Try to close an occassional VM page related panic that is believed to occur
[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 *
a9e5ae82 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.69 2004/09/21 18:09:57 joerg 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) {
a9e5ae82 258#ifdef _KERNEL
f470d0c8 259 kmem_free(kernel_map, (vm_offset_t)stack, td->td_kstack_size);
a9e5ae82
JS
260#else
261 libcaps_free_stack(stack, td->td_kstack_size);
262#endif
f470d0c8
MD
263 stack = NULL;
264 }
265 }
266 if (stack == NULL) {
05220613 267#ifdef _KERNEL
f470d0c8 268 stack = (void *)kmem_alloc(kernel_map, stksize);
05220613 269#else
f470d0c8 270 stack = libcaps_alloc_stack(stksize);
05220613 271#endif
ef0fdad1 272 flags |= TDF_ALLOCATED_STACK;
99df837e 273 }
75cdbe6c 274 if (cpu < 0)
f470d0c8 275 lwkt_init_thread(td, stack, stksize, flags, mycpu);
75cdbe6c 276 else
f470d0c8 277 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 278 return(td);
7d0bac62
MD
279}
280
709799ea
MD
281#ifdef _KERNEL
282
7d0bac62
MD
283/*
284 * Initialize a preexisting thread structure. This function is used by
285 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
286 *
f8c3996b
MD
287 * All threads start out in a critical section at a priority of
288 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
289 * appropriate. This function may send an IPI message when the
290 * requested cpu is not the current cpu and consequently gd_tdallq may
291 * not be initialized synchronously from the point of view of the originating
292 * cpu.
293 *
294 * NOTE! we have to be careful in regards to creating threads for other cpus
295 * if SMP has not yet been activated.
7d0bac62 296 */
41a01a4d
MD
297#ifdef SMP
298
75cdbe6c
MD
299static void
300lwkt_init_thread_remote(void *arg)
301{
302 thread_t td = arg;
303
304 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
305}
306
41a01a4d
MD
307#endif
308
7d0bac62 309void
f470d0c8
MD
310lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
311 struct globaldata *gd)
7d0bac62 312{
37af14fe
MD
313 globaldata_t mygd = mycpu;
314
99df837e
MD
315 bzero(td, sizeof(struct thread));
316 td->td_kstack = stack;
f470d0c8 317 td->td_kstack_size = stksize;
99df837e 318 td->td_flags |= flags;
26a0694b 319 td->td_gd = gd;
f8c3996b 320 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
c95cd171 321 lwkt_initport(&td->td_msgport, td);
99df837e 322 pmap_init_thread(td);
0f7a3396 323#ifdef SMP
5d21b981
MD
324 /*
325 * Normally initializing a thread for a remote cpu requires sending an
326 * IPI. However, the idlethread is setup before the other cpus are
327 * activated so we have to treat it as a special case. XXX manipulation
328 * of gd_tdallq requires the BGL.
329 */
330 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 331 crit_enter_gd(mygd);
75cdbe6c 332 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 333 crit_exit_gd(mygd);
75cdbe6c 334 } else {
2db3b277 335 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 336 }
0f7a3396 337#else
37af14fe 338 crit_enter_gd(mygd);
0f7a3396 339 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 340 crit_exit_gd(mygd);
0f7a3396 341#endif
73e4f7b9
MD
342}
343
2d93b37a
MD
344#endif /* _KERNEL */
345
73e4f7b9
MD
346void
347lwkt_set_comm(thread_t td, const char *ctl, ...)
348{
e2565a42 349 __va_list va;
73e4f7b9 350
e2565a42 351 __va_start(va, ctl);
73e4f7b9 352 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 353 __va_end(va);
7d0bac62
MD
354}
355
99df837e 356void
73e4f7b9 357lwkt_hold(thread_t td)
99df837e 358{
73e4f7b9
MD
359 ++td->td_refs;
360}
361
362void
363lwkt_rele(thread_t td)
364{
365 KKASSERT(td->td_refs > 0);
366 --td->td_refs;
367}
368
c95cd171
MD
369#ifdef _KERNEL
370
73e4f7b9
MD
371void
372lwkt_wait_free(thread_t td)
373{
374 while (td->td_refs)
377d4740 375 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
376}
377
c95cd171
MD
378#endif
379
73e4f7b9
MD
380void
381lwkt_free_thread(thread_t td)
382{
383 struct globaldata *gd = mycpu;
384
d9eea1a5 385 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
386 ("lwkt_free_thread: did not exit! %p", td));
387
37af14fe 388 crit_enter_gd(gd);
73e4f7b9
MD
389 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
390 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
391 (td->td_flags & TDF_ALLOCATED_THREAD)
392 ) {
73e4f7b9
MD
393 ++gd->gd_tdfreecount;
394 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
37af14fe 395 crit_exit_gd(gd);
99df837e 396 } else {
37af14fe 397 crit_exit_gd(gd);
99df837e 398 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613 399#ifdef _KERNEL
f470d0c8 400 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
05220613 401#else
f470d0c8 402 libcaps_free_stack(td->td_kstack, td->td_kstack_size);
05220613 403#endif
73e4f7b9 404 /* gd invalid */
99df837e 405 td->td_kstack = NULL;
f470d0c8 406 td->td_kstack_size = 0;
99df837e 407 }
05220613
MD
408 if (td->td_flags & TDF_ALLOCATED_THREAD) {
409#ifdef _KERNEL
99df837e 410 zfree(thread_zone, td);
05220613
MD
411#else
412 free(td);
413#endif
414 }
99df837e
MD
415 }
416}
417
418
8ad65e08
MD
419/*
420 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
421 * switch to the idlethread. Switching must occur within a critical
422 * section to avoid races with the scheduling queue.
423 *
424 * We always have full control over our cpu's run queue. Other cpus
425 * that wish to manipulate our queue must use the cpu_*msg() calls to
426 * talk to our cpu, so a critical section is all that is needed and
427 * the result is very, very fast thread switching.
428 *
96728c05
MD
429 * The LWKT scheduler uses a fixed priority model and round-robins at
430 * each priority level. User process scheduling is a totally
431 * different beast and LWKT priorities should not be confused with
432 * user process priorities.
f1d1c3fa 433 *
96728c05
MD
434 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
435 * cleans it up. Note that the td_switch() function cannot do anything that
436 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
437 * the target thread (not the current thread). It's nice to have a scheduler
438 * that does not need the MP lock to work because it allows us to do some
439 * really cool high-performance MP lock optimizations.
8ad65e08 440 */
96728c05 441
8ad65e08
MD
442void
443lwkt_switch(void)
444{
37af14fe
MD
445 globaldata_t gd = mycpu;
446 thread_t td = gd->gd_curthread;
8ad65e08 447 thread_t ntd;
8a8d5d85
MD
448#ifdef SMP
449 int mpheld;
450#endif
8ad65e08 451
46a3f46d
MD
452 /*
453 * Switching from within a 'fast' (non thread switched) interrupt is
454 * illegal.
455 */
37af14fe 456 if (gd->gd_intr_nesting_level && panicstr == NULL) {
fc92d4aa 457 panic("lwkt_switch: cannot switch from within a fast interrupt, yet");
96728c05 458 }
ef0fdad1 459
cb973d15
MD
460 /*
461 * Passive release (used to transition from user to kernel mode
462 * when we block or switch rather then when we enter the kernel).
463 * This function is NOT called if we are switching into a preemption
464 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
465 * our current process designation (if we have one) and become a true
466 * LWKT thread, and may also hand the current process designation to
467 * another process and schedule thread.
cb973d15
MD
468 */
469 if (td->td_release)
470 td->td_release(td);
471
37af14fe 472 crit_enter_gd(gd);
4b5f931b 473 ++switch_count;
8a8d5d85
MD
474
475#ifdef SMP
476 /*
477 * td_mpcount cannot be used to determine if we currently hold the
478 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
479 * to get the lock, and switch out if it can't. Our ownership of
480 * the actual lock will remain stable while we are in a critical section
481 * (but, of course, another cpu may own or release the lock so the
482 * actual value of mp_lock is not stable).
8a8d5d85
MD
483 */
484 mpheld = MP_LOCK_HELD();
0f7a3396
MD
485#ifdef INVARIANTS
486 if (td->td_cscount) {
487 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
488 td);
489 if (panic_on_cscount)
490 panic("switching while mastering cpusync");
491 }
492#endif
8a8d5d85 493#endif
99df837e
MD
494 if ((ntd = td->td_preempted) != NULL) {
495 /*
496 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
497 * thread. This occurs transparently, whether the preempted thread
498 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
499 * itself).
500 *
501 * We have to setup the MP lock for the original thread after backing
502 * out the adjustment that was made to curthread when the original
503 * was preempted.
99df837e 504 */
26a0694b 505 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 506#ifdef SMP
96728c05 507 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 508 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
509 td, ntd, td->td_mpcount, ntd->td_mpcount);
510 }
8a8d5d85
MD
511 if (ntd->td_mpcount) {
512 td->td_mpcount -= ntd->td_mpcount;
513 KKASSERT(td->td_mpcount >= 0);
514 }
515#endif
26a0694b 516 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
517
518 /*
519 * XXX. The interrupt may have woken a thread up, we need to properly
520 * set the reschedule flag if the originally interrupted thread is at
521 * a lower priority.
522 */
523 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
524 need_lwkt_resched();
8a8d5d85 525 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 526 } else {
4b5f931b
MD
527 /*
528 * Priority queue / round-robin at each priority. Note that user
529 * processes run at a fixed, low priority and the user process
530 * scheduler deals with interactions between user processes
531 * by scheduling and descheduling them from the LWKT queue as
532 * necessary.
8a8d5d85
MD
533 *
534 * We have to adjust the MP lock for the target thread. If we
535 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
536 * thread that does not need the MP lock. If we cannot, we spin
537 * instead of HLT.
538 *
539 * A similar issue exists for the tokens held by the target thread.
540 * If we cannot obtain ownership of the tokens we cannot immediately
541 * schedule the thread.
542 */
543
544 /*
545 * We are switching threads. If there are any pending requests for
546 * tokens we can satisfy all of them here.
4b5f931b 547 */
41a01a4d
MD
548#ifdef SMP
549 if (gd->gd_tokreqbase)
550 lwkt_drain_token_requests();
551#endif
552
8ec60c3f
MD
553 /*
554 * If an LWKT reschedule was requested, well that is what we are
555 * doing now so clear it.
556 */
557 clear_lwkt_resched();
4b5f931b
MD
558again:
559 if (gd->gd_runqmask) {
560 int nq = bsrl(gd->gd_runqmask);
561 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
562 gd->gd_runqmask &= ~(1 << nq);
563 goto again;
564 }
8a8d5d85 565#ifdef SMP
41a01a4d
MD
566 /*
567 * If the target needs the MP lock and we couldn't get it,
568 * or if the target is holding tokens and we could not
569 * gain ownership of the tokens, continue looking for a
570 * thread to schedule and spin instead of HLT if we can't.
571 */
572 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
573 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
574 ) {
8a8d5d85
MD
575 u_int32_t rqmask = gd->gd_runqmask;
576 while (rqmask) {
577 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
41a01a4d
MD
578 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock())
579 continue;
580 mpheld = MP_LOCK_HELD();
581 if (ntd->td_toks && !lwkt_chktokens(ntd))
582 continue;
583 break;
8a8d5d85
MD
584 }
585 if (ntd)
586 break;
587 rqmask &= ~(1 << nq);
588 nq = bsrl(rqmask);
589 }
590 if (ntd == NULL) {
a2a5ad0d
MD
591 ntd = &gd->gd_idlethread;
592 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
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 {
598 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
599 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
600 }
601#else
4b5f931b
MD
602 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
603 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 604#endif
4b5f931b 605 } else {
3c23a41a 606 /*
60f945af
MD
607 * We have nothing to run but only let the idle loop halt
608 * the cpu if there are no pending interrupts.
3c23a41a 609 */
a2a5ad0d 610 ntd = &gd->gd_idlethread;
60f945af 611 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 612 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 613 }
f1d1c3fa 614 }
26a0694b
MD
615 KASSERT(ntd->td_pri >= TDPRI_CRIT,
616 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
617
618 /*
619 * Do the actual switch. If the new target does not need the MP lock
620 * and we are holding it, release the MP lock. If the new target requires
621 * the MP lock we have already acquired it for the target.
622 */
623#ifdef SMP
624 if (ntd->td_mpcount == 0 ) {
625 if (MP_LOCK_HELD())
626 cpu_rel_mplock();
627 } else {
628 ASSERT_MP_LOCK_HELD();
629 }
630#endif
37af14fe 631 if (td != ntd)
f1d1c3fa 632 td->td_switch(ntd);
37af14fe
MD
633 /* NOTE: current cpu may have changed after switch */
634 crit_exit_quick(td);
8ad65e08
MD
635}
636
b68b7282 637/*
96728c05
MD
638 * Request that the target thread preempt the current thread. Preemption
639 * only works under a specific set of conditions:
b68b7282 640 *
96728c05
MD
641 * - We are not preempting ourselves
642 * - The target thread is owned by the current cpu
643 * - We are not currently being preempted
644 * - The target is not currently being preempted
645 * - We are able to satisfy the target's MP lock requirements (if any).
646 *
647 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
648 * this is called via lwkt_schedule() through the td_preemptable callback.
649 * critpri is the managed critical priority that we should ignore in order
650 * to determine whether preemption is possible (aka usually just the crit
651 * priority of lwkt_schedule() itself).
b68b7282 652 *
26a0694b
MD
653 * XXX at the moment we run the target thread in a critical section during
654 * the preemption in order to prevent the target from taking interrupts
655 * that *WE* can't. Preemption is strictly limited to interrupt threads
656 * and interrupt-like threads, outside of a critical section, and the
657 * preempted source thread will be resumed the instant the target blocks
658 * whether or not the source is scheduled (i.e. preemption is supposed to
659 * be as transparent as possible).
4b5f931b 660 *
8a8d5d85
MD
661 * The target thread inherits our MP count (added to its own) for the
662 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
663 * MP lock during the preemption. Therefore, any preempting targets must be
664 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
665 * out of sync with the physical mp_lock, but we do not have to preserve
666 * the original ownership of the lock if it was out of synch (that is, we
667 * can leave it synchronized on return).
b68b7282
MD
668 */
669void
96728c05 670lwkt_preempt(thread_t ntd, int critpri)
b68b7282 671{
46a3f46d 672 struct globaldata *gd = mycpu;
0a3f9b47 673 thread_t td;
8a8d5d85
MD
674#ifdef SMP
675 int mpheld;
57c254db 676 int savecnt;
8a8d5d85 677#endif
b68b7282 678
26a0694b 679 /*
96728c05
MD
680 * The caller has put us in a critical section. We can only preempt
681 * if the caller of the caller was not in a critical section (basically
0a3f9b47 682 * a local interrupt), as determined by the 'critpri' parameter.
96728c05
MD
683 *
684 * YYY The target thread must be in a critical section (else it must
685 * inherit our critical section? I dunno yet).
41a01a4d
MD
686 *
687 * Any tokens held by the target may not be held by thread(s) being
688 * preempted. We take the easy way out and do not preempt if
689 * the target is holding tokens.
0a3f9b47
MD
690 *
691 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
692 */
693 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 694
0a3f9b47 695 td = gd->gd_curthread;
0a3f9b47 696 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
697 ++preempt_miss;
698 return;
699 }
96728c05
MD
700 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
701 ++preempt_miss;
8ec60c3f 702 need_lwkt_resched();
96728c05
MD
703 return;
704 }
705#ifdef SMP
46a3f46d 706 if (ntd->td_gd != gd) {
96728c05 707 ++preempt_miss;
8ec60c3f 708 need_lwkt_resched();
96728c05
MD
709 return;
710 }
711#endif
41a01a4d
MD
712 /*
713 * Take the easy way out and do not preempt if the target is holding
714 * one or more tokens. We could test whether the thread(s) being
715 * preempted interlock against the target thread's tokens and whether
716 * we can get all the target thread's tokens, but this situation
717 * should not occur very often so its easier to simply not preempt.
718 */
719 if (ntd->td_toks != NULL) {
720 ++preempt_miss;
8ec60c3f 721 need_lwkt_resched();
41a01a4d
MD
722 return;
723 }
26a0694b
MD
724 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
725 ++preempt_weird;
8ec60c3f 726 need_lwkt_resched();
26a0694b
MD
727 return;
728 }
729 if (ntd->td_preempted) {
4b5f931b 730 ++preempt_hit;
8ec60c3f 731 need_lwkt_resched();
26a0694b 732 return;
b68b7282 733 }
8a8d5d85 734#ifdef SMP
a2a5ad0d
MD
735 /*
736 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
737 * to or from the MP lock. In this case td_mpcount will be pre-disposed
738 * (non-zero) but not actually synchronized with the actual state of the
739 * lock. We can use it to imply an MP lock requirement for the
740 * preemption but we cannot use it to test whether we hold the MP lock
741 * or not.
a2a5ad0d 742 */
96728c05 743 savecnt = td->td_mpcount;
71ef2f5c 744 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
745 ntd->td_mpcount += td->td_mpcount;
746 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
747 ntd->td_mpcount -= td->td_mpcount;
748 ++preempt_miss;
8ec60c3f 749 need_lwkt_resched();
8a8d5d85
MD
750 return;
751 }
752#endif
26a0694b 753
8ec60c3f
MD
754 /*
755 * Since we are able to preempt the current thread, there is no need to
756 * call need_lwkt_resched().
757 */
26a0694b
MD
758 ++preempt_hit;
759 ntd->td_preempted = td;
760 td->td_flags |= TDF_PREEMPT_LOCK;
761 td->td_switch(ntd);
762 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
763#ifdef SMP
764 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
765 mpheld = MP_LOCK_HELD();
766 if (mpheld && td->td_mpcount == 0)
96728c05 767 cpu_rel_mplock();
71ef2f5c 768 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
769 panic("lwkt_preempt(): MP lock was not held through");
770#endif
26a0694b
MD
771 ntd->td_preempted = NULL;
772 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
773}
774
f1d1c3fa
MD
775/*
776 * Yield our thread while higher priority threads are pending. This is
777 * typically called when we leave a critical section but it can be safely
778 * called while we are in a critical section.
779 *
780 * This function will not generally yield to equal priority threads but it
781 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 782 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
783 * lwkt_yield_quick().
784 *
235957ed 785 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
786 * came along but was blocked and made pending.
787 *
f1d1c3fa
MD
788 * (self contained on a per cpu basis)
789 */
790void
791lwkt_yield_quick(void)
792{
7966cb69
MD
793 globaldata_t gd = mycpu;
794 thread_t td = gd->gd_curthread;
ef0fdad1 795
a2a5ad0d 796 /*
235957ed 797 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
798 * it with a non-zero cpl then we might not wind up calling splz after
799 * a task switch when the critical section is exited even though the
46a3f46d 800 * new task could accept the interrupt.
a2a5ad0d
MD
801 *
802 * XXX from crit_exit() only called after last crit section is released.
803 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
804 *
805 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
806 * except for this special case, we MUST call splz() here to handle any
807 * pending ints, particularly after we switch, or we might accidently
808 * halt the cpu with interrupts pending.
a2a5ad0d 809 */
46a3f46d 810 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 811 splz();
f1d1c3fa
MD
812
813 /*
814 * YYY enabling will cause wakeup() to task-switch, which really
815 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
816 * preemption and MP without actually doing preemption or MP, because a
817 * lot of code assumes that wakeup() does not block.
f1d1c3fa 818 */
46a3f46d
MD
819 if (untimely_switch && td->td_nest_count == 0 &&
820 gd->gd_intr_nesting_level == 0
821 ) {
37af14fe 822 crit_enter_quick(td);
f1d1c3fa
MD
823 /*
824 * YYY temporary hacks until we disassociate the userland scheduler
825 * from the LWKT scheduler.
826 */
827 if (td->td_flags & TDF_RUNQ) {
828 lwkt_switch(); /* will not reenter yield function */
829 } else {
37af14fe 830 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 831 lwkt_switch(); /* will not reenter yield function */
37af14fe 832 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 833 }
7966cb69 834 crit_exit_noyield(td);
f1d1c3fa 835 }
f1d1c3fa
MD
836}
837
8ad65e08 838/*
f1d1c3fa 839 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 840 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
841 * the crit_exit() call in lwkt_switch().
842 *
843 * (self contained on a per cpu basis)
8ad65e08
MD
844 */
845void
f1d1c3fa 846lwkt_yield(void)
8ad65e08 847{
37af14fe 848 lwkt_schedule_self(curthread);
f1d1c3fa
MD
849 lwkt_switch();
850}
851
8ad65e08 852/*
f1d1c3fa
MD
853 * Generic schedule. Possibly schedule threads belonging to other cpus and
854 * deal with threads that might be blocked on a wait queue.
855 *
0a3f9b47
MD
856 * We have a little helper inline function which does additional work after
857 * the thread has been enqueued, including dealing with preemption and
858 * setting need_lwkt_resched() (which prevents the kernel from returning
859 * to userland until it has processed higher priority threads).
8ad65e08 860 */
0a3f9b47
MD
861static __inline
862void
8ec60c3f 863_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
0a3f9b47
MD
864{
865 if (ntd->td_preemptable) {
866 ntd->td_preemptable(ntd, cpri); /* YYY +token */
8ec60c3f
MD
867 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
868 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
869 ) {
870 need_lwkt_resched();
0a3f9b47
MD
871 }
872}
873
8ad65e08
MD
874void
875lwkt_schedule(thread_t td)
876{
37af14fe
MD
877 globaldata_t mygd = mycpu;
878
96728c05 879#ifdef INVARIANTS
41a01a4d 880 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
26a0694b
MD
881 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
882 && td->td_proc->p_stat == SSLEEP
883 ) {
884 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
885 curthread,
886 curthread->td_proc ? curthread->td_proc->p_pid : -1,
887 curthread->td_proc ? curthread->td_proc->p_stat : -1,
888 td,
889 td->td_proc ? curthread->td_proc->p_pid : -1,
890 td->td_proc ? curthread->td_proc->p_stat : -1
891 );
892 panic("SCHED PANIC");
893 }
96728c05 894#endif
37af14fe
MD
895 crit_enter_gd(mygd);
896 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
897 _lwkt_enqueue(td);
898 } else {
899 lwkt_wait_t w;
900
901 /*
902 * If the thread is on a wait list we have to send our scheduling
903 * request to the owner of the wait structure. Otherwise we send
904 * the scheduling request to the cpu owning the thread. Races
905 * are ok, the target will forward the message as necessary (the
906 * message may chase the thread around before it finally gets
907 * acted upon).
908 *
909 * (remember, wait structures use stable storage)
0a3f9b47
MD
910 *
911 * NOTE: tokens no longer enter a critical section, so we only need
912 * to account for the crit_enter() above when calling
913 * _lwkt_schedule_post().
f1d1c3fa
MD
914 */
915 if ((w = td->td_wait) != NULL) {
41a01a4d
MD
916 lwkt_tokref wref;
917
918 if (lwkt_trytoken(&wref, &w->wa_token)) {
f1d1c3fa
MD
919 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
920 --w->wa_count;
921 td->td_wait = NULL;
0f7a3396 922#ifdef SMP
8ec60c3f 923 if (td->td_gd == mygd) {
f1d1c3fa 924 _lwkt_enqueue(td);
8ec60c3f 925 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 926 } else {
2db3b277 927 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 928 }
0f7a3396
MD
929#else
930 _lwkt_enqueue(td);
8ec60c3f 931 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 932#endif
41a01a4d 933 lwkt_reltoken(&wref);
f1d1c3fa 934 } else {
96728c05 935 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
936 }
937 } else {
938 /*
939 * If the wait structure is NULL and we own the thread, there
940 * is no race (since we are in a critical section). If we
941 * do not own the thread there might be a race but the
942 * target cpu will deal with it.
943 */
0f7a3396 944#ifdef SMP
37af14fe 945 if (td->td_gd == mygd) {
f1d1c3fa 946 _lwkt_enqueue(td);
8ec60c3f 947 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 948 } else {
2db3b277 949 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 950 }
0f7a3396
MD
951#else
952 _lwkt_enqueue(td);
8ec60c3f 953 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 954#endif
f1d1c3fa 955 }
8ad65e08 956 }
37af14fe 957 crit_exit_gd(mygd);
8ad65e08
MD
958}
959
d9eea1a5
MD
960/*
961 * Managed acquisition. This code assumes that the MP lock is held for
962 * the tdallq operation and that the thread has been descheduled from its
963 * original cpu. We also have to wait for the thread to be entirely switched
964 * out on its original cpu (this is usually fast enough that we never loop)
965 * since the LWKT system does not have to hold the MP lock while switching
966 * and the target may have released it before switching.
967 */
a2a5ad0d
MD
968void
969lwkt_acquire(thread_t td)
970{
37af14fe
MD
971 globaldata_t gd;
972 globaldata_t mygd;
a2a5ad0d
MD
973
974 gd = td->td_gd;
37af14fe 975 mygd = mycpu;
a2a5ad0d 976 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5 977 while (td->td_flags & TDF_RUNNING) /* XXX spin */
5d21b981 978 cpu_mb1();
37af14fe
MD
979 if (gd != mygd) {
980 crit_enter_gd(mygd);
a2a5ad0d 981 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
37af14fe
MD
982 td->td_gd = mygd;
983 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
984 crit_exit_gd(mygd);
a2a5ad0d
MD
985 }
986}
987
f1d1c3fa
MD
988/*
989 * Generic deschedule. Descheduling threads other then your own should be
990 * done only in carefully controlled circumstances. Descheduling is
991 * asynchronous.
992 *
993 * This function may block if the cpu has run out of messages.
8ad65e08
MD
994 */
995void
996lwkt_deschedule(thread_t td)
997{
f1d1c3fa
MD
998 crit_enter();
999 if (td == curthread) {
1000 _lwkt_dequeue(td);
1001 } else {
a72187e9 1002 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1003 _lwkt_dequeue(td);
1004 } else {
2db3b277 1005 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
1006 }
1007 }
1008 crit_exit();
1009}
1010
4b5f931b
MD
1011/*
1012 * Set the target thread's priority. This routine does not automatically
1013 * switch to a higher priority thread, LWKT threads are not designed for
1014 * continuous priority changes. Yield if you want to switch.
1015 *
1016 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1017 * of the td_pri field. The specified priority may also indicate zero or
1018 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1019 *
1020 * Note that we requeue the thread whether it winds up on a different runq
1021 * or not. uio_yield() depends on this and the routine is not normally
1022 * called with the same priority otherwise.
4b5f931b
MD
1023 */
1024void
1025lwkt_setpri(thread_t td, int pri)
1026{
26a0694b 1027 KKASSERT(pri >= 0);
a72187e9 1028 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1029 crit_enter();
1030 if (td->td_flags & TDF_RUNQ) {
1031 _lwkt_dequeue(td);
1032 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1033 _lwkt_enqueue(td);
1034 } else {
1035 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1036 }
1037 crit_exit();
1038}
1039
1040void
1041lwkt_setpri_self(int pri)
1042{
1043 thread_t td = curthread;
1044
4b5f931b
MD
1045 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1046 crit_enter();
1047 if (td->td_flags & TDF_RUNQ) {
1048 _lwkt_dequeue(td);
1049 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1050 _lwkt_enqueue(td);
1051 } else {
1052 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1053 }
1054 crit_exit();
1055}
1056
8ec60c3f
MD
1057/*
1058 * Determine if there is a runnable thread at a higher priority then
1059 * the current thread. lwkt_setpri() does not check this automatically.
1060 * Return 1 if there is, 0 if there isn't.
1061 *
1062 * Example: if bit 31 of runqmask is set and the current thread is priority
1063 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1064 *
1065 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1066 * up comparing against 0xffffffff, a comparison that will always be false.
1067 */
1068int
1069lwkt_checkpri_self(void)
1070{
1071 globaldata_t gd = mycpu;
1072 thread_t td = gd->gd_curthread;
1073 int nq = td->td_pri & TDPRI_MASK;
1074
1075 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1076 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1077 return(1);
1078 ++nq;
1079 }
1080 return(0);
1081}
1082
5d21b981
MD
1083/*
1084 * Migrate the current thread to the specified cpu. The BGL must be held
1085 * (for the gd_tdallq manipulation XXX). This is accomplished by
1086 * descheduling ourselves from the current cpu, moving our thread to the
1087 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1088 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1089 */
3d28ff59 1090#ifdef SMP
5d21b981 1091static void lwkt_setcpu_remote(void *arg);
3d28ff59 1092#endif
5d21b981
MD
1093
1094void
1095lwkt_setcpu_self(globaldata_t rgd)
1096{
1097#ifdef SMP
1098 thread_t td = curthread;
1099
1100 if (td->td_gd != rgd) {
1101 crit_enter_quick(td);
1102 td->td_flags |= TDF_MIGRATING;
1103 lwkt_deschedule_self(td);
1104 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1105 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
1106 lwkt_send_ipiq(rgd, (ipifunc_t)lwkt_setcpu_remote, td);
1107 lwkt_switch();
1108 /* we are now on the target cpu */
1109 crit_exit_quick(td);
1110 }
1111#endif
1112}
1113
1114/*
1115 * Remote IPI for cpu migration (called while in a critical section so we
1116 * do not have to enter another one). The thread has already been moved to
1117 * our cpu's allq, but we must wait for the thread to be completely switched
1118 * out on the originating cpu before we schedule it on ours or the stack
1119 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1120 * change to main memory.
1121 *
1122 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1123 * against wakeups. It is best if this interface is used only when there
1124 * are no pending events that might try to schedule the thread.
1125 */
3d28ff59 1126#ifdef SMP
5d21b981
MD
1127static void
1128lwkt_setcpu_remote(void *arg)
1129{
1130 thread_t td = arg;
1131 globaldata_t gd = mycpu;
1132
1133 while (td->td_flags & TDF_RUNNING)
1134 cpu_mb1();
1135 td->td_gd = gd;
1136 cpu_mb2();
1137 td->td_flags &= ~TDF_MIGRATING;
1138 _lwkt_enqueue(td);
1139}
3d28ff59 1140#endif
5d21b981 1141
4b5f931b
MD
1142struct proc *
1143lwkt_preempted_proc(void)
1144{
73e4f7b9 1145 thread_t td = curthread;
4b5f931b
MD
1146 while (td->td_preempted)
1147 td = td->td_preempted;
1148 return(td->td_proc);
1149}
1150
f1d1c3fa 1151/*
41a01a4d
MD
1152 * Block on the specified wait queue until signaled. A generation number
1153 * must be supplied to interlock the wait queue. The function will
1154 * return immediately if the generation number does not match the wait
1155 * structure's generation number.
f1d1c3fa
MD
1156 */
1157void
ae8050a4 1158lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
1159{
1160 thread_t td = curthread;
41a01a4d 1161 lwkt_tokref ilock;
f1d1c3fa 1162
41a01a4d
MD
1163 lwkt_gettoken(&ilock, &w->wa_token);
1164 crit_enter();
ae8050a4 1165 if (w->wa_gen == *gen) {
f1d1c3fa
MD
1166 _lwkt_dequeue(td);
1167 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1168 ++w->wa_count;
1169 td->td_wait = w;
ae8050a4 1170 td->td_wmesg = wmesg;
41a01a4d 1171 again:
f1d1c3fa 1172 lwkt_switch();
ece04fd0
MD
1173 if (td->td_wmesg != NULL) {
1174 _lwkt_dequeue(td);
1175 goto again;
1176 }
8ad65e08 1177 }
41a01a4d 1178 crit_exit();
ae8050a4 1179 *gen = w->wa_gen;
41a01a4d 1180 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1181}
1182
1183/*
1184 * Signal a wait queue. We gain ownership of the wait queue in order to
1185 * signal it. Once a thread is removed from the wait queue we have to
1186 * deal with the cpu owning the thread.
1187 *
1188 * Note: alternatively we could message the target cpu owning the wait
1189 * queue. YYY implement as sysctl.
1190 */
1191void
ece04fd0 1192lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1193{
1194 thread_t td;
41a01a4d 1195 lwkt_tokref ilock;
f1d1c3fa 1196
41a01a4d 1197 lwkt_gettoken(&ilock, &w->wa_token);
f1d1c3fa 1198 ++w->wa_gen;
41a01a4d 1199 crit_enter();
ece04fd0
MD
1200 if (count < 0)
1201 count = w->wa_count;
f1d1c3fa
MD
1202 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1203 --count;
1204 --w->wa_count;
1205 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1206 td->td_wait = NULL;
ae8050a4 1207 td->td_wmesg = NULL;
a72187e9 1208 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1209 _lwkt_enqueue(td);
1210 } else {
2db3b277 1211 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 1212 }
f1d1c3fa 1213 }
41a01a4d
MD
1214 crit_exit();
1215 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1216}
1217
99df837e
MD
1218/*
1219 * Create a kernel process/thread/whatever. It shares it's address space
1220 * with proc0 - ie: kernel only.
1221 *
365fa13f
MD
1222 * NOTE! By default new threads are created with the MP lock held. A
1223 * thread which does not require the MP lock should release it by calling
1224 * rel_mplock() at the start of the new thread.
99df837e
MD
1225 */
1226int
1227lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1228 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1229 const char *fmt, ...)
99df837e 1230{
73e4f7b9 1231 thread_t td;
e2565a42 1232 __va_list ap;
99df837e 1233
f470d0c8 1234 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu);
a2a5ad0d
MD
1235 if (tdp)
1236 *tdp = td;
709799ea 1237 cpu_set_thread_handler(td, lwkt_exit, func, arg);
ef0fdad1 1238 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1239#ifdef SMP
1240 td->td_mpcount = 1;
1241#endif
99df837e
MD
1242
1243 /*
1244 * Set up arg0 for 'ps' etc
1245 */
e2565a42 1246 __va_start(ap, fmt);
99df837e 1247 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1248 __va_end(ap);
99df837e
MD
1249
1250 /*
1251 * Schedule the thread to run
1252 */
ef0fdad1
MD
1253 if ((td->td_flags & TDF_STOPREQ) == 0)
1254 lwkt_schedule(td);
1255 else
1256 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1257 return 0;
1258}
1259
2d93b37a 1260/*
2d93b37a
MD
1261 * kthread_* is specific to the kernel and is not needed by userland.
1262 */
1263#ifdef _KERNEL
1264
99df837e
MD
1265/*
1266 * Destroy an LWKT thread. Warning! This function is not called when
1267 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1268 * uses a different reaping mechanism.
1269 */
1270void
1271lwkt_exit(void)
1272{
1273 thread_t td = curthread;
8826f33a 1274 globaldata_t gd;
99df837e
MD
1275
1276 if (td->td_flags & TDF_VERBOSE)
1277 printf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1278 caps_exit(td);
37af14fe
MD
1279 crit_enter_quick(td);
1280 lwkt_deschedule_self(td);
8826f33a
MD
1281 gd = mycpu;
1282 KKASSERT(gd == td->td_gd);
1283 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1284 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1285 ++gd->gd_tdfreecount;
1286 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1287 }
99df837e
MD
1288 cpu_thread_exit();
1289}
1290
2d93b37a
MD
1291#endif /* _KERNEL */
1292
1293void
1294crit_panic(void)
1295{
1296 thread_t td = curthread;
1297 int lpri = td->td_pri;
1298
1299 td->td_pri = 0;
1300 panic("td_pri is/would-go negative! %p %d", td, lpri);
1301}
1302