Add ref counting to the checkpoint handler and unload function so an unload
[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 *
5d21b981 26 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.61 2004/05/28 08:37:34 dillon 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{
5d21b981 134 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING)) == 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
5d21b981
MD
306 /*
307 * Normally initializing a thread for a remote cpu requires sending an
308 * IPI. However, the idlethread is setup before the other cpus are
309 * activated so we have to treat it as a special case. XXX manipulation
310 * of gd_tdallq requires the BGL.
311 */
312 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 313 crit_enter_gd(mygd);
75cdbe6c 314 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 315 crit_exit_gd(mygd);
75cdbe6c 316 } else {
2db3b277 317 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 318 }
0f7a3396 319#else
37af14fe 320 crit_enter_gd(mygd);
0f7a3396 321 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 322 crit_exit_gd(mygd);
0f7a3396 323#endif
73e4f7b9
MD
324}
325
2d93b37a
MD
326#endif /* _KERNEL */
327
73e4f7b9
MD
328void
329lwkt_set_comm(thread_t td, const char *ctl, ...)
330{
e2565a42 331 __va_list va;
73e4f7b9 332
e2565a42 333 __va_start(va, ctl);
73e4f7b9 334 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 335 __va_end(va);
7d0bac62
MD
336}
337
99df837e 338void
73e4f7b9 339lwkt_hold(thread_t td)
99df837e 340{
73e4f7b9
MD
341 ++td->td_refs;
342}
343
344void
345lwkt_rele(thread_t td)
346{
347 KKASSERT(td->td_refs > 0);
348 --td->td_refs;
349}
350
c95cd171
MD
351#ifdef _KERNEL
352
73e4f7b9
MD
353void
354lwkt_wait_free(thread_t td)
355{
356 while (td->td_refs)
377d4740 357 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
358}
359
c95cd171
MD
360#endif
361
73e4f7b9
MD
362void
363lwkt_free_thread(thread_t td)
364{
365 struct globaldata *gd = mycpu;
366
d9eea1a5 367 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
368 ("lwkt_free_thread: did not exit! %p", td));
369
37af14fe 370 crit_enter_gd(gd);
73e4f7b9
MD
371 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
372 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
373 (td->td_flags & TDF_ALLOCATED_THREAD)
374 ) {
73e4f7b9
MD
375 ++gd->gd_tdfreecount;
376 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
37af14fe 377 crit_exit_gd(gd);
99df837e 378 } else {
37af14fe 379 crit_exit_gd(gd);
99df837e 380 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613
MD
381#ifdef _KERNEL
382 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK);
383#else
fb04f4fd 384 libcaps_free_stack(td->td_kstack, THREAD_STACK);
05220613 385#endif
73e4f7b9 386 /* gd invalid */
99df837e
MD
387 td->td_kstack = NULL;
388 }
05220613
MD
389 if (td->td_flags & TDF_ALLOCATED_THREAD) {
390#ifdef _KERNEL
99df837e 391 zfree(thread_zone, td);
05220613
MD
392#else
393 free(td);
394#endif
395 }
99df837e
MD
396 }
397}
398
399
8ad65e08
MD
400/*
401 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
402 * switch to the idlethread. Switching must occur within a critical
403 * section to avoid races with the scheduling queue.
404 *
405 * We always have full control over our cpu's run queue. Other cpus
406 * that wish to manipulate our queue must use the cpu_*msg() calls to
407 * talk to our cpu, so a critical section is all that is needed and
408 * the result is very, very fast thread switching.
409 *
96728c05
MD
410 * The LWKT scheduler uses a fixed priority model and round-robins at
411 * each priority level. User process scheduling is a totally
412 * different beast and LWKT priorities should not be confused with
413 * user process priorities.
f1d1c3fa 414 *
96728c05
MD
415 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
416 * cleans it up. Note that the td_switch() function cannot do anything that
417 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
418 * the target thread (not the current thread). It's nice to have a scheduler
419 * that does not need the MP lock to work because it allows us to do some
420 * really cool high-performance MP lock optimizations.
8ad65e08 421 */
96728c05 422
8ad65e08
MD
423void
424lwkt_switch(void)
425{
37af14fe
MD
426 globaldata_t gd = mycpu;
427 thread_t td = gd->gd_curthread;
8ad65e08 428 thread_t ntd;
8a8d5d85
MD
429#ifdef SMP
430 int mpheld;
431#endif
8ad65e08 432
46a3f46d
MD
433 /*
434 * Switching from within a 'fast' (non thread switched) interrupt is
435 * illegal.
436 */
37af14fe 437 if (gd->gd_intr_nesting_level && panicstr == NULL) {
fc92d4aa 438 panic("lwkt_switch: cannot switch from within a fast interrupt, yet");
96728c05 439 }
ef0fdad1 440
cb973d15
MD
441 /*
442 * Passive release (used to transition from user to kernel mode
443 * when we block or switch rather then when we enter the kernel).
444 * This function is NOT called if we are switching into a preemption
445 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
446 * our current process designation (if we have one) and become a true
447 * LWKT thread, and may also hand the current process designation to
448 * another process and schedule thread.
cb973d15
MD
449 */
450 if (td->td_release)
451 td->td_release(td);
452
37af14fe 453 crit_enter_gd(gd);
4b5f931b 454 ++switch_count;
8a8d5d85
MD
455
456#ifdef SMP
457 /*
458 * td_mpcount cannot be used to determine if we currently hold the
459 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
460 * to get the lock, and switch out if it can't. Our ownership of
461 * the actual lock will remain stable while we are in a critical section
462 * (but, of course, another cpu may own or release the lock so the
463 * actual value of mp_lock is not stable).
8a8d5d85
MD
464 */
465 mpheld = MP_LOCK_HELD();
0f7a3396
MD
466#ifdef INVARIANTS
467 if (td->td_cscount) {
468 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
469 td);
470 if (panic_on_cscount)
471 panic("switching while mastering cpusync");
472 }
473#endif
8a8d5d85 474#endif
99df837e
MD
475 if ((ntd = td->td_preempted) != NULL) {
476 /*
477 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
478 * thread. This occurs transparently, whether the preempted thread
479 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
480 * itself).
481 *
482 * We have to setup the MP lock for the original thread after backing
483 * out the adjustment that was made to curthread when the original
484 * was preempted.
99df837e 485 */
26a0694b 486 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 487#ifdef SMP
96728c05 488 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 489 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
490 td, ntd, td->td_mpcount, ntd->td_mpcount);
491 }
8a8d5d85
MD
492 if (ntd->td_mpcount) {
493 td->td_mpcount -= ntd->td_mpcount;
494 KKASSERT(td->td_mpcount >= 0);
495 }
496#endif
26a0694b 497 ntd->td_flags |= TDF_PREEMPT_DONE;
8a8d5d85 498 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 499 } else {
4b5f931b
MD
500 /*
501 * Priority queue / round-robin at each priority. Note that user
502 * processes run at a fixed, low priority and the user process
503 * scheduler deals with interactions between user processes
504 * by scheduling and descheduling them from the LWKT queue as
505 * necessary.
8a8d5d85
MD
506 *
507 * We have to adjust the MP lock for the target thread. If we
508 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
509 * thread that does not need the MP lock. If we cannot, we spin
510 * instead of HLT.
511 *
512 * A similar issue exists for the tokens held by the target thread.
513 * If we cannot obtain ownership of the tokens we cannot immediately
514 * schedule the thread.
515 */
516
517 /*
518 * We are switching threads. If there are any pending requests for
519 * tokens we can satisfy all of them here.
4b5f931b 520 */
41a01a4d
MD
521#ifdef SMP
522 if (gd->gd_tokreqbase)
523 lwkt_drain_token_requests();
524#endif
525
4b5f931b
MD
526again:
527 if (gd->gd_runqmask) {
528 int nq = bsrl(gd->gd_runqmask);
529 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
530 gd->gd_runqmask &= ~(1 << nq);
531 goto again;
532 }
8a8d5d85 533#ifdef SMP
41a01a4d
MD
534 /*
535 * If the target needs the MP lock and we couldn't get it,
536 * or if the target is holding tokens and we could not
537 * gain ownership of the tokens, continue looking for a
538 * thread to schedule and spin instead of HLT if we can't.
539 */
540 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
541 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
542 ) {
8a8d5d85
MD
543 u_int32_t rqmask = gd->gd_runqmask;
544 while (rqmask) {
545 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
41a01a4d
MD
546 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock())
547 continue;
548 mpheld = MP_LOCK_HELD();
549 if (ntd->td_toks && !lwkt_chktokens(ntd))
550 continue;
551 break;
8a8d5d85
MD
552 }
553 if (ntd)
554 break;
555 rqmask &= ~(1 << nq);
556 nq = bsrl(rqmask);
557 }
558 if (ntd == NULL) {
a2a5ad0d
MD
559 ntd = &gd->gd_idlethread;
560 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
561 } else {
562 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
563 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
564 }
565 } else {
566 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
567 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
568 }
569#else
4b5f931b
MD
570 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
571 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 572#endif
4b5f931b 573 } else {
3c23a41a 574 /*
60f945af
MD
575 * We have nothing to run but only let the idle loop halt
576 * the cpu if there are no pending interrupts.
3c23a41a 577 */
a2a5ad0d 578 ntd = &gd->gd_idlethread;
60f945af 579 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 580 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 581 }
f1d1c3fa 582 }
26a0694b
MD
583 KASSERT(ntd->td_pri >= TDPRI_CRIT,
584 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
585
586 /*
587 * Do the actual switch. If the new target does not need the MP lock
588 * and we are holding it, release the MP lock. If the new target requires
589 * the MP lock we have already acquired it for the target.
590 */
591#ifdef SMP
592 if (ntd->td_mpcount == 0 ) {
593 if (MP_LOCK_HELD())
594 cpu_rel_mplock();
595 } else {
596 ASSERT_MP_LOCK_HELD();
597 }
598#endif
37af14fe 599 if (td != ntd)
f1d1c3fa 600 td->td_switch(ntd);
37af14fe
MD
601 /* NOTE: current cpu may have changed after switch */
602 crit_exit_quick(td);
8ad65e08
MD
603}
604
b68b7282 605/*
96728c05
MD
606 * Request that the target thread preempt the current thread. Preemption
607 * only works under a specific set of conditions:
b68b7282 608 *
96728c05
MD
609 * - We are not preempting ourselves
610 * - The target thread is owned by the current cpu
611 * - We are not currently being preempted
612 * - The target is not currently being preempted
613 * - We are able to satisfy the target's MP lock requirements (if any).
614 *
615 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
616 * this is called via lwkt_schedule() through the td_preemptable callback.
617 * critpri is the managed critical priority that we should ignore in order
618 * to determine whether preemption is possible (aka usually just the crit
619 * priority of lwkt_schedule() itself).
b68b7282 620 *
26a0694b
MD
621 * XXX at the moment we run the target thread in a critical section during
622 * the preemption in order to prevent the target from taking interrupts
623 * that *WE* can't. Preemption is strictly limited to interrupt threads
624 * and interrupt-like threads, outside of a critical section, and the
625 * preempted source thread will be resumed the instant the target blocks
626 * whether or not the source is scheduled (i.e. preemption is supposed to
627 * be as transparent as possible).
4b5f931b 628 *
8a8d5d85
MD
629 * The target thread inherits our MP count (added to its own) for the
630 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
631 * MP lock during the preemption. Therefore, any preempting targets must be
632 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
633 * out of sync with the physical mp_lock, but we do not have to preserve
634 * the original ownership of the lock if it was out of synch (that is, we
635 * can leave it synchronized on return).
b68b7282
MD
636 */
637void
96728c05 638lwkt_preempt(thread_t ntd, int critpri)
b68b7282 639{
46a3f46d 640 struct globaldata *gd = mycpu;
0a3f9b47 641 thread_t td;
8a8d5d85
MD
642#ifdef SMP
643 int mpheld;
57c254db 644 int savecnt;
8a8d5d85 645#endif
b68b7282 646
26a0694b 647 /*
96728c05
MD
648 * The caller has put us in a critical section. We can only preempt
649 * if the caller of the caller was not in a critical section (basically
0a3f9b47 650 * a local interrupt), as determined by the 'critpri' parameter.
96728c05
MD
651 *
652 * YYY The target thread must be in a critical section (else it must
653 * inherit our critical section? I dunno yet).
41a01a4d
MD
654 *
655 * Any tokens held by the target may not be held by thread(s) being
656 * preempted. We take the easy way out and do not preempt if
657 * the target is holding tokens.
0a3f9b47
MD
658 *
659 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
660 */
661 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 662
0a3f9b47
MD
663 td = gd->gd_curthread;
664 need_lwkt_resched();
665 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
666 ++preempt_miss;
667 return;
668 }
96728c05
MD
669 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
670 ++preempt_miss;
671 return;
672 }
673#ifdef SMP
46a3f46d 674 if (ntd->td_gd != gd) {
96728c05
MD
675 ++preempt_miss;
676 return;
677 }
678#endif
41a01a4d
MD
679 /*
680 * Take the easy way out and do not preempt if the target is holding
681 * one or more tokens. We could test whether the thread(s) being
682 * preempted interlock against the target thread's tokens and whether
683 * we can get all the target thread's tokens, but this situation
684 * should not occur very often so its easier to simply not preempt.
685 */
686 if (ntd->td_toks != NULL) {
687 ++preempt_miss;
688 return;
689 }
26a0694b
MD
690 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
691 ++preempt_weird;
692 return;
693 }
694 if (ntd->td_preempted) {
4b5f931b 695 ++preempt_hit;
26a0694b 696 return;
b68b7282 697 }
8a8d5d85 698#ifdef SMP
a2a5ad0d
MD
699 /*
700 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
701 * to or from the MP lock. In this case td_mpcount will be pre-disposed
702 * (non-zero) but not actually synchronized with the actual state of the
703 * lock. We can use it to imply an MP lock requirement for the
704 * preemption but we cannot use it to test whether we hold the MP lock
705 * or not.
a2a5ad0d 706 */
96728c05 707 savecnt = td->td_mpcount;
71ef2f5c 708 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
709 ntd->td_mpcount += td->td_mpcount;
710 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
711 ntd->td_mpcount -= td->td_mpcount;
712 ++preempt_miss;
713 return;
714 }
715#endif
26a0694b
MD
716
717 ++preempt_hit;
718 ntd->td_preempted = td;
719 td->td_flags |= TDF_PREEMPT_LOCK;
720 td->td_switch(ntd);
721 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
722#ifdef SMP
723 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
724 mpheld = MP_LOCK_HELD();
725 if (mpheld && td->td_mpcount == 0)
96728c05 726 cpu_rel_mplock();
71ef2f5c 727 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
728 panic("lwkt_preempt(): MP lock was not held through");
729#endif
26a0694b
MD
730 ntd->td_preempted = NULL;
731 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
732}
733
f1d1c3fa
MD
734/*
735 * Yield our thread while higher priority threads are pending. This is
736 * typically called when we leave a critical section but it can be safely
737 * called while we are in a critical section.
738 *
739 * This function will not generally yield to equal priority threads but it
740 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 741 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
742 * lwkt_yield_quick().
743 *
235957ed 744 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
745 * came along but was blocked and made pending.
746 *
f1d1c3fa
MD
747 * (self contained on a per cpu basis)
748 */
749void
750lwkt_yield_quick(void)
751{
7966cb69
MD
752 globaldata_t gd = mycpu;
753 thread_t td = gd->gd_curthread;
ef0fdad1 754
a2a5ad0d 755 /*
235957ed 756 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
757 * it with a non-zero cpl then we might not wind up calling splz after
758 * a task switch when the critical section is exited even though the
46a3f46d 759 * new task could accept the interrupt.
a2a5ad0d
MD
760 *
761 * XXX from crit_exit() only called after last crit section is released.
762 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
763 *
764 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
765 * except for this special case, we MUST call splz() here to handle any
766 * pending ints, particularly after we switch, or we might accidently
767 * halt the cpu with interrupts pending.
a2a5ad0d 768 */
46a3f46d 769 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 770 splz();
f1d1c3fa
MD
771
772 /*
773 * YYY enabling will cause wakeup() to task-switch, which really
774 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
775 * preemption and MP without actually doing preemption or MP, because a
776 * lot of code assumes that wakeup() does not block.
f1d1c3fa 777 */
46a3f46d
MD
778 if (untimely_switch && td->td_nest_count == 0 &&
779 gd->gd_intr_nesting_level == 0
780 ) {
37af14fe 781 crit_enter_quick(td);
f1d1c3fa
MD
782 /*
783 * YYY temporary hacks until we disassociate the userland scheduler
784 * from the LWKT scheduler.
785 */
786 if (td->td_flags & TDF_RUNQ) {
787 lwkt_switch(); /* will not reenter yield function */
788 } else {
37af14fe 789 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 790 lwkt_switch(); /* will not reenter yield function */
37af14fe 791 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 792 }
7966cb69 793 crit_exit_noyield(td);
f1d1c3fa 794 }
f1d1c3fa
MD
795}
796
8ad65e08 797/*
f1d1c3fa 798 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 799 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
800 * the crit_exit() call in lwkt_switch().
801 *
802 * (self contained on a per cpu basis)
8ad65e08
MD
803 */
804void
f1d1c3fa 805lwkt_yield(void)
8ad65e08 806{
37af14fe 807 lwkt_schedule_self(curthread);
f1d1c3fa
MD
808 lwkt_switch();
809}
810
8ad65e08 811/*
f1d1c3fa
MD
812 * Generic schedule. Possibly schedule threads belonging to other cpus and
813 * deal with threads that might be blocked on a wait queue.
814 *
0a3f9b47
MD
815 * We have a little helper inline function which does additional work after
816 * the thread has been enqueued, including dealing with preemption and
817 * setting need_lwkt_resched() (which prevents the kernel from returning
818 * to userland until it has processed higher priority threads).
8ad65e08 819 */
0a3f9b47
MD
820static __inline
821void
822_lwkt_schedule_post(thread_t ntd, int cpri)
823{
824 if (ntd->td_preemptable) {
825 ntd->td_preemptable(ntd, cpri); /* YYY +token */
826 } else {
827 if ((ntd->td_flags & TDF_NORESCHED) == 0) {
828 if ((ntd->td_pri & TDPRI_MASK) >= TDPRI_KERN_USER)
829 need_lwkt_resched();
830 }
831 }
832}
833
8ad65e08
MD
834void
835lwkt_schedule(thread_t td)
836{
37af14fe
MD
837 globaldata_t mygd = mycpu;
838
96728c05 839#ifdef INVARIANTS
41a01a4d 840 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
26a0694b
MD
841 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
842 && td->td_proc->p_stat == SSLEEP
843 ) {
844 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
845 curthread,
846 curthread->td_proc ? curthread->td_proc->p_pid : -1,
847 curthread->td_proc ? curthread->td_proc->p_stat : -1,
848 td,
849 td->td_proc ? curthread->td_proc->p_pid : -1,
850 td->td_proc ? curthread->td_proc->p_stat : -1
851 );
852 panic("SCHED PANIC");
853 }
96728c05 854#endif
37af14fe
MD
855 crit_enter_gd(mygd);
856 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
857 _lwkt_enqueue(td);
858 } else {
859 lwkt_wait_t w;
860
861 /*
862 * If the thread is on a wait list we have to send our scheduling
863 * request to the owner of the wait structure. Otherwise we send
864 * the scheduling request to the cpu owning the thread. Races
865 * are ok, the target will forward the message as necessary (the
866 * message may chase the thread around before it finally gets
867 * acted upon).
868 *
869 * (remember, wait structures use stable storage)
0a3f9b47
MD
870 *
871 * NOTE: tokens no longer enter a critical section, so we only need
872 * to account for the crit_enter() above when calling
873 * _lwkt_schedule_post().
f1d1c3fa
MD
874 */
875 if ((w = td->td_wait) != NULL) {
41a01a4d
MD
876 lwkt_tokref wref;
877
878 if (lwkt_trytoken(&wref, &w->wa_token)) {
f1d1c3fa
MD
879 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
880 --w->wa_count;
881 td->td_wait = NULL;
0f7a3396
MD
882#ifdef SMP
883 if (td->td_gd == mycpu) {
f1d1c3fa 884 _lwkt_enqueue(td);
0a3f9b47 885 _lwkt_schedule_post(td, TDPRI_CRIT);
f1d1c3fa 886 } else {
2db3b277 887 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 888 }
0f7a3396
MD
889#else
890 _lwkt_enqueue(td);
0a3f9b47 891 _lwkt_schedule_post(td, TDPRI_CRIT);
0f7a3396 892#endif
41a01a4d 893 lwkt_reltoken(&wref);
f1d1c3fa 894 } else {
96728c05 895 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
896 }
897 } else {
898 /*
899 * If the wait structure is NULL and we own the thread, there
900 * is no race (since we are in a critical section). If we
901 * do not own the thread there might be a race but the
902 * target cpu will deal with it.
903 */
0f7a3396 904#ifdef SMP
37af14fe 905 if (td->td_gd == mygd) {
f1d1c3fa 906 _lwkt_enqueue(td);
0a3f9b47 907 _lwkt_schedule_post(td, TDPRI_CRIT);
f1d1c3fa 908 } else {
2db3b277 909 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 910 }
0f7a3396
MD
911#else
912 _lwkt_enqueue(td);
0a3f9b47 913 _lwkt_schedule_post(td, TDPRI_CRIT);
0f7a3396 914#endif
f1d1c3fa 915 }
8ad65e08 916 }
37af14fe 917 crit_exit_gd(mygd);
8ad65e08
MD
918}
919
d9eea1a5
MD
920/*
921 * Managed acquisition. This code assumes that the MP lock is held for
922 * the tdallq operation and that the thread has been descheduled from its
923 * original cpu. We also have to wait for the thread to be entirely switched
924 * out on its original cpu (this is usually fast enough that we never loop)
925 * since the LWKT system does not have to hold the MP lock while switching
926 * and the target may have released it before switching.
927 */
a2a5ad0d
MD
928void
929lwkt_acquire(thread_t td)
930{
37af14fe
MD
931 globaldata_t gd;
932 globaldata_t mygd;
a2a5ad0d
MD
933
934 gd = td->td_gd;
37af14fe 935 mygd = mycpu;
a2a5ad0d 936 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5 937 while (td->td_flags & TDF_RUNNING) /* XXX spin */
5d21b981 938 cpu_mb1();
37af14fe
MD
939 if (gd != mygd) {
940 crit_enter_gd(mygd);
a2a5ad0d 941 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
37af14fe
MD
942 td->td_gd = mygd;
943 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
944 crit_exit_gd(mygd);
a2a5ad0d
MD
945 }
946}
947
f1d1c3fa
MD
948/*
949 * Generic deschedule. Descheduling threads other then your own should be
950 * done only in carefully controlled circumstances. Descheduling is
951 * asynchronous.
952 *
953 * This function may block if the cpu has run out of messages.
8ad65e08
MD
954 */
955void
956lwkt_deschedule(thread_t td)
957{
f1d1c3fa
MD
958 crit_enter();
959 if (td == curthread) {
960 _lwkt_dequeue(td);
961 } else {
a72187e9 962 if (td->td_gd == mycpu) {
f1d1c3fa
MD
963 _lwkt_dequeue(td);
964 } else {
2db3b277 965 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
966 }
967 }
968 crit_exit();
969}
970
4b5f931b
MD
971/*
972 * Set the target thread's priority. This routine does not automatically
973 * switch to a higher priority thread, LWKT threads are not designed for
974 * continuous priority changes. Yield if you want to switch.
975 *
976 * We have to retain the critical section count which uses the high bits
26a0694b
MD
977 * of the td_pri field. The specified priority may also indicate zero or
978 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
979 *
980 * Note that we requeue the thread whether it winds up on a different runq
981 * or not. uio_yield() depends on this and the routine is not normally
982 * called with the same priority otherwise.
4b5f931b
MD
983 */
984void
985lwkt_setpri(thread_t td, int pri)
986{
26a0694b 987 KKASSERT(pri >= 0);
a72187e9 988 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
989 crit_enter();
990 if (td->td_flags & TDF_RUNQ) {
991 _lwkt_dequeue(td);
992 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
993 _lwkt_enqueue(td);
994 } else {
995 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
996 }
997 crit_exit();
998}
999
1000void
1001lwkt_setpri_self(int pri)
1002{
1003 thread_t td = curthread;
1004
4b5f931b
MD
1005 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1006 crit_enter();
1007 if (td->td_flags & TDF_RUNQ) {
1008 _lwkt_dequeue(td);
1009 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1010 _lwkt_enqueue(td);
1011 } else {
1012 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1013 }
1014 crit_exit();
1015}
1016
5d21b981
MD
1017/*
1018 * Migrate the current thread to the specified cpu. The BGL must be held
1019 * (for the gd_tdallq manipulation XXX). This is accomplished by
1020 * descheduling ourselves from the current cpu, moving our thread to the
1021 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1022 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1023 */
1024static void lwkt_setcpu_remote(void *arg);
1025
1026void
1027lwkt_setcpu_self(globaldata_t rgd)
1028{
1029#ifdef SMP
1030 thread_t td = curthread;
1031
1032 if (td->td_gd != rgd) {
1033 crit_enter_quick(td);
1034 td->td_flags |= TDF_MIGRATING;
1035 lwkt_deschedule_self(td);
1036 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1037 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
1038 lwkt_send_ipiq(rgd, (ipifunc_t)lwkt_setcpu_remote, td);
1039 lwkt_switch();
1040 /* we are now on the target cpu */
1041 crit_exit_quick(td);
1042 }
1043#endif
1044}
1045
1046/*
1047 * Remote IPI for cpu migration (called while in a critical section so we
1048 * do not have to enter another one). The thread has already been moved to
1049 * our cpu's allq, but we must wait for the thread to be completely switched
1050 * out on the originating cpu before we schedule it on ours or the stack
1051 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1052 * change to main memory.
1053 *
1054 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1055 * against wakeups. It is best if this interface is used only when there
1056 * are no pending events that might try to schedule the thread.
1057 */
1058static void
1059lwkt_setcpu_remote(void *arg)
1060{
1061 thread_t td = arg;
1062 globaldata_t gd = mycpu;
1063
1064 while (td->td_flags & TDF_RUNNING)
1065 cpu_mb1();
1066 td->td_gd = gd;
1067 cpu_mb2();
1068 td->td_flags &= ~TDF_MIGRATING;
1069 _lwkt_enqueue(td);
1070}
1071
4b5f931b
MD
1072struct proc *
1073lwkt_preempted_proc(void)
1074{
73e4f7b9 1075 thread_t td = curthread;
4b5f931b
MD
1076 while (td->td_preempted)
1077 td = td->td_preempted;
1078 return(td->td_proc);
1079}
1080
f1d1c3fa 1081/*
41a01a4d
MD
1082 * Block on the specified wait queue until signaled. A generation number
1083 * must be supplied to interlock the wait queue. The function will
1084 * return immediately if the generation number does not match the wait
1085 * structure's generation number.
f1d1c3fa
MD
1086 */
1087void
ae8050a4 1088lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
1089{
1090 thread_t td = curthread;
41a01a4d 1091 lwkt_tokref ilock;
f1d1c3fa 1092
41a01a4d
MD
1093 lwkt_gettoken(&ilock, &w->wa_token);
1094 crit_enter();
ae8050a4 1095 if (w->wa_gen == *gen) {
f1d1c3fa
MD
1096 _lwkt_dequeue(td);
1097 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1098 ++w->wa_count;
1099 td->td_wait = w;
ae8050a4 1100 td->td_wmesg = wmesg;
41a01a4d 1101 again:
f1d1c3fa 1102 lwkt_switch();
ece04fd0
MD
1103 if (td->td_wmesg != NULL) {
1104 _lwkt_dequeue(td);
1105 goto again;
1106 }
8ad65e08 1107 }
41a01a4d 1108 crit_exit();
ae8050a4 1109 *gen = w->wa_gen;
41a01a4d 1110 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1111}
1112
1113/*
1114 * Signal a wait queue. We gain ownership of the wait queue in order to
1115 * signal it. Once a thread is removed from the wait queue we have to
1116 * deal with the cpu owning the thread.
1117 *
1118 * Note: alternatively we could message the target cpu owning the wait
1119 * queue. YYY implement as sysctl.
1120 */
1121void
ece04fd0 1122lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1123{
1124 thread_t td;
41a01a4d 1125 lwkt_tokref ilock;
f1d1c3fa 1126
41a01a4d 1127 lwkt_gettoken(&ilock, &w->wa_token);
f1d1c3fa 1128 ++w->wa_gen;
41a01a4d 1129 crit_enter();
ece04fd0
MD
1130 if (count < 0)
1131 count = w->wa_count;
f1d1c3fa
MD
1132 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1133 --count;
1134 --w->wa_count;
1135 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1136 td->td_wait = NULL;
ae8050a4 1137 td->td_wmesg = NULL;
a72187e9 1138 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1139 _lwkt_enqueue(td);
1140 } else {
2db3b277 1141 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 1142 }
f1d1c3fa 1143 }
41a01a4d
MD
1144 crit_exit();
1145 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1146}
1147
99df837e
MD
1148/*
1149 * Create a kernel process/thread/whatever. It shares it's address space
1150 * with proc0 - ie: kernel only.
1151 *
365fa13f
MD
1152 * NOTE! By default new threads are created with the MP lock held. A
1153 * thread which does not require the MP lock should release it by calling
1154 * rel_mplock() at the start of the new thread.
99df837e
MD
1155 */
1156int
1157lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1158 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1159 const char *fmt, ...)
99df837e 1160{
73e4f7b9 1161 thread_t td;
e2565a42 1162 __va_list ap;
99df837e 1163
75cdbe6c 1164 td = lwkt_alloc_thread(template, cpu);
a2a5ad0d
MD
1165 if (tdp)
1166 *tdp = td;
709799ea 1167 cpu_set_thread_handler(td, lwkt_exit, func, arg);
ef0fdad1 1168 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1169#ifdef SMP
1170 td->td_mpcount = 1;
1171#endif
99df837e
MD
1172
1173 /*
1174 * Set up arg0 for 'ps' etc
1175 */
e2565a42 1176 __va_start(ap, fmt);
99df837e 1177 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1178 __va_end(ap);
99df837e
MD
1179
1180 /*
1181 * Schedule the thread to run
1182 */
ef0fdad1
MD
1183 if ((td->td_flags & TDF_STOPREQ) == 0)
1184 lwkt_schedule(td);
1185 else
1186 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1187 return 0;
1188}
1189
2d93b37a 1190/*
2d93b37a
MD
1191 * kthread_* is specific to the kernel and is not needed by userland.
1192 */
1193#ifdef _KERNEL
1194
99df837e
MD
1195/*
1196 * Destroy an LWKT thread. Warning! This function is not called when
1197 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1198 * uses a different reaping mechanism.
1199 */
1200void
1201lwkt_exit(void)
1202{
1203 thread_t td = curthread;
1204
1205 if (td->td_flags & TDF_VERBOSE)
1206 printf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1207 caps_exit(td);
37af14fe
MD
1208 crit_enter_quick(td);
1209 lwkt_deschedule_self(td);
99df837e
MD
1210 ++mycpu->gd_tdfreecount;
1211 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1212 cpu_thread_exit();
1213}
1214
1215/*
1216 * Create a kernel process/thread/whatever. It shares it's address space
ef0fdad1 1217 * with proc0 - ie: kernel only. 5.x compatible.
365fa13f
MD
1218 *
1219 * NOTE! By default kthreads are created with the MP lock held. A
1220 * thread which does not require the MP lock should release it by calling
1221 * rel_mplock() at the start of the new thread.
99df837e
MD
1222 */
1223int
1224kthread_create(void (*func)(void *), void *arg,
1225 struct thread **tdp, const char *fmt, ...)
1226{
73e4f7b9 1227 thread_t td;
e2565a42 1228 __va_list ap;
99df837e 1229
75cdbe6c 1230 td = lwkt_alloc_thread(NULL, -1);
a2a5ad0d
MD
1231 if (tdp)
1232 *tdp = td;
99df837e
MD
1233 cpu_set_thread_handler(td, kthread_exit, func, arg);
1234 td->td_flags |= TDF_VERBOSE;
8a8d5d85
MD
1235#ifdef SMP
1236 td->td_mpcount = 1;
1237#endif
99df837e
MD
1238
1239 /*
1240 * Set up arg0 for 'ps' etc
1241 */
e2565a42 1242 __va_start(ap, fmt);
99df837e 1243 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1244 __va_end(ap);
99df837e
MD
1245
1246 /*
1247 * Schedule the thread to run
1248 */
1249 lwkt_schedule(td);
1250 return 0;
1251}
1252
1253/*
1254 * Destroy an LWKT thread. Warning! This function is not called when
1255 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1256 * uses a different reaping mechanism.
1257 *
1258 * XXX duplicates lwkt_exit()
1259 */
1260void
1261kthread_exit(void)
1262{
1263 lwkt_exit();
1264}
1265
2d93b37a
MD
1266#endif /* _KERNEL */
1267
1268void
1269crit_panic(void)
1270{
1271 thread_t td = curthread;
1272 int lpri = td->td_pri;
1273
1274 td->td_pri = 0;
1275 panic("td_pri is/would-go negative! %p %d", td, lpri);
1276}
1277