Fix Saved commands bug. The following Makefile would cause make to core.
[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 *
38717797 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.76 2005/07/07 20:28:26 hmp 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>
b950d4b8
EN
88#include <machine/atomic.h>
89#include <machine/cpu.h>
05220613
MD
90
91#endif
92
7d0bac62 93static int untimely_switch = 0;
0f7a3396
MD
94#ifdef INVARIANTS
95static int panic_on_cscount = 0;
96#endif
05220613
MD
97static __int64_t switch_count = 0;
98static __int64_t preempt_hit = 0;
99static __int64_t preempt_miss = 0;
100static __int64_t preempt_weird = 0;
38717797
HP
101static __int64_t token_contention_count = 0;
102static __int64_t mplock_contention_count = 0;
05220613
MD
103
104#ifdef _KERNEL
105
106SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
0f7a3396
MD
107#ifdef INVARIANTS
108SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
109#endif
4b5f931b 110SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 111SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 112SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 113SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
114#ifdef INVARIANTS
115SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
116 &token_contention_count, 0, "spinning due to token contention");
117SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
118 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
119#endif
05220613
MD
120#endif
121
4b5f931b
MD
122/*
123 * These helper procedures handle the runq, they can only be called from
124 * within a critical section.
75cdbe6c
MD
125 *
126 * WARNING! Prior to SMP being brought up it is possible to enqueue and
127 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
128 * instead of 'mycpu' when referencing the globaldata structure. Once
129 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 130 */
f1d1c3fa
MD
131static __inline
132void
133_lwkt_dequeue(thread_t td)
134{
135 if (td->td_flags & TDF_RUNQ) {
4b5f931b 136 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 137 struct globaldata *gd = td->td_gd;
4b5f931b 138
f1d1c3fa 139 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
140 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
141 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
142 }
143}
144
145static __inline
146void
147_lwkt_enqueue(thread_t td)
148{
5d21b981 149 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING)) == 0) {
4b5f931b 150 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 151 struct globaldata *gd = td->td_gd;
4b5f931b 152
f1d1c3fa 153 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
154 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
155 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
156 }
157}
8ad65e08 158
37af14fe
MD
159/*
160 * Schedule a thread to run. As the current thread we can always safely
161 * schedule ourselves, and a shortcut procedure is provided for that
162 * function.
163 *
164 * (non-blocking, self contained on a per cpu basis)
165 */
166void
167lwkt_schedule_self(thread_t td)
168{
169 crit_enter_quick(td);
170 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
171 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
172 _lwkt_enqueue(td);
173#ifdef _KERNEL
174 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
175 panic("SCHED SELF PANIC");
176#endif
177 crit_exit_quick(td);
178}
179
180/*
181 * Deschedule a thread.
182 *
183 * (non-blocking, self contained on a per cpu basis)
184 */
185void
186lwkt_deschedule_self(thread_t td)
187{
188 crit_enter_quick(td);
189 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
190 _lwkt_dequeue(td);
191 crit_exit_quick(td);
192}
193
2d93b37a
MD
194#ifdef _KERNEL
195
8ad65e08
MD
196/*
197 * LWKTs operate on a per-cpu basis
198 *
73e4f7b9 199 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
200 */
201void
202lwkt_gdinit(struct globaldata *gd)
203{
4b5f931b
MD
204 int i;
205
206 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
207 TAILQ_INIT(&gd->gd_tdrunq[i]);
208 gd->gd_runqmask = 0;
73e4f7b9 209 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
210}
211
2d93b37a
MD
212#endif /* _KERNEL */
213
7d0bac62
MD
214/*
215 * Initialize a thread wait structure prior to first use.
216 *
217 * NOTE! called from low level boot code, we cannot do anything fancy!
218 */
219void
41a01a4d 220lwkt_wait_init(lwkt_wait_t w)
7d0bac62 221{
41a01a4d 222 lwkt_token_init(&w->wa_token);
7d0bac62 223 TAILQ_INIT(&w->wa_waitq);
41a01a4d
MD
224 w->wa_gen = 0;
225 w->wa_count = 0;
7d0bac62
MD
226}
227
228/*
229 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
230 * or LWKT start address before it can be scheduled. If the target cpu is
231 * -1 the thread will be created on the current cpu.
0cfcada1
MD
232 *
233 * If you intend to create a thread without a process context this function
234 * does everything except load the startup and switcher function.
7d0bac62
MD
235 */
236thread_t
f470d0c8 237lwkt_alloc_thread(struct thread *td, int stksize, int cpu)
7d0bac62 238{
99df837e 239 void *stack;
ef0fdad1 240 int flags = 0;
37af14fe 241 globaldata_t gd = mycpu;
7d0bac62 242
ef0fdad1 243 if (td == NULL) {
37af14fe
MD
244 crit_enter_gd(gd);
245 if (gd->gd_tdfreecount > 0) {
246 --gd->gd_tdfreecount;
247 td = TAILQ_FIRST(&gd->gd_tdfreeq);
d9eea1a5 248 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1 249 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
37af14fe
MD
250 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq);
251 crit_exit_gd(gd);
ef0fdad1
MD
252 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
253 } else {
37af14fe 254 crit_exit_gd(gd);
05220613 255#ifdef _KERNEL
ef0fdad1 256 td = zalloc(thread_zone);
05220613
MD
257#else
258 td = malloc(sizeof(struct thread));
259#endif
ef0fdad1 260 td->td_kstack = NULL;
f470d0c8 261 td->td_kstack_size = 0;
ef0fdad1
MD
262 flags |= TDF_ALLOCATED_THREAD;
263 }
264 }
f470d0c8
MD
265 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
266 if (flags & TDF_ALLOCATED_STACK) {
a9e5ae82 267#ifdef _KERNEL
f470d0c8 268 kmem_free(kernel_map, (vm_offset_t)stack, td->td_kstack_size);
a9e5ae82
JS
269#else
270 libcaps_free_stack(stack, td->td_kstack_size);
271#endif
f470d0c8
MD
272 stack = NULL;
273 }
274 }
275 if (stack == NULL) {
05220613 276#ifdef _KERNEL
f470d0c8 277 stack = (void *)kmem_alloc(kernel_map, stksize);
05220613 278#else
f470d0c8 279 stack = libcaps_alloc_stack(stksize);
05220613 280#endif
ef0fdad1 281 flags |= TDF_ALLOCATED_STACK;
99df837e 282 }
75cdbe6c 283 if (cpu < 0)
f470d0c8 284 lwkt_init_thread(td, stack, stksize, flags, mycpu);
75cdbe6c 285 else
f470d0c8 286 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 287 return(td);
7d0bac62
MD
288}
289
709799ea
MD
290#ifdef _KERNEL
291
7d0bac62
MD
292/*
293 * Initialize a preexisting thread structure. This function is used by
294 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
295 *
f8c3996b
MD
296 * All threads start out in a critical section at a priority of
297 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
298 * appropriate. This function may send an IPI message when the
299 * requested cpu is not the current cpu and consequently gd_tdallq may
300 * not be initialized synchronously from the point of view of the originating
301 * cpu.
302 *
303 * NOTE! we have to be careful in regards to creating threads for other cpus
304 * if SMP has not yet been activated.
7d0bac62 305 */
41a01a4d
MD
306#ifdef SMP
307
75cdbe6c
MD
308static void
309lwkt_init_thread_remote(void *arg)
310{
311 thread_t td = arg;
312
313 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
314}
315
41a01a4d
MD
316#endif
317
7d0bac62 318void
f470d0c8
MD
319lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
320 struct globaldata *gd)
7d0bac62 321{
37af14fe
MD
322 globaldata_t mygd = mycpu;
323
99df837e
MD
324 bzero(td, sizeof(struct thread));
325 td->td_kstack = stack;
f470d0c8 326 td->td_kstack_size = stksize;
99df837e 327 td->td_flags |= flags;
26a0694b 328 td->td_gd = gd;
f8c3996b 329 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
c95cd171 330 lwkt_initport(&td->td_msgport, td);
99df837e 331 pmap_init_thread(td);
0f7a3396 332#ifdef SMP
5d21b981
MD
333 /*
334 * Normally initializing a thread for a remote cpu requires sending an
335 * IPI. However, the idlethread is setup before the other cpus are
336 * activated so we have to treat it as a special case. XXX manipulation
337 * of gd_tdallq requires the BGL.
338 */
339 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 340 crit_enter_gd(mygd);
75cdbe6c 341 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 342 crit_exit_gd(mygd);
75cdbe6c 343 } else {
2db3b277 344 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 345 }
0f7a3396 346#else
37af14fe 347 crit_enter_gd(mygd);
0f7a3396 348 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 349 crit_exit_gd(mygd);
0f7a3396 350#endif
73e4f7b9
MD
351}
352
2d93b37a
MD
353#endif /* _KERNEL */
354
73e4f7b9
MD
355void
356lwkt_set_comm(thread_t td, const char *ctl, ...)
357{
e2565a42 358 __va_list va;
73e4f7b9 359
e2565a42 360 __va_start(va, ctl);
73e4f7b9 361 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 362 __va_end(va);
7d0bac62
MD
363}
364
99df837e 365void
73e4f7b9 366lwkt_hold(thread_t td)
99df837e 367{
73e4f7b9
MD
368 ++td->td_refs;
369}
370
371void
372lwkt_rele(thread_t td)
373{
374 KKASSERT(td->td_refs > 0);
375 --td->td_refs;
376}
377
c95cd171
MD
378#ifdef _KERNEL
379
73e4f7b9
MD
380void
381lwkt_wait_free(thread_t td)
382{
383 while (td->td_refs)
377d4740 384 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
385}
386
c95cd171
MD
387#endif
388
73e4f7b9
MD
389void
390lwkt_free_thread(thread_t td)
391{
392 struct globaldata *gd = mycpu;
393
d9eea1a5 394 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
395 ("lwkt_free_thread: did not exit! %p", td));
396
37af14fe 397 crit_enter_gd(gd);
73e4f7b9
MD
398 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
399 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
400 (td->td_flags & TDF_ALLOCATED_THREAD)
401 ) {
73e4f7b9
MD
402 ++gd->gd_tdfreecount;
403 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
37af14fe 404 crit_exit_gd(gd);
99df837e 405 } else {
37af14fe 406 crit_exit_gd(gd);
99df837e 407 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613 408#ifdef _KERNEL
f470d0c8 409 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
05220613 410#else
f470d0c8 411 libcaps_free_stack(td->td_kstack, td->td_kstack_size);
05220613 412#endif
73e4f7b9 413 /* gd invalid */
99df837e 414 td->td_kstack = NULL;
f470d0c8 415 td->td_kstack_size = 0;
99df837e 416 }
05220613
MD
417 if (td->td_flags & TDF_ALLOCATED_THREAD) {
418#ifdef _KERNEL
99df837e 419 zfree(thread_zone, td);
05220613
MD
420#else
421 free(td);
422#endif
423 }
99df837e
MD
424 }
425}
426
427
8ad65e08
MD
428/*
429 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
430 * switch to the idlethread. Switching must occur within a critical
431 * section to avoid races with the scheduling queue.
432 *
433 * We always have full control over our cpu's run queue. Other cpus
434 * that wish to manipulate our queue must use the cpu_*msg() calls to
435 * talk to our cpu, so a critical section is all that is needed and
436 * the result is very, very fast thread switching.
437 *
96728c05
MD
438 * The LWKT scheduler uses a fixed priority model and round-robins at
439 * each priority level. User process scheduling is a totally
440 * different beast and LWKT priorities should not be confused with
441 * user process priorities.
f1d1c3fa 442 *
96728c05
MD
443 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
444 * cleans it up. Note that the td_switch() function cannot do anything that
445 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
446 * the target thread (not the current thread). It's nice to have a scheduler
447 * that does not need the MP lock to work because it allows us to do some
448 * really cool high-performance MP lock optimizations.
8ad65e08 449 */
96728c05 450
8ad65e08
MD
451void
452lwkt_switch(void)
453{
37af14fe
MD
454 globaldata_t gd = mycpu;
455 thread_t td = gd->gd_curthread;
8ad65e08 456 thread_t ntd;
8a8d5d85
MD
457#ifdef SMP
458 int mpheld;
459#endif
8ad65e08 460
46a3f46d
MD
461 /*
462 * Switching from within a 'fast' (non thread switched) interrupt is
463 * illegal.
464 */
37af14fe 465 if (gd->gd_intr_nesting_level && panicstr == NULL) {
2cb56a3e 466 panic("lwkt_switch: cannot switch from within a fast interrupt, yet, td %p\n", td);
96728c05 467 }
ef0fdad1 468
cb973d15
MD
469 /*
470 * Passive release (used to transition from user to kernel mode
471 * when we block or switch rather then when we enter the kernel).
472 * This function is NOT called if we are switching into a preemption
473 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
474 * our current process designation (if we have one) and become a true
475 * LWKT thread, and may also hand the current process designation to
476 * another process and schedule thread.
cb973d15
MD
477 */
478 if (td->td_release)
479 td->td_release(td);
480
37af14fe 481 crit_enter_gd(gd);
8a8d5d85
MD
482
483#ifdef SMP
484 /*
485 * td_mpcount cannot be used to determine if we currently hold the
486 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
487 * to get the lock, and switch out if it can't. Our ownership of
488 * the actual lock will remain stable while we are in a critical section
489 * (but, of course, another cpu may own or release the lock so the
490 * actual value of mp_lock is not stable).
8a8d5d85
MD
491 */
492 mpheld = MP_LOCK_HELD();
0f7a3396
MD
493#ifdef INVARIANTS
494 if (td->td_cscount) {
495 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
496 td);
497 if (panic_on_cscount)
498 panic("switching while mastering cpusync");
499 }
500#endif
8a8d5d85 501#endif
99df837e
MD
502 if ((ntd = td->td_preempted) != NULL) {
503 /*
504 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
505 * thread. This occurs transparently, whether the preempted thread
506 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
507 * itself).
508 *
509 * We have to setup the MP lock for the original thread after backing
510 * out the adjustment that was made to curthread when the original
511 * was preempted.
99df837e 512 */
26a0694b 513 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 514#ifdef SMP
96728c05 515 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 516 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
517 td, ntd, td->td_mpcount, ntd->td_mpcount);
518 }
8a8d5d85
MD
519 if (ntd->td_mpcount) {
520 td->td_mpcount -= ntd->td_mpcount;
521 KKASSERT(td->td_mpcount >= 0);
522 }
523#endif
26a0694b 524 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
525
526 /*
527 * XXX. The interrupt may have woken a thread up, we need to properly
528 * set the reschedule flag if the originally interrupted thread is at
529 * a lower priority.
530 */
531 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
532 need_lwkt_resched();
8a8d5d85 533 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 534 } else {
4b5f931b
MD
535 /*
536 * Priority queue / round-robin at each priority. Note that user
537 * processes run at a fixed, low priority and the user process
538 * scheduler deals with interactions between user processes
539 * by scheduling and descheduling them from the LWKT queue as
540 * necessary.
8a8d5d85
MD
541 *
542 * We have to adjust the MP lock for the target thread. If we
543 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
544 * thread that does not need the MP lock. If we cannot, we spin
545 * instead of HLT.
546 *
547 * A similar issue exists for the tokens held by the target thread.
548 * If we cannot obtain ownership of the tokens we cannot immediately
549 * schedule the thread.
550 */
551
552 /*
553 * We are switching threads. If there are any pending requests for
554 * tokens we can satisfy all of them here.
4b5f931b 555 */
41a01a4d
MD
556#ifdef SMP
557 if (gd->gd_tokreqbase)
558 lwkt_drain_token_requests();
559#endif
560
8ec60c3f
MD
561 /*
562 * If an LWKT reschedule was requested, well that is what we are
563 * doing now so clear it.
564 */
565 clear_lwkt_resched();
4b5f931b
MD
566again:
567 if (gd->gd_runqmask) {
568 int nq = bsrl(gd->gd_runqmask);
569 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
570 gd->gd_runqmask &= ~(1 << nq);
571 goto again;
572 }
8a8d5d85 573#ifdef SMP
41a01a4d
MD
574 /*
575 * If the target needs the MP lock and we couldn't get it,
576 * or if the target is holding tokens and we could not
577 * gain ownership of the tokens, continue looking for a
578 * thread to schedule and spin instead of HLT if we can't.
579 */
580 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
581 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
582 ) {
8a8d5d85
MD
583 u_int32_t rqmask = gd->gd_runqmask;
584 while (rqmask) {
585 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797
HP
586 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
587 /* spinning due to MP lock being held */
588#ifdef INVARIANTS
589 ++mplock_contention_count;
590#endif
41a01a4d 591 continue;
38717797 592 }
41a01a4d 593 mpheld = MP_LOCK_HELD();
38717797
HP
594 if (ntd->td_toks && !lwkt_chktokens(ntd)) {
595 /* spinning due to token contention */
596#ifdef INVARIANTS
597 ++token_contention_count;
598#endif
41a01a4d 599 continue;
38717797 600 }
41a01a4d 601 break;
8a8d5d85
MD
602 }
603 if (ntd)
604 break;
605 rqmask &= ~(1 << nq);
606 nq = bsrl(rqmask);
607 }
608 if (ntd == NULL) {
a2a5ad0d
MD
609 ntd = &gd->gd_idlethread;
610 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
611 } else {
612 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
613 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
614 }
615 } else {
616 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
617 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
618 }
619#else
4b5f931b
MD
620 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
621 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 622#endif
4b5f931b 623 } else {
3c23a41a 624 /*
60f945af
MD
625 * We have nothing to run but only let the idle loop halt
626 * the cpu if there are no pending interrupts.
3c23a41a 627 */
a2a5ad0d 628 ntd = &gd->gd_idlethread;
60f945af 629 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 630 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 631 }
f1d1c3fa 632 }
26a0694b
MD
633 KASSERT(ntd->td_pri >= TDPRI_CRIT,
634 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
635
636 /*
637 * Do the actual switch. If the new target does not need the MP lock
638 * and we are holding it, release the MP lock. If the new target requires
639 * the MP lock we have already acquired it for the target.
640 */
641#ifdef SMP
642 if (ntd->td_mpcount == 0 ) {
643 if (MP_LOCK_HELD())
644 cpu_rel_mplock();
645 } else {
646 ASSERT_MP_LOCK_HELD();
647 }
648#endif
94f6d86e
MD
649 if (td != ntd) {
650 ++switch_count;
f1d1c3fa 651 td->td_switch(ntd);
94f6d86e 652 }
37af14fe
MD
653 /* NOTE: current cpu may have changed after switch */
654 crit_exit_quick(td);
8ad65e08
MD
655}
656
b68b7282 657/*
96728c05
MD
658 * Request that the target thread preempt the current thread. Preemption
659 * only works under a specific set of conditions:
b68b7282 660 *
96728c05
MD
661 * - We are not preempting ourselves
662 * - The target thread is owned by the current cpu
663 * - We are not currently being preempted
664 * - The target is not currently being preempted
665 * - We are able to satisfy the target's MP lock requirements (if any).
666 *
667 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
668 * this is called via lwkt_schedule() through the td_preemptable callback.
669 * critpri is the managed critical priority that we should ignore in order
670 * to determine whether preemption is possible (aka usually just the crit
671 * priority of lwkt_schedule() itself).
b68b7282 672 *
26a0694b
MD
673 * XXX at the moment we run the target thread in a critical section during
674 * the preemption in order to prevent the target from taking interrupts
675 * that *WE* can't. Preemption is strictly limited to interrupt threads
676 * and interrupt-like threads, outside of a critical section, and the
677 * preempted source thread will be resumed the instant the target blocks
678 * whether or not the source is scheduled (i.e. preemption is supposed to
679 * be as transparent as possible).
4b5f931b 680 *
8a8d5d85
MD
681 * The target thread inherits our MP count (added to its own) for the
682 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
683 * MP lock during the preemption. Therefore, any preempting targets must be
684 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
685 * out of sync with the physical mp_lock, but we do not have to preserve
686 * the original ownership of the lock if it was out of synch (that is, we
687 * can leave it synchronized on return).
b68b7282
MD
688 */
689void
96728c05 690lwkt_preempt(thread_t ntd, int critpri)
b68b7282 691{
46a3f46d 692 struct globaldata *gd = mycpu;
0a3f9b47 693 thread_t td;
8a8d5d85
MD
694#ifdef SMP
695 int mpheld;
57c254db 696 int savecnt;
8a8d5d85 697#endif
b68b7282 698
26a0694b 699 /*
96728c05
MD
700 * The caller has put us in a critical section. We can only preempt
701 * if the caller of the caller was not in a critical section (basically
0a3f9b47 702 * a local interrupt), as determined by the 'critpri' parameter.
96728c05
MD
703 *
704 * YYY The target thread must be in a critical section (else it must
705 * inherit our critical section? I dunno yet).
41a01a4d
MD
706 *
707 * Any tokens held by the target may not be held by thread(s) being
708 * preempted. We take the easy way out and do not preempt if
709 * the target is holding tokens.
0a3f9b47
MD
710 *
711 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
712 */
713 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 714
0a3f9b47 715 td = gd->gd_curthread;
0a3f9b47 716 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
717 ++preempt_miss;
718 return;
719 }
96728c05
MD
720 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
721 ++preempt_miss;
8ec60c3f 722 need_lwkt_resched();
96728c05
MD
723 return;
724 }
725#ifdef SMP
46a3f46d 726 if (ntd->td_gd != gd) {
96728c05 727 ++preempt_miss;
8ec60c3f 728 need_lwkt_resched();
96728c05
MD
729 return;
730 }
731#endif
41a01a4d
MD
732 /*
733 * Take the easy way out and do not preempt if the target is holding
734 * one or more tokens. We could test whether the thread(s) being
735 * preempted interlock against the target thread's tokens and whether
736 * we can get all the target thread's tokens, but this situation
737 * should not occur very often so its easier to simply not preempt.
738 */
739 if (ntd->td_toks != NULL) {
740 ++preempt_miss;
8ec60c3f 741 need_lwkt_resched();
41a01a4d
MD
742 return;
743 }
26a0694b
MD
744 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
745 ++preempt_weird;
8ec60c3f 746 need_lwkt_resched();
26a0694b
MD
747 return;
748 }
749 if (ntd->td_preempted) {
4b5f931b 750 ++preempt_hit;
8ec60c3f 751 need_lwkt_resched();
26a0694b 752 return;
b68b7282 753 }
8a8d5d85 754#ifdef SMP
a2a5ad0d
MD
755 /*
756 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
757 * to or from the MP lock. In this case td_mpcount will be pre-disposed
758 * (non-zero) but not actually synchronized with the actual state of the
759 * lock. We can use it to imply an MP lock requirement for the
760 * preemption but we cannot use it to test whether we hold the MP lock
761 * or not.
a2a5ad0d 762 */
96728c05 763 savecnt = td->td_mpcount;
71ef2f5c 764 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
765 ntd->td_mpcount += td->td_mpcount;
766 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
767 ntd->td_mpcount -= td->td_mpcount;
768 ++preempt_miss;
8ec60c3f 769 need_lwkt_resched();
8a8d5d85
MD
770 return;
771 }
772#endif
26a0694b 773
8ec60c3f
MD
774 /*
775 * Since we are able to preempt the current thread, there is no need to
776 * call need_lwkt_resched().
777 */
26a0694b
MD
778 ++preempt_hit;
779 ntd->td_preempted = td;
780 td->td_flags |= TDF_PREEMPT_LOCK;
781 td->td_switch(ntd);
782 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
783#ifdef SMP
784 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
785 mpheld = MP_LOCK_HELD();
786 if (mpheld && td->td_mpcount == 0)
96728c05 787 cpu_rel_mplock();
71ef2f5c 788 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
789 panic("lwkt_preempt(): MP lock was not held through");
790#endif
26a0694b
MD
791 ntd->td_preempted = NULL;
792 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
793}
794
f1d1c3fa
MD
795/*
796 * Yield our thread while higher priority threads are pending. This is
797 * typically called when we leave a critical section but it can be safely
798 * called while we are in a critical section.
799 *
800 * This function will not generally yield to equal priority threads but it
801 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 802 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
803 * lwkt_yield_quick().
804 *
235957ed 805 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
806 * came along but was blocked and made pending.
807 *
f1d1c3fa
MD
808 * (self contained on a per cpu basis)
809 */
810void
811lwkt_yield_quick(void)
812{
7966cb69
MD
813 globaldata_t gd = mycpu;
814 thread_t td = gd->gd_curthread;
ef0fdad1 815
a2a5ad0d 816 /*
235957ed 817 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
818 * it with a non-zero cpl then we might not wind up calling splz after
819 * a task switch when the critical section is exited even though the
46a3f46d 820 * new task could accept the interrupt.
a2a5ad0d
MD
821 *
822 * XXX from crit_exit() only called after last crit section is released.
823 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
824 *
825 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
826 * except for this special case, we MUST call splz() here to handle any
827 * pending ints, particularly after we switch, or we might accidently
828 * halt the cpu with interrupts pending.
a2a5ad0d 829 */
46a3f46d 830 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 831 splz();
f1d1c3fa
MD
832
833 /*
834 * YYY enabling will cause wakeup() to task-switch, which really
835 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
836 * preemption and MP without actually doing preemption or MP, because a
837 * lot of code assumes that wakeup() does not block.
f1d1c3fa 838 */
46a3f46d
MD
839 if (untimely_switch && td->td_nest_count == 0 &&
840 gd->gd_intr_nesting_level == 0
841 ) {
37af14fe 842 crit_enter_quick(td);
f1d1c3fa
MD
843 /*
844 * YYY temporary hacks until we disassociate the userland scheduler
845 * from the LWKT scheduler.
846 */
847 if (td->td_flags & TDF_RUNQ) {
848 lwkt_switch(); /* will not reenter yield function */
849 } else {
37af14fe 850 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 851 lwkt_switch(); /* will not reenter yield function */
37af14fe 852 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 853 }
7966cb69 854 crit_exit_noyield(td);
f1d1c3fa 855 }
f1d1c3fa
MD
856}
857
8ad65e08 858/*
f1d1c3fa 859 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 860 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
861 * the crit_exit() call in lwkt_switch().
862 *
863 * (self contained on a per cpu basis)
8ad65e08
MD
864 */
865void
f1d1c3fa 866lwkt_yield(void)
8ad65e08 867{
37af14fe 868 lwkt_schedule_self(curthread);
f1d1c3fa
MD
869 lwkt_switch();
870}
871
8ad65e08 872/*
f1d1c3fa
MD
873 * Generic schedule. Possibly schedule threads belonging to other cpus and
874 * deal with threads that might be blocked on a wait queue.
875 *
0a3f9b47
MD
876 * We have a little helper inline function which does additional work after
877 * the thread has been enqueued, including dealing with preemption and
878 * setting need_lwkt_resched() (which prevents the kernel from returning
879 * to userland until it has processed higher priority threads).
8ad65e08 880 */
0a3f9b47
MD
881static __inline
882void
8ec60c3f 883_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
0a3f9b47
MD
884{
885 if (ntd->td_preemptable) {
886 ntd->td_preemptable(ntd, cpri); /* YYY +token */
8ec60c3f
MD
887 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
888 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
889 ) {
890 need_lwkt_resched();
0a3f9b47
MD
891 }
892}
893
8ad65e08
MD
894void
895lwkt_schedule(thread_t td)
896{
37af14fe
MD
897 globaldata_t mygd = mycpu;
898
96728c05 899#ifdef INVARIANTS
41a01a4d 900 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
26a0694b
MD
901 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
902 && td->td_proc->p_stat == SSLEEP
903 ) {
904 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
905 curthread,
906 curthread->td_proc ? curthread->td_proc->p_pid : -1,
907 curthread->td_proc ? curthread->td_proc->p_stat : -1,
908 td,
909 td->td_proc ? curthread->td_proc->p_pid : -1,
910 td->td_proc ? curthread->td_proc->p_stat : -1
911 );
912 panic("SCHED PANIC");
913 }
96728c05 914#endif
37af14fe
MD
915 crit_enter_gd(mygd);
916 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
917 _lwkt_enqueue(td);
918 } else {
919 lwkt_wait_t w;
920
921 /*
922 * If the thread is on a wait list we have to send our scheduling
923 * request to the owner of the wait structure. Otherwise we send
924 * the scheduling request to the cpu owning the thread. Races
925 * are ok, the target will forward the message as necessary (the
926 * message may chase the thread around before it finally gets
927 * acted upon).
928 *
929 * (remember, wait structures use stable storage)
0a3f9b47 930 *
0c453950
MD
931 * NOTE: we have to account for the number of critical sections
932 * under our control when calling _lwkt_schedule_post() so it
933 * can figure out whether preemption is allowed.
934 *
935 * NOTE: The wait structure algorithms are a mess and need to be
936 * rewritten.
937 *
938 * NOTE: We cannot safely acquire or release a token, even
939 * non-blocking, because this routine may be called in the context
940 * of a thread already holding the token and thus not provide any
941 * interlock protection. We cannot safely manipulate the td_toks
942 * list for the same reason. Instead we depend on our critical
943 * section if the token is owned by our cpu.
f1d1c3fa
MD
944 */
945 if ((w = td->td_wait) != NULL) {
0c453950 946 if (w->wa_token.t_cpu == mygd) {
f1d1c3fa
MD
947 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
948 --w->wa_count;
949 td->td_wait = NULL;
0f7a3396 950#ifdef SMP
8ec60c3f 951 if (td->td_gd == mygd) {
f1d1c3fa 952 _lwkt_enqueue(td);
8ec60c3f 953 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 954 } else {
2db3b277 955 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 956 }
0f7a3396
MD
957#else
958 _lwkt_enqueue(td);
8ec60c3f 959 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 960#endif
f1d1c3fa 961 } else {
96728c05 962 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
963 }
964 } else {
965 /*
966 * If the wait structure is NULL and we own the thread, there
967 * is no race (since we are in a critical section). If we
968 * do not own the thread there might be a race but the
969 * target cpu will deal with it.
970 */
0f7a3396 971#ifdef SMP
37af14fe 972 if (td->td_gd == mygd) {
f1d1c3fa 973 _lwkt_enqueue(td);
8ec60c3f 974 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 975 } else {
2db3b277 976 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 977 }
0f7a3396
MD
978#else
979 _lwkt_enqueue(td);
8ec60c3f 980 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 981#endif
f1d1c3fa 982 }
8ad65e08 983 }
37af14fe 984 crit_exit_gd(mygd);
8ad65e08
MD
985}
986
d9eea1a5
MD
987/*
988 * Managed acquisition. This code assumes that the MP lock is held for
989 * the tdallq operation and that the thread has been descheduled from its
990 * original cpu. We also have to wait for the thread to be entirely switched
991 * out on its original cpu (this is usually fast enough that we never loop)
992 * since the LWKT system does not have to hold the MP lock while switching
993 * and the target may have released it before switching.
994 */
a2a5ad0d
MD
995void
996lwkt_acquire(thread_t td)
997{
37af14fe
MD
998 globaldata_t gd;
999 globaldata_t mygd;
a2a5ad0d
MD
1000
1001 gd = td->td_gd;
37af14fe 1002 mygd = mycpu;
35238fa5 1003 cpu_lfence();
a2a5ad0d 1004 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5 1005 while (td->td_flags & TDF_RUNNING) /* XXX spin */
35238fa5 1006 cpu_lfence();
37af14fe
MD
1007 if (gd != mygd) {
1008 crit_enter_gd(mygd);
a2a5ad0d 1009 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
37af14fe
MD
1010 td->td_gd = mygd;
1011 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
1012 crit_exit_gd(mygd);
a2a5ad0d
MD
1013 }
1014}
1015
f1d1c3fa
MD
1016/*
1017 * Generic deschedule. Descheduling threads other then your own should be
1018 * done only in carefully controlled circumstances. Descheduling is
1019 * asynchronous.
1020 *
1021 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1022 */
1023void
1024lwkt_deschedule(thread_t td)
1025{
f1d1c3fa
MD
1026 crit_enter();
1027 if (td == curthread) {
1028 _lwkt_dequeue(td);
1029 } else {
a72187e9 1030 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1031 _lwkt_dequeue(td);
1032 } else {
2db3b277 1033 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
1034 }
1035 }
1036 crit_exit();
1037}
1038
4b5f931b
MD
1039/*
1040 * Set the target thread's priority. This routine does not automatically
1041 * switch to a higher priority thread, LWKT threads are not designed for
1042 * continuous priority changes. Yield if you want to switch.
1043 *
1044 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1045 * of the td_pri field. The specified priority may also indicate zero or
1046 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1047 *
1048 * Note that we requeue the thread whether it winds up on a different runq
1049 * or not. uio_yield() depends on this and the routine is not normally
1050 * called with the same priority otherwise.
4b5f931b
MD
1051 */
1052void
1053lwkt_setpri(thread_t td, int pri)
1054{
26a0694b 1055 KKASSERT(pri >= 0);
a72187e9 1056 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1057 crit_enter();
1058 if (td->td_flags & TDF_RUNQ) {
1059 _lwkt_dequeue(td);
1060 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1061 _lwkt_enqueue(td);
1062 } else {
1063 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1064 }
1065 crit_exit();
1066}
1067
1068void
1069lwkt_setpri_self(int pri)
1070{
1071 thread_t td = curthread;
1072
4b5f931b
MD
1073 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1074 crit_enter();
1075 if (td->td_flags & TDF_RUNQ) {
1076 _lwkt_dequeue(td);
1077 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1078 _lwkt_enqueue(td);
1079 } else {
1080 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1081 }
1082 crit_exit();
1083}
1084
8ec60c3f
MD
1085/*
1086 * Determine if there is a runnable thread at a higher priority then
1087 * the current thread. lwkt_setpri() does not check this automatically.
1088 * Return 1 if there is, 0 if there isn't.
1089 *
1090 * Example: if bit 31 of runqmask is set and the current thread is priority
1091 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1092 *
1093 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1094 * up comparing against 0xffffffff, a comparison that will always be false.
1095 */
1096int
1097lwkt_checkpri_self(void)
1098{
1099 globaldata_t gd = mycpu;
1100 thread_t td = gd->gd_curthread;
1101 int nq = td->td_pri & TDPRI_MASK;
1102
1103 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1104 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1105 return(1);
1106 ++nq;
1107 }
1108 return(0);
1109}
1110
5d21b981
MD
1111/*
1112 * Migrate the current thread to the specified cpu. The BGL must be held
1113 * (for the gd_tdallq manipulation XXX). This is accomplished by
1114 * descheduling ourselves from the current cpu, moving our thread to the
1115 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1116 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1117 */
3d28ff59 1118#ifdef SMP
5d21b981 1119static void lwkt_setcpu_remote(void *arg);
3d28ff59 1120#endif
5d21b981
MD
1121
1122void
1123lwkt_setcpu_self(globaldata_t rgd)
1124{
1125#ifdef SMP
1126 thread_t td = curthread;
1127
1128 if (td->td_gd != rgd) {
1129 crit_enter_quick(td);
1130 td->td_flags |= TDF_MIGRATING;
1131 lwkt_deschedule_self(td);
1132 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1133 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
1134 lwkt_send_ipiq(rgd, (ipifunc_t)lwkt_setcpu_remote, td);
1135 lwkt_switch();
1136 /* we are now on the target cpu */
1137 crit_exit_quick(td);
1138 }
1139#endif
1140}
1141
1142/*
1143 * Remote IPI for cpu migration (called while in a critical section so we
1144 * do not have to enter another one). The thread has already been moved to
1145 * our cpu's allq, but we must wait for the thread to be completely switched
1146 * out on the originating cpu before we schedule it on ours or the stack
1147 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1148 * change to main memory.
1149 *
1150 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1151 * against wakeups. It is best if this interface is used only when there
1152 * are no pending events that might try to schedule the thread.
1153 */
3d28ff59 1154#ifdef SMP
5d21b981
MD
1155static void
1156lwkt_setcpu_remote(void *arg)
1157{
1158 thread_t td = arg;
1159 globaldata_t gd = mycpu;
1160
1161 while (td->td_flags & TDF_RUNNING)
35238fa5 1162 cpu_lfence();
5d21b981 1163 td->td_gd = gd;
35238fa5 1164 cpu_sfence();
5d21b981
MD
1165 td->td_flags &= ~TDF_MIGRATING;
1166 _lwkt_enqueue(td);
1167}
3d28ff59 1168#endif
5d21b981 1169
4b5f931b
MD
1170struct proc *
1171lwkt_preempted_proc(void)
1172{
73e4f7b9 1173 thread_t td = curthread;
4b5f931b
MD
1174 while (td->td_preempted)
1175 td = td->td_preempted;
1176 return(td->td_proc);
1177}
1178
f1d1c3fa 1179/*
41a01a4d
MD
1180 * Block on the specified wait queue until signaled. A generation number
1181 * must be supplied to interlock the wait queue. The function will
1182 * return immediately if the generation number does not match the wait
1183 * structure's generation number.
f1d1c3fa
MD
1184 */
1185void
ae8050a4 1186lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
1187{
1188 thread_t td = curthread;
41a01a4d 1189 lwkt_tokref ilock;
f1d1c3fa 1190
41a01a4d
MD
1191 lwkt_gettoken(&ilock, &w->wa_token);
1192 crit_enter();
ae8050a4 1193 if (w->wa_gen == *gen) {
f1d1c3fa
MD
1194 _lwkt_dequeue(td);
1195 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1196 ++w->wa_count;
1197 td->td_wait = w;
ae8050a4 1198 td->td_wmesg = wmesg;
41a01a4d 1199 again:
f1d1c3fa 1200 lwkt_switch();
ece04fd0
MD
1201 if (td->td_wmesg != NULL) {
1202 _lwkt_dequeue(td);
1203 goto again;
1204 }
8ad65e08 1205 }
41a01a4d 1206 crit_exit();
ae8050a4 1207 *gen = w->wa_gen;
41a01a4d 1208 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1209}
1210
1211/*
1212 * Signal a wait queue. We gain ownership of the wait queue in order to
1213 * signal it. Once a thread is removed from the wait queue we have to
1214 * deal with the cpu owning the thread.
1215 *
1216 * Note: alternatively we could message the target cpu owning the wait
1217 * queue. YYY implement as sysctl.
1218 */
1219void
ece04fd0 1220lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1221{
1222 thread_t td;
41a01a4d 1223 lwkt_tokref ilock;
f1d1c3fa 1224
41a01a4d 1225 lwkt_gettoken(&ilock, &w->wa_token);
f1d1c3fa 1226 ++w->wa_gen;
41a01a4d 1227 crit_enter();
ece04fd0
MD
1228 if (count < 0)
1229 count = w->wa_count;
f1d1c3fa
MD
1230 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1231 --count;
1232 --w->wa_count;
1233 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1234 td->td_wait = NULL;
ae8050a4 1235 td->td_wmesg = NULL;
a72187e9 1236 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1237 _lwkt_enqueue(td);
1238 } else {
2db3b277 1239 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 1240 }
f1d1c3fa 1241 }
41a01a4d
MD
1242 crit_exit();
1243 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1244}
1245
99df837e
MD
1246/*
1247 * Create a kernel process/thread/whatever. It shares it's address space
1248 * with proc0 - ie: kernel only.
1249 *
365fa13f
MD
1250 * NOTE! By default new threads are created with the MP lock held. A
1251 * thread which does not require the MP lock should release it by calling
1252 * rel_mplock() at the start of the new thread.
99df837e
MD
1253 */
1254int
1255lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1256 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1257 const char *fmt, ...)
99df837e 1258{
73e4f7b9 1259 thread_t td;
e2565a42 1260 __va_list ap;
99df837e 1261
f470d0c8 1262 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu);
a2a5ad0d
MD
1263 if (tdp)
1264 *tdp = td;
709799ea 1265 cpu_set_thread_handler(td, lwkt_exit, func, arg);
ef0fdad1 1266 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1267#ifdef SMP
1268 td->td_mpcount = 1;
1269#endif
99df837e
MD
1270
1271 /*
1272 * Set up arg0 for 'ps' etc
1273 */
e2565a42 1274 __va_start(ap, fmt);
99df837e 1275 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1276 __va_end(ap);
99df837e
MD
1277
1278 /*
1279 * Schedule the thread to run
1280 */
ef0fdad1
MD
1281 if ((td->td_flags & TDF_STOPREQ) == 0)
1282 lwkt_schedule(td);
1283 else
1284 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1285 return 0;
1286}
1287
2d93b37a 1288/*
2d93b37a
MD
1289 * kthread_* is specific to the kernel and is not needed by userland.
1290 */
1291#ifdef _KERNEL
1292
99df837e
MD
1293/*
1294 * Destroy an LWKT thread. Warning! This function is not called when
1295 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1296 * uses a different reaping mechanism.
1297 */
1298void
1299lwkt_exit(void)
1300{
1301 thread_t td = curthread;
8826f33a 1302 globaldata_t gd;
99df837e
MD
1303
1304 if (td->td_flags & TDF_VERBOSE)
1305 printf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1306 caps_exit(td);
37af14fe
MD
1307 crit_enter_quick(td);
1308 lwkt_deschedule_self(td);
8826f33a
MD
1309 gd = mycpu;
1310 KKASSERT(gd == td->td_gd);
1311 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1312 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1313 ++gd->gd_tdfreecount;
1314 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1315 }
99df837e
MD
1316 cpu_thread_exit();
1317}
1318
2d93b37a
MD
1319#endif /* _KERNEL */
1320
1321void
1322crit_panic(void)
1323{
1324 thread_t td = curthread;
1325 int lpri = td->td_pri;
1326
1327 td->td_pri = 0;
1328 panic("td_pri is/would-go negative! %p %d", td, lpri);
1329}
1330