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)
220 * Common code for vnode open operations via a name lookup.
221 * Lookup the vnode and invoke VOP_CREATE if needed.
222 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
224 * Note that this does NOT free nameidata for the successful case,
225 * due to the NDINIT being done elsewhere.
228 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
229 struct ucred *cred, struct file *fp)
234 struct vattr *vap = &vat;
241 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
242 O_EXCL | O_DIRECTORY) ||
243 (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
245 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
246 ndp->ni_cnd.cn_nameiop = CREATE;
247 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
249 * Set NOCACHE to avoid flushing the cache when
250 * rolling in many files at once.
252 * Set NC_KEEPPOSENTRY to keep positive entries if they already
253 * exist despite NOCACHE.
255 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
256 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
257 ndp->ni_cnd.cn_flags |= FOLLOW;
258 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
260 if ((error = namei(ndp)) != 0)
262 if (ndp->ni_vp == NULL) {
265 vap->va_mode = cmode;
267 vap->va_vaflags |= VA_EXCLUSIVE;
268 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
269 NDFREE(ndp, NDF_ONLY_PNBUF);
271 if ((error = vn_start_write(NULL, &mp,
272 V_XSLEEP | PCATCH)) != 0)
277 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
278 ndp->ni_cnd.cn_flags |= MAKEENTRY;
280 error = mac_vnode_check_create(cred, ndp->ni_dvp,
284 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
287 if (error == 0 && (fmode & O_EXCL) != 0 &&
288 (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
290 vp->v_iflag |= VI_FOPENING;
294 VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
296 vn_finished_write(mp);
298 NDFREE(ndp, NDF_ONLY_PNBUF);
299 if (error == ERELOOKUP) {
307 if (ndp->ni_dvp == ndp->ni_vp)
313 if (fmode & O_EXCL) {
317 if (vp->v_type == VDIR) {
324 ndp->ni_cnd.cn_nameiop = LOOKUP;
325 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
326 ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
328 if ((fmode & FWRITE) == 0)
329 ndp->ni_cnd.cn_flags |= LOCKSHARED;
330 if ((error = namei(ndp)) != 0)
334 error = vn_open_vnode(vp, fmode, cred, curthread, fp);
337 vp->v_iflag &= ~VI_FOPENING;
346 NDFREE(ndp, NDF_ONLY_PNBUF);
354 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
357 int error, lock_flags, type;
359 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
360 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
362 KASSERT(fp != NULL, ("open with flock requires fp"));
363 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
366 lock_flags = VOP_ISLOCKED(vp);
369 lf.l_whence = SEEK_SET;
372 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
374 if ((fmode & FNONBLOCK) == 0)
376 if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
378 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
380 fp->f_flag |= FHASLOCK;
382 vn_lock(vp, lock_flags | LK_RETRY);
387 * Common code for vnode open operations once a vnode is located.
388 * Check permissions, and call the VOP_OPEN routine.
391 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
392 struct thread *td, struct file *fp)
397 if (vp->v_type == VLNK) {
398 if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
401 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
405 if ((fmode & O_PATH) == 0) {
406 if (vp->v_type == VSOCK)
408 if ((fmode & (FWRITE | O_TRUNC)) != 0) {
409 if (vp->v_type == VDIR)
413 if ((fmode & FREAD) != 0)
415 if ((fmode & O_APPEND) && (fmode & FWRITE))
418 if ((fmode & O_CREAT) != 0)
422 if ((fmode & FEXEC) != 0)
425 if ((fmode & O_VERIFY) != 0)
427 error = mac_vnode_check_open(cred, vp, accmode);
431 accmode &= ~(VCREAT | VVERIFY);
433 if ((fmode & O_CREAT) == 0 && accmode != 0) {
434 error = VOP_ACCESS(vp, accmode, cred, td);
438 if ((fmode & O_PATH) != 0) {
439 if (vp->v_type != VFIFO && vp->v_type != VSOCK &&
440 VOP_ACCESS(vp, VREAD, cred, td) == 0)
441 fp->f_flag |= FKQALLOWED;
445 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
446 vn_lock(vp, LK_UPGRADE | LK_RETRY);
447 error = VOP_OPEN(vp, fmode, cred, td, fp);
451 error = vn_open_vnode_advlock(vp, fmode, fp);
452 if (error == 0 && (fmode & FWRITE) != 0) {
453 error = VOP_ADD_WRITECOUNT(vp, 1);
455 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
456 __func__, vp, vp->v_writecount);
461 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
462 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
467 * Arrange the call by having fdrop() to use
468 * vn_closefile(). This is to satisfy
469 * filesystems like devfs or tmpfs, which
470 * override fo_close().
472 fp->f_flag |= FOPENFAILED;
474 if (fp->f_ops == &badfileops) {
475 fp->f_type = DTYPE_VNODE;
481 * If there is no fp, due to kernel-mode open,
482 * we can call VOP_CLOSE() now.
484 if (vp->v_type != VFIFO && (fmode & FWRITE) != 0 &&
485 !MNT_EXTENDED_SHARED(vp->v_mount) &&
486 VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
487 vn_lock(vp, LK_UPGRADE | LK_RETRY);
488 (void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
493 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
499 * Check for write permissions on the specified vnode.
500 * Prototype text segments cannot be written.
504 vn_writechk(struct vnode *vp)
507 ASSERT_VOP_LOCKED(vp, "vn_writechk");
509 * If there's shared text associated with
510 * the vnode, try to free it up once. If
511 * we fail, we can't allow writing.
523 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
524 struct thread *td, bool keep_ref)
527 int error, lock_flags;
529 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
530 MNT_EXTENDED_SHARED(vp->v_mount))
531 lock_flags = LK_SHARED;
533 lock_flags = LK_EXCLUSIVE;
535 vn_start_write(vp, &mp, V_WAIT);
536 vn_lock(vp, lock_flags | LK_RETRY);
537 AUDIT_ARG_VNODE1(vp);
538 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
539 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
540 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
541 __func__, vp, vp->v_writecount);
543 error = VOP_CLOSE(vp, flags, file_cred, td);
548 vn_finished_write(mp);
553 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
557 return (vn_close1(vp, flags, file_cred, td, false));
561 * Heuristic to detect sequential operation.
564 sequential_heuristic(struct uio *uio, struct file *fp)
568 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
571 if (fp->f_flag & FRDAHEAD)
572 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
575 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
576 * that the first I/O is normally considered to be slightly
577 * sequential. Seeking to offset 0 doesn't change sequentiality
578 * unless previous seeks have reduced f_seqcount to 0, in which
579 * case offset 0 is not special.
581 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
582 uio->uio_offset == fp->f_nextoff[rw]) {
584 * f_seqcount is in units of fixed-size blocks so that it
585 * depends mainly on the amount of sequential I/O and not
586 * much on the number of sequential I/O's. The fixed size
587 * of 16384 is hard-coded here since it is (not quite) just
588 * a magic size that works well here. This size is more
589 * closely related to the best I/O size for real disks than
590 * to any block size used by software.
592 if (uio->uio_resid >= IO_SEQMAX * 16384)
593 fp->f_seqcount[rw] = IO_SEQMAX;
595 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
596 if (fp->f_seqcount[rw] > IO_SEQMAX)
597 fp->f_seqcount[rw] = IO_SEQMAX;
599 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
602 /* Not sequential. Quickly draw-down sequentiality. */
603 if (fp->f_seqcount[rw] > 1)
604 fp->f_seqcount[rw] = 1;
606 fp->f_seqcount[rw] = 0;
611 * Package up an I/O request on a vnode into a uio and do it.
614 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
615 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
616 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
623 struct vn_io_fault_args args;
624 int error, lock_flags;
626 if (offset < 0 && vp->v_type != VCHR)
628 auio.uio_iov = &aiov;
630 aiov.iov_base = base;
632 auio.uio_resid = len;
633 auio.uio_offset = offset;
634 auio.uio_segflg = segflg;
639 if ((ioflg & IO_NODELOCKED) == 0) {
640 if ((ioflg & IO_RANGELOCKED) == 0) {
641 if (rw == UIO_READ) {
642 rl_cookie = vn_rangelock_rlock(vp, offset,
644 } else if ((ioflg & IO_APPEND) != 0) {
645 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
647 rl_cookie = vn_rangelock_wlock(vp, offset,
653 if (rw == UIO_WRITE) {
654 if (vp->v_type != VCHR &&
655 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
658 lock_flags = vn_lktype_write(mp, vp);
660 lock_flags = LK_SHARED;
661 vn_lock(vp, lock_flags | LK_RETRY);
665 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
667 if ((ioflg & IO_NOMACCHECK) == 0) {
669 error = mac_vnode_check_read(active_cred, file_cred,
672 error = mac_vnode_check_write(active_cred, file_cred,
677 if (file_cred != NULL)
681 if (do_vn_io_fault(vp, &auio)) {
682 args.kind = VN_IO_FAULT_VOP;
685 args.args.vop_args.vp = vp;
686 error = vn_io_fault1(vp, &auio, &args, td);
687 } else if (rw == UIO_READ) {
688 error = VOP_READ(vp, &auio, ioflg, cred);
689 } else /* if (rw == UIO_WRITE) */ {
690 error = VOP_WRITE(vp, &auio, ioflg, cred);
694 *aresid = auio.uio_resid;
696 if (auio.uio_resid && error == 0)
698 if ((ioflg & IO_NODELOCKED) == 0) {
701 vn_finished_write(mp);
704 if (rl_cookie != NULL)
705 vn_rangelock_unlock(vp, rl_cookie);
710 * Package up an I/O request on a vnode into a uio and do it. The I/O
711 * request is split up into smaller chunks and we try to avoid saturating
712 * the buffer cache while potentially holding a vnode locked, so we
713 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
714 * to give other processes a chance to lock the vnode (either other processes
715 * core'ing the same binary, or unrelated processes scanning the directory).
718 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
719 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
720 struct ucred *file_cred, size_t *aresid, struct thread *td)
729 * Force `offset' to a multiple of MAXBSIZE except possibly
730 * for the first chunk, so that filesystems only need to
731 * write full blocks except possibly for the first and last
734 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
738 if (rw != UIO_READ && vp->v_type == VREG)
741 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
742 ioflg, active_cred, file_cred, &iaresid, td);
743 len -= chunk; /* aresid calc already includes length */
747 base = (char *)base + chunk;
748 kern_yield(PRI_USER);
751 *aresid = len + iaresid;
755 #if OFF_MAX <= LONG_MAX
757 foffset_lock(struct file *fp, int flags)
759 volatile short *flagsp;
763 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
765 if ((flags & FOF_NOLOCK) != 0)
766 return (atomic_load_long(&fp->f_offset));
769 * According to McKusick the vn lock was protecting f_offset here.
770 * It is now protected by the FOFFSET_LOCKED flag.
772 flagsp = &fp->f_vnread_flags;
773 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
774 return (atomic_load_long(&fp->f_offset));
776 sleepq_lock(&fp->f_vnread_flags);
777 state = atomic_load_16(flagsp);
779 if ((state & FOFFSET_LOCKED) == 0) {
780 if (!atomic_fcmpset_acq_16(flagsp, &state,
785 if ((state & FOFFSET_LOCK_WAITING) == 0) {
786 if (!atomic_fcmpset_acq_16(flagsp, &state,
787 state | FOFFSET_LOCK_WAITING))
791 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
792 sleepq_wait(&fp->f_vnread_flags, PUSER -1);
794 sleepq_lock(&fp->f_vnread_flags);
795 state = atomic_load_16(flagsp);
797 res = atomic_load_long(&fp->f_offset);
798 sleepq_release(&fp->f_vnread_flags);
803 foffset_unlock(struct file *fp, off_t val, int flags)
805 volatile short *flagsp;
808 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
810 if ((flags & FOF_NOUPDATE) == 0)
811 atomic_store_long(&fp->f_offset, val);
812 if ((flags & FOF_NEXTOFF_R) != 0)
813 fp->f_nextoff[UIO_READ] = val;
814 if ((flags & FOF_NEXTOFF_W) != 0)
815 fp->f_nextoff[UIO_WRITE] = val;
817 if ((flags & FOF_NOLOCK) != 0)
820 flagsp = &fp->f_vnread_flags;
821 state = atomic_load_16(flagsp);
822 if ((state & FOFFSET_LOCK_WAITING) == 0 &&
823 atomic_cmpset_rel_16(flagsp, state, 0))
826 sleepq_lock(&fp->f_vnread_flags);
827 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
828 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
829 fp->f_vnread_flags = 0;
830 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
831 sleepq_release(&fp->f_vnread_flags);
835 foffset_lock(struct file *fp, int flags)
840 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
842 mtxp = mtx_pool_find(mtxpool_sleep, fp);
844 if ((flags & FOF_NOLOCK) == 0) {
845 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
846 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
847 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
850 fp->f_vnread_flags |= FOFFSET_LOCKED;
858 foffset_unlock(struct file *fp, off_t val, int flags)
862 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
864 mtxp = mtx_pool_find(mtxpool_sleep, fp);
866 if ((flags & FOF_NOUPDATE) == 0)
868 if ((flags & FOF_NEXTOFF_R) != 0)
869 fp->f_nextoff[UIO_READ] = val;
870 if ((flags & FOF_NEXTOFF_W) != 0)
871 fp->f_nextoff[UIO_WRITE] = val;
872 if ((flags & FOF_NOLOCK) == 0) {
873 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
874 ("Lost FOFFSET_LOCKED"));
875 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
876 wakeup(&fp->f_vnread_flags);
877 fp->f_vnread_flags = 0;
884 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
887 if ((flags & FOF_OFFSET) == 0)
888 uio->uio_offset = foffset_lock(fp, flags);
892 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
895 if ((flags & FOF_OFFSET) == 0)
896 foffset_unlock(fp, uio->uio_offset, flags);
900 get_advice(struct file *fp, struct uio *uio)
905 ret = POSIX_FADV_NORMAL;
906 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
909 mtxp = mtx_pool_find(mtxpool_sleep, fp);
911 if (fp->f_advice != NULL &&
912 uio->uio_offset >= fp->f_advice->fa_start &&
913 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
914 ret = fp->f_advice->fa_advice;
920 get_write_ioflag(struct file *fp)
928 mp = atomic_load_ptr(&vp->v_mount);
930 if ((fp->f_flag & O_DIRECT) != 0)
933 if ((fp->f_flag & O_FSYNC) != 0 ||
934 (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
938 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
939 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
940 * fall back to full O_SYNC behavior.
942 if ((fp->f_flag & O_DSYNC) != 0)
943 ioflag |= IO_SYNC | IO_DATASYNC;
949 vn_read_from_obj(struct vnode *vp, struct uio *uio)
952 vm_page_t ma[io_hold_cnt + 2];
957 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
958 obj = atomic_load_ptr(&vp->v_object);
960 return (EJUSTRETURN);
963 * Depends on type stability of vm_objects.
965 vm_object_pip_add(obj, 1);
966 if ((obj->flags & OBJ_DEAD) != 0) {
968 * Note that object might be already reused from the
969 * vnode, and the OBJ_DEAD flag cleared. This is fine,
970 * we recheck for DOOMED vnode state after all pages
971 * are busied, and retract then.
973 * But we check for OBJ_DEAD to ensure that we do not
974 * busy pages while vm_object_terminate_pages()
975 * processes the queue.
981 resid = uio->uio_resid;
982 off = uio->uio_offset;
983 for (i = 0; resid > 0; i++) {
984 MPASS(i < io_hold_cnt + 2);
985 ma[i] = vm_page_grab_unlocked(obj, atop(off),
986 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
992 * Skip invalid pages. Valid mask can be partial only
993 * at EOF, and we clip later.
995 if (vm_page_none_valid(ma[i])) {
996 vm_page_sunbusy(ma[i]);
1004 error = EJUSTRETURN;
1009 * Check VIRF_DOOMED after we busied our pages. Since
1010 * vgonel() terminates the vnode' vm_object, it cannot
1011 * process past pages busied by us.
1013 if (VN_IS_DOOMED(vp)) {
1014 error = EJUSTRETURN;
1018 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1019 if (resid > uio->uio_resid)
1020 resid = uio->uio_resid;
1023 * Unlocked read of vnp_size is safe because truncation cannot
1024 * pass busied page. But we load vnp_size into a local
1025 * variable so that possible concurrent extension does not
1026 * break calculation.
1028 #if defined(__powerpc__) && !defined(__powerpc64__)
1029 vsz = obj->un_pager.vnp.vnp_size;
1031 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1033 if (uio->uio_offset >= vsz) {
1034 error = EJUSTRETURN;
1037 if (uio->uio_offset + resid > vsz)
1038 resid = vsz - uio->uio_offset;
1040 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1043 for (j = 0; j < i; j++) {
1045 vm_page_reference(ma[j]);
1046 vm_page_sunbusy(ma[j]);
1049 vm_object_pip_wakeup(obj);
1052 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1056 * File table vnode read routine.
1059 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1067 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1069 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1072 if (fp->f_flag & FNONBLOCK)
1073 ioflag |= IO_NDELAY;
1074 if (fp->f_flag & O_DIRECT)
1075 ioflag |= IO_DIRECT;
1078 * Try to read from page cache. VIRF_DOOMED check is racy but
1079 * allows us to avoid unneeded work outright.
1081 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1082 (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1083 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1085 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1088 if (error != EJUSTRETURN)
1092 advice = get_advice(fp, uio);
1093 vn_lock(vp, LK_SHARED | LK_RETRY);
1096 case POSIX_FADV_NORMAL:
1097 case POSIX_FADV_SEQUENTIAL:
1098 case POSIX_FADV_NOREUSE:
1099 ioflag |= sequential_heuristic(uio, fp);
1101 case POSIX_FADV_RANDOM:
1102 /* Disable read-ahead for random I/O. */
1105 orig_offset = uio->uio_offset;
1108 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1111 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1112 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1114 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1115 orig_offset != uio->uio_offset)
1117 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1118 * for the backing file after a POSIX_FADV_NOREUSE
1121 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1122 POSIX_FADV_DONTNEED);
1127 * File table vnode write routine.
1130 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1138 bool need_finished_write;
1140 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1142 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1144 if (vp->v_type == VREG)
1147 if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1148 ioflag |= IO_APPEND;
1149 if ((fp->f_flag & FNONBLOCK) != 0)
1150 ioflag |= IO_NDELAY;
1151 ioflag |= get_write_ioflag(fp);
1154 need_finished_write = false;
1155 if (vp->v_type != VCHR) {
1156 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1159 need_finished_write = true;
1162 advice = get_advice(fp, uio);
1164 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1166 case POSIX_FADV_NORMAL:
1167 case POSIX_FADV_SEQUENTIAL:
1168 case POSIX_FADV_NOREUSE:
1169 ioflag |= sequential_heuristic(uio, fp);
1171 case POSIX_FADV_RANDOM:
1172 /* XXX: Is this correct? */
1175 orig_offset = uio->uio_offset;
1178 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1181 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1182 fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1184 if (need_finished_write)
1185 vn_finished_write(mp);
1186 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1187 orig_offset != uio->uio_offset)
1189 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1190 * for the backing file after a POSIX_FADV_NOREUSE
1193 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1194 POSIX_FADV_DONTNEED);
1200 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1201 * prevent the following deadlock:
1203 * Assume that the thread A reads from the vnode vp1 into userspace
1204 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
1205 * currently not resident, then system ends up with the call chain
1206 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1207 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1208 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1209 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1210 * backed by the pages of vnode vp1, and some page in buf2 is not
1211 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1213 * To prevent the lock order reversal and deadlock, vn_io_fault() does
1214 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1215 * Instead, it first tries to do the whole range i/o with pagefaults
1216 * disabled. If all pages in the i/o buffer are resident and mapped,
1217 * VOP will succeed (ignoring the genuine filesystem errors).
1218 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1219 * i/o in chunks, with all pages in the chunk prefaulted and held
1220 * using vm_fault_quick_hold_pages().
1222 * Filesystems using this deadlock avoidance scheme should use the
1223 * array of the held pages from uio, saved in the curthread->td_ma,
1224 * instead of doing uiomove(). A helper function
1225 * vn_io_fault_uiomove() converts uiomove request into
1226 * uiomove_fromphys() over td_ma array.
1228 * Since vnode locks do not cover the whole i/o anymore, rangelocks
1229 * make the current i/o request atomic with respect to other i/os and
1234 * Decode vn_io_fault_args and perform the corresponding i/o.
1237 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1243 save = vm_fault_disable_pagefaults();
1244 switch (args->kind) {
1245 case VN_IO_FAULT_FOP:
1246 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1247 uio, args->cred, args->flags, td);
1249 case VN_IO_FAULT_VOP:
1250 if (uio->uio_rw == UIO_READ) {
1251 error = VOP_READ(args->args.vop_args.vp, uio,
1252 args->flags, args->cred);
1253 } else if (uio->uio_rw == UIO_WRITE) {
1254 error = VOP_WRITE(args->args.vop_args.vp, uio,
1255 args->flags, args->cred);
1259 panic("vn_io_fault_doio: unknown kind of io %d %d",
1260 args->kind, uio->uio_rw);
1262 vm_fault_enable_pagefaults(save);
1267 vn_io_fault_touch(char *base, const struct uio *uio)
1272 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1278 vn_io_fault_prefault_user(const struct uio *uio)
1281 const struct iovec *iov;
1286 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1287 ("vn_io_fault_prefault userspace"));
1291 resid = uio->uio_resid;
1292 base = iov->iov_base;
1295 error = vn_io_fault_touch(base, uio);
1298 if (len < PAGE_SIZE) {
1300 error = vn_io_fault_touch(base + len - 1, uio);
1305 if (++i >= uio->uio_iovcnt)
1307 iov = uio->uio_iov + i;
1308 base = iov->iov_base;
1320 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1321 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1322 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1323 * into args and call vn_io_fault1() to handle faults during the user
1324 * mode buffer accesses.
1327 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1330 vm_page_t ma[io_hold_cnt + 2];
1331 struct uio *uio_clone, short_uio;
1332 struct iovec short_iovec[1];
1333 vm_page_t *prev_td_ma;
1335 vm_offset_t addr, end;
1338 int error, cnt, saveheld, prev_td_ma_cnt;
1340 if (vn_io_fault_prefault) {
1341 error = vn_io_fault_prefault_user(uio);
1343 return (error); /* Or ignore ? */
1346 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1349 * The UFS follows IO_UNIT directive and replays back both
1350 * uio_offset and uio_resid if an error is encountered during the
1351 * operation. But, since the iovec may be already advanced,
1352 * uio is still in an inconsistent state.
1354 * Cache a copy of the original uio, which is advanced to the redo
1355 * point using UIO_NOCOPY below.
1357 uio_clone = cloneuio(uio);
1358 resid = uio->uio_resid;
1360 short_uio.uio_segflg = UIO_USERSPACE;
1361 short_uio.uio_rw = uio->uio_rw;
1362 short_uio.uio_td = uio->uio_td;
1364 error = vn_io_fault_doio(args, uio, td);
1365 if (error != EFAULT)
1368 atomic_add_long(&vn_io_faults_cnt, 1);
1369 uio_clone->uio_segflg = UIO_NOCOPY;
1370 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1371 uio_clone->uio_segflg = uio->uio_segflg;
1373 saveheld = curthread_pflags_set(TDP_UIOHELD);
1374 prev_td_ma = td->td_ma;
1375 prev_td_ma_cnt = td->td_ma_cnt;
1377 while (uio_clone->uio_resid != 0) {
1378 len = uio_clone->uio_iov->iov_len;
1380 KASSERT(uio_clone->uio_iovcnt >= 1,
1381 ("iovcnt underflow"));
1382 uio_clone->uio_iov++;
1383 uio_clone->uio_iovcnt--;
1386 if (len > ptoa(io_hold_cnt))
1387 len = ptoa(io_hold_cnt);
1388 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1389 end = round_page(addr + len);
1394 cnt = atop(end - trunc_page(addr));
1396 * A perfectly misaligned address and length could cause
1397 * both the start and the end of the chunk to use partial
1398 * page. +2 accounts for such a situation.
1400 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1401 addr, len, prot, ma, io_hold_cnt + 2);
1406 short_uio.uio_iov = &short_iovec[0];
1407 short_iovec[0].iov_base = (void *)addr;
1408 short_uio.uio_iovcnt = 1;
1409 short_uio.uio_resid = short_iovec[0].iov_len = len;
1410 short_uio.uio_offset = uio_clone->uio_offset;
1412 td->td_ma_cnt = cnt;
1414 error = vn_io_fault_doio(args, &short_uio, td);
1415 vm_page_unhold_pages(ma, cnt);
1416 adv = len - short_uio.uio_resid;
1418 uio_clone->uio_iov->iov_base =
1419 (char *)uio_clone->uio_iov->iov_base + adv;
1420 uio_clone->uio_iov->iov_len -= adv;
1421 uio_clone->uio_resid -= adv;
1422 uio_clone->uio_offset += adv;
1424 uio->uio_resid -= adv;
1425 uio->uio_offset += adv;
1427 if (error != 0 || adv == 0)
1430 td->td_ma = prev_td_ma;
1431 td->td_ma_cnt = prev_td_ma_cnt;
1432 curthread_pflags_restore(saveheld);
1434 free(uio_clone, M_IOV);
1439 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1440 int flags, struct thread *td)
1445 struct vn_io_fault_args args;
1448 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1452 * The ability to read(2) on a directory has historically been
1453 * allowed for all users, but this can and has been the source of
1454 * at least one security issue in the past. As such, it is now hidden
1455 * away behind a sysctl for those that actually need it to use it, and
1456 * restricted to root when it's turned on to make it relatively safe to
1457 * leave on for longer sessions of need.
1459 if (vp->v_type == VDIR) {
1460 KASSERT(uio->uio_rw == UIO_READ,
1461 ("illegal write attempted on a directory"));
1462 if (!vfs_allow_read_dir)
1464 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1468 foffset_lock_uio(fp, uio, flags);
1469 if (do_vn_io_fault(vp, uio)) {
1470 args.kind = VN_IO_FAULT_FOP;
1471 args.args.fop_args.fp = fp;
1472 args.args.fop_args.doio = doio;
1473 args.cred = active_cred;
1474 args.flags = flags | FOF_OFFSET;
1475 if (uio->uio_rw == UIO_READ) {
1476 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1477 uio->uio_offset + uio->uio_resid);
1478 } else if ((fp->f_flag & O_APPEND) != 0 ||
1479 (flags & FOF_OFFSET) == 0) {
1480 /* For appenders, punt and lock the whole range. */
1481 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1483 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1484 uio->uio_offset + uio->uio_resid);
1486 error = vn_io_fault1(vp, uio, &args, td);
1487 vn_rangelock_unlock(vp, rl_cookie);
1489 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1491 foffset_unlock_uio(fp, uio, flags);
1496 * Helper function to perform the requested uiomove operation using
1497 * the held pages for io->uio_iov[0].iov_base buffer instead of
1498 * copyin/copyout. Access to the pages with uiomove_fromphys()
1499 * instead of iov_base prevents page faults that could occur due to
1500 * pmap_collect() invalidating the mapping created by
1501 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1502 * object cleanup revoking the write access from page mappings.
1504 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1505 * instead of plain uiomove().
1508 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1510 struct uio transp_uio;
1511 struct iovec transp_iov[1];
1517 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1518 uio->uio_segflg != UIO_USERSPACE)
1519 return (uiomove(data, xfersize, uio));
1521 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1522 transp_iov[0].iov_base = data;
1523 transp_uio.uio_iov = &transp_iov[0];
1524 transp_uio.uio_iovcnt = 1;
1525 if (xfersize > uio->uio_resid)
1526 xfersize = uio->uio_resid;
1527 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1528 transp_uio.uio_offset = 0;
1529 transp_uio.uio_segflg = UIO_SYSSPACE;
1531 * Since transp_iov points to data, and td_ma page array
1532 * corresponds to original uio->uio_iov, we need to invert the
1533 * direction of the i/o operation as passed to
1534 * uiomove_fromphys().
1536 switch (uio->uio_rw) {
1538 transp_uio.uio_rw = UIO_READ;
1541 transp_uio.uio_rw = UIO_WRITE;
1544 transp_uio.uio_td = uio->uio_td;
1545 error = uiomove_fromphys(td->td_ma,
1546 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1547 xfersize, &transp_uio);
1548 adv = xfersize - transp_uio.uio_resid;
1550 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1551 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1553 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1555 td->td_ma_cnt -= pgadv;
1556 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1557 uio->uio_iov->iov_len -= adv;
1558 uio->uio_resid -= adv;
1559 uio->uio_offset += adv;
1564 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1568 vm_offset_t iov_base;
1572 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1573 uio->uio_segflg != UIO_USERSPACE)
1574 return (uiomove_fromphys(ma, offset, xfersize, uio));
1576 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1577 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1578 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1579 switch (uio->uio_rw) {
1581 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1585 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1589 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1591 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1593 td->td_ma_cnt -= pgadv;
1594 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1595 uio->uio_iov->iov_len -= cnt;
1596 uio->uio_resid -= cnt;
1597 uio->uio_offset += cnt;
1602 * File table truncate routine.
1605 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1617 * Lock the whole range for truncation. Otherwise split i/o
1618 * might happen partly before and partly after the truncation.
1620 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1621 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1624 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1625 AUDIT_ARG_VNODE1(vp);
1626 if (vp->v_type == VDIR) {
1631 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1635 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1639 vn_finished_write(mp);
1641 vn_rangelock_unlock(vp, rl_cookie);
1642 if (error == ERELOOKUP)
1648 * Truncate a file that is already locked.
1651 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1657 error = VOP_ADD_WRITECOUNT(vp, 1);
1660 vattr.va_size = length;
1662 vattr.va_vaflags |= VA_SYNC;
1663 error = VOP_SETATTR(vp, &vattr, cred);
1664 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1670 * File table vnode stat routine.
1673 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred)
1675 struct vnode *vp = fp->f_vnode;
1678 vn_lock(vp, LK_SHARED | LK_RETRY);
1679 error = VOP_STAT(vp, sb, active_cred, fp->f_cred);
1686 * File table vnode ioctl routine.
1689 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1694 struct fiobmap2_arg *bmarg;
1698 switch (vp->v_type) {
1703 vn_lock(vp, LK_SHARED | LK_RETRY);
1704 error = VOP_GETATTR(vp, &vattr, active_cred);
1707 *(int *)data = vattr.va_size - fp->f_offset;
1710 bmarg = (struct fiobmap2_arg *)data;
1711 vn_lock(vp, LK_SHARED | LK_RETRY);
1713 error = mac_vnode_check_read(active_cred, fp->f_cred,
1717 error = VOP_BMAP(vp, bmarg->bn, NULL,
1718 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1725 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1730 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1738 * File table vnode poll routine.
1741 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1748 #if defined(MAC) || defined(AUDIT)
1749 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1750 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1751 AUDIT_ARG_VNODE1(vp);
1752 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1758 error = VOP_POLL(vp, events, fp->f_cred, td);
1763 * Acquire the requested lock and then check for validity. LK_RETRY
1764 * permits vn_lock to return doomed vnodes.
1766 static int __noinline
1767 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1771 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1772 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1775 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1777 if ((flags & LK_RETRY) == 0) {
1788 * Nothing to do if we got the lock.
1794 * Interlock was dropped by the call in _vn_lock.
1796 flags &= ~LK_INTERLOCK;
1798 error = VOP_LOCK1(vp, flags, file, line);
1799 } while (error != 0);
1804 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1808 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1809 ("vn_lock: no locktype (%d passed)", flags));
1810 VNPASS(vp->v_holdcnt > 0, vp);
1811 error = VOP_LOCK1(vp, flags, file, line);
1812 if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1813 return (_vn_lock_fallback(vp, flags, file, line, error));
1818 * File table vnode close routine.
1821 vn_closefile(struct file *fp, struct thread *td)
1829 fp->f_ops = &badfileops;
1830 ref = (fp->f_flag & FHASLOCK) != 0;
1832 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1834 if (__predict_false(ref)) {
1835 lf.l_whence = SEEK_SET;
1838 lf.l_type = F_UNLCK;
1839 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1846 * Preparing to start a filesystem write operation. If the operation is
1847 * permitted, then we bump the count of operations in progress and
1848 * proceed. If a suspend request is in progress, we wait until the
1849 * suspension is over, and then proceed.
1852 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1854 struct mount_pcpu *mpcpu;
1857 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1858 vfs_op_thread_enter(mp, mpcpu)) {
1859 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1860 vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1861 vfs_op_thread_exit(mp, mpcpu);
1866 mtx_assert(MNT_MTX(mp), MA_OWNED);
1873 * Check on status of suspension.
1875 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1876 mp->mnt_susp_owner != curthread) {
1877 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1878 (flags & PCATCH) : 0) | (PUSER - 1);
1879 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1880 if (flags & V_NOWAIT) {
1881 error = EWOULDBLOCK;
1884 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1890 if (flags & V_XSLEEP)
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_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1907 ("V_MNTREF requires mp"));
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().
1932 if (vp == NULL && (flags & V_MNTREF) == 0)
1935 return (vn_start_write_refed(mp, flags, false));
1939 * Secondary suspension. Used by operations such as vop_inactive
1940 * routines that are needed by the higher level functions. These
1941 * are allowed to proceed until all the higher level functions have
1942 * completed (indicated by mnt_writeopcount dropping to zero). At that
1943 * time, these operations are halted until the suspension is over.
1946 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1951 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1952 ("V_MNTREF requires mp"));
1956 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1958 if (error != EOPNOTSUPP)
1964 * If we are not suspended or have not yet reached suspended
1965 * mode, then let the operation proceed.
1967 if ((mp = *mpp) == NULL)
1971 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1973 * As long as a vnode is not provided we need to acquire a
1974 * refcount for the provided mountpoint too, in order to
1975 * emulate a vfs_ref().
1978 if (vp == NULL && (flags & V_MNTREF) == 0)
1980 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1981 mp->mnt_secondary_writes++;
1982 mp->mnt_secondary_accwrites++;
1986 if (flags & V_NOWAIT) {
1989 return (EWOULDBLOCK);
1992 * Wait for the suspension to finish.
1994 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1995 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
2004 * Filesystem write operation has completed. If we are suspending and this
2005 * operation is the last one, notify the suspender that the suspension is
2009 vn_finished_write(struct mount *mp)
2011 struct mount_pcpu *mpcpu;
2017 if (vfs_op_thread_enter(mp, mpcpu)) {
2018 vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2019 vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2020 vfs_op_thread_exit(mp, mpcpu);
2025 vfs_assert_mount_counters(mp);
2027 c = --mp->mnt_writeopcount;
2028 if (mp->mnt_vfs_ops == 0) {
2029 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2034 vfs_dump_mount_counters(mp);
2035 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2036 wakeup(&mp->mnt_writeopcount);
2041 * Filesystem secondary write operation has completed. If we are
2042 * suspending and this operation is the last one, notify the suspender
2043 * that the suspension is now in effect.
2046 vn_finished_secondary_write(struct mount *mp)
2052 mp->mnt_secondary_writes--;
2053 if (mp->mnt_secondary_writes < 0)
2054 panic("vn_finished_secondary_write: neg cnt");
2055 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2056 mp->mnt_secondary_writes <= 0)
2057 wakeup(&mp->mnt_secondary_writes);
2062 * Request a filesystem to suspend write operations.
2065 vfs_write_suspend(struct mount *mp, int flags)
2072 vfs_assert_mount_counters(mp);
2073 if (mp->mnt_susp_owner == curthread) {
2074 vfs_op_exit_locked(mp);
2078 while (mp->mnt_kern_flag & MNTK_SUSPEND)
2079 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2082 * Unmount holds a write reference on the mount point. If we
2083 * own busy reference and drain for writers, we deadlock with
2084 * the reference draining in the unmount path. Callers of
2085 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2086 * vfs_busy() reference is owned and caller is not in the
2089 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2090 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2091 vfs_op_exit_locked(mp);
2096 mp->mnt_kern_flag |= MNTK_SUSPEND;
2097 mp->mnt_susp_owner = curthread;
2098 if (mp->mnt_writeopcount > 0)
2099 (void) msleep(&mp->mnt_writeopcount,
2100 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2103 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2104 vfs_write_resume(mp, 0);
2105 /* vfs_write_resume does vfs_op_exit() for us */
2111 * Request a filesystem to resume write operations.
2114 vfs_write_resume(struct mount *mp, int flags)
2118 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2119 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2120 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2122 mp->mnt_susp_owner = NULL;
2123 wakeup(&mp->mnt_writeopcount);
2124 wakeup(&mp->mnt_flag);
2125 curthread->td_pflags &= ~TDP_IGNSUSP;
2126 if ((flags & VR_START_WRITE) != 0) {
2128 mp->mnt_writeopcount++;
2131 if ((flags & VR_NO_SUSPCLR) == 0)
2134 } else if ((flags & VR_START_WRITE) != 0) {
2136 vn_start_write_refed(mp, 0, true);
2143 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2147 vfs_write_suspend_umnt(struct mount *mp)
2151 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2152 ("vfs_write_suspend_umnt: recursed"));
2154 /* dounmount() already called vn_start_write(). */
2156 vn_finished_write(mp);
2157 error = vfs_write_suspend(mp, 0);
2159 vn_start_write(NULL, &mp, V_WAIT);
2163 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2166 vn_start_write(NULL, &mp, V_WAIT);
2168 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2169 wakeup(&mp->mnt_flag);
2171 curthread->td_pflags |= TDP_IGNSUSP;
2176 * Implement kqueues for files by translating it to vnode operation.
2179 vn_kqfilter(struct file *fp, struct knote *kn)
2182 return (VOP_KQFILTER(fp->f_vnode, kn));
2186 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2188 if ((fp->f_flag & FKQALLOWED) == 0)
2190 return (vn_kqfilter(fp, kn));
2194 * Simplified in-kernel wrapper calls for extended attribute access.
2195 * Both calls pass in a NULL credential, authorizing as "kernel" access.
2196 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2199 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2200 const char *attrname, int *buflen, char *buf, struct thread *td)
2206 iov.iov_len = *buflen;
2209 auio.uio_iov = &iov;
2210 auio.uio_iovcnt = 1;
2211 auio.uio_rw = UIO_READ;
2212 auio.uio_segflg = UIO_SYSSPACE;
2214 auio.uio_offset = 0;
2215 auio.uio_resid = *buflen;
2217 if ((ioflg & IO_NODELOCKED) == 0)
2218 vn_lock(vp, LK_SHARED | LK_RETRY);
2220 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2222 /* authorize attribute retrieval as kernel */
2223 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2226 if ((ioflg & IO_NODELOCKED) == 0)
2230 *buflen = *buflen - auio.uio_resid;
2237 * XXX failure mode if partially written?
2240 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2241 const char *attrname, int buflen, char *buf, struct thread *td)
2248 iov.iov_len = buflen;
2251 auio.uio_iov = &iov;
2252 auio.uio_iovcnt = 1;
2253 auio.uio_rw = UIO_WRITE;
2254 auio.uio_segflg = UIO_SYSSPACE;
2256 auio.uio_offset = 0;
2257 auio.uio_resid = buflen;
2259 if ((ioflg & IO_NODELOCKED) == 0) {
2260 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2262 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2265 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2267 /* authorize attribute setting as kernel */
2268 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2270 if ((ioflg & IO_NODELOCKED) == 0) {
2271 vn_finished_write(mp);
2279 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2280 const char *attrname, struct thread *td)
2285 if ((ioflg & IO_NODELOCKED) == 0) {
2286 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2288 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2291 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2293 /* authorize attribute removal as kernel */
2294 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2295 if (error == EOPNOTSUPP)
2296 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2299 if ((ioflg & IO_NODELOCKED) == 0) {
2300 vn_finished_write(mp);
2308 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2312 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2316 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2319 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2324 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2325 int lkflags, struct vnode **rvp)
2330 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2332 ltype = VOP_ISLOCKED(vp);
2333 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2334 ("vn_vget_ino: vp not locked"));
2335 error = vfs_busy(mp, MBF_NOWAIT);
2339 error = vfs_busy(mp, 0);
2340 vn_lock(vp, ltype | LK_RETRY);
2344 if (VN_IS_DOOMED(vp)) {
2350 error = alloc(mp, alloc_arg, lkflags, rvp);
2352 if (error != 0 || *rvp != vp)
2353 vn_lock(vp, ltype | LK_RETRY);
2354 if (VN_IS_DOOMED(vp)) {
2367 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2377 * There are conditions where the limit is to be ignored.
2378 * However, since it is almost never reached, check it first.
2380 ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2381 lim = lim_cur(td, RLIMIT_FSIZE);
2382 if (__predict_false(ktr_write))
2383 lim = td->td_ktr_io_lim;
2384 if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2388 * The limit is reached.
2390 if (vp->v_type != VREG ||
2391 (td->td_pflags2 & TDP2_ACCT) != 0)
2394 if (!ktr_write || ktr_filesize_limit_signal) {
2395 PROC_LOCK(td->td_proc);
2396 kern_psignal(td->td_proc, SIGXFSZ);
2397 PROC_UNLOCK(td->td_proc);
2403 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2410 vn_lock(vp, LK_SHARED | LK_RETRY);
2411 AUDIT_ARG_VNODE1(vp);
2414 return (setfmode(td, active_cred, vp, mode));
2418 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2425 vn_lock(vp, LK_SHARED | LK_RETRY);
2426 AUDIT_ARG_VNODE1(vp);
2429 return (setfown(td, active_cred, vp, uid, gid));
2433 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2437 if ((object = vp->v_object) == NULL)
2439 VM_OBJECT_WLOCK(object);
2440 vm_object_page_remove(object, start, end, 0);
2441 VM_OBJECT_WUNLOCK(object);
2445 vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off,
2454 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2455 ("%s: Wrong command %lu", __func__, cmd));
2456 ASSERT_VOP_LOCKED(vp, "vn_bmap_seekhole_locked");
2458 if (vp->v_type != VREG) {
2462 error = VOP_GETATTR(vp, &va, cred);
2466 if (noff >= va.va_size) {
2470 bsize = vp->v_mount->mnt_stat.f_iosize;
2471 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2473 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2474 if (error == EOPNOTSUPP) {
2478 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2479 (bnp != -1 && cmd == FIOSEEKDATA)) {
2486 if (noff > va.va_size)
2488 /* noff == va.va_size. There is an implicit hole at the end of file. */
2489 if (cmd == FIOSEEKDATA)
2498 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2502 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2503 ("%s: Wrong command %lu", __func__, cmd));
2505 if (vn_lock(vp, LK_SHARED) != 0)
2507 error = vn_bmap_seekhole_locked(vp, cmd, off, cred);
2513 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2518 off_t foffset, size;
2521 cred = td->td_ucred;
2523 foffset = foffset_lock(fp, 0);
2524 noneg = (vp->v_type != VCHR);
2530 (offset > 0 && foffset > OFF_MAX - offset))) {
2537 vn_lock(vp, LK_SHARED | LK_RETRY);
2538 error = VOP_GETATTR(vp, &vattr, cred);
2544 * If the file references a disk device, then fetch
2545 * the media size and use that to determine the ending
2548 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2549 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2550 vattr.va_size = size;
2552 (vattr.va_size > OFF_MAX ||
2553 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2557 offset += vattr.va_size;
2562 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2563 if (error == ENOTTY)
2567 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2568 if (error == ENOTTY)
2574 if (error == 0 && noneg && offset < 0)
2578 VFS_KNOTE_UNLOCKED(vp, 0);
2579 td->td_uretoff.tdu_off = offset;
2581 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2586 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2592 * Grant permission if the caller is the owner of the file, or
2593 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2594 * on the file. If the time pointer is null, then write
2595 * permission on the file is also sufficient.
2597 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2598 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2599 * will be allowed to set the times [..] to the current
2602 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2603 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2604 error = VOP_ACCESS(vp, VWRITE, cred, td);
2609 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2614 if (fp->f_type == DTYPE_FIFO)
2615 kif->kf_type = KF_TYPE_FIFO;
2617 kif->kf_type = KF_TYPE_VNODE;
2620 FILEDESC_SUNLOCK(fdp);
2621 error = vn_fill_kinfo_vnode(vp, kif);
2623 FILEDESC_SLOCK(fdp);
2628 vn_fill_junk(struct kinfo_file *kif)
2633 * Simulate vn_fullpath returning changing values for a given
2634 * vp during e.g. coredump.
2636 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2637 olen = strlen(kif->kf_path);
2639 strcpy(&kif->kf_path[len - 1], "$");
2641 for (; olen < len; olen++)
2642 strcpy(&kif->kf_path[olen], "A");
2646 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2649 char *fullpath, *freepath;
2652 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2655 error = vn_fullpath(vp, &fullpath, &freepath);
2657 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2659 if (freepath != NULL)
2660 free(freepath, M_TEMP);
2662 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2667 * Retrieve vnode attributes.
2669 va.va_fsid = VNOVAL;
2671 vn_lock(vp, LK_SHARED | LK_RETRY);
2672 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2676 if (va.va_fsid != VNOVAL)
2677 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2679 kif->kf_un.kf_file.kf_file_fsid =
2680 vp->v_mount->mnt_stat.f_fsid.val[0];
2681 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2682 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2683 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2684 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2685 kif->kf_un.kf_file.kf_file_size = va.va_size;
2686 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2687 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2688 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2693 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2694 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2698 struct pmckern_map_in pkm;
2704 boolean_t writecounted;
2707 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2708 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2710 * POSIX shared-memory objects are defined to have
2711 * kernel persistence, and are not defined to support
2712 * read(2)/write(2) -- or even open(2). Thus, we can
2713 * use MAP_ASYNC to trade on-disk coherence for speed.
2714 * The shm_open(3) library routine turns on the FPOSIXSHM
2715 * flag to request this behavior.
2717 if ((fp->f_flag & FPOSIXSHM) != 0)
2718 flags |= MAP_NOSYNC;
2723 * Ensure that file and memory protections are
2724 * compatible. Note that we only worry about
2725 * writability if mapping is shared; in this case,
2726 * current and max prot are dictated by the open file.
2727 * XXX use the vnode instead? Problem is: what
2728 * credentials do we use for determination? What if
2729 * proc does a setuid?
2732 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2733 maxprot = VM_PROT_NONE;
2734 if ((prot & VM_PROT_EXECUTE) != 0)
2737 maxprot = VM_PROT_EXECUTE;
2738 if ((fp->f_flag & FREAD) != 0)
2739 maxprot |= VM_PROT_READ;
2740 else if ((prot & VM_PROT_READ) != 0)
2744 * If we are sharing potential changes via MAP_SHARED and we
2745 * are trying to get write permission although we opened it
2746 * without asking for it, bail out.
2748 if ((flags & MAP_SHARED) != 0) {
2749 if ((fp->f_flag & FWRITE) != 0)
2750 maxprot |= VM_PROT_WRITE;
2751 else if ((prot & VM_PROT_WRITE) != 0)
2754 maxprot |= VM_PROT_WRITE;
2755 cap_maxprot |= VM_PROT_WRITE;
2757 maxprot &= cap_maxprot;
2760 * For regular files and shared memory, POSIX requires that
2761 * the value of foff be a legitimate offset within the data
2762 * object. In particular, negative offsets are invalid.
2763 * Blocking negative offsets and overflows here avoids
2764 * possible wraparound or user-level access into reserved
2765 * ranges of the data object later. In contrast, POSIX does
2766 * not dictate how offsets are used by device drivers, so in
2767 * the case of a device mapping a negative offset is passed
2774 foff > OFF_MAX - size)
2777 writecounted = FALSE;
2778 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2779 &foff, &object, &writecounted);
2782 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2783 foff, writecounted, td);
2786 * If this mapping was accounted for in the vnode's
2787 * writecount, then undo that now.
2790 vm_pager_release_writecount(object, 0, size);
2791 vm_object_deallocate(object);
2794 /* Inform hwpmc(4) if an executable is being mapped. */
2795 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2796 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2798 pkm.pm_address = (uintptr_t) *addr;
2799 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2807 vn_fsid(struct vnode *vp, struct vattr *va)
2811 f = &vp->v_mount->mnt_stat.f_fsid;
2812 va->va_fsid = (uint32_t)f->val[1];
2813 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2814 va->va_fsid += (uint32_t)f->val[0];
2818 vn_fsync_buf(struct vnode *vp, int waitfor)
2820 struct buf *bp, *nbp;
2823 int error, maxretry;
2826 maxretry = 10000; /* large, arbitrarily chosen */
2828 if (vp->v_type == VCHR) {
2830 mp = vp->v_rdev->si_mountpt;
2837 * MARK/SCAN initialization to avoid infinite loops.
2839 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2840 bp->b_vflags &= ~BV_SCANNED;
2845 * Flush all dirty buffers associated with a vnode.
2848 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2849 if ((bp->b_vflags & BV_SCANNED) != 0)
2851 bp->b_vflags |= BV_SCANNED;
2852 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2853 if (waitfor != MNT_WAIT)
2856 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2857 BO_LOCKPTR(bo)) != 0) {
2864 KASSERT(bp->b_bufobj == bo,
2865 ("bp %p wrong b_bufobj %p should be %p",
2866 bp, bp->b_bufobj, bo));
2867 if ((bp->b_flags & B_DELWRI) == 0)
2868 panic("fsync: not dirty");
2869 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2875 if (maxretry < 1000)
2876 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2882 * If synchronous the caller expects us to completely resolve all
2883 * dirty buffers in the system. Wait for in-progress I/O to
2884 * complete (which could include background bitmap writes), then
2885 * retry if dirty blocks still exist.
2887 if (waitfor == MNT_WAIT) {
2888 bufobj_wwait(bo, 0, 0);
2889 if (bo->bo_dirty.bv_cnt > 0) {
2891 * If we are unable to write any of these buffers
2892 * then we fail now rather than trying endlessly
2893 * to write them out.
2895 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2896 if ((error = bp->b_error) != 0)
2898 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2899 (error == 0 && --maxretry >= 0))
2907 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2913 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
2914 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2915 * to do the actual copy.
2916 * vn_generic_copy_file_range() is factored out, so it can be called
2917 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2918 * different file systems.
2921 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2922 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2923 struct ucred *outcred, struct thread *fsize_td)
2930 *lenp = 0; /* For error returns. */
2933 /* Do some sanity checks on the arguments. */
2934 if (invp->v_type == VDIR || outvp->v_type == VDIR)
2936 else if (*inoffp < 0 || *outoffp < 0 ||
2937 invp->v_type != VREG || outvp->v_type != VREG)
2942 /* Ensure offset + len does not wrap around. */
2945 if (uval > INT64_MAX)
2946 len = INT64_MAX - *inoffp;
2949 if (uval > INT64_MAX)
2950 len = INT64_MAX - *outoffp;
2955 * If the two vnode are for the same file system, call
2956 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2957 * which can handle copies across multiple file systems.
2960 if (invp->v_mount == outvp->v_mount)
2961 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2962 lenp, flags, incred, outcred, fsize_td);
2964 error = vn_generic_copy_file_range(invp, inoffp, outvp,
2965 outoffp, lenp, flags, incred, outcred, fsize_td);
2971 * Test len bytes of data starting at dat for all bytes == 0.
2972 * Return true if all bytes are zero, false otherwise.
2973 * Expects dat to be well aligned.
2976 mem_iszero(void *dat, int len)
2982 for (p = dat; len > 0; len -= sizeof(*p), p++) {
2983 if (len >= sizeof(*p)) {
2987 cp = (const char *)p;
2988 for (i = 0; i < len; i++, cp++)
2997 * Look for a hole in the output file and, if found, adjust *outoffp
2998 * and *xferp to skip past the hole.
2999 * *xferp is the entire hole length to be written and xfer2 is how many bytes
3000 * to be written as 0's upon return.
3003 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3004 off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3009 if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3010 *dataoffp = *outoffp;
3011 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3014 *holeoffp = *dataoffp;
3015 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3018 if (error != 0 || *holeoffp == *dataoffp) {
3020 * Since outvp is unlocked, it may be possible for
3021 * another thread to do a truncate(), lseek(), write()
3022 * creating a hole at startoff between the above
3023 * VOP_IOCTL() calls, if the other thread does not do
3025 * If that happens, *holeoffp == *dataoffp and finding
3026 * the hole has failed, so disable vn_skip_hole().
3028 *holeoffp = -1; /* Disable use of vn_skip_hole(). */
3031 KASSERT(*dataoffp >= *outoffp,
3032 ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3033 (intmax_t)*dataoffp, (intmax_t)*outoffp));
3034 KASSERT(*holeoffp > *dataoffp,
3035 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3036 (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3040 * If there is a hole before the data starts, advance *outoffp and
3041 * *xferp past the hole.
3043 if (*dataoffp > *outoffp) {
3044 delta = *dataoffp - *outoffp;
3045 if (delta >= *xferp) {
3046 /* Entire *xferp is a hole. */
3053 xfer2 = MIN(xfer2, *xferp);
3057 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3058 * that the write ends at the start of the hole.
3059 * *holeoffp should always be greater than *outoffp, but for the
3060 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3063 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3064 xfer2 = *holeoffp - *outoffp;
3069 * Write an xfer sized chunk to outvp in blksize blocks from dat.
3070 * dat is a maximum of blksize in length and can be written repeatedly in
3072 * If growfile == true, just grow the file via vn_truncate_locked() instead
3073 * of doing actual writes.
3074 * If checkhole == true, a hole is being punched, so skip over any hole
3075 * already in the output file.
3078 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3079 u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3082 off_t dataoff, holeoff, xfer2;
3086 * Loop around doing writes of blksize until write has been completed.
3087 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3088 * done for each iteration, since the xfer argument can be very
3089 * large if there is a large hole to punch in the output file.
3094 xfer2 = MIN(xfer, blksize);
3097 * Punching a hole. Skip writing if there is
3098 * already a hole in the output file.
3100 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3101 &dataoff, &holeoff, cred);
3106 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3111 error = vn_start_write(outvp, &mp, V_WAIT);
3115 error = vn_lock(outvp, LK_EXCLUSIVE);
3117 error = vn_truncate_locked(outvp, outoff + xfer,
3122 error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3124 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3125 outoff, UIO_SYSSPACE, IO_NODELOCKED,
3126 curthread->td_ucred, cred, NULL, curthread);
3133 vn_finished_write(mp);
3134 } while (!growfile && xfer > 0 && error == 0);
3139 * Copy a byte range of one file to another. This function can handle the
3140 * case where invp and outvp are on different file systems.
3141 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3142 * is no better file system specific way to do it.
3145 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3146 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3147 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3149 struct vattr va, inva;
3152 off_t startoff, endoff, xfer, xfer2;
3154 int error, interrupted;
3155 bool cantseek, readzeros, eof, lastblock, holetoeof;
3157 size_t copylen, len, rem, savlen;
3159 long holein, holeout;
3160 struct timespec curts, endts;
3162 holein = holeout = 0;
3163 savlen = len = *lenp;
3168 error = vn_lock(invp, LK_SHARED);
3171 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3174 error = VOP_GETATTR(invp, &inva, incred);
3180 error = vn_start_write(outvp, &mp, V_WAIT);
3182 error = vn_lock(outvp, LK_EXCLUSIVE);
3185 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3186 * now that outvp is locked.
3188 if (fsize_td != NULL) {
3189 io.uio_offset = *outoffp;
3191 error = vn_rlimit_fsize(outvp, &io, fsize_td);
3195 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3198 * Holes that are past EOF do not need to be written as a block
3199 * of zero bytes. So, truncate the output file as far as
3200 * possible and then use va.va_size to decide if writing 0
3201 * bytes is necessary in the loop below.
3204 error = VOP_GETATTR(outvp, &va, outcred);
3205 if (error == 0 && va.va_size > *outoffp && va.va_size <=
3208 error = mac_vnode_check_write(curthread->td_ucred,
3212 error = vn_truncate_locked(outvp, *outoffp,
3215 va.va_size = *outoffp;
3220 vn_finished_write(mp);
3225 * Set the blksize to the larger of the hole sizes for invp and outvp.
3226 * If hole sizes aren't available, set the blksize to the larger
3227 * f_iosize of invp and outvp.
3228 * This code expects the hole sizes and f_iosizes to be powers of 2.
3229 * This value is clipped at 4Kbytes and 1Mbyte.
3231 blksize = MAX(holein, holeout);
3233 /* Clip len to end at an exact multiple of hole size. */
3235 rem = *inoffp % blksize;
3237 rem = blksize - rem;
3238 if (len > rem && len - rem > blksize)
3239 len = savlen = rounddown(len - rem, blksize) + rem;
3243 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3244 outvp->v_mount->mnt_stat.f_iosize);
3247 else if (blksize > 1024 * 1024)
3248 blksize = 1024 * 1024;
3249 dat = malloc(blksize, M_TEMP, M_WAITOK);
3252 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3253 * to find holes. Otherwise, just scan the read block for all 0s
3254 * in the inner loop where the data copying is done.
3255 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3256 * support holes on the server, but do not support FIOSEEKHOLE.
3257 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3258 * that this function should return after 1second with a partial
3261 if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3262 getnanouptime(&endts);
3265 timespecclear(&endts);
3266 holetoeof = eof = false;
3267 while (len > 0 && error == 0 && !eof && interrupted == 0) {
3268 endoff = 0; /* To shut up compilers. */
3274 * Find the next data area. If there is just a hole to EOF,
3275 * FIOSEEKDATA should fail with ENXIO.
3276 * (I do not know if any file system will report a hole to
3277 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3278 * will fail for those file systems.)
3280 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3281 * the code just falls through to the inner copy loop.
3285 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3287 if (error == ENXIO) {
3288 startoff = endoff = inva.va_size;
3289 eof = holetoeof = true;
3293 if (error == 0 && !holetoeof) {
3295 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3298 * Since invp is unlocked, it may be possible for
3299 * another thread to do a truncate(), lseek(), write()
3300 * creating a hole at startoff between the above
3301 * VOP_IOCTL() calls, if the other thread does not do
3303 * If that happens, startoff == endoff and finding
3304 * the hole has failed, so set an error.
3306 if (error == 0 && startoff == endoff)
3307 error = EINVAL; /* Any error. Reset to 0. */
3310 if (startoff > *inoffp) {
3311 /* Found hole before data block. */
3312 xfer = MIN(startoff - *inoffp, len);
3313 if (*outoffp < va.va_size) {
3314 /* Must write 0s to punch hole. */
3315 xfer2 = MIN(va.va_size - *outoffp,
3317 memset(dat, 0, MIN(xfer2, blksize));
3318 error = vn_write_outvp(outvp, dat,
3319 *outoffp, xfer2, blksize, false,
3320 holeout > 0, outcred);
3323 if (error == 0 && *outoffp + xfer >
3324 va.va_size && (xfer == len || holetoeof)) {
3325 /* Grow output file (hole at end). */
3326 error = vn_write_outvp(outvp, dat,
3327 *outoffp, xfer, blksize, true,
3335 interrupted = sig_intr();
3336 if (timespecisset(&endts) &&
3338 getnanouptime(&curts);
3339 if (timespeccmp(&curts,
3347 copylen = MIN(len, endoff - startoff);
3359 * Set first xfer to end at a block boundary, so that
3360 * holes are more likely detected in the loop below via
3361 * the for all bytes 0 method.
3363 xfer -= (*inoffp % blksize);
3365 /* Loop copying the data block. */
3366 while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3369 error = vn_lock(invp, LK_SHARED);
3372 error = vn_rdwr(UIO_READ, invp, dat, xfer,
3373 startoff, UIO_SYSSPACE, IO_NODELOCKED,
3374 curthread->td_ucred, incred, &aresid,
3378 if (error == 0 && aresid > 0) {
3379 /* Stop the copy at EOF on the input file. */
3386 * Skip the write for holes past the initial EOF
3387 * of the output file, unless this is the last
3388 * write of the output file at EOF.
3390 readzeros = cantseek ? mem_iszero(dat, xfer) :
3394 if (!cantseek || *outoffp < va.va_size ||
3395 lastblock || !readzeros)
3396 error = vn_write_outvp(outvp, dat,
3397 *outoffp, xfer, blksize,
3398 readzeros && lastblock &&
3399 *outoffp >= va.va_size, false,
3408 interrupted = sig_intr();
3409 if (timespecisset(&endts) &&
3411 getnanouptime(&curts);
3412 if (timespeccmp(&curts,
3424 *lenp = savlen - len;
3430 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3434 off_t olen, ooffset;
3437 int audited_vnode1 = 0;
3441 if (vp->v_type != VREG)
3444 /* Allocating blocks may take a long time, so iterate. */
3451 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3454 error = vn_lock(vp, LK_EXCLUSIVE);
3456 vn_finished_write(mp);
3460 if (!audited_vnode1) {
3461 AUDIT_ARG_VNODE1(vp);
3466 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3469 error = VOP_ALLOCATE(vp, &offset, &len, 0,
3472 vn_finished_write(mp);
3474 if (olen + ooffset != offset + len) {
3475 panic("offset + len changed from %jx/%jx to %jx/%jx",
3476 ooffset, olen, offset, len);
3478 if (error != 0 || len == 0)
3480 KASSERT(olen > len, ("Iteration did not make progress?"));
3488 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags,
3489 int ioflag, struct ucred *cred, struct ucred *active_cred,
3490 struct ucred *file_cred)
3497 bool audited_vnode1 = false;
3506 if ((ioflag & (IO_NODELOCKED | IO_RANGELOCKED)) == 0)
3507 rl_cookie = vn_rangelock_wlock(vp, off, off + len);
3508 while (len > 0 && error == 0) {
3510 * Try to deallocate the longest range in one pass.
3511 * In case a pass takes too long to be executed, it returns
3512 * partial result. The residue will be proceeded in the next
3516 if ((ioflag & IO_NODELOCKED) == 0) {
3518 if ((error = vn_start_write(vp, &mp,
3519 V_WAIT | PCATCH)) != 0)
3521 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
3524 if (!audited_vnode1) {
3525 AUDIT_ARG_VNODE1(vp);
3526 audited_vnode1 = true;
3531 if ((ioflag & IO_NOMACCHECK) == 0)
3532 error = mac_vnode_check_write(active_cred, file_cred,
3536 error = VOP_DEALLOCATE(vp, &off, &len, flags, ioflag,
3539 if ((ioflag & IO_NODELOCKED) == 0) {
3542 vn_finished_write(mp);
3546 if (error == 0 && len != 0)
3550 if (rl_cookie != NULL)
3551 vn_rangelock_unlock(vp, rl_cookie);
3558 * This function is supposed to be used in the situations where the deallocation
3559 * is not triggered by a user request.
3562 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags,
3563 int ioflag, struct ucred *active_cred, struct ucred *file_cred)
3567 if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset ||
3570 if (vp->v_type != VREG)
3573 cred = file_cred != NOCRED ? file_cred : active_cred;
3574 return (vn_deallocate_impl(vp, offset, length, flags, ioflag, cred,
3575 active_cred, file_cred));
3579 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
3580 struct ucred *active_cred, struct thread *td)
3588 if (cmd != SPACECTL_DEALLOC || *offset < 0 || *length <= 0 ||
3589 *length > OFF_MAX - *offset || flags != 0)
3591 if (vp->v_type != VREG)
3594 ioflag = get_write_ioflag(fp);
3597 case SPACECTL_DEALLOC:
3598 error = vn_deallocate_impl(vp, offset, length, flags, ioflag,
3599 active_cred, active_cred, fp->f_cred);
3602 panic("vn_fspacectl: unknown cmd %d", cmd);
3608 static u_long vn_lock_pair_pause_cnt;
3609 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3610 &vn_lock_pair_pause_cnt, 0,
3611 "Count of vn_lock_pair deadlocks");
3613 u_int vn_lock_pair_pause_max;
3614 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3615 &vn_lock_pair_pause_max, 0,
3616 "Max ticks for vn_lock_pair deadlock avoidance sleep");
3619 vn_lock_pair_pause(const char *wmesg)
3621 atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3622 pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3626 * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3627 * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3628 * must be unlocked. Same for vp2 and vp2_locked. One of the vnodes
3631 * The function returns with both vnodes exclusively locked, and
3632 * guarantees that it does not create lock order reversal with other
3633 * threads during its execution. Both vnodes could be unlocked
3634 * temporary (and reclaimed).
3637 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3642 if (vp1 == NULL && vp2 == NULL)
3646 ASSERT_VOP_ELOCKED(vp1, "vp1");
3648 ASSERT_VOP_UNLOCKED(vp1, "vp1");
3654 ASSERT_VOP_ELOCKED(vp2, "vp2");
3656 ASSERT_VOP_UNLOCKED(vp2, "vp2");
3660 if (!vp1_locked && !vp2_locked) {
3661 vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3666 if (vp1_locked && vp2_locked)
3668 if (vp1_locked && vp2 != NULL) {
3670 error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3671 __FILE__, __LINE__);
3676 vn_lock_pair_pause("vlp1");
3678 vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3681 if (vp2_locked && vp1 != NULL) {
3683 error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3684 __FILE__, __LINE__);
3689 vn_lock_pair_pause("vlp2");
3691 vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3696 ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3698 ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3702 vn_lktype_write(struct mount *mp, struct vnode *vp)
3704 if (MNT_SHARED_WRITES(mp) ||
3705 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
3707 return (LK_EXCLUSIVE);