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>
74 #include <sys/sleepqueue.h>
75 #include <sys/sysctl.h>
76 #include <sys/ttycom.h>
78 #include <sys/syslog.h>
79 #include <sys/unistd.h>
82 #include <security/audit/audit.h>
83 #include <security/mac/mac_framework.h>
86 #include <vm/vm_extern.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pager.h>
94 #include <sys/pmckern.h>
97 static fo_rdwr_t vn_read;
98 static fo_rdwr_t vn_write;
99 static fo_rdwr_t vn_io_fault;
100 static fo_truncate_t vn_truncate;
101 static fo_ioctl_t vn_ioctl;
102 static fo_poll_t vn_poll;
103 static fo_kqfilter_t vn_kqfilter;
104 static fo_stat_t vn_statfile;
105 static fo_close_t vn_closefile;
106 static fo_mmap_t vn_mmap;
107 static fo_fallocate_t vn_fallocate;
109 struct fileops vnops = {
110 .fo_read = vn_io_fault,
111 .fo_write = vn_io_fault,
112 .fo_truncate = vn_truncate,
113 .fo_ioctl = vn_ioctl,
115 .fo_kqfilter = vn_kqfilter,
116 .fo_stat = vn_statfile,
117 .fo_close = vn_closefile,
118 .fo_chmod = vn_chmod,
119 .fo_chown = vn_chown,
120 .fo_sendfile = vn_sendfile,
122 .fo_fill_kinfo = vn_fill_kinfo,
124 .fo_fallocate = vn_fallocate,
125 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
128 static const int io_hold_cnt = 16;
129 static int vn_io_fault_enable = 1;
130 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW,
131 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
132 static int vn_io_fault_prefault = 0;
133 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RW,
134 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
135 static u_long vn_io_faults_cnt;
136 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
137 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
139 static int vfs_allow_read_dir = 0;
140 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
141 &vfs_allow_read_dir, 0,
142 "Enable read(2) of directory by root for filesystems that support it");
145 * Returns true if vn_io_fault mode of handling the i/o request should
149 do_vn_io_fault(struct vnode *vp, struct uio *uio)
153 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
154 (mp = vp->v_mount) != NULL &&
155 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
159 * Structure used to pass arguments to vn_io_fault1(), to do either
160 * file- or vnode-based I/O calls.
162 struct vn_io_fault_args {
170 struct fop_args_tag {
174 struct vop_args_tag {
180 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
181 struct vn_io_fault_args *args, struct thread *td);
184 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
186 struct thread *td = ndp->ni_cnd.cn_thread;
188 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
192 * Common code for vnode open operations via a name lookup.
193 * Lookup the vnode and invoke VOP_CREATE if needed.
194 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
196 * Note that this does NOT free nameidata for the successful case,
197 * due to the NDINIT being done elsewhere.
200 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
201 struct ucred *cred, struct file *fp)
205 struct thread *td = ndp->ni_cnd.cn_thread;
207 struct vattr *vap = &vat;
212 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
213 O_EXCL | O_DIRECTORY))
215 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
216 ndp->ni_cnd.cn_nameiop = CREATE;
218 * Set NOCACHE to avoid flushing the cache when
219 * rolling in many files at once.
221 ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE;
222 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
223 ndp->ni_cnd.cn_flags |= FOLLOW;
224 if ((fmode & O_BENEATH) != 0)
225 ndp->ni_cnd.cn_flags |= BENEATH;
226 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
227 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
228 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
229 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
230 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
232 if ((error = namei(ndp)) != 0)
234 if (ndp->ni_vp == NULL) {
237 vap->va_mode = cmode;
239 vap->va_vaflags |= VA_EXCLUSIVE;
240 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
241 NDFREE(ndp, NDF_ONLY_PNBUF);
243 if ((error = vn_start_write(NULL, &mp,
244 V_XSLEEP | PCATCH)) != 0)
248 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
249 ndp->ni_cnd.cn_flags |= MAKEENTRY;
251 error = mac_vnode_check_create(cred, ndp->ni_dvp,
255 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
258 vn_finished_write(mp);
260 NDFREE(ndp, NDF_ONLY_PNBUF);
266 if (ndp->ni_dvp == ndp->ni_vp)
272 if (fmode & O_EXCL) {
276 if (vp->v_type == VDIR) {
283 ndp->ni_cnd.cn_nameiop = LOOKUP;
284 ndp->ni_cnd.cn_flags = ISOPEN |
285 ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
286 if (!(fmode & FWRITE))
287 ndp->ni_cnd.cn_flags |= LOCKSHARED;
288 if ((fmode & O_BENEATH) != 0)
289 ndp->ni_cnd.cn_flags |= BENEATH;
290 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
291 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
292 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
293 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
294 if ((error = namei(ndp)) != 0)
298 error = vn_open_vnode(vp, fmode, cred, td, fp);
304 NDFREE(ndp, NDF_ONLY_PNBUF);
312 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
315 int error, lock_flags, type;
317 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
318 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
320 KASSERT(fp != NULL, ("open with flock requires fp"));
321 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
324 lock_flags = VOP_ISLOCKED(vp);
327 lf.l_whence = SEEK_SET;
330 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
332 if ((fmode & FNONBLOCK) == 0)
334 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
336 fp->f_flag |= FHASLOCK;
338 vn_lock(vp, lock_flags | LK_RETRY);
339 if (error == 0 && VN_IS_DOOMED(vp))
345 * Common code for vnode open operations once a vnode is located.
346 * Check permissions, and call the VOP_OPEN routine.
349 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
350 struct thread *td, struct file *fp)
355 if (vp->v_type == VLNK)
357 if (vp->v_type == VSOCK)
359 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
362 if (fmode & (FWRITE | O_TRUNC)) {
363 if (vp->v_type == VDIR)
371 if ((fmode & O_APPEND) && (fmode & FWRITE))
376 if (fmode & O_VERIFY)
378 error = mac_vnode_check_open(cred, vp, accmode);
382 accmode &= ~(VCREAT | VVERIFY);
384 if ((fmode & O_CREAT) == 0 && accmode != 0) {
385 error = VOP_ACCESS(vp, accmode, cred, td);
389 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
390 vn_lock(vp, LK_UPGRADE | LK_RETRY);
391 error = VOP_OPEN(vp, fmode, cred, td, fp);
395 error = vn_open_vnode_advlock(vp, fmode, fp);
396 if (error == 0 && (fmode & FWRITE) != 0) {
397 error = VOP_ADD_WRITECOUNT(vp, 1);
399 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
400 __func__, vp, vp->v_writecount);
405 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
406 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
407 * Arrange for that by having fdrop() to use vn_closefile().
410 fp->f_flag |= FOPENFAILED;
412 if (fp->f_ops == &badfileops) {
413 fp->f_type = DTYPE_VNODE;
419 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
425 * Check for write permissions on the specified vnode.
426 * Prototype text segments cannot be written.
430 vn_writechk(struct vnode *vp)
433 ASSERT_VOP_LOCKED(vp, "vn_writechk");
435 * If there's shared text associated with
436 * the vnode, try to free it up once. If
437 * we fail, we can't allow writing.
449 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
450 struct thread *td, bool keep_ref)
453 int error, lock_flags;
455 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
456 MNT_EXTENDED_SHARED(vp->v_mount))
457 lock_flags = LK_SHARED;
459 lock_flags = LK_EXCLUSIVE;
461 vn_start_write(vp, &mp, V_WAIT);
462 vn_lock(vp, lock_flags | LK_RETRY);
463 AUDIT_ARG_VNODE1(vp);
464 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
465 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
466 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
467 __func__, vp, vp->v_writecount);
469 error = VOP_CLOSE(vp, flags, file_cred, td);
474 vn_finished_write(mp);
479 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
483 return (vn_close1(vp, flags, file_cred, td, false));
487 * Heuristic to detect sequential operation.
490 sequential_heuristic(struct uio *uio, struct file *fp)
494 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
497 if (fp->f_flag & FRDAHEAD)
498 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
501 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
502 * that the first I/O is normally considered to be slightly
503 * sequential. Seeking to offset 0 doesn't change sequentiality
504 * unless previous seeks have reduced f_seqcount to 0, in which
505 * case offset 0 is not special.
507 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
508 uio->uio_offset == fp->f_nextoff[rw]) {
510 * f_seqcount is in units of fixed-size blocks so that it
511 * depends mainly on the amount of sequential I/O and not
512 * much on the number of sequential I/O's. The fixed size
513 * of 16384 is hard-coded here since it is (not quite) just
514 * a magic size that works well here. This size is more
515 * closely related to the best I/O size for real disks than
516 * to any block size used by software.
518 if (uio->uio_resid >= IO_SEQMAX * 16384)
519 fp->f_seqcount[rw] = IO_SEQMAX;
521 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
522 if (fp->f_seqcount[rw] > IO_SEQMAX)
523 fp->f_seqcount[rw] = IO_SEQMAX;
525 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
528 /* Not sequential. Quickly draw-down sequentiality. */
529 if (fp->f_seqcount[rw] > 1)
530 fp->f_seqcount[rw] = 1;
532 fp->f_seqcount[rw] = 0;
537 * Package up an I/O request on a vnode into a uio and do it.
540 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
541 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
542 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
549 struct vn_io_fault_args args;
550 int error, lock_flags;
552 if (offset < 0 && vp->v_type != VCHR)
554 auio.uio_iov = &aiov;
556 aiov.iov_base = base;
558 auio.uio_resid = len;
559 auio.uio_offset = offset;
560 auio.uio_segflg = segflg;
565 if ((ioflg & IO_NODELOCKED) == 0) {
566 if ((ioflg & IO_RANGELOCKED) == 0) {
567 if (rw == UIO_READ) {
568 rl_cookie = vn_rangelock_rlock(vp, offset,
571 rl_cookie = vn_rangelock_wlock(vp, offset,
577 if (rw == UIO_WRITE) {
578 if (vp->v_type != VCHR &&
579 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
582 if (MNT_SHARED_WRITES(mp) ||
583 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
584 lock_flags = LK_SHARED;
586 lock_flags = LK_EXCLUSIVE;
588 lock_flags = LK_SHARED;
589 vn_lock(vp, lock_flags | LK_RETRY);
593 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
595 if ((ioflg & IO_NOMACCHECK) == 0) {
597 error = mac_vnode_check_read(active_cred, file_cred,
600 error = mac_vnode_check_write(active_cred, file_cred,
605 if (file_cred != NULL)
609 if (do_vn_io_fault(vp, &auio)) {
610 args.kind = VN_IO_FAULT_VOP;
613 args.args.vop_args.vp = vp;
614 error = vn_io_fault1(vp, &auio, &args, td);
615 } else if (rw == UIO_READ) {
616 error = VOP_READ(vp, &auio, ioflg, cred);
617 } else /* if (rw == UIO_WRITE) */ {
618 error = VOP_WRITE(vp, &auio, ioflg, cred);
622 *aresid = auio.uio_resid;
624 if (auio.uio_resid && error == 0)
626 if ((ioflg & IO_NODELOCKED) == 0) {
629 vn_finished_write(mp);
632 if (rl_cookie != NULL)
633 vn_rangelock_unlock(vp, rl_cookie);
638 * Package up an I/O request on a vnode into a uio and do it. The I/O
639 * request is split up into smaller chunks and we try to avoid saturating
640 * the buffer cache while potentially holding a vnode locked, so we
641 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
642 * to give other processes a chance to lock the vnode (either other processes
643 * core'ing the same binary, or unrelated processes scanning the directory).
646 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
647 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
648 struct ucred *file_cred, size_t *aresid, struct thread *td)
657 * Force `offset' to a multiple of MAXBSIZE except possibly
658 * for the first chunk, so that filesystems only need to
659 * write full blocks except possibly for the first and last
662 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
666 if (rw != UIO_READ && vp->v_type == VREG)
669 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
670 ioflg, active_cred, file_cred, &iaresid, td);
671 len -= chunk; /* aresid calc already includes length */
675 base = (char *)base + chunk;
676 kern_yield(PRI_USER);
679 *aresid = len + iaresid;
683 #if OFF_MAX <= LONG_MAX
685 foffset_lock(struct file *fp, int flags)
687 volatile short *flagsp;
691 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
693 if ((flags & FOF_NOLOCK) != 0)
694 return (atomic_load_long(&fp->f_offset));
697 * According to McKusick the vn lock was protecting f_offset here.
698 * It is now protected by the FOFFSET_LOCKED flag.
700 flagsp = &fp->f_vnread_flags;
701 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
702 return (atomic_load_long(&fp->f_offset));
704 sleepq_lock(&fp->f_vnread_flags);
705 state = atomic_load_16(flagsp);
707 if ((state & FOFFSET_LOCKED) == 0) {
708 if (!atomic_fcmpset_acq_16(flagsp, &state,
713 if ((state & FOFFSET_LOCK_WAITING) == 0) {
714 if (!atomic_fcmpset_acq_16(flagsp, &state,
715 state | FOFFSET_LOCK_WAITING))
719 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
720 sleepq_wait(&fp->f_vnread_flags, PUSER -1);
722 sleepq_lock(&fp->f_vnread_flags);
723 state = atomic_load_16(flagsp);
725 res = atomic_load_long(&fp->f_offset);
726 sleepq_release(&fp->f_vnread_flags);
731 foffset_unlock(struct file *fp, off_t val, int flags)
733 volatile short *flagsp;
736 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
738 if ((flags & FOF_NOUPDATE) == 0)
739 atomic_store_long(&fp->f_offset, val);
740 if ((flags & FOF_NEXTOFF_R) != 0)
741 fp->f_nextoff[UIO_READ] = val;
742 if ((flags & FOF_NEXTOFF_W) != 0)
743 fp->f_nextoff[UIO_WRITE] = val;
745 if ((flags & FOF_NOLOCK) != 0)
748 flagsp = &fp->f_vnread_flags;
749 state = atomic_load_16(flagsp);
750 if ((state & FOFFSET_LOCK_WAITING) == 0 &&
751 atomic_cmpset_rel_16(flagsp, state, 0))
754 sleepq_lock(&fp->f_vnread_flags);
755 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
756 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
757 fp->f_vnread_flags = 0;
758 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
759 sleepq_release(&fp->f_vnread_flags);
763 foffset_lock(struct file *fp, int flags)
768 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
770 mtxp = mtx_pool_find(mtxpool_sleep, fp);
772 if ((flags & FOF_NOLOCK) == 0) {
773 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
774 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
775 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
778 fp->f_vnread_flags |= FOFFSET_LOCKED;
786 foffset_unlock(struct file *fp, off_t val, int flags)
790 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
792 mtxp = mtx_pool_find(mtxpool_sleep, fp);
794 if ((flags & FOF_NOUPDATE) == 0)
796 if ((flags & FOF_NEXTOFF_R) != 0)
797 fp->f_nextoff[UIO_READ] = val;
798 if ((flags & FOF_NEXTOFF_W) != 0)
799 fp->f_nextoff[UIO_WRITE] = val;
800 if ((flags & FOF_NOLOCK) == 0) {
801 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
802 ("Lost FOFFSET_LOCKED"));
803 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
804 wakeup(&fp->f_vnread_flags);
805 fp->f_vnread_flags = 0;
812 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
815 if ((flags & FOF_OFFSET) == 0)
816 uio->uio_offset = foffset_lock(fp, flags);
820 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
823 if ((flags & FOF_OFFSET) == 0)
824 foffset_unlock(fp, uio->uio_offset, flags);
828 get_advice(struct file *fp, struct uio *uio)
833 ret = POSIX_FADV_NORMAL;
834 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
837 mtxp = mtx_pool_find(mtxpool_sleep, fp);
839 if (fp->f_advice != NULL &&
840 uio->uio_offset >= fp->f_advice->fa_start &&
841 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
842 ret = fp->f_advice->fa_advice;
848 * File table vnode read routine.
851 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
859 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
861 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
864 if (fp->f_flag & FNONBLOCK)
866 if (fp->f_flag & O_DIRECT)
868 advice = get_advice(fp, uio);
869 vn_lock(vp, LK_SHARED | LK_RETRY);
872 case POSIX_FADV_NORMAL:
873 case POSIX_FADV_SEQUENTIAL:
874 case POSIX_FADV_NOREUSE:
875 ioflag |= sequential_heuristic(uio, fp);
877 case POSIX_FADV_RANDOM:
878 /* Disable read-ahead for random I/O. */
881 orig_offset = uio->uio_offset;
884 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
887 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
888 fp->f_nextoff[UIO_READ] = uio->uio_offset;
890 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
891 orig_offset != uio->uio_offset)
893 * Use POSIX_FADV_DONTNEED to flush pages and buffers
894 * for the backing file after a POSIX_FADV_NOREUSE
897 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
898 POSIX_FADV_DONTNEED);
903 * File table vnode write routine.
906 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
912 int error, ioflag, lock_flags;
915 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
917 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
919 if (vp->v_type == VREG)
922 if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
924 if (fp->f_flag & FNONBLOCK)
926 if (fp->f_flag & O_DIRECT)
928 if ((fp->f_flag & O_FSYNC) ||
929 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
932 if (vp->v_type != VCHR &&
933 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
936 advice = get_advice(fp, uio);
938 if (MNT_SHARED_WRITES(mp) ||
939 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
940 lock_flags = LK_SHARED;
942 lock_flags = LK_EXCLUSIVE;
945 vn_lock(vp, lock_flags | LK_RETRY);
947 case POSIX_FADV_NORMAL:
948 case POSIX_FADV_SEQUENTIAL:
949 case POSIX_FADV_NOREUSE:
950 ioflag |= sequential_heuristic(uio, fp);
952 case POSIX_FADV_RANDOM:
953 /* XXX: Is this correct? */
956 orig_offset = uio->uio_offset;
959 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
962 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
963 fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
965 if (vp->v_type != VCHR)
966 vn_finished_write(mp);
967 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
968 orig_offset != uio->uio_offset)
970 * Use POSIX_FADV_DONTNEED to flush pages and buffers
971 * for the backing file after a POSIX_FADV_NOREUSE
974 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
975 POSIX_FADV_DONTNEED);
981 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
982 * prevent the following deadlock:
984 * Assume that the thread A reads from the vnode vp1 into userspace
985 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
986 * currently not resident, then system ends up with the call chain
987 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
988 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
989 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
990 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
991 * backed by the pages of vnode vp1, and some page in buf2 is not
992 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
994 * To prevent the lock order reversal and deadlock, vn_io_fault() does
995 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
996 * Instead, it first tries to do the whole range i/o with pagefaults
997 * disabled. If all pages in the i/o buffer are resident and mapped,
998 * VOP will succeed (ignoring the genuine filesystem errors).
999 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1000 * i/o in chunks, with all pages in the chunk prefaulted and held
1001 * using vm_fault_quick_hold_pages().
1003 * Filesystems using this deadlock avoidance scheme should use the
1004 * array of the held pages from uio, saved in the curthread->td_ma,
1005 * instead of doing uiomove(). A helper function
1006 * vn_io_fault_uiomove() converts uiomove request into
1007 * uiomove_fromphys() over td_ma array.
1009 * Since vnode locks do not cover the whole i/o anymore, rangelocks
1010 * make the current i/o request atomic with respect to other i/os and
1015 * Decode vn_io_fault_args and perform the corresponding i/o.
1018 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1024 save = vm_fault_disable_pagefaults();
1025 switch (args->kind) {
1026 case VN_IO_FAULT_FOP:
1027 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1028 uio, args->cred, args->flags, td);
1030 case VN_IO_FAULT_VOP:
1031 if (uio->uio_rw == UIO_READ) {
1032 error = VOP_READ(args->args.vop_args.vp, uio,
1033 args->flags, args->cred);
1034 } else if (uio->uio_rw == UIO_WRITE) {
1035 error = VOP_WRITE(args->args.vop_args.vp, uio,
1036 args->flags, args->cred);
1040 panic("vn_io_fault_doio: unknown kind of io %d %d",
1041 args->kind, uio->uio_rw);
1043 vm_fault_enable_pagefaults(save);
1048 vn_io_fault_touch(char *base, const struct uio *uio)
1053 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1059 vn_io_fault_prefault_user(const struct uio *uio)
1062 const struct iovec *iov;
1067 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1068 ("vn_io_fault_prefault userspace"));
1072 resid = uio->uio_resid;
1073 base = iov->iov_base;
1076 error = vn_io_fault_touch(base, uio);
1079 if (len < PAGE_SIZE) {
1081 error = vn_io_fault_touch(base + len - 1, uio);
1086 if (++i >= uio->uio_iovcnt)
1088 iov = uio->uio_iov + i;
1089 base = iov->iov_base;
1101 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1102 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1103 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1104 * into args and call vn_io_fault1() to handle faults during the user
1105 * mode buffer accesses.
1108 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1111 vm_page_t ma[io_hold_cnt + 2];
1112 struct uio *uio_clone, short_uio;
1113 struct iovec short_iovec[1];
1114 vm_page_t *prev_td_ma;
1116 vm_offset_t addr, end;
1119 int error, cnt, saveheld, prev_td_ma_cnt;
1121 if (vn_io_fault_prefault) {
1122 error = vn_io_fault_prefault_user(uio);
1124 return (error); /* Or ignore ? */
1127 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1130 * The UFS follows IO_UNIT directive and replays back both
1131 * uio_offset and uio_resid if an error is encountered during the
1132 * operation. But, since the iovec may be already advanced,
1133 * uio is still in an inconsistent state.
1135 * Cache a copy of the original uio, which is advanced to the redo
1136 * point using UIO_NOCOPY below.
1138 uio_clone = cloneuio(uio);
1139 resid = uio->uio_resid;
1141 short_uio.uio_segflg = UIO_USERSPACE;
1142 short_uio.uio_rw = uio->uio_rw;
1143 short_uio.uio_td = uio->uio_td;
1145 error = vn_io_fault_doio(args, uio, td);
1146 if (error != EFAULT)
1149 atomic_add_long(&vn_io_faults_cnt, 1);
1150 uio_clone->uio_segflg = UIO_NOCOPY;
1151 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1152 uio_clone->uio_segflg = uio->uio_segflg;
1154 saveheld = curthread_pflags_set(TDP_UIOHELD);
1155 prev_td_ma = td->td_ma;
1156 prev_td_ma_cnt = td->td_ma_cnt;
1158 while (uio_clone->uio_resid != 0) {
1159 len = uio_clone->uio_iov->iov_len;
1161 KASSERT(uio_clone->uio_iovcnt >= 1,
1162 ("iovcnt underflow"));
1163 uio_clone->uio_iov++;
1164 uio_clone->uio_iovcnt--;
1167 if (len > io_hold_cnt * PAGE_SIZE)
1168 len = io_hold_cnt * PAGE_SIZE;
1169 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1170 end = round_page(addr + len);
1175 cnt = atop(end - trunc_page(addr));
1177 * A perfectly misaligned address and length could cause
1178 * both the start and the end of the chunk to use partial
1179 * page. +2 accounts for such a situation.
1181 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1182 addr, len, prot, ma, io_hold_cnt + 2);
1187 short_uio.uio_iov = &short_iovec[0];
1188 short_iovec[0].iov_base = (void *)addr;
1189 short_uio.uio_iovcnt = 1;
1190 short_uio.uio_resid = short_iovec[0].iov_len = len;
1191 short_uio.uio_offset = uio_clone->uio_offset;
1193 td->td_ma_cnt = cnt;
1195 error = vn_io_fault_doio(args, &short_uio, td);
1196 vm_page_unhold_pages(ma, cnt);
1197 adv = len - short_uio.uio_resid;
1199 uio_clone->uio_iov->iov_base =
1200 (char *)uio_clone->uio_iov->iov_base + adv;
1201 uio_clone->uio_iov->iov_len -= adv;
1202 uio_clone->uio_resid -= adv;
1203 uio_clone->uio_offset += adv;
1205 uio->uio_resid -= adv;
1206 uio->uio_offset += adv;
1208 if (error != 0 || adv == 0)
1211 td->td_ma = prev_td_ma;
1212 td->td_ma_cnt = prev_td_ma_cnt;
1213 curthread_pflags_restore(saveheld);
1215 free(uio_clone, M_IOV);
1220 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1221 int flags, struct thread *td)
1226 struct vn_io_fault_args args;
1229 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1233 * The ability to read(2) on a directory has historically been
1234 * allowed for all users, but this can and has been the source of
1235 * at least one security issue in the past. As such, it is now hidden
1236 * away behind a sysctl for those that actually need it to use it, and
1237 * restricted to root when it's turned on to make it relatively safe to
1238 * leave on for longer sessions of need.
1240 if (vp->v_type == VDIR) {
1241 KASSERT(uio->uio_rw == UIO_READ,
1242 ("illegal write attempted on a directory"));
1243 if (!vfs_allow_read_dir)
1245 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1249 foffset_lock_uio(fp, uio, flags);
1250 if (do_vn_io_fault(vp, uio)) {
1251 args.kind = VN_IO_FAULT_FOP;
1252 args.args.fop_args.fp = fp;
1253 args.args.fop_args.doio = doio;
1254 args.cred = active_cred;
1255 args.flags = flags | FOF_OFFSET;
1256 if (uio->uio_rw == UIO_READ) {
1257 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1258 uio->uio_offset + uio->uio_resid);
1259 } else if ((fp->f_flag & O_APPEND) != 0 ||
1260 (flags & FOF_OFFSET) == 0) {
1261 /* For appenders, punt and lock the whole range. */
1262 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1264 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1265 uio->uio_offset + uio->uio_resid);
1267 error = vn_io_fault1(vp, uio, &args, td);
1268 vn_rangelock_unlock(vp, rl_cookie);
1270 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1272 foffset_unlock_uio(fp, uio, flags);
1277 * Helper function to perform the requested uiomove operation using
1278 * the held pages for io->uio_iov[0].iov_base buffer instead of
1279 * copyin/copyout. Access to the pages with uiomove_fromphys()
1280 * instead of iov_base prevents page faults that could occur due to
1281 * pmap_collect() invalidating the mapping created by
1282 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1283 * object cleanup revoking the write access from page mappings.
1285 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1286 * instead of plain uiomove().
1289 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1291 struct uio transp_uio;
1292 struct iovec transp_iov[1];
1298 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1299 uio->uio_segflg != UIO_USERSPACE)
1300 return (uiomove(data, xfersize, uio));
1302 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1303 transp_iov[0].iov_base = data;
1304 transp_uio.uio_iov = &transp_iov[0];
1305 transp_uio.uio_iovcnt = 1;
1306 if (xfersize > uio->uio_resid)
1307 xfersize = uio->uio_resid;
1308 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1309 transp_uio.uio_offset = 0;
1310 transp_uio.uio_segflg = UIO_SYSSPACE;
1312 * Since transp_iov points to data, and td_ma page array
1313 * corresponds to original uio->uio_iov, we need to invert the
1314 * direction of the i/o operation as passed to
1315 * uiomove_fromphys().
1317 switch (uio->uio_rw) {
1319 transp_uio.uio_rw = UIO_READ;
1322 transp_uio.uio_rw = UIO_WRITE;
1325 transp_uio.uio_td = uio->uio_td;
1326 error = uiomove_fromphys(td->td_ma,
1327 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1328 xfersize, &transp_uio);
1329 adv = xfersize - transp_uio.uio_resid;
1331 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1332 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1334 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1336 td->td_ma_cnt -= pgadv;
1337 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1338 uio->uio_iov->iov_len -= adv;
1339 uio->uio_resid -= adv;
1340 uio->uio_offset += adv;
1345 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1349 vm_offset_t iov_base;
1353 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1354 uio->uio_segflg != UIO_USERSPACE)
1355 return (uiomove_fromphys(ma, offset, xfersize, uio));
1357 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1358 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1359 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1360 switch (uio->uio_rw) {
1362 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1366 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1370 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1372 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1374 td->td_ma_cnt -= pgadv;
1375 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1376 uio->uio_iov->iov_len -= cnt;
1377 uio->uio_resid -= cnt;
1378 uio->uio_offset += cnt;
1383 * File table truncate routine.
1386 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1397 * Lock the whole range for truncation. Otherwise split i/o
1398 * might happen partly before and partly after the truncation.
1400 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1401 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1404 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1405 AUDIT_ARG_VNODE1(vp);
1406 if (vp->v_type == VDIR) {
1411 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1415 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1419 vn_finished_write(mp);
1421 vn_rangelock_unlock(vp, rl_cookie);
1426 * Truncate a file that is already locked.
1429 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1435 error = VOP_ADD_WRITECOUNT(vp, 1);
1438 vattr.va_size = length;
1440 vattr.va_vaflags |= VA_SYNC;
1441 error = VOP_SETATTR(vp, &vattr, cred);
1442 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1448 * File table vnode stat routine.
1451 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1454 struct vnode *vp = fp->f_vnode;
1457 vn_lock(vp, LK_SHARED | LK_RETRY);
1458 error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1465 * Stat a vnode; implementation for the stat syscall
1468 vn_stat(struct vnode *vp, struct stat *sb, struct ucred *active_cred,
1469 struct ucred *file_cred, struct thread *td)
1476 AUDIT_ARG_VNODE1(vp);
1478 error = mac_vnode_check_stat(active_cred, file_cred, vp);
1486 * Initialize defaults for new and unusual fields, so that file
1487 * systems which don't support these fields don't need to know
1490 vap->va_birthtime.tv_sec = -1;
1491 vap->va_birthtime.tv_nsec = 0;
1492 vap->va_fsid = VNOVAL;
1493 vap->va_rdev = NODEV;
1495 error = VOP_GETATTR(vp, vap, active_cred);
1500 * Zero the spare stat fields
1502 bzero(sb, sizeof *sb);
1505 * Copy from vattr table
1507 if (vap->va_fsid != VNOVAL)
1508 sb->st_dev = vap->va_fsid;
1510 sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1511 sb->st_ino = vap->va_fileid;
1512 mode = vap->va_mode;
1513 switch (vap->va_type) {
1539 sb->st_nlink = vap->va_nlink;
1540 sb->st_uid = vap->va_uid;
1541 sb->st_gid = vap->va_gid;
1542 sb->st_rdev = vap->va_rdev;
1543 if (vap->va_size > OFF_MAX)
1545 sb->st_size = vap->va_size;
1546 sb->st_atim.tv_sec = vap->va_atime.tv_sec;
1547 sb->st_atim.tv_nsec = vap->va_atime.tv_nsec;
1548 sb->st_mtim.tv_sec = vap->va_mtime.tv_sec;
1549 sb->st_mtim.tv_nsec = vap->va_mtime.tv_nsec;
1550 sb->st_ctim.tv_sec = vap->va_ctime.tv_sec;
1551 sb->st_ctim.tv_nsec = vap->va_ctime.tv_nsec;
1552 sb->st_birthtim.tv_sec = vap->va_birthtime.tv_sec;
1553 sb->st_birthtim.tv_nsec = vap->va_birthtime.tv_nsec;
1556 * According to www.opengroup.org, the meaning of st_blksize is
1557 * "a filesystem-specific preferred I/O block size for this
1558 * object. In some filesystem types, this may vary from file
1560 * Use minimum/default of PAGE_SIZE (e.g. for VCHR).
1563 sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1565 sb->st_flags = vap->va_flags;
1566 if (priv_check_cred_vfs_generation(td->td_ucred))
1569 sb->st_gen = vap->va_gen;
1571 sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1576 * File table vnode ioctl routine.
1579 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1584 struct fiobmap2_arg *bmarg;
1588 switch (vp->v_type) {
1593 vn_lock(vp, LK_SHARED | LK_RETRY);
1594 error = VOP_GETATTR(vp, &vattr, active_cred);
1597 *(int *)data = vattr.va_size - fp->f_offset;
1600 bmarg = (struct fiobmap2_arg *)data;
1601 vn_lock(vp, LK_SHARED | LK_RETRY);
1603 error = mac_vnode_check_read(active_cred, fp->f_cred,
1607 error = VOP_BMAP(vp, bmarg->bn, NULL,
1608 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1615 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1620 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1628 * File table vnode poll routine.
1631 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1638 #if defined(MAC) || defined(AUDIT)
1639 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1640 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1641 AUDIT_ARG_VNODE1(vp);
1642 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1648 error = VOP_POLL(vp, events, fp->f_cred, td);
1653 * Acquire the requested lock and then check for validity. LK_RETRY
1654 * permits vn_lock to return doomed vnodes.
1656 static int __noinline
1657 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1661 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1662 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1665 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1667 if ((flags & LK_RETRY) == 0) {
1678 * Nothing to do if we got the lock.
1684 * Interlock was dropped by the call in _vn_lock.
1686 flags &= ~LK_INTERLOCK;
1688 error = VOP_LOCK1(vp, flags, file, line);
1689 } while (error != 0);
1694 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1698 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1699 ("vn_lock: no locktype (%d passed)", flags));
1700 VNPASS(vp->v_holdcnt > 0, vp);
1701 error = VOP_LOCK1(vp, flags, file, line);
1702 if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1703 return (_vn_lock_fallback(vp, flags, file, line, error));
1708 * File table vnode close routine.
1711 vn_closefile(struct file *fp, struct thread *td)
1719 fp->f_ops = &badfileops;
1720 ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1722 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1724 if (__predict_false(ref)) {
1725 lf.l_whence = SEEK_SET;
1728 lf.l_type = F_UNLCK;
1729 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1736 * Preparing to start a filesystem write operation. If the operation is
1737 * permitted, then we bump the count of operations in progress and
1738 * proceed. If a suspend request is in progress, we wait until the
1739 * suspension is over, and then proceed.
1742 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1746 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1747 vfs_op_thread_enter(mp)) {
1748 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1749 vfs_mp_count_add_pcpu(mp, writeopcount, 1);
1750 vfs_op_thread_exit(mp);
1755 mtx_assert(MNT_MTX(mp), MA_OWNED);
1762 * Check on status of suspension.
1764 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1765 mp->mnt_susp_owner != curthread) {
1766 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1767 (flags & PCATCH) : 0) | (PUSER - 1);
1768 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1769 if (flags & V_NOWAIT) {
1770 error = EWOULDBLOCK;
1773 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1779 if (flags & V_XSLEEP)
1781 mp->mnt_writeopcount++;
1783 if (error != 0 || (flags & V_XSLEEP) != 0)
1790 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1795 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1796 ("V_MNTREF requires mp"));
1800 * If a vnode is provided, get and return the mount point that
1801 * to which it will write.
1804 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1806 if (error != EOPNOTSUPP)
1811 if ((mp = *mpp) == NULL)
1815 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1817 * As long as a vnode is not provided we need to acquire a
1818 * refcount for the provided mountpoint too, in order to
1819 * emulate a vfs_ref().
1821 if (vp == NULL && (flags & V_MNTREF) == 0)
1824 return (vn_start_write_refed(mp, flags, false));
1828 * Secondary suspension. Used by operations such as vop_inactive
1829 * routines that are needed by the higher level functions. These
1830 * are allowed to proceed until all the higher level functions have
1831 * completed (indicated by mnt_writeopcount dropping to zero). At that
1832 * time, these operations are halted until the suspension is over.
1835 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1840 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1841 ("V_MNTREF requires mp"));
1845 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1847 if (error != EOPNOTSUPP)
1853 * If we are not suspended or have not yet reached suspended
1854 * mode, then let the operation proceed.
1856 if ((mp = *mpp) == NULL)
1860 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1862 * As long as a vnode is not provided we need to acquire a
1863 * refcount for the provided mountpoint too, in order to
1864 * emulate a vfs_ref().
1867 if (vp == NULL && (flags & V_MNTREF) == 0)
1869 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1870 mp->mnt_secondary_writes++;
1871 mp->mnt_secondary_accwrites++;
1875 if (flags & V_NOWAIT) {
1878 return (EWOULDBLOCK);
1881 * Wait for the suspension to finish.
1883 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1884 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1893 * Filesystem write operation has completed. If we are suspending and this
1894 * operation is the last one, notify the suspender that the suspension is
1898 vn_finished_write(struct mount *mp)
1905 if (vfs_op_thread_enter(mp)) {
1906 vfs_mp_count_sub_pcpu(mp, writeopcount, 1);
1907 vfs_mp_count_sub_pcpu(mp, ref, 1);
1908 vfs_op_thread_exit(mp);
1913 vfs_assert_mount_counters(mp);
1915 c = --mp->mnt_writeopcount;
1916 if (mp->mnt_vfs_ops == 0) {
1917 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1922 vfs_dump_mount_counters(mp);
1923 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
1924 wakeup(&mp->mnt_writeopcount);
1929 * Filesystem secondary write operation has completed. If we are
1930 * suspending and this operation is the last one, notify the suspender
1931 * that the suspension is now in effect.
1934 vn_finished_secondary_write(struct mount *mp)
1940 mp->mnt_secondary_writes--;
1941 if (mp->mnt_secondary_writes < 0)
1942 panic("vn_finished_secondary_write: neg cnt");
1943 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1944 mp->mnt_secondary_writes <= 0)
1945 wakeup(&mp->mnt_secondary_writes);
1950 * Request a filesystem to suspend write operations.
1953 vfs_write_suspend(struct mount *mp, int flags)
1960 vfs_assert_mount_counters(mp);
1961 if (mp->mnt_susp_owner == curthread) {
1962 vfs_op_exit_locked(mp);
1966 while (mp->mnt_kern_flag & MNTK_SUSPEND)
1967 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1970 * Unmount holds a write reference on the mount point. If we
1971 * own busy reference and drain for writers, we deadlock with
1972 * the reference draining in the unmount path. Callers of
1973 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1974 * vfs_busy() reference is owned and caller is not in the
1977 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1978 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1979 vfs_op_exit_locked(mp);
1984 mp->mnt_kern_flag |= MNTK_SUSPEND;
1985 mp->mnt_susp_owner = curthread;
1986 if (mp->mnt_writeopcount > 0)
1987 (void) msleep(&mp->mnt_writeopcount,
1988 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1991 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
1992 vfs_write_resume(mp, 0);
1993 /* vfs_write_resume does vfs_op_exit() for us */
1999 * Request a filesystem to resume write operations.
2002 vfs_write_resume(struct mount *mp, int flags)
2006 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2007 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2008 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2010 mp->mnt_susp_owner = NULL;
2011 wakeup(&mp->mnt_writeopcount);
2012 wakeup(&mp->mnt_flag);
2013 curthread->td_pflags &= ~TDP_IGNSUSP;
2014 if ((flags & VR_START_WRITE) != 0) {
2016 mp->mnt_writeopcount++;
2019 if ((flags & VR_NO_SUSPCLR) == 0)
2022 } else if ((flags & VR_START_WRITE) != 0) {
2024 vn_start_write_refed(mp, 0, true);
2031 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2035 vfs_write_suspend_umnt(struct mount *mp)
2039 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2040 ("vfs_write_suspend_umnt: recursed"));
2042 /* dounmount() already called vn_start_write(). */
2044 vn_finished_write(mp);
2045 error = vfs_write_suspend(mp, 0);
2047 vn_start_write(NULL, &mp, V_WAIT);
2051 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2054 vn_start_write(NULL, &mp, V_WAIT);
2056 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2057 wakeup(&mp->mnt_flag);
2059 curthread->td_pflags |= TDP_IGNSUSP;
2064 * Implement kqueues for files by translating it to vnode operation.
2067 vn_kqfilter(struct file *fp, struct knote *kn)
2070 return (VOP_KQFILTER(fp->f_vnode, kn));
2074 * Simplified in-kernel wrapper calls for extended attribute access.
2075 * Both calls pass in a NULL credential, authorizing as "kernel" access.
2076 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2079 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2080 const char *attrname, int *buflen, char *buf, struct thread *td)
2086 iov.iov_len = *buflen;
2089 auio.uio_iov = &iov;
2090 auio.uio_iovcnt = 1;
2091 auio.uio_rw = UIO_READ;
2092 auio.uio_segflg = UIO_SYSSPACE;
2094 auio.uio_offset = 0;
2095 auio.uio_resid = *buflen;
2097 if ((ioflg & IO_NODELOCKED) == 0)
2098 vn_lock(vp, LK_SHARED | LK_RETRY);
2100 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2102 /* authorize attribute retrieval as kernel */
2103 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2106 if ((ioflg & IO_NODELOCKED) == 0)
2110 *buflen = *buflen - auio.uio_resid;
2117 * XXX failure mode if partially written?
2120 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2121 const char *attrname, int buflen, char *buf, struct thread *td)
2128 iov.iov_len = buflen;
2131 auio.uio_iov = &iov;
2132 auio.uio_iovcnt = 1;
2133 auio.uio_rw = UIO_WRITE;
2134 auio.uio_segflg = UIO_SYSSPACE;
2136 auio.uio_offset = 0;
2137 auio.uio_resid = buflen;
2139 if ((ioflg & IO_NODELOCKED) == 0) {
2140 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2142 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2145 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2147 /* authorize attribute setting as kernel */
2148 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2150 if ((ioflg & IO_NODELOCKED) == 0) {
2151 vn_finished_write(mp);
2159 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2160 const char *attrname, struct thread *td)
2165 if ((ioflg & IO_NODELOCKED) == 0) {
2166 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2168 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2171 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2173 /* authorize attribute removal as kernel */
2174 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2175 if (error == EOPNOTSUPP)
2176 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2179 if ((ioflg & IO_NODELOCKED) == 0) {
2180 vn_finished_write(mp);
2188 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2192 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2196 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2199 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2204 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2205 int lkflags, struct vnode **rvp)
2210 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2212 ltype = VOP_ISLOCKED(vp);
2213 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2214 ("vn_vget_ino: vp not locked"));
2215 error = vfs_busy(mp, MBF_NOWAIT);
2219 error = vfs_busy(mp, 0);
2220 vn_lock(vp, ltype | LK_RETRY);
2224 if (VN_IS_DOOMED(vp)) {
2230 error = alloc(mp, alloc_arg, lkflags, rvp);
2232 if (error != 0 || *rvp != vp)
2233 vn_lock(vp, ltype | LK_RETRY);
2234 if (VN_IS_DOOMED(vp)) {
2247 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2251 if (vp->v_type != VREG || td == NULL)
2253 if ((uoff_t)uio->uio_offset + uio->uio_resid >
2254 lim_cur(td, RLIMIT_FSIZE)) {
2255 PROC_LOCK(td->td_proc);
2256 kern_psignal(td->td_proc, SIGXFSZ);
2257 PROC_UNLOCK(td->td_proc);
2264 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2271 vn_lock(vp, LK_SHARED | LK_RETRY);
2272 AUDIT_ARG_VNODE1(vp);
2275 return (setfmode(td, active_cred, vp, mode));
2279 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2286 vn_lock(vp, LK_SHARED | LK_RETRY);
2287 AUDIT_ARG_VNODE1(vp);
2290 return (setfown(td, active_cred, vp, uid, gid));
2294 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2298 if ((object = vp->v_object) == NULL)
2300 VM_OBJECT_WLOCK(object);
2301 vm_object_page_remove(object, start, end, 0);
2302 VM_OBJECT_WUNLOCK(object);
2306 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2314 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2315 ("Wrong command %lu", cmd));
2317 if (vn_lock(vp, LK_SHARED) != 0)
2319 if (vp->v_type != VREG) {
2323 error = VOP_GETATTR(vp, &va, cred);
2327 if (noff >= va.va_size) {
2331 bsize = vp->v_mount->mnt_stat.f_iosize;
2332 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2334 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2335 if (error == EOPNOTSUPP) {
2339 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2340 (bnp != -1 && cmd == FIOSEEKDATA)) {
2347 if (noff > va.va_size)
2349 /* noff == va.va_size. There is an implicit hole at the end of file. */
2350 if (cmd == FIOSEEKDATA)
2360 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2365 off_t foffset, size;
2368 cred = td->td_ucred;
2370 foffset = foffset_lock(fp, 0);
2371 noneg = (vp->v_type != VCHR);
2377 (offset > 0 && foffset > OFF_MAX - offset))) {
2384 vn_lock(vp, LK_SHARED | LK_RETRY);
2385 error = VOP_GETATTR(vp, &vattr, cred);
2391 * If the file references a disk device, then fetch
2392 * the media size and use that to determine the ending
2395 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2396 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2397 vattr.va_size = size;
2399 (vattr.va_size > OFF_MAX ||
2400 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2404 offset += vattr.va_size;
2409 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2410 if (error == ENOTTY)
2414 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2415 if (error == ENOTTY)
2421 if (error == 0 && noneg && offset < 0)
2425 VFS_KNOTE_UNLOCKED(vp, 0);
2426 td->td_uretoff.tdu_off = offset;
2428 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2433 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2439 * Grant permission if the caller is the owner of the file, or
2440 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2441 * on the file. If the time pointer is null, then write
2442 * permission on the file is also sufficient.
2444 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2445 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2446 * will be allowed to set the times [..] to the current
2449 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2450 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2451 error = VOP_ACCESS(vp, VWRITE, cred, td);
2456 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2461 if (fp->f_type == DTYPE_FIFO)
2462 kif->kf_type = KF_TYPE_FIFO;
2464 kif->kf_type = KF_TYPE_VNODE;
2467 FILEDESC_SUNLOCK(fdp);
2468 error = vn_fill_kinfo_vnode(vp, kif);
2470 FILEDESC_SLOCK(fdp);
2475 vn_fill_junk(struct kinfo_file *kif)
2480 * Simulate vn_fullpath returning changing values for a given
2481 * vp during e.g. coredump.
2483 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2484 olen = strlen(kif->kf_path);
2486 strcpy(&kif->kf_path[len - 1], "$");
2488 for (; olen < len; olen++)
2489 strcpy(&kif->kf_path[olen], "A");
2493 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2496 char *fullpath, *freepath;
2499 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2502 error = vn_fullpath(curthread, vp, &fullpath, &freepath);
2504 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2506 if (freepath != NULL)
2507 free(freepath, M_TEMP);
2509 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2514 * Retrieve vnode attributes.
2516 va.va_fsid = VNOVAL;
2518 vn_lock(vp, LK_SHARED | LK_RETRY);
2519 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2523 if (va.va_fsid != VNOVAL)
2524 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2526 kif->kf_un.kf_file.kf_file_fsid =
2527 vp->v_mount->mnt_stat.f_fsid.val[0];
2528 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2529 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2530 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2531 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2532 kif->kf_un.kf_file.kf_file_size = va.va_size;
2533 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2534 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2535 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2540 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2541 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2545 struct pmckern_map_in pkm;
2551 boolean_t writecounted;
2554 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2555 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2557 * POSIX shared-memory objects are defined to have
2558 * kernel persistence, and are not defined to support
2559 * read(2)/write(2) -- or even open(2). Thus, we can
2560 * use MAP_ASYNC to trade on-disk coherence for speed.
2561 * The shm_open(3) library routine turns on the FPOSIXSHM
2562 * flag to request this behavior.
2564 if ((fp->f_flag & FPOSIXSHM) != 0)
2565 flags |= MAP_NOSYNC;
2570 * Ensure that file and memory protections are
2571 * compatible. Note that we only worry about
2572 * writability if mapping is shared; in this case,
2573 * current and max prot are dictated by the open file.
2574 * XXX use the vnode instead? Problem is: what
2575 * credentials do we use for determination? What if
2576 * proc does a setuid?
2579 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2580 maxprot = VM_PROT_NONE;
2581 if ((prot & VM_PROT_EXECUTE) != 0)
2584 maxprot = VM_PROT_EXECUTE;
2585 if ((fp->f_flag & FREAD) != 0)
2586 maxprot |= VM_PROT_READ;
2587 else if ((prot & VM_PROT_READ) != 0)
2591 * If we are sharing potential changes via MAP_SHARED and we
2592 * are trying to get write permission although we opened it
2593 * without asking for it, bail out.
2595 if ((flags & MAP_SHARED) != 0) {
2596 if ((fp->f_flag & FWRITE) != 0)
2597 maxprot |= VM_PROT_WRITE;
2598 else if ((prot & VM_PROT_WRITE) != 0)
2601 maxprot |= VM_PROT_WRITE;
2602 cap_maxprot |= VM_PROT_WRITE;
2604 maxprot &= cap_maxprot;
2607 * For regular files and shared memory, POSIX requires that
2608 * the value of foff be a legitimate offset within the data
2609 * object. In particular, negative offsets are invalid.
2610 * Blocking negative offsets and overflows here avoids
2611 * possible wraparound or user-level access into reserved
2612 * ranges of the data object later. In contrast, POSIX does
2613 * not dictate how offsets are used by device drivers, so in
2614 * the case of a device mapping a negative offset is passed
2621 foff < 0 || foff > OFF_MAX - size)
2624 writecounted = FALSE;
2625 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2626 &foff, &object, &writecounted);
2629 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2630 foff, writecounted, td);
2633 * If this mapping was accounted for in the vnode's
2634 * writecount, then undo that now.
2637 vm_pager_release_writecount(object, 0, size);
2638 vm_object_deallocate(object);
2641 /* Inform hwpmc(4) if an executable is being mapped. */
2642 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2643 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2645 pkm.pm_address = (uintptr_t) *addr;
2646 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2654 vn_fsid(struct vnode *vp, struct vattr *va)
2658 f = &vp->v_mount->mnt_stat.f_fsid;
2659 va->va_fsid = (uint32_t)f->val[1];
2660 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2661 va->va_fsid += (uint32_t)f->val[0];
2665 vn_fsync_buf(struct vnode *vp, int waitfor)
2667 struct buf *bp, *nbp;
2670 int error, maxretry;
2673 maxretry = 10000; /* large, arbitrarily chosen */
2675 if (vp->v_type == VCHR) {
2677 mp = vp->v_rdev->si_mountpt;
2684 * MARK/SCAN initialization to avoid infinite loops.
2686 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2687 bp->b_vflags &= ~BV_SCANNED;
2692 * Flush all dirty buffers associated with a vnode.
2695 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2696 if ((bp->b_vflags & BV_SCANNED) != 0)
2698 bp->b_vflags |= BV_SCANNED;
2699 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2700 if (waitfor != MNT_WAIT)
2703 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2704 BO_LOCKPTR(bo)) != 0) {
2711 KASSERT(bp->b_bufobj == bo,
2712 ("bp %p wrong b_bufobj %p should be %p",
2713 bp, bp->b_bufobj, bo));
2714 if ((bp->b_flags & B_DELWRI) == 0)
2715 panic("fsync: not dirty");
2716 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2722 if (maxretry < 1000)
2723 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2729 * If synchronous the caller expects us to completely resolve all
2730 * dirty buffers in the system. Wait for in-progress I/O to
2731 * complete (which could include background bitmap writes), then
2732 * retry if dirty blocks still exist.
2734 if (waitfor == MNT_WAIT) {
2735 bufobj_wwait(bo, 0, 0);
2736 if (bo->bo_dirty.bv_cnt > 0) {
2738 * If we are unable to write any of these buffers
2739 * then we fail now rather than trying endlessly
2740 * to write them out.
2742 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2743 if ((error = bp->b_error) != 0)
2745 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2746 (error == 0 && --maxretry >= 0))
2754 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2760 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
2761 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2762 * to do the actual copy.
2763 * vn_generic_copy_file_range() is factored out, so it can be called
2764 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2765 * different file systems.
2768 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2769 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2770 struct ucred *outcred, struct thread *fsize_td)
2774 uint64_t uvalin, uvalout;
2777 *lenp = 0; /* For error returns. */
2780 /* Do some sanity checks on the arguments. */
2785 if (invp->v_type == VDIR || outvp->v_type == VDIR)
2787 else if (*inoffp < 0 || uvalin > INT64_MAX || uvalin <
2788 (uint64_t)*inoffp || *outoffp < 0 || uvalout > INT64_MAX ||
2789 uvalout < (uint64_t)*outoffp || invp->v_type != VREG ||
2790 outvp->v_type != VREG)
2796 * If the two vnode are for the same file system, call
2797 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2798 * which can handle copies across multiple file systems.
2801 if (invp->v_mount == outvp->v_mount)
2802 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2803 lenp, flags, incred, outcred, fsize_td);
2805 error = vn_generic_copy_file_range(invp, inoffp, outvp,
2806 outoffp, lenp, flags, incred, outcred, fsize_td);
2812 * Test len bytes of data starting at dat for all bytes == 0.
2813 * Return true if all bytes are zero, false otherwise.
2814 * Expects dat to be well aligned.
2817 mem_iszero(void *dat, int len)
2823 for (p = dat; len > 0; len -= sizeof(*p), p++) {
2824 if (len >= sizeof(*p)) {
2828 cp = (const char *)p;
2829 for (i = 0; i < len; i++, cp++)
2838 * Look for a hole in the output file and, if found, adjust *outoffp
2839 * and *xferp to skip past the hole.
2840 * *xferp is the entire hole length to be written and xfer2 is how many bytes
2841 * to be written as 0's upon return.
2844 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2845 off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2850 if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2851 *dataoffp = *outoffp;
2852 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2855 *holeoffp = *dataoffp;
2856 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2859 if (error != 0 || *holeoffp == *dataoffp) {
2861 * Since outvp is unlocked, it may be possible for
2862 * another thread to do a truncate(), lseek(), write()
2863 * creating a hole at startoff between the above
2864 * VOP_IOCTL() calls, if the other thread does not do
2866 * If that happens, *holeoffp == *dataoffp and finding
2867 * the hole has failed, so disable vn_skip_hole().
2869 *holeoffp = -1; /* Disable use of vn_skip_hole(). */
2872 KASSERT(*dataoffp >= *outoffp,
2873 ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2874 (intmax_t)*dataoffp, (intmax_t)*outoffp));
2875 KASSERT(*holeoffp > *dataoffp,
2876 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2877 (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2881 * If there is a hole before the data starts, advance *outoffp and
2882 * *xferp past the hole.
2884 if (*dataoffp > *outoffp) {
2885 delta = *dataoffp - *outoffp;
2886 if (delta >= *xferp) {
2887 /* Entire *xferp is a hole. */
2894 xfer2 = MIN(xfer2, *xferp);
2898 * If a hole starts before the end of this xfer2, reduce this xfer2 so
2899 * that the write ends at the start of the hole.
2900 * *holeoffp should always be greater than *outoffp, but for the
2901 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
2904 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
2905 xfer2 = *holeoffp - *outoffp;
2910 * Write an xfer sized chunk to outvp in blksize blocks from dat.
2911 * dat is a maximum of blksize in length and can be written repeatedly in
2913 * If growfile == true, just grow the file via vn_truncate_locked() instead
2914 * of doing actual writes.
2915 * If checkhole == true, a hole is being punched, so skip over any hole
2916 * already in the output file.
2919 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
2920 u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
2923 off_t dataoff, holeoff, xfer2;
2927 * Loop around doing writes of blksize until write has been completed.
2928 * Lock/unlock on each loop iteration so that a bwillwrite() can be
2929 * done for each iteration, since the xfer argument can be very
2930 * large if there is a large hole to punch in the output file.
2935 xfer2 = MIN(xfer, blksize);
2938 * Punching a hole. Skip writing if there is
2939 * already a hole in the output file.
2941 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
2942 &dataoff, &holeoff, cred);
2947 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
2952 error = vn_start_write(outvp, &mp, V_WAIT);
2954 if (MNT_SHARED_WRITES(mp))
2957 lckf = LK_EXCLUSIVE;
2958 error = vn_lock(outvp, lckf);
2962 error = vn_truncate_locked(outvp, outoff + xfer,
2965 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
2966 outoff, UIO_SYSSPACE, IO_NODELOCKED,
2967 curthread->td_ucred, cred, NULL, curthread);
2974 vn_finished_write(mp);
2975 } while (!growfile && xfer > 0 && error == 0);
2980 * Copy a byte range of one file to another. This function can handle the
2981 * case where invp and outvp are on different file systems.
2982 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
2983 * is no better file system specific way to do it.
2986 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
2987 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
2988 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
2993 off_t startoff, endoff, xfer, xfer2;
2996 bool cantseek, readzeros, eof, lastblock;
2998 size_t copylen, len, savlen;
3000 long holein, holeout;
3002 holein = holeout = 0;
3003 savlen = len = *lenp;
3007 error = vn_lock(invp, LK_SHARED);
3010 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3015 error = vn_start_write(outvp, &mp, V_WAIT);
3017 error = vn_lock(outvp, LK_EXCLUSIVE);
3020 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3021 * now that outvp is locked.
3023 if (fsize_td != NULL) {
3024 io.uio_offset = *outoffp;
3026 error = vn_rlimit_fsize(outvp, &io, fsize_td);
3030 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3033 * Holes that are past EOF do not need to be written as a block
3034 * of zero bytes. So, truncate the output file as far as
3035 * possible and then use va.va_size to decide if writing 0
3036 * bytes is necessary in the loop below.
3039 error = VOP_GETATTR(outvp, &va, outcred);
3040 if (error == 0 && va.va_size > *outoffp && va.va_size <=
3043 error = mac_vnode_check_write(curthread->td_ucred,
3047 error = vn_truncate_locked(outvp, *outoffp,
3050 va.va_size = *outoffp;
3055 vn_finished_write(mp);
3060 * Set the blksize to the larger of the hole sizes for invp and outvp.
3061 * If hole sizes aren't available, set the blksize to the larger
3062 * f_iosize of invp and outvp.
3063 * This code expects the hole sizes and f_iosizes to be powers of 2.
3064 * This value is clipped at 4Kbytes and 1Mbyte.
3066 blksize = MAX(holein, holeout);
3068 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3069 outvp->v_mount->mnt_stat.f_iosize);
3072 else if (blksize > 1024 * 1024)
3073 blksize = 1024 * 1024;
3074 dat = malloc(blksize, M_TEMP, M_WAITOK);
3077 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3078 * to find holes. Otherwise, just scan the read block for all 0s
3079 * in the inner loop where the data copying is done.
3080 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3081 * support holes on the server, but do not support FIOSEEKHOLE.
3084 while (len > 0 && error == 0 && !eof) {
3085 endoff = 0; /* To shut up compilers. */
3091 * Find the next data area. If there is just a hole to EOF,
3092 * FIOSEEKDATA should fail and then we drop down into the
3093 * inner loop and create the hole on the outvp file.
3094 * (I do not know if any file system will report a hole to
3095 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3096 * will fail for those file systems.)
3098 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3099 * the code just falls through to the inner copy loop.
3103 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3107 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3110 * Since invp is unlocked, it may be possible for
3111 * another thread to do a truncate(), lseek(), write()
3112 * creating a hole at startoff between the above
3113 * VOP_IOCTL() calls, if the other thread does not do
3115 * If that happens, startoff == endoff and finding
3116 * the hole has failed, so set an error.
3118 if (error == 0 && startoff == endoff)
3119 error = EINVAL; /* Any error. Reset to 0. */
3122 if (startoff > *inoffp) {
3123 /* Found hole before data block. */
3124 xfer = MIN(startoff - *inoffp, len);
3125 if (*outoffp < va.va_size) {
3126 /* Must write 0s to punch hole. */
3127 xfer2 = MIN(va.va_size - *outoffp,
3129 memset(dat, 0, MIN(xfer2, blksize));
3130 error = vn_write_outvp(outvp, dat,
3131 *outoffp, xfer2, blksize, false,
3132 holeout > 0, outcred);
3135 if (error == 0 && *outoffp + xfer >
3136 va.va_size && xfer == len)
3137 /* Grow last block. */
3138 error = vn_write_outvp(outvp, dat,
3139 *outoffp, xfer, blksize, true,
3147 copylen = MIN(len, endoff - startoff);
3159 * Set first xfer to end at a block boundary, so that
3160 * holes are more likely detected in the loop below via
3161 * the for all bytes 0 method.
3163 xfer -= (*inoffp % blksize);
3165 /* Loop copying the data block. */
3166 while (copylen > 0 && error == 0 && !eof) {
3169 error = vn_lock(invp, LK_SHARED);
3172 error = vn_rdwr(UIO_READ, invp, dat, xfer,
3173 startoff, UIO_SYSSPACE, IO_NODELOCKED,
3174 curthread->td_ucred, incred, &aresid,
3178 if (error == 0 && aresid > 0) {
3179 /* Stop the copy at EOF on the input file. */
3186 * Skip the write for holes past the initial EOF
3187 * of the output file, unless this is the last
3188 * write of the output file at EOF.
3190 readzeros = cantseek ? mem_iszero(dat, xfer) :
3194 if (!cantseek || *outoffp < va.va_size ||
3195 lastblock || !readzeros)
3196 error = vn_write_outvp(outvp, dat,
3197 *outoffp, xfer, blksize,
3198 readzeros && lastblock &&
3199 *outoffp >= va.va_size, false,
3213 *lenp = savlen - len;
3219 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3223 off_t olen, ooffset;
3226 int audited_vnode1 = 0;
3230 if (vp->v_type != VREG)
3233 /* Allocating blocks may take a long time, so iterate. */
3240 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3243 error = vn_lock(vp, LK_EXCLUSIVE);
3245 vn_finished_write(mp);
3249 if (!audited_vnode1) {
3250 AUDIT_ARG_VNODE1(vp);
3255 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3258 error = VOP_ALLOCATE(vp, &offset, &len);
3260 vn_finished_write(mp);
3262 if (olen + ooffset != offset + len) {
3263 panic("offset + len changed from %jx/%jx to %jx/%jx",
3264 ooffset, olen, offset, len);
3266 if (error != 0 || len == 0)
3268 KASSERT(olen > len, ("Iteration did not make progress?"));