xref: /dflybsd-src/sys/kern/lwkt_token.c (revision c666f28aa7c82e205ee3709528b79a41e8cc5308)
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 <sys/kthread.h>
57 #include <machine/cpu.h>
58 #include <sys/lock.h>
59 #include <sys/caps.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 
87 #define TOKEN_STRING	"REF=%p TOK=%p TD=%p"
88 #define TOKEN_ARGS	lwkt_tokref_t ref, lwkt_token_t tok, struct thread *td
89 #define CONTENDED_STRING	TOKEN_STRING " (contention started)"
90 #define UNCONTENDED_STRING	TOKEN_STRING " (contention stopped)"
91 #if !defined(KTR_TOKENS)
92 #define	KTR_TOKENS	KTR_ALL
93 #endif
94 
95 KTR_INFO_MASTER(tokens);
96 KTR_INFO(KTR_TOKENS, tokens, fail, 0, TOKEN_STRING, TOKEN_ARGS);
97 KTR_INFO(KTR_TOKENS, tokens, succ, 1, TOKEN_STRING, TOKEN_ARGS);
98 #if 0
99 KTR_INFO(KTR_TOKENS, tokens, release, 2, TOKEN_STRING, TOKEN_ARGS);
100 KTR_INFO(KTR_TOKENS, tokens, remote, 3, TOKEN_STRING, TOKEN_ARGS);
101 KTR_INFO(KTR_TOKENS, tokens, reqremote, 4, TOKEN_STRING, TOKEN_ARGS);
102 KTR_INFO(KTR_TOKENS, tokens, reqfail, 5, TOKEN_STRING, TOKEN_ARGS);
103 KTR_INFO(KTR_TOKENS, tokens, drain, 6, TOKEN_STRING, TOKEN_ARGS);
104 KTR_INFO(KTR_TOKENS, tokens, contention_start, 7, CONTENDED_STRING, TOKEN_ARGS);
105 KTR_INFO(KTR_TOKENS, tokens, contention_stop, 7, UNCONTENDED_STRING, TOKEN_ARGS);
106 #endif
107 
108 #define logtoken(name, ref)						\
109 	KTR_LOG(tokens_ ## name, ref, ref->tr_tok, curthread)
110 
111 /*
112  * Global tokens.  These replace the MP lock for major subsystem locking.
113  * These tokens are initially used to lockup both global and individual
114  * operations.
115  *
116  * Once individual structures get their own locks these tokens are used
117  * only to protect global lists & other variables and to interlock
118  * allocations and teardowns and such.
119  *
120  * The UP initializer causes token acquisition to also acquire the MP lock
121  * for maximum compatibility.  The feature may be enabled and disabled at
122  * any time, the MP state is copied to the tokref when the token is acquired
123  * and will not race against sysctl changes.
124  */
125 struct lwkt_token mp_token = LWKT_TOKEN_INITIALIZER(mp_token);
126 struct lwkt_token pmap_token = LWKT_TOKEN_INITIALIZER(pmap_token);
127 struct lwkt_token dev_token = LWKT_TOKEN_INITIALIZER(dev_token);
128 struct lwkt_token vm_token = LWKT_TOKEN_INITIALIZER(vm_token);
129 struct lwkt_token vmspace_token = LWKT_TOKEN_INITIALIZER(vmspace_token);
130 struct lwkt_token kvm_token = LWKT_TOKEN_INITIALIZER(kvm_token);
131 struct lwkt_token proc_token = LWKT_TOKEN_INITIALIZER(proc_token);
132 struct lwkt_token tty_token = LWKT_TOKEN_INITIALIZER(tty_token);
133 struct lwkt_token vnode_token = LWKT_TOKEN_INITIALIZER(vnode_token);
134 struct lwkt_token vmobj_token = LWKT_TOKEN_INITIALIZER(vmobj_token);
135 
136 static int lwkt_token_spin = 5;
137 SYSCTL_INT(_lwkt, OID_AUTO, token_spin, CTLFLAG_RW,
138     &lwkt_token_spin, 0, "Decontention spin loops");
139 static int lwkt_token_delay = 0;
140 SYSCTL_INT(_lwkt, OID_AUTO, token_delay, CTLFLAG_RW,
141     &lwkt_token_delay, 0, "Decontention spin delay in ns");
142 
143 /*
144  * The collision count is bumped every time the LWKT scheduler fails
145  * to acquire needed tokens in addition to a normal lwkt_gettoken()
146  * stall.
147  */
148 SYSCTL_LONG(_lwkt, OID_AUTO, mp_collisions, CTLFLAG_RW,
149     &mp_token.t_collisions, 0, "Collision counter of mp_token");
150 SYSCTL_LONG(_lwkt, OID_AUTO, pmap_collisions, CTLFLAG_RW,
151     &pmap_token.t_collisions, 0, "Collision counter of pmap_token");
152 SYSCTL_LONG(_lwkt, OID_AUTO, dev_collisions, CTLFLAG_RW,
153     &dev_token.t_collisions, 0, "Collision counter of dev_token");
154 SYSCTL_LONG(_lwkt, OID_AUTO, vm_collisions, CTLFLAG_RW,
155     &vm_token.t_collisions, 0, "Collision counter of vm_token");
156 SYSCTL_LONG(_lwkt, OID_AUTO, vmspace_collisions, CTLFLAG_RW,
157     &vmspace_token.t_collisions, 0, "Collision counter of vmspace_token");
158 SYSCTL_LONG(_lwkt, OID_AUTO, kvm_collisions, CTLFLAG_RW,
159     &kvm_token.t_collisions, 0, "Collision counter of kvm_token");
160 SYSCTL_LONG(_lwkt, OID_AUTO, proc_collisions, CTLFLAG_RW,
161     &proc_token.t_collisions, 0, "Collision counter of proc_token");
162 SYSCTL_LONG(_lwkt, OID_AUTO, tty_collisions, CTLFLAG_RW,
163     &tty_token.t_collisions, 0, "Collision counter of tty_token");
164 SYSCTL_LONG(_lwkt, OID_AUTO, vnode_collisions, CTLFLAG_RW,
165     &vnode_token.t_collisions, 0, "Collision counter of vnode_token");
166 
167 #ifdef DEBUG_LOCKS_LATENCY
168 
169 static long tokens_add_latency;
170 SYSCTL_LONG(_debug, OID_AUTO, tokens_add_latency, CTLFLAG_RW,
171 	    &tokens_add_latency, 0,
172 	    "Add spinlock latency");
173 
174 #endif
175 
176 
177 static int _lwkt_getalltokens_sorted(thread_t td);
178 
179 #ifdef SMP
180 /*
181  * Acquire the initial mplock
182  *
183  * (low level boot only)
184  */
185 void
186 cpu_get_initial_mplock(void)
187 {
188 	KKASSERT(mp_token.t_ref == NULL);
189 	if (lwkt_trytoken(&mp_token) == FALSE)
190 		panic("cpu_get_initial_mplock");
191 }
192 #endif
193 
194 /*
195  * Return a pool token given an address.  Use a prime number to reduce
196  * overlaps.
197  */
198 static __inline
199 lwkt_token_t
200 _lwkt_token_pool_lookup(void *ptr)
201 {
202 	u_int i;
203 
204 	i = (u_int)(uintptr_t)ptr % LWKT_NUM_POOL_TOKENS;
205 	return(&pool_tokens[i]);
206 }
207 
208 /*
209  * Initialize a tokref_t prior to making it visible in the thread's
210  * token array.
211  */
212 static __inline
213 void
214 _lwkt_tokref_init(lwkt_tokref_t ref, lwkt_token_t tok, thread_t td, long excl)
215 {
216 	ref->tr_tok = tok;
217 	ref->tr_count = excl;
218 	ref->tr_owner = td;
219 }
220 
221 /*
222  * Attempt to acquire a shared or exclusive token.  Returns TRUE on success,
223  * FALSE on failure.
224  *
225  * If TOK_EXCLUSIVE is set in mode we are attempting to get an exclusive
226  * token, otherwise are attempting to get a shared token.
227  *
228  * If TOK_EXCLREQ is set in mode this is a blocking operation, otherwise
229  * it is a non-blocking operation (for both exclusive or shared acquisions).
230  */
231 static __inline
232 int
233 _lwkt_trytokref(lwkt_tokref_t ref, thread_t td, long mode)
234 {
235 	lwkt_token_t tok;
236 	lwkt_tokref_t oref;
237 	long count;
238 
239 	tok = ref->tr_tok;
240 	KASSERT(((mode & TOK_EXCLREQ) == 0 ||	/* non blocking */
241 		td->td_gd->gd_intr_nesting_level == 0 ||
242 		panic_cpu_gd == mycpu),
243 		("Attempt to acquire token %p not already "
244 		"held in hard code section", tok));
245 
246 	if (mode & TOK_EXCLUSIVE) {
247 		/*
248 		 * Attempt to get an exclusive token
249 		 */
250 		for (;;) {
251 			count = tok->t_count;
252 			oref = tok->t_ref;	/* can be NULL */
253 			cpu_ccfence();
254 			if ((count & ~TOK_EXCLREQ) == 0) {
255 				/*
256 				 * It is possible to get the exclusive bit.
257 				 * We must clear TOK_EXCLREQ on successful
258 				 * acquisition.
259 				 */
260 				if (atomic_cmpset_long(&tok->t_count, count,
261 						       (count & ~TOK_EXCLREQ) |
262 						       TOK_EXCLUSIVE)) {
263 					KKASSERT(tok->t_ref == NULL);
264 					tok->t_ref = ref;
265 					return TRUE;
266 				}
267 				/* retry */
268 			} else if ((count & TOK_EXCLUSIVE) &&
269 				   oref >= &td->td_toks_base &&
270 				   oref < td->td_toks_stop) {
271 				/*
272 				 * Our thread already holds the exclusive
273 				 * bit, we treat this tokref as a shared
274 				 * token (sorta) to make the token release
275 				 * code easier.
276 				 *
277 				 * NOTE: oref cannot race above if it
278 				 *	 happens to be ours, so we're good.
279 				 *	 But we must still have a stable
280 				 *	 variable for both parts of the
281 				 *	 comparison.
282 				 *
283 				 * NOTE: Since we already have an exclusive
284 				 *	 lock and don't need to check EXCLREQ
285 				 *	 we can just use an atomic_add here
286 				 */
287 				atomic_add_long(&tok->t_count, TOK_INCR);
288 				ref->tr_count &= ~TOK_EXCLUSIVE;
289 				return TRUE;
290 			} else if ((mode & TOK_EXCLREQ) &&
291 				   (count & TOK_EXCLREQ) == 0) {
292 				/*
293 				 * Unable to get the exclusive bit but being
294 				 * asked to set the exclusive-request bit.
295 				 * Since we are going to retry anyway just
296 				 * set the bit unconditionally.
297 				 */
298 				atomic_set_long(&tok->t_count, TOK_EXCLREQ);
299 				return FALSE;
300 			} else {
301 				/*
302 				 * Unable to get the exclusive bit and not
303 				 * being asked to set the exclusive-request
304 				 * (aka lwkt_trytoken()), or EXCLREQ was
305 				 * already set.
306 				 */
307 				cpu_pause();
308 				return FALSE;
309 			}
310 			/* retry */
311 		}
312 	} else {
313 		/*
314 		 * Attempt to get a shared token.  Note that TOK_EXCLREQ
315 		 * for shared tokens simply means the caller intends to
316 		 * block.  We never actually set the bit in tok->t_count.
317 		 */
318 		for (;;) {
319 			count = tok->t_count;
320 			oref = tok->t_ref;	/* can be NULL */
321 			cpu_ccfence();
322 			if ((count & (TOK_EXCLUSIVE/*|TOK_EXCLREQ*/)) == 0) {
323 				/* XXX EXCLREQ should work */
324 				/*
325 				 * It is possible to get the token shared.
326 				 */
327 				if (atomic_cmpset_long(&tok->t_count, count,
328 						       count + TOK_INCR)) {
329 					return TRUE;
330 				}
331 				/* retry */
332 			} else if ((count & TOK_EXCLUSIVE) &&
333 				   oref >= &td->td_toks_base &&
334 				   oref < td->td_toks_stop) {
335 				/*
336 				 * We own the exclusive bit on the token so
337 				 * we can in fact also get it shared.
338 				 */
339 				atomic_add_long(&tok->t_count, TOK_INCR);
340 				return TRUE;
341 			} else {
342 				/*
343 				 * We failed to get the token shared
344 				 */
345 				return FALSE;
346 			}
347 			/* retry */
348 		}
349 	}
350 }
351 
352 static __inline
353 int
354 _lwkt_trytokref_spin(lwkt_tokref_t ref, thread_t td, long mode)
355 {
356 	int spin;
357 
358 	if (_lwkt_trytokref(ref, td, mode)) {
359 #ifdef DEBUG_LOCKS_LATENCY
360 		long j;
361 		for (j = tokens_add_latency; j > 0; --j)
362 			cpu_ccfence();
363 #endif
364 		return TRUE;
365 	}
366 	for (spin = lwkt_token_spin; spin > 0; --spin) {
367 		if (lwkt_token_delay)
368 			tsc_delay(lwkt_token_delay);
369 		else
370 			cpu_pause();
371 		if (_lwkt_trytokref(ref, td, mode)) {
372 #ifdef DEBUG_LOCKS_LATENCY
373 			long j;
374 			for (j = tokens_add_latency; j > 0; --j)
375 				cpu_ccfence();
376 #endif
377 			return TRUE;
378 		}
379 	}
380 	return FALSE;
381 }
382 
383 /*
384  * Release a token that we hold.
385  */
386 static __inline
387 void
388 _lwkt_reltokref(lwkt_tokref_t ref, thread_t td)
389 {
390 	lwkt_token_t tok;
391 	long count;
392 
393 	tok = ref->tr_tok;
394 	for (;;) {
395 		count = tok->t_count;
396 		cpu_ccfence();
397 		if (tok->t_ref == ref) {
398 			/*
399 			 * We are an exclusive holder.  We must clear tr_ref
400 			 * before we clear the TOK_EXCLUSIVE bit.  If we are
401 			 * unable to clear the bit we must restore
402 			 * tok->t_ref.
403 			 */
404 			KKASSERT(count & TOK_EXCLUSIVE);
405 			tok->t_ref = NULL;
406 			if (atomic_cmpset_long(&tok->t_count, count,
407 					       count & ~TOK_EXCLUSIVE)) {
408 				return;
409 			}
410 			tok->t_ref = ref;
411 			/* retry */
412 		} else {
413 			/*
414 			 * We are a shared holder
415 			 */
416 			KKASSERT(count & TOK_COUNTMASK);
417 			if (atomic_cmpset_long(&tok->t_count, count,
418 					       count - TOK_INCR)) {
419 				return;
420 			}
421 			/* retry */
422 		}
423 		/* retry */
424 	}
425 }
426 
427 /*
428  * Obtain all the tokens required by the specified thread on the current
429  * cpu, return 0 on failure and non-zero on success.  If a failure occurs
430  * any partially acquired tokens will be released prior to return.
431  *
432  * lwkt_getalltokens is called by the LWKT scheduler to re-acquire all
433  * tokens that the thread had to release when it switched away.
434  *
435  * If spinning is non-zero this function acquires the tokens in a particular
436  * order to deal with potential deadlocks.  We simply use address order for
437  * the case.
438  *
439  * Called from a critical section.
440  */
441 int
442 lwkt_getalltokens(thread_t td, int spinning)
443 {
444 	lwkt_tokref_t scan;
445 	lwkt_token_t tok;
446 
447 	if (spinning)
448 		return(_lwkt_getalltokens_sorted(td));
449 
450 	/*
451 	 * Acquire tokens in forward order, assign or validate tok->t_ref.
452 	 */
453 	for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) {
454 		tok = scan->tr_tok;
455 		for (;;) {
456 			/*
457 			 * Only try really hard on the last token
458 			 */
459 			if (scan == td->td_toks_stop - 1) {
460 			    if (_lwkt_trytokref_spin(scan, td, scan->tr_count))
461 				    break;
462 			} else {
463 			    if (_lwkt_trytokref(scan, td, scan->tr_count))
464 				    break;
465 			}
466 
467 			/*
468 			 * Otherwise we failed to acquire all the tokens.
469 			 * Release whatever we did get.
470 			 */
471 			KASSERT(tok->t_desc, ("token %p is not initialized", tok));
472 			strncpy(td->td_gd->gd_cnt.v_token_name,
473 				tok->t_desc,
474 				sizeof(td->td_gd->gd_cnt.v_token_name) - 1);
475 
476 			if (lwkt_sched_debug > 0) {
477 				--lwkt_sched_debug;
478 				kprintf("toka %p %s %s\n",
479 					tok, tok->t_desc, td->td_comm);
480 			}
481 			td->td_wmesg = tok->t_desc;
482 			++tok->t_collisions;
483 			while (--scan >= &td->td_toks_base)
484 				_lwkt_reltokref(scan, td);
485 			return(FALSE);
486 		}
487 	}
488 	return (TRUE);
489 }
490 
491 /*
492  * Release all tokens owned by the specified thread on the current cpu.
493  *
494  * This code is really simple.  Even in cases where we own all the tokens
495  * note that t_ref may not match the scan for recursively held tokens which
496  * are held deeper in the stack, or for the case where a lwkt_getalltokens()
497  * failed.
498  *
499  * Tokens are released in reverse order to reduce chasing race failures.
500  *
501  * Called from a critical section.
502  */
503 void
504 lwkt_relalltokens(thread_t td)
505 {
506 	lwkt_tokref_t scan;
507 
508 	/*
509 	 * Weird order is to try to avoid a panic loop
510 	 */
511 	if (td->td_toks_have) {
512 		scan = td->td_toks_have;
513 		td->td_toks_have = NULL;
514 	} else {
515 		scan = td->td_toks_stop;
516 	}
517 	while (--scan >= &td->td_toks_base)
518 		_lwkt_reltokref(scan, td);
519 }
520 
521 /*
522  * This is the decontention version of lwkt_getalltokens().  The tokens are
523  * acquired in address-sorted order to deal with any deadlocks.  Ultimately
524  * token failures will spin into the scheduler and get here.
525  *
526  * Called from critical section
527  */
528 static
529 int
530 _lwkt_getalltokens_sorted(thread_t td)
531 {
532 	lwkt_tokref_t sort_array[LWKT_MAXTOKENS];
533 	lwkt_tokref_t scan;
534 	lwkt_token_t tok;
535 	int i;
536 	int j;
537 	int n;
538 
539 	/*
540 	 * Sort the token array.  Yah yah, I know this isn't fun.
541 	 *
542 	 * NOTE: Recursively acquired tokens are ordered the same as in the
543 	 *	 td_toks_array so we can always get the earliest one first.
544 	 */
545 	i = 0;
546 	scan = &td->td_toks_base;
547 	while (scan < td->td_toks_stop) {
548 		for (j = 0; j < i; ++j) {
549 			if (scan->tr_tok < sort_array[j]->tr_tok)
550 				break;
551 		}
552 		if (j != i) {
553 			bcopy(sort_array + j, sort_array + j + 1,
554 			      (i - j) * sizeof(lwkt_tokref_t));
555 		}
556 		sort_array[j] = scan;
557 		++scan;
558 		++i;
559 	}
560 	n = i;
561 
562 	/*
563 	 * Acquire tokens in forward order, assign or validate tok->t_ref.
564 	 */
565 	for (i = 0; i < n; ++i) {
566 		scan = sort_array[i];
567 		tok = scan->tr_tok;
568 		for (;;) {
569 			/*
570 			 * Only try really hard on the last token
571 			 */
572 			if (scan == td->td_toks_stop - 1) {
573 			    if (_lwkt_trytokref_spin(scan, td, scan->tr_count))
574 				    break;
575 			} else {
576 			    if (_lwkt_trytokref(scan, td, scan->tr_count))
577 				    break;
578 			}
579 
580 			/*
581 			 * Otherwise we failed to acquire all the tokens.
582 			 * Release whatever we did get.
583 			 */
584 			if (lwkt_sched_debug > 0) {
585 				--lwkt_sched_debug;
586 				kprintf("tokb %p %s %s\n",
587 					tok, tok->t_desc, td->td_comm);
588 			}
589 			td->td_wmesg = tok->t_desc;
590 			++tok->t_collisions;
591 			while (--i >= 0) {
592 				scan = sort_array[i];
593 				_lwkt_reltokref(scan, td);
594 			}
595 			return(FALSE);
596 		}
597 	}
598 
599 	/*
600 	 * We were successful, there is no need for another core to signal
601 	 * us.
602 	 */
603 	return (TRUE);
604 }
605 
606 /*
607  * Get a serializing token.  This routine can block.
608  */
609 void
610 lwkt_gettoken(lwkt_token_t tok)
611 {
612 	thread_t td = curthread;
613 	lwkt_tokref_t ref;
614 
615 	ref = td->td_toks_stop;
616 	KKASSERT(ref < &td->td_toks_end);
617 	++td->td_toks_stop;
618 	cpu_ccfence();
619 	_lwkt_tokref_init(ref, tok, td, TOK_EXCLUSIVE|TOK_EXCLREQ);
620 
621 #ifdef DEBUG_LOCKS
622 	/*
623 	 * Taking an exclusive token after holding it shared will
624 	 * livelock. Scan for that case and assert.
625 	 */
626 	lwkt_tokref_t tk;
627 	int found = 0;
628 	for (tk = &td->td_toks_base; tk < ref; tk++) {
629 		if (tk->tr_tok != tok)
630 			continue;
631 
632 		found++;
633 		if (tk->tr_count & TOK_EXCLUSIVE)
634 			goto good;
635 	}
636 	/* We found only shared instances of this token if found >0 here */
637 	KASSERT((found == 0), ("Token %p s/x livelock", tok));
638 good:
639 #endif
640 
641 	if (_lwkt_trytokref_spin(ref, td, TOK_EXCLUSIVE|TOK_EXCLREQ))
642 		return;
643 
644 	/*
645 	 * Give up running if we can't acquire the token right now.
646 	 *
647 	 * Since the tokref is already active the scheduler now
648 	 * takes care of acquisition, so we need only call
649 	 * lwkt_switch().
650 	 *
651 	 * Since we failed this was not a recursive token so upon
652 	 * return tr_tok->t_ref should be assigned to this specific
653 	 * ref.
654 	 */
655 	td->td_wmesg = tok->t_desc;
656 	++tok->t_collisions;
657 	logtoken(fail, ref);
658 	td->td_toks_have = td->td_toks_stop - 1;
659 	lwkt_switch();
660 	logtoken(succ, ref);
661 	KKASSERT(tok->t_ref == ref);
662 }
663 
664 /*
665  * Similar to gettoken but we acquire a shared token instead of an exclusive
666  * token.
667  */
668 void
669 lwkt_gettoken_shared(lwkt_token_t tok)
670 {
671 	thread_t td = curthread;
672 	lwkt_tokref_t ref;
673 
674 	ref = td->td_toks_stop;
675 	KKASSERT(ref < &td->td_toks_end);
676 	++td->td_toks_stop;
677 	cpu_ccfence();
678 	_lwkt_tokref_init(ref, tok, td, TOK_EXCLREQ);
679 
680 #ifdef DEBUG_LOCKS
681         /*
682          * Taking a pool token in shared mode is a bad idea; other
683          * addresses deeper in the call stack may hash to the same pool
684          * token and you may end up with an exclusive-shared livelock.
685          * Warn in this condition.
686          */
687         if ((tok >= &pool_tokens[0]) &&
688             (tok < &pool_tokens[LWKT_NUM_POOL_TOKENS]))
689                 kprintf("Warning! Taking pool token %p in shared mode\n", tok);
690 #endif
691 
692 
693 	if (_lwkt_trytokref_spin(ref, td, TOK_EXCLREQ))
694 		return;
695 
696 	/*
697 	 * Give up running if we can't acquire the token right now.
698 	 *
699 	 * Since the tokref is already active the scheduler now
700 	 * takes care of acquisition, so we need only call
701 	 * lwkt_switch().
702 	 *
703 	 * Since we failed this was not a recursive token so upon
704 	 * return tr_tok->t_ref should be assigned to this specific
705 	 * ref.
706 	 */
707 	td->td_wmesg = tok->t_desc;
708 	++tok->t_collisions;
709 	logtoken(fail, ref);
710 	td->td_toks_have = td->td_toks_stop - 1;
711 	lwkt_switch();
712 	logtoken(succ, ref);
713 }
714 
715 /*
716  * Attempt to acquire a token, return TRUE on success, FALSE on failure.
717  *
718  * We setup the tokref in case we actually get the token (if we switch later
719  * it becomes mandatory so we set TOK_EXCLREQ), but we call trytokref without
720  * TOK_EXCLREQ in case we fail.
721  */
722 int
723 lwkt_trytoken(lwkt_token_t tok)
724 {
725 	thread_t td = curthread;
726 	lwkt_tokref_t ref;
727 
728 	ref = td->td_toks_stop;
729 	KKASSERT(ref < &td->td_toks_end);
730 	++td->td_toks_stop;
731 	cpu_ccfence();
732 	_lwkt_tokref_init(ref, tok, td, TOK_EXCLUSIVE|TOK_EXCLREQ);
733 
734 	if (_lwkt_trytokref(ref, td, TOK_EXCLUSIVE))
735 		return TRUE;
736 
737 	/*
738 	 * Failed, unpend the request
739 	 */
740 	cpu_ccfence();
741 	--td->td_toks_stop;
742 	++tok->t_collisions;
743 	return FALSE;
744 }
745 
746 
747 void
748 lwkt_gettoken_hard(lwkt_token_t tok)
749 {
750 	lwkt_gettoken(tok);
751 	crit_enter_hard();
752 }
753 
754 lwkt_token_t
755 lwkt_getpooltoken(void *ptr)
756 {
757 	lwkt_token_t tok;
758 
759 	tok = _lwkt_token_pool_lookup(ptr);
760 	lwkt_gettoken(tok);
761 	return (tok);
762 }
763 
764 /*
765  * Release a serializing token.
766  *
767  * WARNING!  All tokens must be released in reverse order.  This will be
768  *	     asserted.
769  */
770 void
771 lwkt_reltoken(lwkt_token_t tok)
772 {
773 	thread_t td = curthread;
774 	lwkt_tokref_t ref;
775 
776 	/*
777 	 * Remove ref from thread token list and assert that it matches
778 	 * the token passed in.  Tokens must be released in reverse order.
779 	 */
780 	ref = td->td_toks_stop - 1;
781 	KKASSERT(ref >= &td->td_toks_base && ref->tr_tok == tok);
782 	_lwkt_reltokref(ref, td);
783 	cpu_sfence();
784 	td->td_toks_stop = ref;
785 }
786 
787 void
788 lwkt_reltoken_hard(lwkt_token_t tok)
789 {
790 	lwkt_reltoken(tok);
791 	crit_exit_hard();
792 }
793 
794 /*
795  * It is faster for users of lwkt_getpooltoken() to use the returned
796  * token and just call lwkt_reltoken(), but for convenience we provide
797  * this function which looks the token up based on the ident.
798  */
799 void
800 lwkt_relpooltoken(void *ptr)
801 {
802 	lwkt_token_t tok = _lwkt_token_pool_lookup(ptr);
803 	lwkt_reltoken(tok);
804 }
805 
806 /*
807  * Return a count of the number of token refs the thread has to the
808  * specified token, whether it currently owns the token or not.
809  */
810 int
811 lwkt_cnttoken(lwkt_token_t tok, thread_t td)
812 {
813 	lwkt_tokref_t scan;
814 	int count = 0;
815 
816 	for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) {
817 		if (scan->tr_tok == tok)
818 			++count;
819 	}
820 	return(count);
821 }
822 
823 /*
824  * Pool tokens are used to provide a type-stable serializing token
825  * pointer that does not race against disappearing data structures.
826  *
827  * This routine is called in early boot just after we setup the BSP's
828  * globaldata structure.
829  */
830 void
831 lwkt_token_pool_init(void)
832 {
833 	int i;
834 
835 	for (i = 0; i < LWKT_NUM_POOL_TOKENS; ++i)
836 		lwkt_token_init(&pool_tokens[i], "pool");
837 }
838 
839 lwkt_token_t
840 lwkt_token_pool_lookup(void *ptr)
841 {
842 	return (_lwkt_token_pool_lookup(ptr));
843 }
844 
845 /*
846  * Initialize a token.
847  */
848 void
849 lwkt_token_init(lwkt_token_t tok, const char *desc)
850 {
851 	tok->t_count = 0;
852 	tok->t_ref = NULL;
853 	tok->t_collisions = 0;
854 	tok->t_desc = desc;
855 }
856 
857 void
858 lwkt_token_uninit(lwkt_token_t tok)
859 {
860 	/* empty */
861 }
862 
863 /*
864  * Exchange the two most recent tokens on the tokref stack.  This allows
865  * you to release a token out of order.
866  *
867  * We have to be careful about the case where the top two tokens are
868  * the same token.  In this case tok->t_ref will point to the deeper
869  * ref and must remain pointing to the deeper ref.  If we were to swap
870  * it the first release would clear the token even though a second
871  * ref is still present.
872  *
873  * Only exclusively held tokens contain a reference to the tokref which
874  * has to be flipped along with the swap.
875  */
876 void
877 lwkt_token_swap(void)
878 {
879 	lwkt_tokref_t ref1, ref2;
880 	lwkt_token_t tok1, tok2;
881 	long count1, count2;
882 	thread_t td = curthread;
883 
884 	crit_enter();
885 
886 	ref1 = td->td_toks_stop - 1;
887 	ref2 = td->td_toks_stop - 2;
888 	KKASSERT(ref1 >= &td->td_toks_base);
889 	KKASSERT(ref2 >= &td->td_toks_base);
890 
891 	tok1 = ref1->tr_tok;
892 	tok2 = ref2->tr_tok;
893 	count1 = ref1->tr_count;
894 	count2 = ref2->tr_count;
895 
896 	if (tok1 != tok2) {
897 		ref1->tr_tok = tok2;
898 		ref1->tr_count = count2;
899 		ref2->tr_tok = tok1;
900 		ref2->tr_count = count1;
901 		if (tok1->t_ref == ref1)
902 			tok1->t_ref = ref2;
903 		if (tok2->t_ref == ref2)
904 			tok2->t_ref = ref1;
905 	}
906 
907 	crit_exit();
908 }
909