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 #include "opt_hwpmc_hooks.h"
48 #include <sys/param.h>
49 #include <sys/systm.h>
52 #include <sys/fcntl.h>
59 #include <sys/limits.h>
62 #include <sys/mount.h>
63 #include <sys/mutex.h>
64 #include <sys/namei.h>
65 #include <sys/vnode.h>
66 #include <sys/dirent.h>
69 #include <sys/filio.h>
70 #include <sys/resourcevar.h>
71 #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>
81 #include <sys/ktrace.h>
83 #include <security/audit/audit.h>
84 #include <security/mac/mac_framework.h>
87 #include <vm/vm_extern.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_object.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_pager.h>
95 #include <sys/pmckern.h>
98 static fo_rdwr_t vn_read;
99 static fo_rdwr_t vn_write;
100 static fo_rdwr_t vn_io_fault;
101 static fo_truncate_t vn_truncate;
102 static fo_ioctl_t vn_ioctl;
103 static fo_poll_t vn_poll;
104 static fo_kqfilter_t vn_kqfilter;
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 const u_int io_hold_cnt = 16;
129 static int vn_io_fault_enable = 1;
130 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
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_RWTUN,
134 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
135 static int vn_io_pgcache_read_enable = 1;
136 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
137 &vn_io_pgcache_read_enable, 0,
138 "Enable copying from page cache for reads, avoiding fs");
139 static u_long vn_io_faults_cnt;
140 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
141 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
143 static int vfs_allow_read_dir = 0;
144 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
145 &vfs_allow_read_dir, 0,
146 "Enable read(2) of directory by root for filesystems that support it");
149 * Returns true if vn_io_fault mode of handling the i/o request should
153 do_vn_io_fault(struct vnode *vp, struct uio *uio)
157 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
158 (mp = vp->v_mount) != NULL &&
159 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
163 * Structure used to pass arguments to vn_io_fault1(), to do either
164 * file- or vnode-based I/O calls.
166 struct vn_io_fault_args {
174 struct fop_args_tag {
178 struct vop_args_tag {
184 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
185 struct vn_io_fault_args *args, struct thread *td);
188 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
190 struct thread *td = ndp->ni_cnd.cn_thread;
192 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
196 open2nameif(int fmode, u_int vn_open_flags)
200 res = ISOPEN | LOCKLEAF;
201 if ((fmode & O_RESOLVE_BENEATH) != 0)
203 if ((fmode & O_EMPTY_PATH) != 0)
205 if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
207 if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
209 if ((vn_open_flags & VN_OPEN_WANTIOCTLCAPS) != 0)
210 res |= WANTIOCTLCAPS;
215 * Common code for vnode open operations via a name lookup.
216 * Lookup the vnode and invoke VOP_CREATE if needed.
217 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
219 * Note that this does NOT free nameidata for the successful case,
220 * due to the NDINIT being done elsewhere.
223 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
224 struct ucred *cred, struct file *fp)
228 struct thread *td = ndp->ni_cnd.cn_thread;
230 struct vattr *vap = &vat;
237 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
238 O_EXCL | O_DIRECTORY) ||
239 (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
241 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
242 ndp->ni_cnd.cn_nameiop = CREATE;
243 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
245 * Set NOCACHE to avoid flushing the cache when
246 * rolling in many files at once.
248 * Set NC_KEEPPOSENTRY to keep positive entries if they already
249 * exist despite NOCACHE.
251 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
252 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
253 ndp->ni_cnd.cn_flags |= FOLLOW;
254 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
256 if ((error = namei(ndp)) != 0)
258 if (ndp->ni_vp == NULL) {
261 vap->va_mode = cmode;
263 vap->va_vaflags |= VA_EXCLUSIVE;
264 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
265 NDFREE(ndp, NDF_ONLY_PNBUF);
267 if ((error = vn_start_write(NULL, &mp,
268 V_XSLEEP | PCATCH)) != 0)
273 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
274 ndp->ni_cnd.cn_flags |= MAKEENTRY;
276 error = mac_vnode_check_create(cred, ndp->ni_dvp,
280 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
283 if (error == 0 && (fmode & O_EXCL) != 0 &&
284 (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
286 vp->v_iflag |= VI_FOPENING;
290 VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
292 vn_finished_write(mp);
294 NDFREE(ndp, NDF_ONLY_PNBUF);
295 if (error == ERELOOKUP) {
303 if (ndp->ni_dvp == ndp->ni_vp)
309 if (fmode & O_EXCL) {
313 if (vp->v_type == VDIR) {
320 ndp->ni_cnd.cn_nameiop = LOOKUP;
321 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
322 ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
324 if ((fmode & FWRITE) == 0)
325 ndp->ni_cnd.cn_flags |= LOCKSHARED;
326 if ((error = namei(ndp)) != 0)
330 error = vn_open_vnode(vp, fmode, cred, td, fp);
333 vp->v_iflag &= ~VI_FOPENING;
342 NDFREE(ndp, NDF_ONLY_PNBUF);
350 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
353 int error, lock_flags, type;
355 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
356 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
358 KASSERT(fp != NULL, ("open with flock requires fp"));
359 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
362 lock_flags = VOP_ISLOCKED(vp);
365 lf.l_whence = SEEK_SET;
368 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
370 if ((fmode & FNONBLOCK) == 0)
372 if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
374 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
376 fp->f_flag |= FHASLOCK;
378 vn_lock(vp, lock_flags | LK_RETRY);
383 * Common code for vnode open operations once a vnode is located.
384 * Check permissions, and call the VOP_OPEN routine.
387 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
388 struct thread *td, struct file *fp)
393 if (vp->v_type == VLNK) {
394 if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
397 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
401 if ((fmode & O_PATH) == 0) {
402 if (vp->v_type == VSOCK)
404 if ((fmode & (FWRITE | O_TRUNC)) != 0) {
405 if (vp->v_type == VDIR)
409 if ((fmode & FREAD) != 0)
411 if ((fmode & O_APPEND) && (fmode & FWRITE))
414 if ((fmode & O_CREAT) != 0)
418 if ((fmode & FEXEC) != 0)
421 if ((fmode & O_VERIFY) != 0)
423 error = mac_vnode_check_open(cred, vp, accmode);
427 accmode &= ~(VCREAT | VVERIFY);
429 if ((fmode & O_CREAT) == 0 && accmode != 0) {
430 error = VOP_ACCESS(vp, accmode, cred, td);
434 if ((fmode & O_PATH) != 0) {
435 if (vp->v_type != VFIFO && vp->v_type != VSOCK &&
436 VOP_ACCESS(vp, VREAD, cred, td) == 0)
437 fp->f_flag |= FKQALLOWED;
441 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
442 vn_lock(vp, LK_UPGRADE | LK_RETRY);
443 error = VOP_OPEN(vp, fmode, cred, td, fp);
447 error = vn_open_vnode_advlock(vp, fmode, fp);
448 if (error == 0 && (fmode & FWRITE) != 0) {
449 error = VOP_ADD_WRITECOUNT(vp, 1);
451 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
452 __func__, vp, vp->v_writecount);
457 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
458 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
463 * Arrange the call by having fdrop() to use
464 * vn_closefile(). This is to satisfy
465 * filesystems like devfs or tmpfs, which
466 * override fo_close().
468 fp->f_flag |= FOPENFAILED;
470 if (fp->f_ops == &badfileops) {
471 fp->f_type = DTYPE_VNODE;
477 * If there is no fp, due to kernel-mode open,
478 * we can call VOP_CLOSE() now.
480 if ((vp->v_type == VFIFO || (fmode & FWRITE) != 0 ||
481 !MNT_EXTENDED_SHARED(vp->v_mount)) &&
482 VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
483 vn_lock(vp, LK_UPGRADE | LK_RETRY);
484 (void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
489 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
495 * Check for write permissions on the specified vnode.
496 * Prototype text segments cannot be written.
500 vn_writechk(struct vnode *vp)
503 ASSERT_VOP_LOCKED(vp, "vn_writechk");
505 * If there's shared text associated with
506 * the vnode, try to free it up once. If
507 * we fail, we can't allow writing.
519 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
520 struct thread *td, bool keep_ref)
523 int error, lock_flags;
525 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
526 MNT_EXTENDED_SHARED(vp->v_mount))
527 lock_flags = LK_SHARED;
529 lock_flags = LK_EXCLUSIVE;
531 vn_start_write(vp, &mp, V_WAIT);
532 vn_lock(vp, lock_flags | LK_RETRY);
533 AUDIT_ARG_VNODE1(vp);
534 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
535 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
536 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
537 __func__, vp, vp->v_writecount);
539 error = VOP_CLOSE(vp, flags, file_cred, td);
544 vn_finished_write(mp);
549 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
553 return (vn_close1(vp, flags, file_cred, td, false));
557 * Heuristic to detect sequential operation.
560 sequential_heuristic(struct uio *uio, struct file *fp)
564 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
567 if (fp->f_flag & FRDAHEAD)
568 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
571 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
572 * that the first I/O is normally considered to be slightly
573 * sequential. Seeking to offset 0 doesn't change sequentiality
574 * unless previous seeks have reduced f_seqcount to 0, in which
575 * case offset 0 is not special.
577 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
578 uio->uio_offset == fp->f_nextoff[rw]) {
580 * f_seqcount is in units of fixed-size blocks so that it
581 * depends mainly on the amount of sequential I/O and not
582 * much on the number of sequential I/O's. The fixed size
583 * of 16384 is hard-coded here since it is (not quite) just
584 * a magic size that works well here. This size is more
585 * closely related to the best I/O size for real disks than
586 * to any block size used by software.
588 if (uio->uio_resid >= IO_SEQMAX * 16384)
589 fp->f_seqcount[rw] = IO_SEQMAX;
591 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
592 if (fp->f_seqcount[rw] > IO_SEQMAX)
593 fp->f_seqcount[rw] = IO_SEQMAX;
595 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
598 /* Not sequential. Quickly draw-down sequentiality. */
599 if (fp->f_seqcount[rw] > 1)
600 fp->f_seqcount[rw] = 1;
602 fp->f_seqcount[rw] = 0;
607 * Package up an I/O request on a vnode into a uio and do it.
610 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
611 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
612 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
619 struct vn_io_fault_args args;
620 int error, lock_flags;
622 if (offset < 0 && vp->v_type != VCHR)
624 auio.uio_iov = &aiov;
626 aiov.iov_base = base;
628 auio.uio_resid = len;
629 auio.uio_offset = offset;
630 auio.uio_segflg = segflg;
635 if ((ioflg & IO_NODELOCKED) == 0) {
636 if ((ioflg & IO_RANGELOCKED) == 0) {
637 if (rw == UIO_READ) {
638 rl_cookie = vn_rangelock_rlock(vp, offset,
640 } else if ((ioflg & IO_APPEND) != 0) {
641 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
643 rl_cookie = vn_rangelock_wlock(vp, offset,
649 if (rw == UIO_WRITE) {
650 if (vp->v_type != VCHR &&
651 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
654 lock_flags = vn_lktype_write(mp, vp);
656 lock_flags = LK_SHARED;
657 vn_lock(vp, lock_flags | LK_RETRY);
661 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
663 if ((ioflg & IO_NOMACCHECK) == 0) {
665 error = mac_vnode_check_read(active_cred, file_cred,
668 error = mac_vnode_check_write(active_cred, file_cred,
673 if (file_cred != NULL)
677 if (do_vn_io_fault(vp, &auio)) {
678 args.kind = VN_IO_FAULT_VOP;
681 args.args.vop_args.vp = vp;
682 error = vn_io_fault1(vp, &auio, &args, td);
683 } else if (rw == UIO_READ) {
684 error = VOP_READ(vp, &auio, ioflg, cred);
685 } else /* if (rw == UIO_WRITE) */ {
686 error = VOP_WRITE(vp, &auio, ioflg, cred);
690 *aresid = auio.uio_resid;
692 if (auio.uio_resid && error == 0)
694 if ((ioflg & IO_NODELOCKED) == 0) {
697 vn_finished_write(mp);
700 if (rl_cookie != NULL)
701 vn_rangelock_unlock(vp, rl_cookie);
706 * Package up an I/O request on a vnode into a uio and do it. The I/O
707 * request is split up into smaller chunks and we try to avoid saturating
708 * the buffer cache while potentially holding a vnode locked, so we
709 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
710 * to give other processes a chance to lock the vnode (either other processes
711 * core'ing the same binary, or unrelated processes scanning the directory).
714 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
715 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
716 struct ucred *file_cred, size_t *aresid, struct thread *td)
725 * Force `offset' to a multiple of MAXBSIZE except possibly
726 * for the first chunk, so that filesystems only need to
727 * write full blocks except possibly for the first and last
730 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
734 if (rw != UIO_READ && vp->v_type == VREG)
737 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
738 ioflg, active_cred, file_cred, &iaresid, td);
739 len -= chunk; /* aresid calc already includes length */
743 base = (char *)base + chunk;
744 kern_yield(PRI_USER);
747 *aresid = len + iaresid;
751 #if OFF_MAX <= LONG_MAX
753 foffset_lock(struct file *fp, int flags)
755 volatile short *flagsp;
759 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
761 if ((flags & FOF_NOLOCK) != 0)
762 return (atomic_load_long(&fp->f_offset));
765 * According to McKusick the vn lock was protecting f_offset here.
766 * It is now protected by the FOFFSET_LOCKED flag.
768 flagsp = &fp->f_vnread_flags;
769 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
770 return (atomic_load_long(&fp->f_offset));
772 sleepq_lock(&fp->f_vnread_flags);
773 state = atomic_load_16(flagsp);
775 if ((state & FOFFSET_LOCKED) == 0) {
776 if (!atomic_fcmpset_acq_16(flagsp, &state,
781 if ((state & FOFFSET_LOCK_WAITING) == 0) {
782 if (!atomic_fcmpset_acq_16(flagsp, &state,
783 state | FOFFSET_LOCK_WAITING))
787 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
788 sleepq_wait(&fp->f_vnread_flags, PUSER -1);
790 sleepq_lock(&fp->f_vnread_flags);
791 state = atomic_load_16(flagsp);
793 res = atomic_load_long(&fp->f_offset);
794 sleepq_release(&fp->f_vnread_flags);
799 foffset_unlock(struct file *fp, off_t val, int flags)
801 volatile short *flagsp;
804 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
806 if ((flags & FOF_NOUPDATE) == 0)
807 atomic_store_long(&fp->f_offset, val);
808 if ((flags & FOF_NEXTOFF_R) != 0)
809 fp->f_nextoff[UIO_READ] = val;
810 if ((flags & FOF_NEXTOFF_W) != 0)
811 fp->f_nextoff[UIO_WRITE] = val;
813 if ((flags & FOF_NOLOCK) != 0)
816 flagsp = &fp->f_vnread_flags;
817 state = atomic_load_16(flagsp);
818 if ((state & FOFFSET_LOCK_WAITING) == 0 &&
819 atomic_cmpset_rel_16(flagsp, state, 0))
822 sleepq_lock(&fp->f_vnread_flags);
823 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
824 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
825 fp->f_vnread_flags = 0;
826 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
827 sleepq_release(&fp->f_vnread_flags);
831 foffset_lock(struct file *fp, int flags)
836 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
838 mtxp = mtx_pool_find(mtxpool_sleep, fp);
840 if ((flags & FOF_NOLOCK) == 0) {
841 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
842 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
843 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
846 fp->f_vnread_flags |= FOFFSET_LOCKED;
854 foffset_unlock(struct file *fp, off_t val, int flags)
858 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
860 mtxp = mtx_pool_find(mtxpool_sleep, fp);
862 if ((flags & FOF_NOUPDATE) == 0)
864 if ((flags & FOF_NEXTOFF_R) != 0)
865 fp->f_nextoff[UIO_READ] = val;
866 if ((flags & FOF_NEXTOFF_W) != 0)
867 fp->f_nextoff[UIO_WRITE] = val;
868 if ((flags & FOF_NOLOCK) == 0) {
869 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
870 ("Lost FOFFSET_LOCKED"));
871 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
872 wakeup(&fp->f_vnread_flags);
873 fp->f_vnread_flags = 0;
880 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
883 if ((flags & FOF_OFFSET) == 0)
884 uio->uio_offset = foffset_lock(fp, flags);
888 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
891 if ((flags & FOF_OFFSET) == 0)
892 foffset_unlock(fp, uio->uio_offset, flags);
896 get_advice(struct file *fp, struct uio *uio)
901 ret = POSIX_FADV_NORMAL;
902 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
905 mtxp = mtx_pool_find(mtxpool_sleep, fp);
907 if (fp->f_advice != NULL &&
908 uio->uio_offset >= fp->f_advice->fa_start &&
909 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
910 ret = fp->f_advice->fa_advice;
916 get_write_ioflag(struct file *fp)
924 mp = atomic_load_ptr(&vp->v_mount);
926 if ((fp->f_flag & O_DIRECT) != 0)
929 if ((fp->f_flag & O_FSYNC) != 0 ||
930 (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
934 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
935 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
936 * fall back to full O_SYNC behavior.
938 if ((fp->f_flag & O_DSYNC) != 0)
939 ioflag |= IO_SYNC | IO_DATASYNC;
945 vn_read_from_obj(struct vnode *vp, struct uio *uio)
948 vm_page_t ma[io_hold_cnt + 2];
953 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
954 obj = atomic_load_ptr(&vp->v_object);
956 return (EJUSTRETURN);
959 * Depends on type stability of vm_objects.
961 vm_object_pip_add(obj, 1);
962 if ((obj->flags & OBJ_DEAD) != 0) {
964 * Note that object might be already reused from the
965 * vnode, and the OBJ_DEAD flag cleared. This is fine,
966 * we recheck for DOOMED vnode state after all pages
967 * are busied, and retract then.
969 * But we check for OBJ_DEAD to ensure that we do not
970 * busy pages while vm_object_terminate_pages()
971 * processes the queue.
977 resid = uio->uio_resid;
978 off = uio->uio_offset;
979 for (i = 0; resid > 0; i++) {
980 MPASS(i < io_hold_cnt + 2);
981 ma[i] = vm_page_grab_unlocked(obj, atop(off),
982 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
988 * Skip invalid pages. Valid mask can be partial only
989 * at EOF, and we clip later.
991 if (vm_page_none_valid(ma[i])) {
992 vm_page_sunbusy(ma[i]);
1000 error = EJUSTRETURN;
1005 * Check VIRF_DOOMED after we busied our pages. Since
1006 * vgonel() terminates the vnode' vm_object, it cannot
1007 * process past pages busied by us.
1009 if (VN_IS_DOOMED(vp)) {
1010 error = EJUSTRETURN;
1014 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1015 if (resid > uio->uio_resid)
1016 resid = uio->uio_resid;
1019 * Unlocked read of vnp_size is safe because truncation cannot
1020 * pass busied page. But we load vnp_size into a local
1021 * variable so that possible concurrent extension does not
1022 * break calculation.
1024 #if defined(__powerpc__) && !defined(__powerpc64__)
1025 vsz = obj->un_pager.vnp.vnp_size;
1027 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1029 if (uio->uio_offset >= vsz) {
1030 error = EJUSTRETURN;
1033 if (uio->uio_offset + resid > vsz)
1034 resid = vsz - uio->uio_offset;
1036 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1039 for (j = 0; j < i; j++) {
1041 vm_page_reference(ma[j]);
1042 vm_page_sunbusy(ma[j]);
1045 vm_object_pip_wakeup(obj);
1048 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1052 * File table vnode read routine.
1055 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1063 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1065 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1068 if (fp->f_flag & FNONBLOCK)
1069 ioflag |= IO_NDELAY;
1070 if (fp->f_flag & O_DIRECT)
1071 ioflag |= IO_DIRECT;
1074 * Try to read from page cache. VIRF_DOOMED check is racy but
1075 * allows us to avoid unneeded work outright.
1077 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1078 (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1079 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1081 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1084 if (error != EJUSTRETURN)
1088 advice = get_advice(fp, uio);
1089 vn_lock(vp, LK_SHARED | LK_RETRY);
1092 case POSIX_FADV_NORMAL:
1093 case POSIX_FADV_SEQUENTIAL:
1094 case POSIX_FADV_NOREUSE:
1095 ioflag |= sequential_heuristic(uio, fp);
1097 case POSIX_FADV_RANDOM:
1098 /* Disable read-ahead for random I/O. */
1101 orig_offset = uio->uio_offset;
1104 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1107 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1108 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1110 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1111 orig_offset != uio->uio_offset)
1113 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1114 * for the backing file after a POSIX_FADV_NOREUSE
1117 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1118 POSIX_FADV_DONTNEED);
1123 * File table vnode write routine.
1126 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1134 bool need_finished_write;
1136 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1138 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1140 if (vp->v_type == VREG)
1143 if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1144 ioflag |= IO_APPEND;
1145 if ((fp->f_flag & FNONBLOCK) != 0)
1146 ioflag |= IO_NDELAY;
1147 ioflag |= get_write_ioflag(fp);
1150 need_finished_write = false;
1151 if (vp->v_type != VCHR) {
1152 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1155 need_finished_write = true;
1158 advice = get_advice(fp, uio);
1160 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1162 case POSIX_FADV_NORMAL:
1163 case POSIX_FADV_SEQUENTIAL:
1164 case POSIX_FADV_NOREUSE:
1165 ioflag |= sequential_heuristic(uio, fp);
1167 case POSIX_FADV_RANDOM:
1168 /* XXX: Is this correct? */
1171 orig_offset = uio->uio_offset;
1174 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1177 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1178 fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1180 if (need_finished_write)
1181 vn_finished_write(mp);
1182 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1183 orig_offset != uio->uio_offset)
1185 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1186 * for the backing file after a POSIX_FADV_NOREUSE
1189 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1190 POSIX_FADV_DONTNEED);
1196 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1197 * prevent the following deadlock:
1199 * Assume that the thread A reads from the vnode vp1 into userspace
1200 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
1201 * currently not resident, then system ends up with the call chain
1202 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1203 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1204 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1205 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1206 * backed by the pages of vnode vp1, and some page in buf2 is not
1207 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1209 * To prevent the lock order reversal and deadlock, vn_io_fault() does
1210 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1211 * Instead, it first tries to do the whole range i/o with pagefaults
1212 * disabled. If all pages in the i/o buffer are resident and mapped,
1213 * VOP will succeed (ignoring the genuine filesystem errors).
1214 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1215 * i/o in chunks, with all pages in the chunk prefaulted and held
1216 * using vm_fault_quick_hold_pages().
1218 * Filesystems using this deadlock avoidance scheme should use the
1219 * array of the held pages from uio, saved in the curthread->td_ma,
1220 * instead of doing uiomove(). A helper function
1221 * vn_io_fault_uiomove() converts uiomove request into
1222 * uiomove_fromphys() over td_ma array.
1224 * Since vnode locks do not cover the whole i/o anymore, rangelocks
1225 * make the current i/o request atomic with respect to other i/os and
1230 * Decode vn_io_fault_args and perform the corresponding i/o.
1233 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1239 save = vm_fault_disable_pagefaults();
1240 switch (args->kind) {
1241 case VN_IO_FAULT_FOP:
1242 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1243 uio, args->cred, args->flags, td);
1245 case VN_IO_FAULT_VOP:
1246 if (uio->uio_rw == UIO_READ) {
1247 error = VOP_READ(args->args.vop_args.vp, uio,
1248 args->flags, args->cred);
1249 } else if (uio->uio_rw == UIO_WRITE) {
1250 error = VOP_WRITE(args->args.vop_args.vp, uio,
1251 args->flags, args->cred);
1255 panic("vn_io_fault_doio: unknown kind of io %d %d",
1256 args->kind, uio->uio_rw);
1258 vm_fault_enable_pagefaults(save);
1263 vn_io_fault_touch(char *base, const struct uio *uio)
1268 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1274 vn_io_fault_prefault_user(const struct uio *uio)
1277 const struct iovec *iov;
1282 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1283 ("vn_io_fault_prefault userspace"));
1287 resid = uio->uio_resid;
1288 base = iov->iov_base;
1291 error = vn_io_fault_touch(base, uio);
1294 if (len < PAGE_SIZE) {
1296 error = vn_io_fault_touch(base + len - 1, uio);
1301 if (++i >= uio->uio_iovcnt)
1303 iov = uio->uio_iov + i;
1304 base = iov->iov_base;
1316 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1317 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1318 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1319 * into args and call vn_io_fault1() to handle faults during the user
1320 * mode buffer accesses.
1323 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1326 vm_page_t ma[io_hold_cnt + 2];
1327 struct uio *uio_clone, short_uio;
1328 struct iovec short_iovec[1];
1329 vm_page_t *prev_td_ma;
1331 vm_offset_t addr, end;
1334 int error, cnt, saveheld, prev_td_ma_cnt;
1336 if (vn_io_fault_prefault) {
1337 error = vn_io_fault_prefault_user(uio);
1339 return (error); /* Or ignore ? */
1342 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1345 * The UFS follows IO_UNIT directive and replays back both
1346 * uio_offset and uio_resid if an error is encountered during the
1347 * operation. But, since the iovec may be already advanced,
1348 * uio is still in an inconsistent state.
1350 * Cache a copy of the original uio, which is advanced to the redo
1351 * point using UIO_NOCOPY below.
1353 uio_clone = cloneuio(uio);
1354 resid = uio->uio_resid;
1356 short_uio.uio_segflg = UIO_USERSPACE;
1357 short_uio.uio_rw = uio->uio_rw;
1358 short_uio.uio_td = uio->uio_td;
1360 error = vn_io_fault_doio(args, uio, td);
1361 if (error != EFAULT)
1364 atomic_add_long(&vn_io_faults_cnt, 1);
1365 uio_clone->uio_segflg = UIO_NOCOPY;
1366 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1367 uio_clone->uio_segflg = uio->uio_segflg;
1369 saveheld = curthread_pflags_set(TDP_UIOHELD);
1370 prev_td_ma = td->td_ma;
1371 prev_td_ma_cnt = td->td_ma_cnt;
1373 while (uio_clone->uio_resid != 0) {
1374 len = uio_clone->uio_iov->iov_len;
1376 KASSERT(uio_clone->uio_iovcnt >= 1,
1377 ("iovcnt underflow"));
1378 uio_clone->uio_iov++;
1379 uio_clone->uio_iovcnt--;
1382 if (len > ptoa(io_hold_cnt))
1383 len = ptoa(io_hold_cnt);
1384 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1385 end = round_page(addr + len);
1390 cnt = atop(end - trunc_page(addr));
1392 * A perfectly misaligned address and length could cause
1393 * both the start and the end of the chunk to use partial
1394 * page. +2 accounts for such a situation.
1396 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1397 addr, len, prot, ma, io_hold_cnt + 2);
1402 short_uio.uio_iov = &short_iovec[0];
1403 short_iovec[0].iov_base = (void *)addr;
1404 short_uio.uio_iovcnt = 1;
1405 short_uio.uio_resid = short_iovec[0].iov_len = len;
1406 short_uio.uio_offset = uio_clone->uio_offset;
1408 td->td_ma_cnt = cnt;
1410 error = vn_io_fault_doio(args, &short_uio, td);
1411 vm_page_unhold_pages(ma, cnt);
1412 adv = len - short_uio.uio_resid;
1414 uio_clone->uio_iov->iov_base =
1415 (char *)uio_clone->uio_iov->iov_base + adv;
1416 uio_clone->uio_iov->iov_len -= adv;
1417 uio_clone->uio_resid -= adv;
1418 uio_clone->uio_offset += adv;
1420 uio->uio_resid -= adv;
1421 uio->uio_offset += adv;
1423 if (error != 0 || adv == 0)
1426 td->td_ma = prev_td_ma;
1427 td->td_ma_cnt = prev_td_ma_cnt;
1428 curthread_pflags_restore(saveheld);
1430 free(uio_clone, M_IOV);
1435 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1436 int flags, struct thread *td)
1441 struct vn_io_fault_args args;
1443 bool do_io_fault, do_rangelock;
1445 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1449 * The ability to read(2) on a directory has historically been
1450 * allowed for all users, but this can and has been the source of
1451 * at least one security issue in the past. As such, it is now hidden
1452 * away behind a sysctl for those that actually need it to use it, and
1453 * restricted to root when it's turned on to make it relatively safe to
1454 * leave on for longer sessions of need.
1456 if (vp->v_type == VDIR) {
1457 KASSERT(uio->uio_rw == UIO_READ,
1458 ("illegal write attempted on a directory"));
1459 if (!vfs_allow_read_dir)
1461 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1465 do_io_fault = do_vn_io_fault(vp, uio);
1466 do_rangelock = do_io_fault || (vn_irflag_read(vp) & VIRF_PGREAD) != 0;
1467 foffset_lock_uio(fp, uio, flags);
1469 if (uio->uio_rw == UIO_READ) {
1470 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1471 uio->uio_offset + uio->uio_resid);
1472 } else if ((fp->f_flag & O_APPEND) != 0 ||
1473 (flags & FOF_OFFSET) == 0) {
1474 /* For appenders, punt and lock the whole range. */
1475 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1477 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1478 uio->uio_offset + uio->uio_resid);
1482 args.kind = VN_IO_FAULT_FOP;
1483 args.args.fop_args.fp = fp;
1484 args.args.fop_args.doio = doio;
1485 args.cred = active_cred;
1486 args.flags = flags | FOF_OFFSET;
1487 error = vn_io_fault1(vp, uio, &args, td);
1489 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1492 vn_rangelock_unlock(vp, rl_cookie);
1493 foffset_unlock_uio(fp, uio, flags);
1498 * Helper function to perform the requested uiomove operation using
1499 * the held pages for io->uio_iov[0].iov_base buffer instead of
1500 * copyin/copyout. Access to the pages with uiomove_fromphys()
1501 * instead of iov_base prevents page faults that could occur due to
1502 * pmap_collect() invalidating the mapping created by
1503 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1504 * object cleanup revoking the write access from page mappings.
1506 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1507 * instead of plain uiomove().
1510 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1512 struct uio transp_uio;
1513 struct iovec transp_iov[1];
1519 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1520 uio->uio_segflg != UIO_USERSPACE)
1521 return (uiomove(data, xfersize, uio));
1523 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1524 transp_iov[0].iov_base = data;
1525 transp_uio.uio_iov = &transp_iov[0];
1526 transp_uio.uio_iovcnt = 1;
1527 if (xfersize > uio->uio_resid)
1528 xfersize = uio->uio_resid;
1529 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1530 transp_uio.uio_offset = 0;
1531 transp_uio.uio_segflg = UIO_SYSSPACE;
1533 * Since transp_iov points to data, and td_ma page array
1534 * corresponds to original uio->uio_iov, we need to invert the
1535 * direction of the i/o operation as passed to
1536 * uiomove_fromphys().
1538 switch (uio->uio_rw) {
1540 transp_uio.uio_rw = UIO_READ;
1543 transp_uio.uio_rw = UIO_WRITE;
1546 transp_uio.uio_td = uio->uio_td;
1547 error = uiomove_fromphys(td->td_ma,
1548 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1549 xfersize, &transp_uio);
1550 adv = xfersize - transp_uio.uio_resid;
1552 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1553 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1555 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1557 td->td_ma_cnt -= pgadv;
1558 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1559 uio->uio_iov->iov_len -= adv;
1560 uio->uio_resid -= adv;
1561 uio->uio_offset += adv;
1566 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1570 vm_offset_t iov_base;
1574 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1575 uio->uio_segflg != UIO_USERSPACE)
1576 return (uiomove_fromphys(ma, offset, xfersize, uio));
1578 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1579 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1580 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1581 switch (uio->uio_rw) {
1583 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1587 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1591 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1593 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1595 td->td_ma_cnt -= pgadv;
1596 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1597 uio->uio_iov->iov_len -= cnt;
1598 uio->uio_resid -= cnt;
1599 uio->uio_offset += cnt;
1604 * File table truncate routine.
1607 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1619 * Lock the whole range for truncation. Otherwise split i/o
1620 * might happen partly before and partly after the truncation.
1622 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1623 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1626 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1627 AUDIT_ARG_VNODE1(vp);
1628 if (vp->v_type == VDIR) {
1633 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1637 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1641 vn_finished_write(mp);
1643 vn_rangelock_unlock(vp, rl_cookie);
1644 if (error == ERELOOKUP)
1650 * Truncate a file that is already locked.
1653 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1659 error = VOP_ADD_WRITECOUNT(vp, 1);
1662 vattr.va_size = length;
1664 vattr.va_vaflags |= VA_SYNC;
1665 error = VOP_SETATTR(vp, &vattr, cred);
1666 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1672 * File table vnode stat routine.
1675 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1678 struct vnode *vp = fp->f_vnode;
1681 vn_lock(vp, LK_SHARED | LK_RETRY);
1682 error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1689 * File table vnode ioctl routine.
1692 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1697 struct fiobmap2_arg *bmarg;
1701 switch (vp->v_type) {
1706 vn_lock(vp, LK_SHARED | LK_RETRY);
1707 error = VOP_GETATTR(vp, &vattr, active_cred);
1710 *(int *)data = vattr.va_size - fp->f_offset;
1713 bmarg = (struct fiobmap2_arg *)data;
1714 vn_lock(vp, LK_SHARED | LK_RETRY);
1716 error = mac_vnode_check_read(active_cred, fp->f_cred,
1720 error = VOP_BMAP(vp, bmarg->bn, NULL,
1721 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1728 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1733 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1741 * File table vnode poll routine.
1744 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1751 #if defined(MAC) || defined(AUDIT)
1752 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1753 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1754 AUDIT_ARG_VNODE1(vp);
1755 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1761 error = VOP_POLL(vp, events, fp->f_cred, td);
1766 * Acquire the requested lock and then check for validity. LK_RETRY
1767 * permits vn_lock to return doomed vnodes.
1769 static int __noinline
1770 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1774 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1775 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1778 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1780 if ((flags & LK_RETRY) == 0) {
1791 * Nothing to do if we got the lock.
1797 * Interlock was dropped by the call in _vn_lock.
1799 flags &= ~LK_INTERLOCK;
1801 error = VOP_LOCK1(vp, flags, file, line);
1802 } while (error != 0);
1807 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1811 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1812 ("vn_lock: no locktype (%d passed)", flags));
1813 VNPASS(vp->v_holdcnt > 0, vp);
1814 error = VOP_LOCK1(vp, flags, file, line);
1815 if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1816 return (_vn_lock_fallback(vp, flags, file, line, error));
1821 * File table vnode close routine.
1824 vn_closefile(struct file *fp, struct thread *td)
1832 fp->f_ops = &badfileops;
1833 ref = (fp->f_flag & FHASLOCK) != 0;
1835 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1837 if (__predict_false(ref)) {
1838 lf.l_whence = SEEK_SET;
1841 lf.l_type = F_UNLCK;
1842 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1849 * Preparing to start a filesystem write operation. If the operation is
1850 * permitted, then we bump the count of operations in progress and
1851 * proceed. If a suspend request is in progress, we wait until the
1852 * suspension is over, and then proceed.
1855 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1857 struct mount_pcpu *mpcpu;
1860 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1861 vfs_op_thread_enter(mp, mpcpu)) {
1862 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1863 vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1864 vfs_op_thread_exit(mp, mpcpu);
1869 mtx_assert(MNT_MTX(mp), MA_OWNED);
1876 * Check on status of suspension.
1878 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1879 mp->mnt_susp_owner != curthread) {
1880 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1881 (flags & PCATCH) : 0) | (PUSER - 1);
1882 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1883 if (flags & V_NOWAIT) {
1884 error = EWOULDBLOCK;
1887 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1893 if (flags & V_XSLEEP)
1895 mp->mnt_writeopcount++;
1897 if (error != 0 || (flags & V_XSLEEP) != 0)
1904 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1909 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1910 ("V_MNTREF requires mp"));
1914 * If a vnode is provided, get and return the mount point that
1915 * to which it will write.
1918 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1920 if (error != EOPNOTSUPP)
1925 if ((mp = *mpp) == NULL)
1929 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1931 * As long as a vnode is not provided we need to acquire a
1932 * refcount for the provided mountpoint too, in order to
1933 * emulate a vfs_ref().
1935 if (vp == NULL && (flags & V_MNTREF) == 0)
1938 return (vn_start_write_refed(mp, flags, false));
1942 * Secondary suspension. Used by operations such as vop_inactive
1943 * routines that are needed by the higher level functions. These
1944 * are allowed to proceed until all the higher level functions have
1945 * completed (indicated by mnt_writeopcount dropping to zero). At that
1946 * time, these operations are halted until the suspension is over.
1949 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1954 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1955 ("V_MNTREF requires mp"));
1959 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1961 if (error != EOPNOTSUPP)
1967 * If we are not suspended or have not yet reached suspended
1968 * mode, then let the operation proceed.
1970 if ((mp = *mpp) == NULL)
1974 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1976 * As long as a vnode is not provided we need to acquire a
1977 * refcount for the provided mountpoint too, in order to
1978 * emulate a vfs_ref().
1981 if (vp == NULL && (flags & V_MNTREF) == 0)
1983 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1984 mp->mnt_secondary_writes++;
1985 mp->mnt_secondary_accwrites++;
1989 if (flags & V_NOWAIT) {
1992 return (EWOULDBLOCK);
1995 * Wait for the suspension to finish.
1997 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1998 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
2007 * Filesystem write operation has completed. If we are suspending and this
2008 * operation is the last one, notify the suspender that the suspension is
2012 vn_finished_write(struct mount *mp)
2014 struct mount_pcpu *mpcpu;
2020 if (vfs_op_thread_enter(mp, mpcpu)) {
2021 vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2022 vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2023 vfs_op_thread_exit(mp, mpcpu);
2028 vfs_assert_mount_counters(mp);
2030 c = --mp->mnt_writeopcount;
2031 if (mp->mnt_vfs_ops == 0) {
2032 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2037 vfs_dump_mount_counters(mp);
2038 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2039 wakeup(&mp->mnt_writeopcount);
2044 * Filesystem secondary write operation has completed. If we are
2045 * suspending and this operation is the last one, notify the suspender
2046 * that the suspension is now in effect.
2049 vn_finished_secondary_write(struct mount *mp)
2055 mp->mnt_secondary_writes--;
2056 if (mp->mnt_secondary_writes < 0)
2057 panic("vn_finished_secondary_write: neg cnt");
2058 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2059 mp->mnt_secondary_writes <= 0)
2060 wakeup(&mp->mnt_secondary_writes);
2065 * Request a filesystem to suspend write operations.
2068 vfs_write_suspend(struct mount *mp, int flags)
2075 vfs_assert_mount_counters(mp);
2076 if (mp->mnt_susp_owner == curthread) {
2077 vfs_op_exit_locked(mp);
2081 while (mp->mnt_kern_flag & MNTK_SUSPEND)
2082 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2085 * Unmount holds a write reference on the mount point. If we
2086 * own busy reference and drain for writers, we deadlock with
2087 * the reference draining in the unmount path. Callers of
2088 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2089 * vfs_busy() reference is owned and caller is not in the
2092 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2093 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2094 vfs_op_exit_locked(mp);
2099 mp->mnt_kern_flag |= MNTK_SUSPEND;
2100 mp->mnt_susp_owner = curthread;
2101 if (mp->mnt_writeopcount > 0)
2102 (void) msleep(&mp->mnt_writeopcount,
2103 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2106 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2107 vfs_write_resume(mp, 0);
2108 /* vfs_write_resume does vfs_op_exit() for us */
2114 * Request a filesystem to resume write operations.
2117 vfs_write_resume(struct mount *mp, int flags)
2121 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2122 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2123 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2125 mp->mnt_susp_owner = NULL;
2126 wakeup(&mp->mnt_writeopcount);
2127 wakeup(&mp->mnt_flag);
2128 curthread->td_pflags &= ~TDP_IGNSUSP;
2129 if ((flags & VR_START_WRITE) != 0) {
2131 mp->mnt_writeopcount++;
2134 if ((flags & VR_NO_SUSPCLR) == 0)
2137 } else if ((flags & VR_START_WRITE) != 0) {
2139 vn_start_write_refed(mp, 0, true);
2146 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2150 vfs_write_suspend_umnt(struct mount *mp)
2154 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2155 ("vfs_write_suspend_umnt: recursed"));
2157 /* dounmount() already called vn_start_write(). */
2159 vn_finished_write(mp);
2160 error = vfs_write_suspend(mp, 0);
2162 vn_start_write(NULL, &mp, V_WAIT);
2166 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2169 vn_start_write(NULL, &mp, V_WAIT);
2171 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2172 wakeup(&mp->mnt_flag);
2174 curthread->td_pflags |= TDP_IGNSUSP;
2179 * Implement kqueues for files by translating it to vnode operation.
2182 vn_kqfilter(struct file *fp, struct knote *kn)
2185 return (VOP_KQFILTER(fp->f_vnode, kn));
2189 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2191 if ((fp->f_flag & FKQALLOWED) == 0)
2193 return (vn_kqfilter(fp, kn));
2197 * Simplified in-kernel wrapper calls for extended attribute access.
2198 * Both calls pass in a NULL credential, authorizing as "kernel" access.
2199 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2202 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2203 const char *attrname, int *buflen, char *buf, struct thread *td)
2209 iov.iov_len = *buflen;
2212 auio.uio_iov = &iov;
2213 auio.uio_iovcnt = 1;
2214 auio.uio_rw = UIO_READ;
2215 auio.uio_segflg = UIO_SYSSPACE;
2217 auio.uio_offset = 0;
2218 auio.uio_resid = *buflen;
2220 if ((ioflg & IO_NODELOCKED) == 0)
2221 vn_lock(vp, LK_SHARED | LK_RETRY);
2223 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2225 /* authorize attribute retrieval as kernel */
2226 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2229 if ((ioflg & IO_NODELOCKED) == 0)
2233 *buflen = *buflen - auio.uio_resid;
2240 * XXX failure mode if partially written?
2243 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2244 const char *attrname, int buflen, char *buf, struct thread *td)
2251 iov.iov_len = buflen;
2254 auio.uio_iov = &iov;
2255 auio.uio_iovcnt = 1;
2256 auio.uio_rw = UIO_WRITE;
2257 auio.uio_segflg = UIO_SYSSPACE;
2259 auio.uio_offset = 0;
2260 auio.uio_resid = buflen;
2262 if ((ioflg & IO_NODELOCKED) == 0) {
2263 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2265 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2268 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2270 /* authorize attribute setting as kernel */
2271 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2273 if ((ioflg & IO_NODELOCKED) == 0) {
2274 vn_finished_write(mp);
2282 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2283 const char *attrname, struct thread *td)
2288 if ((ioflg & IO_NODELOCKED) == 0) {
2289 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2291 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2294 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2296 /* authorize attribute removal as kernel */
2297 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2298 if (error == EOPNOTSUPP)
2299 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2302 if ((ioflg & IO_NODELOCKED) == 0) {
2303 vn_finished_write(mp);
2311 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2315 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2319 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2322 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2327 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2328 int lkflags, struct vnode **rvp)
2333 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2335 ltype = VOP_ISLOCKED(vp);
2336 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2337 ("vn_vget_ino: vp not locked"));
2338 error = vfs_busy(mp, MBF_NOWAIT);
2342 error = vfs_busy(mp, 0);
2343 vn_lock(vp, ltype | LK_RETRY);
2347 if (VN_IS_DOOMED(vp)) {
2353 error = alloc(mp, alloc_arg, lkflags, rvp);
2355 if (error != 0 || *rvp != vp)
2356 vn_lock(vp, ltype | LK_RETRY);
2357 if (VN_IS_DOOMED(vp)) {
2370 vn_send_sigxfsz(struct proc *p)
2373 kern_psignal(p, SIGXFSZ);
2378 vn_rlimit_trunc(u_quad_t size, struct thread *td)
2380 if (size <= lim_cur(td, RLIMIT_FSIZE))
2382 vn_send_sigxfsz(td->td_proc);
2387 vn_rlimit_fsizex1(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2388 bool adj, struct thread *td)
2393 if (vp->v_type != VREG)
2397 * Handle file system maximum file size.
2399 if (maxfsz != 0 && uio->uio_offset + uio->uio_resid > maxfsz) {
2400 if (!adj || uio->uio_offset >= maxfsz)
2402 uio->uio_resid = maxfsz - uio->uio_offset;
2406 * This is kernel write (e.g. vnode_pager) or accounting
2407 * write, ignore limit.
2409 if (td == NULL || (td->td_pflags2 & TDP2_ACCT) != 0)
2413 * Calculate file size limit.
2415 ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2416 lim = __predict_false(ktr_write) ? td->td_ktr_io_lim :
2417 lim_cur(td, RLIMIT_FSIZE);
2420 * Is the limit reached?
2422 if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2426 * Prepared filesystems can handle writes truncated to the
2429 if (adj && (uoff_t)uio->uio_offset < lim) {
2430 uio->uio_resid = lim - (uoff_t)uio->uio_offset;
2434 if (!ktr_write || ktr_filesize_limit_signal)
2435 vn_send_sigxfsz(td->td_proc);
2440 * Helper for VOP_WRITE() implementations, the common code to
2441 * handle maximum supported file size on the filesystem, and
2442 * RLIMIT_FSIZE, except for special writes from accounting subsystem
2445 * For maximum file size (maxfsz argument):
2446 * - return EFBIG if uio_offset is beyond it
2447 * - otherwise, clamp uio_resid if write would extend file beyond maxfsz.
2450 * - return EFBIG and send SIGXFSZ if uio_offset is beyond the limit
2451 * - otherwise, clamp uio_resid if write would extend file beyond limit.
2453 * If clamping occured, the adjustment for uio_resid is stored in
2454 * *resid_adj, to be re-applied by vn_rlimit_fsizex_res() on return
2458 vn_rlimit_fsizex(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2459 ssize_t *resid_adj, struct thread *td)
2465 resid_orig = uio->uio_resid;
2466 adj = resid_adj != NULL;
2467 error = vn_rlimit_fsizex1(vp, uio, maxfsz, adj, td);
2469 *resid_adj = resid_orig - uio->uio_resid;
2474 vn_rlimit_fsizex_res(struct uio *uio, ssize_t resid_adj)
2476 uio->uio_resid += resid_adj;
2480 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2483 return (vn_rlimit_fsizex(vp, __DECONST(struct uio *, uio), 0, NULL,
2488 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2495 vn_lock(vp, LK_SHARED | LK_RETRY);
2496 AUDIT_ARG_VNODE1(vp);
2499 return (setfmode(td, active_cred, vp, mode));
2503 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2510 vn_lock(vp, LK_SHARED | LK_RETRY);
2511 AUDIT_ARG_VNODE1(vp);
2514 return (setfown(td, active_cred, vp, uid, gid));
2518 * Remove pages in the range ["start", "end") from the vnode's VM object. If
2519 * "end" is 0, then the range extends to the end of the object.
2522 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2526 if ((object = vp->v_object) == NULL)
2528 VM_OBJECT_WLOCK(object);
2529 vm_object_page_remove(object, start, end, 0);
2530 VM_OBJECT_WUNLOCK(object);
2534 * Like vn_pages_remove(), but skips invalid pages, which by definition are not
2535 * mapped into any process' address space. Filesystems may use this in
2536 * preference to vn_pages_remove() to avoid blocking on pages busied in
2537 * preparation for a VOP_GETPAGES.
2540 vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2544 if ((object = vp->v_object) == NULL)
2546 VM_OBJECT_WLOCK(object);
2547 vm_object_page_remove(object, start, end, OBJPR_VALIDONLY);
2548 VM_OBJECT_WUNLOCK(object);
2552 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2561 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2562 ("Wrong command %lu", cmd));
2564 if (vn_lock(vp, LK_EXCLUSIVE) != 0)
2566 if (vp->v_type != VREG) {
2570 error = VOP_GETATTR(vp, &va, cred);
2574 if (noff < 0 || noff >= va.va_size) {
2579 /* See the comment in ufs_bmap_seekdata(). */
2582 VM_OBJECT_WLOCK(obj);
2583 vm_object_page_clean(obj, 0, 0, OBJPC_SYNC);
2584 VM_OBJECT_WUNLOCK(obj);
2587 bsize = vp->v_mount->mnt_stat.f_iosize;
2588 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2590 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2591 if (error == EOPNOTSUPP) {
2595 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2596 (bnp != -1 && cmd == FIOSEEKDATA)) {
2603 if (noff > va.va_size)
2605 /* noff == va.va_size. There is an implicit hole at the end of file. */
2606 if (cmd == FIOSEEKDATA)
2616 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2621 off_t foffset, size;
2624 cred = td->td_ucred;
2626 foffset = foffset_lock(fp, 0);
2627 noneg = (vp->v_type != VCHR);
2633 (offset > 0 && foffset > OFF_MAX - offset))) {
2640 vn_lock(vp, LK_SHARED | LK_RETRY);
2641 error = VOP_GETATTR(vp, &vattr, cred);
2647 * If the file references a disk device, then fetch
2648 * the media size and use that to determine the ending
2651 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2652 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2653 vattr.va_size = size;
2655 (vattr.va_size > OFF_MAX ||
2656 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2660 offset += vattr.va_size;
2665 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2666 if (error == ENOTTY)
2670 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2671 if (error == ENOTTY)
2677 if (error == 0 && noneg && offset < 0)
2681 VFS_KNOTE_UNLOCKED(vp, 0);
2682 td->td_uretoff.tdu_off = offset;
2684 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2689 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2695 * Grant permission if the caller is the owner of the file, or
2696 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2697 * on the file. If the time pointer is null, then write
2698 * permission on the file is also sufficient.
2700 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2701 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2702 * will be allowed to set the times [..] to the current
2705 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2706 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2707 error = VOP_ACCESS(vp, VWRITE, cred, td);
2712 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2717 if (fp->f_type == DTYPE_FIFO)
2718 kif->kf_type = KF_TYPE_FIFO;
2720 kif->kf_type = KF_TYPE_VNODE;
2723 FILEDESC_SUNLOCK(fdp);
2724 error = vn_fill_kinfo_vnode(vp, kif);
2726 FILEDESC_SLOCK(fdp);
2731 vn_fill_junk(struct kinfo_file *kif)
2736 * Simulate vn_fullpath returning changing values for a given
2737 * vp during e.g. coredump.
2739 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2740 olen = strlen(kif->kf_path);
2742 strcpy(&kif->kf_path[len - 1], "$");
2744 for (; olen < len; olen++)
2745 strcpy(&kif->kf_path[olen], "A");
2749 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2752 char *fullpath, *freepath;
2755 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2758 error = vn_fullpath(vp, &fullpath, &freepath);
2760 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2762 if (freepath != NULL)
2763 free(freepath, M_TEMP);
2765 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2770 * Retrieve vnode attributes.
2772 va.va_fsid = VNOVAL;
2774 vn_lock(vp, LK_SHARED | LK_RETRY);
2775 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2779 if (va.va_fsid != VNOVAL)
2780 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2782 kif->kf_un.kf_file.kf_file_fsid =
2783 vp->v_mount->mnt_stat.f_fsid.val[0];
2784 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2785 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2786 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2787 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2788 kif->kf_un.kf_file.kf_file_size = va.va_size;
2789 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2790 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2791 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2792 kif->kf_un.kf_file.kf_file_nlink = va.va_nlink;
2797 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2798 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2802 struct pmckern_map_in pkm;
2808 boolean_t writecounted;
2811 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2812 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2814 * POSIX shared-memory objects are defined to have
2815 * kernel persistence, and are not defined to support
2816 * read(2)/write(2) -- or even open(2). Thus, we can
2817 * use MAP_ASYNC to trade on-disk coherence for speed.
2818 * The shm_open(3) library routine turns on the FPOSIXSHM
2819 * flag to request this behavior.
2821 if ((fp->f_flag & FPOSIXSHM) != 0)
2822 flags |= MAP_NOSYNC;
2827 * Ensure that file and memory protections are
2828 * compatible. Note that we only worry about
2829 * writability if mapping is shared; in this case,
2830 * current and max prot are dictated by the open file.
2831 * XXX use the vnode instead? Problem is: what
2832 * credentials do we use for determination? What if
2833 * proc does a setuid?
2836 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2837 maxprot = VM_PROT_NONE;
2838 if ((prot & VM_PROT_EXECUTE) != 0)
2841 maxprot = VM_PROT_EXECUTE;
2842 if ((fp->f_flag & FREAD) != 0)
2843 maxprot |= VM_PROT_READ;
2844 else if ((prot & VM_PROT_READ) != 0)
2848 * If we are sharing potential changes via MAP_SHARED and we
2849 * are trying to get write permission although we opened it
2850 * without asking for it, bail out.
2852 if ((flags & MAP_SHARED) != 0) {
2853 if ((fp->f_flag & FWRITE) != 0)
2854 maxprot |= VM_PROT_WRITE;
2855 else if ((prot & VM_PROT_WRITE) != 0)
2858 maxprot |= VM_PROT_WRITE;
2859 cap_maxprot |= VM_PROT_WRITE;
2861 maxprot &= cap_maxprot;
2864 * For regular files and shared memory, POSIX requires that
2865 * the value of foff be a legitimate offset within the data
2866 * object. In particular, negative offsets are invalid.
2867 * Blocking negative offsets and overflows here avoids
2868 * possible wraparound or user-level access into reserved
2869 * ranges of the data object later. In contrast, POSIX does
2870 * not dictate how offsets are used by device drivers, so in
2871 * the case of a device mapping a negative offset is passed
2878 foff > OFF_MAX - size)
2881 writecounted = FALSE;
2882 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2883 &foff, &object, &writecounted);
2886 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2887 foff, writecounted, td);
2890 * If this mapping was accounted for in the vnode's
2891 * writecount, then undo that now.
2894 vm_pager_release_writecount(object, 0, size);
2895 vm_object_deallocate(object);
2898 /* Inform hwpmc(4) if an executable is being mapped. */
2899 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2900 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2902 pkm.pm_address = (uintptr_t) *addr;
2903 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2911 vn_fsid(struct vnode *vp, struct vattr *va)
2915 f = &vp->v_mount->mnt_stat.f_fsid;
2916 va->va_fsid = (uint32_t)f->val[1];
2917 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2918 va->va_fsid += (uint32_t)f->val[0];
2922 vn_fsync_buf(struct vnode *vp, int waitfor)
2924 struct buf *bp, *nbp;
2927 int error, maxretry;
2930 maxretry = 10000; /* large, arbitrarily chosen */
2932 if (vp->v_type == VCHR) {
2934 mp = vp->v_rdev->si_mountpt;
2941 * MARK/SCAN initialization to avoid infinite loops.
2943 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2944 bp->b_vflags &= ~BV_SCANNED;
2949 * Flush all dirty buffers associated with a vnode.
2952 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2953 if ((bp->b_vflags & BV_SCANNED) != 0)
2955 bp->b_vflags |= BV_SCANNED;
2956 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2957 if (waitfor != MNT_WAIT)
2960 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2961 BO_LOCKPTR(bo)) != 0) {
2968 KASSERT(bp->b_bufobj == bo,
2969 ("bp %p wrong b_bufobj %p should be %p",
2970 bp, bp->b_bufobj, bo));
2971 if ((bp->b_flags & B_DELWRI) == 0)
2972 panic("fsync: not dirty");
2973 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2979 if (maxretry < 1000)
2980 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2986 * If synchronous the caller expects us to completely resolve all
2987 * dirty buffers in the system. Wait for in-progress I/O to
2988 * complete (which could include background bitmap writes), then
2989 * retry if dirty blocks still exist.
2991 if (waitfor == MNT_WAIT) {
2992 bufobj_wwait(bo, 0, 0);
2993 if (bo->bo_dirty.bv_cnt > 0) {
2995 * If we are unable to write any of these buffers
2996 * then we fail now rather than trying endlessly
2997 * to write them out.
2999 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
3000 if ((error = bp->b_error) != 0)
3002 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
3003 (error == 0 && --maxretry >= 0))
3011 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
3017 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
3018 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
3019 * to do the actual copy.
3020 * vn_generic_copy_file_range() is factored out, so it can be called
3021 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
3022 * different file systems.
3025 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
3026 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
3027 struct ucred *outcred, struct thread *fsize_td)
3029 struct mount *inmp, *outmp;
3030 struct vnode *invpl, *outvpl;
3035 invpl = outvpl = NULL;
3037 *lenp = 0; /* For error returns. */
3040 /* Do some sanity checks on the arguments. */
3041 if (invp->v_type == VDIR || outvp->v_type == VDIR)
3043 else if (*inoffp < 0 || *outoffp < 0 ||
3044 invp->v_type != VREG || outvp->v_type != VREG)
3049 /* Ensure offset + len does not wrap around. */
3052 if (uval > INT64_MAX)
3053 len = INT64_MAX - *inoffp;
3056 if (uval > INT64_MAX)
3057 len = INT64_MAX - *outoffp;
3061 error = VOP_GETLOWVNODE(invp, &invpl, FREAD);
3064 error = VOP_GETLOWVNODE(outvp, &outvpl, FWRITE);
3068 inmp = invpl->v_mount;
3069 outmp = outvpl->v_mount;
3070 if (inmp == NULL || outmp == NULL)
3074 error = vfs_busy(inmp, 0);
3079 error = vfs_busy(outmp, MBF_NOWAIT);
3082 error = vfs_busy(outmp, 0);
3093 * If the two vnode are for the same file system, call
3094 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
3095 * which can handle copies across multiple file systems.
3099 error = VOP_COPY_FILE_RANGE(invpl, inoffp, outvpl, outoffp,
3100 lenp, flags, incred, outcred, fsize_td);
3103 if (error == ENOSYS)
3104 error = vn_generic_copy_file_range(invpl, inoffp, outvpl,
3105 outoffp, lenp, flags, incred, outcred, fsize_td);
3120 * Test len bytes of data starting at dat for all bytes == 0.
3121 * Return true if all bytes are zero, false otherwise.
3122 * Expects dat to be well aligned.
3125 mem_iszero(void *dat, int len)
3131 for (p = dat; len > 0; len -= sizeof(*p), p++) {
3132 if (len >= sizeof(*p)) {
3136 cp = (const char *)p;
3137 for (i = 0; i < len; i++, cp++)
3146 * Look for a hole in the output file and, if found, adjust *outoffp
3147 * and *xferp to skip past the hole.
3148 * *xferp is the entire hole length to be written and xfer2 is how many bytes
3149 * to be written as 0's upon return.
3152 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3153 off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3158 if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3159 *dataoffp = *outoffp;
3160 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3163 *holeoffp = *dataoffp;
3164 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3167 if (error != 0 || *holeoffp == *dataoffp) {
3169 * Since outvp is unlocked, it may be possible for
3170 * another thread to do a truncate(), lseek(), write()
3171 * creating a hole at startoff between the above
3172 * VOP_IOCTL() calls, if the other thread does not do
3174 * If that happens, *holeoffp == *dataoffp and finding
3175 * the hole has failed, so disable vn_skip_hole().
3177 *holeoffp = -1; /* Disable use of vn_skip_hole(). */
3180 KASSERT(*dataoffp >= *outoffp,
3181 ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3182 (intmax_t)*dataoffp, (intmax_t)*outoffp));
3183 KASSERT(*holeoffp > *dataoffp,
3184 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3185 (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3189 * If there is a hole before the data starts, advance *outoffp and
3190 * *xferp past the hole.
3192 if (*dataoffp > *outoffp) {
3193 delta = *dataoffp - *outoffp;
3194 if (delta >= *xferp) {
3195 /* Entire *xferp is a hole. */
3202 xfer2 = MIN(xfer2, *xferp);
3206 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3207 * that the write ends at the start of the hole.
3208 * *holeoffp should always be greater than *outoffp, but for the
3209 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3212 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3213 xfer2 = *holeoffp - *outoffp;
3218 * Write an xfer sized chunk to outvp in blksize blocks from dat.
3219 * dat is a maximum of blksize in length and can be written repeatedly in
3221 * If growfile == true, just grow the file via vn_truncate_locked() instead
3222 * of doing actual writes.
3223 * If checkhole == true, a hole is being punched, so skip over any hole
3224 * already in the output file.
3227 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3228 u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3231 off_t dataoff, holeoff, xfer2;
3235 * Loop around doing writes of blksize until write has been completed.
3236 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3237 * done for each iteration, since the xfer argument can be very
3238 * large if there is a large hole to punch in the output file.
3243 xfer2 = MIN(xfer, blksize);
3246 * Punching a hole. Skip writing if there is
3247 * already a hole in the output file.
3249 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3250 &dataoff, &holeoff, cred);
3255 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3260 error = vn_start_write(outvp, &mp, V_WAIT);
3264 error = vn_lock(outvp, LK_EXCLUSIVE);
3266 error = vn_truncate_locked(outvp, outoff + xfer,
3271 error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3273 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3274 outoff, UIO_SYSSPACE, IO_NODELOCKED,
3275 curthread->td_ucred, cred, NULL, curthread);
3282 vn_finished_write(mp);
3283 } while (!growfile && xfer > 0 && error == 0);
3288 * Copy a byte range of one file to another. This function can handle the
3289 * case where invp and outvp are on different file systems.
3290 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3291 * is no better file system specific way to do it.
3294 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3295 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3296 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3298 struct vattr va, inva;
3300 off_t startoff, endoff, xfer, xfer2;
3302 int error, interrupted;
3303 bool cantseek, readzeros, eof, lastblock, holetoeof;
3304 ssize_t aresid, r = 0;
3305 size_t copylen, len, savlen;
3307 long holein, holeout;
3308 struct timespec curts, endts;
3310 holein = holeout = 0;
3311 savlen = len = *lenp;
3316 error = vn_lock(invp, LK_SHARED);
3319 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3322 error = VOP_GETATTR(invp, &inva, incred);
3328 error = vn_start_write(outvp, &mp, V_WAIT);
3330 error = vn_lock(outvp, LK_EXCLUSIVE);
3333 * If fsize_td != NULL, do a vn_rlimit_fsizex() call,
3334 * now that outvp is locked.
3336 if (fsize_td != NULL) {
3339 io.uio_offset = *outoffp;
3341 error = vn_rlimit_fsizex(outvp, &io, 0, &r, fsize_td);
3342 len = savlen = io.uio_resid;
3344 * No need to call vn_rlimit_fsizex_res before return,
3345 * since the uio is local.
3348 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3351 * Holes that are past EOF do not need to be written as a block
3352 * of zero bytes. So, truncate the output file as far as
3353 * possible and then use va.va_size to decide if writing 0
3354 * bytes is necessary in the loop below.
3357 error = VOP_GETATTR(outvp, &va, outcred);
3358 if (error == 0 && va.va_size > *outoffp && va.va_size <=
3361 error = mac_vnode_check_write(curthread->td_ucred,
3365 error = vn_truncate_locked(outvp, *outoffp,
3368 va.va_size = *outoffp;
3373 vn_finished_write(mp);
3377 if (holein == 0 && holeout > 0) {
3379 * For this special case, the input data will be scanned
3380 * for blocks of all 0 bytes. For these blocks, the
3381 * write can be skipped for the output file to create
3382 * an unallocated region.
3383 * Therefore, use the appropriate size for the output file.
3386 if (blksize <= 512) {
3388 * Use f_iosize, since ZFS reports a _PC_MIN_HOLE_SIZE
3389 * of 512, although it actually only creates
3390 * unallocated regions for blocks >= f_iosize.
3392 blksize = outvp->v_mount->mnt_stat.f_iosize;
3396 * Use the larger of the two f_iosize values. If they are
3397 * not the same size, one will normally be an exact multiple of
3398 * the other, since they are both likely to be a power of 2.
3400 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3401 outvp->v_mount->mnt_stat.f_iosize);
3404 /* Clip to sane limits. */
3407 else if (blksize > maxphys)
3409 dat = malloc(blksize, M_TEMP, M_WAITOK);
3412 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3413 * to find holes. Otherwise, just scan the read block for all 0s
3414 * in the inner loop where the data copying is done.
3415 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3416 * support holes on the server, but do not support FIOSEEKHOLE.
3417 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3418 * that this function should return after 1second with a partial
3421 if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3422 getnanouptime(&endts);
3425 timespecclear(&endts);
3426 holetoeof = eof = false;
3427 while (len > 0 && error == 0 && !eof && interrupted == 0) {
3428 endoff = 0; /* To shut up compilers. */
3434 * Find the next data area. If there is just a hole to EOF,
3435 * FIOSEEKDATA should fail with ENXIO.
3436 * (I do not know if any file system will report a hole to
3437 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3438 * will fail for those file systems.)
3440 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3441 * the code just falls through to the inner copy loop.
3445 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3447 if (error == ENXIO) {
3448 startoff = endoff = inva.va_size;
3449 eof = holetoeof = true;
3453 if (error == 0 && !holetoeof) {
3455 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3458 * Since invp is unlocked, it may be possible for
3459 * another thread to do a truncate(), lseek(), write()
3460 * creating a hole at startoff between the above
3461 * VOP_IOCTL() calls, if the other thread does not do
3463 * If that happens, startoff == endoff and finding
3464 * the hole has failed, so set an error.
3466 if (error == 0 && startoff == endoff)
3467 error = EINVAL; /* Any error. Reset to 0. */
3470 if (startoff > *inoffp) {
3471 /* Found hole before data block. */
3472 xfer = MIN(startoff - *inoffp, len);
3473 if (*outoffp < va.va_size) {
3474 /* Must write 0s to punch hole. */
3475 xfer2 = MIN(va.va_size - *outoffp,
3477 memset(dat, 0, MIN(xfer2, blksize));
3478 error = vn_write_outvp(outvp, dat,
3479 *outoffp, xfer2, blksize, false,
3480 holeout > 0, outcred);
3483 if (error == 0 && *outoffp + xfer >
3484 va.va_size && (xfer == len || holetoeof)) {
3485 /* Grow output file (hole at end). */
3486 error = vn_write_outvp(outvp, dat,
3487 *outoffp, xfer, blksize, true,
3495 interrupted = sig_intr();
3496 if (timespecisset(&endts) &&
3498 getnanouptime(&curts);
3499 if (timespeccmp(&curts,
3507 copylen = MIN(len, endoff - startoff);
3519 * Set first xfer to end at a block boundary, so that
3520 * holes are more likely detected in the loop below via
3521 * the for all bytes 0 method.
3523 xfer -= (*inoffp % blksize);
3525 /* Loop copying the data block. */
3526 while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3529 error = vn_lock(invp, LK_SHARED);
3532 error = vn_rdwr(UIO_READ, invp, dat, xfer,
3533 startoff, UIO_SYSSPACE, IO_NODELOCKED,
3534 curthread->td_ucred, incred, &aresid,
3538 if (error == 0 && aresid > 0) {
3539 /* Stop the copy at EOF on the input file. */
3546 * Skip the write for holes past the initial EOF
3547 * of the output file, unless this is the last
3548 * write of the output file at EOF.
3550 readzeros = cantseek ? mem_iszero(dat, xfer) :
3554 if (!cantseek || *outoffp < va.va_size ||
3555 lastblock || !readzeros)
3556 error = vn_write_outvp(outvp, dat,
3557 *outoffp, xfer, blksize,
3558 readzeros && lastblock &&
3559 *outoffp >= va.va_size, false,
3568 interrupted = sig_intr();
3569 if (timespecisset(&endts) &&
3571 getnanouptime(&curts);
3572 if (timespeccmp(&curts,
3584 *lenp = savlen - len;
3590 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3594 off_t olen, ooffset;
3597 int audited_vnode1 = 0;
3601 if (vp->v_type != VREG)
3604 /* Allocating blocks may take a long time, so iterate. */
3611 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3614 error = vn_lock(vp, LK_EXCLUSIVE);
3616 vn_finished_write(mp);
3620 if (!audited_vnode1) {
3621 AUDIT_ARG_VNODE1(vp);
3626 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3629 error = VOP_ALLOCATE(vp, &offset, &len, 0,
3632 vn_finished_write(mp);
3634 if (olen + ooffset != offset + len) {
3635 panic("offset + len changed from %jx/%jx to %jx/%jx",
3636 ooffset, olen, offset, len);
3638 if (error != 0 || len == 0)
3640 KASSERT(olen > len, ("Iteration did not make progress?"));
3647 #define DIRENT_MINSIZE (sizeof(struct dirent) - (MAXNAMLEN+1) + 4)
3650 * Keep this assert as long as sizeof(struct dirent) is used as the maximum
3653 _Static_assert(_GENERIC_MAXDIRSIZ == sizeof(struct dirent),
3654 "'struct dirent' size must be a multiple of its alignment "
3655 "(see _GENERIC_DIRLEN())");
3658 * Returns successive directory entries through some caller's provided buffer.
3660 * This function automatically refills the provided buffer with calls to
3661 * VOP_READDIR() (after MAC permission checks).
3663 * 'td' is used for credentials and passed to uiomove(). 'dirbuf' is the
3664 * caller's buffer to fill and 'dirbuflen' its allocated size. 'dirbuf' must
3665 * be properly aligned to access 'struct dirent' structures and 'dirbuflen'
3666 * must be greater than GENERIC_MAXDIRSIZ to avoid VOP_READDIR() returning
3667 * EINVAL (the latter is not a strong guarantee (yet); but EINVAL will always
3668 * be returned if this requirement is not verified). '*dpp' points to the
3669 * current directory entry in the buffer and '*len' contains the remaining
3670 * valid bytes in 'dirbuf' after 'dpp' (including the pointed entry).
3672 * At first call (or when restarting the read), '*len' must have been set to 0,
3673 * '*off' to 0 (or any valid start offset) and '*eofflag' to 0. There are no
3674 * more entries as soon as '*len' is 0 after a call that returned 0. Calling
3675 * again this function after such a condition is considered an error and EINVAL
3676 * will be returned. Other possible error codes are those of VOP_READDIR(),
3677 * EINTEGRITY if the returned entries do not pass coherency tests, or EINVAL
3678 * (bad call). All errors are unrecoverable, i.e., the state ('*len', '*off'
3679 * and '*eofflag') must be re-initialized before a subsequent call. On error
3680 * or at end of directory, '*dpp' is reset to NULL.
3682 * '*len', '*off' and '*eofflag' are internal state the caller should not
3683 * tamper with except as explained above. '*off' is the next directory offset
3684 * to read from to refill the buffer. '*eofflag' is set to 0 or 1 by the last
3685 * internal call to VOP_READDIR() that returned without error, indicating
3686 * whether it reached the end of the directory, and to 2 by this function after
3687 * all entries have been read.
3690 vn_dir_next_dirent(struct vnode *vp, struct thread *td,
3691 char *dirbuf, size_t dirbuflen,
3692 struct dirent **dpp, size_t *len, off_t *off, int *eofflag)
3694 struct dirent *dp = NULL;
3700 ASSERT_VOP_LOCKED(vp, "vnode not locked");
3701 VNASSERT(vp->v_type == VDIR, vp, ("vnode is not a directory"));
3702 MPASS2((uintptr_t)dirbuf < (uintptr_t)dirbuf + dirbuflen,
3703 "Address space overflow");
3705 if (__predict_false(dirbuflen < GENERIC_MAXDIRSIZ)) {
3706 /* Don't take any chances in this case */
3715 * The caller continued to call us after an error (we set dp to
3716 * NULL in a previous iteration). Bail out right now.
3718 if (__predict_false(dp == NULL))
3721 MPASS(*len <= dirbuflen);
3722 MPASS2((uintptr_t)dirbuf <= (uintptr_t)dp &&
3723 (uintptr_t)dp + *len <= (uintptr_t)dirbuf + dirbuflen,
3724 "Filled range not inside buffer");
3726 reclen = dp->d_reclen;
3727 if (reclen >= *len) {
3728 /* End of buffer reached */
3731 dp = (struct dirent *)((char *)dp + reclen);
3739 /* Have to refill. */
3745 /* Nothing more to read. */
3746 *eofflag = 2; /* Remember the caller reached EOF. */
3750 /* The caller didn't test for EOF. */
3755 iov.iov_base = dirbuf;
3756 iov.iov_len = dirbuflen;
3760 uio.uio_offset = *off;
3761 uio.uio_resid = dirbuflen;
3762 uio.uio_segflg = UIO_SYSSPACE;
3763 uio.uio_rw = UIO_READ;
3767 error = mac_vnode_check_readdir(td->td_ucred, vp);
3770 error = VOP_READDIR(vp, &uio, td->td_ucred, eofflag,
3775 *len = dirbuflen - uio.uio_resid;
3776 *off = uio.uio_offset;
3779 /* Sanity check on INVARIANTS. */
3780 MPASS(*eofflag != 0);
3786 * Normalize the flag returned by VOP_READDIR(), since we use 2
3787 * as a sentinel value.
3792 dp = (struct dirent *)dirbuf;
3795 if (__predict_false(*len < GENERIC_MINDIRSIZ ||
3796 dp->d_reclen < GENERIC_MINDIRSIZ)) {
3810 * Checks whether a directory is empty or not.
3812 * If the directory is empty, returns 0, and if it is not, ENOTEMPTY. Other
3813 * values are genuine errors preventing the check.
3816 vn_dir_check_empty(struct vnode *vp)
3818 struct thread *const td = curthread;
3820 size_t dirbuflen, len;
3826 ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
3827 VNPASS(vp->v_type == VDIR, vp);
3829 error = VOP_GETATTR(vp, &va, td->td_ucred);
3833 dirbuflen = max(DEV_BSIZE, GENERIC_MAXDIRSIZ);
3834 if (dirbuflen < va.va_blocksize)
3835 dirbuflen = va.va_blocksize;
3836 dirbuf = malloc(dirbuflen, M_TEMP, M_WAITOK);
3843 error = vn_dir_next_dirent(vp, td, dirbuf, dirbuflen,
3844 &dp, &len, &off, &eofflag);
3855 * Skip whiteouts. Unionfs operates on filesystems only and
3856 * not on hierarchies, so these whiteouts would be shadowed on
3857 * the system hierarchy but not for a union using the
3858 * filesystem of their directories as the upper layer.
3859 * Additionally, unionfs currently transparently exposes
3860 * union-specific metadata of its upper layer, meaning that
3861 * whiteouts can be seen through the union view in empty
3862 * directories. Taking into account these whiteouts would then
3863 * prevent mounting another filesystem on such effectively
3864 * empty directories.
3866 if (dp->d_type == DT_WHT)
3870 * Any file in the directory which is not '.' or '..' indicates
3871 * the directory is not empty.
3873 switch (dp->d_namlen) {
3875 if (dp->d_name[1] != '.') {
3876 /* Can't be '..' (nor '.') */
3882 if (dp->d_name[0] != '.') {
3883 /* Can't be '..' nor '.' */
3896 free(dirbuf, M_TEMP);
3901 static u_long vn_lock_pair_pause_cnt;
3902 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3903 &vn_lock_pair_pause_cnt, 0,
3904 "Count of vn_lock_pair deadlocks");
3906 u_int vn_lock_pair_pause_max;
3907 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3908 &vn_lock_pair_pause_max, 0,
3909 "Max ticks for vn_lock_pair deadlock avoidance sleep");
3912 vn_lock_pair_pause(const char *wmesg)
3914 atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3915 pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3919 * Lock pair of (possibly same) vnodes vp1, vp2, avoiding lock order
3920 * reversal. vp1_locked indicates whether vp1 is locked; if not, vp1
3921 * must be unlocked. Same for vp2 and vp2_locked. One of the vnodes
3924 * The function returns with both vnodes exclusively or shared locked,
3925 * according to corresponding lkflags, and guarantees that it does not
3926 * create lock order reversal with other threads during its execution.
3927 * Both vnodes could be unlocked temporary (and reclaimed).
3929 * If requesting shared locking, locked vnode lock must not be recursed.
3931 * Only one of LK_SHARED and LK_EXCLUSIVE must be specified.
3932 * LK_NODDLKTREAT can be optionally passed.
3934 * If vp1 == vp2, only one, most exclusive, lock is obtained on it.
3937 vn_lock_pair(struct vnode *vp1, bool vp1_locked, int lkflags1,
3938 struct vnode *vp2, bool vp2_locked, int lkflags2)
3942 MPASS(((lkflags1 & LK_SHARED) != 0) ^ ((lkflags1 & LK_EXCLUSIVE) != 0));
3943 MPASS((lkflags1 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0);
3944 MPASS(((lkflags2 & LK_SHARED) != 0) ^ ((lkflags2 & LK_EXCLUSIVE) != 0));
3945 MPASS((lkflags2 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0);
3947 if (vp1 == NULL && vp2 == NULL)
3951 MPASS(vp1_locked == vp2_locked);
3953 /* Select the most exclusive mode for lock. */
3954 if ((lkflags1 & LK_TYPE_MASK) != (lkflags2 & LK_TYPE_MASK))
3955 lkflags1 = (lkflags1 & ~LK_SHARED) | LK_EXCLUSIVE;
3958 ASSERT_VOP_LOCKED(vp1, "vp1");
3960 /* No need to relock if any lock is exclusive. */
3961 if ((vp1->v_vnlock->lock_object.lo_flags &
3965 locked1 = VOP_ISLOCKED(vp1);
3966 if (((lkflags1 & LK_SHARED) != 0 &&
3967 locked1 != LK_EXCLUSIVE) ||
3968 ((lkflags1 & LK_EXCLUSIVE) != 0 &&
3969 locked1 == LK_EXCLUSIVE))
3974 ASSERT_VOP_UNLOCKED(vp1, "vp1");
3975 vn_lock(vp1, lkflags1 | LK_RETRY);
3980 if ((lkflags1 & LK_SHARED) != 0 &&
3981 (vp1->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
3982 lkflags1 = (lkflags1 & ~LK_SHARED) | LK_EXCLUSIVE;
3983 if (vp1_locked && VOP_ISLOCKED(vp1) != LK_EXCLUSIVE) {
3984 ASSERT_VOP_LOCKED(vp1, "vp1");
3985 if ((lkflags1 & LK_EXCLUSIVE) != 0) {
3987 ASSERT_VOP_UNLOCKED(vp1,
3988 "vp1 shared recursed");
3991 } else if (!vp1_locked)
3992 ASSERT_VOP_UNLOCKED(vp1, "vp1");
3998 if ((lkflags2 & LK_SHARED) != 0 &&
3999 (vp2->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
4000 lkflags2 = (lkflags2 & ~LK_SHARED) | LK_EXCLUSIVE;
4001 if (vp2_locked && VOP_ISLOCKED(vp2) != LK_EXCLUSIVE) {
4002 ASSERT_VOP_LOCKED(vp2, "vp2");
4003 if ((lkflags2 & LK_EXCLUSIVE) != 0) {
4005 ASSERT_VOP_UNLOCKED(vp2,
4006 "vp2 shared recursed");
4009 } else if (!vp2_locked)
4010 ASSERT_VOP_UNLOCKED(vp2, "vp2");
4015 if (!vp1_locked && !vp2_locked) {
4016 vn_lock(vp1, lkflags1 | LK_RETRY);
4020 while (!vp1_locked || !vp2_locked) {
4021 if (vp1_locked && vp2 != NULL) {
4023 error = VOP_LOCK1(vp2, lkflags2 | LK_NOWAIT,
4024 __FILE__, __LINE__);
4029 vn_lock_pair_pause("vlp1");
4031 vn_lock(vp2, lkflags2 | LK_RETRY);
4034 if (vp2_locked && vp1 != NULL) {
4036 error = VOP_LOCK1(vp1, lkflags1 | LK_NOWAIT,
4037 __FILE__, __LINE__);
4042 vn_lock_pair_pause("vlp2");
4044 vn_lock(vp1, lkflags1 | LK_RETRY);
4049 if (lkflags1 == LK_EXCLUSIVE)
4050 ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
4052 ASSERT_VOP_LOCKED(vp1, "vp1 ret");
4055 if (lkflags2 == LK_EXCLUSIVE)
4056 ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
4058 ASSERT_VOP_LOCKED(vp2, "vp2 ret");
4063 vn_lktype_write(struct mount *mp, struct vnode *vp)
4065 if (MNT_SHARED_WRITES(mp) ||
4066 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
4068 return (LK_EXCLUSIVE);