86
87#include "opt_ddb.h"
88#include "opt_witness.h"
89
90#include <sys/param.h>
91#include <sys/bus.h>
92#include <sys/kdb.h>
93#include <sys/kernel.h>
94#include <sys/ktr.h>
95#include <sys/lock.h>
96#include <sys/malloc.h>
97#include <sys/mutex.h>
98#include <sys/proc.h>
99#include <sys/sysctl.h>
100#include <sys/systm.h>
101
102#include <ddb/ddb.h>
103
104#include <machine/stdarg.h>
105
106/* Define this to check for blessed mutexes */
107#undef BLESSING
108
109#define WITNESS_COUNT 1024
110#define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
111/*
112 * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
113 * will hold LOCK_NCHILDREN * 2 locks. We handle failure ok, and we should
114 * probably be safe for the most part, but it's still a SWAG.
115 */
116#define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
117
118#define WITNESS_NCHILDREN 6
119
120struct witness_child_list_entry;
121
122struct witness {
123 const char *w_name;
124 struct lock_class *w_class;
125 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */
126 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */
127 struct witness_child_list_entry *w_children; /* Great evilness... */
128 const char *w_file;
129 int w_line;
130 u_int w_level;
131 u_int w_refcount;
132 u_char w_Giant_squawked:1;
133 u_char w_other_squawked:1;
134 u_char w_same_squawked:1;
135 u_char w_displayed:1;
136};
137
138struct witness_child_list_entry {
139 struct witness_child_list_entry *wcl_next;
140 struct witness *wcl_children[WITNESS_NCHILDREN];
141 u_int wcl_count;
142};
143
144STAILQ_HEAD(witness_list, witness);
145
146#ifdef BLESSING
147struct witness_blessed {
148 const char *b_lock1;
149 const char *b_lock2;
150};
151#endif
152
153struct witness_order_list_entry {
154 const char *w_name;
155 struct lock_class *w_class;
156};
157
158#ifdef BLESSING
159static int blessed(struct witness *, struct witness *);
160#endif
161static int depart(struct witness *w);
162static struct witness *enroll(const char *description,
163 struct lock_class *lock_class);
164static int insertchild(struct witness *parent, struct witness *child);
165static int isitmychild(struct witness *parent, struct witness *child);
166static int isitmydescendant(struct witness *parent, struct witness *child);
167static int itismychild(struct witness *parent, struct witness *child);
168static int rebalancetree(struct witness_list *list);
169static void removechild(struct witness *parent, struct witness *child);
170static int reparentchildren(struct witness *newparent,
171 struct witness *oldparent);
172static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
173static void witness_displaydescendants(void(*)(const char *fmt, ...),
174 struct witness *, int indent);
175static const char *fixup_filename(const char *file);
176static void witness_leveldescendents(struct witness *parent, int level);
177static void witness_levelall(void);
178static struct witness *witness_get(void);
179static void witness_free(struct witness *m);
180static struct witness_child_list_entry *witness_child_get(void);
181static void witness_child_free(struct witness_child_list_entry *wcl);
182static struct lock_list_entry *witness_lock_list_get(void);
183static void witness_lock_list_free(struct lock_list_entry *lle);
184static struct lock_instance *find_instance(struct lock_list_entry *lock_list,
185 struct lock_object *lock);
186static void witness_list_lock(struct lock_instance *instance);
187#ifdef DDB
188static void witness_list(struct thread *td);
189static void witness_display_list(void(*prnt)(const char *fmt, ...),
190 struct witness_list *list);
191static void witness_display(void(*)(const char *fmt, ...));
192#endif
193
194SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, 0, "Witness Locking");
195
196/*
197 * If set to 0, witness is disabled. If set to 1, witness performs full lock
198 * order checking for all locks. If set to 2 or higher, then witness skips
199 * the full lock order check if the lock being acquired is at a higher level
200 * (i.e. farther down in the tree) than the current lock. This last mode is
201 * somewhat experimental and not considered fully safe. At runtime, this
202 * value may be set to 0 to turn off witness. witness is not allowed be
203 * turned on once it is turned off, however.
204 */
205static int witness_watch = 1;
206TUNABLE_INT("debug.witness.watch", &witness_watch);
207SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
208 sysctl_debug_witness_watch, "I", "witness is watching lock operations");
209
210#ifdef KDB
211/*
212 * When KDB is enabled and witness_kdb is set to 1, it will cause the system
213 * to drop into kdebug() when:
214 * - a lock heirarchy violation occurs
215 * - locks are held when going to sleep.
216 */
217#ifdef WITNESS_KDB
218int witness_kdb = 1;
219#else
220int witness_kdb = 0;
221#endif
222TUNABLE_INT("debug.witness.kdb", &witness_kdb);
223SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
224
225/*
226 * When KDB is enabled and witness_trace is set to 1, it will cause the system
227 * to print a stack trace:
228 * - a lock heirarchy violation occurs
229 * - locks are held when going to sleep.
230 */
231int witness_trace = 1;
232TUNABLE_INT("debug.witness.trace", &witness_trace);
233SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
234#endif /* KDB */
235
236#ifdef WITNESS_SKIPSPIN
237int witness_skipspin = 1;
238#else
239int witness_skipspin = 0;
240#endif
241TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
242SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN,
243 &witness_skipspin, 0, "");
244
245static struct mtx w_mtx;
246static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
247static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
248static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
249static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
250static struct witness_child_list_entry *w_child_free = NULL;
251static struct lock_list_entry *w_lock_list_free = NULL;
252
253static struct witness w_data[WITNESS_COUNT];
254static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
255static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
256
257static struct witness_order_list_entry order_lists[] = {
258 { "proctree", &lock_class_sx },
259 { "allproc", &lock_class_sx },
260 { "Giant", &lock_class_mtx_sleep },
261 { "filedesc structure", &lock_class_mtx_sleep },
262 { "pipe mutex", &lock_class_mtx_sleep },
263 { "sigio lock", &lock_class_mtx_sleep },
264 { "process group", &lock_class_mtx_sleep },
265 { "process lock", &lock_class_mtx_sleep },
266 { "session", &lock_class_mtx_sleep },
267 { "uidinfo hash", &lock_class_mtx_sleep },
268 { "uidinfo struct", &lock_class_mtx_sleep },
269 { "allprison", &lock_class_mtx_sleep },
270 { NULL, NULL },
271 /*
272 * Sockets
273 */
274 { "filedesc structure", &lock_class_mtx_sleep },
275 { "accept", &lock_class_mtx_sleep },
276 { "so_snd", &lock_class_mtx_sleep },
277 { "so_rcv", &lock_class_mtx_sleep },
278 { "sellck", &lock_class_mtx_sleep },
279 { NULL, NULL },
280 /*
281 * Routing
282 */
283 { "so_rcv", &lock_class_mtx_sleep },
284 { "radix node head", &lock_class_mtx_sleep },
285 { "rtentry", &lock_class_mtx_sleep },
286 { "ifaddr", &lock_class_mtx_sleep },
287 { NULL, NULL },
288 /*
289 * UNIX Domain Sockets
290 */
291 { "unp", &lock_class_mtx_sleep },
292 { "so_snd", &lock_class_mtx_sleep },
293 { NULL, NULL },
294 /*
295 * UDP/IP
296 */
297 { "udp", &lock_class_mtx_sleep },
298 { "udpinp", &lock_class_mtx_sleep },
299 { "so_snd", &lock_class_mtx_sleep },
300 { NULL, NULL },
301 /*
302 * TCP/IP
303 */
304 { "tcp", &lock_class_mtx_sleep },
305 { "tcpinp", &lock_class_mtx_sleep },
306 { "so_snd", &lock_class_mtx_sleep },
307 { NULL, NULL },
308 /*
309 * SLIP
310 */
311 { "slip_mtx", &lock_class_mtx_sleep },
312 { "slip sc_mtx", &lock_class_mtx_sleep },
313 { NULL, NULL },
314 /*
315 * netatalk
316 */
317 { "ddp_list_mtx", &lock_class_mtx_sleep },
318 { "ddp_mtx", &lock_class_mtx_sleep },
319 { NULL, NULL },
320 /*
321 * BPF
322 */
323 { "bpf global lock", &lock_class_mtx_sleep },
324 { "bpf interface lock", &lock_class_mtx_sleep },
325 { "bpf cdev lock", &lock_class_mtx_sleep },
326 { NULL, NULL },
327 /*
328 * NFS server
329 */
330 { "nfsd_mtx", &lock_class_mtx_sleep },
331 { "so_snd", &lock_class_mtx_sleep },
332 { NULL, NULL },
333 /*
334 * CDEV
335 */
336 { "system map", &lock_class_mtx_sleep },
337 { "vm page queue mutex", &lock_class_mtx_sleep },
338 { "vnode interlock", &lock_class_mtx_sleep },
339 { "cdev", &lock_class_mtx_sleep },
340 { NULL, NULL },
341 /*
342 * spin locks
343 */
344#ifdef SMP
345 { "ap boot", &lock_class_mtx_spin },
346#endif
347 { "sio", &lock_class_mtx_spin },
348#ifdef __i386__
349 { "cy", &lock_class_mtx_spin },
350#endif
351 { "uart_hwmtx", &lock_class_mtx_spin },
352 { "sabtty", &lock_class_mtx_spin },
353 { "zstty", &lock_class_mtx_spin },
354 { "ng_node", &lock_class_mtx_spin },
355 { "ng_worklist", &lock_class_mtx_spin },
356 { "taskqueue_fast", &lock_class_mtx_spin },
357 { "intr table", &lock_class_mtx_spin },
358 { "ithread table lock", &lock_class_mtx_spin },
359 { "sleepq chain", &lock_class_mtx_spin },
360 { "sched lock", &lock_class_mtx_spin },
361 { "turnstile chain", &lock_class_mtx_spin },
362 { "td_contested", &lock_class_mtx_spin },
363 { "callout", &lock_class_mtx_spin },
364 { "entropy harvest mutex", &lock_class_mtx_spin },
365 /*
366 * leaf locks
367 */
368 { "allpmaps", &lock_class_mtx_spin },
369 { "vm page queue free mutex", &lock_class_mtx_spin },
370 { "icu", &lock_class_mtx_spin },
371#ifdef SMP
372 { "smp rendezvous", &lock_class_mtx_spin },
373#if defined(__i386__) || defined(__amd64__)
374 { "tlb", &lock_class_mtx_spin },
375#endif
376#ifdef __sparc64__
377 { "ipi", &lock_class_mtx_spin },
| 86
87#include "opt_ddb.h"
88#include "opt_witness.h"
89
90#include <sys/param.h>
91#include <sys/bus.h>
92#include <sys/kdb.h>
93#include <sys/kernel.h>
94#include <sys/ktr.h>
95#include <sys/lock.h>
96#include <sys/malloc.h>
97#include <sys/mutex.h>
98#include <sys/proc.h>
99#include <sys/sysctl.h>
100#include <sys/systm.h>
101
102#include <ddb/ddb.h>
103
104#include <machine/stdarg.h>
105
106/* Define this to check for blessed mutexes */
107#undef BLESSING
108
109#define WITNESS_COUNT 1024
110#define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
111/*
112 * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
113 * will hold LOCK_NCHILDREN * 2 locks. We handle failure ok, and we should
114 * probably be safe for the most part, but it's still a SWAG.
115 */
116#define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
117
118#define WITNESS_NCHILDREN 6
119
120struct witness_child_list_entry;
121
122struct witness {
123 const char *w_name;
124 struct lock_class *w_class;
125 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */
126 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */
127 struct witness_child_list_entry *w_children; /* Great evilness... */
128 const char *w_file;
129 int w_line;
130 u_int w_level;
131 u_int w_refcount;
132 u_char w_Giant_squawked:1;
133 u_char w_other_squawked:1;
134 u_char w_same_squawked:1;
135 u_char w_displayed:1;
136};
137
138struct witness_child_list_entry {
139 struct witness_child_list_entry *wcl_next;
140 struct witness *wcl_children[WITNESS_NCHILDREN];
141 u_int wcl_count;
142};
143
144STAILQ_HEAD(witness_list, witness);
145
146#ifdef BLESSING
147struct witness_blessed {
148 const char *b_lock1;
149 const char *b_lock2;
150};
151#endif
152
153struct witness_order_list_entry {
154 const char *w_name;
155 struct lock_class *w_class;
156};
157
158#ifdef BLESSING
159static int blessed(struct witness *, struct witness *);
160#endif
161static int depart(struct witness *w);
162static struct witness *enroll(const char *description,
163 struct lock_class *lock_class);
164static int insertchild(struct witness *parent, struct witness *child);
165static int isitmychild(struct witness *parent, struct witness *child);
166static int isitmydescendant(struct witness *parent, struct witness *child);
167static int itismychild(struct witness *parent, struct witness *child);
168static int rebalancetree(struct witness_list *list);
169static void removechild(struct witness *parent, struct witness *child);
170static int reparentchildren(struct witness *newparent,
171 struct witness *oldparent);
172static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
173static void witness_displaydescendants(void(*)(const char *fmt, ...),
174 struct witness *, int indent);
175static const char *fixup_filename(const char *file);
176static void witness_leveldescendents(struct witness *parent, int level);
177static void witness_levelall(void);
178static struct witness *witness_get(void);
179static void witness_free(struct witness *m);
180static struct witness_child_list_entry *witness_child_get(void);
181static void witness_child_free(struct witness_child_list_entry *wcl);
182static struct lock_list_entry *witness_lock_list_get(void);
183static void witness_lock_list_free(struct lock_list_entry *lle);
184static struct lock_instance *find_instance(struct lock_list_entry *lock_list,
185 struct lock_object *lock);
186static void witness_list_lock(struct lock_instance *instance);
187#ifdef DDB
188static void witness_list(struct thread *td);
189static void witness_display_list(void(*prnt)(const char *fmt, ...),
190 struct witness_list *list);
191static void witness_display(void(*)(const char *fmt, ...));
192#endif
193
194SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, 0, "Witness Locking");
195
196/*
197 * If set to 0, witness is disabled. If set to 1, witness performs full lock
198 * order checking for all locks. If set to 2 or higher, then witness skips
199 * the full lock order check if the lock being acquired is at a higher level
200 * (i.e. farther down in the tree) than the current lock. This last mode is
201 * somewhat experimental and not considered fully safe. At runtime, this
202 * value may be set to 0 to turn off witness. witness is not allowed be
203 * turned on once it is turned off, however.
204 */
205static int witness_watch = 1;
206TUNABLE_INT("debug.witness.watch", &witness_watch);
207SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
208 sysctl_debug_witness_watch, "I", "witness is watching lock operations");
209
210#ifdef KDB
211/*
212 * When KDB is enabled and witness_kdb is set to 1, it will cause the system
213 * to drop into kdebug() when:
214 * - a lock heirarchy violation occurs
215 * - locks are held when going to sleep.
216 */
217#ifdef WITNESS_KDB
218int witness_kdb = 1;
219#else
220int witness_kdb = 0;
221#endif
222TUNABLE_INT("debug.witness.kdb", &witness_kdb);
223SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
224
225/*
226 * When KDB is enabled and witness_trace is set to 1, it will cause the system
227 * to print a stack trace:
228 * - a lock heirarchy violation occurs
229 * - locks are held when going to sleep.
230 */
231int witness_trace = 1;
232TUNABLE_INT("debug.witness.trace", &witness_trace);
233SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
234#endif /* KDB */
235
236#ifdef WITNESS_SKIPSPIN
237int witness_skipspin = 1;
238#else
239int witness_skipspin = 0;
240#endif
241TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
242SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN,
243 &witness_skipspin, 0, "");
244
245static struct mtx w_mtx;
246static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
247static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
248static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
249static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
250static struct witness_child_list_entry *w_child_free = NULL;
251static struct lock_list_entry *w_lock_list_free = NULL;
252
253static struct witness w_data[WITNESS_COUNT];
254static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
255static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
256
257static struct witness_order_list_entry order_lists[] = {
258 { "proctree", &lock_class_sx },
259 { "allproc", &lock_class_sx },
260 { "Giant", &lock_class_mtx_sleep },
261 { "filedesc structure", &lock_class_mtx_sleep },
262 { "pipe mutex", &lock_class_mtx_sleep },
263 { "sigio lock", &lock_class_mtx_sleep },
264 { "process group", &lock_class_mtx_sleep },
265 { "process lock", &lock_class_mtx_sleep },
266 { "session", &lock_class_mtx_sleep },
267 { "uidinfo hash", &lock_class_mtx_sleep },
268 { "uidinfo struct", &lock_class_mtx_sleep },
269 { "allprison", &lock_class_mtx_sleep },
270 { NULL, NULL },
271 /*
272 * Sockets
273 */
274 { "filedesc structure", &lock_class_mtx_sleep },
275 { "accept", &lock_class_mtx_sleep },
276 { "so_snd", &lock_class_mtx_sleep },
277 { "so_rcv", &lock_class_mtx_sleep },
278 { "sellck", &lock_class_mtx_sleep },
279 { NULL, NULL },
280 /*
281 * Routing
282 */
283 { "so_rcv", &lock_class_mtx_sleep },
284 { "radix node head", &lock_class_mtx_sleep },
285 { "rtentry", &lock_class_mtx_sleep },
286 { "ifaddr", &lock_class_mtx_sleep },
287 { NULL, NULL },
288 /*
289 * UNIX Domain Sockets
290 */
291 { "unp", &lock_class_mtx_sleep },
292 { "so_snd", &lock_class_mtx_sleep },
293 { NULL, NULL },
294 /*
295 * UDP/IP
296 */
297 { "udp", &lock_class_mtx_sleep },
298 { "udpinp", &lock_class_mtx_sleep },
299 { "so_snd", &lock_class_mtx_sleep },
300 { NULL, NULL },
301 /*
302 * TCP/IP
303 */
304 { "tcp", &lock_class_mtx_sleep },
305 { "tcpinp", &lock_class_mtx_sleep },
306 { "so_snd", &lock_class_mtx_sleep },
307 { NULL, NULL },
308 /*
309 * SLIP
310 */
311 { "slip_mtx", &lock_class_mtx_sleep },
312 { "slip sc_mtx", &lock_class_mtx_sleep },
313 { NULL, NULL },
314 /*
315 * netatalk
316 */
317 { "ddp_list_mtx", &lock_class_mtx_sleep },
318 { "ddp_mtx", &lock_class_mtx_sleep },
319 { NULL, NULL },
320 /*
321 * BPF
322 */
323 { "bpf global lock", &lock_class_mtx_sleep },
324 { "bpf interface lock", &lock_class_mtx_sleep },
325 { "bpf cdev lock", &lock_class_mtx_sleep },
326 { NULL, NULL },
327 /*
328 * NFS server
329 */
330 { "nfsd_mtx", &lock_class_mtx_sleep },
331 { "so_snd", &lock_class_mtx_sleep },
332 { NULL, NULL },
333 /*
334 * CDEV
335 */
336 { "system map", &lock_class_mtx_sleep },
337 { "vm page queue mutex", &lock_class_mtx_sleep },
338 { "vnode interlock", &lock_class_mtx_sleep },
339 { "cdev", &lock_class_mtx_sleep },
340 { NULL, NULL },
341 /*
342 * spin locks
343 */
344#ifdef SMP
345 { "ap boot", &lock_class_mtx_spin },
346#endif
347 { "sio", &lock_class_mtx_spin },
348#ifdef __i386__
349 { "cy", &lock_class_mtx_spin },
350#endif
351 { "uart_hwmtx", &lock_class_mtx_spin },
352 { "sabtty", &lock_class_mtx_spin },
353 { "zstty", &lock_class_mtx_spin },
354 { "ng_node", &lock_class_mtx_spin },
355 { "ng_worklist", &lock_class_mtx_spin },
356 { "taskqueue_fast", &lock_class_mtx_spin },
357 { "intr table", &lock_class_mtx_spin },
358 { "ithread table lock", &lock_class_mtx_spin },
359 { "sleepq chain", &lock_class_mtx_spin },
360 { "sched lock", &lock_class_mtx_spin },
361 { "turnstile chain", &lock_class_mtx_spin },
362 { "td_contested", &lock_class_mtx_spin },
363 { "callout", &lock_class_mtx_spin },
364 { "entropy harvest mutex", &lock_class_mtx_spin },
365 /*
366 * leaf locks
367 */
368 { "allpmaps", &lock_class_mtx_spin },
369 { "vm page queue free mutex", &lock_class_mtx_spin },
370 { "icu", &lock_class_mtx_spin },
371#ifdef SMP
372 { "smp rendezvous", &lock_class_mtx_spin },
373#if defined(__i386__) || defined(__amd64__)
374 { "tlb", &lock_class_mtx_spin },
375#endif
376#ifdef __sparc64__
377 { "ipi", &lock_class_mtx_spin },
|
378#endif
379#endif
380 { "clk", &lock_class_mtx_spin },
381 { "mutex profiling lock", &lock_class_mtx_spin },
382 { "kse zombie lock", &lock_class_mtx_spin },
383 { "ALD Queue", &lock_class_mtx_spin },
384#ifdef __ia64__
385 { "MCA spin lock", &lock_class_mtx_spin },
386#endif
387#if defined(__i386__) || defined(__amd64__)
388 { "pcicfg", &lock_class_mtx_spin },
389 { "NDIS thread lock", &lock_class_mtx_spin },
390#endif
391 { "tw_osl_io_lock", &lock_class_mtx_spin },
392 { "tw_osl_q_lock", &lock_class_mtx_spin },
393 { "tw_cl_io_lock", &lock_class_mtx_spin },
394 { "tw_cl_intr_lock", &lock_class_mtx_spin },
395 { "tw_cl_gen_lock", &lock_class_mtx_spin },
396 { NULL, NULL },
397 { NULL, NULL }
398};
399
400#ifdef BLESSING
401/*
402 * Pairs of locks which have been blessed
403 * Don't complain about order problems with blessed locks
404 */
405static struct witness_blessed blessed_list[] = {
406};
407static int blessed_count =
408 sizeof(blessed_list) / sizeof(struct witness_blessed);
409#endif
410
411/*
412 * List of all locks in the system.
413 */
414TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks);
415
416static struct mtx all_mtx = {
417 { &lock_class_mtx_sleep, /* mtx_object.lo_class */
418 "All locks list", /* mtx_object.lo_name */
419 "All locks list", /* mtx_object.lo_type */
420 LO_INITIALIZED, /* mtx_object.lo_flags */
421 { NULL, NULL }, /* mtx_object.lo_list */
422 NULL }, /* mtx_object.lo_witness */
423 MTX_UNOWNED, 0 /* mtx_lock, mtx_recurse */
424};
425
426/*
427 * This global is set to 0 once it becomes safe to use the witness code.
428 */
429static int witness_cold = 1;
430
431/*
432 * Global variables for book keeping.
433 */
434static int lock_cur_cnt;
435static int lock_max_cnt;
436
437/*
438 * The WITNESS-enabled diagnostic code.
439 */
440static void
441witness_initialize(void *dummy __unused)
442{
443 struct lock_object *lock;
444 struct witness_order_list_entry *order;
445 struct witness *w, *w1;
446 int i;
447
448 /*
449 * We have to release Giant before initializing its witness
450 * structure so that WITNESS doesn't get confused.
451 */
452 mtx_unlock(&Giant);
453 mtx_assert(&Giant, MA_NOTOWNED);
454
455 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
456 TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
457 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
458 MTX_NOWITNESS);
459 for (i = 0; i < WITNESS_COUNT; i++)
460 witness_free(&w_data[i]);
461 for (i = 0; i < WITNESS_CHILDCOUNT; i++)
462 witness_child_free(&w_childdata[i]);
463 for (i = 0; i < LOCK_CHILDCOUNT; i++)
464 witness_lock_list_free(&w_locklistdata[i]);
465
466 /* First add in all the specified order lists. */
467 for (order = order_lists; order->w_name != NULL; order++) {
468 w = enroll(order->w_name, order->w_class);
469 if (w == NULL)
470 continue;
471 w->w_file = "order list";
472 for (order++; order->w_name != NULL; order++) {
473 w1 = enroll(order->w_name, order->w_class);
474 if (w1 == NULL)
475 continue;
476 w1->w_file = "order list";
477 if (!itismychild(w, w1))
478 panic("Not enough memory for static orders!");
479 w = w1;
480 }
481 }
482
483 /* Iterate through all locks and add them to witness. */
484 mtx_lock(&all_mtx);
485 TAILQ_FOREACH(lock, &all_locks, lo_list) {
486 if (lock->lo_flags & LO_WITNESS)
487 lock->lo_witness = enroll(lock->lo_type,
488 lock->lo_class);
489 else
490 lock->lo_witness = NULL;
491 }
492 mtx_unlock(&all_mtx);
493
494 /* Mark the witness code as being ready for use. */
495 atomic_store_rel_int(&witness_cold, 0);
496
497 mtx_lock(&Giant);
498}
499SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
500
501static int
502sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
503{
504 int error, value;
505
506 value = witness_watch;
507 error = sysctl_handle_int(oidp, &value, 0, req);
508 if (error != 0 || req->newptr == NULL)
509 return (error);
510 error = suser(req->td);
511 if (error != 0)
512 return (error);
513 if (value == witness_watch)
514 return (0);
515 if (value != 0)
516 return (EINVAL);
517 witness_watch = 0;
518 return (0);
519}
520
521void
522witness_init(struct lock_object *lock)
523{
524 struct lock_class *class;
525
526 class = lock->lo_class;
527 if (lock->lo_flags & LO_INITIALIZED)
528 panic("%s: lock (%s) %s is already initialized", __func__,
529 class->lc_name, lock->lo_name);
530 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
531 (class->lc_flags & LC_RECURSABLE) == 0)
532 panic("%s: lock (%s) %s can not be recursable", __func__,
533 class->lc_name, lock->lo_name);
534 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
535 (class->lc_flags & LC_SLEEPABLE) == 0)
536 panic("%s: lock (%s) %s can not be sleepable", __func__,
537 class->lc_name, lock->lo_name);
538 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
539 (class->lc_flags & LC_UPGRADABLE) == 0)
540 panic("%s: lock (%s) %s can not be upgradable", __func__,
541 class->lc_name, lock->lo_name);
542
543 mtx_lock(&all_mtx);
544 TAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
545 lock->lo_flags |= LO_INITIALIZED;
546 lock_cur_cnt++;
547 if (lock_cur_cnt > lock_max_cnt)
548 lock_max_cnt = lock_cur_cnt;
549 mtx_unlock(&all_mtx);
550 if (!witness_cold && witness_watch != 0 && panicstr == NULL &&
551 (lock->lo_flags & LO_WITNESS) != 0)
552 lock->lo_witness = enroll(lock->lo_type, class);
553 else
554 lock->lo_witness = NULL;
555}
556
557void
558witness_destroy(struct lock_object *lock)
559{
560 struct witness *w;
561
562 if (witness_cold)
563 panic("lock (%s) %s destroyed while witness_cold",
564 lock->lo_class->lc_name, lock->lo_name);
565 if ((lock->lo_flags & LO_INITIALIZED) == 0)
566 panic("%s: lock (%s) %s is not initialized", __func__,
567 lock->lo_class->lc_name, lock->lo_name);
568
569 /* XXX: need to verify that no one holds the lock */
570 w = lock->lo_witness;
571 if (w != NULL) {
572 mtx_lock_spin(&w_mtx);
573 MPASS(w->w_refcount > 0);
574 w->w_refcount--;
575
576 /*
577 * Lock is already released if we have an allocation failure
578 * and depart() fails.
579 */
580 if (w->w_refcount != 0 || depart(w))
581 mtx_unlock_spin(&w_mtx);
582 }
583
584 mtx_lock(&all_mtx);
585 lock_cur_cnt--;
586 TAILQ_REMOVE(&all_locks, lock, lo_list);
587 lock->lo_flags &= ~LO_INITIALIZED;
588 mtx_unlock(&all_mtx);
589}
590
591#ifdef DDB
592static void
593witness_display_list(void(*prnt)(const char *fmt, ...),
594 struct witness_list *list)
595{
596 struct witness *w;
597
598 STAILQ_FOREACH(w, list, w_typelist) {
599 if (w->w_file == NULL || w->w_level > 0)
600 continue;
601 /*
602 * This lock has no anscestors, display its descendants.
603 */
604 witness_displaydescendants(prnt, w, 0);
605 }
606}
607
608static void
609witness_display(void(*prnt)(const char *fmt, ...))
610{
611 struct witness *w;
612
613 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
614 witness_levelall();
615
616 /* Clear all the displayed flags. */
617 STAILQ_FOREACH(w, &w_all, w_list) {
618 w->w_displayed = 0;
619 }
620
621 /*
622 * First, handle sleep locks which have been acquired at least
623 * once.
624 */
625 prnt("Sleep locks:\n");
626 witness_display_list(prnt, &w_sleep);
627
628 /*
629 * Now do spin locks which have been acquired at least once.
630 */
631 prnt("\nSpin locks:\n");
632 witness_display_list(prnt, &w_spin);
633
634 /*
635 * Finally, any locks which have not been acquired yet.
636 */
637 prnt("\nLocks which were never acquired:\n");
638 STAILQ_FOREACH(w, &w_all, w_list) {
639 if (w->w_file != NULL || w->w_refcount == 0)
640 continue;
641 prnt("%s\n", w->w_name);
642 }
643}
644#endif /* DDB */
645
646/* Trim useless garbage from filenames. */
647static const char *
648fixup_filename(const char *file)
649{
650
651 if (file == NULL)
652 return (NULL);
653 while (strncmp(file, "../", 3) == 0)
654 file += 3;
655 return (file);
656}
657
658int
659witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
660{
661
662 if (witness_watch == 0 || panicstr != NULL)
663 return (0);
664
665 /* Require locks that witness knows about. */
666 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
667 lock2->lo_witness == NULL)
668 return (EINVAL);
669
670 MPASS(!mtx_owned(&w_mtx));
671 mtx_lock_spin(&w_mtx);
672
673 /*
674 * If we already have either an explicit or implied lock order that
675 * is the other way around, then return an error.
676 */
677 if (isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
678 mtx_unlock_spin(&w_mtx);
679 return (EDOOFUS);
680 }
681
682 /* Try to add the new order. */
683 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
684 lock2->lo_type, lock1->lo_type);
685 if (!itismychild(lock1->lo_witness, lock2->lo_witness))
686 return (ENOMEM);
687 mtx_unlock_spin(&w_mtx);
688 return (0);
689}
690
691void
692witness_checkorder(struct lock_object *lock, int flags, const char *file,
693 int line)
694{
695 struct lock_list_entry **lock_list, *lle;
696 struct lock_instance *lock1, *lock2;
697 struct lock_class *class;
698 struct witness *w, *w1;
699 struct thread *td;
700 int i, j;
701
702 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
703 panicstr != NULL)
704 return;
705
706 /*
707 * Try locks do not block if they fail to acquire the lock, thus
708 * there is no danger of deadlocks or of switching while holding a
709 * spin lock if we acquire a lock via a try operation. This
710 * function shouldn't even be called for try locks, so panic if
711 * that happens.
712 */
713 if (flags & LOP_TRYLOCK)
714 panic("%s should not be called for try lock operations",
715 __func__);
716
717 w = lock->lo_witness;
718 class = lock->lo_class;
719 td = curthread;
720 file = fixup_filename(file);
721
722 if (class->lc_flags & LC_SLEEPLOCK) {
723 /*
724 * Since spin locks include a critical section, this check
725 * implicitly enforces a lock order of all sleep locks before
726 * all spin locks.
727 */
728 if (td->td_critnest != 0 && !kdb_active)
729 panic("blockable sleep lock (%s) %s @ %s:%d",
730 class->lc_name, lock->lo_name, file, line);
731
732 /*
733 * If this is the first lock acquired then just return as
734 * no order checking is needed.
735 */
736 if (td->td_sleeplocks == NULL)
737 return;
738 lock_list = &td->td_sleeplocks;
739 } else {
740 /*
741 * If this is the first lock, just return as no order
742 * checking is needed. We check this in both if clauses
743 * here as unifying the check would require us to use a
744 * critical section to ensure we don't migrate while doing
745 * the check. Note that if this is not the first lock, we
746 * are already in a critical section and are safe for the
747 * rest of the check.
748 */
749 if (PCPU_GET(spinlocks) == NULL)
750 return;
751 lock_list = PCPU_PTR(spinlocks);
752 }
753
754 /*
755 * Check to see if we are recursing on a lock we already own. If
756 * so, make sure that we don't mismatch exclusive and shared lock
757 * acquires.
758 */
759 lock1 = find_instance(*lock_list, lock);
760 if (lock1 != NULL) {
761 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
762 (flags & LOP_EXCLUSIVE) == 0) {
763 printf("shared lock of (%s) %s @ %s:%d\n",
764 class->lc_name, lock->lo_name, file, line);
765 printf("while exclusively locked from %s:%d\n",
766 lock1->li_file, lock1->li_line);
767 panic("share->excl");
768 }
769 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
770 (flags & LOP_EXCLUSIVE) != 0) {
771 printf("exclusive lock of (%s) %s @ %s:%d\n",
772 class->lc_name, lock->lo_name, file, line);
773 printf("while share locked from %s:%d\n",
774 lock1->li_file, lock1->li_line);
775 panic("excl->share");
776 }
777 return;
778 }
779
780 /*
781 * Try locks do not block if they fail to acquire the lock, thus
782 * there is no danger of deadlocks or of switching while holding a
783 * spin lock if we acquire a lock via a try operation.
784 */
785 if (flags & LOP_TRYLOCK)
786 return;
787
788 /*
789 * Check for duplicate locks of the same type. Note that we only
790 * have to check for this on the last lock we just acquired. Any
791 * other cases will be caught as lock order violations.
792 */
793 lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
794 w1 = lock1->li_lock->lo_witness;
795 if (w1 == w) {
796 if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK) ||
797 (flags & LOP_DUPOK))
798 return;
799 w->w_same_squawked = 1;
800 printf("acquiring duplicate lock of same type: \"%s\"\n",
801 lock->lo_type);
802 printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
803 lock1->li_file, lock1->li_line);
804 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
805#ifdef KDB
806 goto debugger;
807#else
808 return;
809#endif
810 }
811 MPASS(!mtx_owned(&w_mtx));
812 mtx_lock_spin(&w_mtx);
813 /*
814 * If we have a known higher number just say ok
815 */
816 if (witness_watch > 1 && w->w_level > w1->w_level) {
817 mtx_unlock_spin(&w_mtx);
818 return;
819 }
820 /*
821 * If we know that the the lock we are acquiring comes after
822 * the lock we most recently acquired in the lock order tree,
823 * then there is no need for any further checks.
824 */
825 if (isitmydescendant(w1, w)) {
826 mtx_unlock_spin(&w_mtx);
827 return;
828 }
829 for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
830 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
831
832 MPASS(j < WITNESS_COUNT);
833 lock1 = &lle->ll_children[i];
834 w1 = lock1->li_lock->lo_witness;
835
836 /*
837 * If this lock doesn't undergo witness checking,
838 * then skip it.
839 */
840 if (w1 == NULL) {
841 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
842 ("lock missing witness structure"));
843 continue;
844 }
845 /*
846 * If we are locking Giant and this is a sleepable
847 * lock, then skip it.
848 */
849 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
850 lock == &Giant.mtx_object)
851 continue;
852 /*
853 * If we are locking a sleepable lock and this lock
854 * is Giant, then skip it.
855 */
856 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
857 lock1->li_lock == &Giant.mtx_object)
858 continue;
859 /*
860 * If we are locking a sleepable lock and this lock
861 * isn't sleepable, we want to treat it as a lock
862 * order violation to enfore a general lock order of
863 * sleepable locks before non-sleepable locks.
864 */
865 if (!((lock->lo_flags & LO_SLEEPABLE) != 0 &&
866 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
867 /*
868 * Check the lock order hierarchy for a reveresal.
869 */
870 if (!isitmydescendant(w, w1))
871 continue;
872 /*
873 * We have a lock order violation, check to see if it
874 * is allowed or has already been yelled about.
875 */
876 mtx_unlock_spin(&w_mtx);
877#ifdef BLESSING
878 /*
879 * If the lock order is blessed, just bail. We don't
880 * look for other lock order violations though, which
881 * may be a bug.
882 */
883 if (blessed(w, w1))
884 return;
885#endif
886 if (lock1->li_lock == &Giant.mtx_object) {
887 if (w1->w_Giant_squawked)
888 return;
889 else
890 w1->w_Giant_squawked = 1;
891 } else {
892 if (w1->w_other_squawked)
893 return;
894 else
895 w1->w_other_squawked = 1;
896 }
897 /*
898 * Ok, yell about it.
899 */
900 printf("lock order reversal\n");
901 /*
902 * Try to locate an earlier lock with
903 * witness w in our list.
904 */
905 do {
906 lock2 = &lle->ll_children[i];
907 MPASS(lock2->li_lock != NULL);
908 if (lock2->li_lock->lo_witness == w)
909 break;
910 if (i == 0 && lle->ll_next != NULL) {
911 lle = lle->ll_next;
912 i = lle->ll_count - 1;
913 MPASS(i >= 0 && i < LOCK_NCHILDREN);
914 } else
915 i--;
916 } while (i >= 0);
917 if (i < 0) {
918 printf(" 1st %p %s (%s) @ %s:%d\n",
919 lock1->li_lock, lock1->li_lock->lo_name,
920 lock1->li_lock->lo_type, lock1->li_file,
921 lock1->li_line);
922 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
923 lock->lo_name, lock->lo_type, file, line);
924 } else {
925 printf(" 1st %p %s (%s) @ %s:%d\n",
926 lock2->li_lock, lock2->li_lock->lo_name,
927 lock2->li_lock->lo_type, lock2->li_file,
928 lock2->li_line);
929 printf(" 2nd %p %s (%s) @ %s:%d\n",
930 lock1->li_lock, lock1->li_lock->lo_name,
931 lock1->li_lock->lo_type, lock1->li_file,
932 lock1->li_line);
933 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
934 lock->lo_name, lock->lo_type, file, line);
935 }
936#ifdef KDB
937 goto debugger;
938#else
939 return;
940#endif
941 }
942 }
943 lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
944 /*
945 * If requested, build a new lock order. However, don't build a new
946 * relationship between a sleepable lock and Giant if it is in the
947 * wrong direction. The correct lock order is that sleepable locks
948 * always come before Giant.
949 */
950 if (flags & LOP_NEWORDER &&
951 !(lock1->li_lock == &Giant.mtx_object &&
952 (lock->lo_flags & LO_SLEEPABLE) != 0)) {
953 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
954 lock->lo_type, lock1->li_lock->lo_type);
955 if (!itismychild(lock1->li_lock->lo_witness, w))
956 /* Witness is dead. */
957 return;
958 }
959 mtx_unlock_spin(&w_mtx);
960 return;
961
962#ifdef KDB
963debugger:
964 if (witness_trace)
965 kdb_backtrace();
966 if (witness_kdb)
967 kdb_enter(__func__);
968#endif
969}
970
971void
972witness_lock(struct lock_object *lock, int flags, const char *file, int line)
973{
974 struct lock_list_entry **lock_list, *lle;
975 struct lock_instance *instance;
976 struct witness *w;
977 struct thread *td;
978
979 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
980 panicstr != NULL)
981 return;
982 w = lock->lo_witness;
983 td = curthread;
984 file = fixup_filename(file);
985
986 /* Determine lock list for this lock. */
987 if (lock->lo_class->lc_flags & LC_SLEEPLOCK)
988 lock_list = &td->td_sleeplocks;
989 else
990 lock_list = PCPU_PTR(spinlocks);
991
992 /* Check to see if we are recursing on a lock we already own. */
993 instance = find_instance(*lock_list, lock);
994 if (instance != NULL) {
995 instance->li_flags++;
996 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
997 td->td_proc->p_pid, lock->lo_name,
998 instance->li_flags & LI_RECURSEMASK);
999 instance->li_file = file;
1000 instance->li_line = line;
1001 return;
1002 }
1003
1004 /* Update per-witness last file and line acquire. */
1005 w->w_file = file;
1006 w->w_line = line;
1007
1008 /* Find the next open lock instance in the list and fill it. */
1009 lle = *lock_list;
1010 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1011 lle = witness_lock_list_get();
1012 if (lle == NULL)
1013 return;
1014 lle->ll_next = *lock_list;
1015 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1016 td->td_proc->p_pid, lle);
1017 *lock_list = lle;
1018 }
1019 instance = &lle->ll_children[lle->ll_count++];
1020 instance->li_lock = lock;
1021 instance->li_line = line;
1022 instance->li_file = file;
1023 if ((flags & LOP_EXCLUSIVE) != 0)
1024 instance->li_flags = LI_EXCLUSIVE;
1025 else
1026 instance->li_flags = 0;
1027 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1028 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1029}
1030
1031void
1032witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1033{
1034 struct lock_instance *instance;
1035 struct lock_class *class;
1036
1037 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1038 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1039 return;
1040 class = lock->lo_class;
1041 file = fixup_filename(file);
1042 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1043 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1044 class->lc_name, lock->lo_name, file, line);
1045 if ((flags & LOP_TRYLOCK) == 0)
1046 panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
1047 lock->lo_name, file, line);
1048 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1049 panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1050 class->lc_name, lock->lo_name, file, line);
1051 instance = find_instance(curthread->td_sleeplocks, lock);
1052 if (instance == NULL)
1053 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1054 class->lc_name, lock->lo_name, file, line);
1055 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1056 panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1057 class->lc_name, lock->lo_name, file, line);
1058 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1059 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1060 class->lc_name, lock->lo_name,
1061 instance->li_flags & LI_RECURSEMASK, file, line);
1062 instance->li_flags |= LI_EXCLUSIVE;
1063}
1064
1065void
1066witness_downgrade(struct lock_object *lock, int flags, const char *file,
1067 int line)
1068{
1069 struct lock_instance *instance;
1070 struct lock_class *class;
1071
1072 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1073 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1074 return;
1075 class = lock->lo_class;
1076 file = fixup_filename(file);
1077 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1078 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1079 class->lc_name, lock->lo_name, file, line);
1080 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1081 panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1082 class->lc_name, lock->lo_name, file, line);
1083 instance = find_instance(curthread->td_sleeplocks, lock);
1084 if (instance == NULL)
1085 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1086 class->lc_name, lock->lo_name, file, line);
1087 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1088 panic("downgrade of shared lock (%s) %s @ %s:%d",
1089 class->lc_name, lock->lo_name, file, line);
1090 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1091 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1092 class->lc_name, lock->lo_name,
1093 instance->li_flags & LI_RECURSEMASK, file, line);
1094 instance->li_flags &= ~LI_EXCLUSIVE;
1095}
1096
1097void
1098witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1099{
1100 struct lock_list_entry **lock_list, *lle;
1101 struct lock_instance *instance;
1102 struct lock_class *class;
1103 struct thread *td;
1104 register_t s;
1105 int i, j;
1106
1107 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
1108 panicstr != NULL)
1109 return;
1110 td = curthread;
1111 class = lock->lo_class;
1112 file = fixup_filename(file);
1113
1114 /* Find lock instance associated with this lock. */
1115 if (class->lc_flags & LC_SLEEPLOCK)
1116 lock_list = &td->td_sleeplocks;
1117 else
1118 lock_list = PCPU_PTR(spinlocks);
1119 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1120 for (i = 0; i < (*lock_list)->ll_count; i++) {
1121 instance = &(*lock_list)->ll_children[i];
1122 if (instance->li_lock == lock)
1123 goto found;
1124 }
1125 panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
1126 file, line);
1127found:
1128
1129 /* First, check for shared/exclusive mismatches. */
1130 if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
1131 (flags & LOP_EXCLUSIVE) == 0) {
1132 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1133 lock->lo_name, file, line);
1134 printf("while exclusively locked from %s:%d\n",
1135 instance->li_file, instance->li_line);
1136 panic("excl->ushare");
1137 }
1138 if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
1139 (flags & LOP_EXCLUSIVE) != 0) {
1140 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1141 lock->lo_name, file, line);
1142 printf("while share locked from %s:%d\n", instance->li_file,
1143 instance->li_line);
1144 panic("share->uexcl");
1145 }
1146
1147 /* If we are recursed, unrecurse. */
1148 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1149 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1150 td->td_proc->p_pid, instance->li_lock->lo_name,
1151 instance->li_flags);
1152 instance->li_flags--;
1153 return;
1154 }
1155
1156 /* Otherwise, remove this item from the list. */
1157 s = intr_disable();
1158 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1159 td->td_proc->p_pid, instance->li_lock->lo_name,
1160 (*lock_list)->ll_count - 1);
1161 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1162 (*lock_list)->ll_children[j] =
1163 (*lock_list)->ll_children[j + 1];
1164 (*lock_list)->ll_count--;
1165 intr_restore(s);
1166
1167 /* If this lock list entry is now empty, free it. */
1168 if ((*lock_list)->ll_count == 0) {
1169 lle = *lock_list;
1170 *lock_list = lle->ll_next;
1171 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1172 td->td_proc->p_pid, lle);
1173 witness_lock_list_free(lle);
1174 }
1175}
1176
1177/*
1178 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1179 * exempt Giant and sleepable locks from the checks as well. If any
1180 * non-exempt locks are held, then a supplied message is printed to the
1181 * console along with a list of the offending locks. If indicated in the
1182 * flags then a failure results in a panic as well.
1183 */
1184int
1185witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1186{
1187 struct lock_list_entry *lle;
1188 struct lock_instance *lock1;
1189 struct thread *td;
1190 va_list ap;
1191 int i, n;
1192
1193 if (witness_cold || witness_watch == 0 || panicstr != NULL)
1194 return (0);
1195 n = 0;
1196 td = curthread;
1197 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1198 for (i = lle->ll_count - 1; i >= 0; i--) {
1199 lock1 = &lle->ll_children[i];
1200 if (lock1->li_lock == lock)
1201 continue;
1202 if (flags & WARN_GIANTOK &&
1203 lock1->li_lock == &Giant.mtx_object)
1204 continue;
1205 if (flags & WARN_SLEEPOK &&
1206 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1207 continue;
1208 if (n == 0) {
1209 va_start(ap, fmt);
1210 vprintf(fmt, ap);
1211 va_end(ap);
1212 printf(" with the following");
1213 if (flags & WARN_SLEEPOK)
1214 printf(" non-sleepable");
1215 printf(" locks held:\n");
1216 }
1217 n++;
1218 witness_list_lock(lock1);
1219 }
1220 if (PCPU_GET(spinlocks) != NULL) {
1221 /*
1222 * Since we already hold a spinlock preemption is
1223 * already blocked.
1224 */
1225 if (n == 0) {
1226 va_start(ap, fmt);
1227 vprintf(fmt, ap);
1228 va_end(ap);
1229 printf(" with the following");
1230 if (flags & WARN_SLEEPOK)
1231 printf(" non-sleepable");
1232 printf(" locks held:\n");
1233 }
1234 n += witness_list_locks(PCPU_PTR(spinlocks));
1235 }
1236 if (flags & WARN_PANIC && n)
1237 panic("witness_warn");
1238#ifdef KDB
1239 else if (witness_kdb && n)
1240 kdb_enter(__func__);
1241 else if (witness_trace && n)
1242 kdb_backtrace();
1243#endif
1244 return (n);
1245}
1246
1247const char *
1248witness_file(struct lock_object *lock)
1249{
1250 struct witness *w;
1251
1252 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1253 return ("?");
1254 w = lock->lo_witness;
1255 return (w->w_file);
1256}
1257
1258int
1259witness_line(struct lock_object *lock)
1260{
1261 struct witness *w;
1262
1263 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1264 return (0);
1265 w = lock->lo_witness;
1266 return (w->w_line);
1267}
1268
1269static struct witness *
1270enroll(const char *description, struct lock_class *lock_class)
1271{
1272 struct witness *w;
1273
1274 if (witness_watch == 0 || panicstr != NULL)
1275 return (NULL);
1276 if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
1277 return (NULL);
1278 mtx_lock_spin(&w_mtx);
1279 STAILQ_FOREACH(w, &w_all, w_list) {
1280 if (w->w_name == description || (w->w_refcount > 0 &&
1281 strcmp(description, w->w_name) == 0)) {
1282 w->w_refcount++;
1283 mtx_unlock_spin(&w_mtx);
1284 if (lock_class != w->w_class)
1285 panic(
1286 "lock (%s) %s does not match earlier (%s) lock",
1287 description, lock_class->lc_name,
1288 w->w_class->lc_name);
1289 return (w);
1290 }
1291 }
1292 /*
1293 * This isn't quite right, as witness_cold is still 0 while we
1294 * enroll all the locks initialized before witness_initialize().
1295 */
1296 if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
1297 mtx_unlock_spin(&w_mtx);
1298 panic("spin lock %s not in order list", description);
1299 }
1300 if ((w = witness_get()) == NULL)
1301 return (NULL);
1302 w->w_name = description;
1303 w->w_class = lock_class;
1304 w->w_refcount = 1;
1305 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1306 if (lock_class->lc_flags & LC_SPINLOCK)
1307 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1308 else if (lock_class->lc_flags & LC_SLEEPLOCK)
1309 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1310 else {
1311 mtx_unlock_spin(&w_mtx);
1312 panic("lock class %s is not sleep or spin",
1313 lock_class->lc_name);
1314 }
1315 mtx_unlock_spin(&w_mtx);
1316 return (w);
1317}
1318
1319/* Don't let the door bang you on the way out... */
1320static int
1321depart(struct witness *w)
1322{
1323 struct witness_child_list_entry *wcl, *nwcl;
1324 struct witness_list *list;
1325 struct witness *parent;
1326
1327 MPASS(w->w_refcount == 0);
1328 if (w->w_class->lc_flags & LC_SLEEPLOCK)
1329 list = &w_sleep;
1330 else
1331 list = &w_spin;
1332 /*
1333 * First, we run through the entire tree looking for any
1334 * witnesses that the outgoing witness is a child of. For
1335 * each parent that we find, we reparent all the direct
1336 * children of the outgoing witness to its parent.
1337 */
1338 STAILQ_FOREACH(parent, list, w_typelist) {
1339 if (!isitmychild(parent, w))
1340 continue;
1341 removechild(parent, w);
1342 if (!reparentchildren(parent, w))
1343 return (0);
1344 }
1345
1346 /*
1347 * Now we go through and free up the child list of the
1348 * outgoing witness.
1349 */
1350 for (wcl = w->w_children; wcl != NULL; wcl = nwcl) {
1351 nwcl = wcl->wcl_next;
1352 witness_child_free(wcl);
1353 }
1354
1355 /*
1356 * Detach from various lists and free.
1357 */
1358 STAILQ_REMOVE(list, w, witness, w_typelist);
1359 STAILQ_REMOVE(&w_all, w, witness, w_list);
1360 witness_free(w);
1361
1362 /* Finally, fixup the tree. */
1363 return (rebalancetree(list));
1364}
1365
1366/*
1367 * Prune an entire lock order tree. We look for cases where a lock
1368 * is now both a descendant and a direct child of a given lock. In
1369 * that case, we want to remove the direct child link from the tree.
1370 *
1371 * Returns false if insertchild() fails.
1372 */
1373static int
1374rebalancetree(struct witness_list *list)
1375{
1376 struct witness *child, *parent;
1377
1378 STAILQ_FOREACH(child, list, w_typelist) {
1379 STAILQ_FOREACH(parent, list, w_typelist) {
1380 if (!isitmychild(parent, child))
1381 continue;
1382 removechild(parent, child);
1383 if (isitmydescendant(parent, child))
1384 continue;
1385 if (!insertchild(parent, child))
1386 return (0);
1387 }
1388 }
1389 witness_levelall();
1390 return (1);
1391}
1392
1393/*
1394 * Add "child" as a direct child of "parent". Returns false if
1395 * we fail due to out of memory.
1396 */
1397static int
1398insertchild(struct witness *parent, struct witness *child)
1399{
1400 struct witness_child_list_entry **wcl;
1401
1402 MPASS(child != NULL && parent != NULL);
1403
1404 /*
1405 * Insert "child" after "parent"
1406 */
1407 wcl = &parent->w_children;
1408 while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1409 wcl = &(*wcl)->wcl_next;
1410 if (*wcl == NULL) {
1411 *wcl = witness_child_get();
1412 if (*wcl == NULL)
1413 return (0);
1414 }
1415 (*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1416
1417 return (1);
1418}
1419
1420/*
1421 * Make all the direct descendants of oldparent be direct descendants
1422 * of newparent.
1423 */
1424static int
1425reparentchildren(struct witness *newparent, struct witness *oldparent)
1426{
1427 struct witness_child_list_entry *wcl;
1428 int i;
1429
1430 /* Avoid making a witness a child of itself. */
1431 MPASS(!isitmychild(oldparent, newparent));
1432
1433 for (wcl = oldparent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1434 for (i = 0; i < wcl->wcl_count; i++)
1435 if (!insertchild(newparent, wcl->wcl_children[i]))
1436 return (0);
1437 return (1);
1438}
1439
1440static int
1441itismychild(struct witness *parent, struct witness *child)
1442{
1443 struct witness_list *list;
1444
1445 MPASS(child != NULL && parent != NULL);
1446 if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
1447 (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
1448 panic(
1449 "%s: parent (%s) and child (%s) are not the same lock type",
1450 __func__, parent->w_class->lc_name,
1451 child->w_class->lc_name);
1452
1453 if (!insertchild(parent, child))
1454 return (0);
1455
1456 if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1457 list = &w_sleep;
1458 else
1459 list = &w_spin;
1460 return (rebalancetree(list));
1461}
1462
1463static void
1464removechild(struct witness *parent, struct witness *child)
1465{
1466 struct witness_child_list_entry **wcl, *wcl1;
1467 int i;
1468
1469 for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1470 for (i = 0; i < (*wcl)->wcl_count; i++)
1471 if ((*wcl)->wcl_children[i] == child)
1472 goto found;
1473 return;
1474found:
1475 (*wcl)->wcl_count--;
1476 if ((*wcl)->wcl_count > i)
1477 (*wcl)->wcl_children[i] =
1478 (*wcl)->wcl_children[(*wcl)->wcl_count];
1479 MPASS((*wcl)->wcl_children[i] != NULL);
1480 if ((*wcl)->wcl_count != 0)
1481 return;
1482 wcl1 = *wcl;
1483 *wcl = wcl1->wcl_next;
1484 witness_child_free(wcl1);
1485}
1486
1487static int
1488isitmychild(struct witness *parent, struct witness *child)
1489{
1490 struct witness_child_list_entry *wcl;
1491 int i;
1492
1493 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1494 for (i = 0; i < wcl->wcl_count; i++) {
1495 if (wcl->wcl_children[i] == child)
1496 return (1);
1497 }
1498 }
1499 return (0);
1500}
1501
1502static int
1503isitmydescendant(struct witness *parent, struct witness *child)
1504{
1505 struct witness_child_list_entry *wcl;
1506 int i, j;
1507
1508 if (isitmychild(parent, child))
1509 return (1);
1510 j = 0;
1511 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1512 MPASS(j < 1000);
1513 for (i = 0; i < wcl->wcl_count; i++) {
1514 if (isitmydescendant(wcl->wcl_children[i], child))
1515 return (1);
1516 }
1517 j++;
1518 }
1519 return (0);
1520}
1521
1522static void
1523witness_levelall (void)
1524{
1525 struct witness_list *list;
1526 struct witness *w, *w1;
1527
1528 /*
1529 * First clear all levels.
1530 */
1531 STAILQ_FOREACH(w, &w_all, w_list) {
1532 w->w_level = 0;
1533 }
1534
1535 /*
1536 * Look for locks with no parent and level all their descendants.
1537 */
1538 STAILQ_FOREACH(w, &w_all, w_list) {
1539 /*
1540 * This is just an optimization, technically we could get
1541 * away just walking the all list each time.
1542 */
1543 if (w->w_class->lc_flags & LC_SLEEPLOCK)
1544 list = &w_sleep;
1545 else
1546 list = &w_spin;
1547 STAILQ_FOREACH(w1, list, w_typelist) {
1548 if (isitmychild(w1, w))
1549 goto skip;
1550 }
1551 witness_leveldescendents(w, 0);
1552 skip:
1553 ; /* silence GCC 3.x */
1554 }
1555}
1556
1557static void
1558witness_leveldescendents(struct witness *parent, int level)
1559{
1560 struct witness_child_list_entry *wcl;
1561 int i;
1562
1563 if (parent->w_level < level)
1564 parent->w_level = level;
1565 level++;
1566 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1567 for (i = 0; i < wcl->wcl_count; i++)
1568 witness_leveldescendents(wcl->wcl_children[i], level);
1569}
1570
1571static void
1572witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1573 struct witness *parent, int indent)
1574{
1575 struct witness_child_list_entry *wcl;
1576 int i, level;
1577
1578 level = parent->w_level;
1579 prnt("%-2d", level);
1580 for (i = 0; i < indent; i++)
1581 prnt(" ");
1582 if (parent->w_refcount > 0)
1583 prnt("%s", parent->w_name);
1584 else
1585 prnt("(dead)");
1586 if (parent->w_displayed) {
1587 prnt(" -- (already displayed)\n");
1588 return;
1589 }
1590 parent->w_displayed = 1;
1591 if (parent->w_refcount > 0) {
1592 if (parent->w_file != NULL)
1593 prnt(" -- last acquired @ %s:%d", parent->w_file,
1594 parent->w_line);
1595 }
1596 prnt("\n");
1597 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1598 for (i = 0; i < wcl->wcl_count; i++)
1599 witness_displaydescendants(prnt,
1600 wcl->wcl_children[i], indent + 1);
1601}
1602
1603#ifdef BLESSING
1604static int
1605blessed(struct witness *w1, struct witness *w2)
1606{
1607 int i;
1608 struct witness_blessed *b;
1609
1610 for (i = 0; i < blessed_count; i++) {
1611 b = &blessed_list[i];
1612 if (strcmp(w1->w_name, b->b_lock1) == 0) {
1613 if (strcmp(w2->w_name, b->b_lock2) == 0)
1614 return (1);
1615 continue;
1616 }
1617 if (strcmp(w1->w_name, b->b_lock2) == 0)
1618 if (strcmp(w2->w_name, b->b_lock1) == 0)
1619 return (1);
1620 }
1621 return (0);
1622}
1623#endif
1624
1625static struct witness *
1626witness_get(void)
1627{
1628 struct witness *w;
1629
1630 if (witness_watch == 0) {
1631 mtx_unlock_spin(&w_mtx);
1632 return (NULL);
1633 }
1634 if (STAILQ_EMPTY(&w_free)) {
1635 witness_watch = 0;
1636 mtx_unlock_spin(&w_mtx);
1637 printf("%s: witness exhausted\n", __func__);
1638 return (NULL);
1639 }
1640 w = STAILQ_FIRST(&w_free);
1641 STAILQ_REMOVE_HEAD(&w_free, w_list);
1642 bzero(w, sizeof(*w));
1643 return (w);
1644}
1645
1646static void
1647witness_free(struct witness *w)
1648{
1649
1650 STAILQ_INSERT_HEAD(&w_free, w, w_list);
1651}
1652
1653static struct witness_child_list_entry *
1654witness_child_get(void)
1655{
1656 struct witness_child_list_entry *wcl;
1657
1658 if (witness_watch == 0) {
1659 mtx_unlock_spin(&w_mtx);
1660 return (NULL);
1661 }
1662 wcl = w_child_free;
1663 if (wcl == NULL) {
1664 witness_watch = 0;
1665 mtx_unlock_spin(&w_mtx);
1666 printf("%s: witness exhausted\n", __func__);
1667 return (NULL);
1668 }
1669 w_child_free = wcl->wcl_next;
1670 bzero(wcl, sizeof(*wcl));
1671 return (wcl);
1672}
1673
1674static void
1675witness_child_free(struct witness_child_list_entry *wcl)
1676{
1677
1678 wcl->wcl_next = w_child_free;
1679 w_child_free = wcl;
1680}
1681
1682static struct lock_list_entry *
1683witness_lock_list_get(void)
1684{
1685 struct lock_list_entry *lle;
1686
1687 if (witness_watch == 0)
1688 return (NULL);
1689 mtx_lock_spin(&w_mtx);
1690 lle = w_lock_list_free;
1691 if (lle == NULL) {
1692 witness_watch = 0;
1693 mtx_unlock_spin(&w_mtx);
1694 printf("%s: witness exhausted\n", __func__);
1695 return (NULL);
1696 }
1697 w_lock_list_free = lle->ll_next;
1698 mtx_unlock_spin(&w_mtx);
1699 bzero(lle, sizeof(*lle));
1700 return (lle);
1701}
1702
1703static void
1704witness_lock_list_free(struct lock_list_entry *lle)
1705{
1706
1707 mtx_lock_spin(&w_mtx);
1708 lle->ll_next = w_lock_list_free;
1709 w_lock_list_free = lle;
1710 mtx_unlock_spin(&w_mtx);
1711}
1712
1713static struct lock_instance *
1714find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1715{
1716 struct lock_list_entry *lle;
1717 struct lock_instance *instance;
1718 int i;
1719
1720 for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1721 for (i = lle->ll_count - 1; i >= 0; i--) {
1722 instance = &lle->ll_children[i];
1723 if (instance->li_lock == lock)
1724 return (instance);
1725 }
1726 return (NULL);
1727}
1728
1729static void
1730witness_list_lock(struct lock_instance *instance)
1731{
1732 struct lock_object *lock;
1733
1734 lock = instance->li_lock;
1735 printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1736 "exclusive" : "shared", lock->lo_class->lc_name, lock->lo_name);
1737 if (lock->lo_type != lock->lo_name)
1738 printf(" (%s)", lock->lo_type);
1739 printf(" r = %d (%p) locked @ %s:%d\n",
1740 instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
1741 instance->li_line);
1742}
1743
1744#ifdef DDB
1745static int
1746witness_thread_has_locks(struct thread *td)
1747{
1748
1749 return (td->td_sleeplocks != NULL);
1750}
1751
1752static int
1753witness_proc_has_locks(struct proc *p)
1754{
1755 struct thread *td;
1756
1757 FOREACH_THREAD_IN_PROC(p, td) {
1758 if (witness_thread_has_locks(td))
1759 return (1);
1760 }
1761 return (0);
1762}
1763#endif
1764
1765int
1766witness_list_locks(struct lock_list_entry **lock_list)
1767{
1768 struct lock_list_entry *lle;
1769 int i, nheld;
1770
1771 nheld = 0;
1772 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1773 for (i = lle->ll_count - 1; i >= 0; i--) {
1774 witness_list_lock(&lle->ll_children[i]);
1775 nheld++;
1776 }
1777 return (nheld);
1778}
1779
1780/*
1781 * This is a bit risky at best. We call this function when we have timed
1782 * out acquiring a spin lock, and we assume that the other CPU is stuck
1783 * with this lock held. So, we go groveling around in the other CPU's
1784 * per-cpu data to try to find the lock instance for this spin lock to
1785 * see when it was last acquired.
1786 */
1787void
1788witness_display_spinlock(struct lock_object *lock, struct thread *owner)
1789{
1790 struct lock_instance *instance;
1791 struct pcpu *pc;
1792
1793 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
1794 return;
1795 pc = pcpu_find(owner->td_oncpu);
1796 instance = find_instance(pc->pc_spinlocks, lock);
1797 if (instance != NULL)
1798 witness_list_lock(instance);
1799}
1800
1801void
1802witness_save(struct lock_object *lock, const char **filep, int *linep)
1803{
1804 struct lock_instance *instance;
1805
1806 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1807 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1808 return;
1809 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1810 panic("%s: lock (%s) %s is not a sleep lock", __func__,
1811 lock->lo_class->lc_name, lock->lo_name);
1812 instance = find_instance(curthread->td_sleeplocks, lock);
1813 if (instance == NULL)
1814 panic("%s: lock (%s) %s not locked", __func__,
1815 lock->lo_class->lc_name, lock->lo_name);
1816 *filep = instance->li_file;
1817 *linep = instance->li_line;
1818}
1819
1820void
1821witness_restore(struct lock_object *lock, const char *file, int line)
1822{
1823 struct lock_instance *instance;
1824
1825 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1826 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1827 return;
1828 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1829 panic("%s: lock (%s) %s is not a sleep lock", __func__,
1830 lock->lo_class->lc_name, lock->lo_name);
1831 instance = find_instance(curthread->td_sleeplocks, lock);
1832 if (instance == NULL)
1833 panic("%s: lock (%s) %s not locked", __func__,
1834 lock->lo_class->lc_name, lock->lo_name);
1835 lock->lo_witness->w_file = file;
1836 lock->lo_witness->w_line = line;
1837 instance->li_file = file;
1838 instance->li_line = line;
1839}
1840
1841void
1842witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1843{
1844#ifdef INVARIANT_SUPPORT
1845 struct lock_instance *instance;
1846
1847 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1848 return;
1849 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1850 instance = find_instance(curthread->td_sleeplocks, lock);
1851 else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1852 instance = find_instance(PCPU_GET(spinlocks), lock);
1853 else {
1854 panic("Lock (%s) %s is not sleep or spin!",
1855 lock->lo_class->lc_name, lock->lo_name);
1856 }
1857 file = fixup_filename(file);
1858 switch (flags) {
1859 case LA_UNLOCKED:
1860 if (instance != NULL)
1861 panic("Lock (%s) %s locked @ %s:%d.",
1862 lock->lo_class->lc_name, lock->lo_name, file, line);
1863 break;
1864 case LA_LOCKED:
1865 case LA_LOCKED | LA_RECURSED:
1866 case LA_LOCKED | LA_NOTRECURSED:
1867 case LA_SLOCKED:
1868 case LA_SLOCKED | LA_RECURSED:
1869 case LA_SLOCKED | LA_NOTRECURSED:
1870 case LA_XLOCKED:
1871 case LA_XLOCKED | LA_RECURSED:
1872 case LA_XLOCKED | LA_NOTRECURSED:
1873 if (instance == NULL) {
1874 panic("Lock (%s) %s not locked @ %s:%d.",
1875 lock->lo_class->lc_name, lock->lo_name, file, line);
1876 break;
1877 }
1878 if ((flags & LA_XLOCKED) != 0 &&
1879 (instance->li_flags & LI_EXCLUSIVE) == 0)
1880 panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1881 lock->lo_class->lc_name, lock->lo_name, file, line);
1882 if ((flags & LA_SLOCKED) != 0 &&
1883 (instance->li_flags & LI_EXCLUSIVE) != 0)
1884 panic("Lock (%s) %s exclusively locked @ %s:%d.",
1885 lock->lo_class->lc_name, lock->lo_name, file, line);
1886 if ((flags & LA_RECURSED) != 0 &&
1887 (instance->li_flags & LI_RECURSEMASK) == 0)
1888 panic("Lock (%s) %s not recursed @ %s:%d.",
1889 lock->lo_class->lc_name, lock->lo_name, file, line);
1890 if ((flags & LA_NOTRECURSED) != 0 &&
1891 (instance->li_flags & LI_RECURSEMASK) != 0)
1892 panic("Lock (%s) %s recursed @ %s:%d.",
1893 lock->lo_class->lc_name, lock->lo_name, file, line);
1894 break;
1895 default:
1896 panic("Invalid lock assertion at %s:%d.", file, line);
1897
1898 }
1899#endif /* INVARIANT_SUPPORT */
1900}
1901
1902#ifdef DDB
1903static void
1904witness_list(struct thread *td)
1905{
1906
1907 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1908 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
1909
1910 if (witness_watch == 0)
1911 return;
1912
1913 witness_list_locks(&td->td_sleeplocks);
1914
1915 /*
1916 * We only handle spinlocks if td == curthread. This is somewhat broken
1917 * if td is currently executing on some other CPU and holds spin locks
1918 * as we won't display those locks. If we had a MI way of getting
1919 * the per-cpu data for a given cpu then we could use
1920 * td->td_oncpu to get the list of spinlocks for this thread
1921 * and "fix" this.
1922 *
1923 * That still wouldn't really fix this unless we locked sched_lock
1924 * or stopped the other CPU to make sure it wasn't changing the list
1925 * out from under us. It is probably best to just not try to handle
1926 * threads on other CPU's for now.
1927 */
1928 if (td == curthread && PCPU_GET(spinlocks) != NULL)
1929 witness_list_locks(PCPU_PTR(spinlocks));
1930}
1931
1932DB_SHOW_COMMAND(locks, db_witness_list)
1933{
1934 struct thread *td;
1935 pid_t pid;
1936 struct proc *p;
1937
1938 if (have_addr) {
1939 pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1940 ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1941 ((addr >> 16) % 16) * 10000;
1942 /* sx_slock(&allproc_lock); */
1943 FOREACH_PROC_IN_SYSTEM(p) {
1944 if (p->p_pid == pid)
1945 break;
1946 }
1947 /* sx_sunlock(&allproc_lock); */
1948 if (p == NULL) {
1949 db_printf("pid %d not found\n", pid);
1950 return;
1951 }
1952 FOREACH_THREAD_IN_PROC(p, td) {
1953 witness_list(td);
1954 }
1955 } else {
1956 td = curthread;
1957 witness_list(td);
1958 }
1959}
1960
1961DB_SHOW_COMMAND(alllocks, db_witness_list_all)
1962{
1963 struct thread *td;
1964 struct proc *p;
1965
1966 /*
1967 * It would be nice to list only threads and processes that actually
1968 * held sleep locks, but that information is currently not exported
1969 * by WITNESS.
1970 */
1971 FOREACH_PROC_IN_SYSTEM(p) {
1972 if (!witness_proc_has_locks(p))
1973 continue;
1974 FOREACH_THREAD_IN_PROC(p, td) {
1975 if (!witness_thread_has_locks(td))
1976 continue;
1977 printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
1978 p->p_comm, td, td->td_tid);
1979 witness_list(td);
1980 }
1981 }
1982}
1983
1984DB_SHOW_COMMAND(witness, db_witness_display)
1985{
1986
1987 witness_display(db_printf);
1988}
1989#endif
| 379#endif
380#endif
381 { "clk", &lock_class_mtx_spin },
382 { "mutex profiling lock", &lock_class_mtx_spin },
383 { "kse zombie lock", &lock_class_mtx_spin },
384 { "ALD Queue", &lock_class_mtx_spin },
385#ifdef __ia64__
386 { "MCA spin lock", &lock_class_mtx_spin },
387#endif
388#if defined(__i386__) || defined(__amd64__)
389 { "pcicfg", &lock_class_mtx_spin },
390 { "NDIS thread lock", &lock_class_mtx_spin },
391#endif
392 { "tw_osl_io_lock", &lock_class_mtx_spin },
393 { "tw_osl_q_lock", &lock_class_mtx_spin },
394 { "tw_cl_io_lock", &lock_class_mtx_spin },
395 { "tw_cl_intr_lock", &lock_class_mtx_spin },
396 { "tw_cl_gen_lock", &lock_class_mtx_spin },
397 { NULL, NULL },
398 { NULL, NULL }
399};
400
401#ifdef BLESSING
402/*
403 * Pairs of locks which have been blessed
404 * Don't complain about order problems with blessed locks
405 */
406static struct witness_blessed blessed_list[] = {
407};
408static int blessed_count =
409 sizeof(blessed_list) / sizeof(struct witness_blessed);
410#endif
411
412/*
413 * List of all locks in the system.
414 */
415TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks);
416
417static struct mtx all_mtx = {
418 { &lock_class_mtx_sleep, /* mtx_object.lo_class */
419 "All locks list", /* mtx_object.lo_name */
420 "All locks list", /* mtx_object.lo_type */
421 LO_INITIALIZED, /* mtx_object.lo_flags */
422 { NULL, NULL }, /* mtx_object.lo_list */
423 NULL }, /* mtx_object.lo_witness */
424 MTX_UNOWNED, 0 /* mtx_lock, mtx_recurse */
425};
426
427/*
428 * This global is set to 0 once it becomes safe to use the witness code.
429 */
430static int witness_cold = 1;
431
432/*
433 * Global variables for book keeping.
434 */
435static int lock_cur_cnt;
436static int lock_max_cnt;
437
438/*
439 * The WITNESS-enabled diagnostic code.
440 */
441static void
442witness_initialize(void *dummy __unused)
443{
444 struct lock_object *lock;
445 struct witness_order_list_entry *order;
446 struct witness *w, *w1;
447 int i;
448
449 /*
450 * We have to release Giant before initializing its witness
451 * structure so that WITNESS doesn't get confused.
452 */
453 mtx_unlock(&Giant);
454 mtx_assert(&Giant, MA_NOTOWNED);
455
456 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
457 TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
458 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
459 MTX_NOWITNESS);
460 for (i = 0; i < WITNESS_COUNT; i++)
461 witness_free(&w_data[i]);
462 for (i = 0; i < WITNESS_CHILDCOUNT; i++)
463 witness_child_free(&w_childdata[i]);
464 for (i = 0; i < LOCK_CHILDCOUNT; i++)
465 witness_lock_list_free(&w_locklistdata[i]);
466
467 /* First add in all the specified order lists. */
468 for (order = order_lists; order->w_name != NULL; order++) {
469 w = enroll(order->w_name, order->w_class);
470 if (w == NULL)
471 continue;
472 w->w_file = "order list";
473 for (order++; order->w_name != NULL; order++) {
474 w1 = enroll(order->w_name, order->w_class);
475 if (w1 == NULL)
476 continue;
477 w1->w_file = "order list";
478 if (!itismychild(w, w1))
479 panic("Not enough memory for static orders!");
480 w = w1;
481 }
482 }
483
484 /* Iterate through all locks and add them to witness. */
485 mtx_lock(&all_mtx);
486 TAILQ_FOREACH(lock, &all_locks, lo_list) {
487 if (lock->lo_flags & LO_WITNESS)
488 lock->lo_witness = enroll(lock->lo_type,
489 lock->lo_class);
490 else
491 lock->lo_witness = NULL;
492 }
493 mtx_unlock(&all_mtx);
494
495 /* Mark the witness code as being ready for use. */
496 atomic_store_rel_int(&witness_cold, 0);
497
498 mtx_lock(&Giant);
499}
500SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
501
502static int
503sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
504{
505 int error, value;
506
507 value = witness_watch;
508 error = sysctl_handle_int(oidp, &value, 0, req);
509 if (error != 0 || req->newptr == NULL)
510 return (error);
511 error = suser(req->td);
512 if (error != 0)
513 return (error);
514 if (value == witness_watch)
515 return (0);
516 if (value != 0)
517 return (EINVAL);
518 witness_watch = 0;
519 return (0);
520}
521
522void
523witness_init(struct lock_object *lock)
524{
525 struct lock_class *class;
526
527 class = lock->lo_class;
528 if (lock->lo_flags & LO_INITIALIZED)
529 panic("%s: lock (%s) %s is already initialized", __func__,
530 class->lc_name, lock->lo_name);
531 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
532 (class->lc_flags & LC_RECURSABLE) == 0)
533 panic("%s: lock (%s) %s can not be recursable", __func__,
534 class->lc_name, lock->lo_name);
535 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
536 (class->lc_flags & LC_SLEEPABLE) == 0)
537 panic("%s: lock (%s) %s can not be sleepable", __func__,
538 class->lc_name, lock->lo_name);
539 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
540 (class->lc_flags & LC_UPGRADABLE) == 0)
541 panic("%s: lock (%s) %s can not be upgradable", __func__,
542 class->lc_name, lock->lo_name);
543
544 mtx_lock(&all_mtx);
545 TAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
546 lock->lo_flags |= LO_INITIALIZED;
547 lock_cur_cnt++;
548 if (lock_cur_cnt > lock_max_cnt)
549 lock_max_cnt = lock_cur_cnt;
550 mtx_unlock(&all_mtx);
551 if (!witness_cold && witness_watch != 0 && panicstr == NULL &&
552 (lock->lo_flags & LO_WITNESS) != 0)
553 lock->lo_witness = enroll(lock->lo_type, class);
554 else
555 lock->lo_witness = NULL;
556}
557
558void
559witness_destroy(struct lock_object *lock)
560{
561 struct witness *w;
562
563 if (witness_cold)
564 panic("lock (%s) %s destroyed while witness_cold",
565 lock->lo_class->lc_name, lock->lo_name);
566 if ((lock->lo_flags & LO_INITIALIZED) == 0)
567 panic("%s: lock (%s) %s is not initialized", __func__,
568 lock->lo_class->lc_name, lock->lo_name);
569
570 /* XXX: need to verify that no one holds the lock */
571 w = lock->lo_witness;
572 if (w != NULL) {
573 mtx_lock_spin(&w_mtx);
574 MPASS(w->w_refcount > 0);
575 w->w_refcount--;
576
577 /*
578 * Lock is already released if we have an allocation failure
579 * and depart() fails.
580 */
581 if (w->w_refcount != 0 || depart(w))
582 mtx_unlock_spin(&w_mtx);
583 }
584
585 mtx_lock(&all_mtx);
586 lock_cur_cnt--;
587 TAILQ_REMOVE(&all_locks, lock, lo_list);
588 lock->lo_flags &= ~LO_INITIALIZED;
589 mtx_unlock(&all_mtx);
590}
591
592#ifdef DDB
593static void
594witness_display_list(void(*prnt)(const char *fmt, ...),
595 struct witness_list *list)
596{
597 struct witness *w;
598
599 STAILQ_FOREACH(w, list, w_typelist) {
600 if (w->w_file == NULL || w->w_level > 0)
601 continue;
602 /*
603 * This lock has no anscestors, display its descendants.
604 */
605 witness_displaydescendants(prnt, w, 0);
606 }
607}
608
609static void
610witness_display(void(*prnt)(const char *fmt, ...))
611{
612 struct witness *w;
613
614 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
615 witness_levelall();
616
617 /* Clear all the displayed flags. */
618 STAILQ_FOREACH(w, &w_all, w_list) {
619 w->w_displayed = 0;
620 }
621
622 /*
623 * First, handle sleep locks which have been acquired at least
624 * once.
625 */
626 prnt("Sleep locks:\n");
627 witness_display_list(prnt, &w_sleep);
628
629 /*
630 * Now do spin locks which have been acquired at least once.
631 */
632 prnt("\nSpin locks:\n");
633 witness_display_list(prnt, &w_spin);
634
635 /*
636 * Finally, any locks which have not been acquired yet.
637 */
638 prnt("\nLocks which were never acquired:\n");
639 STAILQ_FOREACH(w, &w_all, w_list) {
640 if (w->w_file != NULL || w->w_refcount == 0)
641 continue;
642 prnt("%s\n", w->w_name);
643 }
644}
645#endif /* DDB */
646
647/* Trim useless garbage from filenames. */
648static const char *
649fixup_filename(const char *file)
650{
651
652 if (file == NULL)
653 return (NULL);
654 while (strncmp(file, "../", 3) == 0)
655 file += 3;
656 return (file);
657}
658
659int
660witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
661{
662
663 if (witness_watch == 0 || panicstr != NULL)
664 return (0);
665
666 /* Require locks that witness knows about. */
667 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
668 lock2->lo_witness == NULL)
669 return (EINVAL);
670
671 MPASS(!mtx_owned(&w_mtx));
672 mtx_lock_spin(&w_mtx);
673
674 /*
675 * If we already have either an explicit or implied lock order that
676 * is the other way around, then return an error.
677 */
678 if (isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
679 mtx_unlock_spin(&w_mtx);
680 return (EDOOFUS);
681 }
682
683 /* Try to add the new order. */
684 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
685 lock2->lo_type, lock1->lo_type);
686 if (!itismychild(lock1->lo_witness, lock2->lo_witness))
687 return (ENOMEM);
688 mtx_unlock_spin(&w_mtx);
689 return (0);
690}
691
692void
693witness_checkorder(struct lock_object *lock, int flags, const char *file,
694 int line)
695{
696 struct lock_list_entry **lock_list, *lle;
697 struct lock_instance *lock1, *lock2;
698 struct lock_class *class;
699 struct witness *w, *w1;
700 struct thread *td;
701 int i, j;
702
703 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
704 panicstr != NULL)
705 return;
706
707 /*
708 * Try locks do not block if they fail to acquire the lock, thus
709 * there is no danger of deadlocks or of switching while holding a
710 * spin lock if we acquire a lock via a try operation. This
711 * function shouldn't even be called for try locks, so panic if
712 * that happens.
713 */
714 if (flags & LOP_TRYLOCK)
715 panic("%s should not be called for try lock operations",
716 __func__);
717
718 w = lock->lo_witness;
719 class = lock->lo_class;
720 td = curthread;
721 file = fixup_filename(file);
722
723 if (class->lc_flags & LC_SLEEPLOCK) {
724 /*
725 * Since spin locks include a critical section, this check
726 * implicitly enforces a lock order of all sleep locks before
727 * all spin locks.
728 */
729 if (td->td_critnest != 0 && !kdb_active)
730 panic("blockable sleep lock (%s) %s @ %s:%d",
731 class->lc_name, lock->lo_name, file, line);
732
733 /*
734 * If this is the first lock acquired then just return as
735 * no order checking is needed.
736 */
737 if (td->td_sleeplocks == NULL)
738 return;
739 lock_list = &td->td_sleeplocks;
740 } else {
741 /*
742 * If this is the first lock, just return as no order
743 * checking is needed. We check this in both if clauses
744 * here as unifying the check would require us to use a
745 * critical section to ensure we don't migrate while doing
746 * the check. Note that if this is not the first lock, we
747 * are already in a critical section and are safe for the
748 * rest of the check.
749 */
750 if (PCPU_GET(spinlocks) == NULL)
751 return;
752 lock_list = PCPU_PTR(spinlocks);
753 }
754
755 /*
756 * Check to see if we are recursing on a lock we already own. If
757 * so, make sure that we don't mismatch exclusive and shared lock
758 * acquires.
759 */
760 lock1 = find_instance(*lock_list, lock);
761 if (lock1 != NULL) {
762 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
763 (flags & LOP_EXCLUSIVE) == 0) {
764 printf("shared lock of (%s) %s @ %s:%d\n",
765 class->lc_name, lock->lo_name, file, line);
766 printf("while exclusively locked from %s:%d\n",
767 lock1->li_file, lock1->li_line);
768 panic("share->excl");
769 }
770 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
771 (flags & LOP_EXCLUSIVE) != 0) {
772 printf("exclusive lock of (%s) %s @ %s:%d\n",
773 class->lc_name, lock->lo_name, file, line);
774 printf("while share locked from %s:%d\n",
775 lock1->li_file, lock1->li_line);
776 panic("excl->share");
777 }
778 return;
779 }
780
781 /*
782 * Try locks do not block if they fail to acquire the lock, thus
783 * there is no danger of deadlocks or of switching while holding a
784 * spin lock if we acquire a lock via a try operation.
785 */
786 if (flags & LOP_TRYLOCK)
787 return;
788
789 /*
790 * Check for duplicate locks of the same type. Note that we only
791 * have to check for this on the last lock we just acquired. Any
792 * other cases will be caught as lock order violations.
793 */
794 lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
795 w1 = lock1->li_lock->lo_witness;
796 if (w1 == w) {
797 if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK) ||
798 (flags & LOP_DUPOK))
799 return;
800 w->w_same_squawked = 1;
801 printf("acquiring duplicate lock of same type: \"%s\"\n",
802 lock->lo_type);
803 printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
804 lock1->li_file, lock1->li_line);
805 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
806#ifdef KDB
807 goto debugger;
808#else
809 return;
810#endif
811 }
812 MPASS(!mtx_owned(&w_mtx));
813 mtx_lock_spin(&w_mtx);
814 /*
815 * If we have a known higher number just say ok
816 */
817 if (witness_watch > 1 && w->w_level > w1->w_level) {
818 mtx_unlock_spin(&w_mtx);
819 return;
820 }
821 /*
822 * If we know that the the lock we are acquiring comes after
823 * the lock we most recently acquired in the lock order tree,
824 * then there is no need for any further checks.
825 */
826 if (isitmydescendant(w1, w)) {
827 mtx_unlock_spin(&w_mtx);
828 return;
829 }
830 for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
831 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
832
833 MPASS(j < WITNESS_COUNT);
834 lock1 = &lle->ll_children[i];
835 w1 = lock1->li_lock->lo_witness;
836
837 /*
838 * If this lock doesn't undergo witness checking,
839 * then skip it.
840 */
841 if (w1 == NULL) {
842 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
843 ("lock missing witness structure"));
844 continue;
845 }
846 /*
847 * If we are locking Giant and this is a sleepable
848 * lock, then skip it.
849 */
850 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
851 lock == &Giant.mtx_object)
852 continue;
853 /*
854 * If we are locking a sleepable lock and this lock
855 * is Giant, then skip it.
856 */
857 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
858 lock1->li_lock == &Giant.mtx_object)
859 continue;
860 /*
861 * If we are locking a sleepable lock and this lock
862 * isn't sleepable, we want to treat it as a lock
863 * order violation to enfore a general lock order of
864 * sleepable locks before non-sleepable locks.
865 */
866 if (!((lock->lo_flags & LO_SLEEPABLE) != 0 &&
867 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
868 /*
869 * Check the lock order hierarchy for a reveresal.
870 */
871 if (!isitmydescendant(w, w1))
872 continue;
873 /*
874 * We have a lock order violation, check to see if it
875 * is allowed or has already been yelled about.
876 */
877 mtx_unlock_spin(&w_mtx);
878#ifdef BLESSING
879 /*
880 * If the lock order is blessed, just bail. We don't
881 * look for other lock order violations though, which
882 * may be a bug.
883 */
884 if (blessed(w, w1))
885 return;
886#endif
887 if (lock1->li_lock == &Giant.mtx_object) {
888 if (w1->w_Giant_squawked)
889 return;
890 else
891 w1->w_Giant_squawked = 1;
892 } else {
893 if (w1->w_other_squawked)
894 return;
895 else
896 w1->w_other_squawked = 1;
897 }
898 /*
899 * Ok, yell about it.
900 */
901 printf("lock order reversal\n");
902 /*
903 * Try to locate an earlier lock with
904 * witness w in our list.
905 */
906 do {
907 lock2 = &lle->ll_children[i];
908 MPASS(lock2->li_lock != NULL);
909 if (lock2->li_lock->lo_witness == w)
910 break;
911 if (i == 0 && lle->ll_next != NULL) {
912 lle = lle->ll_next;
913 i = lle->ll_count - 1;
914 MPASS(i >= 0 && i < LOCK_NCHILDREN);
915 } else
916 i--;
917 } while (i >= 0);
918 if (i < 0) {
919 printf(" 1st %p %s (%s) @ %s:%d\n",
920 lock1->li_lock, lock1->li_lock->lo_name,
921 lock1->li_lock->lo_type, lock1->li_file,
922 lock1->li_line);
923 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
924 lock->lo_name, lock->lo_type, file, line);
925 } else {
926 printf(" 1st %p %s (%s) @ %s:%d\n",
927 lock2->li_lock, lock2->li_lock->lo_name,
928 lock2->li_lock->lo_type, lock2->li_file,
929 lock2->li_line);
930 printf(" 2nd %p %s (%s) @ %s:%d\n",
931 lock1->li_lock, lock1->li_lock->lo_name,
932 lock1->li_lock->lo_type, lock1->li_file,
933 lock1->li_line);
934 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
935 lock->lo_name, lock->lo_type, file, line);
936 }
937#ifdef KDB
938 goto debugger;
939#else
940 return;
941#endif
942 }
943 }
944 lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
945 /*
946 * If requested, build a new lock order. However, don't build a new
947 * relationship between a sleepable lock and Giant if it is in the
948 * wrong direction. The correct lock order is that sleepable locks
949 * always come before Giant.
950 */
951 if (flags & LOP_NEWORDER &&
952 !(lock1->li_lock == &Giant.mtx_object &&
953 (lock->lo_flags & LO_SLEEPABLE) != 0)) {
954 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
955 lock->lo_type, lock1->li_lock->lo_type);
956 if (!itismychild(lock1->li_lock->lo_witness, w))
957 /* Witness is dead. */
958 return;
959 }
960 mtx_unlock_spin(&w_mtx);
961 return;
962
963#ifdef KDB
964debugger:
965 if (witness_trace)
966 kdb_backtrace();
967 if (witness_kdb)
968 kdb_enter(__func__);
969#endif
970}
971
972void
973witness_lock(struct lock_object *lock, int flags, const char *file, int line)
974{
975 struct lock_list_entry **lock_list, *lle;
976 struct lock_instance *instance;
977 struct witness *w;
978 struct thread *td;
979
980 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
981 panicstr != NULL)
982 return;
983 w = lock->lo_witness;
984 td = curthread;
985 file = fixup_filename(file);
986
987 /* Determine lock list for this lock. */
988 if (lock->lo_class->lc_flags & LC_SLEEPLOCK)
989 lock_list = &td->td_sleeplocks;
990 else
991 lock_list = PCPU_PTR(spinlocks);
992
993 /* Check to see if we are recursing on a lock we already own. */
994 instance = find_instance(*lock_list, lock);
995 if (instance != NULL) {
996 instance->li_flags++;
997 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
998 td->td_proc->p_pid, lock->lo_name,
999 instance->li_flags & LI_RECURSEMASK);
1000 instance->li_file = file;
1001 instance->li_line = line;
1002 return;
1003 }
1004
1005 /* Update per-witness last file and line acquire. */
1006 w->w_file = file;
1007 w->w_line = line;
1008
1009 /* Find the next open lock instance in the list and fill it. */
1010 lle = *lock_list;
1011 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1012 lle = witness_lock_list_get();
1013 if (lle == NULL)
1014 return;
1015 lle->ll_next = *lock_list;
1016 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1017 td->td_proc->p_pid, lle);
1018 *lock_list = lle;
1019 }
1020 instance = &lle->ll_children[lle->ll_count++];
1021 instance->li_lock = lock;
1022 instance->li_line = line;
1023 instance->li_file = file;
1024 if ((flags & LOP_EXCLUSIVE) != 0)
1025 instance->li_flags = LI_EXCLUSIVE;
1026 else
1027 instance->li_flags = 0;
1028 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1029 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1030}
1031
1032void
1033witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1034{
1035 struct lock_instance *instance;
1036 struct lock_class *class;
1037
1038 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1039 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1040 return;
1041 class = lock->lo_class;
1042 file = fixup_filename(file);
1043 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1044 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1045 class->lc_name, lock->lo_name, file, line);
1046 if ((flags & LOP_TRYLOCK) == 0)
1047 panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
1048 lock->lo_name, file, line);
1049 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1050 panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1051 class->lc_name, lock->lo_name, file, line);
1052 instance = find_instance(curthread->td_sleeplocks, lock);
1053 if (instance == NULL)
1054 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1055 class->lc_name, lock->lo_name, file, line);
1056 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1057 panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1058 class->lc_name, lock->lo_name, file, line);
1059 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1060 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1061 class->lc_name, lock->lo_name,
1062 instance->li_flags & LI_RECURSEMASK, file, line);
1063 instance->li_flags |= LI_EXCLUSIVE;
1064}
1065
1066void
1067witness_downgrade(struct lock_object *lock, int flags, const char *file,
1068 int line)
1069{
1070 struct lock_instance *instance;
1071 struct lock_class *class;
1072
1073 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1074 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1075 return;
1076 class = lock->lo_class;
1077 file = fixup_filename(file);
1078 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1079 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1080 class->lc_name, lock->lo_name, file, line);
1081 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1082 panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1083 class->lc_name, lock->lo_name, file, line);
1084 instance = find_instance(curthread->td_sleeplocks, lock);
1085 if (instance == NULL)
1086 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1087 class->lc_name, lock->lo_name, file, line);
1088 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1089 panic("downgrade of shared lock (%s) %s @ %s:%d",
1090 class->lc_name, lock->lo_name, file, line);
1091 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1092 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1093 class->lc_name, lock->lo_name,
1094 instance->li_flags & LI_RECURSEMASK, file, line);
1095 instance->li_flags &= ~LI_EXCLUSIVE;
1096}
1097
1098void
1099witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1100{
1101 struct lock_list_entry **lock_list, *lle;
1102 struct lock_instance *instance;
1103 struct lock_class *class;
1104 struct thread *td;
1105 register_t s;
1106 int i, j;
1107
1108 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
1109 panicstr != NULL)
1110 return;
1111 td = curthread;
1112 class = lock->lo_class;
1113 file = fixup_filename(file);
1114
1115 /* Find lock instance associated with this lock. */
1116 if (class->lc_flags & LC_SLEEPLOCK)
1117 lock_list = &td->td_sleeplocks;
1118 else
1119 lock_list = PCPU_PTR(spinlocks);
1120 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1121 for (i = 0; i < (*lock_list)->ll_count; i++) {
1122 instance = &(*lock_list)->ll_children[i];
1123 if (instance->li_lock == lock)
1124 goto found;
1125 }
1126 panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
1127 file, line);
1128found:
1129
1130 /* First, check for shared/exclusive mismatches. */
1131 if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
1132 (flags & LOP_EXCLUSIVE) == 0) {
1133 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1134 lock->lo_name, file, line);
1135 printf("while exclusively locked from %s:%d\n",
1136 instance->li_file, instance->li_line);
1137 panic("excl->ushare");
1138 }
1139 if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
1140 (flags & LOP_EXCLUSIVE) != 0) {
1141 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1142 lock->lo_name, file, line);
1143 printf("while share locked from %s:%d\n", instance->li_file,
1144 instance->li_line);
1145 panic("share->uexcl");
1146 }
1147
1148 /* If we are recursed, unrecurse. */
1149 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1150 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1151 td->td_proc->p_pid, instance->li_lock->lo_name,
1152 instance->li_flags);
1153 instance->li_flags--;
1154 return;
1155 }
1156
1157 /* Otherwise, remove this item from the list. */
1158 s = intr_disable();
1159 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1160 td->td_proc->p_pid, instance->li_lock->lo_name,
1161 (*lock_list)->ll_count - 1);
1162 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1163 (*lock_list)->ll_children[j] =
1164 (*lock_list)->ll_children[j + 1];
1165 (*lock_list)->ll_count--;
1166 intr_restore(s);
1167
1168 /* If this lock list entry is now empty, free it. */
1169 if ((*lock_list)->ll_count == 0) {
1170 lle = *lock_list;
1171 *lock_list = lle->ll_next;
1172 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1173 td->td_proc->p_pid, lle);
1174 witness_lock_list_free(lle);
1175 }
1176}
1177
1178/*
1179 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1180 * exempt Giant and sleepable locks from the checks as well. If any
1181 * non-exempt locks are held, then a supplied message is printed to the
1182 * console along with a list of the offending locks. If indicated in the
1183 * flags then a failure results in a panic as well.
1184 */
1185int
1186witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1187{
1188 struct lock_list_entry *lle;
1189 struct lock_instance *lock1;
1190 struct thread *td;
1191 va_list ap;
1192 int i, n;
1193
1194 if (witness_cold || witness_watch == 0 || panicstr != NULL)
1195 return (0);
1196 n = 0;
1197 td = curthread;
1198 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1199 for (i = lle->ll_count - 1; i >= 0; i--) {
1200 lock1 = &lle->ll_children[i];
1201 if (lock1->li_lock == lock)
1202 continue;
1203 if (flags & WARN_GIANTOK &&
1204 lock1->li_lock == &Giant.mtx_object)
1205 continue;
1206 if (flags & WARN_SLEEPOK &&
1207 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1208 continue;
1209 if (n == 0) {
1210 va_start(ap, fmt);
1211 vprintf(fmt, ap);
1212 va_end(ap);
1213 printf(" with the following");
1214 if (flags & WARN_SLEEPOK)
1215 printf(" non-sleepable");
1216 printf(" locks held:\n");
1217 }
1218 n++;
1219 witness_list_lock(lock1);
1220 }
1221 if (PCPU_GET(spinlocks) != NULL) {
1222 /*
1223 * Since we already hold a spinlock preemption is
1224 * already blocked.
1225 */
1226 if (n == 0) {
1227 va_start(ap, fmt);
1228 vprintf(fmt, ap);
1229 va_end(ap);
1230 printf(" with the following");
1231 if (flags & WARN_SLEEPOK)
1232 printf(" non-sleepable");
1233 printf(" locks held:\n");
1234 }
1235 n += witness_list_locks(PCPU_PTR(spinlocks));
1236 }
1237 if (flags & WARN_PANIC && n)
1238 panic("witness_warn");
1239#ifdef KDB
1240 else if (witness_kdb && n)
1241 kdb_enter(__func__);
1242 else if (witness_trace && n)
1243 kdb_backtrace();
1244#endif
1245 return (n);
1246}
1247
1248const char *
1249witness_file(struct lock_object *lock)
1250{
1251 struct witness *w;
1252
1253 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1254 return ("?");
1255 w = lock->lo_witness;
1256 return (w->w_file);
1257}
1258
1259int
1260witness_line(struct lock_object *lock)
1261{
1262 struct witness *w;
1263
1264 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1265 return (0);
1266 w = lock->lo_witness;
1267 return (w->w_line);
1268}
1269
1270static struct witness *
1271enroll(const char *description, struct lock_class *lock_class)
1272{
1273 struct witness *w;
1274
1275 if (witness_watch == 0 || panicstr != NULL)
1276 return (NULL);
1277 if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
1278 return (NULL);
1279 mtx_lock_spin(&w_mtx);
1280 STAILQ_FOREACH(w, &w_all, w_list) {
1281 if (w->w_name == description || (w->w_refcount > 0 &&
1282 strcmp(description, w->w_name) == 0)) {
1283 w->w_refcount++;
1284 mtx_unlock_spin(&w_mtx);
1285 if (lock_class != w->w_class)
1286 panic(
1287 "lock (%s) %s does not match earlier (%s) lock",
1288 description, lock_class->lc_name,
1289 w->w_class->lc_name);
1290 return (w);
1291 }
1292 }
1293 /*
1294 * This isn't quite right, as witness_cold is still 0 while we
1295 * enroll all the locks initialized before witness_initialize().
1296 */
1297 if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
1298 mtx_unlock_spin(&w_mtx);
1299 panic("spin lock %s not in order list", description);
1300 }
1301 if ((w = witness_get()) == NULL)
1302 return (NULL);
1303 w->w_name = description;
1304 w->w_class = lock_class;
1305 w->w_refcount = 1;
1306 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1307 if (lock_class->lc_flags & LC_SPINLOCK)
1308 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1309 else if (lock_class->lc_flags & LC_SLEEPLOCK)
1310 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1311 else {
1312 mtx_unlock_spin(&w_mtx);
1313 panic("lock class %s is not sleep or spin",
1314 lock_class->lc_name);
1315 }
1316 mtx_unlock_spin(&w_mtx);
1317 return (w);
1318}
1319
1320/* Don't let the door bang you on the way out... */
1321static int
1322depart(struct witness *w)
1323{
1324 struct witness_child_list_entry *wcl, *nwcl;
1325 struct witness_list *list;
1326 struct witness *parent;
1327
1328 MPASS(w->w_refcount == 0);
1329 if (w->w_class->lc_flags & LC_SLEEPLOCK)
1330 list = &w_sleep;
1331 else
1332 list = &w_spin;
1333 /*
1334 * First, we run through the entire tree looking for any
1335 * witnesses that the outgoing witness is a child of. For
1336 * each parent that we find, we reparent all the direct
1337 * children of the outgoing witness to its parent.
1338 */
1339 STAILQ_FOREACH(parent, list, w_typelist) {
1340 if (!isitmychild(parent, w))
1341 continue;
1342 removechild(parent, w);
1343 if (!reparentchildren(parent, w))
1344 return (0);
1345 }
1346
1347 /*
1348 * Now we go through and free up the child list of the
1349 * outgoing witness.
1350 */
1351 for (wcl = w->w_children; wcl != NULL; wcl = nwcl) {
1352 nwcl = wcl->wcl_next;
1353 witness_child_free(wcl);
1354 }
1355
1356 /*
1357 * Detach from various lists and free.
1358 */
1359 STAILQ_REMOVE(list, w, witness, w_typelist);
1360 STAILQ_REMOVE(&w_all, w, witness, w_list);
1361 witness_free(w);
1362
1363 /* Finally, fixup the tree. */
1364 return (rebalancetree(list));
1365}
1366
1367/*
1368 * Prune an entire lock order tree. We look for cases where a lock
1369 * is now both a descendant and a direct child of a given lock. In
1370 * that case, we want to remove the direct child link from the tree.
1371 *
1372 * Returns false if insertchild() fails.
1373 */
1374static int
1375rebalancetree(struct witness_list *list)
1376{
1377 struct witness *child, *parent;
1378
1379 STAILQ_FOREACH(child, list, w_typelist) {
1380 STAILQ_FOREACH(parent, list, w_typelist) {
1381 if (!isitmychild(parent, child))
1382 continue;
1383 removechild(parent, child);
1384 if (isitmydescendant(parent, child))
1385 continue;
1386 if (!insertchild(parent, child))
1387 return (0);
1388 }
1389 }
1390 witness_levelall();
1391 return (1);
1392}
1393
1394/*
1395 * Add "child" as a direct child of "parent". Returns false if
1396 * we fail due to out of memory.
1397 */
1398static int
1399insertchild(struct witness *parent, struct witness *child)
1400{
1401 struct witness_child_list_entry **wcl;
1402
1403 MPASS(child != NULL && parent != NULL);
1404
1405 /*
1406 * Insert "child" after "parent"
1407 */
1408 wcl = &parent->w_children;
1409 while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1410 wcl = &(*wcl)->wcl_next;
1411 if (*wcl == NULL) {
1412 *wcl = witness_child_get();
1413 if (*wcl == NULL)
1414 return (0);
1415 }
1416 (*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1417
1418 return (1);
1419}
1420
1421/*
1422 * Make all the direct descendants of oldparent be direct descendants
1423 * of newparent.
1424 */
1425static int
1426reparentchildren(struct witness *newparent, struct witness *oldparent)
1427{
1428 struct witness_child_list_entry *wcl;
1429 int i;
1430
1431 /* Avoid making a witness a child of itself. */
1432 MPASS(!isitmychild(oldparent, newparent));
1433
1434 for (wcl = oldparent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1435 for (i = 0; i < wcl->wcl_count; i++)
1436 if (!insertchild(newparent, wcl->wcl_children[i]))
1437 return (0);
1438 return (1);
1439}
1440
1441static int
1442itismychild(struct witness *parent, struct witness *child)
1443{
1444 struct witness_list *list;
1445
1446 MPASS(child != NULL && parent != NULL);
1447 if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
1448 (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
1449 panic(
1450 "%s: parent (%s) and child (%s) are not the same lock type",
1451 __func__, parent->w_class->lc_name,
1452 child->w_class->lc_name);
1453
1454 if (!insertchild(parent, child))
1455 return (0);
1456
1457 if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1458 list = &w_sleep;
1459 else
1460 list = &w_spin;
1461 return (rebalancetree(list));
1462}
1463
1464static void
1465removechild(struct witness *parent, struct witness *child)
1466{
1467 struct witness_child_list_entry **wcl, *wcl1;
1468 int i;
1469
1470 for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1471 for (i = 0; i < (*wcl)->wcl_count; i++)
1472 if ((*wcl)->wcl_children[i] == child)
1473 goto found;
1474 return;
1475found:
1476 (*wcl)->wcl_count--;
1477 if ((*wcl)->wcl_count > i)
1478 (*wcl)->wcl_children[i] =
1479 (*wcl)->wcl_children[(*wcl)->wcl_count];
1480 MPASS((*wcl)->wcl_children[i] != NULL);
1481 if ((*wcl)->wcl_count != 0)
1482 return;
1483 wcl1 = *wcl;
1484 *wcl = wcl1->wcl_next;
1485 witness_child_free(wcl1);
1486}
1487
1488static int
1489isitmychild(struct witness *parent, struct witness *child)
1490{
1491 struct witness_child_list_entry *wcl;
1492 int i;
1493
1494 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1495 for (i = 0; i < wcl->wcl_count; i++) {
1496 if (wcl->wcl_children[i] == child)
1497 return (1);
1498 }
1499 }
1500 return (0);
1501}
1502
1503static int
1504isitmydescendant(struct witness *parent, struct witness *child)
1505{
1506 struct witness_child_list_entry *wcl;
1507 int i, j;
1508
1509 if (isitmychild(parent, child))
1510 return (1);
1511 j = 0;
1512 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1513 MPASS(j < 1000);
1514 for (i = 0; i < wcl->wcl_count; i++) {
1515 if (isitmydescendant(wcl->wcl_children[i], child))
1516 return (1);
1517 }
1518 j++;
1519 }
1520 return (0);
1521}
1522
1523static void
1524witness_levelall (void)
1525{
1526 struct witness_list *list;
1527 struct witness *w, *w1;
1528
1529 /*
1530 * First clear all levels.
1531 */
1532 STAILQ_FOREACH(w, &w_all, w_list) {
1533 w->w_level = 0;
1534 }
1535
1536 /*
1537 * Look for locks with no parent and level all their descendants.
1538 */
1539 STAILQ_FOREACH(w, &w_all, w_list) {
1540 /*
1541 * This is just an optimization, technically we could get
1542 * away just walking the all list each time.
1543 */
1544 if (w->w_class->lc_flags & LC_SLEEPLOCK)
1545 list = &w_sleep;
1546 else
1547 list = &w_spin;
1548 STAILQ_FOREACH(w1, list, w_typelist) {
1549 if (isitmychild(w1, w))
1550 goto skip;
1551 }
1552 witness_leveldescendents(w, 0);
1553 skip:
1554 ; /* silence GCC 3.x */
1555 }
1556}
1557
1558static void
1559witness_leveldescendents(struct witness *parent, int level)
1560{
1561 struct witness_child_list_entry *wcl;
1562 int i;
1563
1564 if (parent->w_level < level)
1565 parent->w_level = level;
1566 level++;
1567 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1568 for (i = 0; i < wcl->wcl_count; i++)
1569 witness_leveldescendents(wcl->wcl_children[i], level);
1570}
1571
1572static void
1573witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1574 struct witness *parent, int indent)
1575{
1576 struct witness_child_list_entry *wcl;
1577 int i, level;
1578
1579 level = parent->w_level;
1580 prnt("%-2d", level);
1581 for (i = 0; i < indent; i++)
1582 prnt(" ");
1583 if (parent->w_refcount > 0)
1584 prnt("%s", parent->w_name);
1585 else
1586 prnt("(dead)");
1587 if (parent->w_displayed) {
1588 prnt(" -- (already displayed)\n");
1589 return;
1590 }
1591 parent->w_displayed = 1;
1592 if (parent->w_refcount > 0) {
1593 if (parent->w_file != NULL)
1594 prnt(" -- last acquired @ %s:%d", parent->w_file,
1595 parent->w_line);
1596 }
1597 prnt("\n");
1598 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1599 for (i = 0; i < wcl->wcl_count; i++)
1600 witness_displaydescendants(prnt,
1601 wcl->wcl_children[i], indent + 1);
1602}
1603
1604#ifdef BLESSING
1605static int
1606blessed(struct witness *w1, struct witness *w2)
1607{
1608 int i;
1609 struct witness_blessed *b;
1610
1611 for (i = 0; i < blessed_count; i++) {
1612 b = &blessed_list[i];
1613 if (strcmp(w1->w_name, b->b_lock1) == 0) {
1614 if (strcmp(w2->w_name, b->b_lock2) == 0)
1615 return (1);
1616 continue;
1617 }
1618 if (strcmp(w1->w_name, b->b_lock2) == 0)
1619 if (strcmp(w2->w_name, b->b_lock1) == 0)
1620 return (1);
1621 }
1622 return (0);
1623}
1624#endif
1625
1626static struct witness *
1627witness_get(void)
1628{
1629 struct witness *w;
1630
1631 if (witness_watch == 0) {
1632 mtx_unlock_spin(&w_mtx);
1633 return (NULL);
1634 }
1635 if (STAILQ_EMPTY(&w_free)) {
1636 witness_watch = 0;
1637 mtx_unlock_spin(&w_mtx);
1638 printf("%s: witness exhausted\n", __func__);
1639 return (NULL);
1640 }
1641 w = STAILQ_FIRST(&w_free);
1642 STAILQ_REMOVE_HEAD(&w_free, w_list);
1643 bzero(w, sizeof(*w));
1644 return (w);
1645}
1646
1647static void
1648witness_free(struct witness *w)
1649{
1650
1651 STAILQ_INSERT_HEAD(&w_free, w, w_list);
1652}
1653
1654static struct witness_child_list_entry *
1655witness_child_get(void)
1656{
1657 struct witness_child_list_entry *wcl;
1658
1659 if (witness_watch == 0) {
1660 mtx_unlock_spin(&w_mtx);
1661 return (NULL);
1662 }
1663 wcl = w_child_free;
1664 if (wcl == NULL) {
1665 witness_watch = 0;
1666 mtx_unlock_spin(&w_mtx);
1667 printf("%s: witness exhausted\n", __func__);
1668 return (NULL);
1669 }
1670 w_child_free = wcl->wcl_next;
1671 bzero(wcl, sizeof(*wcl));
1672 return (wcl);
1673}
1674
1675static void
1676witness_child_free(struct witness_child_list_entry *wcl)
1677{
1678
1679 wcl->wcl_next = w_child_free;
1680 w_child_free = wcl;
1681}
1682
1683static struct lock_list_entry *
1684witness_lock_list_get(void)
1685{
1686 struct lock_list_entry *lle;
1687
1688 if (witness_watch == 0)
1689 return (NULL);
1690 mtx_lock_spin(&w_mtx);
1691 lle = w_lock_list_free;
1692 if (lle == NULL) {
1693 witness_watch = 0;
1694 mtx_unlock_spin(&w_mtx);
1695 printf("%s: witness exhausted\n", __func__);
1696 return (NULL);
1697 }
1698 w_lock_list_free = lle->ll_next;
1699 mtx_unlock_spin(&w_mtx);
1700 bzero(lle, sizeof(*lle));
1701 return (lle);
1702}
1703
1704static void
1705witness_lock_list_free(struct lock_list_entry *lle)
1706{
1707
1708 mtx_lock_spin(&w_mtx);
1709 lle->ll_next = w_lock_list_free;
1710 w_lock_list_free = lle;
1711 mtx_unlock_spin(&w_mtx);
1712}
1713
1714static struct lock_instance *
1715find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1716{
1717 struct lock_list_entry *lle;
1718 struct lock_instance *instance;
1719 int i;
1720
1721 for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1722 for (i = lle->ll_count - 1; i >= 0; i--) {
1723 instance = &lle->ll_children[i];
1724 if (instance->li_lock == lock)
1725 return (instance);
1726 }
1727 return (NULL);
1728}
1729
1730static void
1731witness_list_lock(struct lock_instance *instance)
1732{
1733 struct lock_object *lock;
1734
1735 lock = instance->li_lock;
1736 printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1737 "exclusive" : "shared", lock->lo_class->lc_name, lock->lo_name);
1738 if (lock->lo_type != lock->lo_name)
1739 printf(" (%s)", lock->lo_type);
1740 printf(" r = %d (%p) locked @ %s:%d\n",
1741 instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
1742 instance->li_line);
1743}
1744
1745#ifdef DDB
1746static int
1747witness_thread_has_locks(struct thread *td)
1748{
1749
1750 return (td->td_sleeplocks != NULL);
1751}
1752
1753static int
1754witness_proc_has_locks(struct proc *p)
1755{
1756 struct thread *td;
1757
1758 FOREACH_THREAD_IN_PROC(p, td) {
1759 if (witness_thread_has_locks(td))
1760 return (1);
1761 }
1762 return (0);
1763}
1764#endif
1765
1766int
1767witness_list_locks(struct lock_list_entry **lock_list)
1768{
1769 struct lock_list_entry *lle;
1770 int i, nheld;
1771
1772 nheld = 0;
1773 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1774 for (i = lle->ll_count - 1; i >= 0; i--) {
1775 witness_list_lock(&lle->ll_children[i]);
1776 nheld++;
1777 }
1778 return (nheld);
1779}
1780
1781/*
1782 * This is a bit risky at best. We call this function when we have timed
1783 * out acquiring a spin lock, and we assume that the other CPU is stuck
1784 * with this lock held. So, we go groveling around in the other CPU's
1785 * per-cpu data to try to find the lock instance for this spin lock to
1786 * see when it was last acquired.
1787 */
1788void
1789witness_display_spinlock(struct lock_object *lock, struct thread *owner)
1790{
1791 struct lock_instance *instance;
1792 struct pcpu *pc;
1793
1794 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
1795 return;
1796 pc = pcpu_find(owner->td_oncpu);
1797 instance = find_instance(pc->pc_spinlocks, lock);
1798 if (instance != NULL)
1799 witness_list_lock(instance);
1800}
1801
1802void
1803witness_save(struct lock_object *lock, const char **filep, int *linep)
1804{
1805 struct lock_instance *instance;
1806
1807 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1808 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1809 return;
1810 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1811 panic("%s: lock (%s) %s is not a sleep lock", __func__,
1812 lock->lo_class->lc_name, lock->lo_name);
1813 instance = find_instance(curthread->td_sleeplocks, lock);
1814 if (instance == NULL)
1815 panic("%s: lock (%s) %s not locked", __func__,
1816 lock->lo_class->lc_name, lock->lo_name);
1817 *filep = instance->li_file;
1818 *linep = instance->li_line;
1819}
1820
1821void
1822witness_restore(struct lock_object *lock, const char *file, int line)
1823{
1824 struct lock_instance *instance;
1825
1826 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1827 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1828 return;
1829 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1830 panic("%s: lock (%s) %s is not a sleep lock", __func__,
1831 lock->lo_class->lc_name, lock->lo_name);
1832 instance = find_instance(curthread->td_sleeplocks, lock);
1833 if (instance == NULL)
1834 panic("%s: lock (%s) %s not locked", __func__,
1835 lock->lo_class->lc_name, lock->lo_name);
1836 lock->lo_witness->w_file = file;
1837 lock->lo_witness->w_line = line;
1838 instance->li_file = file;
1839 instance->li_line = line;
1840}
1841
1842void
1843witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1844{
1845#ifdef INVARIANT_SUPPORT
1846 struct lock_instance *instance;
1847
1848 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1849 return;
1850 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1851 instance = find_instance(curthread->td_sleeplocks, lock);
1852 else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1853 instance = find_instance(PCPU_GET(spinlocks), lock);
1854 else {
1855 panic("Lock (%s) %s is not sleep or spin!",
1856 lock->lo_class->lc_name, lock->lo_name);
1857 }
1858 file = fixup_filename(file);
1859 switch (flags) {
1860 case LA_UNLOCKED:
1861 if (instance != NULL)
1862 panic("Lock (%s) %s locked @ %s:%d.",
1863 lock->lo_class->lc_name, lock->lo_name, file, line);
1864 break;
1865 case LA_LOCKED:
1866 case LA_LOCKED | LA_RECURSED:
1867 case LA_LOCKED | LA_NOTRECURSED:
1868 case LA_SLOCKED:
1869 case LA_SLOCKED | LA_RECURSED:
1870 case LA_SLOCKED | LA_NOTRECURSED:
1871 case LA_XLOCKED:
1872 case LA_XLOCKED | LA_RECURSED:
1873 case LA_XLOCKED | LA_NOTRECURSED:
1874 if (instance == NULL) {
1875 panic("Lock (%s) %s not locked @ %s:%d.",
1876 lock->lo_class->lc_name, lock->lo_name, file, line);
1877 break;
1878 }
1879 if ((flags & LA_XLOCKED) != 0 &&
1880 (instance->li_flags & LI_EXCLUSIVE) == 0)
1881 panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1882 lock->lo_class->lc_name, lock->lo_name, file, line);
1883 if ((flags & LA_SLOCKED) != 0 &&
1884 (instance->li_flags & LI_EXCLUSIVE) != 0)
1885 panic("Lock (%s) %s exclusively locked @ %s:%d.",
1886 lock->lo_class->lc_name, lock->lo_name, file, line);
1887 if ((flags & LA_RECURSED) != 0 &&
1888 (instance->li_flags & LI_RECURSEMASK) == 0)
1889 panic("Lock (%s) %s not recursed @ %s:%d.",
1890 lock->lo_class->lc_name, lock->lo_name, file, line);
1891 if ((flags & LA_NOTRECURSED) != 0 &&
1892 (instance->li_flags & LI_RECURSEMASK) != 0)
1893 panic("Lock (%s) %s recursed @ %s:%d.",
1894 lock->lo_class->lc_name, lock->lo_name, file, line);
1895 break;
1896 default:
1897 panic("Invalid lock assertion at %s:%d.", file, line);
1898
1899 }
1900#endif /* INVARIANT_SUPPORT */
1901}
1902
1903#ifdef DDB
1904static void
1905witness_list(struct thread *td)
1906{
1907
1908 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1909 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
1910
1911 if (witness_watch == 0)
1912 return;
1913
1914 witness_list_locks(&td->td_sleeplocks);
1915
1916 /*
1917 * We only handle spinlocks if td == curthread. This is somewhat broken
1918 * if td is currently executing on some other CPU and holds spin locks
1919 * as we won't display those locks. If we had a MI way of getting
1920 * the per-cpu data for a given cpu then we could use
1921 * td->td_oncpu to get the list of spinlocks for this thread
1922 * and "fix" this.
1923 *
1924 * That still wouldn't really fix this unless we locked sched_lock
1925 * or stopped the other CPU to make sure it wasn't changing the list
1926 * out from under us. It is probably best to just not try to handle
1927 * threads on other CPU's for now.
1928 */
1929 if (td == curthread && PCPU_GET(spinlocks) != NULL)
1930 witness_list_locks(PCPU_PTR(spinlocks));
1931}
1932
1933DB_SHOW_COMMAND(locks, db_witness_list)
1934{
1935 struct thread *td;
1936 pid_t pid;
1937 struct proc *p;
1938
1939 if (have_addr) {
1940 pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1941 ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1942 ((addr >> 16) % 16) * 10000;
1943 /* sx_slock(&allproc_lock); */
1944 FOREACH_PROC_IN_SYSTEM(p) {
1945 if (p->p_pid == pid)
1946 break;
1947 }
1948 /* sx_sunlock(&allproc_lock); */
1949 if (p == NULL) {
1950 db_printf("pid %d not found\n", pid);
1951 return;
1952 }
1953 FOREACH_THREAD_IN_PROC(p, td) {
1954 witness_list(td);
1955 }
1956 } else {
1957 td = curthread;
1958 witness_list(td);
1959 }
1960}
1961
1962DB_SHOW_COMMAND(alllocks, db_witness_list_all)
1963{
1964 struct thread *td;
1965 struct proc *p;
1966
1967 /*
1968 * It would be nice to list only threads and processes that actually
1969 * held sleep locks, but that information is currently not exported
1970 * by WITNESS.
1971 */
1972 FOREACH_PROC_IN_SYSTEM(p) {
1973 if (!witness_proc_has_locks(p))
1974 continue;
1975 FOREACH_THREAD_IN_PROC(p, td) {
1976 if (!witness_thread_has_locks(td))
1977 continue;
1978 printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
1979 p->p_comm, td, td->td_tid);
1980 witness_list(td);
1981 }
1982 }
1983}
1984
1985DB_SHOW_COMMAND(witness, db_witness_display)
1986{
1987
1988 witness_display(db_printf);
1989}
1990#endif
|