FreeBSD/Linux Kernel Cross Reference
sys/kern/vfs_vnops.c
1 /*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
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.
11 *
12 * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13 * Copyright (c) 2013, 2014 The FreeBSD Foundation
14 *
15 * Portions of this software were developed by Konstantin Belousov
16 * under sponsorship from the FreeBSD Foundation.
17 *
18 * Redistribution and use in source and binary forms, with or without
19 * modification, are permitted provided that the following conditions
20 * are met:
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.
29 *
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
40 * SUCH DAMAGE.
41 *
42 * @(#)vfs_vnops.c 8.2 (Berkeley) 1/21/94
43 */
44
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47
48 #include "opt_hwpmc_hooks.h"
49
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/disk.h>
53 #include <sys/fail.h>
54 #include <sys/fcntl.h>
55 #include <sys/file.h>
56 #include <sys/kdb.h>
57 #include <sys/ktr.h>
58 #include <sys/stat.h>
59 #include <sys/priv.h>
60 #include <sys/proc.h>
61 #include <sys/limits.h>
62 #include <sys/lock.h>
63 #include <sys/mman.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
68 #include <sys/bio.h>
69 #include <sys/buf.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
73 #include <sys/prng.h>
74 #include <sys/sx.h>
75 #include <sys/sleepqueue.h>
76 #include <sys/sysctl.h>
77 #include <sys/ttycom.h>
78 #include <sys/conf.h>
79 #include <sys/syslog.h>
80 #include <sys/unistd.h>
81 #include <sys/user.h>
82 #include <sys/ktrace.h>
83
84 #include <security/audit/audit.h>
85 #include <security/mac/mac_framework.h>
86
87 #include <vm/vm.h>
88 #include <vm/vm_extern.h>
89 #include <vm/pmap.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_pager.h>
94
95 #ifdef HWPMC_HOOKS
96 #include <sys/pmckern.h>
97 #endif
98
99 static fo_rdwr_t vn_read;
100 static fo_rdwr_t vn_write;
101 static fo_rdwr_t vn_io_fault;
102 static fo_truncate_t vn_truncate;
103 static fo_ioctl_t vn_ioctl;
104 static fo_poll_t vn_poll;
105 static fo_kqfilter_t vn_kqfilter;
106 static fo_close_t vn_closefile;
107 static fo_mmap_t vn_mmap;
108 static fo_fallocate_t vn_fallocate;
109
110 struct fileops vnops = {
111 .fo_read = vn_io_fault,
112 .fo_write = vn_io_fault,
113 .fo_truncate = vn_truncate,
114 .fo_ioctl = vn_ioctl,
115 .fo_poll = vn_poll,
116 .fo_kqfilter = vn_kqfilter,
117 .fo_stat = vn_statfile,
118 .fo_close = vn_closefile,
119 .fo_chmod = vn_chmod,
120 .fo_chown = vn_chown,
121 .fo_sendfile = vn_sendfile,
122 .fo_seek = vn_seek,
123 .fo_fill_kinfo = vn_fill_kinfo,
124 .fo_mmap = vn_mmap,
125 .fo_fallocate = vn_fallocate,
126 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
127 };
128
129 const u_int io_hold_cnt = 16;
130 static int vn_io_fault_enable = 1;
131 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
132 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
133 static int vn_io_fault_prefault = 0;
134 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
135 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
136 static int vn_io_pgcache_read_enable = 1;
137 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
138 &vn_io_pgcache_read_enable, 0,
139 "Enable copying from page cache for reads, avoiding fs");
140 static u_long vn_io_faults_cnt;
141 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
142 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
143
144 static int vfs_allow_read_dir = 0;
145 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
146 &vfs_allow_read_dir, 0,
147 "Enable read(2) of directory by root for filesystems that support it");
148
149 /*
150 * Returns true if vn_io_fault mode of handling the i/o request should
151 * be used.
152 */
153 static bool
154 do_vn_io_fault(struct vnode *vp, struct uio *uio)
155 {
156 struct mount *mp;
157
158 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
159 (mp = vp->v_mount) != NULL &&
160 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
161 }
162
163 /*
164 * Structure used to pass arguments to vn_io_fault1(), to do either
165 * file- or vnode-based I/O calls.
166 */
167 struct vn_io_fault_args {
168 enum {
169 VN_IO_FAULT_FOP,
170 VN_IO_FAULT_VOP
171 } kind;
172 struct ucred *cred;
173 int flags;
174 union {
175 struct fop_args_tag {
176 struct file *fp;
177 fo_rdwr_t *doio;
178 } fop_args;
179 struct vop_args_tag {
180 struct vnode *vp;
181 } vop_args;
182 } args;
183 };
184
185 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
186 struct vn_io_fault_args *args, struct thread *td);
187
188 int
189 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
190 {
191 struct thread *td = ndp->ni_cnd.cn_thread;
192
193 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
194 }
195
196 static uint64_t
197 open2nameif(int fmode, u_int vn_open_flags)
198 {
199 uint64_t res;
200
201 res = ISOPEN | LOCKLEAF;
202 if ((fmode & O_RESOLVE_BENEATH) != 0)
203 res |= RBENEATH;
204 if ((fmode & O_EMPTY_PATH) != 0)
205 res |= EMPTYPATH;
206 if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
207 res |= AUDITVNODE1;
208 if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
209 res |= NOCAPCHECK;
210 return (res);
211 }
212
213 /*
214 * Common code for vnode open operations via a name lookup.
215 * Lookup the vnode and invoke VOP_CREATE if needed.
216 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
217 *
218 * Note that this does NOT free nameidata for the successful case,
219 * due to the NDINIT being done elsewhere.
220 */
221 int
222 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
223 struct ucred *cred, struct file *fp)
224 {
225 struct vnode *vp;
226 struct mount *mp;
227 struct thread *td = ndp->ni_cnd.cn_thread;
228 struct vattr vat;
229 struct vattr *vap = &vat;
230 int fmode, error;
231 bool first_open;
232
233 restart:
234 first_open = false;
235 fmode = *flagp;
236 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
237 O_EXCL | O_DIRECTORY) ||
238 (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
239 return (EINVAL);
240 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
241 ndp->ni_cnd.cn_nameiop = CREATE;
242 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
243 /*
244 * Set NOCACHE to avoid flushing the cache when
245 * rolling in many files at once.
246 *
247 * Set NC_KEEPPOSENTRY to keep positive entries if they already
248 * exist despite NOCACHE.
249 */
250 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
251 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
252 ndp->ni_cnd.cn_flags |= FOLLOW;
253 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
254 bwillwrite();
255 if ((error = namei(ndp)) != 0)
256 return (error);
257 if (ndp->ni_vp == NULL) {
258 VATTR_NULL(vap);
259 vap->va_type = VREG;
260 vap->va_mode = cmode;
261 if (fmode & O_EXCL)
262 vap->va_vaflags |= VA_EXCLUSIVE;
263 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
264 NDFREE(ndp, NDF_ONLY_PNBUF);
265 vput(ndp->ni_dvp);
266 if ((error = vn_start_write(NULL, &mp,
267 V_XSLEEP | PCATCH)) != 0)
268 return (error);
269 NDREINIT(ndp);
270 goto restart;
271 }
272 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
273 ndp->ni_cnd.cn_flags |= MAKEENTRY;
274 #ifdef MAC
275 error = mac_vnode_check_create(cred, ndp->ni_dvp,
276 &ndp->ni_cnd, vap);
277 if (error == 0)
278 #endif
279 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
280 &ndp->ni_cnd, vap);
281 vp = ndp->ni_vp;
282 if (error == 0 && (fmode & O_EXCL) != 0 &&
283 (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
284 VI_LOCK(vp);
285 vp->v_iflag |= VI_FOPENING;
286 VI_UNLOCK(vp);
287 first_open = true;
288 }
289 VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
290 false);
291 vn_finished_write(mp);
292 if (error) {
293 NDFREE(ndp, NDF_ONLY_PNBUF);
294 if (error == ERELOOKUP) {
295 NDREINIT(ndp);
296 goto restart;
297 }
298 return (error);
299 }
300 fmode &= ~O_TRUNC;
301 } else {
302 if (ndp->ni_dvp == ndp->ni_vp)
303 vrele(ndp->ni_dvp);
304 else
305 vput(ndp->ni_dvp);
306 ndp->ni_dvp = NULL;
307 vp = ndp->ni_vp;
308 if (fmode & O_EXCL) {
309 error = EEXIST;
310 goto bad;
311 }
312 if (vp->v_type == VDIR) {
313 error = EISDIR;
314 goto bad;
315 }
316 fmode &= ~O_CREAT;
317 }
318 } else {
319 ndp->ni_cnd.cn_nameiop = LOOKUP;
320 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
321 ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
322 FOLLOW;
323 if ((fmode & FWRITE) == 0)
324 ndp->ni_cnd.cn_flags |= LOCKSHARED;
325 if ((error = namei(ndp)) != 0)
326 return (error);
327 vp = ndp->ni_vp;
328 }
329 error = vn_open_vnode(vp, fmode, cred, td, fp);
330 if (first_open) {
331 VI_LOCK(vp);
332 vp->v_iflag &= ~VI_FOPENING;
333 wakeup(vp);
334 VI_UNLOCK(vp);
335 }
336 if (error)
337 goto bad;
338 *flagp = fmode;
339 return (0);
340 bad:
341 NDFREE(ndp, NDF_ONLY_PNBUF);
342 vput(vp);
343 *flagp = fmode;
344 ndp->ni_vp = NULL;
345 return (error);
346 }
347
348 static int
349 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
350 {
351 struct flock lf;
352 int error, lock_flags, type;
353
354 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
355 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
356 return (0);
357 KASSERT(fp != NULL, ("open with flock requires fp"));
358 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
359 return (EOPNOTSUPP);
360
361 lock_flags = VOP_ISLOCKED(vp);
362 VOP_UNLOCK(vp);
363
364 lf.l_whence = SEEK_SET;
365 lf.l_start = 0;
366 lf.l_len = 0;
367 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
368 type = F_FLOCK;
369 if ((fmode & FNONBLOCK) == 0)
370 type |= F_WAIT;
371 if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
372 type |= F_FIRSTOPEN;
373 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
374 if (error == 0)
375 fp->f_flag |= FHASLOCK;
376
377 vn_lock(vp, lock_flags | LK_RETRY);
378 return (error);
379 }
380
381 /*
382 * Common code for vnode open operations once a vnode is located.
383 * Check permissions, and call the VOP_OPEN routine.
384 */
385 int
386 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
387 struct thread *td, struct file *fp)
388 {
389 accmode_t accmode;
390 int error;
391
392 if (vp->v_type == VLNK) {
393 if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
394 return (EMLINK);
395 }
396 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
397 return (ENOTDIR);
398
399 accmode = 0;
400 if ((fmode & O_PATH) == 0) {
401 if (vp->v_type == VSOCK)
402 return (EOPNOTSUPP);
403 if ((fmode & (FWRITE | O_TRUNC)) != 0) {
404 if (vp->v_type == VDIR)
405 return (EISDIR);
406 accmode |= VWRITE;
407 }
408 if ((fmode & FREAD) != 0)
409 accmode |= VREAD;
410 if ((fmode & O_APPEND) && (fmode & FWRITE))
411 accmode |= VAPPEND;
412 #ifdef MAC
413 if ((fmode & O_CREAT) != 0)
414 accmode |= VCREAT;
415 #endif
416 }
417 if ((fmode & FEXEC) != 0)
418 accmode |= VEXEC;
419 #ifdef MAC
420 if ((fmode & O_VERIFY) != 0)
421 accmode |= VVERIFY;
422 error = mac_vnode_check_open(cred, vp, accmode);
423 if (error != 0)
424 return (error);
425
426 accmode &= ~(VCREAT | VVERIFY);
427 #endif
428 if ((fmode & O_CREAT) == 0 && accmode != 0) {
429 error = VOP_ACCESS(vp, accmode, cred, td);
430 if (error != 0)
431 return (error);
432 }
433 if ((fmode & O_PATH) != 0) {
434 if (vp->v_type != VFIFO && vp->v_type != VSOCK &&
435 VOP_ACCESS(vp, VREAD, cred, td) == 0)
436 fp->f_flag |= FKQALLOWED;
437 return (0);
438 }
439
440 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
441 vn_lock(vp, LK_UPGRADE | LK_RETRY);
442 error = VOP_OPEN(vp, fmode, cred, td, fp);
443 if (error != 0)
444 return (error);
445
446 error = vn_open_vnode_advlock(vp, fmode, fp);
447 if (error == 0 && (fmode & FWRITE) != 0) {
448 error = VOP_ADD_WRITECOUNT(vp, 1);
449 if (error == 0) {
450 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
451 __func__, vp, vp->v_writecount);
452 }
453 }
454
455 /*
456 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
457 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
458 */
459 if (error != 0) {
460 if (fp != NULL) {
461 /*
462 * Arrange the call by having fdrop() to use
463 * vn_closefile(). This is to satisfy
464 * filesystems like devfs or tmpfs, which
465 * override fo_close().
466 */
467 fp->f_flag |= FOPENFAILED;
468 fp->f_vnode = vp;
469 if (fp->f_ops == &badfileops) {
470 fp->f_type = DTYPE_VNODE;
471 fp->f_ops = &vnops;
472 }
473 vref(vp);
474 } else {
475 /*
476 * If there is no fp, due to kernel-mode open,
477 * we can call VOP_CLOSE() now.
478 */
479 if (vp->v_type != VFIFO && (fmode & FWRITE) != 0 &&
480 !MNT_EXTENDED_SHARED(vp->v_mount) &&
481 VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
482 vn_lock(vp, LK_UPGRADE | LK_RETRY);
483 (void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
484 cred, td);
485 }
486 }
487
488 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
489 return (error);
490
491 }
492
493 /*
494 * Check for write permissions on the specified vnode.
495 * Prototype text segments cannot be written.
496 * It is racy.
497 */
498 int
499 vn_writechk(struct vnode *vp)
500 {
501
502 ASSERT_VOP_LOCKED(vp, "vn_writechk");
503 /*
504 * If there's shared text associated with
505 * the vnode, try to free it up once. If
506 * we fail, we can't allow writing.
507 */
508 if (VOP_IS_TEXT(vp))
509 return (ETXTBSY);
510
511 return (0);
512 }
513
514 /*
515 * Vnode close call
516 */
517 static int
518 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
519 struct thread *td, bool keep_ref)
520 {
521 struct mount *mp;
522 int error, lock_flags;
523
524 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
525 MNT_EXTENDED_SHARED(vp->v_mount))
526 lock_flags = LK_SHARED;
527 else
528 lock_flags = LK_EXCLUSIVE;
529
530 vn_start_write(vp, &mp, V_WAIT);
531 vn_lock(vp, lock_flags | LK_RETRY);
532 AUDIT_ARG_VNODE1(vp);
533 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
534 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
535 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
536 __func__, vp, vp->v_writecount);
537 }
538 error = VOP_CLOSE(vp, flags, file_cred, td);
539 if (keep_ref)
540 VOP_UNLOCK(vp);
541 else
542 vput(vp);
543 vn_finished_write(mp);
544 return (error);
545 }
546
547 int
548 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
549 struct thread *td)
550 {
551
552 return (vn_close1(vp, flags, file_cred, td, false));
553 }
554
555 /*
556 * Heuristic to detect sequential operation.
557 */
558 static int
559 sequential_heuristic(struct uio *uio, struct file *fp)
560 {
561 enum uio_rw rw;
562
563 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
564
565 rw = uio->uio_rw;
566 if (fp->f_flag & FRDAHEAD)
567 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
568
569 /*
570 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
571 * that the first I/O is normally considered to be slightly
572 * sequential. Seeking to offset 0 doesn't change sequentiality
573 * unless previous seeks have reduced f_seqcount to 0, in which
574 * case offset 0 is not special.
575 */
576 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
577 uio->uio_offset == fp->f_nextoff[rw]) {
578 /*
579 * f_seqcount is in units of fixed-size blocks so that it
580 * depends mainly on the amount of sequential I/O and not
581 * much on the number of sequential I/O's. The fixed size
582 * of 16384 is hard-coded here since it is (not quite) just
583 * a magic size that works well here. This size is more
584 * closely related to the best I/O size for real disks than
585 * to any block size used by software.
586 */
587 if (uio->uio_resid >= IO_SEQMAX * 16384)
588 fp->f_seqcount[rw] = IO_SEQMAX;
589 else {
590 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
591 if (fp->f_seqcount[rw] > IO_SEQMAX)
592 fp->f_seqcount[rw] = IO_SEQMAX;
593 }
594 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
595 }
596
597 /* Not sequential. Quickly draw-down sequentiality. */
598 if (fp->f_seqcount[rw] > 1)
599 fp->f_seqcount[rw] = 1;
600 else
601 fp->f_seqcount[rw] = 0;
602 return (0);
603 }
604
605 /*
606 * Package up an I/O request on a vnode into a uio and do it.
607 */
608 int
609 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
610 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
611 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
612 {
613 struct uio auio;
614 struct iovec aiov;
615 struct mount *mp;
616 struct ucred *cred;
617 void *rl_cookie;
618 struct vn_io_fault_args args;
619 int error, lock_flags;
620
621 if (offset < 0 && vp->v_type != VCHR)
622 return (EINVAL);
623 auio.uio_iov = &aiov;
624 auio.uio_iovcnt = 1;
625 aiov.iov_base = base;
626 aiov.iov_len = len;
627 auio.uio_resid = len;
628 auio.uio_offset = offset;
629 auio.uio_segflg = segflg;
630 auio.uio_rw = rw;
631 auio.uio_td = td;
632 error = 0;
633
634 if ((ioflg & IO_NODELOCKED) == 0) {
635 if ((ioflg & IO_RANGELOCKED) == 0) {
636 if (rw == UIO_READ) {
637 rl_cookie = vn_rangelock_rlock(vp, offset,
638 offset + len);
639 } else if ((ioflg & IO_APPEND) != 0) {
640 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
641 } else {
642 rl_cookie = vn_rangelock_wlock(vp, offset,
643 offset + len);
644 }
645 } else
646 rl_cookie = NULL;
647 mp = NULL;
648 if (rw == UIO_WRITE) {
649 if (vp->v_type != VCHR &&
650 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
651 != 0)
652 goto out;
653 lock_flags = vn_lktype_write(mp, vp);
654 } else
655 lock_flags = LK_SHARED;
656 vn_lock(vp, lock_flags | LK_RETRY);
657 } else
658 rl_cookie = NULL;
659
660 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
661 #ifdef MAC
662 if ((ioflg & IO_NOMACCHECK) == 0) {
663 if (rw == UIO_READ)
664 error = mac_vnode_check_read(active_cred, file_cred,
665 vp);
666 else
667 error = mac_vnode_check_write(active_cred, file_cred,
668 vp);
669 }
670 #endif
671 if (error == 0) {
672 if (file_cred != NULL)
673 cred = file_cred;
674 else
675 cred = active_cred;
676 if (do_vn_io_fault(vp, &auio)) {
677 args.kind = VN_IO_FAULT_VOP;
678 args.cred = cred;
679 args.flags = ioflg;
680 args.args.vop_args.vp = vp;
681 error = vn_io_fault1(vp, &auio, &args, td);
682 } else if (rw == UIO_READ) {
683 error = VOP_READ(vp, &auio, ioflg, cred);
684 } else /* if (rw == UIO_WRITE) */ {
685 error = VOP_WRITE(vp, &auio, ioflg, cred);
686 }
687 }
688 if (aresid)
689 *aresid = auio.uio_resid;
690 else
691 if (auio.uio_resid && error == 0)
692 error = EIO;
693 if ((ioflg & IO_NODELOCKED) == 0) {
694 VOP_UNLOCK(vp);
695 if (mp != NULL)
696 vn_finished_write(mp);
697 }
698 out:
699 if (rl_cookie != NULL)
700 vn_rangelock_unlock(vp, rl_cookie);
701 return (error);
702 }
703
704 /*
705 * Package up an I/O request on a vnode into a uio and do it. The I/O
706 * request is split up into smaller chunks and we try to avoid saturating
707 * the buffer cache while potentially holding a vnode locked, so we
708 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
709 * to give other processes a chance to lock the vnode (either other processes
710 * core'ing the same binary, or unrelated processes scanning the directory).
711 */
712 int
713 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
714 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
715 struct ucred *file_cred, size_t *aresid, struct thread *td)
716 {
717 int error = 0;
718 ssize_t iaresid;
719
720 do {
721 int chunk;
722
723 /*
724 * Force `offset' to a multiple of MAXBSIZE except possibly
725 * for the first chunk, so that filesystems only need to
726 * write full blocks except possibly for the first and last
727 * chunks.
728 */
729 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
730
731 if (chunk > len)
732 chunk = len;
733 if (rw != UIO_READ && vp->v_type == VREG)
734 bwillwrite();
735 iaresid = 0;
736 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
737 ioflg, active_cred, file_cred, &iaresid, td);
738 len -= chunk; /* aresid calc already includes length */
739 if (error)
740 break;
741 offset += chunk;
742 base = (char *)base + chunk;
743 kern_yield(PRI_USER);
744 } while (len);
745 if (aresid)
746 *aresid = len + iaresid;
747 return (error);
748 }
749
750 #if OFF_MAX <= LONG_MAX
751 off_t
752 foffset_lock(struct file *fp, int flags)
753 {
754 volatile short *flagsp;
755 off_t res;
756 short state;
757
758 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
759
760 if ((flags & FOF_NOLOCK) != 0)
761 return (atomic_load_long(&fp->f_offset));
762
763 /*
764 * According to McKusick the vn lock was protecting f_offset here.
765 * It is now protected by the FOFFSET_LOCKED flag.
766 */
767 flagsp = &fp->f_vnread_flags;
768 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
769 return (atomic_load_long(&fp->f_offset));
770
771 sleepq_lock(&fp->f_vnread_flags);
772 state = atomic_load_16(flagsp);
773 for (;;) {
774 if ((state & FOFFSET_LOCKED) == 0) {
775 if (!atomic_fcmpset_acq_16(flagsp, &state,
776 FOFFSET_LOCKED))
777 continue;
778 break;
779 }
780 if ((state & FOFFSET_LOCK_WAITING) == 0) {
781 if (!atomic_fcmpset_acq_16(flagsp, &state,
782 state | FOFFSET_LOCK_WAITING))
783 continue;
784 }
785 DROP_GIANT();
786 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
787 sleepq_wait(&fp->f_vnread_flags, PUSER -1);
788 PICKUP_GIANT();
789 sleepq_lock(&fp->f_vnread_flags);
790 state = atomic_load_16(flagsp);
791 }
792 res = atomic_load_long(&fp->f_offset);
793 sleepq_release(&fp->f_vnread_flags);
794 return (res);
795 }
796
797 void
798 foffset_unlock(struct file *fp, off_t val, int flags)
799 {
800 volatile short *flagsp;
801 short state;
802
803 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
804
805 if ((flags & FOF_NOUPDATE) == 0)
806 atomic_store_long(&fp->f_offset, val);
807 if ((flags & FOF_NEXTOFF_R) != 0)
808 fp->f_nextoff[UIO_READ] = val;
809 if ((flags & FOF_NEXTOFF_W) != 0)
810 fp->f_nextoff[UIO_WRITE] = val;
811
812 if ((flags & FOF_NOLOCK) != 0)
813 return;
814
815 flagsp = &fp->f_vnread_flags;
816 state = atomic_load_16(flagsp);
817 if ((state & FOFFSET_LOCK_WAITING) == 0 &&
818 atomic_cmpset_rel_16(flagsp, state, 0))
819 return;
820
821 sleepq_lock(&fp->f_vnread_flags);
822 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
823 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
824 fp->f_vnread_flags = 0;
825 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
826 sleepq_release(&fp->f_vnread_flags);
827 }
828 #else
829 off_t
830 foffset_lock(struct file *fp, int flags)
831 {
832 struct mtx *mtxp;
833 off_t res;
834
835 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
836
837 mtxp = mtx_pool_find(mtxpool_sleep, fp);
838 mtx_lock(mtxp);
839 if ((flags & FOF_NOLOCK) == 0) {
840 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
841 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
842 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
843 "vofflock", 0);
844 }
845 fp->f_vnread_flags |= FOFFSET_LOCKED;
846 }
847 res = fp->f_offset;
848 mtx_unlock(mtxp);
849 return (res);
850 }
851
852 void
853 foffset_unlock(struct file *fp, off_t val, int flags)
854 {
855 struct mtx *mtxp;
856
857 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
858
859 mtxp = mtx_pool_find(mtxpool_sleep, fp);
860 mtx_lock(mtxp);
861 if ((flags & FOF_NOUPDATE) == 0)
862 fp->f_offset = val;
863 if ((flags & FOF_NEXTOFF_R) != 0)
864 fp->f_nextoff[UIO_READ] = val;
865 if ((flags & FOF_NEXTOFF_W) != 0)
866 fp->f_nextoff[UIO_WRITE] = val;
867 if ((flags & FOF_NOLOCK) == 0) {
868 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
869 ("Lost FOFFSET_LOCKED"));
870 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
871 wakeup(&fp->f_vnread_flags);
872 fp->f_vnread_flags = 0;
873 }
874 mtx_unlock(mtxp);
875 }
876 #endif
877
878 void
879 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
880 {
881
882 if ((flags & FOF_OFFSET) == 0)
883 uio->uio_offset = foffset_lock(fp, flags);
884 }
885
886 void
887 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
888 {
889
890 if ((flags & FOF_OFFSET) == 0)
891 foffset_unlock(fp, uio->uio_offset, flags);
892 }
893
894 static int
895 get_advice(struct file *fp, struct uio *uio)
896 {
897 struct mtx *mtxp;
898 int ret;
899
900 ret = POSIX_FADV_NORMAL;
901 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
902 return (ret);
903
904 mtxp = mtx_pool_find(mtxpool_sleep, fp);
905 mtx_lock(mtxp);
906 if (fp->f_advice != NULL &&
907 uio->uio_offset >= fp->f_advice->fa_start &&
908 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
909 ret = fp->f_advice->fa_advice;
910 mtx_unlock(mtxp);
911 return (ret);
912 }
913
914 static int
915 get_write_ioflag(struct file *fp)
916 {
917 int ioflag;
918 struct mount *mp;
919 struct vnode *vp;
920
921 ioflag = 0;
922 vp = fp->f_vnode;
923 mp = atomic_load_ptr(&vp->v_mount);
924
925 if ((fp->f_flag & O_DIRECT) != 0)
926 ioflag |= IO_DIRECT;
927
928 if ((fp->f_flag & O_FSYNC) != 0 ||
929 (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
930 ioflag |= IO_SYNC;
931
932 /*
933 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
934 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
935 * fall back to full O_SYNC behavior.
936 */
937 if ((fp->f_flag & O_DSYNC) != 0)
938 ioflag |= IO_SYNC | IO_DATASYNC;
939
940 return (ioflag);
941 }
942
943 int
944 vn_read_from_obj(struct vnode *vp, struct uio *uio)
945 {
946 vm_object_t obj;
947 vm_page_t ma[io_hold_cnt + 2];
948 off_t off, vsz;
949 ssize_t resid;
950 int error, i, j;
951
952 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
953 obj = atomic_load_ptr(&vp->v_object);
954 if (obj == NULL)
955 return (EJUSTRETURN);
956
957 /*
958 * Depends on type stability of vm_objects.
959 */
960 vm_object_pip_add(obj, 1);
961 if ((obj->flags & OBJ_DEAD) != 0) {
962 /*
963 * Note that object might be already reused from the
964 * vnode, and the OBJ_DEAD flag cleared. This is fine,
965 * we recheck for DOOMED vnode state after all pages
966 * are busied, and retract then.
967 *
968 * But we check for OBJ_DEAD to ensure that we do not
969 * busy pages while vm_object_terminate_pages()
970 * processes the queue.
971 */
972 error = EJUSTRETURN;
973 goto out_pip;
974 }
975
976 resid = uio->uio_resid;
977 off = uio->uio_offset;
978 for (i = 0; resid > 0; i++) {
979 MPASS(i < io_hold_cnt + 2);
980 ma[i] = vm_page_grab_unlocked(obj, atop(off),
981 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
982 VM_ALLOC_NOWAIT);
983 if (ma[i] == NULL)
984 break;
985
986 /*
987 * Skip invalid pages. Valid mask can be partial only
988 * at EOF, and we clip later.
989 */
990 if (vm_page_none_valid(ma[i])) {
991 vm_page_sunbusy(ma[i]);
992 break;
993 }
994
995 resid -= PAGE_SIZE;
996 off += PAGE_SIZE;
997 }
998 if (i == 0) {
999 error = EJUSTRETURN;
1000 goto out_pip;
1001 }
1002
1003 /*
1004 * Check VIRF_DOOMED after we busied our pages. Since
1005 * vgonel() terminates the vnode' vm_object, it cannot
1006 * process past pages busied by us.
1007 */
1008 if (VN_IS_DOOMED(vp)) {
1009 error = EJUSTRETURN;
1010 goto out;
1011 }
1012
1013 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1014 if (resid > uio->uio_resid)
1015 resid = uio->uio_resid;
1016
1017 /*
1018 * Unlocked read of vnp_size is safe because truncation cannot
1019 * pass busied page. But we load vnp_size into a local
1020 * variable so that possible concurrent extension does not
1021 * break calculation.
1022 */
1023 #if defined(__powerpc__) && !defined(__powerpc64__)
1024 vsz = obj->un_pager.vnp.vnp_size;
1025 #else
1026 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1027 #endif
1028 if (uio->uio_offset >= vsz) {
1029 error = EJUSTRETURN;
1030 goto out;
1031 }
1032 if (uio->uio_offset + resid > vsz)
1033 resid = vsz - uio->uio_offset;
1034
1035 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1036
1037 out:
1038 for (j = 0; j < i; j++) {
1039 if (error == 0)
1040 vm_page_reference(ma[j]);
1041 vm_page_sunbusy(ma[j]);
1042 }
1043 out_pip:
1044 vm_object_pip_wakeup(obj);
1045 if (error != 0)
1046 return (error);
1047 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1048 }
1049
1050 /*
1051 * File table vnode read routine.
1052 */
1053 static int
1054 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1055 struct thread *td)
1056 {
1057 struct vnode *vp;
1058 off_t orig_offset;
1059 int error, ioflag;
1060 int advice;
1061
1062 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1063 uio->uio_td, td));
1064 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1065 vp = fp->f_vnode;
1066 ioflag = 0;
1067 if (fp->f_flag & FNONBLOCK)
1068 ioflag |= IO_NDELAY;
1069 if (fp->f_flag & O_DIRECT)
1070 ioflag |= IO_DIRECT;
1071
1072 /*
1073 * Try to read from page cache. VIRF_DOOMED check is racy but
1074 * allows us to avoid unneeded work outright.
1075 */
1076 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1077 (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1078 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1079 if (error == 0) {
1080 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1081 return (0);
1082 }
1083 if (error != EJUSTRETURN)
1084 return (error);
1085 }
1086
1087 advice = get_advice(fp, uio);
1088 vn_lock(vp, LK_SHARED | LK_RETRY);
1089
1090 switch (advice) {
1091 case POSIX_FADV_NORMAL:
1092 case POSIX_FADV_SEQUENTIAL:
1093 case POSIX_FADV_NOREUSE:
1094 ioflag |= sequential_heuristic(uio, fp);
1095 break;
1096 case POSIX_FADV_RANDOM:
1097 /* Disable read-ahead for random I/O. */
1098 break;
1099 }
1100 orig_offset = uio->uio_offset;
1101
1102 #ifdef MAC
1103 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1104 if (error == 0)
1105 #endif
1106 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1107 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1108 VOP_UNLOCK(vp);
1109 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1110 orig_offset != uio->uio_offset)
1111 /*
1112 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1113 * for the backing file after a POSIX_FADV_NOREUSE
1114 * read(2).
1115 */
1116 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1117 POSIX_FADV_DONTNEED);
1118 return (error);
1119 }
1120
1121 /*
1122 * File table vnode write routine.
1123 */
1124 static int
1125 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1126 struct thread *td)
1127 {
1128 struct vnode *vp;
1129 struct mount *mp;
1130 off_t orig_offset;
1131 int error, ioflag;
1132 int advice;
1133 bool need_finished_write;
1134
1135 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1136 uio->uio_td, td));
1137 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1138 vp = fp->f_vnode;
1139 if (vp->v_type == VREG)
1140 bwillwrite();
1141 ioflag = IO_UNIT;
1142 if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1143 ioflag |= IO_APPEND;
1144 if ((fp->f_flag & FNONBLOCK) != 0)
1145 ioflag |= IO_NDELAY;
1146 ioflag |= get_write_ioflag(fp);
1147
1148 mp = NULL;
1149 need_finished_write = false;
1150 if (vp->v_type != VCHR) {
1151 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1152 if (error != 0)
1153 goto unlock;
1154 need_finished_write = true;
1155 }
1156
1157 advice = get_advice(fp, uio);
1158
1159 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1160 switch (advice) {
1161 case POSIX_FADV_NORMAL:
1162 case POSIX_FADV_SEQUENTIAL:
1163 case POSIX_FADV_NOREUSE:
1164 ioflag |= sequential_heuristic(uio, fp);
1165 break;
1166 case POSIX_FADV_RANDOM:
1167 /* XXX: Is this correct? */
1168 break;
1169 }
1170 orig_offset = uio->uio_offset;
1171
1172 #ifdef MAC
1173 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1174 if (error == 0)
1175 #endif
1176 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1177 fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1178 VOP_UNLOCK(vp);
1179 if (need_finished_write)
1180 vn_finished_write(mp);
1181 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1182 orig_offset != uio->uio_offset)
1183 /*
1184 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1185 * for the backing file after a POSIX_FADV_NOREUSE
1186 * write(2).
1187 */
1188 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1189 POSIX_FADV_DONTNEED);
1190 unlock:
1191 return (error);
1192 }
1193
1194 /*
1195 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1196 * prevent the following deadlock:
1197 *
1198 * Assume that the thread A reads from the vnode vp1 into userspace
1199 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
1200 * currently not resident, then system ends up with the call chain
1201 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1202 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1203 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1204 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1205 * backed by the pages of vnode vp1, and some page in buf2 is not
1206 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1207 *
1208 * To prevent the lock order reversal and deadlock, vn_io_fault() does
1209 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1210 * Instead, it first tries to do the whole range i/o with pagefaults
1211 * disabled. If all pages in the i/o buffer are resident and mapped,
1212 * VOP will succeed (ignoring the genuine filesystem errors).
1213 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1214 * i/o in chunks, with all pages in the chunk prefaulted and held
1215 * using vm_fault_quick_hold_pages().
1216 *
1217 * Filesystems using this deadlock avoidance scheme should use the
1218 * array of the held pages from uio, saved in the curthread->td_ma,
1219 * instead of doing uiomove(). A helper function
1220 * vn_io_fault_uiomove() converts uiomove request into
1221 * uiomove_fromphys() over td_ma array.
1222 *
1223 * Since vnode locks do not cover the whole i/o anymore, rangelocks
1224 * make the current i/o request atomic with respect to other i/os and
1225 * truncations.
1226 */
1227
1228 /*
1229 * Decode vn_io_fault_args and perform the corresponding i/o.
1230 */
1231 static int
1232 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1233 struct thread *td)
1234 {
1235 int error, save;
1236
1237 error = 0;
1238 save = vm_fault_disable_pagefaults();
1239 switch (args->kind) {
1240 case VN_IO_FAULT_FOP:
1241 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1242 uio, args->cred, args->flags, td);
1243 break;
1244 case VN_IO_FAULT_VOP:
1245 if (uio->uio_rw == UIO_READ) {
1246 error = VOP_READ(args->args.vop_args.vp, uio,
1247 args->flags, args->cred);
1248 } else if (uio->uio_rw == UIO_WRITE) {
1249 error = VOP_WRITE(args->args.vop_args.vp, uio,
1250 args->flags, args->cred);
1251 }
1252 break;
1253 default:
1254 panic("vn_io_fault_doio: unknown kind of io %d %d",
1255 args->kind, uio->uio_rw);
1256 }
1257 vm_fault_enable_pagefaults(save);
1258 return (error);
1259 }
1260
1261 static int
1262 vn_io_fault_touch(char *base, const struct uio *uio)
1263 {
1264 int r;
1265
1266 r = fubyte(base);
1267 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1268 return (EFAULT);
1269 return (0);
1270 }
1271
1272 static int
1273 vn_io_fault_prefault_user(const struct uio *uio)
1274 {
1275 char *base;
1276 const struct iovec *iov;
1277 size_t len;
1278 ssize_t resid;
1279 int error, i;
1280
1281 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1282 ("vn_io_fault_prefault userspace"));
1283
1284 error = i = 0;
1285 iov = uio->uio_iov;
1286 resid = uio->uio_resid;
1287 base = iov->iov_base;
1288 len = iov->iov_len;
1289 while (resid > 0) {
1290 error = vn_io_fault_touch(base, uio);
1291 if (error != 0)
1292 break;
1293 if (len < PAGE_SIZE) {
1294 if (len != 0) {
1295 error = vn_io_fault_touch(base + len - 1, uio);
1296 if (error != 0)
1297 break;
1298 resid -= len;
1299 }
1300 if (++i >= uio->uio_iovcnt)
1301 break;
1302 iov = uio->uio_iov + i;
1303 base = iov->iov_base;
1304 len = iov->iov_len;
1305 } else {
1306 len -= PAGE_SIZE;
1307 base += PAGE_SIZE;
1308 resid -= PAGE_SIZE;
1309 }
1310 }
1311 return (error);
1312 }
1313
1314 /*
1315 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1316 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1317 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1318 * into args and call vn_io_fault1() to handle faults during the user
1319 * mode buffer accesses.
1320 */
1321 static int
1322 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1323 struct thread *td)
1324 {
1325 vm_page_t ma[io_hold_cnt + 2];
1326 struct uio *uio_clone, short_uio;
1327 struct iovec short_iovec[1];
1328 vm_page_t *prev_td_ma;
1329 vm_prot_t prot;
1330 vm_offset_t addr, end;
1331 size_t len, resid;
1332 ssize_t adv;
1333 int error, cnt, saveheld, prev_td_ma_cnt;
1334
1335 if (vn_io_fault_prefault) {
1336 error = vn_io_fault_prefault_user(uio);
1337 if (error != 0)
1338 return (error); /* Or ignore ? */
1339 }
1340
1341 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1342
1343 /*
1344 * The UFS follows IO_UNIT directive and replays back both
1345 * uio_offset and uio_resid if an error is encountered during the
1346 * operation. But, since the iovec may be already advanced,
1347 * uio is still in an inconsistent state.
1348 *
1349 * Cache a copy of the original uio, which is advanced to the redo
1350 * point using UIO_NOCOPY below.
1351 */
1352 uio_clone = cloneuio(uio);
1353 resid = uio->uio_resid;
1354
1355 short_uio.uio_segflg = UIO_USERSPACE;
1356 short_uio.uio_rw = uio->uio_rw;
1357 short_uio.uio_td = uio->uio_td;
1358
1359 error = vn_io_fault_doio(args, uio, td);
1360 if (error != EFAULT)
1361 goto out;
1362
1363 atomic_add_long(&vn_io_faults_cnt, 1);
1364 uio_clone->uio_segflg = UIO_NOCOPY;
1365 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1366 uio_clone->uio_segflg = uio->uio_segflg;
1367
1368 saveheld = curthread_pflags_set(TDP_UIOHELD);
1369 prev_td_ma = td->td_ma;
1370 prev_td_ma_cnt = td->td_ma_cnt;
1371
1372 while (uio_clone->uio_resid != 0) {
1373 len = uio_clone->uio_iov->iov_len;
1374 if (len == 0) {
1375 KASSERT(uio_clone->uio_iovcnt >= 1,
1376 ("iovcnt underflow"));
1377 uio_clone->uio_iov++;
1378 uio_clone->uio_iovcnt--;
1379 continue;
1380 }
1381 if (len > ptoa(io_hold_cnt))
1382 len = ptoa(io_hold_cnt);
1383 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1384 end = round_page(addr + len);
1385 if (end < addr) {
1386 error = EFAULT;
1387 break;
1388 }
1389 cnt = atop(end - trunc_page(addr));
1390 /*
1391 * A perfectly misaligned address and length could cause
1392 * both the start and the end of the chunk to use partial
1393 * page. +2 accounts for such a situation.
1394 */
1395 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1396 addr, len, prot, ma, io_hold_cnt + 2);
1397 if (cnt == -1) {
1398 error = EFAULT;
1399 break;
1400 }
1401 short_uio.uio_iov = &short_iovec[0];
1402 short_iovec[0].iov_base = (void *)addr;
1403 short_uio.uio_iovcnt = 1;
1404 short_uio.uio_resid = short_iovec[0].iov_len = len;
1405 short_uio.uio_offset = uio_clone->uio_offset;
1406 td->td_ma = ma;
1407 td->td_ma_cnt = cnt;
1408
1409 error = vn_io_fault_doio(args, &short_uio, td);
1410 vm_page_unhold_pages(ma, cnt);
1411 adv = len - short_uio.uio_resid;
1412
1413 uio_clone->uio_iov->iov_base =
1414 (char *)uio_clone->uio_iov->iov_base + adv;
1415 uio_clone->uio_iov->iov_len -= adv;
1416 uio_clone->uio_resid -= adv;
1417 uio_clone->uio_offset += adv;
1418
1419 uio->uio_resid -= adv;
1420 uio->uio_offset += adv;
1421
1422 if (error != 0 || adv == 0)
1423 break;
1424 }
1425 td->td_ma = prev_td_ma;
1426 td->td_ma_cnt = prev_td_ma_cnt;
1427 curthread_pflags_restore(saveheld);
1428 out:
1429 free(uio_clone, M_IOV);
1430 return (error);
1431 }
1432
1433 static int
1434 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1435 int flags, struct thread *td)
1436 {
1437 fo_rdwr_t *doio;
1438 struct vnode *vp;
1439 void *rl_cookie;
1440 struct vn_io_fault_args args;
1441 int error;
1442
1443 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1444 vp = fp->f_vnode;
1445
1446 /*
1447 * The ability to read(2) on a directory has historically been
1448 * allowed for all users, but this can and has been the source of
1449 * at least one security issue in the past. As such, it is now hidden
1450 * away behind a sysctl for those that actually need it to use it, and
1451 * restricted to root when it's turned on to make it relatively safe to
1452 * leave on for longer sessions of need.
1453 */
1454 if (vp->v_type == VDIR) {
1455 KASSERT(uio->uio_rw == UIO_READ,
1456 ("illegal write attempted on a directory"));
1457 if (!vfs_allow_read_dir)
1458 return (EISDIR);
1459 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1460 return (EISDIR);
1461 }
1462
1463 foffset_lock_uio(fp, uio, flags);
1464 if (do_vn_io_fault(vp, uio)) {
1465 args.kind = VN_IO_FAULT_FOP;
1466 args.args.fop_args.fp = fp;
1467 args.args.fop_args.doio = doio;
1468 args.cred = active_cred;
1469 args.flags = flags | FOF_OFFSET;
1470 if (uio->uio_rw == UIO_READ) {
1471 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1472 uio->uio_offset + uio->uio_resid);
1473 } else if ((fp->f_flag & O_APPEND) != 0 ||
1474 (flags & FOF_OFFSET) == 0) {
1475 /* For appenders, punt and lock the whole range. */
1476 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1477 } else {
1478 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1479 uio->uio_offset + uio->uio_resid);
1480 }
1481 error = vn_io_fault1(vp, uio, &args, td);
1482 vn_rangelock_unlock(vp, rl_cookie);
1483 } else {
1484 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1485 }
1486 foffset_unlock_uio(fp, uio, flags);
1487 return (error);
1488 }
1489
1490 /*
1491 * Helper function to perform the requested uiomove operation using
1492 * the held pages for io->uio_iov[0].iov_base buffer instead of
1493 * copyin/copyout. Access to the pages with uiomove_fromphys()
1494 * instead of iov_base prevents page faults that could occur due to
1495 * pmap_collect() invalidating the mapping created by
1496 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1497 * object cleanup revoking the write access from page mappings.
1498 *
1499 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1500 * instead of plain uiomove().
1501 */
1502 int
1503 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1504 {
1505 struct uio transp_uio;
1506 struct iovec transp_iov[1];
1507 struct thread *td;
1508 size_t adv;
1509 int error, pgadv;
1510
1511 td = curthread;
1512 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1513 uio->uio_segflg != UIO_USERSPACE)
1514 return (uiomove(data, xfersize, uio));
1515
1516 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1517 transp_iov[0].iov_base = data;
1518 transp_uio.uio_iov = &transp_iov[0];
1519 transp_uio.uio_iovcnt = 1;
1520 if (xfersize > uio->uio_resid)
1521 xfersize = uio->uio_resid;
1522 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1523 transp_uio.uio_offset = 0;
1524 transp_uio.uio_segflg = UIO_SYSSPACE;
1525 /*
1526 * Since transp_iov points to data, and td_ma page array
1527 * corresponds to original uio->uio_iov, we need to invert the
1528 * direction of the i/o operation as passed to
1529 * uiomove_fromphys().
1530 */
1531 switch (uio->uio_rw) {
1532 case UIO_WRITE:
1533 transp_uio.uio_rw = UIO_READ;
1534 break;
1535 case UIO_READ:
1536 transp_uio.uio_rw = UIO_WRITE;
1537 break;
1538 }
1539 transp_uio.uio_td = uio->uio_td;
1540 error = uiomove_fromphys(td->td_ma,
1541 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1542 xfersize, &transp_uio);
1543 adv = xfersize - transp_uio.uio_resid;
1544 pgadv =
1545 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1546 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1547 td->td_ma += pgadv;
1548 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1549 pgadv));
1550 td->td_ma_cnt -= pgadv;
1551 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1552 uio->uio_iov->iov_len -= adv;
1553 uio->uio_resid -= adv;
1554 uio->uio_offset += adv;
1555 return (error);
1556 }
1557
1558 int
1559 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1560 struct uio *uio)
1561 {
1562 struct thread *td;
1563 vm_offset_t iov_base;
1564 int cnt, pgadv;
1565
1566 td = curthread;
1567 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1568 uio->uio_segflg != UIO_USERSPACE)
1569 return (uiomove_fromphys(ma, offset, xfersize, uio));
1570
1571 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1572 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1573 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1574 switch (uio->uio_rw) {
1575 case UIO_WRITE:
1576 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1577 offset, cnt);
1578 break;
1579 case UIO_READ:
1580 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1581 cnt);
1582 break;
1583 }
1584 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1585 td->td_ma += pgadv;
1586 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1587 pgadv));
1588 td->td_ma_cnt -= pgadv;
1589 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1590 uio->uio_iov->iov_len -= cnt;
1591 uio->uio_resid -= cnt;
1592 uio->uio_offset += cnt;
1593 return (0);
1594 }
1595
1596 /*
1597 * File table truncate routine.
1598 */
1599 static int
1600 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1601 struct thread *td)
1602 {
1603 struct mount *mp;
1604 struct vnode *vp;
1605 void *rl_cookie;
1606 int error;
1607
1608 vp = fp->f_vnode;
1609
1610 retry:
1611 /*
1612 * Lock the whole range for truncation. Otherwise split i/o
1613 * might happen partly before and partly after the truncation.
1614 */
1615 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1616 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1617 if (error)
1618 goto out1;
1619 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1620 AUDIT_ARG_VNODE1(vp);
1621 if (vp->v_type == VDIR) {
1622 error = EISDIR;
1623 goto out;
1624 }
1625 #ifdef MAC
1626 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1627 if (error)
1628 goto out;
1629 #endif
1630 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1631 fp->f_cred);
1632 out:
1633 VOP_UNLOCK(vp);
1634 vn_finished_write(mp);
1635 out1:
1636 vn_rangelock_unlock(vp, rl_cookie);
1637 if (error == ERELOOKUP)
1638 goto retry;
1639 return (error);
1640 }
1641
1642 /*
1643 * Truncate a file that is already locked.
1644 */
1645 int
1646 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1647 struct ucred *cred)
1648 {
1649 struct vattr vattr;
1650 int error;
1651
1652 error = VOP_ADD_WRITECOUNT(vp, 1);
1653 if (error == 0) {
1654 VATTR_NULL(&vattr);
1655 vattr.va_size = length;
1656 if (sync)
1657 vattr.va_vaflags |= VA_SYNC;
1658 error = VOP_SETATTR(vp, &vattr, cred);
1659 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1660 }
1661 return (error);
1662 }
1663
1664 /*
1665 * File table vnode stat routine.
1666 */
1667 int
1668 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1669 struct thread *td)
1670 {
1671 struct vnode *vp = fp->f_vnode;
1672 int error;
1673
1674 vn_lock(vp, LK_SHARED | LK_RETRY);
1675 error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1676 VOP_UNLOCK(vp);
1677
1678 return (error);
1679 }
1680
1681 /*
1682 * File table vnode ioctl routine.
1683 */
1684 static int
1685 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1686 struct thread *td)
1687 {
1688 struct vattr vattr;
1689 struct vnode *vp;
1690 struct fiobmap2_arg *bmarg;
1691 int error;
1692
1693 vp = fp->f_vnode;
1694 switch (vp->v_type) {
1695 case VDIR:
1696 case VREG:
1697 switch (com) {
1698 case FIONREAD:
1699 vn_lock(vp, LK_SHARED | LK_RETRY);
1700 error = VOP_GETATTR(vp, &vattr, active_cred);
1701 VOP_UNLOCK(vp);
1702 if (error == 0)
1703 *(int *)data = vattr.va_size - fp->f_offset;
1704 return (error);
1705 case FIOBMAP2:
1706 bmarg = (struct fiobmap2_arg *)data;
1707 vn_lock(vp, LK_SHARED | LK_RETRY);
1708 #ifdef MAC
1709 error = mac_vnode_check_read(active_cred, fp->f_cred,
1710 vp);
1711 if (error == 0)
1712 #endif
1713 error = VOP_BMAP(vp, bmarg->bn, NULL,
1714 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1715 VOP_UNLOCK(vp);
1716 return (error);
1717 case FIONBIO:
1718 case FIOASYNC:
1719 return (0);
1720 default:
1721 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1722 active_cred, td));
1723 }
1724 break;
1725 case VCHR:
1726 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1727 active_cred, td));
1728 default:
1729 return (ENOTTY);
1730 }
1731 }
1732
1733 /*
1734 * File table vnode poll routine.
1735 */
1736 static int
1737 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1738 struct thread *td)
1739 {
1740 struct vnode *vp;
1741 int error;
1742
1743 vp = fp->f_vnode;
1744 #if defined(MAC) || defined(AUDIT)
1745 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1746 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1747 AUDIT_ARG_VNODE1(vp);
1748 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1749 VOP_UNLOCK(vp);
1750 if (error != 0)
1751 return (error);
1752 }
1753 #endif
1754 error = VOP_POLL(vp, events, fp->f_cred, td);
1755 return (error);
1756 }
1757
1758 /*
1759 * Acquire the requested lock and then check for validity. LK_RETRY
1760 * permits vn_lock to return doomed vnodes.
1761 */
1762 static int __noinline
1763 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1764 int error)
1765 {
1766
1767 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1768 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1769
1770 if (error == 0)
1771 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1772
1773 if ((flags & LK_RETRY) == 0) {
1774 if (error == 0) {
1775 VOP_UNLOCK(vp);
1776 error = ENOENT;
1777 }
1778 return (error);
1779 }
1780
1781 /*
1782 * LK_RETRY case.
1783 *
1784 * Nothing to do if we got the lock.
1785 */
1786 if (error == 0)
1787 return (0);
1788
1789 /*
1790 * Interlock was dropped by the call in _vn_lock.
1791 */
1792 flags &= ~LK_INTERLOCK;
1793 do {
1794 error = VOP_LOCK1(vp, flags, file, line);
1795 } while (error != 0);
1796 return (0);
1797 }
1798
1799 int
1800 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1801 {
1802 int error;
1803
1804 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1805 ("vn_lock: no locktype (%d passed)", flags));
1806 VNPASS(vp->v_holdcnt > 0, vp);
1807 error = VOP_LOCK1(vp, flags, file, line);
1808 if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1809 return (_vn_lock_fallback(vp, flags, file, line, error));
1810 return (0);
1811 }
1812
1813 /*
1814 * File table vnode close routine.
1815 */
1816 static int
1817 vn_closefile(struct file *fp, struct thread *td)
1818 {
1819 struct vnode *vp;
1820 struct flock lf;
1821 int error;
1822 bool ref;
1823
1824 vp = fp->f_vnode;
1825 fp->f_ops = &badfileops;
1826 ref = (fp->f_flag & FHASLOCK) != 0;
1827
1828 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1829
1830 if (__predict_false(ref)) {
1831 lf.l_whence = SEEK_SET;
1832 lf.l_start = 0;
1833 lf.l_len = 0;
1834 lf.l_type = F_UNLCK;
1835 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1836 vrele(vp);
1837 }
1838 return (error);
1839 }
1840
1841 /*
1842 * Preparing to start a filesystem write operation. If the operation is
1843 * permitted, then we bump the count of operations in progress and
1844 * proceed. If a suspend request is in progress, we wait until the
1845 * suspension is over, and then proceed.
1846 */
1847 static int
1848 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1849 {
1850 struct mount_pcpu *mpcpu;
1851 int error, mflags;
1852
1853 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1854 vfs_op_thread_enter(mp, mpcpu)) {
1855 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1856 vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1857 vfs_op_thread_exit(mp, mpcpu);
1858 return (0);
1859 }
1860
1861 if (mplocked)
1862 mtx_assert(MNT_MTX(mp), MA_OWNED);
1863 else
1864 MNT_ILOCK(mp);
1865
1866 error = 0;
1867
1868 /*
1869 * Check on status of suspension.
1870 */
1871 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1872 mp->mnt_susp_owner != curthread) {
1873 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1874 (flags & PCATCH) : 0) | (PUSER - 1);
1875 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1876 if (flags & V_NOWAIT) {
1877 error = EWOULDBLOCK;
1878 goto unlock;
1879 }
1880 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1881 "suspfs", 0);
1882 if (error)
1883 goto unlock;
1884 }
1885 }
1886 if (flags & V_XSLEEP)
1887 goto unlock;
1888 mp->mnt_writeopcount++;
1889 unlock:
1890 if (error != 0 || (flags & V_XSLEEP) != 0)
1891 MNT_REL(mp);
1892 MNT_IUNLOCK(mp);
1893 return (error);
1894 }
1895
1896 int
1897 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1898 {
1899 struct mount *mp;
1900 int error;
1901
1902 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1903 ("V_MNTREF requires mp"));
1904
1905 error = 0;
1906 /*
1907 * If a vnode is provided, get and return the mount point that
1908 * to which it will write.
1909 */
1910 if (vp != NULL) {
1911 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1912 *mpp = NULL;
1913 if (error != EOPNOTSUPP)
1914 return (error);
1915 return (0);
1916 }
1917 }
1918 if ((mp = *mpp) == NULL)
1919 return (0);
1920
1921 /*
1922 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1923 * a vfs_ref().
1924 * As long as a vnode is not provided we need to acquire a
1925 * refcount for the provided mountpoint too, in order to
1926 * emulate a vfs_ref().
1927 */
1928 if (vp == NULL && (flags & V_MNTREF) == 0)
1929 vfs_ref(mp);
1930
1931 return (vn_start_write_refed(mp, flags, false));
1932 }
1933
1934 /*
1935 * Secondary suspension. Used by operations such as vop_inactive
1936 * routines that are needed by the higher level functions. These
1937 * are allowed to proceed until all the higher level functions have
1938 * completed (indicated by mnt_writeopcount dropping to zero). At that
1939 * time, these operations are halted until the suspension is over.
1940 */
1941 int
1942 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1943 {
1944 struct mount *mp;
1945 int error;
1946
1947 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1948 ("V_MNTREF requires mp"));
1949
1950 retry:
1951 if (vp != NULL) {
1952 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1953 *mpp = NULL;
1954 if (error != EOPNOTSUPP)
1955 return (error);
1956 return (0);
1957 }
1958 }
1959 /*
1960 * If we are not suspended or have not yet reached suspended
1961 * mode, then let the operation proceed.
1962 */
1963 if ((mp = *mpp) == NULL)
1964 return (0);
1965
1966 /*
1967 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1968 * a vfs_ref().
1969 * As long as a vnode is not provided we need to acquire a
1970 * refcount for the provided mountpoint too, in order to
1971 * emulate a vfs_ref().
1972 */
1973 MNT_ILOCK(mp);
1974 if (vp == NULL && (flags & V_MNTREF) == 0)
1975 MNT_REF(mp);
1976 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1977 mp->mnt_secondary_writes++;
1978 mp->mnt_secondary_accwrites++;
1979 MNT_IUNLOCK(mp);
1980 return (0);
1981 }
1982 if (flags & V_NOWAIT) {
1983 MNT_REL(mp);
1984 MNT_IUNLOCK(mp);
1985 return (EWOULDBLOCK);
1986 }
1987 /*
1988 * Wait for the suspension to finish.
1989 */
1990 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1991 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1992 "suspfs", 0);
1993 vfs_rel(mp);
1994 if (error == 0)
1995 goto retry;
1996 return (error);
1997 }
1998
1999 /*
2000 * Filesystem write operation has completed. If we are suspending and this
2001 * operation is the last one, notify the suspender that the suspension is
2002 * now in effect.
2003 */
2004 void
2005 vn_finished_write(struct mount *mp)
2006 {
2007 struct mount_pcpu *mpcpu;
2008 int c;
2009
2010 if (mp == NULL)
2011 return;
2012
2013 if (vfs_op_thread_enter(mp, mpcpu)) {
2014 vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2015 vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2016 vfs_op_thread_exit(mp, mpcpu);
2017 return;
2018 }
2019
2020 MNT_ILOCK(mp);
2021 vfs_assert_mount_counters(mp);
2022 MNT_REL(mp);
2023 c = --mp->mnt_writeopcount;
2024 if (mp->mnt_vfs_ops == 0) {
2025 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2026 MNT_IUNLOCK(mp);
2027 return;
2028 }
2029 if (c < 0)
2030 vfs_dump_mount_counters(mp);
2031 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2032 wakeup(&mp->mnt_writeopcount);
2033 MNT_IUNLOCK(mp);
2034 }
2035
2036 /*
2037 * Filesystem secondary write operation has completed. If we are
2038 * suspending and this operation is the last one, notify the suspender
2039 * that the suspension is now in effect.
2040 */
2041 void
2042 vn_finished_secondary_write(struct mount *mp)
2043 {
2044 if (mp == NULL)
2045 return;
2046 MNT_ILOCK(mp);
2047 MNT_REL(mp);
2048 mp->mnt_secondary_writes--;
2049 if (mp->mnt_secondary_writes < 0)
2050 panic("vn_finished_secondary_write: neg cnt");
2051 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2052 mp->mnt_secondary_writes <= 0)
2053 wakeup(&mp->mnt_secondary_writes);
2054 MNT_IUNLOCK(mp);
2055 }
2056
2057 /*
2058 * Request a filesystem to suspend write operations.
2059 */
2060 int
2061 vfs_write_suspend(struct mount *mp, int flags)
2062 {
2063 int error;
2064
2065 vfs_op_enter(mp);
2066
2067 MNT_ILOCK(mp);
2068 vfs_assert_mount_counters(mp);
2069 if (mp->mnt_susp_owner == curthread) {
2070 vfs_op_exit_locked(mp);
2071 MNT_IUNLOCK(mp);
2072 return (EALREADY);
2073 }
2074 while (mp->mnt_kern_flag & MNTK_SUSPEND)
2075 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2076
2077 /*
2078 * Unmount holds a write reference on the mount point. If we
2079 * own busy reference and drain for writers, we deadlock with
2080 * the reference draining in the unmount path. Callers of
2081 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2082 * vfs_busy() reference is owned and caller is not in the
2083 * unmount context.
2084 */
2085 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2086 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2087 vfs_op_exit_locked(mp);
2088 MNT_IUNLOCK(mp);
2089 return (EBUSY);
2090 }
2091
2092 mp->mnt_kern_flag |= MNTK_SUSPEND;
2093 mp->mnt_susp_owner = curthread;
2094 if (mp->mnt_writeopcount > 0)
2095 (void) msleep(&mp->mnt_writeopcount,
2096 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2097 else
2098 MNT_IUNLOCK(mp);
2099 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2100 vfs_write_resume(mp, 0);
2101 /* vfs_write_resume does vfs_op_exit() for us */
2102 }
2103 return (error);
2104 }
2105
2106 /*
2107 * Request a filesystem to resume write operations.
2108 */
2109 void
2110 vfs_write_resume(struct mount *mp, int flags)
2111 {
2112
2113 MNT_ILOCK(mp);
2114 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2115 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2116 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2117 MNTK_SUSPENDED);
2118 mp->mnt_susp_owner = NULL;
2119 wakeup(&mp->mnt_writeopcount);
2120 wakeup(&mp->mnt_flag);
2121 curthread->td_pflags &= ~TDP_IGNSUSP;
2122 if ((flags & VR_START_WRITE) != 0) {
2123 MNT_REF(mp);
2124 mp->mnt_writeopcount++;
2125 }
2126 MNT_IUNLOCK(mp);
2127 if ((flags & VR_NO_SUSPCLR) == 0)
2128 VFS_SUSP_CLEAN(mp);
2129 vfs_op_exit(mp);
2130 } else if ((flags & VR_START_WRITE) != 0) {
2131 MNT_REF(mp);
2132 vn_start_write_refed(mp, 0, true);
2133 } else {
2134 MNT_IUNLOCK(mp);
2135 }
2136 }
2137
2138 /*
2139 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2140 * methods.
2141 */
2142 int
2143 vfs_write_suspend_umnt(struct mount *mp)
2144 {
2145 int error;
2146
2147 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2148 ("vfs_write_suspend_umnt: recursed"));
2149
2150 /* dounmount() already called vn_start_write(). */
2151 for (;;) {
2152 vn_finished_write(mp);
2153 error = vfs_write_suspend(mp, 0);
2154 if (error != 0) {
2155 vn_start_write(NULL, &mp, V_WAIT);
2156 return (error);
2157 }
2158 MNT_ILOCK(mp);
2159 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2160 break;
2161 MNT_IUNLOCK(mp);
2162 vn_start_write(NULL, &mp, V_WAIT);
2163 }
2164 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2165 wakeup(&mp->mnt_flag);
2166 MNT_IUNLOCK(mp);
2167 curthread->td_pflags |= TDP_IGNSUSP;
2168 return (0);
2169 }
2170
2171 /*
2172 * Implement kqueues for files by translating it to vnode operation.
2173 */
2174 static int
2175 vn_kqfilter(struct file *fp, struct knote *kn)
2176 {
2177
2178 return (VOP_KQFILTER(fp->f_vnode, kn));
2179 }
2180
2181 int
2182 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2183 {
2184 if ((fp->f_flag & FKQALLOWED) == 0)
2185 return (EBADF);
2186 return (vn_kqfilter(fp, kn));
2187 }
2188
2189 /*
2190 * Simplified in-kernel wrapper calls for extended attribute access.
2191 * Both calls pass in a NULL credential, authorizing as "kernel" access.
2192 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2193 */
2194 int
2195 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2196 const char *attrname, int *buflen, char *buf, struct thread *td)
2197 {
2198 struct uio auio;
2199 struct iovec iov;
2200 int error;
2201
2202 iov.iov_len = *buflen;
2203 iov.iov_base = buf;
2204
2205 auio.uio_iov = &iov;
2206 auio.uio_iovcnt = 1;
2207 auio.uio_rw = UIO_READ;
2208 auio.uio_segflg = UIO_SYSSPACE;
2209 auio.uio_td = td;
2210 auio.uio_offset = 0;
2211 auio.uio_resid = *buflen;
2212
2213 if ((ioflg & IO_NODELOCKED) == 0)
2214 vn_lock(vp, LK_SHARED | LK_RETRY);
2215
2216 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2217
2218 /* authorize attribute retrieval as kernel */
2219 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2220 td);
2221
2222 if ((ioflg & IO_NODELOCKED) == 0)
2223 VOP_UNLOCK(vp);
2224
2225 if (error == 0) {
2226 *buflen = *buflen - auio.uio_resid;
2227 }
2228
2229 return (error);
2230 }
2231
2232 /*
2233 * XXX failure mode if partially written?
2234 */
2235 int
2236 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2237 const char *attrname, int buflen, char *buf, struct thread *td)
2238 {
2239 struct uio auio;
2240 struct iovec iov;
2241 struct mount *mp;
2242 int error;
2243
2244 iov.iov_len = buflen;
2245 iov.iov_base = buf;
2246
2247 auio.uio_iov = &iov;
2248 auio.uio_iovcnt = 1;
2249 auio.uio_rw = UIO_WRITE;
2250 auio.uio_segflg = UIO_SYSSPACE;
2251 auio.uio_td = td;
2252 auio.uio_offset = 0;
2253 auio.uio_resid = buflen;
2254
2255 if ((ioflg & IO_NODELOCKED) == 0) {
2256 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2257 return (error);
2258 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2259 }
2260
2261 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2262
2263 /* authorize attribute setting as kernel */
2264 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2265
2266 if ((ioflg & IO_NODELOCKED) == 0) {
2267 vn_finished_write(mp);
2268 VOP_UNLOCK(vp);
2269 }
2270
2271 return (error);
2272 }
2273
2274 int
2275 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2276 const char *attrname, struct thread *td)
2277 {
2278 struct mount *mp;
2279 int error;
2280
2281 if ((ioflg & IO_NODELOCKED) == 0) {
2282 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2283 return (error);
2284 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2285 }
2286
2287 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2288
2289 /* authorize attribute removal as kernel */
2290 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2291 if (error == EOPNOTSUPP)
2292 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2293 NULL, td);
2294
2295 if ((ioflg & IO_NODELOCKED) == 0) {
2296 vn_finished_write(mp);
2297 VOP_UNLOCK(vp);
2298 }
2299
2300 return (error);
2301 }
2302
2303 static int
2304 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2305 struct vnode **rvp)
2306 {
2307
2308 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2309 }
2310
2311 int
2312 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2313 {
2314
2315 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2316 lkflags, rvp));
2317 }
2318
2319 int
2320 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2321 int lkflags, struct vnode **rvp)
2322 {
2323 struct mount *mp;
2324 int ltype, error;
2325
2326 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2327 mp = vp->v_mount;
2328 ltype = VOP_ISLOCKED(vp);
2329 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2330 ("vn_vget_ino: vp not locked"));
2331 error = vfs_busy(mp, MBF_NOWAIT);
2332 if (error != 0) {
2333 vfs_ref(mp);
2334 VOP_UNLOCK(vp);
2335 error = vfs_busy(mp, 0);
2336 vn_lock(vp, ltype | LK_RETRY);
2337 vfs_rel(mp);
2338 if (error != 0)
2339 return (ENOENT);
2340 if (VN_IS_DOOMED(vp)) {
2341 vfs_unbusy(mp);
2342 return (ENOENT);
2343 }
2344 }
2345 VOP_UNLOCK(vp);
2346 error = alloc(mp, alloc_arg, lkflags, rvp);
2347 vfs_unbusy(mp);
2348 if (error != 0 || *rvp != vp)
2349 vn_lock(vp, ltype | LK_RETRY);
2350 if (VN_IS_DOOMED(vp)) {
2351 if (error == 0) {
2352 if (*rvp == vp)
2353 vunref(vp);
2354 else
2355 vput(*rvp);
2356 }
2357 error = ENOENT;
2358 }
2359 return (error);
2360 }
2361
2362 int
2363 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2364 struct thread *td)
2365 {
2366 off_t lim;
2367 bool ktr_write;
2368
2369 if (td == NULL)
2370 return (0);
2371
2372 /*
2373 * There are conditions where the limit is to be ignored.
2374 * However, since it is almost never reached, check it first.
2375 */
2376 ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2377 lim = lim_cur(td, RLIMIT_FSIZE);
2378 if (__predict_false(ktr_write))
2379 lim = td->td_ktr_io_lim;
2380 if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2381 return (0);
2382
2383 /*
2384 * The limit is reached.
2385 */
2386 if (vp->v_type != VREG ||
2387 (td->td_pflags2 & TDP2_ACCT) != 0)
2388 return (0);
2389
2390 if (!ktr_write || ktr_filesize_limit_signal) {
2391 PROC_LOCK(td->td_proc);
2392 kern_psignal(td->td_proc, SIGXFSZ);
2393 PROC_UNLOCK(td->td_proc);
2394 }
2395 return (EFBIG);
2396 }
2397
2398 int
2399 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2400 struct thread *td)
2401 {
2402 struct vnode *vp;
2403
2404 vp = fp->f_vnode;
2405 #ifdef AUDIT
2406 vn_lock(vp, LK_SHARED | LK_RETRY);
2407 AUDIT_ARG_VNODE1(vp);
2408 VOP_UNLOCK(vp);
2409 #endif
2410 return (setfmode(td, active_cred, vp, mode));
2411 }
2412
2413 int
2414 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2415 struct thread *td)
2416 {
2417 struct vnode *vp;
2418
2419 vp = fp->f_vnode;
2420 #ifdef AUDIT
2421 vn_lock(vp, LK_SHARED | LK_RETRY);
2422 AUDIT_ARG_VNODE1(vp);
2423 VOP_UNLOCK(vp);
2424 #endif
2425 return (setfown(td, active_cred, vp, uid, gid));
2426 }
2427
2428 /*
2429 * Remove pages in the range ["start", "end") from the vnode's VM object. If
2430 * "end" is 0, then the range extends to the end of the object.
2431 */
2432 void
2433 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2434 {
2435 vm_object_t object;
2436
2437 if ((object = vp->v_object) == NULL)
2438 return;
2439 VM_OBJECT_WLOCK(object);
2440 vm_object_page_remove(object, start, end, 0);
2441 VM_OBJECT_WUNLOCK(object);
2442 }
2443
2444 /*
2445 * Like vn_pages_remove(), but skips invalid pages, which by definition are not
2446 * mapped into any process' address space. Filesystems may use this in
2447 * preference to vn_pages_remove() to avoid blocking on pages busied in
2448 * preparation for a VOP_GETPAGES.
2449 */
2450 void
2451 vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2452 {
2453 vm_object_t object;
2454
2455 if ((object = vp->v_object) == NULL)
2456 return;
2457 VM_OBJECT_WLOCK(object);
2458 vm_object_page_remove(object, start, end, OBJPR_VALIDONLY);
2459 VM_OBJECT_WUNLOCK(object);
2460 }
2461
2462 int
2463 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2464 {
2465 struct vattr va;
2466 daddr_t bn, bnp;
2467 uint64_t bsize;
2468 off_t noff;
2469 int error;
2470
2471 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2472 ("Wrong command %lu", cmd));
2473
2474 if (vn_lock(vp, LK_SHARED) != 0)
2475 return (EBADF);
2476 if (vp->v_type != VREG) {
2477 error = ENOTTY;
2478 goto unlock;
2479 }
2480 error = VOP_GETATTR(vp, &va, cred);
2481 if (error != 0)
2482 goto unlock;
2483 noff = *off;
2484 if (noff >= va.va_size) {
2485 error = ENXIO;
2486 goto unlock;
2487 }
2488 bsize = vp->v_mount->mnt_stat.f_iosize;
2489 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2490 noff % bsize) {
2491 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2492 if (error == EOPNOTSUPP) {
2493 error = ENOTTY;
2494 goto unlock;
2495 }
2496 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2497 (bnp != -1 && cmd == FIOSEEKDATA)) {
2498 noff = bn * bsize;
2499 if (noff < *off)
2500 noff = *off;
2501 goto unlock;
2502 }
2503 }
2504 if (noff > va.va_size)
2505 noff = va.va_size;
2506 /* noff == va.va_size. There is an implicit hole at the end of file. */
2507 if (cmd == FIOSEEKDATA)
2508 error = ENXIO;
2509 unlock:
2510 VOP_UNLOCK(vp);
2511 if (error == 0)
2512 *off = noff;
2513 return (error);
2514 }
2515
2516 int
2517 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2518 {
2519 struct ucred *cred;
2520 struct vnode *vp;
2521 struct vattr vattr;
2522 off_t foffset, size;
2523 int error, noneg;
2524
2525 cred = td->td_ucred;
2526 vp = fp->f_vnode;
2527 foffset = foffset_lock(fp, 0);
2528 noneg = (vp->v_type != VCHR);
2529 error = 0;
2530 switch (whence) {
2531 case L_INCR:
2532 if (noneg &&
2533 (foffset < 0 ||
2534 (offset > 0 && foffset > OFF_MAX - offset))) {
2535 error = EOVERFLOW;
2536 break;
2537 }
2538 offset += foffset;
2539 break;
2540 case L_XTND:
2541 vn_lock(vp, LK_SHARED | LK_RETRY);
2542 error = VOP_GETATTR(vp, &vattr, cred);
2543 VOP_UNLOCK(vp);
2544 if (error)
2545 break;
2546
2547 /*
2548 * If the file references a disk device, then fetch
2549 * the media size and use that to determine the ending
2550 * offset.
2551 */
2552 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2553 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2554 vattr.va_size = size;
2555 if (noneg &&
2556 (vattr.va_size > OFF_MAX ||
2557 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2558 error = EOVERFLOW;
2559 break;
2560 }
2561 offset += vattr.va_size;
2562 break;
2563 case L_SET:
2564 break;
2565 case SEEK_DATA:
2566 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2567 if (error == ENOTTY)
2568 error = EINVAL;
2569 break;
2570 case SEEK_HOLE:
2571 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2572 if (error == ENOTTY)
2573 error = EINVAL;
2574 break;
2575 default:
2576 error = EINVAL;
2577 }
2578 if (error == 0 && noneg && offset < 0)
2579 error = EINVAL;
2580 if (error != 0)
2581 goto drop;
2582 VFS_KNOTE_UNLOCKED(vp, 0);
2583 td->td_uretoff.tdu_off = offset;
2584 drop:
2585 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2586 return (error);
2587 }
2588
2589 int
2590 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2591 struct thread *td)
2592 {
2593 int error;
2594
2595 /*
2596 * Grant permission if the caller is the owner of the file, or
2597 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2598 * on the file. If the time pointer is null, then write
2599 * permission on the file is also sufficient.
2600 *
2601 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2602 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2603 * will be allowed to set the times [..] to the current
2604 * server time.
2605 */
2606 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2607 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2608 error = VOP_ACCESS(vp, VWRITE, cred, td);
2609 return (error);
2610 }
2611
2612 int
2613 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2614 {
2615 struct vnode *vp;
2616 int error;
2617
2618 if (fp->f_type == DTYPE_FIFO)
2619 kif->kf_type = KF_TYPE_FIFO;
2620 else
2621 kif->kf_type = KF_TYPE_VNODE;
2622 vp = fp->f_vnode;
2623 vref(vp);
2624 FILEDESC_SUNLOCK(fdp);
2625 error = vn_fill_kinfo_vnode(vp, kif);
2626 vrele(vp);
2627 FILEDESC_SLOCK(fdp);
2628 return (error);
2629 }
2630
2631 static inline void
2632 vn_fill_junk(struct kinfo_file *kif)
2633 {
2634 size_t len, olen;
2635
2636 /*
2637 * Simulate vn_fullpath returning changing values for a given
2638 * vp during e.g. coredump.
2639 */
2640 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2641 olen = strlen(kif->kf_path);
2642 if (len < olen)
2643 strcpy(&kif->kf_path[len - 1], "$");
2644 else
2645 for (; olen < len; olen++)
2646 strcpy(&kif->kf_path[olen], "A");
2647 }
2648
2649 int
2650 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2651 {
2652 struct vattr va;
2653 char *fullpath, *freepath;
2654 int error;
2655
2656 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2657 freepath = NULL;
2658 fullpath = "-";
2659 error = vn_fullpath(vp, &fullpath, &freepath);
2660 if (error == 0) {
2661 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2662 }
2663 if (freepath != NULL)
2664 free(freepath, M_TEMP);
2665
2666 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2667 vn_fill_junk(kif);
2668 );
2669
2670 /*
2671 * Retrieve vnode attributes.
2672 */
2673 va.va_fsid = VNOVAL;
2674 va.va_rdev = NODEV;
2675 vn_lock(vp, LK_SHARED | LK_RETRY);
2676 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2677 VOP_UNLOCK(vp);
2678 if (error != 0)
2679 return (error);
2680 if (va.va_fsid != VNOVAL)
2681 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2682 else
2683 kif->kf_un.kf_file.kf_file_fsid =
2684 vp->v_mount->mnt_stat.f_fsid.val[0];
2685 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2686 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2687 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2688 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2689 kif->kf_un.kf_file.kf_file_size = va.va_size;
2690 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2691 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2692 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2693 return (0);
2694 }
2695
2696 int
2697 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2698 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2699 struct thread *td)
2700 {
2701 #ifdef HWPMC_HOOKS
2702 struct pmckern_map_in pkm;
2703 #endif
2704 struct mount *mp;
2705 struct vnode *vp;
2706 vm_object_t object;
2707 vm_prot_t maxprot;
2708 boolean_t writecounted;
2709 int error;
2710
2711 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2712 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2713 /*
2714 * POSIX shared-memory objects are defined to have
2715 * kernel persistence, and are not defined to support
2716 * read(2)/write(2) -- or even open(2). Thus, we can
2717 * use MAP_ASYNC to trade on-disk coherence for speed.
2718 * The shm_open(3) library routine turns on the FPOSIXSHM
2719 * flag to request this behavior.
2720 */
2721 if ((fp->f_flag & FPOSIXSHM) != 0)
2722 flags |= MAP_NOSYNC;
2723 #endif
2724 vp = fp->f_vnode;
2725
2726 /*
2727 * Ensure that file and memory protections are
2728 * compatible. Note that we only worry about
2729 * writability if mapping is shared; in this case,
2730 * current and max prot are dictated by the open file.
2731 * XXX use the vnode instead? Problem is: what
2732 * credentials do we use for determination? What if
2733 * proc does a setuid?
2734 */
2735 mp = vp->v_mount;
2736 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2737 maxprot = VM_PROT_NONE;
2738 if ((prot & VM_PROT_EXECUTE) != 0)
2739 return (EACCES);
2740 } else
2741 maxprot = VM_PROT_EXECUTE;
2742 if ((fp->f_flag & FREAD) != 0)
2743 maxprot |= VM_PROT_READ;
2744 else if ((prot & VM_PROT_READ) != 0)
2745 return (EACCES);
2746
2747 /*
2748 * If we are sharing potential changes via MAP_SHARED and we
2749 * are trying to get write permission although we opened it
2750 * without asking for it, bail out.
2751 */
2752 if ((flags & MAP_SHARED) != 0) {
2753 if ((fp->f_flag & FWRITE) != 0)
2754 maxprot |= VM_PROT_WRITE;
2755 else if ((prot & VM_PROT_WRITE) != 0)
2756 return (EACCES);
2757 } else {
2758 maxprot |= VM_PROT_WRITE;
2759 cap_maxprot |= VM_PROT_WRITE;
2760 }
2761 maxprot &= cap_maxprot;
2762
2763 /*
2764 * For regular files and shared memory, POSIX requires that
2765 * the value of foff be a legitimate offset within the data
2766 * object. In particular, negative offsets are invalid.
2767 * Blocking negative offsets and overflows here avoids
2768 * possible wraparound or user-level access into reserved
2769 * ranges of the data object later. In contrast, POSIX does
2770 * not dictate how offsets are used by device drivers, so in
2771 * the case of a device mapping a negative offset is passed
2772 * on.
2773 */
2774 if (
2775 #ifdef _LP64
2776 size > OFF_MAX ||
2777 #endif
2778 foff > OFF_MAX - size)
2779 return (EINVAL);
2780
2781 writecounted = FALSE;
2782 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2783 &foff, &object, &writecounted);
2784 if (error != 0)
2785 return (error);
2786 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2787 foff, writecounted, td);
2788 if (error != 0) {
2789 /*
2790 * If this mapping was accounted for in the vnode's
2791 * writecount, then undo that now.
2792 */
2793 if (writecounted)
2794 vm_pager_release_writecount(object, 0, size);
2795 vm_object_deallocate(object);
2796 }
2797 #ifdef HWPMC_HOOKS
2798 /* Inform hwpmc(4) if an executable is being mapped. */
2799 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2800 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2801 pkm.pm_file = vp;
2802 pkm.pm_address = (uintptr_t) *addr;
2803 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2804 }
2805 }
2806 #endif
2807 return (error);
2808 }
2809
2810 void
2811 vn_fsid(struct vnode *vp, struct vattr *va)
2812 {
2813 fsid_t *f;
2814
2815 f = &vp->v_mount->mnt_stat.f_fsid;
2816 va->va_fsid = (uint32_t)f->val[1];
2817 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2818 va->va_fsid += (uint32_t)f->val[0];
2819 }
2820
2821 int
2822 vn_fsync_buf(struct vnode *vp, int waitfor)
2823 {
2824 struct buf *bp, *nbp;
2825 struct bufobj *bo;
2826 struct mount *mp;
2827 int error, maxretry;
2828
2829 error = 0;
2830 maxretry = 10000; /* large, arbitrarily chosen */
2831 mp = NULL;
2832 if (vp->v_type == VCHR) {
2833 VI_LOCK(vp);
2834 mp = vp->v_rdev->si_mountpt;
2835 VI_UNLOCK(vp);
2836 }
2837 bo = &vp->v_bufobj;
2838 BO_LOCK(bo);
2839 loop1:
2840 /*
2841 * MARK/SCAN initialization to avoid infinite loops.
2842 */
2843 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2844 bp->b_vflags &= ~BV_SCANNED;
2845 bp->b_error = 0;
2846 }
2847
2848 /*
2849 * Flush all dirty buffers associated with a vnode.
2850 */
2851 loop2:
2852 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2853 if ((bp->b_vflags & BV_SCANNED) != 0)
2854 continue;
2855 bp->b_vflags |= BV_SCANNED;
2856 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2857 if (waitfor != MNT_WAIT)
2858 continue;
2859 if (BUF_LOCK(bp,
2860 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2861 BO_LOCKPTR(bo)) != 0) {
2862 BO_LOCK(bo);
2863 goto loop1;
2864 }
2865 BO_LOCK(bo);
2866 }
2867 BO_UNLOCK(bo);
2868 KASSERT(bp->b_bufobj == bo,
2869 ("bp %p wrong b_bufobj %p should be %p",
2870 bp, bp->b_bufobj, bo));
2871 if ((bp->b_flags & B_DELWRI) == 0)
2872 panic("fsync: not dirty");
2873 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2874 vfs_bio_awrite(bp);
2875 } else {
2876 bremfree(bp);
2877 bawrite(bp);
2878 }
2879 if (maxretry < 1000)
2880 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2881 BO_LOCK(bo);
2882 goto loop2;
2883 }
2884
2885 /*
2886 * If synchronous the caller expects us to completely resolve all
2887 * dirty buffers in the system. Wait for in-progress I/O to
2888 * complete (which could include background bitmap writes), then
2889 * retry if dirty blocks still exist.
2890 */
2891 if (waitfor == MNT_WAIT) {
2892 bufobj_wwait(bo, 0, 0);
2893 if (bo->bo_dirty.bv_cnt > 0) {
2894 /*
2895 * If we are unable to write any of these buffers
2896 * then we fail now rather than trying endlessly
2897 * to write them out.
2898 */
2899 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2900 if ((error = bp->b_error) != 0)
2901 break;
2902 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2903 (error == 0 && --maxretry >= 0))
2904 goto loop1;
2905 if (error == 0)
2906 error = EAGAIN;
2907 }
2908 }
2909 BO_UNLOCK(bo);
2910 if (error != 0)
2911 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2912
2913 return (error);
2914 }
2915
2916 /*
2917 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
2918 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2919 * to do the actual copy.
2920 * vn_generic_copy_file_range() is factored out, so it can be called
2921 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2922 * different file systems.
2923 */
2924 int
2925 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2926 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2927 struct ucred *outcred, struct thread *fsize_td)
2928 {
2929 int error;
2930 size_t len;
2931 uint64_t uval;
2932
2933 len = *lenp;
2934 *lenp = 0; /* For error returns. */
2935 error = 0;
2936
2937 /* Do some sanity checks on the arguments. */
2938 if (invp->v_type == VDIR || outvp->v_type == VDIR)
2939 error = EISDIR;
2940 else if (*inoffp < 0 || *outoffp < 0 ||
2941 invp->v_type != VREG || outvp->v_type != VREG)
2942 error = EINVAL;
2943 if (error != 0)
2944 goto out;
2945
2946 /* Ensure offset + len does not wrap around. */
2947 uval = *inoffp;
2948 uval += len;
2949 if (uval > INT64_MAX)
2950 len = INT64_MAX - *inoffp;
2951 uval = *outoffp;
2952 uval += len;
2953 if (uval > INT64_MAX)
2954 len = INT64_MAX - *outoffp;
2955 if (len == 0)
2956 goto out;
2957
2958 /*
2959 * If the two vnode are for the same file system, call
2960 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2961 * which can handle copies across multiple file systems.
2962 */
2963 *lenp = len;
2964 if (invp->v_mount == outvp->v_mount)
2965 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2966 lenp, flags, incred, outcred, fsize_td);
2967 else
2968 error = vn_generic_copy_file_range(invp, inoffp, outvp,
2969 outoffp, lenp, flags, incred, outcred, fsize_td);
2970 out:
2971 return (error);
2972 }
2973
2974 /*
2975 * Test len bytes of data starting at dat for all bytes == 0.
2976 * Return true if all bytes are zero, false otherwise.
2977 * Expects dat to be well aligned.
2978 */
2979 static bool
2980 mem_iszero(void *dat, int len)
2981 {
2982 int i;
2983 const u_int *p;
2984 const char *cp;
2985
2986 for (p = dat; len > 0; len -= sizeof(*p), p++) {
2987 if (len >= sizeof(*p)) {
2988 if (*p != 0)
2989 return (false);
2990 } else {
2991 cp = (const char *)p;
2992 for (i = 0; i < len; i++, cp++)
2993 if (*cp != '\0')
2994 return (false);
2995 }
2996 }
2997 return (true);
2998 }
2999
3000 /*
3001 * Look for a hole in the output file and, if found, adjust *outoffp
3002 * and *xferp to skip past the hole.
3003 * *xferp is the entire hole length to be written and xfer2 is how many bytes
3004 * to be written as 0's upon return.
3005 */
3006 static off_t
3007 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3008 off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3009 {
3010 int error;
3011 off_t delta;
3012
3013 if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3014 *dataoffp = *outoffp;
3015 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3016 curthread);
3017 if (error == 0) {
3018 *holeoffp = *dataoffp;
3019 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3020 curthread);
3021 }
3022 if (error != 0 || *holeoffp == *dataoffp) {
3023 /*
3024 * Since outvp is unlocked, it may be possible for
3025 * another thread to do a truncate(), lseek(), write()
3026 * creating a hole at startoff between the above
3027 * VOP_IOCTL() calls, if the other thread does not do
3028 * rangelocking.
3029 * If that happens, *holeoffp == *dataoffp and finding
3030 * the hole has failed, so disable vn_skip_hole().
3031 */
3032 *holeoffp = -1; /* Disable use of vn_skip_hole(). */
3033 return (xfer2);
3034 }
3035 KASSERT(*dataoffp >= *outoffp,
3036 ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3037 (intmax_t)*dataoffp, (intmax_t)*outoffp));
3038 KASSERT(*holeoffp > *dataoffp,
3039 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3040 (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3041 }
3042
3043 /*
3044 * If there is a hole before the data starts, advance *outoffp and
3045 * *xferp past the hole.
3046 */
3047 if (*dataoffp > *outoffp) {
3048 delta = *dataoffp - *outoffp;
3049 if (delta >= *xferp) {
3050 /* Entire *xferp is a hole. */
3051 *outoffp += *xferp;
3052 *xferp = 0;
3053 return (0);
3054 }
3055 *xferp -= delta;
3056 *outoffp += delta;
3057 xfer2 = MIN(xfer2, *xferp);
3058 }
3059
3060 /*
3061 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3062 * that the write ends at the start of the hole.
3063 * *holeoffp should always be greater than *outoffp, but for the
3064 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3065 * value.
3066 */
3067 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3068 xfer2 = *holeoffp - *outoffp;
3069 return (xfer2);
3070 }
3071
3072 /*
3073 * Write an xfer sized chunk to outvp in blksize blocks from dat.
3074 * dat is a maximum of blksize in length and can be written repeatedly in
3075 * the chunk.
3076 * If growfile == true, just grow the file via vn_truncate_locked() instead
3077 * of doing actual writes.
3078 * If checkhole == true, a hole is being punched, so skip over any hole
3079 * already in the output file.
3080 */
3081 static int
3082 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3083 u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3084 {
3085 struct mount *mp;
3086 off_t dataoff, holeoff, xfer2;
3087 int error;
3088
3089 /*
3090 * Loop around doing writes of blksize until write has been completed.
3091 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3092 * done for each iteration, since the xfer argument can be very
3093 * large if there is a large hole to punch in the output file.
3094 */
3095 error = 0;
3096 holeoff = 0;
3097 do {
3098 xfer2 = MIN(xfer, blksize);
3099 if (checkhole) {
3100 /*
3101 * Punching a hole. Skip writing if there is
3102 * already a hole in the output file.
3103 */
3104 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3105 &dataoff, &holeoff, cred);
3106 if (xfer == 0)
3107 break;
3108 if (holeoff < 0)
3109 checkhole = false;
3110 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3111 (intmax_t)xfer2));
3112 }
3113 bwillwrite();
3114 mp = NULL;
3115 error = vn_start_write(outvp, &mp, V_WAIT);
3116 if (error != 0)
3117 break;
3118 if (growfile) {
3119 error = vn_lock(outvp, LK_EXCLUSIVE);
3120 if (error == 0) {
3121 error = vn_truncate_locked(outvp, outoff + xfer,
3122 false, cred);
3123 VOP_UNLOCK(outvp);
3124 }
3125 } else {
3126 error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3127 if (error == 0) {
3128 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3129 outoff, UIO_SYSSPACE, IO_NODELOCKED,
3130 curthread->td_ucred, cred, NULL, curthread);
3131 outoff += xfer2;
3132 xfer -= xfer2;
3133 VOP_UNLOCK(outvp);
3134 }
3135 }
3136 if (mp != NULL)
3137 vn_finished_write(mp);
3138 } while (!growfile && xfer > 0 && error == 0);
3139 return (error);
3140 }
3141
3142 /*
3143 * Copy a byte range of one file to another. This function can handle the
3144 * case where invp and outvp are on different file systems.
3145 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3146 * is no better file system specific way to do it.
3147 */
3148 int
3149 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3150 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3151 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3152 {
3153 struct vattr va, inva;
3154 struct mount *mp;
3155 struct uio io;
3156 off_t startoff, endoff, xfer, xfer2;
3157 u_long blksize;
3158 int error, interrupted;
3159 bool cantseek, readzeros, eof, lastblock, holetoeof;
3160 ssize_t aresid;
3161 size_t copylen, len, rem, savlen;
3162 char *dat;
3163 long holein, holeout;
3164 struct timespec curts, endts;
3165
3166 holein = holeout = 0;
3167 savlen = len = *lenp;
3168 error = 0;
3169 interrupted = 0;
3170 dat = NULL;
3171
3172 error = vn_lock(invp, LK_SHARED);
3173 if (error != 0)
3174 goto out;
3175 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3176 holein = 0;
3177 if (holein > 0)
3178 error = VOP_GETATTR(invp, &inva, incred);
3179 VOP_UNLOCK(invp);
3180 if (error != 0)
3181 goto out;
3182
3183 mp = NULL;
3184 error = vn_start_write(outvp, &mp, V_WAIT);
3185 if (error == 0)
3186 error = vn_lock(outvp, LK_EXCLUSIVE);
3187 if (error == 0) {
3188 /*
3189 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3190 * now that outvp is locked.
3191 */
3192 if (fsize_td != NULL) {
3193 io.uio_offset = *outoffp;
3194 io.uio_resid = len;
3195 error = vn_rlimit_fsize(outvp, &io, fsize_td);
3196 if (error != 0)
3197 error = EFBIG;
3198 }
3199 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3200 holeout = 0;
3201 /*
3202 * Holes that are past EOF do not need to be written as a block
3203 * of zero bytes. So, truncate the output file as far as
3204 * possible and then use va.va_size to decide if writing 0
3205 * bytes is necessary in the loop below.
3206 */
3207 if (error == 0)
3208 error = VOP_GETATTR(outvp, &va, outcred);
3209 if (error == 0 && va.va_size > *outoffp && va.va_size <=
3210 *outoffp + len) {
3211 #ifdef MAC
3212 error = mac_vnode_check_write(curthread->td_ucred,
3213 outcred, outvp);
3214 if (error == 0)
3215 #endif
3216 error = vn_truncate_locked(outvp, *outoffp,
3217 false, outcred);
3218 if (error == 0)
3219 va.va_size = *outoffp;
3220 }
3221 VOP_UNLOCK(outvp);
3222 }
3223 if (mp != NULL)
3224 vn_finished_write(mp);
3225 if (error != 0)
3226 goto out;
3227
3228 /*
3229 * Set the blksize to the larger of the hole sizes for invp and outvp.
3230 * If hole sizes aren't available, set the blksize to the larger
3231 * f_iosize of invp and outvp.
3232 * This code expects the hole sizes and f_iosizes to be powers of 2.
3233 * This value is clipped at 4Kbytes and 1Mbyte.
3234 */
3235 blksize = MAX(holein, holeout);
3236
3237 /* Clip len to end at an exact multiple of hole size. */
3238 if (blksize > 1) {
3239 rem = *inoffp % blksize;
3240 if (rem > 0)
3241 rem = blksize - rem;
3242 if (len > rem && len - rem > blksize)
3243 len = savlen = rounddown(len - rem, blksize) + rem;
3244 }
3245
3246 if (blksize <= 1)
3247 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3248 outvp->v_mount->mnt_stat.f_iosize);
3249 if (blksize < 4096)
3250 blksize = 4096;
3251 else if (blksize > 1024 * 1024)
3252 blksize = 1024 * 1024;
3253 dat = malloc(blksize, M_TEMP, M_WAITOK);
3254
3255 /*
3256 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3257 * to find holes. Otherwise, just scan the read block for all 0s
3258 * in the inner loop where the data copying is done.
3259 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3260 * support holes on the server, but do not support FIOSEEKHOLE.
3261 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3262 * that this function should return after 1second with a partial
3263 * completion.
3264 */
3265 if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3266 getnanouptime(&endts);
3267 endts.tv_sec++;
3268 } else
3269 timespecclear(&endts);
3270 holetoeof = eof = false;
3271 while (len > 0 && error == 0 && !eof && interrupted == 0) {
3272 endoff = 0; /* To shut up compilers. */
3273 cantseek = true;
3274 startoff = *inoffp;
3275 copylen = len;
3276
3277 /*
3278 * Find the next data area. If there is just a hole to EOF,
3279 * FIOSEEKDATA should fail with ENXIO.
3280 * (I do not know if any file system will report a hole to
3281 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3282 * will fail for those file systems.)
3283 *
3284 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3285 * the code just falls through to the inner copy loop.
3286 */
3287 error = EINVAL;
3288 if (holein > 0) {
3289 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3290 incred, curthread);
3291 if (error == ENXIO) {
3292 startoff = endoff = inva.va_size;
3293 eof = holetoeof = true;
3294 error = 0;
3295 }
3296 }
3297 if (error == 0 && !holetoeof) {
3298 endoff = startoff;
3299 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3300 incred, curthread);
3301 /*
3302 * Since invp is unlocked, it may be possible for
3303 * another thread to do a truncate(), lseek(), write()
3304 * creating a hole at startoff between the above
3305 * VOP_IOCTL() calls, if the other thread does not do
3306 * rangelocking.
3307 * If that happens, startoff == endoff and finding
3308 * the hole has failed, so set an error.
3309 */
3310 if (error == 0 && startoff == endoff)
3311 error = EINVAL; /* Any error. Reset to 0. */
3312 }
3313 if (error == 0) {
3314 if (startoff > *inoffp) {
3315 /* Found hole before data block. */
3316 xfer = MIN(startoff - *inoffp, len);
3317 if (*outoffp < va.va_size) {
3318 /* Must write 0s to punch hole. */
3319 xfer2 = MIN(va.va_size - *outoffp,
3320 xfer);
3321 memset(dat, 0, MIN(xfer2, blksize));
3322 error = vn_write_outvp(outvp, dat,
3323 *outoffp, xfer2, blksize, false,
3324 holeout > 0, outcred);
3325 }
3326
3327 if (error == 0 && *outoffp + xfer >
3328 va.va_size && (xfer == len || holetoeof)) {
3329 /* Grow output file (hole at end). */
3330 error = vn_write_outvp(outvp, dat,
3331 *outoffp, xfer, blksize, true,
3332 false, outcred);
3333 }
3334 if (error == 0) {
3335 *inoffp += xfer;
3336 *outoffp += xfer;
3337 len -= xfer;
3338 if (len < savlen) {
3339 interrupted = sig_intr();
3340 if (timespecisset(&endts) &&
3341 interrupted == 0) {
3342 getnanouptime(&curts);
3343 if (timespeccmp(&curts,
3344 &endts, >=))
3345 interrupted =
3346 EINTR;
3347 }
3348 }
3349 }
3350 }
3351 copylen = MIN(len, endoff - startoff);
3352 cantseek = false;
3353 } else {
3354 cantseek = true;
3355 startoff = *inoffp;
3356 copylen = len;
3357 error = 0;
3358 }
3359
3360 xfer = blksize;
3361 if (cantseek) {
3362 /*
3363 * Set first xfer to end at a block boundary, so that
3364 * holes are more likely detected in the loop below via
3365 * the for all bytes 0 method.
3366 */
3367 xfer -= (*inoffp % blksize);
3368 }
3369 /* Loop copying the data block. */
3370 while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3371 if (copylen < xfer)
3372 xfer = copylen;
3373 error = vn_lock(invp, LK_SHARED);
3374 if (error != 0)
3375 goto out;
3376 error = vn_rdwr(UIO_READ, invp, dat, xfer,
3377 startoff, UIO_SYSSPACE, IO_NODELOCKED,
3378 curthread->td_ucred, incred, &aresid,
3379 curthread);
3380 VOP_UNLOCK(invp);
3381 lastblock = false;
3382 if (error == 0 && aresid > 0) {
3383 /* Stop the copy at EOF on the input file. */
3384 xfer -= aresid;
3385 eof = true;
3386 lastblock = true;
3387 }
3388 if (error == 0) {
3389 /*
3390 * Skip the write for holes past the initial EOF
3391 * of the output file, unless this is the last
3392 * write of the output file at EOF.
3393 */
3394 readzeros = cantseek ? mem_iszero(dat, xfer) :
3395 false;
3396 if (xfer == len)
3397 lastblock = true;
3398 if (!cantseek || *outoffp < va.va_size ||
3399 lastblock || !readzeros)
3400 error = vn_write_outvp(outvp, dat,
3401 *outoffp, xfer, blksize,
3402 readzeros && lastblock &&
3403 *outoffp >= va.va_size, false,
3404 outcred);
3405 if (error == 0) {
3406 *inoffp += xfer;
3407 startoff += xfer;
3408 *outoffp += xfer;
3409 copylen -= xfer;
3410 len -= xfer;
3411 if (len < savlen) {
3412 interrupted = sig_intr();
3413 if (timespecisset(&endts) &&
3414 interrupted == 0) {
3415 getnanouptime(&curts);
3416 if (timespeccmp(&curts,
3417 &endts, >=))
3418 interrupted =
3419 EINTR;
3420 }
3421 }
3422 }
3423 }
3424 xfer = blksize;
3425 }
3426 }
3427 out:
3428 *lenp = savlen - len;
3429 free(dat, M_TEMP);
3430 return (error);
3431 }
3432
3433 static int
3434 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3435 {
3436 struct mount *mp;
3437 struct vnode *vp;
3438 off_t olen, ooffset;
3439 int error;
3440 #ifdef AUDIT
3441 int audited_vnode1 = 0;
3442 #endif
3443
3444 vp = fp->f_vnode;
3445 if (vp->v_type != VREG)
3446 return (ENODEV);
3447
3448 /* Allocating blocks may take a long time, so iterate. */
3449 for (;;) {
3450 olen = len;
3451 ooffset = offset;
3452
3453 bwillwrite();
3454 mp = NULL;
3455 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3456 if (error != 0)
3457 break;
3458 error = vn_lock(vp, LK_EXCLUSIVE);
3459 if (error != 0) {
3460 vn_finished_write(mp);
3461 break;
3462 }
3463 #ifdef AUDIT
3464 if (!audited_vnode1) {
3465 AUDIT_ARG_VNODE1(vp);
3466 audited_vnode1 = 1;
3467 }
3468 #endif
3469 #ifdef MAC
3470 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3471 if (error == 0)
3472 #endif
3473 error = VOP_ALLOCATE(vp, &offset, &len);
3474 VOP_UNLOCK(vp);
3475 vn_finished_write(mp);
3476
3477 if (olen + ooffset != offset + len) {
3478 panic("offset + len changed from %jx/%jx to %jx/%jx",
3479 ooffset, olen, offset, len);
3480 }
3481 if (error != 0 || len == 0)
3482 break;
3483 KASSERT(olen > len, ("Iteration did not make progress?"));
3484 maybe_yield();
3485 }
3486
3487 return (error);
3488 }
3489
3490 static u_long vn_lock_pair_pause_cnt;
3491 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3492 &vn_lock_pair_pause_cnt, 0,
3493 "Count of vn_lock_pair deadlocks");
3494
3495 u_int vn_lock_pair_pause_max;
3496 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3497 &vn_lock_pair_pause_max, 0,
3498 "Max ticks for vn_lock_pair deadlock avoidance sleep");
3499
3500 static void
3501 vn_lock_pair_pause(const char *wmesg)
3502 {
3503 atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3504 pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3505 }
3506
3507 /*
3508 * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3509 * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3510 * must be unlocked. Same for vp2 and vp2_locked. One of the vnodes
3511 * can be NULL.
3512 *
3513 * The function returns with both vnodes exclusively locked, and
3514 * guarantees that it does not create lock order reversal with other
3515 * threads during its execution. Both vnodes could be unlocked
3516 * temporary (and reclaimed).
3517 */
3518 void
3519 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3520 bool vp2_locked)
3521 {
3522 int error;
3523
3524 if (vp1 == NULL && vp2 == NULL)
3525 return;
3526 if (vp1 != NULL) {
3527 if (vp1_locked)
3528 ASSERT_VOP_ELOCKED(vp1, "vp1");
3529 else
3530 ASSERT_VOP_UNLOCKED(vp1, "vp1");
3531 } else {
3532 vp1_locked = true;
3533 }
3534 if (vp2 != NULL) {
3535 if (vp2_locked)
3536 ASSERT_VOP_ELOCKED(vp2, "vp2");
3537 else
3538 ASSERT_VOP_UNLOCKED(vp2, "vp2");
3539 } else {
3540 vp2_locked = true;
3541 }
3542 if (!vp1_locked && !vp2_locked) {
3543 vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3544 vp1_locked = true;
3545 }
3546
3547 for (;;) {
3548 if (vp1_locked && vp2_locked)
3549 break;
3550 if (vp1_locked && vp2 != NULL) {
3551 if (vp1 != NULL) {
3552 error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3553 __FILE__, __LINE__);
3554 if (error == 0)
3555 break;
3556 VOP_UNLOCK(vp1);
3557 vp1_locked = false;
3558 vn_lock_pair_pause("vlp1");
3559 }
3560 vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3561 vp2_locked = true;
3562 }
3563 if (vp2_locked && vp1 != NULL) {
3564 if (vp2 != NULL) {
3565 error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3566 __FILE__, __LINE__);
3567 if (error == 0)
3568 break;
3569 VOP_UNLOCK(vp2);
3570 vp2_locked = false;
3571 vn_lock_pair_pause("vlp2");
3572 }
3573 vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3574 vp1_locked = true;
3575 }
3576 }
3577 if (vp1 != NULL)
3578 ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3579 if (vp2 != NULL)
3580 ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3581 }
3582
3583 int
3584 vn_lktype_write(struct mount *mp, struct vnode *vp)
3585 {
3586 if (MNT_SHARED_WRITES(mp) ||
3587 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
3588 return (LK_SHARED);
3589 return (LK_EXCLUSIVE);
3590 }
Cache object: 69d2d222a80ec5b38f73c0d5a9bc03c8
|