xref: /original-bsd/sys/miscfs/union/union_subr.c (revision f4a18198)
1 /*
2  * Copyright (c) 1994 Jan-Simon Pendry
3  * Copyright (c) 1994
4  *	The Regents of the University of California.  All rights reserved.
5  *
6  * This code is derived from software contributed to Berkeley by
7  * Jan-Simon Pendry.
8  *
9  * %sccs.include.redist.c%
10  *
11  *	@(#)union_subr.c	8.11 (Berkeley) 06/17/94
12  */
13 
14 #include <sys/param.h>
15 #include <sys/systm.h>
16 #include <sys/time.h>
17 #include <sys/kernel.h>
18 #include <sys/vnode.h>
19 #include <sys/namei.h>
20 #include <sys/malloc.h>
21 #include <sys/file.h>
22 #include <sys/filedesc.h>
23 #include <sys/queue.h>
24 #include <sys/mount.h>
25 #include <vm/vm.h>		/* for vnode_pager_setsize */
26 #include <miscfs/union/union.h>
27 
28 #ifdef DIAGNOSTIC
29 #include <sys/proc.h>
30 #endif
31 
32 /* must be power of two, otherwise change UNION_HASH() */
33 #define NHASH 32
34 
35 /* unsigned int ... */
36 #define UNION_HASH(u, l) \
37 	(((((unsigned long) (u)) + ((unsigned long) l)) >> 8) & (NHASH-1))
38 
39 static LIST_HEAD(unhead, union_node) unhead[NHASH];
40 static int unvplock[NHASH];
41 
42 int
43 union_init()
44 {
45 	int i;
46 
47 	for (i = 0; i < NHASH; i++)
48 		LIST_INIT(&unhead[i]);
49 	bzero((caddr_t) unvplock, sizeof(unvplock));
50 }
51 
52 static int
53 union_list_lock(ix)
54 	int ix;
55 {
56 
57 	if (unvplock[ix] & UN_LOCKED) {
58 		unvplock[ix] |= UN_WANT;
59 		sleep((caddr_t) &unvplock[ix], PINOD);
60 		return (1);
61 	}
62 
63 	unvplock[ix] |= UN_LOCKED;
64 
65 	return (0);
66 }
67 
68 static void
69 union_list_unlock(ix)
70 	int ix;
71 {
72 
73 	unvplock[ix] &= ~UN_LOCKED;
74 
75 	if (unvplock[ix] & UN_WANT) {
76 		unvplock[ix] &= ~UN_WANT;
77 		wakeup((caddr_t) &unvplock[ix]);
78 	}
79 }
80 
81 void
82 union_updatevp(un, uppervp, lowervp)
83 	struct union_node *un;
84 	struct vnode *uppervp;
85 	struct vnode *lowervp;
86 {
87 	int ohash = UNION_HASH(un->un_uppervp, un->un_lowervp);
88 	int nhash = UNION_HASH(uppervp, lowervp);
89 	int docache = (lowervp != NULLVP || uppervp != NULLVP);
90 
91 	/*
92 	 * Ensure locking is ordered from lower to higher
93 	 * to avoid deadlocks.
94 	 */
95 	if (nhash < ohash) {
96 		int t = ohash;
97 		ohash = nhash;
98 		nhash = t;
99 	}
100 
101 	if (ohash != nhash)
102 		while (union_list_lock(ohash))
103 			continue;
104 
105 	while (union_list_lock(nhash))
106 		continue;
107 
108 	if (ohash != nhash || !docache) {
109 		if (un->un_flags & UN_CACHED) {
110 			un->un_flags &= ~UN_CACHED;
111 			LIST_REMOVE(un, un_cache);
112 		}
113 	}
114 
115 	if (ohash != nhash)
116 		union_list_unlock(ohash);
117 
118 	if (un->un_lowervp != lowervp) {
119 		if (un->un_lowervp) {
120 			vrele(un->un_lowervp);
121 			if (un->un_path) {
122 				free(un->un_path, M_TEMP);
123 				un->un_path = 0;
124 			}
125 			if (un->un_dirvp) {
126 				vrele(un->un_dirvp);
127 				un->un_dirvp = NULLVP;
128 			}
129 		}
130 		un->un_lowervp = lowervp;
131 		un->un_lowersz = VNOVAL;
132 	}
133 
134 	if (un->un_uppervp != uppervp) {
135 		if (un->un_uppervp)
136 			vrele(un->un_uppervp);
137 
138 		un->un_uppervp = uppervp;
139 		un->un_uppersz = VNOVAL;
140 	}
141 
142 	if (docache && (ohash != nhash)) {
143 		LIST_INSERT_HEAD(&unhead[nhash], un, un_cache);
144 		un->un_flags |= UN_CACHED;
145 	}
146 
147 	union_list_unlock(nhash);
148 }
149 
150 void
151 union_newlower(un, lowervp)
152 	struct union_node *un;
153 	struct vnode *lowervp;
154 {
155 
156 	union_updatevp(un, un->un_uppervp, lowervp);
157 }
158 
159 void
160 union_newupper(un, uppervp)
161 	struct union_node *un;
162 	struct vnode *uppervp;
163 {
164 
165 	union_updatevp(un, uppervp, un->un_lowervp);
166 }
167 
168 /*
169  * Keep track of size changes in the underlying vnodes.
170  * If the size changes, then callback to the vm layer
171  * giving priority to the upper layer size.
172  */
173 void
174 union_newsize(vp, uppersz, lowersz)
175 	struct vnode *vp;
176 	off_t uppersz, lowersz;
177 {
178 	struct union_node *un;
179 	off_t sz;
180 
181 	/* only interested in regular files */
182 	if (vp->v_type != VREG)
183 		return;
184 
185 	un = VTOUNION(vp);
186 	sz = VNOVAL;
187 
188 	if ((uppersz != VNOVAL) && (un->un_uppersz != uppersz)) {
189 		un->un_uppersz = uppersz;
190 		if (sz == VNOVAL)
191 			sz = un->un_uppersz;
192 	}
193 
194 	if ((lowersz != VNOVAL) && (un->un_lowersz != lowersz)) {
195 		un->un_lowersz = lowersz;
196 		if (sz == VNOVAL)
197 			sz = un->un_lowersz;
198 	}
199 
200 	if (sz != VNOVAL) {
201 #ifdef UNION_DIAGNOSTIC
202 		printf("union: %s size now %ld\n",
203 			uppersz != VNOVAL ? "upper" : "lower", (long) sz);
204 #endif
205 		vnode_pager_setsize(vp, sz);
206 	}
207 }
208 
209 /*
210  * allocate a union_node/vnode pair.  the vnode is
211  * referenced and locked.  the new vnode is returned
212  * via (vpp).  (mp) is the mountpoint of the union filesystem,
213  * (dvp) is the parent directory where the upper layer object
214  * should exist (but doesn't) and (cnp) is the componentname
215  * information which is partially copied to allow the upper
216  * layer object to be created at a later time.  (uppervp)
217  * and (lowervp) reference the upper and lower layer objects
218  * being mapped.  either, but not both, can be nil.
219  * if supplied, (uppervp) is locked.
220  * the reference is either maintained in the new union_node
221  * object which is allocated, or they are vrele'd.
222  *
223  * all union_nodes are maintained on a singly-linked
224  * list.  new nodes are only allocated when they cannot
225  * be found on this list.  entries on the list are
226  * removed when the vfs reclaim entry is called.
227  *
228  * a single lock is kept for the entire list.  this is
229  * needed because the getnewvnode() function can block
230  * waiting for a vnode to become free, in which case there
231  * may be more than one process trying to get the same
232  * vnode.  this lock is only taken if we are going to
233  * call getnewvnode, since the kernel itself is single-threaded.
234  *
235  * if an entry is found on the list, then call vget() to
236  * take a reference.  this is done because there may be
237  * zero references to it and so it needs to removed from
238  * the vnode free list.
239  */
240 int
241 union_allocvp(vpp, mp, undvp, dvp, cnp, uppervp, lowervp)
242 	struct vnode **vpp;
243 	struct mount *mp;
244 	struct vnode *undvp;		/* parent union vnode */
245 	struct vnode *dvp;		/* may be null */
246 	struct componentname *cnp;	/* may be null */
247 	struct vnode *uppervp;		/* may be null */
248 	struct vnode *lowervp;		/* may be null */
249 {
250 	int error;
251 	struct union_node *un;
252 	struct union_node **pp;
253 	struct vnode *xlowervp = NULLVP;
254 	struct union_mount *um = MOUNTTOUNIONMOUNT(mp);
255 	int hash;
256 	int vflag;
257 	int try;
258 
259 	if (uppervp == NULLVP && lowervp == NULLVP)
260 		panic("union: unidentifiable allocation");
261 
262 	if (uppervp && lowervp && (uppervp->v_type != lowervp->v_type)) {
263 		xlowervp = lowervp;
264 		lowervp = NULLVP;
265 	}
266 
267 	/* detect the root vnode (and aliases) */
268 	vflag = 0;
269 	if ((uppervp == um->um_uppervp) &&
270 	    ((lowervp == NULLVP) || lowervp == um->um_lowervp)) {
271 		if (lowervp == NULLVP) {
272 			lowervp = um->um_lowervp;
273 			if (lowervp != NULLVP)
274 				VREF(lowervp);
275 		}
276 		vflag = VROOT;
277 	}
278 
279 loop:
280 	for (try = 0; try < 3; try++) {
281 		switch (try) {
282 		case 0:
283 			if (lowervp == NULLVP)
284 				continue;
285 			hash = UNION_HASH(uppervp, lowervp);
286 			break;
287 
288 		case 1:
289 			if (uppervp == NULLVP)
290 				continue;
291 			hash = UNION_HASH(uppervp, NULLVP);
292 			break;
293 
294 		case 2:
295 			if (lowervp == NULLVP)
296 				continue;
297 			hash = UNION_HASH(NULLVP, lowervp);
298 			break;
299 		}
300 
301 		while (union_list_lock(hash))
302 			continue;
303 
304 		for (un = unhead[hash].lh_first; un != 0;
305 					un = un->un_cache.le_next) {
306 			if ((un->un_lowervp == lowervp ||
307 			     un->un_lowervp == NULLVP) &&
308 			    (un->un_uppervp == uppervp ||
309 			     un->un_uppervp == NULLVP) &&
310 			    (UNIONTOV(un)->v_mount == mp)) {
311 				if (vget(UNIONTOV(un), 0)) {
312 					union_list_unlock(hash);
313 					goto loop;
314 				}
315 				break;
316 			}
317 		}
318 
319 		union_list_unlock(hash);
320 
321 		if (un)
322 			break;
323 	}
324 
325 	if (un) {
326 		/*
327 		 * Obtain a lock on the union_node.
328 		 * uppervp is locked, though un->un_uppervp
329 		 * may not be.  this doesn't break the locking
330 		 * hierarchy since in the case that un->un_uppervp
331 		 * is not yet locked it will be vrele'd and replaced
332 		 * with uppervp.
333 		 */
334 
335 		if ((dvp != NULLVP) && (uppervp == dvp)) {
336 			/*
337 			 * Access ``.'', so (un) will already
338 			 * be locked.  Since this process has
339 			 * the lock on (uppervp) no other
340 			 * process can hold the lock on (un).
341 			 */
342 #ifdef DIAGNOSTIC
343 			if ((un->un_flags & UN_LOCKED) == 0)
344 				panic("union: . not locked");
345 			else if (curproc && un->un_pid != curproc->p_pid &&
346 				    un->un_pid > -1 && curproc->p_pid > -1)
347 				panic("union: allocvp not lock owner");
348 #endif
349 		} else {
350 			if (un->un_flags & UN_LOCKED) {
351 				vrele(UNIONTOV(un));
352 				un->un_flags |= UN_WANT;
353 				sleep((caddr_t) &un->un_flags, PINOD);
354 				goto loop;
355 			}
356 			un->un_flags |= UN_LOCKED;
357 
358 #ifdef DIAGNOSTIC
359 			if (curproc)
360 				un->un_pid = curproc->p_pid;
361 			else
362 				un->un_pid = -1;
363 #endif
364 		}
365 
366 		/*
367 		 * At this point, the union_node is locked,
368 		 * un->un_uppervp may not be locked, and uppervp
369 		 * is locked or nil.
370 		 */
371 
372 		/*
373 		 * Save information about the upper layer.
374 		 */
375 		if (uppervp != un->un_uppervp) {
376 			union_newupper(un, uppervp);
377 		} else if (uppervp) {
378 			vrele(uppervp);
379 		}
380 
381 		if (un->un_uppervp) {
382 			un->un_flags |= UN_ULOCK;
383 			un->un_flags &= ~UN_KLOCK;
384 		}
385 
386 		/*
387 		 * Save information about the lower layer.
388 		 * This needs to keep track of pathname
389 		 * and directory information which union_vn_create
390 		 * might need.
391 		 */
392 		if (lowervp != un->un_lowervp) {
393 			union_newlower(un, lowervp);
394 			if (cnp && (lowervp != NULLVP) &&
395 			    (lowervp->v_type == VREG)) {
396 				un->un_hash = cnp->cn_hash;
397 				un->un_path = malloc(cnp->cn_namelen+1,
398 						M_TEMP, M_WAITOK);
399 				bcopy(cnp->cn_nameptr, un->un_path,
400 						cnp->cn_namelen);
401 				un->un_path[cnp->cn_namelen] = '\0';
402 				VREF(dvp);
403 				un->un_dirvp = dvp;
404 			}
405 		} else if (lowervp) {
406 			vrele(lowervp);
407 		}
408 		*vpp = UNIONTOV(un);
409 		return (0);
410 	}
411 
412 	/*
413 	 * otherwise lock the vp list while we call getnewvnode
414 	 * since that can block.
415 	 */
416 	hash = UNION_HASH(uppervp, lowervp);
417 
418 	if (union_list_lock(hash))
419 		goto loop;
420 
421 	error = getnewvnode(VT_UNION, mp, union_vnodeop_p, vpp);
422 	if (error) {
423 		if (uppervp) {
424 			if (dvp == uppervp)
425 				vrele(uppervp);
426 			else
427 				vput(uppervp);
428 		}
429 		if (lowervp)
430 			vrele(lowervp);
431 
432 		goto out;
433 	}
434 
435 	MALLOC((*vpp)->v_data, void *, sizeof(struct union_node),
436 		M_TEMP, M_WAITOK);
437 
438 	(*vpp)->v_flag |= vflag;
439 	if (uppervp)
440 		(*vpp)->v_type = uppervp->v_type;
441 	else
442 		(*vpp)->v_type = lowervp->v_type;
443 	un = VTOUNION(*vpp);
444 	un->un_vnode = *vpp;
445 	un->un_uppervp = uppervp;
446 	un->un_uppersz = VNOVAL;
447 	un->un_lowervp = lowervp;
448 	un->un_lowersz = VNOVAL;
449 	un->un_pvp = undvp;
450 	if (undvp != NULLVP)
451 		VREF(undvp);
452 	un->un_openl = 0;
453 	un->un_flags = UN_LOCKED;
454 	if (un->un_uppervp)
455 		un->un_flags |= UN_ULOCK;
456 #ifdef DIAGNOSTIC
457 	if (curproc)
458 		un->un_pid = curproc->p_pid;
459 	else
460 		un->un_pid = -1;
461 #endif
462 	if (cnp && (lowervp != NULLVP) && (lowervp->v_type == VREG)) {
463 		un->un_hash = cnp->cn_hash;
464 		un->un_path = malloc(cnp->cn_namelen+1, M_TEMP, M_WAITOK);
465 		bcopy(cnp->cn_nameptr, un->un_path, cnp->cn_namelen);
466 		un->un_path[cnp->cn_namelen] = '\0';
467 		VREF(dvp);
468 		un->un_dirvp = dvp;
469 	} else {
470 		un->un_hash = 0;
471 		un->un_path = 0;
472 		un->un_dirvp = 0;
473 	}
474 
475 	LIST_INSERT_HEAD(&unhead[hash], un, un_cache);
476 	un->un_flags |= UN_CACHED;
477 
478 	if (xlowervp)
479 		vrele(xlowervp);
480 
481 out:
482 	union_list_unlock(hash);
483 
484 	return (error);
485 }
486 
487 int
488 union_freevp(vp)
489 	struct vnode *vp;
490 {
491 	struct union_node *un = VTOUNION(vp);
492 
493 	if (un->un_flags & UN_CACHED) {
494 		un->un_flags &= ~UN_CACHED;
495 		LIST_REMOVE(un, un_cache);
496 	}
497 
498 	if (un->un_pvp != NULLVP)
499 		vrele(un->un_pvp);
500 	if (un->un_uppervp != NULLVP)
501 		vrele(un->un_uppervp);
502 	if (un->un_lowervp != NULLVP)
503 		vrele(un->un_lowervp);
504 	if (un->un_dirvp != NULLVP)
505 		vrele(un->un_dirvp);
506 	if (un->un_path)
507 		free(un->un_path, M_TEMP);
508 
509 	FREE(vp->v_data, M_TEMP);
510 	vp->v_data = 0;
511 
512 	return (0);
513 }
514 
515 /*
516  * copyfile.  copy the vnode (fvp) to the vnode (tvp)
517  * using a sequence of reads and writes.  both (fvp)
518  * and (tvp) are locked on entry and exit.
519  */
520 int
521 union_copyfile(fvp, tvp, cred, p)
522 	struct vnode *fvp;
523 	struct vnode *tvp;
524 	struct ucred *cred;
525 	struct proc *p;
526 {
527 	char *buf;
528 	struct uio uio;
529 	struct iovec iov;
530 	int error = 0;
531 
532 	/*
533 	 * strategy:
534 	 * allocate a buffer of size MAXBSIZE.
535 	 * loop doing reads and writes, keeping track
536 	 * of the current uio offset.
537 	 * give up at the first sign of trouble.
538 	 */
539 
540 	uio.uio_procp = p;
541 	uio.uio_segflg = UIO_SYSSPACE;
542 	uio.uio_offset = 0;
543 
544 	VOP_UNLOCK(fvp);				/* XXX */
545 	LEASE_CHECK(fvp, p, cred, LEASE_READ);
546 	VOP_LOCK(fvp);					/* XXX */
547 	VOP_UNLOCK(tvp);				/* XXX */
548 	LEASE_CHECK(tvp, p, cred, LEASE_WRITE);
549 	VOP_LOCK(tvp);					/* XXX */
550 
551 	buf = malloc(MAXBSIZE, M_TEMP, M_WAITOK);
552 
553 	/* ugly loop follows... */
554 	do {
555 		off_t offset = uio.uio_offset;
556 
557 		uio.uio_iov = &iov;
558 		uio.uio_iovcnt = 1;
559 		iov.iov_base = buf;
560 		iov.iov_len = MAXBSIZE;
561 		uio.uio_resid = iov.iov_len;
562 		uio.uio_rw = UIO_READ;
563 		error = VOP_READ(fvp, &uio, 0, cred);
564 
565 		if (error == 0) {
566 			uio.uio_iov = &iov;
567 			uio.uio_iovcnt = 1;
568 			iov.iov_base = buf;
569 			iov.iov_len = MAXBSIZE - uio.uio_resid;
570 			uio.uio_offset = offset;
571 			uio.uio_rw = UIO_WRITE;
572 			uio.uio_resid = iov.iov_len;
573 
574 			if (uio.uio_resid == 0)
575 				break;
576 
577 			do {
578 				error = VOP_WRITE(tvp, &uio, 0, cred);
579 			} while ((uio.uio_resid > 0) && (error == 0));
580 		}
581 
582 	} while (error == 0);
583 
584 	free(buf, M_TEMP);
585 	return (error);
586 }
587 
588 /*
589  * (un) is assumed to be locked on entry and remains
590  * locked on exit.
591  */
592 int
593 union_copyup(un, docopy, cred, p)
594 	struct union_node *un;
595 	int docopy;
596 	struct ucred *cred;
597 	struct proc *p;
598 {
599 	int error;
600 	struct vnode *lvp, *uvp;
601 
602 	error = union_vn_create(&uvp, un, p);
603 	if (error)
604 		return (error);
605 
606 	/* at this point, uppervp is locked */
607 	union_newupper(un, uvp);
608 	un->un_flags |= UN_ULOCK;
609 
610 	lvp = un->un_lowervp;
611 
612 	if (docopy) {
613 		/*
614 		 * XX - should not ignore errors
615 		 * from VOP_CLOSE
616 		 */
617 		VOP_LOCK(lvp);
618 		error = VOP_OPEN(lvp, FREAD, cred, p);
619 		if (error == 0) {
620 			error = union_copyfile(lvp, uvp, cred, p);
621 			VOP_UNLOCK(lvp);
622 			(void) VOP_CLOSE(lvp, FREAD);
623 		}
624 #ifdef UNION_DIAGNOSTIC
625 		if (error == 0)
626 			uprintf("union: copied up %s\n", un->un_path);
627 #endif
628 
629 	}
630 	un->un_flags &= ~UN_ULOCK;
631 	VOP_UNLOCK(uvp);
632 	union_vn_close(uvp, FWRITE, cred, p);
633 	VOP_LOCK(uvp);
634 	un->un_flags |= UN_ULOCK;
635 
636 	/*
637 	 * Subsequent IOs will go to the top layer, so
638 	 * call close on the lower vnode and open on the
639 	 * upper vnode to ensure that the filesystem keeps
640 	 * its references counts right.  This doesn't do
641 	 * the right thing with (cred) and (FREAD) though.
642 	 * Ignoring error returns is not right, either.
643 	 */
644 	if (error == 0) {
645 		int i;
646 
647 		for (i = 0; i < un->un_openl; i++) {
648 			(void) VOP_CLOSE(lvp, FREAD);
649 			(void) VOP_OPEN(uvp, FREAD, cred, p);
650 		}
651 		un->un_openl = 0;
652 	}
653 
654 	return (error);
655 
656 }
657 
658 /*
659  * Create a shadow directory in the upper layer.
660  * The new vnode is returned locked.
661  *
662  * (um) points to the union mount structure for access to the
663  * the mounting process's credentials.
664  * (dvp) is the directory in which to create the shadow directory.
665  * it is unlocked on entry and exit.
666  * (cnp) is the componentname to be created.
667  * (vpp) is the returned newly created shadow directory, which
668  * is returned locked.
669  */
670 int
671 union_mkshadow(um, dvp, cnp, vpp)
672 	struct union_mount *um;
673 	struct vnode *dvp;
674 	struct componentname *cnp;
675 	struct vnode **vpp;
676 {
677 	int error;
678 	struct vattr va;
679 	struct proc *p = cnp->cn_proc;
680 	struct componentname cn;
681 
682 	/*
683 	 * policy: when creating the shadow directory in the
684 	 * upper layer, create it owned by the user who did
685 	 * the mount, group from parent directory, and mode
686 	 * 777 modified by umask (ie mostly identical to the
687 	 * mkdir syscall).  (jsp, kb)
688 	 */
689 
690 	/*
691 	 * A new componentname structure must be faked up because
692 	 * there is no way to know where the upper level cnp came
693 	 * from or what it is being used for.  This must duplicate
694 	 * some of the work done by NDINIT, some of the work done
695 	 * by namei, some of the work done by lookup and some of
696 	 * the work done by VOP_LOOKUP when given a CREATE flag.
697 	 * Conclusion: Horrible.
698 	 *
699 	 * The pathname buffer will be FREEed by VOP_MKDIR.
700 	 */
701 	cn.cn_pnbuf = malloc(cnp->cn_namelen+1, M_NAMEI, M_WAITOK);
702 	bcopy(cnp->cn_nameptr, cn.cn_pnbuf, cnp->cn_namelen);
703 	cn.cn_pnbuf[cnp->cn_namelen] = '\0';
704 
705 	cn.cn_nameiop = CREATE;
706 	cn.cn_flags = (LOCKPARENT|HASBUF|SAVENAME|SAVESTART|ISLASTCN);
707 	cn.cn_proc = cnp->cn_proc;
708 	if (um->um_op == UNMNT_ABOVE)
709 		cn.cn_cred = cnp->cn_cred;
710 	else
711 		cn.cn_cred = um->um_cred;
712 	cn.cn_nameptr = cn.cn_pnbuf;
713 	cn.cn_namelen = cnp->cn_namelen;
714 	cn.cn_hash = cnp->cn_hash;
715 	cn.cn_consume = cnp->cn_consume;
716 
717 	VREF(dvp);
718 	if (error = relookup(dvp, vpp, &cn))
719 		return (error);
720 	vrele(dvp);
721 
722 	if (*vpp) {
723 		VOP_ABORTOP(dvp, &cn);
724 		VOP_UNLOCK(dvp);
725 		vrele(*vpp);
726 		*vpp = NULLVP;
727 		return (EEXIST);
728 	}
729 
730 	VATTR_NULL(&va);
731 	va.va_type = VDIR;
732 	va.va_mode = um->um_cmode;
733 
734 	/* LEASE_CHECK: dvp is locked */
735 	LEASE_CHECK(dvp, p, p->p_ucred, LEASE_WRITE);
736 
737 	error = VOP_MKDIR(dvp, vpp, &cn, &va);
738 	return (error);
739 }
740 
741 /*
742  * union_vn_create: creates and opens a new shadow file
743  * on the upper union layer.  this function is similar
744  * in spirit to calling vn_open but it avoids calling namei().
745  * the problem with calling namei is that a) it locks too many
746  * things, and b) it doesn't start at the "right" directory,
747  * whereas relookup is told where to start.
748  */
749 int
750 union_vn_create(vpp, un, p)
751 	struct vnode **vpp;
752 	struct union_node *un;
753 	struct proc *p;
754 {
755 	struct vnode *vp;
756 	struct ucred *cred = p->p_ucred;
757 	struct vattr vat;
758 	struct vattr *vap = &vat;
759 	int fmode = FFLAGS(O_WRONLY|O_CREAT|O_TRUNC|O_EXCL);
760 	int error;
761 	int cmode = UN_FILEMODE & ~p->p_fd->fd_cmask;
762 	char *cp;
763 	struct componentname cn;
764 
765 	*vpp = NULLVP;
766 
767 	/*
768 	 * Build a new componentname structure (for the same
769 	 * reasons outlines in union_mkshadow).
770 	 * The difference here is that the file is owned by
771 	 * the current user, rather than by the person who
772 	 * did the mount, since the current user needs to be
773 	 * able to write the file (that's why it is being
774 	 * copied in the first place).
775 	 */
776 	cn.cn_namelen = strlen(un->un_path);
777 	cn.cn_pnbuf = (caddr_t) malloc(cn.cn_namelen, M_NAMEI, M_WAITOK);
778 	bcopy(un->un_path, cn.cn_pnbuf, cn.cn_namelen+1);
779 	cn.cn_nameiop = CREATE;
780 	cn.cn_flags = (LOCKPARENT|HASBUF|SAVENAME|SAVESTART|ISLASTCN);
781 	cn.cn_proc = p;
782 	cn.cn_cred = p->p_ucred;
783 	cn.cn_nameptr = cn.cn_pnbuf;
784 	cn.cn_hash = un->un_hash;
785 	cn.cn_consume = 0;
786 
787 	VREF(un->un_dirvp);
788 	if (error = relookup(un->un_dirvp, &vp, &cn))
789 		return (error);
790 	vrele(un->un_dirvp);
791 
792 	if (vp) {
793 		VOP_ABORTOP(un->un_dirvp, &cn);
794 		if (un->un_dirvp == vp)
795 			vrele(un->un_dirvp);
796 		else
797 			vput(un->un_dirvp);
798 		vrele(vp);
799 		return (EEXIST);
800 	}
801 
802 	/*
803 	 * Good - there was no race to create the file
804 	 * so go ahead and create it.  The permissions
805 	 * on the file will be 0666 modified by the
806 	 * current user's umask.  Access to the file, while
807 	 * it is unioned, will require access to the top *and*
808 	 * bottom files.  Access when not unioned will simply
809 	 * require access to the top-level file.
810 	 * TODO: confirm choice of access permissions.
811 	 */
812 	VATTR_NULL(vap);
813 	vap->va_type = VREG;
814 	vap->va_mode = cmode;
815 	LEASE_CHECK(un->un_dirvp, p, cred, LEASE_WRITE);
816 	if (error = VOP_CREATE(un->un_dirvp, &vp, &cn, vap))
817 		return (error);
818 
819 	if (error = VOP_OPEN(vp, fmode, cred, p)) {
820 		vput(vp);
821 		return (error);
822 	}
823 
824 	vp->v_writecount++;
825 	*vpp = vp;
826 	return (0);
827 }
828 
829 int
830 union_vn_close(vp, fmode, cred, p)
831 	struct vnode *vp;
832 	int fmode;
833 	struct ucred *cred;
834 	struct proc *p;
835 {
836 	if (fmode & FWRITE)
837 		--vp->v_writecount;
838 	return (VOP_CLOSE(vp, fmode));
839 }
840 
841 void
842 union_removed_upper(un)
843 	struct union_node *un;
844 {
845 
846 	if (un->un_flags & UN_ULOCK) {
847 		un->un_flags &= ~UN_ULOCK;
848 		VOP_UNLOCK(un->un_uppervp);
849 	}
850 
851 	if (un->un_flags & UN_CACHED) {
852 		un->un_flags &= ~UN_CACHED;
853 		LIST_REMOVE(un, un_cache);
854 	}
855 }
856 
857 struct vnode *
858 union_lowervp(vp)
859 	struct vnode *vp;
860 {
861 	struct union_node *un = VTOUNION(vp);
862 
863 	if ((un->un_lowervp != NULLVP) &&
864 	    (vp->v_type == un->un_lowervp->v_type)) {
865 		if (vget(un->un_lowervp, 0) == 0)
866 			return (un->un_lowervp);
867 	}
868 
869 	return (NULLVP);
870 }
871