2 * Copyright (c) 2004 Joerg Sonnenberger <joerg@bec.de>. All rights reserved.
3 * Copyright (c) 2006 Matthew Dillon <dillon@backplane.com>. All rights reserved.
5 * Copyright (c) 1982, 1986, 1989, 1993
6 * The Regents of the University of California. All rights reserved.
8 * This code is derived from software contributed to Berkeley by
9 * Scooter Morris at Genentech Inc.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
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. All advertising materials mentioning features or use of this software
20 * must display the following acknowledgement:
21 * This product includes software developed by the University of
22 * California, Berkeley and its contributors.
23 * 4. Neither the name of the University nor the names of its contributors
24 * may be used to endorse or promote products derived from this software
25 * without specific prior written permission.
27 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
28 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
29 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
30 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
31 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
32 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
33 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
34 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
35 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
36 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39 * @(#)ufs_lockf.c 8.3 (Berkeley) 1/6/94
40 * $FreeBSD: src/sys/kern/kern_lockf.c,v 1.25 1999/11/16 16:28:56 phk Exp $
41 * $DragonFly: src/sys/kern/kern_lockf.c,v 1.28 2006/05/08 00:38:58 dillon Exp $
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/kernel.h>
49 #include <sys/unistd.h>
50 #include <sys/vnode.h>
51 #include <sys/malloc.h>
52 #include <sys/fcntl.h>
53 #include <sys/resourcevar.h>
55 #include <sys/lockf.h>
56 #include <machine/limits.h> /* for LLONG_MAX */
57 #include <machine/stdarg.h>
60 int lf_global_counter = 0;
64 int lf_print_ranges = 0;
66 static void _lf_print_lock(const struct lockf *);
67 static void _lf_printf(const char *, ...);
69 #define lf_print_lock(lock) if (lf_print_ranges) _lf_print_lock(lock)
70 #define lf_printf(ctl, args...) if (lf_print_ranges) _lf_printf(ctl, args)
72 #define lf_print_lock(lock)
73 #define lf_printf(ctl, args...)
76 static MALLOC_DEFINE(M_LOCKF, "lockf", "Byte-range locking structures");
78 static void lf_wakeup(struct lockf *, off_t, off_t);
79 static struct lockf_range *lf_alloc_range(void);
80 static void lf_create_range(struct lockf_range *, struct proc *, int, int,
82 static void lf_insert(struct lockf_range_list *list,
83 struct lockf_range *elm,
84 struct lockf_range *insert_point);
85 static void lf_destroy_range(struct lockf_range *);
87 static int lf_setlock(struct lockf *, struct proc *, int, int,
89 static int lf_getlock(struct flock *, struct lockf *, struct proc *,
90 int, int, off_t, off_t);
92 static int lf_count_change(struct proc *, int);
95 * Return TRUE (non-zero) if the type and posix flags match.
99 lf_match(struct lockf_range *range, int type, int flags)
101 if (range->lf_type != type)
103 if ((range->lf_flags ^ flags) & F_POSIX)
109 * Check whether range and [start, end] overlap.
113 lf_overlap(const struct lockf_range *range, off_t start, off_t end)
115 if (range->lf_start >= start && range->lf_start <= end)
117 else if (start >= range->lf_start && start <= range->lf_end)
125 * Change the POSIX lock accounting for the given process.
128 lf_count_adjust(struct proc *p, int increase)
134 uip = p->p_ucred->cr_uidinfo;
137 uip->ui_posixlocks += p->p_numposixlocks;
139 uip->ui_posixlocks -= p->p_numposixlocks;
141 KASSERT(uip->ui_posixlocks >= 0,
142 ("Negative number of POSIX locks held by %s user: %d.",
143 increase ? "new" : "old", uip->ui_posixlocks));
147 lf_count_change(struct proc *owner, int diff)
152 /* we might actually not have a process context */
156 uip = owner->p_ucred->cr_uidinfo;
158 max = MIN(owner->p_rlimit[RLIMIT_POSIXLOCKS].rlim_cur,
159 maxposixlocksperuid);
160 if (diff > 0 && owner->p_ucred->cr_uid != 0 && max != -1 &&
161 uip->ui_posixlocks >= max ) {
165 uip->ui_posixlocks += diff;
166 owner->p_numposixlocks += diff;
168 KASSERT(uip->ui_posixlocks >= 0,
169 ("Negative number of POSIX locks held by user: %d.",
170 uip->ui_posixlocks));
171 KASSERT(owner->p_numposixlocks >= 0,
172 ("Negative number of POSIX locks held by proc: %d.",
173 uip->ui_posixlocks));
179 * Advisory record locking support
182 lf_advlock(struct vop_advlock_args *ap, struct lockf *lock, u_quad_t size)
184 struct flock *fl = ap->a_fl;
187 int type, flags, error;
191 * Convert the flock structure into a start and end.
193 switch (fl->l_whence) {
197 * Caller is responsible for adding any necessary offset
198 * when SEEK_CUR is used.
204 start = size + fl->l_start;
212 if (fl->l_len == 0) {
216 end = start + fl->l_len - 1;
224 * This isn't really correct for flock-style locks,
225 * but the current handling is somewhat broken anyway.
227 owner = (struct proc *)ap->a_id;
230 * Do the requested operation.
232 lwkt_gettoken(&ilock, lwkt_token_pool_get(lock));
234 if (lock->init_done == 0) {
235 TAILQ_INIT(&lock->lf_range);
236 TAILQ_INIT(&lock->lf_blocked);
242 error = lf_setlock(lock, owner, type, flags, start, end);
246 error = lf_setlock(lock, owner, type, flags, start, end);
250 error = lf_getlock(fl, lock, owner, type, flags, start, end);
257 lwkt_reltoken(&ilock);
262 lf_setlock(struct lockf *lock, struct proc *owner, int type, int flags,
263 off_t start, off_t end)
265 struct lockf_range *range;
266 struct lockf_range *brange;
267 struct lockf_range *next;
268 struct lockf_range *first_match;
269 struct lockf_range *last_match;
270 struct lockf_range *insert_point;
271 struct lockf_range *new_range1;
272 struct lockf_range *new_range2;
277 struct lockf_range_list deadlist;
285 * Preallocate two ranges so we don't have to worry about blocking
286 * in the middle of the lock code.
288 if (new_range1 == NULL)
289 new_range1 = lf_alloc_range();
290 if (new_range2 == NULL)
291 new_range2 = lf_alloc_range();
300 * Locate the insertion point for the new lock (the first range
301 * with an lf_start >= start).
303 * Locate the first and latch ranges owned by us that overlap
304 * the requested range.
306 TAILQ_FOREACH(range, &lock->lf_range, lf_link) {
307 if (insert_point == NULL && range->lf_start >= start)
308 insert_point = range;
311 * Skip non-overlapping locks. Locks are sorted by lf_start
312 * So we can terminate the search when lf_start exceeds the
313 * requested range (insert_point is still guarenteed to be
316 if (range->lf_end < start)
318 if (range->lf_start > end) {
324 * Overlapping lock. Set first_match and last_match if we
327 if (range->lf_owner == owner) {
328 if (first_match == NULL)
335 * If we aren't the owner check for a conflicting lock. Only
338 if (type != F_UNLCK) {
339 if (type == F_WRLCK || range->lf_type == F_WRLCK)
345 * If a conflicting lock was observed, block or fail as appropriate.
346 * (this code is skipped when unlocking)
349 if ((flags & F_WAIT) == 0) {
355 * We are blocked. For POSIX locks we have to check
356 * for deadlocks and return with EDEADLK. This is done
357 * by checking whether range->lf_owner is already
360 * Since flock-style locks cover the whole file, a
361 * deadlock between those is nearly impossible.
362 * This can only occur if a process tries to lock the
363 * same inode exclusively while holding a shared lock
364 * with another descriptor.
365 * XXX How can we cleanly detect this?
366 * XXX The current mixing of flock & fcntl/lockf is evil.
368 * Handle existing locks of flock-style like POSIX locks.
370 if (flags & F_POSIX) {
371 TAILQ_FOREACH(brange, &lock->lf_blocked, lf_link)
372 if (brange->lf_owner == range->lf_owner) {
379 * For flock-style locks, we must first remove
380 * any shared locks that we hold before we sleep
381 * waiting for an exclusive lock.
383 if ((flags & F_POSIX) == 0 && type == F_WRLCK)
384 lf_setlock(lock, owner, F_UNLCK, 0, start, end);
388 lf_create_range(brange, owner, type, 0, start, end);
389 TAILQ_INSERT_TAIL(&lock->lf_blocked, brange, lf_link);
390 error = tsleep(brange, PCATCH, "lockf", 0);
393 * We may have been awaked by a signal and/or by a
394 * debugger continuing us (in which case we must remove
395 * ourselves from the blocked list) and/or by another
396 * process releasing/downgrading a lock (in which case
397 * we have already been removed from the blocked list
398 * and our lf_flags field is 1).
400 * Sleep if it looks like we might be livelocking.
402 if (brange->lf_flags == 0)
403 TAILQ_REMOVE(&lock->lf_blocked, brange, lf_link);
405 tsleep(brange, 0, "lockfz", 2);
408 lf_destroy_range(brange);
416 * If there are no overlapping locks owned by us then creating
417 * the new lock is easy. This is the most common case.
419 if (first_match == NULL) {
422 if (flags & F_POSIX) {
423 if (lf_count_change(owner, 1)) {
430 lf_create_range(range, owner, type, flags, start, end);
431 lf_insert(&lock->lf_range, range, insert_point);
436 * This is a special case that we need to check for in a couple
439 if (first_match == last_match && first_match->lf_start < start &&
440 last_match->lf_end > end) {
447 * Figure out the worst case net increase in POSIX locks and account
448 * for it now before we start modifying things. If neither the
449 * first or last locks match we have an issue. If there is only
450 * one overlapping range which needs to be clipped on both ends
451 * we wind up having to create up to two new locks, else only one.
453 * count represents the 'cleanup' adjustment needed. It starts
454 * negative, is incremented whenever we create a new POSIX lock,
455 * and decremented whenever we delete an existing one. At the
456 * end of the day it had better be <= 0 or we didn't calculate the
457 * worse case properly here.
460 if (flags & F_POSIX) {
461 if (!lf_match(first_match, type, flags) &&
462 !lf_match(last_match, type, flags)
469 if (count && lf_count_change(owner, -count)) {
476 * Create and insert the lock represented the requested range.
477 * Adjust the net POSIX lock count. We have to move our insertion
478 * point since brange now represents the first record >= start.
480 * When unlocking, no new lock is inserted but we still clip.
482 if (type != F_UNLCK) {
485 lf_create_range(brange, owner, type, flags, start, end);
486 lf_insert(&lock->lf_range, brange, insert_point);
487 insert_point = brange;
495 * Handle the double_clip case. This is the only case where
496 * we wind up having to add TWO locks.
499 KKASSERT(first_match == last_match);
500 last_match = new_range2;
502 lf_create_range(last_match, first_match->lf_owner,
503 first_match->lf_type, first_match->lf_flags,
504 end + 1, first_match->lf_end);
505 first_match->lf_end = start - 1;
506 first_match->lf_flags &= ~F_NOEND;
509 * Figure out where to insert the right side clip.
511 lf_insert(&lock->lf_range, last_match, first_match);
512 if (last_match->lf_flags & F_POSIX)
517 * Clip or destroy the locks between first_match and last_match,
518 * inclusive. Ignore the primary lock we created (brange). Note
519 * that if double-clipped, first_match and last_match will be
520 * outside our clipping range. Otherwise first_match and last_match
523 * We have already taken care of any double clipping.
525 * insert_point may become invalid as we delete records, do not
526 * use that pointer any more.
528 * NOTE: brange will be NULL if F_UNLCKing.
530 TAILQ_INIT(&deadlist);
533 while ((range = next) != NULL) {
534 next = TAILQ_NEXT(range, lf_link);
537 * Ignore elements that we do not own and ignore the
538 * primary request range which we just created.
540 if (range->lf_owner != owner || range == brange)
544 * We may have to wakeup a waiter when downgrading a lock.
548 if (type == F_RDLCK && range->lf_type == F_WRLCK)
552 * Clip left. This can only occur on first_match. If
553 * we have already double-clipped it there is nothing to do.
555 if (range->lf_start < start) {
556 KKASSERT(range == first_match);
557 if (range->lf_end >= start) {
558 range->lf_end = start - 1;
559 range->lf_flags &= ~F_NOEND;
561 if (range == last_match)
567 * Clip right. This can only occur on last_match. If
568 * we have already double-clipped it there is nothing to do.
570 * Since we are adjusting lf_start, we have to move the
571 * record to maintain the sorted list. Since lf_start is
572 * only getting larger we can use the next element as the
573 * insert point (we don't have to backtrack).
575 if (range->lf_end > end) {
576 KKASSERT(range == last_match);
577 if (last_match->lf_start <= end) {
578 last_match->lf_start = end + 1;
579 TAILQ_REMOVE(&lock->lf_range, last_match, lf_link);
580 lf_insert(&lock->lf_range, last_match, next);
582 /* range == last_match, we are done */
587 * The record must be entirely enclosed. Note that the
588 * record could be first_match or last_match, and will be
591 KKASSERT(range->lf_start >= start && range->lf_end <= end);
592 TAILQ_REMOVE(&lock->lf_range, range, lf_link);
593 if (range->lf_flags & F_POSIX)
595 TAILQ_INSERT_TAIL(&deadlist, range, lf_link);
596 if (range == last_match)
601 * Attempt to merge locks adjacent to brange. For example, we may
602 * have had to clip first_match and/or last_match, and they might
603 * be adjacent. Or there might simply have been an adjacent lock
606 * Don't get fancy, just check adjacent elements in the list if they
607 * happen to be owned by us.
610 range = TAILQ_PREV(brange, lockf_range_list, lf_link);
612 range->lf_owner == owner &&
613 range->lf_end == brange->lf_start - 1 &&
614 lf_match(range, type, flags)
619 range->lf_end = brange->lf_end;
620 range->lf_flags |= brange->lf_flags & F_NOEND;
621 TAILQ_REMOVE(&lock->lf_range, brange, lf_link);
622 if (brange->lf_flags & F_POSIX)
624 TAILQ_INSERT_TAIL(&deadlist, brange, lf_link);
627 range = TAILQ_NEXT(brange, lf_link);
629 range->lf_owner == owner &&
630 range->lf_start == brange->lf_end + 1 &&
631 lf_match(range, type, flags)
636 brange->lf_end = range->lf_end;
637 brange->lf_flags |= brange->lf_flags & F_NOEND;
638 TAILQ_REMOVE(&lock->lf_range, range, lf_link);
639 if (range->lf_flags & F_POSIX)
641 TAILQ_INSERT_TAIL(&deadlist, range, lf_link);
646 * Destroy deleted elements. We didn't want to do it in the loop
647 * because the free() might have blocked.
649 * Adjust the count for any posix locks we thought we might create
652 while ((range = TAILQ_FIRST(&deadlist)) != NULL) {
653 TAILQ_REMOVE(&deadlist, range, lf_link);
654 lf_destroy_range(range);
657 KKASSERT(count <= 0);
659 lf_count_change(owner, count);
663 lf_wakeup(lock, start, end);
666 if (new_range1 != NULL)
667 lf_destroy_range(new_range1);
668 if (new_range2 != NULL)
669 lf_destroy_range(new_range2);
674 * Check whether there is a blocking lock,
675 * and if so return its process identifier.
678 lf_getlock(struct flock *fl, struct lockf *lock, struct proc *owner,
679 int type, int flags, off_t start, off_t end)
681 struct lockf_range *range;
683 TAILQ_FOREACH(range, &lock->lf_range, lf_link)
684 if (range->lf_owner != owner &&
685 lf_overlap(range, start, end) &&
686 (type == F_WRLCK || range->lf_type == F_WRLCK))
689 fl->l_type = F_UNLCK;
692 fl->l_type = range->lf_type;
693 fl->l_whence = SEEK_SET;
694 fl->l_start = range->lf_start;
695 if (range->lf_flags & F_NOEND)
698 fl->l_len = range->lf_end - range->lf_start + 1;
699 if (range->lf_owner != NULL && (range->lf_flags & F_POSIX))
700 fl->l_pid = range->lf_owner->p_pid;
707 * Wakeup pending lock attempts.
710 lf_wakeup(struct lockf *lock, off_t start, off_t end)
712 struct lockf_range *range, *nrange;
713 TAILQ_FOREACH_MUTABLE(range, &lock->lf_blocked, lf_link, nrange) {
714 if (lf_overlap(range, start, end) == 0)
716 TAILQ_REMOVE(&lock->lf_blocked, range, lf_link);
719 if (range->lf_start >= start && range->lf_end <= end)
725 * Allocate a range structure and initialize it sufficiently such that
726 * lf_destroy_range() does not barf.
728 static struct lockf_range *
731 struct lockf_range *range;
736 range = malloc(sizeof(struct lockf_range), M_LOCKF, M_WAITOK);
737 range->lf_owner = NULL;
742 lf_insert(struct lockf_range_list *list, struct lockf_range *elm,
743 struct lockf_range *insert_point)
745 while (insert_point && insert_point->lf_start < elm->lf_start)
746 insert_point = TAILQ_NEXT(insert_point, lf_link);
747 if (insert_point != NULL)
748 TAILQ_INSERT_BEFORE(insert_point, elm, lf_link);
750 TAILQ_INSERT_TAIL(list, elm, lf_link);
754 lf_create_range(struct lockf_range *range, struct proc *owner, int type,
755 int flags, off_t start, off_t end)
757 KKASSERT(start <= end);
758 range->lf_type = type;
759 range->lf_flags = flags;
760 range->lf_start = start;
762 range->lf_owner = owner;
764 lf_printf("lf_create_range: %lld..%lld\n",
765 range->lf_start, range->lf_end);
769 lf_destroy_range(struct lockf_range *range)
771 lf_printf("lf_destroy_range: %lld..%lld\n",
772 range->lf_start, range->lf_end);
773 free(range, M_LOCKF);
776 KKASSERT(lf_global_counter>=0);
783 _lf_printf(const char *ctl, ...)
788 if (lf_print_ranges) {
789 if ((p = curproc) != NULL)
790 printf("pid %d (%s): ", p->p_pid, p->p_comm);
798 _lf_print_lock(const struct lockf *lock)
800 struct lockf_range *range;
802 if (lf_print_ranges == 0)
805 if (TAILQ_EMPTY(&lock->lf_range)) {
806 lf_printf("lockf %p: no ranges locked\n", lock);
808 lf_printf("lockf %p:\n", lock);
810 TAILQ_FOREACH(range, &lock->lf_range, lf_link)
811 printf("\t%lld..%lld type %s owned by %d\n",
812 range->lf_start, range->lf_end,
813 range->lf_type == F_RDLCK ? "shared" : "exclusive",
814 range->lf_flags & F_POSIX ? range->lf_owner->p_pid : -1);
815 if (TAILQ_EMPTY(&lock->lf_blocked))
816 printf("no process waiting for range\n");
818 printf("blocked locks:");
819 TAILQ_FOREACH(range, &lock->lf_blocked, lf_link)
820 printf("\t%lld..%lld type %s waiting on %p\n",
821 range->lf_start, range->lf_end,
822 range->lf_type == F_RDLCK ? "shared" : "exclusive",
825 #endif /* LOCKF_DEBUG */