Initialize the FP unit earlier in the AP boot sequence. This solves
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08
MD
1/*
2 * Copyright (c) 2003 Matthew Dillon <dillon@backplane.com>
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 *
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24 * SUCH DAMAGE.
25 *
fc92d4aa 26 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.60 2004/05/10 10:51:31 hmp Exp $
75cdbe6c
MD
27 */
28
29/*
30 * Each cpu in a system has its own self-contained light weight kernel
31 * thread scheduler, which means that generally speaking we only need
32 * to use a critical section to avoid problems. Foreign thread
33 * scheduling is queued via (async) IPIs.
8ad65e08
MD
34 */
35
05220613
MD
36#ifdef _KERNEL
37
8ad65e08
MD
38#include <sys/param.h>
39#include <sys/systm.h>
40#include <sys/kernel.h>
41#include <sys/proc.h>
42#include <sys/rtprio.h>
43#include <sys/queue.h>
f1d1c3fa 44#include <sys/thread2.h>
7d0bac62 45#include <sys/sysctl.h>
99df837e 46#include <sys/kthread.h>
f1d1c3fa 47#include <machine/cpu.h>
99df837e 48#include <sys/lock.h>
f6bf3af1 49#include <sys/caps.h>
f1d1c3fa 50
7d0bac62
MD
51#include <vm/vm.h>
52#include <vm/vm_param.h>
53#include <vm/vm_kern.h>
54#include <vm/vm_object.h>
55#include <vm/vm_page.h>
56#include <vm/vm_map.h>
57#include <vm/vm_pager.h>
58#include <vm/vm_extern.h>
59#include <vm/vm_zone.h>
60
99df837e 61#include <machine/stdarg.h>
57c254db 62#include <machine/ipl.h>
96728c05 63#include <machine/smp.h>
99df837e 64
05220613
MD
65#define THREAD_STACK (UPAGES * PAGE_SIZE)
66
67#else
68
69#include <sys/stdint.h>
fb04f4fd 70#include <libcaps/thread.h>
05220613
MD
71#include <sys/thread.h>
72#include <sys/msgport.h>
73#include <sys/errno.h>
fb04f4fd 74#include <libcaps/globaldata.h>
05220613
MD
75#include <sys/thread2.h>
76#include <sys/msgport2.h>
709799ea 77#include <stdio.h>
05220613 78#include <stdlib.h>
709799ea 79#include <string.h>
c95cd171 80#include <machine/cpufunc.h>
709799ea 81#include <machine/lock.h>
05220613
MD
82
83#endif
84
7d0bac62 85static int untimely_switch = 0;
0f7a3396
MD
86#ifdef INVARIANTS
87static int panic_on_cscount = 0;
88#endif
05220613
MD
89static __int64_t switch_count = 0;
90static __int64_t preempt_hit = 0;
91static __int64_t preempt_miss = 0;
92static __int64_t preempt_weird = 0;
05220613
MD
93
94#ifdef _KERNEL
95
96SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
0f7a3396
MD
97#ifdef INVARIANTS
98SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
99#endif
4b5f931b 100SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 101SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 102SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 103SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
7d0bac62 104
05220613
MD
105#endif
106
4b5f931b
MD
107/*
108 * These helper procedures handle the runq, they can only be called from
109 * within a critical section.
75cdbe6c
MD
110 *
111 * WARNING! Prior to SMP being brought up it is possible to enqueue and
112 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
113 * instead of 'mycpu' when referencing the globaldata structure. Once
114 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 115 */
f1d1c3fa
MD
116static __inline
117void
118_lwkt_dequeue(thread_t td)
119{
120 if (td->td_flags & TDF_RUNQ) {
4b5f931b 121 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 122 struct globaldata *gd = td->td_gd;
4b5f931b 123
f1d1c3fa 124 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
125 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
126 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
127 }
128}
129
130static __inline
131void
132_lwkt_enqueue(thread_t td)
133{
134 if ((td->td_flags & TDF_RUNQ) == 0) {
4b5f931b 135 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 136 struct globaldata *gd = td->td_gd;
4b5f931b 137
f1d1c3fa 138 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
139 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
140 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
141 }
142}
8ad65e08 143
37af14fe
MD
144/*
145 * Schedule a thread to run. As the current thread we can always safely
146 * schedule ourselves, and a shortcut procedure is provided for that
147 * function.
148 *
149 * (non-blocking, self contained on a per cpu basis)
150 */
151void
152lwkt_schedule_self(thread_t td)
153{
154 crit_enter_quick(td);
155 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
156 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
157 _lwkt_enqueue(td);
158#ifdef _KERNEL
159 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
160 panic("SCHED SELF PANIC");
161#endif
162 crit_exit_quick(td);
163}
164
165/*
166 * Deschedule a thread.
167 *
168 * (non-blocking, self contained on a per cpu basis)
169 */
170void
171lwkt_deschedule_self(thread_t td)
172{
173 crit_enter_quick(td);
174 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
175 _lwkt_dequeue(td);
176 crit_exit_quick(td);
177}
178
2d93b37a
MD
179#ifdef _KERNEL
180
8ad65e08
MD
181/*
182 * LWKTs operate on a per-cpu basis
183 *
73e4f7b9 184 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
185 */
186void
187lwkt_gdinit(struct globaldata *gd)
188{
4b5f931b
MD
189 int i;
190
191 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
192 TAILQ_INIT(&gd->gd_tdrunq[i]);
193 gd->gd_runqmask = 0;
73e4f7b9 194 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
195}
196
2d93b37a
MD
197#endif /* _KERNEL */
198
7d0bac62
MD
199/*
200 * Initialize a thread wait structure prior to first use.
201 *
202 * NOTE! called from low level boot code, we cannot do anything fancy!
203 */
204void
41a01a4d 205lwkt_wait_init(lwkt_wait_t w)
7d0bac62 206{
41a01a4d 207 lwkt_token_init(&w->wa_token);
7d0bac62 208 TAILQ_INIT(&w->wa_waitq);
41a01a4d
MD
209 w->wa_gen = 0;
210 w->wa_count = 0;
7d0bac62
MD
211}
212
213/*
214 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
215 * or LWKT start address before it can be scheduled. If the target cpu is
216 * -1 the thread will be created on the current cpu.
0cfcada1
MD
217 *
218 * If you intend to create a thread without a process context this function
219 * does everything except load the startup and switcher function.
7d0bac62
MD
220 */
221thread_t
75cdbe6c 222lwkt_alloc_thread(struct thread *td, int cpu)
7d0bac62 223{
99df837e 224 void *stack;
ef0fdad1 225 int flags = 0;
37af14fe 226 globaldata_t gd = mycpu;
7d0bac62 227
ef0fdad1 228 if (td == NULL) {
37af14fe
MD
229 crit_enter_gd(gd);
230 if (gd->gd_tdfreecount > 0) {
231 --gd->gd_tdfreecount;
232 td = TAILQ_FIRST(&gd->gd_tdfreeq);
d9eea1a5 233 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1 234 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
37af14fe
MD
235 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq);
236 crit_exit_gd(gd);
ef0fdad1
MD
237 stack = td->td_kstack;
238 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
239 } else {
37af14fe 240 crit_exit_gd(gd);
05220613 241#ifdef _KERNEL
ef0fdad1 242 td = zalloc(thread_zone);
05220613
MD
243#else
244 td = malloc(sizeof(struct thread));
245#endif
ef0fdad1
MD
246 td->td_kstack = NULL;
247 flags |= TDF_ALLOCATED_THREAD;
248 }
249 }
250 if ((stack = td->td_kstack) == NULL) {
05220613
MD
251#ifdef _KERNEL
252 stack = (void *)kmem_alloc(kernel_map, THREAD_STACK);
253#else
fb04f4fd 254 stack = libcaps_alloc_stack(THREAD_STACK);
05220613 255#endif
ef0fdad1 256 flags |= TDF_ALLOCATED_STACK;
99df837e 257 }
75cdbe6c
MD
258 if (cpu < 0)
259 lwkt_init_thread(td, stack, flags, mycpu);
260 else
261 lwkt_init_thread(td, stack, flags, globaldata_find(cpu));
99df837e 262 return(td);
7d0bac62
MD
263}
264
709799ea
MD
265#ifdef _KERNEL
266
7d0bac62
MD
267/*
268 * Initialize a preexisting thread structure. This function is used by
269 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
270 *
f8c3996b
MD
271 * All threads start out in a critical section at a priority of
272 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
273 * appropriate. This function may send an IPI message when the
274 * requested cpu is not the current cpu and consequently gd_tdallq may
275 * not be initialized synchronously from the point of view of the originating
276 * cpu.
277 *
278 * NOTE! we have to be careful in regards to creating threads for other cpus
279 * if SMP has not yet been activated.
7d0bac62 280 */
41a01a4d
MD
281#ifdef SMP
282
75cdbe6c
MD
283static void
284lwkt_init_thread_remote(void *arg)
285{
286 thread_t td = arg;
287
288 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
289}
290
41a01a4d
MD
291#endif
292
7d0bac62 293void
26a0694b 294lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
7d0bac62 295{
37af14fe
MD
296 globaldata_t mygd = mycpu;
297
99df837e
MD
298 bzero(td, sizeof(struct thread));
299 td->td_kstack = stack;
300 td->td_flags |= flags;
26a0694b 301 td->td_gd = gd;
f8c3996b 302 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
c95cd171 303 lwkt_initport(&td->td_msgport, td);
99df837e 304 pmap_init_thread(td);
0f7a3396 305#ifdef SMP
37af14fe
MD
306 if (gd == mygd) {
307 crit_enter_gd(mygd);
75cdbe6c 308 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 309 crit_exit_gd(mygd);
75cdbe6c 310 } else {
2db3b277 311 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 312 }
0f7a3396 313#else
37af14fe 314 crit_enter_gd(mygd);
0f7a3396 315 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 316 crit_exit_gd(mygd);
0f7a3396 317#endif
73e4f7b9
MD
318}
319
2d93b37a
MD
320#endif /* _KERNEL */
321
73e4f7b9
MD
322void
323lwkt_set_comm(thread_t td, const char *ctl, ...)
324{
e2565a42 325 __va_list va;
73e4f7b9 326
e2565a42 327 __va_start(va, ctl);
73e4f7b9 328 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 329 __va_end(va);
7d0bac62
MD
330}
331
99df837e 332void
73e4f7b9 333lwkt_hold(thread_t td)
99df837e 334{
73e4f7b9
MD
335 ++td->td_refs;
336}
337
338void
339lwkt_rele(thread_t td)
340{
341 KKASSERT(td->td_refs > 0);
342 --td->td_refs;
343}
344
c95cd171
MD
345#ifdef _KERNEL
346
73e4f7b9
MD
347void
348lwkt_wait_free(thread_t td)
349{
350 while (td->td_refs)
377d4740 351 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
352}
353
c95cd171
MD
354#endif
355
73e4f7b9
MD
356void
357lwkt_free_thread(thread_t td)
358{
359 struct globaldata *gd = mycpu;
360
d9eea1a5 361 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
362 ("lwkt_free_thread: did not exit! %p", td));
363
37af14fe 364 crit_enter_gd(gd);
73e4f7b9
MD
365 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
366 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
367 (td->td_flags & TDF_ALLOCATED_THREAD)
368 ) {
73e4f7b9
MD
369 ++gd->gd_tdfreecount;
370 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
37af14fe 371 crit_exit_gd(gd);
99df837e 372 } else {
37af14fe 373 crit_exit_gd(gd);
99df837e 374 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613
MD
375#ifdef _KERNEL
376 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK);
377#else
fb04f4fd 378 libcaps_free_stack(td->td_kstack, THREAD_STACK);
05220613 379#endif
73e4f7b9 380 /* gd invalid */
99df837e
MD
381 td->td_kstack = NULL;
382 }
05220613
MD
383 if (td->td_flags & TDF_ALLOCATED_THREAD) {
384#ifdef _KERNEL
99df837e 385 zfree(thread_zone, td);
05220613
MD
386#else
387 free(td);
388#endif
389 }
99df837e
MD
390 }
391}
392
393
8ad65e08
MD
394/*
395 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
396 * switch to the idlethread. Switching must occur within a critical
397 * section to avoid races with the scheduling queue.
398 *
399 * We always have full control over our cpu's run queue. Other cpus
400 * that wish to manipulate our queue must use the cpu_*msg() calls to
401 * talk to our cpu, so a critical section is all that is needed and
402 * the result is very, very fast thread switching.
403 *
96728c05
MD
404 * The LWKT scheduler uses a fixed priority model and round-robins at
405 * each priority level. User process scheduling is a totally
406 * different beast and LWKT priorities should not be confused with
407 * user process priorities.
f1d1c3fa 408 *
96728c05
MD
409 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
410 * cleans it up. Note that the td_switch() function cannot do anything that
411 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
412 * the target thread (not the current thread). It's nice to have a scheduler
413 * that does not need the MP lock to work because it allows us to do some
414 * really cool high-performance MP lock optimizations.
8ad65e08 415 */
96728c05 416
8ad65e08
MD
417void
418lwkt_switch(void)
419{
37af14fe
MD
420 globaldata_t gd = mycpu;
421 thread_t td = gd->gd_curthread;
8ad65e08 422 thread_t ntd;
8a8d5d85
MD
423#ifdef SMP
424 int mpheld;
425#endif
8ad65e08 426
46a3f46d
MD
427 /*
428 * Switching from within a 'fast' (non thread switched) interrupt is
429 * illegal.
430 */
37af14fe 431 if (gd->gd_intr_nesting_level && panicstr == NULL) {
fc92d4aa 432 panic("lwkt_switch: cannot switch from within a fast interrupt, yet");
96728c05 433 }
ef0fdad1 434
cb973d15
MD
435 /*
436 * Passive release (used to transition from user to kernel mode
437 * when we block or switch rather then when we enter the kernel).
438 * This function is NOT called if we are switching into a preemption
439 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
440 * our current process designation (if we have one) and become a true
441 * LWKT thread, and may also hand the current process designation to
442 * another process and schedule thread.
cb973d15
MD
443 */
444 if (td->td_release)
445 td->td_release(td);
446
37af14fe 447 crit_enter_gd(gd);
4b5f931b 448 ++switch_count;
8a8d5d85
MD
449
450#ifdef SMP
451 /*
452 * td_mpcount cannot be used to determine if we currently hold the
453 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
454 * to get the lock, and switch out if it can't. Our ownership of
455 * the actual lock will remain stable while we are in a critical section
456 * (but, of course, another cpu may own or release the lock so the
457 * actual value of mp_lock is not stable).
8a8d5d85
MD
458 */
459 mpheld = MP_LOCK_HELD();
0f7a3396
MD
460#ifdef INVARIANTS
461 if (td->td_cscount) {
462 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
463 td);
464 if (panic_on_cscount)
465 panic("switching while mastering cpusync");
466 }
467#endif
8a8d5d85 468#endif
99df837e
MD
469 if ((ntd = td->td_preempted) != NULL) {
470 /*
471 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
472 * thread. This occurs transparently, whether the preempted thread
473 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
474 * itself).
475 *
476 * We have to setup the MP lock for the original thread after backing
477 * out the adjustment that was made to curthread when the original
478 * was preempted.
99df837e 479 */
26a0694b 480 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 481#ifdef SMP
96728c05 482 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 483 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
484 td, ntd, td->td_mpcount, ntd->td_mpcount);
485 }
8a8d5d85
MD
486 if (ntd->td_mpcount) {
487 td->td_mpcount -= ntd->td_mpcount;
488 KKASSERT(td->td_mpcount >= 0);
489 }
490#endif
26a0694b 491 ntd->td_flags |= TDF_PREEMPT_DONE;
8a8d5d85 492 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 493 } else {
4b5f931b
MD
494 /*
495 * Priority queue / round-robin at each priority. Note that user
496 * processes run at a fixed, low priority and the user process
497 * scheduler deals with interactions between user processes
498 * by scheduling and descheduling them from the LWKT queue as
499 * necessary.
8a8d5d85
MD
500 *
501 * We have to adjust the MP lock for the target thread. If we
502 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
503 * thread that does not need the MP lock. If we cannot, we spin
504 * instead of HLT.
505 *
506 * A similar issue exists for the tokens held by the target thread.
507 * If we cannot obtain ownership of the tokens we cannot immediately
508 * schedule the thread.
509 */
510
511 /*
512 * We are switching threads. If there are any pending requests for
513 * tokens we can satisfy all of them here.
4b5f931b 514 */
41a01a4d
MD
515#ifdef SMP
516 if (gd->gd_tokreqbase)
517 lwkt_drain_token_requests();
518#endif
519
4b5f931b
MD
520again:
521 if (gd->gd_runqmask) {
522 int nq = bsrl(gd->gd_runqmask);
523 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
524 gd->gd_runqmask &= ~(1 << nq);
525 goto again;
526 }
8a8d5d85 527#ifdef SMP
41a01a4d
MD
528 /*
529 * If the target needs the MP lock and we couldn't get it,
530 * or if the target is holding tokens and we could not
531 * gain ownership of the tokens, continue looking for a
532 * thread to schedule and spin instead of HLT if we can't.
533 */
534 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
535 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
536 ) {
8a8d5d85
MD
537 u_int32_t rqmask = gd->gd_runqmask;
538 while (rqmask) {
539 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
41a01a4d
MD
540 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock())
541 continue;
542 mpheld = MP_LOCK_HELD();
543 if (ntd->td_toks && !lwkt_chktokens(ntd))
544 continue;
545 break;
8a8d5d85
MD
546 }
547 if (ntd)
548 break;
549 rqmask &= ~(1 << nq);
550 nq = bsrl(rqmask);
551 }
552 if (ntd == NULL) {
a2a5ad0d
MD
553 ntd = &gd->gd_idlethread;
554 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
555 } else {
556 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
557 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
558 }
559 } else {
560 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
561 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
562 }
563#else
4b5f931b
MD
564 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
565 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 566#endif
4b5f931b 567 } else {
3c23a41a 568 /*
60f945af
MD
569 * We have nothing to run but only let the idle loop halt
570 * the cpu if there are no pending interrupts.
3c23a41a 571 */
a2a5ad0d 572 ntd = &gd->gd_idlethread;
60f945af 573 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 574 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 575 }
f1d1c3fa 576 }
26a0694b
MD
577 KASSERT(ntd->td_pri >= TDPRI_CRIT,
578 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
579
580 /*
581 * Do the actual switch. If the new target does not need the MP lock
582 * and we are holding it, release the MP lock. If the new target requires
583 * the MP lock we have already acquired it for the target.
584 */
585#ifdef SMP
586 if (ntd->td_mpcount == 0 ) {
587 if (MP_LOCK_HELD())
588 cpu_rel_mplock();
589 } else {
590 ASSERT_MP_LOCK_HELD();
591 }
592#endif
37af14fe 593 if (td != ntd)
f1d1c3fa 594 td->td_switch(ntd);
37af14fe
MD
595 /* NOTE: current cpu may have changed after switch */
596 crit_exit_quick(td);
8ad65e08
MD
597}
598
b68b7282 599/*
96728c05
MD
600 * Request that the target thread preempt the current thread. Preemption
601 * only works under a specific set of conditions:
b68b7282 602 *
96728c05
MD
603 * - We are not preempting ourselves
604 * - The target thread is owned by the current cpu
605 * - We are not currently being preempted
606 * - The target is not currently being preempted
607 * - We are able to satisfy the target's MP lock requirements (if any).
608 *
609 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
610 * this is called via lwkt_schedule() through the td_preemptable callback.
611 * critpri is the managed critical priority that we should ignore in order
612 * to determine whether preemption is possible (aka usually just the crit
613 * priority of lwkt_schedule() itself).
b68b7282 614 *
26a0694b
MD
615 * XXX at the moment we run the target thread in a critical section during
616 * the preemption in order to prevent the target from taking interrupts
617 * that *WE* can't. Preemption is strictly limited to interrupt threads
618 * and interrupt-like threads, outside of a critical section, and the
619 * preempted source thread will be resumed the instant the target blocks
620 * whether or not the source is scheduled (i.e. preemption is supposed to
621 * be as transparent as possible).
4b5f931b 622 *
8a8d5d85
MD
623 * The target thread inherits our MP count (added to its own) for the
624 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
625 * MP lock during the preemption. Therefore, any preempting targets must be
626 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
627 * out of sync with the physical mp_lock, but we do not have to preserve
628 * the original ownership of the lock if it was out of synch (that is, we
629 * can leave it synchronized on return).
b68b7282
MD
630 */
631void
96728c05 632lwkt_preempt(thread_t ntd, int critpri)
b68b7282 633{
46a3f46d 634 struct globaldata *gd = mycpu;
0a3f9b47 635 thread_t td;
8a8d5d85
MD
636#ifdef SMP
637 int mpheld;
57c254db 638 int savecnt;
8a8d5d85 639#endif
b68b7282 640
26a0694b 641 /*
96728c05
MD
642 * The caller has put us in a critical section. We can only preempt
643 * if the caller of the caller was not in a critical section (basically
0a3f9b47 644 * a local interrupt), as determined by the 'critpri' parameter.
96728c05
MD
645 *
646 * YYY The target thread must be in a critical section (else it must
647 * inherit our critical section? I dunno yet).
41a01a4d
MD
648 *
649 * Any tokens held by the target may not be held by thread(s) being
650 * preempted. We take the easy way out and do not preempt if
651 * the target is holding tokens.
0a3f9b47
MD
652 *
653 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
654 */
655 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 656
0a3f9b47
MD
657 td = gd->gd_curthread;
658 need_lwkt_resched();
659 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
660 ++preempt_miss;
661 return;
662 }
96728c05
MD
663 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
664 ++preempt_miss;
665 return;
666 }
667#ifdef SMP
46a3f46d 668 if (ntd->td_gd != gd) {
96728c05
MD
669 ++preempt_miss;
670 return;
671 }
672#endif
41a01a4d
MD
673 /*
674 * Take the easy way out and do not preempt if the target is holding
675 * one or more tokens. We could test whether the thread(s) being
676 * preempted interlock against the target thread's tokens and whether
677 * we can get all the target thread's tokens, but this situation
678 * should not occur very often so its easier to simply not preempt.
679 */
680 if (ntd->td_toks != NULL) {
681 ++preempt_miss;
682 return;
683 }
26a0694b
MD
684 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
685 ++preempt_weird;
686 return;
687 }
688 if (ntd->td_preempted) {
4b5f931b 689 ++preempt_hit;
26a0694b 690 return;
b68b7282 691 }
8a8d5d85 692#ifdef SMP
a2a5ad0d
MD
693 /*
694 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
695 * to or from the MP lock. In this case td_mpcount will be pre-disposed
696 * (non-zero) but not actually synchronized with the actual state of the
697 * lock. We can use it to imply an MP lock requirement for the
698 * preemption but we cannot use it to test whether we hold the MP lock
699 * or not.
a2a5ad0d 700 */
96728c05 701 savecnt = td->td_mpcount;
71ef2f5c 702 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
703 ntd->td_mpcount += td->td_mpcount;
704 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
705 ntd->td_mpcount -= td->td_mpcount;
706 ++preempt_miss;
707 return;
708 }
709#endif
26a0694b
MD
710
711 ++preempt_hit;
712 ntd->td_preempted = td;
713 td->td_flags |= TDF_PREEMPT_LOCK;
714 td->td_switch(ntd);
715 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
716#ifdef SMP
717 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
718 mpheld = MP_LOCK_HELD();
719 if (mpheld && td->td_mpcount == 0)
96728c05 720 cpu_rel_mplock();
71ef2f5c 721 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
722 panic("lwkt_preempt(): MP lock was not held through");
723#endif
26a0694b
MD
724 ntd->td_preempted = NULL;
725 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
726}
727
f1d1c3fa
MD
728/*
729 * Yield our thread while higher priority threads are pending. This is
730 * typically called when we leave a critical section but it can be safely
731 * called while we are in a critical section.
732 *
733 * This function will not generally yield to equal priority threads but it
734 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 735 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
736 * lwkt_yield_quick().
737 *
235957ed 738 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
739 * came along but was blocked and made pending.
740 *
f1d1c3fa
MD
741 * (self contained on a per cpu basis)
742 */
743void
744lwkt_yield_quick(void)
745{
7966cb69
MD
746 globaldata_t gd = mycpu;
747 thread_t td = gd->gd_curthread;
ef0fdad1 748
a2a5ad0d 749 /*
235957ed 750 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
751 * it with a non-zero cpl then we might not wind up calling splz after
752 * a task switch when the critical section is exited even though the
46a3f46d 753 * new task could accept the interrupt.
a2a5ad0d
MD
754 *
755 * XXX from crit_exit() only called after last crit section is released.
756 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
757 *
758 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
759 * except for this special case, we MUST call splz() here to handle any
760 * pending ints, particularly after we switch, or we might accidently
761 * halt the cpu with interrupts pending.
a2a5ad0d 762 */
46a3f46d 763 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 764 splz();
f1d1c3fa
MD
765
766 /*
767 * YYY enabling will cause wakeup() to task-switch, which really
768 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
769 * preemption and MP without actually doing preemption or MP, because a
770 * lot of code assumes that wakeup() does not block.
f1d1c3fa 771 */
46a3f46d
MD
772 if (untimely_switch && td->td_nest_count == 0 &&
773 gd->gd_intr_nesting_level == 0
774 ) {
37af14fe 775 crit_enter_quick(td);
f1d1c3fa
MD
776 /*
777 * YYY temporary hacks until we disassociate the userland scheduler
778 * from the LWKT scheduler.
779 */
780 if (td->td_flags & TDF_RUNQ) {
781 lwkt_switch(); /* will not reenter yield function */
782 } else {
37af14fe 783 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 784 lwkt_switch(); /* will not reenter yield function */
37af14fe 785 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 786 }
7966cb69 787 crit_exit_noyield(td);
f1d1c3fa 788 }
f1d1c3fa
MD
789}
790
8ad65e08 791/*
f1d1c3fa 792 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 793 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
794 * the crit_exit() call in lwkt_switch().
795 *
796 * (self contained on a per cpu basis)
8ad65e08
MD
797 */
798void
f1d1c3fa 799lwkt_yield(void)
8ad65e08 800{
37af14fe 801 lwkt_schedule_self(curthread);
f1d1c3fa
MD
802 lwkt_switch();
803}
804
8ad65e08 805/*
f1d1c3fa
MD
806 * Generic schedule. Possibly schedule threads belonging to other cpus and
807 * deal with threads that might be blocked on a wait queue.
808 *
0a3f9b47
MD
809 * We have a little helper inline function which does additional work after
810 * the thread has been enqueued, including dealing with preemption and
811 * setting need_lwkt_resched() (which prevents the kernel from returning
812 * to userland until it has processed higher priority threads).
8ad65e08 813 */
0a3f9b47
MD
814static __inline
815void
816_lwkt_schedule_post(thread_t ntd, int cpri)
817{
818 if (ntd->td_preemptable) {
819 ntd->td_preemptable(ntd, cpri); /* YYY +token */
820 } else {
821 if ((ntd->td_flags & TDF_NORESCHED) == 0) {
822 if ((ntd->td_pri & TDPRI_MASK) >= TDPRI_KERN_USER)
823 need_lwkt_resched();
824 }
825 }
826}
827
8ad65e08
MD
828void
829lwkt_schedule(thread_t td)
830{
37af14fe
MD
831 globaldata_t mygd = mycpu;
832
96728c05 833#ifdef INVARIANTS
41a01a4d 834 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
26a0694b
MD
835 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
836 && td->td_proc->p_stat == SSLEEP
837 ) {
838 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
839 curthread,
840 curthread->td_proc ? curthread->td_proc->p_pid : -1,
841 curthread->td_proc ? curthread->td_proc->p_stat : -1,
842 td,
843 td->td_proc ? curthread->td_proc->p_pid : -1,
844 td->td_proc ? curthread->td_proc->p_stat : -1
845 );
846 panic("SCHED PANIC");
847 }
96728c05 848#endif
37af14fe
MD
849 crit_enter_gd(mygd);
850 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
851 _lwkt_enqueue(td);
852 } else {
853 lwkt_wait_t w;
854
855 /*
856 * If the thread is on a wait list we have to send our scheduling
857 * request to the owner of the wait structure. Otherwise we send
858 * the scheduling request to the cpu owning the thread. Races
859 * are ok, the target will forward the message as necessary (the
860 * message may chase the thread around before it finally gets
861 * acted upon).
862 *
863 * (remember, wait structures use stable storage)
0a3f9b47
MD
864 *
865 * NOTE: tokens no longer enter a critical section, so we only need
866 * to account for the crit_enter() above when calling
867 * _lwkt_schedule_post().
f1d1c3fa
MD
868 */
869 if ((w = td->td_wait) != NULL) {
41a01a4d
MD
870 lwkt_tokref wref;
871
872 if (lwkt_trytoken(&wref, &w->wa_token)) {
f1d1c3fa
MD
873 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
874 --w->wa_count;
875 td->td_wait = NULL;
0f7a3396
MD
876#ifdef SMP
877 if (td->td_gd == mycpu) {
f1d1c3fa 878 _lwkt_enqueue(td);
0a3f9b47 879 _lwkt_schedule_post(td, TDPRI_CRIT);
f1d1c3fa 880 } else {
2db3b277 881 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 882 }
0f7a3396
MD
883#else
884 _lwkt_enqueue(td);
0a3f9b47 885 _lwkt_schedule_post(td, TDPRI_CRIT);
0f7a3396 886#endif
41a01a4d 887 lwkt_reltoken(&wref);
f1d1c3fa 888 } else {
96728c05 889 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
890 }
891 } else {
892 /*
893 * If the wait structure is NULL and we own the thread, there
894 * is no race (since we are in a critical section). If we
895 * do not own the thread there might be a race but the
896 * target cpu will deal with it.
897 */
0f7a3396 898#ifdef SMP
37af14fe 899 if (td->td_gd == mygd) {
f1d1c3fa 900 _lwkt_enqueue(td);
0a3f9b47 901 _lwkt_schedule_post(td, TDPRI_CRIT);
f1d1c3fa 902 } else {
2db3b277 903 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 904 }
0f7a3396
MD
905#else
906 _lwkt_enqueue(td);
0a3f9b47 907 _lwkt_schedule_post(td, TDPRI_CRIT);
0f7a3396 908#endif
f1d1c3fa 909 }
8ad65e08 910 }
37af14fe 911 crit_exit_gd(mygd);
8ad65e08
MD
912}
913
d9eea1a5
MD
914/*
915 * Managed acquisition. This code assumes that the MP lock is held for
916 * the tdallq operation and that the thread has been descheduled from its
917 * original cpu. We also have to wait for the thread to be entirely switched
918 * out on its original cpu (this is usually fast enough that we never loop)
919 * since the LWKT system does not have to hold the MP lock while switching
920 * and the target may have released it before switching.
921 */
a2a5ad0d
MD
922void
923lwkt_acquire(thread_t td)
924{
37af14fe
MD
925 globaldata_t gd;
926 globaldata_t mygd;
a2a5ad0d
MD
927
928 gd = td->td_gd;
37af14fe 929 mygd = mycpu;
a2a5ad0d 930 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5
MD
931 while (td->td_flags & TDF_RUNNING) /* XXX spin */
932 ;
37af14fe
MD
933 if (gd != mygd) {
934 crit_enter_gd(mygd);
a2a5ad0d 935 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
37af14fe
MD
936 td->td_gd = mygd;
937 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
938 crit_exit_gd(mygd);
a2a5ad0d
MD
939 }
940}
941
f1d1c3fa
MD
942/*
943 * Generic deschedule. Descheduling threads other then your own should be
944 * done only in carefully controlled circumstances. Descheduling is
945 * asynchronous.
946 *
947 * This function may block if the cpu has run out of messages.
8ad65e08
MD
948 */
949void
950lwkt_deschedule(thread_t td)
951{
f1d1c3fa
MD
952 crit_enter();
953 if (td == curthread) {
954 _lwkt_dequeue(td);
955 } else {
a72187e9 956 if (td->td_gd == mycpu) {
f1d1c3fa
MD
957 _lwkt_dequeue(td);
958 } else {
2db3b277 959 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
960 }
961 }
962 crit_exit();
963}
964
4b5f931b
MD
965/*
966 * Set the target thread's priority. This routine does not automatically
967 * switch to a higher priority thread, LWKT threads are not designed for
968 * continuous priority changes. Yield if you want to switch.
969 *
970 * We have to retain the critical section count which uses the high bits
26a0694b
MD
971 * of the td_pri field. The specified priority may also indicate zero or
972 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
973 *
974 * Note that we requeue the thread whether it winds up on a different runq
975 * or not. uio_yield() depends on this and the routine is not normally
976 * called with the same priority otherwise.
4b5f931b
MD
977 */
978void
979lwkt_setpri(thread_t td, int pri)
980{
26a0694b 981 KKASSERT(pri >= 0);
a72187e9 982 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
983 crit_enter();
984 if (td->td_flags & TDF_RUNQ) {
985 _lwkt_dequeue(td);
986 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
987 _lwkt_enqueue(td);
988 } else {
989 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
990 }
991 crit_exit();
992}
993
994void
995lwkt_setpri_self(int pri)
996{
997 thread_t td = curthread;
998
4b5f931b
MD
999 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1000 crit_enter();
1001 if (td->td_flags & TDF_RUNQ) {
1002 _lwkt_dequeue(td);
1003 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1004 _lwkt_enqueue(td);
1005 } else {
1006 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1007 }
1008 crit_exit();
1009}
1010
1011struct proc *
1012lwkt_preempted_proc(void)
1013{
73e4f7b9 1014 thread_t td = curthread;
4b5f931b
MD
1015 while (td->td_preempted)
1016 td = td->td_preempted;
1017 return(td->td_proc);
1018}
1019
f1d1c3fa 1020/*
41a01a4d
MD
1021 * Block on the specified wait queue until signaled. A generation number
1022 * must be supplied to interlock the wait queue. The function will
1023 * return immediately if the generation number does not match the wait
1024 * structure's generation number.
f1d1c3fa
MD
1025 */
1026void
ae8050a4 1027lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
1028{
1029 thread_t td = curthread;
41a01a4d 1030 lwkt_tokref ilock;
f1d1c3fa 1031
41a01a4d
MD
1032 lwkt_gettoken(&ilock, &w->wa_token);
1033 crit_enter();
ae8050a4 1034 if (w->wa_gen == *gen) {
f1d1c3fa
MD
1035 _lwkt_dequeue(td);
1036 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1037 ++w->wa_count;
1038 td->td_wait = w;
ae8050a4 1039 td->td_wmesg = wmesg;
41a01a4d 1040 again:
f1d1c3fa 1041 lwkt_switch();
ece04fd0
MD
1042 if (td->td_wmesg != NULL) {
1043 _lwkt_dequeue(td);
1044 goto again;
1045 }
8ad65e08 1046 }
41a01a4d 1047 crit_exit();
ae8050a4 1048 *gen = w->wa_gen;
41a01a4d 1049 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1050}
1051
1052/*
1053 * Signal a wait queue. We gain ownership of the wait queue in order to
1054 * signal it. Once a thread is removed from the wait queue we have to
1055 * deal with the cpu owning the thread.
1056 *
1057 * Note: alternatively we could message the target cpu owning the wait
1058 * queue. YYY implement as sysctl.
1059 */
1060void
ece04fd0 1061lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1062{
1063 thread_t td;
41a01a4d 1064 lwkt_tokref ilock;
f1d1c3fa 1065
41a01a4d 1066 lwkt_gettoken(&ilock, &w->wa_token);
f1d1c3fa 1067 ++w->wa_gen;
41a01a4d 1068 crit_enter();
ece04fd0
MD
1069 if (count < 0)
1070 count = w->wa_count;
f1d1c3fa
MD
1071 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1072 --count;
1073 --w->wa_count;
1074 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1075 td->td_wait = NULL;
ae8050a4 1076 td->td_wmesg = NULL;
a72187e9 1077 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1078 _lwkt_enqueue(td);
1079 } else {
2db3b277 1080 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 1081 }
f1d1c3fa 1082 }
41a01a4d
MD
1083 crit_exit();
1084 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1085}
1086
99df837e
MD
1087/*
1088 * Create a kernel process/thread/whatever. It shares it's address space
1089 * with proc0 - ie: kernel only.
1090 *
365fa13f
MD
1091 * NOTE! By default new threads are created with the MP lock held. A
1092 * thread which does not require the MP lock should release it by calling
1093 * rel_mplock() at the start of the new thread.
99df837e
MD
1094 */
1095int
1096lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1097 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1098 const char *fmt, ...)
99df837e 1099{
73e4f7b9 1100 thread_t td;
e2565a42 1101 __va_list ap;
99df837e 1102
75cdbe6c 1103 td = lwkt_alloc_thread(template, cpu);
a2a5ad0d
MD
1104 if (tdp)
1105 *tdp = td;
709799ea 1106 cpu_set_thread_handler(td, lwkt_exit, func, arg);
ef0fdad1 1107 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1108#ifdef SMP
1109 td->td_mpcount = 1;
1110#endif
99df837e
MD
1111
1112 /*
1113 * Set up arg0 for 'ps' etc
1114 */
e2565a42 1115 __va_start(ap, fmt);
99df837e 1116 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1117 __va_end(ap);
99df837e
MD
1118
1119 /*
1120 * Schedule the thread to run
1121 */
ef0fdad1
MD
1122 if ((td->td_flags & TDF_STOPREQ) == 0)
1123 lwkt_schedule(td);
1124 else
1125 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1126 return 0;
1127}
1128
2d93b37a 1129/*
2d93b37a
MD
1130 * kthread_* is specific to the kernel and is not needed by userland.
1131 */
1132#ifdef _KERNEL
1133
99df837e
MD
1134/*
1135 * Destroy an LWKT thread. Warning! This function is not called when
1136 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1137 * uses a different reaping mechanism.
1138 */
1139void
1140lwkt_exit(void)
1141{
1142 thread_t td = curthread;
1143
1144 if (td->td_flags & TDF_VERBOSE)
1145 printf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1146 caps_exit(td);
37af14fe
MD
1147 crit_enter_quick(td);
1148 lwkt_deschedule_self(td);
99df837e
MD
1149 ++mycpu->gd_tdfreecount;
1150 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1151 cpu_thread_exit();
1152}
1153
1154/*
1155 * Create a kernel process/thread/whatever. It shares it's address space
ef0fdad1 1156 * with proc0 - ie: kernel only. 5.x compatible.
365fa13f
MD
1157 *
1158 * NOTE! By default kthreads are created with the MP lock held. A
1159 * thread which does not require the MP lock should release it by calling
1160 * rel_mplock() at the start of the new thread.
99df837e
MD
1161 */
1162int
1163kthread_create(void (*func)(void *), void *arg,
1164 struct thread **tdp, const char *fmt, ...)
1165{
73e4f7b9 1166 thread_t td;
e2565a42 1167 __va_list ap;
99df837e 1168
75cdbe6c 1169 td = lwkt_alloc_thread(NULL, -1);
a2a5ad0d
MD
1170 if (tdp)
1171 *tdp = td;
99df837e
MD
1172 cpu_set_thread_handler(td, kthread_exit, func, arg);
1173 td->td_flags |= TDF_VERBOSE;
8a8d5d85
MD
1174#ifdef SMP
1175 td->td_mpcount = 1;
1176#endif
99df837e
MD
1177
1178 /*
1179 * Set up arg0 for 'ps' etc
1180 */
e2565a42 1181 __va_start(ap, fmt);
99df837e 1182 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1183 __va_end(ap);
99df837e
MD
1184
1185 /*
1186 * Schedule the thread to run
1187 */
1188 lwkt_schedule(td);
1189 return 0;
1190}
1191
1192/*
1193 * Destroy an LWKT thread. Warning! This function is not called when
1194 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1195 * uses a different reaping mechanism.
1196 *
1197 * XXX duplicates lwkt_exit()
1198 */
1199void
1200kthread_exit(void)
1201{
1202 lwkt_exit();
1203}
1204
2d93b37a
MD
1205#endif /* _KERNEL */
1206
1207void
1208crit_panic(void)
1209{
1210 thread_t td = curthread;
1211 int lpri = td->td_pri;
1212
1213 td->td_pri = 0;
1214 panic("td_pri is/would-go negative! %p %d", td, lpri);
1215}
1216