Add a new option "TIMER_USE_1", which allows switching the primary heart
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
8c10bfcf
MD
2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 *
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:
8c10bfcf 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.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
8c10bfcf 33 *
496907b0 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.68 2004/07/29 09:02:33 dillon Exp $
75cdbe6c
MD
35 */
36
37/*
38 * Each cpu in a system has its own self-contained light weight kernel
39 * thread scheduler, which means that generally speaking we only need
40 * to use a critical section to avoid problems. Foreign thread
41 * scheduling is queued via (async) IPIs.
8ad65e08
MD
42 */
43
05220613
MD
44#ifdef _KERNEL
45
8ad65e08
MD
46#include <sys/param.h>
47#include <sys/systm.h>
48#include <sys/kernel.h>
49#include <sys/proc.h>
50#include <sys/rtprio.h>
51#include <sys/queue.h>
f1d1c3fa 52#include <sys/thread2.h>
7d0bac62 53#include <sys/sysctl.h>
99df837e 54#include <sys/kthread.h>
f1d1c3fa 55#include <machine/cpu.h>
99df837e 56#include <sys/lock.h>
f6bf3af1 57#include <sys/caps.h>
f1d1c3fa 58
7d0bac62
MD
59#include <vm/vm.h>
60#include <vm/vm_param.h>
61#include <vm/vm_kern.h>
62#include <vm/vm_object.h>
63#include <vm/vm_page.h>
64#include <vm/vm_map.h>
65#include <vm/vm_pager.h>
66#include <vm/vm_extern.h>
67#include <vm/vm_zone.h>
68
99df837e 69#include <machine/stdarg.h>
57c254db 70#include <machine/ipl.h>
96728c05 71#include <machine/smp.h>
99df837e 72
05220613
MD
73#else
74
75#include <sys/stdint.h>
fb04f4fd 76#include <libcaps/thread.h>
05220613
MD
77#include <sys/thread.h>
78#include <sys/msgport.h>
79#include <sys/errno.h>
fb04f4fd 80#include <libcaps/globaldata.h>
7e8303ad 81#include <machine/cpufunc.h>
05220613
MD
82#include <sys/thread2.h>
83#include <sys/msgport2.h>
709799ea 84#include <stdio.h>
05220613 85#include <stdlib.h>
709799ea 86#include <string.h>
709799ea 87#include <machine/lock.h>
05220613
MD
88
89#endif
90
7d0bac62 91static int untimely_switch = 0;
0f7a3396
MD
92#ifdef INVARIANTS
93static int panic_on_cscount = 0;
94#endif
05220613
MD
95static __int64_t switch_count = 0;
96static __int64_t preempt_hit = 0;
97static __int64_t preempt_miss = 0;
98static __int64_t preempt_weird = 0;
05220613
MD
99
100#ifdef _KERNEL
101
102SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
0f7a3396
MD
103#ifdef INVARIANTS
104SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
105#endif
4b5f931b 106SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 107SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 108SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 109SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
7d0bac62 110
05220613
MD
111#endif
112
4b5f931b
MD
113/*
114 * These helper procedures handle the runq, they can only be called from
115 * within a critical section.
75cdbe6c
MD
116 *
117 * WARNING! Prior to SMP being brought up it is possible to enqueue and
118 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
119 * instead of 'mycpu' when referencing the globaldata structure. Once
120 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 121 */
f1d1c3fa
MD
122static __inline
123void
124_lwkt_dequeue(thread_t td)
125{
126 if (td->td_flags & TDF_RUNQ) {
4b5f931b 127 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 128 struct globaldata *gd = td->td_gd;
4b5f931b 129
f1d1c3fa 130 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
131 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
132 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
133 }
134}
135
136static __inline
137void
138_lwkt_enqueue(thread_t td)
139{
5d21b981 140 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING)) == 0) {
4b5f931b 141 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 142 struct globaldata *gd = td->td_gd;
4b5f931b 143
f1d1c3fa 144 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
145 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
146 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
147 }
148}
8ad65e08 149
37af14fe
MD
150/*
151 * Schedule a thread to run. As the current thread we can always safely
152 * schedule ourselves, and a shortcut procedure is provided for that
153 * function.
154 *
155 * (non-blocking, self contained on a per cpu basis)
156 */
157void
158lwkt_schedule_self(thread_t td)
159{
160 crit_enter_quick(td);
161 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
162 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
163 _lwkt_enqueue(td);
164#ifdef _KERNEL
165 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
166 panic("SCHED SELF PANIC");
167#endif
168 crit_exit_quick(td);
169}
170
171/*
172 * Deschedule a thread.
173 *
174 * (non-blocking, self contained on a per cpu basis)
175 */
176void
177lwkt_deschedule_self(thread_t td)
178{
179 crit_enter_quick(td);
180 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
181 _lwkt_dequeue(td);
182 crit_exit_quick(td);
183}
184
2d93b37a
MD
185#ifdef _KERNEL
186
8ad65e08
MD
187/*
188 * LWKTs operate on a per-cpu basis
189 *
73e4f7b9 190 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
191 */
192void
193lwkt_gdinit(struct globaldata *gd)
194{
4b5f931b
MD
195 int i;
196
197 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
198 TAILQ_INIT(&gd->gd_tdrunq[i]);
199 gd->gd_runqmask = 0;
73e4f7b9 200 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
201}
202
2d93b37a
MD
203#endif /* _KERNEL */
204
7d0bac62
MD
205/*
206 * Initialize a thread wait structure prior to first use.
207 *
208 * NOTE! called from low level boot code, we cannot do anything fancy!
209 */
210void
41a01a4d 211lwkt_wait_init(lwkt_wait_t w)
7d0bac62 212{
41a01a4d 213 lwkt_token_init(&w->wa_token);
7d0bac62 214 TAILQ_INIT(&w->wa_waitq);
41a01a4d
MD
215 w->wa_gen = 0;
216 w->wa_count = 0;
7d0bac62
MD
217}
218
219/*
220 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
221 * or LWKT start address before it can be scheduled. If the target cpu is
222 * -1 the thread will be created on the current cpu.
0cfcada1
MD
223 *
224 * If you intend to create a thread without a process context this function
225 * does everything except load the startup and switcher function.
7d0bac62
MD
226 */
227thread_t
f470d0c8 228lwkt_alloc_thread(struct thread *td, int stksize, int cpu)
7d0bac62 229{
99df837e 230 void *stack;
ef0fdad1 231 int flags = 0;
37af14fe 232 globaldata_t gd = mycpu;
7d0bac62 233
ef0fdad1 234 if (td == NULL) {
37af14fe
MD
235 crit_enter_gd(gd);
236 if (gd->gd_tdfreecount > 0) {
237 --gd->gd_tdfreecount;
238 td = TAILQ_FIRST(&gd->gd_tdfreeq);
d9eea1a5 239 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1 240 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
37af14fe
MD
241 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq);
242 crit_exit_gd(gd);
ef0fdad1
MD
243 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
244 } else {
37af14fe 245 crit_exit_gd(gd);
05220613 246#ifdef _KERNEL
ef0fdad1 247 td = zalloc(thread_zone);
05220613
MD
248#else
249 td = malloc(sizeof(struct thread));
250#endif
ef0fdad1 251 td->td_kstack = NULL;
f470d0c8 252 td->td_kstack_size = 0;
ef0fdad1
MD
253 flags |= TDF_ALLOCATED_THREAD;
254 }
255 }
f470d0c8
MD
256 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
257 if (flags & TDF_ALLOCATED_STACK) {
258 kmem_free(kernel_map, (vm_offset_t)stack, td->td_kstack_size);
259 stack = NULL;
260 }
261 }
262 if (stack == NULL) {
05220613 263#ifdef _KERNEL
f470d0c8 264 stack = (void *)kmem_alloc(kernel_map, stksize);
05220613 265#else
f470d0c8 266 stack = libcaps_alloc_stack(stksize);
05220613 267#endif
ef0fdad1 268 flags |= TDF_ALLOCATED_STACK;
99df837e 269 }
75cdbe6c 270 if (cpu < 0)
f470d0c8 271 lwkt_init_thread(td, stack, stksize, flags, mycpu);
75cdbe6c 272 else
f470d0c8 273 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 274 return(td);
7d0bac62
MD
275}
276
709799ea
MD
277#ifdef _KERNEL
278
7d0bac62
MD
279/*
280 * Initialize a preexisting thread structure. This function is used by
281 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
282 *
f8c3996b
MD
283 * All threads start out in a critical section at a priority of
284 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
285 * appropriate. This function may send an IPI message when the
286 * requested cpu is not the current cpu and consequently gd_tdallq may
287 * not be initialized synchronously from the point of view of the originating
288 * cpu.
289 *
290 * NOTE! we have to be careful in regards to creating threads for other cpus
291 * if SMP has not yet been activated.
7d0bac62 292 */
41a01a4d
MD
293#ifdef SMP
294
75cdbe6c
MD
295static void
296lwkt_init_thread_remote(void *arg)
297{
298 thread_t td = arg;
299
300 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
301}
302
41a01a4d
MD
303#endif
304
7d0bac62 305void
f470d0c8
MD
306lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
307 struct globaldata *gd)
7d0bac62 308{
37af14fe
MD
309 globaldata_t mygd = mycpu;
310
99df837e
MD
311 bzero(td, sizeof(struct thread));
312 td->td_kstack = stack;
f470d0c8 313 td->td_kstack_size = stksize;
99df837e 314 td->td_flags |= flags;
26a0694b 315 td->td_gd = gd;
f8c3996b 316 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
c95cd171 317 lwkt_initport(&td->td_msgport, td);
99df837e 318 pmap_init_thread(td);
0f7a3396 319#ifdef SMP
5d21b981
MD
320 /*
321 * Normally initializing a thread for a remote cpu requires sending an
322 * IPI. However, the idlethread is setup before the other cpus are
323 * activated so we have to treat it as a special case. XXX manipulation
324 * of gd_tdallq requires the BGL.
325 */
326 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 327 crit_enter_gd(mygd);
75cdbe6c 328 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 329 crit_exit_gd(mygd);
75cdbe6c 330 } else {
2db3b277 331 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 332 }
0f7a3396 333#else
37af14fe 334 crit_enter_gd(mygd);
0f7a3396 335 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 336 crit_exit_gd(mygd);
0f7a3396 337#endif
73e4f7b9
MD
338}
339
2d93b37a
MD
340#endif /* _KERNEL */
341
73e4f7b9
MD
342void
343lwkt_set_comm(thread_t td, const char *ctl, ...)
344{
e2565a42 345 __va_list va;
73e4f7b9 346
e2565a42 347 __va_start(va, ctl);
73e4f7b9 348 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 349 __va_end(va);
7d0bac62
MD
350}
351
99df837e 352void
73e4f7b9 353lwkt_hold(thread_t td)
99df837e 354{
73e4f7b9
MD
355 ++td->td_refs;
356}
357
358void
359lwkt_rele(thread_t td)
360{
361 KKASSERT(td->td_refs > 0);
362 --td->td_refs;
363}
364
c95cd171
MD
365#ifdef _KERNEL
366
73e4f7b9
MD
367void
368lwkt_wait_free(thread_t td)
369{
370 while (td->td_refs)
377d4740 371 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
372}
373
c95cd171
MD
374#endif
375
73e4f7b9
MD
376void
377lwkt_free_thread(thread_t td)
378{
379 struct globaldata *gd = mycpu;
380
d9eea1a5 381 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
382 ("lwkt_free_thread: did not exit! %p", td));
383
37af14fe 384 crit_enter_gd(gd);
73e4f7b9
MD
385 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
386 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
387 (td->td_flags & TDF_ALLOCATED_THREAD)
388 ) {
73e4f7b9
MD
389 ++gd->gd_tdfreecount;
390 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
37af14fe 391 crit_exit_gd(gd);
99df837e 392 } else {
37af14fe 393 crit_exit_gd(gd);
99df837e 394 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613 395#ifdef _KERNEL
f470d0c8 396 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
05220613 397#else
f470d0c8 398 libcaps_free_stack(td->td_kstack, td->td_kstack_size);
05220613 399#endif
73e4f7b9 400 /* gd invalid */
99df837e 401 td->td_kstack = NULL;
f470d0c8 402 td->td_kstack_size = 0;
99df837e 403 }
05220613
MD
404 if (td->td_flags & TDF_ALLOCATED_THREAD) {
405#ifdef _KERNEL
99df837e 406 zfree(thread_zone, td);
05220613
MD
407#else
408 free(td);
409#endif
410 }
99df837e
MD
411 }
412}
413
414
8ad65e08
MD
415/*
416 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
417 * switch to the idlethread. Switching must occur within a critical
418 * section to avoid races with the scheduling queue.
419 *
420 * We always have full control over our cpu's run queue. Other cpus
421 * that wish to manipulate our queue must use the cpu_*msg() calls to
422 * talk to our cpu, so a critical section is all that is needed and
423 * the result is very, very fast thread switching.
424 *
96728c05
MD
425 * The LWKT scheduler uses a fixed priority model and round-robins at
426 * each priority level. User process scheduling is a totally
427 * different beast and LWKT priorities should not be confused with
428 * user process priorities.
f1d1c3fa 429 *
96728c05
MD
430 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
431 * cleans it up. Note that the td_switch() function cannot do anything that
432 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
433 * the target thread (not the current thread). It's nice to have a scheduler
434 * that does not need the MP lock to work because it allows us to do some
435 * really cool high-performance MP lock optimizations.
8ad65e08 436 */
96728c05 437
8ad65e08
MD
438void
439lwkt_switch(void)
440{
37af14fe
MD
441 globaldata_t gd = mycpu;
442 thread_t td = gd->gd_curthread;
8ad65e08 443 thread_t ntd;
8a8d5d85
MD
444#ifdef SMP
445 int mpheld;
446#endif
8ad65e08 447
46a3f46d
MD
448 /*
449 * Switching from within a 'fast' (non thread switched) interrupt is
450 * illegal.
451 */
37af14fe 452 if (gd->gd_intr_nesting_level && panicstr == NULL) {
fc92d4aa 453 panic("lwkt_switch: cannot switch from within a fast interrupt, yet");
96728c05 454 }
ef0fdad1 455
cb973d15
MD
456 /*
457 * Passive release (used to transition from user to kernel mode
458 * when we block or switch rather then when we enter the kernel).
459 * This function is NOT called if we are switching into a preemption
460 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
461 * our current process designation (if we have one) and become a true
462 * LWKT thread, and may also hand the current process designation to
463 * another process and schedule thread.
cb973d15
MD
464 */
465 if (td->td_release)
466 td->td_release(td);
467
37af14fe 468 crit_enter_gd(gd);
4b5f931b 469 ++switch_count;
8a8d5d85
MD
470
471#ifdef SMP
472 /*
473 * td_mpcount cannot be used to determine if we currently hold the
474 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
475 * to get the lock, and switch out if it can't. Our ownership of
476 * the actual lock will remain stable while we are in a critical section
477 * (but, of course, another cpu may own or release the lock so the
478 * actual value of mp_lock is not stable).
8a8d5d85
MD
479 */
480 mpheld = MP_LOCK_HELD();
0f7a3396
MD
481#ifdef INVARIANTS
482 if (td->td_cscount) {
483 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
484 td);
485 if (panic_on_cscount)
486 panic("switching while mastering cpusync");
487 }
488#endif
8a8d5d85 489#endif
99df837e
MD
490 if ((ntd = td->td_preempted) != NULL) {
491 /*
492 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
493 * thread. This occurs transparently, whether the preempted thread
494 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
495 * itself).
496 *
497 * We have to setup the MP lock for the original thread after backing
498 * out the adjustment that was made to curthread when the original
499 * was preempted.
99df837e 500 */
26a0694b 501 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 502#ifdef SMP
96728c05 503 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 504 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
505 td, ntd, td->td_mpcount, ntd->td_mpcount);
506 }
8a8d5d85
MD
507 if (ntd->td_mpcount) {
508 td->td_mpcount -= ntd->td_mpcount;
509 KKASSERT(td->td_mpcount >= 0);
510 }
511#endif
26a0694b 512 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
513
514 /*
515 * XXX. The interrupt may have woken a thread up, we need to properly
516 * set the reschedule flag if the originally interrupted thread is at
517 * a lower priority.
518 */
519 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
520 need_lwkt_resched();
8a8d5d85 521 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 522 } else {
4b5f931b
MD
523 /*
524 * Priority queue / round-robin at each priority. Note that user
525 * processes run at a fixed, low priority and the user process
526 * scheduler deals with interactions between user processes
527 * by scheduling and descheduling them from the LWKT queue as
528 * necessary.
8a8d5d85
MD
529 *
530 * We have to adjust the MP lock for the target thread. If we
531 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
532 * thread that does not need the MP lock. If we cannot, we spin
533 * instead of HLT.
534 *
535 * A similar issue exists for the tokens held by the target thread.
536 * If we cannot obtain ownership of the tokens we cannot immediately
537 * schedule the thread.
538 */
539
540 /*
541 * We are switching threads. If there are any pending requests for
542 * tokens we can satisfy all of them here.
4b5f931b 543 */
41a01a4d
MD
544#ifdef SMP
545 if (gd->gd_tokreqbase)
546 lwkt_drain_token_requests();
547#endif
548
8ec60c3f
MD
549 /*
550 * If an LWKT reschedule was requested, well that is what we are
551 * doing now so clear it.
552 */
553 clear_lwkt_resched();
4b5f931b
MD
554again:
555 if (gd->gd_runqmask) {
556 int nq = bsrl(gd->gd_runqmask);
557 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
558 gd->gd_runqmask &= ~(1 << nq);
559 goto again;
560 }
8a8d5d85 561#ifdef SMP
41a01a4d
MD
562 /*
563 * If the target needs the MP lock and we couldn't get it,
564 * or if the target is holding tokens and we could not
565 * gain ownership of the tokens, continue looking for a
566 * thread to schedule and spin instead of HLT if we can't.
567 */
568 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
569 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
570 ) {
8a8d5d85
MD
571 u_int32_t rqmask = gd->gd_runqmask;
572 while (rqmask) {
573 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
41a01a4d
MD
574 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock())
575 continue;
576 mpheld = MP_LOCK_HELD();
577 if (ntd->td_toks && !lwkt_chktokens(ntd))
578 continue;
579 break;
8a8d5d85
MD
580 }
581 if (ntd)
582 break;
583 rqmask &= ~(1 << nq);
584 nq = bsrl(rqmask);
585 }
586 if (ntd == NULL) {
a2a5ad0d
MD
587 ntd = &gd->gd_idlethread;
588 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
589 } else {
590 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
591 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
592 }
593 } else {
594 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
595 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
596 }
597#else
4b5f931b
MD
598 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
599 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 600#endif
4b5f931b 601 } else {
3c23a41a 602 /*
60f945af
MD
603 * We have nothing to run but only let the idle loop halt
604 * the cpu if there are no pending interrupts.
3c23a41a 605 */
a2a5ad0d 606 ntd = &gd->gd_idlethread;
60f945af 607 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 608 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 609 }
f1d1c3fa 610 }
26a0694b
MD
611 KASSERT(ntd->td_pri >= TDPRI_CRIT,
612 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
613
614 /*
615 * Do the actual switch. If the new target does not need the MP lock
616 * and we are holding it, release the MP lock. If the new target requires
617 * the MP lock we have already acquired it for the target.
618 */
619#ifdef SMP
620 if (ntd->td_mpcount == 0 ) {
621 if (MP_LOCK_HELD())
622 cpu_rel_mplock();
623 } else {
624 ASSERT_MP_LOCK_HELD();
625 }
626#endif
37af14fe 627 if (td != ntd)
f1d1c3fa 628 td->td_switch(ntd);
37af14fe
MD
629 /* NOTE: current cpu may have changed after switch */
630 crit_exit_quick(td);
8ad65e08
MD
631}
632
b68b7282 633/*
96728c05
MD
634 * Request that the target thread preempt the current thread. Preemption
635 * only works under a specific set of conditions:
b68b7282 636 *
96728c05
MD
637 * - We are not preempting ourselves
638 * - The target thread is owned by the current cpu
639 * - We are not currently being preempted
640 * - The target is not currently being preempted
641 * - We are able to satisfy the target's MP lock requirements (if any).
642 *
643 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
644 * this is called via lwkt_schedule() through the td_preemptable callback.
645 * critpri is the managed critical priority that we should ignore in order
646 * to determine whether preemption is possible (aka usually just the crit
647 * priority of lwkt_schedule() itself).
b68b7282 648 *
26a0694b
MD
649 * XXX at the moment we run the target thread in a critical section during
650 * the preemption in order to prevent the target from taking interrupts
651 * that *WE* can't. Preemption is strictly limited to interrupt threads
652 * and interrupt-like threads, outside of a critical section, and the
653 * preempted source thread will be resumed the instant the target blocks
654 * whether or not the source is scheduled (i.e. preemption is supposed to
655 * be as transparent as possible).
4b5f931b 656 *
8a8d5d85
MD
657 * The target thread inherits our MP count (added to its own) for the
658 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
659 * MP lock during the preemption. Therefore, any preempting targets must be
660 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
661 * out of sync with the physical mp_lock, but we do not have to preserve
662 * the original ownership of the lock if it was out of synch (that is, we
663 * can leave it synchronized on return).
b68b7282
MD
664 */
665void
96728c05 666lwkt_preempt(thread_t ntd, int critpri)
b68b7282 667{
46a3f46d 668 struct globaldata *gd = mycpu;
0a3f9b47 669 thread_t td;
8a8d5d85
MD
670#ifdef SMP
671 int mpheld;
57c254db 672 int savecnt;
8a8d5d85 673#endif
b68b7282 674
26a0694b 675 /*
96728c05
MD
676 * The caller has put us in a critical section. We can only preempt
677 * if the caller of the caller was not in a critical section (basically
0a3f9b47 678 * a local interrupt), as determined by the 'critpri' parameter.
96728c05
MD
679 *
680 * YYY The target thread must be in a critical section (else it must
681 * inherit our critical section? I dunno yet).
41a01a4d
MD
682 *
683 * Any tokens held by the target may not be held by thread(s) being
684 * preempted. We take the easy way out and do not preempt if
685 * the target is holding tokens.
0a3f9b47
MD
686 *
687 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
688 */
689 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 690
0a3f9b47 691 td = gd->gd_curthread;
0a3f9b47 692 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
693 ++preempt_miss;
694 return;
695 }
96728c05
MD
696 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
697 ++preempt_miss;
8ec60c3f 698 need_lwkt_resched();
96728c05
MD
699 return;
700 }
701#ifdef SMP
46a3f46d 702 if (ntd->td_gd != gd) {
96728c05 703 ++preempt_miss;
8ec60c3f 704 need_lwkt_resched();
96728c05
MD
705 return;
706 }
707#endif
41a01a4d
MD
708 /*
709 * Take the easy way out and do not preempt if the target is holding
710 * one or more tokens. We could test whether the thread(s) being
711 * preempted interlock against the target thread's tokens and whether
712 * we can get all the target thread's tokens, but this situation
713 * should not occur very often so its easier to simply not preempt.
714 */
715 if (ntd->td_toks != NULL) {
716 ++preempt_miss;
8ec60c3f 717 need_lwkt_resched();
41a01a4d
MD
718 return;
719 }
26a0694b
MD
720 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
721 ++preempt_weird;
8ec60c3f 722 need_lwkt_resched();
26a0694b
MD
723 return;
724 }
725 if (ntd->td_preempted) {
4b5f931b 726 ++preempt_hit;
8ec60c3f 727 need_lwkt_resched();
26a0694b 728 return;
b68b7282 729 }
8a8d5d85 730#ifdef SMP
a2a5ad0d
MD
731 /*
732 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
733 * to or from the MP lock. In this case td_mpcount will be pre-disposed
734 * (non-zero) but not actually synchronized with the actual state of the
735 * lock. We can use it to imply an MP lock requirement for the
736 * preemption but we cannot use it to test whether we hold the MP lock
737 * or not.
a2a5ad0d 738 */
96728c05 739 savecnt = td->td_mpcount;
71ef2f5c 740 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
741 ntd->td_mpcount += td->td_mpcount;
742 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
743 ntd->td_mpcount -= td->td_mpcount;
744 ++preempt_miss;
8ec60c3f 745 need_lwkt_resched();
8a8d5d85
MD
746 return;
747 }
748#endif
26a0694b 749
8ec60c3f
MD
750 /*
751 * Since we are able to preempt the current thread, there is no need to
752 * call need_lwkt_resched().
753 */
26a0694b
MD
754 ++preempt_hit;
755 ntd->td_preempted = td;
756 td->td_flags |= TDF_PREEMPT_LOCK;
757 td->td_switch(ntd);
758 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
759#ifdef SMP
760 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
761 mpheld = MP_LOCK_HELD();
762 if (mpheld && td->td_mpcount == 0)
96728c05 763 cpu_rel_mplock();
71ef2f5c 764 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
765 panic("lwkt_preempt(): MP lock was not held through");
766#endif
26a0694b
MD
767 ntd->td_preempted = NULL;
768 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
769}
770
f1d1c3fa
MD
771/*
772 * Yield our thread while higher priority threads are pending. This is
773 * typically called when we leave a critical section but it can be safely
774 * called while we are in a critical section.
775 *
776 * This function will not generally yield to equal priority threads but it
777 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 778 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
779 * lwkt_yield_quick().
780 *
235957ed 781 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
782 * came along but was blocked and made pending.
783 *
f1d1c3fa
MD
784 * (self contained on a per cpu basis)
785 */
786void
787lwkt_yield_quick(void)
788{
7966cb69
MD
789 globaldata_t gd = mycpu;
790 thread_t td = gd->gd_curthread;
ef0fdad1 791
a2a5ad0d 792 /*
235957ed 793 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
794 * it with a non-zero cpl then we might not wind up calling splz after
795 * a task switch when the critical section is exited even though the
46a3f46d 796 * new task could accept the interrupt.
a2a5ad0d
MD
797 *
798 * XXX from crit_exit() only called after last crit section is released.
799 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
800 *
801 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
802 * except for this special case, we MUST call splz() here to handle any
803 * pending ints, particularly after we switch, or we might accidently
804 * halt the cpu with interrupts pending.
a2a5ad0d 805 */
46a3f46d 806 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 807 splz();
f1d1c3fa
MD
808
809 /*
810 * YYY enabling will cause wakeup() to task-switch, which really
811 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
812 * preemption and MP without actually doing preemption or MP, because a
813 * lot of code assumes that wakeup() does not block.
f1d1c3fa 814 */
46a3f46d
MD
815 if (untimely_switch && td->td_nest_count == 0 &&
816 gd->gd_intr_nesting_level == 0
817 ) {
37af14fe 818 crit_enter_quick(td);
f1d1c3fa
MD
819 /*
820 * YYY temporary hacks until we disassociate the userland scheduler
821 * from the LWKT scheduler.
822 */
823 if (td->td_flags & TDF_RUNQ) {
824 lwkt_switch(); /* will not reenter yield function */
825 } else {
37af14fe 826 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 827 lwkt_switch(); /* will not reenter yield function */
37af14fe 828 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 829 }
7966cb69 830 crit_exit_noyield(td);
f1d1c3fa 831 }
f1d1c3fa
MD
832}
833
8ad65e08 834/*
f1d1c3fa 835 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 836 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
837 * the crit_exit() call in lwkt_switch().
838 *
839 * (self contained on a per cpu basis)
8ad65e08
MD
840 */
841void
f1d1c3fa 842lwkt_yield(void)
8ad65e08 843{
37af14fe 844 lwkt_schedule_self(curthread);
f1d1c3fa
MD
845 lwkt_switch();
846}
847
8ad65e08 848/*
f1d1c3fa
MD
849 * Generic schedule. Possibly schedule threads belonging to other cpus and
850 * deal with threads that might be blocked on a wait queue.
851 *
0a3f9b47
MD
852 * We have a little helper inline function which does additional work after
853 * the thread has been enqueued, including dealing with preemption and
854 * setting need_lwkt_resched() (which prevents the kernel from returning
855 * to userland until it has processed higher priority threads).
8ad65e08 856 */
0a3f9b47
MD
857static __inline
858void
8ec60c3f 859_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
0a3f9b47
MD
860{
861 if (ntd->td_preemptable) {
862 ntd->td_preemptable(ntd, cpri); /* YYY +token */
8ec60c3f
MD
863 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
864 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
865 ) {
866 need_lwkt_resched();
0a3f9b47
MD
867 }
868}
869
8ad65e08
MD
870void
871lwkt_schedule(thread_t td)
872{
37af14fe
MD
873 globaldata_t mygd = mycpu;
874
96728c05 875#ifdef INVARIANTS
41a01a4d 876 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
26a0694b
MD
877 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
878 && td->td_proc->p_stat == SSLEEP
879 ) {
880 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
881 curthread,
882 curthread->td_proc ? curthread->td_proc->p_pid : -1,
883 curthread->td_proc ? curthread->td_proc->p_stat : -1,
884 td,
885 td->td_proc ? curthread->td_proc->p_pid : -1,
886 td->td_proc ? curthread->td_proc->p_stat : -1
887 );
888 panic("SCHED PANIC");
889 }
96728c05 890#endif
37af14fe
MD
891 crit_enter_gd(mygd);
892 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
893 _lwkt_enqueue(td);
894 } else {
895 lwkt_wait_t w;
896
897 /*
898 * If the thread is on a wait list we have to send our scheduling
899 * request to the owner of the wait structure. Otherwise we send
900 * the scheduling request to the cpu owning the thread. Races
901 * are ok, the target will forward the message as necessary (the
902 * message may chase the thread around before it finally gets
903 * acted upon).
904 *
905 * (remember, wait structures use stable storage)
0a3f9b47
MD
906 *
907 * NOTE: tokens no longer enter a critical section, so we only need
908 * to account for the crit_enter() above when calling
909 * _lwkt_schedule_post().
f1d1c3fa
MD
910 */
911 if ((w = td->td_wait) != NULL) {
41a01a4d
MD
912 lwkt_tokref wref;
913
914 if (lwkt_trytoken(&wref, &w->wa_token)) {
f1d1c3fa
MD
915 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
916 --w->wa_count;
917 td->td_wait = NULL;
0f7a3396 918#ifdef SMP
8ec60c3f 919 if (td->td_gd == mygd) {
f1d1c3fa 920 _lwkt_enqueue(td);
8ec60c3f 921 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 922 } else {
2db3b277 923 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 924 }
0f7a3396
MD
925#else
926 _lwkt_enqueue(td);
8ec60c3f 927 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 928#endif
41a01a4d 929 lwkt_reltoken(&wref);
f1d1c3fa 930 } else {
96728c05 931 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
932 }
933 } else {
934 /*
935 * If the wait structure is NULL and we own the thread, there
936 * is no race (since we are in a critical section). If we
937 * do not own the thread there might be a race but the
938 * target cpu will deal with it.
939 */
0f7a3396 940#ifdef SMP
37af14fe 941 if (td->td_gd == mygd) {
f1d1c3fa 942 _lwkt_enqueue(td);
8ec60c3f 943 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 944 } else {
2db3b277 945 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 946 }
0f7a3396
MD
947#else
948 _lwkt_enqueue(td);
8ec60c3f 949 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 950#endif
f1d1c3fa 951 }
8ad65e08 952 }
37af14fe 953 crit_exit_gd(mygd);
8ad65e08
MD
954}
955
d9eea1a5
MD
956/*
957 * Managed acquisition. This code assumes that the MP lock is held for
958 * the tdallq operation and that the thread has been descheduled from its
959 * original cpu. We also have to wait for the thread to be entirely switched
960 * out on its original cpu (this is usually fast enough that we never loop)
961 * since the LWKT system does not have to hold the MP lock while switching
962 * and the target may have released it before switching.
963 */
a2a5ad0d
MD
964void
965lwkt_acquire(thread_t td)
966{
37af14fe
MD
967 globaldata_t gd;
968 globaldata_t mygd;
a2a5ad0d
MD
969
970 gd = td->td_gd;
37af14fe 971 mygd = mycpu;
a2a5ad0d 972 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5 973 while (td->td_flags & TDF_RUNNING) /* XXX spin */
5d21b981 974 cpu_mb1();
37af14fe
MD
975 if (gd != mygd) {
976 crit_enter_gd(mygd);
a2a5ad0d 977 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
37af14fe
MD
978 td->td_gd = mygd;
979 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
980 crit_exit_gd(mygd);
a2a5ad0d
MD
981 }
982}
983
f1d1c3fa
MD
984/*
985 * Generic deschedule. Descheduling threads other then your own should be
986 * done only in carefully controlled circumstances. Descheduling is
987 * asynchronous.
988 *
989 * This function may block if the cpu has run out of messages.
8ad65e08
MD
990 */
991void
992lwkt_deschedule(thread_t td)
993{
f1d1c3fa
MD
994 crit_enter();
995 if (td == curthread) {
996 _lwkt_dequeue(td);
997 } else {
a72187e9 998 if (td->td_gd == mycpu) {
f1d1c3fa
MD
999 _lwkt_dequeue(td);
1000 } else {
2db3b277 1001 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
1002 }
1003 }
1004 crit_exit();
1005}
1006
4b5f931b
MD
1007/*
1008 * Set the target thread's priority. This routine does not automatically
1009 * switch to a higher priority thread, LWKT threads are not designed for
1010 * continuous priority changes. Yield if you want to switch.
1011 *
1012 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1013 * of the td_pri field. The specified priority may also indicate zero or
1014 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1015 *
1016 * Note that we requeue the thread whether it winds up on a different runq
1017 * or not. uio_yield() depends on this and the routine is not normally
1018 * called with the same priority otherwise.
4b5f931b
MD
1019 */
1020void
1021lwkt_setpri(thread_t td, int pri)
1022{
26a0694b 1023 KKASSERT(pri >= 0);
a72187e9 1024 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1025 crit_enter();
1026 if (td->td_flags & TDF_RUNQ) {
1027 _lwkt_dequeue(td);
1028 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1029 _lwkt_enqueue(td);
1030 } else {
1031 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1032 }
1033 crit_exit();
1034}
1035
1036void
1037lwkt_setpri_self(int pri)
1038{
1039 thread_t td = curthread;
1040
4b5f931b
MD
1041 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1042 crit_enter();
1043 if (td->td_flags & TDF_RUNQ) {
1044 _lwkt_dequeue(td);
1045 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1046 _lwkt_enqueue(td);
1047 } else {
1048 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1049 }
1050 crit_exit();
1051}
1052
8ec60c3f
MD
1053/*
1054 * Determine if there is a runnable thread at a higher priority then
1055 * the current thread. lwkt_setpri() does not check this automatically.
1056 * Return 1 if there is, 0 if there isn't.
1057 *
1058 * Example: if bit 31 of runqmask is set and the current thread is priority
1059 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1060 *
1061 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1062 * up comparing against 0xffffffff, a comparison that will always be false.
1063 */
1064int
1065lwkt_checkpri_self(void)
1066{
1067 globaldata_t gd = mycpu;
1068 thread_t td = gd->gd_curthread;
1069 int nq = td->td_pri & TDPRI_MASK;
1070
1071 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1072 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1073 return(1);
1074 ++nq;
1075 }
1076 return(0);
1077}
1078
5d21b981
MD
1079/*
1080 * Migrate the current thread to the specified cpu. The BGL must be held
1081 * (for the gd_tdallq manipulation XXX). This is accomplished by
1082 * descheduling ourselves from the current cpu, moving our thread to the
1083 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1084 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1085 */
3d28ff59 1086#ifdef SMP
5d21b981 1087static void lwkt_setcpu_remote(void *arg);
3d28ff59 1088#endif
5d21b981
MD
1089
1090void
1091lwkt_setcpu_self(globaldata_t rgd)
1092{
1093#ifdef SMP
1094 thread_t td = curthread;
1095
1096 if (td->td_gd != rgd) {
1097 crit_enter_quick(td);
1098 td->td_flags |= TDF_MIGRATING;
1099 lwkt_deschedule_self(td);
1100 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1101 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
1102 lwkt_send_ipiq(rgd, (ipifunc_t)lwkt_setcpu_remote, td);
1103 lwkt_switch();
1104 /* we are now on the target cpu */
1105 crit_exit_quick(td);
1106 }
1107#endif
1108}
1109
1110/*
1111 * Remote IPI for cpu migration (called while in a critical section so we
1112 * do not have to enter another one). The thread has already been moved to
1113 * our cpu's allq, but we must wait for the thread to be completely switched
1114 * out on the originating cpu before we schedule it on ours or the stack
1115 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1116 * change to main memory.
1117 *
1118 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1119 * against wakeups. It is best if this interface is used only when there
1120 * are no pending events that might try to schedule the thread.
1121 */
3d28ff59 1122#ifdef SMP
5d21b981
MD
1123static void
1124lwkt_setcpu_remote(void *arg)
1125{
1126 thread_t td = arg;
1127 globaldata_t gd = mycpu;
1128
1129 while (td->td_flags & TDF_RUNNING)
1130 cpu_mb1();
1131 td->td_gd = gd;
1132 cpu_mb2();
1133 td->td_flags &= ~TDF_MIGRATING;
1134 _lwkt_enqueue(td);
1135}
3d28ff59 1136#endif
5d21b981 1137
4b5f931b
MD
1138struct proc *
1139lwkt_preempted_proc(void)
1140{
73e4f7b9 1141 thread_t td = curthread;
4b5f931b
MD
1142 while (td->td_preempted)
1143 td = td->td_preempted;
1144 return(td->td_proc);
1145}
1146
f1d1c3fa 1147/*
41a01a4d
MD
1148 * Block on the specified wait queue until signaled. A generation number
1149 * must be supplied to interlock the wait queue. The function will
1150 * return immediately if the generation number does not match the wait
1151 * structure's generation number.
f1d1c3fa
MD
1152 */
1153void
ae8050a4 1154lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
1155{
1156 thread_t td = curthread;
41a01a4d 1157 lwkt_tokref ilock;
f1d1c3fa 1158
41a01a4d
MD
1159 lwkt_gettoken(&ilock, &w->wa_token);
1160 crit_enter();
ae8050a4 1161 if (w->wa_gen == *gen) {
f1d1c3fa
MD
1162 _lwkt_dequeue(td);
1163 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1164 ++w->wa_count;
1165 td->td_wait = w;
ae8050a4 1166 td->td_wmesg = wmesg;
41a01a4d 1167 again:
f1d1c3fa 1168 lwkt_switch();
ece04fd0
MD
1169 if (td->td_wmesg != NULL) {
1170 _lwkt_dequeue(td);
1171 goto again;
1172 }
8ad65e08 1173 }
41a01a4d 1174 crit_exit();
ae8050a4 1175 *gen = w->wa_gen;
41a01a4d 1176 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1177}
1178
1179/*
1180 * Signal a wait queue. We gain ownership of the wait queue in order to
1181 * signal it. Once a thread is removed from the wait queue we have to
1182 * deal with the cpu owning the thread.
1183 *
1184 * Note: alternatively we could message the target cpu owning the wait
1185 * queue. YYY implement as sysctl.
1186 */
1187void
ece04fd0 1188lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1189{
1190 thread_t td;
41a01a4d 1191 lwkt_tokref ilock;
f1d1c3fa 1192
41a01a4d 1193 lwkt_gettoken(&ilock, &w->wa_token);
f1d1c3fa 1194 ++w->wa_gen;
41a01a4d 1195 crit_enter();
ece04fd0
MD
1196 if (count < 0)
1197 count = w->wa_count;
f1d1c3fa
MD
1198 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1199 --count;
1200 --w->wa_count;
1201 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1202 td->td_wait = NULL;
ae8050a4 1203 td->td_wmesg = NULL;
a72187e9 1204 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1205 _lwkt_enqueue(td);
1206 } else {
2db3b277 1207 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 1208 }
f1d1c3fa 1209 }
41a01a4d
MD
1210 crit_exit();
1211 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1212}
1213
99df837e
MD
1214/*
1215 * Create a kernel process/thread/whatever. It shares it's address space
1216 * with proc0 - ie: kernel only.
1217 *
365fa13f
MD
1218 * NOTE! By default new threads are created with the MP lock held. A
1219 * thread which does not require the MP lock should release it by calling
1220 * rel_mplock() at the start of the new thread.
99df837e
MD
1221 */
1222int
1223lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1224 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1225 const char *fmt, ...)
99df837e 1226{
73e4f7b9 1227 thread_t td;
e2565a42 1228 __va_list ap;
99df837e 1229
f470d0c8 1230 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu);
a2a5ad0d
MD
1231 if (tdp)
1232 *tdp = td;
709799ea 1233 cpu_set_thread_handler(td, lwkt_exit, func, arg);
ef0fdad1 1234 td->td_flags |= TDF_VERBOSE | tdflags;
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 */
ef0fdad1
MD
1249 if ((td->td_flags & TDF_STOPREQ) == 0)
1250 lwkt_schedule(td);
1251 else
1252 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1253 return 0;
1254}
1255
2d93b37a 1256/*
2d93b37a
MD
1257 * kthread_* is specific to the kernel and is not needed by userland.
1258 */
1259#ifdef _KERNEL
1260
99df837e
MD
1261/*
1262 * Destroy an LWKT thread. Warning! This function is not called when
1263 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1264 * uses a different reaping mechanism.
1265 */
1266void
1267lwkt_exit(void)
1268{
1269 thread_t td = curthread;
8826f33a 1270 globaldata_t gd;
99df837e
MD
1271
1272 if (td->td_flags & TDF_VERBOSE)
1273 printf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1274 caps_exit(td);
37af14fe
MD
1275 crit_enter_quick(td);
1276 lwkt_deschedule_self(td);
8826f33a
MD
1277 gd = mycpu;
1278 KKASSERT(gd == td->td_gd);
1279 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1280 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1281 ++gd->gd_tdfreecount;
1282 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1283 }
99df837e
MD
1284 cpu_thread_exit();
1285}
1286
2d93b37a
MD
1287#endif /* _KERNEL */
1288
1289void
1290crit_panic(void)
1291{
1292 thread_t td = curthread;
1293 int lpri = td->td_pri;
1294
1295 td->td_pri = 0;
1296 panic("td_pri is/would-go negative! %p %d", td, lpri);
1297}
1298