xref: /dragonfly/sys/kern/kern_lockf.c (revision 92c825d6)
1 /*
2  * Copyright (c) 2004 Joerg Sonnenberger <joerg@bec.de>.  All rights reserved.
3  * Copyright (c) 2006 Matthew Dillon <dillon@backplane.com>.  All rights reserved.
4  *
5  * Copyright (c) 1982, 1986, 1989, 1993
6  *	The Regents of the University of California.  All rights reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * Scooter Morris at Genentech Inc.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  *	@(#)ufs_lockf.c	8.3 (Berkeley) 1/6/94
36  * $FreeBSD: src/sys/kern/kern_lockf.c,v 1.25 1999/11/16 16:28:56 phk Exp $
37  */
38 
39 #include "opt_debug_lockf.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/lock.h>
45 #include <sys/proc.h>
46 #include <sys/unistd.h>
47 #include <sys/vnode.h>
48 #include <sys/malloc.h>
49 #include <sys/fcntl.h>
50 #include <sys/resourcevar.h>
51 
52 #include <sys/lockf.h>
53 #include <machine/limits.h>	/* for LLONG_MAX */
54 #include <machine/stdarg.h>
55 
56 #include <sys/spinlock2.h>
57 
58 #ifdef INVARIANTS
59 int lf_global_counter = 0;
60 #endif
61 
62 #ifdef LOCKF_DEBUG
63 int lf_print_ranges = 0;
64 
65 static void _lf_print_lock(const struct lockf *);
66 static void _lf_printf(const char *, ...) __printflike(1, 2);
67 
68 #define lf_print_lock(lock) if (lf_print_ranges) _lf_print_lock(lock)
69 #define lf_printf(ctl, args...)	if (lf_print_ranges) _lf_printf(ctl, args)
70 #else
71 #define lf_print_lock(lock)
72 #define lf_printf(ctl, args...)
73 #endif
74 
75 static MALLOC_DEFINE(M_LOCKF, "lockf", "Byte-range locking structures");
76 
77 static void	lf_wakeup(struct lockf *, off_t, off_t);
78 static struct lockf_range *lf_alloc_range(void);
79 static void	lf_create_range(struct lockf_range *, struct proc *, int, int,
80 				off_t, off_t);
81 static void	lf_insert(struct lockf_range_list *list,
82 				struct lockf_range *elm,
83 				struct lockf_range *insert_point);
84 static void	lf_destroy_range(struct lockf_range *);
85 
86 static int	lf_setlock(struct lockf *, struct proc *, int, int,
87 			   off_t, off_t);
88 static int	lf_getlock(struct flock *, struct lockf *, struct proc *,
89 			   int, int, off_t, off_t);
90 
91 static int	lf_count_change(struct proc *, int);
92 
93 /*
94  * Return TRUE (non-zero) if the type and posix flags match.
95  */
96 static __inline
97 int
98 lf_match(struct lockf_range *range, int type, int flags)
99 {
100 	if (range->lf_type != type)
101 		return(0);
102 	if ((range->lf_flags ^ flags) & F_POSIX)
103 		return(0);
104 	return(1);
105 }
106 
107 /*
108  * Check whether range and [start, end] overlap.
109  */
110 static __inline
111 int
112 lf_overlap(const struct lockf_range *range, off_t start, off_t end)
113 {
114 	if (range->lf_start >= start && range->lf_start <= end)
115 		return(1);
116 	else if (start >= range->lf_start && start <= range->lf_end)
117 		return(1);
118 	else
119 		return(0);
120 }
121 
122 
123 /*
124  * Change the POSIX lock accounting for the given process.
125  */
126 void
127 lf_count_adjust(struct proc *p, int increase)
128 {
129 	struct uidinfo *uip;
130 
131 	KKASSERT(p != NULL);
132 
133 	uip = p->p_ucred->cr_uidinfo;
134 	if (increase)
135 		atomic_add_int(&uip->ui_posixlocks, p->p_numposixlocks);
136 	else
137 		atomic_add_int(&uip->ui_posixlocks, -p->p_numposixlocks);
138 
139 	KASSERT(uip->ui_posixlocks >= 0,
140 		("Negative number of POSIX locks held by %s user: %d.",
141 		 increase ? "new" : "old", uip->ui_posixlocks));
142 }
143 
144 static int
145 lf_count_change(struct proc *owner, int diff)
146 {
147 	struct uidinfo *uip;
148 	int max, ret;
149 
150 	/* we might actually not have a process context */
151 	if (owner == NULL)
152 		return(0);
153 
154 	uip = owner->p_ucred->cr_uidinfo;
155 
156 	max = MIN(owner->p_rlimit[RLIMIT_POSIXLOCKS].rlim_cur,
157 		  maxposixlocksperuid);
158 
159 	if (diff > 0 && owner->p_ucred->cr_uid != 0 && max != -1 &&
160 	    uip->ui_posixlocks >= max ) {
161 		ret = 1;
162 	} else {
163 		atomic_add_int(&uip->ui_posixlocks, diff);
164 		atomic_add_int(&owner->p_numposixlocks, diff);
165 		KASSERT(uip->ui_posixlocks >= 0,
166 			("Negative number of POSIX locks held by user: %d.",
167 			 uip->ui_posixlocks));
168 		KASSERT(owner->p_numposixlocks >= 0,
169 			("Negative number of POSIX locks held by proc: %d.",
170 			 uip->ui_posixlocks));
171 		ret = 0;
172 	}
173 	return ret;
174 }
175 
176 /*
177  * Advisory record locking support
178  */
179 int
180 lf_advlock(struct vop_advlock_args *ap, struct lockf *lock, u_quad_t size)
181 {
182 	struct flock *fl = ap->a_fl;
183 	struct proc *owner;
184 	off_t start, end;
185 	int type, flags, error;
186 	lwkt_token_t token;
187 
188 	/*
189 	 * Convert the flock structure into a start and end.
190 	 */
191 	switch (fl->l_whence) {
192 	case SEEK_SET:
193 	case SEEK_CUR:
194 		/*
195 		 * Caller is responsible for adding any necessary offset
196 		 * when SEEK_CUR is used.
197 		 */
198 		start = fl->l_start;
199 		break;
200 
201 	case SEEK_END:
202 		start = size + fl->l_start;
203 		break;
204 
205 	default:
206 		return(EINVAL);
207 	}
208 
209 	flags = ap->a_flags;
210 	if (start < 0)
211 		return(EINVAL);
212 	if (fl->l_len == 0) {
213 		flags |= F_NOEND;
214 		end = LLONG_MAX;
215 	} else if (fl->l_len < 0) {
216 		return(EINVAL);
217 	} else {
218 		end = start + fl->l_len - 1;
219 		if (end < start)
220 			return(EINVAL);
221 	}
222 
223 	type = fl->l_type;
224 	/*
225 	 * This isn't really correct for flock-style locks,
226 	 * but the current handling is somewhat broken anyway.
227 	 */
228 	owner = (struct proc *)ap->a_id;
229 
230 	/*
231 	 * Do the requested operation.
232 	 */
233 	token = lwkt_getpooltoken(lock);
234 
235 	if (lock->init_done == 0) {
236 		TAILQ_INIT(&lock->lf_range);
237 		TAILQ_INIT(&lock->lf_blocked);
238 		lock->init_done = 1;
239 	}
240 
241 	switch(ap->a_op) {
242 	case F_SETLK:
243 		/*
244 		 * NOTE: It is possible for both lf_range and lf_blocked to
245 		 * be empty if we block and get woken up, but another process
246 		 * then gets in and issues an unlock.  So VMAYHAVELOCKS must
247 		 * be set after the lf_setlock() operation completes rather
248 		 * then before.
249 		 */
250 		error = lf_setlock(lock, owner, type, flags, start, end);
251 		vsetflags(ap->a_vp, VMAYHAVELOCKS);
252 		break;
253 
254 	case F_UNLCK:
255 		error = lf_setlock(lock, owner, type, flags, start, end);
256 		if (TAILQ_EMPTY(&lock->lf_range) &&
257 		    TAILQ_EMPTY(&lock->lf_blocked)) {
258 			vclrflags(ap->a_vp, VMAYHAVELOCKS);
259 		}
260 		break;
261 
262 	case F_GETLK:
263 		error = lf_getlock(fl, lock, owner, type, flags, start, end);
264 		break;
265 
266 	default:
267 		error = EINVAL;
268 		break;
269 	}
270 	lwkt_reltoken(token);
271 	return(error);
272 }
273 
274 static int
275 lf_setlock(struct lockf *lock, struct proc *owner, int type, int flags,
276 	   off_t start, off_t end)
277 {
278 	struct lockf_range *range;
279 	struct lockf_range *brange;
280 	struct lockf_range *next;
281 	struct lockf_range *first_match;
282 	struct lockf_range *last_match;
283 	struct lockf_range *insert_point;
284 	struct lockf_range *new_range1;
285 	struct lockf_range *new_range2;
286 	int wakeup_needed;
287 	int double_clip;
288 	int unlock_override;
289 	int error = 0;
290 	int count;
291 	struct lockf_range_list deadlist;
292 
293 	new_range1 = NULL;
294 	new_range2 = NULL;
295 	count = 0;
296 
297 restart:
298 	/*
299 	 * Preallocate two ranges so we don't have to worry about blocking
300 	 * in the middle of the lock code.
301 	 */
302 	if (new_range1 == NULL)
303 		new_range1 = lf_alloc_range();
304 	if (new_range2 == NULL)
305 		new_range2 = lf_alloc_range();
306 	first_match = NULL;
307 	last_match = NULL;
308 	insert_point = NULL;
309 	wakeup_needed = 0;
310 
311 	lf_print_lock(lock);
312 
313 	/*
314 	 * Locate the insertion point for the new lock (the first range
315 	 * with an lf_start >= start).
316 	 *
317 	 * Locate the first and latch ranges owned by us that overlap
318 	 * the requested range.
319 	 */
320 	TAILQ_FOREACH(range, &lock->lf_range, lf_link) {
321 		if (insert_point == NULL && range->lf_start >= start)
322 			insert_point = range;
323 
324 		/*
325 		 * Skip non-overlapping locks.  Locks are sorted by lf_start
326 		 * So we can terminate the search when lf_start exceeds the
327 		 * requested range (insert_point is still guarenteed to be
328 		 * set properly).
329 		 */
330 		if (range->lf_end < start)
331 			continue;
332 		if (range->lf_start > end) {
333 			range = NULL;
334 			break;
335 		}
336 
337 		/*
338 		 * Overlapping lock.  Set first_match and last_match if we
339 		 * are the owner.
340 		 */
341 		if (range->lf_owner == owner) {
342 			if (first_match == NULL)
343 				first_match = range;
344 			last_match = range;
345 			continue;
346 		}
347 
348 		/*
349 		 * If we aren't the owner check for a conflicting lock.  Only
350 		 * if not unlocking.
351 		 */
352 		if (type != F_UNLCK) {
353 			if (type == F_WRLCK || range->lf_type == F_WRLCK)
354 				break;
355 		}
356 	}
357 
358 	/*
359 	 * If a conflicting lock was observed, block or fail as appropriate.
360 	 * (this code is skipped when unlocking)
361 	 */
362 	if (range != NULL) {
363 		if ((flags & F_WAIT) == 0) {
364 			error = EAGAIN;
365 			goto do_cleanup;
366 		}
367 
368 		/*
369 		 * We are blocked. For POSIX locks we have to check
370 		 * for deadlocks and return with EDEADLK. This is done
371 		 * by checking whether range->lf_owner is already
372 		 * blocked.
373 		 *
374 		 * Since flock-style locks cover the whole file, a
375 		 * deadlock between those is nearly impossible.
376 		 * This can only occur if a process tries to lock the
377 		 * same inode exclusively while holding a shared lock
378 		 * with another descriptor.
379 		 * XXX How can we cleanly detect this?
380 		 * XXX The current mixing of flock & fcntl/lockf is evil.
381 		 *
382 		 * Handle existing locks of flock-style like POSIX locks.
383 		 */
384 		if (flags & F_POSIX) {
385 			TAILQ_FOREACH(brange, &lock->lf_blocked, lf_link) {
386 				if (brange->lf_owner == range->lf_owner) {
387 					error = EDEADLK;
388 					goto do_cleanup;
389 				}
390 			}
391 		}
392 
393 		/*
394 		 * For flock-style locks, we must first remove
395 		 * any shared locks that we hold before we sleep
396 		 * waiting for an exclusive lock.
397 		 */
398 		if ((flags & F_POSIX) == 0 && type == F_WRLCK)
399 			lf_setlock(lock, owner, F_UNLCK, 0, start, end);
400 
401 		brange = new_range1;
402 		new_range1 = NULL;
403 		lf_create_range(brange, owner, type, 0, start, end);
404 		TAILQ_INSERT_TAIL(&lock->lf_blocked, brange, lf_link);
405 		error = tsleep(brange, PCATCH, "lockf", 0);
406 
407 		/*
408 		 * We may have been awaked by a signal and/or by a
409 		 * debugger continuing us (in which case we must remove
410 		 * ourselves from the blocked list) and/or by another
411 		 * process releasing/downgrading a lock (in which case
412 		 * we have already been removed from the blocked list
413 		 * and our lf_flags field is 1).
414 		 *
415 		 * Sleep if it looks like we might be livelocking.
416 		 */
417 		if (brange->lf_flags == 0)
418 			TAILQ_REMOVE(&lock->lf_blocked, brange, lf_link);
419 		if (count == 2)
420 			tsleep(brange, 0, "lockfz", 2);
421 		else
422 			++count;
423 		lf_destroy_range(brange);
424 
425 		if (error)
426 			goto do_cleanup;
427 		goto restart;
428 	}
429 
430 	/*
431 	 * If there are no overlapping locks owned by us then creating
432 	 * the new lock is easy.  This is the most common case.
433 	 */
434 	if (first_match == NULL) {
435 		if (type == F_UNLCK)
436 			goto do_wakeup;
437 		if (flags & F_POSIX) {
438 			if (lf_count_change(owner, 1)) {
439 				error = ENOLCK;
440 				goto do_cleanup;
441 			}
442 		}
443 		range = new_range1;
444 		new_range1 = NULL;
445 		lf_create_range(range, owner, type, flags, start, end);
446 		lf_insert(&lock->lf_range, range, insert_point);
447 		goto do_wakeup;
448 	}
449 
450 	/*
451 	 * double_clip - Calculate a special case where TWO locks may have
452 	 *		 to be added due to the new lock breaking up an
453 	 *		 existing incompatible lock in the middle.
454 	 *
455 	 * unlock_override - Calculate a special case where NO locks
456 	 *		 need to be created.  This occurs when an unlock
457 	 *		 does not clip any locks at the front and rear.
458 	 *
459 	 * WARNING!  closef() and fdrop() assume that an F_UNLCK of the
460 	 *	     entire range will always succeed so the unlock_override
461 	 *	     case is mandatory.
462 	 */
463 	double_clip = 0;
464 	unlock_override = 0;
465 	if (first_match->lf_start < start) {
466 		if (first_match == last_match && last_match->lf_end > end)
467 			double_clip = 1;
468 	} else if (type == F_UNLCK && last_match->lf_end <= end) {
469 		unlock_override = 1;
470 	}
471 
472 	/*
473 	 * Figure out the worst case net increase in POSIX locks and account
474 	 * for it now before we start modifying things.  If neither the
475 	 * first or last locks match we have an issue.  If there is only
476 	 * one overlapping range which needs to be clipped on both ends
477 	 * we wind up having to create up to two new locks, else only one.
478 	 *
479 	 * When unlocking the worst case is always 1 new lock if our
480 	 * unlock request cuts the middle out of an existing lock range.
481 	 *
482 	 * count represents the 'cleanup' adjustment needed.  It starts
483 	 * negative, is incremented whenever we create a new POSIX lock,
484 	 * and decremented whenever we delete an existing one.  At the
485 	 * end of the day it had better be <= 0 or we didn't calculate the
486 	 * worse case properly here.
487 	 */
488 	count = 0;
489 	if ((flags & F_POSIX) && !unlock_override) {
490 		if (!lf_match(first_match, type, flags) &&
491 		    !lf_match(last_match, type, flags)
492 		) {
493 			if (double_clip && type != F_UNLCK)
494 				count = -2;
495 			else
496 				count = -1;
497 		}
498 		if (count && lf_count_change(owner, -count)) {
499 			error = ENOLCK;
500 			goto do_cleanup;
501 		}
502 	}
503 	/* else flock style lock which encompasses entire range */
504 
505 	/*
506 	 * Create and insert the lock represented the requested range.
507 	 * Adjust the net POSIX lock count.  We have to move our insertion
508 	 * point since brange now represents the first record >= start.
509 	 *
510 	 * When unlocking, no new lock is inserted but we still clip.
511 	 */
512 	if (type != F_UNLCK) {
513 		brange = new_range1;
514 		new_range1 = NULL;
515 		lf_create_range(brange, owner, type, flags, start, end);
516 		lf_insert(&lock->lf_range, brange, insert_point);
517 		insert_point = brange;
518 		if (flags & F_POSIX)
519 			++count;
520 	} else {
521 		brange = NULL;
522 	}
523 
524 	/*
525 	 * Handle the double_clip case.  This is the only case where
526 	 * we wind up having to add TWO locks.
527 	 */
528 	if (double_clip) {
529 		KKASSERT(first_match == last_match);
530 		last_match = new_range2;
531 		new_range2 = NULL;
532 		lf_create_range(last_match, first_match->lf_owner,
533 				first_match->lf_type, first_match->lf_flags,
534 				end + 1, first_match->lf_end);
535 		first_match->lf_end = start - 1;
536 		first_match->lf_flags &= ~F_NOEND;
537 
538 		/*
539 		 * Figure out where to insert the right side clip.
540 		 */
541 		lf_insert(&lock->lf_range, last_match, first_match);
542 		if (last_match->lf_flags & F_POSIX)
543 			++count;
544 	}
545 
546 	/*
547 	 * Clip or destroy the locks between first_match and last_match,
548 	 * inclusive.  Ignore the primary lock we created (brange).  Note
549 	 * that if double-clipped, first_match and last_match will be
550 	 * outside our clipping range.  Otherwise first_match and last_match
551 	 * will be deleted.
552 	 *
553 	 * We have already taken care of any double clipping.
554 	 *
555 	 * The insert_point may become invalid as we delete records, do not
556 	 * use that pointer any more.  Also, when removing something other
557 	 * then 'range' we have to check to see if the item we are removing
558 	 * is 'next' and adjust 'next' properly.
559 	 *
560 	 * NOTE: brange will be NULL if F_UNLCKing.
561 	 */
562 	TAILQ_INIT(&deadlist);
563 	next = first_match;
564 
565 	while ((range = next) != NULL) {
566 		next = TAILQ_NEXT(range, lf_link);
567 
568 		/*
569 		 * Ignore elements that we do not own and ignore the
570 		 * primary request range which we just created.
571 		 */
572 		if (range->lf_owner != owner || range == brange)
573 			continue;
574 
575 		/*
576 		 * We may have to wakeup a waiter when downgrading a lock.
577 		 */
578 		if (type == F_UNLCK)
579 			wakeup_needed = 1;
580 		if (type == F_RDLCK && range->lf_type == F_WRLCK)
581 			wakeup_needed = 1;
582 
583 		/*
584 		 * Clip left.  This can only occur on first_match.
585 		 *
586 		 * Merge the left clip with brange if possible.  This must
587 		 * be done specifically, not in the optimized merge heuristic
588 		 * below, since we may have counted on it in our 'count'
589 		 * calculation above.
590 		 */
591 		if (range->lf_start < start) {
592 			KKASSERT(range == first_match);
593 			if (brange &&
594 			    range->lf_end >= start - 1 &&
595 			    lf_match(range, type, flags)) {
596 				range->lf_end = brange->lf_end;
597 				range->lf_flags |= brange->lf_flags & F_NOEND;
598 				/*
599 				 * Removing something other then 'range',
600 				 * adjust 'next' if necessary.
601 				 */
602 				if (next == brange)
603 					next = TAILQ_NEXT(next, lf_link);
604 				TAILQ_REMOVE(&lock->lf_range, brange, lf_link);
605 				if (brange->lf_flags & F_POSIX)
606 					--count;
607 				TAILQ_INSERT_TAIL(&deadlist, brange, lf_link);
608 				brange = range;
609 			} else if (range->lf_end >= start) {
610 				range->lf_end = start - 1;
611 				if (type != F_UNLCK)
612 					range->lf_flags &= ~F_NOEND;
613 			}
614 			if (range == last_match)
615 				break;
616 			continue;
617 		}
618 
619 		/*
620 		 * Clip right.  This can only occur on last_match.
621 		 *
622 		 * Merge the right clip if possible.  This must be done
623 		 * specifically, not in the optimized merge heuristic
624 		 * below, since we may have counted on it in our 'count'
625 		 * calculation.
626 		 *
627 		 * Since we are adjusting lf_start, we have to move the
628 		 * record to maintain the sorted list.  Since lf_start is
629 		 * only getting larger we can use the next element as the
630 		 * insert point (we don't have to backtrack).
631 		 */
632 		if (range->lf_end > end) {
633 			KKASSERT(range == last_match);
634 			if (brange &&
635 			    range->lf_start <= end + 1 &&
636 			    lf_match(range, type, flags)) {
637 				brange->lf_end = range->lf_end;
638 				brange->lf_flags |= range->lf_flags & F_NOEND;
639 				TAILQ_REMOVE(&lock->lf_range, range, lf_link);
640 				if (range->lf_flags & F_POSIX)
641 					--count;
642 				TAILQ_INSERT_TAIL(&deadlist, range, lf_link);
643 			} else if (range->lf_start <= end) {
644 				range->lf_start = end + 1;
645 				TAILQ_REMOVE(&lock->lf_range, range, lf_link);
646 				lf_insert(&lock->lf_range, range, next);
647 			}
648 			/* range == last_match, we are done */
649 			break;
650 		}
651 
652 		/*
653 		 * The record must be entirely enclosed.  Note that the
654 		 * record could be first_match or last_match, and will be
655 		 * deleted.
656 		 */
657 		KKASSERT(range->lf_start >= start && range->lf_end <= end);
658 		TAILQ_REMOVE(&lock->lf_range, range, lf_link);
659 		if (range->lf_flags & F_POSIX)
660 			--count;
661 		TAILQ_INSERT_TAIL(&deadlist, range, lf_link);
662 		if (range == last_match)
663 			break;
664 	}
665 
666 	/*
667 	 * Attempt to merge locks adjacent to brange.  For example, we may
668 	 * have had to clip first_match and/or last_match, and they might
669 	 * be adjacent.  Or there might simply have been an adjacent lock
670 	 * already there.
671 	 *
672 	 * Don't get fancy, just check adjacent elements in the list if they
673 	 * happen to be owned by us.
674 	 *
675 	 * This case only gets hit if we have a situation where a shared
676 	 * and exclusive lock are adjacent, and the exclusive lock is
677 	 * downgraded to shared or the shared lock is upgraded to exclusive.
678 	 */
679 	if (brange) {
680 		range = TAILQ_PREV(brange, lockf_range_list, lf_link);
681 		if (range &&
682 		    range->lf_owner == owner &&
683 		    range->lf_end == brange->lf_start - 1 &&
684 		    lf_match(range, type, flags)
685 		) {
686 			/*
687 			 * Extend range to cover brange and scrap brange.
688 			 */
689 			range->lf_end = brange->lf_end;
690 			range->lf_flags |= brange->lf_flags & F_NOEND;
691 			TAILQ_REMOVE(&lock->lf_range, brange, lf_link);
692 			if (brange->lf_flags & F_POSIX)
693 				--count;
694 			TAILQ_INSERT_TAIL(&deadlist, brange, lf_link);
695 			brange = range;
696 		}
697 		range = TAILQ_NEXT(brange, lf_link);
698 		if (range &&
699 		    range->lf_owner == owner &&
700 		    range->lf_start == brange->lf_end + 1 &&
701 		    lf_match(range, type, flags)
702 		) {
703 			/*
704 			 * Extend brange to cover range and scrap range.
705 			 */
706 			brange->lf_end = range->lf_end;
707 			brange->lf_flags |= range->lf_flags & F_NOEND;
708 			TAILQ_REMOVE(&lock->lf_range, range, lf_link);
709 			if (range->lf_flags & F_POSIX)
710 				--count;
711 			TAILQ_INSERT_TAIL(&deadlist, range, lf_link);
712 		}
713 	}
714 
715 	/*
716 	 * Destroy deleted elements.  We didn't want to do it in the loop
717 	 * because the free() might have blocked.
718 	 *
719 	 * Adjust the count for any posix locks we thought we might create
720 	 * but didn't.
721 	 */
722 	while ((range = TAILQ_FIRST(&deadlist)) != NULL) {
723 		TAILQ_REMOVE(&deadlist, range, lf_link);
724 		lf_destroy_range(range);
725 	}
726 
727 	KKASSERT(count <= 0);
728 	if (count < 0)
729 		lf_count_change(owner, count);
730 do_wakeup:
731 	lf_print_lock(lock);
732 	if (wakeup_needed)
733 		lf_wakeup(lock, start, end);
734 	error = 0;
735 do_cleanup:
736 	if (new_range1 != NULL)
737 		lf_destroy_range(new_range1);
738 	if (new_range2 != NULL)
739 		lf_destroy_range(new_range2);
740 	return(error);
741 }
742 
743 /*
744  * Check whether there is a blocking lock,
745  * and if so return its process identifier.
746  */
747 static int
748 lf_getlock(struct flock *fl, struct lockf *lock, struct proc *owner,
749 	   int type, int flags, off_t start, off_t end)
750 {
751 	struct lockf_range *range;
752 
753 	TAILQ_FOREACH(range, &lock->lf_range, lf_link)
754 		if (range->lf_owner != owner &&
755 		    lf_overlap(range, start, end) &&
756 		    (type == F_WRLCK || range->lf_type == F_WRLCK))
757 			break;
758 	if (range == NULL) {
759 		fl->l_type = F_UNLCK;
760 		return(0);
761 	}
762 	fl->l_type = range->lf_type;
763 	fl->l_whence = SEEK_SET;
764 	fl->l_start = range->lf_start;
765 	if (range->lf_flags & F_NOEND)
766 		fl->l_len = 0;
767 	else
768 		fl->l_len = range->lf_end - range->lf_start + 1;
769 	if (range->lf_owner != NULL && (range->lf_flags & F_POSIX))
770 		fl->l_pid = range->lf_owner->p_pid;
771 	else
772 		fl->l_pid = -1;
773 	return(0);
774 }
775 
776 /*
777  * Wakeup pending lock attempts.  Theoretically we can stop as soon as
778  * we encounter an exclusive request that covers the whole range (at least
779  * insofar as the sleep code above calls lf_wakeup() if it would otherwise
780  * exit instead of loop), but for now just wakeup all overlapping
781  * requests.  XXX
782  */
783 static void
784 lf_wakeup(struct lockf *lock, off_t start, off_t end)
785 {
786 	struct lockf_range *range, *nrange;
787 
788 	TAILQ_FOREACH_MUTABLE(range, &lock->lf_blocked, lf_link, nrange) {
789 		if (lf_overlap(range, start, end) == 0)
790 			continue;
791 		TAILQ_REMOVE(&lock->lf_blocked, range, lf_link);
792 		range->lf_flags = 1;
793 		wakeup(range);
794 	}
795 }
796 
797 /*
798  * Allocate a range structure and initialize it sufficiently such that
799  * lf_destroy_range() does not barf.
800  */
801 static struct lockf_range *
802 lf_alloc_range(void)
803 {
804 	struct lockf_range *range;
805 
806 #ifdef INVARIANTS
807 	atomic_add_int(&lf_global_counter, 1);
808 #endif
809 	range = kmalloc(sizeof(struct lockf_range), M_LOCKF, M_WAITOK);
810 	range->lf_owner = NULL;
811 	return(range);
812 }
813 
814 static void
815 lf_insert(struct lockf_range_list *list, struct lockf_range *elm,
816 	  struct lockf_range *insert_point)
817 {
818 	while (insert_point && insert_point->lf_start < elm->lf_start)
819 		insert_point = TAILQ_NEXT(insert_point, lf_link);
820 	if (insert_point != NULL)
821 		TAILQ_INSERT_BEFORE(insert_point, elm, lf_link);
822 	else
823 		TAILQ_INSERT_TAIL(list, elm, lf_link);
824 }
825 
826 static void
827 lf_create_range(struct lockf_range *range, struct proc *owner, int type,
828 		int flags, off_t start, off_t end)
829 {
830 	KKASSERT(start <= end);
831 	range->lf_type = type;
832 	range->lf_flags = flags;
833 	range->lf_start = start;
834 	range->lf_end = end;
835 	range->lf_owner = owner;
836 
837 	lf_printf("lf_create_range: %ju..%ju\n",
838 	    (uintmax_t)range->lf_start, (uintmax_t)range->lf_end);
839 }
840 
841 static void
842 lf_destroy_range(struct lockf_range *range)
843 {
844 	lf_printf("lf_destroy_range: %ju..%ju\n",
845 	    (uintmax_t)range->lf_start, (uintmax_t)range->lf_end);
846 	kfree(range, M_LOCKF);
847 #ifdef INVARIANTS
848 	atomic_add_int(&lf_global_counter, -1);
849 	KKASSERT(lf_global_counter >= 0);
850 #endif
851 }
852 
853 #ifdef LOCKF_DEBUG
854 
855 static void
856 _lf_printf(const char *ctl, ...)
857 {
858 	struct proc *p;
859 	__va_list va;
860 
861 	if (lf_print_ranges) {
862 	    if ((p = curproc) != NULL)
863 		kprintf("pid %d (%s): ", p->p_pid, p->p_comm);
864 	}
865 	__va_start(va, ctl);
866 	kvprintf(ctl, va);
867 	__va_end(va);
868 }
869 
870 static void
871 _lf_print_lock(const struct lockf *lock)
872 {
873 	struct lockf_range *range;
874 
875 	if (lf_print_ranges == 0)
876 		return;
877 
878 	if (TAILQ_EMPTY(&lock->lf_range)) {
879 		lf_printf("lockf %p: no ranges locked\n", lock);
880 	} else {
881 		lf_printf("lockf %p:\n", lock);
882 	}
883 	TAILQ_FOREACH(range, &lock->lf_range, lf_link)
884 		kprintf("\t%jd..%jd type %s owned by %d\n",
885 		       (uintmax_t)range->lf_start, (uintmax_t)range->lf_end,
886 		       range->lf_type == F_RDLCK ? "shared" : "exclusive",
887 		       range->lf_flags & F_POSIX ? range->lf_owner->p_pid : -1);
888 	if (TAILQ_EMPTY(&lock->lf_blocked))
889 		kprintf("no process waiting for range\n");
890 	else
891 		kprintf("blocked locks:");
892 	TAILQ_FOREACH(range, &lock->lf_blocked, lf_link)
893 		kprintf("\t%jd..%jd type %s waiting on %p\n",
894 		       (uintmax_t)range->lf_start, (uintmax_t)range->lf_end,
895 		       range->lf_type == F_RDLCK ? "shared" : "exclusive",
896 		       range);
897 }
898 #endif /* LOCKF_DEBUG */
899