xref: /dragonfly/sys/kern/lwkt_token.c (revision 279dd846)
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/spinlock.h>
60 
61 #include <sys/thread2.h>
62 #include <sys/spinlock2.h>
63 #include <sys/mplock2.h>
64 
65 #include <vm/vm.h>
66 #include <vm/vm_param.h>
67 #include <vm/vm_kern.h>
68 #include <vm/vm_object.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_map.h>
71 #include <vm/vm_pager.h>
72 #include <vm/vm_extern.h>
73 #include <vm/vm_zone.h>
74 
75 #include <machine/stdarg.h>
76 #include <machine/smp.h>
77 
78 #include "opt_ddb.h"
79 #ifdef DDB
80 #include <ddb/ddb.h>
81 #endif
82 
83 extern int lwkt_sched_debug;
84 
85 #ifndef LWKT_NUM_POOL_TOKENS
86 #define LWKT_NUM_POOL_TOKENS	4001	/* prime number */
87 #endif
88 
89 static lwkt_token	pool_tokens[LWKT_NUM_POOL_TOKENS];
90 struct spinlock		tok_debug_spin = SPINLOCK_INITIALIZER(&tok_debug_spin, "tok_debug_spin");
91 
92 #define TOKEN_STRING	"REF=%p TOK=%p TD=%p"
93 #define TOKEN_ARGS	lwkt_tokref_t ref, lwkt_token_t tok, struct thread *td
94 #define CONTENDED_STRING	TOKEN_STRING " (contention started)"
95 #define UNCONTENDED_STRING	TOKEN_STRING " (contention stopped)"
96 #if !defined(KTR_TOKENS)
97 #define	KTR_TOKENS	KTR_ALL
98 #endif
99 
100 KTR_INFO_MASTER(tokens);
101 KTR_INFO(KTR_TOKENS, tokens, fail, 0, TOKEN_STRING, TOKEN_ARGS);
102 KTR_INFO(KTR_TOKENS, tokens, succ, 1, TOKEN_STRING, TOKEN_ARGS);
103 #if 0
104 KTR_INFO(KTR_TOKENS, tokens, release, 2, TOKEN_STRING, TOKEN_ARGS);
105 KTR_INFO(KTR_TOKENS, tokens, remote, 3, TOKEN_STRING, TOKEN_ARGS);
106 KTR_INFO(KTR_TOKENS, tokens, reqremote, 4, TOKEN_STRING, TOKEN_ARGS);
107 KTR_INFO(KTR_TOKENS, tokens, reqfail, 5, TOKEN_STRING, TOKEN_ARGS);
108 KTR_INFO(KTR_TOKENS, tokens, drain, 6, TOKEN_STRING, TOKEN_ARGS);
109 KTR_INFO(KTR_TOKENS, tokens, contention_start, 7, CONTENDED_STRING, TOKEN_ARGS);
110 KTR_INFO(KTR_TOKENS, tokens, contention_stop, 7, UNCONTENDED_STRING, TOKEN_ARGS);
111 #endif
112 
113 #define logtoken(name, ref)						\
114 	KTR_LOG(tokens_ ## name, ref, ref->tr_tok, curthread)
115 
116 /*
117  * Global tokens.  These replace the MP lock for major subsystem locking.
118  * These tokens are initially used to lockup both global and individual
119  * operations.
120  *
121  * Once individual structures get their own locks these tokens are used
122  * only to protect global lists & other variables and to interlock
123  * allocations and teardowns and such.
124  *
125  * The UP initializer causes token acquisition to also acquire the MP lock
126  * for maximum compatibility.  The feature may be enabled and disabled at
127  * any time, the MP state is copied to the tokref when the token is acquired
128  * and will not race against sysctl changes.
129  */
130 struct lwkt_token mp_token = LWKT_TOKEN_INITIALIZER(mp_token);
131 struct lwkt_token pmap_token = LWKT_TOKEN_INITIALIZER(pmap_token);
132 struct lwkt_token dev_token = LWKT_TOKEN_INITIALIZER(dev_token);
133 struct lwkt_token vm_token = LWKT_TOKEN_INITIALIZER(vm_token);
134 struct lwkt_token vmspace_token = LWKT_TOKEN_INITIALIZER(vmspace_token);
135 struct lwkt_token kvm_token = LWKT_TOKEN_INITIALIZER(kvm_token);
136 struct lwkt_token sigio_token = LWKT_TOKEN_INITIALIZER(sigio_token);
137 struct lwkt_token tty_token = LWKT_TOKEN_INITIALIZER(tty_token);
138 struct lwkt_token vnode_token = LWKT_TOKEN_INITIALIZER(vnode_token);
139 
140 static int lwkt_token_spin = 5;
141 SYSCTL_INT(_lwkt, OID_AUTO, token_spin, CTLFLAG_RW,
142     &lwkt_token_spin, 0, "Decontention spin loops");
143 static int lwkt_token_delay = 0;
144 SYSCTL_INT(_lwkt, OID_AUTO, token_delay, CTLFLAG_RW,
145     &lwkt_token_delay, 0, "Decontention spin delay in ns");
146 
147 /*
148  * The collision count is bumped every time the LWKT scheduler fails
149  * to acquire needed tokens in addition to a normal lwkt_gettoken()
150  * stall.
151  */
152 SYSCTL_LONG(_lwkt, OID_AUTO, mp_collisions, CTLFLAG_RW,
153     &mp_token.t_collisions, 0, "Collision counter of mp_token");
154 SYSCTL_LONG(_lwkt, OID_AUTO, pmap_collisions, CTLFLAG_RW,
155     &pmap_token.t_collisions, 0, "Collision counter of pmap_token");
156 SYSCTL_LONG(_lwkt, OID_AUTO, dev_collisions, CTLFLAG_RW,
157     &dev_token.t_collisions, 0, "Collision counter of dev_token");
158 SYSCTL_LONG(_lwkt, OID_AUTO, vm_collisions, CTLFLAG_RW,
159     &vm_token.t_collisions, 0, "Collision counter of vm_token");
160 SYSCTL_LONG(_lwkt, OID_AUTO, vmspace_collisions, CTLFLAG_RW,
161     &vmspace_token.t_collisions, 0, "Collision counter of vmspace_token");
162 SYSCTL_LONG(_lwkt, OID_AUTO, kvm_collisions, CTLFLAG_RW,
163     &kvm_token.t_collisions, 0, "Collision counter of kvm_token");
164 SYSCTL_LONG(_lwkt, OID_AUTO, sigio_collisions, CTLFLAG_RW,
165     &sigio_token.t_collisions, 0, "Collision counter of sigio_token");
166 SYSCTL_LONG(_lwkt, OID_AUTO, tty_collisions, CTLFLAG_RW,
167     &tty_token.t_collisions, 0, "Collision counter of tty_token");
168 SYSCTL_LONG(_lwkt, OID_AUTO, vnode_collisions, CTLFLAG_RW,
169     &vnode_token.t_collisions, 0, "Collision counter of vnode_token");
170 
171 int tokens_debug_output;
172 SYSCTL_INT(_lwkt, OID_AUTO, tokens_debug_output, CTLFLAG_RW,
173     &tokens_debug_output, 0, "Generate stack trace N times");
174 
175 
176 #ifdef DEBUG_LOCKS_LATENCY
177 
178 static long tokens_add_latency;
179 SYSCTL_LONG(_debug, OID_AUTO, tokens_add_latency, CTLFLAG_RW,
180 	    &tokens_add_latency, 0,
181 	    "Add spinlock latency");
182 
183 #endif
184 
185 
186 static int _lwkt_getalltokens_sorted(thread_t td);
187 
188 /*
189  * Acquire the initial mplock
190  *
191  * (low level boot only)
192  */
193 void
194 cpu_get_initial_mplock(void)
195 {
196 	KKASSERT(mp_token.t_ref == NULL);
197 	if (lwkt_trytoken(&mp_token) == FALSE)
198 		panic("cpu_get_initial_mplock");
199 }
200 
201 /*
202  * Return a pool token given an address.  Use a prime number to reduce
203  * overlaps.
204  */
205 static __inline
206 lwkt_token_t
207 _lwkt_token_pool_lookup(void *ptr)
208 {
209 	u_int i;
210 
211 	i = (u_int)(uintptr_t)ptr % LWKT_NUM_POOL_TOKENS;
212 	return(&pool_tokens[i]);
213 }
214 
215 /*
216  * Initialize a tokref_t prior to making it visible in the thread's
217  * token array.
218  */
219 static __inline
220 void
221 _lwkt_tokref_init(lwkt_tokref_t ref, lwkt_token_t tok, thread_t td, long excl)
222 {
223 	ref->tr_tok = tok;
224 	ref->tr_count = excl;
225 	ref->tr_owner = td;
226 }
227 
228 /*
229  * Attempt to acquire a shared or exclusive token.  Returns TRUE on success,
230  * FALSE on failure.
231  *
232  * If TOK_EXCLUSIVE is set in mode we are attempting to get an exclusive
233  * token, otherwise are attempting to get a shared token.
234  *
235  * If TOK_EXCLREQ is set in mode this is a blocking operation, otherwise
236  * it is a non-blocking operation (for both exclusive or shared acquisions).
237  */
238 static __inline
239 int
240 _lwkt_trytokref(lwkt_tokref_t ref, thread_t td, long mode)
241 {
242 	lwkt_token_t tok;
243 	lwkt_tokref_t oref;
244 	long count;
245 
246 	tok = ref->tr_tok;
247 	KASSERT(((mode & TOK_EXCLREQ) == 0 ||	/* non blocking */
248 		td->td_gd->gd_intr_nesting_level == 0 ||
249 		panic_cpu_gd == mycpu),
250 		("Attempt to acquire token %p not already "
251 		"held in hard code section", tok));
252 
253 	if (mode & TOK_EXCLUSIVE) {
254 		/*
255 		 * Attempt to get an exclusive token
256 		 */
257 		for (;;) {
258 			count = tok->t_count;
259 			oref = tok->t_ref;	/* can be NULL */
260 			cpu_ccfence();
261 			if ((count & ~TOK_EXCLREQ) == 0) {
262 				/*
263 				 * It is possible to get the exclusive bit.
264 				 * We must clear TOK_EXCLREQ on successful
265 				 * acquisition.
266 				 */
267 				if (atomic_cmpset_long(&tok->t_count, count,
268 						       (count & ~TOK_EXCLREQ) |
269 						       TOK_EXCLUSIVE)) {
270 					KKASSERT(tok->t_ref == NULL);
271 					tok->t_ref = ref;
272 					return TRUE;
273 				}
274 				/* retry */
275 			} else if ((count & TOK_EXCLUSIVE) &&
276 				   oref >= &td->td_toks_base &&
277 				   oref < td->td_toks_stop) {
278 				/*
279 				 * Our thread already holds the exclusive
280 				 * bit, we treat this tokref as a shared
281 				 * token (sorta) to make the token release
282 				 * code easier.
283 				 *
284 				 * NOTE: oref cannot race above if it
285 				 *	 happens to be ours, so we're good.
286 				 *	 But we must still have a stable
287 				 *	 variable for both parts of the
288 				 *	 comparison.
289 				 *
290 				 * NOTE: Since we already have an exclusive
291 				 *	 lock and don't need to check EXCLREQ
292 				 *	 we can just use an atomic_add here
293 				 */
294 				atomic_add_long(&tok->t_count, TOK_INCR);
295 				ref->tr_count &= ~TOK_EXCLUSIVE;
296 				return TRUE;
297 			} else if ((mode & TOK_EXCLREQ) &&
298 				   (count & TOK_EXCLREQ) == 0) {
299 				/*
300 				 * Unable to get the exclusive bit but being
301 				 * asked to set the exclusive-request bit.
302 				 * Since we are going to retry anyway just
303 				 * set the bit unconditionally.
304 				 */
305 				atomic_set_long(&tok->t_count, TOK_EXCLREQ);
306 				return FALSE;
307 			} else {
308 				/*
309 				 * Unable to get the exclusive bit and not
310 				 * being asked to set the exclusive-request
311 				 * (aka lwkt_trytoken()), or EXCLREQ was
312 				 * already set.
313 				 */
314 				cpu_pause();
315 				return FALSE;
316 			}
317 			/* retry */
318 		}
319 	} else {
320 		/*
321 		 * Attempt to get a shared token.  Note that TOK_EXCLREQ
322 		 * for shared tokens simply means the caller intends to
323 		 * block.  We never actually set the bit in tok->t_count.
324 		 */
325 		for (;;) {
326 			count = tok->t_count;
327 			oref = tok->t_ref;	/* can be NULL */
328 			cpu_ccfence();
329 			if ((count & (TOK_EXCLUSIVE/*|TOK_EXCLREQ*/)) == 0) {
330 				/* XXX EXCLREQ should work */
331 				/*
332 				 * It is possible to get the token shared.
333 				 */
334 				if (atomic_cmpset_long(&tok->t_count, count,
335 						       count + TOK_INCR)) {
336 					return TRUE;
337 				}
338 				/* retry */
339 			} else if ((count & TOK_EXCLUSIVE) &&
340 				   oref >= &td->td_toks_base &&
341 				   oref < td->td_toks_stop) {
342 				/*
343 				 * We own the exclusive bit on the token so
344 				 * we can in fact also get it shared.
345 				 */
346 				atomic_add_long(&tok->t_count, TOK_INCR);
347 				return TRUE;
348 			} else {
349 				/*
350 				 * We failed to get the token shared
351 				 */
352 				return FALSE;
353 			}
354 			/* retry */
355 		}
356 	}
357 }
358 
359 static __inline
360 int
361 _lwkt_trytokref_spin(lwkt_tokref_t ref, thread_t td, long mode)
362 {
363 	int spin;
364 
365 	if (_lwkt_trytokref(ref, td, mode)) {
366 #ifdef DEBUG_LOCKS_LATENCY
367 		long j;
368 		for (j = tokens_add_latency; j > 0; --j)
369 			cpu_ccfence();
370 #endif
371 		return TRUE;
372 	}
373 	for (spin = lwkt_token_spin; spin > 0; --spin) {
374 		if (lwkt_token_delay)
375 			tsc_delay(lwkt_token_delay);
376 		else
377 			cpu_pause();
378 		if (_lwkt_trytokref(ref, td, mode)) {
379 #ifdef DEBUG_LOCKS_LATENCY
380 			long j;
381 			for (j = tokens_add_latency; j > 0; --j)
382 				cpu_ccfence();
383 #endif
384 			return TRUE;
385 		}
386 	}
387 	return FALSE;
388 }
389 
390 /*
391  * Release a token that we hold.
392  */
393 static __inline
394 void
395 _lwkt_reltokref(lwkt_tokref_t ref, thread_t td)
396 {
397 	lwkt_token_t tok;
398 	long count;
399 
400 	tok = ref->tr_tok;
401 	for (;;) {
402 		count = tok->t_count;
403 		cpu_ccfence();
404 		if (tok->t_ref == ref) {
405 			/*
406 			 * We are an exclusive holder.  We must clear tr_ref
407 			 * before we clear the TOK_EXCLUSIVE bit.  If we are
408 			 * unable to clear the bit we must restore
409 			 * tok->t_ref.
410 			 */
411 			KKASSERT(count & TOK_EXCLUSIVE);
412 			tok->t_ref = NULL;
413 			if (atomic_cmpset_long(&tok->t_count, count,
414 					       count & ~TOK_EXCLUSIVE)) {
415 				return;
416 			}
417 			tok->t_ref = ref;
418 			/* retry */
419 		} else {
420 			/*
421 			 * We are a shared holder
422 			 */
423 			KKASSERT(count & TOK_COUNTMASK);
424 			if (atomic_cmpset_long(&tok->t_count, count,
425 					       count - TOK_INCR)) {
426 				return;
427 			}
428 			/* retry */
429 		}
430 		/* retry */
431 	}
432 }
433 
434 /*
435  * Obtain all the tokens required by the specified thread on the current
436  * cpu, return 0 on failure and non-zero on success.  If a failure occurs
437  * any partially acquired tokens will be released prior to return.
438  *
439  * lwkt_getalltokens is called by the LWKT scheduler to re-acquire all
440  * tokens that the thread had to release when it switched away.
441  *
442  * If spinning is non-zero this function acquires the tokens in a particular
443  * order to deal with potential deadlocks.  We simply use address order for
444  * the case.
445  *
446  * Called from a critical section.
447  */
448 int
449 lwkt_getalltokens(thread_t td, int spinning)
450 {
451 	lwkt_tokref_t scan;
452 	lwkt_token_t tok;
453 
454 	if (spinning)
455 		return(_lwkt_getalltokens_sorted(td));
456 
457 	/*
458 	 * Acquire tokens in forward order, assign or validate tok->t_ref.
459 	 */
460 	for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) {
461 		tok = scan->tr_tok;
462 		for (;;) {
463 			/*
464 			 * Only try really hard on the last token
465 			 */
466 			if (scan == td->td_toks_stop - 1) {
467 			    if (_lwkt_trytokref_spin(scan, td, scan->tr_count))
468 				    break;
469 			} else {
470 			    if (_lwkt_trytokref(scan, td, scan->tr_count))
471 				    break;
472 			}
473 
474 			/*
475 			 * Otherwise we failed to acquire all the tokens.
476 			 * Release whatever we did get.
477 			 */
478 			KASSERT(tok->t_desc,
479 				("token %p is not initialized", tok));
480 			strncpy(td->td_gd->gd_cnt.v_lock_name,
481 				tok->t_desc,
482 				sizeof(td->td_gd->gd_cnt.v_lock_name) - 1);
483 
484 			if (lwkt_sched_debug > 0) {
485 				--lwkt_sched_debug;
486 				kprintf("toka %p %s %s\n",
487 					tok, tok->t_desc, td->td_comm);
488 			}
489 			td->td_wmesg = tok->t_desc;
490 			++tok->t_collisions;
491 			while (--scan >= &td->td_toks_base)
492 				_lwkt_reltokref(scan, td);
493 			return(FALSE);
494 		}
495 	}
496 	return (TRUE);
497 }
498 
499 /*
500  * Release all tokens owned by the specified thread on the current cpu.
501  *
502  * This code is really simple.  Even in cases where we own all the tokens
503  * note that t_ref may not match the scan for recursively held tokens which
504  * are held deeper in the stack, or for the case where a lwkt_getalltokens()
505  * failed.
506  *
507  * Tokens are released in reverse order to reduce chasing race failures.
508  *
509  * Called from a critical section.
510  */
511 void
512 lwkt_relalltokens(thread_t td)
513 {
514 	lwkt_tokref_t scan;
515 
516 	/*
517 	 * Weird order is to try to avoid a panic loop
518 	 */
519 	if (td->td_toks_have) {
520 		scan = td->td_toks_have;
521 		td->td_toks_have = NULL;
522 	} else {
523 		scan = td->td_toks_stop;
524 	}
525 	while (--scan >= &td->td_toks_base)
526 		_lwkt_reltokref(scan, td);
527 }
528 
529 /*
530  * This is the decontention version of lwkt_getalltokens().  The tokens are
531  * acquired in address-sorted order to deal with any deadlocks.  Ultimately
532  * token failures will spin into the scheduler and get here.
533  *
534  * Called from critical section
535  */
536 static
537 int
538 _lwkt_getalltokens_sorted(thread_t td)
539 {
540 	lwkt_tokref_t sort_array[LWKT_MAXTOKENS];
541 	lwkt_tokref_t scan;
542 	lwkt_token_t tok;
543 	int i;
544 	int j;
545 	int n;
546 
547 	/*
548 	 * Sort the token array.  Yah yah, I know this isn't fun.
549 	 *
550 	 * NOTE: Recursively acquired tokens are ordered the same as in the
551 	 *	 td_toks_array so we can always get the earliest one first.
552 	 */
553 	i = 0;
554 	scan = &td->td_toks_base;
555 	while (scan < td->td_toks_stop) {
556 		for (j = 0; j < i; ++j) {
557 			if (scan->tr_tok < sort_array[j]->tr_tok)
558 				break;
559 		}
560 		if (j != i) {
561 			bcopy(sort_array + j, sort_array + j + 1,
562 			      (i - j) * sizeof(lwkt_tokref_t));
563 		}
564 		sort_array[j] = scan;
565 		++scan;
566 		++i;
567 	}
568 	n = i;
569 
570 	/*
571 	 * Acquire tokens in forward order, assign or validate tok->t_ref.
572 	 */
573 	for (i = 0; i < n; ++i) {
574 		scan = sort_array[i];
575 		tok = scan->tr_tok;
576 		for (;;) {
577 			/*
578 			 * Only try really hard on the last token
579 			 */
580 			if (scan == td->td_toks_stop - 1) {
581 			    if (_lwkt_trytokref_spin(scan, td, scan->tr_count))
582 				    break;
583 			} else {
584 			    if (_lwkt_trytokref(scan, td, scan->tr_count))
585 				    break;
586 			}
587 
588 			/*
589 			 * Otherwise we failed to acquire all the tokens.
590 			 * Release whatever we did get.
591 			 */
592 			if (lwkt_sched_debug > 0) {
593 				--lwkt_sched_debug;
594 				kprintf("tokb %p %s %s\n",
595 					tok, tok->t_desc, td->td_comm);
596 			}
597 			td->td_wmesg = tok->t_desc;
598 			++tok->t_collisions;
599 			while (--i >= 0) {
600 				scan = sort_array[i];
601 				_lwkt_reltokref(scan, td);
602 			}
603 			return(FALSE);
604 		}
605 	}
606 
607 	/*
608 	 * We were successful, there is no need for another core to signal
609 	 * us.
610 	 */
611 	return (TRUE);
612 }
613 
614 /*
615  * Get a serializing token.  This routine can block.
616  */
617 void
618 lwkt_gettoken(lwkt_token_t tok)
619 {
620 	thread_t td = curthread;
621 	lwkt_tokref_t ref;
622 
623 	ref = td->td_toks_stop;
624 	KKASSERT(ref < &td->td_toks_end);
625 	++td->td_toks_stop;
626 	cpu_ccfence();
627 	_lwkt_tokref_init(ref, tok, td, TOK_EXCLUSIVE|TOK_EXCLREQ);
628 
629 #ifdef DEBUG_LOCKS
630 	/*
631 	 * Taking an exclusive token after holding it shared will
632 	 * livelock. Scan for that case and assert.
633 	 */
634 	lwkt_tokref_t tk;
635 	int found = 0;
636 	for (tk = &td->td_toks_base; tk < ref; tk++) {
637 		if (tk->tr_tok != tok)
638 			continue;
639 
640 		found++;
641 		if (tk->tr_count & TOK_EXCLUSIVE)
642 			goto good;
643 	}
644 	/* We found only shared instances of this token if found >0 here */
645 	KASSERT((found == 0), ("Token %p s/x livelock", tok));
646 good:
647 #endif
648 
649 	if (_lwkt_trytokref_spin(ref, td, TOK_EXCLUSIVE|TOK_EXCLREQ))
650 		return;
651 
652 	/*
653 	 * Give up running if we can't acquire the token right now.
654 	 *
655 	 * Since the tokref is already active the scheduler now
656 	 * takes care of acquisition, so we need only call
657 	 * lwkt_switch().
658 	 *
659 	 * Since we failed this was not a recursive token so upon
660 	 * return tr_tok->t_ref should be assigned to this specific
661 	 * ref.
662 	 */
663 	td->td_wmesg = tok->t_desc;
664 	++tok->t_collisions;
665 	logtoken(fail, ref);
666 	td->td_toks_have = td->td_toks_stop - 1;
667 
668 	if (tokens_debug_output > 0) {
669 		--tokens_debug_output;
670 		spin_lock(&tok_debug_spin);
671 		kprintf("Excl Token thread %p %s %s\n",
672 			td, tok->t_desc, td->td_comm);
673 		print_backtrace(6);
674 		kprintf("\n");
675 		spin_unlock(&tok_debug_spin);
676 	}
677 
678 	lwkt_switch();
679 	logtoken(succ, ref);
680 	KKASSERT(tok->t_ref == ref);
681 }
682 
683 /*
684  * Similar to gettoken but we acquire a shared token instead of an exclusive
685  * token.
686  */
687 void
688 lwkt_gettoken_shared(lwkt_token_t tok)
689 {
690 	thread_t td = curthread;
691 	lwkt_tokref_t ref;
692 
693 	ref = td->td_toks_stop;
694 	KKASSERT(ref < &td->td_toks_end);
695 	++td->td_toks_stop;
696 	cpu_ccfence();
697 	_lwkt_tokref_init(ref, tok, td, TOK_EXCLREQ);
698 
699 #ifdef DEBUG_LOCKS
700         /*
701          * Taking a pool token in shared mode is a bad idea; other
702          * addresses deeper in the call stack may hash to the same pool
703          * token and you may end up with an exclusive-shared livelock.
704          * Warn in this condition.
705          */
706         if ((tok >= &pool_tokens[0]) &&
707             (tok < &pool_tokens[LWKT_NUM_POOL_TOKENS]))
708                 kprintf("Warning! Taking pool token %p in shared mode\n", tok);
709 #endif
710 
711 
712 	if (_lwkt_trytokref_spin(ref, td, TOK_EXCLREQ))
713 		return;
714 
715 	/*
716 	 * Give up running if we can't acquire the token right now.
717 	 *
718 	 * Since the tokref is already active the scheduler now
719 	 * takes care of acquisition, so we need only call
720 	 * lwkt_switch().
721 	 *
722 	 * Since we failed this was not a recursive token so upon
723 	 * return tr_tok->t_ref should be assigned to this specific
724 	 * ref.
725 	 */
726 	td->td_wmesg = tok->t_desc;
727 	++tok->t_collisions;
728 	logtoken(fail, ref);
729 	td->td_toks_have = td->td_toks_stop - 1;
730 
731 	if (tokens_debug_output > 0) {
732 		--tokens_debug_output;
733 		spin_lock(&tok_debug_spin);
734 		kprintf("Shar Token thread %p %s %s\n",
735 			td, tok->t_desc, td->td_comm);
736 		print_backtrace(6);
737 		kprintf("\n");
738 		spin_unlock(&tok_debug_spin);
739 	}
740 
741 	lwkt_switch();
742 	logtoken(succ, ref);
743 }
744 
745 /*
746  * Attempt to acquire a token, return TRUE on success, FALSE on failure.
747  *
748  * We setup the tokref in case we actually get the token (if we switch later
749  * it becomes mandatory so we set TOK_EXCLREQ), but we call trytokref without
750  * TOK_EXCLREQ in case we fail.
751  */
752 int
753 lwkt_trytoken(lwkt_token_t tok)
754 {
755 	thread_t td = curthread;
756 	lwkt_tokref_t ref;
757 
758 	ref = td->td_toks_stop;
759 	KKASSERT(ref < &td->td_toks_end);
760 	++td->td_toks_stop;
761 	cpu_ccfence();
762 	_lwkt_tokref_init(ref, tok, td, TOK_EXCLUSIVE|TOK_EXCLREQ);
763 
764 	if (_lwkt_trytokref(ref, td, TOK_EXCLUSIVE))
765 		return TRUE;
766 
767 	/*
768 	 * Failed, unpend the request
769 	 */
770 	cpu_ccfence();
771 	--td->td_toks_stop;
772 	++tok->t_collisions;
773 	return FALSE;
774 }
775 
776 
777 void
778 lwkt_gettoken_hard(lwkt_token_t tok)
779 {
780 	lwkt_gettoken(tok);
781 	crit_enter_hard();
782 }
783 
784 lwkt_token_t
785 lwkt_getpooltoken(void *ptr)
786 {
787 	lwkt_token_t tok;
788 
789 	tok = _lwkt_token_pool_lookup(ptr);
790 	lwkt_gettoken(tok);
791 	return (tok);
792 }
793 
794 /*
795  * Release a serializing token.
796  *
797  * WARNING!  All tokens must be released in reverse order.  This will be
798  *	     asserted.
799  */
800 void
801 lwkt_reltoken(lwkt_token_t tok)
802 {
803 	thread_t td = curthread;
804 	lwkt_tokref_t ref;
805 
806 	/*
807 	 * Remove ref from thread token list and assert that it matches
808 	 * the token passed in.  Tokens must be released in reverse order.
809 	 */
810 	ref = td->td_toks_stop - 1;
811 	KKASSERT(ref >= &td->td_toks_base && ref->tr_tok == tok);
812 	_lwkt_reltokref(ref, td);
813 	cpu_sfence();
814 	td->td_toks_stop = ref;
815 }
816 
817 void
818 lwkt_reltoken_hard(lwkt_token_t tok)
819 {
820 	lwkt_reltoken(tok);
821 	crit_exit_hard();
822 }
823 
824 /*
825  * It is faster for users of lwkt_getpooltoken() to use the returned
826  * token and just call lwkt_reltoken(), but for convenience we provide
827  * this function which looks the token up based on the ident.
828  */
829 void
830 lwkt_relpooltoken(void *ptr)
831 {
832 	lwkt_token_t tok = _lwkt_token_pool_lookup(ptr);
833 	lwkt_reltoken(tok);
834 }
835 
836 /*
837  * Return a count of the number of token refs the thread has to the
838  * specified token, whether it currently owns the token or not.
839  */
840 int
841 lwkt_cnttoken(lwkt_token_t tok, thread_t td)
842 {
843 	lwkt_tokref_t scan;
844 	int count = 0;
845 
846 	for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) {
847 		if (scan->tr_tok == tok)
848 			++count;
849 	}
850 	return(count);
851 }
852 
853 /*
854  * Pool tokens are used to provide a type-stable serializing token
855  * pointer that does not race against disappearing data structures.
856  *
857  * This routine is called in early boot just after we setup the BSP's
858  * globaldata structure.
859  */
860 void
861 lwkt_token_pool_init(void)
862 {
863 	int i;
864 
865 	for (i = 0; i < LWKT_NUM_POOL_TOKENS; ++i)
866 		lwkt_token_init(&pool_tokens[i], "pool");
867 }
868 
869 lwkt_token_t
870 lwkt_token_pool_lookup(void *ptr)
871 {
872 	return (_lwkt_token_pool_lookup(ptr));
873 }
874 
875 /*
876  * Initialize a token.
877  */
878 void
879 lwkt_token_init(lwkt_token_t tok, const char *desc)
880 {
881 	tok->t_count = 0;
882 	tok->t_ref = NULL;
883 	tok->t_collisions = 0;
884 	tok->t_desc = desc;
885 }
886 
887 void
888 lwkt_token_uninit(lwkt_token_t tok)
889 {
890 	/* empty */
891 }
892 
893 /*
894  * Exchange the two most recent tokens on the tokref stack.  This allows
895  * you to release a token out of order.
896  *
897  * We have to be careful about the case where the top two tokens are
898  * the same token.  In this case tok->t_ref will point to the deeper
899  * ref and must remain pointing to the deeper ref.  If we were to swap
900  * it the first release would clear the token even though a second
901  * ref is still present.
902  *
903  * Only exclusively held tokens contain a reference to the tokref which
904  * has to be flipped along with the swap.
905  */
906 void
907 lwkt_token_swap(void)
908 {
909 	lwkt_tokref_t ref1, ref2;
910 	lwkt_token_t tok1, tok2;
911 	long count1, count2;
912 	thread_t td = curthread;
913 
914 	crit_enter();
915 
916 	ref1 = td->td_toks_stop - 1;
917 	ref2 = td->td_toks_stop - 2;
918 	KKASSERT(ref1 >= &td->td_toks_base);
919 	KKASSERT(ref2 >= &td->td_toks_base);
920 
921 	tok1 = ref1->tr_tok;
922 	tok2 = ref2->tr_tok;
923 	count1 = ref1->tr_count;
924 	count2 = ref2->tr_count;
925 
926 	if (tok1 != tok2) {
927 		ref1->tr_tok = tok2;
928 		ref1->tr_count = count2;
929 		ref2->tr_tok = tok1;
930 		ref2->tr_count = count1;
931 		if (tok1->t_ref == ref1)
932 			tok1->t_ref = ref2;
933 		if (tok2->t_ref == ref2)
934 			tok2->t_ref = ref1;
935 	}
936 
937 	crit_exit();
938 }
939 
940 #ifdef DDB
941 DB_SHOW_COMMAND(tokens, db_tok_all)
942 {
943 	struct lwkt_token *tok, **ptr;
944 	struct lwkt_token *toklist[16] = {
945 		&mp_token,
946 		&pmap_token,
947 		&dev_token,
948 		&vm_token,
949 		&vmspace_token,
950 		&kvm_token,
951 		&sigio_token,
952 		&tty_token,
953 		&vnode_token,
954 		NULL
955 	};
956 
957 	ptr = toklist;
958 	for (tok = *ptr; tok; tok = *(++ptr)) {
959 		db_printf("tok=%p tr_owner=%p t_colissions=%ld t_desc=%s\n", tok,
960 		    (tok->t_ref ? tok->t_ref->tr_owner : NULL),
961 		    tok->t_collisions, tok->t_desc);
962 	}
963 }
964 #endif /* DDB */
965