1 /*        $NetBSD: subr_lockdebug.c,v 1.83 2022/09/02 06:01:38 nakayama Exp $   */
2 
3 /*-
4  * Copyright (c) 2006, 2007, 2008, 2020 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Andrew Doran.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*
33  * Basic lock debugging code shared among lock primitives.
34  */
35 
36 #include <sys/cdefs.h>
37 __KERNEL_RCSID(0, "$NetBSD: subr_lockdebug.c,v 1.83 2022/09/02 06:01:38 nakayama Exp $");
38 
39 #ifdef _KERNEL_OPT
40 #include "opt_ddb.h"
41 #endif
42 
43 #include <sys/param.h>
44 #include <sys/proc.h>
45 #include <sys/systm.h>
46 #include <sys/kernel.h>
47 #include <sys/kmem.h>
48 #include <sys/lockdebug.h>
49 #include <sys/sleepq.h>
50 #include <sys/cpu.h>
51 #include <sys/atomic.h>
52 #include <sys/lock.h>
53 #include <sys/rbtree.h>
54 #include <sys/ksyms.h>
55 #include <sys/kcov.h>
56 
57 #include <machine/lock.h>
58 
59 #ifdef DDB
60 #include <machine/db_machdep.h>
61 #include <ddb/db_interface.h>
62 #include <ddb/db_access.h>
63 #include <ddb/db_sym.h>
64 #endif
65 
66 unsigned int                  ld_panic;
67 
68 #ifdef LOCKDEBUG
69 
70 #ifdef __ia64__
71 #define   LD_BATCH_SHIFT      16
72 #else
73 #define   LD_BATCH_SHIFT      9
74 #endif
75 #define   LD_BATCH  (1 << LD_BATCH_SHIFT)
76 #define   LD_BATCH_MASK       (LD_BATCH - 1)
77 #define   LD_MAX_LOCKS        1048576
78 #define   LD_SLOP             16
79 
80 #define   LD_LOCKED 0x01
81 #define   LD_SLEEPER          0x02
82 
83 #define   LD_WRITE_LOCK       0x80000000
84 
85 typedef struct lockdebug {
86           struct rb_node      ld_rb_node;
87           __cpu_simple_lock_t ld_spinlock;
88           _TAILQ_ENTRY(struct lockdebug, volatile) ld_chain;
89           _TAILQ_ENTRY(struct lockdebug, volatile) ld_achain;
90           volatile void       *ld_lock;
91           lockops_t *ld_lockops;
92           struct lwp          *ld_lwp;
93           uintptr_t ld_locked;
94           uintptr_t ld_unlocked;
95           uintptr_t ld_initaddr;
96           uint16_t  ld_shares;
97           uint16_t  ld_cpu;
98           uint8_t             ld_flags;
99           uint8_t             ld_shwant;          /* advisory */
100           uint8_t             ld_exwant;          /* advisory */
101           uint8_t             ld_unused;
102 } volatile lockdebug_t;
103 
104 typedef _TAILQ_HEAD(lockdebuglist, struct lockdebug, volatile) lockdebuglist_t;
105 
106 __cpu_simple_lock_t ld_mod_lk;
107 lockdebuglist_t               ld_free = TAILQ_HEAD_INITIALIZER(ld_free);
108 #ifdef _KERNEL
109 lockdebuglist_t               ld_all = TAILQ_HEAD_INITIALIZER(ld_all);
110 #else
111 extern lockdebuglist_t        ld_all;
112 #define cpu_name(a) "?"
113 #define cpu_index(a)          -1
114 #define curlwp                NULL
115 #endif /* _KERNEL */
116 int                           ld_nfree;
117 int                           ld_freeptr;
118 int                           ld_recurse;
119 bool                          ld_nomore;
120 lockdebug_t                   ld_prime[LD_BATCH];
121 
122 #ifdef _KERNEL
123 static void         lockdebug_abort1(const char *, size_t, lockdebug_t *, int,
124     const char *, bool);
125 static int          lockdebug_more(int);
126 static void         lockdebug_init(void);
127 static void         lockdebug_dump(lwp_t *, lockdebug_t *,
128     void (*)(const char *, ...)
129     __printflike(1, 2));
130 
131 static signed int
ld_rbto_compare_nodes(void * ctx,const void * n1,const void * n2)132 ld_rbto_compare_nodes(void *ctx, const void *n1, const void *n2)
133 {
134           const lockdebug_t *ld1 = n1;
135           const lockdebug_t *ld2 = n2;
136           const uintptr_t a = (uintptr_t)ld1->ld_lock;
137           const uintptr_t b = (uintptr_t)ld2->ld_lock;
138 
139           if (a < b)
140                     return -1;
141           if (a > b)
142                     return 1;
143           return 0;
144 }
145 
146 static signed int
ld_rbto_compare_key(void * ctx,const void * n,const void * key)147 ld_rbto_compare_key(void *ctx, const void *n, const void *key)
148 {
149           const lockdebug_t *ld = n;
150           const uintptr_t a = (uintptr_t)ld->ld_lock;
151           const uintptr_t b = (uintptr_t)key;
152 
153           if (a < b)
154                     return -1;
155           if (a > b)
156                     return 1;
157           return 0;
158 }
159 
160 static rb_tree_t ld_rb_tree;
161 
162 static const rb_tree_ops_t ld_rb_tree_ops = {
163           .rbto_compare_nodes = ld_rbto_compare_nodes,
164           .rbto_compare_key = ld_rbto_compare_key,
165           .rbto_node_offset = offsetof(lockdebug_t, ld_rb_node),
166           .rbto_context = NULL
167 };
168 
169 static inline lockdebug_t *
lockdebug_lookup1(const volatile void * lock)170 lockdebug_lookup1(const volatile void *lock)
171 {
172           lockdebug_t *ld;
173           struct cpu_info *ci;
174 
175           ci = curcpu();
176           __cpu_simple_lock(&ci->ci_data.cpu_ld_lock);
177           ld = rb_tree_find_node(&ld_rb_tree, (void *)(intptr_t)lock);
178           __cpu_simple_unlock(&ci->ci_data.cpu_ld_lock);
179           if (ld == NULL) {
180                     return NULL;
181           }
182           __cpu_simple_lock(&ld->ld_spinlock);
183 
184           return ld;
185 }
186 
187 static void
lockdebug_lock_cpus(void)188 lockdebug_lock_cpus(void)
189 {
190           CPU_INFO_ITERATOR cii;
191           struct cpu_info *ci;
192 
193           for (CPU_INFO_FOREACH(cii, ci)) {
194                     __cpu_simple_lock(&ci->ci_data.cpu_ld_lock);
195           }
196 }
197 
198 static void
lockdebug_unlock_cpus(void)199 lockdebug_unlock_cpus(void)
200 {
201           CPU_INFO_ITERATOR cii;
202           struct cpu_info *ci;
203 
204           for (CPU_INFO_FOREACH(cii, ci)) {
205                     __cpu_simple_unlock(&ci->ci_data.cpu_ld_lock);
206           }
207 }
208 
209 /*
210  * lockdebug_lookup:
211  *
212  *        Find a lockdebug structure by a pointer to a lock and return it locked.
213  */
214 static inline lockdebug_t *
lockdebug_lookup(const char * func,size_t line,const volatile void * lock,uintptr_t where)215 lockdebug_lookup(const char *func, size_t line, const volatile void *lock,
216     uintptr_t where)
217 {
218           lockdebug_t *ld;
219 
220           kcov_silence_enter();
221           ld = lockdebug_lookup1(lock);
222           kcov_silence_leave();
223 
224           if (__predict_false(ld == NULL)) {
225                     panic("%s,%zu: uninitialized lock (lock=%p, from=%08"
226                         PRIxPTR ")", func, line, lock, where);
227           }
228           return ld;
229 }
230 
231 /*
232  * lockdebug_init:
233  *
234  *        Initialize the lockdebug system.  Allocate an initial pool of
235  *        lockdebug structures before the VM system is up and running.
236  */
237 static void
lockdebug_init(void)238 lockdebug_init(void)
239 {
240           lockdebug_t *ld;
241           int i;
242 
243           TAILQ_INIT(&curcpu()->ci_data.cpu_ld_locks);
244           TAILQ_INIT(&curlwp->l_ld_locks);
245           __cpu_simple_lock_init(&curcpu()->ci_data.cpu_ld_lock);
246           __cpu_simple_lock_init(&ld_mod_lk);
247 
248           rb_tree_init(&ld_rb_tree, &ld_rb_tree_ops);
249 
250           ld = ld_prime;
251           for (i = 1, ld++; i < LD_BATCH; i++, ld++) {
252                     __cpu_simple_lock_init(&ld->ld_spinlock);
253                     TAILQ_INSERT_TAIL(&ld_free, ld, ld_chain);
254                     TAILQ_INSERT_TAIL(&ld_all, ld, ld_achain);
255           }
256           ld_freeptr = 1;
257           ld_nfree = LD_BATCH - 1;
258 }
259 
260 /*
261  * lockdebug_alloc:
262  *
263  *        A lock is being initialized, so allocate an associated debug
264  *        structure.
265  */
266 bool
lockdebug_alloc(const char * func,size_t line,volatile void * lock,lockops_t * lo,uintptr_t initaddr)267 lockdebug_alloc(const char *func, size_t line, volatile void *lock,
268     lockops_t *lo, uintptr_t initaddr)
269 {
270           struct cpu_info *ci;
271           lockdebug_t *ld;
272           int s;
273 
274           if (__predict_false(lo == NULL || panicstr != NULL || ld_panic))
275                     return false;
276           if (__predict_false(ld_freeptr == 0))
277                     lockdebug_init();
278 
279           s = splhigh();
280           __cpu_simple_lock(&ld_mod_lk);
281           if (__predict_false((ld = lockdebug_lookup1(lock)) != NULL)) {
282                     __cpu_simple_unlock(&ld_mod_lk);
283                     lockdebug_abort1(func, line, ld, s, "already initialized",
284                         true);
285                     return false;
286           }
287 
288           /*
289            * Pinch a new debug structure.  We may recurse because we call
290            * kmem_alloc(), which may need to initialize new locks somewhere
291            * down the path.  If not recursing, we try to maintain at least
292            * LD_SLOP structures free, which should hopefully be enough to
293            * satisfy kmem_alloc().  If we can't provide a structure, not to
294            * worry: we'll just mark the lock as not having an ID.
295            */
296           ci = curcpu();
297           ci->ci_lkdebug_recurse++;
298           if (TAILQ_EMPTY(&ld_free)) {
299                     if (ci->ci_lkdebug_recurse > 1 || ld_nomore) {
300                               ci->ci_lkdebug_recurse--;
301                               __cpu_simple_unlock(&ld_mod_lk);
302                               splx(s);
303                               return false;
304                     }
305                     s = lockdebug_more(s);
306           } else if (ci->ci_lkdebug_recurse == 1 && ld_nfree < LD_SLOP) {
307                     s = lockdebug_more(s);
308           }
309           if (__predict_false((ld = TAILQ_FIRST(&ld_free)) == NULL)) {
310                     __cpu_simple_unlock(&ld_mod_lk);
311                     splx(s);
312                     return false;
313           }
314           TAILQ_REMOVE(&ld_free, ld, ld_chain);
315           ld_nfree--;
316           ci->ci_lkdebug_recurse--;
317 
318           if (__predict_false(ld->ld_lock != NULL)) {
319                     panic("%s,%zu: corrupt table ld %p", func, line, ld);
320           }
321 
322           /* Initialise the structure. */
323           ld->ld_lock = lock;
324           ld->ld_lockops = lo;
325           ld->ld_locked = 0;
326           ld->ld_unlocked = 0;
327           ld->ld_lwp = NULL;
328           ld->ld_initaddr = initaddr;
329           ld->ld_flags = (lo->lo_type == LOCKOPS_SLEEP ? LD_SLEEPER : 0);
330           lockdebug_lock_cpus();
331           (void)rb_tree_insert_node(&ld_rb_tree, __UNVOLATILE(ld));
332           lockdebug_unlock_cpus();
333           __cpu_simple_unlock(&ld_mod_lk);
334 
335           splx(s);
336           return true;
337 }
338 
339 /*
340  * lockdebug_free:
341  *
342  *        A lock is being destroyed, so release debugging resources.
343  */
344 void
lockdebug_free(const char * func,size_t line,volatile void * lock)345 lockdebug_free(const char *func, size_t line, volatile void *lock)
346 {
347           lockdebug_t *ld;
348           int s;
349 
350           if (__predict_false(panicstr != NULL || ld_panic))
351                     return;
352 
353           s = splhigh();
354           __cpu_simple_lock(&ld_mod_lk);
355           ld = lockdebug_lookup(func, line, lock,
356               (uintptr_t) __builtin_return_address(0));
357           if (__predict_false(ld == NULL)) {
358                     __cpu_simple_unlock(&ld_mod_lk);
359                     panic("%s,%zu: destroying uninitialized object %p"
360                         "(ld_lock=%p)", func, line, lock, ld->ld_lock);
361                     return;
362           }
363           if (__predict_false((ld->ld_flags & LD_LOCKED) != 0 ||
364               ld->ld_shares != 0)) {
365                     __cpu_simple_unlock(&ld_mod_lk);
366                     lockdebug_abort1(func, line, ld, s, "is locked or in use",
367                         true);
368                     return;
369           }
370           lockdebug_lock_cpus();
371           rb_tree_remove_node(&ld_rb_tree, __UNVOLATILE(ld));
372           lockdebug_unlock_cpus();
373           ld->ld_lock = NULL;
374           TAILQ_INSERT_TAIL(&ld_free, ld, ld_chain);
375           ld_nfree++;
376           __cpu_simple_unlock(&ld->ld_spinlock);
377           __cpu_simple_unlock(&ld_mod_lk);
378           splx(s);
379 }
380 
381 /*
382  * lockdebug_more:
383  *
384  *        Allocate a batch of debug structures and add to the free list.
385  *        Must be called with ld_mod_lk held.
386  */
387 static int
lockdebug_more(int s)388 lockdebug_more(int s)
389 {
390           lockdebug_t *ld;
391           void *block;
392           int i, base, m;
393 
394           /*
395            * Can't call kmem_alloc() if in interrupt context.  XXX We could
396            * deadlock, because we don't know which locks the caller holds.
397            */
398           if (cpu_intr_p() || cpu_softintr_p()) {
399                     return s;
400           }
401 
402           while (ld_nfree < LD_SLOP) {
403                     __cpu_simple_unlock(&ld_mod_lk);
404                     splx(s);
405                     block = kmem_zalloc(LD_BATCH * sizeof(lockdebug_t), KM_SLEEP);
406                     s = splhigh();
407                     __cpu_simple_lock(&ld_mod_lk);
408 
409                     if (ld_nfree > LD_SLOP) {
410                               /* Somebody beat us to it. */
411                               __cpu_simple_unlock(&ld_mod_lk);
412                               splx(s);
413                               kmem_free(block, LD_BATCH * sizeof(lockdebug_t));
414                               s = splhigh();
415                               __cpu_simple_lock(&ld_mod_lk);
416                               continue;
417                     }
418 
419                     base = ld_freeptr;
420                     ld_nfree += LD_BATCH;
421                     ld = block;
422                     base <<= LD_BATCH_SHIFT;
423                     m = uimin(LD_MAX_LOCKS, base + LD_BATCH);
424 
425                     if (m == LD_MAX_LOCKS)
426                               ld_nomore = true;
427 
428                     for (i = base; i < m; i++, ld++) {
429                               __cpu_simple_lock_init(&ld->ld_spinlock);
430                               TAILQ_INSERT_TAIL(&ld_free, ld, ld_chain);
431                               TAILQ_INSERT_TAIL(&ld_all, ld, ld_achain);
432                     }
433 
434                     membar_producer();
435           }
436 
437           return s;
438 }
439 
440 /*
441  * lockdebug_wantlock:
442  *
443  *        Process the preamble to a lock acquire.  The "shared"
444  *        parameter controls which ld_{ex,sh}want counter is
445  *        updated; a negative value of shared updates neither.
446  */
447 void
lockdebug_wantlock(const char * func,size_t line,const volatile void * lock,uintptr_t where,int shared)448 lockdebug_wantlock(const char *func, size_t line,
449     const volatile void *lock, uintptr_t where, int shared)
450 {
451           struct lwp *l = curlwp;
452           lockdebug_t *ld;
453           bool recurse;
454           int s;
455 
456           (void)shared;
457           recurse = false;
458 
459           if (__predict_false(panicstr != NULL || ld_panic))
460                     return;
461 
462           s = splhigh();
463           if ((ld = lockdebug_lookup(func, line, lock, where)) == NULL) {
464                     splx(s);
465                     return;
466           }
467           if ((ld->ld_flags & LD_LOCKED) != 0 || ld->ld_shares != 0) {
468                     if ((ld->ld_flags & LD_SLEEPER) != 0) {
469                               if (ld->ld_lwp == l)
470                                         recurse = true;
471                     } else if (ld->ld_cpu == (uint16_t)cpu_index(curcpu()))
472                               recurse = true;
473           }
474           if (cpu_intr_p()) {
475                     if (__predict_false((ld->ld_flags & LD_SLEEPER) != 0)) {
476                               lockdebug_abort1(func, line, ld, s,
477                                   "acquiring sleep lock from interrupt context",
478                                   true);
479                               return;
480                     }
481           }
482           if (shared > 0)
483                     ld->ld_shwant++;
484           else if (shared == 0)
485                     ld->ld_exwant++;
486           if (__predict_false(recurse)) {
487                     lockdebug_abort1(func, line, ld, s, "locking against myself",
488                         true);
489                     return;
490           }
491           if (l->l_ld_wanted == NULL) {
492                     l->l_ld_wanted = ld;
493           }
494           __cpu_simple_unlock(&ld->ld_spinlock);
495           splx(s);
496 }
497 
498 /*
499  * lockdebug_locked:
500  *
501  *        Process a lock acquire operation.
502  */
503 void
lockdebug_locked(const char * func,size_t line,volatile void * lock,void * cvlock,uintptr_t where,int shared)504 lockdebug_locked(const char *func, size_t line,
505     volatile void *lock, void *cvlock, uintptr_t where, int shared)
506 {
507           struct lwp *l = curlwp;
508           lockdebug_t *ld;
509           int s;
510 
511           if (__predict_false(panicstr != NULL || ld_panic))
512                     return;
513 
514           s = splhigh();
515           if ((ld = lockdebug_lookup(func, line, lock, where)) == NULL) {
516                     splx(s);
517                     return;
518           }
519           if (shared) {
520                     l->l_shlocks++;
521                     ld->ld_locked = where;
522                     ld->ld_shares++;
523                     ld->ld_shwant--;
524           } else {
525                     if (__predict_false((ld->ld_flags & LD_LOCKED) != 0)) {
526                               lockdebug_abort1(func, line, ld, s, "already locked",
527                                   true);
528                               return;
529                     }
530                     ld->ld_flags |= LD_LOCKED;
531                     ld->ld_locked = where;
532                     ld->ld_exwant--;
533                     if ((ld->ld_flags & LD_SLEEPER) != 0) {
534                               TAILQ_INSERT_TAIL(&l->l_ld_locks, ld, ld_chain);
535                     } else {
536                               TAILQ_INSERT_TAIL(&curcpu()->ci_data.cpu_ld_locks,
537                                   ld, ld_chain);
538                     }
539           }
540           ld->ld_cpu = (uint16_t)cpu_index(curcpu());
541           ld->ld_lwp = l;
542           __cpu_simple_unlock(&ld->ld_spinlock);
543           if (l->l_ld_wanted == ld) {
544                     l->l_ld_wanted = NULL;
545           }
546           splx(s);
547 }
548 
549 /*
550  * lockdebug_unlocked:
551  *
552  *        Process a lock release operation.
553  */
554 void
lockdebug_unlocked(const char * func,size_t line,volatile void * lock,uintptr_t where,int shared)555 lockdebug_unlocked(const char *func, size_t line,
556     volatile void *lock, uintptr_t where, int shared)
557 {
558           struct lwp *l = curlwp;
559           lockdebug_t *ld;
560           int s;
561 
562           if (__predict_false(panicstr != NULL || ld_panic))
563                     return;
564 
565           s = splhigh();
566           if ((ld = lockdebug_lookup(func, line, lock, where)) == NULL) {
567                     splx(s);
568                     return;
569           }
570           if (shared) {
571                     if (__predict_false(l->l_shlocks == 0)) {
572                               lockdebug_abort1(func, line, ld, s,
573                                   "no shared locks held by LWP", true);
574                               return;
575                     }
576                     if (__predict_false(ld->ld_shares == 0)) {
577                               lockdebug_abort1(func, line, ld, s,
578                                   "no shared holds on this lock", true);
579                               return;
580                     }
581                     l->l_shlocks--;
582                     ld->ld_shares--;
583                     if (ld->ld_lwp == l) {
584                               ld->ld_unlocked = where;
585                               ld->ld_lwp = NULL;
586                     }
587                     if (ld->ld_cpu == (uint16_t)cpu_index(curcpu()))
588                               ld->ld_cpu = (uint16_t)-1;
589           } else {
590                     if (__predict_false((ld->ld_flags & LD_LOCKED) == 0)) {
591                               lockdebug_abort1(func, line, ld, s, "not locked", true);
592                               return;
593                     }
594 
595                     if ((ld->ld_flags & LD_SLEEPER) != 0) {
596                               if (__predict_false(ld->ld_lwp != curlwp)) {
597                                         lockdebug_abort1(func, line, ld, s,
598                                             "not held by current LWP", true);
599                                         return;
600                               }
601                               TAILQ_REMOVE(&l->l_ld_locks, ld, ld_chain);
602                     } else {
603                               uint16_t idx = (uint16_t)cpu_index(curcpu());
604                               if (__predict_false(ld->ld_cpu != idx)) {
605                                         lockdebug_abort1(func, line, ld, s,
606                                             "not held by current CPU", true);
607                                         return;
608                               }
609                               TAILQ_REMOVE(&curcpu()->ci_data.cpu_ld_locks, ld,
610                                   ld_chain);
611                     }
612                     ld->ld_flags &= ~LD_LOCKED;
613                     ld->ld_unlocked = where;
614                     ld->ld_lwp = NULL;
615           }
616           __cpu_simple_unlock(&ld->ld_spinlock);
617           splx(s);
618 }
619 
620 /*
621  * lockdebug_barrier:
622  *
623  *        Panic if we hold more than one specified lock, and optionally, if we
624  *        hold any sleep locks.
625  */
626 void
lockdebug_barrier(const char * func,size_t line,volatile void * onelock,int slplocks)627 lockdebug_barrier(const char *func, size_t line, volatile void *onelock,
628     int slplocks)
629 {
630           struct lwp *l = curlwp;
631           lockdebug_t *ld;
632           int s;
633 
634           if (__predict_false(panicstr != NULL || ld_panic))
635                     return;
636 
637           s = splhigh();
638           if ((l->l_pflag & LP_INTR) == 0) {
639                     TAILQ_FOREACH(ld, &curcpu()->ci_data.cpu_ld_locks, ld_chain) {
640                               if (ld->ld_lock == onelock) {
641                                         continue;
642                               }
643                               __cpu_simple_lock(&ld->ld_spinlock);
644                               lockdebug_abort1(func, line, ld, s,
645                                   "spin lock held", true);
646                               return;
647                     }
648           }
649           if (slplocks) {
650                     splx(s);
651                     return;
652           }
653           ld = TAILQ_FIRST(&l->l_ld_locks);
654           if (__predict_false(ld != NULL && ld->ld_lock != onelock)) {
655                     __cpu_simple_lock(&ld->ld_spinlock);
656                     lockdebug_abort1(func, line, ld, s, "sleep lock held", true);
657                     return;
658           }
659           splx(s);
660           if (l->l_shlocks != 0) {
661                     TAILQ_FOREACH(ld, &ld_all, ld_achain) {
662                               if (ld->ld_lock == onelock) {
663                                         continue;
664                               }
665                               if (ld->ld_lwp == l)
666                                         lockdebug_dump(l, ld, printf);
667                     }
668                     panic("%s,%zu: holding %d shared locks", func, line,
669                         l->l_shlocks);
670           }
671 }
672 
673 /*
674  * lockdebug_mem_check:
675  *
676  *        Check for in-use locks within a memory region that is
677  *        being freed.
678  */
679 void
lockdebug_mem_check(const char * func,size_t line,void * base,size_t sz)680 lockdebug_mem_check(const char *func, size_t line, void *base, size_t sz)
681 {
682           lockdebug_t *ld;
683           struct cpu_info *ci;
684           int s;
685 
686           if (__predict_false(panicstr != NULL || ld_panic))
687                     return;
688 
689           kcov_silence_enter();
690 
691           s = splhigh();
692           ci = curcpu();
693           __cpu_simple_lock(&ci->ci_data.cpu_ld_lock);
694           ld = (lockdebug_t *)rb_tree_find_node_geq(&ld_rb_tree, base);
695           if (ld != NULL) {
696                     const uintptr_t lock = (uintptr_t)ld->ld_lock;
697 
698                     if (__predict_false((uintptr_t)base > lock))
699                               panic("%s,%zu: corrupt tree ld=%p, base=%p, sz=%zu",
700                                   func, line, ld, base, sz);
701                     if (lock >= (uintptr_t)base + sz)
702                               ld = NULL;
703           }
704           __cpu_simple_unlock(&ci->ci_data.cpu_ld_lock);
705           if (__predict_false(ld != NULL)) {
706                     __cpu_simple_lock(&ld->ld_spinlock);
707                     lockdebug_abort1(func, line, ld, s,
708                         "allocation contains active lock", !cold);
709                     kcov_silence_leave();
710                     return;
711           }
712           splx(s);
713 
714           kcov_silence_leave();
715 }
716 #endif /* _KERNEL */
717 
718 /*
719  * lockdebug_dump:
720  *
721  *        Dump information about a lock on panic, or for DDB.
722  */
723 static void
724 lockdebug_dump(lwp_t *l, lockdebug_t *ld, void (*pr)(const char *, ...)
725     __printflike(1, 2))
726 {
727           int sleeper = (ld->ld_flags & LD_SLEEPER);
728           lockops_t *lo = ld->ld_lockops;
729           char locksym[128], initsym[128], lockedsym[128], unlockedsym[128];
730 
731 #ifdef DDB
732           db_symstr(locksym, sizeof(locksym), (db_expr_t)(intptr_t)ld->ld_lock,
733               DB_STGY_ANY);
734           db_symstr(initsym, sizeof(initsym), (db_expr_t)ld->ld_initaddr,
735               DB_STGY_PROC);
736           db_symstr(lockedsym, sizeof(lockedsym), (db_expr_t)ld->ld_locked,
737               DB_STGY_PROC);
738           db_symstr(unlockedsym, sizeof(unlockedsym), (db_expr_t)ld->ld_unlocked,
739               DB_STGY_PROC);
740 #else
741           snprintf(locksym, sizeof(locksym), "%#018lx",
742               (unsigned long)ld->ld_lock);
743           snprintf(initsym, sizeof(initsym), "%#018lx",
744               (unsigned long)ld->ld_initaddr);
745           snprintf(lockedsym, sizeof(lockedsym), "%#018lx",
746               (unsigned long)ld->ld_locked);
747           snprintf(unlockedsym, sizeof(unlockedsym), "%#018lx",
748               (unsigned long)ld->ld_unlocked);
749 #endif
750 
751           (*pr)(
752               "lock address : %s\n"
753               "type         : %s\n"
754               "initialized  : %s",
755               locksym, (sleeper ? "sleep/adaptive" : "spin"),
756               initsym);
757 
758 #ifndef _KERNEL
759           lockops_t los;
760           lo = &los;
761           db_read_bytes((db_addr_t)ld->ld_lockops, sizeof(los), (char *)lo);
762 #endif
763           (*pr)("\n"
764               "shared holds : %18u exclusive: %18u\n"
765               "shares wanted: %18u exclusive: %18u\n"
766               "relevant cpu : %18u last held: %18u\n"
767               "relevant lwp : %#018lx last held: %#018lx\n"
768               "last locked%c : %s\n"
769               "unlocked%c    : %s\n",
770               (unsigned)ld->ld_shares, ((ld->ld_flags & LD_LOCKED) != 0),
771               (unsigned)ld->ld_shwant, (unsigned)ld->ld_exwant,
772               (unsigned)cpu_index(l->l_cpu), (unsigned)ld->ld_cpu,
773               (long)l, (long)ld->ld_lwp,
774               ((ld->ld_flags & LD_LOCKED) ? '*' : ' '),
775               lockedsym,
776               ((ld->ld_flags & LD_LOCKED) ? ' ' : '*'),
777               unlockedsym);
778 
779 #ifdef _KERNEL
780           if (lo->lo_dump != NULL)
781                     (*lo->lo_dump)(ld->ld_lock, pr);
782 
783           if (sleeper) {
784                     turnstile_print(ld->ld_lock, pr);
785           }
786 #endif
787 }
788 
789 #ifdef _KERNEL
790 /*
791  * lockdebug_abort1:
792  *
793  *        An error has been trapped - dump lock info and panic.
794  */
795 static void
lockdebug_abort1(const char * func,size_t line,lockdebug_t * ld,int s,const char * msg,bool dopanic)796 lockdebug_abort1(const char *func, size_t line, lockdebug_t *ld, int s,
797                      const char *msg, bool dopanic)
798 {
799 
800           /*
801            * Don't make the situation worse if the system is already going
802            * down in flames.  Once a panic is triggered, lockdebug state
803            * becomes stale and cannot be trusted.
804            */
805           if (atomic_inc_uint_nv(&ld_panic) != 1) {
806                     __cpu_simple_unlock(&ld->ld_spinlock);
807                     splx(s);
808                     return;
809           }
810 
811           printf("%s error: %s,%zu: %s\n\n", ld->ld_lockops->lo_name,
812               func, line, msg);
813           lockdebug_dump(curlwp, ld, printf);
814           __cpu_simple_unlock(&ld->ld_spinlock);
815           splx(s);
816           printf("\n");
817           if (dopanic)
818                     panic("LOCKDEBUG: %s error: %s,%zu: %s",
819                         ld->ld_lockops->lo_name, func, line, msg);
820 }
821 
822 #endif /* _KERNEL */
823 #endif    /* LOCKDEBUG */
824 
825 /*
826  * lockdebug_lock_print:
827  *
828  *        Handle the DDB 'show lock' command.
829  */
830 #ifdef DDB
831 void
832 lockdebug_lock_print(void *addr,
833     void (*pr)(const char *, ...) __printflike(1, 2))
834 {
835 #ifdef LOCKDEBUG
836           lockdebug_t *ld, lds;
837 
838           TAILQ_FOREACH(ld, &ld_all, ld_achain) {
839                     db_read_bytes((db_addr_t)ld, sizeof(lds), __UNVOLATILE(&lds));
840                     ld = &lds;
841                     if (ld->ld_lock == NULL)
842                               continue;
843                     if (addr == NULL || ld->ld_lock == addr) {
844                               lockdebug_dump(curlwp, ld, pr);
845                               if (addr != NULL)
846                                         return;
847                     }
848           }
849           if (addr != NULL) {
850                     (*pr)("Sorry, no record of a lock with address %p found.\n",
851                         addr);
852           }
853 #else
854           char sym[128];
855           uintptr_t word;
856 
857           (*pr)("WARNING: lock print is unreliable without LOCKDEBUG\n");
858           db_symstr(sym, sizeof(sym), (db_expr_t)(intptr_t)addr, DB_STGY_ANY);
859           db_read_bytes((db_addr_t)addr, sizeof(word), (char *)&word);
860           (*pr)("%s: possible owner: %p, bits: 0x%" PRIxPTR "\n", sym,
861               (void *)(word & ~(uintptr_t)ALIGNBYTES), word & ALIGNBYTES);
862 #endif    /* LOCKDEBUG */
863 }
864 
865 #ifdef _KERNEL
866 #ifdef LOCKDEBUG
867 static void
868 lockdebug_show_one(lwp_t *l, lockdebug_t *ld, int i,
869     void (*pr)(const char *, ...) __printflike(1, 2))
870 {
871           char sym[128];
872 
873 #ifdef DDB
874           db_symstr(sym, sizeof(sym), (db_expr_t)ld->ld_initaddr, DB_STGY_PROC);
875 #else
876           snprintf(sym, sizeof(sym), "%p", (void *)ld->ld_initaddr);
877 #endif
878           (*pr)("* Lock %d (initialized at %s)\n", i++, sym);
879           lockdebug_dump(l, ld, pr);
880 }
881 
882 static void
883 lockdebug_show_trace(const void *ptr,
884     void (*pr)(const char *, ...) __printflike(1, 2))
885 {
886 
887           db_stack_trace_print((db_expr_t)(intptr_t)ptr, true, 32, "a", pr);
888 }
889 
890 static void
891 lockdebug_show_all_locks_lwp(void (*pr)(const char *, ...) __printflike(1, 2),
892     bool show_trace)
893 {
894           struct proc *p;
895 
896           LIST_FOREACH(p, &allproc, p_list) {
897                     struct lwp *l;
898                     LIST_FOREACH(l, &p->p_lwps, l_sibling) {
899                               lockdebug_t *ld;
900                               int i = 0;
901                               if (TAILQ_EMPTY(&l->l_ld_locks) &&
902                                   l->l_ld_wanted == NULL) {
903                                         continue;
904                               }
905                               (*pr)("\n****** LWP %d.%d (%s) @ %p, l_stat=%d\n",
906                                   p->p_pid, l->l_lid,
907                                   l->l_name ? l->l_name : p->p_comm, l, l->l_stat);
908                               if (!TAILQ_EMPTY(&l->l_ld_locks)) {
909                                         (*pr)("\n*** Locks held: \n");
910                                         TAILQ_FOREACH(ld, &l->l_ld_locks, ld_chain) {
911                                                   (*pr)("\n");
912                                                   lockdebug_show_one(l, ld, i++, pr);
913                                         }
914                               } else {
915                                         (*pr)("\n*** Locks held: none\n");
916                               }
917 
918                               if (l->l_ld_wanted != NULL) {
919                                         (*pr)("\n*** Locks wanted: \n\n");
920                                         lockdebug_show_one(l, l->l_ld_wanted, 0, pr);
921                               } else {
922                                         (*pr)("\n*** Locks wanted: none\n");
923                               }
924                               if (show_trace) {
925                                         (*pr)("\n*** Traceback: \n\n");
926                                         lockdebug_show_trace(l, pr);
927                                         (*pr)("\n");
928                               }
929                     }
930           }
931 }
932 
933 static void
934 lockdebug_show_all_locks_cpu(void (*pr)(const char *, ...) __printflike(1, 2),
935     bool show_trace)
936 {
937           lockdebug_t *ld;
938           CPU_INFO_ITERATOR cii;
939           struct cpu_info *ci;
940 
941           for (CPU_INFO_FOREACH(cii, ci)) {
942                     int i = 0;
943                     if (TAILQ_EMPTY(&ci->ci_data.cpu_ld_locks))
944                               continue;
945                     (*pr)("\n******* Locks held on %s:\n", cpu_name(ci));
946                     TAILQ_FOREACH(ld, &ci->ci_data.cpu_ld_locks, ld_chain) {
947                               (*pr)("\n");
948 #ifdef MULTIPROCESSOR
949                               lockdebug_show_one(ci->ci_curlwp, ld, i++, pr);
950                               if (show_trace)
951                                         lockdebug_show_trace(ci->ci_curlwp, pr);
952 #else
953                               lockdebug_show_one(curlwp, ld, i++, pr);
954                               if (show_trace)
955                                         lockdebug_show_trace(curlwp, pr);
956 #endif
957                     }
958           }
959 }
960 #endif /* _KERNEL */
961 #endif    /* LOCKDEBUG */
962 
963 #ifdef _KERNEL
964 void
965 lockdebug_show_all_locks(void (*pr)(const char *, ...) __printflike(1, 2),
966     const char *modif)
967 {
968 #ifdef LOCKDEBUG
969           bool show_trace = false;
970           if (modif[0] == 't')
971                     show_trace = true;
972 
973           (*pr)("[Locks tracked through LWPs]\n");
974           lockdebug_show_all_locks_lwp(pr, show_trace);
975           (*pr)("\n");
976 
977           (*pr)("[Locks tracked through CPUs]\n");
978           lockdebug_show_all_locks_cpu(pr, show_trace);
979           (*pr)("\n");
980 #else
981           (*pr)("Sorry, kernel not built with the LOCKDEBUG option.\n");
982 #endif    /* LOCKDEBUG */
983 }
984 
985 void
986 lockdebug_show_lockstats(void (*pr)(const char *, ...) __printflike(1, 2))
987 {
988 #ifdef LOCKDEBUG
989           lockdebug_t *ld;
990           void *_ld;
991           uint32_t n_null = 0;
992           uint32_t n_spin_mutex = 0;
993           uint32_t n_adaptive_mutex = 0;
994           uint32_t n_rwlock = 0;
995           uint32_t n_others = 0;
996 
997           RB_TREE_FOREACH(_ld, &ld_rb_tree) {
998                     ld = _ld;
999                     if (ld->ld_lock == NULL) {
1000                               n_null++;
1001                               continue;
1002                     }
1003                     if (ld->ld_lockops->lo_name[0] == 'M') {
1004                               if (ld->ld_lockops->lo_type == LOCKOPS_SLEEP)
1005                                         n_adaptive_mutex++;
1006                               else
1007                                         n_spin_mutex++;
1008                               continue;
1009                     }
1010                     if (ld->ld_lockops->lo_name[0] == 'R') {
1011                               n_rwlock++;
1012                               continue;
1013                     }
1014                     n_others++;
1015           }
1016           (*pr)(
1017               "spin mutex: %u\n"
1018               "adaptive mutex: %u\n"
1019               "rwlock: %u\n"
1020               "null locks: %u\n"
1021               "others: %u\n",
1022               n_spin_mutex, n_adaptive_mutex, n_rwlock,
1023               n_null, n_others);
1024 #else
1025           (*pr)("Sorry, kernel not built with the LOCKDEBUG option.\n");
1026 #endif    /* LOCKDEBUG */
1027 }
1028 #endif /* _KERNEL */
1029 #endif    /* DDB */
1030 
1031 #ifdef _KERNEL
1032 /*
1033  * lockdebug_dismiss:
1034  *
1035  *      The system is rebooting, and potentially from an unsafe
1036  *      place so avoid any future aborts.
1037  */
1038 void
lockdebug_dismiss(void)1039 lockdebug_dismiss(void)
1040 {
1041 
1042           atomic_inc_uint_nv(&ld_panic);
1043 }
1044 
1045 /*
1046  * lockdebug_abort:
1047  *
1048  *        An error has been trapped - dump lock info and call panic().
1049  */
1050 void
lockdebug_abort(const char * func,size_t line,const volatile void * lock,lockops_t * ops,const char * msg)1051 lockdebug_abort(const char *func, size_t line, const volatile void *lock,
1052     lockops_t *ops, const char *msg)
1053 {
1054 #ifdef LOCKDEBUG
1055           lockdebug_t *ld;
1056           int s;
1057 
1058           s = splhigh();
1059           if ((ld = lockdebug_lookup(func, line, lock,
1060                               (uintptr_t) __builtin_return_address(0))) != NULL) {
1061                     lockdebug_abort1(func, line, ld, s, msg, true);
1062                     return;
1063           }
1064           splx(s);
1065 #endif    /* LOCKDEBUG */
1066 
1067           /*
1068            * Don't make the situation worse if the system is already going
1069            * down in flames.  Once a panic is triggered, lockdebug state
1070            * becomes stale and cannot be trusted.
1071            */
1072           if (atomic_inc_uint_nv(&ld_panic) > 1)
1073                     return;
1074 
1075           char locksym[128];
1076 
1077 #ifdef DDB
1078           db_symstr(locksym, sizeof(locksym), (db_expr_t)(intptr_t)lock,
1079               DB_STGY_ANY);
1080 #else
1081           snprintf(locksym, sizeof(locksym), "%#018lx", (unsigned long)lock);
1082 #endif
1083 
1084           printf("%s error: %s,%zu: %s\n\n"
1085               "lock address : %s\n"
1086               "current cpu  : %18d\n"
1087               "current lwp  : %#018lx\n",
1088               ops->lo_name, func, line, msg, locksym,
1089               (int)cpu_index(curcpu()), (long)curlwp);
1090           (*ops->lo_dump)(lock, printf);
1091           printf("\n");
1092 
1093           panic("lock error: %s: %s,%zu: %s: lock %p cpu %d lwp %p",
1094               ops->lo_name, func, line, msg, lock, cpu_index(curcpu()), curlwp);
1095 }
1096 #endif /* _KERNEL */
1097