1 /*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
4 * Copyright (c) 2001 Dag-Erling Smørgrav
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer
12 * in this position and unchanged.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. The name of the author may not be used to endorse or promote products
17 * derived from this software without specific prior written permission.
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
20 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
21 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
22 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
23 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
24 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
28 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 */
30
31 #include <sys/cdefs.h>
32 #include "opt_pseudofs.h"
33
34 #include <sys/param.h>
35 #include <sys/kernel.h>
36 #include <sys/systm.h>
37 #include <sys/ctype.h>
38 #include <sys/dirent.h>
39 #include <sys/fcntl.h>
40 #include <sys/imgact.h>
41 #include <sys/limits.h>
42 #include <sys/lock.h>
43 #include <sys/malloc.h>
44 #include <sys/mount.h>
45 #include <sys/mutex.h>
46 #include <sys/namei.h>
47 #include <sys/proc.h>
48 #include <sys/sbuf.h>
49 #include <sys/sx.h>
50 #include <sys/sysctl.h>
51 #include <sys/vnode.h>
52
53 #include <fs/pseudofs/pseudofs.h>
54 #include <fs/pseudofs/pseudofs_internal.h>
55
56 #define KASSERT_PN_IS_DIR(pn) \
57 KASSERT((pn)->pn_type == pfstype_root || \
58 (pn)->pn_type == pfstype_dir || \
59 (pn)->pn_type == pfstype_procdir, \
60 ("%s(): VDIR vnode refers to non-directory pfs_node", __func__))
61
62 #define KASSERT_PN_IS_FILE(pn) \
63 KASSERT((pn)->pn_type == pfstype_file, \
64 ("%s(): VREG vnode refers to non-file pfs_node", __func__))
65
66 #define KASSERT_PN_IS_LINK(pn) \
67 KASSERT((pn)->pn_type == pfstype_symlink, \
68 ("%s(): VLNK vnode refers to non-link pfs_node", __func__))
69
70 #define PFS_MAXBUFSIZ 1024 * 1024
71
72 /*
73 * Returns the fileno, adjusted for target pid
74 */
75 static uint32_t
pn_fileno(struct pfs_node * pn,pid_t pid)76 pn_fileno(struct pfs_node *pn, pid_t pid)
77 {
78
79 KASSERT(pn->pn_fileno > 0,
80 ("%s(): no fileno allocated", __func__));
81 if (pid != NO_PID)
82 return (pn->pn_fileno * NO_PID + pid);
83 return (pn->pn_fileno);
84 }
85
86 /*
87 * Returns non-zero if given file is visible to given thread.
88 */
89 static int
pfs_visible_proc(struct thread * td,struct pfs_node * pn,struct proc * proc)90 pfs_visible_proc(struct thread *td, struct pfs_node *pn, struct proc *proc)
91 {
92
93 if (proc == NULL)
94 return (0);
95
96 PROC_LOCK_ASSERT(proc, MA_OWNED);
97
98 if ((proc->p_flag & P_WEXIT) != 0)
99 return (0);
100 if (p_cansee(td, proc) != 0)
101 return (0);
102 return (pn_vis(td, proc, pn));
103 }
104
105 static int
pfs_visible(struct thread * td,struct pfs_node * pn,pid_t pid,struct proc ** p)106 pfs_visible(struct thread *td, struct pfs_node *pn, pid_t pid,
107 struct proc **p)
108 {
109 struct proc *proc;
110
111 PFS_TRACE(("%s (pid: %d, req: %d)",
112 pn->pn_name, pid, td->td_proc->p_pid));
113
114 if (p)
115 *p = NULL;
116 if (pid == NO_PID)
117 PFS_RETURN (pn_vis(td, NULL, pn));
118 proc = pfind(pid);
119 if (proc == NULL)
120 PFS_RETURN (0);
121 if (pfs_visible_proc(td, pn, proc)) {
122 if (p)
123 *p = proc;
124 else
125 PROC_UNLOCK(proc);
126 PFS_RETURN (1);
127 }
128 PROC_UNLOCK(proc);
129 PFS_RETURN (0);
130 }
131
132 static int
pfs_lookup_proc(pid_t pid,struct proc ** p)133 pfs_lookup_proc(pid_t pid, struct proc **p)
134 {
135 struct proc *proc;
136 struct thread *td;
137
138 proc = pfind(pid);
139 if (proc == NULL)
140 return (0);
141 if ((proc->p_flag & P_WEXIT) != 0) {
142 PROC_UNLOCK(proc);
143 return (0);
144 }
145 _PHOLD(proc);
146 td = curthread;
147 execve_block_wait(td, proc);
148 *p = proc;
149 PROC_UNLOCK(proc);
150 return (1);
151 }
152
153 /*
154 * Verify permissions
155 */
156 static int
pfs_access(struct vop_access_args * va)157 pfs_access(struct vop_access_args *va)
158 {
159 struct vnode *vn = va->a_vp;
160 struct pfs_vdata *pvd = vn->v_data;
161 struct vattr vattr;
162 int error;
163
164 PFS_TRACE(("%s", pvd->pvd_pn->pn_name));
165 (void)pvd;
166
167 error = VOP_GETATTR(vn, &vattr, va->a_cred);
168 if (error)
169 PFS_RETURN (error);
170 error = vaccess(vn->v_type, vattr.va_mode, vattr.va_uid, vattr.va_gid,
171 va->a_accmode, va->a_cred);
172 PFS_RETURN (error);
173 }
174
175 /*
176 * Close a file or directory
177 */
178 static int
pfs_close(struct vop_close_args * va)179 pfs_close(struct vop_close_args *va)
180 {
181 struct vnode *vn = va->a_vp;
182 struct pfs_vdata *pvd = vn->v_data;
183 struct pfs_node *pn = pvd->pvd_pn;
184 struct proc *proc;
185 int error;
186
187 PFS_TRACE(("%s", pn->pn_name));
188 pfs_assert_not_owned(pn);
189
190 /*
191 * Do nothing unless this is the last close and the node has a
192 * last-close handler.
193 */
194 if (vrefcnt(vn) > 1 || pn->pn_close == NULL)
195 PFS_RETURN (0);
196
197 if (pvd->pvd_pid != NO_PID) {
198 proc = pfind(pvd->pvd_pid);
199 } else {
200 proc = NULL;
201 }
202
203 error = pn_close(va->a_td, proc, pn);
204
205 if (proc != NULL)
206 PROC_UNLOCK(proc);
207
208 PFS_RETURN (error);
209 }
210
211 /*
212 * Get file attributes
213 */
214 static int
pfs_getattr(struct vop_getattr_args * va)215 pfs_getattr(struct vop_getattr_args *va)
216 {
217 struct vnode *vn = va->a_vp;
218 struct pfs_vdata *pvd = vn->v_data;
219 struct pfs_node *pn = pvd->pvd_pn;
220 struct vattr *vap = va->a_vap;
221 struct proc *proc;
222 int error = 0;
223
224 PFS_TRACE(("%s", pn->pn_name));
225 pfs_assert_not_owned(pn);
226
227 if (!pfs_visible(curthread, pn, pvd->pvd_pid, &proc))
228 PFS_RETURN (ENOENT);
229
230 vap->va_type = vn->v_type;
231 vap->va_fileid = pn_fileno(pn, pvd->pvd_pid);
232 vap->va_flags = 0;
233 vap->va_blocksize = PAGE_SIZE;
234 vap->va_bytes = vap->va_size = 0;
235 vap->va_filerev = 0;
236 vap->va_fsid = vn->v_mount->mnt_stat.f_fsid.val[0];
237 vap->va_nlink = 1;
238 nanotime(&vap->va_ctime);
239 vap->va_atime = vap->va_mtime = vap->va_ctime;
240
241 switch (pn->pn_type) {
242 case pfstype_procdir:
243 case pfstype_root:
244 case pfstype_dir:
245 #if 0
246 pfs_lock(pn);
247 /* compute link count */
248 pfs_unlock(pn);
249 #endif
250 vap->va_mode = 0555;
251 break;
252 case pfstype_file:
253 case pfstype_symlink:
254 vap->va_mode = 0444;
255 break;
256 default:
257 printf("shouldn't be here!\n");
258 vap->va_mode = 0;
259 break;
260 }
261
262 if (proc != NULL) {
263 vap->va_uid = proc->p_ucred->cr_ruid;
264 vap->va_gid = proc->p_ucred->cr_rgid;
265 } else {
266 vap->va_uid = 0;
267 vap->va_gid = 0;
268 }
269
270 if (pn->pn_attr != NULL)
271 error = pn_attr(curthread, proc, pn, vap);
272
273 if(proc != NULL)
274 PROC_UNLOCK(proc);
275
276 PFS_RETURN (error);
277 }
278
279 /*
280 * Perform an ioctl
281 */
282 static int
pfs_ioctl(struct vop_ioctl_args * va)283 pfs_ioctl(struct vop_ioctl_args *va)
284 {
285 struct vnode *vn;
286 struct pfs_vdata *pvd;
287 struct pfs_node *pn;
288 struct proc *proc;
289 int error;
290
291 vn = va->a_vp;
292 vn_lock(vn, LK_SHARED | LK_RETRY);
293 if (VN_IS_DOOMED(vn)) {
294 VOP_UNLOCK(vn);
295 return (EBADF);
296 }
297 pvd = vn->v_data;
298 pn = pvd->pvd_pn;
299
300 PFS_TRACE(("%s: %lx", pn->pn_name, va->a_command));
301 pfs_assert_not_owned(pn);
302
303 if (vn->v_type != VREG) {
304 VOP_UNLOCK(vn);
305 PFS_RETURN (EINVAL);
306 }
307 KASSERT_PN_IS_FILE(pn);
308
309 if (pn->pn_ioctl == NULL) {
310 VOP_UNLOCK(vn);
311 PFS_RETURN (ENOTTY);
312 }
313
314 /*
315 * This is necessary because process' privileges may
316 * have changed since the open() call.
317 */
318 if (!pfs_visible(curthread, pn, pvd->pvd_pid, &proc)) {
319 VOP_UNLOCK(vn);
320 PFS_RETURN (EIO);
321 }
322
323 error = pn_ioctl(curthread, proc, pn, va->a_command, va->a_data);
324
325 if (proc != NULL)
326 PROC_UNLOCK(proc);
327
328 VOP_UNLOCK(vn);
329 PFS_RETURN (error);
330 }
331
332 /*
333 * Perform getextattr
334 */
335 static int
pfs_getextattr(struct vop_getextattr_args * va)336 pfs_getextattr(struct vop_getextattr_args *va)
337 {
338 struct vnode *vn = va->a_vp;
339 struct pfs_vdata *pvd = vn->v_data;
340 struct pfs_node *pn = pvd->pvd_pn;
341 struct proc *proc;
342 int error;
343
344 PFS_TRACE(("%s", pn->pn_name));
345 pfs_assert_not_owned(pn);
346
347 /*
348 * This is necessary because either process' privileges may
349 * have changed since the open() call.
350 */
351 if (!pfs_visible(curthread, pn, pvd->pvd_pid, &proc))
352 PFS_RETURN (EIO);
353
354 if (pn->pn_getextattr == NULL)
355 error = EOPNOTSUPP;
356 else
357 error = pn_getextattr(curthread, proc, pn,
358 va->a_attrnamespace, va->a_name, va->a_uio,
359 va->a_size, va->a_cred);
360
361 if (proc != NULL)
362 PROC_UNLOCK(proc);
363
364 PFS_RETURN (error);
365 }
366
367 /*
368 * Convert a vnode to its component name
369 */
370 static int
pfs_vptocnp(struct vop_vptocnp_args * ap)371 pfs_vptocnp(struct vop_vptocnp_args *ap)
372 {
373 struct vnode *vp = ap->a_vp;
374 struct vnode **dvp = ap->a_vpp;
375 struct pfs_vdata *pvd = vp->v_data;
376 struct pfs_node *pd = pvd->pvd_pn;
377 struct pfs_node *pn;
378 struct mount *mp;
379 char *buf = ap->a_buf;
380 size_t *buflen = ap->a_buflen;
381 char pidbuf[PFS_NAMELEN];
382 pid_t pid = pvd->pvd_pid;
383 int len, i, error, locked;
384
385 i = *buflen;
386 error = 0;
387
388 pfs_lock(pd);
389
390 if (vp->v_type == VDIR && pd->pn_type == pfstype_root) {
391 *dvp = vp;
392 vhold(*dvp);
393 pfs_unlock(pd);
394 PFS_RETURN (0);
395 } else if (vp->v_type == VDIR && pd->pn_type == pfstype_procdir) {
396 len = snprintf(pidbuf, sizeof(pidbuf), "%d", pid);
397 i -= len;
398 if (i < 0) {
399 error = ENOMEM;
400 goto failed;
401 }
402 bcopy(pidbuf, buf + i, len);
403 } else {
404 len = strlen(pd->pn_name);
405 i -= len;
406 if (i < 0) {
407 error = ENOMEM;
408 goto failed;
409 }
410 bcopy(pd->pn_name, buf + i, len);
411 }
412
413 pn = pd->pn_parent;
414 pfs_unlock(pd);
415
416 mp = vp->v_mount;
417 error = vfs_busy(mp, 0);
418 if (error)
419 return (error);
420
421 /*
422 * vp is held by caller.
423 */
424 locked = VOP_ISLOCKED(vp);
425 VOP_UNLOCK(vp);
426
427 error = pfs_vncache_alloc(mp, dvp, pn, pid);
428 if (error) {
429 vn_lock(vp, locked | LK_RETRY);
430 vfs_unbusy(mp);
431 PFS_RETURN(error);
432 }
433
434 *buflen = i;
435 VOP_UNLOCK(*dvp);
436 vn_lock(vp, locked | LK_RETRY);
437 vfs_unbusy(mp);
438
439 PFS_RETURN (0);
440 failed:
441 pfs_unlock(pd);
442 PFS_RETURN(error);
443 }
444
445 /*
446 * Look up a file or directory
447 */
448 static int
pfs_lookup(struct vop_cachedlookup_args * va)449 pfs_lookup(struct vop_cachedlookup_args *va)
450 {
451 struct vnode *vn = va->a_dvp;
452 struct vnode **vpp = va->a_vpp;
453 struct componentname *cnp = va->a_cnp;
454 struct pfs_vdata *pvd = vn->v_data;
455 struct pfs_node *pd = pvd->pvd_pn;
456 struct pfs_node *pn, *pdn = NULL;
457 struct mount *mp;
458 pid_t pid = pvd->pvd_pid;
459 char *pname;
460 int error, i, namelen, visible;
461
462 PFS_TRACE(("%.*s", (int)cnp->cn_namelen, cnp->cn_nameptr));
463 pfs_assert_not_owned(pd);
464
465 if (vn->v_type != VDIR)
466 PFS_RETURN (ENOTDIR);
467 KASSERT_PN_IS_DIR(pd);
468
469 /*
470 * Don't support DELETE or RENAME. CREATE is supported so
471 * that O_CREAT will work, but the lookup will still fail if
472 * the file does not exist.
473 */
474 if ((cnp->cn_flags & ISLASTCN) &&
475 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
476 PFS_RETURN (EOPNOTSUPP);
477
478 /* shortcut: check if the name is too long */
479 if (cnp->cn_namelen >= PFS_NAMELEN)
480 PFS_RETURN (ENOENT);
481
482 /* check that parent directory is visible... */
483 if (!pfs_visible(curthread, pd, pvd->pvd_pid, NULL))
484 PFS_RETURN (ENOENT);
485
486 /* self */
487 namelen = cnp->cn_namelen;
488 pname = cnp->cn_nameptr;
489 if (namelen == 1 && pname[0] == '.') {
490 pn = pd;
491 *vpp = vn;
492 vref(vn);
493 PFS_RETURN (0);
494 }
495
496 mp = vn->v_mount;
497
498 /* parent */
499 if (cnp->cn_flags & ISDOTDOT) {
500 if (pd->pn_type == pfstype_root)
501 PFS_RETURN (EIO);
502 error = vfs_busy(mp, MBF_NOWAIT);
503 if (error != 0) {
504 vfs_ref(mp);
505 VOP_UNLOCK(vn);
506 error = vfs_busy(mp, 0);
507 vn_lock(vn, LK_EXCLUSIVE | LK_RETRY);
508 vfs_rel(mp);
509 if (error != 0)
510 PFS_RETURN(ENOENT);
511 if (VN_IS_DOOMED(vn)) {
512 vfs_unbusy(mp);
513 PFS_RETURN(ENOENT);
514 }
515 }
516 VOP_UNLOCK(vn);
517 KASSERT(pd->pn_parent != NULL,
518 ("%s(): non-root directory has no parent", __func__));
519 /*
520 * This one is tricky. Descendents of procdir nodes
521 * inherit their parent's process affinity, but
522 * there's no easy reverse mapping. For simplicity,
523 * we assume that if this node is a procdir, its
524 * parent isn't (which is correct as long as
525 * descendents of procdir nodes are never procdir
526 * nodes themselves)
527 */
528 if (pd->pn_type == pfstype_procdir)
529 pid = NO_PID;
530 pfs_lock(pd);
531 pn = pd->pn_parent;
532 pfs_unlock(pd);
533 goto got_pnode;
534 }
535
536 pfs_lock(pd);
537
538 /* named node */
539 for (pn = pd->pn_nodes; pn != NULL; pn = pn->pn_next)
540 if (pn->pn_type == pfstype_procdir)
541 pdn = pn;
542 else if (strncmp(pname, pn->pn_name, namelen) == 0 &&
543 pn->pn_name[namelen] == '\0') {
544 pfs_unlock(pd);
545 goto got_pnode;
546 }
547
548 /* process dependent node */
549 if ((pn = pdn) != NULL) {
550 pid = 0;
551 for (pid = 0, i = 0; i < namelen && isdigit(pname[i]); ++i)
552 if ((pid = pid * 10 + pname[i] - '0') > PID_MAX)
553 break;
554 if (i == cnp->cn_namelen) {
555 pfs_unlock(pd);
556 goto got_pnode;
557 }
558 }
559
560 pfs_unlock(pd);
561
562 PFS_RETURN (ENOENT);
563
564 got_pnode:
565 pfs_assert_not_owned(pd);
566 pfs_assert_not_owned(pn);
567 visible = pfs_visible(curthread, pn, pid, NULL);
568 if (!visible) {
569 error = ENOENT;
570 goto failed;
571 }
572
573 error = pfs_vncache_alloc(mp, vpp, pn, pid);
574 if (error)
575 goto failed;
576
577 if (cnp->cn_flags & ISDOTDOT) {
578 vfs_unbusy(mp);
579 vn_lock(vn, LK_EXCLUSIVE | LK_RETRY);
580 if (VN_IS_DOOMED(vn)) {
581 vput(*vpp);
582 *vpp = NULL;
583 PFS_RETURN(ENOENT);
584 }
585 }
586 if (cnp->cn_flags & MAKEENTRY && !VN_IS_DOOMED(vn))
587 cache_enter(vn, *vpp, cnp);
588 PFS_RETURN (0);
589 failed:
590 if (cnp->cn_flags & ISDOTDOT) {
591 vfs_unbusy(mp);
592 vn_lock(vn, LK_EXCLUSIVE | LK_RETRY);
593 *vpp = NULL;
594 }
595 PFS_RETURN(error);
596 }
597
598 /*
599 * Open a file or directory.
600 */
601 static int
pfs_open(struct vop_open_args * va)602 pfs_open(struct vop_open_args *va)
603 {
604 struct vnode *vn = va->a_vp;
605 struct pfs_vdata *pvd = vn->v_data;
606 struct pfs_node *pn = pvd->pvd_pn;
607 int mode = va->a_mode;
608
609 PFS_TRACE(("%s (mode 0x%x)", pn->pn_name, mode));
610 pfs_assert_not_owned(pn);
611
612 /* check if the requested mode is permitted */
613 if (((mode & FREAD) && !(mode & PFS_RD)) ||
614 ((mode & FWRITE) && !(mode & PFS_WR)))
615 PFS_RETURN (EPERM);
616
617 /* we don't support locking */
618 if ((mode & O_SHLOCK) || (mode & O_EXLOCK))
619 PFS_RETURN (EOPNOTSUPP);
620
621 PFS_RETURN (0);
622 }
623
624 struct sbuf_seek_helper {
625 off_t skip_bytes;
626 struct uio *uio;
627 };
628
629 static int
pfs_sbuf_uio_drain(void * arg,const char * data,int len)630 pfs_sbuf_uio_drain(void *arg, const char *data, int len)
631 {
632 struct sbuf_seek_helper *ssh;
633 struct uio *uio;
634 int error, skipped;
635
636 ssh = arg;
637 uio = ssh->uio;
638 skipped = 0;
639
640 /* Need to discard first uio_offset bytes. */
641 if (ssh->skip_bytes > 0) {
642 if (ssh->skip_bytes >= len) {
643 ssh->skip_bytes -= len;
644 return (len);
645 }
646
647 data += ssh->skip_bytes;
648 len -= ssh->skip_bytes;
649 skipped = ssh->skip_bytes;
650 ssh->skip_bytes = 0;
651 }
652
653 error = uiomove(__DECONST(void *, data), len, uio);
654 if (error != 0)
655 return (-error);
656
657 /*
658 * The fill function has more to emit, but the reader is finished.
659 * This is similar to the truncated read case for non-draining PFS
660 * sbufs, and should be handled appropriately in fill-routines.
661 */
662 if (uio->uio_resid == 0)
663 return (-ENOBUFS);
664
665 return (skipped + len);
666 }
667
668 /*
669 * Read from a file
670 */
671 static int
pfs_read(struct vop_read_args * va)672 pfs_read(struct vop_read_args *va)
673 {
674 struct vnode *vn = va->a_vp;
675 struct pfs_vdata *pvd = vn->v_data;
676 struct pfs_node *pn = pvd->pvd_pn;
677 struct uio *uio = va->a_uio;
678 struct proc *proc;
679 struct thread *td;
680 struct sbuf *sb = NULL;
681 int error, locked;
682 off_t buflen, buflim;
683 struct sbuf_seek_helper ssh;
684
685 PFS_TRACE(("%s", pn->pn_name));
686 pfs_assert_not_owned(pn);
687
688 if (vn->v_type != VREG)
689 PFS_RETURN (EINVAL);
690 KASSERT_PN_IS_FILE(pn);
691
692 if (!(pn->pn_flags & PFS_RD))
693 PFS_RETURN (EBADF);
694
695 if (pn->pn_fill == NULL)
696 PFS_RETURN (EIO);
697
698 td = curthread;
699
700 /*
701 * This is necessary because either process' privileges may
702 * have changed since the open() call.
703 */
704 if (!pfs_visible(td, pn, pvd->pvd_pid, &proc))
705 PFS_RETURN (EIO);
706
707 vhold(vn);
708 locked = VOP_ISLOCKED(vn);
709 VOP_UNLOCK(vn);
710
711 if (proc != NULL) {
712 _PHOLD(proc);
713 execve_block_wait(td, proc);
714 if (!pfs_visible_proc(td, pn, proc)) {
715 PROC_UNLOCK(proc);
716 error = EIO;
717 goto ret;
718 }
719 PROC_UNLOCK(proc);
720 }
721
722 if (pn->pn_flags & PFS_RAWRD) {
723 PFS_TRACE(("%zd resid", uio->uio_resid));
724 error = pn_fill(curthread, proc, pn, NULL, uio);
725 PFS_TRACE(("%zd resid", uio->uio_resid));
726 goto ret;
727 }
728
729 if (uio->uio_resid < 0 || uio->uio_offset < 0 ||
730 uio->uio_resid > OFF_MAX - uio->uio_offset) {
731 error = EINVAL;
732 goto ret;
733 }
734 buflen = uio->uio_offset + uio->uio_resid + 1;
735 if (pn->pn_flags & PFS_AUTODRAIN)
736 /*
737 * We can use a smaller buffer if we can stream output to the
738 * consumer.
739 */
740 buflim = PAGE_SIZE;
741 else
742 buflim = PFS_MAXBUFSIZ;
743 if (buflen > buflim)
744 buflen = buflim;
745
746 sb = sbuf_new(sb, NULL, buflen, 0);
747 if (sb == NULL) {
748 error = EIO;
749 goto ret;
750 }
751
752 if (pn->pn_flags & PFS_AUTODRAIN) {
753 ssh.skip_bytes = uio->uio_offset;
754 ssh.uio = uio;
755 sbuf_set_drain(sb, pfs_sbuf_uio_drain, &ssh);
756 }
757
758 error = pn_fill(curthread, proc, pn, sb, uio);
759
760 if (error) {
761 sbuf_delete(sb);
762 goto ret;
763 }
764
765 /*
766 * XXX: If the buffer overflowed, sbuf_len() will not return
767 * the data length. Then just use the full length because an
768 * overflowed sbuf must be full.
769 */
770 error = sbuf_finish(sb);
771 if ((pn->pn_flags & PFS_AUTODRAIN)) {
772 /*
773 * ENOBUFS just indicates early termination of the fill
774 * function as the caller's buffer was already filled. Squash
775 * to zero.
776 */
777 if (uio->uio_resid == 0 && error == ENOBUFS)
778 error = 0;
779 } else {
780 if (error == 0)
781 buflen = sbuf_len(sb);
782 else
783 /* The trailing byte is not valid. */
784 buflen--;
785 error = uiomove_frombuf(sbuf_data(sb), buflen, uio);
786 }
787 sbuf_delete(sb);
788 ret:
789 vn_lock(vn, locked | LK_RETRY);
790 vdrop(vn);
791 if (proc != NULL) {
792 PROC_LOCK(proc);
793 execve_unblock(td, proc);
794 _PRELE(proc);
795 PROC_UNLOCK(proc);
796 }
797 PFS_RETURN (error);
798 }
799
800 /*
801 * Iterate through directory entries
802 */
803 static int
pfs_iterate(struct thread * td,struct proc * proc,struct pfs_node * pd,struct pfs_node ** pn,struct proc ** p)804 pfs_iterate(struct thread *td, struct proc *proc, struct pfs_node *pd,
805 struct pfs_node **pn, struct proc **p)
806 {
807 int visible;
808
809 sx_assert(&allproc_lock, SX_SLOCKED);
810 pfs_assert_owned(pd);
811 again:
812 if (*pn == NULL) {
813 /* first node */
814 *pn = pd->pn_nodes;
815 } else if ((*pn)->pn_type != pfstype_procdir) {
816 /* next node */
817 *pn = (*pn)->pn_next;
818 }
819 if (*pn != NULL && (*pn)->pn_type == pfstype_procdir) {
820 /* next process */
821 if (*p == NULL)
822 *p = LIST_FIRST(&allproc);
823 else
824 *p = LIST_NEXT(*p, p_list);
825 /* out of processes: next node */
826 if (*p == NULL)
827 *pn = (*pn)->pn_next;
828 else
829 PROC_LOCK(*p);
830 }
831
832 if ((*pn) == NULL)
833 return (-1);
834
835 if (*p != NULL) {
836 visible = pfs_visible_proc(td, *pn, *p);
837 PROC_UNLOCK(*p);
838 } else if (proc != NULL) {
839 visible = pfs_visible_proc(td, *pn, proc);
840 } else {
841 visible = pn_vis(td, NULL, *pn);
842 }
843 if (!visible)
844 goto again;
845
846 return (0);
847 }
848
849 /* Directory entry list */
850 struct pfsentry {
851 STAILQ_ENTRY(pfsentry) link;
852 struct dirent entry;
853 };
854 STAILQ_HEAD(pfsdirentlist, pfsentry);
855
856 /*
857 * Return directory entries.
858 */
859 static int
pfs_readdir(struct vop_readdir_args * va)860 pfs_readdir(struct vop_readdir_args *va)
861 {
862 struct vnode *vn = va->a_vp;
863 struct pfs_vdata *pvd = vn->v_data;
864 struct pfs_node *pd = pvd->pvd_pn;
865 pid_t pid = pvd->pvd_pid;
866 struct proc *p, *proc;
867 struct thread *td;
868 struct pfs_node *pn;
869 struct uio *uio;
870 struct pfsentry *pfsent, *pfsent2;
871 struct pfsdirentlist lst;
872 off_t coffset, offset;
873 int error, i, resid;
874
875 STAILQ_INIT(&lst);
876 error = 0;
877 KASSERT(pd->pn_info == vn->v_mount->mnt_data,
878 ("%s(): pn_info does not match mountpoint", __func__));
879 PFS_TRACE(("%s pid %lu", pd->pn_name, (unsigned long)pid));
880 pfs_assert_not_owned(pd);
881
882 if (va->a_eofflag != NULL)
883 *va->a_eofflag = 0;
884
885 if (vn->v_type != VDIR)
886 PFS_RETURN (ENOTDIR);
887 KASSERT_PN_IS_DIR(pd);
888 uio = va->a_uio;
889
890 /* only allow reading entire entries */
891 offset = uio->uio_offset;
892 resid = uio->uio_resid;
893 if (offset < 0 || offset % PFS_DELEN != 0 ||
894 (resid && resid < PFS_DELEN))
895 PFS_RETURN (EINVAL);
896 if (resid == 0)
897 PFS_RETURN (0);
898
899 proc = NULL;
900 if (pid != NO_PID && !pfs_lookup_proc(pid, &proc))
901 PFS_RETURN (ENOENT);
902
903 /*
904 * The allproc lock is required in pfs_iterate() for procdir
905 * directories.
906 */
907 sx_slock(&allproc_lock);
908 pfs_lock(pd);
909
910 KASSERT(pid == NO_PID || proc != NULL,
911 ("%s(): no process for pid %lu", __func__, (unsigned long)pid));
912
913 td = curthread;
914 if (pid != NO_PID) {
915 PROC_LOCK(proc);
916
917 /* check if the directory is visible to the caller */
918 if (!pfs_visible_proc(curthread, pd, proc)) {
919 execve_unblock(td, proc);
920 _PRELE(proc);
921 PROC_UNLOCK(proc);
922 pfs_unlock(pd);
923 sx_sunlock(&allproc_lock);
924 PFS_RETURN (ENOENT);
925 }
926 }
927
928 for (pn = NULL, p = NULL, coffset = 0; resid >= PFS_DELEN;
929 coffset += PFS_DELEN) {
930 if (pfs_iterate(curthread, proc, pd, &pn, &p) == -1) {
931 if (va->a_eofflag != NULL)
932 *va->a_eofflag = 1;
933 break;
934 }
935 if (coffset < offset)
936 continue;
937 if ((pfsent = malloc(sizeof(struct pfsentry), M_IOV,
938 M_NOWAIT | M_ZERO)) == NULL) {
939 error = ENOMEM;
940 break;
941 }
942 pfsent->entry.d_reclen = PFS_DELEN;
943 pfsent->entry.d_fileno = pn_fileno(pn, pid);
944 /* PFS_DELEN was picked to fit PFS_NAMLEN */
945 for (i = 0; i < PFS_NAMELEN - 1 && pn->pn_name[i] != '\0'; ++i)
946 pfsent->entry.d_name[i] = pn->pn_name[i];
947 pfsent->entry.d_namlen = i;
948 /* NOTE: d_off is the offset of the *next* entry. */
949 pfsent->entry.d_off = offset + PFS_DELEN;
950 switch (pn->pn_type) {
951 case pfstype_procdir:
952 KASSERT(p != NULL,
953 ("reached procdir node with p == NULL"));
954 pfsent->entry.d_namlen = snprintf(pfsent->entry.d_name,
955 PFS_NAMELEN, "%d", p->p_pid);
956 /* fall through */
957 case pfstype_root:
958 case pfstype_dir:
959 case pfstype_this:
960 case pfstype_parent:
961 pfsent->entry.d_type = DT_DIR;
962 break;
963 case pfstype_file:
964 pfsent->entry.d_type = DT_REG;
965 break;
966 case pfstype_symlink:
967 pfsent->entry.d_type = DT_LNK;
968 break;
969 default:
970 panic("%s has unexpected node type: %d", pn->pn_name, pn->pn_type);
971 }
972 PFS_TRACE(("%s", pfsent->entry.d_name));
973 dirent_terminate(&pfsent->entry);
974 STAILQ_INSERT_TAIL(&lst, pfsent, link);
975 offset += PFS_DELEN;
976 resid -= PFS_DELEN;
977 }
978 if (proc != NULL) {
979 execve_unblock(td, proc);
980 _PRELE(proc);
981 PROC_UNLOCK(proc);
982 }
983 pfs_unlock(pd);
984 sx_sunlock(&allproc_lock);
985 i = 0;
986 STAILQ_FOREACH_SAFE(pfsent, &lst, link, pfsent2) {
987 if (error == 0)
988 error = uiomove(&pfsent->entry, PFS_DELEN, uio);
989 free(pfsent, M_IOV);
990 i++;
991 }
992 PFS_TRACE(("%ju bytes", (uintmax_t)(i * PFS_DELEN)));
993 PFS_RETURN (error);
994 }
995
996 /*
997 * Read a symbolic link
998 */
999 static int
pfs_readlink(struct vop_readlink_args * va)1000 pfs_readlink(struct vop_readlink_args *va)
1001 {
1002 struct vnode *vn = va->a_vp;
1003 struct pfs_vdata *pvd = vn->v_data;
1004 struct pfs_node *pn = pvd->pvd_pn;
1005 struct uio *uio = va->a_uio;
1006 struct proc *proc = NULL;
1007 char buf[PATH_MAX];
1008 struct sbuf sb;
1009 int error, locked;
1010
1011 PFS_TRACE(("%s", pn->pn_name));
1012 pfs_assert_not_owned(pn);
1013
1014 if (vn->v_type != VLNK)
1015 PFS_RETURN (EINVAL);
1016 KASSERT_PN_IS_LINK(pn);
1017
1018 if (pn->pn_fill == NULL)
1019 PFS_RETURN (EIO);
1020
1021 if (pvd->pvd_pid != NO_PID) {
1022 if ((proc = pfind(pvd->pvd_pid)) == NULL)
1023 PFS_RETURN (EIO);
1024 if (proc->p_flag & P_WEXIT) {
1025 PROC_UNLOCK(proc);
1026 PFS_RETURN (EIO);
1027 }
1028 _PHOLD(proc);
1029 PROC_UNLOCK(proc);
1030 }
1031 vhold(vn);
1032 locked = VOP_ISLOCKED(vn);
1033 VOP_UNLOCK(vn);
1034
1035 /* sbuf_new() can't fail with a static buffer */
1036 sbuf_new(&sb, buf, sizeof buf, 0);
1037
1038 error = pn_fill(curthread, proc, pn, &sb, NULL);
1039
1040 if (proc != NULL)
1041 PRELE(proc);
1042 vn_lock(vn, locked | LK_RETRY);
1043 vdrop(vn);
1044
1045 if (error) {
1046 sbuf_delete(&sb);
1047 PFS_RETURN (error);
1048 }
1049
1050 if (sbuf_finish(&sb) != 0) {
1051 sbuf_delete(&sb);
1052 PFS_RETURN (ENAMETOOLONG);
1053 }
1054
1055 error = uiomove_frombuf(sbuf_data(&sb), sbuf_len(&sb), uio);
1056 sbuf_delete(&sb);
1057 PFS_RETURN (error);
1058 }
1059
1060 /*
1061 * Reclaim a vnode
1062 */
1063 static int
pfs_reclaim(struct vop_reclaim_args * va)1064 pfs_reclaim(struct vop_reclaim_args *va)
1065 {
1066 struct vnode *vn = va->a_vp;
1067 struct pfs_vdata *pvd = vn->v_data;
1068 struct pfs_node *pn = pvd->pvd_pn;
1069
1070 PFS_TRACE(("%s", pn->pn_name));
1071 pfs_assert_not_owned(pn);
1072
1073 return (pfs_vncache_free(va->a_vp));
1074 }
1075
1076 /*
1077 * Set attributes
1078 */
1079 static int
pfs_setattr(struct vop_setattr_args * va)1080 pfs_setattr(struct vop_setattr_args *va)
1081 {
1082 struct vnode *vn = va->a_vp;
1083 struct pfs_vdata *pvd = vn->v_data;
1084 struct pfs_node *pn = pvd->pvd_pn;
1085
1086 PFS_TRACE(("%s", pn->pn_name));
1087 pfs_assert_not_owned(pn);
1088
1089 /* Silently ignore unchangeable attributes. */
1090 PFS_RETURN (0);
1091 }
1092
1093 /*
1094 * Write to a file
1095 */
1096 static int
pfs_write(struct vop_write_args * va)1097 pfs_write(struct vop_write_args *va)
1098 {
1099 struct vnode *vn = va->a_vp;
1100 struct pfs_vdata *pvd = vn->v_data;
1101 struct pfs_node *pn = pvd->pvd_pn;
1102 struct uio *uio = va->a_uio;
1103 struct proc *proc;
1104 struct thread *td;
1105 struct sbuf sb;
1106 int error;
1107
1108 PFS_TRACE(("%s", pn->pn_name));
1109 pfs_assert_not_owned(pn);
1110
1111 if (vn->v_type != VREG)
1112 PFS_RETURN (EINVAL);
1113 KASSERT_PN_IS_FILE(pn);
1114
1115 if (!(pn->pn_flags & PFS_WR))
1116 PFS_RETURN (EBADF);
1117
1118 if (pn->pn_fill == NULL)
1119 PFS_RETURN (EIO);
1120
1121 if (uio->uio_resid > PFS_MAXBUFSIZ)
1122 PFS_RETURN (EIO);
1123
1124 td = curthread;
1125
1126 /*
1127 * This is necessary because either process' privileges may
1128 * have changed since the open() call.
1129 */
1130 if (!pfs_visible(td, pn, pvd->pvd_pid, &proc))
1131 PFS_RETURN (EIO);
1132 if (proc != NULL) {
1133 _PHOLD(proc);
1134 execve_block_wait(td, proc);
1135 if (!pfs_visible_proc(td, pn, proc)) {
1136 PROC_UNLOCK(proc);
1137 error = EIO;
1138 goto out;
1139 }
1140 PROC_UNLOCK(proc);
1141 }
1142
1143 if (pn->pn_flags & PFS_RAWWR) {
1144 error = pn_fill(curthread, proc, pn, NULL, uio);
1145 goto out;
1146 }
1147
1148 sbuf_uionew(&sb, uio, &error);
1149 if (error != 0)
1150 goto out;
1151
1152 error = pn_fill(curthread, proc, pn, &sb, uio);
1153
1154 sbuf_delete(&sb);
1155 out:
1156 if (proc != NULL) {
1157 PROC_LOCK(proc);
1158 execve_unblock(td, proc);
1159 _PRELE(proc);
1160 PROC_UNLOCK(proc);
1161 }
1162 PFS_RETURN (error);
1163 }
1164
1165 /*
1166 * Vnode operations
1167 */
1168 struct vop_vector pfs_vnodeops = {
1169 .vop_default = &default_vnodeops,
1170
1171 .vop_access = pfs_access,
1172 .vop_cachedlookup = pfs_lookup,
1173 .vop_close = pfs_close,
1174 .vop_create = VOP_EOPNOTSUPP,
1175 .vop_getattr = pfs_getattr,
1176 .vop_getextattr = pfs_getextattr,
1177 .vop_ioctl = pfs_ioctl,
1178 .vop_link = VOP_EOPNOTSUPP,
1179 .vop_lookup = vfs_cache_lookup,
1180 .vop_mkdir = VOP_EOPNOTSUPP,
1181 .vop_mknod = VOP_EOPNOTSUPP,
1182 .vop_open = pfs_open,
1183 .vop_read = pfs_read,
1184 .vop_readdir = pfs_readdir,
1185 .vop_readlink = pfs_readlink,
1186 .vop_reclaim = pfs_reclaim,
1187 .vop_remove = VOP_EOPNOTSUPP,
1188 .vop_rename = VOP_EOPNOTSUPP,
1189 .vop_rmdir = VOP_EOPNOTSUPP,
1190 .vop_setattr = pfs_setattr,
1191 .vop_symlink = VOP_EOPNOTSUPP,
1192 .vop_vptocnp = pfs_vptocnp,
1193 .vop_write = pfs_write,
1194 .vop_add_writecount = vop_stdadd_writecount_nomsync,
1195 /* XXX I've probably forgotten a few that need VOP_EOPNOTSUPP */
1196 };
1197 VFS_VOP_VECTOR_REGISTER(pfs_vnodeops);
1198