2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 2008 Isilon Inc http://www.isilon.com/
5 * Authors: Doug Rabson <dfr@rabson.org>
6 * Developed with Red Inc: Alfred Perlstein <alfred@freebsd.org>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * Copyright (c) 1982, 1986, 1989, 1993
31 * The Regents of the University of California. All rights reserved.
33 * This code is derived from software contributed to Berkeley by
34 * Scooter Morris at Genentech Inc.
36 * Redistribution and use in source and binary forms, with or without
37 * modification, are permitted provided that the following conditions
39 * 1. Redistributions of source code must retain the above copyright
40 * notice, this list of conditions and the following disclaimer.
41 * 2. Redistributions in binary form must reproduce the above copyright
42 * notice, this list of conditions and the following disclaimer in the
43 * documentation and/or other materials provided with the distribution.
44 * 3. Neither the name of the University nor the names of its contributors
45 * may be used to endorse or promote products derived from this software
46 * without specific prior written permission.
48 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
49 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
50 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
51 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
52 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
53 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
54 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
55 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
56 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
57 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
60 * @(#)ufs_lockf.c 8.3 (Berkeley) 1/6/94
63 #include <sys/cdefs.h>
64 #include "opt_debug_lockf.h"
66 #include <sys/param.h>
67 #include <sys/systm.h>
70 #include <sys/kernel.h>
71 #include <sys/limits.h>
73 #include <sys/mount.h>
74 #include <sys/mutex.h>
79 #include <sys/unistd.h>
81 #include <sys/vnode.h>
82 #include <sys/malloc.h>
83 #include <sys/fcntl.h>
84 #include <sys/lockf.h>
85 #include <sys/taskqueue.h>
88 #include <sys/sysctl.h>
90 static int lockf_debug = 0; /* control debug output */
91 SYSCTL_INT(_debug, OID_AUTO, lockf_debug, CTLFLAG_RW, &lockf_debug, 0, "");
94 static MALLOC_DEFINE(M_LOCKF, "lockf", "Byte-range locking structures");
98 struct owner_vertex_list;
101 #define NOLOCKF (struct lockf_entry *)0
104 static void lf_init(void *);
105 static int lf_hash_owner(caddr_t, struct vnode *, struct flock *, int);
106 static int lf_owner_matches(struct lock_owner *, caddr_t, struct flock *,
108 static struct lockf_entry *
109 lf_alloc_lock(struct lock_owner *);
110 static int lf_free_lock(struct lockf_entry *);
111 static int lf_clearlock(struct lockf *, struct lockf_entry *);
112 static int lf_overlaps(struct lockf_entry *, struct lockf_entry *);
113 static int lf_blocks(struct lockf_entry *, struct lockf_entry *);
114 static void lf_free_edge(struct lockf_edge *);
115 static struct lockf_edge *
117 static void lf_alloc_vertex(struct lockf_entry *);
118 static int lf_add_edge(struct lockf_entry *, struct lockf_entry *);
119 static void lf_remove_edge(struct lockf_edge *);
120 static void lf_remove_outgoing(struct lockf_entry *);
121 static void lf_remove_incoming(struct lockf_entry *);
122 static int lf_add_outgoing(struct lockf *, struct lockf_entry *);
123 static int lf_add_incoming(struct lockf *, struct lockf_entry *);
124 static int lf_findoverlap(struct lockf_entry **, struct lockf_entry *,
126 static struct lockf_entry *
127 lf_getblock(struct lockf *, struct lockf_entry *);
128 static int lf_getlock(struct lockf *, struct lockf_entry *, struct flock *);
129 static void lf_insert_lock(struct lockf *, struct lockf_entry *);
130 static void lf_wakeup_lock(struct lockf *, struct lockf_entry *);
131 static void lf_update_dependancies(struct lockf *, struct lockf_entry *,
132 int all, struct lockf_entry_list *);
133 static void lf_set_start(struct lockf *, struct lockf_entry *, off_t,
134 struct lockf_entry_list*);
135 static void lf_set_end(struct lockf *, struct lockf_entry *, off_t,
136 struct lockf_entry_list*);
137 static int lf_setlock(struct lockf *, struct lockf_entry *,
138 struct vnode *, void **cookiep);
139 static int lf_cancel(struct lockf *, struct lockf_entry *, void *);
140 static void lf_split(struct lockf *, struct lockf_entry *,
141 struct lockf_entry *, struct lockf_entry_list *);
143 static int graph_reaches(struct owner_vertex *x, struct owner_vertex *y,
144 struct owner_vertex_list *path);
145 static void graph_check(struct owner_graph *g, int checkorder);
146 static void graph_print_vertices(struct owner_vertex_list *set);
148 static int graph_delta_forward(struct owner_graph *g,
149 struct owner_vertex *x, struct owner_vertex *y,
150 struct owner_vertex_list *delta);
151 static int graph_delta_backward(struct owner_graph *g,
152 struct owner_vertex *x, struct owner_vertex *y,
153 struct owner_vertex_list *delta);
154 static int graph_add_indices(int *indices, int n,
155 struct owner_vertex_list *set);
156 static int graph_assign_indices(struct owner_graph *g, int *indices,
157 int nextunused, struct owner_vertex_list *set);
158 static int graph_add_edge(struct owner_graph *g,
159 struct owner_vertex *x, struct owner_vertex *y);
160 static void graph_remove_edge(struct owner_graph *g,
161 struct owner_vertex *x, struct owner_vertex *y);
162 static struct owner_vertex *graph_alloc_vertex(struct owner_graph *g,
163 struct lock_owner *lo);
164 static void graph_free_vertex(struct owner_graph *g,
165 struct owner_vertex *v);
166 static struct owner_graph * graph_init(struct owner_graph *g);
168 static void lf_print(char *, struct lockf_entry *);
169 static void lf_printlist(char *, struct lockf_entry *);
170 static void lf_print_owner(struct lock_owner *);
174 * This structure is used to keep track of both local and remote lock
175 * owners. The lf_owner field of the struct lockf_entry points back at
176 * the lock owner structure. Each possible lock owner (local proc for
177 * POSIX fcntl locks, local file for BSD flock locks or <pid,sysid>
178 * pair for remote locks) is represented by a unique instance of
181 * If a lock owner has a lock that blocks some other lock or a lock
182 * that is waiting for some other lock, it also has a vertex in the
186 * (s) locked by state->ls_lock
187 * (S) locked by lf_lock_states_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 passed 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 int lo_hash; /* (c) Used to lock the appropriate chain */
201 struct owner_vertex *lo_vertex; /* (g) entry in deadlock graph */
204 LIST_HEAD(lock_owner_list, lock_owner);
206 struct lock_owner_chain {
208 struct lock_owner_list list;
211 static struct sx lf_lock_states_lock;
212 static struct lockf_list lf_lock_states; /* (S) */
213 static struct lock_owner_chain lf_lock_owners[LOCK_OWNER_HASH_SIZE];
216 * Structures for deadlock detection.
218 * We have two types of directed graph, the first is the set of locks,
219 * both active and pending on a vnode. Within this graph, active locks
220 * are terminal nodes in the graph (i.e. have no out-going
221 * edges). Pending locks have out-going edges to each blocking active
222 * lock that prevents the lock from being granted and also to each
223 * older pending lock that would block them if it was active. The
224 * graph for each vnode is naturally acyclic; new edges are only ever
225 * added to or from new nodes (either new pending locks which only add
226 * out-going edges or new active locks which only add in-coming edges)
227 * therefore they cannot create loops in the lock graph.
229 * The second graph is a global graph of lock owners. Each lock owner
230 * is a vertex in that graph and an edge is added to the graph
231 * whenever an edge is added to a vnode graph, with end points
232 * corresponding to owner of the new pending lock and the owner of the
233 * lock upon which it waits. In order to prevent deadlock, we only add
234 * an edge to this graph if the new edge would not create a cycle.
236 * The lock owner graph is topologically sorted, i.e. if a node has
237 * any outgoing edges, then it has an order strictly less than any
238 * node to which it has an outgoing edge. We preserve this ordering
239 * (and detect cycles) on edge insertion using Algorithm PK from the
240 * paper "A Dynamic Topological Sort Algorithm for Directed Acyclic
241 * Graphs" (ACM Journal of Experimental Algorithms, Vol 11, Article
247 LIST_ENTRY(owner_edge) e_outlink; /* (g) link from's out-edge list */
248 LIST_ENTRY(owner_edge) e_inlink; /* (g) link to's in-edge list */
249 int e_refs; /* (g) number of times added */
250 struct owner_vertex *e_from; /* (c) out-going from here */
251 struct owner_vertex *e_to; /* (c) in-coming to here */
253 LIST_HEAD(owner_edge_list, owner_edge);
255 struct owner_vertex {
256 TAILQ_ENTRY(owner_vertex) v_link; /* (g) workspace for edge insertion */
257 uint32_t v_gen; /* (g) workspace for edge insertion */
258 int v_order; /* (g) order of vertex in graph */
259 struct owner_edge_list v_outedges;/* (g) list of out-edges */
260 struct owner_edge_list v_inedges; /* (g) list of in-edges */
261 struct lock_owner *v_owner; /* (c) corresponding lock owner */
263 TAILQ_HEAD(owner_vertex_list, owner_vertex);
266 struct owner_vertex** g_vertices; /* (g) pointers to vertices */
267 int g_size; /* (g) number of vertices */
268 int g_space; /* (g) space allocated for vertices */
269 int *g_indexbuf; /* (g) workspace for loop detection */
270 uint32_t g_gen; /* (g) increment when re-ordering */
273 static struct sx lf_owner_graph_lock;
274 static struct owner_graph lf_owner_graph;
277 * Initialise various structures and locks.
284 sx_init(&lf_lock_states_lock, "lock states lock");
285 LIST_INIT(&lf_lock_states);
287 for (i = 0; i < LOCK_OWNER_HASH_SIZE; i++) {
288 sx_init(&lf_lock_owners[i].lock, "lock owners lock");
289 LIST_INIT(&lf_lock_owners[i].list);
292 sx_init(&lf_owner_graph_lock, "owner graph lock");
293 graph_init(&lf_owner_graph);
295 SYSINIT(lf_init, SI_SUB_LOCK, SI_ORDER_FIRST, lf_init, NULL);
298 * Generate a hash value for a lock owner.
301 lf_hash_owner(caddr_t id, struct vnode *vp, struct flock *fl, int flags)
305 if (flags & F_REMOTE) {
306 h = HASHSTEP(0, fl->l_pid);
307 h = HASHSTEP(h, fl->l_sysid);
308 } else if (flags & F_FLOCK) {
309 h = ((uintptr_t) id) >> 7;
311 h = ((uintptr_t) vp) >> 7;
314 return (h % LOCK_OWNER_HASH_SIZE);
318 * Return true if a lock owner matches the details passed to
322 lf_owner_matches(struct lock_owner *lo, caddr_t id, struct flock *fl,
325 if (flags & F_REMOTE) {
326 return lo->lo_pid == fl->l_pid
327 && lo->lo_sysid == fl->l_sysid;
329 return lo->lo_id == id;
333 static struct lockf_entry *
334 lf_alloc_lock(struct lock_owner *lo)
336 struct lockf_entry *lf;
338 lf = malloc(sizeof(struct lockf_entry), M_LOCKF, M_WAITOK|M_ZERO);
342 printf("Allocated lock %p\n", lf);
345 sx_xlock(&lf_lock_owners[lo->lo_hash].lock);
347 sx_xunlock(&lf_lock_owners[lo->lo_hash].lock);
355 lf_free_lock(struct lockf_entry *lock)
357 struct sx *chainlock;
359 KASSERT(lock->lf_refs > 0, ("lockf_entry negative ref count %p", lock));
360 if (--lock->lf_refs > 0)
363 * Adjust the lock_owner reference count and
364 * reclaim the entry if this is the last lock
367 struct lock_owner *lo = lock->lf_owner;
369 KASSERT(LIST_EMPTY(&lock->lf_outedges),
370 ("freeing lock with dependencies"));
371 KASSERT(LIST_EMPTY(&lock->lf_inedges),
372 ("freeing lock with dependants"));
373 chainlock = &lf_lock_owners[lo->lo_hash].lock;
375 KASSERT(lo->lo_refs > 0, ("lock owner refcount"));
377 if (lo->lo_refs == 0) {
380 printf("lf_free_lock: freeing lock owner %p\n",
384 sx_xlock(&lf_owner_graph_lock);
385 graph_free_vertex(&lf_owner_graph,
387 sx_xunlock(&lf_owner_graph_lock);
389 LIST_REMOVE(lo, lo_link);
393 printf("Freed lock owner %p\n", lo);
396 sx_unlock(chainlock);
398 if ((lock->lf_flags & F_REMOTE) && lock->lf_vnode) {
399 vrele(lock->lf_vnode);
400 lock->lf_vnode = NULL;
404 printf("Freed lock %p\n", lock);
411 * Advisory record locking support
414 lf_advlockasync(struct vop_advlockasync_args *ap, struct lockf **statep,
418 struct flock *fl = ap->a_fl;
419 struct lockf_entry *lock;
420 struct vnode *vp = ap->a_vp;
421 caddr_t id = ap->a_id;
422 int flags = ap->a_flags;
424 struct lock_owner *lo;
425 off_t start, end, oadd;
429 * Handle the F_UNLKSYS case first - no need to mess about
430 * creating a lock owner for this one.
432 if (ap->a_op == F_UNLCKSYS) {
433 lf_clearremotesys(fl->l_sysid);
438 * Convert the flock structure into a start and end.
440 switch (fl->l_whence) {
444 * Caller is responsible for adding any necessary offset
445 * when SEEK_CUR is used.
451 if (size > OFF_MAX ||
452 (fl->l_start > 0 && size > OFF_MAX - fl->l_start))
454 start = size + fl->l_start;
469 } else if (fl->l_len == 0) {
472 oadd = fl->l_len - 1;
473 if (oadd > OFF_MAX - start)
481 * Avoid the common case of unlocking when inode has no locks.
483 if (ap->a_op != F_SETLK && (*statep) == NULL) {
485 if ((*statep) == NULL) {
486 fl->l_type = F_UNLCK;
494 * Map our arguments to an existing lock owner or create one
495 * if this is the first time we have seen this owner.
497 hash = lf_hash_owner(id, vp, fl, flags);
498 sx_xlock(&lf_lock_owners[hash].lock);
499 LIST_FOREACH(lo, &lf_lock_owners[hash].list, lo_link)
500 if (lf_owner_matches(lo, id, fl, flags))
504 * We initialise the lock with a reference
505 * count which matches the new lockf_entry
506 * structure created below.
508 lo = malloc(sizeof(struct lock_owner), M_LOCKF,
512 printf("Allocated lock owner %p\n", lo);
516 lo->lo_flags = flags;
519 if (flags & F_REMOTE) {
520 lo->lo_pid = fl->l_pid;
521 lo->lo_sysid = fl->l_sysid;
522 } else if (flags & F_FLOCK) {
526 struct proc *p = (struct proc *) id;
527 lo->lo_pid = p->p_pid;
530 lo->lo_vertex = NULL;
533 if (lockf_debug & 1) {
534 printf("lf_advlockasync: new lock owner %p ", lo);
540 LIST_INSERT_HEAD(&lf_lock_owners[hash].list, lo, lo_link);
543 * We have seen this lock owner before, increase its
544 * reference count to account for the new lockf_entry
545 * structure we create below.
549 sx_xunlock(&lf_lock_owners[hash].lock);
552 * Create the lockf structure. We initialise the lf_owner
553 * field here instead of in lf_alloc_lock() to avoid paying
554 * the lf_lock_owners_lock tax twice.
556 lock = lf_alloc_lock(NULL);
558 lock->lf_start = start;
562 if (flags & F_REMOTE) {
564 * For remote locks, the caller may release its ref to
565 * the vnode at any time - we have to ref it here to
566 * prevent it from being recycled unexpectedly.
571 lock->lf_type = fl->l_type;
572 LIST_INIT(&lock->lf_outedges);
573 LIST_INIT(&lock->lf_inedges);
574 lock->lf_async_task = ap->a_task;
575 lock->lf_flags = ap->a_flags;
578 * Do the requested operation. First find our state structure
579 * and create a new one if necessary - the caller's *statep
580 * variable and the state's ls_threads count is protected by
581 * the vnode interlock.
584 if (VN_IS_DOOMED(vp)) {
591 * Allocate a state structure if necessary.
599 ls = malloc(sizeof(struct lockf), M_LOCKF, M_WAITOK|M_ZERO);
600 sx_init(&ls->ls_lock, "ls_lock");
601 LIST_INIT(&ls->ls_active);
602 LIST_INIT(&ls->ls_pending);
605 sx_xlock(&lf_lock_states_lock);
606 LIST_INSERT_HEAD(&lf_lock_states, ls, ls_link);
607 sx_xunlock(&lf_lock_states_lock);
610 * Cope if we lost a race with some other thread while
611 * trying to allocate memory.
614 if (VN_IS_DOOMED(vp)) {
616 sx_xlock(&lf_lock_states_lock);
617 LIST_REMOVE(ls, ls_link);
618 sx_xunlock(&lf_lock_states_lock);
619 sx_destroy(&ls->ls_lock);
624 if ((*statep) == NULL) {
625 state = *statep = ls;
629 MPASS(state->ls_threads >= 0);
633 sx_xlock(&lf_lock_states_lock);
634 LIST_REMOVE(ls, ls_link);
635 sx_xunlock(&lf_lock_states_lock);
636 sx_destroy(&ls->ls_lock);
640 MPASS(state->ls_threads >= 0);
645 sx_xlock(&state->ls_lock);
647 * Recheck the doomed vnode after state->ls_lock is
648 * locked. lf_purgelocks() requires that no new threads add
649 * pending locks when vnode is marked by VIRF_DOOMED flag.
651 if (VN_IS_DOOMED(vp)) {
653 MPASS(state->ls_threads > 0);
657 sx_xunlock(&state->ls_lock);
664 error = lf_setlock(state, lock, vp, ap->a_cookiep);
668 error = lf_clearlock(state, lock);
673 error = lf_getlock(state, lock, fl);
679 error = lf_cancel(state, lock, *ap->a_cookiep);
693 * Check for some can't happen stuff. In this case, the active
694 * lock list becoming disordered or containing mutually
695 * blocking locks. We also check the pending list for locks
696 * which should be active (i.e. have no out-going edges).
698 LIST_FOREACH(lock, &state->ls_active, lf_link) {
699 struct lockf_entry *lf;
700 if (LIST_NEXT(lock, lf_link))
701 KASSERT((lock->lf_start
702 <= LIST_NEXT(lock, lf_link)->lf_start),
703 ("locks disordered"));
704 LIST_FOREACH(lf, &state->ls_active, lf_link) {
707 KASSERT(!lf_blocks(lock, lf),
708 ("two conflicting active locks"));
709 if (lock->lf_owner == lf->lf_owner)
710 KASSERT(!lf_overlaps(lock, lf),
711 ("two overlapping locks from same owner"));
714 LIST_FOREACH(lock, &state->ls_pending, lf_link) {
715 KASSERT(!LIST_EMPTY(&lock->lf_outedges),
716 ("pending lock which should be active"));
719 sx_xunlock(&state->ls_lock);
722 MPASS(state->ls_threads > 0);
724 if (state->ls_threads != 0) {
729 if (error == EDOOFUS) {
730 KASSERT(ap->a_op == F_SETLK, ("EDOOFUS"));
737 lf_advlock(struct vop_advlock_args *ap, struct lockf **statep, u_quad_t size)
739 struct vop_advlockasync_args a;
745 a.a_flags = ap->a_flags;
749 return (lf_advlockasync(&a, statep, size));
753 lf_purgelocks(struct vnode *vp, struct lockf **statep)
756 struct lockf_entry *lock, *nlock;
759 * For this to work correctly, the caller must ensure that no
760 * other threads enter the locking system for this vnode,
761 * e.g. by checking VIRF_DOOMED. We wake up any threads that are
762 * sleeping waiting for locks on this vnode and then free all
763 * the remaining locks.
766 KASSERT(VN_IS_DOOMED(vp),
767 ("lf_purgelocks: vp %p has not vgone yet", vp));
774 if (LIST_EMPTY(&state->ls_active) && state->ls_threads == 0) {
775 KASSERT(LIST_EMPTY(&state->ls_pending),
776 ("freeing state with pending locks"));
780 MPASS(state->ls_threads >= 0);
784 sx_xlock(&state->ls_lock);
785 sx_xlock(&lf_owner_graph_lock);
786 LIST_FOREACH_SAFE(lock, &state->ls_pending, lf_link, nlock) {
787 LIST_REMOVE(lock, lf_link);
788 lf_remove_outgoing(lock);
789 lf_remove_incoming(lock);
792 * If its an async lock, we can just free it
793 * here, otherwise we let the sleeping thread
796 if (lock->lf_async_task) {
799 lock->lf_flags |= F_INTR;
803 sx_xunlock(&lf_owner_graph_lock);
804 sx_xunlock(&state->ls_lock);
807 * Wait for all other threads, sleeping and otherwise
811 while (state->ls_threads > 1)
812 msleep(state, VI_MTX(vp), 0, "purgelocks", 0);
816 * We can just free all the active locks since they
817 * will have no dependencies (we removed them all
818 * above). We don't need to bother locking since we
819 * are the last thread using this state structure.
821 KASSERT(LIST_EMPTY(&state->ls_pending),
822 ("lock pending for %p", state));
823 LIST_FOREACH_SAFE(lock, &state->ls_active, lf_link, nlock) {
824 LIST_REMOVE(lock, lf_link);
828 sx_xlock(&lf_lock_states_lock);
829 LIST_REMOVE(state, ls_link);
830 sx_xunlock(&lf_lock_states_lock);
831 sx_destroy(&state->ls_lock);
832 free(state, M_LOCKF);
836 * Return non-zero if locks 'x' and 'y' overlap.
839 lf_overlaps(struct lockf_entry *x, struct lockf_entry *y)
842 return (x->lf_start <= y->lf_end && x->lf_end >= y->lf_start);
846 * Return non-zero if lock 'x' is blocked by lock 'y' (or vice versa).
849 lf_blocks(struct lockf_entry *x, struct lockf_entry *y)
852 return x->lf_owner != y->lf_owner
853 && (x->lf_type == F_WRLCK || y->lf_type == F_WRLCK)
854 && lf_overlaps(x, y);
858 * Allocate a lock edge from the free list
860 static struct lockf_edge *
864 return (malloc(sizeof(struct lockf_edge), M_LOCKF, M_WAITOK|M_ZERO));
871 lf_free_edge(struct lockf_edge *e)
878 * Ensure that the lock's owner has a corresponding vertex in the
882 lf_alloc_vertex(struct lockf_entry *lock)
884 struct owner_graph *g = &lf_owner_graph;
886 if (!lock->lf_owner->lo_vertex)
887 lock->lf_owner->lo_vertex =
888 graph_alloc_vertex(g, lock->lf_owner);
892 * Attempt to record an edge from lock x to lock y. Return EDEADLK if
893 * the new edge would cause a cycle in the owner graph.
896 lf_add_edge(struct lockf_entry *x, struct lockf_entry *y)
898 struct owner_graph *g = &lf_owner_graph;
899 struct lockf_edge *e;
903 LIST_FOREACH(e, &x->lf_outedges, le_outlink)
904 KASSERT(e->le_to != y, ("adding lock edge twice"));
908 * Make sure the two owners have entries in the owner graph.
913 error = graph_add_edge(g, x->lf_owner->lo_vertex,
914 y->lf_owner->lo_vertex);
919 LIST_INSERT_HEAD(&x->lf_outedges, e, le_outlink);
920 LIST_INSERT_HEAD(&y->lf_inedges, e, le_inlink);
928 * Remove an edge from the lock graph.
931 lf_remove_edge(struct lockf_edge *e)
933 struct owner_graph *g = &lf_owner_graph;
934 struct lockf_entry *x = e->le_from;
935 struct lockf_entry *y = e->le_to;
937 graph_remove_edge(g, x->lf_owner->lo_vertex, y->lf_owner->lo_vertex);
938 LIST_REMOVE(e, le_outlink);
939 LIST_REMOVE(e, le_inlink);
946 * Remove all out-going edges from lock x.
949 lf_remove_outgoing(struct lockf_entry *x)
951 struct lockf_edge *e;
953 while ((e = LIST_FIRST(&x->lf_outedges)) != NULL) {
959 * Remove all in-coming edges from lock x.
962 lf_remove_incoming(struct lockf_entry *x)
964 struct lockf_edge *e;
966 while ((e = LIST_FIRST(&x->lf_inedges)) != NULL) {
972 * Walk the list of locks for the file and create an out-going edge
973 * from lock to each blocking lock.
976 lf_add_outgoing(struct lockf *state, struct lockf_entry *lock)
978 struct lockf_entry *overlap;
981 LIST_FOREACH(overlap, &state->ls_active, lf_link) {
983 * We may assume that the active list is sorted by
986 if (overlap->lf_start > lock->lf_end)
988 if (!lf_blocks(lock, overlap))
992 * We've found a blocking lock. Add the corresponding
993 * edge to the graphs and see if it would cause a
996 error = lf_add_edge(lock, overlap);
999 * The only error that lf_add_edge returns is EDEADLK.
1000 * Remove any edges we added and return the error.
1003 lf_remove_outgoing(lock);
1009 * We also need to add edges to sleeping locks that block
1010 * us. This ensures that lf_wakeup_lock cannot grant two
1011 * mutually blocking locks simultaneously and also enforces a
1012 * 'first come, first served' fairness model. Note that this
1013 * only happens if we are blocked by at least one active lock
1014 * due to the call to lf_getblock in lf_setlock below.
1016 LIST_FOREACH(overlap, &state->ls_pending, lf_link) {
1017 if (!lf_blocks(lock, overlap))
1020 * We've found a blocking lock. Add the corresponding
1021 * edge to the graphs and see if it would cause a
1024 error = lf_add_edge(lock, overlap);
1027 * The only error that lf_add_edge returns is EDEADLK.
1028 * Remove any edges we added and return the error.
1031 lf_remove_outgoing(lock);
1040 * Walk the list of pending locks for the file and create an in-coming
1041 * edge from lock to each blocking lock.
1044 lf_add_incoming(struct lockf *state, struct lockf_entry *lock)
1046 struct lockf_entry *overlap;
1049 sx_assert(&state->ls_lock, SX_XLOCKED);
1050 if (LIST_EMPTY(&state->ls_pending))
1054 sx_xlock(&lf_owner_graph_lock);
1055 LIST_FOREACH(overlap, &state->ls_pending, lf_link) {
1056 if (!lf_blocks(lock, overlap))
1060 * We've found a blocking lock. Add the corresponding
1061 * edge to the graphs and see if it would cause a
1064 error = lf_add_edge(overlap, lock);
1067 * The only error that lf_add_edge returns is EDEADLK.
1068 * Remove any edges we added and return the error.
1071 lf_remove_incoming(lock);
1075 sx_xunlock(&lf_owner_graph_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 dependencies 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 dependent 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 dependencies, 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 dependencies 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 dependencies 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 dependencies 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 Dependencies
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 error, priority, stops_deferred;
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 stops_deferred = sigdeferstop(SIGDEFERSTOP_ERESTART);
1461 error = sx_sleep(lock, &state->ls_lock, priority, lockstr, 0);
1462 sigallowstop(stops_deferred);
1463 if (lf_free_lock(lock)) {
1469 * We may have been awakened by a signal and/or by a
1470 * debugger continuing us (in which cases we must
1471 * remove our lock graph edges) and/or by another
1472 * process releasing a lock (in which case our edges
1473 * have already been removed and we have been moved to
1474 * the active list). We may also have been woken by
1475 * lf_purgelocks which we report to the caller as
1476 * EINTR. In that case, lf_purgelocks will have
1477 * removed our lock graph edges.
1479 * Note that it is possible to receive a signal after
1480 * we were successfully woken (and moved to the active
1481 * list) but before we resumed execution. In this
1482 * case, our lf_outedges list will be clear. We
1483 * pretend there was no error.
1485 * Note also, if we have been sleeping long enough, we
1486 * may now have incoming edges from some newer lock
1487 * which is waiting behind us in the queue.
1489 if (lock->lf_flags & F_INTR) {
1494 if (LIST_EMPTY(&lock->lf_outedges)) {
1497 lf_cancel_lock(state, lock);
1501 if (lockf_debug & 1) {
1502 lf_print("lf_setlock: granted", lock);
1508 * It looks like we are going to grant the lock. First add
1509 * edges from any currently pending lock that the new lock
1512 error = lf_add_incoming(state, 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 dependencies 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 lf_add_incoming(state, splitlock);
1843 lf_set_end(state, lock1, lock2->lf_start - 1, granted);
1846 * OK, now link it in
1848 lf_insert_lock(state, splitlock);
1852 STAILQ_ENTRY(lockdesc) link;
1856 STAILQ_HEAD(lockdesclist, lockdesc);
1859 lf_iteratelocks_sysid(int sysid, lf_iterator *fn, void *arg)
1862 struct lockf_entry *lf;
1863 struct lockdesc *ldesc;
1864 struct lockdesclist locks;
1868 * In order to keep the locking simple, we iterate over the
1869 * active lock lists to build a list of locks that need
1870 * releasing. We then call the iterator for each one in turn.
1872 * We take an extra reference to the vnode for the duration to
1873 * make sure it doesn't go away before we are finished.
1875 STAILQ_INIT(&locks);
1876 sx_xlock(&lf_lock_states_lock);
1877 LIST_FOREACH(ls, &lf_lock_states, ls_link) {
1878 sx_xlock(&ls->ls_lock);
1879 LIST_FOREACH(lf, &ls->ls_active, lf_link) {
1880 if (lf->lf_owner->lo_sysid != sysid)
1883 ldesc = malloc(sizeof(struct lockdesc), M_LOCKF,
1885 ldesc->vp = lf->lf_vnode;
1887 ldesc->fl.l_start = lf->lf_start;
1888 if (lf->lf_end == OFF_MAX)
1889 ldesc->fl.l_len = 0;
1892 lf->lf_end - lf->lf_start + 1;
1893 ldesc->fl.l_whence = SEEK_SET;
1894 ldesc->fl.l_type = F_UNLCK;
1895 ldesc->fl.l_pid = lf->lf_owner->lo_pid;
1896 ldesc->fl.l_sysid = sysid;
1897 STAILQ_INSERT_TAIL(&locks, ldesc, link);
1899 sx_xunlock(&ls->ls_lock);
1901 sx_xunlock(&lf_lock_states_lock);
1904 * Call the iterator function for each lock in turn. If the
1905 * iterator returns an error code, just free the rest of the
1906 * lockdesc structures.
1909 while ((ldesc = STAILQ_FIRST(&locks)) != NULL) {
1910 STAILQ_REMOVE_HEAD(&locks, link);
1912 error = fn(ldesc->vp, &ldesc->fl, arg);
1914 free(ldesc, M_LOCKF);
1921 lf_iteratelocks_vnode(struct vnode *vp, lf_iterator *fn, void *arg)
1924 struct lockf_entry *lf;
1925 struct lockdesc *ldesc;
1926 struct lockdesclist locks;
1930 * In order to keep the locking simple, we iterate over the
1931 * active lock lists to build a list of locks that need
1932 * releasing. We then call the iterator for each one in turn.
1934 * We take an extra reference to the vnode for the duration to
1935 * make sure it doesn't go away before we are finished.
1937 STAILQ_INIT(&locks);
1944 MPASS(ls->ls_threads >= 0);
1948 sx_xlock(&ls->ls_lock);
1949 LIST_FOREACH(lf, &ls->ls_active, lf_link) {
1950 ldesc = malloc(sizeof(struct lockdesc), M_LOCKF,
1952 ldesc->vp = lf->lf_vnode;
1954 ldesc->fl.l_start = lf->lf_start;
1955 if (lf->lf_end == OFF_MAX)
1956 ldesc->fl.l_len = 0;
1959 lf->lf_end - lf->lf_start + 1;
1960 ldesc->fl.l_whence = SEEK_SET;
1961 ldesc->fl.l_type = F_UNLCK;
1962 ldesc->fl.l_pid = lf->lf_owner->lo_pid;
1963 ldesc->fl.l_sysid = lf->lf_owner->lo_sysid;
1964 STAILQ_INSERT_TAIL(&locks, ldesc, link);
1966 sx_xunlock(&ls->ls_lock);
1968 MPASS(ls->ls_threads > 0);
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 for (i = 0; i < LOCK_OWNER_HASH_SIZE; i++) {
2015 sx_xlock(&lf_lock_owners[i].lock);
2016 LIST_FOREACH(lo, &lf_lock_owners[i].list, lo_link)
2017 if (lo->lo_sysid == sysid)
2018 count += lo->lo_refs;
2019 sx_xunlock(&lf_lock_owners[i].lock);
2028 * Return non-zero if y is reachable from x using a brute force
2029 * search. If reachable and path is non-null, return the route taken
2033 graph_reaches(struct owner_vertex *x, struct owner_vertex *y,
2034 struct owner_vertex_list *path)
2036 struct owner_edge *e;
2040 TAILQ_INSERT_HEAD(path, x, v_link);
2044 LIST_FOREACH(e, &x->v_outedges, e_outlink) {
2045 if (graph_reaches(e->e_to, y, path)) {
2047 TAILQ_INSERT_HEAD(path, x, v_link);
2055 * Perform consistency checks on the graph. Make sure the values of
2056 * v_order are correct. If checkorder is non-zero, check no vertex can
2057 * reach any other vertex with a smaller order.
2060 graph_check(struct owner_graph *g, int checkorder)
2064 for (i = 0; i < g->g_size; i++) {
2065 if (!g->g_vertices[i]->v_owner)
2067 KASSERT(g->g_vertices[i]->v_order == i,
2068 ("lock graph vertices disordered"));
2070 for (j = 0; j < i; j++) {
2071 if (!g->g_vertices[j]->v_owner)
2073 KASSERT(!graph_reaches(g->g_vertices[i],
2074 g->g_vertices[j], NULL),
2075 ("lock graph vertices disordered"));
2082 graph_print_vertices(struct owner_vertex_list *set)
2084 struct owner_vertex *v;
2087 TAILQ_FOREACH(v, set, v_link) {
2088 printf("%d:", v->v_order);
2089 lf_print_owner(v->v_owner);
2090 if (TAILQ_NEXT(v, v_link))
2099 * Calculate the sub-set of vertices v from the affected region [y..x]
2100 * where v is reachable from y. Return -1 if a loop was detected
2101 * (i.e. x is reachable from y, otherwise the number of vertices in
2105 graph_delta_forward(struct owner_graph *g, struct owner_vertex *x,
2106 struct owner_vertex *y, struct owner_vertex_list *delta)
2109 struct owner_vertex *v;
2110 struct owner_edge *e;
2114 * We start with a set containing just y. Then for each vertex
2115 * v in the set so far unprocessed, we add each vertex that v
2116 * has an out-edge to and that is within the affected region
2117 * [y..x]. If we see the vertex x on our travels, stop
2121 TAILQ_INSERT_TAIL(delta, y, v_link);
2126 LIST_FOREACH(e, &v->v_outedges, e_outlink) {
2129 if (e->e_to->v_order < x->v_order
2130 && e->e_to->v_gen != gen) {
2131 e->e_to->v_gen = gen;
2132 TAILQ_INSERT_TAIL(delta, e->e_to, v_link);
2136 v = TAILQ_NEXT(v, v_link);
2143 * Calculate the sub-set of vertices v from the affected region [y..x]
2144 * where v reaches x. Return the number of vertices in this subset.
2147 graph_delta_backward(struct owner_graph *g, struct owner_vertex *x,
2148 struct owner_vertex *y, struct owner_vertex_list *delta)
2151 struct owner_vertex *v;
2152 struct owner_edge *e;
2156 * We start with a set containing just x. Then for each vertex
2157 * v in the set so far unprocessed, we add each vertex that v
2158 * has an in-edge from and that is within the affected region
2162 TAILQ_INSERT_TAIL(delta, x, v_link);
2167 LIST_FOREACH(e, &v->v_inedges, e_inlink) {
2168 if (e->e_from->v_order > y->v_order
2169 && e->e_from->v_gen != gen) {
2170 e->e_from->v_gen = gen;
2171 TAILQ_INSERT_HEAD(delta, e->e_from, v_link);
2175 v = TAILQ_PREV(v, owner_vertex_list, v_link);
2182 graph_add_indices(int *indices, int n, struct owner_vertex_list *set)
2184 struct owner_vertex *v;
2187 TAILQ_FOREACH(v, set, v_link) {
2189 i > 0 && indices[i - 1] > v->v_order; i--)
2191 for (j = n - 1; j >= i; j--)
2192 indices[j + 1] = indices[j];
2193 indices[i] = v->v_order;
2201 graph_assign_indices(struct owner_graph *g, int *indices, int nextunused,
2202 struct owner_vertex_list *set)
2204 struct owner_vertex *v, *vlowest;
2206 while (!TAILQ_EMPTY(set)) {
2208 TAILQ_FOREACH(v, set, v_link) {
2209 if (!vlowest || v->v_order < vlowest->v_order)
2212 TAILQ_REMOVE(set, vlowest, v_link);
2213 vlowest->v_order = indices[nextunused];
2214 g->g_vertices[vlowest->v_order] = vlowest;
2218 return (nextunused);
2222 graph_add_edge(struct owner_graph *g, struct owner_vertex *x,
2223 struct owner_vertex *y)
2225 struct owner_edge *e;
2226 struct owner_vertex_list deltaF, deltaB;
2231 sx_assert(&lf_owner_graph_lock, SX_XLOCKED);
2233 LIST_FOREACH(e, &x->v_outedges, e_outlink) {
2241 if (lockf_debug & 8) {
2242 printf("adding edge %d:", x->v_order);
2243 lf_print_owner(x->v_owner);
2244 printf(" -> %d:", y->v_order);
2245 lf_print_owner(y->v_owner);
2249 if (y->v_order < x->v_order) {
2251 * The new edge violates the order. First find the set
2252 * of affected vertices reachable from y (deltaF) and
2253 * the set of affect vertices affected that reach x
2254 * (deltaB), using the graph generation number to
2255 * detect whether we have visited a given vertex
2256 * already. We re-order the graph so that each vertex
2257 * in deltaB appears before each vertex in deltaF.
2259 * If x is a member of deltaF, then the new edge would
2260 * create a cycle. Otherwise, we may assume that
2261 * deltaF and deltaB are disjoint.
2264 if (g->g_gen == 0) {
2268 for (vi = 0; vi < g->g_size; vi++) {
2269 g->g_vertices[vi]->v_gen = 0;
2273 nF = graph_delta_forward(g, x, y, &deltaF);
2276 if (lockf_debug & 8) {
2277 struct owner_vertex_list path;
2278 printf("deadlock: ");
2280 graph_reaches(y, x, &path);
2281 graph_print_vertices(&path);
2288 if (lockf_debug & 8) {
2289 printf("re-ordering graph vertices\n");
2290 printf("deltaF = ");
2291 graph_print_vertices(&deltaF);
2295 nB = graph_delta_backward(g, x, y, &deltaB);
2298 if (lockf_debug & 8) {
2299 printf("deltaB = ");
2300 graph_print_vertices(&deltaB);
2305 * We first build a set of vertex indices (vertex
2306 * order values) that we may use, then we re-assign
2307 * orders first to those vertices in deltaB, then to
2308 * deltaF. Note that the contents of deltaF and deltaB
2309 * may be partially disordered - we perform an
2310 * insertion sort while building our index set.
2312 indices = g->g_indexbuf;
2313 n = graph_add_indices(indices, 0, &deltaF);
2314 graph_add_indices(indices, n, &deltaB);
2317 * We must also be sure to maintain the relative
2318 * ordering of deltaF and deltaB when re-assigning
2319 * vertices. We do this by iteratively removing the
2320 * lowest ordered element from the set and assigning
2321 * it the next value from our new ordering.
2323 i = graph_assign_indices(g, indices, 0, &deltaB);
2324 graph_assign_indices(g, indices, i, &deltaF);
2327 if (lockf_debug & 8) {
2328 struct owner_vertex_list set;
2330 for (i = 0; i < nB + nF; i++)
2331 TAILQ_INSERT_TAIL(&set,
2332 g->g_vertices[indices[i]], v_link);
2333 printf("new ordering = ");
2334 graph_print_vertices(&set);
2339 KASSERT(x->v_order < y->v_order, ("Failed to re-order graph"));
2342 if (lockf_debug & 8) {
2343 graph_check(g, TRUE);
2347 e = malloc(sizeof(struct owner_edge), M_LOCKF, M_WAITOK);
2349 LIST_INSERT_HEAD(&x->v_outedges, e, e_outlink);
2350 LIST_INSERT_HEAD(&y->v_inedges, e, e_inlink);
2359 * Remove an edge x->y from the graph.
2362 graph_remove_edge(struct owner_graph *g, struct owner_vertex *x,
2363 struct owner_vertex *y)
2365 struct owner_edge *e;
2367 sx_assert(&lf_owner_graph_lock, SX_XLOCKED);
2369 LIST_FOREACH(e, &x->v_outedges, e_outlink) {
2373 KASSERT(e, ("Removing non-existent edge from deadlock graph"));
2376 if (e->e_refs == 0) {
2378 if (lockf_debug & 8) {
2379 printf("removing edge %d:", x->v_order);
2380 lf_print_owner(x->v_owner);
2381 printf(" -> %d:", y->v_order);
2382 lf_print_owner(y->v_owner);
2386 LIST_REMOVE(e, e_outlink);
2387 LIST_REMOVE(e, e_inlink);
2393 * Allocate a vertex from the free list. Return ENOMEM if there are
2396 static struct owner_vertex *
2397 graph_alloc_vertex(struct owner_graph *g, struct lock_owner *lo)
2399 struct owner_vertex *v;
2401 sx_assert(&lf_owner_graph_lock, SX_XLOCKED);
2403 v = malloc(sizeof(struct owner_vertex), M_LOCKF, M_WAITOK);
2404 if (g->g_size == g->g_space) {
2405 g->g_vertices = realloc(g->g_vertices,
2406 2 * g->g_space * sizeof(struct owner_vertex *),
2408 free(g->g_indexbuf, M_LOCKF);
2409 g->g_indexbuf = malloc(2 * g->g_space * sizeof(int),
2411 g->g_space = 2 * g->g_space;
2413 v->v_order = g->g_size;
2414 v->v_gen = g->g_gen;
2415 g->g_vertices[g->g_size] = v;
2418 LIST_INIT(&v->v_outedges);
2419 LIST_INIT(&v->v_inedges);
2426 graph_free_vertex(struct owner_graph *g, struct owner_vertex *v)
2428 struct owner_vertex *w;
2431 sx_assert(&lf_owner_graph_lock, SX_XLOCKED);
2433 KASSERT(LIST_EMPTY(&v->v_outedges), ("Freeing vertex with edges"));
2434 KASSERT(LIST_EMPTY(&v->v_inedges), ("Freeing vertex with edges"));
2437 * Remove from the graph's array and close up the gap,
2438 * renumbering the other vertices.
2440 for (i = v->v_order + 1; i < g->g_size; i++) {
2441 w = g->g_vertices[i];
2443 g->g_vertices[i - 1] = w;
2450 static struct owner_graph *
2451 graph_init(struct owner_graph *g)
2454 g->g_vertices = malloc(10 * sizeof(struct owner_vertex *),
2458 g->g_indexbuf = malloc(g->g_space * sizeof(int), M_LOCKF, M_WAITOK);
2464 struct kinfo_lockf_linked {
2465 struct kinfo_lockf kl;
2467 STAILQ_ENTRY(kinfo_lockf_linked) link;
2471 vfs_report_lockf(struct mount *mp, struct sbuf *sb)
2474 struct lockf_entry *lf;
2475 struct kinfo_lockf_linked *klf;
2477 struct ucred *ucred;
2478 char *fullpath, *freepath;
2480 STAILQ_HEAD(, kinfo_lockf_linked) locks;
2483 STAILQ_INIT(&locks);
2484 sx_slock(&lf_lock_states_lock);
2485 LIST_FOREACH(ls, &lf_lock_states, ls_link) {
2486 sx_slock(&ls->ls_lock);
2487 LIST_FOREACH(lf, &ls->ls_active, lf_link) {
2489 if (VN_IS_DOOMED(vp) || vp->v_mount != mp)
2492 klf = malloc(sizeof(struct kinfo_lockf_linked),
2493 M_LOCKF, M_WAITOK | M_ZERO);
2495 klf->kl.kl_structsize = sizeof(struct kinfo_lockf);
2496 klf->kl.kl_start = lf->lf_start;
2497 klf->kl.kl_len = lf->lf_end == OFF_MAX ? 0 :
2498 lf->lf_end - lf->lf_start + 1;
2499 klf->kl.kl_rw = lf->lf_type == F_RDLCK ?
2500 KLOCKF_RW_READ : KLOCKF_RW_WRITE;
2501 if (lf->lf_owner->lo_sysid != 0) {
2502 klf->kl.kl_pid = lf->lf_owner->lo_pid;
2503 klf->kl.kl_sysid = lf->lf_owner->lo_sysid;
2504 klf->kl.kl_type = KLOCKF_TYPE_REMOTE;
2505 } else if (lf->lf_owner->lo_pid == -1) {
2506 klf->kl.kl_pid = -1;
2507 klf->kl.kl_sysid = 0;
2508 klf->kl.kl_type = KLOCKF_TYPE_FLOCK;
2510 klf->kl.kl_pid = lf->lf_owner->lo_pid;
2511 klf->kl.kl_sysid = 0;
2512 klf->kl.kl_type = KLOCKF_TYPE_PID;
2514 STAILQ_INSERT_TAIL(&locks, klf, link);
2516 sx_sunlock(&ls->ls_lock);
2518 sx_sunlock(&lf_lock_states_lock);
2521 ucred = curthread->td_ucred;
2522 while ((klf = STAILQ_FIRST(&locks)) != NULL) {
2523 STAILQ_REMOVE_HEAD(&locks, link);
2525 if (gerror == 0 && vn_lock(vp, LK_SHARED) == 0) {
2526 error = prison_canseemount(ucred, vp->v_mount);
2528 error = VOP_STAT(vp, &stt, ucred, NOCRED,
2532 klf->kl.kl_file_fsid = stt.st_dev;
2533 klf->kl.kl_file_rdev = stt.st_rdev;
2534 klf->kl.kl_file_fileid = stt.st_ino;
2537 error = vn_fullpath(vp, &fullpath, &freepath);
2539 strlcpy(klf->kl.kl_path, fullpath,
2540 sizeof(klf->kl.kl_path));
2541 free(freepath, M_TEMP);
2542 if (sbuf_bcat(sb, &klf->kl,
2543 klf->kl.kl_structsize) != 0) {
2544 gerror = sbuf_error(sb);
2556 sysctl_kern_lockf_run(struct sbuf *sb)
2562 mtx_lock(&mountlist_mtx);
2563 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2564 error = vfs_busy(mp, MBF_MNTLSTLOCK);
2567 error = mp->mnt_op->vfs_report_lockf(mp, sb);
2568 mtx_lock(&mountlist_mtx);
2573 mtx_unlock(&mountlist_mtx);
2578 sysctl_kern_lockf(SYSCTL_HANDLER_ARGS)
2583 sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_lockf) * 5, req);
2584 sbuf_clear_flags(&sb, SBUF_INCLUDENUL);
2585 error = sysctl_kern_lockf_run(&sb);
2586 error2 = sbuf_finish(&sb);
2588 return (error != 0 ? error : error2);
2590 SYSCTL_PROC(_kern, KERN_LOCKF, lockf,
2591 CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE,
2592 0, 0, sysctl_kern_lockf, "S,lockf",
2593 "Advisory locks table");
2597 * Print description of a lock owner
2600 lf_print_owner(struct lock_owner *lo)
2603 if (lo->lo_flags & F_REMOTE) {
2604 printf("remote pid %d, system %d",
2605 lo->lo_pid, lo->lo_sysid);
2606 } else if (lo->lo_flags & F_FLOCK) {
2607 printf("file %p", lo->lo_id);
2609 printf("local pid %d", lo->lo_pid);
2617 lf_print(char *tag, struct lockf_entry *lock)
2620 printf("%s: lock %p for ", tag, (void *)lock);
2621 lf_print_owner(lock->lf_owner);
2622 printf("\nvnode %p", lock->lf_vnode);
2623 VOP_PRINT(lock->lf_vnode);
2624 printf(" %s, start %jd, end ",
2625 lock->lf_type == F_RDLCK ? "shared" :
2626 lock->lf_type == F_WRLCK ? "exclusive" :
2627 lock->lf_type == F_UNLCK ? "unlock" : "unknown",
2628 (intmax_t)lock->lf_start);
2629 if (lock->lf_end == OFF_MAX)
2632 printf("%jd", (intmax_t)lock->lf_end);
2633 if (!LIST_EMPTY(&lock->lf_outedges))
2634 printf(" block %p\n",
2635 (void *)LIST_FIRST(&lock->lf_outedges)->le_to);
2641 lf_printlist(char *tag, struct lockf_entry *lock)
2643 struct lockf_entry *lf, *blk;
2644 struct lockf_edge *e;
2646 printf("%s: Lock list for vnode %p:\n", tag, lock->lf_vnode);
2647 LIST_FOREACH(lf, &lock->lf_vnode->v_lockf->ls_active, lf_link) {
2648 printf("\tlock %p for ",(void *)lf);
2649 lf_print_owner(lock->lf_owner);
2650 printf(", %s, start %jd, end %jd",
2651 lf->lf_type == F_RDLCK ? "shared" :
2652 lf->lf_type == F_WRLCK ? "exclusive" :
2653 lf->lf_type == F_UNLCK ? "unlock" :
2654 "unknown", (intmax_t)lf->lf_start, (intmax_t)lf->lf_end);
2655 LIST_FOREACH(e, &lf->lf_outedges, le_outlink) {
2657 printf("\n\t\tlock request %p for ", (void *)blk);
2658 lf_print_owner(blk->lf_owner);
2659 printf(", %s, start %jd, end %jd",
2660 blk->lf_type == F_RDLCK ? "shared" :
2661 blk->lf_type == F_WRLCK ? "exclusive" :
2662 blk->lf_type == F_UNLCK ? "unlock" :
2663 "unknown", (intmax_t)blk->lf_start,
2664 (intmax_t)blk->lf_end);
2665 if (!LIST_EMPTY(&blk->lf_inedges))
2666 panic("lf_printlist: bad list");
2671 #endif /* LOCKF_DEBUG */