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 if ((vn_open_flags & VN_OPEN_WANTIOCTLCAPS) != 0)
211 res |= WANTIOCTLCAPS;
212 return (res);
213 }
214
215 /*
216 * Common code for vnode open operations via a name lookup.
217 * Lookup the vnode and invoke VOP_CREATE if needed.
218 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
219 *
220 * Note that this does NOT free nameidata for the successful case,
221 * due to the NDINIT being done elsewhere.
222 */
223 int
224 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
225 struct ucred *cred, struct file *fp)
226 {
227 struct vnode *vp;
228 struct mount *mp;
229 struct thread *td = ndp->ni_cnd.cn_thread;
230 struct vattr vat;
231 struct vattr *vap = &vat;
232 int fmode, error;
233 bool first_open;
234
235 restart:
236 first_open = false;
237 fmode = *flagp;
238 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
239 O_EXCL | O_DIRECTORY) ||
240 (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
241 return (EINVAL);
242 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
243 ndp->ni_cnd.cn_nameiop = CREATE;
244 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
245 /*
246 * Set NOCACHE to avoid flushing the cache when
247 * rolling in many files at once.
248 *
249 * Set NC_KEEPPOSENTRY to keep positive entries if they already
250 * exist despite NOCACHE.
251 */
252 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
253 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
254 ndp->ni_cnd.cn_flags |= FOLLOW;
255 if ((vn_open_flags & VN_OPEN_INVFS) == 0)
256 bwillwrite();
257 if ((error = namei(ndp)) != 0)
258 return (error);
259 if (ndp->ni_vp == NULL) {
260 VATTR_NULL(vap);
261 vap->va_type = VREG;
262 vap->va_mode = cmode;
263 if (fmode & O_EXCL)
264 vap->va_vaflags |= VA_EXCLUSIVE;
265 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
266 NDFREE(ndp, NDF_ONLY_PNBUF);
267 vput(ndp->ni_dvp);
268 if ((error = vn_start_write(NULL, &mp,
269 V_XSLEEP | PCATCH)) != 0)
270 return (error);
271 NDREINIT(ndp);
272 goto restart;
273 }
274 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
275 ndp->ni_cnd.cn_flags |= MAKEENTRY;
276 #ifdef MAC
277 error = mac_vnode_check_create(cred, ndp->ni_dvp,
278 &ndp->ni_cnd, vap);
279 if (error == 0)
280 #endif
281 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
282 &ndp->ni_cnd, vap);
283 vp = ndp->ni_vp;
284 if (error == 0 && (fmode & O_EXCL) != 0 &&
285 (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
286 VI_LOCK(vp);
287 vp->v_iflag |= VI_FOPENING;
288 VI_UNLOCK(vp);
289 first_open = true;
290 }
291 VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
292 false);
293 vn_finished_write(mp);
294 if (error) {
295 NDFREE(ndp, NDF_ONLY_PNBUF);
296 if (error == ERELOOKUP) {
297 NDREINIT(ndp);
298 goto restart;
299 }
300 return (error);
301 }
302 fmode &= ~O_TRUNC;
303 } else {
304 if (ndp->ni_dvp == ndp->ni_vp)
305 vrele(ndp->ni_dvp);
306 else
307 vput(ndp->ni_dvp);
308 ndp->ni_dvp = NULL;
309 vp = ndp->ni_vp;
310 if (fmode & O_EXCL) {
311 error = EEXIST;
312 goto bad;
313 }
314 if (vp->v_type == VDIR) {
315 error = EISDIR;
316 goto bad;
317 }
318 fmode &= ~O_CREAT;
319 }
320 } else {
321 ndp->ni_cnd.cn_nameiop = LOOKUP;
322 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
323 ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
324 FOLLOW;
325 if ((fmode & FWRITE) == 0)
326 ndp->ni_cnd.cn_flags |= LOCKSHARED;
327 if ((error = namei(ndp)) != 0)
328 return (error);
329 vp = ndp->ni_vp;
330 }
331 error = vn_open_vnode(vp, fmode, cred, td, fp);
332 if (first_open) {
333 VI_LOCK(vp);
334 vp->v_iflag &= ~VI_FOPENING;
335 wakeup(vp);
336 VI_UNLOCK(vp);
337 }
338 if (error)
339 goto bad;
340 *flagp = fmode;
341 return (0);
342 bad:
343 NDFREE(ndp, NDF_ONLY_PNBUF);
344 vput(vp);
345 *flagp = fmode;
346 ndp->ni_vp = NULL;
347 return (error);
348 }
349
350 static int
351 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
352 {
353 struct flock lf;
354 int error, lock_flags, type;
355
356 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
357 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
358 return (0);
359 KASSERT(fp != NULL, ("open with flock requires fp"));
360 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
361 return (EOPNOTSUPP);
362
363 lock_flags = VOP_ISLOCKED(vp);
364 VOP_UNLOCK(vp);
365
366 lf.l_whence = SEEK_SET;
367 lf.l_start = 0;
368 lf.l_len = 0;
369 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
370 type = F_FLOCK;
371 if ((fmode & FNONBLOCK) == 0)
372 type |= F_WAIT;
373 if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
374 type |= F_FIRSTOPEN;
375 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
376 if (error == 0)
377 fp->f_flag |= FHASLOCK;
378
379 vn_lock(vp, lock_flags | LK_RETRY);
380 return (error);
381 }
382
383 /*
384 * Common code for vnode open operations once a vnode is located.
385 * Check permissions, and call the VOP_OPEN routine.
386 */
387 int
388 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
389 struct thread *td, struct file *fp)
390 {
391 accmode_t accmode;
392 int error;
393
394 if (vp->v_type == VLNK) {
395 if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
396 return (EMLINK);
397 }
398 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
399 return (ENOTDIR);
400
401 accmode = 0;
402 if ((fmode & O_PATH) == 0) {
403 if (vp->v_type == VSOCK)
404 return (EOPNOTSUPP);
405 if ((fmode & (FWRITE | O_TRUNC)) != 0) {
406 if (vp->v_type == VDIR)
407 return (EISDIR);
408 accmode |= VWRITE;
409 }
410 if ((fmode & FREAD) != 0)
411 accmode |= VREAD;
412 if ((fmode & O_APPEND) && (fmode & FWRITE))
413 accmode |= VAPPEND;
414 #ifdef MAC
415 if ((fmode & O_CREAT) != 0)
416 accmode |= VCREAT;
417 #endif
418 }
419 if ((fmode & FEXEC) != 0)
420 accmode |= VEXEC;
421 #ifdef MAC
422 if ((fmode & O_VERIFY) != 0)
423 accmode |= VVERIFY;
424 error = mac_vnode_check_open(cred, vp, accmode);
425 if (error != 0)
426 return (error);
427
428 accmode &= ~(VCREAT | VVERIFY);
429 #endif
430 if ((fmode & O_CREAT) == 0 && accmode != 0) {
431 error = VOP_ACCESS(vp, accmode, cred, td);
432 if (error != 0)
433 return (error);
434 }
435 if ((fmode & O_PATH) != 0) {
436 if (vp->v_type != VFIFO && vp->v_type != VSOCK &&
437 VOP_ACCESS(vp, VREAD, cred, td) == 0)
438 fp->f_flag |= FKQALLOWED;
439 return (0);
440 }
441
442 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
443 vn_lock(vp, LK_UPGRADE | LK_RETRY);
444 error = VOP_OPEN(vp, fmode, cred, td, fp);
445 if (error != 0)
446 return (error);
447
448 error = vn_open_vnode_advlock(vp, fmode, fp);
449 if (error == 0 && (fmode & FWRITE) != 0) {
450 error = VOP_ADD_WRITECOUNT(vp, 1);
451 if (error == 0) {
452 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
453 __func__, vp, vp->v_writecount);
454 }
455 }
456
457 /*
458 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
459 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
460 */
461 if (error != 0) {
462 if (fp != NULL) {
463 /*
464 * Arrange the call by having fdrop() to use
465 * vn_closefile(). This is to satisfy
466 * filesystems like devfs or tmpfs, which
467 * override fo_close().
468 */
469 fp->f_flag |= FOPENFAILED;
470 fp->f_vnode = vp;
471 if (fp->f_ops == &badfileops) {
472 fp->f_type = DTYPE_VNODE;
473 fp->f_ops = &vnops;
474 }
475 vref(vp);
476 } else {
477 /*
478 * If there is no fp, due to kernel-mode open,
479 * we can call VOP_CLOSE() now.
480 */
481 if (vp->v_type != VFIFO && (fmode & FWRITE) != 0 &&
482 !MNT_EXTENDED_SHARED(vp->v_mount) &&
483 VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
484 vn_lock(vp, LK_UPGRADE | LK_RETRY);
485 (void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
486 cred, td);
487 }
488 }
489
490 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
491 return (error);
492
493 }
494
495 /*
496 * Check for write permissions on the specified vnode.
497 * Prototype text segments cannot be written.
498 * It is racy.
499 */
500 int
501 vn_writechk(struct vnode *vp)
502 {
503
504 ASSERT_VOP_LOCKED(vp, "vn_writechk");
505 /*
506 * If there's shared text associated with
507 * the vnode, try to free it up once. If
508 * we fail, we can't allow writing.
509 */
510 if (VOP_IS_TEXT(vp))
511 return (ETXTBSY);
512
513 return (0);
514 }
515
516 /*
517 * Vnode close call
518 */
519 static int
520 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
521 struct thread *td, bool keep_ref)
522 {
523 struct mount *mp;
524 int error, lock_flags;
525
526 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
527 MNT_EXTENDED_SHARED(vp->v_mount))
528 lock_flags = LK_SHARED;
529 else
530 lock_flags = LK_EXCLUSIVE;
531
532 vn_start_write(vp, &mp, V_WAIT);
533 vn_lock(vp, lock_flags | LK_RETRY);
534 AUDIT_ARG_VNODE1(vp);
535 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
536 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
537 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
538 __func__, vp, vp->v_writecount);
539 }
540 error = VOP_CLOSE(vp, flags, file_cred, td);
541 if (keep_ref)
542 VOP_UNLOCK(vp);
543 else
544 vput(vp);
545 vn_finished_write(mp);
546 return (error);
547 }
548
549 int
550 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
551 struct thread *td)
552 {
553
554 return (vn_close1(vp, flags, file_cred, td, false));
555 }
556
557 /*
558 * Heuristic to detect sequential operation.
559 */
560 static int
561 sequential_heuristic(struct uio *uio, struct file *fp)
562 {
563 enum uio_rw rw;
564
565 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
566
567 rw = uio->uio_rw;
568 if (fp->f_flag & FRDAHEAD)
569 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
570
571 /*
572 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
573 * that the first I/O is normally considered to be slightly
574 * sequential. Seeking to offset 0 doesn't change sequentiality
575 * unless previous seeks have reduced f_seqcount to 0, in which
576 * case offset 0 is not special.
577 */
578 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
579 uio->uio_offset == fp->f_nextoff[rw]) {
580 /*
581 * f_seqcount is in units of fixed-size blocks so that it
582 * depends mainly on the amount of sequential I/O and not
583 * much on the number of sequential I/O's. The fixed size
584 * of 16384 is hard-coded here since it is (not quite) just
585 * a magic size that works well here. This size is more
586 * closely related to the best I/O size for real disks than
587 * to any block size used by software.
588 */
589 if (uio->uio_resid >= IO_SEQMAX * 16384)
590 fp->f_seqcount[rw] = IO_SEQMAX;
591 else {
592 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
593 if (fp->f_seqcount[rw] > IO_SEQMAX)
594 fp->f_seqcount[rw] = IO_SEQMAX;
595 }
596 return (fp->f_seqcount[rw] << IO_SEQSHIFT);
597 }
598
599 /* Not sequential. Quickly draw-down sequentiality. */
600 if (fp->f_seqcount[rw] > 1)
601 fp->f_seqcount[rw] = 1;
602 else
603 fp->f_seqcount[rw] = 0;
604 return (0);
605 }
606
607 /*
608 * Package up an I/O request on a vnode into a uio and do it.
609 */
610 int
611 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
612 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
613 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
614 {
615 struct uio auio;
616 struct iovec aiov;
617 struct mount *mp;
618 struct ucred *cred;
619 void *rl_cookie;
620 struct vn_io_fault_args args;
621 int error, lock_flags;
622
623 if (offset < 0 && vp->v_type != VCHR)
624 return (EINVAL);
625 auio.uio_iov = &aiov;
626 auio.uio_iovcnt = 1;
627 aiov.iov_base = base;
628 aiov.iov_len = len;
629 auio.uio_resid = len;
630 auio.uio_offset = offset;
631 auio.uio_segflg = segflg;
632 auio.uio_rw = rw;
633 auio.uio_td = td;
634 error = 0;
635
636 if ((ioflg & IO_NODELOCKED) == 0) {
637 if ((ioflg & IO_RANGELOCKED) == 0) {
638 if (rw == UIO_READ) {
639 rl_cookie = vn_rangelock_rlock(vp, offset,
640 offset + len);
641 } else if ((ioflg & IO_APPEND) != 0) {
642 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
643 } else {
644 rl_cookie = vn_rangelock_wlock(vp, offset,
645 offset + len);
646 }
647 } else
648 rl_cookie = NULL;
649 mp = NULL;
650 if (rw == UIO_WRITE) {
651 if (vp->v_type != VCHR &&
652 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
653 != 0)
654 goto out;
655 lock_flags = vn_lktype_write(mp, vp);
656 } else
657 lock_flags = LK_SHARED;
658 vn_lock(vp, lock_flags | LK_RETRY);
659 } else
660 rl_cookie = NULL;
661
662 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
663 #ifdef MAC
664 if ((ioflg & IO_NOMACCHECK) == 0) {
665 if (rw == UIO_READ)
666 error = mac_vnode_check_read(active_cred, file_cred,
667 vp);
668 else
669 error = mac_vnode_check_write(active_cred, file_cred,
670 vp);
671 }
672 #endif
673 if (error == 0) {
674 if (file_cred != NULL)
675 cred = file_cred;
676 else
677 cred = active_cred;
678 if (do_vn_io_fault(vp, &auio)) {
679 args.kind = VN_IO_FAULT_VOP;
680 args.cred = cred;
681 args.flags = ioflg;
682 args.args.vop_args.vp = vp;
683 error = vn_io_fault1(vp, &auio, &args, td);
684 } else if (rw == UIO_READ) {
685 error = VOP_READ(vp, &auio, ioflg, cred);
686 } else /* if (rw == UIO_WRITE) */ {
687 error = VOP_WRITE(vp, &auio, ioflg, cred);
688 }
689 }
690 if (aresid)
691 *aresid = auio.uio_resid;
692 else
693 if (auio.uio_resid && error == 0)
694 error = EIO;
695 if ((ioflg & IO_NODELOCKED) == 0) {
696 VOP_UNLOCK(vp);
697 if (mp != NULL)
698 vn_finished_write(mp);
699 }
700 out:
701 if (rl_cookie != NULL)
702 vn_rangelock_unlock(vp, rl_cookie);
703 return (error);
704 }
705
706 /*
707 * Package up an I/O request on a vnode into a uio and do it. The I/O
708 * request is split up into smaller chunks and we try to avoid saturating
709 * the buffer cache while potentially holding a vnode locked, so we
710 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
711 * to give other processes a chance to lock the vnode (either other processes
712 * core'ing the same binary, or unrelated processes scanning the directory).
713 */
714 int
715 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
716 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
717 struct ucred *file_cred, size_t *aresid, struct thread *td)
718 {
719 int error = 0;
720 ssize_t iaresid;
721
722 do {
723 int chunk;
724
725 /*
726 * Force `offset' to a multiple of MAXBSIZE except possibly
727 * for the first chunk, so that filesystems only need to
728 * write full blocks except possibly for the first and last
729 * chunks.
730 */
731 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
732
733 if (chunk > len)
734 chunk = len;
735 if (rw != UIO_READ && vp->v_type == VREG)
736 bwillwrite();
737 iaresid = 0;
738 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
739 ioflg, active_cred, file_cred, &iaresid, td);
740 len -= chunk; /* aresid calc already includes length */
741 if (error)
742 break;
743 offset += chunk;
744 base = (char *)base + chunk;
745 kern_yield(PRI_USER);
746 } while (len);
747 if (aresid)
748 *aresid = len + iaresid;
749 return (error);
750 }
751
752 #if OFF_MAX <= LONG_MAX
753 off_t
754 foffset_lock(struct file *fp, int flags)
755 {
756 volatile short *flagsp;
757 off_t res;
758 short state;
759
760 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
761
762 if ((flags & FOF_NOLOCK) != 0)
763 return (atomic_load_long(&fp->f_offset));
764
765 /*
766 * According to McKusick the vn lock was protecting f_offset here.
767 * It is now protected by the FOFFSET_LOCKED flag.
768 */
769 flagsp = &fp->f_vnread_flags;
770 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
771 return (atomic_load_long(&fp->f_offset));
772
773 sleepq_lock(&fp->f_vnread_flags);
774 state = atomic_load_16(flagsp);
775 for (;;) {
776 if ((state & FOFFSET_LOCKED) == 0) {
777 if (!atomic_fcmpset_acq_16(flagsp, &state,
778 FOFFSET_LOCKED))
779 continue;
780 break;
781 }
782 if ((state & FOFFSET_LOCK_WAITING) == 0) {
783 if (!atomic_fcmpset_acq_16(flagsp, &state,
784 state | FOFFSET_LOCK_WAITING))
785 continue;
786 }
787 DROP_GIANT();
788 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
789 sleepq_wait(&fp->f_vnread_flags, PUSER -1);
790 PICKUP_GIANT();
791 sleepq_lock(&fp->f_vnread_flags);
792 state = atomic_load_16(flagsp);
793 }
794 res = atomic_load_long(&fp->f_offset);
795 sleepq_release(&fp->f_vnread_flags);
796 return (res);
797 }
798
799 void
800 foffset_unlock(struct file *fp, off_t val, int flags)
801 {
802 volatile short *flagsp;
803 short state;
804
805 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
806
807 if ((flags & FOF_NOUPDATE) == 0)
808 atomic_store_long(&fp->f_offset, val);
809 if ((flags & FOF_NEXTOFF_R) != 0)
810 fp->f_nextoff[UIO_READ] = val;
811 if ((flags & FOF_NEXTOFF_W) != 0)
812 fp->f_nextoff[UIO_WRITE] = val;
813
814 if ((flags & FOF_NOLOCK) != 0)
815 return;
816
817 flagsp = &fp->f_vnread_flags;
818 state = atomic_load_16(flagsp);
819 if ((state & FOFFSET_LOCK_WAITING) == 0 &&
820 atomic_cmpset_rel_16(flagsp, state, 0))
821 return;
822
823 sleepq_lock(&fp->f_vnread_flags);
824 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
825 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
826 fp->f_vnread_flags = 0;
827 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
828 sleepq_release(&fp->f_vnread_flags);
829 }
830 #else
831 off_t
832 foffset_lock(struct file *fp, int flags)
833 {
834 struct mtx *mtxp;
835 off_t res;
836
837 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
838
839 mtxp = mtx_pool_find(mtxpool_sleep, fp);
840 mtx_lock(mtxp);
841 if ((flags & FOF_NOLOCK) == 0) {
842 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
843 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
844 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
845 "vofflock", 0);
846 }
847 fp->f_vnread_flags |= FOFFSET_LOCKED;
848 }
849 res = fp->f_offset;
850 mtx_unlock(mtxp);
851 return (res);
852 }
853
854 void
855 foffset_unlock(struct file *fp, off_t val, int flags)
856 {
857 struct mtx *mtxp;
858
859 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
860
861 mtxp = mtx_pool_find(mtxpool_sleep, fp);
862 mtx_lock(mtxp);
863 if ((flags & FOF_NOUPDATE) == 0)
864 fp->f_offset = val;
865 if ((flags & FOF_NEXTOFF_R) != 0)
866 fp->f_nextoff[UIO_READ] = val;
867 if ((flags & FOF_NEXTOFF_W) != 0)
868 fp->f_nextoff[UIO_WRITE] = val;
869 if ((flags & FOF_NOLOCK) == 0) {
870 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
871 ("Lost FOFFSET_LOCKED"));
872 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
873 wakeup(&fp->f_vnread_flags);
874 fp->f_vnread_flags = 0;
875 }
876 mtx_unlock(mtxp);
877 }
878 #endif
879
880 void
881 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
882 {
883
884 if ((flags & FOF_OFFSET) == 0)
885 uio->uio_offset = foffset_lock(fp, flags);
886 }
887
888 void
889 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
890 {
891
892 if ((flags & FOF_OFFSET) == 0)
893 foffset_unlock(fp, uio->uio_offset, flags);
894 }
895
896 static int
897 get_advice(struct file *fp, struct uio *uio)
898 {
899 struct mtx *mtxp;
900 int ret;
901
902 ret = POSIX_FADV_NORMAL;
903 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
904 return (ret);
905
906 mtxp = mtx_pool_find(mtxpool_sleep, fp);
907 mtx_lock(mtxp);
908 if (fp->f_advice != NULL &&
909 uio->uio_offset >= fp->f_advice->fa_start &&
910 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
911 ret = fp->f_advice->fa_advice;
912 mtx_unlock(mtxp);
913 return (ret);
914 }
915
916 static int
917 get_write_ioflag(struct file *fp)
918 {
919 int ioflag;
920 struct mount *mp;
921 struct vnode *vp;
922
923 ioflag = 0;
924 vp = fp->f_vnode;
925 mp = atomic_load_ptr(&vp->v_mount);
926
927 if ((fp->f_flag & O_DIRECT) != 0)
928 ioflag |= IO_DIRECT;
929
930 if ((fp->f_flag & O_FSYNC) != 0 ||
931 (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
932 ioflag |= IO_SYNC;
933
934 /*
935 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
936 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
937 * fall back to full O_SYNC behavior.
938 */
939 if ((fp->f_flag & O_DSYNC) != 0)
940 ioflag |= IO_SYNC | IO_DATASYNC;
941
942 return (ioflag);
943 }
944
945 int
946 vn_read_from_obj(struct vnode *vp, struct uio *uio)
947 {
948 vm_object_t obj;
949 vm_page_t ma[io_hold_cnt + 2];
950 off_t off, vsz;
951 ssize_t resid;
952 int error, i, j;
953
954 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
955 obj = atomic_load_ptr(&vp->v_object);
956 if (obj == NULL)
957 return (EJUSTRETURN);
958
959 /*
960 * Depends on type stability of vm_objects.
961 */
962 vm_object_pip_add(obj, 1);
963 if ((obj->flags & OBJ_DEAD) != 0) {
964 /*
965 * Note that object might be already reused from the
966 * vnode, and the OBJ_DEAD flag cleared. This is fine,
967 * we recheck for DOOMED vnode state after all pages
968 * are busied, and retract then.
969 *
970 * But we check for OBJ_DEAD to ensure that we do not
971 * busy pages while vm_object_terminate_pages()
972 * processes the queue.
973 */
974 error = EJUSTRETURN;
975 goto out_pip;
976 }
977
978 resid = uio->uio_resid;
979 off = uio->uio_offset;
980 for (i = 0; resid > 0; i++) {
981 MPASS(i < io_hold_cnt + 2);
982 ma[i] = vm_page_grab_unlocked(obj, atop(off),
983 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
984 VM_ALLOC_NOWAIT);
985 if (ma[i] == NULL)
986 break;
987
988 /*
989 * Skip invalid pages. Valid mask can be partial only
990 * at EOF, and we clip later.
991 */
992 if (vm_page_none_valid(ma[i])) {
993 vm_page_sunbusy(ma[i]);
994 break;
995 }
996
997 resid -= PAGE_SIZE;
998 off += PAGE_SIZE;
999 }
1000 if (i == 0) {
1001 error = EJUSTRETURN;
1002 goto out_pip;
1003 }
1004
1005 /*
1006 * Check VIRF_DOOMED after we busied our pages. Since
1007 * vgonel() terminates the vnode' vm_object, it cannot
1008 * process past pages busied by us.
1009 */
1010 if (VN_IS_DOOMED(vp)) {
1011 error = EJUSTRETURN;
1012 goto out;
1013 }
1014
1015 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1016 if (resid > uio->uio_resid)
1017 resid = uio->uio_resid;
1018
1019 /*
1020 * Unlocked read of vnp_size is safe because truncation cannot
1021 * pass busied page. But we load vnp_size into a local
1022 * variable so that possible concurrent extension does not
1023 * break calculation.
1024 */
1025 #if defined(__powerpc__) && !defined(__powerpc64__)
1026 vsz = obj->un_pager.vnp.vnp_size;
1027 #else
1028 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1029 #endif
1030 if (uio->uio_offset >= vsz) {
1031 error = EJUSTRETURN;
1032 goto out;
1033 }
1034 if (uio->uio_offset + resid > vsz)
1035 resid = vsz - uio->uio_offset;
1036
1037 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1038
1039 out:
1040 for (j = 0; j < i; j++) {
1041 if (error == 0)
1042 vm_page_reference(ma[j]);
1043 vm_page_sunbusy(ma[j]);
1044 }
1045 out_pip:
1046 vm_object_pip_wakeup(obj);
1047 if (error != 0)
1048 return (error);
1049 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1050 }
1051
1052 /*
1053 * File table vnode read routine.
1054 */
1055 static int
1056 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1057 struct thread *td)
1058 {
1059 struct vnode *vp;
1060 off_t orig_offset;
1061 int error, ioflag;
1062 int advice;
1063
1064 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1065 uio->uio_td, td));
1066 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1067 vp = fp->f_vnode;
1068 ioflag = 0;
1069 if (fp->f_flag & FNONBLOCK)
1070 ioflag |= IO_NDELAY;
1071 if (fp->f_flag & O_DIRECT)
1072 ioflag |= IO_DIRECT;
1073
1074 /*
1075 * Try to read from page cache. VIRF_DOOMED check is racy but
1076 * allows us to avoid unneeded work outright.
1077 */
1078 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1079 (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1080 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1081 if (error == 0) {
1082 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1083 return (0);
1084 }
1085 if (error != EJUSTRETURN)
1086 return (error);
1087 }
1088
1089 advice = get_advice(fp, uio);
1090 vn_lock(vp, LK_SHARED | LK_RETRY);
1091
1092 switch (advice) {
1093 case POSIX_FADV_NORMAL:
1094 case POSIX_FADV_SEQUENTIAL:
1095 case POSIX_FADV_NOREUSE:
1096 ioflag |= sequential_heuristic(uio, fp);
1097 break;
1098 case POSIX_FADV_RANDOM:
1099 /* Disable read-ahead for random I/O. */
1100 break;
1101 }
1102 orig_offset = uio->uio_offset;
1103
1104 #ifdef MAC
1105 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1106 if (error == 0)
1107 #endif
1108 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1109 fp->f_nextoff[UIO_READ] = uio->uio_offset;
1110 VOP_UNLOCK(vp);
1111 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1112 orig_offset != uio->uio_offset)
1113 /*
1114 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1115 * for the backing file after a POSIX_FADV_NOREUSE
1116 * read(2).
1117 */
1118 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1119 POSIX_FADV_DONTNEED);
1120 return (error);
1121 }
1122
1123 /*
1124 * File table vnode write routine.
1125 */
1126 static int
1127 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1128 struct thread *td)
1129 {
1130 struct vnode *vp;
1131 struct mount *mp;
1132 off_t orig_offset;
1133 int error, ioflag;
1134 int advice;
1135 bool need_finished_write;
1136
1137 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1138 uio->uio_td, td));
1139 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1140 vp = fp->f_vnode;
1141 if (vp->v_type == VREG)
1142 bwillwrite();
1143 ioflag = IO_UNIT;
1144 if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1145 ioflag |= IO_APPEND;
1146 if ((fp->f_flag & FNONBLOCK) != 0)
1147 ioflag |= IO_NDELAY;
1148 ioflag |= get_write_ioflag(fp);
1149
1150 mp = NULL;
1151 need_finished_write = false;
1152 if (vp->v_type != VCHR) {
1153 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1154 if (error != 0)
1155 goto unlock;
1156 need_finished_write = true;
1157 }
1158
1159 advice = get_advice(fp, uio);
1160
1161 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1162 switch (advice) {
1163 case POSIX_FADV_NORMAL:
1164 case POSIX_FADV_SEQUENTIAL:
1165 case POSIX_FADV_NOREUSE:
1166 ioflag |= sequential_heuristic(uio, fp);
1167 break;
1168 case POSIX_FADV_RANDOM:
1169 /* XXX: Is this correct? */
1170 break;
1171 }
1172 orig_offset = uio->uio_offset;
1173
1174 #ifdef MAC
1175 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1176 if (error == 0)
1177 #endif
1178 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1179 fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1180 VOP_UNLOCK(vp);
1181 if (need_finished_write)
1182 vn_finished_write(mp);
1183 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1184 orig_offset != uio->uio_offset)
1185 /*
1186 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1187 * for the backing file after a POSIX_FADV_NOREUSE
1188 * write(2).
1189 */
1190 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1191 POSIX_FADV_DONTNEED);
1192 unlock:
1193 return (error);
1194 }
1195
1196 /*
1197 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1198 * prevent the following deadlock:
1199 *
1200 * Assume that the thread A reads from the vnode vp1 into userspace
1201 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
1202 * currently not resident, then system ends up with the call chain
1203 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1204 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1205 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1206 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1207 * backed by the pages of vnode vp1, and some page in buf2 is not
1208 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1209 *
1210 * To prevent the lock order reversal and deadlock, vn_io_fault() does
1211 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1212 * Instead, it first tries to do the whole range i/o with pagefaults
1213 * disabled. If all pages in the i/o buffer are resident and mapped,
1214 * VOP will succeed (ignoring the genuine filesystem errors).
1215 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1216 * i/o in chunks, with all pages in the chunk prefaulted and held
1217 * using vm_fault_quick_hold_pages().
1218 *
1219 * Filesystems using this deadlock avoidance scheme should use the
1220 * array of the held pages from uio, saved in the curthread->td_ma,
1221 * instead of doing uiomove(). A helper function
1222 * vn_io_fault_uiomove() converts uiomove request into
1223 * uiomove_fromphys() over td_ma array.
1224 *
1225 * Since vnode locks do not cover the whole i/o anymore, rangelocks
1226 * make the current i/o request atomic with respect to other i/os and
1227 * truncations.
1228 */
1229
1230 /*
1231 * Decode vn_io_fault_args and perform the corresponding i/o.
1232 */
1233 static int
1234 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1235 struct thread *td)
1236 {
1237 int error, save;
1238
1239 error = 0;
1240 save = vm_fault_disable_pagefaults();
1241 switch (args->kind) {
1242 case VN_IO_FAULT_FOP:
1243 error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1244 uio, args->cred, args->flags, td);
1245 break;
1246 case VN_IO_FAULT_VOP:
1247 if (uio->uio_rw == UIO_READ) {
1248 error = VOP_READ(args->args.vop_args.vp, uio,
1249 args->flags, args->cred);
1250 } else if (uio->uio_rw == UIO_WRITE) {
1251 error = VOP_WRITE(args->args.vop_args.vp, uio,
1252 args->flags, args->cred);
1253 }
1254 break;
1255 default:
1256 panic("vn_io_fault_doio: unknown kind of io %d %d",
1257 args->kind, uio->uio_rw);
1258 }
1259 vm_fault_enable_pagefaults(save);
1260 return (error);
1261 }
1262
1263 static int
1264 vn_io_fault_touch(char *base, const struct uio *uio)
1265 {
1266 int r;
1267
1268 r = fubyte(base);
1269 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1270 return (EFAULT);
1271 return (0);
1272 }
1273
1274 static int
1275 vn_io_fault_prefault_user(const struct uio *uio)
1276 {
1277 char *base;
1278 const struct iovec *iov;
1279 size_t len;
1280 ssize_t resid;
1281 int error, i;
1282
1283 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1284 ("vn_io_fault_prefault userspace"));
1285
1286 error = i = 0;
1287 iov = uio->uio_iov;
1288 resid = uio->uio_resid;
1289 base = iov->iov_base;
1290 len = iov->iov_len;
1291 while (resid > 0) {
1292 error = vn_io_fault_touch(base, uio);
1293 if (error != 0)
1294 break;
1295 if (len < PAGE_SIZE) {
1296 if (len != 0) {
1297 error = vn_io_fault_touch(base + len - 1, uio);
1298 if (error != 0)
1299 break;
1300 resid -= len;
1301 }
1302 if (++i >= uio->uio_iovcnt)
1303 break;
1304 iov = uio->uio_iov + i;
1305 base = iov->iov_base;
1306 len = iov->iov_len;
1307 } else {
1308 len -= PAGE_SIZE;
1309 base += PAGE_SIZE;
1310 resid -= PAGE_SIZE;
1311 }
1312 }
1313 return (error);
1314 }
1315
1316 /*
1317 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1318 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1319 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1320 * into args and call vn_io_fault1() to handle faults during the user
1321 * mode buffer accesses.
1322 */
1323 static int
1324 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1325 struct thread *td)
1326 {
1327 vm_page_t ma[io_hold_cnt + 2];
1328 struct uio *uio_clone, short_uio;
1329 struct iovec short_iovec[1];
1330 vm_page_t *prev_td_ma;
1331 vm_prot_t prot;
1332 vm_offset_t addr, end;
1333 size_t len, resid;
1334 ssize_t adv;
1335 int error, cnt, saveheld, prev_td_ma_cnt;
1336
1337 if (vn_io_fault_prefault) {
1338 error = vn_io_fault_prefault_user(uio);
1339 if (error != 0)
1340 return (error); /* Or ignore ? */
1341 }
1342
1343 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1344
1345 /*
1346 * The UFS follows IO_UNIT directive and replays back both
1347 * uio_offset and uio_resid if an error is encountered during the
1348 * operation. But, since the iovec may be already advanced,
1349 * uio is still in an inconsistent state.
1350 *
1351 * Cache a copy of the original uio, which is advanced to the redo
1352 * point using UIO_NOCOPY below.
1353 */
1354 uio_clone = cloneuio(uio);
1355 resid = uio->uio_resid;
1356
1357 short_uio.uio_segflg = UIO_USERSPACE;
1358 short_uio.uio_rw = uio->uio_rw;
1359 short_uio.uio_td = uio->uio_td;
1360
1361 error = vn_io_fault_doio(args, uio, td);
1362 if (error != EFAULT)
1363 goto out;
1364
1365 atomic_add_long(&vn_io_faults_cnt, 1);
1366 uio_clone->uio_segflg = UIO_NOCOPY;
1367 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1368 uio_clone->uio_segflg = uio->uio_segflg;
1369
1370 saveheld = curthread_pflags_set(TDP_UIOHELD);
1371 prev_td_ma = td->td_ma;
1372 prev_td_ma_cnt = td->td_ma_cnt;
1373
1374 while (uio_clone->uio_resid != 0) {
1375 len = uio_clone->uio_iov->iov_len;
1376 if (len == 0) {
1377 KASSERT(uio_clone->uio_iovcnt >= 1,
1378 ("iovcnt underflow"));
1379 uio_clone->uio_iov++;
1380 uio_clone->uio_iovcnt--;
1381 continue;
1382 }
1383 if (len > ptoa(io_hold_cnt))
1384 len = ptoa(io_hold_cnt);
1385 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1386 end = round_page(addr + len);
1387 if (end < addr) {
1388 error = EFAULT;
1389 break;
1390 }
1391 cnt = atop(end - trunc_page(addr));
1392 /*
1393 * A perfectly misaligned address and length could cause
1394 * both the start and the end of the chunk to use partial
1395 * page. +2 accounts for such a situation.
1396 */
1397 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1398 addr, len, prot, ma, io_hold_cnt + 2);
1399 if (cnt == -1) {
1400 error = EFAULT;
1401 break;
1402 }
1403 short_uio.uio_iov = &short_iovec[0];
1404 short_iovec[0].iov_base = (void *)addr;
1405 short_uio.uio_iovcnt = 1;
1406 short_uio.uio_resid = short_iovec[0].iov_len = len;
1407 short_uio.uio_offset = uio_clone->uio_offset;
1408 td->td_ma = ma;
1409 td->td_ma_cnt = cnt;
1410
1411 error = vn_io_fault_doio(args, &short_uio, td);
1412 vm_page_unhold_pages(ma, cnt);
1413 adv = len - short_uio.uio_resid;
1414
1415 uio_clone->uio_iov->iov_base =
1416 (char *)uio_clone->uio_iov->iov_base + adv;
1417 uio_clone->uio_iov->iov_len -= adv;
1418 uio_clone->uio_resid -= adv;
1419 uio_clone->uio_offset += adv;
1420
1421 uio->uio_resid -= adv;
1422 uio->uio_offset += adv;
1423
1424 if (error != 0 || adv == 0)
1425 break;
1426 }
1427 td->td_ma = prev_td_ma;
1428 td->td_ma_cnt = prev_td_ma_cnt;
1429 curthread_pflags_restore(saveheld);
1430 out:
1431 free(uio_clone, M_IOV);
1432 return (error);
1433 }
1434
1435 static int
1436 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1437 int flags, struct thread *td)
1438 {
1439 fo_rdwr_t *doio;
1440 struct vnode *vp;
1441 void *rl_cookie;
1442 struct vn_io_fault_args args;
1443 int error;
1444
1445 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1446 vp = fp->f_vnode;
1447
1448 /*
1449 * The ability to read(2) on a directory has historically been
1450 * allowed for all users, but this can and has been the source of
1451 * at least one security issue in the past. As such, it is now hidden
1452 * away behind a sysctl for those that actually need it to use it, and
1453 * restricted to root when it's turned on to make it relatively safe to
1454 * leave on for longer sessions of need.
1455 */
1456 if (vp->v_type == VDIR) {
1457 KASSERT(uio->uio_rw == UIO_READ,
1458 ("illegal write attempted on a directory"));
1459 if (!vfs_allow_read_dir)
1460 return (EISDIR);
1461 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1462 return (EISDIR);
1463 }
1464
1465 foffset_lock_uio(fp, uio, flags);
1466 if (do_vn_io_fault(vp, uio)) {
1467 args.kind = VN_IO_FAULT_FOP;
1468 args.args.fop_args.fp = fp;
1469 args.args.fop_args.doio = doio;
1470 args.cred = active_cred;
1471 args.flags = flags | FOF_OFFSET;
1472 if (uio->uio_rw == UIO_READ) {
1473 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1474 uio->uio_offset + uio->uio_resid);
1475 } else if ((fp->f_flag & O_APPEND) != 0 ||
1476 (flags & FOF_OFFSET) == 0) {
1477 /* For appenders, punt and lock the whole range. */
1478 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1479 } else {
1480 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1481 uio->uio_offset + uio->uio_resid);
1482 }
1483 error = vn_io_fault1(vp, uio, &args, td);
1484 vn_rangelock_unlock(vp, rl_cookie);
1485 } else {
1486 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1487 }
1488 foffset_unlock_uio(fp, uio, flags);
1489 return (error);
1490 }
1491
1492 /*
1493 * Helper function to perform the requested uiomove operation using
1494 * the held pages for io->uio_iov[0].iov_base buffer instead of
1495 * copyin/copyout. Access to the pages with uiomove_fromphys()
1496 * instead of iov_base prevents page faults that could occur due to
1497 * pmap_collect() invalidating the mapping created by
1498 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1499 * object cleanup revoking the write access from page mappings.
1500 *
1501 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1502 * instead of plain uiomove().
1503 */
1504 int
1505 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1506 {
1507 struct uio transp_uio;
1508 struct iovec transp_iov[1];
1509 struct thread *td;
1510 size_t adv;
1511 int error, pgadv;
1512
1513 td = curthread;
1514 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1515 uio->uio_segflg != UIO_USERSPACE)
1516 return (uiomove(data, xfersize, uio));
1517
1518 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1519 transp_iov[0].iov_base = data;
1520 transp_uio.uio_iov = &transp_iov[0];
1521 transp_uio.uio_iovcnt = 1;
1522 if (xfersize > uio->uio_resid)
1523 xfersize = uio->uio_resid;
1524 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1525 transp_uio.uio_offset = 0;
1526 transp_uio.uio_segflg = UIO_SYSSPACE;
1527 /*
1528 * Since transp_iov points to data, and td_ma page array
1529 * corresponds to original uio->uio_iov, we need to invert the
1530 * direction of the i/o operation as passed to
1531 * uiomove_fromphys().
1532 */
1533 switch (uio->uio_rw) {
1534 case UIO_WRITE:
1535 transp_uio.uio_rw = UIO_READ;
1536 break;
1537 case UIO_READ:
1538 transp_uio.uio_rw = UIO_WRITE;
1539 break;
1540 }
1541 transp_uio.uio_td = uio->uio_td;
1542 error = uiomove_fromphys(td->td_ma,
1543 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1544 xfersize, &transp_uio);
1545 adv = xfersize - transp_uio.uio_resid;
1546 pgadv =
1547 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1548 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1549 td->td_ma += pgadv;
1550 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1551 pgadv));
1552 td->td_ma_cnt -= pgadv;
1553 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1554 uio->uio_iov->iov_len -= adv;
1555 uio->uio_resid -= adv;
1556 uio->uio_offset += adv;
1557 return (error);
1558 }
1559
1560 int
1561 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1562 struct uio *uio)
1563 {
1564 struct thread *td;
1565 vm_offset_t iov_base;
1566 int cnt, pgadv;
1567
1568 td = curthread;
1569 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1570 uio->uio_segflg != UIO_USERSPACE)
1571 return (uiomove_fromphys(ma, offset, xfersize, uio));
1572
1573 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1574 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1575 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1576 switch (uio->uio_rw) {
1577 case UIO_WRITE:
1578 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1579 offset, cnt);
1580 break;
1581 case UIO_READ:
1582 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1583 cnt);
1584 break;
1585 }
1586 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1587 td->td_ma += pgadv;
1588 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1589 pgadv));
1590 td->td_ma_cnt -= pgadv;
1591 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1592 uio->uio_iov->iov_len -= cnt;
1593 uio->uio_resid -= cnt;
1594 uio->uio_offset += cnt;
1595 return (0);
1596 }
1597
1598 /*
1599 * File table truncate routine.
1600 */
1601 static int
1602 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1603 struct thread *td)
1604 {
1605 struct mount *mp;
1606 struct vnode *vp;
1607 void *rl_cookie;
1608 int error;
1609
1610 vp = fp->f_vnode;
1611
1612 retry:
1613 /*
1614 * Lock the whole range for truncation. Otherwise split i/o
1615 * might happen partly before and partly after the truncation.
1616 */
1617 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1618 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1619 if (error)
1620 goto out1;
1621 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1622 AUDIT_ARG_VNODE1(vp);
1623 if (vp->v_type == VDIR) {
1624 error = EISDIR;
1625 goto out;
1626 }
1627 #ifdef MAC
1628 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1629 if (error)
1630 goto out;
1631 #endif
1632 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1633 fp->f_cred);
1634 out:
1635 VOP_UNLOCK(vp);
1636 vn_finished_write(mp);
1637 out1:
1638 vn_rangelock_unlock(vp, rl_cookie);
1639 if (error == ERELOOKUP)
1640 goto retry;
1641 return (error);
1642 }
1643
1644 /*
1645 * Truncate a file that is already locked.
1646 */
1647 int
1648 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1649 struct ucred *cred)
1650 {
1651 struct vattr vattr;
1652 int error;
1653
1654 error = VOP_ADD_WRITECOUNT(vp, 1);
1655 if (error == 0) {
1656 VATTR_NULL(&vattr);
1657 vattr.va_size = length;
1658 if (sync)
1659 vattr.va_vaflags |= VA_SYNC;
1660 error = VOP_SETATTR(vp, &vattr, cred);
1661 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1662 }
1663 return (error);
1664 }
1665
1666 /*
1667 * File table vnode stat routine.
1668 */
1669 int
1670 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1671 struct thread *td)
1672 {
1673 struct vnode *vp = fp->f_vnode;
1674 int error;
1675
1676 vn_lock(vp, LK_SHARED | LK_RETRY);
1677 error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1678 VOP_UNLOCK(vp);
1679
1680 return (error);
1681 }
1682
1683 /*
1684 * File table vnode ioctl routine.
1685 */
1686 static int
1687 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1688 struct thread *td)
1689 {
1690 struct vattr vattr;
1691 struct vnode *vp;
1692 struct fiobmap2_arg *bmarg;
1693 int error;
1694
1695 vp = fp->f_vnode;
1696 switch (vp->v_type) {
1697 case VDIR:
1698 case VREG:
1699 switch (com) {
1700 case FIONREAD:
1701 vn_lock(vp, LK_SHARED | LK_RETRY);
1702 error = VOP_GETATTR(vp, &vattr, active_cred);
1703 VOP_UNLOCK(vp);
1704 if (error == 0)
1705 *(int *)data = vattr.va_size - fp->f_offset;
1706 return (error);
1707 case FIOBMAP2:
1708 bmarg = (struct fiobmap2_arg *)data;
1709 vn_lock(vp, LK_SHARED | LK_RETRY);
1710 #ifdef MAC
1711 error = mac_vnode_check_read(active_cred, fp->f_cred,
1712 vp);
1713 if (error == 0)
1714 #endif
1715 error = VOP_BMAP(vp, bmarg->bn, NULL,
1716 &bmarg->bn, &bmarg->runp, &bmarg->runb);
1717 VOP_UNLOCK(vp);
1718 return (error);
1719 case FIONBIO:
1720 case FIOASYNC:
1721 return (0);
1722 default:
1723 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1724 active_cred, td));
1725 }
1726 break;
1727 case VCHR:
1728 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1729 active_cred, td));
1730 default:
1731 return (ENOTTY);
1732 }
1733 }
1734
1735 /*
1736 * File table vnode poll routine.
1737 */
1738 static int
1739 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1740 struct thread *td)
1741 {
1742 struct vnode *vp;
1743 int error;
1744
1745 vp = fp->f_vnode;
1746 #if defined(MAC) || defined(AUDIT)
1747 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1748 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1749 AUDIT_ARG_VNODE1(vp);
1750 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1751 VOP_UNLOCK(vp);
1752 if (error != 0)
1753 return (error);
1754 }
1755 #endif
1756 error = VOP_POLL(vp, events, fp->f_cred, td);
1757 return (error);
1758 }
1759
1760 /*
1761 * Acquire the requested lock and then check for validity. LK_RETRY
1762 * permits vn_lock to return doomed vnodes.
1763 */
1764 static int __noinline
1765 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1766 int error)
1767 {
1768
1769 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1770 ("vn_lock: error %d incompatible with flags %#x", error, flags));
1771
1772 if (error == 0)
1773 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1774
1775 if ((flags & LK_RETRY) == 0) {
1776 if (error == 0) {
1777 VOP_UNLOCK(vp);
1778 error = ENOENT;
1779 }
1780 return (error);
1781 }
1782
1783 /*
1784 * LK_RETRY case.
1785 *
1786 * Nothing to do if we got the lock.
1787 */
1788 if (error == 0)
1789 return (0);
1790
1791 /*
1792 * Interlock was dropped by the call in _vn_lock.
1793 */
1794 flags &= ~LK_INTERLOCK;
1795 do {
1796 error = VOP_LOCK1(vp, flags, file, line);
1797 } while (error != 0);
1798 return (0);
1799 }
1800
1801 int
1802 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1803 {
1804 int error;
1805
1806 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1807 ("vn_lock: no locktype (%d passed)", flags));
1808 VNPASS(vp->v_holdcnt > 0, vp);
1809 error = VOP_LOCK1(vp, flags, file, line);
1810 if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1811 return (_vn_lock_fallback(vp, flags, file, line, error));
1812 return (0);
1813 }
1814
1815 /*
1816 * File table vnode close routine.
1817 */
1818 static int
1819 vn_closefile(struct file *fp, struct thread *td)
1820 {
1821 struct vnode *vp;
1822 struct flock lf;
1823 int error;
1824 bool ref;
1825
1826 vp = fp->f_vnode;
1827 fp->f_ops = &badfileops;
1828 ref = (fp->f_flag & FHASLOCK) != 0;
1829
1830 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1831
1832 if (__predict_false(ref)) {
1833 lf.l_whence = SEEK_SET;
1834 lf.l_start = 0;
1835 lf.l_len = 0;
1836 lf.l_type = F_UNLCK;
1837 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1838 vrele(vp);
1839 }
1840 return (error);
1841 }
1842
1843 /*
1844 * Preparing to start a filesystem write operation. If the operation is
1845 * permitted, then we bump the count of operations in progress and
1846 * proceed. If a suspend request is in progress, we wait until the
1847 * suspension is over, and then proceed.
1848 */
1849 static int
1850 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1851 {
1852 struct mount_pcpu *mpcpu;
1853 int error, mflags;
1854
1855 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1856 vfs_op_thread_enter(mp, mpcpu)) {
1857 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1858 vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1859 vfs_op_thread_exit(mp, mpcpu);
1860 return (0);
1861 }
1862
1863 if (mplocked)
1864 mtx_assert(MNT_MTX(mp), MA_OWNED);
1865 else
1866 MNT_ILOCK(mp);
1867
1868 error = 0;
1869
1870 /*
1871 * Check on status of suspension.
1872 */
1873 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1874 mp->mnt_susp_owner != curthread) {
1875 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1876 (flags & PCATCH) : 0) | (PUSER - 1);
1877 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1878 if (flags & V_NOWAIT) {
1879 error = EWOULDBLOCK;
1880 goto unlock;
1881 }
1882 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1883 "suspfs", 0);
1884 if (error)
1885 goto unlock;
1886 }
1887 }
1888 if (flags & V_XSLEEP)
1889 goto unlock;
1890 mp->mnt_writeopcount++;
1891 unlock:
1892 if (error != 0 || (flags & V_XSLEEP) != 0)
1893 MNT_REL(mp);
1894 MNT_IUNLOCK(mp);
1895 return (error);
1896 }
1897
1898 int
1899 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1900 {
1901 struct mount *mp;
1902 int error;
1903
1904 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1905 ("V_MNTREF requires mp"));
1906
1907 error = 0;
1908 /*
1909 * If a vnode is provided, get and return the mount point that
1910 * to which it will write.
1911 */
1912 if (vp != NULL) {
1913 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1914 *mpp = NULL;
1915 if (error != EOPNOTSUPP)
1916 return (error);
1917 return (0);
1918 }
1919 }
1920 if ((mp = *mpp) == NULL)
1921 return (0);
1922
1923 /*
1924 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1925 * a vfs_ref().
1926 * As long as a vnode is not provided we need to acquire a
1927 * refcount for the provided mountpoint too, in order to
1928 * emulate a vfs_ref().
1929 */
1930 if (vp == NULL && (flags & V_MNTREF) == 0)
1931 vfs_ref(mp);
1932
1933 return (vn_start_write_refed(mp, flags, false));
1934 }
1935
1936 /*
1937 * Secondary suspension. Used by operations such as vop_inactive
1938 * routines that are needed by the higher level functions. These
1939 * are allowed to proceed until all the higher level functions have
1940 * completed (indicated by mnt_writeopcount dropping to zero). At that
1941 * time, these operations are halted until the suspension is over.
1942 */
1943 int
1944 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1945 {
1946 struct mount *mp;
1947 int error;
1948
1949 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1950 ("V_MNTREF requires mp"));
1951
1952 retry:
1953 if (vp != NULL) {
1954 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1955 *mpp = NULL;
1956 if (error != EOPNOTSUPP)
1957 return (error);
1958 return (0);
1959 }
1960 }
1961 /*
1962 * If we are not suspended or have not yet reached suspended
1963 * mode, then let the operation proceed.
1964 */
1965 if ((mp = *mpp) == NULL)
1966 return (0);
1967
1968 /*
1969 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1970 * a vfs_ref().
1971 * As long as a vnode is not provided we need to acquire a
1972 * refcount for the provided mountpoint too, in order to
1973 * emulate a vfs_ref().
1974 */
1975 MNT_ILOCK(mp);
1976 if (vp == NULL && (flags & V_MNTREF) == 0)
1977 MNT_REF(mp);
1978 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1979 mp->mnt_secondary_writes++;
1980 mp->mnt_secondary_accwrites++;
1981 MNT_IUNLOCK(mp);
1982 return (0);
1983 }
1984 if (flags & V_NOWAIT) {
1985 MNT_REL(mp);
1986 MNT_IUNLOCK(mp);
1987 return (EWOULDBLOCK);
1988 }
1989 /*
1990 * Wait for the suspension to finish.
1991 */
1992 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1993 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1994 "suspfs", 0);
1995 vfs_rel(mp);
1996 if (error == 0)
1997 goto retry;
1998 return (error);
1999 }
2000
2001 /*
2002 * Filesystem write operation has completed. If we are suspending and this
2003 * operation is the last one, notify the suspender that the suspension is
2004 * now in effect.
2005 */
2006 void
2007 vn_finished_write(struct mount *mp)
2008 {
2009 struct mount_pcpu *mpcpu;
2010 int c;
2011
2012 if (mp == NULL)
2013 return;
2014
2015 if (vfs_op_thread_enter(mp, mpcpu)) {
2016 vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2017 vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2018 vfs_op_thread_exit(mp, mpcpu);
2019 return;
2020 }
2021
2022 MNT_ILOCK(mp);
2023 vfs_assert_mount_counters(mp);
2024 MNT_REL(mp);
2025 c = --mp->mnt_writeopcount;
2026 if (mp->mnt_vfs_ops == 0) {
2027 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2028 MNT_IUNLOCK(mp);
2029 return;
2030 }
2031 if (c < 0)
2032 vfs_dump_mount_counters(mp);
2033 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2034 wakeup(&mp->mnt_writeopcount);
2035 MNT_IUNLOCK(mp);
2036 }
2037
2038 /*
2039 * Filesystem secondary write operation has completed. If we are
2040 * suspending and this operation is the last one, notify the suspender
2041 * that the suspension is now in effect.
2042 */
2043 void
2044 vn_finished_secondary_write(struct mount *mp)
2045 {
2046 if (mp == NULL)
2047 return;
2048 MNT_ILOCK(mp);
2049 MNT_REL(mp);
2050 mp->mnt_secondary_writes--;
2051 if (mp->mnt_secondary_writes < 0)
2052 panic("vn_finished_secondary_write: neg cnt");
2053 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2054 mp->mnt_secondary_writes <= 0)
2055 wakeup(&mp->mnt_secondary_writes);
2056 MNT_IUNLOCK(mp);
2057 }
2058
2059 /*
2060 * Request a filesystem to suspend write operations.
2061 */
2062 int
2063 vfs_write_suspend(struct mount *mp, int flags)
2064 {
2065 int error;
2066
2067 vfs_op_enter(mp);
2068
2069 MNT_ILOCK(mp);
2070 vfs_assert_mount_counters(mp);
2071 if (mp->mnt_susp_owner == curthread) {
2072 vfs_op_exit_locked(mp);
2073 MNT_IUNLOCK(mp);
2074 return (EALREADY);
2075 }
2076 while (mp->mnt_kern_flag & MNTK_SUSPEND)
2077 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2078
2079 /*
2080 * Unmount holds a write reference on the mount point. If we
2081 * own busy reference and drain for writers, we deadlock with
2082 * the reference draining in the unmount path. Callers of
2083 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2084 * vfs_busy() reference is owned and caller is not in the
2085 * unmount context.
2086 */
2087 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2088 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2089 vfs_op_exit_locked(mp);
2090 MNT_IUNLOCK(mp);
2091 return (EBUSY);
2092 }
2093
2094 mp->mnt_kern_flag |= MNTK_SUSPEND;
2095 mp->mnt_susp_owner = curthread;
2096 if (mp->mnt_writeopcount > 0)
2097 (void) msleep(&mp->mnt_writeopcount,
2098 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2099 else
2100 MNT_IUNLOCK(mp);
2101 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2102 vfs_write_resume(mp, 0);
2103 /* vfs_write_resume does vfs_op_exit() for us */
2104 }
2105 return (error);
2106 }
2107
2108 /*
2109 * Request a filesystem to resume write operations.
2110 */
2111 void
2112 vfs_write_resume(struct mount *mp, int flags)
2113 {
2114
2115 MNT_ILOCK(mp);
2116 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2117 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2118 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2119 MNTK_SUSPENDED);
2120 mp->mnt_susp_owner = NULL;
2121 wakeup(&mp->mnt_writeopcount);
2122 wakeup(&mp->mnt_flag);
2123 curthread->td_pflags &= ~TDP_IGNSUSP;
2124 if ((flags & VR_START_WRITE) != 0) {
2125 MNT_REF(mp);
2126 mp->mnt_writeopcount++;
2127 }
2128 MNT_IUNLOCK(mp);
2129 if ((flags & VR_NO_SUSPCLR) == 0)
2130 VFS_SUSP_CLEAN(mp);
2131 vfs_op_exit(mp);
2132 } else if ((flags & VR_START_WRITE) != 0) {
2133 MNT_REF(mp);
2134 vn_start_write_refed(mp, 0, true);
2135 } else {
2136 MNT_IUNLOCK(mp);
2137 }
2138 }
2139
2140 /*
2141 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2142 * methods.
2143 */
2144 int
2145 vfs_write_suspend_umnt(struct mount *mp)
2146 {
2147 int error;
2148
2149 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2150 ("vfs_write_suspend_umnt: recursed"));
2151
2152 /* dounmount() already called vn_start_write(). */
2153 for (;;) {
2154 vn_finished_write(mp);
2155 error = vfs_write_suspend(mp, 0);
2156 if (error != 0) {
2157 vn_start_write(NULL, &mp, V_WAIT);
2158 return (error);
2159 }
2160 MNT_ILOCK(mp);
2161 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2162 break;
2163 MNT_IUNLOCK(mp);
2164 vn_start_write(NULL, &mp, V_WAIT);
2165 }
2166 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2167 wakeup(&mp->mnt_flag);
2168 MNT_IUNLOCK(mp);
2169 curthread->td_pflags |= TDP_IGNSUSP;
2170 return (0);
2171 }
2172
2173 /*
2174 * Implement kqueues for files by translating it to vnode operation.
2175 */
2176 static int
2177 vn_kqfilter(struct file *fp, struct knote *kn)
2178 {
2179
2180 return (VOP_KQFILTER(fp->f_vnode, kn));
2181 }
2182
2183 int
2184 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2185 {
2186 if ((fp->f_flag & FKQALLOWED) == 0)
2187 return (EBADF);
2188 return (vn_kqfilter(fp, kn));
2189 }
2190
2191 /*
2192 * Simplified in-kernel wrapper calls for extended attribute access.
2193 * Both calls pass in a NULL credential, authorizing as "kernel" access.
2194 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2195 */
2196 int
2197 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2198 const char *attrname, int *buflen, char *buf, struct thread *td)
2199 {
2200 struct uio auio;
2201 struct iovec iov;
2202 int error;
2203
2204 iov.iov_len = *buflen;
2205 iov.iov_base = buf;
2206
2207 auio.uio_iov = &iov;
2208 auio.uio_iovcnt = 1;
2209 auio.uio_rw = UIO_READ;
2210 auio.uio_segflg = UIO_SYSSPACE;
2211 auio.uio_td = td;
2212 auio.uio_offset = 0;
2213 auio.uio_resid = *buflen;
2214
2215 if ((ioflg & IO_NODELOCKED) == 0)
2216 vn_lock(vp, LK_SHARED | LK_RETRY);
2217
2218 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2219
2220 /* authorize attribute retrieval as kernel */
2221 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2222 td);
2223
2224 if ((ioflg & IO_NODELOCKED) == 0)
2225 VOP_UNLOCK(vp);
2226
2227 if (error == 0) {
2228 *buflen = *buflen - auio.uio_resid;
2229 }
2230
2231 return (error);
2232 }
2233
2234 /*
2235 * XXX failure mode if partially written?
2236 */
2237 int
2238 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2239 const char *attrname, int buflen, char *buf, struct thread *td)
2240 {
2241 struct uio auio;
2242 struct iovec iov;
2243 struct mount *mp;
2244 int error;
2245
2246 iov.iov_len = buflen;
2247 iov.iov_base = buf;
2248
2249 auio.uio_iov = &iov;
2250 auio.uio_iovcnt = 1;
2251 auio.uio_rw = UIO_WRITE;
2252 auio.uio_segflg = UIO_SYSSPACE;
2253 auio.uio_td = td;
2254 auio.uio_offset = 0;
2255 auio.uio_resid = buflen;
2256
2257 if ((ioflg & IO_NODELOCKED) == 0) {
2258 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2259 return (error);
2260 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2261 }
2262
2263 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2264
2265 /* authorize attribute setting as kernel */
2266 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2267
2268 if ((ioflg & IO_NODELOCKED) == 0) {
2269 vn_finished_write(mp);
2270 VOP_UNLOCK(vp);
2271 }
2272
2273 return (error);
2274 }
2275
2276 int
2277 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2278 const char *attrname, struct thread *td)
2279 {
2280 struct mount *mp;
2281 int error;
2282
2283 if ((ioflg & IO_NODELOCKED) == 0) {
2284 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2285 return (error);
2286 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2287 }
2288
2289 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2290
2291 /* authorize attribute removal as kernel */
2292 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2293 if (error == EOPNOTSUPP)
2294 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2295 NULL, td);
2296
2297 if ((ioflg & IO_NODELOCKED) == 0) {
2298 vn_finished_write(mp);
2299 VOP_UNLOCK(vp);
2300 }
2301
2302 return (error);
2303 }
2304
2305 static int
2306 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2307 struct vnode **rvp)
2308 {
2309
2310 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2311 }
2312
2313 int
2314 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2315 {
2316
2317 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2318 lkflags, rvp));
2319 }
2320
2321 int
2322 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2323 int lkflags, struct vnode **rvp)
2324 {
2325 struct mount *mp;
2326 int ltype, error;
2327
2328 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2329 mp = vp->v_mount;
2330 ltype = VOP_ISLOCKED(vp);
2331 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2332 ("vn_vget_ino: vp not locked"));
2333 error = vfs_busy(mp, MBF_NOWAIT);
2334 if (error != 0) {
2335 vfs_ref(mp);
2336 VOP_UNLOCK(vp);
2337 error = vfs_busy(mp, 0);
2338 vn_lock(vp, ltype | LK_RETRY);
2339 vfs_rel(mp);
2340 if (error != 0)
2341 return (ENOENT);
2342 if (VN_IS_DOOMED(vp)) {
2343 vfs_unbusy(mp);
2344 return (ENOENT);
2345 }
2346 }
2347 VOP_UNLOCK(vp);
2348 error = alloc(mp, alloc_arg, lkflags, rvp);
2349 vfs_unbusy(mp);
2350 if (error != 0 || *rvp != vp)
2351 vn_lock(vp, ltype | LK_RETRY);
2352 if (VN_IS_DOOMED(vp)) {
2353 if (error == 0) {
2354 if (*rvp == vp)
2355 vunref(vp);
2356 else
2357 vput(*rvp);
2358 }
2359 error = ENOENT;
2360 }
2361 return (error);
2362 }
2363
2364 static void
2365 vn_send_sigxfsz(struct proc *p)
2366 {
2367 PROC_LOCK(p);
2368 kern_psignal(p, SIGXFSZ);
2369 PROC_UNLOCK(p);
2370 }
2371
2372 int
2373 vn_rlimit_trunc(u_quad_t size, struct thread *td)
2374 {
2375 if (size <= lim_cur(td, RLIMIT_FSIZE))
2376 return (0);
2377 vn_send_sigxfsz(td->td_proc);
2378 return (EFBIG);
2379 }
2380
2381 static int
2382 vn_rlimit_fsizex1(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2383 bool adj, struct thread *td)
2384 {
2385 off_t lim;
2386 bool ktr_write;
2387
2388 if (vp->v_type != VREG)
2389 return (0);
2390
2391 /*
2392 * Handle file system maximum file size.
2393 */
2394 if (maxfsz != 0 && uio->uio_offset + uio->uio_resid > maxfsz) {
2395 if (!adj || uio->uio_offset >= maxfsz)
2396 return (EFBIG);
2397 uio->uio_resid = maxfsz - uio->uio_offset;
2398 }
2399
2400 /*
2401 * This is kernel write (e.g. vnode_pager) or accounting
2402 * write, ignore limit.
2403 */
2404 if (td == NULL || (td->td_pflags2 & TDP2_ACCT) != 0)
2405 return (0);
2406
2407 /*
2408 * Calculate file size limit.
2409 */
2410 ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2411 lim = __predict_false(ktr_write) ? td->td_ktr_io_lim :
2412 lim_cur(td, RLIMIT_FSIZE);
2413
2414 /*
2415 * Is the limit reached?
2416 */
2417 if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2418 return (0);
2419
2420 /*
2421 * Prepared filesystems can handle writes truncated to the
2422 * file size limit.
2423 */
2424 if (adj && (uoff_t)uio->uio_offset < lim) {
2425 uio->uio_resid = lim - (uoff_t)uio->uio_offset;
2426 return (0);
2427 }
2428
2429 if (!ktr_write || ktr_filesize_limit_signal)
2430 vn_send_sigxfsz(td->td_proc);
2431 return (EFBIG);
2432 }
2433
2434 /*
2435 * Helper for VOP_WRITE() implementations, the common code to
2436 * handle maximum supported file size on the filesystem, and
2437 * RLIMIT_FSIZE, except for special writes from accounting subsystem
2438 * and ktrace.
2439 *
2440 * For maximum file size (maxfsz argument):
2441 * - return EFBIG if uio_offset is beyond it
2442 * - otherwise, clamp uio_resid if write would extend file beyond maxfsz.
2443 *
2444 * For RLIMIT_FSIZE:
2445 * - return EFBIG and send SIGXFSZ if uio_offset is beyond the limit
2446 * - otherwise, clamp uio_resid if write would extend file beyond limit.
2447 *
2448 * If clamping occured, the adjustment for uio_resid is stored in
2449 * *resid_adj, to be re-applied by vn_rlimit_fsizex_res() on return
2450 * from the VOP.
2451 */
2452 int
2453 vn_rlimit_fsizex(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2454 ssize_t *resid_adj, struct thread *td)
2455 {
2456 ssize_t resid_orig;
2457 int error;
2458 bool adj;
2459
2460 resid_orig = uio->uio_resid;
2461 adj = resid_adj != NULL;
2462 error = vn_rlimit_fsizex1(vp, uio, maxfsz, adj, td);
2463 if (adj)
2464 *resid_adj = resid_orig - uio->uio_resid;
2465 return (error);
2466 }
2467
2468 void
2469 vn_rlimit_fsizex_res(struct uio *uio, ssize_t resid_adj)
2470 {
2471 uio->uio_resid += resid_adj;
2472 }
2473
2474 int
2475 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2476 struct thread *td)
2477 {
2478 return (vn_rlimit_fsizex(vp, __DECONST(struct uio *, uio), 0, NULL,
2479 td));
2480 }
2481
2482 int
2483 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2484 struct thread *td)
2485 {
2486 struct vnode *vp;
2487
2488 vp = fp->f_vnode;
2489 #ifdef AUDIT
2490 vn_lock(vp, LK_SHARED | LK_RETRY);
2491 AUDIT_ARG_VNODE1(vp);
2492 VOP_UNLOCK(vp);
2493 #endif
2494 return (setfmode(td, active_cred, vp, mode));
2495 }
2496
2497 int
2498 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2499 struct thread *td)
2500 {
2501 struct vnode *vp;
2502
2503 vp = fp->f_vnode;
2504 #ifdef AUDIT
2505 vn_lock(vp, LK_SHARED | LK_RETRY);
2506 AUDIT_ARG_VNODE1(vp);
2507 VOP_UNLOCK(vp);
2508 #endif
2509 return (setfown(td, active_cred, vp, uid, gid));
2510 }
2511
2512 /*
2513 * Remove pages in the range ["start", "end") from the vnode's VM object. If
2514 * "end" is 0, then the range extends to the end of the object.
2515 */
2516 void
2517 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2518 {
2519 vm_object_t object;
2520
2521 if ((object = vp->v_object) == NULL)
2522 return;
2523 VM_OBJECT_WLOCK(object);
2524 vm_object_page_remove(object, start, end, 0);
2525 VM_OBJECT_WUNLOCK(object);
2526 }
2527
2528 /*
2529 * Like vn_pages_remove(), but skips invalid pages, which by definition are not
2530 * mapped into any process' address space. Filesystems may use this in
2531 * preference to vn_pages_remove() to avoid blocking on pages busied in
2532 * preparation for a VOP_GETPAGES.
2533 */
2534 void
2535 vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2536 {
2537 vm_object_t object;
2538
2539 if ((object = vp->v_object) == NULL)
2540 return;
2541 VM_OBJECT_WLOCK(object);
2542 vm_object_page_remove(object, start, end, OBJPR_VALIDONLY);
2543 VM_OBJECT_WUNLOCK(object);
2544 }
2545
2546 int
2547 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2548 {
2549 struct vattr va;
2550 daddr_t bn, bnp;
2551 uint64_t bsize;
2552 off_t noff;
2553 int error;
2554
2555 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2556 ("Wrong command %lu", cmd));
2557
2558 if (vn_lock(vp, LK_SHARED) != 0)
2559 return (EBADF);
2560 if (vp->v_type != VREG) {
2561 error = ENOTTY;
2562 goto unlock;
2563 }
2564 error = VOP_GETATTR(vp, &va, cred);
2565 if (error != 0)
2566 goto unlock;
2567 noff = *off;
2568 if (noff < 0 || noff >= va.va_size) {
2569 error = ENXIO;
2570 goto unlock;
2571 }
2572 bsize = vp->v_mount->mnt_stat.f_iosize;
2573 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2574 noff % bsize) {
2575 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2576 if (error == EOPNOTSUPP) {
2577 error = ENOTTY;
2578 goto unlock;
2579 }
2580 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2581 (bnp != -1 && cmd == FIOSEEKDATA)) {
2582 noff = bn * bsize;
2583 if (noff < *off)
2584 noff = *off;
2585 goto unlock;
2586 }
2587 }
2588 if (noff > va.va_size)
2589 noff = va.va_size;
2590 /* noff == va.va_size. There is an implicit hole at the end of file. */
2591 if (cmd == FIOSEEKDATA)
2592 error = ENXIO;
2593 unlock:
2594 VOP_UNLOCK(vp);
2595 if (error == 0)
2596 *off = noff;
2597 return (error);
2598 }
2599
2600 int
2601 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2602 {
2603 struct ucred *cred;
2604 struct vnode *vp;
2605 struct vattr vattr;
2606 off_t foffset, size;
2607 int error, noneg;
2608
2609 cred = td->td_ucred;
2610 vp = fp->f_vnode;
2611 foffset = foffset_lock(fp, 0);
2612 noneg = (vp->v_type != VCHR);
2613 error = 0;
2614 switch (whence) {
2615 case L_INCR:
2616 if (noneg &&
2617 (foffset < 0 ||
2618 (offset > 0 && foffset > OFF_MAX - offset))) {
2619 error = EOVERFLOW;
2620 break;
2621 }
2622 offset += foffset;
2623 break;
2624 case L_XTND:
2625 vn_lock(vp, LK_SHARED | LK_RETRY);
2626 error = VOP_GETATTR(vp, &vattr, cred);
2627 VOP_UNLOCK(vp);
2628 if (error)
2629 break;
2630
2631 /*
2632 * If the file references a disk device, then fetch
2633 * the media size and use that to determine the ending
2634 * offset.
2635 */
2636 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2637 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2638 vattr.va_size = size;
2639 if (noneg &&
2640 (vattr.va_size > OFF_MAX ||
2641 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2642 error = EOVERFLOW;
2643 break;
2644 }
2645 offset += vattr.va_size;
2646 break;
2647 case L_SET:
2648 break;
2649 case SEEK_DATA:
2650 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2651 if (error == ENOTTY)
2652 error = EINVAL;
2653 break;
2654 case SEEK_HOLE:
2655 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2656 if (error == ENOTTY)
2657 error = EINVAL;
2658 break;
2659 default:
2660 error = EINVAL;
2661 }
2662 if (error == 0 && noneg && offset < 0)
2663 error = EINVAL;
2664 if (error != 0)
2665 goto drop;
2666 VFS_KNOTE_UNLOCKED(vp, 0);
2667 td->td_uretoff.tdu_off = offset;
2668 drop:
2669 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2670 return (error);
2671 }
2672
2673 int
2674 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2675 struct thread *td)
2676 {
2677 int error;
2678
2679 /*
2680 * Grant permission if the caller is the owner of the file, or
2681 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2682 * on the file. If the time pointer is null, then write
2683 * permission on the file is also sufficient.
2684 *
2685 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2686 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2687 * will be allowed to set the times [..] to the current
2688 * server time.
2689 */
2690 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2691 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2692 error = VOP_ACCESS(vp, VWRITE, cred, td);
2693 return (error);
2694 }
2695
2696 int
2697 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2698 {
2699 struct vnode *vp;
2700 int error;
2701
2702 if (fp->f_type == DTYPE_FIFO)
2703 kif->kf_type = KF_TYPE_FIFO;
2704 else
2705 kif->kf_type = KF_TYPE_VNODE;
2706 vp = fp->f_vnode;
2707 vref(vp);
2708 FILEDESC_SUNLOCK(fdp);
2709 error = vn_fill_kinfo_vnode(vp, kif);
2710 vrele(vp);
2711 FILEDESC_SLOCK(fdp);
2712 return (error);
2713 }
2714
2715 static inline void
2716 vn_fill_junk(struct kinfo_file *kif)
2717 {
2718 size_t len, olen;
2719
2720 /*
2721 * Simulate vn_fullpath returning changing values for a given
2722 * vp during e.g. coredump.
2723 */
2724 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2725 olen = strlen(kif->kf_path);
2726 if (len < olen)
2727 strcpy(&kif->kf_path[len - 1], "$");
2728 else
2729 for (; olen < len; olen++)
2730 strcpy(&kif->kf_path[olen], "A");
2731 }
2732
2733 int
2734 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2735 {
2736 struct vattr va;
2737 char *fullpath, *freepath;
2738 int error;
2739
2740 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2741 freepath = NULL;
2742 fullpath = "-";
2743 error = vn_fullpath(vp, &fullpath, &freepath);
2744 if (error == 0) {
2745 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2746 }
2747 if (freepath != NULL)
2748 free(freepath, M_TEMP);
2749
2750 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2751 vn_fill_junk(kif);
2752 );
2753
2754 /*
2755 * Retrieve vnode attributes.
2756 */
2757 va.va_fsid = VNOVAL;
2758 va.va_rdev = NODEV;
2759 vn_lock(vp, LK_SHARED | LK_RETRY);
2760 error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2761 VOP_UNLOCK(vp);
2762 if (error != 0)
2763 return (error);
2764 if (va.va_fsid != VNOVAL)
2765 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2766 else
2767 kif->kf_un.kf_file.kf_file_fsid =
2768 vp->v_mount->mnt_stat.f_fsid.val[0];
2769 kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2770 kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2771 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2772 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2773 kif->kf_un.kf_file.kf_file_size = va.va_size;
2774 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2775 kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2776 kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2777 return (0);
2778 }
2779
2780 int
2781 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2782 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2783 struct thread *td)
2784 {
2785 #ifdef HWPMC_HOOKS
2786 struct pmckern_map_in pkm;
2787 #endif
2788 struct mount *mp;
2789 struct vnode *vp;
2790 vm_object_t object;
2791 vm_prot_t maxprot;
2792 boolean_t writecounted;
2793 int error;
2794
2795 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2796 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2797 /*
2798 * POSIX shared-memory objects are defined to have
2799 * kernel persistence, and are not defined to support
2800 * read(2)/write(2) -- or even open(2). Thus, we can
2801 * use MAP_ASYNC to trade on-disk coherence for speed.
2802 * The shm_open(3) library routine turns on the FPOSIXSHM
2803 * flag to request this behavior.
2804 */
2805 if ((fp->f_flag & FPOSIXSHM) != 0)
2806 flags |= MAP_NOSYNC;
2807 #endif
2808 vp = fp->f_vnode;
2809
2810 /*
2811 * Ensure that file and memory protections are
2812 * compatible. Note that we only worry about
2813 * writability if mapping is shared; in this case,
2814 * current and max prot are dictated by the open file.
2815 * XXX use the vnode instead? Problem is: what
2816 * credentials do we use for determination? What if
2817 * proc does a setuid?
2818 */
2819 mp = vp->v_mount;
2820 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2821 maxprot = VM_PROT_NONE;
2822 if ((prot & VM_PROT_EXECUTE) != 0)
2823 return (EACCES);
2824 } else
2825 maxprot = VM_PROT_EXECUTE;
2826 if ((fp->f_flag & FREAD) != 0)
2827 maxprot |= VM_PROT_READ;
2828 else if ((prot & VM_PROT_READ) != 0)
2829 return (EACCES);
2830
2831 /*
2832 * If we are sharing potential changes via MAP_SHARED and we
2833 * are trying to get write permission although we opened it
2834 * without asking for it, bail out.
2835 */
2836 if ((flags & MAP_SHARED) != 0) {
2837 if ((fp->f_flag & FWRITE) != 0)
2838 maxprot |= VM_PROT_WRITE;
2839 else if ((prot & VM_PROT_WRITE) != 0)
2840 return (EACCES);
2841 } else {
2842 maxprot |= VM_PROT_WRITE;
2843 cap_maxprot |= VM_PROT_WRITE;
2844 }
2845 maxprot &= cap_maxprot;
2846
2847 /*
2848 * For regular files and shared memory, POSIX requires that
2849 * the value of foff be a legitimate offset within the data
2850 * object. In particular, negative offsets are invalid.
2851 * Blocking negative offsets and overflows here avoids
2852 * possible wraparound or user-level access into reserved
2853 * ranges of the data object later. In contrast, POSIX does
2854 * not dictate how offsets are used by device drivers, so in
2855 * the case of a device mapping a negative offset is passed
2856 * on.
2857 */
2858 if (
2859 #ifdef _LP64
2860 size > OFF_MAX ||
2861 #endif
2862 foff > OFF_MAX - size)
2863 return (EINVAL);
2864
2865 writecounted = FALSE;
2866 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2867 &foff, &object, &writecounted);
2868 if (error != 0)
2869 return (error);
2870 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2871 foff, writecounted, td);
2872 if (error != 0) {
2873 /*
2874 * If this mapping was accounted for in the vnode's
2875 * writecount, then undo that now.
2876 */
2877 if (writecounted)
2878 vm_pager_release_writecount(object, 0, size);
2879 vm_object_deallocate(object);
2880 }
2881 #ifdef HWPMC_HOOKS
2882 /* Inform hwpmc(4) if an executable is being mapped. */
2883 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2884 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2885 pkm.pm_file = vp;
2886 pkm.pm_address = (uintptr_t) *addr;
2887 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2888 }
2889 }
2890 #endif
2891 return (error);
2892 }
2893
2894 void
2895 vn_fsid(struct vnode *vp, struct vattr *va)
2896 {
2897 fsid_t *f;
2898
2899 f = &vp->v_mount->mnt_stat.f_fsid;
2900 va->va_fsid = (uint32_t)f->val[1];
2901 va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2902 va->va_fsid += (uint32_t)f->val[0];
2903 }
2904
2905 int
2906 vn_fsync_buf(struct vnode *vp, int waitfor)
2907 {
2908 struct buf *bp, *nbp;
2909 struct bufobj *bo;
2910 struct mount *mp;
2911 int error, maxretry;
2912
2913 error = 0;
2914 maxretry = 10000; /* large, arbitrarily chosen */
2915 mp = NULL;
2916 if (vp->v_type == VCHR) {
2917 VI_LOCK(vp);
2918 mp = vp->v_rdev->si_mountpt;
2919 VI_UNLOCK(vp);
2920 }
2921 bo = &vp->v_bufobj;
2922 BO_LOCK(bo);
2923 loop1:
2924 /*
2925 * MARK/SCAN initialization to avoid infinite loops.
2926 */
2927 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2928 bp->b_vflags &= ~BV_SCANNED;
2929 bp->b_error = 0;
2930 }
2931
2932 /*
2933 * Flush all dirty buffers associated with a vnode.
2934 */
2935 loop2:
2936 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2937 if ((bp->b_vflags & BV_SCANNED) != 0)
2938 continue;
2939 bp->b_vflags |= BV_SCANNED;
2940 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2941 if (waitfor != MNT_WAIT)
2942 continue;
2943 if (BUF_LOCK(bp,
2944 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2945 BO_LOCKPTR(bo)) != 0) {
2946 BO_LOCK(bo);
2947 goto loop1;
2948 }
2949 BO_LOCK(bo);
2950 }
2951 BO_UNLOCK(bo);
2952 KASSERT(bp->b_bufobj == bo,
2953 ("bp %p wrong b_bufobj %p should be %p",
2954 bp, bp->b_bufobj, bo));
2955 if ((bp->b_flags & B_DELWRI) == 0)
2956 panic("fsync: not dirty");
2957 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2958 vfs_bio_awrite(bp);
2959 } else {
2960 bremfree(bp);
2961 bawrite(bp);
2962 }
2963 if (maxretry < 1000)
2964 pause("dirty", hz < 1000 ? 1 : hz / 1000);
2965 BO_LOCK(bo);
2966 goto loop2;
2967 }
2968
2969 /*
2970 * If synchronous the caller expects us to completely resolve all
2971 * dirty buffers in the system. Wait for in-progress I/O to
2972 * complete (which could include background bitmap writes), then
2973 * retry if dirty blocks still exist.
2974 */
2975 if (waitfor == MNT_WAIT) {
2976 bufobj_wwait(bo, 0, 0);
2977 if (bo->bo_dirty.bv_cnt > 0) {
2978 /*
2979 * If we are unable to write any of these buffers
2980 * then we fail now rather than trying endlessly
2981 * to write them out.
2982 */
2983 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2984 if ((error = bp->b_error) != 0)
2985 break;
2986 if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2987 (error == 0 && --maxretry >= 0))
2988 goto loop1;
2989 if (error == 0)
2990 error = EAGAIN;
2991 }
2992 }
2993 BO_UNLOCK(bo);
2994 if (error != 0)
2995 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2996
2997 return (error);
2998 }
2999
3000 /*
3001 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE()
3002 * or vn_generic_copy_file_range() after rangelocking the byte ranges,
3003 * to do the actual copy.
3004 * vn_generic_copy_file_range() is factored out, so it can be called
3005 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
3006 * different file systems.
3007 */
3008 int
3009 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
3010 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
3011 struct ucred *outcred, struct thread *fsize_td)
3012 {
3013 int error;
3014 size_t len;
3015 uint64_t uval;
3016
3017 len = *lenp;
3018 *lenp = 0; /* For error returns. */
3019 error = 0;
3020
3021 /* Do some sanity checks on the arguments. */
3022 if (invp->v_type == VDIR || outvp->v_type == VDIR)
3023 error = EISDIR;
3024 else if (*inoffp < 0 || *outoffp < 0 ||
3025 invp->v_type != VREG || outvp->v_type != VREG)
3026 error = EINVAL;
3027 if (error != 0)
3028 goto out;
3029
3030 /* Ensure offset + len does not wrap around. */
3031 uval = *inoffp;
3032 uval += len;
3033 if (uval > INT64_MAX)
3034 len = INT64_MAX - *inoffp;
3035 uval = *outoffp;
3036 uval += len;
3037 if (uval > INT64_MAX)
3038 len = INT64_MAX - *outoffp;
3039 if (len == 0)
3040 goto out;
3041
3042 /*
3043 * If the two vnode are for the same file system, call
3044 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
3045 * which can handle copies across multiple file systems.
3046 */
3047 *lenp = len;
3048 if (invp->v_mount == outvp->v_mount)
3049 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
3050 lenp, flags, incred, outcred, fsize_td);
3051 else
3052 error = vn_generic_copy_file_range(invp, inoffp, outvp,
3053 outoffp, lenp, flags, incred, outcred, fsize_td);
3054 out:
3055 return (error);
3056 }
3057
3058 /*
3059 * Test len bytes of data starting at dat for all bytes == 0.
3060 * Return true if all bytes are zero, false otherwise.
3061 * Expects dat to be well aligned.
3062 */
3063 static bool
3064 mem_iszero(void *dat, int len)
3065 {
3066 int i;
3067 const u_int *p;
3068 const char *cp;
3069
3070 for (p = dat; len > 0; len -= sizeof(*p), p++) {
3071 if (len >= sizeof(*p)) {
3072 if (*p != 0)
3073 return (false);
3074 } else {
3075 cp = (const char *)p;
3076 for (i = 0; i < len; i++, cp++)
3077 if (*cp != '\0')
3078 return (false);
3079 }
3080 }
3081 return (true);
3082 }
3083
3084 /*
3085 * Look for a hole in the output file and, if found, adjust *outoffp
3086 * and *xferp to skip past the hole.
3087 * *xferp is the entire hole length to be written and xfer2 is how many bytes
3088 * to be written as 0's upon return.
3089 */
3090 static off_t
3091 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3092 off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3093 {
3094 int error;
3095 off_t delta;
3096
3097 if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3098 *dataoffp = *outoffp;
3099 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3100 curthread);
3101 if (error == 0) {
3102 *holeoffp = *dataoffp;
3103 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3104 curthread);
3105 }
3106 if (error != 0 || *holeoffp == *dataoffp) {
3107 /*
3108 * Since outvp is unlocked, it may be possible for
3109 * another thread to do a truncate(), lseek(), write()
3110 * creating a hole at startoff between the above
3111 * VOP_IOCTL() calls, if the other thread does not do
3112 * rangelocking.
3113 * If that happens, *holeoffp == *dataoffp and finding
3114 * the hole has failed, so disable vn_skip_hole().
3115 */
3116 *holeoffp = -1; /* Disable use of vn_skip_hole(). */
3117 return (xfer2);
3118 }
3119 KASSERT(*dataoffp >= *outoffp,
3120 ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3121 (intmax_t)*dataoffp, (intmax_t)*outoffp));
3122 KASSERT(*holeoffp > *dataoffp,
3123 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3124 (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3125 }
3126
3127 /*
3128 * If there is a hole before the data starts, advance *outoffp and
3129 * *xferp past the hole.
3130 */
3131 if (*dataoffp > *outoffp) {
3132 delta = *dataoffp - *outoffp;
3133 if (delta >= *xferp) {
3134 /* Entire *xferp is a hole. */
3135 *outoffp += *xferp;
3136 *xferp = 0;
3137 return (0);
3138 }
3139 *xferp -= delta;
3140 *outoffp += delta;
3141 xfer2 = MIN(xfer2, *xferp);
3142 }
3143
3144 /*
3145 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3146 * that the write ends at the start of the hole.
3147 * *holeoffp should always be greater than *outoffp, but for the
3148 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3149 * value.
3150 */
3151 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3152 xfer2 = *holeoffp - *outoffp;
3153 return (xfer2);
3154 }
3155
3156 /*
3157 * Write an xfer sized chunk to outvp in blksize blocks from dat.
3158 * dat is a maximum of blksize in length and can be written repeatedly in
3159 * the chunk.
3160 * If growfile == true, just grow the file via vn_truncate_locked() instead
3161 * of doing actual writes.
3162 * If checkhole == true, a hole is being punched, so skip over any hole
3163 * already in the output file.
3164 */
3165 static int
3166 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3167 u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3168 {
3169 struct mount *mp;
3170 off_t dataoff, holeoff, xfer2;
3171 int error;
3172
3173 /*
3174 * Loop around doing writes of blksize until write has been completed.
3175 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3176 * done for each iteration, since the xfer argument can be very
3177 * large if there is a large hole to punch in the output file.
3178 */
3179 error = 0;
3180 holeoff = 0;
3181 do {
3182 xfer2 = MIN(xfer, blksize);
3183 if (checkhole) {
3184 /*
3185 * Punching a hole. Skip writing if there is
3186 * already a hole in the output file.
3187 */
3188 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3189 &dataoff, &holeoff, cred);
3190 if (xfer == 0)
3191 break;
3192 if (holeoff < 0)
3193 checkhole = false;
3194 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3195 (intmax_t)xfer2));
3196 }
3197 bwillwrite();
3198 mp = NULL;
3199 error = vn_start_write(outvp, &mp, V_WAIT);
3200 if (error != 0)
3201 break;
3202 if (growfile) {
3203 error = vn_lock(outvp, LK_EXCLUSIVE);
3204 if (error == 0) {
3205 error = vn_truncate_locked(outvp, outoff + xfer,
3206 false, cred);
3207 VOP_UNLOCK(outvp);
3208 }
3209 } else {
3210 error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3211 if (error == 0) {
3212 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3213 outoff, UIO_SYSSPACE, IO_NODELOCKED,
3214 curthread->td_ucred, cred, NULL, curthread);
3215 outoff += xfer2;
3216 xfer -= xfer2;
3217 VOP_UNLOCK(outvp);
3218 }
3219 }
3220 if (mp != NULL)
3221 vn_finished_write(mp);
3222 } while (!growfile && xfer > 0 && error == 0);
3223 return (error);
3224 }
3225
3226 /*
3227 * Copy a byte range of one file to another. This function can handle the
3228 * case where invp and outvp are on different file systems.
3229 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3230 * is no better file system specific way to do it.
3231 */
3232 int
3233 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3234 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3235 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3236 {
3237 struct vattr va, inva;
3238 struct mount *mp;
3239 off_t startoff, endoff, xfer, xfer2;
3240 u_long blksize;
3241 int error, interrupted;
3242 bool cantseek, readzeros, eof, lastblock, holetoeof;
3243 ssize_t aresid, r = 0;
3244 size_t copylen, len, savlen;
3245 char *dat;
3246 long holein, holeout;
3247 struct timespec curts, endts;
3248
3249 holein = holeout = 0;
3250 savlen = len = *lenp;
3251 error = 0;
3252 interrupted = 0;
3253 dat = NULL;
3254
3255 error = vn_lock(invp, LK_SHARED);
3256 if (error != 0)
3257 goto out;
3258 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3259 holein = 0;
3260 if (holein > 0)
3261 error = VOP_GETATTR(invp, &inva, incred);
3262 VOP_UNLOCK(invp);
3263 if (error != 0)
3264 goto out;
3265
3266 mp = NULL;
3267 error = vn_start_write(outvp, &mp, V_WAIT);
3268 if (error == 0)
3269 error = vn_lock(outvp, LK_EXCLUSIVE);
3270 if (error == 0) {
3271 /*
3272 * If fsize_td != NULL, do a vn_rlimit_fsizex() call,
3273 * now that outvp is locked.
3274 */
3275 if (fsize_td != NULL) {
3276 struct uio io;
3277
3278 io.uio_offset = *outoffp;
3279 io.uio_resid = len;
3280 error = vn_rlimit_fsizex(outvp, &io, 0, &r, fsize_td);
3281 len = savlen = io.uio_resid;
3282 /*
3283 * No need to call vn_rlimit_fsizex_res before return,
3284 * since the uio is local.
3285 */
3286 }
3287 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3288 holeout = 0;
3289 /*
3290 * Holes that are past EOF do not need to be written as a block
3291 * of zero bytes. So, truncate the output file as far as
3292 * possible and then use va.va_size to decide if writing 0
3293 * bytes is necessary in the loop below.
3294 */
3295 if (error == 0)
3296 error = VOP_GETATTR(outvp, &va, outcred);
3297 if (error == 0 && va.va_size > *outoffp && va.va_size <=
3298 *outoffp + len) {
3299 #ifdef MAC
3300 error = mac_vnode_check_write(curthread->td_ucred,
3301 outcred, outvp);
3302 if (error == 0)
3303 #endif
3304 error = vn_truncate_locked(outvp, *outoffp,
3305 false, outcred);
3306 if (error == 0)
3307 va.va_size = *outoffp;
3308 }
3309 VOP_UNLOCK(outvp);
3310 }
3311 if (mp != NULL)
3312 vn_finished_write(mp);
3313 if (error != 0)
3314 goto out;
3315
3316 if (holein == 0 && holeout > 0) {
3317 /*
3318 * For this special case, the input data will be scanned
3319 * for blocks of all 0 bytes. For these blocks, the
3320 * write can be skipped for the output file to create
3321 * an unallocated region.
3322 * Therefore, use the appropriate size for the output file.
3323 */
3324 blksize = holeout;
3325 if (blksize <= 512) {
3326 /*
3327 * Use f_iosize, since ZFS reports a _PC_MIN_HOLE_SIZE
3328 * of 512, although it actually only creates
3329 * unallocated regions for blocks >= f_iosize.
3330 */
3331 blksize = outvp->v_mount->mnt_stat.f_iosize;
3332 }
3333 } else {
3334 /*
3335 * Use the larger of the two f_iosize values. If they are
3336 * not the same size, one will normally be an exact multiple of
3337 * the other, since they are both likely to be a power of 2.
3338 */
3339 blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3340 outvp->v_mount->mnt_stat.f_iosize);
3341 }
3342
3343 /* Clip to sane limits. */
3344 if (blksize < 4096)
3345 blksize = 4096;
3346 else if (blksize > maxphys)
3347 blksize = maxphys;
3348 dat = malloc(blksize, M_TEMP, M_WAITOK);
3349
3350 /*
3351 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3352 * to find holes. Otherwise, just scan the read block for all 0s
3353 * in the inner loop where the data copying is done.
3354 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3355 * support holes on the server, but do not support FIOSEEKHOLE.
3356 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3357 * that this function should return after 1second with a partial
3358 * completion.
3359 */
3360 if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3361 getnanouptime(&endts);
3362 endts.tv_sec++;
3363 } else
3364 timespecclear(&endts);
3365 holetoeof = eof = false;
3366 while (len > 0 && error == 0 && !eof && interrupted == 0) {
3367 endoff = 0; /* To shut up compilers. */
3368 cantseek = true;
3369 startoff = *inoffp;
3370 copylen = len;
3371
3372 /*
3373 * Find the next data area. If there is just a hole to EOF,
3374 * FIOSEEKDATA should fail with ENXIO.
3375 * (I do not know if any file system will report a hole to
3376 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3377 * will fail for those file systems.)
3378 *
3379 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3380 * the code just falls through to the inner copy loop.
3381 */
3382 error = EINVAL;
3383 if (holein > 0) {
3384 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3385 incred, curthread);
3386 if (error == ENXIO) {
3387 startoff = endoff = inva.va_size;
3388 eof = holetoeof = true;
3389 error = 0;
3390 }
3391 }
3392 if (error == 0 && !holetoeof) {
3393 endoff = startoff;
3394 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3395 incred, curthread);
3396 /*
3397 * Since invp is unlocked, it may be possible for
3398 * another thread to do a truncate(), lseek(), write()
3399 * creating a hole at startoff between the above
3400 * VOP_IOCTL() calls, if the other thread does not do
3401 * rangelocking.
3402 * If that happens, startoff == endoff and finding
3403 * the hole has failed, so set an error.
3404 */
3405 if (error == 0 && startoff == endoff)
3406 error = EINVAL; /* Any error. Reset to 0. */
3407 }
3408 if (error == 0) {
3409 if (startoff > *inoffp) {
3410 /* Found hole before data block. */
3411 xfer = MIN(startoff - *inoffp, len);
3412 if (*outoffp < va.va_size) {
3413 /* Must write 0s to punch hole. */
3414 xfer2 = MIN(va.va_size - *outoffp,
3415 xfer);
3416 memset(dat, 0, MIN(xfer2, blksize));
3417 error = vn_write_outvp(outvp, dat,
3418 *outoffp, xfer2, blksize, false,
3419 holeout > 0, outcred);
3420 }
3421
3422 if (error == 0 && *outoffp + xfer >
3423 va.va_size && (xfer == len || holetoeof)) {
3424 /* Grow output file (hole at end). */
3425 error = vn_write_outvp(outvp, dat,
3426 *outoffp, xfer, blksize, true,
3427 false, outcred);
3428 }
3429 if (error == 0) {
3430 *inoffp += xfer;
3431 *outoffp += xfer;
3432 len -= xfer;
3433 if (len < savlen) {
3434 interrupted = sig_intr();
3435 if (timespecisset(&endts) &&
3436 interrupted == 0) {
3437 getnanouptime(&curts);
3438 if (timespeccmp(&curts,
3439 &endts, >=))
3440 interrupted =
3441 EINTR;
3442 }
3443 }
3444 }
3445 }
3446 copylen = MIN(len, endoff - startoff);
3447 cantseek = false;
3448 } else {
3449 cantseek = true;
3450 startoff = *inoffp;
3451 copylen = len;
3452 error = 0;
3453 }
3454
3455 xfer = blksize;
3456 if (cantseek) {
3457 /*
3458 * Set first xfer to end at a block boundary, so that
3459 * holes are more likely detected in the loop below via
3460 * the for all bytes 0 method.
3461 */
3462 xfer -= (*inoffp % blksize);
3463 }
3464 /* Loop copying the data block. */
3465 while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3466 if (copylen < xfer)
3467 xfer = copylen;
3468 error = vn_lock(invp, LK_SHARED);
3469 if (error != 0)
3470 goto out;
3471 error = vn_rdwr(UIO_READ, invp, dat, xfer,
3472 startoff, UIO_SYSSPACE, IO_NODELOCKED,
3473 curthread->td_ucred, incred, &aresid,
3474 curthread);
3475 VOP_UNLOCK(invp);
3476 lastblock = false;
3477 if (error == 0 && aresid > 0) {
3478 /* Stop the copy at EOF on the input file. */
3479 xfer -= aresid;
3480 eof = true;
3481 lastblock = true;
3482 }
3483 if (error == 0) {
3484 /*
3485 * Skip the write for holes past the initial EOF
3486 * of the output file, unless this is the last
3487 * write of the output file at EOF.
3488 */
3489 readzeros = cantseek ? mem_iszero(dat, xfer) :
3490 false;
3491 if (xfer == len)
3492 lastblock = true;
3493 if (!cantseek || *outoffp < va.va_size ||
3494 lastblock || !readzeros)
3495 error = vn_write_outvp(outvp, dat,
3496 *outoffp, xfer, blksize,
3497 readzeros && lastblock &&
3498 *outoffp >= va.va_size, false,
3499 outcred);
3500 if (error == 0) {
3501 *inoffp += xfer;
3502 startoff += xfer;
3503 *outoffp += xfer;
3504 copylen -= xfer;
3505 len -= xfer;
3506 if (len < savlen) {
3507 interrupted = sig_intr();
3508 if (timespecisset(&endts) &&
3509 interrupted == 0) {
3510 getnanouptime(&curts);
3511 if (timespeccmp(&curts,
3512 &endts, >=))
3513 interrupted =
3514 EINTR;
3515 }
3516 }
3517 }
3518 }
3519 xfer = blksize;
3520 }
3521 }
3522 out:
3523 *lenp = savlen - len;
3524 free(dat, M_TEMP);
3525 return (error);
3526 }
3527
3528 static int
3529 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3530 {
3531 struct mount *mp;
3532 struct vnode *vp;
3533 off_t olen, ooffset;
3534 int error;
3535 #ifdef AUDIT
3536 int audited_vnode1 = 0;
3537 #endif
3538
3539 vp = fp->f_vnode;
3540 if (vp->v_type != VREG)
3541 return (ENODEV);
3542
3543 /* Allocating blocks may take a long time, so iterate. */
3544 for (;;) {
3545 olen = len;
3546 ooffset = offset;
3547
3548 bwillwrite();
3549 mp = NULL;
3550 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3551 if (error != 0)
3552 break;
3553 error = vn_lock(vp, LK_EXCLUSIVE);
3554 if (error != 0) {
3555 vn_finished_write(mp);
3556 break;
3557 }
3558 #ifdef AUDIT
3559 if (!audited_vnode1) {
3560 AUDIT_ARG_VNODE1(vp);
3561 audited_vnode1 = 1;
3562 }
3563 #endif
3564 #ifdef MAC
3565 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3566 if (error == 0)
3567 #endif
3568 error = VOP_ALLOCATE(vp, &offset, &len, 0,
3569 td->td_ucred);
3570 VOP_UNLOCK(vp);
3571 vn_finished_write(mp);
3572
3573 if (olen + ooffset != offset + len) {
3574 panic("offset + len changed from %jx/%jx to %jx/%jx",
3575 ooffset, olen, offset, len);
3576 }
3577 if (error != 0 || len == 0)
3578 break;
3579 KASSERT(olen > len, ("Iteration did not make progress?"));
3580 maybe_yield();
3581 }
3582
3583 return (error);
3584 }
3585
3586 static u_long vn_lock_pair_pause_cnt;
3587 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3588 &vn_lock_pair_pause_cnt, 0,
3589 "Count of vn_lock_pair deadlocks");
3590
3591 u_int vn_lock_pair_pause_max;
3592 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3593 &vn_lock_pair_pause_max, 0,
3594 "Max ticks for vn_lock_pair deadlock avoidance sleep");
3595
3596 static void
3597 vn_lock_pair_pause(const char *wmesg)
3598 {
3599 atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3600 pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3601 }
3602
3603 /*
3604 * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3605 * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3606 * must be unlocked. Same for vp2 and vp2_locked. One of the vnodes
3607 * can be NULL.
3608 *
3609 * The function returns with both vnodes exclusively locked, and
3610 * guarantees that it does not create lock order reversal with other
3611 * threads during its execution. Both vnodes could be unlocked
3612 * temporary (and reclaimed).
3613 */
3614 void
3615 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3616 bool vp2_locked)
3617 {
3618 int error;
3619
3620 if (vp1 == NULL && vp2 == NULL)
3621 return;
3622 if (vp1 != NULL) {
3623 if (vp1_locked)
3624 ASSERT_VOP_ELOCKED(vp1, "vp1");
3625 else
3626 ASSERT_VOP_UNLOCKED(vp1, "vp1");
3627 } else {
3628 vp1_locked = true;
3629 }
3630 if (vp2 != NULL) {
3631 if (vp2_locked)
3632 ASSERT_VOP_ELOCKED(vp2, "vp2");
3633 else
3634 ASSERT_VOP_UNLOCKED(vp2, "vp2");
3635 } else {
3636 vp2_locked = true;
3637 }
3638 if (!vp1_locked && !vp2_locked) {
3639 vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3640 vp1_locked = true;
3641 }
3642
3643 for (;;) {
3644 if (vp1_locked && vp2_locked)
3645 break;
3646 if (vp1_locked && vp2 != NULL) {
3647 if (vp1 != NULL) {
3648 error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3649 __FILE__, __LINE__);
3650 if (error == 0)
3651 break;
3652 VOP_UNLOCK(vp1);
3653 vp1_locked = false;
3654 vn_lock_pair_pause("vlp1");
3655 }
3656 vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3657 vp2_locked = true;
3658 }
3659 if (vp2_locked && vp1 != NULL) {
3660 if (vp2 != NULL) {
3661 error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3662 __FILE__, __LINE__);
3663 if (error == 0)
3664 break;
3665 VOP_UNLOCK(vp2);
3666 vp2_locked = false;
3667 vn_lock_pair_pause("vlp2");
3668 }
3669 vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3670 vp1_locked = true;
3671 }
3672 }
3673 if (vp1 != NULL)
3674 ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3675 if (vp2 != NULL)
3676 ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3677 }
3678
3679 int
3680 vn_lktype_write(struct mount *mp, struct vnode *vp)
3681 {
3682 if (MNT_SHARED_WRITES(mp) ||
3683 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
3684 return (LK_SHARED);
3685 return (LK_EXCLUSIVE);
3686 }
Cache object: 2f5efc3c06f105b1fb30f8a21e773892
|