2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1982, 1986, 1989, 1993
5 * The Regents of the University of California. All rights reserved.
6 * (c) UNIX System Laboratories, Inc.
7 * All or some portions of this file are derived from material licensed
8 * to the University of California by American Telephone and Telegraph
9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10 * the permission of UNIX System Laboratories, Inc.
12 * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13 * Copyright (c) 2013, 2014 The FreeBSD Foundation
15 * Portions of this software were developed by Konstantin Belousov
16 * under sponsorship from the FreeBSD Foundation.
18 * Redistribution and use in source and binary forms, with or without
19 * modification, are permitted provided that the following conditions
21 * 1. Redistributions of source code must retain the above copyright
22 * notice, this list of conditions and the following disclaimer.
23 * 2. Redistributions in binary form must reproduce the above copyright
24 * notice, this list of conditions and the following disclaimer in the
25 * documentation and/or other materials provided with the distribution.
26 * 3. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42 * @(#)vfs_vnops.c 8.2 (Berkeley) 1/21/94
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
48 #include "opt_hwpmc_hooks.h"
50 #include <sys/param.h>
51 #include <sys/systm.h>
54 #include <sys/fcntl.h>
61 #include <sys/limits.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
68 #include <sys/dirent.h>
71 #include <sys/filio.h>
72 #include <sys/resourcevar.h>
73 #include <sys/rwlock.h>
76 #include <sys/sleepqueue.h>
77 #include <sys/sysctl.h>
78 #include <sys/ttycom.h>
80 #include <sys/syslog.h>
81 #include <sys/unistd.h>
83 #include <sys/ktrace.h>
85 #include <security/audit/audit.h>
86 #include <security/mac/mac_framework.h>
89 #include <vm/vm_extern.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_page.h>
94 #include <vm/vm_pager.h>
97 #include <sys/pmckern.h>
100 static fo_rdwr_t vn_read;
101 static fo_rdwr_t vn_write;
102 static fo_rdwr_t vn_io_fault;
103 static fo_truncate_t vn_truncate;
104 static fo_ioctl_t vn_ioctl;
105 static fo_poll_t vn_poll;
106 static fo_kqfilter_t vn_kqfilter;
107 static fo_close_t vn_closefile;
108 static fo_mmap_t vn_mmap;
109 static fo_fallocate_t vn_fallocate;
111 struct fileops vnops = {
112 .fo_read = vn_io_fault,
113 .fo_write = vn_io_fault,
114 .fo_truncate = vn_truncate,
115 .fo_ioctl = vn_ioctl,
117 .fo_kqfilter = vn_kqfilter,
118 .fo_stat = vn_statfile,
119 .fo_close = vn_closefile,
120 .fo_chmod = vn_chmod,
121 .fo_chown = vn_chown,
122 .fo_sendfile = vn_sendfile,
124 .fo_fill_kinfo = vn_fill_kinfo,
126 .fo_fallocate = vn_fallocate,
127 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
130 const u_int io_hold_cnt = 16;
131 static int vn_io_fault_enable = 1;
132 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
133 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
134 static int vn_io_fault_prefault = 0;
135 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
136 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
137 static int vn_io_pgcache_read_enable = 1;
138 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
139 &vn_io_pgcache_read_enable, 0,
140 "Enable copying from page cache for reads, avoiding fs");
141 static u_long vn_io_faults_cnt;
142 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
143 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
145 static int vfs_allow_read_dir = 0;
146 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
147 &vfs_allow_read_dir, 0,
148 "Enable read(2) of directory by root for filesystems that support it");
151 * Returns true if vn_io_fault mode of handling the i/o request should
155 do_vn_io_fault(struct vnode *vp, struct uio *uio)
159 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
160 (mp = vp->v_mount) != NULL &&
161 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
165 * Structure used to pass arguments to vn_io_fault1(), to do either
166 * file- or vnode-based I/O calls.
168 struct vn_io_fault_args {
176 struct fop_args_tag {
180 struct vop_args_tag {
186 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
187 struct vn_io_fault_args *args, struct thread *td);
190 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
192 struct thread *td = ndp->ni_cnd.cn_thread;
194 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
198 open2nameif(int fmode, u_int vn_open_flags)
202 res = ISOPEN | LOCKLEAF;
203 if ((fmode & O_RESOLVE_BENEATH) != 0)
205 if ((fmode & O_EMPTY_PATH) != 0)
207 if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
209 if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
211 if ((vn_open_flags & VN_OPEN_WANTIOCTLCAPS) != 0)
212 res |= WANTIOCTLCAPS;
217 * Common code for vnode open operations via a name lookup.
218 * Lookup the vnode and invoke VOP_CREATE if needed.
219 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
221 * Note that this does NOT free nameidata for the successful case,
222 * due to the NDINIT being done elsewhere.
225 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
226 struct ucred *cred, struct file *fp)
230 struct thread *td = ndp->ni_cnd.cn_thread;
232 struct vattr *vap = &vat;
239 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
240 O_EXCL | O_DIRECTORY) ||
241 (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
243 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
244 ndp->ni_cnd.cn_nameiop = CREATE;
245 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
247 * Set NOCACHE to avoid flushing the cache when
248 * rolling in many files at once.
250 * Set NC_KEEPPOSENTRY to keep positive entries if they already
251 * exist despite NOCACHE.
253 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
254 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
255 ndp->ni_cnd.cn_flags |= FOLLOW;
256 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
258 if ((error = namei(ndp)) != 0)
260 if (ndp->ni_vp == NULL) {
263 vap->va_mode = cmode;
265 vap->va_vaflags |= VA_EXCLUSIVE;
266 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
267 NDFREE(ndp, NDF_ONLY_PNBUF);
269 if ((error = vn_start_write(NULL, &mp,
270 V_XSLEEP | PCATCH)) != 0)
275 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
276 ndp->ni_cnd.cn_flags |= MAKEENTRY;
278 error = mac_vnode_check_create(cred, ndp->ni_dvp,
282 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
285 if (error == 0 && (fmode & O_EXCL) != 0 &&
286 (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
288 vp->v_iflag |= VI_FOPENING;
292 VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
294 vn_finished_write(mp);
296 NDFREE(ndp, NDF_ONLY_PNBUF);
297 if (error == ERELOOKUP) {
305 if (ndp->ni_dvp == ndp->ni_vp)
311 if (fmode & O_EXCL) {
315 if (vp->v_type == VDIR) {
322 ndp->ni_cnd.cn_nameiop = LOOKUP;
323 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
324 ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
326 if ((fmode & FWRITE) == 0)
327 ndp->ni_cnd.cn_flags |= LOCKSHARED;
328 if ((error = namei(ndp)) != 0)
332 error = vn_open_vnode(vp, fmode, cred, td, fp);
335 vp->v_iflag &= ~VI_FOPENING;
344 NDFREE(ndp, NDF_ONLY_PNBUF);
352 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
355 int error, lock_flags, type;
357 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
358 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
360 KASSERT(fp != NULL, ("open with flock requires fp"));
361 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
364 lock_flags = VOP_ISLOCKED(vp);
367 lf.l_whence = SEEK_SET;
370 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
372 if ((fmode & FNONBLOCK) == 0)
374 if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
376 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
378 fp->f_flag |= FHASLOCK;
380 vn_lock(vp, lock_flags | LK_RETRY);
385 * Common code for vnode open operations once a vnode is located.
386 * Check permissions, and call the VOP_OPEN routine.
389 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
390 struct thread *td, struct file *fp)
395 if (vp->v_type == VLNK) {
396 if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
399 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
403 if ((fmode & O_PATH) == 0) {
404 if (vp->v_type == VSOCK)
406 if ((fmode & (FWRITE | O_TRUNC)) != 0) {
407 if (vp->v_type == VDIR)
411 if ((fmode & FREAD) != 0)
413 if ((fmode & O_APPEND) && (fmode & FWRITE))
416 if ((fmode & O_CREAT) != 0)
420 if ((fmode & FEXEC) != 0)
423 if ((fmode & O_VERIFY) != 0)
425 error = mac_vnode_check_open(cred, vp, accmode);
429 accmode &= ~(VCREAT | VVERIFY);
431 if ((fmode & O_CREAT) == 0 && accmode != 0) {
432 error = VOP_ACCESS(vp, accmode, cred, td);
436 if ((fmode & O_PATH) != 0) {
437 if (vp->v_type != VFIFO && vp->v_type != VSOCK &&
438 VOP_ACCESS(vp, VREAD, cred, td) == 0)
439 fp->f_flag |= FKQALLOWED;
443 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
444 vn_lock(vp, LK_UPGRADE | LK_RETRY);
445 error = VOP_OPEN(vp, fmode, cred, td, fp);
449 error = vn_open_vnode_advlock(vp, fmode, fp);
450 if (error == 0 && (fmode & FWRITE) != 0) {
451 error = VOP_ADD_WRITECOUNT(vp, 1);
453 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
454 __func__, vp, vp->v_writecount);
459 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
460 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
465 * Arrange the call by having fdrop() to use
466 * vn_closefile(). This is to satisfy
467 * filesystems like devfs or tmpfs, which
468 * override fo_close().
470 fp->f_flag |= FOPENFAILED;
472 if (fp->f_ops == &badfileops) {
473 fp->f_type = DTYPE_VNODE;
479 * If there is no fp, due to kernel-mode open,
480 * we can call VOP_CLOSE() now.
482 if ((vp->v_type == VFIFO || (fmode & FWRITE) != 0 ||
483 !MNT_EXTENDED_SHARED(vp->v_mount)) &&
484 VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
485 vn_lock(vp, LK_UPGRADE | LK_RETRY);
486 (void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
491 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
497 * Check for write permissions on the specified vnode.
498 * Prototype text segments cannot be written.
502 vn_writechk(struct vnode *vp)
505 ASSERT_VOP_LOCKED(vp, "vn_writechk");
507 * If there's shared text associated with
508 * the vnode, try to free it up once. If
509 * we fail, we can't allow writing.
521 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
522 struct thread *td, bool keep_ref)
525 int error, lock_flags;
527 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
528 MNT_EXTENDED_SHARED(vp->v_mount))
529 lock_flags = LK_SHARED;
531 lock_flags = LK_EXCLUSIVE;
533 vn_start_write(vp, &mp, V_WAIT);
534 vn_lock(vp, lock_flags | LK_RETRY);
535 AUDIT_ARG_VNODE1(vp);
536 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
537 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
538 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
539 __func__, vp, vp->v_writecount);
541 error = VOP_CLOSE(vp, flags, file_cred, td);
546 vn_finished_write(mp);
551 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
555 return (vn_close1(vp, flags, file_cred, td, false));
559 * Heuristic to detect sequential operation.
562 sequential_heuristic(struct uio *uio, struct file *fp)
566 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
569 if (fp->f_flag & FRDAHEAD)
570 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
573 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
574 * that the first I/O is normally considered to be slightly
575 * sequential. Seeking to offset 0 doesn't change sequentiality
576 * unless previous seeks have reduced f_seqcount to 0, in which
577 * case offset 0 is not special.
579 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
580 uio->uio_offset == fp->f_nextoff[rw]) {
582 * f_seqcount is in units of fixed-size blocks so that it
583 * depends mainly on the amount of sequential I/O and not
584 * much on the number of sequential I/O's. The fixed size
585 * of 16384 is hard-coded here since it is (not quite) just
586 * a magic size that works well here. This size is more
587 * closely related to the best I/O size for real disks than
588 * to any block size used by software.
590 if (uio->uio_resid >= IO_SEQMAX * 16384)
591 fp->f_seqcount[rw] = IO_SEQMAX;
593 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
594 if (fp->f_seqcount[rw] > IO_SEQMAX)
595 fp->f_seqcount[rw] = IO_SEQMAX;
597 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
600 /* Not sequential. Quickly draw-down sequentiality. */
601 if (fp->f_seqcount[rw] > 1)
602 fp->f_seqcount[rw] = 1;
604 fp->f_seqcount[rw] = 0;
609 * Package up an I/O request on a vnode into a uio and do it.
612 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
613 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
614 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
621 struct vn_io_fault_args args;
622 int error, lock_flags;
624 if (offset < 0 && vp->v_type != VCHR)
626 auio.uio_iov = &aiov;
628 aiov.iov_base = base;
630 auio.uio_resid = len;
631 auio.uio_offset = offset;
632 auio.uio_segflg = segflg;
637 if ((ioflg & IO_NODELOCKED) == 0) {
638 if ((ioflg & IO_RANGELOCKED) == 0) {
639 if (rw == UIO_READ) {
640 rl_cookie = vn_rangelock_rlock(vp, offset,
642 } else if ((ioflg & IO_APPEND) != 0) {
643 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
645 rl_cookie = vn_rangelock_wlock(vp, offset,
651 if (rw == UIO_WRITE) {
652 if (vp->v_type != VCHR &&
653 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
656 lock_flags = vn_lktype_write(mp, vp);
658 lock_flags = LK_SHARED;
659 vn_lock(vp, lock_flags | LK_RETRY);
663 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
665 if ((ioflg & IO_NOMACCHECK) == 0) {
667 error = mac_vnode_check_read(active_cred, file_cred,
670 error = mac_vnode_check_write(active_cred, file_cred,
675 if (file_cred != NULL)
679 if (do_vn_io_fault(vp, &auio)) {
680 args.kind = VN_IO_FAULT_VOP;
683 args.args.vop_args.vp = vp;
684 error = vn_io_fault1(vp, &auio, &args, td);
685 } else if (rw == UIO_READ) {
686 error = VOP_READ(vp, &auio, ioflg, cred);
687 } else /* if (rw == UIO_WRITE) */ {
688 error = VOP_WRITE(vp, &auio, ioflg, cred);
692 *aresid = auio.uio_resid;
694 if (auio.uio_resid && error == 0)
696 if ((ioflg & IO_NODELOCKED) == 0) {
699 vn_finished_write(mp);
702 if (rl_cookie != NULL)
703 vn_rangelock_unlock(vp, rl_cookie);
708 * Package up an I/O request on a vnode into a uio and do it. The I/O
709 * request is split up into smaller chunks and we try to avoid saturating
710 * the buffer cache while potentially holding a vnode locked, so we
711 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
712 * to give other processes a chance to lock the vnode (either other processes
713 * core'ing the same binary, or unrelated processes scanning the directory).
716 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
717 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
718 struct ucred *file_cred, size_t *aresid, struct thread *td)
727 * Force `offset' to a multiple of MAXBSIZE except possibly
728 * for the first chunk, so that filesystems only need to
729 * write full blocks except possibly for the first and last
732 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
736 if (rw != UIO_READ && vp->v_type == VREG)
739 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
740 ioflg, active_cred, file_cred, &iaresid, td);
741 len -= chunk; /* aresid calc already includes length */
745 base = (char *)base + chunk;
746 kern_yield(PRI_USER);
749 *aresid = len + iaresid;
753 #if OFF_MAX <= LONG_MAX
755 foffset_lock(struct file *fp, int flags)
757 volatile short *flagsp;
761 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
763 if ((flags & FOF_NOLOCK) != 0)
764 return (atomic_load_long(&fp->f_offset));
767 * According to McKusick the vn lock was protecting f_offset here.
768 * It is now protected by the FOFFSET_LOCKED flag.
770 flagsp = &fp->f_vnread_flags;
771 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
772 return (atomic_load_long(&fp->f_offset));
774 sleepq_lock(&fp->f_vnread_flags);
775 state = atomic_load_16(flagsp);
777 if ((state & FOFFSET_LOCKED) == 0) {
778 if (!atomic_fcmpset_acq_16(flagsp, &state,
783 if ((state & FOFFSET_LOCK_WAITING) == 0) {
784 if (!atomic_fcmpset_acq_16(flagsp, &state,
785 state | FOFFSET_LOCK_WAITING))
789 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
790 sleepq_wait(&fp->f_vnread_flags, PUSER -1);
792 sleepq_lock(&fp->f_vnread_flags);
793 state = atomic_load_16(flagsp);
795 res = atomic_load_long(&fp->f_offset);
796 sleepq_release(&fp->f_vnread_flags);
801 foffset_unlock(struct file *fp, off_t val, int flags)
803 volatile short *flagsp;
806 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
808 if ((flags & FOF_NOUPDATE) == 0)
809 atomic_store_long(&fp->f_offset, val);
810 if ((flags & FOF_NEXTOFF_R) != 0)
811 fp->f_nextoff[UIO_READ] = val;
812 if ((flags & FOF_NEXTOFF_W) != 0)
813 fp->f_nextoff[UIO_WRITE] = val;
815 if ((flags & FOF_NOLOCK) != 0)
818 flagsp = &fp->f_vnread_flags;
819 state = atomic_load_16(flagsp);
820 if ((state & FOFFSET_LOCK_WAITING) == 0 &&
821 atomic_cmpset_rel_16(flagsp, state, 0))
824 sleepq_lock(&fp->f_vnread_flags);
825 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
826 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
827 fp->f_vnread_flags = 0;
828 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
829 sleepq_release(&fp->f_vnread_flags);
833 foffset_lock(struct file *fp, int flags)
838 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
840 mtxp = mtx_pool_find(mtxpool_sleep, fp);
842 if ((flags & FOF_NOLOCK) == 0) {
843 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
844 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
845 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
848 fp->f_vnread_flags |= FOFFSET_LOCKED;
856 foffset_unlock(struct file *fp, off_t val, int flags)
860 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
862 mtxp = mtx_pool_find(mtxpool_sleep, fp);
864 if ((flags & FOF_NOUPDATE) == 0)
866 if ((flags & FOF_NEXTOFF_R) != 0)
867 fp->f_nextoff[UIO_READ] = val;
868 if ((flags & FOF_NEXTOFF_W) != 0)
869 fp->f_nextoff[UIO_WRITE] = val;
870 if ((flags & FOF_NOLOCK) == 0) {
871 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
872 ("Lost FOFFSET_LOCKED"));
873 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
874 wakeup(&fp->f_vnread_flags);
875 fp->f_vnread_flags = 0;
882 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
885 if ((flags & FOF_OFFSET) == 0)
886 uio->uio_offset = foffset_lock(fp, flags);
890 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
893 if ((flags & FOF_OFFSET) == 0)
894 foffset_unlock(fp, uio->uio_offset, flags);
898 get_advice(struct file *fp, struct uio *uio)
903 ret = POSIX_FADV_NORMAL;
904 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
907 mtxp = mtx_pool_find(mtxpool_sleep, fp);
909 if (fp->f_advice != NULL &&
910 uio->uio_offset >= fp->f_advice->fa_start &&
911 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
912 ret = fp->f_advice->fa_advice;
918 get_write_ioflag(struct file *fp)
926 mp = atomic_load_ptr(&vp->v_mount);
928 if ((fp->f_flag & O_DIRECT) != 0)
931 if ((fp->f_flag & O_FSYNC) != 0 ||
932 (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
936 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
937 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
938 * fall back to full O_SYNC behavior.
940 if ((fp->f_flag & O_DSYNC) != 0)
941 ioflag |= IO_SYNC | IO_DATASYNC;
947 vn_read_from_obj(struct vnode *vp, struct uio *uio)
950 vm_page_t ma[io_hold_cnt + 2];
955 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
956 obj = atomic_load_ptr(&vp->v_object);
958 return (EJUSTRETURN);
961 * Depends on type stability of vm_objects.
963 vm_object_pip_add(obj, 1);
964 if ((obj->flags & OBJ_DEAD) != 0) {
966 * Note that object might be already reused from the
967 * vnode, and the OBJ_DEAD flag cleared. This is fine,
968 * we recheck for DOOMED vnode state after all pages
969 * are busied, and retract then.
971 * But we check for OBJ_DEAD to ensure that we do not
972 * busy pages while vm_object_terminate_pages()
973 * processes the queue.
979 resid = uio->uio_resid;
980 off = uio->uio_offset;
981 for (i = 0; resid > 0; i++) {
982 MPASS(i < io_hold_cnt + 2);
983 ma[i] = vm_page_grab_unlocked(obj, atop(off),
984 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
990 * Skip invalid pages. Valid mask can be partial only
991 * at EOF, and we clip later.
993 if (vm_page_none_valid(ma[i])) {
994 vm_page_sunbusy(ma[i]);
1002 error = EJUSTRETURN;
1007 * Check VIRF_DOOMED after we busied our pages. Since
1008 * vgonel() terminates the vnode' vm_object, it cannot
1009 * process past pages busied by us.
1011 if (VN_IS_DOOMED(vp)) {
1012 error = EJUSTRETURN;
1016 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1017 if (resid > uio->uio_resid)
1018 resid = uio->uio_resid;
1021 * Unlocked read of vnp_size is safe because truncation cannot
1022 * pass busied page. But we load vnp_size into a local
1023 * variable so that possible concurrent extension does not
1024 * break calculation.
1026 #if defined(__powerpc__) && !defined(__powerpc64__)
1027 vsz = obj->un_pager.vnp.vnp_size;
1029 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1031 if (uio->uio_offset >= vsz) {
1032 error = EJUSTRETURN;
1035 if (uio->uio_offset + resid > vsz)
1036 resid = vsz - uio->uio_offset;
1038 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1041 for (j = 0; j < i; j++) {
1043 vm_page_reference(ma[j]);
1044 vm_page_sunbusy(ma[j]);
1047 vm_object_pip_wakeup(obj);
1050 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1054 * File table vnode read routine.
1057 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1065 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1067 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1070 if (fp->f_flag & FNONBLOCK)
1071 ioflag |= IO_NDELAY;
1072 if (fp->f_flag & O_DIRECT)
1073 ioflag |= IO_DIRECT;
1076 * Try to read from page cache. VIRF_DOOMED check is racy but
1077 * allows us to avoid unneeded work outright.
1079 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1080 (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1081 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1083 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1086 if (error != EJUSTRETURN)
1090 advice = get_advice(fp, uio);
1091 vn_lock(vp, LK_SHARED | LK_RETRY);
1094 case POSIX_FADV_NORMAL:
1095 case POSIX_FADV_SEQUENTIAL:
1096 case POSIX_FADV_NOREUSE:
1097 ioflag |= sequential_heuristic(uio, fp);
1099 case POSIX_FADV_RANDOM:
1100 /* Disable read-ahead for random I/O. */
1103 orig_offset = uio->uio_offset;
1106 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1109 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1110 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1112 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1113 orig_offset != uio->uio_offset)
1115 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1116 * for the backing file after a POSIX_FADV_NOREUSE
1119 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1120 POSIX_FADV_DONTNEED);
1125 * File table vnode write routine.
1128 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1136 bool need_finished_write;
1138 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1140 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1142 if (vp->v_type == VREG)
1145 if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1146 ioflag |= IO_APPEND;
1147 if ((fp->f_flag & FNONBLOCK) != 0)
1148 ioflag |= IO_NDELAY;
1149 ioflag |= get_write_ioflag(fp);
1152 need_finished_write = false;
1153 if (vp->v_type != VCHR) {
1154 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1157 need_finished_write = true;
1160 advice = get_advice(fp, uio);
1162 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1164 case POSIX_FADV_NORMAL:
1165 case POSIX_FADV_SEQUENTIAL:
1166 case POSIX_FADV_NOREUSE:
1167 ioflag |= sequential_heuristic(uio, fp);
1169 case POSIX_FADV_RANDOM:
1170 /* XXX: Is this correct? */
1173 orig_offset = uio->uio_offset;
1176 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1179 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1180 fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1182 if (need_finished_write)
1183 vn_finished_write(mp);
1184 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1185 orig_offset != uio->uio_offset)
1187 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1188 * for the backing file after a POSIX_FADV_NOREUSE
1191 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1192 POSIX_FADV_DONTNEED);
1198 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1199 * prevent the following deadlock:
1201 * Assume that the thread A reads from the vnode vp1 into userspace
1202 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
1203 * currently not resident, then system ends up with the call chain
1204 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1205 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1206 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1207 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1208 * backed by the pages of vnode vp1, and some page in buf2 is not
1209 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1211 * To prevent the lock order reversal and deadlock, vn_io_fault() does
1212 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1213 * Instead, it first tries to do the whole range i/o with pagefaults
1214 * disabled. If all pages in the i/o buffer are resident and mapped,
1215 * VOP will succeed (ignoring the genuine filesystem errors).
1216 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1217 * i/o in chunks, with all pages in the chunk prefaulted and held
1218 * using vm_fault_quick_hold_pages().
1220 * Filesystems using this deadlock avoidance scheme should use the
1221 * array of the held pages from uio, saved in the curthread->td_ma,
1222 * instead of doing uiomove(). A helper function
1223 * vn_io_fault_uiomove() converts uiomove request into
1224 * uiomove_fromphys() over td_ma array.
1226 * Since vnode locks do not cover the whole i/o anymore, rangelocks
1227 * make the current i/o request atomic with respect to other i/os and
1232 * Decode vn_io_fault_args and perform the corresponding i/o.
1235 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1241 save = vm_fault_disable_pagefaults();
1242 switch (args->kind) {
1243 case VN_IO_FAULT_FOP:
1244 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1245 uio, args->cred, args->flags, td);
1247 case VN_IO_FAULT_VOP:
1248 if (uio->uio_rw == UIO_READ) {
1249 error = VOP_READ(args->args.vop_args.vp, uio,
1250 args->flags, args->cred);
1251 } else if (uio->uio_rw == UIO_WRITE) {
1252 error = VOP_WRITE(args->args.vop_args.vp, uio,
1253 args->flags, args->cred);
1257 panic("vn_io_fault_doio: unknown kind of io %d %d",
1258 args->kind, uio->uio_rw);
1260 vm_fault_enable_pagefaults(save);
1265 vn_io_fault_touch(char *base, const struct uio *uio)
1270 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1276 vn_io_fault_prefault_user(const struct uio *uio)
1279 const struct iovec *iov;
1284 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1285 ("vn_io_fault_prefault userspace"));
1289 resid = uio->uio_resid;
1290 base = iov->iov_base;
1293 error = vn_io_fault_touch(base, uio);
1296 if (len < PAGE_SIZE) {
1298 error = vn_io_fault_touch(base + len - 1, uio);
1303 if (++i >= uio->uio_iovcnt)
1305 iov = uio->uio_iov + i;
1306 base = iov->iov_base;
1318 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1319 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1320 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1321 * into args and call vn_io_fault1() to handle faults during the user
1322 * mode buffer accesses.
1325 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1328 vm_page_t ma[io_hold_cnt + 2];
1329 struct uio *uio_clone, short_uio;
1330 struct iovec short_iovec[1];
1331 vm_page_t *prev_td_ma;
1333 vm_offset_t addr, end;
1336 int error, cnt, saveheld, prev_td_ma_cnt;
1338 if (vn_io_fault_prefault) {
1339 error = vn_io_fault_prefault_user(uio);
1341 return (error); /* Or ignore ? */
1344 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1347 * The UFS follows IO_UNIT directive and replays back both
1348 * uio_offset and uio_resid if an error is encountered during the
1349 * operation. But, since the iovec may be already advanced,
1350 * uio is still in an inconsistent state.
1352 * Cache a copy of the original uio, which is advanced to the redo
1353 * point using UIO_NOCOPY below.
1355 uio_clone = cloneuio(uio);
1356 resid = uio->uio_resid;
1358 short_uio.uio_segflg = UIO_USERSPACE;
1359 short_uio.uio_rw = uio->uio_rw;
1360 short_uio.uio_td = uio->uio_td;
1362 error = vn_io_fault_doio(args, uio, td);
1363 if (error != EFAULT)
1366 atomic_add_long(&vn_io_faults_cnt, 1);
1367 uio_clone->uio_segflg = UIO_NOCOPY;
1368 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1369 uio_clone->uio_segflg = uio->uio_segflg;
1371 saveheld = curthread_pflags_set(TDP_UIOHELD);
1372 prev_td_ma = td->td_ma;
1373 prev_td_ma_cnt = td->td_ma_cnt;
1375 while (uio_clone->uio_resid != 0) {
1376 len = uio_clone->uio_iov->iov_len;
1378 KASSERT(uio_clone->uio_iovcnt >= 1,
1379 ("iovcnt underflow"));
1380 uio_clone->uio_iov++;
1381 uio_clone->uio_iovcnt--;
1384 if (len > ptoa(io_hold_cnt))
1385 len = ptoa(io_hold_cnt);
1386 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1387 end = round_page(addr + len);
1392 cnt = atop(end - trunc_page(addr));
1394 * A perfectly misaligned address and length could cause
1395 * both the start and the end of the chunk to use partial
1396 * page. +2 accounts for such a situation.
1398 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1399 addr, len, prot, ma, io_hold_cnt + 2);
1404 short_uio.uio_iov = &short_iovec[0];
1405 short_iovec[0].iov_base = (void *)addr;
1406 short_uio.uio_iovcnt = 1;
1407 short_uio.uio_resid = short_iovec[0].iov_len = len;
1408 short_uio.uio_offset = uio_clone->uio_offset;
1410 td->td_ma_cnt = cnt;
1412 error = vn_io_fault_doio(args, &short_uio, td);
1413 vm_page_unhold_pages(ma, cnt);
1414 adv = len - short_uio.uio_resid;
1416 uio_clone->uio_iov->iov_base =
1417 (char *)uio_clone->uio_iov->iov_base + adv;
1418 uio_clone->uio_iov->iov_len -= adv;
1419 uio_clone->uio_resid -= adv;
1420 uio_clone->uio_offset += adv;
1422 uio->uio_resid -= adv;
1423 uio->uio_offset += adv;
1425 if (error != 0 || adv == 0)
1428 td->td_ma = prev_td_ma;
1429 td->td_ma_cnt = prev_td_ma_cnt;
1430 curthread_pflags_restore(saveheld);
1432 free(uio_clone, M_IOV);
1437 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1438 int flags, struct thread *td)
1443 struct vn_io_fault_args args;
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 foffset_lock_uio(fp, uio, flags);
1467 if (do_vn_io_fault(vp, uio)) {
1468 args.kind = VN_IO_FAULT_FOP;
1469 args.args.fop_args.fp = fp;
1470 args.args.fop_args.doio = doio;
1471 args.cred = active_cred;
1472 args.flags = flags | FOF_OFFSET;
1473 if (uio->uio_rw == UIO_READ) {
1474 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1475 uio->uio_offset + uio->uio_resid);
1476 } else if ((fp->f_flag & O_APPEND) != 0 ||
1477 (flags & FOF_OFFSET) == 0) {
1478 /* For appenders, punt and lock the whole range. */
1479 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1481 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1482 uio->uio_offset + uio->uio_resid);
1484 error = vn_io_fault1(vp, uio, &args, td);
1485 vn_rangelock_unlock(vp, rl_cookie);
1487 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1489 foffset_unlock_uio(fp, uio, flags);
1494 * Helper function to perform the requested uiomove operation using
1495 * the held pages for io->uio_iov[0].iov_base buffer instead of
1496 * copyin/copyout. Access to the pages with uiomove_fromphys()
1497 * instead of iov_base prevents page faults that could occur due to
1498 * pmap_collect() invalidating the mapping created by
1499 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1500 * object cleanup revoking the write access from page mappings.
1502 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1503 * instead of plain uiomove().
1506 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1508 struct uio transp_uio;
1509 struct iovec transp_iov[1];
1515 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1516 uio->uio_segflg != UIO_USERSPACE)
1517 return (uiomove(data, xfersize, uio));
1519 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1520 transp_iov[0].iov_base = data;
1521 transp_uio.uio_iov = &transp_iov[0];
1522 transp_uio.uio_iovcnt = 1;
1523 if (xfersize > uio->uio_resid)
1524 xfersize = uio->uio_resid;
1525 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1526 transp_uio.uio_offset = 0;
1527 transp_uio.uio_segflg = UIO_SYSSPACE;
1529 * Since transp_iov points to data, and td_ma page array
1530 * corresponds to original uio->uio_iov, we need to invert the
1531 * direction of the i/o operation as passed to
1532 * uiomove_fromphys().
1534 switch (uio->uio_rw) {
1536 transp_uio.uio_rw = UIO_READ;
1539 transp_uio.uio_rw = UIO_WRITE;
1542 transp_uio.uio_td = uio->uio_td;
1543 error = uiomove_fromphys(td->td_ma,
1544 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1545 xfersize, &transp_uio);
1546 adv = xfersize - transp_uio.uio_resid;
1548 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1549 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1551 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1553 td->td_ma_cnt -= pgadv;
1554 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1555 uio->uio_iov->iov_len -= adv;
1556 uio->uio_resid -= adv;
1557 uio->uio_offset += adv;
1562 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1566 vm_offset_t iov_base;
1570 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1571 uio->uio_segflg != UIO_USERSPACE)
1572 return (uiomove_fromphys(ma, offset, xfersize, uio));
1574 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1575 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1576 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1577 switch (uio->uio_rw) {
1579 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1583 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1587 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1589 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1591 td->td_ma_cnt -= pgadv;
1592 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1593 uio->uio_iov->iov_len -= cnt;
1594 uio->uio_resid -= cnt;
1595 uio->uio_offset += cnt;
1600 * File table truncate routine.
1603 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1615 * Lock the whole range for truncation. Otherwise split i/o
1616 * might happen partly before and partly after the truncation.
1618 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1619 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1622 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1623 AUDIT_ARG_VNODE1(vp);
1624 if (vp->v_type == VDIR) {
1629 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1633 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1637 vn_finished_write(mp);
1639 vn_rangelock_unlock(vp, rl_cookie);
1640 if (error == ERELOOKUP)
1646 * Truncate a file that is already locked.
1649 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1655 error = VOP_ADD_WRITECOUNT(vp, 1);
1658 vattr.va_size = length;
1660 vattr.va_vaflags |= VA_SYNC;
1661 error = VOP_SETATTR(vp, &vattr, cred);
1662 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1668 * File table vnode stat routine.
1671 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1674 struct vnode *vp = fp->f_vnode;
1677 vn_lock(vp, LK_SHARED | LK_RETRY);
1678 error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1685 * File table vnode ioctl routine.
1688 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1693 struct fiobmap2_arg *bmarg;
1697 switch (vp->v_type) {
1702 vn_lock(vp, LK_SHARED | LK_RETRY);
1703 error = VOP_GETATTR(vp, &vattr, active_cred);
1706 *(int *)data = vattr.va_size - fp->f_offset;
1709 bmarg = (struct fiobmap2_arg *)data;
1710 vn_lock(vp, LK_SHARED | LK_RETRY);
1712 error = mac_vnode_check_read(active_cred, fp->f_cred,
1716 error = VOP_BMAP(vp, bmarg->bn, NULL,
1717 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1724 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1729 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1737 * File table vnode poll routine.
1740 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1747 #if defined(MAC) || defined(AUDIT)
1748 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1749 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1750 AUDIT_ARG_VNODE1(vp);
1751 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1757 error = VOP_POLL(vp, events, fp->f_cred, td);
1762 * Acquire the requested lock and then check for validity. LK_RETRY
1763 * permits vn_lock to return doomed vnodes.
1765 static int __noinline
1766 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1770 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1771 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1774 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1776 if ((flags & LK_RETRY) == 0) {
1787 * Nothing to do if we got the lock.
1793 * Interlock was dropped by the call in _vn_lock.
1795 flags &= ~LK_INTERLOCK;
1797 error = VOP_LOCK1(vp, flags, file, line);
1798 } while (error != 0);
1803 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1807 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1808 ("vn_lock: no locktype (%d passed)", flags));
1809 VNPASS(vp->v_holdcnt > 0, vp);
1810 error = VOP_LOCK1(vp, flags, file, line);
1811 if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1812 return (_vn_lock_fallback(vp, flags, file, line, error));
1817 * File table vnode close routine.
1820 vn_closefile(struct file *fp, struct thread *td)
1828 fp->f_ops = &badfileops;
1829 ref = (fp->f_flag & FHASLOCK) != 0;
1831 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1833 if (__predict_false(ref)) {
1834 lf.l_whence = SEEK_SET;
1837 lf.l_type = F_UNLCK;
1838 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1845 * Preparing to start a filesystem write operation. If the operation is
1846 * permitted, then we bump the count of operations in progress and
1847 * proceed. If a suspend request is in progress, we wait until the
1848 * suspension is over, and then proceed.
1851 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1853 struct mount_pcpu *mpcpu;
1856 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1857 vfs_op_thread_enter(mp, mpcpu)) {
1858 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1859 vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1860 vfs_op_thread_exit(mp, mpcpu);
1865 mtx_assert(MNT_MTX(mp), MA_OWNED);
1872 * Check on status of suspension.
1874 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1875 mp->mnt_susp_owner != curthread) {
1876 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1877 (flags & PCATCH) : 0) | (PUSER - 1);
1878 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1879 if (flags & V_NOWAIT) {
1880 error = EWOULDBLOCK;
1883 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1889 if (flags & V_XSLEEP)
1891 mp->mnt_writeopcount++;
1893 if (error != 0 || (flags & V_XSLEEP) != 0)
1900 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1905 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1906 ("V_MNTREF requires mp"));
1910 * If a vnode is provided, get and return the mount point that
1911 * to which it will write.
1914 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1916 if (error != EOPNOTSUPP)
1921 if ((mp = *mpp) == NULL)
1925 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1927 * As long as a vnode is not provided we need to acquire a
1928 * refcount for the provided mountpoint too, in order to
1929 * emulate a vfs_ref().
1931 if (vp == NULL && (flags & V_MNTREF) == 0)
1934 return (vn_start_write_refed(mp, flags, false));
1938 * Secondary suspension. Used by operations such as vop_inactive
1939 * routines that are needed by the higher level functions. These
1940 * are allowed to proceed until all the higher level functions have
1941 * completed (indicated by mnt_writeopcount dropping to zero). At that
1942 * time, these operations are halted until the suspension is over.
1945 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1950 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1951 ("V_MNTREF requires mp"));
1955 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1957 if (error != EOPNOTSUPP)
1963 * If we are not suspended or have not yet reached suspended
1964 * mode, then let the operation proceed.
1966 if ((mp = *mpp) == NULL)
1970 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1972 * As long as a vnode is not provided we need to acquire a
1973 * refcount for the provided mountpoint too, in order to
1974 * emulate a vfs_ref().
1977 if (vp == NULL && (flags & V_MNTREF) == 0)
1979 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1980 mp->mnt_secondary_writes++;
1981 mp->mnt_secondary_accwrites++;
1985 if (flags & V_NOWAIT) {
1988 return (EWOULDBLOCK);
1991 * Wait for the suspension to finish.
1993 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1994 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
2003 * Filesystem write operation has completed. If we are suspending and this
2004 * operation is the last one, notify the suspender that the suspension is
2008 vn_finished_write(struct mount *mp)
2010 struct mount_pcpu *mpcpu;
2016 if (vfs_op_thread_enter(mp, mpcpu)) {
2017 vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2018 vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2019 vfs_op_thread_exit(mp, mpcpu);
2024 vfs_assert_mount_counters(mp);
2026 c = --mp->mnt_writeopcount;
2027 if (mp->mnt_vfs_ops == 0) {
2028 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2033 vfs_dump_mount_counters(mp);
2034 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2035 wakeup(&mp->mnt_writeopcount);
2040 * Filesystem secondary write operation has completed. If we are
2041 * suspending and this operation is the last one, notify the suspender
2042 * that the suspension is now in effect.
2045 vn_finished_secondary_write(struct mount *mp)
2051 mp->mnt_secondary_writes--;
2052 if (mp->mnt_secondary_writes < 0)
2053 panic("vn_finished_secondary_write: neg cnt");
2054 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2055 mp->mnt_secondary_writes <= 0)
2056 wakeup(&mp->mnt_secondary_writes);
2061 * Request a filesystem to suspend write operations.
2064 vfs_write_suspend(struct mount *mp, int flags)
2071 vfs_assert_mount_counters(mp);
2072 if (mp->mnt_susp_owner == curthread) {
2073 vfs_op_exit_locked(mp);
2077 while (mp->mnt_kern_flag & MNTK_SUSPEND)
2078 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2081 * Unmount holds a write reference on the mount point. If we
2082 * own busy reference and drain for writers, we deadlock with
2083 * the reference draining in the unmount path. Callers of
2084 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2085 * vfs_busy() reference is owned and caller is not in the
2088 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2089 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2090 vfs_op_exit_locked(mp);
2095 mp->mnt_kern_flag |= MNTK_SUSPEND;
2096 mp->mnt_susp_owner = curthread;
2097 if (mp->mnt_writeopcount > 0)
2098 (void) msleep(&mp->mnt_writeopcount,
2099 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2102 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2103 vfs_write_resume(mp, 0);
2104 /* vfs_write_resume does vfs_op_exit() for us */
2110 * Request a filesystem to resume write operations.
2113 vfs_write_resume(struct mount *mp, int flags)
2117 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2118 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2119 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2121 mp->mnt_susp_owner = NULL;
2122 wakeup(&mp->mnt_writeopcount);
2123 wakeup(&mp->mnt_flag);
2124 curthread->td_pflags &= ~TDP_IGNSUSP;
2125 if ((flags & VR_START_WRITE) != 0) {
2127 mp->mnt_writeopcount++;
2130 if ((flags & VR_NO_SUSPCLR) == 0)
2133 } else if ((flags & VR_START_WRITE) != 0) {
2135 vn_start_write_refed(mp, 0, true);
2142 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2146 vfs_write_suspend_umnt(struct mount *mp)
2150 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2151 ("vfs_write_suspend_umnt: recursed"));
2153 /* dounmount() already called vn_start_write(). */
2155 vn_finished_write(mp);
2156 error = vfs_write_suspend(mp, 0);
2158 vn_start_write(NULL, &mp, V_WAIT);
2162 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2165 vn_start_write(NULL, &mp, V_WAIT);
2167 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2168 wakeup(&mp->mnt_flag);
2170 curthread->td_pflags |= TDP_IGNSUSP;
2175 * Implement kqueues for files by translating it to vnode operation.
2178 vn_kqfilter(struct file *fp, struct knote *kn)
2181 return (VOP_KQFILTER(fp->f_vnode, kn));
2185 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2187 if ((fp->f_flag & FKQALLOWED) == 0)
2189 return (vn_kqfilter(fp, kn));
2193 * Simplified in-kernel wrapper calls for extended attribute access.
2194 * Both calls pass in a NULL credential, authorizing as "kernel" access.
2195 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2198 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2199 const char *attrname, int *buflen, char *buf, struct thread *td)
2205 iov.iov_len = *buflen;
2208 auio.uio_iov = &iov;
2209 auio.uio_iovcnt = 1;
2210 auio.uio_rw = UIO_READ;
2211 auio.uio_segflg = UIO_SYSSPACE;
2213 auio.uio_offset = 0;
2214 auio.uio_resid = *buflen;
2216 if ((ioflg & IO_NODELOCKED) == 0)
2217 vn_lock(vp, LK_SHARED | LK_RETRY);
2219 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2221 /* authorize attribute retrieval as kernel */
2222 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2225 if ((ioflg & IO_NODELOCKED) == 0)
2229 *buflen = *buflen - auio.uio_resid;
2236 * XXX failure mode if partially written?
2239 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2240 const char *attrname, int buflen, char *buf, struct thread *td)
2247 iov.iov_len = buflen;
2250 auio.uio_iov = &iov;
2251 auio.uio_iovcnt = 1;
2252 auio.uio_rw = UIO_WRITE;
2253 auio.uio_segflg = UIO_SYSSPACE;
2255 auio.uio_offset = 0;
2256 auio.uio_resid = buflen;
2258 if ((ioflg & IO_NODELOCKED) == 0) {
2259 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2261 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2264 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2266 /* authorize attribute setting as kernel */
2267 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2269 if ((ioflg & IO_NODELOCKED) == 0) {
2270 vn_finished_write(mp);
2278 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2279 const char *attrname, struct thread *td)
2284 if ((ioflg & IO_NODELOCKED) == 0) {
2285 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2287 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2290 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2292 /* authorize attribute removal as kernel */
2293 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2294 if (error == EOPNOTSUPP)
2295 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2298 if ((ioflg & IO_NODELOCKED) == 0) {
2299 vn_finished_write(mp);
2307 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2311 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2315 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2318 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2323 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2324 int lkflags, struct vnode **rvp)
2329 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2331 ltype = VOP_ISLOCKED(vp);
2332 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2333 ("vn_vget_ino: vp not locked"));
2334 error = vfs_busy(mp, MBF_NOWAIT);
2338 error = vfs_busy(mp, 0);
2339 vn_lock(vp, ltype | LK_RETRY);
2343 if (VN_IS_DOOMED(vp)) {
2349 error = alloc(mp, alloc_arg, lkflags, rvp);
2351 if (error != 0 || *rvp != vp)
2352 vn_lock(vp, ltype | LK_RETRY);
2353 if (VN_IS_DOOMED(vp)) {
2366 vn_send_sigxfsz(struct proc *p)
2369 kern_psignal(p, SIGXFSZ);
2374 vn_rlimit_trunc(u_quad_t size, struct thread *td)
2376 if (size <= lim_cur(td, RLIMIT_FSIZE))
2378 vn_send_sigxfsz(td->td_proc);
2383 vn_rlimit_fsizex1(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2384 bool adj, struct thread *td)
2389 if (vp->v_type != VREG)
2393 * Handle file system maximum file size.
2395 if (maxfsz != 0 && uio->uio_offset + uio->uio_resid > maxfsz) {
2396 if (!adj || uio->uio_offset >= maxfsz)
2398 uio->uio_resid = maxfsz - uio->uio_offset;
2402 * This is kernel write (e.g. vnode_pager) or accounting
2403 * write, ignore limit.
2405 if (td == NULL || (td->td_pflags2 & TDP2_ACCT) != 0)
2409 * Calculate file size limit.
2411 ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2412 lim = __predict_false(ktr_write) ? td->td_ktr_io_lim :
2413 lim_cur(td, RLIMIT_FSIZE);
2416 * Is the limit reached?
2418 if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2422 * Prepared filesystems can handle writes truncated to the
2425 if (adj && (uoff_t)uio->uio_offset < lim) {
2426 uio->uio_resid = lim - (uoff_t)uio->uio_offset;
2430 if (!ktr_write || ktr_filesize_limit_signal)
2431 vn_send_sigxfsz(td->td_proc);
2436 * Helper for VOP_WRITE() implementations, the common code to
2437 * handle maximum supported file size on the filesystem, and
2438 * RLIMIT_FSIZE, except for special writes from accounting subsystem
2441 * For maximum file size (maxfsz argument):
2442 * - return EFBIG if uio_offset is beyond it
2443 * - otherwise, clamp uio_resid if write would extend file beyond maxfsz.
2446 * - return EFBIG and send SIGXFSZ if uio_offset is beyond the limit
2447 * - otherwise, clamp uio_resid if write would extend file beyond limit.
2449 * If clamping occured, the adjustment for uio_resid is stored in
2450 * *resid_adj, to be re-applied by vn_rlimit_fsizex_res() on return
2454 vn_rlimit_fsizex(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2455 ssize_t *resid_adj, struct thread *td)
2461 resid_orig = uio->uio_resid;
2462 adj = resid_adj != NULL;
2463 error = vn_rlimit_fsizex1(vp, uio, maxfsz, adj, td);
2465 *resid_adj = resid_orig - uio->uio_resid;
2470 vn_rlimit_fsizex_res(struct uio *uio, ssize_t resid_adj)
2472 uio->uio_resid += resid_adj;
2476 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2479 return (vn_rlimit_fsizex(vp, __DECONST(struct uio *, uio), 0, NULL,
2484 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2491 vn_lock(vp, LK_SHARED | LK_RETRY);
2492 AUDIT_ARG_VNODE1(vp);
2495 return (setfmode(td, active_cred, vp, mode));
2499 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2506 vn_lock(vp, LK_SHARED | LK_RETRY);
2507 AUDIT_ARG_VNODE1(vp);
2510 return (setfown(td, active_cred, vp, uid, gid));
2514 * Remove pages in the range ["start", "end") from the vnode's VM object. If
2515 * "end" is 0, then the range extends to the end of the object.
2518 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2522 if ((object = vp->v_object) == NULL)
2524 VM_OBJECT_WLOCK(object);
2525 vm_object_page_remove(object, start, end, 0);
2526 VM_OBJECT_WUNLOCK(object);
2530 * Like vn_pages_remove(), but skips invalid pages, which by definition are not
2531 * mapped into any process' address space. Filesystems may use this in
2532 * preference to vn_pages_remove() to avoid blocking on pages busied in
2533 * preparation for a VOP_GETPAGES.
2536 vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2540 if ((object = vp->v_object) == NULL)
2542 VM_OBJECT_WLOCK(object);
2543 vm_object_page_remove(object, start, end, OBJPR_VALIDONLY);
2544 VM_OBJECT_WUNLOCK(object);
2548 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2557 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2558 ("Wrong command %lu", cmd));
2560 if (vn_lock(vp, LK_EXCLUSIVE) != 0)
2562 if (vp->v_type != VREG) {
2566 error = VOP_GETATTR(vp, &va, cred);
2570 if (noff < 0 || noff >= va.va_size) {
2575 /* See the comment in ufs_bmap_seekdata(). */
2578 VM_OBJECT_WLOCK(obj);
2579 vm_object_page_clean(obj, 0, 0, OBJPC_SYNC);
2580 VM_OBJECT_WUNLOCK(obj);
2583 bsize = vp->v_mount->mnt_stat.f_iosize;
2584 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2586 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2587 if (error == EOPNOTSUPP) {
2591 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2592 (bnp != -1 && cmd == FIOSEEKDATA)) {
2599 if (noff > va.va_size)
2601 /* noff == va.va_size. There is an implicit hole at the end of file. */
2602 if (cmd == FIOSEEKDATA)
2612 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2617 off_t foffset, size;
2620 cred = td->td_ucred;
2622 foffset = foffset_lock(fp, 0);
2623 noneg = (vp->v_type != VCHR);
2629 (offset > 0 && foffset > OFF_MAX - offset))) {
2636 vn_lock(vp, LK_SHARED | LK_RETRY);
2637 error = VOP_GETATTR(vp, &vattr, cred);
2643 * If the file references a disk device, then fetch
2644 * the media size and use that to determine the ending
2647 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2648 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2649 vattr.va_size = size;
2651 (vattr.va_size > OFF_MAX ||
2652 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2656 offset += vattr.va_size;
2661 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2662 if (error == ENOTTY)
2666 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2667 if (error == ENOTTY)
2673 if (error == 0 && noneg && offset < 0)
2677 VFS_KNOTE_UNLOCKED(vp, 0);
2678 td->td_uretoff.tdu_off = offset;
2680 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2685 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2691 * Grant permission if the caller is the owner of the file, or
2692 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2693 * on the file. If the time pointer is null, then write
2694 * permission on the file is also sufficient.
2696 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2697 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2698 * will be allowed to set the times [..] to the current
2701 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2702 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2703 error = VOP_ACCESS(vp, VWRITE, cred, td);
2708 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2713 if (fp->f_type == DTYPE_FIFO)
2714 kif->kf_type = KF_TYPE_FIFO;
2716 kif->kf_type = KF_TYPE_VNODE;
2719 FILEDESC_SUNLOCK(fdp);
2720 error = vn_fill_kinfo_vnode(vp, kif);
2722 FILEDESC_SLOCK(fdp);
2727 vn_fill_junk(struct kinfo_file *kif)
2732 * Simulate vn_fullpath returning changing values for a given
2733 * vp during e.g. coredump.
2735 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2736 olen = strlen(kif->kf_path);
2738 strcpy(&kif->kf_path[len - 1], "$");
2740 for (; olen < len; olen++)
2741 strcpy(&kif->kf_path[olen], "A");
2745 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2748 char *fullpath, *freepath;
2751 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2754 error = vn_fullpath(vp, &fullpath, &freepath);
2756 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2758 if (freepath != NULL)
2759 free(freepath, M_TEMP);
2761 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2766 * Retrieve vnode attributes.
2768 va.va_fsid = VNOVAL;
2770 vn_lock(vp, LK_SHARED | LK_RETRY);
2771 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2775 if (va.va_fsid != VNOVAL)
2776 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2778 kif->kf_un.kf_file.kf_file_fsid =
2779 vp->v_mount->mnt_stat.f_fsid.val[0];
2780 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2781 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2782 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2783 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2784 kif->kf_un.kf_file.kf_file_size = va.va_size;
2785 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2786 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2787 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2788 kif->kf_un.kf_file.kf_file_nlink = va.va_nlink;
2793 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2794 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2798 struct pmckern_map_in pkm;
2804 boolean_t writecounted;
2807 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2808 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2810 * POSIX shared-memory objects are defined to have
2811 * kernel persistence, and are not defined to support
2812 * read(2)/write(2) -- or even open(2). Thus, we can
2813 * use MAP_ASYNC to trade on-disk coherence for speed.
2814 * The shm_open(3) library routine turns on the FPOSIXSHM
2815 * flag to request this behavior.
2817 if ((fp->f_flag & FPOSIXSHM) != 0)
2818 flags |= MAP_NOSYNC;
2823 * Ensure that file and memory protections are
2824 * compatible. Note that we only worry about
2825 * writability if mapping is shared; in this case,
2826 * current and max prot are dictated by the open file.
2827 * XXX use the vnode instead? Problem is: what
2828 * credentials do we use for determination? What if
2829 * proc does a setuid?
2832 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2833 maxprot = VM_PROT_NONE;
2834 if ((prot & VM_PROT_EXECUTE) != 0)
2837 maxprot = VM_PROT_EXECUTE;
2838 if ((fp->f_flag & FREAD) != 0)
2839 maxprot |= VM_PROT_READ;
2840 else if ((prot & VM_PROT_READ) != 0)
2844 * If we are sharing potential changes via MAP_SHARED and we
2845 * are trying to get write permission although we opened it
2846 * without asking for it, bail out.
2848 if ((flags & MAP_SHARED) != 0) {
2849 if ((fp->f_flag & FWRITE) != 0)
2850 maxprot |= VM_PROT_WRITE;
2851 else if ((prot & VM_PROT_WRITE) != 0)
2854 maxprot |= VM_PROT_WRITE;
2855 cap_maxprot |= VM_PROT_WRITE;
2857 maxprot &= cap_maxprot;
2860 * For regular files and shared memory, POSIX requires that
2861 * the value of foff be a legitimate offset within the data
2862 * object. In particular, negative offsets are invalid.
2863 * Blocking negative offsets and overflows here avoids
2864 * possible wraparound or user-level access into reserved
2865 * ranges of the data object later. In contrast, POSIX does
2866 * not dictate how offsets are used by device drivers, so in
2867 * the case of a device mapping a negative offset is passed
2874 foff > OFF_MAX - size)
2877 writecounted = FALSE;
2878 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2879 &foff, &object, &writecounted);
2882 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2883 foff, writecounted, td);
2886 * If this mapping was accounted for in the vnode's
2887 * writecount, then undo that now.
2890 vm_pager_release_writecount(object, 0, size);
2891 vm_object_deallocate(object);
2894 /* Inform hwpmc(4) if an executable is being mapped. */
2895 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2896 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2898 pkm.pm_address = (uintptr_t) *addr;
2899 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2907 vn_fsid(struct vnode *vp, struct vattr *va)
2911 f = &vp->v_mount->mnt_stat.f_fsid;
2912 va->va_fsid = (uint32_t)f->val[1];
2913 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2914 va->va_fsid += (uint32_t)f->val[0];
2918 vn_fsync_buf(struct vnode *vp, int waitfor)
2920 struct buf *bp, *nbp;
2923 int error, maxretry;
2926 maxretry = 10000; /* large, arbitrarily chosen */
2928 if (vp->v_type == VCHR) {
2930 mp = vp->v_rdev->si_mountpt;
2937 * MARK/SCAN initialization to avoid infinite loops.
2939 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2940 bp->b_vflags &= ~BV_SCANNED;
2945 * Flush all dirty buffers associated with a vnode.
2948 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2949 if ((bp->b_vflags & BV_SCANNED) != 0)
2951 bp->b_vflags |= BV_SCANNED;
2952 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2953 if (waitfor != MNT_WAIT)
2956 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2957 BO_LOCKPTR(bo)) != 0) {
2964 KASSERT(bp->b_bufobj == bo,
2965 ("bp %p wrong b_bufobj %p should be %p",
2966 bp, bp->b_bufobj, bo));
2967 if ((bp->b_flags & B_DELWRI) == 0)
2968 panic("fsync: not dirty");
2969 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2975 if (maxretry < 1000)
2976 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2982 * If synchronous the caller expects us to completely resolve all
2983 * dirty buffers in the system. Wait for in-progress I/O to
2984 * complete (which could include background bitmap writes), then
2985 * retry if dirty blocks still exist.
2987 if (waitfor == MNT_WAIT) {
2988 bufobj_wwait(bo, 0, 0);
2989 if (bo->bo_dirty.bv_cnt > 0) {
2991 * If we are unable to write any of these buffers
2992 * then we fail now rather than trying endlessly
2993 * to write them out.
2995 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2996 if ((error = bp->b_error) != 0)
2998 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2999 (error == 0 && --maxretry >= 0))
3007 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
3013 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
3014 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
3015 * to do the actual copy.
3016 * vn_generic_copy_file_range() is factored out, so it can be called
3017 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
3018 * different file systems.
3021 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
3022 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
3023 struct ucred *outcred, struct thread *fsize_td)
3030 *lenp = 0; /* For error returns. */
3033 /* Do some sanity checks on the arguments. */
3034 if (invp->v_type == VDIR || outvp->v_type == VDIR)
3036 else if (*inoffp < 0 || *outoffp < 0 ||
3037 invp->v_type != VREG || outvp->v_type != VREG)
3042 /* Ensure offset + len does not wrap around. */
3045 if (uval > INT64_MAX)
3046 len = INT64_MAX - *inoffp;
3049 if (uval > INT64_MAX)
3050 len = INT64_MAX - *outoffp;
3055 * If the two vnode are for the same file system, call
3056 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
3057 * which can handle copies across multiple file systems.
3060 if (invp->v_mount == outvp->v_mount)
3061 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
3062 lenp, flags, incred, outcred, fsize_td);
3064 error = vn_generic_copy_file_range(invp, inoffp, outvp,
3065 outoffp, lenp, flags, incred, outcred, fsize_td);
3071 * Test len bytes of data starting at dat for all bytes == 0.
3072 * Return true if all bytes are zero, false otherwise.
3073 * Expects dat to be well aligned.
3076 mem_iszero(void *dat, int len)
3082 for (p = dat; len > 0; len -= sizeof(*p), p++) {
3083 if (len >= sizeof(*p)) {
3087 cp = (const char *)p;
3088 for (i = 0; i < len; i++, cp++)
3097 * Look for a hole in the output file and, if found, adjust *outoffp
3098 * and *xferp to skip past the hole.
3099 * *xferp is the entire hole length to be written and xfer2 is how many bytes
3100 * to be written as 0's upon return.
3103 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3104 off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3109 if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3110 *dataoffp = *outoffp;
3111 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3114 *holeoffp = *dataoffp;
3115 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3118 if (error != 0 || *holeoffp == *dataoffp) {
3120 * Since outvp is unlocked, it may be possible for
3121 * another thread to do a truncate(), lseek(), write()
3122 * creating a hole at startoff between the above
3123 * VOP_IOCTL() calls, if the other thread does not do
3125 * If that happens, *holeoffp == *dataoffp and finding
3126 * the hole has failed, so disable vn_skip_hole().
3128 *holeoffp = -1; /* Disable use of vn_skip_hole(). */
3131 KASSERT(*dataoffp >= *outoffp,
3132 ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3133 (intmax_t)*dataoffp, (intmax_t)*outoffp));
3134 KASSERT(*holeoffp > *dataoffp,
3135 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3136 (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3140 * If there is a hole before the data starts, advance *outoffp and
3141 * *xferp past the hole.
3143 if (*dataoffp > *outoffp) {
3144 delta = *dataoffp - *outoffp;
3145 if (delta >= *xferp) {
3146 /* Entire *xferp is a hole. */
3153 xfer2 = MIN(xfer2, *xferp);
3157 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3158 * that the write ends at the start of the hole.
3159 * *holeoffp should always be greater than *outoffp, but for the
3160 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3163 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3164 xfer2 = *holeoffp - *outoffp;
3169 * Write an xfer sized chunk to outvp in blksize blocks from dat.
3170 * dat is a maximum of blksize in length and can be written repeatedly in
3172 * If growfile == true, just grow the file via vn_truncate_locked() instead
3173 * of doing actual writes.
3174 * If checkhole == true, a hole is being punched, so skip over any hole
3175 * already in the output file.
3178 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3179 u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3182 off_t dataoff, holeoff, xfer2;
3186 * Loop around doing writes of blksize until write has been completed.
3187 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3188 * done for each iteration, since the xfer argument can be very
3189 * large if there is a large hole to punch in the output file.
3194 xfer2 = MIN(xfer, blksize);
3197 * Punching a hole. Skip writing if there is
3198 * already a hole in the output file.
3200 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3201 &dataoff, &holeoff, cred);
3206 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3211 error = vn_start_write(outvp, &mp, V_WAIT);
3215 error = vn_lock(outvp, LK_EXCLUSIVE);
3217 error = vn_truncate_locked(outvp, outoff + xfer,
3222 error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3224 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3225 outoff, UIO_SYSSPACE, IO_NODELOCKED,
3226 curthread->td_ucred, cred, NULL, curthread);
3233 vn_finished_write(mp);
3234 } while (!growfile && xfer > 0 && error == 0);
3239 * Copy a byte range of one file to another. This function can handle the
3240 * case where invp and outvp are on different file systems.
3241 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3242 * is no better file system specific way to do it.
3245 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3246 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3247 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3249 struct vattr va, inva;
3251 off_t startoff, endoff, xfer, xfer2;
3253 int error, interrupted;
3254 bool cantseek, readzeros, eof, lastblock, holetoeof;
3255 ssize_t aresid, r = 0;
3256 size_t copylen, len, savlen;
3258 long holein, holeout;
3259 struct timespec curts, endts;
3261 holein = holeout = 0;
3262 savlen = len = *lenp;
3267 error = vn_lock(invp, LK_SHARED);
3270 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3273 error = VOP_GETATTR(invp, &inva, incred);
3279 error = vn_start_write(outvp, &mp, V_WAIT);
3281 error = vn_lock(outvp, LK_EXCLUSIVE);
3284 * If fsize_td != NULL, do a vn_rlimit_fsizex() call,
3285 * now that outvp is locked.
3287 if (fsize_td != NULL) {
3290 io.uio_offset = *outoffp;
3292 error = vn_rlimit_fsizex(outvp, &io, 0, &r, fsize_td);
3293 len = savlen = io.uio_resid;
3295 * No need to call vn_rlimit_fsizex_res before return,
3296 * since the uio is local.
3299 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3302 * Holes that are past EOF do not need to be written as a block
3303 * of zero bytes. So, truncate the output file as far as
3304 * possible and then use va.va_size to decide if writing 0
3305 * bytes is necessary in the loop below.
3308 error = VOP_GETATTR(outvp, &va, outcred);
3309 if (error == 0 && va.va_size > *outoffp && va.va_size <=
3312 error = mac_vnode_check_write(curthread->td_ucred,
3316 error = vn_truncate_locked(outvp, *outoffp,
3319 va.va_size = *outoffp;
3324 vn_finished_write(mp);
3328 if (holein == 0 && holeout > 0) {
3330 * For this special case, the input data will be scanned
3331 * for blocks of all 0 bytes. For these blocks, the
3332 * write can be skipped for the output file to create
3333 * an unallocated region.
3334 * Therefore, use the appropriate size for the output file.
3337 if (blksize <= 512) {
3339 * Use f_iosize, since ZFS reports a _PC_MIN_HOLE_SIZE
3340 * of 512, although it actually only creates
3341 * unallocated regions for blocks >= f_iosize.
3343 blksize = outvp->v_mount->mnt_stat.f_iosize;
3347 * Use the larger of the two f_iosize values. If they are
3348 * not the same size, one will normally be an exact multiple of
3349 * the other, since they are both likely to be a power of 2.
3351 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3352 outvp->v_mount->mnt_stat.f_iosize);
3355 /* Clip to sane limits. */
3358 else if (blksize > maxphys)
3360 dat = malloc(blksize, M_TEMP, M_WAITOK);
3363 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3364 * to find holes. Otherwise, just scan the read block for all 0s
3365 * in the inner loop where the data copying is done.
3366 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3367 * support holes on the server, but do not support FIOSEEKHOLE.
3368 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3369 * that this function should return after 1second with a partial
3372 if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3373 getnanouptime(&endts);
3376 timespecclear(&endts);
3377 holetoeof = eof = false;
3378 while (len > 0 && error == 0 && !eof && interrupted == 0) {
3379 endoff = 0; /* To shut up compilers. */
3385 * Find the next data area. If there is just a hole to EOF,
3386 * FIOSEEKDATA should fail with ENXIO.
3387 * (I do not know if any file system will report a hole to
3388 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3389 * will fail for those file systems.)
3391 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3392 * the code just falls through to the inner copy loop.
3396 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3398 if (error == ENXIO) {
3399 startoff = endoff = inva.va_size;
3400 eof = holetoeof = true;
3404 if (error == 0 && !holetoeof) {
3406 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3409 * Since invp is unlocked, it may be possible for
3410 * another thread to do a truncate(), lseek(), write()
3411 * creating a hole at startoff between the above
3412 * VOP_IOCTL() calls, if the other thread does not do
3414 * If that happens, startoff == endoff and finding
3415 * the hole has failed, so set an error.
3417 if (error == 0 && startoff == endoff)
3418 error = EINVAL; /* Any error. Reset to 0. */
3421 if (startoff > *inoffp) {
3422 /* Found hole before data block. */
3423 xfer = MIN(startoff - *inoffp, len);
3424 if (*outoffp < va.va_size) {
3425 /* Must write 0s to punch hole. */
3426 xfer2 = MIN(va.va_size - *outoffp,
3428 memset(dat, 0, MIN(xfer2, blksize));
3429 error = vn_write_outvp(outvp, dat,
3430 *outoffp, xfer2, blksize, false,
3431 holeout > 0, outcred);
3434 if (error == 0 && *outoffp + xfer >
3435 va.va_size && (xfer == len || holetoeof)) {
3436 /* Grow output file (hole at end). */
3437 error = vn_write_outvp(outvp, dat,
3438 *outoffp, xfer, blksize, true,
3446 interrupted = sig_intr();
3447 if (timespecisset(&endts) &&
3449 getnanouptime(&curts);
3450 if (timespeccmp(&curts,
3458 copylen = MIN(len, endoff - startoff);
3470 * Set first xfer to end at a block boundary, so that
3471 * holes are more likely detected in the loop below via
3472 * the for all bytes 0 method.
3474 xfer -= (*inoffp % blksize);
3476 /* Loop copying the data block. */
3477 while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3480 error = vn_lock(invp, LK_SHARED);
3483 error = vn_rdwr(UIO_READ, invp, dat, xfer,
3484 startoff, UIO_SYSSPACE, IO_NODELOCKED,
3485 curthread->td_ucred, incred, &aresid,
3489 if (error == 0 && aresid > 0) {
3490 /* Stop the copy at EOF on the input file. */
3497 * Skip the write for holes past the initial EOF
3498 * of the output file, unless this is the last
3499 * write of the output file at EOF.
3501 readzeros = cantseek ? mem_iszero(dat, xfer) :
3505 if (!cantseek || *outoffp < va.va_size ||
3506 lastblock || !readzeros)
3507 error = vn_write_outvp(outvp, dat,
3508 *outoffp, xfer, blksize,
3509 readzeros && lastblock &&
3510 *outoffp >= va.va_size, false,
3519 interrupted = sig_intr();
3520 if (timespecisset(&endts) &&
3522 getnanouptime(&curts);
3523 if (timespeccmp(&curts,
3535 *lenp = savlen - len;
3541 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3545 off_t olen, ooffset;
3548 int audited_vnode1 = 0;
3552 if (vp->v_type != VREG)
3555 /* Allocating blocks may take a long time, so iterate. */
3562 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3565 error = vn_lock(vp, LK_EXCLUSIVE);
3567 vn_finished_write(mp);
3571 if (!audited_vnode1) {
3572 AUDIT_ARG_VNODE1(vp);
3577 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3580 error = VOP_ALLOCATE(vp, &offset, &len, 0,
3583 vn_finished_write(mp);
3585 if (olen + ooffset != offset + len) {
3586 panic("offset + len changed from %jx/%jx to %jx/%jx",
3587 ooffset, olen, offset, len);
3589 if (error != 0 || len == 0)
3591 KASSERT(olen > len, ("Iteration did not make progress?"));
3598 #define DIRENT_MINSIZE (sizeof(struct dirent) - (MAXNAMLEN+1) + 4)
3601 * Keep this assert as long as sizeof(struct dirent) is used as the maximum
3604 _Static_assert(_GENERIC_MAXDIRSIZ == sizeof(struct dirent),
3605 "'struct dirent' size must be a multiple of its alignment "
3606 "(see _GENERIC_DIRLEN())");
3609 * Returns successive directory entries through some caller's provided buffer.
3611 * This function automatically refills the provided buffer with calls to
3612 * VOP_READDIR() (after MAC permission checks).
3614 * 'td' is used for credentials and passed to uiomove(). 'dirbuf' is the
3615 * caller's buffer to fill and 'dirbuflen' its allocated size. 'dirbuf' must
3616 * be properly aligned to access 'struct dirent' structures and 'dirbuflen'
3617 * must be greater than GENERIC_MAXDIRSIZ to avoid VOP_READDIR() returning
3618 * EINVAL (the latter is not a strong guarantee (yet); but EINVAL will always
3619 * be returned if this requirement is not verified). '*dpp' points to the
3620 * current directory entry in the buffer and '*len' contains the remaining
3621 * valid bytes in 'dirbuf' after 'dpp' (including the pointed entry).
3623 * At first call (or when restarting the read), '*len' must have been set to 0,
3624 * '*off' to 0 (or any valid start offset) and '*eofflag' to 0. There are no
3625 * more entries as soon as '*len' is 0 after a call that returned 0. Calling
3626 * again this function after such a condition is considered an error and EINVAL
3627 * will be returned. Other possible error codes are those of VOP_READDIR(),
3628 * EINTEGRITY if the returned entries do not pass coherency tests, or EINVAL
3629 * (bad call). All errors are unrecoverable, i.e., the state ('*len', '*off'
3630 * and '*eofflag') must be re-initialized before a subsequent call. On error
3631 * or at end of directory, '*dpp' is reset to NULL.
3633 * '*len', '*off' and '*eofflag' are internal state the caller should not
3634 * tamper with except as explained above. '*off' is the next directory offset
3635 * to read from to refill the buffer. '*eofflag' is set to 0 or 1 by the last
3636 * internal call to VOP_READDIR() that returned without error, indicating
3637 * whether it reached the end of the directory, and to 2 by this function after
3638 * all entries have been read.
3641 vn_dir_next_dirent(struct vnode *vp, struct thread *td,
3642 char *dirbuf, size_t dirbuflen,
3643 struct dirent **dpp, size_t *len, off_t *off, int *eofflag)
3645 struct dirent *dp = NULL;
3651 ASSERT_VOP_LOCKED(vp, "vnode not locked");
3652 VNASSERT(vp->v_type == VDIR, vp, ("vnode is not a directory"));
3653 MPASS2((uintptr_t)dirbuf < (uintptr_t)dirbuf + dirbuflen,
3654 "Address space overflow");
3656 if (__predict_false(dirbuflen < GENERIC_MAXDIRSIZ)) {
3657 /* Don't take any chances in this case */
3666 * The caller continued to call us after an error (we set dp to
3667 * NULL in a previous iteration). Bail out right now.
3669 if (__predict_false(dp == NULL))
3672 MPASS(*len <= dirbuflen);
3673 MPASS2((uintptr_t)dirbuf <= (uintptr_t)dp &&
3674 (uintptr_t)dp + *len <= (uintptr_t)dirbuf + dirbuflen,
3675 "Filled range not inside buffer");
3677 reclen = dp->d_reclen;
3678 if (reclen >= *len) {
3679 /* End of buffer reached */
3682 dp = (struct dirent *)((char *)dp + reclen);
3690 /* Have to refill. */
3696 /* Nothing more to read. */
3697 *eofflag = 2; /* Remember the caller reached EOF. */
3701 /* The caller didn't test for EOF. */
3706 iov.iov_base = dirbuf;
3707 iov.iov_len = dirbuflen;
3711 uio.uio_offset = *off;
3712 uio.uio_resid = dirbuflen;
3713 uio.uio_segflg = UIO_SYSSPACE;
3714 uio.uio_rw = UIO_READ;
3718 error = mac_vnode_check_readdir(td->td_ucred, vp);
3721 error = VOP_READDIR(vp, &uio, td->td_ucred, eofflag,
3726 *len = dirbuflen - uio.uio_resid;
3727 *off = uio.uio_offset;
3730 /* Sanity check on INVARIANTS. */
3731 MPASS(*eofflag != 0);
3737 * Normalize the flag returned by VOP_READDIR(), since we use 2
3738 * as a sentinel value.
3743 dp = (struct dirent *)dirbuf;
3746 if (__predict_false(*len < GENERIC_MINDIRSIZ ||
3747 dp->d_reclen < GENERIC_MINDIRSIZ)) {
3761 * Checks whether a directory is empty or not.
3763 * If the directory is empty, returns 0, and if it is not, ENOTEMPTY. Other
3764 * values are genuine errors preventing the check.
3767 vn_dir_check_empty(struct vnode *vp)
3769 struct thread *const td = curthread;
3771 size_t dirbuflen, len;
3777 ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
3778 VNPASS(vp->v_type == VDIR, vp);
3780 error = VOP_GETATTR(vp, &va, td->td_ucred);
3784 dirbuflen = max(DEV_BSIZE, GENERIC_MAXDIRSIZ);
3785 if (dirbuflen < va.va_blocksize)
3786 dirbuflen = va.va_blocksize;
3787 dirbuf = malloc(dirbuflen, M_TEMP, M_WAITOK);
3794 error = vn_dir_next_dirent(vp, td, dirbuf, dirbuflen,
3795 &dp, &len, &off, &eofflag);
3806 * Skip whiteouts. Unionfs operates on filesystems only and
3807 * not on hierarchies, so these whiteouts would be shadowed on
3808 * the system hierarchy but not for a union using the
3809 * filesystem of their directories as the upper layer.
3810 * Additionally, unionfs currently transparently exposes
3811 * union-specific metadata of its upper layer, meaning that
3812 * whiteouts can be seen through the union view in empty
3813 * directories. Taking into account these whiteouts would then
3814 * prevent mounting another filesystem on such effectively
3815 * empty directories.
3817 if (dp->d_type == DT_WHT)
3821 * Any file in the directory which is not '.' or '..' indicates
3822 * the directory is not empty.
3824 switch (dp->d_namlen) {
3826 if (dp->d_name[1] != '.') {
3827 /* Can't be '..' (nor '.') */
3833 if (dp->d_name[0] != '.') {
3834 /* Can't be '..' nor '.' */
3847 free(dirbuf, M_TEMP);
3852 static u_long vn_lock_pair_pause_cnt;
3853 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3854 &vn_lock_pair_pause_cnt, 0,
3855 "Count of vn_lock_pair deadlocks");
3857 u_int vn_lock_pair_pause_max;
3858 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3859 &vn_lock_pair_pause_max, 0,
3860 "Max ticks for vn_lock_pair deadlock avoidance sleep");
3863 vn_lock_pair_pause(const char *wmesg)
3865 atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3866 pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3870 * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3871 * vp1_locked indicates whether vp1 is locked; if not, vp1 must be
3872 * unlocked. Same for vp2 and vp2_locked. One of the vnodes can be
3875 * The function returns with both vnodes exclusively or shared locked,
3876 * according to corresponding lkflags, and guarantees that it does not
3877 * create lock order reversal with other threads during its execution.
3878 * Both vnodes could be unlocked temporary (and reclaimed).
3880 * If requesting shared locking, locked vnode lock must not be recursed.
3882 * Only one of LK_SHARED and LK_EXCLUSIVE must be specified.
3883 * LK_NODDLKTREAT can be optionally passed.
3886 vn_lock_pair(struct vnode *vp1, bool vp1_locked, int lkflags1,
3887 struct vnode *vp2, bool vp2_locked, int lkflags2)
3891 MPASS(((lkflags1 & LK_SHARED) != 0) ^ ((lkflags1 & LK_EXCLUSIVE) != 0));
3892 MPASS((lkflags1 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0);
3893 MPASS(((lkflags2 & LK_SHARED) != 0) ^ ((lkflags2 & LK_EXCLUSIVE) != 0));
3894 MPASS((lkflags2 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0);
3896 if (vp1 == NULL && vp2 == NULL)
3900 if ((lkflags1 & LK_SHARED) != 0 &&
3901 (vp1->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
3902 lkflags1 = (lkflags1 & ~LK_SHARED) | LK_EXCLUSIVE;
3903 if (vp1_locked && VOP_ISLOCKED(vp1) != LK_EXCLUSIVE) {
3904 ASSERT_VOP_LOCKED(vp1, "vp1");
3905 if ((lkflags1 & LK_EXCLUSIVE) != 0) {
3907 ASSERT_VOP_UNLOCKED(vp1,
3908 "vp1 shared recursed");
3911 } else if (!vp1_locked)
3912 ASSERT_VOP_UNLOCKED(vp1, "vp1");
3918 if ((lkflags2 & LK_SHARED) != 0 &&
3919 (vp2->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
3920 lkflags2 = (lkflags2 & ~LK_SHARED) | LK_EXCLUSIVE;
3921 if (vp2_locked && VOP_ISLOCKED(vp2) != LK_EXCLUSIVE) {
3922 ASSERT_VOP_LOCKED(vp2, "vp2");
3923 if ((lkflags2 & LK_EXCLUSIVE) != 0) {
3925 ASSERT_VOP_UNLOCKED(vp2,
3926 "vp2 shared recursed");
3929 } else if (!vp2_locked)
3930 ASSERT_VOP_UNLOCKED(vp2, "vp2");
3935 if (!vp1_locked && !vp2_locked) {
3936 vn_lock(vp1, lkflags1 | LK_RETRY);
3940 while (!vp1_locked || !vp2_locked) {
3941 if (vp1_locked && vp2 != NULL) {
3943 error = VOP_LOCK1(vp2, lkflags2 | LK_NOWAIT,
3944 __FILE__, __LINE__);
3949 vn_lock_pair_pause("vlp1");
3951 vn_lock(vp2, lkflags2 | LK_RETRY);
3954 if (vp2_locked && vp1 != NULL) {
3956 error = VOP_LOCK1(vp1, lkflags1 | LK_NOWAIT,
3957 __FILE__, __LINE__);
3962 vn_lock_pair_pause("vlp2");
3964 vn_lock(vp1, lkflags1 | LK_RETRY);
3969 if (lkflags1 == LK_EXCLUSIVE)
3970 ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3972 ASSERT_VOP_LOCKED(vp1, "vp1 ret");
3975 if (lkflags2 == LK_EXCLUSIVE)
3976 ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3978 ASSERT_VOP_LOCKED(vp2, "vp2 ret");
3983 vn_lktype_write(struct mount *mp, struct vnode *vp)
3985 if (MNT_SHARED_WRITES(mp) ||
3986 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
3988 return (LK_EXCLUSIVE);