1 /*
2 * Copyright (c) 2003-2006,2009-2019 The DragonFly Project.
3 * All rights reserved.
4 *
5 * This code is derived from software contributed to The DragonFly Project
6 * by Matthew Dillon <dillon@backplane.com>
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 *
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in
16 * the documentation and/or other materials provided with the
17 * distribution.
18 * 3. Neither the name of The DragonFly Project nor the names of its
19 * contributors may be used to endorse or promote products derived
20 * from this software without specific, prior written permission.
21 *
22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * SUCH DAMAGE.
34 */
35
36 /*
37 * lwkt_token - Implement soft token locks.
38 *
39 * Tokens are locks which serialize a thread only while the thread is
40 * running. If the thread blocks all tokens are released, then reacquired
41 * when the thread resumes.
42 *
43 * This implementation requires no critical sections or spin locks, but
44 * does use atomic_cmpset_ptr().
45 *
46 * Tokens may be recursively acquired by the same thread. However the
47 * caller must be sure to release such tokens in reverse order.
48 */
49 #include <sys/param.h>
50 #include <sys/systm.h>
51 #include <sys/kernel.h>
52 #include <sys/proc.h>
53 #include <sys/rtprio.h>
54 #include <sys/queue.h>
55 #include <sys/sysctl.h>
56 #include <sys/ktr.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 #include "opt_ddb.h"
80 #ifdef DDB
81 #include <ddb/ddb.h>
82 #endif
83
84 extern int lwkt_sched_debug;
85
86 #define LWKT_POOL_TOKENS 16384 /* must be power of 2 */
87 #define LWKT_POOL_MASK (LWKT_POOL_TOKENS - 1)
88
89 struct lwkt_pool_token {
90 struct lwkt_token token;
91 } __cachealign;
92
93 static struct lwkt_pool_token pool_tokens[LWKT_POOL_TOKENS];
94 static struct spinlock tok_debug_spin =
95 SPINLOCK_INITIALIZER(&tok_debug_spin, "tok_debug_spin");
96
97 #define TOKEN_STRING "REF=%p TOK=%p TD=%p"
98 #define TOKEN_ARGS lwkt_tokref_t ref, lwkt_token_t tok, struct thread *td
99 #define CONTENDED_STRING TOKEN_STRING " (contention started)"
100 #define UNCONTENDED_STRING TOKEN_STRING " (contention stopped)"
101 #if !defined(KTR_TOKENS)
102 #define KTR_TOKENS KTR_ALL
103 #endif
104
105 KTR_INFO_MASTER(tokens);
106 KTR_INFO(KTR_TOKENS, tokens, fail, 0, TOKEN_STRING, TOKEN_ARGS);
107 KTR_INFO(KTR_TOKENS, tokens, succ, 1, TOKEN_STRING, TOKEN_ARGS);
108 #if 0
109 KTR_INFO(KTR_TOKENS, tokens, release, 2, TOKEN_STRING, TOKEN_ARGS);
110 KTR_INFO(KTR_TOKENS, tokens, remote, 3, TOKEN_STRING, TOKEN_ARGS);
111 KTR_INFO(KTR_TOKENS, tokens, reqremote, 4, TOKEN_STRING, TOKEN_ARGS);
112 KTR_INFO(KTR_TOKENS, tokens, reqfail, 5, TOKEN_STRING, TOKEN_ARGS);
113 KTR_INFO(KTR_TOKENS, tokens, drain, 6, TOKEN_STRING, TOKEN_ARGS);
114 KTR_INFO(KTR_TOKENS, tokens, contention_start, 7, CONTENDED_STRING, TOKEN_ARGS);
115 KTR_INFO(KTR_TOKENS, tokens, contention_stop, 7, UNCONTENDED_STRING, TOKEN_ARGS);
116 #endif
117
118 #define logtoken(name, ref) \
119 KTR_LOG(tokens_ ## name, ref, ref->tr_tok, curthread)
120
121 /*
122 * Global tokens. These replace the MP lock for major subsystem locking.
123 * These tokens are initially used to lockup both global and individual
124 * operations.
125 *
126 * Once individual structures get their own locks these tokens are used
127 * only to protect global lists & other variables and to interlock
128 * allocations and teardowns and such.
129 *
130 * The UP initializer causes token acquisition to also acquire the MP lock
131 * for maximum compatibility. The feature may be enabled and disabled at
132 * any time, the MP state is copied to the tokref when the token is acquired
133 * and will not race against sysctl changes.
134 */
135 struct lwkt_token mp_token = LWKT_TOKEN_INITIALIZER(mp_token);
136 struct lwkt_token pmap_token = LWKT_TOKEN_INITIALIZER(pmap_token);
137 struct lwkt_token dev_token = LWKT_TOKEN_INITIALIZER(dev_token);
138 struct lwkt_token vm_token = LWKT_TOKEN_INITIALIZER(vm_token);
139 struct lwkt_token vmspace_token = LWKT_TOKEN_INITIALIZER(vmspace_token);
140 struct lwkt_token kvm_token = LWKT_TOKEN_INITIALIZER(kvm_token);
141 struct lwkt_token sigio_token = LWKT_TOKEN_INITIALIZER(sigio_token);
142 struct lwkt_token tty_token = LWKT_TOKEN_INITIALIZER(tty_token);
143 struct lwkt_token vnode_token = LWKT_TOKEN_INITIALIZER(vnode_token);
144 struct lwkt_token vga_token = LWKT_TOKEN_INITIALIZER(vga_token);
145 struct lwkt_token kbd_token = LWKT_TOKEN_INITIALIZER(kbd_token);
146
147 /*
148 * Exponential backoff (exclusive tokens) and TSC windowing (shared tokens)
149 * parameters. Remember that tokens backoff to the scheduler. This is a bit
150 * of trade-off. Smaller values like 128 work better in some situations,
151 * but under extreme loads larger values like 4096 seem to provide the most
152 * determinism.
153 */
154 static int token_backoff_max __cachealign = 4096;
155 SYSCTL_INT(_lwkt, OID_AUTO, token_backoff_max, CTLFLAG_RW,
156 &token_backoff_max, 0, "Tokens exponential backoff");
157 static int token_window_shift __cachealign = 8;
158 SYSCTL_INT(_lwkt, OID_AUTO, token_window_shift, CTLFLAG_RW,
159 &token_window_shift, 0, "Tokens TSC windowing shift");
160
161 /*
162 * The collision count is bumped every time the LWKT scheduler fails
163 * to acquire needed tokens in addition to a normal lwkt_gettoken()
164 * stall.
165 */
166 SYSCTL_LONG(_lwkt, OID_AUTO, mp_collisions, CTLFLAG_RW,
167 &mp_token.t_collisions, 0, "Collision counter of mp_token");
168 SYSCTL_LONG(_lwkt, OID_AUTO, pmap_collisions, CTLFLAG_RW,
169 &pmap_token.t_collisions, 0, "Collision counter of pmap_token");
170 SYSCTL_LONG(_lwkt, OID_AUTO, dev_collisions, CTLFLAG_RW,
171 &dev_token.t_collisions, 0, "Collision counter of dev_token");
172 SYSCTL_LONG(_lwkt, OID_AUTO, vm_collisions, CTLFLAG_RW,
173 &vm_token.t_collisions, 0, "Collision counter of vm_token");
174 SYSCTL_LONG(_lwkt, OID_AUTO, vmspace_collisions, CTLFLAG_RW,
175 &vmspace_token.t_collisions, 0, "Collision counter of vmspace_token");
176 SYSCTL_LONG(_lwkt, OID_AUTO, kvm_collisions, CTLFLAG_RW,
177 &kvm_token.t_collisions, 0, "Collision counter of kvm_token");
178 SYSCTL_LONG(_lwkt, OID_AUTO, sigio_collisions, CTLFLAG_RW,
179 &sigio_token.t_collisions, 0, "Collision counter of sigio_token");
180 SYSCTL_LONG(_lwkt, OID_AUTO, tty_collisions, CTLFLAG_RW,
181 &tty_token.t_collisions, 0, "Collision counter of tty_token");
182 SYSCTL_LONG(_lwkt, OID_AUTO, vnode_collisions, CTLFLAG_RW,
183 &vnode_token.t_collisions, 0, "Collision counter of vnode_token");
184
185 static int tokens_debug_output;
186 SYSCTL_INT(_lwkt, OID_AUTO, tokens_debug_output, CTLFLAG_RW,
187 &tokens_debug_output, 0, "Generate stack trace N times");
188
189 static int _lwkt_getalltokens_sorted(thread_t td);
190
191 /*
192 * Acquire the initial mplock
193 *
194 * (low level boot only)
195 */
196 void
cpu_get_initial_mplock(void)197 cpu_get_initial_mplock(void)
198 {
199 KKASSERT(mp_token.t_ref == NULL);
200 if (lwkt_trytoken(&mp_token) == FALSE)
201 panic("cpu_get_initial_mplock");
202 }
203
204 /*
205 * Return a pool token given an address. Use a prime number to reduce
206 * overlaps.
207 */
208 #define POOL_HASH_PRIME1 66555444443333333ULL
209 #define POOL_HASH_PRIME2 989042931893ULL
210
211 static __inline
212 lwkt_token_t
_lwkt_token_pool_lookup(void * ptr)213 _lwkt_token_pool_lookup(void *ptr)
214 {
215 uintptr_t hash1;
216 uintptr_t hash2;
217
218 hash1 = (uintptr_t)ptr + ((uintptr_t)ptr >> 18);
219 hash1 %= POOL_HASH_PRIME1;
220 hash2 = ((uintptr_t)ptr >> 8) + ((uintptr_t)ptr >> 24);
221 hash2 %= POOL_HASH_PRIME2;
222 return (&pool_tokens[(hash1 ^ hash2) & LWKT_POOL_MASK].token);
223 }
224
225 /*
226 * Initialize a tokref_t prior to making it visible in the thread's
227 * token array.
228 */
229 static __inline
230 void
_lwkt_tokref_init(lwkt_tokref_t ref,lwkt_token_t tok,thread_t td,long excl)231 _lwkt_tokref_init(lwkt_tokref_t ref, lwkt_token_t tok, thread_t td, long excl)
232 {
233 ref->tr_tok = tok;
234 ref->tr_count = excl;
235 ref->tr_owner = td;
236 }
237
238 /*
239 * Attempt to acquire a shared or exclusive token. Returns TRUE on success,
240 * FALSE on failure.
241 *
242 * If TOK_EXCLUSIVE is set in mode we are attempting to get an exclusive
243 * token, otherwise are attempting to get a shared token.
244 *
245 * If TOK_EXCLREQ is set in mode this is a blocking operation, otherwise
246 * it is a non-blocking operation (for both exclusive or shared acquisions).
247 */
248 static __inline
249 int
_lwkt_trytokref(lwkt_tokref_t ref,thread_t td,long mode)250 _lwkt_trytokref(lwkt_tokref_t ref, thread_t td, long mode)
251 {
252 lwkt_token_t tok;
253 lwkt_tokref_t oref;
254 long count;
255
256 tok = ref->tr_tok;
257 KASSERT(((mode & TOK_EXCLREQ) == 0 || /* non blocking */
258 td->td_gd->gd_intr_nesting_level == 0 ||
259 panic_cpu_gd == mycpu),
260 ("Attempt to acquire token %p not already "
261 "held in hard code section", tok));
262
263 if (mode & TOK_EXCLUSIVE) {
264 /*
265 * Attempt to get an exclusive token
266 */
267 count = tok->t_count;
268
269 for (;;) {
270 oref = tok->t_ref; /* can be NULL */
271 cpu_ccfence();
272 if ((count & ~TOK_EXCLREQ) == 0) {
273 /*
274 * It is possible to get the exclusive bit.
275 * We must clear TOK_EXCLREQ on successful
276 * acquisition.
277 */
278 if (atomic_fcmpset_long(&tok->t_count, &count,
279 (count & ~TOK_EXCLREQ) |
280 TOK_EXCLUSIVE)) {
281 KKASSERT(tok->t_ref == NULL);
282 tok->t_ref = ref;
283 return TRUE;
284 }
285 /* retry */
286 } else if ((count & TOK_EXCLUSIVE) &&
287 oref >= &td->td_toks_base &&
288 oref < td->td_toks_stop) {
289 /*
290 * Our thread already holds the exclusive
291 * bit, we treat this tokref as a shared
292 * token (sorta) to make the token release
293 * code easier. Treating this as a shared
294 * token allows us to simply increment the
295 * count field.
296 *
297 * NOTE: oref cannot race above if it
298 * happens to be ours, so we're good.
299 * But we must still have a stable
300 * variable for both parts of the
301 * comparison.
302 *
303 * NOTE: Since we already have an exclusive
304 * lock and don't need to check EXCLREQ
305 * we can just use an atomic_add here
306 */
307 atomic_add_long(&tok->t_count, TOK_INCR);
308 ref->tr_count &= ~TOK_EXCLUSIVE;
309 return TRUE;
310 } else if ((mode & TOK_EXCLREQ) &&
311 (count & TOK_EXCLREQ) == 0) {
312 /*
313 * Unable to get the exclusive bit but being
314 * asked to set the exclusive-request bit.
315 * Since we are going to retry anyway just
316 * set the bit unconditionally.
317 */
318 atomic_set_long(&tok->t_count, TOK_EXCLREQ);
319 return FALSE;
320 } else {
321 /*
322 * Unable to get the exclusive bit and not
323 * being asked to set the exclusive-request
324 * (aka lwkt_trytoken()), or EXCLREQ was
325 * already set.
326 */
327 cpu_pause();
328 return FALSE;
329 }
330 /* retry */
331 }
332 } else {
333 /*
334 * Attempt to get a shared token. Note that TOK_EXCLREQ
335 * for shared tokens simply means the caller intends to
336 * block. We never actually set the bit in tok->t_count.
337 *
338 * Due to the token's no-deadlock guarantee, and complications
339 * created by the sorted reacquisition code, we can only
340 * give exclusive requests priority over shared requests
341 * in situations where the thread holds only one token.
342 */
343 count = tok->t_count;
344
345 for (;;) {
346 oref = tok->t_ref; /* can be NULL */
347 cpu_ccfence();
348 if ((count & (TOK_EXCLUSIVE|mode)) == 0 ||
349 ((count & TOK_EXCLUSIVE) == 0 &&
350 td->td_toks_stop != &td->td_toks_base + 1)
351 ) {
352 /*
353 * It may be possible to get the token shared.
354 */
355 if ((atomic_fetchadd_long(&tok->t_count, TOK_INCR) & TOK_EXCLUSIVE) == 0) {
356 return TRUE;
357 }
358 count = atomic_fetchadd_long(&tok->t_count,
359 -TOK_INCR);
360 count -= TOK_INCR;
361 /* retry */
362 } else if ((count & TOK_EXCLUSIVE) &&
363 oref >= &td->td_toks_base &&
364 oref < td->td_toks_stop) {
365 /*
366 * We own the exclusive bit on the token so
367 * we can in fact also get it shared.
368 */
369 atomic_add_long(&tok->t_count, TOK_INCR);
370 return TRUE;
371 } else {
372 /*
373 * We failed to get the token shared
374 */
375 return FALSE;
376 }
377 /* retry */
378 }
379 }
380 }
381
382 static __inline
383 int
_lwkt_trytokref_spin(lwkt_tokref_t ref,thread_t td,long mode)384 _lwkt_trytokref_spin(lwkt_tokref_t ref, thread_t td, long mode)
385 {
386 if (_lwkt_trytokref(ref, td, mode))
387 return TRUE;
388
389 if (mode & TOK_EXCLUSIVE) {
390 /*
391 * Contested exclusive token, use exponential backoff
392 * algorithm.
393 */
394 long expbackoff;
395 long loop;
396
397 expbackoff = 0;
398 while (expbackoff < 6 + token_backoff_max) {
399 expbackoff = (expbackoff + 1) * 3 / 2;
400 if ((rdtsc() >> token_window_shift) % ncpus != mycpuid) {
401 for (loop = expbackoff; loop; --loop)
402 cpu_pause();
403 }
404 if (_lwkt_trytokref(ref, td, mode))
405 return TRUE;
406 }
407 } else {
408 /*
409 * Contested shared token, use TSC windowing. Note that
410 * exclusive tokens have priority over shared tokens only
411 * for the first token.
412 */
413 if ((rdtsc() >> token_window_shift) % ncpus == mycpuid) {
414 if (_lwkt_trytokref(ref, td, mode & ~TOK_EXCLREQ))
415 return TRUE;
416 } else {
417 if (_lwkt_trytokref(ref, td, mode))
418 return TRUE;
419 }
420
421 }
422 ++mycpu->gd_cnt.v_lock_colls;
423
424 return FALSE;
425 }
426
427 /*
428 * Release a token that we hold.
429 *
430 * Since tokens are polled, we don't have to deal with wakeups and releasing
431 * is really easy.
432 */
433 static __inline
434 void
_lwkt_reltokref(lwkt_tokref_t ref,thread_t td)435 _lwkt_reltokref(lwkt_tokref_t ref, thread_t td)
436 {
437 lwkt_token_t tok;
438 long count;
439
440 tok = ref->tr_tok;
441 if (tok->t_ref == ref) {
442 /*
443 * We are an exclusive holder. We must clear tr_ref
444 * before we clear the TOK_EXCLUSIVE bit. If we are
445 * unable to clear the bit we must restore
446 * tok->t_ref.
447 */
448 #if 0
449 KKASSERT(count & TOK_EXCLUSIVE);
450 #endif
451 tok->t_ref = NULL;
452 atomic_clear_long(&tok->t_count, TOK_EXCLUSIVE);
453 } else {
454 /*
455 * We are a shared holder
456 */
457 count = atomic_fetchadd_long(&tok->t_count, -TOK_INCR);
458 KKASSERT(count & TOK_COUNTMASK); /* count prior */
459 }
460 }
461
462 /*
463 * Obtain all the tokens required by the specified thread on the current
464 * cpu, return 0 on failure and non-zero on success. If a failure occurs
465 * any partially acquired tokens will be released prior to return.
466 *
467 * lwkt_getalltokens is called by the LWKT scheduler to re-acquire all
468 * tokens that the thread had to release when it switched away.
469 *
470 * If spinning is non-zero this function acquires the tokens in a particular
471 * order to deal with potential deadlocks. We simply use address order for
472 * the case.
473 *
474 * Called from a critical section.
475 */
476 int
lwkt_getalltokens(thread_t td,int spinning)477 lwkt_getalltokens(thread_t td, int spinning)
478 {
479 lwkt_tokref_t scan;
480 lwkt_token_t tok;
481
482 if (spinning)
483 return(_lwkt_getalltokens_sorted(td));
484
485 /*
486 * Acquire tokens in forward order, assign or validate tok->t_ref.
487 */
488 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) {
489 tok = scan->tr_tok;
490 for (;;) {
491 /*
492 * Only try really hard on the last token
493 */
494 if (scan == td->td_toks_stop - 1) {
495 if (_lwkt_trytokref_spin(scan, td, scan->tr_count))
496 break;
497 } else {
498 if (_lwkt_trytokref(scan, td, scan->tr_count))
499 break;
500 }
501
502 /*
503 * Otherwise we failed to acquire all the tokens.
504 * Release whatever we did get.
505 */
506 KASSERT(tok->t_desc,
507 ("token %p is not initialized", tok));
508 td->td_gd->gd_cnt.v_lock_addr = tok;
509 td->td_gd->gd_cnt.v_lock_name[0] = 't';
510 strncpy(td->td_gd->gd_cnt.v_lock_name + 1,
511 tok->t_desc,
512 sizeof(td->td_gd->gd_cnt.v_lock_name) - 2);
513 if (lwkt_sched_debug > 0) {
514 --lwkt_sched_debug;
515 kprintf("toka %p %s %s\n",
516 tok, tok->t_desc, td->td_comm);
517 }
518 td->td_wmesg = tok->t_desc;
519 ++tok->t_collisions;
520 while (--scan >= &td->td_toks_base)
521 _lwkt_reltokref(scan, td);
522 return(FALSE);
523 }
524 }
525 return (TRUE);
526 }
527
528 /*
529 * Release all tokens owned by the specified thread on the current cpu.
530 *
531 * This code is really simple. Even in cases where we own all the tokens
532 * note that t_ref may not match the scan for recursively held tokens which
533 * are held deeper in the stack, or for the case where a lwkt_getalltokens()
534 * failed.
535 *
536 * Tokens are released in reverse order to reduce chasing race failures.
537 *
538 * Called from a critical section.
539 */
540 void
lwkt_relalltokens(thread_t td)541 lwkt_relalltokens(thread_t td)
542 {
543 lwkt_tokref_t scan;
544
545 /*
546 * Weird order is to try to avoid a panic loop
547 */
548 if (td->td_toks_have) {
549 scan = td->td_toks_have;
550 td->td_toks_have = NULL;
551 } else {
552 scan = td->td_toks_stop;
553 }
554 while (--scan >= &td->td_toks_base)
555 _lwkt_reltokref(scan, td);
556 }
557
558 /*
559 * This is the decontention version of lwkt_getalltokens(). The tokens are
560 * acquired in address-sorted order to deal with any deadlocks. Ultimately
561 * token failures will spin into the scheduler and get here.
562 *
563 * Called from critical section
564 */
565 static
566 int
_lwkt_getalltokens_sorted(thread_t td)567 _lwkt_getalltokens_sorted(thread_t td)
568 {
569 lwkt_tokref_t sort_array[LWKT_MAXTOKENS];
570 lwkt_tokref_t scan;
571 lwkt_token_t tok;
572 int i;
573 int j;
574 int n;
575
576 /*
577 * Sort the token array. Yah yah, I know this isn't fun.
578 *
579 * NOTE: Recursively acquired tokens are ordered the same as in the
580 * td_toks_array so we can always get the earliest one first.
581 * This is particularly important when a token is acquired
582 * exclusively multiple times, as only the first acquisition
583 * is treated as an exclusive token.
584 */
585 i = 0;
586 scan = &td->td_toks_base;
587 while (scan < td->td_toks_stop) {
588 for (j = 0; j < i; ++j) {
589 if (scan->tr_tok < sort_array[j]->tr_tok)
590 break;
591 }
592 if (j != i) {
593 bcopy(sort_array + j, sort_array + j + 1,
594 (i - j) * sizeof(lwkt_tokref_t));
595 }
596 sort_array[j] = scan;
597 ++scan;
598 ++i;
599 }
600 n = i;
601
602 /*
603 * Acquire tokens in forward order, assign or validate tok->t_ref.
604 */
605 for (i = 0; i < n; ++i) {
606 scan = sort_array[i];
607 tok = scan->tr_tok;
608 for (;;) {
609 /*
610 * Only try really hard on the last token
611 */
612 if (scan == td->td_toks_stop - 1) {
613 if (_lwkt_trytokref_spin(scan, td, scan->tr_count))
614 break;
615 } else {
616 if (_lwkt_trytokref(scan, td, scan->tr_count))
617 break;
618 }
619
620 /*
621 * Otherwise we failed to acquire all the tokens.
622 * Release whatever we did get.
623 */
624 td->td_gd->gd_cnt.v_lock_addr = tok;
625 td->td_gd->gd_cnt.v_lock_name[0] = 't';
626 strncpy(td->td_gd->gd_cnt.v_lock_name + 1,
627 tok->t_desc,
628 sizeof(td->td_gd->gd_cnt.v_lock_name) - 2);
629 if (lwkt_sched_debug > 0) {
630 --lwkt_sched_debug;
631 kprintf("tokb %p %s %s\n",
632 tok, tok->t_desc, td->td_comm);
633 }
634 td->td_wmesg = tok->t_desc;
635 ++tok->t_collisions;
636 while (--i >= 0) {
637 scan = sort_array[i];
638 _lwkt_reltokref(scan, td);
639 }
640 return(FALSE);
641 }
642 }
643
644 /*
645 * We were successful, there is no need for another core to signal
646 * us.
647 */
648 return (TRUE);
649 }
650
651 /*
652 * Get a serializing token. This routine can block.
653 */
654 void
lwkt_gettoken(lwkt_token_t tok)655 lwkt_gettoken(lwkt_token_t tok)
656 {
657 thread_t td = curthread;
658 lwkt_tokref_t ref;
659
660 ref = td->td_toks_stop;
661 KKASSERT(ref < &td->td_toks_end);
662 ++td->td_toks_stop;
663 cpu_ccfence();
664 _lwkt_tokref_init(ref, tok, td, TOK_EXCLUSIVE|TOK_EXCLREQ);
665
666 #ifdef DEBUG_LOCKS
667 /*
668 * Taking an exclusive token after holding it shared will
669 * livelock. Scan for that case and assert.
670 */
671 lwkt_tokref_t tk;
672 int found = 0;
673 for (tk = &td->td_toks_base; tk < ref; tk++) {
674 if (tk->tr_tok != tok)
675 continue;
676
677 found++;
678 if (tk->tr_count & TOK_EXCLUSIVE)
679 goto good;
680 }
681 /* We found only shared instances of this token if found >0 here */
682 KASSERT((found == 0), ("Token %p s/x livelock", tok));
683 good:
684 #endif
685
686 if (_lwkt_trytokref_spin(ref, td, TOK_EXCLUSIVE|TOK_EXCLREQ))
687 return;
688
689 /*
690 * Give up running if we can't acquire the token right now.
691 *
692 * Since the tokref is already active the scheduler now
693 * takes care of acquisition, so we need only call
694 * lwkt_switch().
695 *
696 * Since we failed this was not a recursive token so upon
697 * return tr_tok->t_ref should be assigned to this specific
698 * ref.
699 */
700 td->td_wmesg = tok->t_desc;
701 ++tok->t_collisions;
702 logtoken(fail, ref);
703 td->td_toks_have = td->td_toks_stop - 1;
704
705 if (tokens_debug_output > 0) {
706 --tokens_debug_output;
707 spin_lock(&tok_debug_spin);
708 kprintf("Excl Token %p thread %p %s %s\n",
709 tok, td, tok->t_desc, td->td_comm);
710 print_backtrace(6);
711 kprintf("\n");
712 spin_unlock(&tok_debug_spin);
713 }
714
715 atomic_set_int(&td->td_mpflags, TDF_MP_DIDYIELD);
716 lwkt_switch();
717 logtoken(succ, ref);
718 KKASSERT(tok->t_ref == ref);
719 }
720
721 /*
722 * Similar to gettoken but we acquire a shared token instead of an exclusive
723 * token.
724 */
725 void
lwkt_gettoken_shared(lwkt_token_t tok)726 lwkt_gettoken_shared(lwkt_token_t tok)
727 {
728 thread_t td = curthread;
729 lwkt_tokref_t ref;
730
731 ref = td->td_toks_stop;
732 KKASSERT(ref < &td->td_toks_end);
733 ++td->td_toks_stop;
734 cpu_ccfence();
735 _lwkt_tokref_init(ref, tok, td, TOK_EXCLREQ);
736
737 #ifdef DEBUG_LOCKS
738 /*
739 * Taking a pool token in shared mode is a bad idea; other
740 * addresses deeper in the call stack may hash to the same pool
741 * token and you may end up with an exclusive-shared livelock.
742 * Warn in this condition.
743 */
744 if ((tok >= &pool_tokens[0].token) &&
745 (tok < &pool_tokens[LWKT_POOL_TOKENS].token))
746 kprintf("Warning! Taking pool token %p in shared mode\n", tok);
747 #endif
748
749
750 if (_lwkt_trytokref_spin(ref, td, TOK_EXCLREQ))
751 return;
752
753 /*
754 * Give up running if we can't acquire the token right now.
755 *
756 * Since the tokref is already active the scheduler now
757 * takes care of acquisition, so we need only call
758 * lwkt_switch().
759 *
760 * Since we failed this was not a recursive token so upon
761 * return tr_tok->t_ref should be assigned to this specific
762 * ref.
763 */
764 td->td_wmesg = tok->t_desc;
765 ++tok->t_collisions;
766 logtoken(fail, ref);
767 td->td_toks_have = td->td_toks_stop - 1;
768
769 if (tokens_debug_output > 0) {
770 --tokens_debug_output;
771 spin_lock(&tok_debug_spin);
772 kprintf("Shar Token %p thread %p %s %s\n",
773 tok, td, tok->t_desc, td->td_comm);
774 print_backtrace(6);
775 kprintf("\n");
776 spin_unlock(&tok_debug_spin);
777 }
778
779 atomic_set_int(&td->td_mpflags, TDF_MP_DIDYIELD);
780 lwkt_switch();
781 logtoken(succ, ref);
782 }
783
784 /*
785 * Attempt to acquire a token, return TRUE on success, FALSE on failure.
786 *
787 * We setup the tokref in case we actually get the token (if we switch later
788 * it becomes mandatory so we set TOK_EXCLREQ), but we call trytokref without
789 * TOK_EXCLREQ in case we fail.
790 */
791 int
lwkt_trytoken(lwkt_token_t tok)792 lwkt_trytoken(lwkt_token_t tok)
793 {
794 thread_t td = curthread;
795 lwkt_tokref_t ref;
796
797 ref = td->td_toks_stop;
798 KKASSERT(ref < &td->td_toks_end);
799 ++td->td_toks_stop;
800 cpu_ccfence();
801 _lwkt_tokref_init(ref, tok, td, TOK_EXCLUSIVE|TOK_EXCLREQ);
802
803 if (_lwkt_trytokref(ref, td, TOK_EXCLUSIVE))
804 return TRUE;
805
806 /*
807 * Failed, unpend the request
808 */
809 cpu_ccfence();
810 --td->td_toks_stop;
811 ++tok->t_collisions;
812 return FALSE;
813 }
814
815 lwkt_token_t
lwkt_getpooltoken(void * ptr)816 lwkt_getpooltoken(void *ptr)
817 {
818 lwkt_token_t tok;
819
820 tok = _lwkt_token_pool_lookup(ptr);
821 lwkt_gettoken(tok);
822 return (tok);
823 }
824
825 /*
826 * Release a serializing token.
827 *
828 * WARNING! All tokens must be released in reverse order. This will be
829 * asserted.
830 */
831 void
lwkt_reltoken(lwkt_token_t tok)832 lwkt_reltoken(lwkt_token_t tok)
833 {
834 thread_t td = curthread;
835 lwkt_tokref_t ref;
836
837 /*
838 * Remove ref from thread token list and assert that it matches
839 * the token passed in. Tokens must be released in reverse order.
840 */
841 ref = td->td_toks_stop - 1;
842 if (__predict_false(ref < &td->td_toks_base || ref->tr_tok != tok)) {
843 kprintf("LWKT_RELTOKEN ASSERTION td %p tok %p ref %p/%p\n",
844 td, tok, &td->td_toks_base, ref);
845 kprintf("REF CONTENT: tok=%p count=%016lx owner=%p\n",
846 ref->tr_tok, ref->tr_count, ref->tr_owner);
847 if (ref < &td->td_toks_base) {
848 kprintf("lwkt_reltoken: no tokens to release\n");
849 } else {
850 kprintf("lwkt_reltoken: release wants %s and got %s\n",
851 tok->t_desc, ref->tr_tok->t_desc);
852 }
853 panic("lwkt_reltoken: illegal release");
854 }
855 _lwkt_reltokref(ref, td);
856 cpu_sfence();
857 td->td_toks_stop = ref;
858 }
859
860 /*
861 * It is faster for users of lwkt_getpooltoken() to use the returned
862 * token and just call lwkt_reltoken(), but for convenience we provide
863 * this function which looks the token up based on the ident.
864 */
865 void
lwkt_relpooltoken(void * ptr)866 lwkt_relpooltoken(void *ptr)
867 {
868 lwkt_token_t tok = _lwkt_token_pool_lookup(ptr);
869 lwkt_reltoken(tok);
870 }
871
872 /*
873 * Return a count of the number of token refs the thread has to the
874 * specified token, whether it currently owns the token or not.
875 */
876 int
lwkt_cnttoken(lwkt_token_t tok,thread_t td)877 lwkt_cnttoken(lwkt_token_t tok, thread_t td)
878 {
879 lwkt_tokref_t scan;
880 int count = 0;
881
882 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) {
883 if (scan->tr_tok == tok)
884 ++count;
885 }
886 return(count);
887 }
888
889 /*
890 * Pool tokens are used to provide a type-stable serializing token
891 * pointer that does not race against disappearing data structures.
892 *
893 * This routine is called in early boot just after we setup the BSP's
894 * globaldata structure.
895 */
896 void
lwkt_token_pool_init(void)897 lwkt_token_pool_init(void)
898 {
899 int i;
900
901 for (i = 0; i < LWKT_POOL_TOKENS; ++i)
902 lwkt_token_init(&pool_tokens[i].token, "pool");
903 }
904
905 lwkt_token_t
lwkt_token_pool_lookup(void * ptr)906 lwkt_token_pool_lookup(void *ptr)
907 {
908 return (_lwkt_token_pool_lookup(ptr));
909 }
910
911 /*
912 * Initialize a token.
913 */
914 void
lwkt_token_init(lwkt_token_t tok,const char * desc)915 lwkt_token_init(lwkt_token_t tok, const char *desc)
916 {
917 tok->t_count = 0;
918 tok->t_ref = NULL;
919 tok->t_collisions = 0;
920 tok->t_desc = desc;
921 }
922
923 void
lwkt_token_uninit(lwkt_token_t tok)924 lwkt_token_uninit(lwkt_token_t tok)
925 {
926 /* empty */
927 }
928
929 /*
930 * Exchange the two most recent tokens on the tokref stack. This allows
931 * you to release a token out of order.
932 *
933 * We have to be careful about the case where the top two tokens are
934 * the same token. In this case tok->t_ref will point to the deeper
935 * ref and must remain pointing to the deeper ref. If we were to swap
936 * it the first release would clear the token even though a second
937 * ref is still present.
938 *
939 * Only exclusively held tokens contain a reference to the tokref which
940 * has to be flipped along with the swap.
941 */
942 void
lwkt_token_swap(void)943 lwkt_token_swap(void)
944 {
945 lwkt_tokref_t ref1, ref2;
946 lwkt_token_t tok1, tok2;
947 long count1, count2;
948 thread_t td = curthread;
949
950 crit_enter();
951
952 ref1 = td->td_toks_stop - 1;
953 ref2 = td->td_toks_stop - 2;
954 KKASSERT(ref1 >= &td->td_toks_base);
955 KKASSERT(ref2 >= &td->td_toks_base);
956
957 tok1 = ref1->tr_tok;
958 tok2 = ref2->tr_tok;
959 count1 = ref1->tr_count;
960 count2 = ref2->tr_count;
961
962 if (tok1 != tok2) {
963 ref1->tr_tok = tok2;
964 ref1->tr_count = count2;
965 ref2->tr_tok = tok1;
966 ref2->tr_count = count1;
967 if (tok1->t_ref == ref1)
968 tok1->t_ref = ref2;
969 if (tok2->t_ref == ref2)
970 tok2->t_ref = ref1;
971 }
972
973 crit_exit();
974 }
975
976 #ifdef DDB
DB_SHOW_COMMAND(tokens,db_tok_all)977 DB_SHOW_COMMAND(tokens, db_tok_all)
978 {
979 struct lwkt_token *tok, **ptr;
980 struct lwkt_token *toklist[16] = {
981 &mp_token,
982 &pmap_token,
983 &dev_token,
984 &vm_token,
985 &vmspace_token,
986 &kvm_token,
987 &sigio_token,
988 &tty_token,
989 &vnode_token,
990 NULL
991 };
992
993 ptr = toklist;
994 for (tok = *ptr; tok; tok = *(++ptr)) {
995 db_printf("tok=%p tr_owner=%p t_colissions=%ld t_desc=%s\n", tok,
996 (tok->t_ref ? tok->t_ref->tr_owner : NULL),
997 tok->t_collisions, tok->t_desc);
998 }
999 }
1000 #endif /* DDB */
1001