kernel - All lwkt thread now start out mpsafe part 1/2
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
3b998fa9 2 * Copyright (c) 2003-2010 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
60f60350 10 *
8ad65e08
MD
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
8c10bfcf
MD
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.
60f60350 20 *
8c10bfcf
MD
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
8ad65e08 32 * SUCH DAMAGE.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
b37f18d6 47#include <sys/kinfo.h>
8ad65e08 48#include <sys/queue.h>
7d0bac62 49#include <sys/sysctl.h>
99df837e 50#include <sys/kthread.h>
f1d1c3fa 51#include <machine/cpu.h>
99df837e 52#include <sys/lock.h>
f6bf3af1 53#include <sys/caps.h>
9d265729 54#include <sys/spinlock.h>
57aa743c 55#include <sys/ktr.h>
9d265729
MD
56
57#include <sys/thread2.h>
58#include <sys/spinlock2.h>
684a93c4 59#include <sys/mplock2.h>
f1d1c3fa 60
8c72e3d5
AH
61#include <sys/dsched.h>
62
7d0bac62
MD
63#include <vm/vm.h>
64#include <vm/vm_param.h>
65#include <vm/vm_kern.h>
66#include <vm/vm_object.h>
67#include <vm/vm_page.h>
68#include <vm/vm_map.h>
69#include <vm/vm_pager.h>
70#include <vm/vm_extern.h>
7d0bac62 71
99df837e 72#include <machine/stdarg.h>
96728c05 73#include <machine/smp.h>
99df837e 74
d850923c
AE
75#if !defined(KTR_CTXSW)
76#define KTR_CTXSW KTR_ALL
77#endif
78KTR_INFO_MASTER(ctxsw);
a1f0fb66
AE
79KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
80 sizeof(int) + sizeof(struct thread *));
81KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
82 sizeof(int) + sizeof(struct thread *));
83KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
84 sizeof (struct thread *) + sizeof(char *));
85KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
1541028a 86
40aaf5fc
NT
87static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
88
0f7a3396
MD
89#ifdef INVARIANTS
90static int panic_on_cscount = 0;
91#endif
05220613
MD
92static __int64_t switch_count = 0;
93static __int64_t preempt_hit = 0;
94static __int64_t preempt_miss = 0;
95static __int64_t preempt_weird = 0;
f64b567c 96static __int64_t token_contention_count __debugvar = 0;
fb0f29c4 97static int lwkt_use_spin_port;
40aaf5fc 98static struct objcache *thread_cache;
05220613 99
88ebb169 100#ifdef SMP
e381e77c 101static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
88ebb169 102#endif
f9235b6d 103static void lwkt_fairq_accumulate(globaldata_t gd, thread_t td);
e381e77c 104
0855a2af
JG
105extern void cpu_heavy_restore(void);
106extern void cpu_lwkt_restore(void);
107extern void cpu_kthread_restore(void);
108extern void cpu_idle_restore(void);
109
b2b3ffcd 110#ifdef __x86_64__
85514115
MD
111
112static int
0855a2af
JG
113jg_tos_ok(struct thread *td)
114{
85514115
MD
115 void *tos;
116 int tos_ok;
117
0855a2af
JG
118 if (td == NULL) {
119 return 1;
120 }
121 KKASSERT(td->td_sp != NULL);
85514115
MD
122 tos = ((void **)td->td_sp)[0];
123 tos_ok = 0;
124 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
125 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
0855a2af
JG
126 tos_ok = 1;
127 }
128 return tos_ok;
129}
130
85514115
MD
131#endif
132
fb0f29c4
MD
133/*
134 * We can make all thread ports use the spin backend instead of the thread
135 * backend. This should only be set to debug the spin backend.
136 */
137TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
138
0f7a3396
MD
139#ifdef INVARIANTS
140SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
141#endif
4b5f931b 142SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
9733f757
VS
143SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
144 "Successful preemption events");
145SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
146 "Failed preemption events");
26a0694b 147SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
148#ifdef INVARIANTS
149SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
150 &token_contention_count, 0, "spinning due to token contention");
38717797 151#endif
f9235b6d
MD
152static int fairq_enable = 1;
153SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, &fairq_enable, 0, "");
77912481
MD
154static int user_pri_sched = 0;
155SYSCTL_INT(_lwkt, OID_AUTO, user_pri_sched, CTLFLAG_RW, &user_pri_sched, 0, "");
05220613 156
4b5f931b
MD
157/*
158 * These helper procedures handle the runq, they can only be called from
159 * within a critical section.
75cdbe6c
MD
160 *
161 * WARNING! Prior to SMP being brought up it is possible to enqueue and
162 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
163 * instead of 'mycpu' when referencing the globaldata structure. Once
164 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 165 */
f1d1c3fa
MD
166static __inline
167void
168_lwkt_dequeue(thread_t td)
169{
170 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 171 struct globaldata *gd = td->td_gd;
4b5f931b 172
f1d1c3fa 173 td->td_flags &= ~TDF_RUNQ;
f9235b6d
MD
174 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
175 gd->gd_fairq_total_pri -= td->td_pri;
176 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
177 atomic_clear_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
178 }
179}
180
f9235b6d
MD
181/*
182 * Priority enqueue.
183 *
184 * NOTE: There are a limited number of lwkt threads runnable since user
185 * processes only schedule one at a time per cpu.
186 */
f1d1c3fa
MD
187static __inline
188void
189_lwkt_enqueue(thread_t td)
190{
f9235b6d
MD
191 thread_t xtd;
192
7f5d7ed7 193 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 194 struct globaldata *gd = td->td_gd;
4b5f931b 195
f1d1c3fa 196 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
197 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
198 if (xtd == NULL) {
199 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
200 atomic_set_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
201 } else {
202 while (xtd && xtd->td_pri > td->td_pri)
203 xtd = TAILQ_NEXT(xtd, td_threadq);
204 if (xtd)
205 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
206 else
207 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
208 }
209 gd->gd_fairq_total_pri += td->td_pri;
f1d1c3fa
MD
210 }
211}
8ad65e08 212
40aaf5fc
NT
213static __boolean_t
214_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
215{
216 struct thread *td = (struct thread *)obj;
217
218 td->td_kstack = NULL;
219 td->td_kstack_size = 0;
220 td->td_flags = TDF_ALLOCATED_THREAD;
221 return (1);
222}
223
224static void
225_lwkt_thread_dtor(void *obj, void *privdata)
226{
227 struct thread *td = (struct thread *)obj;
228
229 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
230 ("_lwkt_thread_dtor: not allocated from objcache"));
231 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
232 td->td_kstack_size > 0,
233 ("_lwkt_thread_dtor: corrupted stack"));
234 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
235}
236
237/*
238 * Initialize the lwkt s/system.
239 */
240void
241lwkt_init(void)
242{
243 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
244 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
245 NULL, CACHE_NTHREADS/2,
246 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
247}
248
37af14fe
MD
249/*
250 * Schedule a thread to run. As the current thread we can always safely
251 * schedule ourselves, and a shortcut procedure is provided for that
252 * function.
253 *
254 * (non-blocking, self contained on a per cpu basis)
255 */
256void
257lwkt_schedule_self(thread_t td)
258{
259 crit_enter_quick(td);
f9235b6d
MD
260 KASSERT(td != &td->td_gd->gd_idlethread,
261 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 262 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 263 _lwkt_enqueue(td);
37af14fe
MD
264 crit_exit_quick(td);
265}
266
267/*
268 * Deschedule a thread.
269 *
270 * (non-blocking, self contained on a per cpu basis)
271 */
272void
273lwkt_deschedule_self(thread_t td)
274{
275 crit_enter_quick(td);
37af14fe
MD
276 _lwkt_dequeue(td);
277 crit_exit_quick(td);
278}
279
8ad65e08
MD
280/*
281 * LWKTs operate on a per-cpu basis
282 *
73e4f7b9 283 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
284 */
285void
286lwkt_gdinit(struct globaldata *gd)
287{
f9235b6d 288 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 289 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
290}
291
292/*
7d0bac62 293 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
294 * or LWKT start address before it can be scheduled. If the target cpu is
295 * -1 the thread will be created on the current cpu.
0cfcada1
MD
296 *
297 * If you intend to create a thread without a process context this function
298 * does everything except load the startup and switcher function.
7d0bac62
MD
299 */
300thread_t
d3d32139 301lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 302{
c070746a 303 globaldata_t gd = mycpu;
99df837e 304 void *stack;
7d0bac62 305
c070746a
MD
306 /*
307 * If static thread storage is not supplied allocate a thread. Reuse
308 * a cached free thread if possible. gd_freetd is used to keep an exiting
309 * thread intact through the exit.
310 */
ef0fdad1 311 if (td == NULL) {
c070746a
MD
312 if ((td = gd->gd_freetd) != NULL)
313 gd->gd_freetd = NULL;
314 else
315 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
316 KASSERT((td->td_flags &
317 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
318 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
319 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 320 }
c070746a
MD
321
322 /*
323 * Try to reuse cached stack.
324 */
f470d0c8
MD
325 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
326 if (flags & TDF_ALLOCATED_STACK) {
e4846942 327 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
328 stack = NULL;
329 }
330 }
331 if (stack == NULL) {
e4846942 332 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 333 flags |= TDF_ALLOCATED_STACK;
99df837e 334 }
75cdbe6c 335 if (cpu < 0)
c070746a 336 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 337 else
f470d0c8 338 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 339 return(td);
7d0bac62
MD
340}
341
342/*
343 * Initialize a preexisting thread structure. This function is used by
344 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
345 *
f8c3996b
MD
346 * All threads start out in a critical section at a priority of
347 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
348 * appropriate. This function may send an IPI message when the
349 * requested cpu is not the current cpu and consequently gd_tdallq may
350 * not be initialized synchronously from the point of view of the originating
351 * cpu.
352 *
353 * NOTE! we have to be careful in regards to creating threads for other cpus
354 * if SMP has not yet been activated.
7d0bac62 355 */
41a01a4d
MD
356#ifdef SMP
357
75cdbe6c
MD
358static void
359lwkt_init_thread_remote(void *arg)
360{
361 thread_t td = arg;
362
52eedfb5
MD
363 /*
364 * Protected by critical section held by IPI dispatch
365 */
75cdbe6c
MD
366 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
367}
368
41a01a4d
MD
369#endif
370
7d0bac62 371void
f470d0c8
MD
372lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
373 struct globaldata *gd)
7d0bac62 374{
37af14fe
MD
375 globaldata_t mygd = mycpu;
376
c9e9fb21
MD
377 /* all threads start mpsafe now */
378 KKASSERT(flags & TDF_MPSAFE);
379
99df837e
MD
380 bzero(td, sizeof(struct thread));
381 td->td_kstack = stack;
f470d0c8 382 td->td_kstack_size = stksize;
d3d32139 383 td->td_flags = flags;
26a0694b 384 td->td_gd = gd;
f9235b6d
MD
385 td->td_pri = TDPRI_KERN_DAEMON;
386 td->td_critcount = 1;
3b998fa9 387 td->td_toks_stop = &td->td_toks_base;
d3d32139
MD
388#ifdef SMP
389 if ((flags & TDF_MPSAFE) == 0)
390 td->td_mpcount = 1;
391#endif
fb0f29c4
MD
392 if (lwkt_use_spin_port)
393 lwkt_initport_spin(&td->td_msgport);
394 else
395 lwkt_initport_thread(&td->td_msgport, td);
99df837e 396 pmap_init_thread(td);
0f7a3396 397#ifdef SMP
5d21b981
MD
398 /*
399 * Normally initializing a thread for a remote cpu requires sending an
400 * IPI. However, the idlethread is setup before the other cpus are
401 * activated so we have to treat it as a special case. XXX manipulation
402 * of gd_tdallq requires the BGL.
403 */
404 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 405 crit_enter_gd(mygd);
75cdbe6c 406 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 407 crit_exit_gd(mygd);
75cdbe6c 408 } else {
2db3b277 409 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 410 }
0f7a3396 411#else
37af14fe 412 crit_enter_gd(mygd);
0f7a3396 413 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 414 crit_exit_gd(mygd);
0f7a3396 415#endif
8c72e3d5
AH
416
417 dsched_new_thread(td);
73e4f7b9
MD
418}
419
420void
421lwkt_set_comm(thread_t td, const char *ctl, ...)
422{
e2565a42 423 __va_list va;
73e4f7b9 424
e2565a42 425 __va_start(va, ctl);
379210cb 426 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 427 __va_end(va);
e7c0dbba 428 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
429}
430
99df837e 431void
73e4f7b9 432lwkt_hold(thread_t td)
99df837e 433{
73e4f7b9
MD
434 ++td->td_refs;
435}
436
437void
438lwkt_rele(thread_t td)
439{
440 KKASSERT(td->td_refs > 0);
441 --td->td_refs;
442}
443
444void
445lwkt_wait_free(thread_t td)
446{
447 while (td->td_refs)
377d4740 448 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
449}
450
451void
452lwkt_free_thread(thread_t td)
453{
d9eea1a5 454 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
455 ("lwkt_free_thread: did not exit! %p", td));
456
40aaf5fc
NT
457 if (td->td_flags & TDF_ALLOCATED_THREAD) {
458 objcache_put(thread_cache, td);
459 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
460 /* client-allocated struct with internally allocated stack */
461 KASSERT(td->td_kstack && td->td_kstack_size > 0,
462 ("lwkt_free_thread: corrupted stack"));
463 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
464 td->td_kstack = NULL;
465 td->td_kstack_size = 0;
99df837e 466 }
e7c0dbba 467 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
468}
469
470
7d0bac62 471/*
8ad65e08 472 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
473 * switch to the idlethread. Switching must occur within a critical
474 * section to avoid races with the scheduling queue.
475 *
476 * We always have full control over our cpu's run queue. Other cpus
477 * that wish to manipulate our queue must use the cpu_*msg() calls to
478 * talk to our cpu, so a critical section is all that is needed and
479 * the result is very, very fast thread switching.
480 *
96728c05
MD
481 * The LWKT scheduler uses a fixed priority model and round-robins at
482 * each priority level. User process scheduling is a totally
483 * different beast and LWKT priorities should not be confused with
484 * user process priorities.
f1d1c3fa 485 *
96728c05
MD
486 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
487 * cleans it up. Note that the td_switch() function cannot do anything that
488 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
489 * the target thread (not the current thread). It's nice to have a scheduler
490 * that does not need the MP lock to work because it allows us to do some
491 * really cool high-performance MP lock optimizations.
69d78e99
MD
492 *
493 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
494 * is not called by the current thread in the preemption case, only when
495 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
496 */
497void
498lwkt_switch(void)
499{
37af14fe
MD
500 globaldata_t gd = mycpu;
501 thread_t td = gd->gd_curthread;
8ad65e08 502 thread_t ntd;
f9235b6d
MD
503 thread_t xtd;
504 thread_t nlast;
f9235b6d 505 int nquserok;
6f207a2c 506#ifdef SMP
8a8d5d85
MD
507 int mpheld;
508#endif
f9235b6d 509 int didaccumulate;
b37f18d6
MD
510 const char *lmsg; /* diagnostic - 'systat -pv 1' */
511 const void *laddr;
8ad65e08 512
46a3f46d 513 /*
27e88a6e
MD
514 * Switching from within a 'fast' (non thread switched) interrupt or IPI
515 * is illegal. However, we may have to do it anyway if we hit a fatal
516 * kernel trap or we have paniced.
517 *
518 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 519 */
27e88a6e
MD
520 if (gd->gd_intr_nesting_level) {
521 int savegdnest;
522 int savegdtrap;
523
524 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
525 panic("lwkt_switch: cannot switch from within "
526 "a fast interrupt, yet, td %p\n", td);
527 } else {
528 savegdnest = gd->gd_intr_nesting_level;
529 savegdtrap = gd->gd_trap_nesting_level;
530 gd->gd_intr_nesting_level = 0;
531 gd->gd_trap_nesting_level = 0;
a7422615
MD
532 if ((td->td_flags & TDF_PANICWARN) == 0) {
533 td->td_flags |= TDF_PANICWARN;
6ea70f76 534 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615 535 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 536 print_backtrace(-1);
a7422615 537 }
27e88a6e
MD
538 lwkt_switch();
539 gd->gd_intr_nesting_level = savegdnest;
540 gd->gd_trap_nesting_level = savegdtrap;
541 return;
542 }
96728c05 543 }
ef0fdad1 544
cb973d15
MD
545 /*
546 * Passive release (used to transition from user to kernel mode
547 * when we block or switch rather then when we enter the kernel).
548 * This function is NOT called if we are switching into a preemption
549 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
550 * our current process designation (if we have one) and become a true
551 * LWKT thread, and may also hand the current process designation to
552 * another process and schedule thread.
cb973d15
MD
553 */
554 if (td->td_release)
555 td->td_release(td);
556
37af14fe 557 crit_enter_gd(gd);
3b998fa9 558 if (TD_TOKS_HELD(td))
9d265729
MD
559 lwkt_relalltokens(td);
560
561 /*
b02926de
MD
562 * We had better not be holding any spin locks, but don't get into an
563 * endless panic loop.
9d265729 564 */
bbb31c5d
MD
565 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
566 ("lwkt_switch: still holding a shared spinlock %p!",
567 gd->gd_spinlock_rd));
d666840a
MD
568 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
569 ("lwkt_switch: still holding %d exclusive spinlocks!",
570 gd->gd_spinlocks_wr));
9d265729 571
8a8d5d85
MD
572
573#ifdef SMP
574 /*
575 * td_mpcount cannot be used to determine if we currently hold the
576 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
577 * to get the lock, and switch out if it can't. Our ownership of
578 * the actual lock will remain stable while we are in a critical section
579 * (but, of course, another cpu may own or release the lock so the
580 * actual value of mp_lock is not stable).
8a8d5d85 581 */
c5724852 582 mpheld = MP_LOCK_HELD(gd);
0f7a3396
MD
583#ifdef INVARIANTS
584 if (td->td_cscount) {
6ea70f76 585 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
586 td);
587 if (panic_on_cscount)
588 panic("switching while mastering cpusync");
589 }
590#endif
8a8d5d85 591#endif
f9235b6d
MD
592
593 /*
594 * If we had preempted another thread on this cpu, resume the preempted
595 * thread. This occurs transparently, whether the preempted thread
596 * was scheduled or not (it may have been preempted after descheduling
597 * itself).
598 *
599 * We have to setup the MP lock for the original thread after backing
600 * out the adjustment that was made to curthread when the original
601 * was preempted.
602 */
99df837e 603 if ((ntd = td->td_preempted) != NULL) {
26a0694b 604 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 605#ifdef SMP
96728c05 606 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 607 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
608 td, ntd, td->td_mpcount, ntd->td_mpcount);
609 }
8a8d5d85
MD
610 if (ntd->td_mpcount) {
611 td->td_mpcount -= ntd->td_mpcount;
612 KKASSERT(td->td_mpcount >= 0);
613 }
614#endif
26a0694b 615 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
616
617 /*
b9eb1c19
MD
618 * The interrupt may have woken a thread up, we need to properly
619 * set the reschedule flag if the originally interrupted thread is
620 * at a lower priority.
8ec60c3f 621 */
f9235b6d
MD
622 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
623 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
8ec60c3f 624 need_lwkt_resched();
f9235b6d 625 }
8a8d5d85 626 /* YYY release mp lock on switchback if original doesn't need it */
f9235b6d
MD
627 goto havethread_preempted;
628 }
629
630 /*
631 * Implement round-robin fairq with priority insertion. The priority
632 * insertion is handled by _lwkt_enqueue()
633 *
634 * We have to adjust the MP lock for the target thread. If we
635 * need the MP lock and cannot obtain it we try to locate a
636 * thread that does not need the MP lock. If we cannot, we spin
637 * instead of HLT.
638 *
639 * A similar issue exists for the tokens held by the target thread.
640 * If we cannot obtain ownership of the tokens we cannot immediately
641 * schedule the thread.
642 */
643 for (;;) {
644 clear_lwkt_resched();
645 didaccumulate = 0;
646 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
647
4b5f931b 648 /*
f9235b6d 649 * Hotpath if we can get all necessary resources.
41a01a4d 650 *
f9235b6d 651 * If nothing is runnable switch to the idle thread
41a01a4d 652 */
f9235b6d
MD
653 if (ntd == NULL) {
654 ntd = &gd->gd_idlethread;
655 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
656 ntd->td_flags |= TDF_IDLE_NOHLT;
6f207a2c 657#ifdef SMP
f9235b6d
MD
658 if (ntd->td_mpcount) {
659 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
660 panic("Idle thread %p was holding the BGL!", ntd);
661 if (mpheld == 0) {
c5724852
MD
662 set_cpu_contention_mask(gd);
663 handle_cpu_contention_mask();
664 cpu_try_mplock();
665 mpheld = MP_LOCK_HELD(gd);
f9235b6d
MD
666 cpu_pause();
667 continue;
668 }
669 }
c5724852 670 clr_cpu_contention_mask(gd);
6f207a2c 671#endif
b37f18d6
MD
672 cpu_time.cp_msg[0] = 0;
673 cpu_time.cp_stallpc = 0;
f9235b6d
MD
674 goto haveidle;
675 }
41a01a4d
MD
676
677 /*
f9235b6d 678 * Hotpath schedule
6f207a2c
MD
679 *
680 * NOTE: For UP there is no mplock and lwkt_getalltokens()
681 * always succeeds.
8ec60c3f 682 */
f9235b6d
MD
683 if (ntd->td_fairq_accum >= 0 &&
684#ifdef SMP
685 (ntd->td_mpcount == 0 || mpheld || cpu_try_mplock()) &&
686#endif
b37f18d6 687 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
f9235b6d 688 ) {
8a8d5d85 689#ifdef SMP
c5724852 690 clr_cpu_contention_mask(gd);
f9235b6d
MD
691#endif
692 goto havethread;
693 }
694
b37f18d6
MD
695 lmsg = NULL;
696 laddr = NULL;
697
f9235b6d 698#ifdef SMP
c5724852
MD
699 if (ntd->td_fairq_accum >= 0)
700 set_cpu_contention_mask(gd);
f9235b6d 701 /* Reload mpheld (it become stale after mplock/token ops) */
c5724852 702 mpheld = MP_LOCK_HELD(gd);
b37f18d6
MD
703 if (ntd->td_mpcount && mpheld == 0) {
704 lmsg = "mplock";
705 laddr = ntd->td_mplock_stallpc;
706 }
f9235b6d
MD
707#endif
708
709 /*
710 * Coldpath - unable to schedule ntd, continue looking for threads
711 * to schedule. This is only allowed of the (presumably) kernel
712 * thread exhausted its fair share. A kernel thread stuck on
713 * resources does not currently allow a user thread to get in
714 * front of it.
715 */
716#ifdef SMP
717 nquserok = ((ntd->td_pri < TDPRI_KERN_LPSCHED) ||
718 (ntd->td_fairq_accum < 0));
6f207a2c
MD
719#else
720 nquserok = 1;
f9235b6d
MD
721#endif
722 nlast = NULL;
723
724 for (;;) {
41a01a4d 725 /*
f9235b6d
MD
726 * If the fair-share scheduler ran out ntd gets moved to the
727 * end and its accumulator will be bumped, if it didn't we
728 * maintain the same queue position.
df6b8ba0 729 *
f9235b6d 730 * nlast keeps track of the last element prior to any moves.
41a01a4d 731 */
f9235b6d 732 if (ntd->td_fairq_accum < 0) {
f9235b6d
MD
733 lwkt_fairq_accumulate(gd, ntd);
734 didaccumulate = 1;
c5724852
MD
735
736 /*
737 * Move to end
738 */
739 xtd = TAILQ_NEXT(ntd, td_threadq);
f9235b6d
MD
740 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
741 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
c5724852
MD
742
743 /*
744 * Set terminal element (nlast)
745 */
f9235b6d
MD
746 if (nlast == NULL) {
747 nlast = ntd;
748 if (xtd == NULL)
749 xtd = ntd;
750 }
751 ntd = xtd;
752 } else {
753 ntd = TAILQ_NEXT(ntd, td_threadq);
754 }
a453459d 755
f9235b6d
MD
756 /*
757 * If we exhausted the run list switch to the idle thread.
758 * Since one or more threads had resource acquisition issues
759 * we do not allow the idle thread to halt.
760 *
761 * NOTE: nlast can be NULL.
762 */
763 if (ntd == nlast) {
e0a90d3b 764 cpu_pause();
f9235b6d
MD
765 ntd = &gd->gd_idlethread;
766 ntd->td_flags |= TDF_IDLE_NOHLT;
6f207a2c 767#ifdef SMP
f9235b6d 768 if (ntd->td_mpcount) {
c5724852 769 mpheld = MP_LOCK_HELD(gd);
f9235b6d
MD
770 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
771 panic("Idle thread %p was holding the BGL!", ntd);
772 if (mpheld == 0) {
c5724852
MD
773 set_cpu_contention_mask(gd);
774 handle_cpu_contention_mask();
775 cpu_try_mplock();
776 mpheld = MP_LOCK_HELD(gd);
f9235b6d
MD
777 cpu_pause();
778 break; /* try again from the top, almost */
b9eb1c19 779 }
8a8d5d85 780 }
6f207a2c 781#endif
684a93c4
MD
782
783 /*
f9235b6d
MD
784 * If fairq accumulations occured we do not schedule the
785 * idle thread. This will cause us to try again from
786 * the (almost) top.
684a93c4 787 */
f9235b6d 788 if (didaccumulate)
b37f18d6
MD
789 break; /* try again from the top, almost */
790 if (lmsg)
791 strlcpy(cpu_time.cp_msg, lmsg, sizeof(cpu_time.cp_msg));
792 cpu_time.cp_stallpc = (uintptr_t)laddr;
f9235b6d 793 goto haveidle;
8a8d5d85 794 }
f9235b6d 795
df6b8ba0 796 /*
f9235b6d 797 * Try to switch to this thread.
6f207a2c
MD
798 *
799 * NOTE: For UP there is no mplock and lwkt_getalltokens()
800 * always succeeds.
df6b8ba0 801 */
77912481
MD
802 if ((ntd->td_pri >= TDPRI_KERN_LPSCHED || nquserok ||
803 user_pri_sched) && ntd->td_fairq_accum >= 0 &&
f9235b6d
MD
804#ifdef SMP
805 (ntd->td_mpcount == 0 || mpheld || cpu_try_mplock()) &&
8a8d5d85 806#endif
b37f18d6 807 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
f9235b6d 808 ) {
a453459d 809#ifdef SMP
c5724852 810 clr_cpu_contention_mask(gd);
f9235b6d
MD
811#endif
812 goto havethread;
df6b8ba0 813 }
f9235b6d 814#ifdef SMP
c5724852
MD
815 if (ntd->td_fairq_accum >= 0)
816 set_cpu_contention_mask(gd);
817 /*
818 * Reload mpheld (it become stale after mplock/token ops).
819 */
820 mpheld = MP_LOCK_HELD(gd);
b37f18d6
MD
821 if (ntd->td_mpcount && mpheld == 0) {
822 lmsg = "mplock";
823 laddr = ntd->td_mplock_stallpc;
824 }
f9235b6d
MD
825 if (ntd->td_pri >= TDPRI_KERN_LPSCHED && ntd->td_fairq_accum >= 0)
826 nquserok = 0;
a453459d 827#endif
4b5f931b 828 }
c5724852
MD
829
830 /*
831 * All threads exhausted but we can loop due to a negative
832 * accumulator.
833 *
834 * While we are looping in the scheduler be sure to service
835 * any interrupts which were made pending due to our critical
836 * section, otherwise we could livelock (e.g.) IPIs.
837 *
838 * NOTE: splz can enter and exit the mplock so mpheld is
839 * stale after this call.
840 */
841 splz_check();
842
843#ifdef SMP
844 /*
845 * Our mplock can be cached and cause other cpus to livelock
846 * if we loop due to e.g. not being able to acquire tokens.
847 */
848 if (MP_LOCK_HELD(gd))
849 cpu_rel_mplock(gd->gd_cpuid);
850 mpheld = 0;
851#endif
f1d1c3fa 852 }
8a8d5d85
MD
853
854 /*
f9235b6d
MD
855 * Do the actual switch. WARNING: mpheld is stale here.
856 *
857 * We must always decrement td_fairq_accum on non-idle threads just
858 * in case a thread never gets a tick due to being in a continuous
859 * critical section. The page-zeroing code does that.
860 *
861 * If the thread we came up with is a higher or equal priority verses
862 * the thread at the head of the queue we move our thread to the
863 * front. This way we can always check the front of the queue.
864 */
865havethread:
866 ++gd->gd_cnt.v_swtch;
867 --ntd->td_fairq_accum;
868 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
869 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
870 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
871 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
872 }
873havethread_preempted:
874 ;
875 /*
876 * If the new target does not need the MP lock and we are holding it,
877 * release the MP lock. If the new target requires the MP lock we have
878 * already acquired it for the target.
879 *
880 * WARNING: mpheld is stale here.
8a8d5d85 881 */
f9235b6d
MD
882haveidle:
883 KASSERT(ntd->td_critcount,
884 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
885#ifdef SMP
886 if (ntd->td_mpcount == 0 ) {
c5724852
MD
887 if (MP_LOCK_HELD(gd))
888 cpu_rel_mplock(gd->gd_cpuid);
8a8d5d85 889 } else {
a453459d 890 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
891 }
892#endif
94f6d86e
MD
893 if (td != ntd) {
894 ++switch_count;
b2b3ffcd 895#ifdef __x86_64__
f9235b6d
MD
896 {
897 int tos_ok __debugvar = jg_tos_ok(ntd);
898 KKASSERT(tos_ok);
899 }
85514115 900#endif
a1f0fb66 901 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
f1d1c3fa 902 td->td_switch(ntd);
94f6d86e 903 }
37af14fe
MD
904 /* NOTE: current cpu may have changed after switch */
905 crit_exit_quick(td);
8ad65e08
MD
906}
907
f1d1c3fa 908/*
96728c05
MD
909 * Request that the target thread preempt the current thread. Preemption
910 * only works under a specific set of conditions:
b68b7282 911 *
96728c05
MD
912 * - We are not preempting ourselves
913 * - The target thread is owned by the current cpu
914 * - We are not currently being preempted
915 * - The target is not currently being preempted
d3d1cbc8
MD
916 * - We are not holding any spin locks
917 * - The target thread is not holding any tokens
96728c05
MD
918 * - We are able to satisfy the target's MP lock requirements (if any).
919 *
920 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
921 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 922 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
923 * to determine whether preemption is possible (aka usually just the crit
924 * priority of lwkt_schedule() itself).
b68b7282 925 *
26a0694b
MD
926 * XXX at the moment we run the target thread in a critical section during
927 * the preemption in order to prevent the target from taking interrupts
928 * that *WE* can't. Preemption is strictly limited to interrupt threads
929 * and interrupt-like threads, outside of a critical section, and the
930 * preempted source thread will be resumed the instant the target blocks
931 * whether or not the source is scheduled (i.e. preemption is supposed to
932 * be as transparent as possible).
4b5f931b 933 *
8a8d5d85
MD
934 * The target thread inherits our MP count (added to its own) for the
935 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
936 * MP lock during the preemption. Therefore, any preempting targets must be
937 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
938 * out of sync with the physical mp_lock, but we do not have to preserve
939 * the original ownership of the lock if it was out of synch (that is, we
940 * can leave it synchronized on return).
b68b7282
MD
941 */
942void
f9235b6d 943lwkt_preempt(thread_t ntd, int critcount)
b68b7282 944{
46a3f46d 945 struct globaldata *gd = mycpu;
0a3f9b47 946 thread_t td;
8a8d5d85
MD
947#ifdef SMP
948 int mpheld;
57c254db 949 int savecnt;
8a8d5d85 950#endif
b68b7282 951
26a0694b 952 /*
96728c05
MD
953 * The caller has put us in a critical section. We can only preempt
954 * if the caller of the caller was not in a critical section (basically
f9235b6d 955 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 956 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 957 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
958 *
959 * YYY The target thread must be in a critical section (else it must
960 * inherit our critical section? I dunno yet).
41a01a4d 961 *
0a3f9b47 962 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b 963 */
f9235b6d 964 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 965
0a3f9b47 966 td = gd->gd_curthread;
f9235b6d 967 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
968 ++preempt_miss;
969 return;
970 }
f9235b6d 971 if (td->td_critcount > critcount) {
96728c05 972 ++preempt_miss;
8ec60c3f 973 need_lwkt_resched();
96728c05
MD
974 return;
975 }
976#ifdef SMP
46a3f46d 977 if (ntd->td_gd != gd) {
96728c05 978 ++preempt_miss;
8ec60c3f 979 need_lwkt_resched();
96728c05
MD
980 return;
981 }
982#endif
41a01a4d 983 /*
77912481
MD
984 * We don't have to check spinlocks here as they will also bump
985 * td_critcount.
d3d1cbc8
MD
986 *
987 * Do not try to preempt if the target thread is holding any tokens.
988 * We could try to acquire the tokens but this case is so rare there
989 * is no need to support it.
41a01a4d 990 */
77912481
MD
991 KKASSERT(gd->gd_spinlock_rd == NULL);
992 KKASSERT(gd->gd_spinlocks_wr == 0);
993
3b998fa9 994 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8
MD
995 ++preempt_miss;
996 need_lwkt_resched();
997 return;
998 }
26a0694b
MD
999 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1000 ++preempt_weird;
8ec60c3f 1001 need_lwkt_resched();
26a0694b
MD
1002 return;
1003 }
1004 if (ntd->td_preempted) {
4b5f931b 1005 ++preempt_hit;
8ec60c3f 1006 need_lwkt_resched();
26a0694b 1007 return;
b68b7282 1008 }
8a8d5d85 1009#ifdef SMP
a2a5ad0d
MD
1010 /*
1011 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
1012 * to or from the MP lock. In this case td_mpcount will be pre-disposed
1013 * (non-zero) but not actually synchronized with the actual state of the
1014 * lock. We can use it to imply an MP lock requirement for the
1015 * preemption but we cannot use it to test whether we hold the MP lock
1016 * or not.
a2a5ad0d 1017 */
96728c05 1018 savecnt = td->td_mpcount;
c5724852 1019 mpheld = MP_LOCK_HELD(gd);
8a8d5d85
MD
1020 ntd->td_mpcount += td->td_mpcount;
1021 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
1022 ntd->td_mpcount -= td->td_mpcount;
1023 ++preempt_miss;
8ec60c3f 1024 need_lwkt_resched();
8a8d5d85
MD
1025 return;
1026 }
1027#endif
26a0694b 1028
8ec60c3f
MD
1029 /*
1030 * Since we are able to preempt the current thread, there is no need to
1031 * call need_lwkt_resched().
1032 */
26a0694b
MD
1033 ++preempt_hit;
1034 ntd->td_preempted = td;
1035 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1036 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
26a0694b 1037 td->td_switch(ntd);
b9eb1c19 1038
26a0694b 1039 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
1040#ifdef SMP
1041 KKASSERT(savecnt == td->td_mpcount);
c5724852 1042 mpheld = MP_LOCK_HELD(gd);
71ef2f5c 1043 if (mpheld && td->td_mpcount == 0)
c5724852 1044 cpu_rel_mplock(gd->gd_cpuid);
71ef2f5c 1045 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
1046 panic("lwkt_preempt(): MP lock was not held through");
1047#endif
26a0694b
MD
1048 ntd->td_preempted = NULL;
1049 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1050}
1051
1052/*
faaeffac 1053 * Conditionally call splz() if gd_reqflags indicates work is pending.
f1d1c3fa 1054 *
faaeffac
MD
1055 * td_nest_count prevents deep nesting via splz() or doreti() which
1056 * might otherwise blow out the kernel stack. Note that except for
1057 * this special case, we MUST call splz() here to handle any
1058 * pending ints, particularly after we switch, or we might accidently
1059 * halt the cpu with interrupts pending.
ef0fdad1 1060 *
f1d1c3fa
MD
1061 * (self contained on a per cpu basis)
1062 */
1063void
faaeffac 1064splz_check(void)
f1d1c3fa 1065{
7966cb69
MD
1066 globaldata_t gd = mycpu;
1067 thread_t td = gd->gd_curthread;
ef0fdad1 1068
f9235b6d 1069 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
f1d1c3fa 1070 splz();
f1d1c3fa
MD
1071}
1072
8ad65e08 1073/*
f9235b6d
MD
1074 * This function is used to negotiate a passive release of the current
1075 * process/lwp designation with the user scheduler, allowing the user
1076 * scheduler to schedule another user thread. The related kernel thread
1077 * (curthread) continues running in the released state.
8ad65e08
MD
1078 */
1079void
f9235b6d 1080lwkt_passive_release(struct thread *td)
8ad65e08 1081{
f9235b6d
MD
1082 struct lwp *lp = td->td_lwp;
1083
1084 td->td_release = NULL;
1085 lwkt_setpri_self(TDPRI_KERN_USER);
1086 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1087}
1088
f9235b6d 1089
f1d1c3fa 1090/*
f9235b6d
MD
1091 * This implements a normal yield. This routine is virtually a nop if
1092 * there is nothing to yield to but it will always run any pending interrupts
1093 * if called from a critical section.
1094 *
1095 * This yield is designed for kernel threads without a user context.
1096 *
1097 * (self contained on a per cpu basis)
3824f392
MD
1098 */
1099void
f9235b6d 1100lwkt_yield(void)
3824f392 1101{
f9235b6d
MD
1102 globaldata_t gd = mycpu;
1103 thread_t td = gd->gd_curthread;
1104 thread_t xtd;
3824f392 1105
f9235b6d
MD
1106 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1107 splz();
1108 if (td->td_fairq_accum < 0) {
1109 lwkt_schedule_self(curthread);
1110 lwkt_switch();
1111 } else {
1112 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1113 if (xtd && xtd->td_pri > td->td_pri) {
1114 lwkt_schedule_self(curthread);
1115 lwkt_switch();
1116 }
1117 }
3824f392
MD
1118}
1119
1120/*
f9235b6d
MD
1121 * This yield is designed for kernel threads with a user context.
1122 *
1123 * The kernel acting on behalf of the user is potentially cpu-bound,
1124 * this function will efficiently allow other threads to run and also
1125 * switch to other processes by releasing.
3824f392
MD
1126 *
1127 * The lwkt_user_yield() function is designed to have very low overhead
1128 * if no yield is determined to be needed.
1129 */
1130void
1131lwkt_user_yield(void)
1132{
f9235b6d
MD
1133 globaldata_t gd = mycpu;
1134 thread_t td = gd->gd_curthread;
1135
1136 /*
1137 * Always run any pending interrupts in case we are in a critical
1138 * section.
1139 */
1140 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1141 splz();
3824f392
MD
1142
1143#ifdef SMP
1144 /*
1145 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1146 * kernel can prevent other cpus from servicing interrupt threads
1147 * which still require the MP lock (which is a lot of them). This
1148 * has a chaining effect since if the interrupt is blocked, so is
1149 * the event, so normal scheduling will not pick up on the problem.
1150 */
c5724852 1151 if (cpu_contention_mask && td->td_mpcount) {
684a93c4 1152 yield_mplock(td);
3824f392
MD
1153 }
1154#endif
1155
1156 /*
f9235b6d
MD
1157 * Switch (which forces a release) if another kernel thread needs
1158 * the cpu, if userland wants us to resched, or if our kernel
1159 * quantum has run out.
3824f392 1160 */
f9235b6d
MD
1161 if (lwkt_resched_wanted() ||
1162 user_resched_wanted() ||
1163 td->td_fairq_accum < 0)
1164 {
3824f392 1165 lwkt_switch();
3824f392
MD
1166 }
1167
f9235b6d 1168#if 0
3824f392 1169 /*
f9235b6d
MD
1170 * Reacquire the current process if we are released.
1171 *
1172 * XXX not implemented atm. The kernel may be holding locks and such,
1173 * so we want the thread to continue to receive cpu.
3824f392 1174 */
f9235b6d
MD
1175 if (td->td_release == NULL && lp) {
1176 lp->lwp_proc->p_usched->acquire_curproc(lp);
1177 td->td_release = lwkt_passive_release;
1178 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1179 }
f9235b6d 1180#endif
b9eb1c19
MD
1181}
1182
1183/*
f1d1c3fa
MD
1184 * Generic schedule. Possibly schedule threads belonging to other cpus and
1185 * deal with threads that might be blocked on a wait queue.
1186 *
0a3f9b47
MD
1187 * We have a little helper inline function which does additional work after
1188 * the thread has been enqueued, including dealing with preemption and
1189 * setting need_lwkt_resched() (which prevents the kernel from returning
1190 * to userland until it has processed higher priority threads).
6330a558
MD
1191 *
1192 * It is possible for this routine to be called after a failed _enqueue
1193 * (due to the target thread migrating, sleeping, or otherwise blocked).
1194 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1195 *
1196 * reschedok is an optimized constant propagated from lwkt_schedule() or
1197 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1198 * reschedule to be requested if the target thread has a higher priority.
1199 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1200 * be 0, prevented undesired reschedules.
8ad65e08 1201 */
0a3f9b47
MD
1202static __inline
1203void
f9235b6d 1204_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
0a3f9b47 1205{
b9eb1c19 1206 thread_t otd;
c730be20 1207
6330a558 1208 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1209 if (ntd->td_preemptable && reschedok) {
f9235b6d 1210 ntd->td_preemptable(ntd, ccount); /* YYY +token */
361d01dd 1211 } else if (reschedok) {
b9eb1c19 1212 otd = curthread;
f9235b6d 1213 if (ntd->td_pri > otd->td_pri)
c730be20 1214 need_lwkt_resched();
6330a558 1215 }
f9235b6d
MD
1216
1217 /*
1218 * Give the thread a little fair share scheduler bump if it
1219 * has been asleep for a while. This is primarily to avoid
1220 * a degenerate case for interrupt threads where accumulator
1221 * crosses into negative territory unnecessarily.
1222 */
1223 if (ntd->td_fairq_lticks != ticks) {
1224 ntd->td_fairq_lticks = ticks;
1225 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1226 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1227 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1228 }
0a3f9b47
MD
1229 }
1230}
1231
361d01dd 1232static __inline
8ad65e08 1233void
361d01dd 1234_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1235{
37af14fe
MD
1236 globaldata_t mygd = mycpu;
1237
41a01a4d 1238 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1239 crit_enter_gd(mygd);
9388413d 1240 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1241 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1242 _lwkt_enqueue(td);
1243 } else {
f1d1c3fa 1244 /*
7cd8d145
MD
1245 * If we own the thread, there is no race (since we are in a
1246 * critical section). If we do not own the thread there might
1247 * be a race but the target cpu will deal with it.
f1d1c3fa 1248 */
0f7a3396 1249#ifdef SMP
7cd8d145 1250 if (td->td_gd == mygd) {
9d265729 1251 _lwkt_enqueue(td);
f9235b6d 1252 _lwkt_schedule_post(mygd, td, 1, reschedok);
f1d1c3fa 1253 } else {
e381e77c 1254 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1255 }
0f7a3396 1256#else
7cd8d145 1257 _lwkt_enqueue(td);
f9235b6d 1258 _lwkt_schedule_post(mygd, td, 1, reschedok);
0f7a3396 1259#endif
8ad65e08 1260 }
37af14fe 1261 crit_exit_gd(mygd);
8ad65e08
MD
1262}
1263
361d01dd
MD
1264void
1265lwkt_schedule(thread_t td)
1266{
1267 _lwkt_schedule(td, 1);
1268}
1269
1270void
1271lwkt_schedule_noresched(thread_t td)
1272{
1273 _lwkt_schedule(td, 0);
1274}
1275
88ebb169
SW
1276#ifdef SMP
1277
e381e77c
MD
1278/*
1279 * When scheduled remotely if frame != NULL the IPIQ is being
1280 * run via doreti or an interrupt then preemption can be allowed.
1281 *
1282 * To allow preemption we have to drop the critical section so only
1283 * one is present in _lwkt_schedule_post.
1284 */
1285static void
1286lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1287{
1288 thread_t td = curthread;
1289 thread_t ntd = arg;
1290
1291 if (frame && ntd->td_preemptable) {
1292 crit_exit_noyield(td);
1293 _lwkt_schedule(ntd, 1);
1294 crit_enter_quick(td);
1295 } else {
1296 _lwkt_schedule(ntd, 1);
1297 }
1298}
1299
d9eea1a5 1300/*
52eedfb5
MD
1301 * Thread migration using a 'Pull' method. The thread may or may not be
1302 * the current thread. It MUST be descheduled and in a stable state.
1303 * lwkt_giveaway() must be called on the cpu owning the thread.
1304 *
1305 * At any point after lwkt_giveaway() is called, the target cpu may
1306 * 'pull' the thread by calling lwkt_acquire().
1307 *
ae8e83e6
MD
1308 * We have to make sure the thread is not sitting on a per-cpu tsleep
1309 * queue or it will blow up when it moves to another cpu.
1310 *
52eedfb5 1311 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1312 */
a2a5ad0d 1313void
52eedfb5
MD
1314lwkt_giveaway(thread_t td)
1315{
3b4192fb 1316 globaldata_t gd = mycpu;
52eedfb5 1317
3b4192fb
MD
1318 crit_enter_gd(gd);
1319 if (td->td_flags & TDF_TSLEEPQ)
1320 tsleep_remove(td);
1321 KKASSERT(td->td_gd == gd);
1322 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1323 td->td_flags |= TDF_MIGRATING;
1324 crit_exit_gd(gd);
52eedfb5
MD
1325}
1326
1327void
a2a5ad0d
MD
1328lwkt_acquire(thread_t td)
1329{
37af14fe
MD
1330 globaldata_t gd;
1331 globaldata_t mygd;
a2a5ad0d 1332
52eedfb5 1333 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1334 gd = td->td_gd;
37af14fe 1335 mygd = mycpu;
52eedfb5 1336 if (gd != mycpu) {
35238fa5 1337 cpu_lfence();
52eedfb5 1338 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1339 crit_enter_gd(mygd);
df910c23
MD
1340 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1341#ifdef SMP
1342 lwkt_process_ipiq();
1343#endif
52eedfb5 1344 cpu_lfence();
df910c23 1345 }
37af14fe 1346 td->td_gd = mygd;
52eedfb5
MD
1347 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1348 td->td_flags &= ~TDF_MIGRATING;
1349 crit_exit_gd(mygd);
1350 } else {
1351 crit_enter_gd(mygd);
1352 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1353 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1354 crit_exit_gd(mygd);
a2a5ad0d
MD
1355 }
1356}
1357
52eedfb5
MD
1358#endif
1359
8ad65e08 1360/*
f1d1c3fa
MD
1361 * Generic deschedule. Descheduling threads other then your own should be
1362 * done only in carefully controlled circumstances. Descheduling is
1363 * asynchronous.
1364 *
1365 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1366 */
1367void
1368lwkt_deschedule(thread_t td)
1369{
f1d1c3fa 1370 crit_enter();
b8a98473 1371#ifdef SMP
f1d1c3fa
MD
1372 if (td == curthread) {
1373 _lwkt_dequeue(td);
1374 } else {
a72187e9 1375 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1376 _lwkt_dequeue(td);
1377 } else {
b8a98473 1378 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1379 }
1380 }
b8a98473
MD
1381#else
1382 _lwkt_dequeue(td);
1383#endif
f1d1c3fa
MD
1384 crit_exit();
1385}
1386
1387/*
4b5f931b
MD
1388 * Set the target thread's priority. This routine does not automatically
1389 * switch to a higher priority thread, LWKT threads are not designed for
1390 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1391 */
1392void
1393lwkt_setpri(thread_t td, int pri)
1394{
a72187e9 1395 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1396 if (td->td_pri != pri) {
1397 KKASSERT(pri >= 0);
1398 crit_enter();
1399 if (td->td_flags & TDF_RUNQ) {
1400 _lwkt_dequeue(td);
1401 td->td_pri = pri;
1402 _lwkt_enqueue(td);
1403 } else {
1404 td->td_pri = pri;
1405 }
1406 crit_exit();
26a0694b 1407 }
26a0694b
MD
1408}
1409
03bd0a5e
MD
1410/*
1411 * Set the initial priority for a thread prior to it being scheduled for
1412 * the first time. The thread MUST NOT be scheduled before or during
1413 * this call. The thread may be assigned to a cpu other then the current
1414 * cpu.
1415 *
1416 * Typically used after a thread has been created with TDF_STOPPREQ,
1417 * and before the thread is initially scheduled.
1418 */
1419void
1420lwkt_setpri_initial(thread_t td, int pri)
1421{
1422 KKASSERT(pri >= 0);
1423 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1424 td->td_pri = pri;
03bd0a5e
MD
1425}
1426
26a0694b
MD
1427void
1428lwkt_setpri_self(int pri)
1429{
1430 thread_t td = curthread;
1431
4b5f931b
MD
1432 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1433 crit_enter();
1434 if (td->td_flags & TDF_RUNQ) {
1435 _lwkt_dequeue(td);
f9235b6d 1436 td->td_pri = pri;
4b5f931b
MD
1437 _lwkt_enqueue(td);
1438 } else {
f9235b6d 1439 td->td_pri = pri;
4b5f931b
MD
1440 }
1441 crit_exit();
1442}
1443
5d21b981 1444/*
f9235b6d
MD
1445 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1446 *
1447 * Example: two competing threads, same priority N. decrement by (2*N)
1448 * increment by N*8, each thread will get 4 ticks.
1449 */
1450void
1451lwkt_fairq_schedulerclock(thread_t td)
1452{
1453 if (fairq_enable) {
1454 while (td) {
1455 if (td != &td->td_gd->gd_idlethread) {
1456 td->td_fairq_accum -= td->td_gd->gd_fairq_total_pri;
1457 if (td->td_fairq_accum < -TDFAIRQ_MAX(td->td_gd))
1458 td->td_fairq_accum = -TDFAIRQ_MAX(td->td_gd);
1459 if (td->td_fairq_accum < 0)
1460 need_lwkt_resched();
1461 td->td_fairq_lticks = ticks;
1462 }
1463 td = td->td_preempted;
1464 }
1465 }
1466}
1467
1468static void
1469lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1470{
1471 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1472 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1473 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1474}
1475
1476/*
52eedfb5
MD
1477 * Migrate the current thread to the specified cpu.
1478 *
1479 * This is accomplished by descheduling ourselves from the current cpu,
1480 * moving our thread to the tdallq of the target cpu, IPI messaging the
1481 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1482 * races while the thread is being migrated.
ae8e83e6
MD
1483 *
1484 * We must be sure to remove ourselves from the current cpu's tsleepq
1485 * before potentially moving to another queue. The thread can be on
1486 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1487 */
3d28ff59 1488#ifdef SMP
5d21b981 1489static void lwkt_setcpu_remote(void *arg);
3d28ff59 1490#endif
5d21b981
MD
1491
1492void
1493lwkt_setcpu_self(globaldata_t rgd)
1494{
1495#ifdef SMP
1496 thread_t td = curthread;
1497
1498 if (td->td_gd != rgd) {
1499 crit_enter_quick(td);
ae8e83e6 1500 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1501 tsleep_remove(td);
5d21b981
MD
1502 td->td_flags |= TDF_MIGRATING;
1503 lwkt_deschedule_self(td);
52eedfb5 1504 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1505 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1506 lwkt_switch();
1507 /* we are now on the target cpu */
52eedfb5 1508 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1509 crit_exit_quick(td);
1510 }
1511#endif
1512}
1513
ecdefdda
MD
1514void
1515lwkt_migratecpu(int cpuid)
1516{
1517#ifdef SMP
1518 globaldata_t rgd;
1519
1520 rgd = globaldata_find(cpuid);
1521 lwkt_setcpu_self(rgd);
1522#endif
1523}
1524
5d21b981
MD
1525/*
1526 * Remote IPI for cpu migration (called while in a critical section so we
1527 * do not have to enter another one). The thread has already been moved to
1528 * our cpu's allq, but we must wait for the thread to be completely switched
1529 * out on the originating cpu before we schedule it on ours or the stack
1530 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1531 * change to main memory.
1532 *
1533 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1534 * against wakeups. It is best if this interface is used only when there
1535 * are no pending events that might try to schedule the thread.
1536 */
3d28ff59 1537#ifdef SMP
5d21b981
MD
1538static void
1539lwkt_setcpu_remote(void *arg)
1540{
1541 thread_t td = arg;
1542 globaldata_t gd = mycpu;
1543
df910c23
MD
1544 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1545#ifdef SMP
1546 lwkt_process_ipiq();
1547#endif
35238fa5 1548 cpu_lfence();
df910c23 1549 }
5d21b981 1550 td->td_gd = gd;
35238fa5 1551 cpu_sfence();
5d21b981 1552 td->td_flags &= ~TDF_MIGRATING;
9388413d 1553 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1554 _lwkt_enqueue(td);
1555}
3d28ff59 1556#endif
5d21b981 1557
553ea3c8 1558struct lwp *
4b5f931b
MD
1559lwkt_preempted_proc(void)
1560{
73e4f7b9 1561 thread_t td = curthread;
4b5f931b
MD
1562 while (td->td_preempted)
1563 td = td->td_preempted;
553ea3c8 1564 return(td->td_lwp);
4b5f931b
MD
1565}
1566
4b5f931b 1567/*
99df837e
MD
1568 * Create a kernel process/thread/whatever. It shares it's address space
1569 * with proc0 - ie: kernel only.
1570 *
365fa13f
MD
1571 * NOTE! By default new threads are created with the MP lock held. A
1572 * thread which does not require the MP lock should release it by calling
1573 * rel_mplock() at the start of the new thread.
99df837e
MD
1574 */
1575int
c9e9fb21
MD
1576lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1577 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1578{
73e4f7b9 1579 thread_t td;
e2565a42 1580 __va_list ap;
99df837e 1581
d3d32139 1582 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1583 tdflags);
a2a5ad0d
MD
1584 if (tdp)
1585 *tdp = td;
709799ea 1586 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1587
1588 /*
1589 * Set up arg0 for 'ps' etc
1590 */
e2565a42 1591 __va_start(ap, fmt);
379210cb 1592 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1593 __va_end(ap);
99df837e
MD
1594
1595 /*
1596 * Schedule the thread to run
1597 */
ef0fdad1
MD
1598 if ((td->td_flags & TDF_STOPREQ) == 0)
1599 lwkt_schedule(td);
1600 else
1601 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1602 return 0;
1603}
1604
1605/*
1606 * Destroy an LWKT thread. Warning! This function is not called when
1607 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1608 * uses a different reaping mechanism.
1609 */
1610void
1611lwkt_exit(void)
1612{
1613 thread_t td = curthread;
c070746a 1614 thread_t std;
8826f33a 1615 globaldata_t gd;
99df837e
MD
1616
1617 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1618 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1619 caps_exit(td);
c070746a
MD
1620
1621 /*
1622 * Get us into a critical section to interlock gd_freetd and loop
1623 * until we can get it freed.
1624 *
1625 * We have to cache the current td in gd_freetd because objcache_put()ing
1626 * it would rip it out from under us while our thread is still active.
1627 */
1628 gd = mycpu;
37af14fe 1629 crit_enter_quick(td);
c070746a
MD
1630 while ((std = gd->gd_freetd) != NULL) {
1631 gd->gd_freetd = NULL;
1632 objcache_put(thread_cache, std);
1633 }
3b4192fb
MD
1634
1635 /*
1636 * Remove thread resources from kernel lists and deschedule us for
1637 * the last time.
1638 */
1639 if (td->td_flags & TDF_TSLEEPQ)
1640 tsleep_remove(td);
79eae878 1641 biosched_done(td);
f8abf63c 1642 dsched_exit_thread(td);
37af14fe 1643 lwkt_deschedule_self(td);
e56e4dea 1644 lwkt_remove_tdallq(td);
c070746a
MD
1645 if (td->td_flags & TDF_ALLOCATED_THREAD)
1646 gd->gd_freetd = td;
99df837e
MD
1647 cpu_thread_exit();
1648}
1649
e56e4dea
MD
1650void
1651lwkt_remove_tdallq(thread_t td)
1652{
1653 KKASSERT(td->td_gd == mycpu);
1654 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1655}
1656
2d93b37a
MD
1657void
1658crit_panic(void)
1659{
1660 thread_t td = curthread;
850634cc 1661 int lcrit = td->td_critcount;
2d93b37a 1662
850634cc
AH
1663 td->td_critcount = 0;
1664 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
2d93b37a
MD
1665}
1666
d165e668
MD
1667#ifdef SMP
1668
bd8015ca
MD
1669/*
1670 * Called from debugger/panic on cpus which have been stopped. We must still
1671 * process the IPIQ while stopped, even if we were stopped while in a critical
1672 * section (XXX).
1673 *
1674 * If we are dumping also try to process any pending interrupts. This may
1675 * or may not work depending on the state of the cpu at the point it was
1676 * stopped.
1677 */
1678void
1679lwkt_smp_stopped(void)
1680{
1681 globaldata_t gd = mycpu;
1682
1683 crit_enter_gd(gd);
1684 if (dumping) {
1685 lwkt_process_ipiq();
1686 splz();
1687 } else {
1688 lwkt_process_ipiq();
1689 }
1690 crit_exit_gd(gd);
1691}
1692
d165e668 1693#endif