pass in a NULL credential, authorizing as "kernel" access.
2093 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2096 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2097 const char *attrname, int *buflen, char *buf, struct thread *td)
2103 iov.iov_len = *buflen;
2106 auio.uio_iov = &iov;
2107 auio.uio_iovcnt = 1;
2108 auio.uio_rw = UIO_READ;
2109 auio.uio_segflg = UIO_SYSSPACE;
2111 auio.uio_offset = 0;
2112 auio.uio_resid = *buflen;
2114 if ((ioflg & IO_NODELOCKED) == 0)
2115 vn_lock(vp, LK_SHARED | LK_RETRY);
2117 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2119 /* authorize attribute retrieval as kernel */
2120 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2123 if ((ioflg & IO_NODELOCKED) == 0)
2127 *buflen = *buflen - auio.uio_resid;
2134 * XXX failure mode if partially written?
2137 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2138 const char *attrname, int buflen, char *buf, struct thread *td)
2145 iov.iov_len = buflen;
2148 auio.uio_iov = &iov;
2149 auio.uio_iovcnt = 1;
2150 auio.uio_rw = UIO_WRITE;
2151 auio.uio_segflg = UIO_SYSSPACE;
2153 auio.uio_offset = 0;
2154 auio.uio_resid = buflen;
2156 if ((ioflg & IO_NODELOCKED) == 0) {
2157 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2159 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2162 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2164 /* authorize attribute setting as kernel */
2165 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2167 if ((ioflg & IO_NODELOCKED) == 0) {
2168 vn_finished_write(mp);
2176 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2177 const char *attrname, struct thread *td)
2182 if ((ioflg & IO_NODELOCKED) == 0) {
2183 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2185 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2188 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2190 /* authorize attribute removal as kernel */
2191 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2192 if (error == EOPNOTSUPP)
2193 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2196 if ((ioflg & IO_NODELOCKED) == 0) {
2197 vn_finished_write(mp);
2205 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2209 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2213 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2216 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2221 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2222 int lkflags, struct vnode **rvp)
2227 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2229 ltype = VOP_ISLOCKED(vp);
2230 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2231 ("vn_vget_ino: vp not locked"));
2232 error = vfs_busy(mp, MBF_NOWAIT);
2236 error = vfs_busy(mp, 0);
2237 vn_lock(vp, ltype | LK_RETRY);
2241 if (VN_IS_DOOMED(vp)) {
2247 error = alloc(mp, alloc_arg, lkflags, rvp);
2249 if (error != 0 || *rvp != vp)
2250 vn_lock(vp, ltype | LK_RETRY);
2251 if (VN_IS_DOOMED(vp)) {
2264 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2268 if (vp->v_type != VREG || td == NULL)
2270 if ((uoff_t)uio->uio_offset + uio->uio_resid >
2271 lim_cur(td, RLIMIT_FSIZE)) {
2272 PROC_LOCK(td->td_proc);
2273 kern_psignal(td->td_proc, SIGXFSZ);
2274 PROC_UNLOCK(td->td_proc);
2281 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2288 vn_lock(vp, LK_SHARED | LK_RETRY);
2289 AUDIT_ARG_VNODE1(vp);
2292 return (setfmode(td, active_cred, vp, mode));
2296 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2303 vn_lock(vp, LK_SHARED | LK_RETRY);
2304 AUDIT_ARG_VNODE1(vp);
2307 return (setfown(td, active_cred, vp, uid, gid));
2311 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2315 if ((object = vp->v_object) == NULL)
2317 VM_OBJECT_WLOCK(object);
2318 vm_object_page_remove(object, start, end, 0);
2319 VM_OBJECT_WUNLOCK(object);
2323 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2331 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2332 ("Wrong command %lu", cmd));
2334 if (vn_lock(vp, LK_SHARED) != 0)
2336 if (vp->v_type != VREG) {
2340 error = VOP_GETATTR(vp, &va, cred);
2344 if (noff >= va.va_size) {
2348 bsize = vp->v_mount->mnt_stat.f_iosize;
2349 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2351 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2352 if (error == EOPNOTSUPP) {
2356 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2357 (bnp != -1 && cmd == FIOSEEKDATA)) {
2364 if (noff > va.va_size)
2366 /* noff == va.va_size. There is an implicit hole at the end of file. */
2367 if (cmd == FIOSEEKDATA)
2377 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2382 off_t foffset, size;
2385 cred = td->td_ucred;
2387 foffset = foffset_lock(fp, 0);
2388 noneg = (vp->v_type != VCHR);
2394 (offset > 0 && foffset > OFF_MAX - offset))) {
2401 vn_lock(vp, LK_SHARED | LK_RETRY);
2402 error = VOP_GETATTR(vp, &vattr, cred);
2408 * If the file references a disk device, then fetch
2409 * the media size and use that to determine the ending
2412 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2413 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2414 vattr.va_size = size;
2416 (vattr.va_size > OFF_MAX ||
2417 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2421 offset += vattr.va_size;
2426 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2427 if (error == ENOTTY)
2431 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2432 if (error == ENOTTY)
2438 if (error == 0 && noneg && offset < 0)
2442 VFS_KNOTE_UNLOCKED(vp, 0);
2443 td->td_uretoff.tdu_off = offset;
2445 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2450 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2456 * Grant permission if the caller is the owner of the file, or
2457 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2458 * on the file. If the time pointer is null, then write
2459 * permission on the file is also sufficient.
2461 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2462 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2463 * will be allowed to set the times [..] to the current
2466 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2467 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2468 error = VOP_ACCESS(vp, VWRITE, cred, td);
2473 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2478 if (fp->f_type == DTYPE_FIFO)
2479 kif->kf_type = KF_TYPE_FIFO;
2481 kif->kf_type = KF_TYPE_VNODE;
2484 FILEDESC_SUNLOCK(fdp);
2485 error = vn_fill_kinfo_vnode(vp, kif);
2487 FILEDESC_SLOCK(fdp);
2492 vn_fill_junk(struct kinfo_file *kif)
2497 * Simulate vn_fullpath returning changing values for a given
2498 * vp during e.g. coredump.
2500 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2501 olen = strlen(kif->kf_path);
2503 strcpy(&kif->kf_path[len - 1], "$");
2505 for (; olen < len; olen++)
2506 strcpy(&kif->kf_path[olen], "A");
2510 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2513 char *fullpath, *freepath;
2516 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2519 error = vn_fullpath(vp, &fullpath, &freepath);
2521 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2523 if (freepath != NULL)
2524 free(freepath, M_TEMP);
2526 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2531 * Retrieve vnode attributes.
2533 va.va_fsid = VNOVAL;
2535 vn_lock(vp, LK_SHARED | LK_RETRY);
2536 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2540 if (va.va_fsid != VNOVAL)
2541 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2543 kif->kf_un.kf_file.kf_file_fsid =
2544 vp->v_mount->mnt_stat.f_fsid.val[0];
2545 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2546 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2547 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2548 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2549 kif->kf_un.kf_file.kf_file_size = va.va_size;
2550 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2551 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2552 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2557 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2558 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2562 struct pmckern_map_in pkm;
2568 boolean_t writecounted;
2571 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2572 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2574 * POSIX shared-memory objects are defined to have
2575 * kernel persistence, and are not defined to support
2576 * read(2)/write(2) -- or even open(2). Thus, we can
2577 * use MAP_ASYNC to trade on-disk coherence for speed.
2578 * The shm_open(3) library routine turns on the FPOSIXSHM
2579 * flag to request this behavior.
2581 if ((fp->f_flag & FPOSIXSHM) != 0)
2582 flags |= MAP_NOSYNC;
2587 * Ensure that file and memory protections are
2588 * compatible. Note that we only worry about
2589 * writability if mapping is shared; in this case,
2590 * current and max prot are dictated by the open file.
2591 * XXX use the vnode instead? Problem is: what
2592 * credentials do we use for determination? What if
2593 * proc does a setuid?
2596 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2597 maxprot = VM_PROT_NONE;
2598 if ((prot & VM_PROT_EXECUTE) != 0)
2601 maxprot = VM_PROT_EXECUTE;
2602 if ((fp->f_flag & FREAD) != 0)
2603 maxprot |= VM_PROT_READ;
2604 else if ((prot & VM_PROT_READ) != 0)
2608 * If we are sharing potential changes via MAP_SHARED and we
2609 * are trying to get write permission although we opened it
2610 * without asking for it, bail out.
2612 if ((flags & MAP_SHARED) != 0) {
2613 if ((fp->f_flag & FWRITE) != 0)
2614 maxprot |= VM_PROT_WRITE;
2615 else if ((prot & VM_PROT_WRITE) != 0)
2618 maxprot |= VM_PROT_WRITE;
2619 cap_maxprot |= VM_PROT_WRITE;
2621 maxprot &= cap_maxprot;
2624 * For regular files and shared memory, POSIX requires that
2625 * the value of foff be a legitimate offset within the data
2626 * object. In particular, negative offsets are invalid.
2627 * Blocking negative offsets and overflows here avoids
2628 * possible wraparound or user-level access into reserved
2629 * ranges of the data object later. In contrast, POSIX does
2630 * not dictate how offsets are used by device drivers, so in
2631 * the case of a device mapping a negative offset is passed
2638 foff > OFF_MAX - size)
2641 writecounted = FALSE;
2642 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2643 &foff, &object, &writecounted);
2646 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2647 foff, writecounted, td);
2650 * If this mapping was accounted for in the vnode's
2651 * writecount, then undo that now.
2654 vm_pager_release_writecount(object, 0, size);
2655 vm_object_deallocate(object);
2658 /* Inform hwpmc(4) if an executable is being mapped. */
2659 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2660 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2662 pkm.pm_address = (uintptr_t) *addr;
2663 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2671 vn_fsid(struct vnode *vp, struct vattr *va)
2675 f = &vp->v_mount->mnt_stat.f_fsid;
2676 va->va_fsid = (uint32_t)f->val[1];
2677 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2678 va->va_fsid += (uint32_t)f->val[0];
2682 vn_fsync_buf(struct vnode *vp, int waitfor)
2684 struct buf *bp, *nbp;
2687 int error, maxretry;
2690 maxretry = 10000; /* large, arbitrarily chosen */
2692 if (vp->v_type == VCHR) {
2694 mp = vp->v_rdev->si_mountpt;
2701 * MARK/SCAN initialization to avoid infinite loops.
2703 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2704 bp->b_vflags &= ~BV_SCANNED;
2709 * Flush all dirty buffers associated with a vnode.
2712 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2713 if ((bp->b_vflags & BV_SCANNED) != 0)
2715 bp->b_vflags |= BV_SCANNED;
2716 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2717 if (waitfor != MNT_WAIT)
2720 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2721 BO_LOCKPTR(bo)) != 0) {
2728 KASSERT(bp->b_bufobj == bo,
2729 ("bp %p wrong b_bufobj %p should be %p",
2730 bp, bp->b_bufobj, bo));
2731 if ((bp->b_flags & B_DELWRI) == 0)
2732 panic("fsync: not dirty");
2733 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2739 if (maxretry < 1000)
2740 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2746 * If synchronous the caller expects us to completely resolve all
2747 * dirty buffers in the system. Wait for in-progress I/O to
2748 * complete (which could include background bitmap writes), then
2749 * retry if dirty blocks still exist.
2751 if (waitfor == MNT_WAIT) {
2752 bufobj_wwait(bo, 0, 0);
2753 if (bo->bo_dirty.bv_cnt > 0) {
2755 * If we are unable to write any of these buffers
2756 * then we fail now rather than trying endlessly
2757 * to write them out.
2759 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2760 if ((error = bp->b_error) != 0)
2762 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2763 (error == 0 && --maxretry >= 0))
2771 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2777 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
2778 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2779 * to do the actual copy.
2780 * vn_generic_copy_file_range() is factored out, so it can be called
2781 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2782 * different file systems.
2785 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2786 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2787 struct ucred *outcred, struct thread *fsize_td)
2791 uint64_t uvalin, uvalout;
2794 *lenp = 0; /* For error returns. */
2797 /* Do some sanity checks on the arguments. */
2802 if (invp->v_type == VDIR || outvp->v_type == VDIR)
2804 else if (*inoffp < 0 || uvalin > INT64_MAX || uvalin <
2805 (uint64_t)*inoffp || *outoffp < 0 || uvalout > INT64_MAX ||
2806 uvalout < (uint64_t)*outoffp || invp->v_type != VREG ||
2807 outvp->v_type != VREG)
2813 * If the two vnode are for the same file system, call
2814 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2815 * which can handle copies across multiple file systems.
2818 if (invp->v_mount == outvp->v_mount)
2819 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2820 lenp, flags, incred, outcred, fsize_td);
2822 error = vn_generic_copy_file_range(invp, inoffp, outvp,
2823 outoffp, lenp, flags, incred, outcred, fsize_td);
2829 * Test len bytes of data starting at dat for all bytes == 0.
2830 * Return true if all bytes are zero, false otherwise.
2831 * Expects dat to be well aligned.
2834 mem_iszero(void *dat, int len)
2840 for (p = dat; len > 0; len -= sizeof(*p), p++) {
2841 if (len >= sizeof(*p)) {
2845 cp = (const char *)p;
2846 for (i = 0; i < len; i++, cp++)
2855 * Look for a hole in the output file and, if found, adjust *outoffp
2856 * and *xferp to skip past the hole.
2857 * *xferp is the entire hole length to be written and xfer2 is how many bytes
2858 * to be written as 0's upon return.
2861 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2862 off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2867 if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2868 *dataoffp = *outoffp;
2869 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2872 *holeoffp = *dataoffp;
2873 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2876 if (error != 0 || *holeoffp == *dataoffp) {
2878 * Since outvp is unlocked, it may be possible for
2879 * another thread to do a truncate(), lseek(), write()
2880 * creating a hole at startoff between the above
2881 * VOP_IOCTL() calls, if the other thread does not do
2883 * If that happens, *holeoffp == *dataoffp and finding
2884 * the hole has failed, so disable vn_skip_hole().
2886 *holeoffp = -1; /* Disable use of vn_skip_hole(). */
2889 KASSERT(*dataoffp >= *outoffp,
2890 ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2891 (intmax_t)*dataoffp, (intmax_t)*outoffp));
2892 KASSERT(*holeoffp > *dataoffp,
2893 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2894 (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2898 * If there is a hole before the data starts, advance *outoffp and
2899 * *xferp past the hole.
2901 if (*dataoffp > *outoffp) {
2902 delta = *dataoffp - *outoffp;
2903 if (delta >= *xferp) {
2904 /* Entire *xferp is a hole. */
2911 xfer2 = MIN(xfer2, *xferp);
2915 * If a hole starts before the end of this xfer2, reduce this xfer2 so
2916 * that the write ends at the start of the hole.
2917 * *holeoffp should always be greater than *outoffp, but for the
2918 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
2921 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
2922 xfer2 = *holeoffp - *outoffp;
2927 * Write an xfer sized chunk to outvp in blksize blocks from dat.
2928 * dat is a maximum of blksize in length and can be written repeatedly in
2930 * If growfile == true, just grow the file via vn_truncate_locked() instead
2931 * of doing actual writes.
2932 * If checkhole == true, a hole is being punched, so skip over any hole
2933 * already in the output file.
2936 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
2937 u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
2940 off_t dataoff, holeoff, xfer2;
2944 * Loop around doing writes of blksize until write has been completed.
2945 * Lock/unlock on each loop iteration so that a bwillwrite() can be
2946 * done for each iteration, since the xfer argument can be very
2947 * large if there is a large hole to punch in the output file.
2952 xfer2 = MIN(xfer, blksize);
2955 * Punching a hole. Skip writing if there is
2956 * already a hole in the output file.
2958 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
2959 &dataoff, &holeoff, cred);
2964 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
2969 error = vn_start_write(outvp, &mp, V_WAIT);
2971 if (MNT_SHARED_WRITES(mp))
2974 lckf = LK_EXCLUSIVE;
2975 error = vn_lock(outvp, lckf);
2979 error = vn_truncate_locked(outvp, outoff + xfer,
2982 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
2983 outoff, UIO_SYSSPACE, IO_NODELOCKED,
2984 curthread->td_ucred, cred, NULL, curthread);
2991 vn_finished_write(mp);
2992 } while (!growfile && xfer > 0 && error == 0);
2997 * Copy a byte range of one file to another. This function can handle the
2998 * case where invp and outvp are on different file systems.
2999 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3000 * is no better file system specific way to do it.
3003 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3004 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3005 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3010 off_t startoff, endoff, xfer, xfer2;
3013 bool cantseek, readzeros, eof, lastblock;
3015 size_t copylen, len, savlen;
3017 long holein, holeout;
3019 holein = holeout = 0;
3020 savlen = len = *lenp;
3024 error = vn_lock(invp, LK_SHARED);
3027 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3032 error = vn_start_write(outvp, &mp, V_WAIT);
3034 error = vn_lock(outvp, LK_EXCLUSIVE);
3037 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3038 * now that outvp is locked.
3040 if (fsize_td != NULL) {
3041 io.uio_offset = *outoffp;
3043 error = vn_rlimit_fsize(outvp, &io, fsize_td);
3047 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3050 * Holes that are past EOF do not need to be written as a block
3051 * of zero bytes. So, truncate the output file as far as
3052 * possible and then use va.va_size to decide if writing 0
3053 * bytes is necessary in the loop below.
3056 error = VOP_GETATTR(outvp, &va, outcred);
3057 if (error == 0 && va.va_size > *outoffp && va.va_size <=
3060 error = mac_vnode_check_write(curthread->td_ucred,
3064 error = vn_truncate_locked(outvp, *outoffp,
3067 va.va_size = *outoffp;
3072 vn_finished_write(mp);
3077 * Set the blksize to the larger of the hole sizes for invp and outvp.
3078 * If hole sizes aren't available, set the blksize to the larger
3079 * f_iosize of invp and outvp.
3080 * This code expects the hole sizes and f_iosizes to be powers of 2.
3081 * This value is clipped at 4Kbytes and 1Mbyte.
3083 blksize = MAX(holein, holeout);
3085 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3086 outvp->v_mount->mnt_stat.f_iosize);
3089 else if (blksize > 1024 * 1024)
3090 blksize = 1024 * 1024;
3091 dat = malloc(blksize, M_TEMP, M_WAITOK);
3094 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3095 * to find holes. Otherwise, just scan the read block for all 0s
3096 * in the inner loop where the data copying is done.
3097 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3098 * support holes on the server, but do not support FIOSEEKHOLE.
3101 while (len > 0 && error == 0 && !eof) {
3102 endoff = 0; /* To shut up compilers. */
3108 * Find the next data area. If there is just a hole to EOF,
3109 * FIOSEEKDATA should fail and then we drop down into the
3110 * inner loop and create the hole on the outvp file.
3111 * (I do not know if any file system will report a hole to
3112 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3113 * will fail for those file systems.)
3115 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3116 * the code just falls through to the inner copy loop.
3120 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3124 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3127 * Since invp is unlocked, it may be possible for
3128 * another thread to do a truncate(), lseek(), write()
3129 * creating a hole at startoff between the above
3130 * VOP_IOCTL() calls, if the other thread does not do
3132 * If that happens, startoff == endoff and finding
3133 * the hole has failed, so set an error.
3135 if (error == 0 && startoff == endoff)
3136 error = EINVAL; /* Any error. Reset to 0. */
3139 if (startoff > *inoffp) {
3140 /* Found hole before data block. */
3141 xfer = MIN(startoff - *inoffp, len);
3142 if (*outoffp < va.va_size) {
3143 /* Must write 0s to punch hole. */
3144 xfer2 = MIN(va.va_size - *outoffp,
3146 memset(dat, 0, MIN(xfer2, blksize));
3147 error = vn_write_outvp(outvp, dat,
3148 *outoffp, xfer2, blksize, false,
3149 holeout > 0, outcred);
3152 if (error == 0 && *outoffp + xfer >
3153 va.va_size && xfer == len)
3154 /* Grow last block. */
3155 error = vn_write_outvp(outvp, dat,
3156 *outoffp, xfer, blksize, true,
3164 copylen = MIN(len, endoff - startoff);
3176 * Set first xfer to end at a block boundary, so that
3177 * holes are more likely detected in the loop below via
3178 * the for all bytes 0 method.
3180 xfer -= (*inoffp % blksize);
3182 /* Loop copying the data block. */
3183 while (copylen > 0 && error == 0 && !eof) {
3186 error = vn_lock(invp, LK_SHARED);
3189 error = vn_rdwr(UIO_READ, invp, dat, xfer,
3190 startoff, UIO_SYSSPACE, IO_NODELOCKED,
3191 curthread->td_ucred, incred, &aresid,
3195 if (error == 0 && aresid > 0) {
3196 /* Stop the copy at EOF on the input file. */
3203 * Skip the write for holes past the initial EOF
3204 * of the output file, unless this is the last
3205 * write of the output file at EOF.
3207 readzeros = cantseek ? mem_iszero(dat, xfer) :
3211 if (!cantseek || *outoffp < va.va_size ||
3212 lastblock || !readzeros)
3213 error = vn_write_outvp(outvp, dat,
3214 *outoffp, xfer, blksize,
3215 readzeros && lastblock &&
3216 *outoffp >= va.va_size, false,
3230 *lenp = savlen - len;
3236 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3240 off_t olen, ooffset;
3243 int audited_vnode1 = 0;
3247 if (vp->v_type != VREG)
3250 /* Allocating blocks may take a long time, so iterate. */
3257 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3260 error = vn_lock(vp, LK_EXCLUSIVE);
3262 vn_finished_write(mp);
3266 if (!audited_vnode1) {
3267 AUDIT_ARG_VNODE1(vp);
3272 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3275 error = VOP_ALLOCATE(vp, &offset, &len);
3277 vn_finished_write(mp);
3279 if (olen + ooffset != offset + len) {
3280 panic("offset + len changed from %jx/%jx to %jx/%jx",
3281 ooffset, olen, offset, len);
3283 if (error != 0 || len == 0)
3285 KASSERT(olen > len, ("Iteration did not make progress?"));