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 = ndp->ni_cnd.cn_thread;
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)
233 struct thread *td = ndp->ni_cnd.cn_thread;
235 struct vattr *vap = &vat;
242 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
243 O_EXCL | O_DIRECTORY) ||
244 (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
246 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
247 ndp->ni_cnd.cn_nameiop = CREATE;
248 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
250 * Set NOCACHE to avoid flushing the cache when
251 * rolling in many files at once.
253 * Set NC_KEEPPOSENTRY to keep positive entries if they already
254 * exist despite NOCACHE.
256 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
257 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
258 ndp->ni_cnd.cn_flags |= FOLLOW;
259 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
261 if ((error = namei(ndp)) != 0)
263 if (ndp->ni_vp == NULL) {
266 vap->va_mode = cmode;
268 vap->va_vaflags |= VA_EXCLUSIVE;
269 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
270 NDFREE(ndp, NDF_ONLY_PNBUF);
272 if ((error = vn_start_write(NULL, &mp,
273 V_XSLEEP | PCATCH)) != 0)
278 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
279 ndp->ni_cnd.cn_flags |= MAKEENTRY;
281 error = mac_vnode_check_create(cred, ndp->ni_dvp,
285 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
288 if (error == 0 && (fmode & O_EXCL) != 0 &&
289 (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
291 vp->v_iflag |= VI_FOPENING;
295 VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
297 vn_finished_write(mp);
299 NDFREE(ndp, NDF_ONLY_PNBUF);
300 if (error == ERELOOKUP) {
308 if (ndp->ni_dvp == ndp->ni_vp)
314 if (fmode & O_EXCL) {
318 if (vp->v_type == VDIR) {
325 ndp->ni_cnd.cn_nameiop = LOOKUP;
326 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
327 ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
329 if ((fmode & FWRITE) == 0)
330 ndp->ni_cnd.cn_flags |= LOCKSHARED;
331 if ((error = namei(ndp)) != 0)
335 error = vn_open_vnode(vp, fmode, cred, td, fp);
338 vp->v_iflag &= ~VI_FOPENING;
347 NDFREE(ndp, NDF_ONLY_PNBUF);
355 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
358 int error, lock_flags, type;
360 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
361 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
363 KASSERT(fp != NULL, ("open with flock requires fp"));
364 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
367 lock_flags = VOP_ISLOCKED(vp);
370 lf.l_whence = SEEK_SET;
373 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
375 if ((fmode & FNONBLOCK) == 0)
377 if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
379 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
381 fp->f_flag |= FHASLOCK;
383 vn_lock(vp, lock_flags | LK_RETRY);
388 * Common code for vnode open operations once a vnode is located.
389 * Check permissions, and call the VOP_OPEN routine.
392 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
393 struct thread *td, struct file *fp)
398 if (vp->v_type == VLNK) {
399 if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
402 if (vp->v_type == VSOCK)
404 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
408 if ((fmode & O_PATH) == 0) {
409 if ((fmode & (FWRITE | O_TRUNC)) != 0) {
410 if (vp->v_type == VDIR)
414 if ((fmode & FREAD) != 0)
416 if ((fmode & O_APPEND) && (fmode & FWRITE))
419 if ((fmode & O_CREAT) != 0)
423 if ((fmode & FEXEC) != 0)
426 if ((fmode & O_VERIFY) != 0)
428 error = mac_vnode_check_open(cred, vp, accmode);
432 accmode &= ~(VCREAT | VVERIFY);
434 if ((fmode & O_CREAT) == 0 && accmode != 0) {
435 error = VOP_ACCESS(vp, accmode, cred, td);
439 if ((fmode & O_PATH) != 0) {
440 if (vp->v_type == VFIFO)
443 error = VOP_ACCESS(vp, VREAD, cred, td);
445 fp->f_flag |= FKQALLOWED;
449 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
450 vn_lock(vp, LK_UPGRADE | LK_RETRY);
451 error = VOP_OPEN(vp, fmode, cred, td, fp);
455 error = vn_open_vnode_advlock(vp, fmode, fp);
456 if (error == 0 && (fmode & FWRITE) != 0) {
457 error = VOP_ADD_WRITECOUNT(vp, 1);
459 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
460 __func__, vp, vp->v_writecount);
465 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
466 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
471 * Arrange the call by having fdrop() to use
472 * vn_closefile(). This is to satisfy
473 * filesystems like devfs or tmpfs, which
474 * override fo_close().
476 fp->f_flag |= FOPENFAILED;
478 if (fp->f_ops == &badfileops) {
479 fp->f_type = DTYPE_VNODE;
485 * If there is no fp, due to kernel-mode open,
486 * we can call VOP_CLOSE() now.
488 if (vp->v_type != VFIFO && (fmode & FWRITE) != 0 &&
489 !MNT_EXTENDED_SHARED(vp->v_mount) &&
490 VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
491 vn_lock(vp, LK_UPGRADE | LK_RETRY);
492 (void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
497 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
503 * Check for write permissions on the specified vnode.
504 * Prototype text segments cannot be written.
508 vn_writechk(struct vnode *vp)
511 ASSERT_VOP_LOCKED(vp, "vn_writechk");
513 * If there's shared text associated with
514 * the vnode, try to free it up once. If
515 * we fail, we can't allow writing.
527 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
528 struct thread *td, bool keep_ref)
531 int error, lock_flags;
533 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
534 MNT_EXTENDED_SHARED(vp->v_mount))
535 lock_flags = LK_SHARED;
537 lock_flags = LK_EXCLUSIVE;
539 vn_start_write(vp, &mp, V_WAIT);
540 vn_lock(vp, lock_flags | LK_RETRY);
541 AUDIT_ARG_VNODE1(vp);
542 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
543 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
544 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
545 __func__, vp, vp->v_writecount);
547 error = VOP_CLOSE(vp, flags, file_cred, td);
552 vn_finished_write(mp);
557 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
561 return (vn_close1(vp, flags, file_cred, td, false));
565 * Heuristic to detect sequential operation.
568 sequential_heuristic(struct uio *uio, struct file *fp)
572 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
575 if (fp->f_flag & FRDAHEAD)
576 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
579 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
580 * that the first I/O is normally considered to be slightly
581 * sequential. Seeking to offset 0 doesn't change sequentiality
582 * unless previous seeks have reduced f_seqcount to 0, in which
583 * case offset 0 is not special.
585 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
586 uio->uio_offset == fp->f_nextoff[rw]) {
588 * f_seqcount is in units of fixed-size blocks so that it
589 * depends mainly on the amount of sequential I/O and not
590 * much on the number of sequential I/O's. The fixed size
591 * of 16384 is hard-coded here since it is (not quite) just
592 * a magic size that works well here. This size is more
593 * closely related to the best I/O size for real disks than
594 * to any block size used by software.
596 if (uio->uio_resid >= IO_SEQMAX * 16384)
597 fp->f_seqcount[rw] = IO_SEQMAX;
599 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
600 if (fp->f_seqcount[rw] > IO_SEQMAX)
601 fp->f_seqcount[rw] = IO_SEQMAX;
603 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
606 /* Not sequential. Quickly draw-down sequentiality. */
607 if (fp->f_seqcount[rw] > 1)
608 fp->f_seqcount[rw] = 1;
610 fp->f_seqcount[rw] = 0;
615 * Package up an I/O request on a vnode into a uio and do it.
618 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
619 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
620 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
627 struct vn_io_fault_args args;
628 int error, lock_flags;
630 if (offset < 0 && vp->v_type != VCHR)
632 auio.uio_iov = &aiov;
634 aiov.iov_base = base;
636 auio.uio_resid = len;
637 auio.uio_offset = offset;
638 auio.uio_segflg = segflg;
643 if ((ioflg & IO_NODELOCKED) == 0) {
644 if ((ioflg & IO_RANGELOCKED) == 0) {
645 if (rw == UIO_READ) {
646 rl_cookie = vn_rangelock_rlock(vp, offset,
648 } else if ((ioflg & IO_APPEND) != 0) {
649 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
651 rl_cookie = vn_rangelock_wlock(vp, offset,
657 if (rw == UIO_WRITE) {
658 if (vp->v_type != VCHR &&
659 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
662 lock_flags = vn_lktype_write(mp, vp);
664 lock_flags = LK_SHARED;
665 vn_lock(vp, lock_flags | LK_RETRY);
669 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
671 if ((ioflg & IO_NOMACCHECK) == 0) {
673 error = mac_vnode_check_read(active_cred, file_cred,
676 error = mac_vnode_check_write(active_cred, file_cred,
681 if (file_cred != NULL)
685 if (do_vn_io_fault(vp, &auio)) {
686 args.kind = VN_IO_FAULT_VOP;
689 args.args.vop_args.vp = vp;
690 error = vn_io_fault1(vp, &auio, &args, td);
691 } else if (rw == UIO_READ) {
692 error = VOP_READ(vp, &auio, ioflg, cred);
693 } else /* if (rw == UIO_WRITE) */ {
694 error = VOP_WRITE(vp, &auio, ioflg, cred);
698 *aresid = auio.uio_resid;
700 if (auio.uio_resid && error == 0)
702 if ((ioflg & IO_NODELOCKED) == 0) {
705 vn_finished_write(mp);
708 if (rl_cookie != NULL)
709 vn_rangelock_unlock(vp, rl_cookie);
714 * Package up an I/O request on a vnode into a uio and do it. The I/O
715 * request is split up into smaller chunks and we try to avoid saturating
716 * the buffer cache while potentially holding a vnode locked, so we
717 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
718 * to give other processes a chance to lock the vnode (either other processes
719 * core'ing the same binary, or unrelated processes scanning the directory).
722 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
723 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
724 struct ucred *file_cred, size_t *aresid, struct thread *td)
733 * Force `offset' to a multiple of MAXBSIZE except possibly
734 * for the first chunk, so that filesystems only need to
735 * write full blocks except possibly for the first and last
738 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
742 if (rw != UIO_READ && vp->v_type == VREG)
745 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
746 ioflg, active_cred, file_cred, &iaresid, td);
747 len -= chunk; /* aresid calc already includes length */
751 base = (char *)base + chunk;
752 kern_yield(PRI_USER);
755 *aresid = len + iaresid;
759 #if OFF_MAX <= LONG_MAX
761 foffset_lock(struct file *fp, int flags)
763 volatile short *flagsp;
767 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
769 if ((flags & FOF_NOLOCK) != 0)
770 return (atomic_load_long(&fp->f_offset));
773 * According to McKusick the vn lock was protecting f_offset here.
774 * It is now protected by the FOFFSET_LOCKED flag.
776 flagsp = &fp->f_vnread_flags;
777 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
778 return (atomic_load_long(&fp->f_offset));
780 sleepq_lock(&fp->f_vnread_flags);
781 state = atomic_load_16(flagsp);
783 if ((state & FOFFSET_LOCKED) == 0) {
784 if (!atomic_fcmpset_acq_16(flagsp, &state,
789 if ((state & FOFFSET_LOCK_WAITING) == 0) {
790 if (!atomic_fcmpset_acq_16(flagsp, &state,
791 state | FOFFSET_LOCK_WAITING))
795 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
796 sleepq_wait(&fp->f_vnread_flags, PUSER -1);
798 sleepq_lock(&fp->f_vnread_flags);
799 state = atomic_load_16(flagsp);
801 res = atomic_load_long(&fp->f_offset);
802 sleepq_release(&fp->f_vnread_flags);
807 foffset_unlock(struct file *fp, off_t val, int flags)
809 volatile short *flagsp;
812 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
814 if ((flags & FOF_NOUPDATE) == 0)
815 atomic_store_long(&fp->f_offset, val);
816 if ((flags & FOF_NEXTOFF_R) != 0)
817 fp->f_nextoff[UIO_READ] = val;
818 if ((flags & FOF_NEXTOFF_W) != 0)
819 fp->f_nextoff[UIO_WRITE] = val;
821 if ((flags & FOF_NOLOCK) != 0)
824 flagsp = &fp->f_vnread_flags;
825 state = atomic_load_16(flagsp);
826 if ((state & FOFFSET_LOCK_WAITING) == 0 &&
827 atomic_cmpset_rel_16(flagsp, state, 0))
830 sleepq_lock(&fp->f_vnread_flags);
831 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
832 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
833 fp->f_vnread_flags = 0;
834 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
835 sleepq_release(&fp->f_vnread_flags);
839 foffset_lock(struct file *fp, int flags)
844 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
846 mtxp = mtx_pool_find(mtxpool_sleep, fp);
848 if ((flags & FOF_NOLOCK) == 0) {
849 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
850 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
851 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
854 fp->f_vnread_flags |= FOFFSET_LOCKED;
862 foffset_unlock(struct file *fp, off_t val, int flags)
866 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
868 mtxp = mtx_pool_find(mtxpool_sleep, fp);
870 if ((flags & FOF_NOUPDATE) == 0)
872 if ((flags & FOF_NEXTOFF_R) != 0)
873 fp->f_nextoff[UIO_READ] = val;
874 if ((flags & FOF_NEXTOFF_W) != 0)
875 fp->f_nextoff[UIO_WRITE] = val;
876 if ((flags & FOF_NOLOCK) == 0) {
877 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
878 ("Lost FOFFSET_LOCKED"));
879 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
880 wakeup(&fp->f_vnread_flags);
881 fp->f_vnread_flags = 0;
888 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
891 if ((flags & FOF_OFFSET) == 0)
892 uio->uio_offset = foffset_lock(fp, flags);
896 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
899 if ((flags & FOF_OFFSET) == 0)
900 foffset_unlock(fp, uio->uio_offset, flags);
904 get_advice(struct file *fp, struct uio *uio)
909 ret = POSIX_FADV_NORMAL;
910 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
913 mtxp = mtx_pool_find(mtxpool_sleep, fp);
915 if (fp->f_advice != NULL &&
916 uio->uio_offset >= fp->f_advice->fa_start &&
917 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
918 ret = fp->f_advice->fa_advice;
924 get_write_ioflag(struct file *fp)
932 mp = atomic_load_ptr(&vp->v_mount);
934 if ((fp->f_flag & O_DIRECT) != 0)
937 if ((fp->f_flag & O_FSYNC) != 0 ||
938 (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
942 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
943 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
944 * fall back to full O_SYNC behavior.
946 if ((fp->f_flag & O_DSYNC) != 0)
947 ioflag |= IO_SYNC | IO_DATASYNC;
953 vn_read_from_obj(struct vnode *vp, struct uio *uio)
956 vm_page_t ma[io_hold_cnt + 2];
961 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
962 obj = atomic_load_ptr(&vp->v_object);
964 return (EJUSTRETURN);
967 * Depends on type stability of vm_objects.
969 vm_object_pip_add(obj, 1);
970 if ((obj->flags & OBJ_DEAD) != 0) {
972 * Note that object might be already reused from the
973 * vnode, and the OBJ_DEAD flag cleared. This is fine,
974 * we recheck for DOOMED vnode state after all pages
975 * are busied, and retract then.
977 * But we check for OBJ_DEAD to ensure that we do not
978 * busy pages while vm_object_terminate_pages()
979 * processes the queue.
985 resid = uio->uio_resid;
986 off = uio->uio_offset;
987 for (i = 0; resid > 0; i++) {
988 MPASS(i < io_hold_cnt + 2);
989 ma[i] = vm_page_grab_unlocked(obj, atop(off),
990 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
996 * Skip invalid pages. Valid mask can be partial only
997 * at EOF, and we clip later.
999 if (vm_page_none_valid(ma[i])) {
1000 vm_page_sunbusy(ma[i]);
1008 error = EJUSTRETURN;
1013 * Check VIRF_DOOMED after we busied our pages. Since
1014 * vgonel() terminates the vnode' vm_object, it cannot
1015 * process past pages busied by us.
1017 if (VN_IS_DOOMED(vp)) {
1018 error = EJUSTRETURN;
1022 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1023 if (resid > uio->uio_resid)
1024 resid = uio->uio_resid;
1027 * Unlocked read of vnp_size is safe because truncation cannot
1028 * pass busied page. But we load vnp_size into a local
1029 * variable so that possible concurrent extension does not
1030 * break calculation.
1032 #if defined(__powerpc__) && !defined(__powerpc64__)
1033 vsz = obj->un_pager.vnp.vnp_size;
1035 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1037 if (uio->uio_offset >= vsz) {
1038 error = EJUSTRETURN;
1041 if (uio->uio_offset + resid > vsz)
1042 resid = vsz - uio->uio_offset;
1044 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1047 for (j = 0; j < i; j++) {
1049 vm_page_reference(ma[j]);
1050 vm_page_sunbusy(ma[j]);
1053 vm_object_pip_wakeup(obj);
1056 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1060 * File table vnode read routine.
1063 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1071 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1073 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1076 if (fp->f_flag & FNONBLOCK)
1077 ioflag |= IO_NDELAY;
1078 if (fp->f_flag & O_DIRECT)
1079 ioflag |= IO_DIRECT;
1082 * Try to read from page cache. VIRF_DOOMED check is racy but
1083 * allows us to avoid unneeded work outright.
1085 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1086 (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1087 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1089 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1092 if (error != EJUSTRETURN)
1096 advice = get_advice(fp, uio);
1097 vn_lock(vp, LK_SHARED | LK_RETRY);
1100 case POSIX_FADV_NORMAL:
1101 case POSIX_FADV_SEQUENTIAL:
1102 case POSIX_FADV_NOREUSE:
1103 ioflag |= sequential_heuristic(uio, fp);
1105 case POSIX_FADV_RANDOM:
1106 /* Disable read-ahead for random I/O. */
1109 orig_offset = uio->uio_offset;
1112 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1115 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1116 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1118 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1119 orig_offset != uio->uio_offset)
1121 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1122 * for the backing file after a POSIX_FADV_NOREUSE
1125 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1126 POSIX_FADV_DONTNEED);
1131 * File table vnode write routine.
1134 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1142 bool need_finished_write;
1144 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1146 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1148 if (vp->v_type == VREG)
1151 if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1152 ioflag |= IO_APPEND;
1153 if ((fp->f_flag & FNONBLOCK) != 0)
1154 ioflag |= IO_NDELAY;
1155 ioflag |= get_write_ioflag(fp);
1158 need_finished_write = false;
1159 if (vp->v_type != VCHR) {
1160 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1163 need_finished_write = true;
1166 advice = get_advice(fp, uio);
1168 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1170 case POSIX_FADV_NORMAL:
1171 case POSIX_FADV_SEQUENTIAL:
1172 case POSIX_FADV_NOREUSE:
1173 ioflag |= sequential_heuristic(uio, fp);
1175 case POSIX_FADV_RANDOM:
1176 /* XXX: Is this correct? */
1179 orig_offset = uio->uio_offset;
1182 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1185 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1186 fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1188 if (need_finished_write)
1189 vn_finished_write(mp);
1190 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1191 orig_offset != uio->uio_offset)
1193 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1194 * for the backing file after a POSIX_FADV_NOREUSE
1197 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1198 POSIX_FADV_DONTNEED);
1204 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1205 * prevent the following deadlock:
1207 * Assume that the thread A reads from the vnode vp1 into userspace
1208 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
1209 * currently not resident, then system ends up with the call chain
1210 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1211 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1212 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1213 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1214 * backed by the pages of vnode vp1, and some page in buf2 is not
1215 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1217 * To prevent the lock order reversal and deadlock, vn_io_fault() does
1218 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1219 * Instead, it first tries to do the whole range i/o with pagefaults
1220 * disabled. If all pages in the i/o buffer are resident and mapped,
1221 * VOP will succeed (ignoring the genuine filesystem errors).
1222 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1223 * i/o in chunks, with all pages in the chunk prefaulted and held
1224 * using vm_fault_quick_hold_pages().
1226 * Filesystems using this deadlock avoidance scheme should use the
1227 * array of the held pages from uio, saved in the curthread->td_ma,
1228 * instead of doing uiomove(). A helper function
1229 * vn_io_fault_uiomove() converts uiomove request into
1230 * uiomove_fromphys() over td_ma array.
1232 * Since vnode locks do not cover the whole i/o anymore, rangelocks
1233 * make the current i/o request atomic with respect to other i/os and
1238 * Decode vn_io_fault_args and perform the corresponding i/o.
1241 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1247 save = vm_fault_disable_pagefaults();
1248 switch (args->kind) {
1249 case VN_IO_FAULT_FOP:
1250 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1251 uio, args->cred, args->flags, td);
1253 case VN_IO_FAULT_VOP:
1254 if (uio->uio_rw == UIO_READ) {
1255 error = VOP_READ(args->args.vop_args.vp, uio,
1256 args->flags, args->cred);
1257 } else if (uio->uio_rw == UIO_WRITE) {
1258 error = VOP_WRITE(args->args.vop_args.vp, uio,
1259 args->flags, args->cred);
1263 panic("vn_io_fault_doio: unknown kind of io %d %d",
1264 args->kind, uio->uio_rw);
1266 vm_fault_enable_pagefaults(save);
1271 vn_io_fault_touch(char *base, const struct uio *uio)
1276 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1282 vn_io_fault_prefault_user(const struct uio *uio)
1285 const struct iovec *iov;
1290 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1291 ("vn_io_fault_prefault userspace"));
1295 resid = uio->uio_resid;
1296 base = iov->iov_base;
1299 error = vn_io_fault_touch(base, uio);
1302 if (len < PAGE_SIZE) {
1304 error = vn_io_fault_touch(base + len - 1, uio);
1309 if (++i >= uio->uio_iovcnt)
1311 iov = uio->uio_iov + i;
1312 base = iov->iov_base;
1324 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1325 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1326 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1327 * into args and call vn_io_fault1() to handle faults during the user
1328 * mode buffer accesses.
1331 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1334 vm_page_t ma[io_hold_cnt + 2];
1335 struct uio *uio_clone, short_uio;
1336 struct iovec short_iovec[1];
1337 vm_page_t *prev_td_ma;
1339 vm_offset_t addr, end;
1342 int error, cnt, saveheld, prev_td_ma_cnt;
1344 if (vn_io_fault_prefault) {
1345 error = vn_io_fault_prefault_user(uio);
1347 return (error); /* Or ignore ? */
1350 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1353 * The UFS follows IO_UNIT directive and replays back both
1354 * uio_offset and uio_resid if an error is encountered during the
1355 * operation. But, since the iovec may be already advanced,
1356 * uio is still in an inconsistent state.
1358 * Cache a copy of the original uio, which is advanced to the redo
1359 * point using UIO_NOCOPY below.
1361 uio_clone = cloneuio(uio);
1362 resid = uio->uio_resid;
1364 short_uio.uio_segflg = UIO_USERSPACE;
1365 short_uio.uio_rw = uio->uio_rw;
1366 short_uio.uio_td = uio->uio_td;
1368 error = vn_io_fault_doio(args, uio, td);
1369 if (error != EFAULT)
1372 atomic_add_long(&vn_io_faults_cnt, 1);
1373 uio_clone->uio_segflg = UIO_NOCOPY;
1374 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1375 uio_clone->uio_segflg = uio->uio_segflg;
1377 saveheld = curthread_pflags_set(TDP_UIOHELD);
1378 prev_td_ma = td->td_ma;
1379 prev_td_ma_cnt = td->td_ma_cnt;
1381 while (uio_clone->uio_resid != 0) {
1382 len = uio_clone->uio_iov->iov_len;
1384 KASSERT(uio_clone->uio_iovcnt >= 1,
1385 ("iovcnt underflow"));
1386 uio_clone->uio_iov++;
1387 uio_clone->uio_iovcnt--;
1390 if (len > ptoa(io_hold_cnt))
1391 len = ptoa(io_hold_cnt);
1392 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1393 end = round_page(addr + len);
1398 cnt = atop(end - trunc_page(addr));
1400 * A perfectly misaligned address and length could cause
1401 * both the start and the end of the chunk to use partial
1402 * page. +2 accounts for such a situation.
1404 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1405 addr, len, prot, ma, io_hold_cnt + 2);
1410 short_uio.uio_iov = &short_iovec[0];
1411 short_iovec[0].iov_base = (void *)addr;
1412 short_uio.uio_iovcnt = 1;
1413 short_uio.uio_resid = short_iovec[0].iov_len = len;
1414 short_uio.uio_offset = uio_clone->uio_offset;
1416 td->td_ma_cnt = cnt;
1418 error = vn_io_fault_doio(args, &short_uio, td);
1419 vm_page_unhold_pages(ma, cnt);
1420 adv = len - short_uio.uio_resid;
1422 uio_clone->uio_iov->iov_base =
1423 (char *)uio_clone->uio_iov->iov_base + adv;
1424 uio_clone->uio_iov->iov_len -= adv;
1425 uio_clone->uio_resid -= adv;
1426 uio_clone->uio_offset += adv;
1428 uio->uio_resid -= adv;
1429 uio->uio_offset += adv;
1431 if (error != 0 || adv == 0)
1434 td->td_ma = prev_td_ma;
1435 td->td_ma_cnt = prev_td_ma_cnt;
1436 curthread_pflags_restore(saveheld);
1438 free(uio_clone, M_IOV);
1443 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1444 int flags, struct thread *td)
1449 struct vn_io_fault_args args;
1452 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1456 * The ability to read(2) on a directory has historically been
1457 * allowed for all users, but this can and has been the source of
1458 * at least one security issue in the past. As such, it is now hidden
1459 * away behind a sysctl for those that actually need it to use it, and
1460 * restricted to root when it's turned on to make it relatively safe to
1461 * leave on for longer sessions of need.
1463 if (vp->v_type == VDIR) {
1464 KASSERT(uio->uio_rw == UIO_READ,
1465 ("illegal write attempted on a directory"));
1466 if (!vfs_allow_read_dir)
1468 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1472 foffset_lock_uio(fp, uio, flags);
1473 if (do_vn_io_fault(vp, uio)) {
1474 args.kind = VN_IO_FAULT_FOP;
1475 args.args.fop_args.fp = fp;
1476 args.args.fop_args.doio = doio;
1477 args.cred = active_cred;
1478 args.flags = flags | FOF_OFFSET;
1479 if (uio->uio_rw == UIO_READ) {
1480 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1481 uio->uio_offset + uio->uio_resid);
1482 } else if ((fp->f_flag & O_APPEND) != 0 ||
1483 (flags & FOF_OFFSET) == 0) {
1484 /* For appenders, punt and lock the whole range. */
1485 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1487 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1488 uio->uio_offset + uio->uio_resid);
1490 error = vn_io_fault1(vp, uio, &args, td);
1491 vn_rangelock_unlock(vp, rl_cookie);
1493 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1495 foffset_unlock_uio(fp, uio, flags);
1500 * Helper function to perform the requested uiomove operation using
1501 * the held pages for io->uio_iov[0].iov_base buffer instead of
1502 * copyin/copyout. Access to the pages with uiomove_fromphys()
1503 * instead of iov_base prevents page faults that could occur due to
1504 * pmap_collect() invalidating the mapping created by
1505 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1506 * object cleanup revoking the write access from page mappings.
1508 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1509 * instead of plain uiomove().
1512 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1514 struct uio transp_uio;
1515 struct iovec transp_iov[1];
1521 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1522 uio->uio_segflg != UIO_USERSPACE)
1523 return (uiomove(data, xfersize, uio));
1525 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1526 transp_iov[0].iov_base = data;
1527 transp_uio.uio_iov = &transp_iov[0];
1528 transp_uio.uio_iovcnt = 1;
1529 if (xfersize > uio->uio_resid)
1530 xfersize = uio->uio_resid;
1531 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1532 transp_uio.uio_offset = 0;
1533 transp_uio.uio_segflg = UIO_SYSSPACE;
1535 * Since transp_iov points to data, and td_ma page array
1536 * corresponds to original uio->uio_iov, we need to invert the
1537 * direction of the i/o operation as passed to
1538 * uiomove_fromphys().
1540 switch (uio->uio_rw) {
1542 transp_uio.uio_rw = UIO_READ;
1545 transp_uio.uio_rw = UIO_WRITE;
1548 transp_uio.uio_td = uio->uio_td;
1549 error = uiomove_fromphys(td->td_ma,
1550 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1551 xfersize, &transp_uio);
1552 adv = xfersize - transp_uio.uio_resid;
1554 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1555 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1557 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1559 td->td_ma_cnt -= pgadv;
1560 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1561 uio->uio_iov->iov_len -= adv;
1562 uio->uio_resid -= adv;
1563 uio->uio_offset += adv;
1568 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1572 vm_offset_t iov_base;
1576 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1577 uio->uio_segflg != UIO_USERSPACE)
1578 return (uiomove_fromphys(ma, offset, xfersize, uio));
1580 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1581 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1582 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1583 switch (uio->uio_rw) {
1585 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1589 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1593 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1595 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1597 td->td_ma_cnt -= pgadv;
1598 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1599 uio->uio_iov->iov_len -= cnt;
1600 uio->uio_resid -= cnt;
1601 uio->uio_offset += cnt;
1606 * File table truncate routine.
1609 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1621 * Lock the whole range for truncation. Otherwise split i/o
1622 * might happen partly before and partly after the truncation.
1624 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1625 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1628 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1629 AUDIT_ARG_VNODE1(vp);
1630 if (vp->v_type == VDIR) {
1635 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1639 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1643 vn_finished_write(mp);
1645 vn_rangelock_unlock(vp, rl_cookie);
1646 if (error == ERELOOKUP)
1652 * Truncate a file that is already locked.
1655 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1661 error = VOP_ADD_WRITECOUNT(vp, 1);
1664 vattr.va_size = length;
1666 vattr.va_vaflags |= VA_SYNC;
1667 error = VOP_SETATTR(vp, &vattr, cred);
1668 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1674 * File table vnode stat routine.
1677 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1680 struct vnode *vp = fp->f_vnode;
1683 vn_lock(vp, LK_SHARED | LK_RETRY);
1684 error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1691 * File table vnode ioctl routine.
1694 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1699 struct fiobmap2_arg *bmarg;
1703 switch (vp->v_type) {
1708 vn_lock(vp, LK_SHARED | LK_RETRY);
1709 error = VOP_GETATTR(vp, &vattr, active_cred);
1712 *(int *)data = vattr.va_size - fp->f_offset;
1715 bmarg = (struct fiobmap2_arg *)data;
1716 vn_lock(vp, LK_SHARED | LK_RETRY);
1718 error = mac_vnode_check_read(active_cred, fp->f_cred,
1722 error = VOP_BMAP(vp, bmarg->bn, NULL,
1723 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1730 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1735 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1743 * File table vnode poll routine.
1746 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1753 #if defined(MAC) || defined(AUDIT)
1754 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1755 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1756 AUDIT_ARG_VNODE1(vp);
1757 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1763 error = VOP_POLL(vp, events, fp->f_cred, td);
1768 * Acquire the requested lock and then check for validity. LK_RETRY
1769 * permits vn_lock to return doomed vnodes.
1771 static int __noinline
1772 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1776 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1777 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1780 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1782 if ((flags & LK_RETRY) == 0) {
1793 * Nothing to do if we got the lock.
1799 * Interlock was dropped by the call in _vn_lock.
1801 flags &= ~LK_INTERLOCK;
1803 error = VOP_LOCK1(vp, flags, file, line);
1804 } while (error != 0);
1809 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1813 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1814 ("vn_lock: no locktype (%d passed)", flags));
1815 VNPASS(vp->v_holdcnt > 0, vp);
1816 error = VOP_LOCK1(vp, flags, file, line);
1817 if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1818 return (_vn_lock_fallback(vp, flags, file, line, error));
1823 * File table vnode close routine.
1826 vn_closefile(struct file *fp, struct thread *td)
1834 fp->f_ops = &badfileops;
1835 ref = (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1837 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1839 if (__predict_false(ref)) {
1840 lf.l_whence = SEEK_SET;
1843 lf.l_type = F_UNLCK;
1844 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1851 * Preparing to start a filesystem write operation. If the operation is
1852 * permitted, then we bump the count of operations in progress and
1853 * proceed. If a suspend request is in progress, we wait until the
1854 * suspension is over, and then proceed.
1857 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1859 struct mount_pcpu *mpcpu;
1862 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1863 vfs_op_thread_enter(mp, mpcpu)) {
1864 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1865 vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1866 vfs_op_thread_exit(mp, mpcpu);
1871 mtx_assert(MNT_MTX(mp), MA_OWNED);
1878 * Check on status of suspension.
1880 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1881 mp->mnt_susp_owner != curthread) {
1882 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1883 (flags & PCATCH) : 0) | (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_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1912 ("V_MNTREF requires mp"));
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().
1937 if (vp == NULL && (flags & V_MNTREF) == 0)
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_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1957 ("V_MNTREF requires mp"));
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().
1983 if (vp == NULL && (flags & V_MNTREF) == 0)
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.
1999 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
2000 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
2009 * Filesystem write operation has completed. If we are suspending and this
2010 * operation is the last one, notify the suspender that the suspension is
2014 vn_finished_write(struct mount *mp)
2016 struct mount_pcpu *mpcpu;
2022 if (vfs_op_thread_enter(mp, mpcpu)) {
2023 vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2024 vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2025 vfs_op_thread_exit(mp, mpcpu);
2030 vfs_assert_mount_counters(mp);
2032 c = --mp->mnt_writeopcount;
2033 if (mp->mnt_vfs_ops == 0) {
2034 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2039 vfs_dump_mount_counters(mp);
2040 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2041 wakeup(&mp->mnt_writeopcount);
2046 * Filesystem secondary write operation has completed. If we are
2047 * suspending and this operation is the last one, notify the suspender
2048 * that the suspension is now in effect.
2051 vn_finished_secondary_write(struct mount *mp)
2057 mp->mnt_secondary_writes--;
2058 if (mp->mnt_secondary_writes < 0)
2059 panic("vn_finished_secondary_write: neg cnt");
2060 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2061 mp->mnt_secondary_writes <= 0)
2062 wakeup(&mp->mnt_secondary_writes);
2067 * Request a filesystem to suspend write operations.
2070 vfs_write_suspend(struct mount *mp, int flags)
2077 vfs_assert_mount_counters(mp);
2078 if (mp->mnt_susp_owner == curthread) {
2079 vfs_op_exit_locked(mp);
2083 while (mp->mnt_kern_flag & MNTK_SUSPEND)
2084 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2087 * Unmount holds a write reference on the mount point. If we
2088 * own busy reference and drain for writers, we deadlock with
2089 * the reference draining in the unmount path. Callers of
2090 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2091 * vfs_busy() reference is owned and caller is not in the
2094 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2095 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2096 vfs_op_exit_locked(mp);
2101 mp->mnt_kern_flag |= MNTK_SUSPEND;
2102 mp->mnt_susp_owner = curthread;
2103 if (mp->mnt_writeopcount > 0)
2104 (void) msleep(&mp->mnt_writeopcount,
2105 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2108 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2109 vfs_write_resume(mp, 0);
2110 /* vfs_write_resume does vfs_op_exit() for us */
2116 * Request a filesystem to resume write operations.
2119 vfs_write_resume(struct mount *mp, int flags)
2123 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2124 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2125 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2127 mp->mnt_susp_owner = NULL;
2128 wakeup(&mp->mnt_writeopcount);
2129 wakeup(&mp->mnt_flag);
2130 curthread->td_pflags &= ~TDP_IGNSUSP;
2131 if ((flags & VR_START_WRITE) != 0) {
2133 mp->mnt_writeopcount++;
2136 if ((flags & VR_NO_SUSPCLR) == 0)
2139 } else if ((flags & VR_START_WRITE) != 0) {
2141 vn_start_write_refed(mp, 0, true);
2148 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2152 vfs_write_suspend_umnt(struct mount *mp)
2156 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2157 ("vfs_write_suspend_umnt: recursed"));
2159 /* dounmount() already called vn_start_write(). */
2161 vn_finished_write(mp);
2162 error = vfs_write_suspend(mp, 0);
2164 vn_start_write(NULL, &mp, V_WAIT);
2168 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2171 vn_start_write(NULL, &mp, V_WAIT);
2173 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2174 wakeup(&mp->mnt_flag);
2176 curthread->td_pflags |= TDP_IGNSUSP;
2181 * Implement kqueues for files by translating it to vnode operation.
2184 vn_kqfilter(struct file *fp, struct knote *kn)
2187 return (VOP_KQFILTER(fp->f_vnode, kn));
2191 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2193 if ((fp->f_flag & FKQALLOWED) == 0)
2195 return (vn_kqfilter(fp, kn));
2199 * Simplified in-kernel wrapper calls for extended attribute access.
2200 * Both calls pass in a NULL credential, authorizing as "kernel" access.
2201 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2204 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2205 const char *attrname, int *buflen, char *buf, struct thread *td)
2211 iov.iov_len = *buflen;
2214 auio.uio_iov = &iov;
2215 auio.uio_iovcnt = 1;
2216 auio.uio_rw = UIO_READ;
2217 auio.uio_segflg = UIO_SYSSPACE;
2219 auio.uio_offset = 0;
2220 auio.uio_resid = *buflen;
2222 if ((ioflg & IO_NODELOCKED) == 0)
2223 vn_lock(vp, LK_SHARED | LK_RETRY);
2225 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2227 /* authorize attribute retrieval as kernel */
2228 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2231 if ((ioflg & IO_NODELOCKED) == 0)
2235 *buflen = *buflen - auio.uio_resid;
2242 * XXX failure mode if partially written?
2245 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2246 const char *attrname, int buflen, char *buf, struct thread *td)
2253 iov.iov_len = buflen;
2256 auio.uio_iov = &iov;
2257 auio.uio_iovcnt = 1;
2258 auio.uio_rw = UIO_WRITE;
2259 auio.uio_segflg = UIO_SYSSPACE;
2261 auio.uio_offset = 0;
2262 auio.uio_resid = buflen;
2264 if ((ioflg & IO_NODELOCKED) == 0) {
2265 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2267 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2270 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2272 /* authorize attribute setting as kernel */
2273 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2275 if ((ioflg & IO_NODELOCKED) == 0) {
2276 vn_finished_write(mp);
2284 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2285 const char *attrname, struct thread *td)
2290 if ((ioflg & IO_NODELOCKED) == 0) {
2291 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2293 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2296 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2298 /* authorize attribute removal as kernel */
2299 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2300 if (error == EOPNOTSUPP)
2301 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2304 if ((ioflg & IO_NODELOCKED) == 0) {
2305 vn_finished_write(mp);
2313 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2317 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2321 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2324 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2329 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2330 int lkflags, struct vnode **rvp)
2335 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2337 ltype = VOP_ISLOCKED(vp);
2338 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2339 ("vn_vget_ino: vp not locked"));
2340 error = vfs_busy(mp, MBF_NOWAIT);
2344 error = vfs_busy(mp, 0);
2345 vn_lock(vp, ltype | LK_RETRY);
2349 if (VN_IS_DOOMED(vp)) {
2355 error = alloc(mp, alloc_arg, lkflags, rvp);
2357 if (error != 0 || *rvp != vp)
2358 vn_lock(vp, ltype | LK_RETRY);
2359 if (VN_IS_DOOMED(vp)) {
2372 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2382 * There are conditions where the limit is to be ignored.
2383 * However, since it is almost never reached, check it first.
2385 ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2386 lim = lim_cur(td, RLIMIT_FSIZE);
2387 if (__predict_false(ktr_write))
2388 lim = td->td_ktr_io_lim;
2389 if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2393 * The limit is reached.
2395 if (vp->v_type != VREG ||
2396 (td->td_pflags2 & TDP2_ACCT) != 0)
2399 if (!ktr_write || ktr_filesize_limit_signal) {
2400 PROC_LOCK(td->td_proc);
2401 kern_psignal(td->td_proc, SIGXFSZ);
2402 PROC_UNLOCK(td->td_proc);
2408 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2415 vn_lock(vp, LK_SHARED | LK_RETRY);
2416 AUDIT_ARG_VNODE1(vp);
2419 return (setfmode(td, active_cred, vp, mode));
2423 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2430 vn_lock(vp, LK_SHARED | LK_RETRY);
2431 AUDIT_ARG_VNODE1(vp);
2434 return (setfown(td, active_cred, vp, uid, gid));
2438 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2442 if ((object = vp->v_object) == NULL)
2444 VM_OBJECT_WLOCK(object);
2445 vm_object_page_remove(object, start, end, 0);
2446 VM_OBJECT_WUNLOCK(object);
2450 vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off,
2459 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2460 ("%s: Wrong command %lu", __func__, cmd));
2461 ASSERT_VOP_LOCKED(vp, "vn_bmap_seekhole_locked");
2463 if (vp->v_type != VREG) {
2467 error = VOP_GETATTR(vp, &va, cred);
2471 if (noff >= va.va_size) {
2475 bsize = vp->v_mount->mnt_stat.f_iosize;
2476 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2478 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2479 if (error == EOPNOTSUPP) {
2483 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2484 (bnp != -1 && cmd == FIOSEEKDATA)) {
2491 if (noff > va.va_size)
2493 /* noff == va.va_size. There is an implicit hole at the end of file. */
2494 if (cmd == FIOSEEKDATA)
2503 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2507 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2508 ("%s: Wrong command %lu", __func__, cmd));
2510 if (vn_lock(vp, LK_SHARED) != 0)
2512 error = vn_bmap_seekhole_locked(vp, cmd, off, cred);
2518 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2523 off_t foffset, size;
2526 cred = td->td_ucred;
2528 foffset = foffset_lock(fp, 0);
2529 noneg = (vp->v_type != VCHR);
2535 (offset > 0 && foffset > OFF_MAX - offset))) {
2542 vn_lock(vp, LK_SHARED | LK_RETRY);
2543 error = VOP_GETATTR(vp, &vattr, cred);
2549 * If the file references a disk device, then fetch
2550 * the media size and use that to determine the ending
2553 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2554 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2555 vattr.va_size = size;
2557 (vattr.va_size > OFF_MAX ||
2558 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2562 offset += vattr.va_size;
2567 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2568 if (error == ENOTTY)
2572 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2573 if (error == ENOTTY)
2579 if (error == 0 && noneg && offset < 0)
2583 VFS_KNOTE_UNLOCKED(vp, 0);
2584 td->td_uretoff.tdu_off = offset;
2586 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2591 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2597 * Grant permission if the caller is the owner of the file, or
2598 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2599 * on the file. If the time pointer is null, then write
2600 * permission on the file is also sufficient.
2602 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2603 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2604 * will be allowed to set the times [..] to the current
2607 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2608 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2609 error = VOP_ACCESS(vp, VWRITE, cred, td);
2614 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2619 if (fp->f_type == DTYPE_FIFO)
2620 kif->kf_type = KF_TYPE_FIFO;
2622 kif->kf_type = KF_TYPE_VNODE;
2625 FILEDESC_SUNLOCK(fdp);
2626 error = vn_fill_kinfo_vnode(vp, kif);
2628 FILEDESC_SLOCK(fdp);
2633 vn_fill_junk(struct kinfo_file *kif)
2638 * Simulate vn_fullpath returning changing values for a given
2639 * vp during e.g. coredump.
2641 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2642 olen = strlen(kif->kf_path);
2644 strcpy(&kif->kf_path[len - 1], "$");
2646 for (; olen < len; olen++)
2647 strcpy(&kif->kf_path[olen], "A");
2651 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2654 char *fullpath, *freepath;
2657 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2660 error = vn_fullpath(vp, &fullpath, &freepath);
2662 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2664 if (freepath != NULL)
2665 free(freepath, M_TEMP);
2667 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2672 * Retrieve vnode attributes.
2674 va.va_fsid = VNOVAL;
2676 vn_lock(vp, LK_SHARED | LK_RETRY);
2677 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2681 if (va.va_fsid != VNOVAL)
2682 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2684 kif->kf_un.kf_file.kf_file_fsid =
2685 vp->v_mount->mnt_stat.f_fsid.val[0];
2686 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2687 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2688 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2689 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2690 kif->kf_un.kf_file.kf_file_size = va.va_size;
2691 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2692 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2693 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2698 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2699 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2703 struct pmckern_map_in pkm;
2709 boolean_t writecounted;
2712 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2713 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2715 * POSIX shared-memory objects are defined to have
2716 * kernel persistence, and are not defined to support
2717 * read(2)/write(2) -- or even open(2). Thus, we can
2718 * use MAP_ASYNC to trade on-disk coherence for speed.
2719 * The shm_open(3) library routine turns on the FPOSIXSHM
2720 * flag to request this behavior.
2722 if ((fp->f_flag & FPOSIXSHM) != 0)
2723 flags |= MAP_NOSYNC;
2728 * Ensure that file and memory protections are
2729 * compatible. Note that we only worry about
2730 * writability if mapping is shared; in this case,
2731 * current and max prot are dictated by the open file.
2732 * XXX use the vnode instead? Problem is: what
2733 * credentials do we use for determination? What if
2734 * proc does a setuid?
2737 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2738 maxprot = VM_PROT_NONE;
2739 if ((prot & VM_PROT_EXECUTE) != 0)
2742 maxprot = VM_PROT_EXECUTE;
2743 if ((fp->f_flag & FREAD) != 0)
2744 maxprot |= VM_PROT_READ;
2745 else if ((prot & VM_PROT_READ) != 0)
2749 * If we are sharing potential changes via MAP_SHARED and we
2750 * are trying to get write permission although we opened it
2751 * without asking for it, bail out.
2753 if ((flags & MAP_SHARED) != 0) {
2754 if ((fp->f_flag & FWRITE) != 0)
2755 maxprot |= VM_PROT_WRITE;
2756 else if ((prot & VM_PROT_WRITE) != 0)
2759 maxprot |= VM_PROT_WRITE;
2760 cap_maxprot |= VM_PROT_WRITE;
2762 maxprot &= cap_maxprot;
2765 * For regular files and shared memory, POSIX requires that
2766 * the value of foff be a legitimate offset within the data
2767 * object. In particular, negative offsets are invalid.
2768 * Blocking negative offsets and overflows here avoids
2769 * possible wraparound or user-level access into reserved
2770 * ranges of the data object later. In contrast, POSIX does
2771 * not dictate how offsets are used by device drivers, so in
2772 * the case of a device mapping a negative offset is passed
2779 foff > OFF_MAX - size)
2782 writecounted = FALSE;
2783 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2784 &foff, &object, &writecounted);
2787 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2788 foff, writecounted, td);
2791 * If this mapping was accounted for in the vnode's
2792 * writecount, then undo that now.
2795 vm_pager_release_writecount(object, 0, size);
2796 vm_object_deallocate(object);
2799 /* Inform hwpmc(4) if an executable is being mapped. */
2800 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2801 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2803 pkm.pm_address = (uintptr_t) *addr;
2804 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2812 vn_fsid(struct vnode *vp, struct vattr *va)
2816 f = &vp->v_mount->mnt_stat.f_fsid;
2817 va->va_fsid = (uint32_t)f->val[1];
2818 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2819 va->va_fsid += (uint32_t)f->val[0];
2823 vn_fsync_buf(struct vnode *vp, int waitfor)
2825 struct buf *bp, *nbp;
2828 int error, maxretry;
2831 maxretry = 10000; /* large, arbitrarily chosen */
2833 if (vp->v_type == VCHR) {
2835 mp = vp->v_rdev->si_mountpt;
2842 * MARK/SCAN initialization to avoid infinite loops.
2844 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2845 bp->b_vflags &= ~BV_SCANNED;
2850 * Flush all dirty buffers associated with a vnode.
2853 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2854 if ((bp->b_vflags & BV_SCANNED) != 0)
2856 bp->b_vflags |= BV_SCANNED;
2857 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2858 if (waitfor != MNT_WAIT)
2861 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2862 BO_LOCKPTR(bo)) != 0) {
2869 KASSERT(bp->b_bufobj == bo,
2870 ("bp %p wrong b_bufobj %p should be %p",
2871 bp, bp->b_bufobj, bo));
2872 if ((bp->b_flags & B_DELWRI) == 0)
2873 panic("fsync: not dirty");
2874 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2880 if (maxretry < 1000)
2881 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2887 * If synchronous the caller expects us to completely resolve all
2888 * dirty buffers in the system. Wait for in-progress I/O to
2889 * complete (which could include background bitmap writes), then
2890 * retry if dirty blocks still exist.
2892 if (waitfor == MNT_WAIT) {
2893 bufobj_wwait(bo, 0, 0);
2894 if (bo->bo_dirty.bv_cnt > 0) {
2896 * If we are unable to write any of these buffers
2897 * then we fail now rather than trying endlessly
2898 * to write them out.
2900 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2901 if ((error = bp->b_error) != 0)
2903 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2904 (error == 0 && --maxretry >= 0))
2912 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2918 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
2919 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2920 * to do the actual copy.
2921 * vn_generic_copy_file_range() is factored out, so it can be called
2922 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2923 * different file systems.
2926 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2927 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2928 struct ucred *outcred, struct thread *fsize_td)
2935 *lenp = 0; /* For error returns. */
2938 /* Do some sanity checks on the arguments. */
2939 if (invp->v_type == VDIR || outvp->v_type == VDIR)
2941 else if (*inoffp < 0 || *outoffp < 0 ||
2942 invp->v_type != VREG || outvp->v_type != VREG)
2947 /* Ensure offset + len does not wrap around. */
2950 if (uval > INT64_MAX)
2951 len = INT64_MAX - *inoffp;
2954 if (uval > INT64_MAX)
2955 len = INT64_MAX - *outoffp;
2960 * If the two vnode are for the same file system, call
2961 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2962 * which can handle copies across multiple file systems.
2965 if (invp->v_mount == outvp->v_mount)
2966 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2967 lenp, flags, incred, outcred, fsize_td);
2969 error = vn_generic_copy_file_range(invp, inoffp, outvp,
2970 outoffp, lenp, flags, incred, outcred, fsize_td);
2976 * Test len bytes of data starting at dat for all bytes == 0.
2977 * Return true if all bytes are zero, false otherwise.
2978 * Expects dat to be well aligned.
2981 mem_iszero(void *dat, int len)
2987 for (p = dat; len > 0; len -= sizeof(*p), p++) {
2988 if (len >= sizeof(*p)) {
2992 cp = (const char *)p;
2993 for (i = 0; i < len; i++, cp++)
3002 * Look for a hole in the output file and, if found, adjust *outoffp
3003 * and *xferp to skip past the hole.
3004 * *xferp is the entire hole length to be written and xfer2 is how many bytes
3005 * to be written as 0's upon return.
3008 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3009 off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3014 if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3015 *dataoffp = *outoffp;
3016 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3019 *holeoffp = *dataoffp;
3020 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3023 if (error != 0 || *holeoffp == *dataoffp) {
3025 * Since outvp is unlocked, it may be possible for
3026 * another thread to do a truncate(), lseek(), write()
3027 * creating a hole at startoff between the above
3028 * VOP_IOCTL() calls, if the other thread does not do
3030 * If that happens, *holeoffp == *dataoffp and finding
3031 * the hole has failed, so disable vn_skip_hole().
3033 *holeoffp = -1; /* Disable use of vn_skip_hole(). */
3036 KASSERT(*dataoffp >= *outoffp,
3037 ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3038 (intmax_t)*dataoffp, (intmax_t)*outoffp));
3039 KASSERT(*holeoffp > *dataoffp,
3040 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3041 (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3045 * If there is a hole before the data starts, advance *outoffp and
3046 * *xferp past the hole.
3048 if (*dataoffp > *outoffp) {
3049 delta = *dataoffp - *outoffp;
3050 if (delta >= *xferp) {
3051 /* Entire *xferp is a hole. */
3058 xfer2 = MIN(xfer2, *xferp);
3062 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3063 * that the write ends at the start of the hole.
3064 * *holeoffp should always be greater than *outoffp, but for the
3065 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3068 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3069 xfer2 = *holeoffp - *outoffp;
3074 * Write an xfer sized chunk to outvp in blksize blocks from dat.
3075 * dat is a maximum of blksize in length and can be written repeatedly in
3077 * If growfile == true, just grow the file via vn_truncate_locked() instead
3078 * of doing actual writes.
3079 * If checkhole == true, a hole is being punched, so skip over any hole
3080 * already in the output file.
3083 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3084 u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3087 off_t dataoff, holeoff, xfer2;
3091 * Loop around doing writes of blksize until write has been completed.
3092 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3093 * done for each iteration, since the xfer argument can be very
3094 * large if there is a large hole to punch in the output file.
3099 xfer2 = MIN(xfer, blksize);
3102 * Punching a hole. Skip writing if there is
3103 * already a hole in the output file.
3105 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3106 &dataoff, &holeoff, cred);
3111 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3116 error = vn_start_write(outvp, &mp, V_WAIT);
3120 error = vn_lock(outvp, LK_EXCLUSIVE);
3122 error = vn_truncate_locked(outvp, outoff + xfer,
3127 error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3129 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3130 outoff, UIO_SYSSPACE, IO_NODELOCKED,
3131 curthread->td_ucred, cred, NULL, curthread);
3138 vn_finished_write(mp);
3139 } while (!growfile && xfer > 0 && error == 0);
3144 * Copy a byte range of one file to another. This function can handle the
3145 * case where invp and outvp are on different file systems.
3146 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3147 * is no better file system specific way to do it.
3150 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3151 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3152 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3154 struct vattr va, inva;
3157 off_t startoff, endoff, xfer, xfer2;
3159 int error, interrupted;
3160 bool cantseek, readzeros, eof, lastblock, holetoeof;
3162 size_t copylen, len, rem, savlen;
3164 long holein, holeout;
3165 struct timespec curts, endts;
3167 holein = holeout = 0;
3168 savlen = len = *lenp;
3173 error = vn_lock(invp, LK_SHARED);
3176 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3179 error = VOP_GETATTR(invp, &inva, incred);
3185 error = vn_start_write(outvp, &mp, V_WAIT);
3187 error = vn_lock(outvp, LK_EXCLUSIVE);
3190 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3191 * now that outvp is locked.
3193 if (fsize_td != NULL) {
3194 io.uio_offset = *outoffp;
3196 error = vn_rlimit_fsize(outvp, &io, fsize_td);
3200 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3203 * Holes that are past EOF do not need to be written as a block
3204 * of zero bytes. So, truncate the output file as far as
3205 * possible and then use va.va_size to decide if writing 0
3206 * bytes is necessary in the loop below.
3209 error = VOP_GETATTR(outvp, &va, outcred);
3210 if (error == 0 && va.va_size > *outoffp && va.va_size <=
3213 error = mac_vnode_check_write(curthread->td_ucred,
3217 error = vn_truncate_locked(outvp, *outoffp,
3220 va.va_size = *outoffp;
3225 vn_finished_write(mp);
3230 * Set the blksize to the larger of the hole sizes for invp and outvp.
3231 * If hole sizes aren't available, set the blksize to the larger
3232 * f_iosize of invp and outvp.
3233 * This code expects the hole sizes and f_iosizes to be powers of 2.
3234 * This value is clipped at 4Kbytes and 1Mbyte.
3236 blksize = MAX(holein, holeout);
3238 /* Clip len to end at an exact multiple of hole size. */
3240 rem = *inoffp % blksize;
3242 rem = blksize - rem;
3243 if (len > rem && len - rem > blksize)
3244 len = savlen = rounddown(len - rem, blksize) + rem;
3248 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3249 outvp->v_mount->mnt_stat.f_iosize);
3252 else if (blksize > 1024 * 1024)
3253 blksize = 1024 * 1024;
3254 dat = malloc(blksize, M_TEMP, M_WAITOK);
3257 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3258 * to find holes. Otherwise, just scan the read block for all 0s
3259 * in the inner loop where the data copying is done.
3260 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3261 * support holes on the server, but do not support FIOSEEKHOLE.
3262 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3263 * that this function should return after 1second with a partial
3266 if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3267 getnanouptime(&endts);
3270 timespecclear(&endts);
3271 holetoeof = eof = false;
3272 while (len > 0 && error == 0 && !eof && interrupted == 0) {
3273 endoff = 0; /* To shut up compilers. */
3279 * Find the next data area. If there is just a hole to EOF,
3280 * FIOSEEKDATA should fail with ENXIO.
3281 * (I do not know if any file system will report a hole to
3282 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3283 * will fail for those file systems.)
3285 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3286 * the code just falls through to the inner copy loop.
3290 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3292 if (error == ENXIO) {
3293 startoff = endoff = inva.va_size;
3294 eof = holetoeof = true;
3298 if (error == 0 && !holetoeof) {
3300 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3303 * Since invp is unlocked, it may be possible for
3304 * another thread to do a truncate(), lseek(), write()
3305 * creating a hole at startoff between the above
3306 * VOP_IOCTL() calls, if the other thread does not do
3308 * If that happens, startoff == endoff and finding
3309 * the hole has failed, so set an error.
3311 if (error == 0 && startoff == endoff)
3312 error = EINVAL; /* Any error. Reset to 0. */
3315 if (startoff > *inoffp) {
3316 /* Found hole before data block. */
3317 xfer = MIN(startoff - *inoffp, len);
3318 if (*outoffp < va.va_size) {
3319 /* Must write 0s to punch hole. */
3320 xfer2 = MIN(va.va_size - *outoffp,
3322 memset(dat, 0, MIN(xfer2, blksize));
3323 error = vn_write_outvp(outvp, dat,
3324 *outoffp, xfer2, blksize, false,
3325 holeout > 0, outcred);
3328 if (error == 0 && *outoffp + xfer >
3329 va.va_size && (xfer == len || holetoeof)) {
3330 /* Grow output file (hole at end). */
3331 error = vn_write_outvp(outvp, dat,
3332 *outoffp, xfer, blksize, true,
3340 interrupted = sig_intr();
3341 if (timespecisset(&endts) &&
3343 getnanouptime(&curts);
3344 if (timespeccmp(&curts,
3352 copylen = MIN(len, endoff - startoff);
3364 * Set first xfer to end at a block boundary, so that
3365 * holes are more likely detected in the loop below via
3366 * the for all bytes 0 method.
3368 xfer -= (*inoffp % blksize);
3370 /* Loop copying the data block. */
3371 while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3374 error = vn_lock(invp, LK_SHARED);
3377 error = vn_rdwr(UIO_READ, invp, dat, xfer,
3378 startoff, UIO_SYSSPACE, IO_NODELOCKED,
3379 curthread->td_ucred, incred, &aresid,
3383 if (error == 0 && aresid > 0) {
3384 /* Stop the copy at EOF on the input file. */
3391 * Skip the write for holes past the initial EOF
3392 * of the output file, unless this is the last
3393 * write of the output file at EOF.
3395 readzeros = cantseek ? mem_iszero(dat, xfer) :
3399 if (!cantseek || *outoffp < va.va_size ||
3400 lastblock || !readzeros)
3401 error = vn_write_outvp(outvp, dat,
3402 *outoffp, xfer, blksize,
3403 readzeros && lastblock &&
3404 *outoffp >= va.va_size, false,
3413 interrupted = sig_intr();
3414 if (timespecisset(&endts) &&
3416 getnanouptime(&curts);
3417 if (timespeccmp(&curts,
3429 *lenp = savlen - len;
3435 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3439 off_t olen, ooffset;
3442 int audited_vnode1 = 0;
3446 if (vp->v_type != VREG)
3449 /* Allocating blocks may take a long time, so iterate. */
3456 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3459 error = vn_lock(vp, LK_EXCLUSIVE);
3461 vn_finished_write(mp);
3465 if (!audited_vnode1) {
3466 AUDIT_ARG_VNODE1(vp);
3471 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3474 error = VOP_ALLOCATE(vp, &offset, &len);
3476 vn_finished_write(mp);
3478 if (olen + ooffset != offset + len) {
3479 panic("offset + len changed from %jx/%jx to %jx/%jx",
3480 ooffset, olen, offset, len);
3482 if (error != 0 || len == 0)
3484 KASSERT(olen > len, ("Iteration did not make progress?"));
3492 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags,
3493 int ioflag, struct ucred *cred, struct ucred *active_cred,
3494 struct ucred *file_cred)
3501 bool audited_vnode1 = false;
3510 if ((ioflag & (IO_NODELOCKED | IO_RANGELOCKED)) == 0)
3511 rl_cookie = vn_rangelock_wlock(vp, off, off + len);
3512 while (len > 0 && error == 0) {
3514 * Try to deallocate the longest range in one pass.
3515 * In case a pass takes too long to be executed, it returns
3516 * partial result. The residue will be proceeded in the next
3520 if ((ioflag & IO_NODELOCKED) == 0) {
3522 if ((error = vn_start_write(vp, &mp,
3523 V_WAIT | PCATCH)) != 0)
3525 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
3528 if (!audited_vnode1) {
3529 AUDIT_ARG_VNODE1(vp);
3530 audited_vnode1 = true;
3535 if ((ioflag & IO_NOMACCHECK) == 0)
3536 error = mac_vnode_check_write(active_cred, file_cred,
3540 error = VOP_DEALLOCATE(vp, &off, &len, flags, ioflag,
3543 if ((ioflag & IO_NODELOCKED) == 0) {
3546 vn_finished_write(mp);
3550 if (error == 0 && len != 0)
3554 if (rl_cookie != NULL)
3555 vn_rangelock_unlock(vp, rl_cookie);
3562 * This function is supposed to be used in the situations where the deallocation
3563 * is not triggered by a user request.
3566 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags,
3567 int ioflag, struct ucred *active_cred, struct ucred *file_cred)
3571 if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset ||
3574 if (vp->v_type != VREG)
3577 cred = file_cred != NOCRED ? file_cred : active_cred;
3578 return (vn_deallocate_impl(vp, offset, length, flags, ioflag, cred,
3579 active_cred, file_cred));
3583 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
3584 struct ucred *active_cred, struct thread *td)
3592 if (cmd != SPACECTL_DEALLOC || *offset < 0 || *length <= 0 ||
3593 *length > OFF_MAX - *offset || flags != 0)
3595 if (vp->v_type != VREG)
3598 ioflag = get_write_ioflag(fp);
3601 case SPACECTL_DEALLOC:
3602 error = vn_deallocate_impl(vp, offset, length, flags, ioflag,
3603 active_cred, active_cred, fp->f_cred);
3606 panic("vn_fspacectl: unknown cmd %d", cmd);
3612 static u_long vn_lock_pair_pause_cnt;
3613 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3614 &vn_lock_pair_pause_cnt, 0,
3615 "Count of vn_lock_pair deadlocks");
3617 u_int vn_lock_pair_pause_max;
3618 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3619 &vn_lock_pair_pause_max, 0,
3620 "Max ticks for vn_lock_pair deadlock avoidance sleep");
3623 vn_lock_pair_pause(const char *wmesg)
3625 atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3626 pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3630 * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3631 * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3632 * must be unlocked. Same for vp2 and vp2_locked. One of the vnodes
3635 * The function returns with both vnodes exclusively locked, and
3636 * guarantees that it does not create lock order reversal with other
3637 * threads during its execution. Both vnodes could be unlocked
3638 * temporary (and reclaimed).
3641 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3646 if (vp1 == NULL && vp2 == NULL)
3650 ASSERT_VOP_ELOCKED(vp1, "vp1");
3652 ASSERT_VOP_UNLOCKED(vp1, "vp1");
3658 ASSERT_VOP_ELOCKED(vp2, "vp2");
3660 ASSERT_VOP_UNLOCKED(vp2, "vp2");
3664 if (!vp1_locked && !vp2_locked) {
3665 vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3670 if (vp1_locked && vp2_locked)
3672 if (vp1_locked && vp2 != NULL) {
3674 error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3675 __FILE__, __LINE__);
3680 vn_lock_pair_pause("vlp1");
3682 vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3685 if (vp2_locked && vp1 != NULL) {
3687 error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3688 __FILE__, __LINE__);
3693 vn_lock_pair_pause("vlp2");
3695 vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3700 ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3702 ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3706 vn_lktype_write(struct mount *mp, struct vnode *vp)
3708 if (MNT_SHARED_WRITES(mp) ||
3709 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
3711 return (LK_EXCLUSIVE);