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
43 #include <sys/cdefs.h>
44 #include "opt_hwpmc_hooks.h"
46 #include <sys/param.h>
47 #include <sys/systm.h>
50 #include <sys/fcntl.h>
57 #include <sys/limits.h>
60 #include <sys/mount.h>
61 #include <sys/mutex.h>
62 #include <sys/namei.h>
63 #include <sys/vnode.h>
64 #include <sys/dirent.h>
67 #include <sys/filio.h>
68 #include <sys/resourcevar.h>
69 #include <sys/rwlock.h>
72 #include <sys/sleepqueue.h>
73 #include <sys/sysctl.h>
74 #include <sys/ttycom.h>
76 #include <sys/syslog.h>
77 #include <sys/unistd.h>
79 #include <sys/ktrace.h>
81 #include <security/audit/audit.h>
82 #include <security/mac/mac_framework.h>
85 #include <vm/vm_extern.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pager.h>
91 #include <vm/vnode_pager.h>
94 #include <sys/pmckern.h>
97 static fo_rdwr_t vn_read;
98 static fo_rdwr_t vn_write;
99 static fo_rdwr_t vn_io_fault;
100 static fo_truncate_t vn_truncate;
101 static fo_ioctl_t vn_ioctl;
102 static fo_poll_t vn_poll;
103 static fo_kqfilter_t vn_kqfilter;
104 static fo_close_t vn_closefile;
105 static fo_mmap_t vn_mmap;
106 static fo_fallocate_t vn_fallocate;
107 static fo_fspacectl_t vn_fspacectl;
109 struct fileops vnops = {
110 .fo_read = vn_io_fault,
111 .fo_write = vn_io_fault,
112 .fo_truncate = vn_truncate,
113 .fo_ioctl = vn_ioctl,
115 .fo_kqfilter = vn_kqfilter,
116 .fo_stat = vn_statfile,
117 .fo_close = vn_closefile,
118 .fo_chmod = vn_chmod,
119 .fo_chown = vn_chown,
120 .fo_sendfile = vn_sendfile,
122 .fo_fill_kinfo = vn_fill_kinfo,
124 .fo_fallocate = vn_fallocate,
125 .fo_fspacectl = vn_fspacectl,
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 = curthread;
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 ((fmode & FREAD) != 0)
209 if ((fmode & FWRITE) != 0)
211 if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
213 if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
215 if ((vn_open_flags & VN_OPEN_WANTIOCTLCAPS) != 0)
216 res |= WANTIOCTLCAPS;
221 * Common code for vnode open operations via a name lookup.
222 * Lookup the vnode and invoke VOP_CREATE if needed.
223 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
225 * Note that this does NOT free nameidata for the successful case,
226 * due to the NDINIT being done elsewhere.
229 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
230 struct ucred *cred, struct file *fp)
235 struct vattr *vap = &vat;
242 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
243 O_EXCL | O_DIRECTORY) ||
244 (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
246 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
247 ndp->ni_cnd.cn_nameiop = CREATE;
248 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
250 * Set NOCACHE to avoid flushing the cache when
251 * rolling in many files at once.
253 * Set NC_KEEPPOSENTRY to keep positive entries if they already
254 * exist despite NOCACHE.
256 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
257 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
258 ndp->ni_cnd.cn_flags |= FOLLOW;
259 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
261 if ((error = namei(ndp)) != 0)
263 if (ndp->ni_vp == NULL) {
266 vap->va_mode = cmode;
268 vap->va_vaflags |= VA_EXCLUSIVE;
269 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
272 if ((error = vn_start_write(NULL, &mp,
273 V_XSLEEP | V_PCATCH)) != 0)
278 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
279 ndp->ni_cnd.cn_flags |= MAKEENTRY;
281 error = mac_vnode_check_create(cred, ndp->ni_dvp,
285 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
288 if (error == 0 && (fmode & O_EXCL) != 0 &&
289 (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
291 vp->v_iflag |= VI_FOPENING;
295 VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
297 vn_finished_write(mp);
300 if (error == ERELOOKUP) {
308 if (ndp->ni_dvp == ndp->ni_vp)
314 if (fmode & O_EXCL) {
318 if (vp->v_type == VDIR) {
325 ndp->ni_cnd.cn_nameiop = LOOKUP;
326 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
327 ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
329 if ((fmode & FWRITE) == 0)
330 ndp->ni_cnd.cn_flags |= LOCKSHARED;
331 if ((error = namei(ndp)) != 0)
335 error = vn_open_vnode(vp, fmode, cred, curthread, fp);
338 vp->v_iflag &= ~VI_FOPENING;
355 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
358 int error, lock_flags, type;
360 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
361 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
363 KASSERT(fp != NULL, ("open with flock requires fp"));
364 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
367 lock_flags = VOP_ISLOCKED(vp);
370 lf.l_whence = SEEK_SET;
373 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
375 if ((fmode & FNONBLOCK) == 0)
377 if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
379 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
381 fp->f_flag |= FHASLOCK;
383 vn_lock(vp, lock_flags | LK_RETRY);
388 * Common code for vnode open operations once a vnode is located.
389 * Check permissions, and call the VOP_OPEN routine.
392 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
393 struct thread *td, struct file *fp)
398 if (vp->v_type == VLNK) {
399 if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
402 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
406 if ((fmode & O_PATH) == 0) {
407 if (vp->v_type == VSOCK)
409 if ((fmode & (FWRITE | O_TRUNC)) != 0) {
410 if (vp->v_type == VDIR)
414 if ((fmode & FREAD) != 0)
416 if ((fmode & O_APPEND) && (fmode & FWRITE))
419 if ((fmode & O_CREAT) != 0)
423 if ((fmode & FEXEC) != 0)
426 if ((fmode & O_VERIFY) != 0)
428 error = mac_vnode_check_open(cred, vp, accmode);
432 accmode &= ~(VCREAT | VVERIFY);
434 if ((fmode & O_CREAT) == 0 && accmode != 0) {
435 error = VOP_ACCESS(vp, accmode, cred, td);
439 if ((fmode & O_PATH) != 0) {
440 if (vp->v_type != VFIFO && vp->v_type != VSOCK &&
441 VOP_ACCESS(vp, VREAD, cred, td) == 0)
442 fp->f_flag |= FKQALLOWED;
446 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
447 vn_lock(vp, LK_UPGRADE | LK_RETRY);
448 error = VOP_OPEN(vp, fmode, cred, td, fp);
452 error = vn_open_vnode_advlock(vp, fmode, fp);
453 if (error == 0 && (fmode & FWRITE) != 0) {
454 error = VOP_ADD_WRITECOUNT(vp, 1);
456 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
457 __func__, vp, vp->v_writecount);
462 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
463 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
468 * Arrange the call by having fdrop() to use
469 * vn_closefile(). This is to satisfy
470 * filesystems like devfs or tmpfs, which
471 * override fo_close().
473 fp->f_flag |= FOPENFAILED;
475 if (fp->f_ops == &badfileops) {
476 fp->f_type = DTYPE_VNODE;
482 * If there is no fp, due to kernel-mode open,
483 * we can call VOP_CLOSE() now.
485 if ((vp->v_type == VFIFO ||
486 !MNT_EXTENDED_SHARED(vp->v_mount)) &&
487 VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
488 vn_lock(vp, LK_UPGRADE | LK_RETRY);
489 (void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
494 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
500 * Check for write permissions on the specified vnode.
501 * Prototype text segments cannot be written.
505 vn_writechk(struct vnode *vp)
508 ASSERT_VOP_LOCKED(vp, "vn_writechk");
510 * If there's shared text associated with
511 * the vnode, try to free it up once. If
512 * we fail, we can't allow writing.
524 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
525 struct thread *td, bool keep_ref)
528 int error, lock_flags;
530 lock_flags = vp->v_type != VFIFO && MNT_EXTENDED_SHARED(vp->v_mount) ?
531 LK_SHARED : 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 | V_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_read(struct file *fp)
836 return (atomic_load_long(&fp->f_offset));
840 foffset_lock(struct file *fp, int flags)
845 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
847 mtxp = mtx_pool_find(mtxpool_sleep, fp);
849 if ((flags & FOF_NOLOCK) == 0) {
850 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
851 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
852 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
855 fp->f_vnread_flags |= FOFFSET_LOCKED;
863 foffset_unlock(struct file *fp, off_t val, int flags)
867 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
869 mtxp = mtx_pool_find(mtxpool_sleep, fp);
871 if ((flags & FOF_NOUPDATE) == 0)
873 if ((flags & FOF_NEXTOFF_R) != 0)
874 fp->f_nextoff[UIO_READ] = val;
875 if ((flags & FOF_NEXTOFF_W) != 0)
876 fp->f_nextoff[UIO_WRITE] = val;
877 if ((flags & FOF_NOLOCK) == 0) {
878 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
879 ("Lost FOFFSET_LOCKED"));
880 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
881 wakeup(&fp->f_vnread_flags);
882 fp->f_vnread_flags = 0;
888 foffset_read(struct file *fp)
891 return (foffset_lock(fp, FOF_NOLOCK));
896 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
899 if ((flags & FOF_OFFSET) == 0)
900 uio->uio_offset = foffset_lock(fp, flags);
904 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
907 if ((flags & FOF_OFFSET) == 0)
908 foffset_unlock(fp, uio->uio_offset, flags);
912 get_advice(struct file *fp, struct uio *uio)
917 ret = POSIX_FADV_NORMAL;
918 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
921 mtxp = mtx_pool_find(mtxpool_sleep, fp);
923 if (fp->f_advice != NULL &&
924 uio->uio_offset >= fp->f_advice->fa_start &&
925 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
926 ret = fp->f_advice->fa_advice;
932 get_write_ioflag(struct file *fp)
940 mp = atomic_load_ptr(&vp->v_mount);
942 if ((fp->f_flag & O_DIRECT) != 0)
945 if ((fp->f_flag & O_FSYNC) != 0 ||
946 (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
950 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
951 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
952 * fall back to full O_SYNC behavior.
954 if ((fp->f_flag & O_DSYNC) != 0)
955 ioflag |= IO_SYNC | IO_DATASYNC;
961 vn_read_from_obj(struct vnode *vp, struct uio *uio)
964 vm_page_t ma[io_hold_cnt + 2];
969 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
970 obj = atomic_load_ptr(&vp->v_object);
972 return (EJUSTRETURN);
975 * Depends on type stability of vm_objects.
977 vm_object_pip_add(obj, 1);
978 if ((obj->flags & OBJ_DEAD) != 0) {
980 * Note that object might be already reused from the
981 * vnode, and the OBJ_DEAD flag cleared. This is fine,
982 * we recheck for DOOMED vnode state after all pages
983 * are busied, and retract then.
985 * But we check for OBJ_DEAD to ensure that we do not
986 * busy pages while vm_object_terminate_pages()
987 * processes the queue.
993 resid = uio->uio_resid;
994 off = uio->uio_offset;
995 for (i = 0; resid > 0; i++) {
996 MPASS(i < io_hold_cnt + 2);
997 ma[i] = vm_page_grab_unlocked(obj, atop(off),
998 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
1004 * Skip invalid pages. Valid mask can be partial only
1005 * at EOF, and we clip later.
1007 if (vm_page_none_valid(ma[i])) {
1008 vm_page_sunbusy(ma[i]);
1016 error = EJUSTRETURN;
1021 * Check VIRF_DOOMED after we busied our pages. Since
1022 * vgonel() terminates the vnode' vm_object, it cannot
1023 * process past pages busied by us.
1025 if (VN_IS_DOOMED(vp)) {
1026 error = EJUSTRETURN;
1030 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1031 if (resid > uio->uio_resid)
1032 resid = uio->uio_resid;
1035 * Unlocked read of vnp_size is safe because truncation cannot
1036 * pass busied page. But we load vnp_size into a local
1037 * variable so that possible concurrent extension does not
1038 * break calculation.
1040 #if defined(__powerpc__) && !defined(__powerpc64__)
1041 vsz = obj->un_pager.vnp.vnp_size;
1043 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1045 if (uio->uio_offset >= vsz) {
1046 error = EJUSTRETURN;
1049 if (uio->uio_offset + resid > vsz)
1050 resid = vsz - uio->uio_offset;
1052 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1055 for (j = 0; j < i; j++) {
1057 vm_page_reference(ma[j]);
1058 vm_page_sunbusy(ma[j]);
1061 vm_object_pip_wakeup(obj);
1064 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1068 * File table vnode read routine.
1071 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1079 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1081 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1084 if (fp->f_flag & FNONBLOCK)
1085 ioflag |= IO_NDELAY;
1086 if (fp->f_flag & O_DIRECT)
1087 ioflag |= IO_DIRECT;
1090 * Try to read from page cache. VIRF_DOOMED check is racy but
1091 * allows us to avoid unneeded work outright.
1093 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1094 (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1095 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1097 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1100 if (error != EJUSTRETURN)
1104 advice = get_advice(fp, uio);
1105 vn_lock(vp, LK_SHARED | LK_RETRY);
1108 case POSIX_FADV_NORMAL:
1109 case POSIX_FADV_SEQUENTIAL:
1110 case POSIX_FADV_NOREUSE:
1111 ioflag |= sequential_heuristic(uio, fp);
1113 case POSIX_FADV_RANDOM:
1114 /* Disable read-ahead for random I/O. */
1117 orig_offset = uio->uio_offset;
1120 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1123 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1124 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1126 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1127 orig_offset != uio->uio_offset)
1129 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1130 * for the backing file after a POSIX_FADV_NOREUSE
1133 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1134 POSIX_FADV_DONTNEED);
1139 * File table vnode write routine.
1142 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1150 bool need_finished_write;
1152 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1154 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1156 if (vp->v_type == VREG)
1159 if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1160 ioflag |= IO_APPEND;
1161 if ((fp->f_flag & FNONBLOCK) != 0)
1162 ioflag |= IO_NDELAY;
1163 ioflag |= get_write_ioflag(fp);
1166 need_finished_write = false;
1167 if (vp->v_type != VCHR) {
1168 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
1171 need_finished_write = true;
1174 advice = get_advice(fp, uio);
1176 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1178 case POSIX_FADV_NORMAL:
1179 case POSIX_FADV_SEQUENTIAL:
1180 case POSIX_FADV_NOREUSE:
1181 ioflag |= sequential_heuristic(uio, fp);
1183 case POSIX_FADV_RANDOM:
1184 /* XXX: Is this correct? */
1187 orig_offset = uio->uio_offset;
1190 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1193 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1194 fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1196 if (need_finished_write)
1197 vn_finished_write(mp);
1198 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1199 orig_offset != uio->uio_offset)
1201 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1202 * for the backing file after a POSIX_FADV_NOREUSE
1205 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1206 POSIX_FADV_DONTNEED);
1212 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1213 * prevent the following deadlock:
1215 * Assume that the thread A reads from the vnode vp1 into userspace
1216 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
1217 * currently not resident, then system ends up with the call chain
1218 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1219 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1220 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1221 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1222 * backed by the pages of vnode vp1, and some page in buf2 is not
1223 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1225 * To prevent the lock order reversal and deadlock, vn_io_fault() does
1226 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1227 * Instead, it first tries to do the whole range i/o with pagefaults
1228 * disabled. If all pages in the i/o buffer are resident and mapped,
1229 * VOP will succeed (ignoring the genuine filesystem errors).
1230 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1231 * i/o in chunks, with all pages in the chunk prefaulted and held
1232 * using vm_fault_quick_hold_pages().
1234 * Filesystems using this deadlock avoidance scheme should use the
1235 * array of the held pages from uio, saved in the curthread->td_ma,
1236 * instead of doing uiomove(). A helper function
1237 * vn_io_fault_uiomove() converts uiomove request into
1238 * uiomove_fromphys() over td_ma array.
1240 * Since vnode locks do not cover the whole i/o anymore, rangelocks
1241 * make the current i/o request atomic with respect to other i/os and
1246 * Decode vn_io_fault_args and perform the corresponding i/o.
1249 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1255 save = vm_fault_disable_pagefaults();
1256 switch (args->kind) {
1257 case VN_IO_FAULT_FOP:
1258 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1259 uio, args->cred, args->flags, td);
1261 case VN_IO_FAULT_VOP:
1262 if (uio->uio_rw == UIO_READ) {
1263 error = VOP_READ(args->args.vop_args.vp, uio,
1264 args->flags, args->cred);
1265 } else if (uio->uio_rw == UIO_WRITE) {
1266 error = VOP_WRITE(args->args.vop_args.vp, uio,
1267 args->flags, args->cred);
1271 panic("vn_io_fault_doio: unknown kind of io %d %d",
1272 args->kind, uio->uio_rw);
1274 vm_fault_enable_pagefaults(save);
1279 vn_io_fault_touch(char *base, const struct uio *uio)
1284 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1290 vn_io_fault_prefault_user(const struct uio *uio)
1293 const struct iovec *iov;
1298 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1299 ("vn_io_fault_prefault userspace"));
1303 resid = uio->uio_resid;
1304 base = iov->iov_base;
1307 error = vn_io_fault_touch(base, uio);
1310 if (len < PAGE_SIZE) {
1312 error = vn_io_fault_touch(base + len - 1, uio);
1317 if (++i >= uio->uio_iovcnt)
1319 iov = uio->uio_iov + i;
1320 base = iov->iov_base;
1332 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1333 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1334 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1335 * into args and call vn_io_fault1() to handle faults during the user
1336 * mode buffer accesses.
1339 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1342 vm_page_t ma[io_hold_cnt + 2];
1343 struct uio *uio_clone, short_uio;
1344 struct iovec short_iovec[1];
1345 vm_page_t *prev_td_ma;
1347 vm_offset_t addr, end;
1350 int error, cnt, saveheld, prev_td_ma_cnt;
1352 if (vn_io_fault_prefault) {
1353 error = vn_io_fault_prefault_user(uio);
1355 return (error); /* Or ignore ? */
1358 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1361 * The UFS follows IO_UNIT directive and replays back both
1362 * uio_offset and uio_resid if an error is encountered during the
1363 * operation. But, since the iovec may be already advanced,
1364 * uio is still in an inconsistent state.
1366 * Cache a copy of the original uio, which is advanced to the redo
1367 * point using UIO_NOCOPY below.
1369 uio_clone = cloneuio(uio);
1370 resid = uio->uio_resid;
1372 short_uio.uio_segflg = UIO_USERSPACE;
1373 short_uio.uio_rw = uio->uio_rw;
1374 short_uio.uio_td = uio->uio_td;
1376 error = vn_io_fault_doio(args, uio, td);
1377 if (error != EFAULT)
1380 atomic_add_long(&vn_io_faults_cnt, 1);
1381 uio_clone->uio_segflg = UIO_NOCOPY;
1382 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1383 uio_clone->uio_segflg = uio->uio_segflg;
1385 saveheld = curthread_pflags_set(TDP_UIOHELD);
1386 prev_td_ma = td->td_ma;
1387 prev_td_ma_cnt = td->td_ma_cnt;
1389 while (uio_clone->uio_resid != 0) {
1390 len = uio_clone->uio_iov->iov_len;
1392 KASSERT(uio_clone->uio_iovcnt >= 1,
1393 ("iovcnt underflow"));
1394 uio_clone->uio_iov++;
1395 uio_clone->uio_iovcnt--;
1398 if (len > ptoa(io_hold_cnt))
1399 len = ptoa(io_hold_cnt);
1400 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1401 end = round_page(addr + len);
1407 * A perfectly misaligned address and length could cause
1408 * both the start and the end of the chunk to use partial
1409 * page. +2 accounts for such a situation.
1411 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1412 addr, len, prot, ma, io_hold_cnt + 2);
1417 short_uio.uio_iov = &short_iovec[0];
1418 short_iovec[0].iov_base = (void *)addr;
1419 short_uio.uio_iovcnt = 1;
1420 short_uio.uio_resid = short_iovec[0].iov_len = len;
1421 short_uio.uio_offset = uio_clone->uio_offset;
1423 td->td_ma_cnt = cnt;
1425 error = vn_io_fault_doio(args, &short_uio, td);
1426 vm_page_unhold_pages(ma, cnt);
1427 adv = len - short_uio.uio_resid;
1429 uio_clone->uio_iov->iov_base =
1430 (char *)uio_clone->uio_iov->iov_base + adv;
1431 uio_clone->uio_iov->iov_len -= adv;
1432 uio_clone->uio_resid -= adv;
1433 uio_clone->uio_offset += adv;
1435 uio->uio_resid -= adv;
1436 uio->uio_offset += adv;
1438 if (error != 0 || adv == 0)
1441 td->td_ma = prev_td_ma;
1442 td->td_ma_cnt = prev_td_ma_cnt;
1443 curthread_pflags_restore(saveheld);
1450 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1451 int flags, struct thread *td)
1456 struct vn_io_fault_args args;
1458 bool do_io_fault, do_rangelock;
1460 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1464 * The ability to read(2) on a directory has historically been
1465 * allowed for all users, but this can and has been the source of
1466 * at least one security issue in the past. As such, it is now hidden
1467 * away behind a sysctl for those that actually need it to use it, and
1468 * restricted to root when it's turned on to make it relatively safe to
1469 * leave on for longer sessions of need.
1471 if (vp->v_type == VDIR) {
1472 KASSERT(uio->uio_rw == UIO_READ,
1473 ("illegal write attempted on a directory"));
1474 if (!vfs_allow_read_dir)
1476 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1480 do_io_fault = do_vn_io_fault(vp, uio);
1481 do_rangelock = do_io_fault || (vn_irflag_read(vp) & VIRF_PGREAD) != 0;
1482 foffset_lock_uio(fp, uio, flags);
1484 if (uio->uio_rw == UIO_READ) {
1485 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1486 uio->uio_offset + uio->uio_resid);
1487 } else if ((fp->f_flag & O_APPEND) != 0 ||
1488 (flags & FOF_OFFSET) == 0) {
1489 /* For appenders, punt and lock the whole range. */
1490 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1492 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1493 uio->uio_offset + uio->uio_resid);
1497 args.kind = VN_IO_FAULT_FOP;
1498 args.args.fop_args.fp = fp;
1499 args.args.fop_args.doio = doio;
1500 args.cred = active_cred;
1501 args.flags = flags | FOF_OFFSET;
1502 error = vn_io_fault1(vp, uio, &args, td);
1504 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1507 vn_rangelock_unlock(vp, rl_cookie);
1508 foffset_unlock_uio(fp, uio, flags);
1513 * Helper function to perform the requested uiomove operation using
1514 * the held pages for io->uio_iov[0].iov_base buffer instead of
1515 * copyin/copyout. Access to the pages with uiomove_fromphys()
1516 * instead of iov_base prevents page faults that could occur due to
1517 * pmap_collect() invalidating the mapping created by
1518 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1519 * object cleanup revoking the write access from page mappings.
1521 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1522 * instead of plain uiomove().
1525 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1527 struct uio transp_uio;
1528 struct iovec transp_iov[1];
1534 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1535 uio->uio_segflg != UIO_USERSPACE)
1536 return (uiomove(data, xfersize, uio));
1538 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1539 transp_iov[0].iov_base = data;
1540 transp_uio.uio_iov = &transp_iov[0];
1541 transp_uio.uio_iovcnt = 1;
1542 if (xfersize > uio->uio_resid)
1543 xfersize = uio->uio_resid;
1544 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1545 transp_uio.uio_offset = 0;
1546 transp_uio.uio_segflg = UIO_SYSSPACE;
1548 * Since transp_iov points to data, and td_ma page array
1549 * corresponds to original uio->uio_iov, we need to invert the
1550 * direction of the i/o operation as passed to
1551 * uiomove_fromphys().
1553 switch (uio->uio_rw) {
1555 transp_uio.uio_rw = UIO_READ;
1558 transp_uio.uio_rw = UIO_WRITE;
1561 transp_uio.uio_td = uio->uio_td;
1562 error = uiomove_fromphys(td->td_ma,
1563 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1564 xfersize, &transp_uio);
1565 adv = xfersize - transp_uio.uio_resid;
1567 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1568 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1570 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1572 td->td_ma_cnt -= pgadv;
1573 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1574 uio->uio_iov->iov_len -= adv;
1575 uio->uio_resid -= adv;
1576 uio->uio_offset += adv;
1581 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1585 vm_offset_t iov_base;
1589 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1590 uio->uio_segflg != UIO_USERSPACE)
1591 return (uiomove_fromphys(ma, offset, xfersize, uio));
1593 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1594 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1595 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1596 switch (uio->uio_rw) {
1598 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1602 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1606 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1608 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1610 td->td_ma_cnt -= pgadv;
1611 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1612 uio->uio_iov->iov_len -= cnt;
1613 uio->uio_resid -= cnt;
1614 uio->uio_offset += cnt;
1619 * File table truncate routine.
1622 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1634 * Lock the whole range for truncation. Otherwise split i/o
1635 * might happen partly before and partly after the truncation.
1637 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1638 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
1641 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1642 AUDIT_ARG_VNODE1(vp);
1643 if (vp->v_type == VDIR) {
1648 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1652 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1656 vn_finished_write(mp);
1658 vn_rangelock_unlock(vp, rl_cookie);
1659 if (error == ERELOOKUP)
1665 * Truncate a file that is already locked.
1668 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1674 error = VOP_ADD_WRITECOUNT(vp, 1);
1677 vattr.va_size = length;
1679 vattr.va_vaflags |= VA_SYNC;
1680 error = VOP_SETATTR(vp, &vattr, cred);
1681 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1687 * File table vnode stat routine.
1690 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred)
1692 struct vnode *vp = fp->f_vnode;
1695 vn_lock(vp, LK_SHARED | LK_RETRY);
1696 error = VOP_STAT(vp, sb, active_cred, fp->f_cred);
1703 * File table vnode ioctl routine.
1706 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1710 struct fiobmap2_arg *bmarg;
1715 switch (vp->v_type) {
1720 error = vn_getsize(vp, &size, active_cred);
1722 *(int *)data = size - fp->f_offset;
1725 bmarg = (struct fiobmap2_arg *)data;
1726 vn_lock(vp, LK_SHARED | LK_RETRY);
1728 error = mac_vnode_check_read(active_cred, fp->f_cred,
1732 error = VOP_BMAP(vp, bmarg->bn, NULL,
1733 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1740 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1745 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1753 * File table vnode poll routine.
1756 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1763 #if defined(MAC) || defined(AUDIT)
1764 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1765 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1766 AUDIT_ARG_VNODE1(vp);
1767 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1773 error = VOP_POLL(vp, events, fp->f_cred, td);
1778 * Acquire the requested lock and then check for validity. LK_RETRY
1779 * permits vn_lock to return doomed vnodes.
1781 static int __noinline
1782 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1786 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1787 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1790 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1792 if ((flags & LK_RETRY) == 0) {
1803 * Nothing to do if we got the lock.
1809 * Interlock was dropped by the call in _vn_lock.
1811 flags &= ~LK_INTERLOCK;
1813 error = VOP_LOCK1(vp, flags, file, line);
1814 } while (error != 0);
1819 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1823 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1824 ("vn_lock: no locktype (%d passed)", flags));
1825 VNPASS(vp->v_holdcnt > 0, vp);
1826 error = VOP_LOCK1(vp, flags, file, line);
1827 if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1828 return (_vn_lock_fallback(vp, flags, file, line, error));
1833 * File table vnode close routine.
1836 vn_closefile(struct file *fp, struct thread *td)
1844 fp->f_ops = &badfileops;
1845 ref = (fp->f_flag & FHASLOCK) != 0;
1847 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1849 if (__predict_false(ref)) {
1850 lf.l_whence = SEEK_SET;
1853 lf.l_type = F_UNLCK;
1854 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1861 * Preparing to start a filesystem write operation. If the operation is
1862 * permitted, then we bump the count of operations in progress and
1863 * proceed. If a suspend request is in progress, we wait until the
1864 * suspension is over, and then proceed.
1867 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1869 struct mount_pcpu *mpcpu;
1872 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1873 vfs_op_thread_enter(mp, mpcpu)) {
1874 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1875 vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1876 vfs_op_thread_exit(mp, mpcpu);
1881 mtx_assert(MNT_MTX(mp), MA_OWNED);
1888 * Check on status of suspension.
1890 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1891 mp->mnt_susp_owner != curthread) {
1893 if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) {
1894 if (flags & V_PCATCH)
1897 mflags |= (PUSER - 1);
1898 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1899 if ((flags & V_NOWAIT) != 0) {
1900 error = EWOULDBLOCK;
1903 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1909 if ((flags & V_XSLEEP) != 0)
1911 mp->mnt_writeopcount++;
1913 if (error != 0 || (flags & V_XSLEEP) != 0)
1920 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1925 KASSERT((flags & ~V_VALID_FLAGS) == 0,
1926 ("%s: invalid flags passed %d\n", __func__, flags));
1930 * If a vnode is provided, get and return the mount point that
1931 * to which it will write.
1934 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1936 if (error != EOPNOTSUPP)
1941 if ((mp = *mpp) == NULL)
1945 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1947 * As long as a vnode is not provided we need to acquire a
1948 * refcount for the provided mountpoint too, in order to
1949 * emulate a vfs_ref().
1954 error = vn_start_write_refed(mp, flags, false);
1955 if (error != 0 && (flags & V_NOWAIT) == 0)
1961 * Secondary suspension. Used by operations such as vop_inactive
1962 * routines that are needed by the higher level functions. These
1963 * are allowed to proceed until all the higher level functions have
1964 * completed (indicated by mnt_writeopcount dropping to zero). At that
1965 * time, these operations are halted until the suspension is over.
1968 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1973 KASSERT((flags & (~V_VALID_FLAGS | V_XSLEEP)) == 0,
1974 ("%s: invalid flags passed %d\n", __func__, flags));
1978 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1980 if (error != EOPNOTSUPP)
1986 * If we are not suspended or have not yet reached suspended
1987 * mode, then let the operation proceed.
1989 if ((mp = *mpp) == NULL)
1993 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1995 * As long as a vnode is not provided we need to acquire a
1996 * refcount for the provided mountpoint too, in order to
1997 * emulate a vfs_ref().
2002 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
2003 mp->mnt_secondary_writes++;
2004 mp->mnt_secondary_accwrites++;
2008 if ((flags & V_NOWAIT) != 0) {
2012 return (EWOULDBLOCK);
2015 * Wait for the suspension to finish.
2018 if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) {
2019 if ((flags & V_PCATCH) != 0)
2022 mflags |= (PUSER - 1) | PDROP;
2023 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, "suspfs", 0);
2032 * Filesystem write operation has completed. If we are suspending and this
2033 * operation is the last one, notify the suspender that the suspension is
2037 vn_finished_write(struct mount *mp)
2039 struct mount_pcpu *mpcpu;
2045 if (vfs_op_thread_enter(mp, mpcpu)) {
2046 vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2047 vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2048 vfs_op_thread_exit(mp, mpcpu);
2053 vfs_assert_mount_counters(mp);
2055 c = --mp->mnt_writeopcount;
2056 if (mp->mnt_vfs_ops == 0) {
2057 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2062 vfs_dump_mount_counters(mp);
2063 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2064 wakeup(&mp->mnt_writeopcount);
2069 * Filesystem secondary write operation has completed. If we are
2070 * suspending and this operation is the last one, notify the suspender
2071 * that the suspension is now in effect.
2074 vn_finished_secondary_write(struct mount *mp)
2080 mp->mnt_secondary_writes--;
2081 if (mp->mnt_secondary_writes < 0)
2082 panic("vn_finished_secondary_write: neg cnt");
2083 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2084 mp->mnt_secondary_writes <= 0)
2085 wakeup(&mp->mnt_secondary_writes);
2090 * Request a filesystem to suspend write operations.
2093 vfs_write_suspend(struct mount *mp, int flags)
2100 vfs_assert_mount_counters(mp);
2101 if (mp->mnt_susp_owner == curthread) {
2102 vfs_op_exit_locked(mp);
2106 while (mp->mnt_kern_flag & MNTK_SUSPEND)
2107 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2110 * Unmount holds a write reference on the mount point. If we
2111 * own busy reference and drain for writers, we deadlock with
2112 * the reference draining in the unmount path. Callers of
2113 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2114 * vfs_busy() reference is owned and caller is not in the
2117 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2118 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2119 vfs_op_exit_locked(mp);
2124 mp->mnt_kern_flag |= MNTK_SUSPEND;
2125 mp->mnt_susp_owner = curthread;
2126 if (mp->mnt_writeopcount > 0)
2127 (void) msleep(&mp->mnt_writeopcount,
2128 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2131 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2132 vfs_write_resume(mp, 0);
2133 /* vfs_write_resume does vfs_op_exit() for us */
2139 * Request a filesystem to resume write operations.
2142 vfs_write_resume(struct mount *mp, int flags)
2146 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2147 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2148 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2150 mp->mnt_susp_owner = NULL;
2151 wakeup(&mp->mnt_writeopcount);
2152 wakeup(&mp->mnt_flag);
2153 curthread->td_pflags &= ~TDP_IGNSUSP;
2154 if ((flags & VR_START_WRITE) != 0) {
2156 mp->mnt_writeopcount++;
2159 if ((flags & VR_NO_SUSPCLR) == 0)
2162 } else if ((flags & VR_START_WRITE) != 0) {
2164 vn_start_write_refed(mp, 0, true);
2171 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2175 vfs_write_suspend_umnt(struct mount *mp)
2179 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2180 ("vfs_write_suspend_umnt: recursed"));
2182 /* dounmount() already called vn_start_write(). */
2184 vn_finished_write(mp);
2185 error = vfs_write_suspend(mp, 0);
2187 vn_start_write(NULL, &mp, V_WAIT);
2191 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2194 vn_start_write(NULL, &mp, V_WAIT);
2196 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2197 wakeup(&mp->mnt_flag);
2199 curthread->td_pflags |= TDP_IGNSUSP;
2204 * Implement kqueues for files by translating it to vnode operation.
2207 vn_kqfilter(struct file *fp, struct knote *kn)
2210 return (VOP_KQFILTER(fp->f_vnode, kn));
2214 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2216 if ((fp->f_flag & FKQALLOWED) == 0)
2218 return (vn_kqfilter(fp, kn));
2222 * Simplified in-kernel wrapper calls for extended attribute access.
2223 * Both calls pass in a NULL credential, authorizing as "kernel" access.
2224 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2227 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2228 const char *attrname, int *buflen, char *buf, struct thread *td)
2234 iov.iov_len = *buflen;
2237 auio.uio_iov = &iov;
2238 auio.uio_iovcnt = 1;
2239 auio.uio_rw = UIO_READ;
2240 auio.uio_segflg = UIO_SYSSPACE;
2242 auio.uio_offset = 0;
2243 auio.uio_resid = *buflen;
2245 if ((ioflg & IO_NODELOCKED) == 0)
2246 vn_lock(vp, LK_SHARED | LK_RETRY);
2248 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2250 /* authorize attribute retrieval as kernel */
2251 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2254 if ((ioflg & IO_NODELOCKED) == 0)
2258 *buflen = *buflen - auio.uio_resid;
2265 * XXX failure mode if partially written?
2268 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2269 const char *attrname, int buflen, char *buf, struct thread *td)
2276 iov.iov_len = buflen;
2279 auio.uio_iov = &iov;
2280 auio.uio_iovcnt = 1;
2281 auio.uio_rw = UIO_WRITE;
2282 auio.uio_segflg = UIO_SYSSPACE;
2284 auio.uio_offset = 0;
2285 auio.uio_resid = buflen;
2287 if ((ioflg & IO_NODELOCKED) == 0) {
2288 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2290 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2293 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2295 /* authorize attribute setting as kernel */
2296 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2298 if ((ioflg & IO_NODELOCKED) == 0) {
2299 vn_finished_write(mp);
2307 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2308 const char *attrname, struct thread *td)
2313 if ((ioflg & IO_NODELOCKED) == 0) {
2314 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2316 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2319 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2321 /* authorize attribute removal as kernel */
2322 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2323 if (error == EOPNOTSUPP)
2324 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2327 if ((ioflg & IO_NODELOCKED) == 0) {
2328 vn_finished_write(mp);
2336 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2340 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2344 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2347 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2352 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2353 int lkflags, struct vnode **rvp)
2358 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2360 ltype = VOP_ISLOCKED(vp);
2361 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2362 ("vn_vget_ino: vp not locked"));
2363 error = vfs_busy(mp, MBF_NOWAIT);
2367 error = vfs_busy(mp, 0);
2368 vn_lock(vp, ltype | LK_RETRY);
2372 if (VN_IS_DOOMED(vp)) {
2378 error = alloc(mp, alloc_arg, lkflags, rvp);
2380 if (error != 0 || *rvp != vp)
2381 vn_lock(vp, ltype | LK_RETRY);
2382 if (VN_IS_DOOMED(vp)) {
2395 vn_send_sigxfsz(struct proc *p)
2398 kern_psignal(p, SIGXFSZ);
2403 vn_rlimit_trunc(u_quad_t size, struct thread *td)
2405 if (size <= lim_cur(td, RLIMIT_FSIZE))
2407 vn_send_sigxfsz(td->td_proc);
2412 vn_rlimit_fsizex1(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2413 bool adj, struct thread *td)
2418 if (vp->v_type != VREG)
2422 * Handle file system maximum file size.
2424 if (maxfsz != 0 && uio->uio_offset + uio->uio_resid > maxfsz) {
2425 if (!adj || uio->uio_offset >= maxfsz)
2427 uio->uio_resid = maxfsz - uio->uio_offset;
2431 * This is kernel write (e.g. vnode_pager) or accounting
2432 * write, ignore limit.
2434 if (td == NULL || (td->td_pflags2 & TDP2_ACCT) != 0)
2438 * Calculate file size limit.
2440 ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2441 lim = __predict_false(ktr_write) ? td->td_ktr_io_lim :
2442 lim_cur(td, RLIMIT_FSIZE);
2445 * Is the limit reached?
2447 if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2451 * Prepared filesystems can handle writes truncated to the
2454 if (adj && (uoff_t)uio->uio_offset < lim) {
2455 uio->uio_resid = lim - (uoff_t)uio->uio_offset;
2459 if (!ktr_write || ktr_filesize_limit_signal)
2460 vn_send_sigxfsz(td->td_proc);
2465 * Helper for VOP_WRITE() implementations, the common code to
2466 * handle maximum supported file size on the filesystem, and
2467 * RLIMIT_FSIZE, except for special writes from accounting subsystem
2470 * For maximum file size (maxfsz argument):
2471 * - return EFBIG if uio_offset is beyond it
2472 * - otherwise, clamp uio_resid if write would extend file beyond maxfsz.
2475 * - return EFBIG and send SIGXFSZ if uio_offset is beyond the limit
2476 * - otherwise, clamp uio_resid if write would extend file beyond limit.
2478 * If clamping occured, the adjustment for uio_resid is stored in
2479 * *resid_adj, to be re-applied by vn_rlimit_fsizex_res() on return
2483 vn_rlimit_fsizex(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2484 ssize_t *resid_adj, struct thread *td)
2490 resid_orig = uio->uio_resid;
2491 adj = resid_adj != NULL;
2492 error = vn_rlimit_fsizex1(vp, uio, maxfsz, adj, td);
2494 *resid_adj = resid_orig - uio->uio_resid;
2499 vn_rlimit_fsizex_res(struct uio *uio, ssize_t resid_adj)
2501 uio->uio_resid += resid_adj;
2505 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2508 return (vn_rlimit_fsizex(vp, __DECONST(struct uio *, uio), 0, NULL,
2513 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2520 vn_lock(vp, LK_SHARED | LK_RETRY);
2521 AUDIT_ARG_VNODE1(vp);
2524 return (setfmode(td, active_cred, vp, mode));
2528 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2535 vn_lock(vp, LK_SHARED | LK_RETRY);
2536 AUDIT_ARG_VNODE1(vp);
2539 return (setfown(td, active_cred, vp, uid, gid));
2543 * Remove pages in the range ["start", "end") from the vnode's VM object. If
2544 * "end" is 0, then the range extends to the end of the object.
2547 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2551 if ((object = vp->v_object) == NULL)
2553 VM_OBJECT_WLOCK(object);
2554 vm_object_page_remove(object, start, end, 0);
2555 VM_OBJECT_WUNLOCK(object);
2559 * Like vn_pages_remove(), but skips invalid pages, which by definition are not
2560 * mapped into any process' address space. Filesystems may use this in
2561 * preference to vn_pages_remove() to avoid blocking on pages busied in
2562 * preparation for a VOP_GETPAGES.
2565 vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2569 if ((object = vp->v_object) == NULL)
2571 VM_OBJECT_WLOCK(object);
2572 vm_object_page_remove(object, start, end, OBJPR_VALIDONLY);
2573 VM_OBJECT_WUNLOCK(object);
2577 vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off,
2586 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2587 ("%s: Wrong command %lu", __func__, cmd));
2588 ASSERT_VOP_ELOCKED(vp, "vn_bmap_seekhole_locked");
2590 if (vp->v_type != VREG) {
2594 error = vn_getsize_locked(vp, &size, cred);
2598 if (noff < 0 || noff >= size) {
2603 /* See the comment in ufs_bmap_seekdata(). */
2604 vnode_pager_clean_sync(vp);
2606 bsize = vp->v_mount->mnt_stat.f_iosize;
2607 for (bn = noff / bsize; noff < size; bn++, noff += bsize -
2609 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2610 if (error == EOPNOTSUPP) {
2614 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2615 (bnp != -1 && cmd == FIOSEEKDATA)) {
2624 /* noff == size. There is an implicit hole at the end of file. */
2625 if (cmd == FIOSEEKDATA)
2634 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2638 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2639 ("%s: Wrong command %lu", __func__, cmd));
2641 if (vn_lock(vp, LK_EXCLUSIVE) != 0)
2643 error = vn_bmap_seekhole_locked(vp, cmd, off, cred);
2649 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2653 off_t foffset, fsize, size;
2656 cred = td->td_ucred;
2658 noneg = (vp->v_type != VCHR);
2660 * Try to dodge locking for common case of querying the offset.
2662 if (whence == L_INCR && offset == 0) {
2663 foffset = foffset_read(fp);
2664 if (__predict_false(foffset < 0 && noneg)) {
2667 td->td_uretoff.tdu_off = foffset;
2670 foffset = foffset_lock(fp, 0);
2676 (offset > 0 && foffset > OFF_MAX - offset))) {
2683 error = vn_getsize(vp, &fsize, cred);
2688 * If the file references a disk device, then fetch
2689 * the media size and use that to determine the ending
2692 if (fsize == 0 && vp->v_type == VCHR &&
2693 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2695 if (noneg && offset > 0 && fsize > OFF_MAX - offset) {
2704 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2705 if (error == ENOTTY)
2709 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2710 if (error == ENOTTY)
2716 if (error == 0 && noneg && offset < 0)
2720 VFS_KNOTE_UNLOCKED(vp, 0);
2721 td->td_uretoff.tdu_off = offset;
2723 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2728 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2734 * Grant permission if the caller is the owner of the file, or
2735 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2736 * on the file. If the time pointer is null, then write
2737 * permission on the file is also sufficient.
2739 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2740 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2741 * will be allowed to set the times [..] to the current
2744 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2745 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2746 error = VOP_ACCESS(vp, VWRITE, cred, td);
2751 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2756 if (fp->f_type == DTYPE_FIFO)
2757 kif->kf_type = KF_TYPE_FIFO;
2759 kif->kf_type = KF_TYPE_VNODE;
2762 FILEDESC_SUNLOCK(fdp);
2763 error = vn_fill_kinfo_vnode(vp, kif);
2765 FILEDESC_SLOCK(fdp);
2770 vn_fill_junk(struct kinfo_file *kif)
2775 * Simulate vn_fullpath returning changing values for a given
2776 * vp during e.g. coredump.
2778 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2779 olen = strlen(kif->kf_path);
2781 strcpy(&kif->kf_path[len - 1], "$");
2783 for (; olen < len; olen++)
2784 strcpy(&kif->kf_path[olen], "A");
2788 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2791 char *fullpath, *freepath;
2794 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2797 error = vn_fullpath(vp, &fullpath, &freepath);
2799 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2801 if (freepath != NULL)
2802 free(freepath, M_TEMP);
2804 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2809 * Retrieve vnode attributes.
2811 va.va_fsid = VNOVAL;
2813 vn_lock(vp, LK_SHARED | LK_RETRY);
2814 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2818 if (va.va_fsid != VNOVAL)
2819 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2821 kif->kf_un.kf_file.kf_file_fsid =
2822 vp->v_mount->mnt_stat.f_fsid.val[0];
2823 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2824 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2825 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2826 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2827 kif->kf_un.kf_file.kf_file_size = va.va_size;
2828 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2829 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2830 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2831 kif->kf_un.kf_file.kf_file_nlink = va.va_nlink;
2836 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2837 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2841 struct pmckern_map_in pkm;
2847 boolean_t writecounted;
2850 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2851 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2853 * POSIX shared-memory objects are defined to have
2854 * kernel persistence, and are not defined to support
2855 * read(2)/write(2) -- or even open(2). Thus, we can
2856 * use MAP_ASYNC to trade on-disk coherence for speed.
2857 * The shm_open(3) library routine turns on the FPOSIXSHM
2858 * flag to request this behavior.
2860 if ((fp->f_flag & FPOSIXSHM) != 0)
2861 flags |= MAP_NOSYNC;
2866 * Ensure that file and memory protections are
2867 * compatible. Note that we only worry about
2868 * writability if mapping is shared; in this case,
2869 * current and max prot are dictated by the open file.
2870 * XXX use the vnode instead? Problem is: what
2871 * credentials do we use for determination? What if
2872 * proc does a setuid?
2875 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2876 maxprot = VM_PROT_NONE;
2877 if ((prot & VM_PROT_EXECUTE) != 0)
2880 maxprot = VM_PROT_EXECUTE;
2881 if ((fp->f_flag & FREAD) != 0)
2882 maxprot |= VM_PROT_READ;
2883 else if ((prot & VM_PROT_READ) != 0)
2887 * If we are sharing potential changes via MAP_SHARED and we
2888 * are trying to get write permission although we opened it
2889 * without asking for it, bail out.
2891 if ((flags & MAP_SHARED) != 0) {
2892 if ((fp->f_flag & FWRITE) != 0)
2893 maxprot |= VM_PROT_WRITE;
2894 else if ((prot & VM_PROT_WRITE) != 0)
2897 maxprot |= VM_PROT_WRITE;
2898 cap_maxprot |= VM_PROT_WRITE;
2900 maxprot &= cap_maxprot;
2903 * For regular files and shared memory, POSIX requires that
2904 * the value of foff be a legitimate offset within the data
2905 * object. In particular, negative offsets are invalid.
2906 * Blocking negative offsets and overflows here avoids
2907 * possible wraparound or user-level access into reserved
2908 * ranges of the data object later. In contrast, POSIX does
2909 * not dictate how offsets are used by device drivers, so in
2910 * the case of a device mapping a negative offset is passed
2917 foff > OFF_MAX - size)
2920 writecounted = FALSE;
2921 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2922 &foff, &object, &writecounted);
2925 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2926 foff, writecounted, td);
2929 * If this mapping was accounted for in the vnode's
2930 * writecount, then undo that now.
2933 vm_pager_release_writecount(object, 0, size);
2934 vm_object_deallocate(object);
2937 /* Inform hwpmc(4) if an executable is being mapped. */
2938 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2939 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2941 pkm.pm_address = (uintptr_t) *addr;
2942 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2950 vn_fsid(struct vnode *vp, struct vattr *va)
2954 f = &vp->v_mount->mnt_stat.f_fsid;
2955 va->va_fsid = (uint32_t)f->val[1];
2956 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2957 va->va_fsid += (uint32_t)f->val[0];
2961 vn_fsync_buf(struct vnode *vp, int waitfor)
2963 struct buf *bp, *nbp;
2966 int error, maxretry;
2969 maxretry = 10000; /* large, arbitrarily chosen */
2971 if (vp->v_type == VCHR) {
2973 mp = vp->v_rdev->si_mountpt;
2980 * MARK/SCAN initialization to avoid infinite loops.
2982 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2983 bp->b_vflags &= ~BV_SCANNED;
2988 * Flush all dirty buffers associated with a vnode.
2991 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2992 if ((bp->b_vflags & BV_SCANNED) != 0)
2994 bp->b_vflags |= BV_SCANNED;
2995 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2996 if (waitfor != MNT_WAIT)
2999 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
3000 BO_LOCKPTR(bo)) != 0) {
3007 KASSERT(bp->b_bufobj == bo,
3008 ("bp %p wrong b_bufobj %p should be %p",
3009 bp, bp->b_bufobj, bo));
3010 if ((bp->b_flags & B_DELWRI) == 0)
3011 panic("fsync: not dirty");
3012 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
3018 if (maxretry < 1000)
3019 pause("dirty", hz < 1000 ? 1 : hz / 1000);
3025 * If synchronous the caller expects us to completely resolve all
3026 * dirty buffers in the system. Wait for in-progress I/O to
3027 * complete (which could include background bitmap writes), then
3028 * retry if dirty blocks still exist.
3030 if (waitfor == MNT_WAIT) {
3031 bufobj_wwait(bo, 0, 0);
3032 if (bo->bo_dirty.bv_cnt > 0) {
3034 * If we are unable to write any of these buffers
3035 * then we fail now rather than trying endlessly
3036 * to write them out.
3038 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
3039 if ((error = bp->b_error) != 0)
3041 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
3042 (error == 0 && --maxretry >= 0))
3050 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
3056 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
3057 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
3058 * to do the actual copy.
3059 * vn_generic_copy_file_range() is factored out, so it can be called
3060 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
3061 * different file systems.
3064 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
3065 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
3066 struct ucred *outcred, struct thread *fsize_td)
3068 struct mount *inmp, *outmp;
3069 struct vnode *invpl, *outvpl;
3074 invpl = outvpl = NULL;
3076 *lenp = 0; /* For error returns. */
3079 /* Do some sanity checks on the arguments. */
3080 if (invp->v_type == VDIR || outvp->v_type == VDIR)
3082 else if (*inoffp < 0 || *outoffp < 0 ||
3083 invp->v_type != VREG || outvp->v_type != VREG)
3088 /* Ensure offset + len does not wrap around. */
3091 if (uval > INT64_MAX)
3092 len = INT64_MAX - *inoffp;
3095 if (uval > INT64_MAX)
3096 len = INT64_MAX - *outoffp;
3100 error = VOP_GETLOWVNODE(invp, &invpl, FREAD);
3103 error = VOP_GETLOWVNODE(outvp, &outvpl, FWRITE);
3107 inmp = invpl->v_mount;
3108 outmp = outvpl->v_mount;
3109 if (inmp == NULL || outmp == NULL)
3113 error = vfs_busy(inmp, 0);
3118 error = vfs_busy(outmp, MBF_NOWAIT);
3121 error = vfs_busy(outmp, 0);
3132 * If the two vnodes are for the same file system type, call
3133 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
3134 * which can handle copies across multiple file system types.
3137 if (inmp == outmp || inmp->mnt_vfc == outmp->mnt_vfc)
3138 error = VOP_COPY_FILE_RANGE(invpl, inoffp, outvpl, outoffp,
3139 lenp, flags, incred, outcred, fsize_td);
3142 if (error == ENOSYS)
3143 error = vn_generic_copy_file_range(invpl, inoffp, outvpl,
3144 outoffp, lenp, flags, incred, outcred, fsize_td);
3159 * Test len bytes of data starting at dat for all bytes == 0.
3160 * Return true if all bytes are zero, false otherwise.
3161 * Expects dat to be well aligned.
3164 mem_iszero(void *dat, int len)
3170 for (p = dat; len > 0; len -= sizeof(*p), p++) {
3171 if (len >= sizeof(*p)) {
3175 cp = (const char *)p;
3176 for (i = 0; i < len; i++, cp++)
3185 * Look for a hole in the output file and, if found, adjust *outoffp
3186 * and *xferp to skip past the hole.
3187 * *xferp is the entire hole length to be written and xfer2 is how many bytes
3188 * to be written as 0's upon return.
3191 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3192 off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3197 if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3198 *dataoffp = *outoffp;
3199 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3202 *holeoffp = *dataoffp;
3203 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3206 if (error != 0 || *holeoffp == *dataoffp) {
3208 * Since outvp is unlocked, it may be possible for
3209 * another thread to do a truncate(), lseek(), write()
3210 * creating a hole at startoff between the above
3211 * VOP_IOCTL() calls, if the other thread does not do
3213 * If that happens, *holeoffp == *dataoffp and finding
3214 * the hole has failed, so disable vn_skip_hole().
3216 *holeoffp = -1; /* Disable use of vn_skip_hole(). */
3219 KASSERT(*dataoffp >= *outoffp,
3220 ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3221 (intmax_t)*dataoffp, (intmax_t)*outoffp));
3222 KASSERT(*holeoffp > *dataoffp,
3223 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3224 (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3228 * If there is a hole before the data starts, advance *outoffp and
3229 * *xferp past the hole.
3231 if (*dataoffp > *outoffp) {
3232 delta = *dataoffp - *outoffp;
3233 if (delta >= *xferp) {
3234 /* Entire *xferp is a hole. */
3241 xfer2 = MIN(xfer2, *xferp);
3245 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3246 * that the write ends at the start of the hole.
3247 * *holeoffp should always be greater than *outoffp, but for the
3248 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3251 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3252 xfer2 = *holeoffp - *outoffp;
3257 * Write an xfer sized chunk to outvp in blksize blocks from dat.
3258 * dat is a maximum of blksize in length and can be written repeatedly in
3260 * If growfile == true, just grow the file via vn_truncate_locked() instead
3261 * of doing actual writes.
3262 * If checkhole == true, a hole is being punched, so skip over any hole
3263 * already in the output file.
3266 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3267 u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3270 off_t dataoff, holeoff, xfer2;
3274 * Loop around doing writes of blksize until write has been completed.
3275 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3276 * done for each iteration, since the xfer argument can be very
3277 * large if there is a large hole to punch in the output file.
3282 xfer2 = MIN(xfer, blksize);
3285 * Punching a hole. Skip writing if there is
3286 * already a hole in the output file.
3288 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3289 &dataoff, &holeoff, cred);
3294 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3299 error = vn_start_write(outvp, &mp, V_WAIT);
3303 error = vn_lock(outvp, LK_EXCLUSIVE);
3305 error = vn_truncate_locked(outvp, outoff + xfer,
3310 error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3312 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3313 outoff, UIO_SYSSPACE, IO_NODELOCKED,
3314 curthread->td_ucred, cred, NULL, curthread);
3321 vn_finished_write(mp);
3322 } while (!growfile && xfer > 0 && error == 0);
3327 * Copy a byte range of one file to another. This function can handle the
3328 * case where invp and outvp are on different file systems.
3329 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3330 * is no better file system specific way to do it.
3333 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3334 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3335 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3339 off_t startoff, endoff, xfer, xfer2;
3341 int error, interrupted;
3342 bool cantseek, readzeros, eof, first, lastblock, holetoeof, sparse;
3343 ssize_t aresid, r = 0;
3344 size_t copylen, len, savlen;
3347 long holein, holeout;
3348 struct timespec curts, endts;
3350 holein = holeout = 0;
3351 savlen = len = *lenp;
3356 error = vn_lock(invp, LK_SHARED);
3359 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3361 error = VOP_GETATTR(invp, &inva, incred);
3362 if (error == 0 && inva.va_size > OFF_MAX)
3369 * Use va_bytes >= va_size as a hint that the file does not have
3370 * sufficient holes to justify the overhead of doing FIOSEEKHOLE.
3371 * This hint does not work well for file systems doing compression
3372 * and may fail when allocations for extended attributes increases
3373 * the value of va_bytes to >= va_size.
3376 if (holein != 0 && inva.va_bytes >= inva.va_size) {
3382 error = vn_start_write(outvp, &mp, V_WAIT);
3384 error = vn_lock(outvp, LK_EXCLUSIVE);
3387 * If fsize_td != NULL, do a vn_rlimit_fsizex() call,
3388 * now that outvp is locked.
3390 if (fsize_td != NULL) {
3393 io.uio_offset = *outoffp;
3395 error = vn_rlimit_fsizex(outvp, &io, 0, &r, fsize_td);
3396 len = savlen = io.uio_resid;
3398 * No need to call vn_rlimit_fsizex_res before return,
3399 * since the uio is local.
3402 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3405 * Holes that are past EOF do not need to be written as a block
3406 * of zero bytes. So, truncate the output file as far as
3407 * possible and then use size to decide if writing 0
3408 * bytes is necessary in the loop below.
3411 error = vn_getsize_locked(outvp, &outsize, outcred);
3412 if (error == 0 && outsize > *outoffp &&
3413 *outoffp <= OFF_MAX - len && outsize <= *outoffp + len &&
3414 *inoffp < inva.va_size &&
3415 *outoffp <= OFF_MAX - (inva.va_size - *inoffp) &&
3416 outsize <= *outoffp + (inva.va_size - *inoffp)) {
3418 error = mac_vnode_check_write(curthread->td_ucred,
3422 error = vn_truncate_locked(outvp, *outoffp,
3430 vn_finished_write(mp);
3434 if (sparse && holein == 0 && holeout > 0) {
3436 * For this special case, the input data will be scanned
3437 * for blocks of all 0 bytes. For these blocks, the
3438 * write can be skipped for the output file to create
3439 * an unallocated region.
3440 * Therefore, use the appropriate size for the output file.
3443 if (blksize <= 512) {
3445 * Use f_iosize, since ZFS reports a _PC_MIN_HOLE_SIZE
3446 * of 512, although it actually only creates
3447 * unallocated regions for blocks >= f_iosize.
3449 blksize = outvp->v_mount->mnt_stat.f_iosize;
3453 * Use the larger of the two f_iosize values. If they are
3454 * not the same size, one will normally be an exact multiple of
3455 * the other, since they are both likely to be a power of 2.
3457 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3458 outvp->v_mount->mnt_stat.f_iosize);
3461 /* Clip to sane limits. */
3464 else if (blksize > maxphys)
3466 dat = malloc(blksize, M_TEMP, M_WAITOK);
3469 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3470 * to find holes. Otherwise, just scan the read block for all 0s
3471 * in the inner loop where the data copying is done.
3472 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3473 * support holes on the server, but do not support FIOSEEKHOLE.
3474 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3475 * that this function should return after 1second with a partial
3478 if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3479 getnanouptime(&endts);
3482 timespecclear(&endts);
3484 holetoeof = eof = false;
3485 while (len > 0 && error == 0 && !eof && interrupted == 0) {
3486 endoff = 0; /* To shut up compilers. */
3492 * Find the next data area. If there is just a hole to EOF,
3493 * FIOSEEKDATA should fail with ENXIO.
3494 * (I do not know if any file system will report a hole to
3495 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3496 * will fail for those file systems.)
3498 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3499 * the code just falls through to the inner copy loop.
3503 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3505 if (error == ENXIO) {
3506 startoff = endoff = inva.va_size;
3507 eof = holetoeof = true;
3511 if (error == 0 && !holetoeof) {
3513 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3516 * Since invp is unlocked, it may be possible for
3517 * another thread to do a truncate(), lseek(), write()
3518 * creating a hole at startoff between the above
3519 * VOP_IOCTL() calls, if the other thread does not do
3521 * If that happens, startoff == endoff and finding
3522 * the hole has failed, so set an error.
3524 if (error == 0 && startoff == endoff)
3525 error = EINVAL; /* Any error. Reset to 0. */
3528 if (startoff > *inoffp) {
3529 /* Found hole before data block. */
3530 xfer = MIN(startoff - *inoffp, len);
3531 if (*outoffp < outsize) {
3532 /* Must write 0s to punch hole. */
3533 xfer2 = MIN(outsize - *outoffp,
3535 memset(dat, 0, MIN(xfer2, blksize));
3536 error = vn_write_outvp(outvp, dat,
3537 *outoffp, xfer2, blksize, false,
3538 holeout > 0, outcred);
3541 if (error == 0 && *outoffp + xfer >
3542 outsize && (xfer == len || holetoeof)) {
3543 /* Grow output file (hole at end). */
3544 error = vn_write_outvp(outvp, dat,
3545 *outoffp, xfer, blksize, true,
3553 interrupted = sig_intr();
3554 if (timespecisset(&endts) &&
3556 getnanouptime(&curts);
3557 if (timespeccmp(&curts,
3565 copylen = MIN(len, endoff - startoff);
3579 * Set first xfer to end at a block boundary, so that
3580 * holes are more likely detected in the loop below via
3581 * the for all bytes 0 method.
3583 xfer -= (*inoffp % blksize);
3587 * Loop copying the data block. If this was our first attempt
3588 * to copy anything, allow a zero-length block so that the VOPs
3589 * get a chance to update metadata, specifically the atime.
3591 while (error == 0 && ((copylen > 0 && !eof) || first) &&
3596 error = vn_lock(invp, LK_SHARED);
3599 error = vn_rdwr(UIO_READ, invp, dat, xfer,
3600 startoff, UIO_SYSSPACE, IO_NODELOCKED,
3601 curthread->td_ucred, incred, &aresid,
3605 if (error == 0 && (xfer == 0 || aresid > 0)) {
3606 /* Stop the copy at EOF on the input file. */
3613 * Skip the write for holes past the initial EOF
3614 * of the output file, unless this is the last
3615 * write of the output file at EOF.
3617 readzeros = cantseek ? mem_iszero(dat, xfer) :
3621 if (!cantseek || *outoffp < outsize ||
3622 lastblock || !readzeros)
3623 error = vn_write_outvp(outvp, dat,
3624 *outoffp, xfer, blksize,
3625 readzeros && lastblock &&
3626 *outoffp >= outsize, false,
3635 interrupted = sig_intr();
3636 if (timespecisset(&endts) &&
3638 getnanouptime(&curts);
3639 if (timespeccmp(&curts,
3651 *lenp = savlen - len;
3657 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3661 off_t olen, ooffset;
3664 int audited_vnode1 = 0;
3668 if (vp->v_type != VREG)
3671 /* Allocating blocks may take a long time, so iterate. */
3678 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
3681 error = vn_lock(vp, LK_EXCLUSIVE);
3683 vn_finished_write(mp);
3687 if (!audited_vnode1) {
3688 AUDIT_ARG_VNODE1(vp);
3693 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3696 error = VOP_ALLOCATE(vp, &offset, &len, 0,
3699 vn_finished_write(mp);
3701 if (olen + ooffset != offset + len) {
3702 panic("offset + len changed from %jx/%jx to %jx/%jx",
3703 ooffset, olen, offset, len);
3705 if (error != 0 || len == 0)
3707 KASSERT(olen > len, ("Iteration did not make progress?"));
3715 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags,
3716 int ioflag, struct ucred *cred, struct ucred *active_cred,
3717 struct ucred *file_cred)
3724 bool audited_vnode1 = false;
3733 if ((ioflag & (IO_NODELOCKED | IO_RANGELOCKED)) == 0)
3734 rl_cookie = vn_rangelock_wlock(vp, off, off + len);
3735 while (len > 0 && error == 0) {
3737 * Try to deallocate the longest range in one pass.
3738 * In case a pass takes too long to be executed, it returns
3739 * partial result. The residue will be proceeded in the next
3743 if ((ioflag & IO_NODELOCKED) == 0) {
3745 if ((error = vn_start_write(vp, &mp,
3746 V_WAIT | V_PCATCH)) != 0)
3748 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
3751 if (!audited_vnode1) {
3752 AUDIT_ARG_VNODE1(vp);
3753 audited_vnode1 = true;
3758 if ((ioflag & IO_NOMACCHECK) == 0)
3759 error = mac_vnode_check_write(active_cred, file_cred,
3763 error = VOP_DEALLOCATE(vp, &off, &len, flags, ioflag,
3766 if ((ioflag & IO_NODELOCKED) == 0) {
3769 vn_finished_write(mp);
3773 if (error == 0 && len != 0)
3777 if (rl_cookie != NULL)
3778 vn_rangelock_unlock(vp, rl_cookie);
3785 * This function is supposed to be used in the situations where the deallocation
3786 * is not triggered by a user request.
3789 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags,
3790 int ioflag, struct ucred *active_cred, struct ucred *file_cred)
3794 if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset ||
3797 if (vp->v_type != VREG)
3800 cred = file_cred != NOCRED ? file_cred : active_cred;
3801 return (vn_deallocate_impl(vp, offset, length, flags, ioflag, cred,
3802 active_cred, file_cred));
3806 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
3807 struct ucred *active_cred, struct thread *td)
3813 KASSERT(cmd == SPACECTL_DEALLOC, ("vn_fspacectl: Invalid cmd"));
3814 KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0,
3815 ("vn_fspacectl: non-zero flags"));
3816 KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset,
3817 ("vn_fspacectl: offset/length overflow or underflow"));
3820 if (vp->v_type != VREG)
3823 ioflag = get_write_ioflag(fp);
3826 case SPACECTL_DEALLOC:
3827 error = vn_deallocate_impl(vp, offset, length, flags, ioflag,
3828 active_cred, active_cred, fp->f_cred);
3831 panic("vn_fspacectl: unknown cmd %d", cmd);
3838 * Keep this assert as long as sizeof(struct dirent) is used as the maximum
3841 _Static_assert(_GENERIC_MAXDIRSIZ == sizeof(struct dirent),
3842 "'struct dirent' size must be a multiple of its alignment "
3843 "(see _GENERIC_DIRLEN())");
3846 * Returns successive directory entries through some caller's provided buffer.
3848 * This function automatically refills the provided buffer with calls to
3849 * VOP_READDIR() (after MAC permission checks).
3851 * 'td' is used for credentials and passed to uiomove(). 'dirbuf' is the
3852 * caller's buffer to fill and 'dirbuflen' its allocated size. 'dirbuf' must
3853 * be properly aligned to access 'struct dirent' structures and 'dirbuflen'
3854 * must be greater than GENERIC_MAXDIRSIZ to avoid VOP_READDIR() returning
3855 * EINVAL (the latter is not a strong guarantee (yet); but EINVAL will always
3856 * be returned if this requirement is not verified). '*dpp' points to the
3857 * current directory entry in the buffer and '*len' contains the remaining
3858 * valid bytes in 'dirbuf' after 'dpp' (including the pointed entry).
3860 * At first call (or when restarting the read), '*len' must have been set to 0,
3861 * '*off' to 0 (or any valid start offset) and '*eofflag' to 0. There are no
3862 * more entries as soon as '*len' is 0 after a call that returned 0. Calling
3863 * again this function after such a condition is considered an error and EINVAL
3864 * will be returned. Other possible error codes are those of VOP_READDIR(),
3865 * EINTEGRITY if the returned entries do not pass coherency tests, or EINVAL
3866 * (bad call). All errors are unrecoverable, i.e., the state ('*len', '*off'
3867 * and '*eofflag') must be re-initialized before a subsequent call. On error
3868 * or at end of directory, '*dpp' is reset to NULL.
3870 * '*len', '*off' and '*eofflag' are internal state the caller should not
3871 * tamper with except as explained above. '*off' is the next directory offset
3872 * to read from to refill the buffer. '*eofflag' is set to 0 or 1 by the last
3873 * internal call to VOP_READDIR() that returned without error, indicating
3874 * whether it reached the end of the directory, and to 2 by this function after
3875 * all entries have been read.
3878 vn_dir_next_dirent(struct vnode *vp, struct thread *td,
3879 char *dirbuf, size_t dirbuflen,
3880 struct dirent **dpp, size_t *len, off_t *off, int *eofflag)
3882 struct dirent *dp = NULL;
3888 ASSERT_VOP_LOCKED(vp, "vnode not locked");
3889 VNASSERT(vp->v_type == VDIR, vp, ("vnode is not a directory"));
3890 MPASS2((uintptr_t)dirbuf < (uintptr_t)dirbuf + dirbuflen,
3891 "Address space overflow");
3893 if (__predict_false(dirbuflen < GENERIC_MAXDIRSIZ)) {
3894 /* Don't take any chances in this case */
3903 * The caller continued to call us after an error (we set dp to
3904 * NULL in a previous iteration). Bail out right now.
3906 if (__predict_false(dp == NULL))
3909 MPASS(*len <= dirbuflen);
3910 MPASS2((uintptr_t)dirbuf <= (uintptr_t)dp &&
3911 (uintptr_t)dp + *len <= (uintptr_t)dirbuf + dirbuflen,
3912 "Filled range not inside buffer");
3914 reclen = dp->d_reclen;
3915 if (reclen >= *len) {
3916 /* End of buffer reached */
3919 dp = (struct dirent *)((char *)dp + reclen);
3927 /* Have to refill. */
3933 /* Nothing more to read. */
3934 *eofflag = 2; /* Remember the caller reached EOF. */
3938 /* The caller didn't test for EOF. */
3943 iov.iov_base = dirbuf;
3944 iov.iov_len = dirbuflen;
3948 uio.uio_offset = *off;
3949 uio.uio_resid = dirbuflen;
3950 uio.uio_segflg = UIO_SYSSPACE;
3951 uio.uio_rw = UIO_READ;
3955 error = mac_vnode_check_readdir(td->td_ucred, vp);
3958 error = VOP_READDIR(vp, &uio, td->td_ucred, eofflag,
3963 *len = dirbuflen - uio.uio_resid;
3964 *off = uio.uio_offset;
3967 /* Sanity check on INVARIANTS. */
3968 MPASS(*eofflag != 0);
3974 * Normalize the flag returned by VOP_READDIR(), since we use 2
3975 * as a sentinel value.
3980 dp = (struct dirent *)dirbuf;
3983 if (__predict_false(*len < GENERIC_MINDIRSIZ ||
3984 dp->d_reclen < GENERIC_MINDIRSIZ)) {
3998 * Checks whether a directory is empty or not.
4000 * If the directory is empty, returns 0, and if it is not, ENOTEMPTY. Other
4001 * values are genuine errors preventing the check.
4004 vn_dir_check_empty(struct vnode *vp)
4006 struct thread *const td = curthread;
4008 size_t dirbuflen, len;
4014 ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
4015 VNPASS(vp->v_type == VDIR, vp);
4017 error = VOP_GETATTR(vp, &va, td->td_ucred);
4021 dirbuflen = max(DEV_BSIZE, GENERIC_MAXDIRSIZ);
4022 if (dirbuflen < va.va_blocksize)
4023 dirbuflen = va.va_blocksize;
4024 dirbuf = malloc(dirbuflen, M_TEMP, M_WAITOK);
4031 error = vn_dir_next_dirent(vp, td, dirbuf, dirbuflen,
4032 &dp, &len, &off, &eofflag);
4043 * Skip whiteouts. Unionfs operates on filesystems only and
4044 * not on hierarchies, so these whiteouts would be shadowed on
4045 * the system hierarchy but not for a union using the
4046 * filesystem of their directories as the upper layer.
4047 * Additionally, unionfs currently transparently exposes
4048 * union-specific metadata of its upper layer, meaning that
4049 * whiteouts can be seen through the union view in empty
4050 * directories. Taking into account these whiteouts would then
4051 * prevent mounting another filesystem on such effectively
4052 * empty directories.
4054 if (dp->d_type == DT_WHT)
4058 * Any file in the directory which is not '.' or '..' indicates
4059 * the directory is not empty.
4061 switch (dp->d_namlen) {
4063 if (dp->d_name[1] != '.') {
4064 /* Can't be '..' (nor '.') */
4070 if (dp->d_name[0] != '.') {
4071 /* Can't be '..' nor '.' */
4084 free(dirbuf, M_TEMP);
4089 static u_long vn_lock_pair_pause_cnt;
4090 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
4091 &vn_lock_pair_pause_cnt, 0,
4092 "Count of vn_lock_pair deadlocks");
4094 u_int vn_lock_pair_pause_max;
4095 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
4096 &vn_lock_pair_pause_max, 0,
4097 "Max ticks for vn_lock_pair deadlock avoidance sleep");
4100 vn_lock_pair_pause(const char *wmesg)
4102 atomic_add_long(&vn_lock_pair_pause_cnt, 1);
4103 pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
4107 * Lock pair of (possibly same) vnodes vp1, vp2, avoiding lock order
4108 * reversal. vp1_locked indicates whether vp1 is locked; if not, vp1
4109 * must be unlocked. Same for vp2 and vp2_locked. One of the vnodes
4112 * The function returns with both vnodes exclusively or shared locked,
4113 * according to corresponding lkflags, and guarantees that it does not
4114 * create lock order reversal with other threads during its execution.
4115 * Both vnodes could be unlocked temporary (and reclaimed).
4117 * If requesting shared locking, locked vnode lock must not be recursed.
4119 * Only one of LK_SHARED and LK_EXCLUSIVE must be specified.
4120 * LK_NODDLKTREAT can be optionally passed.
4122 * If vp1 == vp2, only one, most exclusive, lock is obtained on it.
4125 vn_lock_pair(struct vnode *vp1, bool vp1_locked, int lkflags1,
4126 struct vnode *vp2, bool vp2_locked, int lkflags2)
4130 MPASS((((lkflags1 & LK_SHARED) != 0) ^ ((lkflags1 & LK_EXCLUSIVE) != 0)) ||
4131 (vp1 == NULL && lkflags1 == 0));
4132 MPASS((lkflags1 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0);
4133 MPASS((((lkflags2 & LK_SHARED) != 0) ^ ((lkflags2 & LK_EXCLUSIVE) != 0)) ||
4134 (vp2 == NULL && lkflags2 == 0));
4135 MPASS((lkflags2 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0);
4137 if (vp1 == NULL && vp2 == NULL)
4141 MPASS(vp1_locked == vp2_locked);
4143 /* Select the most exclusive mode for lock. */
4144 if ((lkflags1 & LK_TYPE_MASK) != (lkflags2 & LK_TYPE_MASK))
4145 lkflags1 = (lkflags1 & ~LK_SHARED) | LK_EXCLUSIVE;
4148 ASSERT_VOP_LOCKED(vp1, "vp1");
4150 /* No need to relock if any lock is exclusive. */
4151 if ((vp1->v_vnlock->lock_object.lo_flags &
4155 locked1 = VOP_ISLOCKED(vp1);
4156 if (((lkflags1 & LK_SHARED) != 0 &&
4157 locked1 != LK_EXCLUSIVE) ||
4158 ((lkflags1 & LK_EXCLUSIVE) != 0 &&
4159 locked1 == LK_EXCLUSIVE))
4164 ASSERT_VOP_UNLOCKED(vp1, "vp1");
4165 vn_lock(vp1, lkflags1 | LK_RETRY);
4170 if ((lkflags1 & LK_SHARED) != 0 &&
4171 (vp1->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
4172 lkflags1 = (lkflags1 & ~LK_SHARED) | LK_EXCLUSIVE;
4173 if (vp1_locked && VOP_ISLOCKED(vp1) != LK_EXCLUSIVE) {
4174 ASSERT_VOP_LOCKED(vp1, "vp1");
4175 if ((lkflags1 & LK_EXCLUSIVE) != 0) {
4177 ASSERT_VOP_UNLOCKED(vp1,
4178 "vp1 shared recursed");
4181 } else if (!vp1_locked)
4182 ASSERT_VOP_UNLOCKED(vp1, "vp1");
4188 if ((lkflags2 & LK_SHARED) != 0 &&
4189 (vp2->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
4190 lkflags2 = (lkflags2 & ~LK_SHARED) | LK_EXCLUSIVE;
4191 if (vp2_locked && VOP_ISLOCKED(vp2) != LK_EXCLUSIVE) {
4192 ASSERT_VOP_LOCKED(vp2, "vp2");
4193 if ((lkflags2 & LK_EXCLUSIVE) != 0) {
4195 ASSERT_VOP_UNLOCKED(vp2,
4196 "vp2 shared recursed");
4199 } else if (!vp2_locked)
4200 ASSERT_VOP_UNLOCKED(vp2, "vp2");
4205 if (!vp1_locked && !vp2_locked) {
4206 vn_lock(vp1, lkflags1 | LK_RETRY);
4210 while (!vp1_locked || !vp2_locked) {
4211 if (vp1_locked && vp2 != NULL) {
4213 error = VOP_LOCK1(vp2, lkflags2 | LK_NOWAIT,
4214 __FILE__, __LINE__);
4219 vn_lock_pair_pause("vlp1");
4221 vn_lock(vp2, lkflags2 | LK_RETRY);
4224 if (vp2_locked && vp1 != NULL) {
4226 error = VOP_LOCK1(vp1, lkflags1 | LK_NOWAIT,
4227 __FILE__, __LINE__);
4232 vn_lock_pair_pause("vlp2");
4234 vn_lock(vp1, lkflags1 | LK_RETRY);
4239 if (lkflags1 == LK_EXCLUSIVE)
4240 ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
4242 ASSERT_VOP_LOCKED(vp1, "vp1 ret");
4245 if (lkflags2 == LK_EXCLUSIVE)
4246 ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
4248 ASSERT_VOP_LOCKED(vp2, "vp2 ret");
4253 vn_lktype_write(struct mount *mp, struct vnode *vp)
4255 if (MNT_SHARED_WRITES(mp) ||
4256 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
4258 return (LK_EXCLUSIVE);
4262 vn_cmp(struct file *fp1, struct file *fp2, struct thread *td)
4264 if (fp2->f_type != DTYPE_VNODE)
4266 return (kcmp_cmp((uintptr_t)fp1->f_vnode, (uintptr_t)fp2->f_vnode));