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 #include "opt_hwpmc_hooks.h"
48 #include <sys/param.h>
49 #include <sys/systm.h>
52 #include <sys/fcntl.h>
59 #include <sys/limits.h>
62 #include <sys/mount.h>
63 #include <sys/mutex.h>
64 #include <sys/namei.h>
65 #include <sys/vnode.h>
66 #include <sys/dirent.h>
69 #include <sys/filio.h>
70 #include <sys/resourcevar.h>
71 #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>
81 #include <sys/ktrace.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>
93 #include <vm/vnode_pager.h>
96 #include <sys/pmckern.h>
99 static fo_rdwr_t vn_read;
100 static fo_rdwr_t vn_write;
101 static fo_rdwr_t vn_io_fault;
102 static fo_truncate_t vn_truncate;
103 static fo_ioctl_t vn_ioctl;
104 static fo_poll_t vn_poll;
105 static fo_kqfilter_t vn_kqfilter;
106 static fo_close_t vn_closefile;
107 static fo_mmap_t vn_mmap;
108 static fo_fallocate_t vn_fallocate;
110 struct fileops vnops = {
111 .fo_read = vn_io_fault,
112 .fo_write = vn_io_fault,
113 .fo_truncate = vn_truncate,
114 .fo_ioctl = vn_ioctl,
116 .fo_kqfilter = vn_kqfilter,
117 .fo_stat = vn_statfile,
118 .fo_close = vn_closefile,
119 .fo_chmod = vn_chmod,
120 .fo_chown = vn_chown,
121 .fo_sendfile = vn_sendfile,
123 .fo_fill_kinfo = vn_fill_kinfo,
125 .fo_fallocate = vn_fallocate,
126 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
129 const u_int io_hold_cnt = 16;
130 static int vn_io_fault_enable = 1;
131 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
132 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
133 static int vn_io_fault_prefault = 0;
134 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
135 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
136 static int vn_io_pgcache_read_enable = 1;
137 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
138 &vn_io_pgcache_read_enable, 0,
139 "Enable copying from page cache for reads, avoiding fs");
140 static u_long vn_io_faults_cnt;
141 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
142 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
144 static int vfs_allow_read_dir = 0;
145 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
146 &vfs_allow_read_dir, 0,
147 "Enable read(2) of directory by root for filesystems that support it");
150 * Returns true if vn_io_fault mode of handling the i/o request should
154 do_vn_io_fault(struct vnode *vp, struct uio *uio)
158 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
159 (mp = vp->v_mount) != NULL &&
160 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
164 * Structure used to pass arguments to vn_io_fault1(), to do either
165 * file- or vnode-based I/O calls.
167 struct vn_io_fault_args {
175 struct fop_args_tag {
179 struct vop_args_tag {
185 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
186 struct vn_io_fault_args *args, struct thread *td);
189 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
191 struct thread *td = ndp->ni_cnd.cn_thread;
193 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
197 open2nameif(int fmode, u_int vn_open_flags)
201 res = ISOPEN | LOCKLEAF;
202 if ((fmode & O_RESOLVE_BENEATH) != 0)
204 if ((fmode & O_EMPTY_PATH) != 0)
206 if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
208 if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
210 if ((vn_open_flags & VN_OPEN_WANTIOCTLCAPS) != 0)
211 res |= WANTIOCTLCAPS;
216 * Common code for vnode open operations via a name lookup.
217 * Lookup the vnode and invoke VOP_CREATE if needed.
218 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
220 * Note that this does NOT free nameidata for the successful case,
221 * due to the NDINIT being done elsewhere.
224 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
225 struct ucred *cred, struct file *fp)
229 struct thread *td = ndp->ni_cnd.cn_thread;
231 struct vattr *vap = &vat;
238 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
239 O_EXCL | O_DIRECTORY) ||
240 (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
242 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
243 ndp->ni_cnd.cn_nameiop = CREATE;
244 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
246 * Set NOCACHE to avoid flushing the cache when
247 * rolling in many files at once.
249 * Set NC_KEEPPOSENTRY to keep positive entries if they already
250 * exist despite NOCACHE.
252 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
253 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
254 ndp->ni_cnd.cn_flags |= FOLLOW;
255 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
257 if ((error = namei(ndp)) != 0)
259 if (ndp->ni_vp == NULL) {
262 vap->va_mode = cmode;
264 vap->va_vaflags |= VA_EXCLUSIVE;
265 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
266 NDFREE(ndp, NDF_ONLY_PNBUF);
268 if ((error = vn_start_write(NULL, &mp,
269 V_XSLEEP | PCATCH)) != 0)
274 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
275 ndp->ni_cnd.cn_flags |= MAKEENTRY;
277 error = mac_vnode_check_create(cred, ndp->ni_dvp,
281 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
284 if (error == 0 && (fmode & O_EXCL) != 0 &&
285 (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
287 vp->v_iflag |= VI_FOPENING;
291 VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
293 vn_finished_write(mp);
295 NDFREE(ndp, NDF_ONLY_PNBUF);
296 if (error == ERELOOKUP) {
304 if (ndp->ni_dvp == ndp->ni_vp)
310 if (fmode & O_EXCL) {
314 if (vp->v_type == VDIR) {
321 ndp->ni_cnd.cn_nameiop = LOOKUP;
322 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
323 ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
325 if ((fmode & FWRITE) == 0)
326 ndp->ni_cnd.cn_flags |= LOCKSHARED;
327 if ((error = namei(ndp)) != 0)
331 error = vn_open_vnode(vp, fmode, cred, td, fp);
334 vp->v_iflag &= ~VI_FOPENING;
343 NDFREE(ndp, NDF_ONLY_PNBUF);
351 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
354 int error, lock_flags, type;
356 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
357 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
359 KASSERT(fp != NULL, ("open with flock requires fp"));
360 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
363 lock_flags = VOP_ISLOCKED(vp);
366 lf.l_whence = SEEK_SET;
369 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
371 if ((fmode & FNONBLOCK) == 0)
373 if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
375 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
377 fp->f_flag |= FHASLOCK;
379 vn_lock(vp, lock_flags | LK_RETRY);
384 * Common code for vnode open operations once a vnode is located.
385 * Check permissions, and call the VOP_OPEN routine.
388 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
389 struct thread *td, struct file *fp)
394 if (vp->v_type == VLNK) {
395 if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
398 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
402 if ((fmode & O_PATH) == 0) {
403 if (vp->v_type == VSOCK)
405 if ((fmode & (FWRITE | O_TRUNC)) != 0) {
406 if (vp->v_type == VDIR)
410 if ((fmode & FREAD) != 0)
412 if ((fmode & O_APPEND) && (fmode & FWRITE))
415 if ((fmode & O_CREAT) != 0)
419 if ((fmode & FEXEC) != 0)
422 if ((fmode & O_VERIFY) != 0)
424 error = mac_vnode_check_open(cred, vp, accmode);
428 accmode &= ~(VCREAT | VVERIFY);
430 if ((fmode & O_CREAT) == 0 && accmode != 0) {
431 error = VOP_ACCESS(vp, accmode, cred, td);
435 if ((fmode & O_PATH) != 0) {
436 if (vp->v_type != VFIFO && vp->v_type != VSOCK &&
437 VOP_ACCESS(vp, VREAD, cred, td) == 0)
438 fp->f_flag |= FKQALLOWED;
442 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
443 vn_lock(vp, LK_UPGRADE | LK_RETRY);
444 error = VOP_OPEN(vp, fmode, cred, td, fp);
448 error = vn_open_vnode_advlock(vp, fmode, fp);
449 if (error == 0 && (fmode & FWRITE) != 0) {
450 error = VOP_ADD_WRITECOUNT(vp, 1);
452 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
453 __func__, vp, vp->v_writecount);
458 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
459 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
464 * Arrange the call by having fdrop() to use
465 * vn_closefile(). This is to satisfy
466 * filesystems like devfs or tmpfs, which
467 * override fo_close().
469 fp->f_flag |= FOPENFAILED;
471 if (fp->f_ops == &badfileops) {
472 fp->f_type = DTYPE_VNODE;
478 * If there is no fp, due to kernel-mode open,
479 * we can call VOP_CLOSE() now.
481 if ((vp->v_type == VFIFO || (fmode & FWRITE) != 0 ||
482 !MNT_EXTENDED_SHARED(vp->v_mount)) &&
483 VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
484 vn_lock(vp, LK_UPGRADE | LK_RETRY);
485 (void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
490 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
496 * Check for write permissions on the specified vnode.
497 * Prototype text segments cannot be written.
501 vn_writechk(struct vnode *vp)
504 ASSERT_VOP_LOCKED(vp, "vn_writechk");
506 * If there's shared text associated with
507 * the vnode, try to free it up once. If
508 * we fail, we can't allow writing.
520 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
521 struct thread *td, bool keep_ref)
524 int error, lock_flags;
526 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
527 MNT_EXTENDED_SHARED(vp->v_mount))
528 lock_flags = LK_SHARED;
530 lock_flags = LK_EXCLUSIVE;
532 vn_start_write(vp, &mp, V_WAIT);
533 vn_lock(vp, lock_flags | LK_RETRY);
534 AUDIT_ARG_VNODE1(vp);
535 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
536 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
537 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
538 __func__, vp, vp->v_writecount);
540 error = VOP_CLOSE(vp, flags, file_cred, td);
545 vn_finished_write(mp);
550 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
554 return (vn_close1(vp, flags, file_cred, td, false));
558 * Heuristic to detect sequential operation.
561 sequential_heuristic(struct uio *uio, struct file *fp)
565 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
568 if (fp->f_flag & FRDAHEAD)
569 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
572 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
573 * that the first I/O is normally considered to be slightly
574 * sequential. Seeking to offset 0 doesn't change sequentiality
575 * unless previous seeks have reduced f_seqcount to 0, in which
576 * case offset 0 is not special.
578 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
579 uio->uio_offset == fp->f_nextoff[rw]) {
581 * f_seqcount is in units of fixed-size blocks so that it
582 * depends mainly on the amount of sequential I/O and not
583 * much on the number of sequential I/O's. The fixed size
584 * of 16384 is hard-coded here since it is (not quite) just
585 * a magic size that works well here. This size is more
586 * closely related to the best I/O size for real disks than
587 * to any block size used by software.
589 if (uio->uio_resid >= IO_SEQMAX * 16384)
590 fp->f_seqcount[rw] = IO_SEQMAX;
592 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
593 if (fp->f_seqcount[rw] > IO_SEQMAX)
594 fp->f_seqcount[rw] = IO_SEQMAX;
596 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
599 /* Not sequential. Quickly draw-down sequentiality. */
600 if (fp->f_seqcount[rw] > 1)
601 fp->f_seqcount[rw] = 1;
603 fp->f_seqcount[rw] = 0;
608 * Package up an I/O request on a vnode into a uio and do it.
611 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
612 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
613 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
620 struct vn_io_fault_args args;
621 int error, lock_flags;
623 if (offset < 0 && vp->v_type != VCHR)
625 auio.uio_iov = &aiov;
627 aiov.iov_base = base;
629 auio.uio_resid = len;
630 auio.uio_offset = offset;
631 auio.uio_segflg = segflg;
636 if ((ioflg & IO_NODELOCKED) == 0) {
637 if ((ioflg & IO_RANGELOCKED) == 0) {
638 if (rw == UIO_READ) {
639 rl_cookie = vn_rangelock_rlock(vp, offset,
641 } else if ((ioflg & IO_APPEND) != 0) {
642 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
644 rl_cookie = vn_rangelock_wlock(vp, offset,
650 if (rw == UIO_WRITE) {
651 if (vp->v_type != VCHR &&
652 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
655 lock_flags = vn_lktype_write(mp, vp);
657 lock_flags = LK_SHARED;
658 vn_lock(vp, lock_flags | LK_RETRY);
662 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
664 if ((ioflg & IO_NOMACCHECK) == 0) {
666 error = mac_vnode_check_read(active_cred, file_cred,
669 error = mac_vnode_check_write(active_cred, file_cred,
674 if (file_cred != NULL)
678 if (do_vn_io_fault(vp, &auio)) {
679 args.kind = VN_IO_FAULT_VOP;
682 args.args.vop_args.vp = vp;
683 error = vn_io_fault1(vp, &auio, &args, td);
684 } else if (rw == UIO_READ) {
685 error = VOP_READ(vp, &auio, ioflg, cred);
686 } else /* if (rw == UIO_WRITE) */ {
687 error = VOP_WRITE(vp, &auio, ioflg, cred);
691 *aresid = auio.uio_resid;
693 if (auio.uio_resid && error == 0)
695 if ((ioflg & IO_NODELOCKED) == 0) {
698 vn_finished_write(mp);
701 if (rl_cookie != NULL)
702 vn_rangelock_unlock(vp, rl_cookie);
707 * Package up an I/O request on a vnode into a uio and do it. The I/O
708 * request is split up into smaller chunks and we try to avoid saturating
709 * the buffer cache while potentially holding a vnode locked, so we
710 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
711 * to give other processes a chance to lock the vnode (either other processes
712 * core'ing the same binary, or unrelated processes scanning the directory).
715 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
716 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
717 struct ucred *file_cred, size_t *aresid, struct thread *td)
726 * Force `offset' to a multiple of MAXBSIZE except possibly
727 * for the first chunk, so that filesystems only need to
728 * write full blocks except possibly for the first and last
731 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
735 if (rw != UIO_READ && vp->v_type == VREG)
738 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
739 ioflg, active_cred, file_cred, &iaresid, td);
740 len -= chunk; /* aresid calc already includes length */
744 base = (char *)base + chunk;
745 kern_yield(PRI_USER);
748 *aresid = len + iaresid;
752 #if OFF_MAX <= LONG_MAX
754 foffset_lock(struct file *fp, int flags)
756 volatile short *flagsp;
760 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
762 if ((flags & FOF_NOLOCK) != 0)
763 return (atomic_load_long(&fp->f_offset));
766 * According to McKusick the vn lock was protecting f_offset here.
767 * It is now protected by the FOFFSET_LOCKED flag.
769 flagsp = &fp->f_vnread_flags;
770 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
771 return (atomic_load_long(&fp->f_offset));
773 sleepq_lock(&fp->f_vnread_flags);
774 state = atomic_load_16(flagsp);
776 if ((state & FOFFSET_LOCKED) == 0) {
777 if (!atomic_fcmpset_acq_16(flagsp, &state,
782 if ((state & FOFFSET_LOCK_WAITING) == 0) {
783 if (!atomic_fcmpset_acq_16(flagsp, &state,
784 state | FOFFSET_LOCK_WAITING))
788 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
789 sleepq_wait(&fp->f_vnread_flags, PUSER -1);
791 sleepq_lock(&fp->f_vnread_flags);
792 state = atomic_load_16(flagsp);
794 res = atomic_load_long(&fp->f_offset);
795 sleepq_release(&fp->f_vnread_flags);
800 foffset_unlock(struct file *fp, off_t val, int flags)
802 volatile short *flagsp;
805 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
807 if ((flags & FOF_NOUPDATE) == 0)
808 atomic_store_long(&fp->f_offset, val);
809 if ((flags & FOF_NEXTOFF_R) != 0)
810 fp->f_nextoff[UIO_READ] = val;
811 if ((flags & FOF_NEXTOFF_W) != 0)
812 fp->f_nextoff[UIO_WRITE] = val;
814 if ((flags & FOF_NOLOCK) != 0)
817 flagsp = &fp->f_vnread_flags;
818 state = atomic_load_16(flagsp);
819 if ((state & FOFFSET_LOCK_WAITING) == 0 &&
820 atomic_cmpset_rel_16(flagsp, state, 0))
823 sleepq_lock(&fp->f_vnread_flags);
824 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
825 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
826 fp->f_vnread_flags = 0;
827 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
828 sleepq_release(&fp->f_vnread_flags);
832 foffset_lock(struct file *fp, int flags)
837 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
839 mtxp = mtx_pool_find(mtxpool_sleep, fp);
841 if ((flags & FOF_NOLOCK) == 0) {
842 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
843 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
844 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
847 fp->f_vnread_flags |= FOFFSET_LOCKED;
855 foffset_unlock(struct file *fp, off_t val, int flags)
859 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
861 mtxp = mtx_pool_find(mtxpool_sleep, fp);
863 if ((flags & FOF_NOUPDATE) == 0)
865 if ((flags & FOF_NEXTOFF_R) != 0)
866 fp->f_nextoff[UIO_READ] = val;
867 if ((flags & FOF_NEXTOFF_W) != 0)
868 fp->f_nextoff[UIO_WRITE] = val;
869 if ((flags & FOF_NOLOCK) == 0) {
870 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
871 ("Lost FOFFSET_LOCKED"));
872 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
873 wakeup(&fp->f_vnread_flags);
874 fp->f_vnread_flags = 0;
881 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
884 if ((flags & FOF_OFFSET) == 0)
885 uio->uio_offset = foffset_lock(fp, flags);
889 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
892 if ((flags & FOF_OFFSET) == 0)
893 foffset_unlock(fp, uio->uio_offset, flags);
897 get_advice(struct file *fp, struct uio *uio)
902 ret = POSIX_FADV_NORMAL;
903 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
906 mtxp = mtx_pool_find(mtxpool_sleep, fp);
908 if (fp->f_advice != NULL &&
909 uio->uio_offset >= fp->f_advice->fa_start &&
910 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
911 ret = fp->f_advice->fa_advice;
917 get_write_ioflag(struct file *fp)
925 mp = atomic_load_ptr(&vp->v_mount);
927 if ((fp->f_flag & O_DIRECT) != 0)
930 if ((fp->f_flag & O_FSYNC) != 0 ||
931 (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
935 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
936 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
937 * fall back to full O_SYNC behavior.
939 if ((fp->f_flag & O_DSYNC) != 0)
940 ioflag |= IO_SYNC | IO_DATASYNC;
946 vn_read_from_obj(struct vnode *vp, struct uio *uio)
949 vm_page_t ma[io_hold_cnt + 2];
954 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
955 obj = atomic_load_ptr(&vp->v_object);
957 return (EJUSTRETURN);
960 * Depends on type stability of vm_objects.
962 vm_object_pip_add(obj, 1);
963 if ((obj->flags & OBJ_DEAD) != 0) {
965 * Note that object might be already reused from the
966 * vnode, and the OBJ_DEAD flag cleared. This is fine,
967 * we recheck for DOOMED vnode state after all pages
968 * are busied, and retract then.
970 * But we check for OBJ_DEAD to ensure that we do not
971 * busy pages while vm_object_terminate_pages()
972 * processes the queue.
978 resid = uio->uio_resid;
979 off = uio->uio_offset;
980 for (i = 0; resid > 0; i++) {
981 MPASS(i < io_hold_cnt + 2);
982 ma[i] = vm_page_grab_unlocked(obj, atop(off),
983 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
989 * Skip invalid pages. Valid mask can be partial only
990 * at EOF, and we clip later.
992 if (vm_page_none_valid(ma[i])) {
993 vm_page_sunbusy(ma[i]);
1001 error = EJUSTRETURN;
1006 * Check VIRF_DOOMED after we busied our pages. Since
1007 * vgonel() terminates the vnode' vm_object, it cannot
1008 * process past pages busied by us.
1010 if (VN_IS_DOOMED(vp)) {
1011 error = EJUSTRETURN;
1015 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1016 if (resid > uio->uio_resid)
1017 resid = uio->uio_resid;
1020 * Unlocked read of vnp_size is safe because truncation cannot
1021 * pass busied page. But we load vnp_size into a local
1022 * variable so that possible concurrent extension does not
1023 * break calculation.
1025 #if defined(__powerpc__) && !defined(__powerpc64__)
1026 vsz = obj->un_pager.vnp.vnp_size;
1028 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1030 if (uio->uio_offset >= vsz) {
1031 error = EJUSTRETURN;
1034 if (uio->uio_offset + resid > vsz)
1035 resid = vsz - uio->uio_offset;
1037 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1040 for (j = 0; j < i; j++) {
1042 vm_page_reference(ma[j]);
1043 vm_page_sunbusy(ma[j]);
1046 vm_object_pip_wakeup(obj);
1049 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1053 * File table vnode read routine.
1056 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1064 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1066 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1069 if (fp->f_flag & FNONBLOCK)
1070 ioflag |= IO_NDELAY;
1071 if (fp->f_flag & O_DIRECT)
1072 ioflag |= IO_DIRECT;
1075 * Try to read from page cache. VIRF_DOOMED check is racy but
1076 * allows us to avoid unneeded work outright.
1078 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1079 (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1080 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1082 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1085 if (error != EJUSTRETURN)
1089 advice = get_advice(fp, uio);
1090 vn_lock(vp, LK_SHARED | LK_RETRY);
1093 case POSIX_FADV_NORMAL:
1094 case POSIX_FADV_SEQUENTIAL:
1095 case POSIX_FADV_NOREUSE:
1096 ioflag |= sequential_heuristic(uio, fp);
1098 case POSIX_FADV_RANDOM:
1099 /* Disable read-ahead for random I/O. */
1102 orig_offset = uio->uio_offset;
1105 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1108 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1109 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1111 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1112 orig_offset != uio->uio_offset)
1114 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1115 * for the backing file after a POSIX_FADV_NOREUSE
1118 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1119 POSIX_FADV_DONTNEED);
1124 * File table vnode write routine.
1127 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1135 bool need_finished_write;
1137 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1139 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1141 if (vp->v_type == VREG)
1144 if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1145 ioflag |= IO_APPEND;
1146 if ((fp->f_flag & FNONBLOCK) != 0)
1147 ioflag |= IO_NDELAY;
1148 ioflag |= get_write_ioflag(fp);
1151 need_finished_write = false;
1152 if (vp->v_type != VCHR) {
1153 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1156 need_finished_write = true;
1159 advice = get_advice(fp, uio);
1161 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1163 case POSIX_FADV_NORMAL:
1164 case POSIX_FADV_SEQUENTIAL:
1165 case POSIX_FADV_NOREUSE:
1166 ioflag |= sequential_heuristic(uio, fp);
1168 case POSIX_FADV_RANDOM:
1169 /* XXX: Is this correct? */
1172 orig_offset = uio->uio_offset;
1175 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1178 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1179 fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1181 if (need_finished_write)
1182 vn_finished_write(mp);
1183 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1184 orig_offset != uio->uio_offset)
1186 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1187 * for the backing file after a POSIX_FADV_NOREUSE
1190 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1191 POSIX_FADV_DONTNEED);
1197 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1198 * prevent the following deadlock:
1200 * Assume that the thread A reads from the vnode vp1 into userspace
1201 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
1202 * currently not resident, then system ends up with the call chain
1203 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1204 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1205 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1206 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1207 * backed by the pages of vnode vp1, and some page in buf2 is not
1208 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1210 * To prevent the lock order reversal and deadlock, vn_io_fault() does
1211 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1212 * Instead, it first tries to do the whole range i/o with pagefaults
1213 * disabled. If all pages in the i/o buffer are resident and mapped,
1214 * VOP will succeed (ignoring the genuine filesystem errors).
1215 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1216 * i/o in chunks, with all pages in the chunk prefaulted and held
1217 * using vm_fault_quick_hold_pages().
1219 * Filesystems using this deadlock avoidance scheme should use the
1220 * array of the held pages from uio, saved in the curthread->td_ma,
1221 * instead of doing uiomove(). A helper function
1222 * vn_io_fault_uiomove() converts uiomove request into
1223 * uiomove_fromphys() over td_ma array.
1225 * Since vnode locks do not cover the whole i/o anymore, rangelocks
1226 * make the current i/o request atomic with respect to other i/os and
1231 * Decode vn_io_fault_args and perform the corresponding i/o.
1234 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1240 save = vm_fault_disable_pagefaults();
1241 switch (args->kind) {
1242 case VN_IO_FAULT_FOP:
1243 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1244 uio, args->cred, args->flags, td);
1246 case VN_IO_FAULT_VOP:
1247 if (uio->uio_rw == UIO_READ) {
1248 error = VOP_READ(args->args.vop_args.vp, uio,
1249 args->flags, args->cred);
1250 } else if (uio->uio_rw == UIO_WRITE) {
1251 error = VOP_WRITE(args->args.vop_args.vp, uio,
1252 args->flags, args->cred);
1256 panic("vn_io_fault_doio: unknown kind of io %d %d",
1257 args->kind, uio->uio_rw);
1259 vm_fault_enable_pagefaults(save);
1264 vn_io_fault_touch(char *base, const struct uio *uio)
1269 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1275 vn_io_fault_prefault_user(const struct uio *uio)
1278 const struct iovec *iov;
1283 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1284 ("vn_io_fault_prefault userspace"));
1288 resid = uio->uio_resid;
1289 base = iov->iov_base;
1292 error = vn_io_fault_touch(base, uio);
1295 if (len < PAGE_SIZE) {
1297 error = vn_io_fault_touch(base + len - 1, uio);
1302 if (++i >= uio->uio_iovcnt)
1304 iov = uio->uio_iov + i;
1305 base = iov->iov_base;
1317 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1318 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1319 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1320 * into args and call vn_io_fault1() to handle faults during the user
1321 * mode buffer accesses.
1324 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1327 vm_page_t ma[io_hold_cnt + 2];
1328 struct uio *uio_clone, short_uio;
1329 struct iovec short_iovec[1];
1330 vm_page_t *prev_td_ma;
1332 vm_offset_t addr, end;
1335 int error, cnt, saveheld, prev_td_ma_cnt;
1337 if (vn_io_fault_prefault) {
1338 error = vn_io_fault_prefault_user(uio);
1340 return (error); /* Or ignore ? */
1343 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1346 * The UFS follows IO_UNIT directive and replays back both
1347 * uio_offset and uio_resid if an error is encountered during the
1348 * operation. But, since the iovec may be already advanced,
1349 * uio is still in an inconsistent state.
1351 * Cache a copy of the original uio, which is advanced to the redo
1352 * point using UIO_NOCOPY below.
1354 uio_clone = cloneuio(uio);
1355 resid = uio->uio_resid;
1357 short_uio.uio_segflg = UIO_USERSPACE;
1358 short_uio.uio_rw = uio->uio_rw;
1359 short_uio.uio_td = uio->uio_td;
1361 error = vn_io_fault_doio(args, uio, td);
1362 if (error != EFAULT)
1365 atomic_add_long(&vn_io_faults_cnt, 1);
1366 uio_clone->uio_segflg = UIO_NOCOPY;
1367 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1368 uio_clone->uio_segflg = uio->uio_segflg;
1370 saveheld = curthread_pflags_set(TDP_UIOHELD);
1371 prev_td_ma = td->td_ma;
1372 prev_td_ma_cnt = td->td_ma_cnt;
1374 while (uio_clone->uio_resid != 0) {
1375 len = uio_clone->uio_iov->iov_len;
1377 KASSERT(uio_clone->uio_iovcnt >= 1,
1378 ("iovcnt underflow"));
1379 uio_clone->uio_iov++;
1380 uio_clone->uio_iovcnt--;
1383 if (len > ptoa(io_hold_cnt))
1384 len = ptoa(io_hold_cnt);
1385 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1386 end = round_page(addr + len);
1391 cnt = atop(end - trunc_page(addr));
1393 * A perfectly misaligned address and length could cause
1394 * both the start and the end of the chunk to use partial
1395 * page. +2 accounts for such a situation.
1397 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1398 addr, len, prot, ma, io_hold_cnt + 2);
1403 short_uio.uio_iov = &short_iovec[0];
1404 short_iovec[0].iov_base = (void *)addr;
1405 short_uio.uio_iovcnt = 1;
1406 short_uio.uio_resid = short_iovec[0].iov_len = len;
1407 short_uio.uio_offset = uio_clone->uio_offset;
1409 td->td_ma_cnt = cnt;
1411 error = vn_io_fault_doio(args, &short_uio, td);
1412 vm_page_unhold_pages(ma, cnt);
1413 adv = len - short_uio.uio_resid;
1415 uio_clone->uio_iov->iov_base =
1416 (char *)uio_clone->uio_iov->iov_base + adv;
1417 uio_clone->uio_iov->iov_len -= adv;
1418 uio_clone->uio_resid -= adv;
1419 uio_clone->uio_offset += adv;
1421 uio->uio_resid -= adv;
1422 uio->uio_offset += adv;
1424 if (error != 0 || adv == 0)
1427 td->td_ma = prev_td_ma;
1428 td->td_ma_cnt = prev_td_ma_cnt;
1429 curthread_pflags_restore(saveheld);
1431 free(uio_clone, M_IOV);
1436 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1437 int flags, struct thread *td)
1442 struct vn_io_fault_args args;
1444 bool do_io_fault, do_rangelock;
1446 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1450 * The ability to read(2) on a directory has historically been
1451 * allowed for all users, but this can and has been the source of
1452 * at least one security issue in the past. As such, it is now hidden
1453 * away behind a sysctl for those that actually need it to use it, and
1454 * restricted to root when it's turned on to make it relatively safe to
1455 * leave on for longer sessions of need.
1457 if (vp->v_type == VDIR) {
1458 KASSERT(uio->uio_rw == UIO_READ,
1459 ("illegal write attempted on a directory"));
1460 if (!vfs_allow_read_dir)
1462 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1466 do_io_fault = do_vn_io_fault(vp, uio);
1467 do_rangelock = do_io_fault || (vn_irflag_read(vp) & VIRF_PGREAD) != 0;
1468 foffset_lock_uio(fp, uio, flags);
1470 if (uio->uio_rw == UIO_READ) {
1471 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1472 uio->uio_offset + uio->uio_resid);
1473 } else if ((fp->f_flag & O_APPEND) != 0 ||
1474 (flags & FOF_OFFSET) == 0) {
1475 /* For appenders, punt and lock the whole range. */
1476 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1478 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1479 uio->uio_offset + uio->uio_resid);
1483 args.kind = VN_IO_FAULT_FOP;
1484 args.args.fop_args.fp = fp;
1485 args.args.fop_args.doio = doio;
1486 args.cred = active_cred;
1487 args.flags = flags | FOF_OFFSET;
1488 error = vn_io_fault1(vp, uio, &args, td);
1490 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1493 vn_rangelock_unlock(vp, rl_cookie);
1494 foffset_unlock_uio(fp, uio, flags);
1499 * Helper function to perform the requested uiomove operation using
1500 * the held pages for io->uio_iov[0].iov_base buffer instead of
1501 * copyin/copyout. Access to the pages with uiomove_fromphys()
1502 * instead of iov_base prevents page faults that could occur due to
1503 * pmap_collect() invalidating the mapping created by
1504 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1505 * object cleanup revoking the write access from page mappings.
1507 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1508 * instead of plain uiomove().
1511 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1513 struct uio transp_uio;
1514 struct iovec transp_iov[1];
1520 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1521 uio->uio_segflg != UIO_USERSPACE)
1522 return (uiomove(data, xfersize, uio));
1524 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1525 transp_iov[0].iov_base = data;
1526 transp_uio.uio_iov = &transp_iov[0];
1527 transp_uio.uio_iovcnt = 1;
1528 if (xfersize > uio->uio_resid)
1529 xfersize = uio->uio_resid;
1530 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1531 transp_uio.uio_offset = 0;
1532 transp_uio.uio_segflg = UIO_SYSSPACE;
1534 * Since transp_iov points to data, and td_ma page array
1535 * corresponds to original uio->uio_iov, we need to invert the
1536 * direction of the i/o operation as passed to
1537 * uiomove_fromphys().
1539 switch (uio->uio_rw) {
1541 transp_uio.uio_rw = UIO_READ;
1544 transp_uio.uio_rw = UIO_WRITE;
1547 transp_uio.uio_td = uio->uio_td;
1548 error = uiomove_fromphys(td->td_ma,
1549 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1550 xfersize, &transp_uio);
1551 adv = xfersize - transp_uio.uio_resid;
1553 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1554 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1556 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1558 td->td_ma_cnt -= pgadv;
1559 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1560 uio->uio_iov->iov_len -= adv;
1561 uio->uio_resid -= adv;
1562 uio->uio_offset += adv;
1567 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1571 vm_offset_t iov_base;
1575 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1576 uio->uio_segflg != UIO_USERSPACE)
1577 return (uiomove_fromphys(ma, offset, xfersize, uio));
1579 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1580 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1581 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1582 switch (uio->uio_rw) {
1584 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1588 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1592 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1594 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1596 td->td_ma_cnt -= pgadv;
1597 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1598 uio->uio_iov->iov_len -= cnt;
1599 uio->uio_resid -= cnt;
1600 uio->uio_offset += cnt;
1605 * File table truncate routine.
1608 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1620 * Lock the whole range for truncation. Otherwise split i/o
1621 * might happen partly before and partly after the truncation.
1623 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1624 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1627 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1628 AUDIT_ARG_VNODE1(vp);
1629 if (vp->v_type == VDIR) {
1634 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1638 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1642 vn_finished_write(mp);
1644 vn_rangelock_unlock(vp, rl_cookie);
1645 if (error == ERELOOKUP)
1651 * Truncate a file that is already locked.
1654 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1660 error = VOP_ADD_WRITECOUNT(vp, 1);
1663 vattr.va_size = length;
1665 vattr.va_vaflags |= VA_SYNC;
1666 error = VOP_SETATTR(vp, &vattr, cred);
1667 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1673 * File table vnode stat routine.
1676 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1679 struct vnode *vp = fp->f_vnode;
1682 vn_lock(vp, LK_SHARED | LK_RETRY);
1683 error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1690 * File table vnode ioctl routine.
1693 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1698 struct fiobmap2_arg *bmarg;
1702 switch (vp->v_type) {
1707 vn_lock(vp, LK_SHARED | LK_RETRY);
1708 error = VOP_GETATTR(vp, &vattr, active_cred);
1711 *(int *)data = vattr.va_size - fp->f_offset;
1714 bmarg = (struct fiobmap2_arg *)data;
1715 vn_lock(vp, LK_SHARED | LK_RETRY);
1717 error = mac_vnode_check_read(active_cred, fp->f_cred,
1721 error = VOP_BMAP(vp, bmarg->bn, NULL,
1722 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1729 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1734 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1742 * File table vnode poll routine.
1745 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1752 #if defined(MAC) || defined(AUDIT)
1753 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1754 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1755 AUDIT_ARG_VNODE1(vp);
1756 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1762 error = VOP_POLL(vp, events, fp->f_cred, td);
1767 * Acquire the requested lock and then check for validity. LK_RETRY
1768 * permits vn_lock to return doomed vnodes.
1770 static int __noinline
1771 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1775 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1776 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1779 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1781 if ((flags & LK_RETRY) == 0) {
1792 * Nothing to do if we got the lock.
1798 * Interlock was dropped by the call in _vn_lock.
1800 flags &= ~LK_INTERLOCK;
1802 error = VOP_LOCK1(vp, flags, file, line);
1803 } while (error != 0);
1808 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1812 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1813 ("vn_lock: no locktype (%d passed)", flags));
1814 VNPASS(vp->v_holdcnt > 0, vp);
1815 error = VOP_LOCK1(vp, flags, file, line);
1816 if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1817 return (_vn_lock_fallback(vp, flags, file, line, error));
1822 * File table vnode close routine.
1825 vn_closefile(struct file *fp, struct thread *td)
1833 fp->f_ops = &badfileops;
1834 ref = (fp->f_flag & FHASLOCK) != 0;
1836 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1838 if (__predict_false(ref)) {
1839 lf.l_whence = SEEK_SET;
1842 lf.l_type = F_UNLCK;
1843 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1850 * Preparing to start a filesystem write operation. If the operation is
1851 * permitted, then we bump the count of operations in progress and
1852 * proceed. If a suspend request is in progress, we wait until the
1853 * suspension is over, and then proceed.
1856 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1858 struct mount_pcpu *mpcpu;
1861 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1862 vfs_op_thread_enter(mp, mpcpu)) {
1863 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1864 vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1865 vfs_op_thread_exit(mp, mpcpu);
1870 mtx_assert(MNT_MTX(mp), MA_OWNED);
1877 * Check on status of suspension.
1879 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1880 mp->mnt_susp_owner != curthread) {
1881 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1882 (flags & PCATCH) : 0) | (PUSER - 1);
1883 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1884 if (flags & V_NOWAIT) {
1885 error = EWOULDBLOCK;
1888 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1894 if (flags & V_XSLEEP)
1896 mp->mnt_writeopcount++;
1898 if (error != 0 || (flags & V_XSLEEP) != 0)
1905 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1910 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1911 ("V_MNTREF requires mp"));
1915 * If a vnode is provided, get and return the mount point that
1916 * to which it will write.
1919 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1921 if (error != EOPNOTSUPP)
1926 if ((mp = *mpp) == NULL)
1930 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1932 * As long as a vnode is not provided we need to acquire a
1933 * refcount for the provided mountpoint too, in order to
1934 * emulate a vfs_ref().
1936 if (vp == NULL && (flags & V_MNTREF) == 0)
1939 return (vn_start_write_refed(mp, flags, false));
1943 * Secondary suspension. Used by operations such as vop_inactive
1944 * routines that are needed by the higher level functions. These
1945 * are allowed to proceed until all the higher level functions have
1946 * completed (indicated by mnt_writeopcount dropping to zero). At that
1947 * time, these operations are halted until the suspension is over.
1950 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1955 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1956 ("V_MNTREF requires mp"));
1960 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1962 if (error != EOPNOTSUPP)
1968 * If we are not suspended or have not yet reached suspended
1969 * mode, then let the operation proceed.
1971 if ((mp = *mpp) == NULL)
1975 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1977 * As long as a vnode is not provided we need to acquire a
1978 * refcount for the provided mountpoint too, in order to
1979 * emulate a vfs_ref().
1982 if (vp == NULL && (flags & V_MNTREF) == 0)
1984 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1985 mp->mnt_secondary_writes++;
1986 mp->mnt_secondary_accwrites++;
1990 if (flags & V_NOWAIT) {
1993 return (EWOULDBLOCK);
1996 * Wait for the suspension to finish.
1998 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1999 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
2008 * Filesystem write operation has completed. If we are suspending and this
2009 * operation is the last one, notify the suspender that the suspension is
2013 vn_finished_write(struct mount *mp)
2015 struct mount_pcpu *mpcpu;
2021 if (vfs_op_thread_enter(mp, mpcpu)) {
2022 vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2023 vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2024 vfs_op_thread_exit(mp, mpcpu);
2029 vfs_assert_mount_counters(mp);
2031 c = --mp->mnt_writeopcount;
2032 if (mp->mnt_vfs_ops == 0) {
2033 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2038 vfs_dump_mount_counters(mp);
2039 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2040 wakeup(&mp->mnt_writeopcount);
2045 * Filesystem secondary write operation has completed. If we are
2046 * suspending and this operation is the last one, notify the suspender
2047 * that the suspension is now in effect.
2050 vn_finished_secondary_write(struct mount *mp)
2056 mp->mnt_secondary_writes--;
2057 if (mp->mnt_secondary_writes < 0)
2058 panic("vn_finished_secondary_write: neg cnt");
2059 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2060 mp->mnt_secondary_writes <= 0)
2061 wakeup(&mp->mnt_secondary_writes);
2066 * Request a filesystem to suspend write operations.
2069 vfs_write_suspend(struct mount *mp, int flags)
2076 vfs_assert_mount_counters(mp);
2077 if (mp->mnt_susp_owner == curthread) {
2078 vfs_op_exit_locked(mp);
2082 while (mp->mnt_kern_flag & MNTK_SUSPEND)
2083 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2086 * Unmount holds a write reference on the mount point. If we
2087 * own busy reference and drain for writers, we deadlock with
2088 * the reference draining in the unmount path. Callers of
2089 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2090 * vfs_busy() reference is owned and caller is not in the
2093 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2094 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2095 vfs_op_exit_locked(mp);
2100 mp->mnt_kern_flag |= MNTK_SUSPEND;
2101 mp->mnt_susp_owner = curthread;
2102 if (mp->mnt_writeopcount > 0)
2103 (void) msleep(&mp->mnt_writeopcount,
2104 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2107 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2108 vfs_write_resume(mp, 0);
2109 /* vfs_write_resume does vfs_op_exit() for us */
2115 * Request a filesystem to resume write operations.
2118 vfs_write_resume(struct mount *mp, int flags)
2122 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2123 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2124 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2126 mp->mnt_susp_owner = NULL;
2127 wakeup(&mp->mnt_writeopcount);
2128 wakeup(&mp->mnt_flag);
2129 curthread->td_pflags &= ~TDP_IGNSUSP;
2130 if ((flags & VR_START_WRITE) != 0) {
2132 mp->mnt_writeopcount++;
2135 if ((flags & VR_NO_SUSPCLR) == 0)
2138 } else if ((flags & VR_START_WRITE) != 0) {
2140 vn_start_write_refed(mp, 0, true);
2147 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2151 vfs_write_suspend_umnt(struct mount *mp)
2155 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2156 ("vfs_write_suspend_umnt: recursed"));
2158 /* dounmount() already called vn_start_write(). */
2160 vn_finished_write(mp);
2161 error = vfs_write_suspend(mp, 0);
2163 vn_start_write(NULL, &mp, V_WAIT);
2167 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2170 vn_start_write(NULL, &mp, V_WAIT);
2172 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2173 wakeup(&mp->mnt_flag);
2175 curthread->td_pflags |= TDP_IGNSUSP;
2180 * Implement kqueues for files by translating it to vnode operation.
2183 vn_kqfilter(struct file *fp, struct knote *kn)
2186 return (VOP_KQFILTER(fp->f_vnode, kn));
2190 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2192 if ((fp->f_flag & FKQALLOWED) == 0)
2194 return (vn_kqfilter(fp, kn));
2198 * Simplified in-kernel wrapper calls for extended attribute access.
2199 * Both calls pass in a NULL credential, authorizing as "kernel" access.
2200 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2203 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2204 const char *attrname, int *buflen, char *buf, struct thread *td)
2210 iov.iov_len = *buflen;
2213 auio.uio_iov = &iov;
2214 auio.uio_iovcnt = 1;
2215 auio.uio_rw = UIO_READ;
2216 auio.uio_segflg = UIO_SYSSPACE;
2218 auio.uio_offset = 0;
2219 auio.uio_resid = *buflen;
2221 if ((ioflg & IO_NODELOCKED) == 0)
2222 vn_lock(vp, LK_SHARED | LK_RETRY);
2224 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2226 /* authorize attribute retrieval as kernel */
2227 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2230 if ((ioflg & IO_NODELOCKED) == 0)
2234 *buflen = *buflen - auio.uio_resid;
2241 * XXX failure mode if partially written?
2244 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2245 const char *attrname, int buflen, char *buf, struct thread *td)
2252 iov.iov_len = buflen;
2255 auio.uio_iov = &iov;
2256 auio.uio_iovcnt = 1;
2257 auio.uio_rw = UIO_WRITE;
2258 auio.uio_segflg = UIO_SYSSPACE;
2260 auio.uio_offset = 0;
2261 auio.uio_resid = buflen;
2263 if ((ioflg & IO_NODELOCKED) == 0) {
2264 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2266 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2269 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2271 /* authorize attribute setting as kernel */
2272 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2274 if ((ioflg & IO_NODELOCKED) == 0) {
2275 vn_finished_write(mp);
2283 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2284 const char *attrname, struct thread *td)
2289 if ((ioflg & IO_NODELOCKED) == 0) {
2290 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2292 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2295 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2297 /* authorize attribute removal as kernel */
2298 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2299 if (error == EOPNOTSUPP)
2300 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2303 if ((ioflg & IO_NODELOCKED) == 0) {
2304 vn_finished_write(mp);
2312 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2316 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2320 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2323 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2328 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2329 int lkflags, struct vnode **rvp)
2334 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2336 ltype = VOP_ISLOCKED(vp);
2337 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2338 ("vn_vget_ino: vp not locked"));
2339 error = vfs_busy(mp, MBF_NOWAIT);
2343 error = vfs_busy(mp, 0);
2344 vn_lock(vp, ltype | LK_RETRY);
2348 if (VN_IS_DOOMED(vp)) {
2354 error = alloc(mp, alloc_arg, lkflags, rvp);
2356 if (error != 0 || *rvp != vp)
2357 vn_lock(vp, ltype | LK_RETRY);
2358 if (VN_IS_DOOMED(vp)) {
2371 vn_send_sigxfsz(struct proc *p)
2374 kern_psignal(p, SIGXFSZ);
2379 vn_rlimit_trunc(u_quad_t size, struct thread *td)
2381 if (size <= lim_cur(td, RLIMIT_FSIZE))
2383 vn_send_sigxfsz(td->td_proc);
2388 vn_rlimit_fsizex1(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2389 bool adj, struct thread *td)
2394 if (vp->v_type != VREG)
2398 * Handle file system maximum file size.
2400 if (maxfsz != 0 && uio->uio_offset + uio->uio_resid > maxfsz) {
2401 if (!adj || uio->uio_offset >= maxfsz)
2403 uio->uio_resid = maxfsz - uio->uio_offset;
2407 * This is kernel write (e.g. vnode_pager) or accounting
2408 * write, ignore limit.
2410 if (td == NULL || (td->td_pflags2 & TDP2_ACCT) != 0)
2414 * Calculate file size limit.
2416 ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2417 lim = __predict_false(ktr_write) ? td->td_ktr_io_lim :
2418 lim_cur(td, RLIMIT_FSIZE);
2421 * Is the limit reached?
2423 if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2427 * Prepared filesystems can handle writes truncated to the
2430 if (adj && (uoff_t)uio->uio_offset < lim) {
2431 uio->uio_resid = lim - (uoff_t)uio->uio_offset;
2435 if (!ktr_write || ktr_filesize_limit_signal)
2436 vn_send_sigxfsz(td->td_proc);
2441 * Helper for VOP_WRITE() implementations, the common code to
2442 * handle maximum supported file size on the filesystem, and
2443 * RLIMIT_FSIZE, except for special writes from accounting subsystem
2446 * For maximum file size (maxfsz argument):
2447 * - return EFBIG if uio_offset is beyond it
2448 * - otherwise, clamp uio_resid if write would extend file beyond maxfsz.
2451 * - return EFBIG and send SIGXFSZ if uio_offset is beyond the limit
2452 * - otherwise, clamp uio_resid if write would extend file beyond limit.
2454 * If clamping occured, the adjustment for uio_resid is stored in
2455 * *resid_adj, to be re-applied by vn_rlimit_fsizex_res() on return
2459 vn_rlimit_fsizex(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2460 ssize_t *resid_adj, struct thread *td)
2466 resid_orig = uio->uio_resid;
2467 adj = resid_adj != NULL;
2468 error = vn_rlimit_fsizex1(vp, uio, maxfsz, adj, td);
2470 *resid_adj = resid_orig - uio->uio_resid;
2475 vn_rlimit_fsizex_res(struct uio *uio, ssize_t resid_adj)
2477 uio->uio_resid += resid_adj;
2481 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2484 return (vn_rlimit_fsizex(vp, __DECONST(struct uio *, uio), 0, NULL,
2489 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2496 vn_lock(vp, LK_SHARED | LK_RETRY);
2497 AUDIT_ARG_VNODE1(vp);
2500 return (setfmode(td, active_cred, vp, mode));
2504 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2511 vn_lock(vp, LK_SHARED | LK_RETRY);
2512 AUDIT_ARG_VNODE1(vp);
2515 return (setfown(td, active_cred, vp, uid, gid));
2519 * Remove pages in the range ["start", "end") from the vnode's VM object. If
2520 * "end" is 0, then the range extends to the end of the object.
2523 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2527 if ((object = vp->v_object) == NULL)
2529 VM_OBJECT_WLOCK(object);
2530 vm_object_page_remove(object, start, end, 0);
2531 VM_OBJECT_WUNLOCK(object);
2535 * Like vn_pages_remove(), but skips invalid pages, which by definition are not
2536 * mapped into any process' address space. Filesystems may use this in
2537 * preference to vn_pages_remove() to avoid blocking on pages busied in
2538 * preparation for a VOP_GETPAGES.
2541 vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2545 if ((object = vp->v_object) == NULL)
2547 VM_OBJECT_WLOCK(object);
2548 vm_object_page_remove(object, start, end, OBJPR_VALIDONLY);
2549 VM_OBJECT_WUNLOCK(object);
2553 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2561 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2562 ("Wrong command %lu", cmd));
2564 if (vn_lock(vp, LK_EXCLUSIVE) != 0)
2566 if (vp->v_type != VREG) {
2570 error = VOP_GETATTR(vp, &va, cred);
2574 if (noff < 0 || noff >= va.va_size) {
2579 /* See the comment in ufs_bmap_seekdata(). */
2580 vnode_pager_clean_sync(vp);
2582 bsize = vp->v_mount->mnt_stat.f_iosize;
2583 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2585 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2586 if (error == EOPNOTSUPP) {
2590 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2591 (bnp != -1 && cmd == FIOSEEKDATA)) {
2598 if (noff > va.va_size)
2600 /* noff == va.va_size. There is an implicit hole at the end of file. */
2601 if (cmd == FIOSEEKDATA)
2611 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2616 off_t foffset, size;
2619 cred = td->td_ucred;
2621 foffset = foffset_lock(fp, 0);
2622 noneg = (vp->v_type != VCHR);
2628 (offset > 0 && foffset > OFF_MAX - offset))) {
2635 vn_lock(vp, LK_SHARED | LK_RETRY);
2636 error = VOP_GETATTR(vp, &vattr, cred);
2642 * If the file references a disk device, then fetch
2643 * the media size and use that to determine the ending
2646 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2647 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2648 vattr.va_size = size;
2650 (vattr.va_size > OFF_MAX ||
2651 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2655 offset += vattr.va_size;
2660 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2661 if (error == ENOTTY)
2665 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2666 if (error == ENOTTY)
2672 if (error == 0 && noneg && offset < 0)
2676 VFS_KNOTE_UNLOCKED(vp, 0);
2677 td->td_uretoff.tdu_off = offset;
2679 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2684 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2690 * Grant permission if the caller is the owner of the file, or
2691 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2692 * on the file. If the time pointer is null, then write
2693 * permission on the file is also sufficient.
2695 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2696 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2697 * will be allowed to set the times [..] to the current
2700 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2701 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2702 error = VOP_ACCESS(vp, VWRITE, cred, td);
2707 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2712 if (fp->f_type == DTYPE_FIFO)
2713 kif->kf_type = KF_TYPE_FIFO;
2715 kif->kf_type = KF_TYPE_VNODE;
2718 FILEDESC_SUNLOCK(fdp);
2719 error = vn_fill_kinfo_vnode(vp, kif);
2721 FILEDESC_SLOCK(fdp);
2726 vn_fill_junk(struct kinfo_file *kif)
2731 * Simulate vn_fullpath returning changing values for a given
2732 * vp during e.g. coredump.
2734 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2735 olen = strlen(kif->kf_path);
2737 strcpy(&kif->kf_path[len - 1], "$");
2739 for (; olen < len; olen++)
2740 strcpy(&kif->kf_path[olen], "A");
2744 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2747 char *fullpath, *freepath;
2750 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2753 error = vn_fullpath(vp, &fullpath, &freepath);
2755 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2757 if (freepath != NULL)
2758 free(freepath, M_TEMP);
2760 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2765 * Retrieve vnode attributes.
2767 va.va_fsid = VNOVAL;
2769 vn_lock(vp, LK_SHARED | LK_RETRY);
2770 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2774 if (va.va_fsid != VNOVAL)
2775 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2777 kif->kf_un.kf_file.kf_file_fsid =
2778 vp->v_mount->mnt_stat.f_fsid.val[0];
2779 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2780 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2781 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2782 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2783 kif->kf_un.kf_file.kf_file_size = va.va_size;
2784 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2785 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2786 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2787 kif->kf_un.kf_file.kf_file_nlink = va.va_nlink;
2792 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2793 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2797 struct pmckern_map_in pkm;
2803 boolean_t writecounted;
2806 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2807 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2809 * POSIX shared-memory objects are defined to have
2810 * kernel persistence, and are not defined to support
2811 * read(2)/write(2) -- or even open(2). Thus, we can
2812 * use MAP_ASYNC to trade on-disk coherence for speed.
2813 * The shm_open(3) library routine turns on the FPOSIXSHM
2814 * flag to request this behavior.
2816 if ((fp->f_flag & FPOSIXSHM) != 0)
2817 flags |= MAP_NOSYNC;
2822 * Ensure that file and memory protections are
2823 * compatible. Note that we only worry about
2824 * writability if mapping is shared; in this case,
2825 * current and max prot are dictated by the open file.
2826 * XXX use the vnode instead? Problem is: what
2827 * credentials do we use for determination? What if
2828 * proc does a setuid?
2831 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2832 maxprot = VM_PROT_NONE;
2833 if ((prot & VM_PROT_EXECUTE) != 0)
2836 maxprot = VM_PROT_EXECUTE;
2837 if ((fp->f_flag & FREAD) != 0)
2838 maxprot |= VM_PROT_READ;
2839 else if ((prot & VM_PROT_READ) != 0)
2843 * If we are sharing potential changes via MAP_SHARED and we
2844 * are trying to get write permission although we opened it
2845 * without asking for it, bail out.
2847 if ((flags & MAP_SHARED) != 0) {
2848 if ((fp->f_flag & FWRITE) != 0)
2849 maxprot |= VM_PROT_WRITE;
2850 else if ((prot & VM_PROT_WRITE) != 0)
2853 maxprot |= VM_PROT_WRITE;
2854 cap_maxprot |= VM_PROT_WRITE;
2856 maxprot &= cap_maxprot;
2859 * For regular files and shared memory, POSIX requires that
2860 * the value of foff be a legitimate offset within the data
2861 * object. In particular, negative offsets are invalid.
2862 * Blocking negative offsets and overflows here avoids
2863 * possible wraparound or user-level access into reserved
2864 * ranges of the data object later. In contrast, POSIX does
2865 * not dictate how offsets are used by device drivers, so in
2866 * the case of a device mapping a negative offset is passed
2873 foff > OFF_MAX - size)
2876 writecounted = FALSE;
2877 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2878 &foff, &object, &writecounted);
2881 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2882 foff, writecounted, td);
2885 * If this mapping was accounted for in the vnode's
2886 * writecount, then undo that now.
2889 vm_pager_release_writecount(object, 0, size);
2890 vm_object_deallocate(object);
2893 /* Inform hwpmc(4) if an executable is being mapped. */
2894 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2895 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2897 pkm.pm_address = (uintptr_t) *addr;
2898 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2906 vn_fsid(struct vnode *vp, struct vattr *va)
2910 f = &vp->v_mount->mnt_stat.f_fsid;
2911 va->va_fsid = (uint32_t)f->val[1];
2912 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2913 va->va_fsid += (uint32_t)f->val[0];
2917 vn_fsync_buf(struct vnode *vp, int waitfor)
2919 struct buf *bp, *nbp;
2922 int error, maxretry;
2925 maxretry = 10000; /* large, arbitrarily chosen */
2927 if (vp->v_type == VCHR) {
2929 mp = vp->v_rdev->si_mountpt;
2936 * MARK/SCAN initialization to avoid infinite loops.
2938 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2939 bp->b_vflags &= ~BV_SCANNED;
2944 * Flush all dirty buffers associated with a vnode.
2947 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2948 if ((bp->b_vflags & BV_SCANNED) != 0)
2950 bp->b_vflags |= BV_SCANNED;
2951 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2952 if (waitfor != MNT_WAIT)
2955 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2956 BO_LOCKPTR(bo)) != 0) {
2963 KASSERT(bp->b_bufobj == bo,
2964 ("bp %p wrong b_bufobj %p should be %p",
2965 bp, bp->b_bufobj, bo));
2966 if ((bp->b_flags & B_DELWRI) == 0)
2967 panic("fsync: not dirty");
2968 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2974 if (maxretry < 1000)
2975 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2981 * If synchronous the caller expects us to completely resolve all
2982 * dirty buffers in the system. Wait for in-progress I/O to
2983 * complete (which could include background bitmap writes), then
2984 * retry if dirty blocks still exist.
2986 if (waitfor == MNT_WAIT) {
2987 bufobj_wwait(bo, 0, 0);
2988 if (bo->bo_dirty.bv_cnt > 0) {
2990 * If we are unable to write any of these buffers
2991 * then we fail now rather than trying endlessly
2992 * to write them out.
2994 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2995 if ((error = bp->b_error) != 0)
2997 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2998 (error == 0 && --maxretry >= 0))
3006 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
3012 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
3013 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
3014 * to do the actual copy.
3015 * vn_generic_copy_file_range() is factored out, so it can be called
3016 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
3017 * different file systems.
3020 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
3021 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
3022 struct ucred *outcred, struct thread *fsize_td)
3024 struct mount *inmp, *outmp;
3025 struct vnode *invpl, *outvpl;
3030 invpl = outvpl = NULL;
3032 *lenp = 0; /* For error returns. */
3035 /* Do some sanity checks on the arguments. */
3036 if (invp->v_type == VDIR || outvp->v_type == VDIR)
3038 else if (*inoffp < 0 || *outoffp < 0 ||
3039 invp->v_type != VREG || outvp->v_type != VREG)
3044 /* Ensure offset + len does not wrap around. */
3047 if (uval > INT64_MAX)
3048 len = INT64_MAX - *inoffp;
3051 if (uval > INT64_MAX)
3052 len = INT64_MAX - *outoffp;
3056 error = VOP_GETLOWVNODE(invp, &invpl, FREAD);
3059 error = VOP_GETLOWVNODE(outvp, &outvpl, FWRITE);
3063 inmp = invpl->v_mount;
3064 outmp = outvpl->v_mount;
3065 if (inmp == NULL || outmp == NULL)
3069 error = vfs_busy(inmp, 0);
3074 error = vfs_busy(outmp, MBF_NOWAIT);
3077 error = vfs_busy(outmp, 0);
3088 * If the two vnode are for the same file system, call
3089 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
3090 * which can handle copies across multiple file systems.
3094 error = VOP_COPY_FILE_RANGE(invpl, inoffp, outvpl, outoffp,
3095 lenp, flags, incred, outcred, fsize_td);
3098 if (error == ENOSYS)
3099 error = vn_generic_copy_file_range(invpl, inoffp, outvpl,
3100 outoffp, lenp, flags, incred, outcred, fsize_td);
3115 * Test len bytes of data starting at dat for all bytes == 0.
3116 * Return true if all bytes are zero, false otherwise.
3117 * Expects dat to be well aligned.
3120 mem_iszero(void *dat, int len)
3126 for (p = dat; len > 0; len -= sizeof(*p), p++) {
3127 if (len >= sizeof(*p)) {
3131 cp = (const char *)p;
3132 for (i = 0; i < len; i++, cp++)
3141 * Look for a hole in the output file and, if found, adjust *outoffp
3142 * and *xferp to skip past the hole.
3143 * *xferp is the entire hole length to be written and xfer2 is how many bytes
3144 * to be written as 0's upon return.
3147 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3148 off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3153 if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3154 *dataoffp = *outoffp;
3155 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3158 *holeoffp = *dataoffp;
3159 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3162 if (error != 0 || *holeoffp == *dataoffp) {
3164 * Since outvp is unlocked, it may be possible for
3165 * another thread to do a truncate(), lseek(), write()
3166 * creating a hole at startoff between the above
3167 * VOP_IOCTL() calls, if the other thread does not do
3169 * If that happens, *holeoffp == *dataoffp and finding
3170 * the hole has failed, so disable vn_skip_hole().
3172 *holeoffp = -1; /* Disable use of vn_skip_hole(). */
3175 KASSERT(*dataoffp >= *outoffp,
3176 ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3177 (intmax_t)*dataoffp, (intmax_t)*outoffp));
3178 KASSERT(*holeoffp > *dataoffp,
3179 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3180 (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3184 * If there is a hole before the data starts, advance *outoffp and
3185 * *xferp past the hole.
3187 if (*dataoffp > *outoffp) {
3188 delta = *dataoffp - *outoffp;
3189 if (delta >= *xferp) {
3190 /* Entire *xferp is a hole. */
3197 xfer2 = MIN(xfer2, *xferp);
3201 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3202 * that the write ends at the start of the hole.
3203 * *holeoffp should always be greater than *outoffp, but for the
3204 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3207 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3208 xfer2 = *holeoffp - *outoffp;
3213 * Write an xfer sized chunk to outvp in blksize blocks from dat.
3214 * dat is a maximum of blksize in length and can be written repeatedly in
3216 * If growfile == true, just grow the file via vn_truncate_locked() instead
3217 * of doing actual writes.
3218 * If checkhole == true, a hole is being punched, so skip over any hole
3219 * already in the output file.
3222 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3223 u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3226 off_t dataoff, holeoff, xfer2;
3230 * Loop around doing writes of blksize until write has been completed.
3231 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3232 * done for each iteration, since the xfer argument can be very
3233 * large if there is a large hole to punch in the output file.
3238 xfer2 = MIN(xfer, blksize);
3241 * Punching a hole. Skip writing if there is
3242 * already a hole in the output file.
3244 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3245 &dataoff, &holeoff, cred);
3250 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3255 error = vn_start_write(outvp, &mp, V_WAIT);
3259 error = vn_lock(outvp, LK_EXCLUSIVE);
3261 error = vn_truncate_locked(outvp, outoff + xfer,
3266 error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3268 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3269 outoff, UIO_SYSSPACE, IO_NODELOCKED,
3270 curthread->td_ucred, cred, NULL, curthread);
3277 vn_finished_write(mp);
3278 } while (!growfile && xfer > 0 && error == 0);
3283 * Copy a byte range of one file to another. This function can handle the
3284 * case where invp and outvp are on different file systems.
3285 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3286 * is no better file system specific way to do it.
3289 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3290 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3291 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3293 struct vattr va, inva;
3295 off_t startoff, endoff, xfer, xfer2;
3297 int error, interrupted;
3298 bool cantseek, readzeros, eof, lastblock, holetoeof;
3299 ssize_t aresid, r = 0;
3300 size_t copylen, len, savlen;
3302 long holein, holeout;
3303 struct timespec curts, endts;
3305 holein = holeout = 0;
3306 savlen = len = *lenp;
3311 error = vn_lock(invp, LK_SHARED);
3314 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3316 error = VOP_GETATTR(invp, &inva, incred);
3322 error = vn_start_write(outvp, &mp, V_WAIT);
3324 error = vn_lock(outvp, LK_EXCLUSIVE);
3327 * If fsize_td != NULL, do a vn_rlimit_fsizex() call,
3328 * now that outvp is locked.
3330 if (fsize_td != NULL) {
3333 io.uio_offset = *outoffp;
3335 error = vn_rlimit_fsizex(outvp, &io, 0, &r, fsize_td);
3336 len = savlen = io.uio_resid;
3338 * No need to call vn_rlimit_fsizex_res before return,
3339 * since the uio is local.
3342 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3345 * Holes that are past EOF do not need to be written as a block
3346 * of zero bytes. So, truncate the output file as far as
3347 * possible and then use va.va_size to decide if writing 0
3348 * bytes is necessary in the loop below.
3351 error = VOP_GETATTR(outvp, &va, outcred);
3352 if (error == 0 && va.va_size > *outoffp &&
3353 *outoffp <= OFF_MAX - len && va.va_size <= *outoffp + len &&
3354 *inoffp < inva.va_size &&
3355 *outoffp <= OFF_MAX - (inva.va_size - *inoffp) &&
3356 va.va_size <= *outoffp + (inva.va_size - *inoffp)) {
3358 error = mac_vnode_check_write(curthread->td_ucred,
3362 error = vn_truncate_locked(outvp, *outoffp,
3365 va.va_size = *outoffp;
3370 vn_finished_write(mp);
3374 if (holein == 0 && holeout > 0) {
3376 * For this special case, the input data will be scanned
3377 * for blocks of all 0 bytes. For these blocks, the
3378 * write can be skipped for the output file to create
3379 * an unallocated region.
3380 * Therefore, use the appropriate size for the output file.
3383 if (blksize <= 512) {
3385 * Use f_iosize, since ZFS reports a _PC_MIN_HOLE_SIZE
3386 * of 512, although it actually only creates
3387 * unallocated regions for blocks >= f_iosize.
3389 blksize = outvp->v_mount->mnt_stat.f_iosize;
3393 * Use the larger of the two f_iosize values. If they are
3394 * not the same size, one will normally be an exact multiple of
3395 * the other, since they are both likely to be a power of 2.
3397 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3398 outvp->v_mount->mnt_stat.f_iosize);
3401 /* Clip to sane limits. */
3404 else if (blksize > maxphys)
3406 dat = malloc(blksize, M_TEMP, M_WAITOK);
3409 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3410 * to find holes. Otherwise, just scan the read block for all 0s
3411 * in the inner loop where the data copying is done.
3412 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3413 * support holes on the server, but do not support FIOSEEKHOLE.
3414 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3415 * that this function should return after 1second with a partial
3418 if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3419 getnanouptime(&endts);
3422 timespecclear(&endts);
3423 holetoeof = eof = false;
3424 while (len > 0 && error == 0 && !eof && interrupted == 0) {
3425 endoff = 0; /* To shut up compilers. */
3431 * Find the next data area. If there is just a hole to EOF,
3432 * FIOSEEKDATA should fail with ENXIO.
3433 * (I do not know if any file system will report a hole to
3434 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3435 * will fail for those file systems.)
3437 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3438 * the code just falls through to the inner copy loop.
3442 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3444 if (error == ENXIO) {
3445 startoff = endoff = inva.va_size;
3446 eof = holetoeof = true;
3450 if (error == 0 && !holetoeof) {
3452 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3455 * Since invp is unlocked, it may be possible for
3456 * another thread to do a truncate(), lseek(), write()
3457 * creating a hole at startoff between the above
3458 * VOP_IOCTL() calls, if the other thread does not do
3460 * If that happens, startoff == endoff and finding
3461 * the hole has failed, so set an error.
3463 if (error == 0 && startoff == endoff)
3464 error = EINVAL; /* Any error. Reset to 0. */
3467 if (startoff > *inoffp) {
3468 /* Found hole before data block. */
3469 xfer = MIN(startoff - *inoffp, len);
3470 if (*outoffp < va.va_size) {
3471 /* Must write 0s to punch hole. */
3472 xfer2 = MIN(va.va_size - *outoffp,
3474 memset(dat, 0, MIN(xfer2, blksize));
3475 error = vn_write_outvp(outvp, dat,
3476 *outoffp, xfer2, blksize, false,
3477 holeout > 0, outcred);
3480 if (error == 0 && *outoffp + xfer >
3481 va.va_size && (xfer == len || holetoeof)) {
3482 /* Grow output file (hole at end). */
3483 error = vn_write_outvp(outvp, dat,
3484 *outoffp, xfer, blksize, true,
3492 interrupted = sig_intr();
3493 if (timespecisset(&endts) &&
3495 getnanouptime(&curts);
3496 if (timespeccmp(&curts,
3504 copylen = MIN(len, endoff - startoff);
3516 * Set first xfer to end at a block boundary, so that
3517 * holes are more likely detected in the loop below via
3518 * the for all bytes 0 method.
3520 xfer -= (*inoffp % blksize);
3522 /* Loop copying the data block. */
3523 while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3526 error = vn_lock(invp, LK_SHARED);
3529 error = vn_rdwr(UIO_READ, invp, dat, xfer,
3530 startoff, UIO_SYSSPACE, IO_NODELOCKED,
3531 curthread->td_ucred, incred, &aresid,
3535 if (error == 0 && aresid > 0) {
3536 /* Stop the copy at EOF on the input file. */
3543 * Skip the write for holes past the initial EOF
3544 * of the output file, unless this is the last
3545 * write of the output file at EOF.
3547 readzeros = cantseek ? mem_iszero(dat, xfer) :
3551 if (!cantseek || *outoffp < va.va_size ||
3552 lastblock || !readzeros)
3553 error = vn_write_outvp(outvp, dat,
3554 *outoffp, xfer, blksize,
3555 readzeros && lastblock &&
3556 *outoffp >= va.va_size, false,
3565 interrupted = sig_intr();
3566 if (timespecisset(&endts) &&
3568 getnanouptime(&curts);
3569 if (timespeccmp(&curts,
3581 *lenp = savlen - len;
3587 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3591 off_t olen, ooffset;
3594 int audited_vnode1 = 0;
3598 if (vp->v_type != VREG)
3601 /* Allocating blocks may take a long time, so iterate. */
3608 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3611 error = vn_lock(vp, LK_EXCLUSIVE);
3613 vn_finished_write(mp);
3617 if (!audited_vnode1) {
3618 AUDIT_ARG_VNODE1(vp);
3623 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3626 error = VOP_ALLOCATE(vp, &offset, &len, 0,
3629 vn_finished_write(mp);
3631 if (olen + ooffset != offset + len) {
3632 panic("offset + len changed from %jx/%jx to %jx/%jx",
3633 ooffset, olen, offset, len);
3635 if (error != 0 || len == 0)
3637 KASSERT(olen > len, ("Iteration did not make progress?"));
3644 #define DIRENT_MINSIZE (sizeof(struct dirent) - (MAXNAMLEN+1) + 4)
3647 * Keep this assert as long as sizeof(struct dirent) is used as the maximum
3650 _Static_assert(_GENERIC_MAXDIRSIZ == sizeof(struct dirent),
3651 "'struct dirent' size must be a multiple of its alignment "
3652 "(see _GENERIC_DIRLEN())");
3655 * Returns successive directory entries through some caller's provided buffer.
3657 * This function automatically refills the provided buffer with calls to
3658 * VOP_READDIR() (after MAC permission checks).
3660 * 'td' is used for credentials and passed to uiomove(). 'dirbuf' is the
3661 * caller's buffer to fill and 'dirbuflen' its allocated size. 'dirbuf' must
3662 * be properly aligned to access 'struct dirent' structures and 'dirbuflen'
3663 * must be greater than GENERIC_MAXDIRSIZ to avoid VOP_READDIR() returning
3664 * EINVAL (the latter is not a strong guarantee (yet); but EINVAL will always
3665 * be returned if this requirement is not verified). '*dpp' points to the
3666 * current directory entry in the buffer and '*len' contains the remaining
3667 * valid bytes in 'dirbuf' after 'dpp' (including the pointed entry).
3669 * At first call (or when restarting the read), '*len' must have been set to 0,
3670 * '*off' to 0 (or any valid start offset) and '*eofflag' to 0. There are no
3671 * more entries as soon as '*len' is 0 after a call that returned 0. Calling
3672 * again this function after such a condition is considered an error and EINVAL
3673 * will be returned. Other possible error codes are those of VOP_READDIR(),
3674 * EINTEGRITY if the returned entries do not pass coherency tests, or EINVAL
3675 * (bad call). All errors are unrecoverable, i.e., the state ('*len', '*off'
3676 * and '*eofflag') must be re-initialized before a subsequent call. On error
3677 * or at end of directory, '*dpp' is reset to NULL.
3679 * '*len', '*off' and '*eofflag' are internal state the caller should not
3680 * tamper with except as explained above. '*off' is the next directory offset
3681 * to read from to refill the buffer. '*eofflag' is set to 0 or 1 by the last
3682 * internal call to VOP_READDIR() that returned without error, indicating
3683 * whether it reached the end of the directory, and to 2 by this function after
3684 * all entries have been read.
3687 vn_dir_next_dirent(struct vnode *vp, struct thread *td,
3688 char *dirbuf, size_t dirbuflen,
3689 struct dirent **dpp, size_t *len, off_t *off, int *eofflag)
3691 struct dirent *dp = NULL;
3697 ASSERT_VOP_LOCKED(vp, "vnode not locked");
3698 VNASSERT(vp->v_type == VDIR, vp, ("vnode is not a directory"));
3699 MPASS2((uintptr_t)dirbuf < (uintptr_t)dirbuf + dirbuflen,
3700 "Address space overflow");
3702 if (__predict_false(dirbuflen < GENERIC_MAXDIRSIZ)) {
3703 /* Don't take any chances in this case */
3712 * The caller continued to call us after an error (we set dp to
3713 * NULL in a previous iteration). Bail out right now.
3715 if (__predict_false(dp == NULL))
3718 MPASS(*len <= dirbuflen);
3719 MPASS2((uintptr_t)dirbuf <= (uintptr_t)dp &&
3720 (uintptr_t)dp + *len <= (uintptr_t)dirbuf + dirbuflen,
3721 "Filled range not inside buffer");
3723 reclen = dp->d_reclen;
3724 if (reclen >= *len) {
3725 /* End of buffer reached */
3728 dp = (struct dirent *)((char *)dp + reclen);
3736 /* Have to refill. */
3742 /* Nothing more to read. */
3743 *eofflag = 2; /* Remember the caller reached EOF. */
3747 /* The caller didn't test for EOF. */
3752 iov.iov_base = dirbuf;
3753 iov.iov_len = dirbuflen;
3757 uio.uio_offset = *off;
3758 uio.uio_resid = dirbuflen;
3759 uio.uio_segflg = UIO_SYSSPACE;
3760 uio.uio_rw = UIO_READ;
3764 error = mac_vnode_check_readdir(td->td_ucred, vp);
3767 error = VOP_READDIR(vp, &uio, td->td_ucred, eofflag,
3772 *len = dirbuflen - uio.uio_resid;
3773 *off = uio.uio_offset;
3776 /* Sanity check on INVARIANTS. */
3777 MPASS(*eofflag != 0);
3783 * Normalize the flag returned by VOP_READDIR(), since we use 2
3784 * as a sentinel value.
3789 dp = (struct dirent *)dirbuf;
3792 if (__predict_false(*len < GENERIC_MINDIRSIZ ||
3793 dp->d_reclen < GENERIC_MINDIRSIZ)) {
3807 * Checks whether a directory is empty or not.
3809 * If the directory is empty, returns 0, and if it is not, ENOTEMPTY. Other
3810 * values are genuine errors preventing the check.
3813 vn_dir_check_empty(struct vnode *vp)
3815 struct thread *const td = curthread;
3817 size_t dirbuflen, len;
3823 ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
3824 VNPASS(vp->v_type == VDIR, vp);
3826 error = VOP_GETATTR(vp, &va, td->td_ucred);
3830 dirbuflen = max(DEV_BSIZE, GENERIC_MAXDIRSIZ);
3831 if (dirbuflen < va.va_blocksize)
3832 dirbuflen = va.va_blocksize;
3833 dirbuf = malloc(dirbuflen, M_TEMP, M_WAITOK);
3840 error = vn_dir_next_dirent(vp, td, dirbuf, dirbuflen,
3841 &dp, &len, &off, &eofflag);
3852 * Skip whiteouts. Unionfs operates on filesystems only and
3853 * not on hierarchies, so these whiteouts would be shadowed on
3854 * the system hierarchy but not for a union using the
3855 * filesystem of their directories as the upper layer.
3856 * Additionally, unionfs currently transparently exposes
3857 * union-specific metadata of its upper layer, meaning that
3858 * whiteouts can be seen through the union view in empty
3859 * directories. Taking into account these whiteouts would then
3860 * prevent mounting another filesystem on such effectively
3861 * empty directories.
3863 if (dp->d_type == DT_WHT)
3867 * Any file in the directory which is not '.' or '..' indicates
3868 * the directory is not empty.
3870 switch (dp->d_namlen) {
3872 if (dp->d_name[1] != '.') {
3873 /* Can't be '..' (nor '.') */
3879 if (dp->d_name[0] != '.') {
3880 /* Can't be '..' nor '.' */
3893 free(dirbuf, M_TEMP);
3898 static u_long vn_lock_pair_pause_cnt;
3899 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3900 &vn_lock_pair_pause_cnt, 0,
3901 "Count of vn_lock_pair deadlocks");
3903 u_int vn_lock_pair_pause_max;
3904 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3905 &vn_lock_pair_pause_max, 0,
3906 "Max ticks for vn_lock_pair deadlock avoidance sleep");
3909 vn_lock_pair_pause(const char *wmesg)
3911 atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3912 pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3916 * Lock pair of (possibly same) vnodes vp1, vp2, avoiding lock order
3917 * reversal. vp1_locked indicates whether vp1 is locked; if not, vp1
3918 * must be unlocked. Same for vp2 and vp2_locked. One of the vnodes
3921 * The function returns with both vnodes exclusively or shared locked,
3922 * according to corresponding lkflags, and guarantees that it does not
3923 * create lock order reversal with other threads during its execution.
3924 * Both vnodes could be unlocked temporary (and reclaimed).
3926 * If requesting shared locking, locked vnode lock must not be recursed.
3928 * Only one of LK_SHARED and LK_EXCLUSIVE must be specified.
3929 * LK_NODDLKTREAT can be optionally passed.
3931 * If vp1 == vp2, only one, most exclusive, lock is obtained on it.
3934 vn_lock_pair(struct vnode *vp1, bool vp1_locked, int lkflags1,
3935 struct vnode *vp2, bool vp2_locked, int lkflags2)
3939 MPASS(((lkflags1 & LK_SHARED) != 0) ^ ((lkflags1 & LK_EXCLUSIVE) != 0));
3940 MPASS((lkflags1 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0);
3941 MPASS(((lkflags2 & LK_SHARED) != 0) ^ ((lkflags2 & LK_EXCLUSIVE) != 0));
3942 MPASS((lkflags2 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0);
3944 if (vp1 == NULL && vp2 == NULL)
3948 MPASS(vp1_locked == vp2_locked);
3950 /* Select the most exclusive mode for lock. */
3951 if ((lkflags1 & LK_TYPE_MASK) != (lkflags2 & LK_TYPE_MASK))
3952 lkflags1 = (lkflags1 & ~LK_SHARED) | LK_EXCLUSIVE;
3955 ASSERT_VOP_LOCKED(vp1, "vp1");
3957 /* No need to relock if any lock is exclusive. */
3958 if ((vp1->v_vnlock->lock_object.lo_flags &
3962 locked1 = VOP_ISLOCKED(vp1);
3963 if (((lkflags1 & LK_SHARED) != 0 &&
3964 locked1 != LK_EXCLUSIVE) ||
3965 ((lkflags1 & LK_EXCLUSIVE) != 0 &&
3966 locked1 == LK_EXCLUSIVE))
3971 ASSERT_VOP_UNLOCKED(vp1, "vp1");
3972 vn_lock(vp1, lkflags1 | LK_RETRY);
3977 if ((lkflags1 & LK_SHARED) != 0 &&
3978 (vp1->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
3979 lkflags1 = (lkflags1 & ~LK_SHARED) | LK_EXCLUSIVE;
3980 if (vp1_locked && VOP_ISLOCKED(vp1) != LK_EXCLUSIVE) {
3981 ASSERT_VOP_LOCKED(vp1, "vp1");
3982 if ((lkflags1 & LK_EXCLUSIVE) != 0) {
3984 ASSERT_VOP_UNLOCKED(vp1,
3985 "vp1 shared recursed");
3988 } else if (!vp1_locked)
3989 ASSERT_VOP_UNLOCKED(vp1, "vp1");
3995 if ((lkflags2 & LK_SHARED) != 0 &&
3996 (vp2->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
3997 lkflags2 = (lkflags2 & ~LK_SHARED) | LK_EXCLUSIVE;
3998 if (vp2_locked && VOP_ISLOCKED(vp2) != LK_EXCLUSIVE) {
3999 ASSERT_VOP_LOCKED(vp2, "vp2");
4000 if ((lkflags2 & LK_EXCLUSIVE) != 0) {
4002 ASSERT_VOP_UNLOCKED(vp2,
4003 "vp2 shared recursed");
4006 } else if (!vp2_locked)
4007 ASSERT_VOP_UNLOCKED(vp2, "vp2");
4012 if (!vp1_locked && !vp2_locked) {
4013 vn_lock(vp1, lkflags1 | LK_RETRY);
4017 while (!vp1_locked || !vp2_locked) {
4018 if (vp1_locked && vp2 != NULL) {
4020 error = VOP_LOCK1(vp2, lkflags2 | LK_NOWAIT,
4021 __FILE__, __LINE__);
4026 vn_lock_pair_pause("vlp1");
4028 vn_lock(vp2, lkflags2 | LK_RETRY);
4031 if (vp2_locked && vp1 != NULL) {
4033 error = VOP_LOCK1(vp1, lkflags1 | LK_NOWAIT,
4034 __FILE__, __LINE__);
4039 vn_lock_pair_pause("vlp2");
4041 vn_lock(vp1, lkflags1 | LK_RETRY);
4046 if (lkflags1 == LK_EXCLUSIVE)
4047 ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
4049 ASSERT_VOP_LOCKED(vp1, "vp1 ret");
4052 if (lkflags2 == LK_EXCLUSIVE)
4053 ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
4055 ASSERT_VOP_LOCKED(vp2, "vp2 ret");
4060 vn_lktype_write(struct mount *mp, struct vnode *vp)
4062 if (MNT_SHARED_WRITES(mp) ||
4063 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
4065 return (LK_EXCLUSIVE);