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 1.37 2007/11/01 22:48:16 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>
59 #include <sys/spinlock2.h>
62 int lf_global_counter = 0;
66 int lf_print_ranges = 0;
68 static void _lf_print_lock(const struct lockf *);
69 static void _lf_printf(const char *, ...);
71 #define lf_print_lock(lock) if (lf_print_ranges) _lf_print_lock(lock)
72 #define lf_printf(ctl, args...) if (lf_print_ranges) _lf_printf(ctl, args)
74 #define lf_print_lock(lock)
75 #define lf_printf(ctl, args...)
78 static MALLOC_DEFINE(M_LOCKF, "lockf", "Byte-range locking structures");
80 static void lf_wakeup(struct lockf *, off_t, off_t);
81 static struct lockf_range *lf_alloc_range(void);
82 static void lf_create_range(struct lockf_range *, struct proc *, int, int,
84 static void lf_insert(struct lockf_range_list *list,
85 struct lockf_range *elm,
86 struct lockf_range *insert_point);
87 static void lf_destroy_range(struct lockf_range *);
89 static int lf_setlock(struct lockf *, struct proc *, int, int,
91 static int lf_getlock(struct flock *, struct lockf *, struct proc *,
92 int, int, off_t, off_t);
94 static int lf_count_change(struct proc *, int);
97 * Return TRUE (non-zero) if the type and posix flags match.
101 lf_match(struct lockf_range *range, int type, int flags)
103 if (range->lf_type != type)
105 if ((range->lf_flags ^ flags) & F_POSIX)
111 * Check whether range and [start, end] overlap.
115 lf_overlap(const struct lockf_range *range, off_t start, off_t end)
117 if (range->lf_start >= start && range->lf_start <= end)
119 else if (start >= range->lf_start && start <= range->lf_end)
127 * Change the POSIX lock accounting for the given process.
130 lf_count_adjust(struct proc *p, int increase)
136 uip = p->p_ucred->cr_uidinfo;
137 spin_lock_wr(&uip->ui_lock);
140 uip->ui_posixlocks += p->p_numposixlocks;
142 uip->ui_posixlocks -= p->p_numposixlocks;
144 KASSERT(uip->ui_posixlocks >= 0,
145 ("Negative number of POSIX locks held by %s user: %d.",
146 increase ? "new" : "old", uip->ui_posixlocks));
147 spin_unlock_wr(&uip->ui_lock);
151 lf_count_change(struct proc *owner, int diff)
156 /* we might actually not have a process context */
160 uip = owner->p_ucred->cr_uidinfo;
162 max = MIN(owner->p_rlimit[RLIMIT_POSIXLOCKS].rlim_cur,
163 maxposixlocksperuid);
165 spin_lock_wr(&uip->ui_lock);
166 if (diff > 0 && owner->p_ucred->cr_uid != 0 && max != -1 &&
167 uip->ui_posixlocks >= max ) {
170 uip->ui_posixlocks += diff;
171 owner->p_numposixlocks += diff;
172 KASSERT(uip->ui_posixlocks >= 0,
173 ("Negative number of POSIX locks held by user: %d.",
174 uip->ui_posixlocks));
175 KASSERT(owner->p_numposixlocks >= 0,
176 ("Negative number of POSIX locks held by proc: %d.",
177 uip->ui_posixlocks));
180 spin_unlock_wr(&uip->ui_lock);
185 * Advisory record locking support
188 lf_advlock(struct vop_advlock_args *ap, struct lockf *lock, u_quad_t size)
190 struct flock *fl = ap->a_fl;
193 int type, flags, error;
197 * Convert the flock structure into a start and end.
199 switch (fl->l_whence) {
203 * Caller is responsible for adding any necessary offset
204 * when SEEK_CUR is used.
210 start = size + fl->l_start;
220 if (fl->l_len == 0) {
224 end = start + fl->l_len - 1;
231 * This isn't really correct for flock-style locks,
232 * but the current handling is somewhat broken anyway.
234 owner = (struct proc *)ap->a_id;
237 * Do the requested operation.
239 lwkt_gettoken(&ilock, lwkt_token_pool_get(lock));
241 if (lock->init_done == 0) {
242 TAILQ_INIT(&lock->lf_range);
243 TAILQ_INIT(&lock->lf_blocked);
250 * NOTE: It is possible for both lf_range and lf_blocked to
251 * be empty if we block and get woken up, but another process
252 * then gets in and issues an unlock. So VMAYHAVELOCKS must
253 * be set after the lf_setlock() operation completes rather
256 error = lf_setlock(lock, owner, type, flags, start, end);
257 ap->a_vp->v_flag |= VMAYHAVELOCKS;
261 error = lf_setlock(lock, owner, type, flags, start, end);
262 if (TAILQ_EMPTY(&lock->lf_range) &&
263 TAILQ_EMPTY(&lock->lf_blocked)) {
264 ap->a_vp->v_flag &= ~VMAYHAVELOCKS;
269 error = lf_getlock(fl, lock, owner, type, flags, start, end);
276 lwkt_reltoken(&ilock);
281 lf_setlock(struct lockf *lock, struct proc *owner, int type, int flags,
282 off_t start, off_t end)
284 struct lockf_range *range;
285 struct lockf_range *brange;
286 struct lockf_range *next;
287 struct lockf_range *first_match;
288 struct lockf_range *last_match;
289 struct lockf_range *insert_point;
290 struct lockf_range *new_range1;
291 struct lockf_range *new_range2;
296 struct lockf_range_list deadlist;
304 * Preallocate two ranges so we don't have to worry about blocking
305 * in the middle of the lock code.
307 if (new_range1 == NULL)
308 new_range1 = lf_alloc_range();
309 if (new_range2 == NULL)
310 new_range2 = lf_alloc_range();
319 * Locate the insertion point for the new lock (the first range
320 * with an lf_start >= start).
322 * Locate the first and latch ranges owned by us that overlap
323 * the requested range.
325 TAILQ_FOREACH(range, &lock->lf_range, lf_link) {
326 if (insert_point == NULL && range->lf_start >= start)
327 insert_point = range;
330 * Skip non-overlapping locks. Locks are sorted by lf_start
331 * So we can terminate the search when lf_start exceeds the
332 * requested range (insert_point is still guarenteed to be
335 if (range->lf_end < start)
337 if (range->lf_start > end) {
343 * Overlapping lock. Set first_match and last_match if we
346 if (range->lf_owner == owner) {
347 if (first_match == NULL)
354 * If we aren't the owner check for a conflicting lock. Only
357 if (type != F_UNLCK) {
358 if (type == F_WRLCK || range->lf_type == F_WRLCK)
364 * If a conflicting lock was observed, block or fail as appropriate.
365 * (this code is skipped when unlocking)
368 if ((flags & F_WAIT) == 0) {
374 * We are blocked. For POSIX locks we have to check
375 * for deadlocks and return with EDEADLK. This is done
376 * by checking whether range->lf_owner is already
379 * Since flock-style locks cover the whole file, a
380 * deadlock between those is nearly impossible.
381 * This can only occur if a process tries to lock the
382 * same inode exclusively while holding a shared lock
383 * with another descriptor.
384 * XXX How can we cleanly detect this?
385 * XXX The current mixing of flock & fcntl/lockf is evil.
387 * Handle existing locks of flock-style like POSIX locks.
389 if (flags & F_POSIX) {
390 TAILQ_FOREACH(brange, &lock->lf_blocked, lf_link)
391 if (brange->lf_owner == range->lf_owner) {
398 * For flock-style locks, we must first remove
399 * any shared locks that we hold before we sleep
400 * waiting for an exclusive lock.
402 if ((flags & F_POSIX) == 0 && type == F_WRLCK)
403 lf_setlock(lock, owner, F_UNLCK, 0, start, end);
407 lf_create_range(brange, owner, type, 0, start, end);
408 TAILQ_INSERT_TAIL(&lock->lf_blocked, brange, lf_link);
409 error = tsleep(brange, PCATCH, "lockf", 0);
412 * We may have been awaked by a signal and/or by a
413 * debugger continuing us (in which case we must remove
414 * ourselves from the blocked list) and/or by another
415 * process releasing/downgrading a lock (in which case
416 * we have already been removed from the blocked list
417 * and our lf_flags field is 1).
419 * Sleep if it looks like we might be livelocking.
421 if (brange->lf_flags == 0)
422 TAILQ_REMOVE(&lock->lf_blocked, brange, lf_link);
424 tsleep(brange, 0, "lockfz", 2);
427 lf_destroy_range(brange);
435 * If there are no overlapping locks owned by us then creating
436 * the new lock is easy. This is the most common case.
438 if (first_match == NULL) {
441 if (flags & F_POSIX) {
442 if (lf_count_change(owner, 1)) {
449 lf_create_range(range, owner, type, flags, start, end);
450 lf_insert(&lock->lf_range, range, insert_point);
455 * This is a special case that we need to check for in a couple
458 if (first_match == last_match && first_match->lf_start < start &&
459 last_match->lf_end > end) {
466 * Figure out the worst case net increase in POSIX locks and account
467 * for it now before we start modifying things. If neither the
468 * first or last locks match we have an issue. If there is only
469 * one overlapping range which needs to be clipped on both ends
470 * we wind up having to create up to two new locks, else only one.
472 * When unlocking the worst case is always 1 new lock if our
473 * unlock request cuts the middle out of an existing lock range.
475 * count represents the 'cleanup' adjustment needed. It starts
476 * negative, is incremented whenever we create a new POSIX lock,
477 * and decremented whenever we delete an existing one. At the
478 * end of the day it had better be <= 0 or we didn't calculate the
479 * worse case properly here.
482 if (flags & F_POSIX) {
483 if (!lf_match(first_match, type, flags) &&
484 !lf_match(last_match, type, flags)
486 if (double_clip && type != F_UNLCK)
491 if (count && lf_count_change(owner, -count)) {
496 /* else flock style lock which encompasses entire range */
499 * Create and insert the lock represented the requested range.
500 * Adjust the net POSIX lock count. We have to move our insertion
501 * point since brange now represents the first record >= start.
503 * When unlocking, no new lock is inserted but we still clip.
505 if (type != F_UNLCK) {
508 lf_create_range(brange, owner, type, flags, start, end);
509 lf_insert(&lock->lf_range, brange, insert_point);
510 insert_point = brange;
518 * Handle the double_clip case. This is the only case where
519 * we wind up having to add TWO locks.
522 KKASSERT(first_match == last_match);
523 last_match = new_range2;
525 lf_create_range(last_match, first_match->lf_owner,
526 first_match->lf_type, first_match->lf_flags,
527 end + 1, first_match->lf_end);
528 first_match->lf_end = start - 1;
529 first_match->lf_flags &= ~F_NOEND;
532 * Figure out where to insert the right side clip.
534 lf_insert(&lock->lf_range, last_match, first_match);
535 if (last_match->lf_flags & F_POSIX)
540 * Clip or destroy the locks between first_match and last_match,
541 * inclusive. Ignore the primary lock we created (brange). Note
542 * that if double-clipped, first_match and last_match will be
543 * outside our clipping range. Otherwise first_match and last_match
546 * We have already taken care of any double clipping.
548 * The insert_point may become invalid as we delete records, do not
549 * use that pointer any more. Also, when removing something other
550 * then 'range' we have to check to see if the item we are removing
551 * is 'next' and adjust 'next' properly.
553 * NOTE: brange will be NULL if F_UNLCKing.
555 TAILQ_INIT(&deadlist);
558 while ((range = next) != NULL) {
559 next = TAILQ_NEXT(range, lf_link);
562 * Ignore elements that we do not own and ignore the
563 * primary request range which we just created.
565 if (range->lf_owner != owner || range == brange)
569 * We may have to wakeup a waiter when downgrading a lock.
573 if (type == F_RDLCK && range->lf_type == F_WRLCK)
577 * Clip left. This can only occur on first_match.
579 * Merge the left clip with brange if possible. This must
580 * be done specifically, not in the optimized merge heuristic
581 * below, since we may have counted on it in our 'count'
584 if (range->lf_start < start) {
585 KKASSERT(range == first_match);
587 range->lf_end >= start - 1 &&
588 lf_match(range, type, flags)) {
589 range->lf_end = brange->lf_end;
590 range->lf_flags |= brange->lf_flags & F_NOEND;
592 * Removing something other then 'range',
593 * adjust 'next' if necessary.
596 next = TAILQ_NEXT(next, lf_link);
597 TAILQ_REMOVE(&lock->lf_range, brange, lf_link);
598 if (brange->lf_flags & F_POSIX)
600 TAILQ_INSERT_TAIL(&deadlist, brange, lf_link);
602 } else if (range->lf_end >= start) {
603 range->lf_end = start - 1;
605 range->lf_flags &= ~F_NOEND;
607 if (range == last_match)
613 * Clip right. This can only occur on last_match.
615 * Merge the right clip if possible. This must be done
616 * specifically, not in the optimized merge heuristic
617 * below, since we may have counted on it in our 'count'
620 * Since we are adjusting lf_start, we have to move the
621 * record to maintain the sorted list. Since lf_start is
622 * only getting larger we can use the next element as the
623 * insert point (we don't have to backtrack).
625 if (range->lf_end > end) {
626 KKASSERT(range == last_match);
628 range->lf_start <= end + 1 &&
629 lf_match(range, type, flags)) {
630 brange->lf_end = range->lf_end;
631 brange->lf_flags |= range->lf_flags & F_NOEND;
632 TAILQ_REMOVE(&lock->lf_range, range, lf_link);
633 if (range->lf_flags & F_POSIX)
635 TAILQ_INSERT_TAIL(&deadlist, range, lf_link);
636 } else if (range->lf_start <= end) {
637 range->lf_start = end + 1;
638 TAILQ_REMOVE(&lock->lf_range, range, lf_link);
639 lf_insert(&lock->lf_range, range, next);
641 /* range == last_match, we are done */
646 * The record must be entirely enclosed. Note that the
647 * record could be first_match or last_match, and will be
650 KKASSERT(range->lf_start >= start && range->lf_end <= end);
651 TAILQ_REMOVE(&lock->lf_range, range, lf_link);
652 if (range->lf_flags & F_POSIX)
654 TAILQ_INSERT_TAIL(&deadlist, range, lf_link);
655 if (range == last_match)
660 * Attempt to merge locks adjacent to brange. For example, we may
661 * have had to clip first_match and/or last_match, and they might
662 * be adjacent. Or there might simply have been an adjacent lock
665 * Don't get fancy, just check adjacent elements in the list if they
666 * happen to be owned by us.
668 * This case only gets hit if we have a situation where a shared
669 * and exclusive lock are adjacent, and the exclusive lock is
670 * downgraded to shared or the shared lock is upgraded to exclusive.
673 range = TAILQ_PREV(brange, lockf_range_list, lf_link);
675 range->lf_owner == owner &&
676 range->lf_end == brange->lf_start - 1 &&
677 lf_match(range, type, flags)
680 * Extend range to cover brange and scrap brange.
682 range->lf_end = brange->lf_end;
683 range->lf_flags |= brange->lf_flags & F_NOEND;
684 TAILQ_REMOVE(&lock->lf_range, brange, lf_link);
685 if (brange->lf_flags & F_POSIX)
687 TAILQ_INSERT_TAIL(&deadlist, brange, lf_link);
690 range = TAILQ_NEXT(brange, lf_link);
692 range->lf_owner == owner &&
693 range->lf_start == brange->lf_end + 1 &&
694 lf_match(range, type, flags)
697 * Extend brange to cover range and scrap range.
699 brange->lf_end = range->lf_end;
700 brange->lf_flags |= range->lf_flags & F_NOEND;
701 TAILQ_REMOVE(&lock->lf_range, range, lf_link);
702 if (range->lf_flags & F_POSIX)
704 TAILQ_INSERT_TAIL(&deadlist, range, lf_link);
709 * Destroy deleted elements. We didn't want to do it in the loop
710 * because the free() might have blocked.
712 * Adjust the count for any posix locks we thought we might create
715 while ((range = TAILQ_FIRST(&deadlist)) != NULL) {
716 TAILQ_REMOVE(&deadlist, range, lf_link);
717 lf_destroy_range(range);
720 KKASSERT(count <= 0);
722 lf_count_change(owner, count);
726 lf_wakeup(lock, start, end);
729 if (new_range1 != NULL)
730 lf_destroy_range(new_range1);
731 if (new_range2 != NULL)
732 lf_destroy_range(new_range2);
737 * Check whether there is a blocking lock,
738 * and if so return its process identifier.
741 lf_getlock(struct flock *fl, struct lockf *lock, struct proc *owner,
742 int type, int flags, off_t start, off_t end)
744 struct lockf_range *range;
746 TAILQ_FOREACH(range, &lock->lf_range, lf_link)
747 if (range->lf_owner != owner &&
748 lf_overlap(range, start, end) &&
749 (type == F_WRLCK || range->lf_type == F_WRLCK))
752 fl->l_type = F_UNLCK;
755 fl->l_type = range->lf_type;
756 fl->l_whence = SEEK_SET;
757 fl->l_start = range->lf_start;
758 if (range->lf_flags & F_NOEND)
761 fl->l_len = range->lf_end - range->lf_start + 1;
762 if (range->lf_owner != NULL && (range->lf_flags & F_POSIX))
763 fl->l_pid = range->lf_owner->p_pid;
770 * Wakeup pending lock attempts. Theoretically we can stop as soon as
771 * we encounter an exclusive request that covers the whole range (at least
772 * insofar as the sleep code above calls lf_wakeup() if it would otherwise
773 * exit instead of loop), but for now just wakeup all overlapping
777 lf_wakeup(struct lockf *lock, off_t start, off_t end)
779 struct lockf_range *range, *nrange;
781 TAILQ_FOREACH_MUTABLE(range, &lock->lf_blocked, lf_link, nrange) {
782 if (lf_overlap(range, start, end) == 0)
784 TAILQ_REMOVE(&lock->lf_blocked, range, lf_link);
791 * Allocate a range structure and initialize it sufficiently such that
792 * lf_destroy_range() does not barf.
794 static struct lockf_range *
797 struct lockf_range *range;
802 range = kmalloc(sizeof(struct lockf_range), M_LOCKF, M_WAITOK);
803 range->lf_owner = NULL;
808 lf_insert(struct lockf_range_list *list, struct lockf_range *elm,
809 struct lockf_range *insert_point)
811 while (insert_point && insert_point->lf_start < elm->lf_start)
812 insert_point = TAILQ_NEXT(insert_point, lf_link);
813 if (insert_point != NULL)
814 TAILQ_INSERT_BEFORE(insert_point, elm, lf_link);
816 TAILQ_INSERT_TAIL(list, elm, lf_link);
820 lf_create_range(struct lockf_range *range, struct proc *owner, int type,
821 int flags, off_t start, off_t end)
823 KKASSERT(start <= end);
824 range->lf_type = type;
825 range->lf_flags = flags;
826 range->lf_start = start;
828 range->lf_owner = owner;
830 lf_printf("lf_create_range: %lld..%lld\n",
831 range->lf_start, range->lf_end);
835 lf_destroy_range(struct lockf_range *range)
837 lf_printf("lf_destroy_range: %lld..%lld\n",
838 range->lf_start, range->lf_end);
839 kfree(range, M_LOCKF);
842 KKASSERT(lf_global_counter>=0);
849 _lf_printf(const char *ctl, ...)
854 if (lf_print_ranges) {
855 if ((p = curproc) != NULL)
856 kprintf("pid %d (%s): ", p->p_pid, p->p_comm);
864 _lf_print_lock(const struct lockf *lock)
866 struct lockf_range *range;
868 if (lf_print_ranges == 0)
871 if (TAILQ_EMPTY(&lock->lf_range)) {
872 lf_printf("lockf %p: no ranges locked\n", lock);
874 lf_printf("lockf %p:\n", lock);
876 TAILQ_FOREACH(range, &lock->lf_range, lf_link)
877 kprintf("\t%lld..%lld type %s owned by %d\n",
878 range->lf_start, range->lf_end,
879 range->lf_type == F_RDLCK ? "shared" : "exclusive",
880 range->lf_flags & F_POSIX ? range->lf_owner->p_pid : -1);
881 if (TAILQ_EMPTY(&lock->lf_blocked))
882 kprintf("no process waiting for range\n");
884 kprintf("blocked locks:");
885 TAILQ_FOREACH(range, &lock->lf_blocked, lf_link)
886 kprintf("\t%lld..%lld type %s waiting on %p\n",
887 range->lf_start, range->lf_end,
888 range->lf_type == F_RDLCK ? "shared" : "exclusive",
891 #endif /* LOCKF_DEBUG */