sys/dev __FreeBSD__ -> __DragonFly__ cleanups.
[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 *
c31b1324 26 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.51 2004/02/09 21:13:18 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.
f1d1c3fa 34 *
75cdbe6c
MD
35 * NOTE: on UP machines smp_active is defined to be 0. On SMP machines
36 * smp_active is 0 prior to SMP activation, then it is 1. The LWKT module
37 * uses smp_active to optimize UP builds and to avoid sending IPIs during
38 * early boot (primarily interrupt and network thread initialization).
8ad65e08
MD
39 */
40
05220613
MD
41#ifdef _KERNEL
42
8ad65e08
MD
43#include <sys/param.h>
44#include <sys/systm.h>
45#include <sys/kernel.h>
46#include <sys/proc.h>
47#include <sys/rtprio.h>
48#include <sys/queue.h>
f1d1c3fa 49#include <sys/thread2.h>
7d0bac62 50#include <sys/sysctl.h>
99df837e 51#include <sys/kthread.h>
f1d1c3fa 52#include <machine/cpu.h>
99df837e 53#include <sys/lock.h>
f6bf3af1 54#include <sys/caps.h>
f1d1c3fa 55
7d0bac62
MD
56#include <vm/vm.h>
57#include <vm/vm_param.h>
58#include <vm/vm_kern.h>
59#include <vm/vm_object.h>
60#include <vm/vm_page.h>
61#include <vm/vm_map.h>
62#include <vm/vm_pager.h>
63#include <vm/vm_extern.h>
64#include <vm/vm_zone.h>
65
99df837e 66#include <machine/stdarg.h>
57c254db 67#include <machine/ipl.h>
96728c05 68#include <machine/smp.h>
99df837e 69
05220613
MD
70#define THREAD_STACK (UPAGES * PAGE_SIZE)
71
72#else
73
74#include <sys/stdint.h>
fb04f4fd 75#include <libcaps/thread.h>
05220613
MD
76#include <sys/thread.h>
77#include <sys/msgport.h>
78#include <sys/errno.h>
fb04f4fd 79#include <libcaps/globaldata.h>
05220613
MD
80#include <sys/thread2.h>
81#include <sys/msgport2.h>
709799ea 82#include <stdio.h>
05220613 83#include <stdlib.h>
709799ea 84#include <string.h>
c95cd171 85#include <machine/cpufunc.h>
709799ea 86#include <machine/lock.h>
05220613
MD
87
88#endif
89
7d0bac62 90static int untimely_switch = 0;
05220613
MD
91static __int64_t switch_count = 0;
92static __int64_t preempt_hit = 0;
93static __int64_t preempt_miss = 0;
94static __int64_t preempt_weird = 0;
709799ea 95#ifdef SMP
05220613
MD
96static __int64_t ipiq_count = 0;
97static __int64_t ipiq_fifofull = 0;
709799ea 98#endif
05220613
MD
99
100#ifdef _KERNEL
101
102SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
4b5f931b 103SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 104SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 105SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 106SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
709799ea 107#ifdef SMP
96728c05 108SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
96728c05 109SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
709799ea 110#endif
7d0bac62 111
05220613
MD
112#endif
113
4b5f931b
MD
114/*
115 * These helper procedures handle the runq, they can only be called from
116 * within a critical section.
75cdbe6c
MD
117 *
118 * WARNING! Prior to SMP being brought up it is possible to enqueue and
119 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
120 * instead of 'mycpu' when referencing the globaldata structure. Once
121 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 122 */
f1d1c3fa
MD
123static __inline
124void
125_lwkt_dequeue(thread_t td)
126{
127 if (td->td_flags & TDF_RUNQ) {
4b5f931b 128 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 129 struct globaldata *gd = td->td_gd;
4b5f931b 130
f1d1c3fa 131 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
132 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
133 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
134 }
135}
136
137static __inline
138void
139_lwkt_enqueue(thread_t td)
140{
141 if ((td->td_flags & TDF_RUNQ) == 0) {
4b5f931b 142 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 143 struct globaldata *gd = td->td_gd;
4b5f931b 144
f1d1c3fa 145 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
146 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
147 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
148 }
149}
8ad65e08 150
57c254db
MD
151static __inline
152int
153_lwkt_wantresched(thread_t ntd, thread_t cur)
154{
155 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
156}
157
2d93b37a
MD
158#ifdef _KERNEL
159
8ad65e08
MD
160/*
161 * LWKTs operate on a per-cpu basis
162 *
73e4f7b9 163 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
164 */
165void
166lwkt_gdinit(struct globaldata *gd)
167{
4b5f931b
MD
168 int i;
169
170 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
171 TAILQ_INIT(&gd->gd_tdrunq[i]);
172 gd->gd_runqmask = 0;
73e4f7b9 173 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
174}
175
2d93b37a
MD
176#endif /* _KERNEL */
177
7d0bac62
MD
178/*
179 * Initialize a thread wait structure prior to first use.
180 *
181 * NOTE! called from low level boot code, we cannot do anything fancy!
182 */
183void
184lwkt_init_wait(lwkt_wait_t w)
185{
186 TAILQ_INIT(&w->wa_waitq);
187}
188
189/*
190 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
191 * or LWKT start address before it can be scheduled. If the target cpu is
192 * -1 the thread will be created on the current cpu.
0cfcada1
MD
193 *
194 * If you intend to create a thread without a process context this function
195 * does everything except load the startup and switcher function.
7d0bac62
MD
196 */
197thread_t
75cdbe6c 198lwkt_alloc_thread(struct thread *td, int cpu)
7d0bac62 199{
99df837e 200 void *stack;
ef0fdad1 201 int flags = 0;
7d0bac62 202
ef0fdad1 203 if (td == NULL) {
26a0694b 204 crit_enter();
ef0fdad1
MD
205 if (mycpu->gd_tdfreecount > 0) {
206 --mycpu->gd_tdfreecount;
207 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
d9eea1a5 208 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1
MD
209 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
210 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
211 crit_exit();
212 stack = td->td_kstack;
213 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
214 } else {
215 crit_exit();
05220613 216#ifdef _KERNEL
ef0fdad1 217 td = zalloc(thread_zone);
05220613
MD
218#else
219 td = malloc(sizeof(struct thread));
220#endif
ef0fdad1
MD
221 td->td_kstack = NULL;
222 flags |= TDF_ALLOCATED_THREAD;
223 }
224 }
225 if ((stack = td->td_kstack) == NULL) {
05220613
MD
226#ifdef _KERNEL
227 stack = (void *)kmem_alloc(kernel_map, THREAD_STACK);
228#else
fb04f4fd 229 stack = libcaps_alloc_stack(THREAD_STACK);
05220613 230#endif
ef0fdad1 231 flags |= TDF_ALLOCATED_STACK;
99df837e 232 }
75cdbe6c
MD
233 if (cpu < 0)
234 lwkt_init_thread(td, stack, flags, mycpu);
235 else
236 lwkt_init_thread(td, stack, flags, globaldata_find(cpu));
99df837e 237 return(td);
7d0bac62
MD
238}
239
709799ea
MD
240#ifdef _KERNEL
241
7d0bac62
MD
242/*
243 * Initialize a preexisting thread structure. This function is used by
244 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
245 *
f8c3996b
MD
246 * All threads start out in a critical section at a priority of
247 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
248 * appropriate. This function may send an IPI message when the
249 * requested cpu is not the current cpu and consequently gd_tdallq may
250 * not be initialized synchronously from the point of view of the originating
251 * cpu.
252 *
253 * NOTE! we have to be careful in regards to creating threads for other cpus
254 * if SMP has not yet been activated.
7d0bac62 255 */
75cdbe6c
MD
256static void
257lwkt_init_thread_remote(void *arg)
258{
259 thread_t td = arg;
260
261 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
262}
263
7d0bac62 264void
26a0694b 265lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
7d0bac62 266{
99df837e
MD
267 bzero(td, sizeof(struct thread));
268 td->td_kstack = stack;
269 td->td_flags |= flags;
26a0694b 270 td->td_gd = gd;
f8c3996b 271 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
c95cd171 272 lwkt_initport(&td->td_msgport, td);
99df837e 273 pmap_init_thread(td);
75cdbe6c
MD
274 if (smp_active == 0 || gd == mycpu) {
275 crit_enter();
276 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
277 crit_exit();
278 } else {
279 lwkt_send_ipiq(gd->gd_cpuid, lwkt_init_thread_remote, td);
280 }
73e4f7b9
MD
281}
282
2d93b37a
MD
283#endif /* _KERNEL */
284
73e4f7b9
MD
285void
286lwkt_set_comm(thread_t td, const char *ctl, ...)
287{
e2565a42 288 __va_list va;
73e4f7b9 289
e2565a42 290 __va_start(va, ctl);
73e4f7b9 291 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 292 __va_end(va);
7d0bac62
MD
293}
294
99df837e 295void
73e4f7b9 296lwkt_hold(thread_t td)
99df837e 297{
73e4f7b9
MD
298 ++td->td_refs;
299}
300
301void
302lwkt_rele(thread_t td)
303{
304 KKASSERT(td->td_refs > 0);
305 --td->td_refs;
306}
307
c95cd171
MD
308#ifdef _KERNEL
309
73e4f7b9
MD
310void
311lwkt_wait_free(thread_t td)
312{
313 while (td->td_refs)
377d4740 314 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
315}
316
c95cd171
MD
317#endif
318
73e4f7b9
MD
319void
320lwkt_free_thread(thread_t td)
321{
322 struct globaldata *gd = mycpu;
323
d9eea1a5 324 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
325 ("lwkt_free_thread: did not exit! %p", td));
326
327 crit_enter();
73e4f7b9
MD
328 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
329 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
330 (td->td_flags & TDF_ALLOCATED_THREAD)
331 ) {
73e4f7b9
MD
332 ++gd->gd_tdfreecount;
333 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
99df837e
MD
334 crit_exit();
335 } else {
336 crit_exit();
337 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613
MD
338#ifdef _KERNEL
339 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK);
340#else
fb04f4fd 341 libcaps_free_stack(td->td_kstack, THREAD_STACK);
05220613 342#endif
73e4f7b9 343 /* gd invalid */
99df837e
MD
344 td->td_kstack = NULL;
345 }
05220613
MD
346 if (td->td_flags & TDF_ALLOCATED_THREAD) {
347#ifdef _KERNEL
99df837e 348 zfree(thread_zone, td);
05220613
MD
349#else
350 free(td);
351#endif
352 }
99df837e
MD
353 }
354}
355
356
8ad65e08
MD
357/*
358 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
359 * switch to the idlethread. Switching must occur within a critical
360 * section to avoid races with the scheduling queue.
361 *
362 * We always have full control over our cpu's run queue. Other cpus
363 * that wish to manipulate our queue must use the cpu_*msg() calls to
364 * talk to our cpu, so a critical section is all that is needed and
365 * the result is very, very fast thread switching.
366 *
96728c05
MD
367 * The LWKT scheduler uses a fixed priority model and round-robins at
368 * each priority level. User process scheduling is a totally
369 * different beast and LWKT priorities should not be confused with
370 * user process priorities.
f1d1c3fa 371 *
96728c05
MD
372 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
373 * cleans it up. Note that the td_switch() function cannot do anything that
374 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
375 * the target thread (not the current thread). It's nice to have a scheduler
376 * that does not need the MP lock to work because it allows us to do some
377 * really cool high-performance MP lock optimizations.
8ad65e08 378 */
96728c05 379
8ad65e08
MD
380void
381lwkt_switch(void)
382{
4b5f931b 383 struct globaldata *gd;
f1d1c3fa 384 thread_t td = curthread;
8ad65e08 385 thread_t ntd;
8a8d5d85
MD
386#ifdef SMP
387 int mpheld;
388#endif
8ad65e08 389
46a3f46d
MD
390 /*
391 * Switching from within a 'fast' (non thread switched) interrupt is
392 * illegal.
393 */
394 if (mycpu->gd_intr_nesting_level && panicstr == NULL) {
03aa8d99 395 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
96728c05 396 }
ef0fdad1 397
cb973d15
MD
398 /*
399 * Passive release (used to transition from user to kernel mode
400 * when we block or switch rather then when we enter the kernel).
401 * This function is NOT called if we are switching into a preemption
402 * or returning from a preemption. Typically this causes us to lose
403 * our P_CURPROC designation (if we have one) and become a true LWKT
404 * thread, and may also hand P_CURPROC to another process and schedule
405 * its thread.
406 */
407 if (td->td_release)
408 td->td_release(td);
409
f1d1c3fa 410 crit_enter();
4b5f931b 411 ++switch_count;
8a8d5d85
MD
412
413#ifdef SMP
414 /*
415 * td_mpcount cannot be used to determine if we currently hold the
416 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
417 * to get the lock, and switch out if it can't. Our ownership of
418 * the actual lock will remain stable while we are in a critical section
419 * (but, of course, another cpu may own or release the lock so the
420 * actual value of mp_lock is not stable).
8a8d5d85
MD
421 */
422 mpheld = MP_LOCK_HELD();
423#endif
99df837e
MD
424 if ((ntd = td->td_preempted) != NULL) {
425 /*
426 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
427 * thread. This occurs transparently, whether the preempted thread
428 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
429 * itself).
430 *
431 * We have to setup the MP lock for the original thread after backing
432 * out the adjustment that was made to curthread when the original
433 * was preempted.
99df837e 434 */
26a0694b 435 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 436#ifdef SMP
96728c05
MD
437 if (ntd->td_mpcount && mpheld == 0) {
438 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
439 td, ntd, td->td_mpcount, ntd->td_mpcount);
440 }
8a8d5d85
MD
441 if (ntd->td_mpcount) {
442 td->td_mpcount -= ntd->td_mpcount;
443 KKASSERT(td->td_mpcount >= 0);
444 }
445#endif
26a0694b 446 ntd->td_flags |= TDF_PREEMPT_DONE;
8a8d5d85 447 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 448 } else {
4b5f931b
MD
449 /*
450 * Priority queue / round-robin at each priority. Note that user
451 * processes run at a fixed, low priority and the user process
452 * scheduler deals with interactions between user processes
453 * by scheduling and descheduling them from the LWKT queue as
454 * necessary.
8a8d5d85
MD
455 *
456 * We have to adjust the MP lock for the target thread. If we
457 * need the MP lock and cannot obtain it we try to locate a
458 * thread that does not need the MP lock.
4b5f931b
MD
459 */
460 gd = mycpu;
4b5f931b
MD
461again:
462 if (gd->gd_runqmask) {
463 int nq = bsrl(gd->gd_runqmask);
464 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
465 gd->gd_runqmask &= ~(1 << nq);
466 goto again;
467 }
8a8d5d85
MD
468#ifdef SMP
469 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
470 /*
96728c05
MD
471 * Target needs MP lock and we couldn't get it, try
472 * to locate a thread which does not need the MP lock
3c23a41a 473 * to run. If we cannot locate a thread spin in idle.
8a8d5d85
MD
474 */
475 u_int32_t rqmask = gd->gd_runqmask;
476 while (rqmask) {
477 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
478 if (ntd->td_mpcount == 0)
479 break;
480 }
481 if (ntd)
482 break;
483 rqmask &= ~(1 << nq);
484 nq = bsrl(rqmask);
485 }
486 if (ntd == NULL) {
a2a5ad0d
MD
487 ntd = &gd->gd_idlethread;
488 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
489 } else {
490 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
491 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
492 }
493 } else {
494 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
495 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
496 }
497#else
4b5f931b
MD
498 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
499 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 500#endif
4b5f931b 501 } else {
3c23a41a 502 /*
60f945af
MD
503 * We have nothing to run but only let the idle loop halt
504 * the cpu if there are no pending interrupts.
3c23a41a 505 */
a2a5ad0d 506 ntd = &gd->gd_idlethread;
60f945af 507 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 508 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 509 }
f1d1c3fa 510 }
26a0694b
MD
511 KASSERT(ntd->td_pri >= TDPRI_CRIT,
512 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
513
514 /*
515 * Do the actual switch. If the new target does not need the MP lock
516 * and we are holding it, release the MP lock. If the new target requires
517 * the MP lock we have already acquired it for the target.
518 */
519#ifdef SMP
520 if (ntd->td_mpcount == 0 ) {
521 if (MP_LOCK_HELD())
522 cpu_rel_mplock();
523 } else {
524 ASSERT_MP_LOCK_HELD();
525 }
526#endif
8a8d5d85 527 if (td != ntd) {
f1d1c3fa 528 td->td_switch(ntd);
8a8d5d85 529 }
96728c05 530
f1d1c3fa 531 crit_exit();
8ad65e08
MD
532}
533
cb973d15
MD
534/*
535 * Switch if another thread has a higher priority. Do not switch to other
536 * threads at the same priority.
537 */
538void
539lwkt_maybe_switch()
540{
541 struct globaldata *gd = mycpu;
542 struct thread *td = gd->gd_curthread;
543
544 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
545 lwkt_switch();
546 }
547}
548
b68b7282 549/*
96728c05
MD
550 * Request that the target thread preempt the current thread. Preemption
551 * only works under a specific set of conditions:
b68b7282 552 *
96728c05
MD
553 * - We are not preempting ourselves
554 * - The target thread is owned by the current cpu
555 * - We are not currently being preempted
556 * - The target is not currently being preempted
557 * - We are able to satisfy the target's MP lock requirements (if any).
558 *
559 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
560 * this is called via lwkt_schedule() through the td_preemptable callback.
561 * critpri is the managed critical priority that we should ignore in order
562 * to determine whether preemption is possible (aka usually just the crit
563 * priority of lwkt_schedule() itself).
b68b7282 564 *
26a0694b
MD
565 * XXX at the moment we run the target thread in a critical section during
566 * the preemption in order to prevent the target from taking interrupts
567 * that *WE* can't. Preemption is strictly limited to interrupt threads
568 * and interrupt-like threads, outside of a critical section, and the
569 * preempted source thread will be resumed the instant the target blocks
570 * whether or not the source is scheduled (i.e. preemption is supposed to
571 * be as transparent as possible).
4b5f931b 572 *
8a8d5d85
MD
573 * The target thread inherits our MP count (added to its own) for the
574 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
575 * MP lock during the preemption. Therefore, any preempting targets must be
576 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
577 * out of sync with the physical mp_lock, but we do not have to preserve
578 * the original ownership of the lock if it was out of synch (that is, we
579 * can leave it synchronized on return).
b68b7282
MD
580 */
581void
96728c05 582lwkt_preempt(thread_t ntd, int critpri)
b68b7282 583{
46a3f46d
MD
584 struct globaldata *gd = mycpu;
585 thread_t td = gd->gd_curthread;
8a8d5d85
MD
586#ifdef SMP
587 int mpheld;
57c254db 588 int savecnt;
8a8d5d85 589#endif
b68b7282 590
26a0694b 591 /*
96728c05
MD
592 * The caller has put us in a critical section. We can only preempt
593 * if the caller of the caller was not in a critical section (basically
57c254db
MD
594 * a local interrupt), as determined by the 'critpri' parameter. If
595 * we are unable to preempt
96728c05
MD
596 *
597 * YYY The target thread must be in a critical section (else it must
598 * inherit our critical section? I dunno yet).
26a0694b
MD
599 */
600 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 601
cb973d15 602 need_resched();
57c254db
MD
603 if (!_lwkt_wantresched(ntd, td)) {
604 ++preempt_miss;
605 return;
606 }
96728c05
MD
607 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
608 ++preempt_miss;
609 return;
610 }
611#ifdef SMP
46a3f46d 612 if (ntd->td_gd != gd) {
96728c05
MD
613 ++preempt_miss;
614 return;
615 }
616#endif
26a0694b
MD
617 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
618 ++preempt_weird;
619 return;
620 }
621 if (ntd->td_preempted) {
4b5f931b 622 ++preempt_hit;
26a0694b 623 return;
b68b7282 624 }
8a8d5d85 625#ifdef SMP
a2a5ad0d
MD
626 /*
627 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
628 * to or from the MP lock. In this case td_mpcount will be pre-disposed
629 * (non-zero) but not actually synchronized with the actual state of the
630 * lock. We can use it to imply an MP lock requirement for the
631 * preemption but we cannot use it to test whether we hold the MP lock
632 * or not.
a2a5ad0d 633 */
96728c05 634 savecnt = td->td_mpcount;
71ef2f5c 635 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
636 ntd->td_mpcount += td->td_mpcount;
637 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
638 ntd->td_mpcount -= td->td_mpcount;
639 ++preempt_miss;
640 return;
641 }
642#endif
26a0694b
MD
643
644 ++preempt_hit;
645 ntd->td_preempted = td;
646 td->td_flags |= TDF_PREEMPT_LOCK;
647 td->td_switch(ntd);
648 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
649#ifdef SMP
650 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
651 mpheld = MP_LOCK_HELD();
652 if (mpheld && td->td_mpcount == 0)
96728c05 653 cpu_rel_mplock();
71ef2f5c 654 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
655 panic("lwkt_preempt(): MP lock was not held through");
656#endif
26a0694b
MD
657 ntd->td_preempted = NULL;
658 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
659}
660
f1d1c3fa
MD
661/*
662 * Yield our thread while higher priority threads are pending. This is
663 * typically called when we leave a critical section but it can be safely
664 * called while we are in a critical section.
665 *
666 * This function will not generally yield to equal priority threads but it
667 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 668 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
669 * lwkt_yield_quick().
670 *
235957ed 671 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
672 * came along but was blocked and made pending.
673 *
f1d1c3fa
MD
674 * (self contained on a per cpu basis)
675 */
676void
677lwkt_yield_quick(void)
678{
7966cb69
MD
679 globaldata_t gd = mycpu;
680 thread_t td = gd->gd_curthread;
ef0fdad1 681
a2a5ad0d 682 /*
235957ed 683 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
684 * it with a non-zero cpl then we might not wind up calling splz after
685 * a task switch when the critical section is exited even though the
46a3f46d 686 * new task could accept the interrupt.
a2a5ad0d
MD
687 *
688 * XXX from crit_exit() only called after last crit section is released.
689 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
690 *
691 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
692 * except for this special case, we MUST call splz() here to handle any
693 * pending ints, particularly after we switch, or we might accidently
694 * halt the cpu with interrupts pending.
a2a5ad0d 695 */
46a3f46d 696 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 697 splz();
f1d1c3fa
MD
698
699 /*
700 * YYY enabling will cause wakeup() to task-switch, which really
701 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
702 * preemption and MP without actually doing preemption or MP, because a
703 * lot of code assumes that wakeup() does not block.
f1d1c3fa 704 */
46a3f46d
MD
705 if (untimely_switch && td->td_nest_count == 0 &&
706 gd->gd_intr_nesting_level == 0
707 ) {
f1d1c3fa
MD
708 crit_enter();
709 /*
710 * YYY temporary hacks until we disassociate the userland scheduler
711 * from the LWKT scheduler.
712 */
713 if (td->td_flags & TDF_RUNQ) {
714 lwkt_switch(); /* will not reenter yield function */
715 } else {
716 lwkt_schedule_self(); /* make sure we are scheduled */
717 lwkt_switch(); /* will not reenter yield function */
718 lwkt_deschedule_self(); /* make sure we are descheduled */
719 }
7966cb69 720 crit_exit_noyield(td);
f1d1c3fa 721 }
f1d1c3fa
MD
722}
723
8ad65e08 724/*
f1d1c3fa 725 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 726 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
727 * the crit_exit() call in lwkt_switch().
728 *
729 * (self contained on a per cpu basis)
8ad65e08
MD
730 */
731void
f1d1c3fa 732lwkt_yield(void)
8ad65e08 733{
f1d1c3fa
MD
734 lwkt_schedule_self();
735 lwkt_switch();
736}
737
738/*
739 * Schedule a thread to run. As the current thread we can always safely
740 * schedule ourselves, and a shortcut procedure is provided for that
741 * function.
742 *
743 * (non-blocking, self contained on a per cpu basis)
744 */
745void
746lwkt_schedule_self(void)
747{
748 thread_t td = curthread;
749
750 crit_enter();
751 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa 752 _lwkt_enqueue(td);
05220613 753#ifdef _KERNEL
26a0694b
MD
754 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
755 panic("SCHED SELF PANIC");
05220613 756#endif
f1d1c3fa 757 crit_exit();
8ad65e08 758}
8ad65e08
MD
759
760/*
f1d1c3fa
MD
761 * Generic schedule. Possibly schedule threads belonging to other cpus and
762 * deal with threads that might be blocked on a wait queue.
763 *
96728c05 764 * YYY this is one of the best places to implement load balancing code.
f1d1c3fa
MD
765 * Load balancing can be accomplished by requesting other sorts of actions
766 * for the thread in question.
8ad65e08
MD
767 */
768void
769lwkt_schedule(thread_t td)
770{
96728c05 771#ifdef INVARIANTS
26a0694b
MD
772 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
773 && td->td_proc->p_stat == SSLEEP
774 ) {
775 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
776 curthread,
777 curthread->td_proc ? curthread->td_proc->p_pid : -1,
778 curthread->td_proc ? curthread->td_proc->p_stat : -1,
779 td,
780 td->td_proc ? curthread->td_proc->p_pid : -1,
781 td->td_proc ? curthread->td_proc->p_stat : -1
782 );
783 panic("SCHED PANIC");
784 }
96728c05 785#endif
f1d1c3fa
MD
786 crit_enter();
787 if (td == curthread) {
788 _lwkt_enqueue(td);
789 } else {
790 lwkt_wait_t w;
791
792 /*
793 * If the thread is on a wait list we have to send our scheduling
794 * request to the owner of the wait structure. Otherwise we send
795 * the scheduling request to the cpu owning the thread. Races
796 * are ok, the target will forward the message as necessary (the
797 * message may chase the thread around before it finally gets
798 * acted upon).
799 *
800 * (remember, wait structures use stable storage)
801 */
802 if ((w = td->td_wait) != NULL) {
96728c05 803 if (lwkt_trytoken(&w->wa_token)) {
f1d1c3fa
MD
804 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
805 --w->wa_count;
806 td->td_wait = NULL;
75cdbe6c 807 if (smp_active == 0 || td->td_gd == mycpu) {
f1d1c3fa 808 _lwkt_enqueue(td);
57c254db 809 if (td->td_preemptable) {
96728c05 810 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
57c254db
MD
811 } else if (_lwkt_wantresched(td, curthread)) {
812 need_resched();
813 }
f1d1c3fa 814 } else {
a72187e9 815 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 816 }
96728c05 817 lwkt_reltoken(&w->wa_token);
f1d1c3fa 818 } else {
96728c05 819 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
820 }
821 } else {
822 /*
823 * If the wait structure is NULL and we own the thread, there
824 * is no race (since we are in a critical section). If we
825 * do not own the thread there might be a race but the
826 * target cpu will deal with it.
827 */
75cdbe6c 828 if (smp_active == 0 || td->td_gd == mycpu) {
f1d1c3fa 829 _lwkt_enqueue(td);
57c254db 830 if (td->td_preemptable) {
96728c05 831 td->td_preemptable(td, TDPRI_CRIT);
57c254db
MD
832 } else if (_lwkt_wantresched(td, curthread)) {
833 need_resched();
834 }
f1d1c3fa 835 } else {
a72187e9 836 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
837 }
838 }
8ad65e08 839 }
f1d1c3fa 840 crit_exit();
8ad65e08
MD
841}
842
d9eea1a5
MD
843/*
844 * Managed acquisition. This code assumes that the MP lock is held for
845 * the tdallq operation and that the thread has been descheduled from its
846 * original cpu. We also have to wait for the thread to be entirely switched
847 * out on its original cpu (this is usually fast enough that we never loop)
848 * since the LWKT system does not have to hold the MP lock while switching
849 * and the target may have released it before switching.
850 */
a2a5ad0d
MD
851void
852lwkt_acquire(thread_t td)
853{
854 struct globaldata *gd;
855
856 gd = td->td_gd;
857 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5
MD
858 while (td->td_flags & TDF_RUNNING) /* XXX spin */
859 ;
a2a5ad0d
MD
860 if (gd != mycpu) {
861 crit_enter();
862 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
863 gd = mycpu;
864 td->td_gd = gd;
a2a5ad0d
MD
865 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
866 crit_exit();
867 }
868}
869
8ad65e08 870/*
f1d1c3fa
MD
871 * Deschedule a thread.
872 *
873 * (non-blocking, self contained on a per cpu basis)
874 */
875void
876lwkt_deschedule_self(void)
877{
878 thread_t td = curthread;
879
880 crit_enter();
881 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa
MD
882 _lwkt_dequeue(td);
883 crit_exit();
884}
885
886/*
887 * Generic deschedule. Descheduling threads other then your own should be
888 * done only in carefully controlled circumstances. Descheduling is
889 * asynchronous.
890 *
891 * This function may block if the cpu has run out of messages.
8ad65e08
MD
892 */
893void
894lwkt_deschedule(thread_t td)
895{
f1d1c3fa
MD
896 crit_enter();
897 if (td == curthread) {
898 _lwkt_dequeue(td);
899 } else {
a72187e9 900 if (td->td_gd == mycpu) {
f1d1c3fa
MD
901 _lwkt_dequeue(td);
902 } else {
a72187e9 903 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
904 }
905 }
906 crit_exit();
907}
908
4b5f931b
MD
909/*
910 * Set the target thread's priority. This routine does not automatically
911 * switch to a higher priority thread, LWKT threads are not designed for
912 * continuous priority changes. Yield if you want to switch.
913 *
914 * We have to retain the critical section count which uses the high bits
26a0694b
MD
915 * of the td_pri field. The specified priority may also indicate zero or
916 * more critical sections by adding TDPRI_CRIT*N.
4b5f931b
MD
917 */
918void
919lwkt_setpri(thread_t td, int pri)
920{
26a0694b 921 KKASSERT(pri >= 0);
a72187e9 922 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
923 crit_enter();
924 if (td->td_flags & TDF_RUNQ) {
925 _lwkt_dequeue(td);
926 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
927 _lwkt_enqueue(td);
928 } else {
929 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
930 }
931 crit_exit();
932}
933
934void
935lwkt_setpri_self(int pri)
936{
937 thread_t td = curthread;
938
4b5f931b
MD
939 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
940 crit_enter();
941 if (td->td_flags & TDF_RUNQ) {
942 _lwkt_dequeue(td);
943 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
944 _lwkt_enqueue(td);
945 } else {
946 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
947 }
948 crit_exit();
949}
950
951struct proc *
952lwkt_preempted_proc(void)
953{
73e4f7b9 954 thread_t td = curthread;
4b5f931b
MD
955 while (td->td_preempted)
956 td = td->td_preempted;
957 return(td->td_proc);
958}
959
ece04fd0 960#if 0
4b5f931b 961
f1d1c3fa
MD
962/*
963 * This function deschedules the current thread and blocks on the specified
964 * wait queue. We obtain ownership of the wait queue in order to block
965 * on it. A generation number is used to interlock the wait queue in case
966 * it gets signalled while we are blocked waiting on the token.
967 *
968 * Note: alternatively we could dequeue our thread and then message the
969 * target cpu owning the wait queue. YYY implement as sysctl.
970 *
971 * Note: wait queue signals normally ping-pong the cpu as an optimization.
972 */
96728c05 973
f1d1c3fa 974void
ae8050a4 975lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
976{
977 thread_t td = curthread;
f1d1c3fa 978
f1d1c3fa 979 lwkt_gettoken(&w->wa_token);
ae8050a4 980 if (w->wa_gen == *gen) {
f1d1c3fa
MD
981 _lwkt_dequeue(td);
982 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
983 ++w->wa_count;
984 td->td_wait = w;
ae8050a4 985 td->td_wmesg = wmesg;
ece04fd0 986again:
f1d1c3fa 987 lwkt_switch();
ece04fd0
MD
988 lwkt_regettoken(&w->wa_token);
989 if (td->td_wmesg != NULL) {
990 _lwkt_dequeue(td);
991 goto again;
992 }
8ad65e08 993 }
ae8050a4
MD
994 /* token might be lost, doesn't matter for gen update */
995 *gen = w->wa_gen;
f1d1c3fa
MD
996 lwkt_reltoken(&w->wa_token);
997}
998
999/*
1000 * Signal a wait queue. We gain ownership of the wait queue in order to
1001 * signal it. Once a thread is removed from the wait queue we have to
1002 * deal with the cpu owning the thread.
1003 *
1004 * Note: alternatively we could message the target cpu owning the wait
1005 * queue. YYY implement as sysctl.
1006 */
1007void
ece04fd0 1008lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1009{
1010 thread_t td;
1011 int count;
1012
1013 lwkt_gettoken(&w->wa_token);
1014 ++w->wa_gen;
ece04fd0
MD
1015 if (count < 0)
1016 count = w->wa_count;
f1d1c3fa
MD
1017 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1018 --count;
1019 --w->wa_count;
1020 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1021 td->td_wait = NULL;
ae8050a4 1022 td->td_wmesg = NULL;
a72187e9 1023 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1024 _lwkt_enqueue(td);
1025 } else {
a72187e9 1026 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
1027 }
1028 lwkt_regettoken(&w->wa_token);
1029 }
1030 lwkt_reltoken(&w->wa_token);
1031}
1032
ece04fd0
MD
1033#endif
1034
99df837e
MD
1035/*
1036 * Create a kernel process/thread/whatever. It shares it's address space
1037 * with proc0 - ie: kernel only.
1038 *
365fa13f
MD
1039 * NOTE! By default new threads are created with the MP lock held. A
1040 * thread which does not require the MP lock should release it by calling
1041 * rel_mplock() at the start of the new thread.
99df837e
MD
1042 */
1043int
1044lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1045 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1046 const char *fmt, ...)
99df837e 1047{
73e4f7b9 1048 thread_t td;
e2565a42 1049 __va_list ap;
99df837e 1050
75cdbe6c 1051 td = lwkt_alloc_thread(template, cpu);
a2a5ad0d
MD
1052 if (tdp)
1053 *tdp = td;
709799ea 1054 cpu_set_thread_handler(td, lwkt_exit, func, arg);
ef0fdad1 1055 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1056#ifdef SMP
1057 td->td_mpcount = 1;
1058#endif
99df837e
MD
1059
1060 /*
1061 * Set up arg0 for 'ps' etc
1062 */
e2565a42 1063 __va_start(ap, fmt);
99df837e 1064 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1065 __va_end(ap);
99df837e
MD
1066
1067 /*
1068 * Schedule the thread to run
1069 */
ef0fdad1
MD
1070 if ((td->td_flags & TDF_STOPREQ) == 0)
1071 lwkt_schedule(td);
1072 else
1073 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1074 return 0;
1075}
1076
2d93b37a 1077/*
2d93b37a
MD
1078 * kthread_* is specific to the kernel and is not needed by userland.
1079 */
1080#ifdef _KERNEL
1081
99df837e
MD
1082/*
1083 * Destroy an LWKT thread. Warning! This function is not called when
1084 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1085 * uses a different reaping mechanism.
1086 */
1087void
1088lwkt_exit(void)
1089{
1090 thread_t td = curthread;
1091
1092 if (td->td_flags & TDF_VERBOSE)
1093 printf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1094 caps_exit(td);
99df837e
MD
1095 crit_enter();
1096 lwkt_deschedule_self();
1097 ++mycpu->gd_tdfreecount;
1098 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1099 cpu_thread_exit();
1100}
1101
1102/*
1103 * Create a kernel process/thread/whatever. It shares it's address space
ef0fdad1 1104 * with proc0 - ie: kernel only. 5.x compatible.
365fa13f
MD
1105 *
1106 * NOTE! By default kthreads are created with the MP lock held. A
1107 * thread which does not require the MP lock should release it by calling
1108 * rel_mplock() at the start of the new thread.
99df837e
MD
1109 */
1110int
1111kthread_create(void (*func)(void *), void *arg,
1112 struct thread **tdp, const char *fmt, ...)
1113{
73e4f7b9 1114 thread_t td;
e2565a42 1115 __va_list ap;
99df837e 1116
75cdbe6c 1117 td = lwkt_alloc_thread(NULL, -1);
a2a5ad0d
MD
1118 if (tdp)
1119 *tdp = td;
99df837e
MD
1120 cpu_set_thread_handler(td, kthread_exit, func, arg);
1121 td->td_flags |= TDF_VERBOSE;
8a8d5d85
MD
1122#ifdef SMP
1123 td->td_mpcount = 1;
1124#endif
99df837e
MD
1125
1126 /*
1127 * Set up arg0 for 'ps' etc
1128 */
e2565a42 1129 __va_start(ap, fmt);
99df837e 1130 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1131 __va_end(ap);
99df837e
MD
1132
1133 /*
1134 * Schedule the thread to run
1135 */
1136 lwkt_schedule(td);
1137 return 0;
1138}
1139
1140/*
1141 * Destroy an LWKT thread. Warning! This function is not called when
1142 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1143 * uses a different reaping mechanism.
1144 *
1145 * XXX duplicates lwkt_exit()
1146 */
1147void
1148kthread_exit(void)
1149{
1150 lwkt_exit();
1151}
1152
2d93b37a
MD
1153#endif /* _KERNEL */
1154
1155void
1156crit_panic(void)
1157{
1158 thread_t td = curthread;
1159 int lpri = td->td_pri;
1160
1161 td->td_pri = 0;
1162 panic("td_pri is/would-go negative! %p %d", td, lpri);
1163}
1164
96728c05
MD
1165#ifdef SMP
1166
1167/*
1168 * Send a function execution request to another cpu. The request is queued
1169 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1170 * possible target cpu. The FIFO can be written.
1171 *
1172 * YYY If the FIFO fills up we have to enable interrupts and process the
1173 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1174 * Create a CPU_*() function to do this!
1175 *
46a3f46d
MD
1176 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
1177 * end will take care of any pending interrupts.
1178 *
96728c05
MD
1179 * Must be called from a critical section.
1180 */
1181int
1182lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1183{
1184 lwkt_ipiq_t ip;
1185 int windex;
a2a5ad0d 1186 struct globaldata *gd = mycpu;
96728c05 1187
a2a5ad0d 1188 if (dcpu == gd->gd_cpuid) {
96728c05
MD
1189 func(arg);
1190 return(0);
1191 }
cb973d15 1192 crit_enter();
a2a5ad0d
MD
1193 ++gd->gd_intr_nesting_level;
1194#ifdef INVARIANTS
1195 if (gd->gd_intr_nesting_level > 20)
1196 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1197#endif
96728c05
MD
1198 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1199 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1200 ++ipiq_count;
a2a5ad0d 1201 ip = &gd->gd_ipiq[dcpu];
cb973d15
MD
1202
1203 /*
1204 * We always drain before the FIFO becomes full so it should never
1205 * become full. We need to leave enough entries to deal with
1206 * reentrancy.
1207 */
1208 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1209 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1210 ip->ip_func[windex] = func;
1211 ip->ip_arg[windex] = arg;
1212 /* YYY memory barrier */
1213 ++ip->ip_windex;
96728c05
MD
1214 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1215 unsigned int eflags = read_eflags();
1216 cpu_enable_intr();
1217 ++ipiq_fifofull;
cb973d15 1218 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
96728c05
MD
1219 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1220 lwkt_process_ipiq();
1221 }
1222 write_eflags(eflags);
1223 }
a2a5ad0d 1224 --gd->gd_intr_nesting_level;
96728c05 1225 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
cb973d15 1226 crit_exit();
96728c05
MD
1227 return(ip->ip_windex);
1228}
1229
cb973d15
MD
1230/*
1231 * Send a message to several target cpus. Typically used for scheduling.
435ff993 1232 * The message will not be sent to stopped cpus.
cb973d15
MD
1233 */
1234void
1235lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1236{
1237 int cpuid;
1238
435ff993 1239 mask &= ~stopped_cpus;
cb973d15
MD
1240 while (mask) {
1241 cpuid = bsfl(mask);
1242 lwkt_send_ipiq(cpuid, func, arg);
1243 mask &= ~(1 << cpuid);
1244 }
1245}
1246
96728c05
MD
1247/*
1248 * Wait for the remote cpu to finish processing a function.
1249 *
1250 * YYY we have to enable interrupts and process the IPIQ while waiting
1251 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1252 * function to do this! YYY we really should 'block' here.
1253 *
1254 * Must be called from a critical section. Thsi routine may be called
1255 * from an interrupt (for example, if an interrupt wakes a foreign thread
1256 * up).
1257 */
1258void
1259lwkt_wait_ipiq(int dcpu, int seq)
1260{
1261 lwkt_ipiq_t ip;
a2a5ad0d 1262 int maxc = 100000000;
96728c05
MD
1263
1264 if (dcpu != mycpu->gd_cpuid) {
1265 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1266 ip = &mycpu->gd_ipiq[dcpu];
cb973d15 1267 if ((int)(ip->ip_xindex - seq) < 0) {
96728c05
MD
1268 unsigned int eflags = read_eflags();
1269 cpu_enable_intr();
cb973d15 1270 while ((int)(ip->ip_xindex - seq) < 0) {
96728c05 1271 lwkt_process_ipiq();
a2a5ad0d 1272 if (--maxc == 0)
cb973d15 1273 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
a2a5ad0d
MD
1274 if (maxc < -1000000)
1275 panic("LWKT_WAIT_IPIQ");
96728c05
MD
1276 }
1277 write_eflags(eflags);
1278 }
1279 }
1280}
1281
1282/*
1283 * Called from IPI interrupt (like a fast interrupt), which has placed
1284 * us in a critical section. The MP lock may or may not be held.
cb973d15
MD
1285 * May also be called from doreti or splz, or be reentrantly called
1286 * indirectly through the ip_func[] we run.
88c4d2f6
MD
1287 *
1288 * There are two versions, one where no interrupt frame is available (when
1289 * called from the send code and from splz, and one where an interrupt
1290 * frame is available.
96728c05
MD
1291 */
1292void
1293lwkt_process_ipiq(void)
1294{
1295 int n;
1296 int cpuid = mycpu->gd_cpuid;
1297
1298 for (n = 0; n < ncpus; ++n) {
1299 lwkt_ipiq_t ip;
1300 int ri;
1301
1302 if (n == cpuid)
1303 continue;
1304 ip = globaldata_find(n)->gd_ipiq;
1305 if (ip == NULL)
1306 continue;
1307 ip = &ip[cpuid];
cb973d15
MD
1308
1309 /*
1310 * Note: xindex is only updated after we are sure the function has
1311 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1312 * function may send an IPI which may block/drain.
1313 */
96728c05
MD
1314 while (ip->ip_rindex != ip->ip_windex) {
1315 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
96728c05 1316 ++ip->ip_rindex;
88c4d2f6
MD
1317 ip->ip_func[ri](ip->ip_arg[ri], NULL);
1318 /* YYY memory barrier */
1319 ip->ip_xindex = ip->ip_rindex;
1320 }
1321 }
1322}
1323
0df69494 1324#ifdef _KERNEL
88c4d2f6
MD
1325void
1326lwkt_process_ipiq_frame(struct intrframe frame)
1327{
1328 int n;
1329 int cpuid = mycpu->gd_cpuid;
1330
1331 for (n = 0; n < ncpus; ++n) {
1332 lwkt_ipiq_t ip;
1333 int ri;
1334
1335 if (n == cpuid)
1336 continue;
1337 ip = globaldata_find(n)->gd_ipiq;
1338 if (ip == NULL)
1339 continue;
1340 ip = &ip[cpuid];
1341
1342 /*
1343 * Note: xindex is only updated after we are sure the function has
1344 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1345 * function may send an IPI which may block/drain.
1346 */
1347 while (ip->ip_rindex != ip->ip_windex) {
1348 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
1349 ++ip->ip_rindex;
1350 ip->ip_func[ri](ip->ip_arg[ri], &frame);
cb973d15
MD
1351 /* YYY memory barrier */
1352 ip->ip_xindex = ip->ip_rindex;
96728c05
MD
1353 }
1354 }
1355}
0df69494 1356#endif
96728c05
MD
1357
1358#else
1359
1360int
1361lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1362{
1363 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1364 return(0); /* NOT REACHED */
1365}
1366
1367void
1368lwkt_wait_ipiq(int dcpu, int seq)
1369{
1370 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);
1371}
1372
1373#endif