xref: /freebsd/sys/fs/unionfs/union_subr.c (revision 1b6c2589164a3a7b2f62d4c28c2ffa1be860959e)
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  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. All advertising materials mentioning features or use of this software
18  *    must display the following acknowledgement:
19  *	This product includes software developed by the University of
20  *	California, Berkeley and its contributors.
21  * 4. Neither the name of the University nor the names of its contributors
22  *    may be used to endorse or promote products derived from this software
23  *    without specific prior written permission.
24  *
25  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
26  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
27  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
28  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
29  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
30  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
31  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
32  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
33  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
34  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35  * SUCH DAMAGE.
36  *
37  *	@(#)union_subr.c	8.20 (Berkeley) 5/20/95
38  * $FreeBSD$
39  */
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/fcntl.h>
44 #include <sys/file.h>
45 #include <sys/filedesc.h>
46 #include <sys/kernel.h>
47 #include <sys/lock.h>
48 #include <sys/malloc.h>
49 #include <sys/module.h>
50 #include <sys/mount.h>
51 #include <sys/mutex.h>
52 #include <sys/namei.h>
53 #include <sys/stat.h>
54 #include <sys/vnode.h>
55 
56 #include <vm/vm.h>
57 #include <vm/vm_extern.h>	/* for vnode_pager_setsize */
58 #include <vm/vm_zone.h>
59 #include <vm/vm_object.h>	/* for vm cache coherency */
60 
61 #include <fs/unionfs/union.h>
62 
63 #include <sys/proc.h>
64 
65 extern int	union_init __P((void));
66 
67 /* must be power of two, otherwise change UNION_HASH() */
68 #define NHASH 32
69 
70 /* unsigned int ... */
71 #define UNION_HASH(u, l) \
72 	(((((uintptr_t) (u)) + ((uintptr_t) l)) >> 8) & (NHASH-1))
73 
74 static LIST_HEAD(unhead, union_node) unhead[NHASH];
75 static int unvplock[NHASH];
76 
77 static void	union_dircache_r __P((struct vnode *vp, struct vnode ***vppp,
78 				      int *cntp));
79 static int	union_list_lock __P((int ix));
80 static void	union_list_unlock __P((int ix));
81 static int	union_relookup __P((struct union_mount *um, struct vnode *dvp,
82 				    struct vnode **vpp,
83 				    struct componentname *cnp,
84 				    struct componentname *cn, char *path,
85 				    int pathlen));
86 static void	union_updatevp __P((struct union_node *un,
87 				    struct vnode *uppervp,
88 				    struct vnode *lowervp));
89 static void union_newlower __P((struct union_node *, struct vnode *));
90 static void union_newupper __P((struct union_node *, struct vnode *));
91 static int union_copyfile __P((struct vnode *, struct vnode *,
92 					struct ucred *, struct thread *));
93 static int union_vn_create __P((struct vnode **, struct union_node *,
94 				struct thread *));
95 static int union_vn_close __P((struct vnode *, int, struct ucred *,
96 				struct thread *));
97 
98 int
99 union_init()
100 {
101 	int i;
102 
103 	for (i = 0; i < NHASH; i++)
104 		LIST_INIT(&unhead[i]);
105 	bzero((caddr_t)unvplock, sizeof(unvplock));
106 	return (0);
107 }
108 
109 static int
110 union_list_lock(ix)
111 	int ix;
112 {
113 	if (unvplock[ix] & UNVP_LOCKED) {
114 		unvplock[ix] |= UNVP_WANT;
115 		(void) tsleep((caddr_t) &unvplock[ix], PINOD, "unllck", 0);
116 		return (1);
117 	}
118 	unvplock[ix] |= UNVP_LOCKED;
119 	return (0);
120 }
121 
122 static void
123 union_list_unlock(ix)
124 	int ix;
125 {
126 	unvplock[ix] &= ~UNVP_LOCKED;
127 
128 	if (unvplock[ix] & UNVP_WANT) {
129 		unvplock[ix] &= ~UNVP_WANT;
130 		wakeup((caddr_t) &unvplock[ix]);
131 	}
132 }
133 
134 /*
135  *	union_updatevp:
136  *
137  *	The uppervp, if not NULL, must be referenced and not locked by us
138  *	The lowervp, if not NULL, must be referenced.
139  *
140  *	if uppervp and lowervp match pointers already installed, nothing
141  *	happens. The passed vp's (when matching) are not adjusted.  This
142  *	routine may only be called by union_newupper() and union_newlower().
143  */
144 
145 static void
146 union_updatevp(un, uppervp, lowervp)
147 	struct union_node *un;
148 	struct vnode *uppervp;
149 	struct vnode *lowervp;
150 {
151 	int ohash = UNION_HASH(un->un_uppervp, un->un_lowervp);
152 	int nhash = UNION_HASH(uppervp, lowervp);
153 	int docache = (lowervp != NULLVP || uppervp != NULLVP);
154 	int lhash, uhash;
155 
156 	/*
157 	 * Ensure locking is ordered from lower to higher
158 	 * to avoid deadlocks.
159 	 */
160 	if (nhash < ohash) {
161 		lhash = nhash;
162 		uhash = ohash;
163 	} else {
164 		lhash = ohash;
165 		uhash = nhash;
166 	}
167 
168 	if (lhash != uhash) {
169 		while (union_list_lock(lhash))
170 			continue;
171 	}
172 
173 	while (union_list_lock(uhash))
174 		continue;
175 
176 	if (ohash != nhash || !docache) {
177 		if (un->un_flags & UN_CACHED) {
178 			un->un_flags &= ~UN_CACHED;
179 			LIST_REMOVE(un, un_cache);
180 		}
181 	}
182 
183 	if (ohash != nhash)
184 		union_list_unlock(ohash);
185 
186 	if (un->un_lowervp != lowervp) {
187 		if (un->un_lowervp) {
188 			vrele(un->un_lowervp);
189 			if (un->un_path) {
190 				free(un->un_path, M_TEMP);
191 				un->un_path = 0;
192 			}
193 		}
194 		un->un_lowervp = lowervp;
195 		un->un_lowersz = VNOVAL;
196 	}
197 
198 	if (un->un_uppervp != uppervp) {
199 		if (un->un_uppervp)
200 			vrele(un->un_uppervp);
201 		un->un_uppervp = uppervp;
202 		un->un_uppersz = VNOVAL;
203 	}
204 
205 	if (docache && (ohash != nhash)) {
206 		LIST_INSERT_HEAD(&unhead[nhash], un, un_cache);
207 		un->un_flags |= UN_CACHED;
208 	}
209 
210 	union_list_unlock(nhash);
211 }
212 
213 /*
214  * Set a new lowervp.  The passed lowervp must be referenced and will be
215  * stored in the vp in a referenced state.
216  */
217 
218 static void
219 union_newlower(un, lowervp)
220 	struct union_node *un;
221 	struct vnode *lowervp;
222 {
223 	union_updatevp(un, un->un_uppervp, lowervp);
224 }
225 
226 /*
227  * Set a new uppervp.  The passed uppervp must be locked and will be
228  * stored in the vp in a locked state.  The caller should not unlock
229  * uppervp.
230  */
231 
232 static void
233 union_newupper(un, uppervp)
234 	struct union_node *un;
235 	struct vnode *uppervp;
236 {
237 	union_updatevp(un, uppervp, un->un_lowervp);
238 }
239 
240 /*
241  * Keep track of size changes in the underlying vnodes.
242  * If the size changes, then callback to the vm layer
243  * giving priority to the upper layer size.
244  */
245 void
246 union_newsize(vp, uppersz, lowersz)
247 	struct vnode *vp;
248 	off_t uppersz, lowersz;
249 {
250 	struct union_node *un;
251 	off_t sz;
252 
253 	/* only interested in regular files */
254 	if (vp->v_type != VREG)
255 		return;
256 
257 	un = VTOUNION(vp);
258 	sz = VNOVAL;
259 
260 	if ((uppersz != VNOVAL) && (un->un_uppersz != uppersz)) {
261 		un->un_uppersz = uppersz;
262 		if (sz == VNOVAL)
263 			sz = un->un_uppersz;
264 	}
265 
266 	if ((lowersz != VNOVAL) && (un->un_lowersz != lowersz)) {
267 		un->un_lowersz = lowersz;
268 		if (sz == VNOVAL)
269 			sz = un->un_lowersz;
270 	}
271 
272 	if (sz != VNOVAL) {
273 		UDEBUG(("union: %s size now %ld\n",
274 			(uppersz != VNOVAL ? "upper" : "lower"), (long)sz));
275 		/*
276 		 * There is no need to change size of non-existent object
277 		 */
278 		/* vnode_pager_setsize(vp, sz); */
279 	}
280 }
281 
282 /*
283  *	union_allocvp:	allocate a union_node and associate it with a
284  *			parent union_node and one or two vnodes.
285  *
286  *	vpp	Holds the returned vnode locked and referenced if no
287  *		error occurs.
288  *
289  *	mp	Holds the mount point.  mp may or may not be busied.
290  *		allocvp makes no changes to mp.
291  *
292  *	dvp	Holds the parent union_node to the one we wish to create.
293  *		XXX may only be used to traverse an uncopied lowervp-based
294  *		tree?  XXX
295  *
296  *		dvp may or may not be locked.  allocvp makes no changes
297  *		to dvp.
298  *
299  *	upperdvp Holds the parent vnode to uppervp, generally used along
300  *		with path component information to create a shadow of
301  *		lowervp when uppervp does not exist.
302  *
303  *		upperdvp is referenced but unlocked on entry, and will be
304  *		dereferenced on return.
305  *
306  *	uppervp	Holds the new uppervp vnode to be stored in the
307  *		union_node we are allocating.  uppervp is referenced but
308  *		not locked, and will be dereferenced on return.
309  *
310  *	lowervp	Holds the new lowervp vnode to be stored in the
311  *		union_node we are allocating.  lowervp is referenced but
312  *		not locked, and will be dereferenced on return.
313  *
314  *	cnp	Holds path component information to be coupled with
315  *		lowervp and upperdvp to allow unionfs to create an uppervp
316  *		later on.  Only used if lowervp is valid.  The conents
317  *		of cnp is only valid for the duration of the call.
318  *
319  *	docache	Determine whether this node should be entered in the
320  *		cache or whether it should be destroyed as soon as possible.
321  *
322  * all union_nodes are maintained on a singly-linked
323  * list.  new nodes are only allocated when they cannot
324  * be found on this list.  entries on the list are
325  * removed when the vfs reclaim entry is called.
326  *
327  * a single lock is kept for the entire list.  this is
328  * needed because the getnewvnode() function can block
329  * waiting for a vnode to become free, in which case there
330  * may be more than one process trying to get the same
331  * vnode.  this lock is only taken if we are going to
332  * call getnewvnode, since the kernel itself is single-threaded.
333  *
334  * if an entry is found on the list, then call vget() to
335  * take a reference.  this is done because there may be
336  * zero references to it and so it needs to removed from
337  * the vnode free list.
338  */
339 
340 int
341 union_allocvp(vpp, mp, dvp, upperdvp, cnp, uppervp, lowervp, docache)
342 	struct vnode **vpp;
343 	struct mount *mp;
344 	struct vnode *dvp;		/* parent union vnode */
345 	struct vnode *upperdvp;		/* parent vnode of uppervp */
346 	struct componentname *cnp;	/* may be null */
347 	struct vnode *uppervp;		/* may be null */
348 	struct vnode *lowervp;		/* may be null */
349 	int docache;
350 {
351 	int error;
352 	struct union_node *un = 0;
353 	struct union_mount *um = MOUNTTOUNIONMOUNT(mp);
354 	struct thread *td = (cnp) ? cnp->cn_thread : curthread;
355 	int hash = 0;
356 	int vflag;
357 	int try;
358 
359 	if (uppervp == NULLVP && lowervp == NULLVP)
360 		panic("union: unidentifiable allocation");
361 
362 	if (uppervp && lowervp && (uppervp->v_type != lowervp->v_type)) {
363 		vrele(lowervp);
364 		lowervp = NULLVP;
365 	}
366 
367 	/* detect the root vnode (and aliases) */
368 	vflag = 0;
369 	if ((uppervp == um->um_uppervp) &&
370 	    ((lowervp == NULLVP) || lowervp == um->um_lowervp)) {
371 		if (lowervp == NULLVP) {
372 			lowervp = um->um_lowervp;
373 			if (lowervp != NULLVP)
374 				VREF(lowervp);
375 		}
376 		vflag = VROOT;
377 	}
378 
379 loop:
380 	if (!docache) {
381 		un = 0;
382 	} else for (try = 0; try < 3; try++) {
383 		switch (try) {
384 		case 0:
385 			if (lowervp == NULLVP)
386 				continue;
387 			hash = UNION_HASH(uppervp, lowervp);
388 			break;
389 
390 		case 1:
391 			if (uppervp == NULLVP)
392 				continue;
393 			hash = UNION_HASH(uppervp, NULLVP);
394 			break;
395 
396 		case 2:
397 			if (lowervp == NULLVP)
398 				continue;
399 			hash = UNION_HASH(NULLVP, lowervp);
400 			break;
401 		}
402 
403 		while (union_list_lock(hash))
404 			continue;
405 
406 		LIST_FOREACH(un, &unhead[hash], un_cache) {
407 			if ((un->un_lowervp == lowervp ||
408 			     un->un_lowervp == NULLVP) &&
409 			    (un->un_uppervp == uppervp ||
410 			     un->un_uppervp == NULLVP) &&
411 			    (UNIONTOV(un)->v_mount == mp)) {
412 				if (vget(UNIONTOV(un), 0,
413 				    cnp ? cnp->cn_thread : NULL)) {
414 					union_list_unlock(hash);
415 					goto loop;
416 				}
417 				break;
418 			}
419 		}
420 
421 		union_list_unlock(hash);
422 
423 		if (un)
424 			break;
425 	}
426 
427 	if (un) {
428 		/*
429 		 * Obtain a lock on the union_node.  Everything is unlocked
430 		 * except for dvp, so check that case.  If they match, our
431 		 * new un is already locked.  Otherwise we have to lock our
432 		 * new un.
433 		 *
434 		 * A potential deadlock situation occurs when we are holding
435 		 * one lock while trying to get another.  We must follow
436 		 * strict ordering rules to avoid it.  We try to locate dvp
437 		 * by scanning up from un_vnode, since the most likely
438 		 * scenario is un being under dvp.
439 		 */
440 
441 		if (dvp && un->un_vnode != dvp) {
442 			struct vnode *scan = un->un_vnode;
443 
444 			do {
445 				scan = VTOUNION(scan)->un_pvp;
446 			} while (scan && scan->v_tag == VT_UNION && scan != dvp);
447 			if (scan != dvp) {
448 				/*
449 				 * our new un is above dvp (we never saw dvp
450 				 * while moving up the tree).
451 				 */
452 				VREF(dvp);
453 				VOP_UNLOCK(dvp, 0, td);
454 				error = vn_lock(un->un_vnode, LK_EXCLUSIVE, td);
455 				vn_lock(dvp, LK_EXCLUSIVE | LK_RETRY, td);
456 				vrele(dvp);
457 			} else {
458 				/*
459 				 * our new un is under dvp
460 				 */
461 				error = vn_lock(un->un_vnode, LK_EXCLUSIVE, td);
462 			}
463 		} else if (dvp == NULLVP) {
464 			/*
465 			 * dvp is NULL, we need to lock un.
466 			 */
467 			error = vn_lock(un->un_vnode, LK_EXCLUSIVE, td);
468 		} else {
469 			/*
470 			 * dvp == un->un_vnode, we are already locked.
471 			 */
472 			error = 0;
473 		}
474 
475 		if (error)
476 			goto loop;
477 
478 		/*
479 		 * At this point, the union_node is locked and referenced.
480 		 *
481 		 * uppervp is locked and referenced or NULL, lowervp is
482 		 * referenced or NULL.
483 		 */
484 		UDEBUG(("Modify existing un %p vn %p upper %p(refs %d) -> %p(refs %d)\n",
485 			un, un->un_vnode, un->un_uppervp,
486 			(un->un_uppervp ? un->un_uppervp->v_usecount : -99),
487 			uppervp,
488 			(uppervp ? uppervp->v_usecount : -99)
489 		));
490 
491 		if (uppervp != un->un_uppervp) {
492 			KASSERT(uppervp == NULL || uppervp->v_usecount > 0, ("union_allocvp: too few refs %d (at least 1 required) on uppervp", uppervp->v_usecount));
493 			union_newupper(un, uppervp);
494 		} else if (uppervp) {
495 			KASSERT(uppervp->v_usecount > 1, ("union_allocvp: too few refs %d (at least 2 required) on uppervp", uppervp->v_usecount));
496 			vrele(uppervp);
497 		}
498 
499 		/*
500 		 * Save information about the lower layer.
501 		 * This needs to keep track of pathname
502 		 * and directory information which union_vn_create
503 		 * might need.
504 		 */
505 		if (lowervp != un->un_lowervp) {
506 			union_newlower(un, lowervp);
507 			if (cnp && (lowervp != NULLVP)) {
508 				un->un_path = malloc(cnp->cn_namelen+1,
509 						M_TEMP, M_WAITOK);
510 				bcopy(cnp->cn_nameptr, un->un_path,
511 						cnp->cn_namelen);
512 				un->un_path[cnp->cn_namelen] = '\0';
513 			}
514 		} else if (lowervp) {
515 			vrele(lowervp);
516 		}
517 
518 		/*
519 		 * and upperdvp
520 		 */
521 		if (upperdvp != un->un_dirvp) {
522 			if (un->un_dirvp)
523 				vrele(un->un_dirvp);
524 			un->un_dirvp = upperdvp;
525 		} else if (upperdvp) {
526 			vrele(upperdvp);
527 		}
528 
529 		*vpp = UNIONTOV(un);
530 		return (0);
531 	}
532 
533 	if (docache) {
534 		/*
535 		 * otherwise lock the vp list while we call getnewvnode
536 		 * since that can block.
537 		 */
538 		hash = UNION_HASH(uppervp, lowervp);
539 
540 		if (union_list_lock(hash))
541 			goto loop;
542 	}
543 
544 	/*
545 	 * Create new node rather then replace old node
546 	 */
547 
548 	error = getnewvnode(VT_UNION, mp, union_vnodeop_p, vpp);
549 	if (error) {
550 		/*
551 		 * If an error occurs clear out vnodes.
552 		 */
553 		if (lowervp)
554 			vrele(lowervp);
555 		if (uppervp)
556 			vrele(uppervp);
557 		if (upperdvp)
558 			vrele(upperdvp);
559 		*vpp = NULL;
560 		goto out;
561 	}
562 
563 	MALLOC((*vpp)->v_data, void *, sizeof(struct union_node),
564 		M_TEMP, M_WAITOK);
565 
566 	(*vpp)->v_flag |= vflag;
567 	if (uppervp)
568 		(*vpp)->v_type = uppervp->v_type;
569 	else
570 		(*vpp)->v_type = lowervp->v_type;
571 
572 	un = VTOUNION(*vpp);
573 	bzero(un, sizeof(*un));
574 
575 	lockinit(&un->un_lock, PVFS, "unlock", 0, 0);
576 	vn_lock(*vpp, LK_EXCLUSIVE | LK_RETRY, td);
577 
578 	un->un_vnode = *vpp;
579 	un->un_uppervp = uppervp;
580 	un->un_uppersz = VNOVAL;
581 	un->un_lowervp = lowervp;
582 	un->un_lowersz = VNOVAL;
583 	un->un_dirvp = upperdvp;
584 	un->un_pvp = dvp;		/* only parent dir in new allocation */
585 	if (dvp != NULLVP)
586 		VREF(dvp);
587 	un->un_dircache = 0;
588 	un->un_openl = 0;
589 
590 	if (cnp && (lowervp != NULLVP)) {
591 		un->un_path = malloc(cnp->cn_namelen+1, M_TEMP, M_WAITOK);
592 		bcopy(cnp->cn_nameptr, un->un_path, cnp->cn_namelen);
593 		un->un_path[cnp->cn_namelen] = '\0';
594 	} else {
595 		un->un_path = 0;
596 		un->un_dirvp = NULL;
597 	}
598 
599 	if (docache) {
600 		LIST_INSERT_HEAD(&unhead[hash], un, un_cache);
601 		un->un_flags |= UN_CACHED;
602 	}
603 
604 out:
605 	if (docache)
606 		union_list_unlock(hash);
607 
608 	return (error);
609 }
610 
611 int
612 union_freevp(vp)
613 	struct vnode *vp;
614 {
615 	struct union_node *un = VTOUNION(vp);
616 
617 	if (un->un_flags & UN_CACHED) {
618 		un->un_flags &= ~UN_CACHED;
619 		LIST_REMOVE(un, un_cache);
620 	}
621 
622 	if (un->un_pvp != NULLVP) {
623 		vrele(un->un_pvp);
624 		un->un_pvp = NULL;
625 	}
626 	if (un->un_uppervp != NULLVP) {
627 		vrele(un->un_uppervp);
628 		un->un_uppervp = NULL;
629 	}
630 	if (un->un_lowervp != NULLVP) {
631 		vrele(un->un_lowervp);
632 		un->un_lowervp = NULL;
633 	}
634 	if (un->un_dirvp != NULLVP) {
635 		vrele(un->un_dirvp);
636 		un->un_dirvp = NULL;
637 	}
638 	if (un->un_path) {
639 		free(un->un_path, M_TEMP);
640 		un->un_path = NULL;
641 	}
642 	lockdestroy(&un->un_lock);
643 
644 	FREE(vp->v_data, M_TEMP);
645 	vp->v_data = 0;
646 
647 	return (0);
648 }
649 
650 /*
651  * copyfile.  copy the vnode (fvp) to the vnode (tvp)
652  * using a sequence of reads and writes.  both (fvp)
653  * and (tvp) are locked on entry and exit.
654  *
655  * fvp and tvp are both exclusive locked on call, but their refcount's
656  * haven't been bumped at all.
657  */
658 static int
659 union_copyfile(fvp, tvp, cred, td)
660 	struct vnode *fvp;
661 	struct vnode *tvp;
662 	struct ucred *cred;
663 	struct thread *td;
664 {
665 	char *buf;
666 	struct uio uio;
667 	struct iovec iov;
668 	int error = 0;
669 
670 	/*
671 	 * strategy:
672 	 * allocate a buffer of size MAXBSIZE.
673 	 * loop doing reads and writes, keeping track
674 	 * of the current uio offset.
675 	 * give up at the first sign of trouble.
676 	 */
677 
678 	bzero(&uio, sizeof(uio));
679 
680 	uio.uio_td = td;
681 	uio.uio_segflg = UIO_SYSSPACE;
682 	uio.uio_offset = 0;
683 
684 	VOP_LEASE(fvp, td, cred, LEASE_READ);
685 	VOP_LEASE(tvp, td, cred, LEASE_WRITE);
686 
687 	buf = malloc(MAXBSIZE, M_TEMP, M_WAITOK);
688 
689 	/* ugly loop follows... */
690 	do {
691 		off_t offset = uio.uio_offset;
692 		int count;
693 		int bufoffset;
694 
695 		/*
696 		 * Setup for big read
697 		 */
698 		uio.uio_iov = &iov;
699 		uio.uio_iovcnt = 1;
700 		iov.iov_base = buf;
701 		iov.iov_len = MAXBSIZE;
702 		uio.uio_resid = iov.iov_len;
703 		uio.uio_rw = UIO_READ;
704 
705 		if ((error = VOP_READ(fvp, &uio, 0, cred)) != 0)
706 			break;
707 
708 		/*
709 		 * Get bytes read, handle read eof case and setup for
710 		 * write loop
711 		 */
712 		if ((count = MAXBSIZE - uio.uio_resid) == 0)
713 			break;
714 		bufoffset = 0;
715 
716 		/*
717 		 * Write until an error occurs or our buffer has been
718 		 * exhausted, then update the offset for the next read.
719 		 */
720 		while (bufoffset < count) {
721 			uio.uio_iov = &iov;
722 			uio.uio_iovcnt = 1;
723 			iov.iov_base = buf + bufoffset;
724 			iov.iov_len = count - bufoffset;
725 			uio.uio_offset = offset + bufoffset;
726 			uio.uio_rw = UIO_WRITE;
727 			uio.uio_resid = iov.iov_len;
728 
729 			if ((error = VOP_WRITE(tvp, &uio, 0, cred)) != 0)
730 				break;
731 			bufoffset += (count - bufoffset) - uio.uio_resid;
732 		}
733 		uio.uio_offset = offset + bufoffset;
734 	} while (error == 0);
735 
736 	free(buf, M_TEMP);
737 	return (error);
738 }
739 
740 /*
741  *
742  * un's vnode is assumed to be locked on entry and remains locked on exit.
743  */
744 
745 int
746 union_copyup(un, docopy, cred, td)
747 	struct union_node *un;
748 	int docopy;
749 	struct ucred *cred;
750 	struct thread *td;
751 {
752 	int error;
753 	struct mount *mp;
754 	struct vnode *lvp, *uvp;
755 
756 	/*
757 	 * If the user does not have read permission, the vnode should not
758 	 * be copied to upper layer.
759 	 */
760 	vn_lock(un->un_lowervp, LK_EXCLUSIVE | LK_RETRY, td);
761 	error = VOP_ACCESS(un->un_lowervp, VREAD, cred, td);
762 	VOP_UNLOCK(un->un_lowervp, 0, td);
763 	if (error)
764 		return (error);
765 
766 	if ((error = vn_start_write(un->un_dirvp, &mp, V_WAIT | PCATCH)) != 0)
767 		return (error);
768 	if ((error = union_vn_create(&uvp, un, td)) != 0) {
769 		vn_finished_write(mp);
770 		return (error);
771 	}
772 
773 	lvp = un->un_lowervp;
774 
775 	KASSERT(uvp->v_usecount > 0, ("copy: uvp refcount 0: %d", uvp->v_usecount));
776 	if (docopy) {
777 		/*
778 		 * XX - should not ignore errors
779 		 * from VOP_CLOSE
780 		 */
781 		vn_lock(lvp, LK_EXCLUSIVE | LK_RETRY, td);
782 		error = VOP_OPEN(lvp, FREAD, cred, td);
783 		if (error == 0 && vn_canvmio(lvp) == TRUE)
784 			error = vfs_object_create(lvp, td, cred);
785 		if (error == 0) {
786 			error = union_copyfile(lvp, uvp, cred, td);
787 			VOP_UNLOCK(lvp, 0, td);
788 			(void) VOP_CLOSE(lvp, FREAD, cred, td);
789 		}
790 		if (error == 0)
791 			UDEBUG(("union: copied up %s\n", un->un_path));
792 
793 	}
794 	VOP_UNLOCK(uvp, 0, td);
795 	vn_finished_write(mp);
796 	union_newupper(un, uvp);
797 	KASSERT(uvp->v_usecount > 0, ("copy: uvp refcount 0: %d", uvp->v_usecount));
798 	union_vn_close(uvp, FWRITE, cred, td);
799 	KASSERT(uvp->v_usecount > 0, ("copy: uvp refcount 0: %d", uvp->v_usecount));
800 	/*
801 	 * Subsequent IOs will go to the top layer, so
802 	 * call close on the lower vnode and open on the
803 	 * upper vnode to ensure that the filesystem keeps
804 	 * its references counts right.  This doesn't do
805 	 * the right thing with (cred) and (FREAD) though.
806 	 * Ignoring error returns is not right, either.
807 	 */
808 	if (error == 0) {
809 		int i;
810 
811 		for (i = 0; i < un->un_openl; i++) {
812 			(void) VOP_CLOSE(lvp, FREAD, cred, td);
813 			(void) VOP_OPEN(uvp, FREAD, cred, td);
814 		}
815 		if (un->un_openl) {
816 			if (vn_canvmio(uvp) == TRUE)
817 				error = vfs_object_create(uvp, td, cred);
818 		}
819 		un->un_openl = 0;
820 	}
821 
822 	return (error);
823 
824 }
825 
826 /*
827  *	union_relookup:
828  *
829  *	dvp should be locked on entry and will be locked on return.  No
830  *	net change in the ref count will occur.
831  *
832  *	If an error is returned, *vpp will be invalid, otherwise it
833  *	will hold a locked, referenced vnode.  If *vpp == dvp then
834  *	remember that only one exclusive lock is held.
835  */
836 
837 static int
838 union_relookup(um, dvp, vpp, cnp, cn, path, pathlen)
839 	struct union_mount *um;
840 	struct vnode *dvp;
841 	struct vnode **vpp;
842 	struct componentname *cnp;
843 	struct componentname *cn;
844 	char *path;
845 	int pathlen;
846 {
847 	int error;
848 
849 	/*
850 	 * A new componentname structure must be faked up because
851 	 * there is no way to know where the upper level cnp came
852 	 * from or what it is being used for.  This must duplicate
853 	 * some of the work done by NDINIT, some of the work done
854 	 * by namei, some of the work done by lookup and some of
855 	 * the work done by VOP_LOOKUP when given a CREATE flag.
856 	 * Conclusion: Horrible.
857 	 */
858 	cn->cn_namelen = pathlen;
859 	cn->cn_pnbuf = zalloc(namei_zone);
860 	bcopy(path, cn->cn_pnbuf, cn->cn_namelen);
861 	cn->cn_pnbuf[cn->cn_namelen] = '\0';
862 
863 	cn->cn_nameiop = CREATE;
864 	cn->cn_flags = (LOCKPARENT|LOCKLEAF|HASBUF|SAVENAME|ISLASTCN);
865 	cn->cn_thread = cnp->cn_thread;
866 	if (um->um_op == UNMNT_ABOVE)
867 		cn->cn_cred = cnp->cn_cred;
868 	else
869 		cn->cn_cred = um->um_cred;
870 	cn->cn_nameptr = cn->cn_pnbuf;
871 	cn->cn_consume = cnp->cn_consume;
872 
873 	VREF(dvp);
874 	VOP_UNLOCK(dvp, 0, cnp->cn_thread);
875 
876 	/*
877 	 * Pass dvp unlocked and referenced on call to relookup().
878 	 *
879 	 * If an error occurs, dvp will be returned unlocked and dereferenced.
880 	 */
881 
882 	if ((error = relookup(dvp, vpp, cn)) != 0) {
883 		vn_lock(dvp, LK_EXCLUSIVE | LK_RETRY, cnp->cn_thread);
884 		return(error);
885 	}
886 
887 	/*
888 	 * If no error occurs, dvp will be returned locked with the reference
889 	 * left as before, and vpp will be returned referenced and locked.
890 	 *
891 	 * We want to return with dvp as it was passed to us, so we get
892 	 * rid of our reference.
893 	 */
894 	vrele(dvp);
895 	return (0);
896 }
897 
898 /*
899  * Create a shadow directory in the upper layer.
900  * The new vnode is returned locked.
901  *
902  * (um) points to the union mount structure for access to the
903  * the mounting process's credentials.
904  * (dvp) is the directory in which to create the shadow directory,
905  * it is locked (but not ref'd) on entry and return.
906  * (cnp) is the componentname to be created.
907  * (vpp) is the returned newly created shadow directory, which
908  * is returned locked and ref'd
909  */
910 int
911 union_mkshadow(um, dvp, cnp, vpp)
912 	struct union_mount *um;
913 	struct vnode *dvp;
914 	struct componentname *cnp;
915 	struct vnode **vpp;
916 {
917 	int error;
918 	struct vattr va;
919 	struct thread *td = cnp->cn_thread;
920 	struct componentname cn;
921 	struct mount *mp;
922 
923 	if ((error = vn_start_write(dvp, &mp, V_WAIT | PCATCH)) != 0)
924 		return (error);
925 	if ((error = union_relookup(um, dvp, vpp, cnp, &cn,
926 			cnp->cn_nameptr, cnp->cn_namelen)) != 0) {
927 		vn_finished_write(mp);
928 		return (error);
929 	}
930 
931 	if (*vpp) {
932 		if (cn.cn_flags & HASBUF) {
933 			zfree(namei_zone, cn.cn_pnbuf);
934 			cn.cn_flags &= ~HASBUF;
935 		}
936 		if (dvp == *vpp)
937 			vrele(*vpp);
938 		else
939 			vput(*vpp);
940 		vn_finished_write(mp);
941 		*vpp = NULLVP;
942 		return (EEXIST);
943 	}
944 
945 	/*
946 	 * policy: when creating the shadow directory in the
947 	 * upper layer, create it owned by the user who did
948 	 * the mount, group from parent directory, and mode
949 	 * 777 modified by umask (ie mostly identical to the
950 	 * mkdir syscall).  (jsp, kb)
951 	 */
952 
953 	VATTR_NULL(&va);
954 	va.va_type = VDIR;
955 	va.va_mode = um->um_cmode;
956 
957 	/* VOP_LEASE: dvp is locked */
958 	VOP_LEASE(dvp, td, cn.cn_cred, LEASE_WRITE);
959 
960 	error = VOP_MKDIR(dvp, vpp, &cn, &va);
961 	if (cn.cn_flags & HASBUF) {
962 		zfree(namei_zone, cn.cn_pnbuf);
963 		cn.cn_flags &= ~HASBUF;
964 	}
965 	/*vput(dvp);*/
966 	vn_finished_write(mp);
967 	return (error);
968 }
969 
970 /*
971  * Create a whiteout entry in the upper layer.
972  *
973  * (um) points to the union mount structure for access to the
974  * the mounting process's credentials.
975  * (dvp) is the directory in which to create the whiteout.
976  * it is locked on entry and return.
977  * (cnp) is the componentname to be created.
978  */
979 int
980 union_mkwhiteout(um, dvp, cnp, path)
981 	struct union_mount *um;
982 	struct vnode *dvp;
983 	struct componentname *cnp;
984 	char *path;
985 {
986 	int error;
987 	struct thread *td = cnp->cn_thread;
988 	struct vnode *wvp;
989 	struct componentname cn;
990 	struct mount *mp;
991 
992 	if ((error = vn_start_write(dvp, &mp, V_WAIT | PCATCH)) != 0)
993 		return (error);
994 	error = union_relookup(um, dvp, &wvp, cnp, &cn, path, strlen(path));
995 	if (error) {
996 		vn_finished_write(mp);
997 		return (error);
998 	}
999 
1000 	if (wvp) {
1001 		if (cn.cn_flags & HASBUF) {
1002 			zfree(namei_zone, cn.cn_pnbuf);
1003 			cn.cn_flags &= ~HASBUF;
1004 		}
1005 		if (wvp == dvp)
1006 			vrele(wvp);
1007 		else
1008 			vput(wvp);
1009 		vn_finished_write(mp);
1010 		return (EEXIST);
1011 	}
1012 
1013 	/* VOP_LEASE: dvp is locked */
1014 	VOP_LEASE(dvp, td, td->td_proc->p_ucred, LEASE_WRITE);
1015 
1016 	error = VOP_WHITEOUT(dvp, &cn, CREATE);
1017 	if (cn.cn_flags & HASBUF) {
1018 		zfree(namei_zone, cn.cn_pnbuf);
1019 		cn.cn_flags &= ~HASBUF;
1020 	}
1021 	vn_finished_write(mp);
1022 	return (error);
1023 }
1024 
1025 /*
1026  * union_vn_create: creates and opens a new shadow file
1027  * on the upper union layer.  this function is similar
1028  * in spirit to calling vn_open but it avoids calling namei().
1029  * the problem with calling namei is that a) it locks too many
1030  * things, and b) it doesn't start at the "right" directory,
1031  * whereas relookup is told where to start.
1032  *
1033  * On entry, the vnode associated with un is locked.  It remains locked
1034  * on return.
1035  *
1036  * If no error occurs, *vpp contains a locked referenced vnode for your
1037  * use.  If an error occurs *vpp iis undefined.
1038  */
1039 static int
1040 union_vn_create(vpp, un, td)
1041 	struct vnode **vpp;
1042 	struct union_node *un;
1043 	struct thread *td;
1044 {
1045 	struct vnode *vp;
1046 	struct ucred *cred = td->td_proc->p_ucred;
1047 	struct vattr vat;
1048 	struct vattr *vap = &vat;
1049 	int fmode = FFLAGS(O_WRONLY|O_CREAT|O_TRUNC|O_EXCL);
1050 	int error;
1051 	int cmode = UN_FILEMODE & ~td->td_proc->p_fd->fd_cmask;
1052 	struct componentname cn;
1053 
1054 	*vpp = NULLVP;
1055 
1056 	/*
1057 	 * Build a new componentname structure (for the same
1058 	 * reasons outlines in union_mkshadow).
1059 	 * The difference here is that the file is owned by
1060 	 * the current user, rather than by the person who
1061 	 * did the mount, since the current user needs to be
1062 	 * able to write the file (that's why it is being
1063 	 * copied in the first place).
1064 	 */
1065 	cn.cn_namelen = strlen(un->un_path);
1066 	cn.cn_pnbuf = zalloc(namei_zone);
1067 	bcopy(un->un_path, cn.cn_pnbuf, cn.cn_namelen+1);
1068 	cn.cn_nameiop = CREATE;
1069 	cn.cn_flags = (LOCKPARENT|LOCKLEAF|HASBUF|SAVENAME|ISLASTCN);
1070 	cn.cn_thread = td;
1071 	cn.cn_cred = td->td_proc->p_ucred;
1072 	cn.cn_nameptr = cn.cn_pnbuf;
1073 	cn.cn_consume = 0;
1074 
1075 	/*
1076 	 * Pass dvp unlocked and referenced on call to relookup().
1077 	 *
1078 	 * If an error occurs, dvp will be returned unlocked and dereferenced.
1079 	 */
1080 	VREF(un->un_dirvp);
1081 	error = relookup(un->un_dirvp, &vp, &cn);
1082 	if (error)
1083 		return (error);
1084 
1085 	/*
1086 	 * If no error occurs, dvp will be returned locked with the reference
1087 	 * left as before, and vpp will be returned referenced and locked.
1088 	 */
1089 	if (vp) {
1090 		vput(un->un_dirvp);
1091 		if (cn.cn_flags & HASBUF) {
1092 			zfree(namei_zone, cn.cn_pnbuf);
1093 			cn.cn_flags &= ~HASBUF;
1094 		}
1095 		if (vp == un->un_dirvp)
1096 			vrele(vp);
1097 		else
1098 			vput(vp);
1099 		return (EEXIST);
1100 	}
1101 
1102 	/*
1103 	 * Good - there was no race to create the file
1104 	 * so go ahead and create it.  The permissions
1105 	 * on the file will be 0666 modified by the
1106 	 * current user's umask.  Access to the file, while
1107 	 * it is unioned, will require access to the top *and*
1108 	 * bottom files.  Access when not unioned will simply
1109 	 * require access to the top-level file.
1110 	 * TODO: confirm choice of access permissions.
1111 	 */
1112 	VATTR_NULL(vap);
1113 	vap->va_type = VREG;
1114 	vap->va_mode = cmode;
1115 	VOP_LEASE(un->un_dirvp, td, cred, LEASE_WRITE);
1116 	error = VOP_CREATE(un->un_dirvp, &vp, &cn, vap);
1117 	if (cn.cn_flags & HASBUF) {
1118 		zfree(namei_zone, cn.cn_pnbuf);
1119 		cn.cn_flags &= ~HASBUF;
1120 	}
1121 	vput(un->un_dirvp);
1122 	if (error)
1123 		return (error);
1124 
1125 	error = VOP_OPEN(vp, fmode, cred, td);
1126 	if (error == 0 && vn_canvmio(vp) == TRUE)
1127 		error = vfs_object_create(vp, td, cred);
1128 	if (error) {
1129 		vput(vp);
1130 		return (error);
1131 	}
1132 	vp->v_writecount++;
1133 	*vpp = vp;
1134 	return (0);
1135 }
1136 
1137 static int
1138 union_vn_close(vp, fmode, cred, td)
1139 	struct vnode *vp;
1140 	int fmode;
1141 	struct ucred *cred;
1142 	struct thread *td;
1143 {
1144 
1145 	if (fmode & FWRITE)
1146 		--vp->v_writecount;
1147 	return (VOP_CLOSE(vp, fmode, cred, td));
1148 }
1149 
1150 #if 0
1151 
1152 /*
1153  *	union_removed_upper:
1154  *
1155  *	called with union_node unlocked. XXX
1156  */
1157 
1158 void
1159 union_removed_upper(un)
1160 	struct union_node *un;
1161 {
1162 	struct thread *td = curthread;	/* XXX */
1163 	struct vnode **vpp;
1164 
1165 	/*
1166 	 * Do not set the uppervp to NULLVP.  If lowervp is NULLVP,
1167 	 * union node will have neither uppervp nor lowervp.  We remove
1168 	 * the union node from cache, so that it will not be referrenced.
1169 	 */
1170 	union_newupper(un, NULLVP);
1171 	if (un->un_dircache != 0) {
1172 		for (vpp = un->un_dircache; *vpp != NULLVP; vpp++)
1173 			vrele(*vpp);
1174 		free(un->un_dircache, M_TEMP);
1175 		un->un_dircache = 0;
1176 	}
1177 
1178 	if (un->un_flags & UN_CACHED) {
1179 		un->un_flags &= ~UN_CACHED;
1180 		LIST_REMOVE(un, un_cache);
1181 	}
1182 }
1183 
1184 #endif
1185 
1186 /*
1187  * determine whether a whiteout is needed
1188  * during a remove/rmdir operation.
1189  */
1190 int
1191 union_dowhiteout(un, cred, td)
1192 	struct union_node *un;
1193 	struct ucred *cred;
1194 	struct thread *td;
1195 {
1196 	struct vattr va;
1197 
1198 	if (un->un_lowervp != NULLVP)
1199 		return (1);
1200 
1201 	if (VOP_GETATTR(un->un_uppervp, &va, cred, td) == 0 &&
1202 	    (va.va_flags & OPAQUE))
1203 		return (1);
1204 
1205 	return (0);
1206 }
1207 
1208 static void
1209 union_dircache_r(vp, vppp, cntp)
1210 	struct vnode *vp;
1211 	struct vnode ***vppp;
1212 	int *cntp;
1213 {
1214 	struct union_node *un;
1215 
1216 	if (vp->v_op != union_vnodeop_p) {
1217 		if (vppp) {
1218 			VREF(vp);
1219 			*(*vppp)++ = vp;
1220 			if (--(*cntp) == 0)
1221 				panic("union: dircache table too small");
1222 		} else {
1223 			(*cntp)++;
1224 		}
1225 
1226 		return;
1227 	}
1228 
1229 	un = VTOUNION(vp);
1230 	if (un->un_uppervp != NULLVP)
1231 		union_dircache_r(un->un_uppervp, vppp, cntp);
1232 	if (un->un_lowervp != NULLVP)
1233 		union_dircache_r(un->un_lowervp, vppp, cntp);
1234 }
1235 
1236 struct vnode *
1237 union_dircache(vp, td)
1238 	struct vnode *vp;
1239 	struct thread *td;
1240 {
1241 	int cnt;
1242 	struct vnode *nvp;
1243 	struct vnode **vpp;
1244 	struct vnode **dircache;
1245 	struct union_node *un;
1246 	int error;
1247 
1248 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
1249 	dircache = VTOUNION(vp)->un_dircache;
1250 
1251 	nvp = NULLVP;
1252 
1253 	if (dircache == NULL) {
1254 		cnt = 0;
1255 		union_dircache_r(vp, 0, &cnt);
1256 		cnt++;
1257 		dircache = malloc(cnt * sizeof(struct vnode *),
1258 				M_TEMP, M_WAITOK);
1259 		vpp = dircache;
1260 		union_dircache_r(vp, &vpp, &cnt);
1261 		*vpp = NULLVP;
1262 		vpp = dircache + 1;
1263 	} else {
1264 		vpp = dircache;
1265 		do {
1266 			if (*vpp++ == VTOUNION(vp)->un_uppervp)
1267 				break;
1268 		} while (*vpp != NULLVP);
1269 	}
1270 
1271 	if (*vpp == NULLVP)
1272 		goto out;
1273 
1274 	/*vn_lock(*vpp, LK_EXCLUSIVE | LK_RETRY, td);*/
1275 	UDEBUG(("ALLOCVP-3 %p ref %d\n", *vpp, (*vpp ? (*vpp)->v_usecount : -99)));
1276 	VREF(*vpp);
1277 	error = union_allocvp(&nvp, vp->v_mount, NULLVP, NULLVP, NULL, *vpp, NULLVP, 0);
1278 	UDEBUG(("ALLOCVP-3B %p ref %d\n", nvp, (*vpp ? (*vpp)->v_usecount : -99)));
1279 	if (error)
1280 		goto out;
1281 
1282 	VTOUNION(vp)->un_dircache = 0;
1283 	un = VTOUNION(nvp);
1284 	un->un_dircache = dircache;
1285 
1286 out:
1287 	VOP_UNLOCK(vp, 0, td);
1288 	return (nvp);
1289 }
1290 
1291 /*
1292  * Module glue to remove #ifdef UNION from vfs_syscalls.c
1293  */
1294 static int
1295 union_dircheck(struct thread *td, struct vnode **vp, struct file *fp)
1296 {
1297 	int error = 0;
1298 
1299 	if ((*vp)->v_op == union_vnodeop_p) {
1300 		struct vnode *lvp;
1301 
1302 		lvp = union_dircache(*vp, td);
1303 		if (lvp != NULLVP) {
1304 			struct vattr va;
1305 
1306 			/*
1307 			 * If the directory is opaque,
1308 			 * then don't show lower entries
1309 			 */
1310 			error = VOP_GETATTR(*vp, &va, fp->f_cred, td);
1311 			if (va.va_flags & OPAQUE) {
1312 				vput(lvp);
1313 				lvp = NULL;
1314 			}
1315 		}
1316 
1317 		if (lvp != NULLVP) {
1318 			error = VOP_OPEN(lvp, FREAD, fp->f_cred, td);
1319 			if (error == 0 && vn_canvmio(lvp) == TRUE)
1320 				error = vfs_object_create(lvp, td, fp->f_cred);
1321 			if (error) {
1322 				vput(lvp);
1323 				return (error);
1324 			}
1325 			VOP_UNLOCK(lvp, 0, td);
1326 			fp->f_data = (caddr_t) lvp;
1327 			fp->f_offset = 0;
1328 			error = vn_close(*vp, FREAD, fp->f_cred, td);
1329 			if (error)
1330 				return (error);
1331 			*vp = lvp;
1332 			return -1;	/* goto unionread */
1333 		}
1334 	}
1335 	return error;
1336 }
1337 
1338 static int
1339 union_modevent(module_t mod, int type, void *data)
1340 {
1341 	switch (type) {
1342 	case MOD_LOAD:
1343 		union_dircheckp = union_dircheck;
1344 		break;
1345 	case MOD_UNLOAD:
1346 		union_dircheckp = NULL;
1347 		break;
1348 	default:
1349 		break;
1350 	}
1351 	return 0;
1352 }
1353 
1354 static moduledata_t union_mod = {
1355 	"union_dircheck",
1356 	union_modevent,
1357 	NULL
1358 };
1359 
1360 DECLARE_MODULE(union_dircheck, union_mod, SI_SUB_VFS, SI_ORDER_ANY);
1361