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 && (fmode & FWRITE) != 0 &&
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 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
532 MNT_EXTENDED_SHARED(vp->v_mount))
533 lock_flags = LK_SHARED;
535 lock_flags = LK_EXCLUSIVE;
537 vn_start_write(vp, &mp, V_WAIT);
538 vn_lock(vp, lock_flags | LK_RETRY);
539 AUDIT_ARG_VNODE1(vp);
540 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
541 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
542 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
543 __func__, vp, vp->v_writecount);
545 error = VOP_CLOSE(vp, flags, file_cred, td);
550 vn_finished_write(mp);
555 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
559 return (vn_close1(vp, flags, file_cred, td, false));
563 * Heuristic to detect sequential operation.
566 sequential_heuristic(struct uio *uio, struct file *fp)
570 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
573 if (fp->f_flag & FRDAHEAD)
574 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
577 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
578 * that the first I/O is normally considered to be slightly
579 * sequential. Seeking to offset 0 doesn't change sequentiality
580 * unless previous seeks have reduced f_seqcount to 0, in which
581 * case offset 0 is not special.
583 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
584 uio->uio_offset == fp->f_nextoff[rw]) {
586 * f_seqcount is in units of fixed-size blocks so that it
587 * depends mainly on the amount of sequential I/O and not
588 * much on the number of sequential I/O's. The fixed size
589 * of 16384 is hard-coded here since it is (not quite) just
590 * a magic size that works well here. This size is more
591 * closely related to the best I/O size for real disks than
592 * to any block size used by software.
594 if (uio->uio_resid >= IO_SEQMAX * 16384)
595 fp->f_seqcount[rw] = IO_SEQMAX;
597 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
598 if (fp->f_seqcount[rw] > IO_SEQMAX)
599 fp->f_seqcount[rw] = IO_SEQMAX;
601 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
604 /* Not sequential. Quickly draw-down sequentiality. */
605 if (fp->f_seqcount[rw] > 1)
606 fp->f_seqcount[rw] = 1;
608 fp->f_seqcount[rw] = 0;
613 * Package up an I/O request on a vnode into a uio and do it.
616 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
617 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
618 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
625 struct vn_io_fault_args args;
626 int error, lock_flags;
628 if (offset < 0 && vp->v_type != VCHR)
630 auio.uio_iov = &aiov;
632 aiov.iov_base = base;
634 auio.uio_resid = len;
635 auio.uio_offset = offset;
636 auio.uio_segflg = segflg;
641 if ((ioflg & IO_NODELOCKED) == 0) {
642 if ((ioflg & IO_RANGELOCKED) == 0) {
643 if (rw == UIO_READ) {
644 rl_cookie = vn_rangelock_rlock(vp, offset,
646 } else if ((ioflg & IO_APPEND) != 0) {
647 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
649 rl_cookie = vn_rangelock_wlock(vp, offset,
655 if (rw == UIO_WRITE) {
656 if (vp->v_type != VCHR &&
657 (error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH))
660 lock_flags = vn_lktype_write(mp, vp);
662 lock_flags = LK_SHARED;
663 vn_lock(vp, lock_flags | LK_RETRY);
667 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
669 if ((ioflg & IO_NOMACCHECK) == 0) {
671 error = mac_vnode_check_read(active_cred, file_cred,
674 error = mac_vnode_check_write(active_cred, file_cred,
679 if (file_cred != NULL)
683 if (do_vn_io_fault(vp, &auio)) {
684 args.kind = VN_IO_FAULT_VOP;
687 args.args.vop_args.vp = vp;
688 error = vn_io_fault1(vp, &auio, &args, td);
689 } else if (rw == UIO_READ) {
690 error = VOP_READ(vp, &auio, ioflg, cred);
691 } else /* if (rw == UIO_WRITE) */ {
692 error = VOP_WRITE(vp, &auio, ioflg, cred);
696 *aresid = auio.uio_resid;
698 if (auio.uio_resid && error == 0)
700 if ((ioflg & IO_NODELOCKED) == 0) {
703 vn_finished_write(mp);
706 if (rl_cookie != NULL)
707 vn_rangelock_unlock(vp, rl_cookie);
712 * Package up an I/O request on a vnode into a uio and do it. The I/O
713 * request is split up into smaller chunks and we try to avoid saturating
714 * the buffer cache while potentially holding a vnode locked, so we
715 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
716 * to give other processes a chance to lock the vnode (either other processes
717 * core'ing the same binary, or unrelated processes scanning the directory).
720 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
721 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
722 struct ucred *file_cred, size_t *aresid, struct thread *td)
731 * Force `offset' to a multiple of MAXBSIZE except possibly
732 * for the first chunk, so that filesystems only need to
733 * write full blocks except possibly for the first and last
736 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
740 if (rw != UIO_READ && vp->v_type == VREG)
743 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
744 ioflg, active_cred, file_cred, &iaresid, td);
745 len -= chunk; /* aresid calc already includes length */
749 base = (char *)base + chunk;
750 kern_yield(PRI_USER);
753 *aresid = len + iaresid;
757 #if OFF_MAX <= LONG_MAX
759 foffset_lock(struct file *fp, int flags)
761 volatile short *flagsp;
765 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
767 if ((flags & FOF_NOLOCK) != 0)
768 return (atomic_load_long(&fp->f_offset));
771 * According to McKusick the vn lock was protecting f_offset here.
772 * It is now protected by the FOFFSET_LOCKED flag.
774 flagsp = &fp->f_vnread_flags;
775 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
776 return (atomic_load_long(&fp->f_offset));
778 sleepq_lock(&fp->f_vnread_flags);
779 state = atomic_load_16(flagsp);
781 if ((state & FOFFSET_LOCKED) == 0) {
782 if (!atomic_fcmpset_acq_16(flagsp, &state,
787 if ((state & FOFFSET_LOCK_WAITING) == 0) {
788 if (!atomic_fcmpset_acq_16(flagsp, &state,
789 state | FOFFSET_LOCK_WAITING))
793 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
794 sleepq_wait(&fp->f_vnread_flags, PUSER -1);
796 sleepq_lock(&fp->f_vnread_flags);
797 state = atomic_load_16(flagsp);
799 res = atomic_load_long(&fp->f_offset);
800 sleepq_release(&fp->f_vnread_flags);
805 foffset_unlock(struct file *fp, off_t val, int flags)
807 volatile short *flagsp;
810 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
812 if ((flags & FOF_NOUPDATE) == 0)
813 atomic_store_long(&fp->f_offset, val);
814 if ((flags & FOF_NEXTOFF_R) != 0)
815 fp->f_nextoff[UIO_READ] = val;
816 if ((flags & FOF_NEXTOFF_W) != 0)
817 fp->f_nextoff[UIO_WRITE] = val;
819 if ((flags & FOF_NOLOCK) != 0)
822 flagsp = &fp->f_vnread_flags;
823 state = atomic_load_16(flagsp);
824 if ((state & FOFFSET_LOCK_WAITING) == 0 &&
825 atomic_cmpset_rel_16(flagsp, state, 0))
828 sleepq_lock(&fp->f_vnread_flags);
829 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
830 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
831 fp->f_vnread_flags = 0;
832 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
833 sleepq_release(&fp->f_vnread_flags);
837 foffset_lock(struct file *fp, int flags)
842 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
844 mtxp = mtx_pool_find(mtxpool_sleep, fp);
846 if ((flags & FOF_NOLOCK) == 0) {
847 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
848 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
849 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
852 fp->f_vnread_flags |= FOFFSET_LOCKED;
860 foffset_unlock(struct file *fp, off_t val, int flags)
864 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
866 mtxp = mtx_pool_find(mtxpool_sleep, fp);
868 if ((flags & FOF_NOUPDATE) == 0)
870 if ((flags & FOF_NEXTOFF_R) != 0)
871 fp->f_nextoff[UIO_READ] = val;
872 if ((flags & FOF_NEXTOFF_W) != 0)
873 fp->f_nextoff[UIO_WRITE] = val;
874 if ((flags & FOF_NOLOCK) == 0) {
875 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
876 ("Lost FOFFSET_LOCKED"));
877 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
878 wakeup(&fp->f_vnread_flags);
879 fp->f_vnread_flags = 0;
886 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
889 if ((flags & FOF_OFFSET) == 0)
890 uio->uio_offset = foffset_lock(fp, flags);
894 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
897 if ((flags & FOF_OFFSET) == 0)
898 foffset_unlock(fp, uio->uio_offset, flags);
902 get_advice(struct file *fp, struct uio *uio)
907 ret = POSIX_FADV_NORMAL;
908 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
911 mtxp = mtx_pool_find(mtxpool_sleep, fp);
913 if (fp->f_advice != NULL &&
914 uio->uio_offset >= fp->f_advice->fa_start &&
915 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
916 ret = fp->f_advice->fa_advice;
922 get_write_ioflag(struct file *fp)
930 mp = atomic_load_ptr(&vp->v_mount);
932 if ((fp->f_flag & O_DIRECT) != 0)
935 if ((fp->f_flag & O_FSYNC) != 0 ||
936 (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
940 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
941 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
942 * fall back to full O_SYNC behavior.
944 if ((fp->f_flag & O_DSYNC) != 0)
945 ioflag |= IO_SYNC | IO_DATASYNC;
951 vn_read_from_obj(struct vnode *vp, struct uio *uio)
954 vm_page_t ma[io_hold_cnt + 2];
959 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
960 obj = atomic_load_ptr(&vp->v_object);
962 return (EJUSTRETURN);
965 * Depends on type stability of vm_objects.
967 vm_object_pip_add(obj, 1);
968 if ((obj->flags & OBJ_DEAD) != 0) {
970 * Note that object might be already reused from the
971 * vnode, and the OBJ_DEAD flag cleared. This is fine,
972 * we recheck for DOOMED vnode state after all pages
973 * are busied, and retract then.
975 * But we check for OBJ_DEAD to ensure that we do not
976 * busy pages while vm_object_terminate_pages()
977 * processes the queue.
983 resid = uio->uio_resid;
984 off = uio->uio_offset;
985 for (i = 0; resid > 0; i++) {
986 MPASS(i < io_hold_cnt + 2);
987 ma[i] = vm_page_grab_unlocked(obj, atop(off),
988 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
994 * Skip invalid pages. Valid mask can be partial only
995 * at EOF, and we clip later.
997 if (vm_page_none_valid(ma[i])) {
998 vm_page_sunbusy(ma[i]);
1006 error = EJUSTRETURN;
1011 * Check VIRF_DOOMED after we busied our pages. Since
1012 * vgonel() terminates the vnode' vm_object, it cannot
1013 * process past pages busied by us.
1015 if (VN_IS_DOOMED(vp)) {
1016 error = EJUSTRETURN;
1020 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1021 if (resid > uio->uio_resid)
1022 resid = uio->uio_resid;
1025 * Unlocked read of vnp_size is safe because truncation cannot
1026 * pass busied page. But we load vnp_size into a local
1027 * variable so that possible concurrent extension does not
1028 * break calculation.
1030 #if defined(__powerpc__) && !defined(__powerpc64__)
1031 vsz = obj->un_pager.vnp.vnp_size;
1033 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1035 if (uio->uio_offset >= vsz) {
1036 error = EJUSTRETURN;
1039 if (uio->uio_offset + resid > vsz)
1040 resid = vsz - uio->uio_offset;
1042 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1045 for (j = 0; j < i; j++) {
1047 vm_page_reference(ma[j]);
1048 vm_page_sunbusy(ma[j]);
1051 vm_object_pip_wakeup(obj);
1054 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1058 * File table vnode read routine.
1061 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1069 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1071 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1074 if (fp->f_flag & FNONBLOCK)
1075 ioflag |= IO_NDELAY;
1076 if (fp->f_flag & O_DIRECT)
1077 ioflag |= IO_DIRECT;
1080 * Try to read from page cache. VIRF_DOOMED check is racy but
1081 * allows us to avoid unneeded work outright.
1083 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1084 (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1085 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1087 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1090 if (error != EJUSTRETURN)
1094 advice = get_advice(fp, uio);
1095 vn_lock(vp, LK_SHARED | LK_RETRY);
1098 case POSIX_FADV_NORMAL:
1099 case POSIX_FADV_SEQUENTIAL:
1100 case POSIX_FADV_NOREUSE:
1101 ioflag |= sequential_heuristic(uio, fp);
1103 case POSIX_FADV_RANDOM:
1104 /* Disable read-ahead for random I/O. */
1107 orig_offset = uio->uio_offset;
1110 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1113 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1114 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1116 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1117 orig_offset != uio->uio_offset)
1119 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1120 * for the backing file after a POSIX_FADV_NOREUSE
1123 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1124 POSIX_FADV_DONTNEED);
1129 * File table vnode write routine.
1132 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1140 bool need_finished_write;
1142 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1144 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1146 if (vp->v_type == VREG)
1149 if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1150 ioflag |= IO_APPEND;
1151 if ((fp->f_flag & FNONBLOCK) != 0)
1152 ioflag |= IO_NDELAY;
1153 ioflag |= get_write_ioflag(fp);
1156 need_finished_write = false;
1157 if (vp->v_type != VCHR) {
1158 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
1161 need_finished_write = true;
1164 advice = get_advice(fp, uio);
1166 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1168 case POSIX_FADV_NORMAL:
1169 case POSIX_FADV_SEQUENTIAL:
1170 case POSIX_FADV_NOREUSE:
1171 ioflag |= sequential_heuristic(uio, fp);
1173 case POSIX_FADV_RANDOM:
1174 /* XXX: Is this correct? */
1177 orig_offset = uio->uio_offset;
1180 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1183 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1184 fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1186 if (need_finished_write)
1187 vn_finished_write(mp);
1188 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1189 orig_offset != uio->uio_offset)
1191 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1192 * for the backing file after a POSIX_FADV_NOREUSE
1195 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1196 POSIX_FADV_DONTNEED);
1202 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1203 * prevent the following deadlock:
1205 * Assume that the thread A reads from the vnode vp1 into userspace
1206 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
1207 * currently not resident, then system ends up with the call chain
1208 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1209 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1210 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1211 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1212 * backed by the pages of vnode vp1, and some page in buf2 is not
1213 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1215 * To prevent the lock order reversal and deadlock, vn_io_fault() does
1216 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1217 * Instead, it first tries to do the whole range i/o with pagefaults
1218 * disabled. If all pages in the i/o buffer are resident and mapped,
1219 * VOP will succeed (ignoring the genuine filesystem errors).
1220 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1221 * i/o in chunks, with all pages in the chunk prefaulted and held
1222 * using vm_fault_quick_hold_pages().
1224 * Filesystems using this deadlock avoidance scheme should use the
1225 * array of the held pages from uio, saved in the curthread->td_ma,
1226 * instead of doing uiomove(). A helper function
1227 * vn_io_fault_uiomove() converts uiomove request into
1228 * uiomove_fromphys() over td_ma array.
1230 * Since vnode locks do not cover the whole i/o anymore, rangelocks
1231 * make the current i/o request atomic with respect to other i/os and
1236 * Decode vn_io_fault_args and perform the corresponding i/o.
1239 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1245 save = vm_fault_disable_pagefaults();
1246 switch (args->kind) {
1247 case VN_IO_FAULT_FOP:
1248 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1249 uio, args->cred, args->flags, td);
1251 case VN_IO_FAULT_VOP:
1252 if (uio->uio_rw == UIO_READ) {
1253 error = VOP_READ(args->args.vop_args.vp, uio,
1254 args->flags, args->cred);
1255 } else if (uio->uio_rw == UIO_WRITE) {
1256 error = VOP_WRITE(args->args.vop_args.vp, uio,
1257 args->flags, args->cred);
1261 panic("vn_io_fault_doio: unknown kind of io %d %d",
1262 args->kind, uio->uio_rw);
1264 vm_fault_enable_pagefaults(save);
1269 vn_io_fault_touch(char *base, const struct uio *uio)
1274 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1280 vn_io_fault_prefault_user(const struct uio *uio)
1283 const struct iovec *iov;
1288 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1289 ("vn_io_fault_prefault userspace"));
1293 resid = uio->uio_resid;
1294 base = iov->iov_base;
1297 error = vn_io_fault_touch(base, uio);
1300 if (len < PAGE_SIZE) {
1302 error = vn_io_fault_touch(base + len - 1, uio);
1307 if (++i >= uio->uio_iovcnt)
1309 iov = uio->uio_iov + i;
1310 base = iov->iov_base;
1322 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1323 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1324 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1325 * into args and call vn_io_fault1() to handle faults during the user
1326 * mode buffer accesses.
1329 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1332 vm_page_t ma[io_hold_cnt + 2];
1333 struct uio *uio_clone, short_uio;
1334 struct iovec short_iovec[1];
1335 vm_page_t *prev_td_ma;
1337 vm_offset_t addr, end;
1340 int error, cnt, saveheld, prev_td_ma_cnt;
1342 if (vn_io_fault_prefault) {
1343 error = vn_io_fault_prefault_user(uio);
1345 return (error); /* Or ignore ? */
1348 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1351 * The UFS follows IO_UNIT directive and replays back both
1352 * uio_offset and uio_resid if an error is encountered during the
1353 * operation. But, since the iovec may be already advanced,
1354 * uio is still in an inconsistent state.
1356 * Cache a copy of the original uio, which is advanced to the redo
1357 * point using UIO_NOCOPY below.
1359 uio_clone = cloneuio(uio);
1360 resid = uio->uio_resid;
1362 short_uio.uio_segflg = UIO_USERSPACE;
1363 short_uio.uio_rw = uio->uio_rw;
1364 short_uio.uio_td = uio->uio_td;
1366 error = vn_io_fault_doio(args, uio, td);
1367 if (error != EFAULT)
1370 atomic_add_long(&vn_io_faults_cnt, 1);
1371 uio_clone->uio_segflg = UIO_NOCOPY;
1372 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1373 uio_clone->uio_segflg = uio->uio_segflg;
1375 saveheld = curthread_pflags_set(TDP_UIOHELD);
1376 prev_td_ma = td->td_ma;
1377 prev_td_ma_cnt = td->td_ma_cnt;
1379 while (uio_clone->uio_resid != 0) {
1380 len = uio_clone->uio_iov->iov_len;
1382 KASSERT(uio_clone->uio_iovcnt >= 1,
1383 ("iovcnt underflow"));
1384 uio_clone->uio_iov++;
1385 uio_clone->uio_iovcnt--;
1388 if (len > ptoa(io_hold_cnt))
1389 len = ptoa(io_hold_cnt);
1390 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1391 end = round_page(addr + len);
1397 * A perfectly misaligned address and length could cause
1398 * both the start and the end of the chunk to use partial
1399 * page. +2 accounts for such a situation.
1401 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1402 addr, len, prot, ma, io_hold_cnt + 2);
1407 short_uio.uio_iov = &short_iovec[0];
1408 short_iovec[0].iov_base = (void *)addr;
1409 short_uio.uio_iovcnt = 1;
1410 short_uio.uio_resid = short_iovec[0].iov_len = len;
1411 short_uio.uio_offset = uio_clone->uio_offset;
1413 td->td_ma_cnt = cnt;
1415 error = vn_io_fault_doio(args, &short_uio, td);
1416 vm_page_unhold_pages(ma, cnt);
1417 adv = len - short_uio.uio_resid;
1419 uio_clone->uio_iov->iov_base =
1420 (char *)uio_clone->uio_iov->iov_base + adv;
1421 uio_clone->uio_iov->iov_len -= adv;
1422 uio_clone->uio_resid -= adv;
1423 uio_clone->uio_offset += adv;
1425 uio->uio_resid -= adv;
1426 uio->uio_offset += adv;
1428 if (error != 0 || adv == 0)
1431 td->td_ma = prev_td_ma;
1432 td->td_ma_cnt = prev_td_ma_cnt;
1433 curthread_pflags_restore(saveheld);
1435 free(uio_clone, M_IOV);
1440 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1441 int flags, struct thread *td)
1446 struct vn_io_fault_args args;
1449 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1453 * The ability to read(2) on a directory has historically been
1454 * allowed for all users, but this can and has been the source of
1455 * at least one security issue in the past. As such, it is now hidden
1456 * away behind a sysctl for those that actually need it to use it, and
1457 * restricted to root when it's turned on to make it relatively safe to
1458 * leave on for longer sessions of need.
1460 if (vp->v_type == VDIR) {
1461 KASSERT(uio->uio_rw == UIO_READ,
1462 ("illegal write attempted on a directory"));
1463 if (!vfs_allow_read_dir)
1465 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1469 foffset_lock_uio(fp, uio, flags);
1470 if (do_vn_io_fault(vp, uio)) {
1471 args.kind = VN_IO_FAULT_FOP;
1472 args.args.fop_args.fp = fp;
1473 args.args.fop_args.doio = doio;
1474 args.cred = active_cred;
1475 args.flags = flags | FOF_OFFSET;
1476 if (uio->uio_rw == UIO_READ) {
1477 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1478 uio->uio_offset + uio->uio_resid);
1479 } else if ((fp->f_flag & O_APPEND) != 0 ||
1480 (flags & FOF_OFFSET) == 0) {
1481 /* For appenders, punt and lock the whole range. */
1482 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1484 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1485 uio->uio_offset + uio->uio_resid);
1487 error = vn_io_fault1(vp, uio, &args, td);
1488 vn_rangelock_unlock(vp, rl_cookie);
1490 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1492 foffset_unlock_uio(fp, uio, flags);
1497 * Helper function to perform the requested uiomove operation using
1498 * the held pages for io->uio_iov[0].iov_base buffer instead of
1499 * copyin/copyout. Access to the pages with uiomove_fromphys()
1500 * instead of iov_base prevents page faults that could occur due to
1501 * pmap_collect() invalidating the mapping created by
1502 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1503 * object cleanup revoking the write access from page mappings.
1505 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1506 * instead of plain uiomove().
1509 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1511 struct uio transp_uio;
1512 struct iovec transp_iov[1];
1518 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1519 uio->uio_segflg != UIO_USERSPACE)
1520 return (uiomove(data, xfersize, uio));
1522 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1523 transp_iov[0].iov_base = data;
1524 transp_uio.uio_iov = &transp_iov[0];
1525 transp_uio.uio_iovcnt = 1;
1526 if (xfersize > uio->uio_resid)
1527 xfersize = uio->uio_resid;
1528 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1529 transp_uio.uio_offset = 0;
1530 transp_uio.uio_segflg = UIO_SYSSPACE;
1532 * Since transp_iov points to data, and td_ma page array
1533 * corresponds to original uio->uio_iov, we need to invert the
1534 * direction of the i/o operation as passed to
1535 * uiomove_fromphys().
1537 switch (uio->uio_rw) {
1539 transp_uio.uio_rw = UIO_READ;
1542 transp_uio.uio_rw = UIO_WRITE;
1545 transp_uio.uio_td = uio->uio_td;
1546 error = uiomove_fromphys(td->td_ma,
1547 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1548 xfersize, &transp_uio);
1549 adv = xfersize - transp_uio.uio_resid;
1551 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1552 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1554 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1556 td->td_ma_cnt -= pgadv;
1557 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1558 uio->uio_iov->iov_len -= adv;
1559 uio->uio_resid -= adv;
1560 uio->uio_offset += adv;
1565 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1569 vm_offset_t iov_base;
1573 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1574 uio->uio_segflg != UIO_USERSPACE)
1575 return (uiomove_fromphys(ma, offset, xfersize, uio));
1577 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1578 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1579 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1580 switch (uio->uio_rw) {
1582 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1586 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1590 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1592 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1594 td->td_ma_cnt -= pgadv;
1595 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1596 uio->uio_iov->iov_len -= cnt;
1597 uio->uio_resid -= cnt;
1598 uio->uio_offset += cnt;
1603 * File table truncate routine.
1606 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1618 * Lock the whole range for truncation. Otherwise split i/o
1619 * might happen partly before and partly after the truncation.
1621 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1622 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
1625 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1626 AUDIT_ARG_VNODE1(vp);
1627 if (vp->v_type == VDIR) {
1632 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1636 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1640 vn_finished_write(mp);
1642 vn_rangelock_unlock(vp, rl_cookie);
1643 if (error == ERELOOKUP)
1649 * Truncate a file that is already locked.
1652 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1658 error = VOP_ADD_WRITECOUNT(vp, 1);
1661 vattr.va_size = length;
1663 vattr.va_vaflags |= VA_SYNC;
1664 error = VOP_SETATTR(vp, &vattr, cred);
1665 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1671 * File table vnode stat routine.
1674 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred)
1676 struct vnode *vp = fp->f_vnode;
1679 vn_lock(vp, LK_SHARED | LK_RETRY);
1680 error = VOP_STAT(vp, sb, active_cred, fp->f_cred);
1687 * File table vnode ioctl routine.
1690 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1695 struct fiobmap2_arg *bmarg;
1699 switch (vp->v_type) {
1704 vn_lock(vp, LK_SHARED | LK_RETRY);
1705 error = VOP_GETATTR(vp, &vattr, active_cred);
1708 *(int *)data = vattr.va_size - fp->f_offset;
1711 bmarg = (struct fiobmap2_arg *)data;
1712 vn_lock(vp, LK_SHARED | LK_RETRY);
1714 error = mac_vnode_check_read(active_cred, fp->f_cred,
1718 error = VOP_BMAP(vp, bmarg->bn, NULL,
1719 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1726 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1731 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1739 * File table vnode poll routine.
1742 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1749 #if defined(MAC) || defined(AUDIT)
1750 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1751 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1752 AUDIT_ARG_VNODE1(vp);
1753 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1759 error = VOP_POLL(vp, events, fp->f_cred, td);
1764 * Acquire the requested lock and then check for validity. LK_RETRY
1765 * permits vn_lock to return doomed vnodes.
1767 static int __noinline
1768 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1772 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1773 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1776 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1778 if ((flags & LK_RETRY) == 0) {
1789 * Nothing to do if we got the lock.
1795 * Interlock was dropped by the call in _vn_lock.
1797 flags &= ~LK_INTERLOCK;
1799 error = VOP_LOCK1(vp, flags, file, line);
1800 } while (error != 0);
1805 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1809 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1810 ("vn_lock: no locktype (%d passed)", flags));
1811 VNPASS(vp->v_holdcnt > 0, vp);
1812 error = VOP_LOCK1(vp, flags, file, line);
1813 if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1814 return (_vn_lock_fallback(vp, flags, file, line, error));
1819 * File table vnode close routine.
1822 vn_closefile(struct file *fp, struct thread *td)
1830 fp->f_ops = &badfileops;
1831 ref = (fp->f_flag & FHASLOCK) != 0;
1833 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1835 if (__predict_false(ref)) {
1836 lf.l_whence = SEEK_SET;
1839 lf.l_type = F_UNLCK;
1840 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1847 * Preparing to start a filesystem write operation. If the operation is
1848 * permitted, then we bump the count of operations in progress and
1849 * proceed. If a suspend request is in progress, we wait until the
1850 * suspension is over, and then proceed.
1853 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1855 struct mount_pcpu *mpcpu;
1858 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1859 vfs_op_thread_enter(mp, mpcpu)) {
1860 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1861 vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1862 vfs_op_thread_exit(mp, mpcpu);
1867 mtx_assert(MNT_MTX(mp), MA_OWNED);
1874 * Check on status of suspension.
1876 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1877 mp->mnt_susp_owner != curthread) {
1879 if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) {
1880 if (flags & V_PCATCH)
1883 mflags |= (PUSER - 1);
1884 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1885 if (flags & V_NOWAIT) {
1886 error = EWOULDBLOCK;
1889 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1895 if (flags & V_XSLEEP)
1897 mp->mnt_writeopcount++;
1899 if (error != 0 || (flags & V_XSLEEP) != 0)
1906 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1911 KASSERT((flags & ~V_VALID_FLAGS) == 0,
1912 ("%s: invalid flags passed %d\n", __func__, flags));
1916 * If a vnode is provided, get and return the mount point that
1917 * to which it will write.
1920 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1922 if (error != EOPNOTSUPP)
1927 if ((mp = *mpp) == NULL)
1931 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1933 * As long as a vnode is not provided we need to acquire a
1934 * refcount for the provided mountpoint too, in order to
1935 * emulate a vfs_ref().
1940 return (vn_start_write_refed(mp, flags, false));
1944 * Secondary suspension. Used by operations such as vop_inactive
1945 * routines that are needed by the higher level functions. These
1946 * are allowed to proceed until all the higher level functions have
1947 * completed (indicated by mnt_writeopcount dropping to zero). At that
1948 * time, these operations are halted until the suspension is over.
1951 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1956 KASSERT((flags & ~V_VALID_FLAGS) == 0,
1957 ("%s: invalid flags passed %d\n", __func__, flags));
1961 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1963 if (error != EOPNOTSUPP)
1969 * If we are not suspended or have not yet reached suspended
1970 * mode, then let the operation proceed.
1972 if ((mp = *mpp) == NULL)
1976 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1978 * As long as a vnode is not provided we need to acquire a
1979 * refcount for the provided mountpoint too, in order to
1980 * emulate a vfs_ref().
1985 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1986 mp->mnt_secondary_writes++;
1987 mp->mnt_secondary_accwrites++;
1991 if (flags & V_NOWAIT) {
1994 return (EWOULDBLOCK);
1997 * Wait for the suspension to finish.
2000 if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) {
2001 if (flags & V_PCATCH)
2004 mflags |= (PUSER - 1) | PDROP;
2005 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,
2567 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2568 ("%s: Wrong command %lu", __func__, cmd));
2569 ASSERT_VOP_LOCKED(vp, "vn_bmap_seekhole_locked");
2571 if (vp->v_type != VREG) {
2575 error = VOP_GETATTR(vp, &va, cred);
2579 if (noff < 0 || noff >= va.va_size) {
2583 bsize = vp->v_mount->mnt_stat.f_iosize;
2584 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2586 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2587 if (error == EOPNOTSUPP) {
2591 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2592 (bnp != -1 && cmd == FIOSEEKDATA)) {
2599 if (noff > va.va_size)
2601 /* noff == va.va_size. There is an implicit hole at the end of file. */
2602 if (cmd == FIOSEEKDATA)
2611 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2615 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2616 ("%s: Wrong command %lu", __func__, cmd));
2618 if (vn_lock(vp, LK_SHARED) != 0)
2620 error = vn_bmap_seekhole_locked(vp, cmd, off, cred);
2626 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2631 off_t foffset, size;
2634 cred = td->td_ucred;
2636 foffset = foffset_lock(fp, 0);
2637 noneg = (vp->v_type != VCHR);
2643 (offset > 0 && foffset > OFF_MAX - offset))) {
2650 vn_lock(vp, LK_SHARED | LK_RETRY);
2651 error = VOP_GETATTR(vp, &vattr, cred);
2657 * If the file references a disk device, then fetch
2658 * the media size and use that to determine the ending
2661 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2662 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2663 vattr.va_size = size;
2665 (vattr.va_size > OFF_MAX ||
2666 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2670 offset += vattr.va_size;
2675 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2676 if (error == ENOTTY)
2680 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2681 if (error == ENOTTY)
2687 if (error == 0 && noneg && offset < 0)
2691 VFS_KNOTE_UNLOCKED(vp, 0);
2692 td->td_uretoff.tdu_off = offset;
2694 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2699 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2705 * Grant permission if the caller is the owner of the file, or
2706 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2707 * on the file. If the time pointer is null, then write
2708 * permission on the file is also sufficient.
2710 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2711 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2712 * will be allowed to set the times [..] to the current
2715 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2716 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2717 error = VOP_ACCESS(vp, VWRITE, cred, td);
2722 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2727 if (fp->f_type == DTYPE_FIFO)
2728 kif->kf_type = KF_TYPE_FIFO;
2730 kif->kf_type = KF_TYPE_VNODE;
2733 FILEDESC_SUNLOCK(fdp);
2734 error = vn_fill_kinfo_vnode(vp, kif);
2736 FILEDESC_SLOCK(fdp);
2741 vn_fill_junk(struct kinfo_file *kif)
2746 * Simulate vn_fullpath returning changing values for a given
2747 * vp during e.g. coredump.
2749 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2750 olen = strlen(kif->kf_path);
2752 strcpy(&kif->kf_path[len - 1], "$");
2754 for (; olen < len; olen++)
2755 strcpy(&kif->kf_path[olen], "A");
2759 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2762 char *fullpath, *freepath;
2765 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2768 error = vn_fullpath(vp, &fullpath, &freepath);
2770 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2772 if (freepath != NULL)
2773 free(freepath, M_TEMP);
2775 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2780 * Retrieve vnode attributes.
2782 va.va_fsid = VNOVAL;
2784 vn_lock(vp, LK_SHARED | LK_RETRY);
2785 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2789 if (va.va_fsid != VNOVAL)
2790 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2792 kif->kf_un.kf_file.kf_file_fsid =
2793 vp->v_mount->mnt_stat.f_fsid.val[0];
2794 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2795 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2796 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2797 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2798 kif->kf_un.kf_file.kf_file_size = va.va_size;
2799 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2800 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2801 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2806 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2807 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2811 struct pmckern_map_in pkm;
2817 boolean_t writecounted;
2820 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2821 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2823 * POSIX shared-memory objects are defined to have
2824 * kernel persistence, and are not defined to support
2825 * read(2)/write(2) -- or even open(2). Thus, we can
2826 * use MAP_ASYNC to trade on-disk coherence for speed.
2827 * The shm_open(3) library routine turns on the FPOSIXSHM
2828 * flag to request this behavior.
2830 if ((fp->f_flag & FPOSIXSHM) != 0)
2831 flags |= MAP_NOSYNC;
2836 * Ensure that file and memory protections are
2837 * compatible. Note that we only worry about
2838 * writability if mapping is shared; in this case,
2839 * current and max prot are dictated by the open file.
2840 * XXX use the vnode instead? Problem is: what
2841 * credentials do we use for determination? What if
2842 * proc does a setuid?
2845 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2846 maxprot = VM_PROT_NONE;
2847 if ((prot & VM_PROT_EXECUTE) != 0)
2850 maxprot = VM_PROT_EXECUTE;
2851 if ((fp->f_flag & FREAD) != 0)
2852 maxprot |= VM_PROT_READ;
2853 else if ((prot & VM_PROT_READ) != 0)
2857 * If we are sharing potential changes via MAP_SHARED and we
2858 * are trying to get write permission although we opened it
2859 * without asking for it, bail out.
2861 if ((flags & MAP_SHARED) != 0) {
2862 if ((fp->f_flag & FWRITE) != 0)
2863 maxprot |= VM_PROT_WRITE;
2864 else if ((prot & VM_PROT_WRITE) != 0)
2867 maxprot |= VM_PROT_WRITE;
2868 cap_maxprot |= VM_PROT_WRITE;
2870 maxprot &= cap_maxprot;
2873 * For regular files and shared memory, POSIX requires that
2874 * the value of foff be a legitimate offset within the data
2875 * object. In particular, negative offsets are invalid.
2876 * Blocking negative offsets and overflows here avoids
2877 * possible wraparound or user-level access into reserved
2878 * ranges of the data object later. In contrast, POSIX does
2879 * not dictate how offsets are used by device drivers, so in
2880 * the case of a device mapping a negative offset is passed
2887 foff > OFF_MAX - size)
2890 writecounted = FALSE;
2891 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2892 &foff, &object, &writecounted);
2895 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2896 foff, writecounted, td);
2899 * If this mapping was accounted for in the vnode's
2900 * writecount, then undo that now.
2903 vm_pager_release_writecount(object, 0, size);
2904 vm_object_deallocate(object);
2907 /* Inform hwpmc(4) if an executable is being mapped. */
2908 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2909 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2911 pkm.pm_address = (uintptr_t) *addr;
2912 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2920 vn_fsid(struct vnode *vp, struct vattr *va)
2924 f = &vp->v_mount->mnt_stat.f_fsid;
2925 va->va_fsid = (uint32_t)f->val[1];
2926 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2927 va->va_fsid += (uint32_t)f->val[0];
2931 vn_fsync_buf(struct vnode *vp, int waitfor)
2933 struct buf *bp, *nbp;
2936 int error, maxretry;
2939 maxretry = 10000; /* large, arbitrarily chosen */
2941 if (vp->v_type == VCHR) {
2943 mp = vp->v_rdev->si_mountpt;
2950 * MARK/SCAN initialization to avoid infinite loops.
2952 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2953 bp->b_vflags &= ~BV_SCANNED;
2958 * Flush all dirty buffers associated with a vnode.
2961 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2962 if ((bp->b_vflags & BV_SCANNED) != 0)
2964 bp->b_vflags |= BV_SCANNED;
2965 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2966 if (waitfor != MNT_WAIT)
2969 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2970 BO_LOCKPTR(bo)) != 0) {
2977 KASSERT(bp->b_bufobj == bo,
2978 ("bp %p wrong b_bufobj %p should be %p",
2979 bp, bp->b_bufobj, bo));
2980 if ((bp->b_flags & B_DELWRI) == 0)
2981 panic("fsync: not dirty");
2982 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2988 if (maxretry < 1000)
2989 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2995 * If synchronous the caller expects us to completely resolve all
2996 * dirty buffers in the system. Wait for in-progress I/O to
2997 * complete (which could include background bitmap writes), then
2998 * retry if dirty blocks still exist.
3000 if (waitfor == MNT_WAIT) {
3001 bufobj_wwait(bo, 0, 0);
3002 if (bo->bo_dirty.bv_cnt > 0) {
3004 * If we are unable to write any of these buffers
3005 * then we fail now rather than trying endlessly
3006 * to write them out.
3008 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
3009 if ((error = bp->b_error) != 0)
3011 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
3012 (error == 0 && --maxretry >= 0))
3020 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
3026 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
3027 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
3028 * to do the actual copy.
3029 * vn_generic_copy_file_range() is factored out, so it can be called
3030 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
3031 * different file systems.
3034 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
3035 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
3036 struct ucred *outcred, struct thread *fsize_td)
3043 *lenp = 0; /* For error returns. */
3046 /* Do some sanity checks on the arguments. */
3047 if (invp->v_type == VDIR || outvp->v_type == VDIR)
3049 else if (*inoffp < 0 || *outoffp < 0 ||
3050 invp->v_type != VREG || outvp->v_type != VREG)
3055 /* Ensure offset + len does not wrap around. */
3058 if (uval > INT64_MAX)
3059 len = INT64_MAX - *inoffp;
3062 if (uval > INT64_MAX)
3063 len = INT64_MAX - *outoffp;
3068 * If the two vnode are for the same file system, call
3069 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
3070 * which can handle copies across multiple file systems.
3073 if (invp->v_mount == outvp->v_mount)
3074 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
3075 lenp, flags, incred, outcred, fsize_td);
3077 error = vn_generic_copy_file_range(invp, inoffp, outvp,
3078 outoffp, lenp, flags, incred, outcred, fsize_td);
3084 * Test len bytes of data starting at dat for all bytes == 0.
3085 * Return true if all bytes are zero, false otherwise.
3086 * Expects dat to be well aligned.
3089 mem_iszero(void *dat, int len)
3095 for (p = dat; len > 0; len -= sizeof(*p), p++) {
3096 if (len >= sizeof(*p)) {
3100 cp = (const char *)p;
3101 for (i = 0; i < len; i++, cp++)
3110 * Look for a hole in the output file and, if found, adjust *outoffp
3111 * and *xferp to skip past the hole.
3112 * *xferp is the entire hole length to be written and xfer2 is how many bytes
3113 * to be written as 0's upon return.
3116 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3117 off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3122 if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3123 *dataoffp = *outoffp;
3124 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3127 *holeoffp = *dataoffp;
3128 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3131 if (error != 0 || *holeoffp == *dataoffp) {
3133 * Since outvp is unlocked, it may be possible for
3134 * another thread to do a truncate(), lseek(), write()
3135 * creating a hole at startoff between the above
3136 * VOP_IOCTL() calls, if the other thread does not do
3138 * If that happens, *holeoffp == *dataoffp and finding
3139 * the hole has failed, so disable vn_skip_hole().
3141 *holeoffp = -1; /* Disable use of vn_skip_hole(). */
3144 KASSERT(*dataoffp >= *outoffp,
3145 ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3146 (intmax_t)*dataoffp, (intmax_t)*outoffp));
3147 KASSERT(*holeoffp > *dataoffp,
3148 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3149 (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3153 * If there is a hole before the data starts, advance *outoffp and
3154 * *xferp past the hole.
3156 if (*dataoffp > *outoffp) {
3157 delta = *dataoffp - *outoffp;
3158 if (delta >= *xferp) {
3159 /* Entire *xferp is a hole. */
3166 xfer2 = MIN(xfer2, *xferp);
3170 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3171 * that the write ends at the start of the hole.
3172 * *holeoffp should always be greater than *outoffp, but for the
3173 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3176 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3177 xfer2 = *holeoffp - *outoffp;
3182 * Write an xfer sized chunk to outvp in blksize blocks from dat.
3183 * dat is a maximum of blksize in length and can be written repeatedly in
3185 * If growfile == true, just grow the file via vn_truncate_locked() instead
3186 * of doing actual writes.
3187 * If checkhole == true, a hole is being punched, so skip over any hole
3188 * already in the output file.
3191 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3192 u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3195 off_t dataoff, holeoff, xfer2;
3199 * Loop around doing writes of blksize until write has been completed.
3200 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3201 * done for each iteration, since the xfer argument can be very
3202 * large if there is a large hole to punch in the output file.
3207 xfer2 = MIN(xfer, blksize);
3210 * Punching a hole. Skip writing if there is
3211 * already a hole in the output file.
3213 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3214 &dataoff, &holeoff, cred);
3219 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3224 error = vn_start_write(outvp, &mp, V_WAIT);
3228 error = vn_lock(outvp, LK_EXCLUSIVE);
3230 error = vn_truncate_locked(outvp, outoff + xfer,
3235 error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3237 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3238 outoff, UIO_SYSSPACE, IO_NODELOCKED,
3239 curthread->td_ucred, cred, NULL, curthread);
3246 vn_finished_write(mp);
3247 } while (!growfile && xfer > 0 && error == 0);
3252 * Copy a byte range of one file to another. This function can handle the
3253 * case where invp and outvp are on different file systems.
3254 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3255 * is no better file system specific way to do it.
3258 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3259 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3260 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3262 struct vattr va, inva;
3264 off_t startoff, endoff, xfer, xfer2;
3266 int error, interrupted;
3267 bool cantseek, readzeros, eof, lastblock, holetoeof;
3268 ssize_t aresid, r = 0;
3269 size_t copylen, len, rem, savlen;
3271 long holein, holeout;
3272 struct timespec curts, endts;
3274 holein = holeout = 0;
3275 savlen = len = *lenp;
3280 error = vn_lock(invp, LK_SHARED);
3283 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3286 error = VOP_GETATTR(invp, &inva, incred);
3292 error = vn_start_write(outvp, &mp, V_WAIT);
3294 error = vn_lock(outvp, LK_EXCLUSIVE);
3297 * If fsize_td != NULL, do a vn_rlimit_fsizex() call,
3298 * now that outvp is locked.
3300 if (fsize_td != NULL) {
3303 io.uio_offset = *outoffp;
3305 error = vn_rlimit_fsizex(outvp, &io, 0, &r, fsize_td);
3306 len = savlen = io.uio_resid;
3308 * No need to call vn_rlimit_fsizex_res before return,
3309 * since the uio is local.
3312 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3315 * Holes that are past EOF do not need to be written as a block
3316 * of zero bytes. So, truncate the output file as far as
3317 * possible and then use va.va_size to decide if writing 0
3318 * bytes is necessary in the loop below.
3321 error = VOP_GETATTR(outvp, &va, outcred);
3322 if (error == 0 && va.va_size > *outoffp && va.va_size <=
3325 error = mac_vnode_check_write(curthread->td_ucred,
3329 error = vn_truncate_locked(outvp, *outoffp,
3332 va.va_size = *outoffp;
3337 vn_finished_write(mp);
3342 * Set the blksize to the larger of the hole sizes for invp and outvp.
3343 * If hole sizes aren't available, set the blksize to the larger
3344 * f_iosize of invp and outvp.
3345 * This code expects the hole sizes and f_iosizes to be powers of 2.
3346 * This value is clipped at 4Kbytes and 1Mbyte.
3348 blksize = MAX(holein, holeout);
3350 /* Clip len to end at an exact multiple of hole size. */
3352 rem = *inoffp % blksize;
3354 rem = blksize - rem;
3355 if (len > rem && len - rem > blksize)
3356 len = savlen = rounddown(len - rem, blksize) + rem;
3360 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3361 outvp->v_mount->mnt_stat.f_iosize);
3364 else if (blksize > 1024 * 1024)
3365 blksize = 1024 * 1024;
3366 dat = malloc(blksize, M_TEMP, M_WAITOK);
3369 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3370 * to find holes. Otherwise, just scan the read block for all 0s
3371 * in the inner loop where the data copying is done.
3372 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3373 * support holes on the server, but do not support FIOSEEKHOLE.
3374 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3375 * that this function should return after 1second with a partial
3378 if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3379 getnanouptime(&endts);
3382 timespecclear(&endts);
3383 holetoeof = eof = false;
3384 while (len > 0 && error == 0 && !eof && interrupted == 0) {
3385 endoff = 0; /* To shut up compilers. */
3391 * Find the next data area. If there is just a hole to EOF,
3392 * FIOSEEKDATA should fail with ENXIO.
3393 * (I do not know if any file system will report a hole to
3394 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3395 * will fail for those file systems.)
3397 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3398 * the code just falls through to the inner copy loop.
3402 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3404 if (error == ENXIO) {
3405 startoff = endoff = inva.va_size;
3406 eof = holetoeof = true;
3410 if (error == 0 && !holetoeof) {
3412 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3415 * Since invp is unlocked, it may be possible for
3416 * another thread to do a truncate(), lseek(), write()
3417 * creating a hole at startoff between the above
3418 * VOP_IOCTL() calls, if the other thread does not do
3420 * If that happens, startoff == endoff and finding
3421 * the hole has failed, so set an error.
3423 if (error == 0 && startoff == endoff)
3424 error = EINVAL; /* Any error. Reset to 0. */
3427 if (startoff > *inoffp) {
3428 /* Found hole before data block. */
3429 xfer = MIN(startoff - *inoffp, len);
3430 if (*outoffp < va.va_size) {
3431 /* Must write 0s to punch hole. */
3432 xfer2 = MIN(va.va_size - *outoffp,
3434 memset(dat, 0, MIN(xfer2, blksize));
3435 error = vn_write_outvp(outvp, dat,
3436 *outoffp, xfer2, blksize, false,
3437 holeout > 0, outcred);
3440 if (error == 0 && *outoffp + xfer >
3441 va.va_size && (xfer == len || holetoeof)) {
3442 /* Grow output file (hole at end). */
3443 error = vn_write_outvp(outvp, dat,
3444 *outoffp, xfer, blksize, true,
3452 interrupted = sig_intr();
3453 if (timespecisset(&endts) &&
3455 getnanouptime(&curts);
3456 if (timespeccmp(&curts,
3464 copylen = MIN(len, endoff - startoff);
3476 * Set first xfer to end at a block boundary, so that
3477 * holes are more likely detected in the loop below via
3478 * the for all bytes 0 method.
3480 xfer -= (*inoffp % blksize);
3482 /* Loop copying the data block. */
3483 while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3486 error = vn_lock(invp, LK_SHARED);
3489 error = vn_rdwr(UIO_READ, invp, dat, xfer,
3490 startoff, UIO_SYSSPACE, IO_NODELOCKED,
3491 curthread->td_ucred, incred, &aresid,
3495 if (error == 0 && aresid > 0) {
3496 /* Stop the copy at EOF on the input file. */
3503 * Skip the write for holes past the initial EOF
3504 * of the output file, unless this is the last
3505 * write of the output file at EOF.
3507 readzeros = cantseek ? mem_iszero(dat, xfer) :
3511 if (!cantseek || *outoffp < va.va_size ||
3512 lastblock || !readzeros)
3513 error = vn_write_outvp(outvp, dat,
3514 *outoffp, xfer, blksize,
3515 readzeros && lastblock &&
3516 *outoffp >= va.va_size, false,
3525 interrupted = sig_intr();
3526 if (timespecisset(&endts) &&
3528 getnanouptime(&curts);
3529 if (timespeccmp(&curts,
3541 *lenp = savlen - len;
3547 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3551 off_t olen, ooffset;
3554 int audited_vnode1 = 0;
3558 if (vp->v_type != VREG)
3561 /* Allocating blocks may take a long time, so iterate. */
3568 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
3571 error = vn_lock(vp, LK_EXCLUSIVE);
3573 vn_finished_write(mp);
3577 if (!audited_vnode1) {
3578 AUDIT_ARG_VNODE1(vp);
3583 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3586 error = VOP_ALLOCATE(vp, &offset, &len, 0,
3589 vn_finished_write(mp);
3591 if (olen + ooffset != offset + len) {
3592 panic("offset + len changed from %jx/%jx to %jx/%jx",
3593 ooffset, olen, offset, len);
3595 if (error != 0 || len == 0)
3597 KASSERT(olen > len, ("Iteration did not make progress?"));
3605 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags,
3606 int ioflag, struct ucred *cred, struct ucred *active_cred,
3607 struct ucred *file_cred)
3614 bool audited_vnode1 = false;
3623 if ((ioflag & (IO_NODELOCKED | IO_RANGELOCKED)) == 0)
3624 rl_cookie = vn_rangelock_wlock(vp, off, off + len);
3625 while (len > 0 && error == 0) {
3627 * Try to deallocate the longest range in one pass.
3628 * In case a pass takes too long to be executed, it returns
3629 * partial result. The residue will be proceeded in the next
3633 if ((ioflag & IO_NODELOCKED) == 0) {
3635 if ((error = vn_start_write(vp, &mp,
3636 V_WAIT | V_PCATCH)) != 0)
3638 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
3641 if (!audited_vnode1) {
3642 AUDIT_ARG_VNODE1(vp);
3643 audited_vnode1 = true;
3648 if ((ioflag & IO_NOMACCHECK) == 0)
3649 error = mac_vnode_check_write(active_cred, file_cred,
3653 error = VOP_DEALLOCATE(vp, &off, &len, flags, ioflag,
3656 if ((ioflag & IO_NODELOCKED) == 0) {
3659 vn_finished_write(mp);
3663 if (error == 0 && len != 0)
3667 if (rl_cookie != NULL)
3668 vn_rangelock_unlock(vp, rl_cookie);
3675 * This function is supposed to be used in the situations where the deallocation
3676 * is not triggered by a user request.
3679 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags,
3680 int ioflag, struct ucred *active_cred, struct ucred *file_cred)
3684 if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset ||
3687 if (vp->v_type != VREG)
3690 cred = file_cred != NOCRED ? file_cred : active_cred;
3691 return (vn_deallocate_impl(vp, offset, length, flags, ioflag, cred,
3692 active_cred, file_cred));
3696 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
3697 struct ucred *active_cred, struct thread *td)
3703 KASSERT(cmd == SPACECTL_DEALLOC, ("vn_fspacectl: Invalid cmd"));
3704 KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0,
3705 ("vn_fspacectl: non-zero flags"));
3706 KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset,
3707 ("vn_fspacectl: offset/length overflow or underflow"));
3710 if (vp->v_type != VREG)
3713 ioflag = get_write_ioflag(fp);
3716 case SPACECTL_DEALLOC:
3717 error = vn_deallocate_impl(vp, offset, length, flags, ioflag,
3718 active_cred, active_cred, fp->f_cred);
3721 panic("vn_fspacectl: unknown cmd %d", cmd);
3727 static u_long vn_lock_pair_pause_cnt;
3728 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3729 &vn_lock_pair_pause_cnt, 0,
3730 "Count of vn_lock_pair deadlocks");
3732 u_int vn_lock_pair_pause_max;
3733 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3734 &vn_lock_pair_pause_max, 0,
3735 "Max ticks for vn_lock_pair deadlock avoidance sleep");
3738 vn_lock_pair_pause(const char *wmesg)
3740 atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3741 pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3745 * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3746 * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3747 * must be unlocked. Same for vp2 and vp2_locked. One of the vnodes
3750 * The function returns with both vnodes exclusively locked, and
3751 * guarantees that it does not create lock order reversal with other
3752 * threads during its execution. Both vnodes could be unlocked
3753 * temporary (and reclaimed).
3756 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3761 if (vp1 == NULL && vp2 == NULL)
3765 ASSERT_VOP_ELOCKED(vp1, "vp1");
3767 ASSERT_VOP_UNLOCKED(vp1, "vp1");
3773 ASSERT_VOP_ELOCKED(vp2, "vp2");
3775 ASSERT_VOP_UNLOCKED(vp2, "vp2");
3779 if (!vp1_locked && !vp2_locked) {
3780 vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3785 if (vp1_locked && vp2_locked)
3787 if (vp1_locked && vp2 != NULL) {
3789 error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3790 __FILE__, __LINE__);
3795 vn_lock_pair_pause("vlp1");
3797 vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3800 if (vp2_locked && vp1 != NULL) {
3802 error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3803 __FILE__, __LINE__);
3808 vn_lock_pair_pause("vlp2");
3810 vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3815 ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3817 ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3821 vn_lktype_write(struct mount *mp, struct vnode *vp)
3823 if (MNT_SHARED_WRITES(mp) ||
3824 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
3826 return (LK_EXCLUSIVE);