Implement a kernel strdup() function (API synch with FreeBSD).
[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 *
8ec60c3f 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.66 2004/07/24 20:21:35 dillon Exp $
75cdbe6c
MD
35 */
36
37/*
38 * Each cpu in a system has its own self-contained light weight kernel
39 * thread scheduler, which means that generally speaking we only need
40 * to use a critical section to avoid problems. Foreign thread
41 * scheduling is queued via (async) IPIs.
8ad65e08
MD
42 */
43
05220613
MD
44#ifdef _KERNEL
45
8ad65e08
MD
46#include <sys/param.h>
47#include <sys/systm.h>
48#include <sys/kernel.h>
49#include <sys/proc.h>
50#include <sys/rtprio.h>
51#include <sys/queue.h>
f1d1c3fa 52#include <sys/thread2.h>
7d0bac62 53#include <sys/sysctl.h>
99df837e 54#include <sys/kthread.h>
f1d1c3fa 55#include <machine/cpu.h>
99df837e 56#include <sys/lock.h>
f6bf3af1 57#include <sys/caps.h>
f1d1c3fa 58
7d0bac62
MD
59#include <vm/vm.h>
60#include <vm/vm_param.h>
61#include <vm/vm_kern.h>
62#include <vm/vm_object.h>
63#include <vm/vm_page.h>
64#include <vm/vm_map.h>
65#include <vm/vm_pager.h>
66#include <vm/vm_extern.h>
67#include <vm/vm_zone.h>
68
99df837e 69#include <machine/stdarg.h>
57c254db 70#include <machine/ipl.h>
96728c05 71#include <machine/smp.h>
99df837e 72
05220613
MD
73#define THREAD_STACK (UPAGES * PAGE_SIZE)
74
75#else
76
77#include <sys/stdint.h>
fb04f4fd 78#include <libcaps/thread.h>
05220613
MD
79#include <sys/thread.h>
80#include <sys/msgport.h>
81#include <sys/errno.h>
fb04f4fd 82#include <libcaps/globaldata.h>
7e8303ad 83#include <machine/cpufunc.h>
05220613
MD
84#include <sys/thread2.h>
85#include <sys/msgport2.h>
709799ea 86#include <stdio.h>
05220613 87#include <stdlib.h>
709799ea 88#include <string.h>
709799ea 89#include <machine/lock.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;
05220613
MD
101
102#ifdef _KERNEL
103
104SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
0f7a3396
MD
105#ifdef INVARIANTS
106SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
107#endif
4b5f931b 108SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 109SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 110SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 111SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
7d0bac62 112
05220613
MD
113#endif
114
4b5f931b
MD
115/*
116 * These helper procedures handle the runq, they can only be called from
117 * within a critical section.
75cdbe6c
MD
118 *
119 * WARNING! Prior to SMP being brought up it is possible to enqueue and
120 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
121 * instead of 'mycpu' when referencing the globaldata structure. Once
122 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 123 */
f1d1c3fa
MD
124static __inline
125void
126_lwkt_dequeue(thread_t td)
127{
128 if (td->td_flags & TDF_RUNQ) {
4b5f931b 129 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 130 struct globaldata *gd = td->td_gd;
4b5f931b 131
f1d1c3fa 132 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
133 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
134 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
135 }
136}
137
138static __inline
139void
140_lwkt_enqueue(thread_t td)
141{
5d21b981 142 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING)) == 0) {
4b5f931b 143 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 144 struct globaldata *gd = td->td_gd;
4b5f931b 145
f1d1c3fa 146 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
147 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
148 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
149 }
150}
8ad65e08 151
37af14fe
MD
152/*
153 * Schedule a thread to run. As the current thread we can always safely
154 * schedule ourselves, and a shortcut procedure is provided for that
155 * function.
156 *
157 * (non-blocking, self contained on a per cpu basis)
158 */
159void
160lwkt_schedule_self(thread_t td)
161{
162 crit_enter_quick(td);
163 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
164 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
165 _lwkt_enqueue(td);
166#ifdef _KERNEL
167 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
168 panic("SCHED SELF PANIC");
169#endif
170 crit_exit_quick(td);
171}
172
173/*
174 * Deschedule a thread.
175 *
176 * (non-blocking, self contained on a per cpu basis)
177 */
178void
179lwkt_deschedule_self(thread_t td)
180{
181 crit_enter_quick(td);
182 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
183 _lwkt_dequeue(td);
184 crit_exit_quick(td);
185}
186
2d93b37a
MD
187#ifdef _KERNEL
188
8ad65e08
MD
189/*
190 * LWKTs operate on a per-cpu basis
191 *
73e4f7b9 192 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
193 */
194void
195lwkt_gdinit(struct globaldata *gd)
196{
4b5f931b
MD
197 int i;
198
199 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
200 TAILQ_INIT(&gd->gd_tdrunq[i]);
201 gd->gd_runqmask = 0;
73e4f7b9 202 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
203}
204
2d93b37a
MD
205#endif /* _KERNEL */
206
7d0bac62
MD
207/*
208 * Initialize a thread wait structure prior to first use.
209 *
210 * NOTE! called from low level boot code, we cannot do anything fancy!
211 */
212void
41a01a4d 213lwkt_wait_init(lwkt_wait_t w)
7d0bac62 214{
41a01a4d 215 lwkt_token_init(&w->wa_token);
7d0bac62 216 TAILQ_INIT(&w->wa_waitq);
41a01a4d
MD
217 w->wa_gen = 0;
218 w->wa_count = 0;
7d0bac62
MD
219}
220
221/*
222 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
223 * or LWKT start address before it can be scheduled. If the target cpu is
224 * -1 the thread will be created on the current cpu.
0cfcada1
MD
225 *
226 * If you intend to create a thread without a process context this function
227 * does everything except load the startup and switcher function.
7d0bac62
MD
228 */
229thread_t
75cdbe6c 230lwkt_alloc_thread(struct thread *td, int cpu)
7d0bac62 231{
99df837e 232 void *stack;
ef0fdad1 233 int flags = 0;
37af14fe 234 globaldata_t gd = mycpu;
7d0bac62 235
ef0fdad1 236 if (td == NULL) {
37af14fe
MD
237 crit_enter_gd(gd);
238 if (gd->gd_tdfreecount > 0) {
239 --gd->gd_tdfreecount;
240 td = TAILQ_FIRST(&gd->gd_tdfreeq);
d9eea1a5 241 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1 242 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
37af14fe
MD
243 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq);
244 crit_exit_gd(gd);
ef0fdad1
MD
245 stack = td->td_kstack;
246 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
247 } else {
37af14fe 248 crit_exit_gd(gd);
05220613 249#ifdef _KERNEL
ef0fdad1 250 td = zalloc(thread_zone);
05220613
MD
251#else
252 td = malloc(sizeof(struct thread));
253#endif
ef0fdad1
MD
254 td->td_kstack = NULL;
255 flags |= TDF_ALLOCATED_THREAD;
256 }
257 }
258 if ((stack = td->td_kstack) == NULL) {
05220613
MD
259#ifdef _KERNEL
260 stack = (void *)kmem_alloc(kernel_map, THREAD_STACK);
261#else
fb04f4fd 262 stack = libcaps_alloc_stack(THREAD_STACK);
05220613 263#endif
ef0fdad1 264 flags |= TDF_ALLOCATED_STACK;
99df837e 265 }
75cdbe6c
MD
266 if (cpu < 0)
267 lwkt_init_thread(td, stack, flags, mycpu);
268 else
269 lwkt_init_thread(td, stack, flags, globaldata_find(cpu));
99df837e 270 return(td);
7d0bac62
MD
271}
272
709799ea
MD
273#ifdef _KERNEL
274
7d0bac62
MD
275/*
276 * Initialize a preexisting thread structure. This function is used by
277 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
278 *
f8c3996b
MD
279 * All threads start out in a critical section at a priority of
280 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
281 * appropriate. This function may send an IPI message when the
282 * requested cpu is not the current cpu and consequently gd_tdallq may
283 * not be initialized synchronously from the point of view of the originating
284 * cpu.
285 *
286 * NOTE! we have to be careful in regards to creating threads for other cpus
287 * if SMP has not yet been activated.
7d0bac62 288 */
41a01a4d
MD
289#ifdef SMP
290
75cdbe6c
MD
291static void
292lwkt_init_thread_remote(void *arg)
293{
294 thread_t td = arg;
295
296 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
297}
298
41a01a4d
MD
299#endif
300
7d0bac62 301void
26a0694b 302lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
7d0bac62 303{
37af14fe
MD
304 globaldata_t mygd = mycpu;
305
99df837e
MD
306 bzero(td, sizeof(struct thread));
307 td->td_kstack = stack;
308 td->td_flags |= flags;
26a0694b 309 td->td_gd = gd;
f8c3996b 310 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
c95cd171 311 lwkt_initport(&td->td_msgport, td);
99df837e 312 pmap_init_thread(td);
0f7a3396 313#ifdef SMP
5d21b981
MD
314 /*
315 * Normally initializing a thread for a remote cpu requires sending an
316 * IPI. However, the idlethread is setup before the other cpus are
317 * activated so we have to treat it as a special case. XXX manipulation
318 * of gd_tdallq requires the BGL.
319 */
320 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 321 crit_enter_gd(mygd);
75cdbe6c 322 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 323 crit_exit_gd(mygd);
75cdbe6c 324 } else {
2db3b277 325 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 326 }
0f7a3396 327#else
37af14fe 328 crit_enter_gd(mygd);
0f7a3396 329 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 330 crit_exit_gd(mygd);
0f7a3396 331#endif
73e4f7b9
MD
332}
333
2d93b37a
MD
334#endif /* _KERNEL */
335
73e4f7b9
MD
336void
337lwkt_set_comm(thread_t td, const char *ctl, ...)
338{
e2565a42 339 __va_list va;
73e4f7b9 340
e2565a42 341 __va_start(va, ctl);
73e4f7b9 342 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 343 __va_end(va);
7d0bac62
MD
344}
345
99df837e 346void
73e4f7b9 347lwkt_hold(thread_t td)
99df837e 348{
73e4f7b9
MD
349 ++td->td_refs;
350}
351
352void
353lwkt_rele(thread_t td)
354{
355 KKASSERT(td->td_refs > 0);
356 --td->td_refs;
357}
358
c95cd171
MD
359#ifdef _KERNEL
360
73e4f7b9
MD
361void
362lwkt_wait_free(thread_t td)
363{
364 while (td->td_refs)
377d4740 365 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
366}
367
c95cd171
MD
368#endif
369
73e4f7b9
MD
370void
371lwkt_free_thread(thread_t td)
372{
373 struct globaldata *gd = mycpu;
374
d9eea1a5 375 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
376 ("lwkt_free_thread: did not exit! %p", td));
377
37af14fe 378 crit_enter_gd(gd);
73e4f7b9
MD
379 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
380 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
381 (td->td_flags & TDF_ALLOCATED_THREAD)
382 ) {
73e4f7b9
MD
383 ++gd->gd_tdfreecount;
384 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
37af14fe 385 crit_exit_gd(gd);
99df837e 386 } else {
37af14fe 387 crit_exit_gd(gd);
99df837e 388 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613
MD
389#ifdef _KERNEL
390 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK);
391#else
fb04f4fd 392 libcaps_free_stack(td->td_kstack, THREAD_STACK);
05220613 393#endif
73e4f7b9 394 /* gd invalid */
99df837e
MD
395 td->td_kstack = NULL;
396 }
05220613
MD
397 if (td->td_flags & TDF_ALLOCATED_THREAD) {
398#ifdef _KERNEL
99df837e 399 zfree(thread_zone, td);
05220613
MD
400#else
401 free(td);
402#endif
403 }
99df837e
MD
404 }
405}
406
407
8ad65e08
MD
408/*
409 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
410 * switch to the idlethread. Switching must occur within a critical
411 * section to avoid races with the scheduling queue.
412 *
413 * We always have full control over our cpu's run queue. Other cpus
414 * that wish to manipulate our queue must use the cpu_*msg() calls to
415 * talk to our cpu, so a critical section is all that is needed and
416 * the result is very, very fast thread switching.
417 *
96728c05
MD
418 * The LWKT scheduler uses a fixed priority model and round-robins at
419 * each priority level. User process scheduling is a totally
420 * different beast and LWKT priorities should not be confused with
421 * user process priorities.
f1d1c3fa 422 *
96728c05
MD
423 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
424 * cleans it up. Note that the td_switch() function cannot do anything that
425 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
426 * the target thread (not the current thread). It's nice to have a scheduler
427 * that does not need the MP lock to work because it allows us to do some
428 * really cool high-performance MP lock optimizations.
8ad65e08 429 */
96728c05 430
8ad65e08
MD
431void
432lwkt_switch(void)
433{
37af14fe
MD
434 globaldata_t gd = mycpu;
435 thread_t td = gd->gd_curthread;
8ad65e08 436 thread_t ntd;
8a8d5d85
MD
437#ifdef SMP
438 int mpheld;
439#endif
8ad65e08 440
46a3f46d
MD
441 /*
442 * Switching from within a 'fast' (non thread switched) interrupt is
443 * illegal.
444 */
37af14fe 445 if (gd->gd_intr_nesting_level && panicstr == NULL) {
fc92d4aa 446 panic("lwkt_switch: cannot switch from within a fast interrupt, yet");
96728c05 447 }
ef0fdad1 448
cb973d15
MD
449 /*
450 * Passive release (used to transition from user to kernel mode
451 * when we block or switch rather then when we enter the kernel).
452 * This function is NOT called if we are switching into a preemption
453 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
454 * our current process designation (if we have one) and become a true
455 * LWKT thread, and may also hand the current process designation to
456 * another process and schedule thread.
cb973d15
MD
457 */
458 if (td->td_release)
459 td->td_release(td);
460
37af14fe 461 crit_enter_gd(gd);
4b5f931b 462 ++switch_count;
8a8d5d85
MD
463
464#ifdef SMP
465 /*
466 * td_mpcount cannot be used to determine if we currently hold the
467 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
468 * to get the lock, and switch out if it can't. Our ownership of
469 * the actual lock will remain stable while we are in a critical section
470 * (but, of course, another cpu may own or release the lock so the
471 * actual value of mp_lock is not stable).
8a8d5d85
MD
472 */
473 mpheld = MP_LOCK_HELD();
0f7a3396
MD
474#ifdef INVARIANTS
475 if (td->td_cscount) {
476 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
477 td);
478 if (panic_on_cscount)
479 panic("switching while mastering cpusync");
480 }
481#endif
8a8d5d85 482#endif
99df837e
MD
483 if ((ntd = td->td_preempted) != NULL) {
484 /*
485 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
486 * thread. This occurs transparently, whether the preempted thread
487 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
488 * itself).
489 *
490 * We have to setup the MP lock for the original thread after backing
491 * out the adjustment that was made to curthread when the original
492 * was preempted.
99df837e 493 */
26a0694b 494 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 495#ifdef SMP
96728c05 496 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 497 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
498 td, ntd, td->td_mpcount, ntd->td_mpcount);
499 }
8a8d5d85
MD
500 if (ntd->td_mpcount) {
501 td->td_mpcount -= ntd->td_mpcount;
502 KKASSERT(td->td_mpcount >= 0);
503 }
504#endif
26a0694b 505 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
506
507 /*
508 * XXX. The interrupt may have woken a thread up, we need to properly
509 * set the reschedule flag if the originally interrupted thread is at
510 * a lower priority.
511 */
512 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
513 need_lwkt_resched();
8a8d5d85 514 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 515 } else {
4b5f931b
MD
516 /*
517 * Priority queue / round-robin at each priority. Note that user
518 * processes run at a fixed, low priority and the user process
519 * scheduler deals with interactions between user processes
520 * by scheduling and descheduling them from the LWKT queue as
521 * necessary.
8a8d5d85
MD
522 *
523 * We have to adjust the MP lock for the target thread. If we
524 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
525 * thread that does not need the MP lock. If we cannot, we spin
526 * instead of HLT.
527 *
528 * A similar issue exists for the tokens held by the target thread.
529 * If we cannot obtain ownership of the tokens we cannot immediately
530 * schedule the thread.
531 */
532
533 /*
534 * We are switching threads. If there are any pending requests for
535 * tokens we can satisfy all of them here.
4b5f931b 536 */
41a01a4d
MD
537#ifdef SMP
538 if (gd->gd_tokreqbase)
539 lwkt_drain_token_requests();
540#endif
541
8ec60c3f
MD
542 /*
543 * If an LWKT reschedule was requested, well that is what we are
544 * doing now so clear it.
545 */
546 clear_lwkt_resched();
4b5f931b
MD
547again:
548 if (gd->gd_runqmask) {
549 int nq = bsrl(gd->gd_runqmask);
550 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
551 gd->gd_runqmask &= ~(1 << nq);
552 goto again;
553 }
8a8d5d85 554#ifdef SMP
41a01a4d
MD
555 /*
556 * If the target needs the MP lock and we couldn't get it,
557 * or if the target is holding tokens and we could not
558 * gain ownership of the tokens, continue looking for a
559 * thread to schedule and spin instead of HLT if we can't.
560 */
561 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
562 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
563 ) {
8a8d5d85
MD
564 u_int32_t rqmask = gd->gd_runqmask;
565 while (rqmask) {
566 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
41a01a4d
MD
567 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock())
568 continue;
569 mpheld = MP_LOCK_HELD();
570 if (ntd->td_toks && !lwkt_chktokens(ntd))
571 continue;
572 break;
8a8d5d85
MD
573 }
574 if (ntd)
575 break;
576 rqmask &= ~(1 << nq);
577 nq = bsrl(rqmask);
578 }
579 if (ntd == NULL) {
a2a5ad0d
MD
580 ntd = &gd->gd_idlethread;
581 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
582 } else {
583 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
584 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
585 }
586 } else {
587 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
588 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
589 }
590#else
4b5f931b
MD
591 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
592 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 593#endif
4b5f931b 594 } else {
3c23a41a 595 /*
60f945af
MD
596 * We have nothing to run but only let the idle loop halt
597 * the cpu if there are no pending interrupts.
3c23a41a 598 */
a2a5ad0d 599 ntd = &gd->gd_idlethread;
60f945af 600 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 601 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 602 }
f1d1c3fa 603 }
26a0694b
MD
604 KASSERT(ntd->td_pri >= TDPRI_CRIT,
605 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
606
607 /*
608 * Do the actual switch. If the new target does not need the MP lock
609 * and we are holding it, release the MP lock. If the new target requires
610 * the MP lock we have already acquired it for the target.
611 */
612#ifdef SMP
613 if (ntd->td_mpcount == 0 ) {
614 if (MP_LOCK_HELD())
615 cpu_rel_mplock();
616 } else {
617 ASSERT_MP_LOCK_HELD();
618 }
619#endif
37af14fe 620 if (td != ntd)
f1d1c3fa 621 td->td_switch(ntd);
37af14fe
MD
622 /* NOTE: current cpu may have changed after switch */
623 crit_exit_quick(td);
8ad65e08
MD
624}
625
b68b7282 626/*
96728c05
MD
627 * Request that the target thread preempt the current thread. Preemption
628 * only works under a specific set of conditions:
b68b7282 629 *
96728c05
MD
630 * - We are not preempting ourselves
631 * - The target thread is owned by the current cpu
632 * - We are not currently being preempted
633 * - The target is not currently being preempted
634 * - We are able to satisfy the target's MP lock requirements (if any).
635 *
636 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
637 * this is called via lwkt_schedule() through the td_preemptable callback.
638 * critpri is the managed critical priority that we should ignore in order
639 * to determine whether preemption is possible (aka usually just the crit
640 * priority of lwkt_schedule() itself).
b68b7282 641 *
26a0694b
MD
642 * XXX at the moment we run the target thread in a critical section during
643 * the preemption in order to prevent the target from taking interrupts
644 * that *WE* can't. Preemption is strictly limited to interrupt threads
645 * and interrupt-like threads, outside of a critical section, and the
646 * preempted source thread will be resumed the instant the target blocks
647 * whether or not the source is scheduled (i.e. preemption is supposed to
648 * be as transparent as possible).
4b5f931b 649 *
8a8d5d85
MD
650 * The target thread inherits our MP count (added to its own) for the
651 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
652 * MP lock during the preemption. Therefore, any preempting targets must be
653 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
654 * out of sync with the physical mp_lock, but we do not have to preserve
655 * the original ownership of the lock if it was out of synch (that is, we
656 * can leave it synchronized on return).
b68b7282
MD
657 */
658void
96728c05 659lwkt_preempt(thread_t ntd, int critpri)
b68b7282 660{
46a3f46d 661 struct globaldata *gd = mycpu;
0a3f9b47 662 thread_t td;
8a8d5d85
MD
663#ifdef SMP
664 int mpheld;
57c254db 665 int savecnt;
8a8d5d85 666#endif
b68b7282 667
26a0694b 668 /*
96728c05
MD
669 * The caller has put us in a critical section. We can only preempt
670 * if the caller of the caller was not in a critical section (basically
0a3f9b47 671 * a local interrupt), as determined by the 'critpri' parameter.
96728c05
MD
672 *
673 * YYY The target thread must be in a critical section (else it must
674 * inherit our critical section? I dunno yet).
41a01a4d
MD
675 *
676 * Any tokens held by the target may not be held by thread(s) being
677 * preempted. We take the easy way out and do not preempt if
678 * the target is holding tokens.
0a3f9b47
MD
679 *
680 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
681 */
682 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 683
0a3f9b47 684 td = gd->gd_curthread;
0a3f9b47 685 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
686 ++preempt_miss;
687 return;
688 }
96728c05
MD
689 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
690 ++preempt_miss;
8ec60c3f 691 need_lwkt_resched();
96728c05
MD
692 return;
693 }
694#ifdef SMP
46a3f46d 695 if (ntd->td_gd != gd) {
96728c05 696 ++preempt_miss;
8ec60c3f 697 need_lwkt_resched();
96728c05
MD
698 return;
699 }
700#endif
41a01a4d
MD
701 /*
702 * Take the easy way out and do not preempt if the target is holding
703 * one or more tokens. We could test whether the thread(s) being
704 * preempted interlock against the target thread's tokens and whether
705 * we can get all the target thread's tokens, but this situation
706 * should not occur very often so its easier to simply not preempt.
707 */
708 if (ntd->td_toks != NULL) {
709 ++preempt_miss;
8ec60c3f 710 need_lwkt_resched();
41a01a4d
MD
711 return;
712 }
26a0694b
MD
713 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
714 ++preempt_weird;
8ec60c3f 715 need_lwkt_resched();
26a0694b
MD
716 return;
717 }
718 if (ntd->td_preempted) {
4b5f931b 719 ++preempt_hit;
8ec60c3f 720 need_lwkt_resched();
26a0694b 721 return;
b68b7282 722 }
8a8d5d85 723#ifdef SMP
a2a5ad0d
MD
724 /*
725 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
726 * to or from the MP lock. In this case td_mpcount will be pre-disposed
727 * (non-zero) but not actually synchronized with the actual state of the
728 * lock. We can use it to imply an MP lock requirement for the
729 * preemption but we cannot use it to test whether we hold the MP lock
730 * or not.
a2a5ad0d 731 */
96728c05 732 savecnt = td->td_mpcount;
71ef2f5c 733 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
734 ntd->td_mpcount += td->td_mpcount;
735 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
736 ntd->td_mpcount -= td->td_mpcount;
737 ++preempt_miss;
8ec60c3f 738 need_lwkt_resched();
8a8d5d85
MD
739 return;
740 }
741#endif
26a0694b 742
8ec60c3f
MD
743 /*
744 * Since we are able to preempt the current thread, there is no need to
745 * call need_lwkt_resched().
746 */
26a0694b
MD
747 ++preempt_hit;
748 ntd->td_preempted = td;
749 td->td_flags |= TDF_PREEMPT_LOCK;
750 td->td_switch(ntd);
751 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
752#ifdef SMP
753 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
754 mpheld = MP_LOCK_HELD();
755 if (mpheld && td->td_mpcount == 0)
96728c05 756 cpu_rel_mplock();
71ef2f5c 757 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
758 panic("lwkt_preempt(): MP lock was not held through");
759#endif
26a0694b
MD
760 ntd->td_preempted = NULL;
761 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
762}
763
f1d1c3fa
MD
764/*
765 * Yield our thread while higher priority threads are pending. This is
766 * typically called when we leave a critical section but it can be safely
767 * called while we are in a critical section.
768 *
769 * This function will not generally yield to equal priority threads but it
770 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 771 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
772 * lwkt_yield_quick().
773 *
235957ed 774 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
775 * came along but was blocked and made pending.
776 *
f1d1c3fa
MD
777 * (self contained on a per cpu basis)
778 */
779void
780lwkt_yield_quick(void)
781{
7966cb69
MD
782 globaldata_t gd = mycpu;
783 thread_t td = gd->gd_curthread;
ef0fdad1 784
a2a5ad0d 785 /*
235957ed 786 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
787 * it with a non-zero cpl then we might not wind up calling splz after
788 * a task switch when the critical section is exited even though the
46a3f46d 789 * new task could accept the interrupt.
a2a5ad0d
MD
790 *
791 * XXX from crit_exit() only called after last crit section is released.
792 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
793 *
794 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
795 * except for this special case, we MUST call splz() here to handle any
796 * pending ints, particularly after we switch, or we might accidently
797 * halt the cpu with interrupts pending.
a2a5ad0d 798 */
46a3f46d 799 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 800 splz();
f1d1c3fa
MD
801
802 /*
803 * YYY enabling will cause wakeup() to task-switch, which really
804 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
805 * preemption and MP without actually doing preemption or MP, because a
806 * lot of code assumes that wakeup() does not block.
f1d1c3fa 807 */
46a3f46d
MD
808 if (untimely_switch && td->td_nest_count == 0 &&
809 gd->gd_intr_nesting_level == 0
810 ) {
37af14fe 811 crit_enter_quick(td);
f1d1c3fa
MD
812 /*
813 * YYY temporary hacks until we disassociate the userland scheduler
814 * from the LWKT scheduler.
815 */
816 if (td->td_flags & TDF_RUNQ) {
817 lwkt_switch(); /* will not reenter yield function */
818 } else {
37af14fe 819 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 820 lwkt_switch(); /* will not reenter yield function */
37af14fe 821 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 822 }
7966cb69 823 crit_exit_noyield(td);
f1d1c3fa 824 }
f1d1c3fa
MD
825}
826
8ad65e08 827/*
f1d1c3fa 828 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 829 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
830 * the crit_exit() call in lwkt_switch().
831 *
832 * (self contained on a per cpu basis)
8ad65e08
MD
833 */
834void
f1d1c3fa 835lwkt_yield(void)
8ad65e08 836{
37af14fe 837 lwkt_schedule_self(curthread);
f1d1c3fa
MD
838 lwkt_switch();
839}
840
8ad65e08 841/*
f1d1c3fa
MD
842 * Generic schedule. Possibly schedule threads belonging to other cpus and
843 * deal with threads that might be blocked on a wait queue.
844 *
0a3f9b47
MD
845 * We have a little helper inline function which does additional work after
846 * the thread has been enqueued, including dealing with preemption and
847 * setting need_lwkt_resched() (which prevents the kernel from returning
848 * to userland until it has processed higher priority threads).
8ad65e08 849 */
0a3f9b47
MD
850static __inline
851void
8ec60c3f 852_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
0a3f9b47
MD
853{
854 if (ntd->td_preemptable) {
855 ntd->td_preemptable(ntd, cpri); /* YYY +token */
8ec60c3f
MD
856 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
857 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
858 ) {
859 need_lwkt_resched();
0a3f9b47
MD
860 }
861}
862
8ad65e08
MD
863void
864lwkt_schedule(thread_t td)
865{
37af14fe
MD
866 globaldata_t mygd = mycpu;
867
96728c05 868#ifdef INVARIANTS
41a01a4d 869 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
26a0694b
MD
870 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
871 && td->td_proc->p_stat == SSLEEP
872 ) {
873 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
874 curthread,
875 curthread->td_proc ? curthread->td_proc->p_pid : -1,
876 curthread->td_proc ? curthread->td_proc->p_stat : -1,
877 td,
878 td->td_proc ? curthread->td_proc->p_pid : -1,
879 td->td_proc ? curthread->td_proc->p_stat : -1
880 );
881 panic("SCHED PANIC");
882 }
96728c05 883#endif
37af14fe
MD
884 crit_enter_gd(mygd);
885 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
886 _lwkt_enqueue(td);
887 } else {
888 lwkt_wait_t w;
889
890 /*
891 * If the thread is on a wait list we have to send our scheduling
892 * request to the owner of the wait structure. Otherwise we send
893 * the scheduling request to the cpu owning the thread. Races
894 * are ok, the target will forward the message as necessary (the
895 * message may chase the thread around before it finally gets
896 * acted upon).
897 *
898 * (remember, wait structures use stable storage)
0a3f9b47
MD
899 *
900 * NOTE: tokens no longer enter a critical section, so we only need
901 * to account for the crit_enter() above when calling
902 * _lwkt_schedule_post().
f1d1c3fa
MD
903 */
904 if ((w = td->td_wait) != NULL) {
41a01a4d
MD
905 lwkt_tokref wref;
906
907 if (lwkt_trytoken(&wref, &w->wa_token)) {
f1d1c3fa
MD
908 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
909 --w->wa_count;
910 td->td_wait = NULL;
0f7a3396 911#ifdef SMP
8ec60c3f 912 if (td->td_gd == mygd) {
f1d1c3fa 913 _lwkt_enqueue(td);
8ec60c3f 914 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 915 } else {
2db3b277 916 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 917 }
0f7a3396
MD
918#else
919 _lwkt_enqueue(td);
8ec60c3f 920 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 921#endif
41a01a4d 922 lwkt_reltoken(&wref);
f1d1c3fa 923 } else {
96728c05 924 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
925 }
926 } else {
927 /*
928 * If the wait structure is NULL and we own the thread, there
929 * is no race (since we are in a critical section). If we
930 * do not own the thread there might be a race but the
931 * target cpu will deal with it.
932 */
0f7a3396 933#ifdef SMP
37af14fe 934 if (td->td_gd == mygd) {
f1d1c3fa 935 _lwkt_enqueue(td);
8ec60c3f 936 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 937 } else {
2db3b277 938 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 939 }
0f7a3396
MD
940#else
941 _lwkt_enqueue(td);
8ec60c3f 942 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 943#endif
f1d1c3fa 944 }
8ad65e08 945 }
37af14fe 946 crit_exit_gd(mygd);
8ad65e08
MD
947}
948
d9eea1a5
MD
949/*
950 * Managed acquisition. This code assumes that the MP lock is held for
951 * the tdallq operation and that the thread has been descheduled from its
952 * original cpu. We also have to wait for the thread to be entirely switched
953 * out on its original cpu (this is usually fast enough that we never loop)
954 * since the LWKT system does not have to hold the MP lock while switching
955 * and the target may have released it before switching.
956 */
a2a5ad0d
MD
957void
958lwkt_acquire(thread_t td)
959{
37af14fe
MD
960 globaldata_t gd;
961 globaldata_t mygd;
a2a5ad0d
MD
962
963 gd = td->td_gd;
37af14fe 964 mygd = mycpu;
a2a5ad0d 965 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5 966 while (td->td_flags & TDF_RUNNING) /* XXX spin */
5d21b981 967 cpu_mb1();
37af14fe
MD
968 if (gd != mygd) {
969 crit_enter_gd(mygd);
a2a5ad0d 970 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
37af14fe
MD
971 td->td_gd = mygd;
972 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
973 crit_exit_gd(mygd);
a2a5ad0d
MD
974 }
975}
976
f1d1c3fa
MD
977/*
978 * Generic deschedule. Descheduling threads other then your own should be
979 * done only in carefully controlled circumstances. Descheduling is
980 * asynchronous.
981 *
982 * This function may block if the cpu has run out of messages.
8ad65e08
MD
983 */
984void
985lwkt_deschedule(thread_t td)
986{
f1d1c3fa
MD
987 crit_enter();
988 if (td == curthread) {
989 _lwkt_dequeue(td);
990 } else {
a72187e9 991 if (td->td_gd == mycpu) {
f1d1c3fa
MD
992 _lwkt_dequeue(td);
993 } else {
2db3b277 994 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
995 }
996 }
997 crit_exit();
998}
999
4b5f931b
MD
1000/*
1001 * Set the target thread's priority. This routine does not automatically
1002 * switch to a higher priority thread, LWKT threads are not designed for
1003 * continuous priority changes. Yield if you want to switch.
1004 *
1005 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1006 * of the td_pri field. The specified priority may also indicate zero or
1007 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1008 *
1009 * Note that we requeue the thread whether it winds up on a different runq
1010 * or not. uio_yield() depends on this and the routine is not normally
1011 * called with the same priority otherwise.
4b5f931b
MD
1012 */
1013void
1014lwkt_setpri(thread_t td, int pri)
1015{
26a0694b 1016 KKASSERT(pri >= 0);
a72187e9 1017 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1018 crit_enter();
1019 if (td->td_flags & TDF_RUNQ) {
1020 _lwkt_dequeue(td);
1021 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1022 _lwkt_enqueue(td);
1023 } else {
1024 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1025 }
1026 crit_exit();
1027}
1028
1029void
1030lwkt_setpri_self(int pri)
1031{
1032 thread_t td = curthread;
1033
4b5f931b
MD
1034 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1035 crit_enter();
1036 if (td->td_flags & TDF_RUNQ) {
1037 _lwkt_dequeue(td);
1038 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1039 _lwkt_enqueue(td);
1040 } else {
1041 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1042 }
1043 crit_exit();
1044}
1045
8ec60c3f
MD
1046/*
1047 * Determine if there is a runnable thread at a higher priority then
1048 * the current thread. lwkt_setpri() does not check this automatically.
1049 * Return 1 if there is, 0 if there isn't.
1050 *
1051 * Example: if bit 31 of runqmask is set and the current thread is priority
1052 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1053 *
1054 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1055 * up comparing against 0xffffffff, a comparison that will always be false.
1056 */
1057int
1058lwkt_checkpri_self(void)
1059{
1060 globaldata_t gd = mycpu;
1061 thread_t td = gd->gd_curthread;
1062 int nq = td->td_pri & TDPRI_MASK;
1063
1064 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1065 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1066 return(1);
1067 ++nq;
1068 }
1069 return(0);
1070}
1071
5d21b981
MD
1072/*
1073 * Migrate the current thread to the specified cpu. The BGL must be held
1074 * (for the gd_tdallq manipulation XXX). This is accomplished by
1075 * descheduling ourselves from the current cpu, moving our thread to the
1076 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1077 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1078 */
3d28ff59 1079#ifdef SMP
5d21b981 1080static void lwkt_setcpu_remote(void *arg);
3d28ff59 1081#endif
5d21b981
MD
1082
1083void
1084lwkt_setcpu_self(globaldata_t rgd)
1085{
1086#ifdef SMP
1087 thread_t td = curthread;
1088
1089 if (td->td_gd != rgd) {
1090 crit_enter_quick(td);
1091 td->td_flags |= TDF_MIGRATING;
1092 lwkt_deschedule_self(td);
1093 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1094 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
1095 lwkt_send_ipiq(rgd, (ipifunc_t)lwkt_setcpu_remote, td);
1096 lwkt_switch();
1097 /* we are now on the target cpu */
1098 crit_exit_quick(td);
1099 }
1100#endif
1101}
1102
1103/*
1104 * Remote IPI for cpu migration (called while in a critical section so we
1105 * do not have to enter another one). The thread has already been moved to
1106 * our cpu's allq, but we must wait for the thread to be completely switched
1107 * out on the originating cpu before we schedule it on ours or the stack
1108 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1109 * change to main memory.
1110 *
1111 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1112 * against wakeups. It is best if this interface is used only when there
1113 * are no pending events that might try to schedule the thread.
1114 */
3d28ff59 1115#ifdef SMP
5d21b981
MD
1116static void
1117lwkt_setcpu_remote(void *arg)
1118{
1119 thread_t td = arg;
1120 globaldata_t gd = mycpu;
1121
1122 while (td->td_flags & TDF_RUNNING)
1123 cpu_mb1();
1124 td->td_gd = gd;
1125 cpu_mb2();
1126 td->td_flags &= ~TDF_MIGRATING;
1127 _lwkt_enqueue(td);
1128}
3d28ff59 1129#endif
5d21b981 1130
4b5f931b
MD
1131struct proc *
1132lwkt_preempted_proc(void)
1133{
73e4f7b9 1134 thread_t td = curthread;
4b5f931b
MD
1135 while (td->td_preempted)
1136 td = td->td_preempted;
1137 return(td->td_proc);
1138}
1139
f1d1c3fa 1140/*
41a01a4d
MD
1141 * Block on the specified wait queue until signaled. A generation number
1142 * must be supplied to interlock the wait queue. The function will
1143 * return immediately if the generation number does not match the wait
1144 * structure's generation number.
f1d1c3fa
MD
1145 */
1146void
ae8050a4 1147lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
1148{
1149 thread_t td = curthread;
41a01a4d 1150 lwkt_tokref ilock;
f1d1c3fa 1151
41a01a4d
MD
1152 lwkt_gettoken(&ilock, &w->wa_token);
1153 crit_enter();
ae8050a4 1154 if (w->wa_gen == *gen) {
f1d1c3fa
MD
1155 _lwkt_dequeue(td);
1156 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1157 ++w->wa_count;
1158 td->td_wait = w;
ae8050a4 1159 td->td_wmesg = wmesg;
41a01a4d 1160 again:
f1d1c3fa 1161 lwkt_switch();
ece04fd0
MD
1162 if (td->td_wmesg != NULL) {
1163 _lwkt_dequeue(td);
1164 goto again;
1165 }
8ad65e08 1166 }
41a01a4d 1167 crit_exit();
ae8050a4 1168 *gen = w->wa_gen;
41a01a4d 1169 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1170}
1171
1172/*
1173 * Signal a wait queue. We gain ownership of the wait queue in order to
1174 * signal it. Once a thread is removed from the wait queue we have to
1175 * deal with the cpu owning the thread.
1176 *
1177 * Note: alternatively we could message the target cpu owning the wait
1178 * queue. YYY implement as sysctl.
1179 */
1180void
ece04fd0 1181lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1182{
1183 thread_t td;
41a01a4d 1184 lwkt_tokref ilock;
f1d1c3fa 1185
41a01a4d 1186 lwkt_gettoken(&ilock, &w->wa_token);
f1d1c3fa 1187 ++w->wa_gen;
41a01a4d 1188 crit_enter();
ece04fd0
MD
1189 if (count < 0)
1190 count = w->wa_count;
f1d1c3fa
MD
1191 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1192 --count;
1193 --w->wa_count;
1194 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1195 td->td_wait = NULL;
ae8050a4 1196 td->td_wmesg = NULL;
a72187e9 1197 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1198 _lwkt_enqueue(td);
1199 } else {
2db3b277 1200 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 1201 }
f1d1c3fa 1202 }
41a01a4d
MD
1203 crit_exit();
1204 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1205}
1206
99df837e
MD
1207/*
1208 * Create a kernel process/thread/whatever. It shares it's address space
1209 * with proc0 - ie: kernel only.
1210 *
365fa13f
MD
1211 * NOTE! By default new threads are created with the MP lock held. A
1212 * thread which does not require the MP lock should release it by calling
1213 * rel_mplock() at the start of the new thread.
99df837e
MD
1214 */
1215int
1216lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1217 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1218 const char *fmt, ...)
99df837e 1219{
73e4f7b9 1220 thread_t td;
e2565a42 1221 __va_list ap;
99df837e 1222
75cdbe6c 1223 td = lwkt_alloc_thread(template, cpu);
a2a5ad0d
MD
1224 if (tdp)
1225 *tdp = td;
709799ea 1226 cpu_set_thread_handler(td, lwkt_exit, func, arg);
ef0fdad1 1227 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1228#ifdef SMP
1229 td->td_mpcount = 1;
1230#endif
99df837e
MD
1231
1232 /*
1233 * Set up arg0 for 'ps' etc
1234 */
e2565a42 1235 __va_start(ap, fmt);
99df837e 1236 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1237 __va_end(ap);
99df837e
MD
1238
1239 /*
1240 * Schedule the thread to run
1241 */
ef0fdad1
MD
1242 if ((td->td_flags & TDF_STOPREQ) == 0)
1243 lwkt_schedule(td);
1244 else
1245 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1246 return 0;
1247}
1248
2d93b37a 1249/*
2d93b37a
MD
1250 * kthread_* is specific to the kernel and is not needed by userland.
1251 */
1252#ifdef _KERNEL
1253
99df837e
MD
1254/*
1255 * Destroy an LWKT thread. Warning! This function is not called when
1256 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1257 * uses a different reaping mechanism.
1258 */
1259void
1260lwkt_exit(void)
1261{
1262 thread_t td = curthread;
8826f33a 1263 globaldata_t gd;
99df837e
MD
1264
1265 if (td->td_flags & TDF_VERBOSE)
1266 printf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1267 caps_exit(td);
37af14fe
MD
1268 crit_enter_quick(td);
1269 lwkt_deschedule_self(td);
8826f33a
MD
1270 gd = mycpu;
1271 KKASSERT(gd == td->td_gd);
1272 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1273 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1274 ++gd->gd_tdfreecount;
1275 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1276 }
99df837e
MD
1277 cpu_thread_exit();
1278}
1279
1280/*
1281 * Create a kernel process/thread/whatever. It shares it's address space
ef0fdad1 1282 * with proc0 - ie: kernel only. 5.x compatible.
365fa13f
MD
1283 *
1284 * NOTE! By default kthreads are created with the MP lock held. A
1285 * thread which does not require the MP lock should release it by calling
1286 * rel_mplock() at the start of the new thread.
99df837e
MD
1287 */
1288int
1289kthread_create(void (*func)(void *), void *arg,
1290 struct thread **tdp, const char *fmt, ...)
1291{
73e4f7b9 1292 thread_t td;
e2565a42 1293 __va_list ap;
99df837e 1294
75cdbe6c 1295 td = lwkt_alloc_thread(NULL, -1);
a2a5ad0d
MD
1296 if (tdp)
1297 *tdp = td;
99df837e
MD
1298 cpu_set_thread_handler(td, kthread_exit, func, arg);
1299 td->td_flags |= TDF_VERBOSE;
8a8d5d85
MD
1300#ifdef SMP
1301 td->td_mpcount = 1;
1302#endif
99df837e
MD
1303
1304 /*
1305 * Set up arg0 for 'ps' etc
1306 */
e2565a42 1307 __va_start(ap, fmt);
99df837e 1308 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1309 __va_end(ap);
99df837e
MD
1310
1311 /*
1312 * Schedule the thread to run
1313 */
1314 lwkt_schedule(td);
1315 return 0;
1316}
1317
1318/*
1319 * Destroy an LWKT thread. Warning! This function is not called when
1320 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1321 * uses a different reaping mechanism.
1322 *
1323 * XXX duplicates lwkt_exit()
1324 */
1325void
1326kthread_exit(void)
1327{
1328 lwkt_exit();
1329}
1330
2d93b37a
MD
1331#endif /* _KERNEL */
1332
1333void
1334crit_panic(void)
1335{
1336 thread_t td = curthread;
1337 int lpri = td->td_pri;
1338
1339 td->td_pri = 0;
1340 panic("td_pri is/would-go negative! %p %d", td, lpri);
1341}
1342