2 * Copyright (c) 2008 Isilon Inc http://www.isilon.com/
3 * Authors: Doug Rabson <dfr@rabson.org>
4 * Developed with Red Inc: Alfred Perlstein <alfred@freebsd.org>
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 * Copyright (c) 1982, 1986, 1989, 1993
29 * The Regents of the University of California. All rights reserved.
31 * This code is derived from software contributed to Berkeley by
32 * Scooter Morris at Genentech Inc.
34 * Redistribution and use in source and binary forms, with or without
35 * modification, are permitted provided that the following conditions
37 * 1. Redistributions of source code must retain the above copyright
38 * notice, this list of conditions and the following disclaimer.
39 * 2. Redistributions in binary form must reproduce the above copyright
40 * notice, this list of conditions and the following disclaimer in the
41 * documentation and/or other materials provided with the distribution.
42 * 4. Neither the name of the University nor the names of its contributors
43 * may be used to endorse or promote products derived from this software
44 * without specific prior written permission.
46 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
47 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
48 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
49 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
50 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
51 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
52 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
53 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
54 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
55 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
58 * @(#)ufs_lockf.c 8.3 (Berkeley) 1/6/94
61 #include <sys/cdefs.h>
62 __FBSDID("$FreeBSD$");
64 #include "opt_debug_lockf.h"
66 #include <sys/param.h>
67 #include <sys/systm.h>
69 #include <sys/kernel.h>
70 #include <sys/limits.h>
72 #include <sys/mount.h>
73 #include <sys/mutex.h>
76 #include <sys/unistd.h>
77 #include <sys/vnode.h>
78 #include <sys/malloc.h>
79 #include <sys/fcntl.h>
80 #include <sys/lockf.h>
81 #include <sys/taskqueue.h>
84 #include <sys/sysctl.h>
86 #include <ufs/ufs/quota.h>
87 #include <ufs/ufs/inode.h>
89 static int lockf_debug = 0; /* control debug output */
90 SYSCTL_INT(_debug, OID_AUTO, lockf_debug, CTLFLAG_RW, &lockf_debug, 0, "");
93 MALLOC_DEFINE(M_LOCKF, "lockf", "Byte-range locking structures");
97 struct owner_vertex_list;
100 #define NOLOCKF (struct lockf_entry *)0
103 static void lf_init(void *);
104 static int lf_hash_owner(caddr_t, struct flock *, int);
105 static int lf_owner_matches(struct lock_owner *, caddr_t, struct flock *,
107 static struct lockf_entry *
108 lf_alloc_lock(struct lock_owner *);
109 static int lf_free_lock(struct lockf_entry *);
110 static int lf_clearlock(struct lockf *, struct lockf_entry *);
111 static int lf_overlaps(struct lockf_entry *, struct lockf_entry *);
112 static int lf_blocks(struct lockf_entry *, struct lockf_entry *);
113 static void lf_free_edge(struct lockf_edge *);
114 static struct lockf_edge *
116 static void lf_alloc_vertex(struct lockf_entry *);
117 static int lf_add_edge(struct lockf_entry *, struct lockf_entry *);
118 static void lf_remove_edge(struct lockf_edge *);
119 static void lf_remove_outgoing(struct lockf_entry *);
120 static void lf_remove_incoming(struct lockf_entry *);
121 static int lf_add_outgoing(struct lockf *, struct lockf_entry *);
122 static int lf_add_incoming(struct lockf *, struct lockf_entry *);
123 static int lf_findoverlap(struct lockf_entry **, struct lockf_entry *,
125 static struct lockf_entry *
126 lf_getblock(struct lockf *, struct lockf_entry *);
127 static int lf_getlock(struct lockf *, struct lockf_entry *, struct flock *);
128 static void lf_insert_lock(struct lockf *, struct lockf_entry *);
129 static void lf_wakeup_lock(struct lockf *, struct lockf_entry *);
130 static void lf_update_dependancies(struct lockf *, struct lockf_entry *,
131 int all, struct lockf_entry_list *);
132 static void lf_set_start(struct lockf *, struct lockf_entry *, off_t,
133 struct lockf_entry_list*);
134 static void lf_set_end(struct lockf *, struct lockf_entry *, off_t,
135 struct lockf_entry_list*);
136 static int lf_setlock(struct lockf *, struct lockf_entry *,
137 struct vnode *, void **cookiep);
138 static int lf_cancel(struct lockf *, struct lockf_entry *, void *);
139 static void lf_split(struct lockf *, struct lockf_entry *,
140 struct lockf_entry *, struct lockf_entry_list *);
142 static int graph_reaches(struct owner_vertex *x, struct owner_vertex *y,
143 struct owner_vertex_list *path);
144 static void graph_check(struct owner_graph *g, int checkorder);
145 static void graph_print_vertices(struct owner_vertex_list *set);
147 static int graph_delta_forward(struct owner_graph *g,
148 struct owner_vertex *x, struct owner_vertex *y,
149 struct owner_vertex_list *delta);
150 static int graph_delta_backward(struct owner_graph *g,
151 struct owner_vertex *x, struct owner_vertex *y,
152 struct owner_vertex_list *delta);
153 static int graph_add_indices(int *indices, int n,
154 struct owner_vertex_list *set);
155 static int graph_assign_indices(struct owner_graph *g, int *indices,
156 int nextunused, struct owner_vertex_list *set);
157 static int graph_add_edge(struct owner_graph *g,
158 struct owner_vertex *x, struct owner_vertex *y);
159 static void graph_remove_edge(struct owner_graph *g,
160 struct owner_vertex *x, struct owner_vertex *y);
161 static struct owner_vertex *graph_alloc_vertex(struct owner_graph *g,
162 struct lock_owner *lo);
163 static void graph_free_vertex(struct owner_graph *g,
164 struct owner_vertex *v);
165 static struct owner_graph * graph_init(struct owner_graph *g);
167 static void lf_print(char *, struct lockf_entry *);
168 static void lf_printlist(char *, struct lockf_entry *);
169 static void lf_print_owner(struct lock_owner *);
173 * This structure is used to keep track of both local and remote lock
174 * owners. The lf_owner field of the struct lockf_entry points back at
175 * the lock owner structure. Each possible lock owner (local proc for
176 * POSIX fcntl locks, local file for BSD flock locks or <pid,sysid>
177 * pair for remote locks) is represented by a unique instance of
180 * If a lock owner has a lock that blocks some other lock or a lock
181 * that is waiting for some other lock, it also has a vertex in the
185 * (s) locked by state->ls_lock
186 * (S) locked by lf_lock_states_lock
187 * (l) locked by lf_lock_owners_lock
188 * (g) locked by lf_owner_graph_lock
189 * (c) const until freeing
191 #define LOCK_OWNER_HASH_SIZE 256
194 LIST_ENTRY(lock_owner) lo_link; /* (l) hash chain */
195 int lo_refs; /* (l) Number of locks referring to this */
196 int lo_flags; /* (c) Flags passwd to lf_advlock */
197 caddr_t lo_id; /* (c) Id value passed to lf_advlock */
198 pid_t lo_pid; /* (c) Process Id of the lock owner */
199 int lo_sysid; /* (c) System Id of the lock owner */
200 struct owner_vertex *lo_vertex; /* (g) entry in deadlock graph */
203 LIST_HEAD(lock_owner_list, lock_owner);
205 static struct sx lf_lock_states_lock;
206 static struct lockf_list lf_lock_states; /* (S) */
207 static struct sx lf_lock_owners_lock;
208 static struct lock_owner_list lf_lock_owners[LOCK_OWNER_HASH_SIZE]; /* (l) */
211 * Structures for deadlock detection.
213 * We have two types of directed graph, the first is the set of locks,
214 * both active and pending on a vnode. Within this graph, active locks
215 * are terminal nodes in the graph (i.e. have no out-going
216 * edges). Pending locks have out-going edges to each blocking active
217 * lock that prevents the lock from being granted and also to each
218 * older pending lock that would block them if it was active. The
219 * graph for each vnode is naturally acyclic; new edges are only ever
220 * added to or from new nodes (either new pending locks which only add
221 * out-going edges or new active locks which only add in-coming edges)
222 * therefore they cannot create loops in the lock graph.
224 * The second graph is a global graph of lock owners. Each lock owner
225 * is a vertex in that graph and an edge is added to the graph
226 * whenever an edge is added to a vnode graph, with end points
227 * corresponding to owner of the new pending lock and the owner of the
228 * lock upon which it waits. In order to prevent deadlock, we only add
229 * an edge to this graph if the new edge would not create a cycle.
231 * The lock owner graph is topologically sorted, i.e. if a node has
232 * any outgoing edges, then it has an order strictly less than any
233 * node to which it has an outgoing edge. We preserve this ordering
234 * (and detect cycles) on edge insertion using Algorithm PK from the
235 * paper "A Dynamic Topological Sort Algorithm for Directed Acyclic
236 * Graphs" (ACM Journal of Experimental Algorithms, Vol 11, Article
242 LIST_ENTRY(owner_edge) e_outlink; /* (g) link from's out-edge list */
243 LIST_ENTRY(owner_edge) e_inlink; /* (g) link to's in-edge list */
244 int e_refs; /* (g) number of times added */
245 struct owner_vertex *e_from; /* (c) out-going from here */
246 struct owner_vertex *e_to; /* (c) in-coming to here */
248 LIST_HEAD(owner_edge_list, owner_edge);
250 struct owner_vertex {
251 TAILQ_ENTRY(owner_vertex) v_link; /* (g) workspace for edge insertion */
252 uint32_t v_gen; /* (g) workspace for edge insertion */
253 int v_order; /* (g) order of vertex in graph */
254 struct owner_edge_list v_outedges;/* (g) list of out-edges */
255 struct owner_edge_list v_inedges; /* (g) list of in-edges */
256 struct lock_owner *v_owner; /* (c) corresponding lock owner */
258 TAILQ_HEAD(owner_vertex_list, owner_vertex);
261 struct owner_vertex** g_vertices; /* (g) pointers to vertices */
262 int g_size; /* (g) number of vertices */
263 int g_space; /* (g) space allocated for vertices */
264 int *g_indexbuf; /* (g) workspace for loop detection */
265 uint32_t g_gen; /* (g) increment when re-ordering */
268 static struct sx lf_owner_graph_lock;
269 static struct owner_graph lf_owner_graph;
272 * Initialise various structures and locks.
279 sx_init(&lf_lock_states_lock, "lock states lock");
280 LIST_INIT(&lf_lock_states);
282 sx_init(&lf_lock_owners_lock, "lock owners lock");
283 for (i = 0; i < LOCK_OWNER_HASH_SIZE; i++)
284 LIST_INIT(&lf_lock_owners[i]);
286 sx_init(&lf_owner_graph_lock, "owner graph lock");
287 graph_init(&lf_owner_graph);
289 SYSINIT(lf_init, SI_SUB_LOCK, SI_ORDER_FIRST, lf_init, NULL);
292 * Generate a hash value for a lock owner.
295 lf_hash_owner(caddr_t id, struct flock *fl, int flags)
299 if (flags & F_REMOTE) {
300 h = HASHSTEP(0, fl->l_pid);
301 h = HASHSTEP(h, fl->l_sysid);
302 } else if (flags & F_FLOCK) {
303 h = ((uintptr_t) id) >> 7;
305 struct proc *p = (struct proc *) id;
306 h = HASHSTEP(0, p->p_pid);
310 return (h % LOCK_OWNER_HASH_SIZE);
314 * Return true if a lock owner matches the details passed to
318 lf_owner_matches(struct lock_owner *lo, caddr_t id, struct flock *fl,
321 if (flags & F_REMOTE) {
322 return lo->lo_pid == fl->l_pid
323 && lo->lo_sysid == fl->l_sysid;
325 return lo->lo_id == id;
329 static struct lockf_entry *
330 lf_alloc_lock(struct lock_owner *lo)
332 struct lockf_entry *lf;
334 lf = malloc(sizeof(struct lockf_entry), M_LOCKF, M_WAITOK|M_ZERO);
338 printf("Allocated lock %p\n", lf);
341 sx_xlock(&lf_lock_owners_lock);
343 sx_xunlock(&lf_lock_owners_lock);
351 lf_free_lock(struct lockf_entry *lock)
354 KASSERT(lock->lf_refs > 0, ("lockf_entry negative ref count %p", lock));
355 if (--lock->lf_refs > 0)
358 * Adjust the lock_owner reference count and
359 * reclaim the entry if this is the last lock
362 struct lock_owner *lo = lock->lf_owner;
364 KASSERT(LIST_EMPTY(&lock->lf_outedges),
365 ("freeing lock with dependancies"));
366 KASSERT(LIST_EMPTY(&lock->lf_inedges),
367 ("freeing lock with dependants"));
368 sx_xlock(&lf_lock_owners_lock);
369 KASSERT(lo->lo_refs > 0, ("lock owner refcount"));
371 if (lo->lo_refs == 0) {
374 printf("lf_free_lock: freeing lock owner %p\n",
378 sx_xlock(&lf_owner_graph_lock);
379 graph_free_vertex(&lf_owner_graph,
381 sx_xunlock(&lf_owner_graph_lock);
383 LIST_REMOVE(lo, lo_link);
387 printf("Freed lock owner %p\n", lo);
390 sx_unlock(&lf_lock_owners_lock);
392 if ((lock->lf_flags & F_REMOTE) && lock->lf_vnode) {
393 vrele(lock->lf_vnode);
394 lock->lf_vnode = NULL;
398 printf("Freed lock %p\n", lock);
405 * Advisory record locking support
408 lf_advlockasync(struct vop_advlockasync_args *ap, struct lockf **statep,
411 struct lockf *state, *freestate = NULL;
412 struct flock *fl = ap->a_fl;
413 struct lockf_entry *lock;
414 struct vnode *vp = ap->a_vp;
415 caddr_t id = ap->a_id;
416 int flags = ap->a_flags;
418 struct lock_owner *lo;
419 off_t start, end, oadd;
423 * Handle the F_UNLKSYS case first - no need to mess about
424 * creating a lock owner for this one.
426 if (ap->a_op == F_UNLCKSYS) {
427 lf_clearremotesys(fl->l_sysid);
432 * Convert the flock structure into a start and end.
434 switch (fl->l_whence) {
439 * Caller is responsible for adding any necessary offset
440 * when SEEK_CUR is used.
446 if (size > OFF_MAX ||
447 (fl->l_start > 0 && size > OFF_MAX - fl->l_start))
449 start = size + fl->l_start;
464 } else if (fl->l_len == 0) {
467 oadd = fl->l_len - 1;
468 if (oadd > OFF_MAX - start)
473 * Avoid the common case of unlocking when inode has no locks.
476 if ((*statep) == NULL) {
477 if (ap->a_op != F_SETLK) {
478 fl->l_type = F_UNLCK;
486 * Map our arguments to an existing lock owner or create one
487 * if this is the first time we have seen this owner.
489 hash = lf_hash_owner(id, fl, flags);
490 sx_xlock(&lf_lock_owners_lock);
491 LIST_FOREACH(lo, &lf_lock_owners[hash], lo_link)
492 if (lf_owner_matches(lo, id, fl, flags))
496 * We initialise the lock with a reference
497 * count which matches the new lockf_entry
498 * structure created below.
500 lo = malloc(sizeof(struct lock_owner), M_LOCKF,
504 printf("Allocated lock owner %p\n", lo);
508 lo->lo_flags = flags;
510 if (flags & F_REMOTE) {
511 lo->lo_pid = fl->l_pid;
512 lo->lo_sysid = fl->l_sysid;
513 } else if (flags & F_FLOCK) {
517 struct proc *p = (struct proc *) id;
518 lo->lo_pid = p->p_pid;
521 lo->lo_vertex = NULL;
524 if (lockf_debug & 1) {
525 printf("lf_advlockasync: new lock owner %p ", lo);
531 LIST_INSERT_HEAD(&lf_lock_owners[hash], lo, lo_link);
534 * We have seen this lock owner before, increase its
535 * reference count to account for the new lockf_entry
536 * structure we create below.
540 sx_xunlock(&lf_lock_owners_lock);
543 * Create the lockf structure. We initialise the lf_owner
544 * field here instead of in lf_alloc_lock() to avoid paying
545 * the lf_lock_owners_lock tax twice.
547 lock = lf_alloc_lock(NULL);
549 lock->lf_start = start;
553 if (flags & F_REMOTE) {
555 * For remote locks, the caller may release its ref to
556 * the vnode at any time - we have to ref it here to
557 * prevent it from being recycled unexpectedly.
563 * XXX The problem is that VTOI is ufs specific, so it will
564 * break LOCKF_DEBUG for all other FS's other than UFS because
565 * it casts the vnode->data ptr to struct inode *.
567 /* lock->lf_inode = VTOI(ap->a_vp); */
568 lock->lf_inode = (struct inode *)0;
569 lock->lf_type = fl->l_type;
570 LIST_INIT(&lock->lf_outedges);
571 LIST_INIT(&lock->lf_inedges);
572 lock->lf_async_task = ap->a_task;
573 lock->lf_flags = ap->a_flags;
576 * Do the requested operation. First find our state structure
577 * and create a new one if necessary - the caller's *statep
578 * variable and the state's ls_threads count is protected by
579 * the vnode interlock.
582 if (vp->v_iflag & VI_DOOMED) {
589 * Allocate a state structure if necessary.
597 ls = malloc(sizeof(struct lockf), M_LOCKF, M_WAITOK|M_ZERO);
598 sx_init(&ls->ls_lock, "ls_lock");
599 LIST_INIT(&ls->ls_active);
600 LIST_INIT(&ls->ls_pending);
603 sx_xlock(&lf_lock_states_lock);
604 LIST_INSERT_HEAD(&lf_lock_states, ls, ls_link);
605 sx_xunlock(&lf_lock_states_lock);
608 * Cope if we lost a race with some other thread while
609 * trying to allocate memory.
612 if (vp->v_iflag & VI_DOOMED) {
614 sx_xlock(&lf_lock_states_lock);
615 LIST_REMOVE(ls, ls_link);
616 sx_xunlock(&lf_lock_states_lock);
617 sx_destroy(&ls->ls_lock);
622 if ((*statep) == NULL) {
623 state = *statep = ls;
630 sx_xlock(&lf_lock_states_lock);
631 LIST_REMOVE(ls, ls_link);
632 sx_xunlock(&lf_lock_states_lock);
633 sx_destroy(&ls->ls_lock);
641 sx_xlock(&state->ls_lock);
643 * Recheck the doomed vnode after state->ls_lock is
644 * locked. lf_purgelocks() requires that no new threads add
645 * pending locks when vnode is marked by VI_DOOMED flag.
648 if (vp->v_iflag & VI_DOOMED) {
652 sx_xunlock(&state->ls_lock);
660 error = lf_setlock(state, lock, vp, ap->a_cookiep);
664 error = lf_clearlock(state, lock);
669 error = lf_getlock(state, lock, fl);
675 error = lf_cancel(state, lock, *ap->a_cookiep);
689 * Check for some can't happen stuff. In this case, the active
690 * lock list becoming disordered or containing mutually
691 * blocking locks. We also check the pending list for locks
692 * which should be active (i.e. have no out-going edges).
694 LIST_FOREACH(lock, &state->ls_active, lf_link) {
695 struct lockf_entry *lf;
696 if (LIST_NEXT(lock, lf_link))
697 KASSERT((lock->lf_start
698 <= LIST_NEXT(lock, lf_link)->lf_start),
699 ("locks disordered"));
700 LIST_FOREACH(lf, &state->ls_active, lf_link) {
703 KASSERT(!lf_blocks(lock, lf),
704 ("two conflicting active locks"));
705 if (lock->lf_owner == lf->lf_owner)
706 KASSERT(!lf_overlaps(lock, lf),
707 ("two overlapping locks from same owner"));
710 LIST_FOREACH(lock, &state->ls_pending, lf_link) {
711 KASSERT(!LIST_EMPTY(&lock->lf_outedges),
712 ("pending lock which should be active"));
715 sx_xunlock(&state->ls_lock);
718 * If we have removed the last active lock on the vnode and
719 * this is the last thread that was in-progress, we can free
720 * the state structure. We update the caller's pointer inside
721 * the vnode interlock but call free outside.
723 * XXX alternatively, keep the state structure around until
724 * the filesystem recycles - requires a callback from the
731 if (LIST_EMPTY(&state->ls_active) && state->ls_threads == 0) {
732 KASSERT(LIST_EMPTY(&state->ls_pending),
733 ("freeing state with pending locks"));
741 sx_xlock(&lf_lock_states_lock);
742 LIST_REMOVE(freestate, ls_link);
743 sx_xunlock(&lf_lock_states_lock);
744 sx_destroy(&freestate->ls_lock);
745 free(freestate, M_LOCKF);
751 lf_advlock(struct vop_advlock_args *ap, struct lockf **statep, u_quad_t size)
753 struct vop_advlockasync_args a;
759 a.a_flags = ap->a_flags;
763 return (lf_advlockasync(&a, statep, size));
767 lf_purgelocks(struct vnode *vp, struct lockf **statep)
770 struct lockf_entry *lock, *nlock;
773 * For this to work correctly, the caller must ensure that no
774 * other threads enter the locking system for this vnode,
775 * e.g. by checking VI_DOOMED. We wake up any threads that are
776 * sleeping waiting for locks on this vnode and then free all
777 * the remaining locks.
780 KASSERT(vp->v_iflag & VI_DOOMED,
781 ("lf_purgelocks: vp %p has not vgone yet", vp));
788 sx_xlock(&state->ls_lock);
789 sx_xlock(&lf_owner_graph_lock);
790 LIST_FOREACH_SAFE(lock, &state->ls_pending, lf_link, nlock) {
791 LIST_REMOVE(lock, lf_link);
792 lf_remove_outgoing(lock);
793 lf_remove_incoming(lock);
796 * If its an async lock, we can just free it
797 * here, otherwise we let the sleeping thread
800 if (lock->lf_async_task) {
803 lock->lf_flags |= F_INTR;
807 sx_xunlock(&lf_owner_graph_lock);
808 sx_xunlock(&state->ls_lock);
811 * Wait for all other threads, sleeping and otherwise
815 while (state->ls_threads > 1)
816 msleep(state, VI_MTX(vp), 0, "purgelocks", 0);
820 * We can just free all the active locks since they
821 * will have no dependancies (we removed them all
822 * above). We don't need to bother locking since we
823 * are the last thread using this state structure.
825 KASSERT(LIST_EMPTY(&state->ls_pending),
826 ("lock pending for %p", state));
827 LIST_FOREACH_SAFE(lock, &state->ls_active, lf_link, nlock) {
828 LIST_REMOVE(lock, lf_link);
831 sx_xlock(&lf_lock_states_lock);
832 LIST_REMOVE(state, ls_link);
833 sx_xunlock(&lf_lock_states_lock);
834 sx_destroy(&state->ls_lock);
835 free(state, M_LOCKF);
842 * Return non-zero if locks 'x' and 'y' overlap.
845 lf_overlaps(struct lockf_entry *x, struct lockf_entry *y)
848 return (x->lf_start <= y->lf_end && x->lf_end >= y->lf_start);
852 * Return non-zero if lock 'x' is blocked by lock 'y' (or vice versa).
855 lf_blocks(struct lockf_entry *x, struct lockf_entry *y)
858 return x->lf_owner != y->lf_owner
859 && (x->lf_type == F_WRLCK || y->lf_type == F_WRLCK)
860 && lf_overlaps(x, y);
864 * Allocate a lock edge from the free list
866 static struct lockf_edge *
870 return (malloc(sizeof(struct lockf_edge), M_LOCKF, M_WAITOK|M_ZERO));
877 lf_free_edge(struct lockf_edge *e)
885 * Ensure that the lock's owner has a corresponding vertex in the
889 lf_alloc_vertex(struct lockf_entry *lock)
891 struct owner_graph *g = &lf_owner_graph;
893 if (!lock->lf_owner->lo_vertex)
894 lock->lf_owner->lo_vertex =
895 graph_alloc_vertex(g, lock->lf_owner);
899 * Attempt to record an edge from lock x to lock y. Return EDEADLK if
900 * the new edge would cause a cycle in the owner graph.
903 lf_add_edge(struct lockf_entry *x, struct lockf_entry *y)
905 struct owner_graph *g = &lf_owner_graph;
906 struct lockf_edge *e;
910 LIST_FOREACH(e, &x->lf_outedges, le_outlink)
911 KASSERT(e->le_to != y, ("adding lock edge twice"));
915 * Make sure the two owners have entries in the owner graph.
920 error = graph_add_edge(g, x->lf_owner->lo_vertex,
921 y->lf_owner->lo_vertex);
926 LIST_INSERT_HEAD(&x->lf_outedges, e, le_outlink);
927 LIST_INSERT_HEAD(&y->lf_inedges, e, le_inlink);
935 * Remove an edge from the lock graph.
938 lf_remove_edge(struct lockf_edge *e)
940 struct owner_graph *g = &lf_owner_graph;
941 struct lockf_entry *x = e->le_from;
942 struct lockf_entry *y = e->le_to;
944 graph_remove_edge(g, x->lf_owner->lo_vertex, y->lf_owner->lo_vertex);
945 LIST_REMOVE(e, le_outlink);
946 LIST_REMOVE(e, le_inlink);
953 * Remove all out-going edges from lock x.
956 lf_remove_outgoing(struct lockf_entry *x)
958 struct lockf_edge *e;
960 while ((e = LIST_FIRST(&x->lf_outedges)) != NULL) {
966 * Remove all in-coming edges from lock x.
969 lf_remove_incoming(struct lockf_entry *x)
971 struct lockf_edge *e;
973 while ((e = LIST_FIRST(&x->lf_inedges)) != NULL) {
979 * Walk the list of locks for the file and create an out-going edge
980 * from lock to each blocking lock.
983 lf_add_outgoing(struct lockf *state, struct lockf_entry *lock)
985 struct lockf_entry *overlap;
988 LIST_FOREACH(overlap, &state->ls_active, lf_link) {
990 * We may assume that the active list is sorted by
993 if (overlap->lf_start > lock->lf_end)
995 if (!lf_blocks(lock, overlap))
999 * We've found a blocking lock. Add the corresponding
1000 * edge to the graphs and see if it would cause a
1003 error = lf_add_edge(lock, overlap);
1006 * The only error that lf_add_edge returns is EDEADLK.
1007 * Remove any edges we added and return the error.
1010 lf_remove_outgoing(lock);
1016 * We also need to add edges to sleeping locks that block
1017 * us. This ensures that lf_wakeup_lock cannot grant two
1018 * mutually blocking locks simultaneously and also enforces a
1019 * 'first come, first served' fairness model. Note that this
1020 * only happens if we are blocked by at least one active lock
1021 * due to the call to lf_getblock in lf_setlock below.
1023 LIST_FOREACH(overlap, &state->ls_pending, lf_link) {
1024 if (!lf_blocks(lock, overlap))
1027 * We've found a blocking lock. Add the corresponding
1028 * edge to the graphs and see if it would cause a
1031 error = lf_add_edge(lock, overlap);
1034 * The only error that lf_add_edge returns is EDEADLK.
1035 * Remove any edges we added and return the error.
1038 lf_remove_outgoing(lock);
1047 * Walk the list of pending locks for the file and create an in-coming
1048 * edge from lock to each blocking lock.
1051 lf_add_incoming(struct lockf *state, struct lockf_entry *lock)
1053 struct lockf_entry *overlap;
1056 LIST_FOREACH(overlap, &state->ls_pending, lf_link) {
1057 if (!lf_blocks(lock, overlap))
1061 * We've found a blocking lock. Add the corresponding
1062 * edge to the graphs and see if it would cause a
1065 error = lf_add_edge(overlap, lock);
1068 * The only error that lf_add_edge returns is EDEADLK.
1069 * Remove any edges we added and return the error.
1072 lf_remove_incoming(lock);
1080 * Insert lock into the active list, keeping list entries ordered by
1081 * increasing values of lf_start.
1084 lf_insert_lock(struct lockf *state, struct lockf_entry *lock)
1086 struct lockf_entry *lf, *lfprev;
1088 if (LIST_EMPTY(&state->ls_active)) {
1089 LIST_INSERT_HEAD(&state->ls_active, lock, lf_link);
1094 LIST_FOREACH(lf, &state->ls_active, lf_link) {
1095 if (lf->lf_start > lock->lf_start) {
1096 LIST_INSERT_BEFORE(lf, lock, lf_link);
1101 LIST_INSERT_AFTER(lfprev, lock, lf_link);
1105 * Wake up a sleeping lock and remove it from the pending list now
1106 * that all its dependancies have been resolved. The caller should
1107 * arrange for the lock to be added to the active list, adjusting any
1108 * existing locks for the same owner as needed.
1111 lf_wakeup_lock(struct lockf *state, struct lockf_entry *wakelock)
1115 * Remove from ls_pending list and wake up the caller
1116 * or start the async notification, as appropriate.
1118 LIST_REMOVE(wakelock, lf_link);
1120 if (lockf_debug & 1)
1121 lf_print("lf_wakeup_lock: awakening", wakelock);
1122 #endif /* LOCKF_DEBUG */
1123 if (wakelock->lf_async_task) {
1124 taskqueue_enqueue(taskqueue_thread, wakelock->lf_async_task);
1131 * Re-check all dependant locks and remove edges to locks that we no
1132 * longer block. If 'all' is non-zero, the lock has been removed and
1133 * we must remove all the dependancies, otherwise it has simply been
1134 * reduced but remains active. Any pending locks which have been been
1135 * unblocked are added to 'granted'
1138 lf_update_dependancies(struct lockf *state, struct lockf_entry *lock, int all,
1139 struct lockf_entry_list *granted)
1141 struct lockf_edge *e, *ne;
1142 struct lockf_entry *deplock;
1144 LIST_FOREACH_SAFE(e, &lock->lf_inedges, le_inlink, ne) {
1145 deplock = e->le_from;
1146 if (all || !lf_blocks(lock, deplock)) {
1147 sx_xlock(&lf_owner_graph_lock);
1149 sx_xunlock(&lf_owner_graph_lock);
1150 if (LIST_EMPTY(&deplock->lf_outedges)) {
1151 lf_wakeup_lock(state, deplock);
1152 LIST_INSERT_HEAD(granted, deplock, lf_link);
1159 * Set the start of an existing active lock, updating dependancies and
1160 * adding any newly woken locks to 'granted'.
1163 lf_set_start(struct lockf *state, struct lockf_entry *lock, off_t new_start,
1164 struct lockf_entry_list *granted)
1167 KASSERT(new_start >= lock->lf_start, ("can't increase lock"));
1168 lock->lf_start = new_start;
1169 LIST_REMOVE(lock, lf_link);
1170 lf_insert_lock(state, lock);
1171 lf_update_dependancies(state, lock, FALSE, granted);
1175 * Set the end of an existing active lock, updating dependancies and
1176 * adding any newly woken locks to 'granted'.
1179 lf_set_end(struct lockf *state, struct lockf_entry *lock, off_t new_end,
1180 struct lockf_entry_list *granted)
1183 KASSERT(new_end <= lock->lf_end, ("can't increase lock"));
1184 lock->lf_end = new_end;
1185 lf_update_dependancies(state, lock, FALSE, granted);
1189 * Add a lock to the active list, updating or removing any current
1190 * locks owned by the same owner and processing any pending locks that
1191 * become unblocked as a result. This code is also used for unlock
1192 * since the logic for updating existing locks is identical.
1194 * As a result of processing the new lock, we may unblock existing
1195 * pending locks as a result of downgrading/unlocking. We simply
1196 * activate the newly granted locks by looping.
1198 * Since the new lock already has its dependancies set up, we always
1199 * add it to the list (unless its an unlock request). This may
1200 * fragment the lock list in some pathological cases but its probably
1201 * not a real problem.
1204 lf_activate_lock(struct lockf *state, struct lockf_entry *lock)
1206 struct lockf_entry *overlap, *lf;
1207 struct lockf_entry_list granted;
1210 LIST_INIT(&granted);
1211 LIST_INSERT_HEAD(&granted, lock, lf_link);
1213 while (!LIST_EMPTY(&granted)) {
1214 lock = LIST_FIRST(&granted);
1215 LIST_REMOVE(lock, lf_link);
1218 * Skip over locks owned by other processes. Handle
1219 * any locks that overlap and are owned by ourselves.
1221 overlap = LIST_FIRST(&state->ls_active);
1223 ovcase = lf_findoverlap(&overlap, lock, SELF);
1226 if (ovcase && (lockf_debug & 2)) {
1227 printf("lf_setlock: overlap %d", ovcase);
1228 lf_print("", overlap);
1234 * 1) overlap == lock
1235 * 2) overlap contains lock
1236 * 3) lock contains overlap
1237 * 4) overlap starts before lock
1238 * 5) overlap ends after lock
1241 case 0: /* no overlap */
1244 case 1: /* overlap == lock */
1246 * We have already setup the
1247 * dependants for the new lock, taking
1248 * into account a possible downgrade
1249 * or unlock. Remove the old lock.
1251 LIST_REMOVE(overlap, lf_link);
1252 lf_update_dependancies(state, overlap, TRUE,
1254 lf_free_lock(overlap);
1257 case 2: /* overlap contains lock */
1259 * Just split the existing lock.
1261 lf_split(state, overlap, lock, &granted);
1264 case 3: /* lock contains overlap */
1266 * Delete the overlap and advance to
1267 * the next entry in the list.
1269 lf = LIST_NEXT(overlap, lf_link);
1270 LIST_REMOVE(overlap, lf_link);
1271 lf_update_dependancies(state, overlap, TRUE,
1273 lf_free_lock(overlap);
1277 case 4: /* overlap starts before lock */
1279 * Just update the overlap end and
1282 lf_set_end(state, overlap, lock->lf_start - 1,
1284 overlap = LIST_NEXT(overlap, lf_link);
1287 case 5: /* overlap ends after lock */
1289 * Change the start of overlap and
1292 lf_set_start(state, overlap, lock->lf_end + 1,
1299 if (lockf_debug & 1) {
1300 if (lock->lf_type != F_UNLCK)
1301 lf_print("lf_activate_lock: activated", lock);
1303 lf_print("lf_activate_lock: unlocked", lock);
1304 lf_printlist("lf_activate_lock", lock);
1306 #endif /* LOCKF_DEBUG */
1307 if (lock->lf_type != F_UNLCK)
1308 lf_insert_lock(state, lock);
1313 * Cancel a pending lock request, either as a result of a signal or a
1314 * cancel request for an async lock.
1317 lf_cancel_lock(struct lockf *state, struct lockf_entry *lock)
1319 struct lockf_entry_list granted;
1322 * Note it is theoretically possible that cancelling this lock
1323 * may allow some other pending lock to become
1324 * active. Consider this case:
1326 * Owner Action Result Dependancies
1328 * A: lock [0..0] succeeds
1329 * B: lock [2..2] succeeds
1330 * C: lock [1..2] blocked C->B
1331 * D: lock [0..1] blocked C->B,D->A,D->C
1332 * A: unlock [0..0] C->B,D->C
1336 LIST_REMOVE(lock, lf_link);
1339 * Removing out-going edges is simple.
1341 sx_xlock(&lf_owner_graph_lock);
1342 lf_remove_outgoing(lock);
1343 sx_xunlock(&lf_owner_graph_lock);
1346 * Removing in-coming edges may allow some other lock to
1347 * become active - we use lf_update_dependancies to figure
1350 LIST_INIT(&granted);
1351 lf_update_dependancies(state, lock, TRUE, &granted);
1355 * Feed any newly active locks to lf_activate_lock.
1357 while (!LIST_EMPTY(&granted)) {
1358 lock = LIST_FIRST(&granted);
1359 LIST_REMOVE(lock, lf_link);
1360 lf_activate_lock(state, lock);
1365 * Set a byte-range lock.
1368 lf_setlock(struct lockf *state, struct lockf_entry *lock, struct vnode *vp,
1371 static char lockstr[] = "lockf";
1372 int priority, error;
1375 if (lockf_debug & 1)
1376 lf_print("lf_setlock", lock);
1377 #endif /* LOCKF_DEBUG */
1383 if (lock->lf_type == F_WRLCK)
1385 if (!(lock->lf_flags & F_NOINTR))
1388 * Scan lock list for this file looking for locks that would block us.
1390 if (lf_getblock(state, lock)) {
1392 * Free the structure and return if nonblocking.
1394 if ((lock->lf_flags & F_WAIT) == 0
1395 && lock->lf_async_task == NULL) {
1402 * For flock type locks, we must first remove
1403 * any shared locks that we hold before we sleep
1404 * waiting for an exclusive lock.
1406 if ((lock->lf_flags & F_FLOCK) &&
1407 lock->lf_type == F_WRLCK) {
1408 lock->lf_type = F_UNLCK;
1409 lf_activate_lock(state, lock);
1410 lock->lf_type = F_WRLCK;
1414 * We are blocked. Create edges to each blocking lock,
1415 * checking for deadlock using the owner graph. For
1416 * simplicity, we run deadlock detection for all
1417 * locks, posix and otherwise.
1419 sx_xlock(&lf_owner_graph_lock);
1420 error = lf_add_outgoing(state, lock);
1421 sx_xunlock(&lf_owner_graph_lock);
1425 if (lockf_debug & 1)
1426 lf_print("lf_setlock: deadlock", lock);
1433 * We have added edges to everything that blocks
1434 * us. Sleep until they all go away.
1436 LIST_INSERT_HEAD(&state->ls_pending, lock, lf_link);
1438 if (lockf_debug & 1) {
1439 struct lockf_edge *e;
1440 LIST_FOREACH(e, &lock->lf_outedges, le_outlink) {
1441 lf_print("lf_setlock: blocking on", e->le_to);
1442 lf_printlist("lf_setlock", e->le_to);
1445 #endif /* LOCKF_DEBUG */
1447 if ((lock->lf_flags & F_WAIT) == 0) {
1449 * The caller requested async notification -
1450 * this callback happens when the blocking
1451 * lock is released, allowing the caller to
1452 * make another attempt to take the lock.
1454 *cookiep = (void *) lock;
1455 error = EINPROGRESS;
1460 error = sx_sleep(lock, &state->ls_lock, priority, lockstr, 0);
1461 if (lf_free_lock(lock)) {
1467 * We may have been awakened by a signal and/or by a
1468 * debugger continuing us (in which cases we must
1469 * remove our lock graph edges) and/or by another
1470 * process releasing a lock (in which case our edges
1471 * have already been removed and we have been moved to
1472 * the active list). We may also have been woken by
1473 * lf_purgelocks which we report to the caller as
1474 * EINTR. In that case, lf_purgelocks will have
1475 * removed our lock graph edges.
1477 * Note that it is possible to receive a signal after
1478 * we were successfully woken (and moved to the active
1479 * list) but before we resumed execution. In this
1480 * case, our lf_outedges list will be clear. We
1481 * pretend there was no error.
1483 * Note also, if we have been sleeping long enough, we
1484 * may now have incoming edges from some newer lock
1485 * which is waiting behind us in the queue.
1487 if (lock->lf_flags & F_INTR) {
1492 if (LIST_EMPTY(&lock->lf_outedges)) {
1495 lf_cancel_lock(state, lock);
1499 if (lockf_debug & 1) {
1500 lf_print("lf_setlock: granted", lock);
1506 * It looks like we are going to grant the lock. First add
1507 * edges from any currently pending lock that the new lock
1510 sx_xlock(&lf_owner_graph_lock);
1511 error = lf_add_incoming(state, lock);
1512 sx_xunlock(&lf_owner_graph_lock);
1515 if (lockf_debug & 1)
1516 lf_print("lf_setlock: deadlock", lock);
1523 * No blocks!! Add the lock. Note that we will
1524 * downgrade or upgrade any overlapping locks this
1525 * process already owns.
1527 lf_activate_lock(state, lock);
1534 * Remove a byte-range lock on an inode.
1536 * Generally, find the lock (or an overlap to that lock)
1537 * and remove it (or shrink it), then wakeup anyone we can.
1540 lf_clearlock(struct lockf *state, struct lockf_entry *unlock)
1542 struct lockf_entry *overlap;
1544 overlap = LIST_FIRST(&state->ls_active);
1546 if (overlap == NOLOCKF)
1549 if (unlock->lf_type != F_UNLCK)
1550 panic("lf_clearlock: bad type");
1551 if (lockf_debug & 1)
1552 lf_print("lf_clearlock", unlock);
1553 #endif /* LOCKF_DEBUG */
1555 lf_activate_lock(state, unlock);
1561 * Check whether there is a blocking lock, and if so return its
1565 lf_getlock(struct lockf *state, struct lockf_entry *lock, struct flock *fl)
1567 struct lockf_entry *block;
1570 if (lockf_debug & 1)
1571 lf_print("lf_getlock", lock);
1572 #endif /* LOCKF_DEBUG */
1574 if ((block = lf_getblock(state, lock))) {
1575 fl->l_type = block->lf_type;
1576 fl->l_whence = SEEK_SET;
1577 fl->l_start = block->lf_start;
1578 if (block->lf_end == OFF_MAX)
1581 fl->l_len = block->lf_end - block->lf_start + 1;
1582 fl->l_pid = block->lf_owner->lo_pid;
1583 fl->l_sysid = block->lf_owner->lo_sysid;
1585 fl->l_type = F_UNLCK;
1591 * Cancel an async lock request.
1594 lf_cancel(struct lockf *state, struct lockf_entry *lock, void *cookie)
1596 struct lockf_entry *reallock;
1599 * We need to match this request with an existing lock
1602 LIST_FOREACH(reallock, &state->ls_pending, lf_link) {
1603 if ((void *) reallock == cookie) {
1605 * Double-check that this lock looks right
1606 * (maybe use a rolling ID for the cancel
1609 if (!(reallock->lf_vnode == lock->lf_vnode
1610 && reallock->lf_start == lock->lf_start
1611 && reallock->lf_end == lock->lf_end)) {
1616 * Make sure this lock was async and then just
1617 * remove it from its wait lists.
1619 if (!reallock->lf_async_task) {
1624 * Note that since any other thread must take
1625 * state->ls_lock before it can possibly
1626 * trigger the async callback, we are safe
1627 * from a race with lf_wakeup_lock, i.e. we
1628 * can free the lock (actually our caller does
1631 lf_cancel_lock(state, reallock);
1637 * We didn't find a matching lock - not much we can do here.
1643 * Walk the list of locks for an inode and
1644 * return the first blocking lock.
1646 static struct lockf_entry *
1647 lf_getblock(struct lockf *state, struct lockf_entry *lock)
1649 struct lockf_entry *overlap;
1651 LIST_FOREACH(overlap, &state->ls_active, lf_link) {
1653 * We may assume that the active list is sorted by
1656 if (overlap->lf_start > lock->lf_end)
1658 if (!lf_blocks(lock, overlap))
1666 * Walk the list of locks for an inode to find an overlapping lock (if
1667 * any) and return a classification of that overlap.
1670 * *overlap The place in the lock list to start looking
1671 * lock The lock which is being tested
1672 * type Pass 'SELF' to test only locks with the same
1673 * owner as lock, or 'OTHER' to test only locks
1674 * with a different owner
1676 * Returns one of six values:
1678 * 1) overlap == lock
1679 * 2) overlap contains lock
1680 * 3) lock contains overlap
1681 * 4) overlap starts before lock
1682 * 5) overlap ends after lock
1684 * If there is an overlapping lock, '*overlap' is set to point at the
1687 * NOTE: this returns only the FIRST overlapping lock. There
1688 * may be more than one.
1691 lf_findoverlap(struct lockf_entry **overlap, struct lockf_entry *lock, int type)
1693 struct lockf_entry *lf;
1697 if ((*overlap) == NOLOCKF) {
1701 if (lockf_debug & 2)
1702 lf_print("lf_findoverlap: looking for overlap in", lock);
1703 #endif /* LOCKF_DEBUG */
1704 start = lock->lf_start;
1709 if (lf->lf_start > end)
1711 if (((type & SELF) && lf->lf_owner != lock->lf_owner) ||
1712 ((type & OTHERS) && lf->lf_owner == lock->lf_owner)) {
1713 *overlap = LIST_NEXT(lf, lf_link);
1717 if (lockf_debug & 2)
1718 lf_print("\tchecking", lf);
1719 #endif /* LOCKF_DEBUG */
1721 * OK, check for overlap
1725 * 1) overlap == lock
1726 * 2) overlap contains lock
1727 * 3) lock contains overlap
1728 * 4) overlap starts before lock
1729 * 5) overlap ends after lock
1731 if (start > lf->lf_end) {
1734 if (lockf_debug & 2)
1735 printf("no overlap\n");
1736 #endif /* LOCKF_DEBUG */
1737 *overlap = LIST_NEXT(lf, lf_link);
1740 if (lf->lf_start == start && lf->lf_end == end) {
1743 if (lockf_debug & 2)
1744 printf("overlap == lock\n");
1745 #endif /* LOCKF_DEBUG */
1749 if (lf->lf_start <= start && lf->lf_end >= end) {
1752 if (lockf_debug & 2)
1753 printf("overlap contains lock\n");
1754 #endif /* LOCKF_DEBUG */
1758 if (start <= lf->lf_start && end >= lf->lf_end) {
1761 if (lockf_debug & 2)
1762 printf("lock contains overlap\n");
1763 #endif /* LOCKF_DEBUG */
1767 if (lf->lf_start < start && lf->lf_end >= start) {
1770 if (lockf_debug & 2)
1771 printf("overlap starts before lock\n");
1772 #endif /* LOCKF_DEBUG */
1776 if (lf->lf_start > start && lf->lf_end > end) {
1779 if (lockf_debug & 2)
1780 printf("overlap ends after lock\n");
1781 #endif /* LOCKF_DEBUG */
1785 panic("lf_findoverlap: default");
1791 * Split an the existing 'lock1', based on the extent of the lock
1792 * described by 'lock2'. The existing lock should cover 'lock2'
1795 * Any pending locks which have been been unblocked are added to
1799 lf_split(struct lockf *state, struct lockf_entry *lock1,
1800 struct lockf_entry *lock2, struct lockf_entry_list *granted)
1802 struct lockf_entry *splitlock;
1805 if (lockf_debug & 2) {
1806 lf_print("lf_split", lock1);
1807 lf_print("splitting from", lock2);
1809 #endif /* LOCKF_DEBUG */
1811 * Check to see if we don't need to split at all.
1813 if (lock1->lf_start == lock2->lf_start) {
1814 lf_set_start(state, lock1, lock2->lf_end + 1, granted);
1817 if (lock1->lf_end == lock2->lf_end) {
1818 lf_set_end(state, lock1, lock2->lf_start - 1, granted);
1822 * Make a new lock consisting of the last part of
1823 * the encompassing lock.
1825 splitlock = lf_alloc_lock(lock1->lf_owner);
1826 memcpy(splitlock, lock1, sizeof *splitlock);
1827 splitlock->lf_refs = 1;
1828 if (splitlock->lf_flags & F_REMOTE)
1829 vref(splitlock->lf_vnode);
1832 * This cannot cause a deadlock since any edges we would add
1833 * to splitlock already exist in lock1. We must be sure to add
1834 * necessary dependancies to splitlock before we reduce lock1
1835 * otherwise we may accidentally grant a pending lock that
1836 * was blocked by the tail end of lock1.
1838 splitlock->lf_start = lock2->lf_end + 1;
1839 LIST_INIT(&splitlock->lf_outedges);
1840 LIST_INIT(&splitlock->lf_inedges);
1841 sx_xlock(&lf_owner_graph_lock);
1842 lf_add_incoming(state, splitlock);
1843 sx_xunlock(&lf_owner_graph_lock);
1845 lf_set_end(state, lock1, lock2->lf_start - 1, granted);
1848 * OK, now link it in
1850 lf_insert_lock(state, splitlock);
1854 STAILQ_ENTRY(lockdesc) link;
1858 STAILQ_HEAD(lockdesclist, lockdesc);
1861 lf_iteratelocks_sysid(int sysid, lf_iterator *fn, void *arg)
1864 struct lockf_entry *lf;
1865 struct lockdesc *ldesc;
1866 struct lockdesclist locks;
1870 * In order to keep the locking simple, we iterate over the
1871 * active lock lists to build a list of locks that need
1872 * releasing. We then call the iterator for each one in turn.
1874 * We take an extra reference to the vnode for the duration to
1875 * make sure it doesn't go away before we are finished.
1877 STAILQ_INIT(&locks);
1878 sx_xlock(&lf_lock_states_lock);
1879 LIST_FOREACH(ls, &lf_lock_states, ls_link) {
1880 sx_xlock(&ls->ls_lock);
1881 LIST_FOREACH(lf, &ls->ls_active, lf_link) {
1882 if (lf->lf_owner->lo_sysid != sysid)
1885 ldesc = malloc(sizeof(struct lockdesc), M_LOCKF,
1887 ldesc->vp = lf->lf_vnode;
1889 ldesc->fl.l_start = lf->lf_start;
1890 if (lf->lf_end == OFF_MAX)
1891 ldesc->fl.l_len = 0;
1894 lf->lf_end - lf->lf_start + 1;
1895 ldesc->fl.l_whence = SEEK_SET;
1896 ldesc->fl.l_type = F_UNLCK;
1897 ldesc->fl.l_pid = lf->lf_owner->lo_pid;
1898 ldesc->fl.l_sysid = sysid;
1899 STAILQ_INSERT_TAIL(&locks, ldesc, link);
1901 sx_xunlock(&ls->ls_lock);
1903 sx_xunlock(&lf_lock_states_lock);
1906 * Call the iterator function for each lock in turn. If the
1907 * iterator returns an error code, just free the rest of the
1908 * lockdesc structures.
1911 while ((ldesc = STAILQ_FIRST(&locks)) != NULL) {
1912 STAILQ_REMOVE_HEAD(&locks, link);
1914 error = fn(ldesc->vp, &ldesc->fl, arg);
1916 free(ldesc, M_LOCKF);
1923 lf_iteratelocks_vnode(struct vnode *vp, lf_iterator *fn, void *arg)
1926 struct lockf_entry *lf;
1927 struct lockdesc *ldesc;
1928 struct lockdesclist locks;
1932 * In order to keep the locking simple, we iterate over the
1933 * active lock lists to build a list of locks that need
1934 * releasing. We then call the iterator for each one in turn.
1936 * We take an extra reference to the vnode for the duration to
1937 * make sure it doesn't go away before we are finished.
1939 STAILQ_INIT(&locks);
1949 sx_xlock(&ls->ls_lock);
1950 LIST_FOREACH(lf, &ls->ls_active, lf_link) {
1951 ldesc = malloc(sizeof(struct lockdesc), M_LOCKF,
1953 ldesc->vp = lf->lf_vnode;
1955 ldesc->fl.l_start = lf->lf_start;
1956 if (lf->lf_end == OFF_MAX)
1957 ldesc->fl.l_len = 0;
1960 lf->lf_end - lf->lf_start + 1;
1961 ldesc->fl.l_whence = SEEK_SET;
1962 ldesc->fl.l_type = F_UNLCK;
1963 ldesc->fl.l_pid = lf->lf_owner->lo_pid;
1964 ldesc->fl.l_sysid = lf->lf_owner->lo_sysid;
1965 STAILQ_INSERT_TAIL(&locks, ldesc, link);
1967 sx_xunlock(&ls->ls_lock);
1974 * Call the iterator function for each lock in turn. If the
1975 * iterator returns an error code, just free the rest of the
1976 * lockdesc structures.
1979 while ((ldesc = STAILQ_FIRST(&locks)) != NULL) {
1980 STAILQ_REMOVE_HEAD(&locks, link);
1982 error = fn(ldesc->vp, &ldesc->fl, arg);
1984 free(ldesc, M_LOCKF);
1991 lf_clearremotesys_iterator(struct vnode *vp, struct flock *fl, void *arg)
1994 VOP_ADVLOCK(vp, 0, F_UNLCK, fl, F_REMOTE);
1999 lf_clearremotesys(int sysid)
2002 KASSERT(sysid != 0, ("Can't clear local locks with F_UNLCKSYS"));
2003 lf_iteratelocks_sysid(sysid, lf_clearremotesys_iterator, NULL);
2007 lf_countlocks(int sysid)
2010 struct lock_owner *lo;
2014 sx_xlock(&lf_lock_owners_lock);
2015 for (i = 0; i < LOCK_OWNER_HASH_SIZE; i++)
2016 LIST_FOREACH(lo, &lf_lock_owners[i], lo_link)
2017 if (lo->lo_sysid == sysid)
2018 count += lo->lo_refs;
2019 sx_xunlock(&lf_lock_owners_lock);
2027 * Return non-zero if y is reachable from x using a brute force
2028 * search. If reachable and path is non-null, return the route taken
2032 graph_reaches(struct owner_vertex *x, struct owner_vertex *y,
2033 struct owner_vertex_list *path)
2035 struct owner_edge *e;
2039 TAILQ_INSERT_HEAD(path, x, v_link);
2043 LIST_FOREACH(e, &x->v_outedges, e_outlink) {
2044 if (graph_reaches(e->e_to, y, path)) {
2046 TAILQ_INSERT_HEAD(path, x, v_link);
2054 * Perform consistency checks on the graph. Make sure the values of
2055 * v_order are correct. If checkorder is non-zero, check no vertex can
2056 * reach any other vertex with a smaller order.
2059 graph_check(struct owner_graph *g, int checkorder)
2063 for (i = 0; i < g->g_size; i++) {
2064 if (!g->g_vertices[i]->v_owner)
2066 KASSERT(g->g_vertices[i]->v_order == i,
2067 ("lock graph vertices disordered"));
2069 for (j = 0; j < i; j++) {
2070 if (!g->g_vertices[j]->v_owner)
2072 KASSERT(!graph_reaches(g->g_vertices[i],
2073 g->g_vertices[j], NULL),
2074 ("lock graph vertices disordered"));
2081 graph_print_vertices(struct owner_vertex_list *set)
2083 struct owner_vertex *v;
2086 TAILQ_FOREACH(v, set, v_link) {
2087 printf("%d:", v->v_order);
2088 lf_print_owner(v->v_owner);
2089 if (TAILQ_NEXT(v, v_link))
2098 * Calculate the sub-set of vertices v from the affected region [y..x]
2099 * where v is reachable from y. Return -1 if a loop was detected
2100 * (i.e. x is reachable from y, otherwise the number of vertices in
2104 graph_delta_forward(struct owner_graph *g, struct owner_vertex *x,
2105 struct owner_vertex *y, struct owner_vertex_list *delta)
2108 struct owner_vertex *v;
2109 struct owner_edge *e;
2113 * We start with a set containing just y. Then for each vertex
2114 * v in the set so far unprocessed, we add each vertex that v
2115 * has an out-edge to and that is within the affected region
2116 * [y..x]. If we see the vertex x on our travels, stop
2120 TAILQ_INSERT_TAIL(delta, y, v_link);
2125 LIST_FOREACH(e, &v->v_outedges, e_outlink) {
2128 if (e->e_to->v_order < x->v_order
2129 && e->e_to->v_gen != gen) {
2130 e->e_to->v_gen = gen;
2131 TAILQ_INSERT_TAIL(delta, e->e_to, v_link);
2135 v = TAILQ_NEXT(v, v_link);
2142 * Calculate the sub-set of vertices v from the affected region [y..x]
2143 * where v reaches x. Return the number of vertices in this subset.
2146 graph_delta_backward(struct owner_graph *g, struct owner_vertex *x,
2147 struct owner_vertex *y, struct owner_vertex_list *delta)
2150 struct owner_vertex *v;
2151 struct owner_edge *e;
2155 * We start with a set containing just x. Then for each vertex
2156 * v in the set so far unprocessed, we add each vertex that v
2157 * has an in-edge from and that is within the affected region
2161 TAILQ_INSERT_TAIL(delta, x, v_link);
2166 LIST_FOREACH(e, &v->v_inedges, e_inlink) {
2167 if (e->e_from->v_order > y->v_order
2168 && e->e_from->v_gen != gen) {
2169 e->e_from->v_gen = gen;
2170 TAILQ_INSERT_HEAD(delta, e->e_from, v_link);
2174 v = TAILQ_PREV(v, owner_vertex_list, v_link);
2181 graph_add_indices(int *indices, int n, struct owner_vertex_list *set)
2183 struct owner_vertex *v;
2186 TAILQ_FOREACH(v, set, v_link) {
2188 i > 0 && indices[i - 1] > v->v_order; i--)
2190 for (j = n - 1; j >= i; j--)
2191 indices[j + 1] = indices[j];
2192 indices[i] = v->v_order;
2200 graph_assign_indices(struct owner_graph *g, int *indices, int nextunused,
2201 struct owner_vertex_list *set)
2203 struct owner_vertex *v, *vlowest;
2205 while (!TAILQ_EMPTY(set)) {
2207 TAILQ_FOREACH(v, set, v_link) {
2208 if (!vlowest || v->v_order < vlowest->v_order)
2211 TAILQ_REMOVE(set, vlowest, v_link);
2212 vlowest->v_order = indices[nextunused];
2213 g->g_vertices[vlowest->v_order] = vlowest;
2217 return (nextunused);
2221 graph_add_edge(struct owner_graph *g, struct owner_vertex *x,
2222 struct owner_vertex *y)
2224 struct owner_edge *e;
2225 struct owner_vertex_list deltaF, deltaB;
2226 int nF, nB, n, vi, i;
2229 sx_assert(&lf_owner_graph_lock, SX_XLOCKED);
2231 LIST_FOREACH(e, &x->v_outedges, e_outlink) {
2239 if (lockf_debug & 8) {
2240 printf("adding edge %d:", x->v_order);
2241 lf_print_owner(x->v_owner);
2242 printf(" -> %d:", y->v_order);
2243 lf_print_owner(y->v_owner);
2247 if (y->v_order < x->v_order) {
2249 * The new edge violates the order. First find the set
2250 * of affected vertices reachable from y (deltaF) and
2251 * the set of affect vertices affected that reach x
2252 * (deltaB), using the graph generation number to
2253 * detect whether we have visited a given vertex
2254 * already. We re-order the graph so that each vertex
2255 * in deltaB appears before each vertex in deltaF.
2257 * If x is a member of deltaF, then the new edge would
2258 * create a cycle. Otherwise, we may assume that
2259 * deltaF and deltaB are disjoint.
2262 if (g->g_gen == 0) {
2266 for (vi = 0; vi < g->g_size; vi++) {
2267 g->g_vertices[vi]->v_gen = 0;
2271 nF = graph_delta_forward(g, x, y, &deltaF);
2274 if (lockf_debug & 8) {
2275 struct owner_vertex_list path;
2276 printf("deadlock: ");
2278 graph_reaches(y, x, &path);
2279 graph_print_vertices(&path);
2286 if (lockf_debug & 8) {
2287 printf("re-ordering graph vertices\n");
2288 printf("deltaF = ");
2289 graph_print_vertices(&deltaF);
2293 nB = graph_delta_backward(g, x, y, &deltaB);
2296 if (lockf_debug & 8) {
2297 printf("deltaB = ");
2298 graph_print_vertices(&deltaB);
2303 * We first build a set of vertex indices (vertex
2304 * order values) that we may use, then we re-assign
2305 * orders first to those vertices in deltaB, then to
2306 * deltaF. Note that the contents of deltaF and deltaB
2307 * may be partially disordered - we perform an
2308 * insertion sort while building our index set.
2310 indices = g->g_indexbuf;
2311 n = graph_add_indices(indices, 0, &deltaF);
2312 graph_add_indices(indices, n, &deltaB);
2315 * We must also be sure to maintain the relative
2316 * ordering of deltaF and deltaB when re-assigning
2317 * vertices. We do this by iteratively removing the
2318 * lowest ordered element from the set and assigning
2319 * it the next value from our new ordering.
2321 i = graph_assign_indices(g, indices, 0, &deltaB);
2322 graph_assign_indices(g, indices, i, &deltaF);
2325 if (lockf_debug & 8) {
2326 struct owner_vertex_list set;
2328 for (i = 0; i < nB + nF; i++)
2329 TAILQ_INSERT_TAIL(&set,
2330 g->g_vertices[indices[i]], v_link);
2331 printf("new ordering = ");
2332 graph_print_vertices(&set);
2337 KASSERT(x->v_order < y->v_order, ("Failed to re-order graph"));
2340 if (lockf_debug & 8) {
2341 graph_check(g, TRUE);
2345 e = malloc(sizeof(struct owner_edge), M_LOCKF, M_WAITOK);
2347 LIST_INSERT_HEAD(&x->v_outedges, e, e_outlink);
2348 LIST_INSERT_HEAD(&y->v_inedges, e, e_inlink);
2357 * Remove an edge x->y from the graph.
2360 graph_remove_edge(struct owner_graph *g, struct owner_vertex *x,
2361 struct owner_vertex *y)
2363 struct owner_edge *e;
2365 sx_assert(&lf_owner_graph_lock, SX_XLOCKED);
2367 LIST_FOREACH(e, &x->v_outedges, e_outlink) {
2371 KASSERT(e, ("Removing non-existent edge from deadlock graph"));
2374 if (e->e_refs == 0) {
2376 if (lockf_debug & 8) {
2377 printf("removing edge %d:", x->v_order);
2378 lf_print_owner(x->v_owner);
2379 printf(" -> %d:", y->v_order);
2380 lf_print_owner(y->v_owner);
2384 LIST_REMOVE(e, e_outlink);
2385 LIST_REMOVE(e, e_inlink);
2391 * Allocate a vertex from the free list. Return ENOMEM if there are
2394 static struct owner_vertex *
2395 graph_alloc_vertex(struct owner_graph *g, struct lock_owner *lo)
2397 struct owner_vertex *v;
2399 sx_assert(&lf_owner_graph_lock, SX_XLOCKED);
2401 v = malloc(sizeof(struct owner_vertex), M_LOCKF, M_WAITOK);
2402 if (g->g_size == g->g_space) {
2403 g->g_vertices = realloc(g->g_vertices,
2404 2 * g->g_space * sizeof(struct owner_vertex *),
2406 free(g->g_indexbuf, M_LOCKF);
2407 g->g_indexbuf = malloc(2 * g->g_space * sizeof(int),
2409 g->g_space = 2 * g->g_space;
2411 v->v_order = g->g_size;
2412 v->v_gen = g->g_gen;
2413 g->g_vertices[g->g_size] = v;
2416 LIST_INIT(&v->v_outedges);
2417 LIST_INIT(&v->v_inedges);
2424 graph_free_vertex(struct owner_graph *g, struct owner_vertex *v)
2426 struct owner_vertex *w;
2429 sx_assert(&lf_owner_graph_lock, SX_XLOCKED);
2431 KASSERT(LIST_EMPTY(&v->v_outedges), ("Freeing vertex with edges"));
2432 KASSERT(LIST_EMPTY(&v->v_inedges), ("Freeing vertex with edges"));
2435 * Remove from the graph's array and close up the gap,
2436 * renumbering the other vertices.
2438 for (i = v->v_order + 1; i < g->g_size; i++) {
2439 w = g->g_vertices[i];
2441 g->g_vertices[i - 1] = w;
2448 static struct owner_graph *
2449 graph_init(struct owner_graph *g)
2452 g->g_vertices = malloc(10 * sizeof(struct owner_vertex *),
2456 g->g_indexbuf = malloc(g->g_space * sizeof(int), M_LOCKF, M_WAITOK);
2464 * Print description of a lock owner
2467 lf_print_owner(struct lock_owner *lo)
2470 if (lo->lo_flags & F_REMOTE) {
2471 printf("remote pid %d, system %d",
2472 lo->lo_pid, lo->lo_sysid);
2473 } else if (lo->lo_flags & F_FLOCK) {
2474 printf("file %p", lo->lo_id);
2476 printf("local pid %d", lo->lo_pid);
2484 lf_print(char *tag, struct lockf_entry *lock)
2487 printf("%s: lock %p for ", tag, (void *)lock);
2488 lf_print_owner(lock->lf_owner);
2489 if (lock->lf_inode != (struct inode *)0)
2490 printf(" in ino %ju on dev <%s>,",
2491 (uintmax_t)lock->lf_inode->i_number,
2492 devtoname(lock->lf_inode->i_dev));
2493 printf(" %s, start %jd, end ",
2494 lock->lf_type == F_RDLCK ? "shared" :
2495 lock->lf_type == F_WRLCK ? "exclusive" :
2496 lock->lf_type == F_UNLCK ? "unlock" : "unknown",
2497 (intmax_t)lock->lf_start);
2498 if (lock->lf_end == OFF_MAX)
2501 printf("%jd", (intmax_t)lock->lf_end);
2502 if (!LIST_EMPTY(&lock->lf_outedges))
2503 printf(" block %p\n",
2504 (void *)LIST_FIRST(&lock->lf_outedges)->le_to);
2510 lf_printlist(char *tag, struct lockf_entry *lock)
2512 struct lockf_entry *lf, *blk;
2513 struct lockf_edge *e;
2515 if (lock->lf_inode == (struct inode *)0)
2518 printf("%s: Lock list for ino %ju on dev <%s>:\n",
2519 tag, (uintmax_t)lock->lf_inode->i_number,
2520 devtoname(lock->lf_inode->i_dev));
2521 LIST_FOREACH(lf, &lock->lf_vnode->v_lockf->ls_active, lf_link) {
2522 printf("\tlock %p for ",(void *)lf);
2523 lf_print_owner(lock->lf_owner);
2524 printf(", %s, start %jd, end %jd",
2525 lf->lf_type == F_RDLCK ? "shared" :
2526 lf->lf_type == F_WRLCK ? "exclusive" :
2527 lf->lf_type == F_UNLCK ? "unlock" :
2528 "unknown", (intmax_t)lf->lf_start, (intmax_t)lf->lf_end);
2529 LIST_FOREACH(e, &lf->lf_outedges, le_outlink) {
2531 printf("\n\t\tlock request %p for ", (void *)blk);
2532 lf_print_owner(blk->lf_owner);
2533 printf(", %s, start %jd, end %jd",
2534 blk->lf_type == F_RDLCK ? "shared" :
2535 blk->lf_type == F_WRLCK ? "exclusive" :
2536 blk->lf_type == F_UNLCK ? "unlock" :
2537 "unknown", (intmax_t)blk->lf_start,
2538 (intmax_t)blk->lf_end);
2539 if (!LIST_EMPTY(&blk->lf_inedges))
2540 panic("lf_printlist: bad list");
2545 #endif /* LOCKF_DEBUG */