2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1982, 1986, 1989, 1993
5 * The Regents of the University of California. All rights reserved.
6 * (c) UNIX System Laboratories, Inc.
7 * All or some portions of this file are derived from material licensed
8 * to the University of California by American Telephone and Telegraph
9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10 * the permission of UNIX System Laboratories, Inc.
12 * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13 * Copyright (c) 2013, 2014 The FreeBSD Foundation
15 * Portions of this software were developed by Konstantin Belousov
16 * under sponsorship from the FreeBSD Foundation.
18 * Redistribution and use in source and binary forms, with or without
19 * modification, are permitted provided that the following conditions
21 * 1. Redistributions of source code must retain the above copyright
22 * notice, this list of conditions and the following disclaimer.
23 * 2. Redistributions in binary form must reproduce the above copyright
24 * notice, this list of conditions and the following disclaimer in the
25 * documentation and/or other materials provided with the distribution.
26 * 3. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42 * @(#)vfs_vnops.c 8.2 (Berkeley) 1/21/94
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
48 #include "opt_hwpmc_hooks.h"
50 #include <sys/param.h>
51 #include <sys/systm.h>
54 #include <sys/fcntl.h>
61 #include <sys/limits.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
75 #include <sys/sleepqueue.h>
76 #include <sys/sysctl.h>
77 #include <sys/ttycom.h>
79 #include <sys/syslog.h>
80 #include <sys/unistd.h>
82 #include <sys/ktrace.h>
84 #include <security/audit/audit.h>
85 #include <security/mac/mac_framework.h>
88 #include <vm/vm_extern.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_pager.h>
96 #include <sys/pmckern.h>
99 static fo_rdwr_t vn_read;
100 static fo_rdwr_t vn_write;
101 static fo_rdwr_t vn_io_fault;
102 static fo_truncate_t vn_truncate;
103 static fo_ioctl_t vn_ioctl;
104 static fo_poll_t vn_poll;
105 static fo_kqfilter_t vn_kqfilter;
106 static fo_close_t vn_closefile;
107 static fo_mmap_t vn_mmap;
108 static fo_fallocate_t vn_fallocate;
109 static fo_fspacectl_t vn_fspacectl;
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_fspacectl = vn_fspacectl,
128 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
131 const u_int io_hold_cnt = 16;
132 static int vn_io_fault_enable = 1;
133 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
134 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
135 static int vn_io_fault_prefault = 0;
136 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
137 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
138 static int vn_io_pgcache_read_enable = 1;
139 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
140 &vn_io_pgcache_read_enable, 0,
141 "Enable copying from page cache for reads, avoiding fs");
142 static u_long vn_io_faults_cnt;
143 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
144 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
146 static int vfs_allow_read_dir = 0;
147 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
148 &vfs_allow_read_dir, 0,
149 "Enable read(2) of directory by root for filesystems that support it");
152 * Returns true if vn_io_fault mode of handling the i/o request should
156 do_vn_io_fault(struct vnode *vp, struct uio *uio)
160 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
161 (mp = vp->v_mount) != NULL &&
162 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
166 * Structure used to pass arguments to vn_io_fault1(), to do either
167 * file- or vnode-based I/O calls.
169 struct vn_io_fault_args {
177 struct fop_args_tag {
181 struct vop_args_tag {
187 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
188 struct vn_io_fault_args *args, struct thread *td);
191 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
193 struct thread *td = curthread;
195 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
199 open2nameif(int fmode, u_int vn_open_flags)
203 res = ISOPEN | LOCKLEAF;
204 if ((fmode & O_RESOLVE_BENEATH) != 0)
206 if ((fmode & O_EMPTY_PATH) != 0)
208 if ((fmode & FREAD) != 0)
210 if ((fmode & FWRITE) != 0)
212 if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
214 if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
216 if ((vn_open_flags & VN_OPEN_WANTIOCTLCAPS) != 0)
217 res |= WANTIOCTLCAPS;
222 * Common code for vnode open operations via a name lookup.
223 * Lookup the vnode and invoke VOP_CREATE if needed.
224 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
226 * Note that this does NOT free nameidata for the successful case,
227 * due to the NDINIT being done elsewhere.
230 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
231 struct ucred *cred, struct file *fp)
236 struct vattr *vap = &vat;
243 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
244 O_EXCL | O_DIRECTORY) ||
245 (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
247 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
248 ndp->ni_cnd.cn_nameiop = CREATE;
249 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
251 * Set NOCACHE to avoid flushing the cache when
252 * rolling in many files at once.
254 * Set NC_KEEPPOSENTRY to keep positive entries if they already
255 * exist despite NOCACHE.
257 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
258 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
259 ndp->ni_cnd.cn_flags |= FOLLOW;
260 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
262 if ((error = namei(ndp)) != 0)
264 if (ndp->ni_vp == NULL) {
267 vap->va_mode = cmode;
269 vap->va_vaflags |= VA_EXCLUSIVE;
270 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
273 if ((error = vn_start_write(NULL, &mp,
274 V_XSLEEP | V_PCATCH)) != 0)
279 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
280 ndp->ni_cnd.cn_flags |= MAKEENTRY;
282 error = mac_vnode_check_create(cred, ndp->ni_dvp,
286 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
289 if (error == 0 && (fmode & O_EXCL) != 0 &&
290 (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
292 vp->v_iflag |= VI_FOPENING;
296 VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
298 vn_finished_write(mp);
301 if (error == ERELOOKUP) {
309 if (ndp->ni_dvp == ndp->ni_vp)
315 if (fmode & O_EXCL) {
319 if (vp->v_type == VDIR) {
326 ndp->ni_cnd.cn_nameiop = LOOKUP;
327 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
328 ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
330 if ((fmode & FWRITE) == 0)
331 ndp->ni_cnd.cn_flags |= LOCKSHARED;
332 if ((error = namei(ndp)) != 0)
336 error = vn_open_vnode(vp, fmode, cred, curthread, fp);
339 vp->v_iflag &= ~VI_FOPENING;
356 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
359 int error, lock_flags, type;
361 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
362 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
364 KASSERT(fp != NULL, ("open with flock requires fp"));
365 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
368 lock_flags = VOP_ISLOCKED(vp);
371 lf.l_whence = SEEK_SET;
374 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
376 if ((fmode & FNONBLOCK) == 0)
378 if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
380 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
382 fp->f_flag |= FHASLOCK;
384 vn_lock(vp, lock_flags | LK_RETRY);
389 * Common code for vnode open operations once a vnode is located.
390 * Check permissions, and call the VOP_OPEN routine.
393 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
394 struct thread *td, struct file *fp)
399 if (vp->v_type == VLNK) {
400 if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
403 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
407 if ((fmode & O_PATH) == 0) {
408 if (vp->v_type == VSOCK)
410 if ((fmode & (FWRITE | O_TRUNC)) != 0) {
411 if (vp->v_type == VDIR)
415 if ((fmode & FREAD) != 0)
417 if ((fmode & O_APPEND) && (fmode & FWRITE))
420 if ((fmode & O_CREAT) != 0)
424 if ((fmode & FEXEC) != 0)
427 if ((fmode & O_VERIFY) != 0)
429 error = mac_vnode_check_open(cred, vp, accmode);
433 accmode &= ~(VCREAT | VVERIFY);
435 if ((fmode & O_CREAT) == 0 && accmode != 0) {
436 error = VOP_ACCESS(vp, accmode, cred, td);
440 if ((fmode & O_PATH) != 0) {
441 if (vp->v_type != VFIFO && vp->v_type != VSOCK &&
442 VOP_ACCESS(vp, VREAD, cred, td) == 0)
443 fp->f_flag |= FKQALLOWED;
447 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
448 vn_lock(vp, LK_UPGRADE | LK_RETRY);
449 error = VOP_OPEN(vp, fmode, cred, td, fp);
453 error = vn_open_vnode_advlock(vp, fmode, fp);
454 if (error == 0 && (fmode & FWRITE) != 0) {
455 error = VOP_ADD_WRITECOUNT(vp, 1);
457 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
458 __func__, vp, vp->v_writecount);
463 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
464 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
469 * Arrange the call by having fdrop() to use
470 * vn_closefile(). This is to satisfy
471 * filesystems like devfs or tmpfs, which
472 * override fo_close().
474 fp->f_flag |= FOPENFAILED;
476 if (fp->f_ops == &badfileops) {
477 fp->f_type = DTYPE_VNODE;
483 * If there is no fp, due to kernel-mode open,
484 * we can call VOP_CLOSE() now.
486 if ((vp->v_type == VFIFO ||
487 !MNT_EXTENDED_SHARED(vp->v_mount)) &&
488 VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
489 vn_lock(vp, LK_UPGRADE | LK_RETRY);
490 (void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
495 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
501 * Check for write permissions on the specified vnode.
502 * Prototype text segments cannot be written.
506 vn_writechk(struct vnode *vp)
509 ASSERT_VOP_LOCKED(vp, "vn_writechk");
511 * If there's shared text associated with
512 * the vnode, try to free it up once. If
513 * we fail, we can't allow writing.
525 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
526 struct thread *td, bool keep_ref)
529 int error, lock_flags;
531 lock_flags = vp->v_type != VFIFO && MNT_EXTENDED_SHARED(vp->v_mount) ?
532 LK_SHARED : LK_EXCLUSIVE;
534 vn_start_write(vp, &mp, V_WAIT);
535 vn_lock(vp, lock_flags | LK_RETRY);
536 AUDIT_ARG_VNODE1(vp);
537 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
538 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
539 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
540 __func__, vp, vp->v_writecount);
542 error = VOP_CLOSE(vp, flags, file_cred, td);
547 vn_finished_write(mp);
552 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
556 return (vn_close1(vp, flags, file_cred, td, false));
560 * Heuristic to detect sequential operation.
563 sequential_heuristic(struct uio *uio, struct file *fp)
567 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
570 if (fp->f_flag & FRDAHEAD)
571 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
574 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
575 * that the first I/O is normally considered to be slightly
576 * sequential. Seeking to offset 0 doesn't change sequentiality
577 * unless previous seeks have reduced f_seqcount to 0, in which
578 * case offset 0 is not special.
580 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
581 uio->uio_offset == fp->f_nextoff[rw]) {
583 * f_seqcount is in units of fixed-size blocks so that it
584 * depends mainly on the amount of sequential I/O and not
585 * much on the number of sequential I/O's. The fixed size
586 * of 16384 is hard-coded here since it is (not quite) just
587 * a magic size that works well here. This size is more
588 * closely related to the best I/O size for real disks than
589 * to any block size used by software.
591 if (uio->uio_resid >= IO_SEQMAX * 16384)
592 fp->f_seqcount[rw] = IO_SEQMAX;
594 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
595 if (fp->f_seqcount[rw] > IO_SEQMAX)
596 fp->f_seqcount[rw] = IO_SEQMAX;
598 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
601 /* Not sequential. Quickly draw-down sequentiality. */
602 if (fp->f_seqcount[rw] > 1)
603 fp->f_seqcount[rw] = 1;
605 fp->f_seqcount[rw] = 0;
610 * Package up an I/O request on a vnode into a uio and do it.
613 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
614 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
615 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
622 struct vn_io_fault_args args;
623 int error, lock_flags;
625 if (offset < 0 && vp->v_type != VCHR)
627 auio.uio_iov = &aiov;
629 aiov.iov_base = base;
631 auio.uio_resid = len;
632 auio.uio_offset = offset;
633 auio.uio_segflg = segflg;
638 if ((ioflg & IO_NODELOCKED) == 0) {
639 if ((ioflg & IO_RANGELOCKED) == 0) {
640 if (rw == UIO_READ) {
641 rl_cookie = vn_rangelock_rlock(vp, offset,
643 } else if ((ioflg & IO_APPEND) != 0) {
644 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
646 rl_cookie = vn_rangelock_wlock(vp, offset,
652 if (rw == UIO_WRITE) {
653 if (vp->v_type != VCHR &&
654 (error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH))
657 lock_flags = vn_lktype_write(mp, vp);
659 lock_flags = LK_SHARED;
660 vn_lock(vp, lock_flags | LK_RETRY);
664 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
666 if ((ioflg & IO_NOMACCHECK) == 0) {
668 error = mac_vnode_check_read(active_cred, file_cred,
671 error = mac_vnode_check_write(active_cred, file_cred,
676 if (file_cred != NULL)
680 if (do_vn_io_fault(vp, &auio)) {
681 args.kind = VN_IO_FAULT_VOP;
684 args.args.vop_args.vp = vp;
685 error = vn_io_fault1(vp, &auio, &args, td);
686 } else if (rw == UIO_READ) {
687 error = VOP_READ(vp, &auio, ioflg, cred);
688 } else /* if (rw == UIO_WRITE) */ {
689 error = VOP_WRITE(vp, &auio, ioflg, cred);
693 *aresid = auio.uio_resid;
695 if (auio.uio_resid && error == 0)
697 if ((ioflg & IO_NODELOCKED) == 0) {
700 vn_finished_write(mp);
703 if (rl_cookie != NULL)
704 vn_rangelock_unlock(vp, rl_cookie);
709 * Package up an I/O request on a vnode into a uio and do it. The I/O
710 * request is split up into smaller chunks and we try to avoid saturating
711 * the buffer cache while potentially holding a vnode locked, so we
712 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
713 * to give other processes a chance to lock the vnode (either other processes
714 * core'ing the same binary, or unrelated processes scanning the directory).
717 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
718 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
719 struct ucred *file_cred, size_t *aresid, struct thread *td)
728 * Force `offset' to a multiple of MAXBSIZE except possibly
729 * for the first chunk, so that filesystems only need to
730 * write full blocks except possibly for the first and last
733 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
737 if (rw != UIO_READ && vp->v_type == VREG)
740 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
741 ioflg, active_cred, file_cred, &iaresid, td);
742 len -= chunk; /* aresid calc already includes length */
746 base = (char *)base + chunk;
747 kern_yield(PRI_USER);
750 *aresid = len + iaresid;
754 #if OFF_MAX <= LONG_MAX
756 foffset_lock(struct file *fp, int flags)
758 volatile short *flagsp;
762 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
764 if ((flags & FOF_NOLOCK) != 0)
765 return (atomic_load_long(&fp->f_offset));
768 * According to McKusick the vn lock was protecting f_offset here.
769 * It is now protected by the FOFFSET_LOCKED flag.
771 flagsp = &fp->f_vnread_flags;
772 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
773 return (atomic_load_long(&fp->f_offset));
775 sleepq_lock(&fp->f_vnread_flags);
776 state = atomic_load_16(flagsp);
778 if ((state & FOFFSET_LOCKED) == 0) {
779 if (!atomic_fcmpset_acq_16(flagsp, &state,
784 if ((state & FOFFSET_LOCK_WAITING) == 0) {
785 if (!atomic_fcmpset_acq_16(flagsp, &state,
786 state | FOFFSET_LOCK_WAITING))
790 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
791 sleepq_wait(&fp->f_vnread_flags, PUSER -1);
793 sleepq_lock(&fp->f_vnread_flags);
794 state = atomic_load_16(flagsp);
796 res = atomic_load_long(&fp->f_offset);
797 sleepq_release(&fp->f_vnread_flags);
802 foffset_unlock(struct file *fp, off_t val, int flags)
804 volatile short *flagsp;
807 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
809 if ((flags & FOF_NOUPDATE) == 0)
810 atomic_store_long(&fp->f_offset, val);
811 if ((flags & FOF_NEXTOFF_R) != 0)
812 fp->f_nextoff[UIO_READ] = val;
813 if ((flags & FOF_NEXTOFF_W) != 0)
814 fp->f_nextoff[UIO_WRITE] = val;
816 if ((flags & FOF_NOLOCK) != 0)
819 flagsp = &fp->f_vnread_flags;
820 state = atomic_load_16(flagsp);
821 if ((state & FOFFSET_LOCK_WAITING) == 0 &&
822 atomic_cmpset_rel_16(flagsp, state, 0))
825 sleepq_lock(&fp->f_vnread_flags);
826 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
827 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
828 fp->f_vnread_flags = 0;
829 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
830 sleepq_release(&fp->f_vnread_flags);
834 foffset_lock(struct file *fp, int flags)
839 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
841 mtxp = mtx_pool_find(mtxpool_sleep, fp);
843 if ((flags & FOF_NOLOCK) == 0) {
844 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
845 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
846 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
849 fp->f_vnread_flags |= FOFFSET_LOCKED;
857 foffset_unlock(struct file *fp, off_t val, int flags)
861 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
863 mtxp = mtx_pool_find(mtxpool_sleep, fp);
865 if ((flags & FOF_NOUPDATE) == 0)
867 if ((flags & FOF_NEXTOFF_R) != 0)
868 fp->f_nextoff[UIO_READ] = val;
869 if ((flags & FOF_NEXTOFF_W) != 0)
870 fp->f_nextoff[UIO_WRITE] = val;
871 if ((flags & FOF_NOLOCK) == 0) {
872 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
873 ("Lost FOFFSET_LOCKED"));
874 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
875 wakeup(&fp->f_vnread_flags);
876 fp->f_vnread_flags = 0;
883 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
886 if ((flags & FOF_OFFSET) == 0)
887 uio->uio_offset = foffset_lock(fp, flags);
891 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
894 if ((flags & FOF_OFFSET) == 0)
895 foffset_unlock(fp, uio->uio_offset, flags);
899 get_advice(struct file *fp, struct uio *uio)
904 ret = POSIX_FADV_NORMAL;
905 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
908 mtxp = mtx_pool_find(mtxpool_sleep, fp);
910 if (fp->f_advice != NULL &&
911 uio->uio_offset >= fp->f_advice->fa_start &&
912 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
913 ret = fp->f_advice->fa_advice;
919 get_write_ioflag(struct file *fp)
927 mp = atomic_load_ptr(&vp->v_mount);
929 if ((fp->f_flag & O_DIRECT) != 0)
932 if ((fp->f_flag & O_FSYNC) != 0 ||
933 (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
937 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
938 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
939 * fall back to full O_SYNC behavior.
941 if ((fp->f_flag & O_DSYNC) != 0)
942 ioflag |= IO_SYNC | IO_DATASYNC;
948 vn_read_from_obj(struct vnode *vp, struct uio *uio)
951 vm_page_t ma[io_hold_cnt + 2];
956 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
957 obj = atomic_load_ptr(&vp->v_object);
959 return (EJUSTRETURN);
962 * Depends on type stability of vm_objects.
964 vm_object_pip_add(obj, 1);
965 if ((obj->flags & OBJ_DEAD) != 0) {
967 * Note that object might be already reused from the
968 * vnode, and the OBJ_DEAD flag cleared. This is fine,
969 * we recheck for DOOMED vnode state after all pages
970 * are busied, and retract then.
972 * But we check for OBJ_DEAD to ensure that we do not
973 * busy pages while vm_object_terminate_pages()
974 * processes the queue.
980 resid = uio->uio_resid;
981 off = uio->uio_offset;
982 for (i = 0; resid > 0; i++) {
983 MPASS(i < io_hold_cnt + 2);
984 ma[i] = vm_page_grab_unlocked(obj, atop(off),
985 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
991 * Skip invalid pages. Valid mask can be partial only
992 * at EOF, and we clip later.
994 if (vm_page_none_valid(ma[i])) {
995 vm_page_sunbusy(ma[i]);
1003 error = EJUSTRETURN;
1008 * Check VIRF_DOOMED after we busied our pages. Since
1009 * vgonel() terminates the vnode' vm_object, it cannot
1010 * process past pages busied by us.
1012 if (VN_IS_DOOMED(vp)) {
1013 error = EJUSTRETURN;
1017 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1018 if (resid > uio->uio_resid)
1019 resid = uio->uio_resid;
1022 * Unlocked read of vnp_size is safe because truncation cannot
1023 * pass busied page. But we load vnp_size into a local
1024 * variable so that possible concurrent extension does not
1025 * break calculation.
1027 #if defined(__powerpc__) && !defined(__powerpc64__)
1028 vsz = obj->un_pager.vnp.vnp_size;
1030 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1032 if (uio->uio_offset >= vsz) {
1033 error = EJUSTRETURN;
1036 if (uio->uio_offset + resid > vsz)
1037 resid = vsz - uio->uio_offset;
1039 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1042 for (j = 0; j < i; j++) {
1044 vm_page_reference(ma[j]);
1045 vm_page_sunbusy(ma[j]);
1048 vm_object_pip_wakeup(obj);
1051 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1055 * File table vnode read routine.
1058 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1066 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1068 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1071 if (fp->f_flag & FNONBLOCK)
1072 ioflag |= IO_NDELAY;
1073 if (fp->f_flag & O_DIRECT)
1074 ioflag |= IO_DIRECT;
1077 * Try to read from page cache. VIRF_DOOMED check is racy but
1078 * allows us to avoid unneeded work outright.
1080 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1081 (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1082 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1084 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1087 if (error != EJUSTRETURN)
1091 advice = get_advice(fp, uio);
1092 vn_lock(vp, LK_SHARED | LK_RETRY);
1095 case POSIX_FADV_NORMAL:
1096 case POSIX_FADV_SEQUENTIAL:
1097 case POSIX_FADV_NOREUSE:
1098 ioflag |= sequential_heuristic(uio, fp);
1100 case POSIX_FADV_RANDOM:
1101 /* Disable read-ahead for random I/O. */
1104 orig_offset = uio->uio_offset;
1107 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1110 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1111 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1113 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1114 orig_offset != uio->uio_offset)
1116 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1117 * for the backing file after a POSIX_FADV_NOREUSE
1120 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1121 POSIX_FADV_DONTNEED);
1126 * File table vnode write routine.
1129 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1137 bool need_finished_write;
1139 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1141 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1143 if (vp->v_type == VREG)
1146 if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1147 ioflag |= IO_APPEND;
1148 if ((fp->f_flag & FNONBLOCK) != 0)
1149 ioflag |= IO_NDELAY;
1150 ioflag |= get_write_ioflag(fp);
1153 need_finished_write = false;
1154 if (vp->v_type != VCHR) {
1155 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
1158 need_finished_write = true;
1161 advice = get_advice(fp, uio);
1163 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1165 case POSIX_FADV_NORMAL:
1166 case POSIX_FADV_SEQUENTIAL:
1167 case POSIX_FADV_NOREUSE:
1168 ioflag |= sequential_heuristic(uio, fp);
1170 case POSIX_FADV_RANDOM:
1171 /* XXX: Is this correct? */
1174 orig_offset = uio->uio_offset;
1177 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1180 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1181 fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1183 if (need_finished_write)
1184 vn_finished_write(mp);
1185 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1186 orig_offset != uio->uio_offset)
1188 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1189 * for the backing file after a POSIX_FADV_NOREUSE
1192 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1193 POSIX_FADV_DONTNEED);
1199 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1200 * prevent the following deadlock:
1202 * Assume that the thread A reads from the vnode vp1 into userspace
1203 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
1204 * currently not resident, then system ends up with the call chain
1205 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1206 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1207 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1208 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1209 * backed by the pages of vnode vp1, and some page in buf2 is not
1210 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1212 * To prevent the lock order reversal and deadlock, vn_io_fault() does
1213 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1214 * Instead, it first tries to do the whole range i/o with pagefaults
1215 * disabled. If all pages in the i/o buffer are resident and mapped,
1216 * VOP will succeed (ignoring the genuine filesystem errors).
1217 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1218 * i/o in chunks, with all pages in the chunk prefaulted and held
1219 * using vm_fault_quick_hold_pages().
1221 * Filesystems using this deadlock avoidance scheme should use the
1222 * array of the held pages from uio, saved in the curthread->td_ma,
1223 * instead of doing uiomove(). A helper function
1224 * vn_io_fault_uiomove() converts uiomove request into
1225 * uiomove_fromphys() over td_ma array.
1227 * Since vnode locks do not cover the whole i/o anymore, rangelocks
1228 * make the current i/o request atomic with respect to other i/os and
1233 * Decode vn_io_fault_args and perform the corresponding i/o.
1236 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1242 save = vm_fault_disable_pagefaults();
1243 switch (args->kind) {
1244 case VN_IO_FAULT_FOP:
1245 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1246 uio, args->cred, args->flags, td);
1248 case VN_IO_FAULT_VOP:
1249 if (uio->uio_rw == UIO_READ) {
1250 error = VOP_READ(args->args.vop_args.vp, uio,
1251 args->flags, args->cred);
1252 } else if (uio->uio_rw == UIO_WRITE) {
1253 error = VOP_WRITE(args->args.vop_args.vp, uio,
1254 args->flags, args->cred);
1258 panic("vn_io_fault_doio: unknown kind of io %d %d",
1259 args->kind, uio->uio_rw);
1261 vm_fault_enable_pagefaults(save);
1266 vn_io_fault_touch(char *base, const struct uio *uio)
1271 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1277 vn_io_fault_prefault_user(const struct uio *uio)
1280 const struct iovec *iov;
1285 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1286 ("vn_io_fault_prefault userspace"));
1290 resid = uio->uio_resid;
1291 base = iov->iov_base;
1294 error = vn_io_fault_touch(base, uio);
1297 if (len < PAGE_SIZE) {
1299 error = vn_io_fault_touch(base + len - 1, uio);
1304 if (++i >= uio->uio_iovcnt)
1306 iov = uio->uio_iov + i;
1307 base = iov->iov_base;
1319 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1320 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1321 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1322 * into args and call vn_io_fault1() to handle faults during the user
1323 * mode buffer accesses.
1326 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1329 vm_page_t ma[io_hold_cnt + 2];
1330 struct uio *uio_clone, short_uio;
1331 struct iovec short_iovec[1];
1332 vm_page_t *prev_td_ma;
1334 vm_offset_t addr, end;
1337 int error, cnt, saveheld, prev_td_ma_cnt;
1339 if (vn_io_fault_prefault) {
1340 error = vn_io_fault_prefault_user(uio);
1342 return (error); /* Or ignore ? */
1345 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1348 * The UFS follows IO_UNIT directive and replays back both
1349 * uio_offset and uio_resid if an error is encountered during the
1350 * operation. But, since the iovec may be already advanced,
1351 * uio is still in an inconsistent state.
1353 * Cache a copy of the original uio, which is advanced to the redo
1354 * point using UIO_NOCOPY below.
1356 uio_clone = cloneuio(uio);
1357 resid = uio->uio_resid;
1359 short_uio.uio_segflg = UIO_USERSPACE;
1360 short_uio.uio_rw = uio->uio_rw;
1361 short_uio.uio_td = uio->uio_td;
1363 error = vn_io_fault_doio(args, uio, td);
1364 if (error != EFAULT)
1367 atomic_add_long(&vn_io_faults_cnt, 1);
1368 uio_clone->uio_segflg = UIO_NOCOPY;
1369 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1370 uio_clone->uio_segflg = uio->uio_segflg;
1372 saveheld = curthread_pflags_set(TDP_UIOHELD);
1373 prev_td_ma = td->td_ma;
1374 prev_td_ma_cnt = td->td_ma_cnt;
1376 while (uio_clone->uio_resid != 0) {
1377 len = uio_clone->uio_iov->iov_len;
1379 KASSERT(uio_clone->uio_iovcnt >= 1,
1380 ("iovcnt underflow"));
1381 uio_clone->uio_iov++;
1382 uio_clone->uio_iovcnt--;
1385 if (len > ptoa(io_hold_cnt))
1386 len = ptoa(io_hold_cnt);
1387 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1388 end = round_page(addr + len);
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 | V_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)
1673 struct vnode *vp = fp->f_vnode;
1676 vn_lock(vp, LK_SHARED | LK_RETRY);
1677 error = VOP_STAT(vp, sb, active_cred, fp->f_cred);
1684 * File table vnode ioctl routine.
1687 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1691 struct fiobmap2_arg *bmarg;
1696 switch (vp->v_type) {
1701 error = vn_getsize(vp, &size, active_cred);
1703 *(int *)data = size - fp->f_offset;
1706 bmarg = (struct fiobmap2_arg *)data;
1707 vn_lock(vp, LK_SHARED | LK_RETRY);
1709 error = mac_vnode_check_read(active_cred, fp->f_cred,
1713 error = VOP_BMAP(vp, bmarg->bn, NULL,
1714 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1721 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1726 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1734 * File table vnode poll routine.
1737 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1744 #if defined(MAC) || defined(AUDIT)
1745 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1746 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1747 AUDIT_ARG_VNODE1(vp);
1748 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1754 error = VOP_POLL(vp, events, fp->f_cred, td);
1759 * Acquire the requested lock and then check for validity. LK_RETRY
1760 * permits vn_lock to return doomed vnodes.
1762 static int __noinline
1763 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1767 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1768 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1771 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1773 if ((flags & LK_RETRY) == 0) {
1784 * Nothing to do if we got the lock.
1790 * Interlock was dropped by the call in _vn_lock.
1792 flags &= ~LK_INTERLOCK;
1794 error = VOP_LOCK1(vp, flags, file, line);
1795 } while (error != 0);
1800 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1804 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1805 ("vn_lock: no locktype (%d passed)", flags));
1806 VNPASS(vp->v_holdcnt > 0, vp);
1807 error = VOP_LOCK1(vp, flags, file, line);
1808 if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1809 return (_vn_lock_fallback(vp, flags, file, line, error));
1814 * File table vnode close routine.
1817 vn_closefile(struct file *fp, struct thread *td)
1825 fp->f_ops = &badfileops;
1826 ref = (fp->f_flag & FHASLOCK) != 0;
1828 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1830 if (__predict_false(ref)) {
1831 lf.l_whence = SEEK_SET;
1834 lf.l_type = F_UNLCK;
1835 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1842 * Preparing to start a filesystem write operation. If the operation is
1843 * permitted, then we bump the count of operations in progress and
1844 * proceed. If a suspend request is in progress, we wait until the
1845 * suspension is over, and then proceed.
1848 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1850 struct mount_pcpu *mpcpu;
1853 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1854 vfs_op_thread_enter(mp, mpcpu)) {
1855 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1856 vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1857 vfs_op_thread_exit(mp, mpcpu);
1862 mtx_assert(MNT_MTX(mp), MA_OWNED);
1869 * Check on status of suspension.
1871 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1872 mp->mnt_susp_owner != curthread) {
1874 if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) {
1875 if (flags & V_PCATCH)
1878 mflags |= (PUSER - 1);
1879 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1880 if ((flags & V_NOWAIT) != 0) {
1881 error = EWOULDBLOCK;
1884 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1890 if ((flags & V_XSLEEP) != 0)
1892 mp->mnt_writeopcount++;
1894 if (error != 0 || (flags & V_XSLEEP) != 0)
1901 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1906 KASSERT((flags & ~V_VALID_FLAGS) == 0,
1907 ("%s: invalid flags passed %d\n", __func__, flags));
1911 * If a vnode is provided, get and return the mount point that
1912 * to which it will write.
1915 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1917 if (error != EOPNOTSUPP)
1922 if ((mp = *mpp) == NULL)
1926 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1928 * As long as a vnode is not provided we need to acquire a
1929 * refcount for the provided mountpoint too, in order to
1930 * emulate a vfs_ref().
1935 error = vn_start_write_refed(mp, flags, false);
1936 if (error != 0 && (flags & V_NOWAIT) == 0)
1942 * Secondary suspension. Used by operations such as vop_inactive
1943 * routines that are needed by the higher level functions. These
1944 * are allowed to proceed until all the higher level functions have
1945 * completed (indicated by mnt_writeopcount dropping to zero). At that
1946 * time, these operations are halted until the suspension is over.
1949 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1954 KASSERT((flags & (~V_VALID_FLAGS | V_XSLEEP)) == 0,
1955 ("%s: invalid flags passed %d\n", __func__, flags));
1959 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1961 if (error != EOPNOTSUPP)
1967 * If we are not suspended or have not yet reached suspended
1968 * mode, then let the operation proceed.
1970 if ((mp = *mpp) == NULL)
1974 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1976 * As long as a vnode is not provided we need to acquire a
1977 * refcount for the provided mountpoint too, in order to
1978 * emulate a vfs_ref().
1983 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1984 mp->mnt_secondary_writes++;
1985 mp->mnt_secondary_accwrites++;
1989 if ((flags & V_NOWAIT) != 0) {
1993 return (EWOULDBLOCK);
1996 * Wait for the suspension to finish.
1999 if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) {
2000 if ((flags & V_PCATCH) != 0)
2003 mflags |= (PUSER - 1) | PDROP;
2004 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, "suspfs", 0);
2013 * Filesystem write operation has completed. If we are suspending and this
2014 * operation is the last one, notify the suspender that the suspension is
2018 vn_finished_write(struct mount *mp)
2020 struct mount_pcpu *mpcpu;
2026 if (vfs_op_thread_enter(mp, mpcpu)) {
2027 vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2028 vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2029 vfs_op_thread_exit(mp, mpcpu);
2034 vfs_assert_mount_counters(mp);
2036 c = --mp->mnt_writeopcount;
2037 if (mp->mnt_vfs_ops == 0) {
2038 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2043 vfs_dump_mount_counters(mp);
2044 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2045 wakeup(&mp->mnt_writeopcount);
2050 * Filesystem secondary write operation has completed. If we are
2051 * suspending and this operation is the last one, notify the suspender
2052 * that the suspension is now in effect.
2055 vn_finished_secondary_write(struct mount *mp)
2061 mp->mnt_secondary_writes--;
2062 if (mp->mnt_secondary_writes < 0)
2063 panic("vn_finished_secondary_write: neg cnt");
2064 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2065 mp->mnt_secondary_writes <= 0)
2066 wakeup(&mp->mnt_secondary_writes);
2071 * Request a filesystem to suspend write operations.
2074 vfs_write_suspend(struct mount *mp, int flags)
2081 vfs_assert_mount_counters(mp);
2082 if (mp->mnt_susp_owner == curthread) {
2083 vfs_op_exit_locked(mp);
2087 while (mp->mnt_kern_flag & MNTK_SUSPEND)
2088 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2091 * Unmount holds a write reference on the mount point. If we
2092 * own busy reference and drain for writers, we deadlock with
2093 * the reference draining in the unmount path. Callers of
2094 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2095 * vfs_busy() reference is owned and caller is not in the
2098 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2099 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2100 vfs_op_exit_locked(mp);
2105 mp->mnt_kern_flag |= MNTK_SUSPEND;
2106 mp->mnt_susp_owner = curthread;
2107 if (mp->mnt_writeopcount > 0)
2108 (void) msleep(&mp->mnt_writeopcount,
2109 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2112 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2113 vfs_write_resume(mp, 0);
2114 /* vfs_write_resume does vfs_op_exit() for us */
2120 * Request a filesystem to resume write operations.
2123 vfs_write_resume(struct mount *mp, int flags)
2127 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2128 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2129 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2131 mp->mnt_susp_owner = NULL;
2132 wakeup(&mp->mnt_writeopcount);
2133 wakeup(&mp->mnt_flag);
2134 curthread->td_pflags &= ~TDP_IGNSUSP;
2135 if ((flags & VR_START_WRITE) != 0) {
2137 mp->mnt_writeopcount++;
2140 if ((flags & VR_NO_SUSPCLR) == 0)
2143 } else if ((flags & VR_START_WRITE) != 0) {
2145 vn_start_write_refed(mp, 0, true);
2152 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2156 vfs_write_suspend_umnt(struct mount *mp)
2160 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2161 ("vfs_write_suspend_umnt: recursed"));
2163 /* dounmount() already called vn_start_write(). */
2165 vn_finished_write(mp);
2166 error = vfs_write_suspend(mp, 0);
2168 vn_start_write(NULL, &mp, V_WAIT);
2172 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2175 vn_start_write(NULL, &mp, V_WAIT);
2177 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2178 wakeup(&mp->mnt_flag);
2180 curthread->td_pflags |= TDP_IGNSUSP;
2185 * Implement kqueues for files by translating it to vnode operation.
2188 vn_kqfilter(struct file *fp, struct knote *kn)
2191 return (VOP_KQFILTER(fp->f_vnode, kn));
2195 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2197 if ((fp->f_flag & FKQALLOWED) == 0)
2199 return (vn_kqfilter(fp, kn));
2203 * Simplified in-kernel wrapper calls for extended attribute access.
2204 * Both calls pass in a NULL credential, authorizing as "kernel" access.
2205 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2208 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2209 const char *attrname, int *buflen, char *buf, struct thread *td)
2215 iov.iov_len = *buflen;
2218 auio.uio_iov = &iov;
2219 auio.uio_iovcnt = 1;
2220 auio.uio_rw = UIO_READ;
2221 auio.uio_segflg = UIO_SYSSPACE;
2223 auio.uio_offset = 0;
2224 auio.uio_resid = *buflen;
2226 if ((ioflg & IO_NODELOCKED) == 0)
2227 vn_lock(vp, LK_SHARED | LK_RETRY);
2229 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2231 /* authorize attribute retrieval as kernel */
2232 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2235 if ((ioflg & IO_NODELOCKED) == 0)
2239 *buflen = *buflen - auio.uio_resid;
2246 * XXX failure mode if partially written?
2249 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2250 const char *attrname, int buflen, char *buf, struct thread *td)
2257 iov.iov_len = buflen;
2260 auio.uio_iov = &iov;
2261 auio.uio_iovcnt = 1;
2262 auio.uio_rw = UIO_WRITE;
2263 auio.uio_segflg = UIO_SYSSPACE;
2265 auio.uio_offset = 0;
2266 auio.uio_resid = buflen;
2268 if ((ioflg & IO_NODELOCKED) == 0) {
2269 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2271 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2274 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2276 /* authorize attribute setting as kernel */
2277 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2279 if ((ioflg & IO_NODELOCKED) == 0) {
2280 vn_finished_write(mp);
2288 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2289 const char *attrname, struct thread *td)
2294 if ((ioflg & IO_NODELOCKED) == 0) {
2295 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2297 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2300 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2302 /* authorize attribute removal as kernel */
2303 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2304 if (error == EOPNOTSUPP)
2305 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2308 if ((ioflg & IO_NODELOCKED) == 0) {
2309 vn_finished_write(mp);
2317 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2321 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2325 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2328 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2333 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2334 int lkflags, struct vnode **rvp)
2339 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2341 ltype = VOP_ISLOCKED(vp);
2342 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2343 ("vn_vget_ino: vp not locked"));
2344 error = vfs_busy(mp, MBF_NOWAIT);
2348 error = vfs_busy(mp, 0);
2349 vn_lock(vp, ltype | LK_RETRY);
2353 if (VN_IS_DOOMED(vp)) {
2359 error = alloc(mp, alloc_arg, lkflags, rvp);
2361 if (error != 0 || *rvp != vp)
2362 vn_lock(vp, ltype | LK_RETRY);
2363 if (VN_IS_DOOMED(vp)) {
2376 vn_send_sigxfsz(struct proc *p)
2379 kern_psignal(p, SIGXFSZ);
2384 vn_rlimit_trunc(u_quad_t size, struct thread *td)
2386 if (size <= lim_cur(td, RLIMIT_FSIZE))
2388 vn_send_sigxfsz(td->td_proc);
2393 vn_rlimit_fsizex1(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2394 bool adj, struct thread *td)
2399 if (vp->v_type != VREG)
2403 * Handle file system maximum file size.
2405 if (maxfsz != 0 && uio->uio_offset + uio->uio_resid > maxfsz) {
2406 if (!adj || uio->uio_offset >= maxfsz)
2408 uio->uio_resid = maxfsz - uio->uio_offset;
2412 * This is kernel write (e.g. vnode_pager) or accounting
2413 * write, ignore limit.
2415 if (td == NULL || (td->td_pflags2 & TDP2_ACCT) != 0)
2419 * Calculate file size limit.
2421 ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2422 lim = __predict_false(ktr_write) ? td->td_ktr_io_lim :
2423 lim_cur(td, RLIMIT_FSIZE);
2426 * Is the limit reached?
2428 if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2432 * Prepared filesystems can handle writes truncated to the
2435 if (adj && (uoff_t)uio->uio_offset < lim) {
2436 uio->uio_resid = lim - (uoff_t)uio->uio_offset;
2440 if (!ktr_write || ktr_filesize_limit_signal)
2441 vn_send_sigxfsz(td->td_proc);
2446 * Helper for VOP_WRITE() implementations, the common code to
2447 * handle maximum supported file size on the filesystem, and
2448 * RLIMIT_FSIZE, except for special writes from accounting subsystem
2451 * For maximum file size (maxfsz argument):
2452 * - return EFBIG if uio_offset is beyond it
2453 * - otherwise, clamp uio_resid if write would extend file beyond maxfsz.
2456 * - return EFBIG and send SIGXFSZ if uio_offset is beyond the limit
2457 * - otherwise, clamp uio_resid if write would extend file beyond limit.
2459 * If clamping occured, the adjustment for uio_resid is stored in
2460 * *resid_adj, to be re-applied by vn_rlimit_fsizex_res() on return
2464 vn_rlimit_fsizex(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2465 ssize_t *resid_adj, struct thread *td)
2471 resid_orig = uio->uio_resid;
2472 adj = resid_adj != NULL;
2473 error = vn_rlimit_fsizex1(vp, uio, maxfsz, adj, td);
2475 *resid_adj = resid_orig - uio->uio_resid;
2480 vn_rlimit_fsizex_res(struct uio *uio, ssize_t resid_adj)
2482 uio->uio_resid += resid_adj;
2486 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2489 return (vn_rlimit_fsizex(vp, __DECONST(struct uio *, uio), 0, NULL,
2494 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2501 vn_lock(vp, LK_SHARED | LK_RETRY);
2502 AUDIT_ARG_VNODE1(vp);
2505 return (setfmode(td, active_cred, vp, mode));
2509 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2516 vn_lock(vp, LK_SHARED | LK_RETRY);
2517 AUDIT_ARG_VNODE1(vp);
2520 return (setfown(td, active_cred, vp, uid, gid));
2524 * Remove pages in the range ["start", "end") from the vnode's VM object. If
2525 * "end" is 0, then the range extends to the end of the object.
2528 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2532 if ((object = vp->v_object) == NULL)
2534 VM_OBJECT_WLOCK(object);
2535 vm_object_page_remove(object, start, end, 0);
2536 VM_OBJECT_WUNLOCK(object);
2540 * Like vn_pages_remove(), but skips invalid pages, which by definition are not
2541 * mapped into any process' address space. Filesystems may use this in
2542 * preference to vn_pages_remove() to avoid blocking on pages busied in
2543 * preparation for a VOP_GETPAGES.
2546 vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2550 if ((object = vp->v_object) == NULL)
2552 VM_OBJECT_WLOCK(object);
2553 vm_object_page_remove(object, start, end, OBJPR_VALIDONLY);
2554 VM_OBJECT_WUNLOCK(object);
2558 vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off,
2568 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2569 ("%s: Wrong command %lu", __func__, cmd));
2570 ASSERT_VOP_ELOCKED(vp, "vn_bmap_seekhole_locked");
2572 if (vp->v_type != VREG) {
2576 error = vn_getsize_locked(vp, &size, cred);
2580 if (noff < 0 || noff >= size) {
2585 /* See the comment in ufs_bmap_seekdata(). */
2588 VM_OBJECT_WLOCK(obj);
2589 vm_object_page_clean(obj, 0, 0, OBJPC_SYNC);
2590 VM_OBJECT_WUNLOCK(obj);
2593 bsize = vp->v_mount->mnt_stat.f_iosize;
2594 for (bn = noff / bsize; noff < size; bn++, noff += bsize -
2596 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2597 if (error == EOPNOTSUPP) {
2601 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2602 (bnp != -1 && cmd == FIOSEEKDATA)) {
2611 /* noff == size. There is an implicit hole at the end of file. */
2612 if (cmd == FIOSEEKDATA)
2621 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2625 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2626 ("%s: Wrong command %lu", __func__, cmd));
2628 if (vn_lock(vp, LK_EXCLUSIVE) != 0)
2630 error = vn_bmap_seekhole_locked(vp, cmd, off, cred);
2636 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2640 off_t foffset, fsize, size;
2643 cred = td->td_ucred;
2645 foffset = foffset_lock(fp, 0);
2646 noneg = (vp->v_type != VCHR);
2652 (offset > 0 && foffset > OFF_MAX - offset))) {
2659 error = vn_getsize(vp, &fsize, cred);
2664 * If the file references a disk device, then fetch
2665 * the media size and use that to determine the ending
2668 if (fsize == 0 && vp->v_type == VCHR &&
2669 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2671 if (noneg && offset > 0 && fsize > OFF_MAX - offset) {
2680 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2681 if (error == ENOTTY)
2685 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2686 if (error == ENOTTY)
2692 if (error == 0 && noneg && offset < 0)
2696 VFS_KNOTE_UNLOCKED(vp, 0);
2697 td->td_uretoff.tdu_off = offset;
2699 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2704 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2710 * Grant permission if the caller is the owner of the file, or
2711 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2712 * on the file. If the time pointer is null, then write
2713 * permission on the file is also sufficient.
2715 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2716 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2717 * will be allowed to set the times [..] to the current
2720 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2721 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2722 error = VOP_ACCESS(vp, VWRITE, cred, td);
2727 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2732 if (fp->f_type == DTYPE_FIFO)
2733 kif->kf_type = KF_TYPE_FIFO;
2735 kif->kf_type = KF_TYPE_VNODE;
2738 FILEDESC_SUNLOCK(fdp);
2739 error = vn_fill_kinfo_vnode(vp, kif);
2741 FILEDESC_SLOCK(fdp);
2746 vn_fill_junk(struct kinfo_file *kif)
2751 * Simulate vn_fullpath returning changing values for a given
2752 * vp during e.g. coredump.
2754 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2755 olen = strlen(kif->kf_path);
2757 strcpy(&kif->kf_path[len - 1], "$");
2759 for (; olen < len; olen++)
2760 strcpy(&kif->kf_path[olen], "A");
2764 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2767 char *fullpath, *freepath;
2770 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2773 error = vn_fullpath(vp, &fullpath, &freepath);
2775 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2777 if (freepath != NULL)
2778 free(freepath, M_TEMP);
2780 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2785 * Retrieve vnode attributes.
2787 va.va_fsid = VNOVAL;
2789 vn_lock(vp, LK_SHARED | LK_RETRY);
2790 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2794 if (va.va_fsid != VNOVAL)
2795 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2797 kif->kf_un.kf_file.kf_file_fsid =
2798 vp->v_mount->mnt_stat.f_fsid.val[0];
2799 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2800 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2801 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2802 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2803 kif->kf_un.kf_file.kf_file_size = va.va_size;
2804 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2805 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2806 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2807 kif->kf_un.kf_file.kf_file_nlink = va.va_nlink;
2812 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2813 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2817 struct pmckern_map_in pkm;
2823 boolean_t writecounted;
2826 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2827 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2829 * POSIX shared-memory objects are defined to have
2830 * kernel persistence, and are not defined to support
2831 * read(2)/write(2) -- or even open(2). Thus, we can
2832 * use MAP_ASYNC to trade on-disk coherence for speed.
2833 * The shm_open(3) library routine turns on the FPOSIXSHM
2834 * flag to request this behavior.
2836 if ((fp->f_flag & FPOSIXSHM) != 0)
2837 flags |= MAP_NOSYNC;
2842 * Ensure that file and memory protections are
2843 * compatible. Note that we only worry about
2844 * writability if mapping is shared; in this case,
2845 * current and max prot are dictated by the open file.
2846 * XXX use the vnode instead? Problem is: what
2847 * credentials do we use for determination? What if
2848 * proc does a setuid?
2851 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2852 maxprot = VM_PROT_NONE;
2853 if ((prot & VM_PROT_EXECUTE) != 0)
2856 maxprot = VM_PROT_EXECUTE;
2857 if ((fp->f_flag & FREAD) != 0)
2858 maxprot |= VM_PROT_READ;
2859 else if ((prot & VM_PROT_READ) != 0)
2863 * If we are sharing potential changes via MAP_SHARED and we
2864 * are trying to get write permission although we opened it
2865 * without asking for it, bail out.
2867 if ((flags & MAP_SHARED) != 0) {
2868 if ((fp->f_flag & FWRITE) != 0)
2869 maxprot |= VM_PROT_WRITE;
2870 else if ((prot & VM_PROT_WRITE) != 0)
2873 maxprot |= VM_PROT_WRITE;
2874 cap_maxprot |= VM_PROT_WRITE;
2876 maxprot &= cap_maxprot;
2879 * For regular files and shared memory, POSIX requires that
2880 * the value of foff be a legitimate offset within the data
2881 * object. In particular, negative offsets are invalid.
2882 * Blocking negative offsets and overflows here avoids
2883 * possible wraparound or user-level access into reserved
2884 * ranges of the data object later. In contrast, POSIX does
2885 * not dictate how offsets are used by device drivers, so in
2886 * the case of a device mapping a negative offset is passed
2893 foff > OFF_MAX - size)
2896 writecounted = FALSE;
2897 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2898 &foff, &object, &writecounted);
2901 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2902 foff, writecounted, td);
2905 * If this mapping was accounted for in the vnode's
2906 * writecount, then undo that now.
2909 vm_pager_release_writecount(object, 0, size);
2910 vm_object_deallocate(object);
2913 /* Inform hwpmc(4) if an executable is being mapped. */
2914 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2915 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2917 pkm.pm_address = (uintptr_t) *addr;
2918 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2926 vn_fsid(struct vnode *vp, struct vattr *va)
2930 f = &vp->v_mount->mnt_stat.f_fsid;
2931 va->va_fsid = (uint32_t)f->val[1];
2932 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2933 va->va_fsid += (uint32_t)f->val[0];
2937 vn_fsync_buf(struct vnode *vp, int waitfor)
2939 struct buf *bp, *nbp;
2942 int error, maxretry;
2945 maxretry = 10000; /* large, arbitrarily chosen */
2947 if (vp->v_type == VCHR) {
2949 mp = vp->v_rdev->si_mountpt;
2956 * MARK/SCAN initialization to avoid infinite loops.
2958 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2959 bp->b_vflags &= ~BV_SCANNED;
2964 * Flush all dirty buffers associated with a vnode.
2967 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2968 if ((bp->b_vflags & BV_SCANNED) != 0)
2970 bp->b_vflags |= BV_SCANNED;
2971 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2972 if (waitfor != MNT_WAIT)
2975 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2976 BO_LOCKPTR(bo)) != 0) {
2983 KASSERT(bp->b_bufobj == bo,
2984 ("bp %p wrong b_bufobj %p should be %p",
2985 bp, bp->b_bufobj, bo));
2986 if ((bp->b_flags & B_DELWRI) == 0)
2987 panic("fsync: not dirty");
2988 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2994 if (maxretry < 1000)
2995 pause("dirty", hz < 1000 ? 1 : hz / 1000);
3001 * If synchronous the caller expects us to completely resolve all
3002 * dirty buffers in the system. Wait for in-progress I/O to
3003 * complete (which could include background bitmap writes), then
3004 * retry if dirty blocks still exist.
3006 if (waitfor == MNT_WAIT) {
3007 bufobj_wwait(bo, 0, 0);
3008 if (bo->bo_dirty.bv_cnt > 0) {
3010 * If we are unable to write any of these buffers
3011 * then we fail now rather than trying endlessly
3012 * to write them out.
3014 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
3015 if ((error = bp->b_error) != 0)
3017 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
3018 (error == 0 && --maxretry >= 0))
3026 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
3032 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
3033 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
3034 * to do the actual copy.
3035 * vn_generic_copy_file_range() is factored out, so it can be called
3036 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
3037 * different file systems.
3040 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
3041 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
3042 struct ucred *outcred, struct thread *fsize_td)
3049 *lenp = 0; /* For error returns. */
3052 /* Do some sanity checks on the arguments. */
3053 if (invp->v_type == VDIR || outvp->v_type == VDIR)
3055 else if (*inoffp < 0 || *outoffp < 0 ||
3056 invp->v_type != VREG || outvp->v_type != VREG)
3061 /* Ensure offset + len does not wrap around. */
3064 if (uval > INT64_MAX)
3065 len = INT64_MAX - *inoffp;
3068 if (uval > INT64_MAX)
3069 len = INT64_MAX - *outoffp;
3074 * If the two vnode are for the same file system, call
3075 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
3076 * which can handle copies across multiple file systems.
3079 if (invp->v_mount == outvp->v_mount)
3080 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
3081 lenp, flags, incred, outcred, fsize_td);
3083 error = vn_generic_copy_file_range(invp, inoffp, outvp,
3084 outoffp, lenp, flags, incred, outcred, fsize_td);
3090 * Test len bytes of data starting at dat for all bytes == 0.
3091 * Return true if all bytes are zero, false otherwise.
3092 * Expects dat to be well aligned.
3095 mem_iszero(void *dat, int len)
3101 for (p = dat; len > 0; len -= sizeof(*p), p++) {
3102 if (len >= sizeof(*p)) {
3106 cp = (const char *)p;
3107 for (i = 0; i < len; i++, cp++)
3116 * Look for a hole in the output file and, if found, adjust *outoffp
3117 * and *xferp to skip past the hole.
3118 * *xferp is the entire hole length to be written and xfer2 is how many bytes
3119 * to be written as 0's upon return.
3122 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3123 off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3128 if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3129 *dataoffp = *outoffp;
3130 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3133 *holeoffp = *dataoffp;
3134 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3137 if (error != 0 || *holeoffp == *dataoffp) {
3139 * Since outvp is unlocked, it may be possible for
3140 * another thread to do a truncate(), lseek(), write()
3141 * creating a hole at startoff between the above
3142 * VOP_IOCTL() calls, if the other thread does not do
3144 * If that happens, *holeoffp == *dataoffp and finding
3145 * the hole has failed, so disable vn_skip_hole().
3147 *holeoffp = -1; /* Disable use of vn_skip_hole(). */
3150 KASSERT(*dataoffp >= *outoffp,
3151 ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3152 (intmax_t)*dataoffp, (intmax_t)*outoffp));
3153 KASSERT(*holeoffp > *dataoffp,
3154 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3155 (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3159 * If there is a hole before the data starts, advance *outoffp and
3160 * *xferp past the hole.
3162 if (*dataoffp > *outoffp) {
3163 delta = *dataoffp - *outoffp;
3164 if (delta >= *xferp) {
3165 /* Entire *xferp is a hole. */
3172 xfer2 = MIN(xfer2, *xferp);
3176 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3177 * that the write ends at the start of the hole.
3178 * *holeoffp should always be greater than *outoffp, but for the
3179 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3182 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3183 xfer2 = *holeoffp - *outoffp;
3188 * Write an xfer sized chunk to outvp in blksize blocks from dat.
3189 * dat is a maximum of blksize in length and can be written repeatedly in
3191 * If growfile == true, just grow the file via vn_truncate_locked() instead
3192 * of doing actual writes.
3193 * If checkhole == true, a hole is being punched, so skip over any hole
3194 * already in the output file.
3197 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3198 u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3201 off_t dataoff, holeoff, xfer2;
3205 * Loop around doing writes of blksize until write has been completed.
3206 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3207 * done for each iteration, since the xfer argument can be very
3208 * large if there is a large hole to punch in the output file.
3213 xfer2 = MIN(xfer, blksize);
3216 * Punching a hole. Skip writing if there is
3217 * already a hole in the output file.
3219 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3220 &dataoff, &holeoff, cred);
3225 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3230 error = vn_start_write(outvp, &mp, V_WAIT);
3234 error = vn_lock(outvp, LK_EXCLUSIVE);
3236 error = vn_truncate_locked(outvp, outoff + xfer,
3241 error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3243 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3244 outoff, UIO_SYSSPACE, IO_NODELOCKED,
3245 curthread->td_ucred, cred, NULL, curthread);
3252 vn_finished_write(mp);
3253 } while (!growfile && xfer > 0 && error == 0);
3258 * Copy a byte range of one file to another. This function can handle the
3259 * case where invp and outvp are on different file systems.
3260 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3261 * is no better file system specific way to do it.
3264 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3265 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3266 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3269 off_t startoff, endoff, xfer, xfer2;
3271 int error, interrupted;
3272 bool cantseek, readzeros, eof, lastblock, holetoeof;
3273 ssize_t aresid, r = 0;
3274 size_t copylen, len, savlen;
3275 off_t insize, outsize;
3277 long holein, holeout;
3278 struct timespec curts, endts;
3280 holein = holeout = 0;
3281 savlen = len = *lenp;
3286 error = vn_lock(invp, LK_SHARED);
3289 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3292 error = vn_getsize_locked(invp, &insize, incred);
3298 error = vn_start_write(outvp, &mp, V_WAIT);
3300 error = vn_lock(outvp, LK_EXCLUSIVE);
3303 * If fsize_td != NULL, do a vn_rlimit_fsizex() call,
3304 * now that outvp is locked.
3306 if (fsize_td != NULL) {
3309 io.uio_offset = *outoffp;
3311 error = vn_rlimit_fsizex(outvp, &io, 0, &r, fsize_td);
3312 len = savlen = io.uio_resid;
3314 * No need to call vn_rlimit_fsizex_res before return,
3315 * since the uio is local.
3318 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3321 * Holes that are past EOF do not need to be written as a block
3322 * of zero bytes. So, truncate the output file as far as
3323 * possible and then use size to decide if writing 0
3324 * bytes is necessary in the loop below.
3327 error = vn_getsize_locked(outvp, &outsize, outcred);
3328 if (error == 0 && outsize > *outoffp && outsize <= *outoffp + len) {
3330 error = mac_vnode_check_write(curthread->td_ucred,
3334 error = vn_truncate_locked(outvp, *outoffp,
3342 vn_finished_write(mp);
3346 if (holein == 0 && holeout > 0) {
3348 * For this special case, the input data will be scanned
3349 * for blocks of all 0 bytes. For these blocks, the
3350 * write can be skipped for the output file to create
3351 * an unallocated region.
3352 * Therefore, use the appropriate size for the output file.
3355 if (blksize <= 512) {
3357 * Use f_iosize, since ZFS reports a _PC_MIN_HOLE_SIZE
3358 * of 512, although it actually only creates
3359 * unallocated regions for blocks >= f_iosize.
3361 blksize = outvp->v_mount->mnt_stat.f_iosize;
3365 * Use the larger of the two f_iosize values. If they are
3366 * not the same size, one will normally be an exact multiple of
3367 * the other, since they are both likely to be a power of 2.
3369 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3370 outvp->v_mount->mnt_stat.f_iosize);
3373 /* Clip to sane limits. */
3376 else if (blksize > maxphys)
3378 dat = malloc(blksize, M_TEMP, M_WAITOK);
3381 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3382 * to find holes. Otherwise, just scan the read block for all 0s
3383 * in the inner loop where the data copying is done.
3384 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3385 * support holes on the server, but do not support FIOSEEKHOLE.
3386 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3387 * that this function should return after 1second with a partial
3390 if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3391 getnanouptime(&endts);
3394 timespecclear(&endts);
3395 holetoeof = eof = false;
3396 while (len > 0 && error == 0 && !eof && interrupted == 0) {
3397 endoff = 0; /* To shut up compilers. */
3403 * Find the next data area. If there is just a hole to EOF,
3404 * FIOSEEKDATA should fail with ENXIO.
3405 * (I do not know if any file system will report a hole to
3406 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3407 * will fail for those file systems.)
3409 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3410 * the code just falls through to the inner copy loop.
3414 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3416 if (error == ENXIO) {
3417 startoff = endoff = insize;
3418 eof = holetoeof = true;
3422 if (error == 0 && !holetoeof) {
3424 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3427 * Since invp is unlocked, it may be possible for
3428 * another thread to do a truncate(), lseek(), write()
3429 * creating a hole at startoff between the above
3430 * VOP_IOCTL() calls, if the other thread does not do
3432 * If that happens, startoff == endoff and finding
3433 * the hole has failed, so set an error.
3435 if (error == 0 && startoff == endoff)
3436 error = EINVAL; /* Any error. Reset to 0. */
3439 if (startoff > *inoffp) {
3440 /* Found hole before data block. */
3441 xfer = MIN(startoff - *inoffp, len);
3442 if (*outoffp < outsize) {
3443 /* Must write 0s to punch hole. */
3444 xfer2 = MIN(outsize - *outoffp,
3446 memset(dat, 0, MIN(xfer2, blksize));
3447 error = vn_write_outvp(outvp, dat,
3448 *outoffp, xfer2, blksize, false,
3449 holeout > 0, outcred);
3452 if (error == 0 && *outoffp + xfer >
3453 outsize && (xfer == len || holetoeof)) {
3454 /* Grow output file (hole at end). */
3455 error = vn_write_outvp(outvp, dat,
3456 *outoffp, xfer, blksize, true,
3464 interrupted = sig_intr();
3465 if (timespecisset(&endts) &&
3467 getnanouptime(&curts);
3468 if (timespeccmp(&curts,
3476 copylen = MIN(len, endoff - startoff);
3488 * Set first xfer to end at a block boundary, so that
3489 * holes are more likely detected in the loop below via
3490 * the for all bytes 0 method.
3492 xfer -= (*inoffp % blksize);
3494 /* Loop copying the data block. */
3495 while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3498 error = vn_lock(invp, LK_SHARED);
3501 error = vn_rdwr(UIO_READ, invp, dat, xfer,
3502 startoff, UIO_SYSSPACE, IO_NODELOCKED,
3503 curthread->td_ucred, incred, &aresid,
3507 if (error == 0 && aresid > 0) {
3508 /* Stop the copy at EOF on the input file. */
3515 * Skip the write for holes past the initial EOF
3516 * of the output file, unless this is the last
3517 * write of the output file at EOF.
3519 readzeros = cantseek ? mem_iszero(dat, xfer) :
3523 if (!cantseek || *outoffp < outsize ||
3524 lastblock || !readzeros)
3525 error = vn_write_outvp(outvp, dat,
3526 *outoffp, xfer, blksize,
3527 readzeros && lastblock &&
3528 *outoffp >= outsize, false,
3537 interrupted = sig_intr();
3538 if (timespecisset(&endts) &&
3540 getnanouptime(&curts);
3541 if (timespeccmp(&curts,
3553 *lenp = savlen - len;
3559 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3563 off_t olen, ooffset;
3566 int audited_vnode1 = 0;
3570 if (vp->v_type != VREG)
3573 /* Allocating blocks may take a long time, so iterate. */
3580 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
3583 error = vn_lock(vp, LK_EXCLUSIVE);
3585 vn_finished_write(mp);
3589 if (!audited_vnode1) {
3590 AUDIT_ARG_VNODE1(vp);
3595 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3598 error = VOP_ALLOCATE(vp, &offset, &len, 0,
3601 vn_finished_write(mp);
3603 if (olen + ooffset != offset + len) {
3604 panic("offset + len changed from %jx/%jx to %jx/%jx",
3605 ooffset, olen, offset, len);
3607 if (error != 0 || len == 0)
3609 KASSERT(olen > len, ("Iteration did not make progress?"));
3617 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags,
3618 int ioflag, struct ucred *cred, struct ucred *active_cred,
3619 struct ucred *file_cred)
3626 bool audited_vnode1 = false;
3635 if ((ioflag & (IO_NODELOCKED | IO_RANGELOCKED)) == 0)
3636 rl_cookie = vn_rangelock_wlock(vp, off, off + len);
3637 while (len > 0 && error == 0) {
3639 * Try to deallocate the longest range in one pass.
3640 * In case a pass takes too long to be executed, it returns
3641 * partial result. The residue will be proceeded in the next
3645 if ((ioflag & IO_NODELOCKED) == 0) {
3647 if ((error = vn_start_write(vp, &mp,
3648 V_WAIT | V_PCATCH)) != 0)
3650 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
3653 if (!audited_vnode1) {
3654 AUDIT_ARG_VNODE1(vp);
3655 audited_vnode1 = true;
3660 if ((ioflag & IO_NOMACCHECK) == 0)
3661 error = mac_vnode_check_write(active_cred, file_cred,
3665 error = VOP_DEALLOCATE(vp, &off, &len, flags, ioflag,
3668 if ((ioflag & IO_NODELOCKED) == 0) {
3671 vn_finished_write(mp);
3675 if (error == 0 && len != 0)
3679 if (rl_cookie != NULL)
3680 vn_rangelock_unlock(vp, rl_cookie);
3687 * This function is supposed to be used in the situations where the deallocation
3688 * is not triggered by a user request.
3691 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags,
3692 int ioflag, struct ucred *active_cred, struct ucred *file_cred)
3696 if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset ||
3699 if (vp->v_type != VREG)
3702 cred = file_cred != NOCRED ? file_cred : active_cred;
3703 return (vn_deallocate_impl(vp, offset, length, flags, ioflag, cred,
3704 active_cred, file_cred));
3708 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
3709 struct ucred *active_cred, struct thread *td)
3715 KASSERT(cmd == SPACECTL_DEALLOC, ("vn_fspacectl: Invalid cmd"));
3716 KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0,
3717 ("vn_fspacectl: non-zero flags"));
3718 KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset,
3719 ("vn_fspacectl: offset/length overflow or underflow"));
3722 if (vp->v_type != VREG)
3725 ioflag = get_write_ioflag(fp);
3728 case SPACECTL_DEALLOC:
3729 error = vn_deallocate_impl(vp, offset, length, flags, ioflag,
3730 active_cred, active_cred, fp->f_cred);
3733 panic("vn_fspacectl: unknown cmd %d", cmd);
3739 static u_long vn_lock_pair_pause_cnt;
3740 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3741 &vn_lock_pair_pause_cnt, 0,
3742 "Count of vn_lock_pair deadlocks");
3744 u_int vn_lock_pair_pause_max;
3745 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3746 &vn_lock_pair_pause_max, 0,
3747 "Max ticks for vn_lock_pair deadlock avoidance sleep");
3750 vn_lock_pair_pause(const char *wmesg)
3752 atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3753 pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3757 * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3758 * vp1_locked indicates whether vp1 is locked; if not, vp1 must be
3759 * unlocked. Same for vp2 and vp2_locked. One of the vnodes can be
3762 * The function returns with both vnodes exclusively or shared locked,
3763 * according to corresponding lkflags, and guarantees that it does not
3764 * create lock order reversal with other threads during its execution.
3765 * Both vnodes could be unlocked temporary (and reclaimed).
3767 * If requesting shared locking, locked vnode lock must not be recursed.
3770 vn_lock_pair(struct vnode *vp1, bool vp1_locked, int lkflags1,
3771 struct vnode *vp2, bool vp2_locked, int lkflags2)
3775 MPASS(lkflags1 == LK_SHARED || lkflags1 == LK_EXCLUSIVE);
3776 MPASS(lkflags2 == LK_SHARED || lkflags2 == LK_EXCLUSIVE);
3778 if (vp1 == NULL && vp2 == NULL)
3782 if (lkflags1 == LK_SHARED &&
3783 (vp1->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
3784 lkflags1 = LK_EXCLUSIVE;
3785 if (vp1_locked && VOP_ISLOCKED(vp1) != LK_EXCLUSIVE) {
3786 ASSERT_VOP_LOCKED(vp1, "vp1");
3787 if (lkflags1 == LK_EXCLUSIVE) {
3789 ASSERT_VOP_UNLOCKED(vp1,
3790 "vp1 shared recursed");
3793 } else if (!vp1_locked)
3794 ASSERT_VOP_UNLOCKED(vp1, "vp1");
3800 if (lkflags2 == LK_SHARED &&
3801 (vp2->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
3802 lkflags2 = LK_EXCLUSIVE;
3803 if (vp2_locked && VOP_ISLOCKED(vp2) != LK_EXCLUSIVE) {
3804 ASSERT_VOP_LOCKED(vp2, "vp2");
3805 if (lkflags2 == LK_EXCLUSIVE) {
3807 ASSERT_VOP_UNLOCKED(vp2,
3808 "vp2 shared recursed");
3811 } else if (!vp2_locked)
3812 ASSERT_VOP_UNLOCKED(vp2, "vp2");
3817 if (!vp1_locked && !vp2_locked) {
3818 vn_lock(vp1, lkflags1 | LK_RETRY);
3822 while (!vp1_locked || !vp2_locked) {
3823 if (vp1_locked && vp2 != NULL) {
3825 error = VOP_LOCK1(vp2, lkflags2 | LK_NOWAIT,
3826 __FILE__, __LINE__);
3831 vn_lock_pair_pause("vlp1");
3833 vn_lock(vp2, lkflags2 | LK_RETRY);
3836 if (vp2_locked && vp1 != NULL) {
3838 error = VOP_LOCK1(vp1, lkflags1 | LK_NOWAIT,
3839 __FILE__, __LINE__);
3844 vn_lock_pair_pause("vlp2");
3846 vn_lock(vp1, lkflags1 | LK_RETRY);
3851 if (lkflags1 == LK_EXCLUSIVE)
3852 ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3854 ASSERT_VOP_LOCKED(vp1, "vp1 ret");
3857 if (lkflags2 == LK_EXCLUSIVE)
3858 ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3860 ASSERT_VOP_LOCKED(vp2, "vp2 ret");
3865 vn_lktype_write(struct mount *mp, struct vnode *vp)
3867 if (MNT_SHARED_WRITES(mp) ||
3868 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
3870 return (LK_EXCLUSIVE);