xref: /dflybsd-src/sys/kern/lwkt_token.c (revision d88b9605be186eb2c93c4f2683c997a128758b5a)
1 /*
2  * Copyright (c) 2003,2004,2009 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 
35 /*
36  * lwkt_token - Implement soft token locks.
37  *
38  * Tokens are locks which serialize a thread only while the thread is
39  * running.  If the thread blocks all tokens are released, then reacquired
40  * when the thread resumes.
41  *
42  * This implementation requires no critical sections or spin locks, but
43  * does use atomic_cmpset_ptr().
44  *
45  * Tokens may be recursively acquired by the same thread.  However the
46  * caller must be sure to release such tokens in reverse order.
47  */
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/kernel.h>
51 #include <sys/proc.h>
52 #include <sys/rtprio.h>
53 #include <sys/queue.h>
54 #include <sys/sysctl.h>
55 #include <sys/ktr.h>
56 #include <ddb/ddb.h>
57 #include <sys/kthread.h>
58 #include <machine/cpu.h>
59 #include <sys/lock.h>
60 #include <sys/spinlock.h>
61 
62 #include <sys/thread2.h>
63 #include <sys/spinlock2.h>
64 #include <sys/mplock2.h>
65 
66 #include <vm/vm.h>
67 #include <vm/vm_param.h>
68 #include <vm/vm_kern.h>
69 #include <vm/vm_object.h>
70 #include <vm/vm_page.h>
71 #include <vm/vm_map.h>
72 #include <vm/vm_pager.h>
73 #include <vm/vm_extern.h>
74 #include <vm/vm_zone.h>
75 
76 #include <machine/stdarg.h>
77 #include <machine/smp.h>
78 
79 extern int lwkt_sched_debug;
80 
81 #ifndef LWKT_NUM_POOL_TOKENS
82 #define LWKT_NUM_POOL_TOKENS	4001	/* prime number */
83 #endif
84 
85 static lwkt_token	pool_tokens[LWKT_NUM_POOL_TOKENS];
86 struct spinlock		tok_debug_spin = SPINLOCK_INITIALIZER(&tok_debug_spin);
87 
88 #define TOKEN_STRING	"REF=%p TOK=%p TD=%p"
89 #define TOKEN_ARGS	lwkt_tokref_t ref, lwkt_token_t tok, struct thread *td
90 #define CONTENDED_STRING	TOKEN_STRING " (contention started)"
91 #define UNCONTENDED_STRING	TOKEN_STRING " (contention stopped)"
92 #if !defined(KTR_TOKENS)
93 #define	KTR_TOKENS	KTR_ALL
94 #endif
95 
96 KTR_INFO_MASTER(tokens);
97 KTR_INFO(KTR_TOKENS, tokens, fail, 0, TOKEN_STRING, TOKEN_ARGS);
98 KTR_INFO(KTR_TOKENS, tokens, succ, 1, TOKEN_STRING, TOKEN_ARGS);
99 #if 0
100 KTR_INFO(KTR_TOKENS, tokens, release, 2, TOKEN_STRING, TOKEN_ARGS);
101 KTR_INFO(KTR_TOKENS, tokens, remote, 3, TOKEN_STRING, TOKEN_ARGS);
102 KTR_INFO(KTR_TOKENS, tokens, reqremote, 4, TOKEN_STRING, TOKEN_ARGS);
103 KTR_INFO(KTR_TOKENS, tokens, reqfail, 5, TOKEN_STRING, TOKEN_ARGS);
104 KTR_INFO(KTR_TOKENS, tokens, drain, 6, TOKEN_STRING, TOKEN_ARGS);
105 KTR_INFO(KTR_TOKENS, tokens, contention_start, 7, CONTENDED_STRING, TOKEN_ARGS);
106 KTR_INFO(KTR_TOKENS, tokens, contention_stop, 7, UNCONTENDED_STRING, TOKEN_ARGS);
107 #endif
108 
109 #define logtoken(name, ref)						\
110 	KTR_LOG(tokens_ ## name, ref, ref->tr_tok, curthread)
111 
112 /*
113  * Global tokens.  These replace the MP lock for major subsystem locking.
114  * These tokens are initially used to lockup both global and individual
115  * operations.
116  *
117  * Once individual structures get their own locks these tokens are used
118  * only to protect global lists & other variables and to interlock
119  * allocations and teardowns and such.
120  *
121  * The UP initializer causes token acquisition to also acquire the MP lock
122  * for maximum compatibility.  The feature may be enabled and disabled at
123  * any time, the MP state is copied to the tokref when the token is acquired
124  * and will not race against sysctl changes.
125  */
126 struct lwkt_token mp_token = LWKT_TOKEN_INITIALIZER(mp_token);
127 struct lwkt_token pmap_token = LWKT_TOKEN_INITIALIZER(pmap_token);
128 struct lwkt_token dev_token = LWKT_TOKEN_INITIALIZER(dev_token);
129 struct lwkt_token vm_token = LWKT_TOKEN_INITIALIZER(vm_token);
130 struct lwkt_token vmspace_token = LWKT_TOKEN_INITIALIZER(vmspace_token);
131 struct lwkt_token kvm_token = LWKT_TOKEN_INITIALIZER(kvm_token);
132 struct lwkt_token proc_token = LWKT_TOKEN_INITIALIZER(proc_token);
133 struct lwkt_token tty_token = LWKT_TOKEN_INITIALIZER(tty_token);
134 struct lwkt_token vnode_token = LWKT_TOKEN_INITIALIZER(vnode_token);
135 struct lwkt_token vmobj_token = LWKT_TOKEN_INITIALIZER(vmobj_token);
136 
137 static int lwkt_token_spin = 5;
138 SYSCTL_INT(_lwkt, OID_AUTO, token_spin, CTLFLAG_RW,
139     &lwkt_token_spin, 0, "Decontention spin loops");
140 static int lwkt_token_delay = 0;
141 SYSCTL_INT(_lwkt, OID_AUTO, token_delay, CTLFLAG_RW,
142     &lwkt_token_delay, 0, "Decontention spin delay in ns");
143 
144 /*
145  * The collision count is bumped every time the LWKT scheduler fails
146  * to acquire needed tokens in addition to a normal lwkt_gettoken()
147  * stall.
148  */
149 SYSCTL_LONG(_lwkt, OID_AUTO, mp_collisions, CTLFLAG_RW,
150     &mp_token.t_collisions, 0, "Collision counter of mp_token");
151 SYSCTL_LONG(_lwkt, OID_AUTO, pmap_collisions, CTLFLAG_RW,
152     &pmap_token.t_collisions, 0, "Collision counter of pmap_token");
153 SYSCTL_LONG(_lwkt, OID_AUTO, dev_collisions, CTLFLAG_RW,
154     &dev_token.t_collisions, 0, "Collision counter of dev_token");
155 SYSCTL_LONG(_lwkt, OID_AUTO, vm_collisions, CTLFLAG_RW,
156     &vm_token.t_collisions, 0, "Collision counter of vm_token");
157 SYSCTL_LONG(_lwkt, OID_AUTO, vmspace_collisions, CTLFLAG_RW,
158     &vmspace_token.t_collisions, 0, "Collision counter of vmspace_token");
159 SYSCTL_LONG(_lwkt, OID_AUTO, kvm_collisions, CTLFLAG_RW,
160     &kvm_token.t_collisions, 0, "Collision counter of kvm_token");
161 SYSCTL_LONG(_lwkt, OID_AUTO, proc_collisions, CTLFLAG_RW,
162     &proc_token.t_collisions, 0, "Collision counter of proc_token");
163 SYSCTL_LONG(_lwkt, OID_AUTO, tty_collisions, CTLFLAG_RW,
164     &tty_token.t_collisions, 0, "Collision counter of tty_token");
165 SYSCTL_LONG(_lwkt, OID_AUTO, vnode_collisions, CTLFLAG_RW,
166     &vnode_token.t_collisions, 0, "Collision counter of vnode_token");
167 
168 int tokens_debug_output;
169 SYSCTL_INT(_lwkt, OID_AUTO, tokens_debug_output, CTLFLAG_RW,
170     &tokens_debug_output, 0, "Generate stack trace N times");
171 
172 
173 #ifdef DEBUG_LOCKS_LATENCY
174 
175 static long tokens_add_latency;
176 SYSCTL_LONG(_debug, OID_AUTO, tokens_add_latency, CTLFLAG_RW,
177 	    &tokens_add_latency, 0,
178 	    "Add spinlock latency");
179 
180 #endif
181 
182 
183 static int _lwkt_getalltokens_sorted(thread_t td);
184 
185 /*
186  * Acquire the initial mplock
187  *
188  * (low level boot only)
189  */
190 void
191 cpu_get_initial_mplock(void)
192 {
193 	KKASSERT(mp_token.t_ref == NULL);
194 	if (lwkt_trytoken(&mp_token) == FALSE)
195 		panic("cpu_get_initial_mplock");
196 }
197 
198 /*
199  * Return a pool token given an address.  Use a prime number to reduce
200  * overlaps.
201  */
202 static __inline
203 lwkt_token_t
204 _lwkt_token_pool_lookup(void *ptr)
205 {
206 	u_int i;
207 
208 	i = (u_int)(uintptr_t)ptr % LWKT_NUM_POOL_TOKENS;
209 	return(&pool_tokens[i]);
210 }
211 
212 /*
213  * Initialize a tokref_t prior to making it visible in the thread's
214  * token array.
215  */
216 static __inline
217 void
218 _lwkt_tokref_init(lwkt_tokref_t ref, lwkt_token_t tok, thread_t td, long excl)
219 {
220 	ref->tr_tok = tok;
221 	ref->tr_count = excl;
222 	ref->tr_owner = td;
223 }
224 
225 /*
226  * Attempt to acquire a shared or exclusive token.  Returns TRUE on success,
227  * FALSE on failure.
228  *
229  * If TOK_EXCLUSIVE is set in mode we are attempting to get an exclusive
230  * token, otherwise are attempting to get a shared token.
231  *
232  * If TOK_EXCLREQ is set in mode this is a blocking operation, otherwise
233  * it is a non-blocking operation (for both exclusive or shared acquisions).
234  */
235 static __inline
236 int
237 _lwkt_trytokref(lwkt_tokref_t ref, thread_t td, long mode)
238 {
239 	lwkt_token_t tok;
240 	lwkt_tokref_t oref;
241 	long count;
242 
243 	tok = ref->tr_tok;
244 	KASSERT(((mode & TOK_EXCLREQ) == 0 ||	/* non blocking */
245 		td->td_gd->gd_intr_nesting_level == 0 ||
246 		panic_cpu_gd == mycpu),
247 		("Attempt to acquire token %p not already "
248 		"held in hard code section", tok));
249 
250 	if (mode & TOK_EXCLUSIVE) {
251 		/*
252 		 * Attempt to get an exclusive token
253 		 */
254 		for (;;) {
255 			count = tok->t_count;
256 			oref = tok->t_ref;	/* can be NULL */
257 			cpu_ccfence();
258 			if ((count & ~TOK_EXCLREQ) == 0) {
259 				/*
260 				 * It is possible to get the exclusive bit.
261 				 * We must clear TOK_EXCLREQ on successful
262 				 * acquisition.
263 				 */
264 				if (atomic_cmpset_long(&tok->t_count, count,
265 						       (count & ~TOK_EXCLREQ) |
266 						       TOK_EXCLUSIVE)) {
267 					KKASSERT(tok->t_ref == NULL);
268 					tok->t_ref = ref;
269 					return TRUE;
270 				}
271 				/* retry */
272 			} else if ((count & TOK_EXCLUSIVE) &&
273 				   oref >= &td->td_toks_base &&
274 				   oref < td->td_toks_stop) {
275 				/*
276 				 * Our thread already holds the exclusive
277 				 * bit, we treat this tokref as a shared
278 				 * token (sorta) to make the token release
279 				 * code easier.
280 				 *
281 				 * NOTE: oref cannot race above if it
282 				 *	 happens to be ours, so we're good.
283 				 *	 But we must still have a stable
284 				 *	 variable for both parts of the
285 				 *	 comparison.
286 				 *
287 				 * NOTE: Since we already have an exclusive
288 				 *	 lock and don't need to check EXCLREQ
289 				 *	 we can just use an atomic_add here
290 				 */
291 				atomic_add_long(&tok->t_count, TOK_INCR);
292 				ref->tr_count &= ~TOK_EXCLUSIVE;
293 				return TRUE;
294 			} else if ((mode & TOK_EXCLREQ) &&
295 				   (count & TOK_EXCLREQ) == 0) {
296 				/*
297 				 * Unable to get the exclusive bit but being
298 				 * asked to set the exclusive-request bit.
299 				 * Since we are going to retry anyway just
300 				 * set the bit unconditionally.
301 				 */
302 				atomic_set_long(&tok->t_count, TOK_EXCLREQ);
303 				return FALSE;
304 			} else {
305 				/*
306 				 * Unable to get the exclusive bit and not
307 				 * being asked to set the exclusive-request
308 				 * (aka lwkt_trytoken()), or EXCLREQ was
309 				 * already set.
310 				 */
311 				cpu_pause();
312 				return FALSE;
313 			}
314 			/* retry */
315 		}
316 	} else {
317 		/*
318 		 * Attempt to get a shared token.  Note that TOK_EXCLREQ
319 		 * for shared tokens simply means the caller intends to
320 		 * block.  We never actually set the bit in tok->t_count.
321 		 */
322 		for (;;) {
323 			count = tok->t_count;
324 			oref = tok->t_ref;	/* can be NULL */
325 			cpu_ccfence();
326 			if ((count & (TOK_EXCLUSIVE/*|TOK_EXCLREQ*/)) == 0) {
327 				/* XXX EXCLREQ should work */
328 				/*
329 				 * It is possible to get the token shared.
330 				 */
331 				if (atomic_cmpset_long(&tok->t_count, count,
332 						       count + TOK_INCR)) {
333 					return TRUE;
334 				}
335 				/* retry */
336 			} else if ((count & TOK_EXCLUSIVE) &&
337 				   oref >= &td->td_toks_base &&
338 				   oref < td->td_toks_stop) {
339 				/*
340 				 * We own the exclusive bit on the token so
341 				 * we can in fact also get it shared.
342 				 */
343 				atomic_add_long(&tok->t_count, TOK_INCR);
344 				return TRUE;
345 			} else {
346 				/*
347 				 * We failed to get the token shared
348 				 */
349 				return FALSE;
350 			}
351 			/* retry */
352 		}
353 	}
354 }
355 
356 static __inline
357 int
358 _lwkt_trytokref_spin(lwkt_tokref_t ref, thread_t td, long mode)
359 {
360 	int spin;
361 
362 	if (_lwkt_trytokref(ref, td, mode)) {
363 #ifdef DEBUG_LOCKS_LATENCY
364 		long j;
365 		for (j = tokens_add_latency; j > 0; --j)
366 			cpu_ccfence();
367 #endif
368 		return TRUE;
369 	}
370 	for (spin = lwkt_token_spin; spin > 0; --spin) {
371 		if (lwkt_token_delay)
372 			tsc_delay(lwkt_token_delay);
373 		else
374 			cpu_pause();
375 		if (_lwkt_trytokref(ref, td, mode)) {
376 #ifdef DEBUG_LOCKS_LATENCY
377 			long j;
378 			for (j = tokens_add_latency; j > 0; --j)
379 				cpu_ccfence();
380 #endif
381 			return TRUE;
382 		}
383 	}
384 	return FALSE;
385 }
386 
387 /*
388  * Release a token that we hold.
389  */
390 static __inline
391 void
392 _lwkt_reltokref(lwkt_tokref_t ref, thread_t td)
393 {
394 	lwkt_token_t tok;
395 	long count;
396 
397 	tok = ref->tr_tok;
398 	for (;;) {
399 		count = tok->t_count;
400 		cpu_ccfence();
401 		if (tok->t_ref == ref) {
402 			/*
403 			 * We are an exclusive holder.  We must clear tr_ref
404 			 * before we clear the TOK_EXCLUSIVE bit.  If we are
405 			 * unable to clear the bit we must restore
406 			 * tok->t_ref.
407 			 */
408 			KKASSERT(count & TOK_EXCLUSIVE);
409 			tok->t_ref = NULL;
410 			if (atomic_cmpset_long(&tok->t_count, count,
411 					       count & ~TOK_EXCLUSIVE)) {
412 				return;
413 			}
414 			tok->t_ref = ref;
415 			/* retry */
416 		} else {
417 			/*
418 			 * We are a shared holder
419 			 */
420 			KKASSERT(count & TOK_COUNTMASK);
421 			if (atomic_cmpset_long(&tok->t_count, count,
422 					       count - TOK_INCR)) {
423 				return;
424 			}
425 			/* retry */
426 		}
427 		/* retry */
428 	}
429 }
430 
431 /*
432  * Obtain all the tokens required by the specified thread on the current
433  * cpu, return 0 on failure and non-zero on success.  If a failure occurs
434  * any partially acquired tokens will be released prior to return.
435  *
436  * lwkt_getalltokens is called by the LWKT scheduler to re-acquire all
437  * tokens that the thread had to release when it switched away.
438  *
439  * If spinning is non-zero this function acquires the tokens in a particular
440  * order to deal with potential deadlocks.  We simply use address order for
441  * the case.
442  *
443  * Called from a critical section.
444  */
445 int
446 lwkt_getalltokens(thread_t td, int spinning)
447 {
448 	lwkt_tokref_t scan;
449 	lwkt_token_t tok;
450 
451 	if (spinning)
452 		return(_lwkt_getalltokens_sorted(td));
453 
454 	/*
455 	 * Acquire tokens in forward order, assign or validate tok->t_ref.
456 	 */
457 	for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) {
458 		tok = scan->tr_tok;
459 		for (;;) {
460 			/*
461 			 * Only try really hard on the last token
462 			 */
463 			if (scan == td->td_toks_stop - 1) {
464 			    if (_lwkt_trytokref_spin(scan, td, scan->tr_count))
465 				    break;
466 			} else {
467 			    if (_lwkt_trytokref(scan, td, scan->tr_count))
468 				    break;
469 			}
470 
471 			/*
472 			 * Otherwise we failed to acquire all the tokens.
473 			 * Release whatever we did get.
474 			 */
475 			KASSERT(tok->t_desc, ("token %p is not initialized", tok));
476 			strncpy(td->td_gd->gd_cnt.v_token_name,
477 				tok->t_desc,
478 				sizeof(td->td_gd->gd_cnt.v_token_name) - 1);
479 
480 			if (lwkt_sched_debug > 0) {
481 				--lwkt_sched_debug;
482 				kprintf("toka %p %s %s\n",
483 					tok, tok->t_desc, td->td_comm);
484 			}
485 			td->td_wmesg = tok->t_desc;
486 			++tok->t_collisions;
487 			while (--scan >= &td->td_toks_base)
488 				_lwkt_reltokref(scan, td);
489 			return(FALSE);
490 		}
491 	}
492 	return (TRUE);
493 }
494 
495 /*
496  * Release all tokens owned by the specified thread on the current cpu.
497  *
498  * This code is really simple.  Even in cases where we own all the tokens
499  * note that t_ref may not match the scan for recursively held tokens which
500  * are held deeper in the stack, or for the case where a lwkt_getalltokens()
501  * failed.
502  *
503  * Tokens are released in reverse order to reduce chasing race failures.
504  *
505  * Called from a critical section.
506  */
507 void
508 lwkt_relalltokens(thread_t td)
509 {
510 	lwkt_tokref_t scan;
511 
512 	/*
513 	 * Weird order is to try to avoid a panic loop
514 	 */
515 	if (td->td_toks_have) {
516 		scan = td->td_toks_have;
517 		td->td_toks_have = NULL;
518 	} else {
519 		scan = td->td_toks_stop;
520 	}
521 	while (--scan >= &td->td_toks_base)
522 		_lwkt_reltokref(scan, td);
523 }
524 
525 /*
526  * This is the decontention version of lwkt_getalltokens().  The tokens are
527  * acquired in address-sorted order to deal with any deadlocks.  Ultimately
528  * token failures will spin into the scheduler and get here.
529  *
530  * Called from critical section
531  */
532 static
533 int
534 _lwkt_getalltokens_sorted(thread_t td)
535 {
536 	lwkt_tokref_t sort_array[LWKT_MAXTOKENS];
537 	lwkt_tokref_t scan;
538 	lwkt_token_t tok;
539 	int i;
540 	int j;
541 	int n;
542 
543 	/*
544 	 * Sort the token array.  Yah yah, I know this isn't fun.
545 	 *
546 	 * NOTE: Recursively acquired tokens are ordered the same as in the
547 	 *	 td_toks_array so we can always get the earliest one first.
548 	 */
549 	i = 0;
550 	scan = &td->td_toks_base;
551 	while (scan < td->td_toks_stop) {
552 		for (j = 0; j < i; ++j) {
553 			if (scan->tr_tok < sort_array[j]->tr_tok)
554 				break;
555 		}
556 		if (j != i) {
557 			bcopy(sort_array + j, sort_array + j + 1,
558 			      (i - j) * sizeof(lwkt_tokref_t));
559 		}
560 		sort_array[j] = scan;
561 		++scan;
562 		++i;
563 	}
564 	n = i;
565 
566 	/*
567 	 * Acquire tokens in forward order, assign or validate tok->t_ref.
568 	 */
569 	for (i = 0; i < n; ++i) {
570 		scan = sort_array[i];
571 		tok = scan->tr_tok;
572 		for (;;) {
573 			/*
574 			 * Only try really hard on the last token
575 			 */
576 			if (scan == td->td_toks_stop - 1) {
577 			    if (_lwkt_trytokref_spin(scan, td, scan->tr_count))
578 				    break;
579 			} else {
580 			    if (_lwkt_trytokref(scan, td, scan->tr_count))
581 				    break;
582 			}
583 
584 			/*
585 			 * Otherwise we failed to acquire all the tokens.
586 			 * Release whatever we did get.
587 			 */
588 			if (lwkt_sched_debug > 0) {
589 				--lwkt_sched_debug;
590 				kprintf("tokb %p %s %s\n",
591 					tok, tok->t_desc, td->td_comm);
592 			}
593 			td->td_wmesg = tok->t_desc;
594 			++tok->t_collisions;
595 			while (--i >= 0) {
596 				scan = sort_array[i];
597 				_lwkt_reltokref(scan, td);
598 			}
599 			return(FALSE);
600 		}
601 	}
602 
603 	/*
604 	 * We were successful, there is no need for another core to signal
605 	 * us.
606 	 */
607 	return (TRUE);
608 }
609 
610 /*
611  * Get a serializing token.  This routine can block.
612  */
613 void
614 lwkt_gettoken(lwkt_token_t tok)
615 {
616 	thread_t td = curthread;
617 	lwkt_tokref_t ref;
618 
619 	ref = td->td_toks_stop;
620 	KKASSERT(ref < &td->td_toks_end);
621 	++td->td_toks_stop;
622 	cpu_ccfence();
623 	_lwkt_tokref_init(ref, tok, td, TOK_EXCLUSIVE|TOK_EXCLREQ);
624 
625 #ifdef DEBUG_LOCKS
626 	/*
627 	 * Taking an exclusive token after holding it shared will
628 	 * livelock. Scan for that case and assert.
629 	 */
630 	lwkt_tokref_t tk;
631 	int found = 0;
632 	for (tk = &td->td_toks_base; tk < ref; tk++) {
633 		if (tk->tr_tok != tok)
634 			continue;
635 
636 		found++;
637 		if (tk->tr_count & TOK_EXCLUSIVE)
638 			goto good;
639 	}
640 	/* We found only shared instances of this token if found >0 here */
641 	KASSERT((found == 0), ("Token %p s/x livelock", tok));
642 good:
643 #endif
644 
645 	if (_lwkt_trytokref_spin(ref, td, TOK_EXCLUSIVE|TOK_EXCLREQ))
646 		return;
647 
648 	/*
649 	 * Give up running if we can't acquire the token right now.
650 	 *
651 	 * Since the tokref is already active the scheduler now
652 	 * takes care of acquisition, so we need only call
653 	 * lwkt_switch().
654 	 *
655 	 * Since we failed this was not a recursive token so upon
656 	 * return tr_tok->t_ref should be assigned to this specific
657 	 * ref.
658 	 */
659 	td->td_wmesg = tok->t_desc;
660 	++tok->t_collisions;
661 	logtoken(fail, ref);
662 	td->td_toks_have = td->td_toks_stop - 1;
663 
664 	if (tokens_debug_output > 0) {
665 		--tokens_debug_output;
666 		spin_lock(&tok_debug_spin);
667 		kprintf("Excl Token thread %p %s %s\n",
668 			td, tok->t_desc, td->td_comm);
669 		print_backtrace(6);
670 		kprintf("\n");
671 		spin_unlock(&tok_debug_spin);
672 	}
673 
674 	lwkt_switch();
675 	logtoken(succ, ref);
676 	KKASSERT(tok->t_ref == ref);
677 }
678 
679 /*
680  * Similar to gettoken but we acquire a shared token instead of an exclusive
681  * token.
682  */
683 void
684 lwkt_gettoken_shared(lwkt_token_t tok)
685 {
686 	thread_t td = curthread;
687 	lwkt_tokref_t ref;
688 
689 	ref = td->td_toks_stop;
690 	KKASSERT(ref < &td->td_toks_end);
691 	++td->td_toks_stop;
692 	cpu_ccfence();
693 	_lwkt_tokref_init(ref, tok, td, TOK_EXCLREQ);
694 
695 #ifdef DEBUG_LOCKS
696         /*
697          * Taking a pool token in shared mode is a bad idea; other
698          * addresses deeper in the call stack may hash to the same pool
699          * token and you may end up with an exclusive-shared livelock.
700          * Warn in this condition.
701          */
702         if ((tok >= &pool_tokens[0]) &&
703             (tok < &pool_tokens[LWKT_NUM_POOL_TOKENS]))
704                 kprintf("Warning! Taking pool token %p in shared mode\n", tok);
705 #endif
706 
707 
708 	if (_lwkt_trytokref_spin(ref, td, TOK_EXCLREQ))
709 		return;
710 
711 	/*
712 	 * Give up running if we can't acquire the token right now.
713 	 *
714 	 * Since the tokref is already active the scheduler now
715 	 * takes care of acquisition, so we need only call
716 	 * lwkt_switch().
717 	 *
718 	 * Since we failed this was not a recursive token so upon
719 	 * return tr_tok->t_ref should be assigned to this specific
720 	 * ref.
721 	 */
722 	td->td_wmesg = tok->t_desc;
723 	++tok->t_collisions;
724 	logtoken(fail, ref);
725 	td->td_toks_have = td->td_toks_stop - 1;
726 
727 	if (tokens_debug_output > 0) {
728 		--tokens_debug_output;
729 		spin_lock(&tok_debug_spin);
730 		kprintf("Shar Token thread %p %s %s\n",
731 			td, tok->t_desc, td->td_comm);
732 		print_backtrace(6);
733 		kprintf("\n");
734 		spin_unlock(&tok_debug_spin);
735 	}
736 
737 	lwkt_switch();
738 	logtoken(succ, ref);
739 }
740 
741 /*
742  * Attempt to acquire a token, return TRUE on success, FALSE on failure.
743  *
744  * We setup the tokref in case we actually get the token (if we switch later
745  * it becomes mandatory so we set TOK_EXCLREQ), but we call trytokref without
746  * TOK_EXCLREQ in case we fail.
747  */
748 int
749 lwkt_trytoken(lwkt_token_t tok)
750 {
751 	thread_t td = curthread;
752 	lwkt_tokref_t ref;
753 
754 	ref = td->td_toks_stop;
755 	KKASSERT(ref < &td->td_toks_end);
756 	++td->td_toks_stop;
757 	cpu_ccfence();
758 	_lwkt_tokref_init(ref, tok, td, TOK_EXCLUSIVE|TOK_EXCLREQ);
759 
760 	if (_lwkt_trytokref(ref, td, TOK_EXCLUSIVE))
761 		return TRUE;
762 
763 	/*
764 	 * Failed, unpend the request
765 	 */
766 	cpu_ccfence();
767 	--td->td_toks_stop;
768 	++tok->t_collisions;
769 	return FALSE;
770 }
771 
772 
773 void
774 lwkt_gettoken_hard(lwkt_token_t tok)
775 {
776 	lwkt_gettoken(tok);
777 	crit_enter_hard();
778 }
779 
780 lwkt_token_t
781 lwkt_getpooltoken(void *ptr)
782 {
783 	lwkt_token_t tok;
784 
785 	tok = _lwkt_token_pool_lookup(ptr);
786 	lwkt_gettoken(tok);
787 	return (tok);
788 }
789 
790 /*
791  * Release a serializing token.
792  *
793  * WARNING!  All tokens must be released in reverse order.  This will be
794  *	     asserted.
795  */
796 void
797 lwkt_reltoken(lwkt_token_t tok)
798 {
799 	thread_t td = curthread;
800 	lwkt_tokref_t ref;
801 
802 	/*
803 	 * Remove ref from thread token list and assert that it matches
804 	 * the token passed in.  Tokens must be released in reverse order.
805 	 */
806 	ref = td->td_toks_stop - 1;
807 	KKASSERT(ref >= &td->td_toks_base && ref->tr_tok == tok);
808 	_lwkt_reltokref(ref, td);
809 	cpu_sfence();
810 	td->td_toks_stop = ref;
811 }
812 
813 void
814 lwkt_reltoken_hard(lwkt_token_t tok)
815 {
816 	lwkt_reltoken(tok);
817 	crit_exit_hard();
818 }
819 
820 /*
821  * It is faster for users of lwkt_getpooltoken() to use the returned
822  * token and just call lwkt_reltoken(), but for convenience we provide
823  * this function which looks the token up based on the ident.
824  */
825 void
826 lwkt_relpooltoken(void *ptr)
827 {
828 	lwkt_token_t tok = _lwkt_token_pool_lookup(ptr);
829 	lwkt_reltoken(tok);
830 }
831 
832 /*
833  * Return a count of the number of token refs the thread has to the
834  * specified token, whether it currently owns the token or not.
835  */
836 int
837 lwkt_cnttoken(lwkt_token_t tok, thread_t td)
838 {
839 	lwkt_tokref_t scan;
840 	int count = 0;
841 
842 	for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) {
843 		if (scan->tr_tok == tok)
844 			++count;
845 	}
846 	return(count);
847 }
848 
849 /*
850  * Pool tokens are used to provide a type-stable serializing token
851  * pointer that does not race against disappearing data structures.
852  *
853  * This routine is called in early boot just after we setup the BSP's
854  * globaldata structure.
855  */
856 void
857 lwkt_token_pool_init(void)
858 {
859 	int i;
860 
861 	for (i = 0; i < LWKT_NUM_POOL_TOKENS; ++i)
862 		lwkt_token_init(&pool_tokens[i], "pool");
863 }
864 
865 lwkt_token_t
866 lwkt_token_pool_lookup(void *ptr)
867 {
868 	return (_lwkt_token_pool_lookup(ptr));
869 }
870 
871 /*
872  * Initialize a token.
873  */
874 void
875 lwkt_token_init(lwkt_token_t tok, const char *desc)
876 {
877 	tok->t_count = 0;
878 	tok->t_ref = NULL;
879 	tok->t_collisions = 0;
880 	tok->t_desc = desc;
881 }
882 
883 void
884 lwkt_token_uninit(lwkt_token_t tok)
885 {
886 	/* empty */
887 }
888 
889 /*
890  * Exchange the two most recent tokens on the tokref stack.  This allows
891  * you to release a token out of order.
892  *
893  * We have to be careful about the case where the top two tokens are
894  * the same token.  In this case tok->t_ref will point to the deeper
895  * ref and must remain pointing to the deeper ref.  If we were to swap
896  * it the first release would clear the token even though a second
897  * ref is still present.
898  *
899  * Only exclusively held tokens contain a reference to the tokref which
900  * has to be flipped along with the swap.
901  */
902 void
903 lwkt_token_swap(void)
904 {
905 	lwkt_tokref_t ref1, ref2;
906 	lwkt_token_t tok1, tok2;
907 	long count1, count2;
908 	thread_t td = curthread;
909 
910 	crit_enter();
911 
912 	ref1 = td->td_toks_stop - 1;
913 	ref2 = td->td_toks_stop - 2;
914 	KKASSERT(ref1 >= &td->td_toks_base);
915 	KKASSERT(ref2 >= &td->td_toks_base);
916 
917 	tok1 = ref1->tr_tok;
918 	tok2 = ref2->tr_tok;
919 	count1 = ref1->tr_count;
920 	count2 = ref2->tr_count;
921 
922 	if (tok1 != tok2) {
923 		ref1->tr_tok = tok2;
924 		ref1->tr_count = count2;
925 		ref2->tr_tok = tok1;
926 		ref2->tr_count = count1;
927 		if (tok1->t_ref == ref1)
928 			tok1->t_ref = ref2;
929 		if (tok2->t_ref == ref2)
930 			tok2->t_ref = ref1;
931 	}
932 
933 	crit_exit();
934 }
935 
936 DB_SHOW_COMMAND(tokens, db_tok_all)
937 {
938 	struct lwkt_token *tok, **ptr;
939 	struct lwkt_token *toklist[16] = {
940 		&mp_token,
941 		&pmap_token,
942 		&dev_token,
943 		&vm_token,
944 		&vmspace_token,
945 		&kvm_token,
946 		&proc_token,
947 		&tty_token,
948 		&vnode_token,
949 		&vmobj_token,
950 		NULL
951 	};
952 
953 	ptr = toklist;
954 	for (tok = *ptr; tok; tok = *(++ptr)) {
955 		db_printf("tok=%p tr_owner=%p t_colissions=%ld t_desc=%s\n", tok,
956 		    (tok->t_ref ? tok->t_ref->tr_owner : NULL),
957 		    tok->t_collisions, tok->t_desc);
958 	}
959 }
960