Fix a MP lock race. The MP locking state can change when lwkt_chktokens()
[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 *
a453459d 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.77 2005/07/19 19:08:05 dillon Exp $
75cdbe6c
MD
35 */
36
37/*
38 * Each cpu in a system has its own self-contained light weight kernel
39 * thread scheduler, which means that generally speaking we only need
40 * to use a critical section to avoid problems. Foreign thread
41 * scheduling is queued via (async) IPIs.
8ad65e08
MD
42 */
43
05220613
MD
44#ifdef _KERNEL
45
8ad65e08
MD
46#include <sys/param.h>
47#include <sys/systm.h>
48#include <sys/kernel.h>
49#include <sys/proc.h>
50#include <sys/rtprio.h>
51#include <sys/queue.h>
f1d1c3fa 52#include <sys/thread2.h>
7d0bac62 53#include <sys/sysctl.h>
99df837e 54#include <sys/kthread.h>
f1d1c3fa 55#include <machine/cpu.h>
99df837e 56#include <sys/lock.h>
f6bf3af1 57#include <sys/caps.h>
f1d1c3fa 58
7d0bac62
MD
59#include <vm/vm.h>
60#include <vm/vm_param.h>
61#include <vm/vm_kern.h>
62#include <vm/vm_object.h>
63#include <vm/vm_page.h>
64#include <vm/vm_map.h>
65#include <vm/vm_pager.h>
66#include <vm/vm_extern.h>
67#include <vm/vm_zone.h>
68
99df837e 69#include <machine/stdarg.h>
57c254db 70#include <machine/ipl.h>
96728c05 71#include <machine/smp.h>
99df837e 72
05220613
MD
73#else
74
75#include <sys/stdint.h>
fb04f4fd 76#include <libcaps/thread.h>
05220613
MD
77#include <sys/thread.h>
78#include <sys/msgport.h>
79#include <sys/errno.h>
fb04f4fd 80#include <libcaps/globaldata.h>
7e8303ad 81#include <machine/cpufunc.h>
05220613
MD
82#include <sys/thread2.h>
83#include <sys/msgport2.h>
709799ea 84#include <stdio.h>
05220613 85#include <stdlib.h>
709799ea 86#include <string.h>
709799ea 87#include <machine/lock.h>
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.
a453459d
MD
579 *
580 * NOTE: the mpheld variable invalid after this conditional, it
581 * can change due to both cpu_try_mplock() returning success
582 * AND interactions in lwkt_chktokens() due to the fact that
583 * we are trying to check the mpcount of a thread other then
584 * the current thread. Because of this, if the current thread
585 * is not holding td_mpcount, an IPI indirectly run via
586 * lwkt_chktokens() can obtain and release the MP lock and
587 * cause the core MP lock to be released.
41a01a4d
MD
588 */
589 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
590 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
591 ) {
8a8d5d85 592 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
593
594 mpheld = MP_LOCK_HELD();
595 ntd = NULL;
8a8d5d85
MD
596 while (rqmask) {
597 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 598 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 599 /* spinning due to MP lock being held */
38717797 600#ifdef INVARIANTS
a453459d 601 ++mplock_contention_count;
38717797 602#endif
a453459d 603 /* mplock still not held, 'mpheld' still valid */
41a01a4d 604 continue;
38717797 605 }
a453459d
MD
606
607 /*
608 * mpheld state invalid after chktokens call returns
609 * failure, but the variable is only needed for
610 * the loop.
611 */
38717797 612 if (ntd->td_toks && !lwkt_chktokens(ntd)) {
a453459d 613 /* spinning due to token contention */
38717797 614#ifdef INVARIANTS
a453459d 615 ++token_contention_count;
38717797 616#endif
a453459d 617 mpheld = MP_LOCK_HELD();
41a01a4d 618 continue;
38717797 619 }
41a01a4d 620 break;
8a8d5d85
MD
621 }
622 if (ntd)
623 break;
624 rqmask &= ~(1 << nq);
625 nq = bsrl(rqmask);
626 }
627 if (ntd == NULL) {
a2a5ad0d
MD
628 ntd = &gd->gd_idlethread;
629 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 630 KASSERT(ntd->td_mpcount == 0, ("Idlex thread %p was holding the BGL!", ntd));
8a8d5d85
MD
631 } else {
632 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
633 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
634 }
635 } else {
636 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
637 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
638 }
639#else
4b5f931b
MD
640 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
641 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 642#endif
4b5f931b 643 } else {
3c23a41a 644 /*
60f945af
MD
645 * We have nothing to run but only let the idle loop halt
646 * the cpu if there are no pending interrupts.
3c23a41a 647 */
a2a5ad0d 648 ntd = &gd->gd_idlethread;
60f945af 649 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 650 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d
MD
651#ifdef SMP
652 KASSERT(ntd->td_mpcount == 0, ("Idley thread %p was holding the BGL!", ntd));
653#endif
4b5f931b 654 }
f1d1c3fa 655 }
26a0694b
MD
656 KASSERT(ntd->td_pri >= TDPRI_CRIT,
657 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
658
659 /*
660 * Do the actual switch. If the new target does not need the MP lock
661 * and we are holding it, release the MP lock. If the new target requires
662 * the MP lock we have already acquired it for the target.
663 */
664#ifdef SMP
665 if (ntd->td_mpcount == 0 ) {
666 if (MP_LOCK_HELD())
667 cpu_rel_mplock();
668 } else {
a453459d 669 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
670 }
671#endif
94f6d86e
MD
672 if (td != ntd) {
673 ++switch_count;
f1d1c3fa 674 td->td_switch(ntd);
94f6d86e 675 }
37af14fe
MD
676 /* NOTE: current cpu may have changed after switch */
677 crit_exit_quick(td);
8ad65e08
MD
678}
679
b68b7282 680/*
96728c05
MD
681 * Request that the target thread preempt the current thread. Preemption
682 * only works under a specific set of conditions:
b68b7282 683 *
96728c05
MD
684 * - We are not preempting ourselves
685 * - The target thread is owned by the current cpu
686 * - We are not currently being preempted
687 * - The target is not currently being preempted
688 * - We are able to satisfy the target's MP lock requirements (if any).
689 *
690 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
691 * this is called via lwkt_schedule() through the td_preemptable callback.
692 * critpri is the managed critical priority that we should ignore in order
693 * to determine whether preemption is possible (aka usually just the crit
694 * priority of lwkt_schedule() itself).
b68b7282 695 *
26a0694b
MD
696 * XXX at the moment we run the target thread in a critical section during
697 * the preemption in order to prevent the target from taking interrupts
698 * that *WE* can't. Preemption is strictly limited to interrupt threads
699 * and interrupt-like threads, outside of a critical section, and the
700 * preempted source thread will be resumed the instant the target blocks
701 * whether or not the source is scheduled (i.e. preemption is supposed to
702 * be as transparent as possible).
4b5f931b 703 *
8a8d5d85
MD
704 * The target thread inherits our MP count (added to its own) for the
705 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
706 * MP lock during the preemption. Therefore, any preempting targets must be
707 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
708 * out of sync with the physical mp_lock, but we do not have to preserve
709 * the original ownership of the lock if it was out of synch (that is, we
710 * can leave it synchronized on return).
b68b7282
MD
711 */
712void
96728c05 713lwkt_preempt(thread_t ntd, int critpri)
b68b7282 714{
46a3f46d 715 struct globaldata *gd = mycpu;
0a3f9b47 716 thread_t td;
8a8d5d85
MD
717#ifdef SMP
718 int mpheld;
57c254db 719 int savecnt;
8a8d5d85 720#endif
b68b7282 721
26a0694b 722 /*
96728c05
MD
723 * The caller has put us in a critical section. We can only preempt
724 * if the caller of the caller was not in a critical section (basically
0a3f9b47 725 * a local interrupt), as determined by the 'critpri' parameter.
96728c05
MD
726 *
727 * YYY The target thread must be in a critical section (else it must
728 * inherit our critical section? I dunno yet).
41a01a4d
MD
729 *
730 * Any tokens held by the target may not be held by thread(s) being
731 * preempted. We take the easy way out and do not preempt if
732 * the target is holding tokens.
0a3f9b47
MD
733 *
734 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
735 */
736 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 737
0a3f9b47 738 td = gd->gd_curthread;
0a3f9b47 739 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
740 ++preempt_miss;
741 return;
742 }
96728c05
MD
743 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
744 ++preempt_miss;
8ec60c3f 745 need_lwkt_resched();
96728c05
MD
746 return;
747 }
748#ifdef SMP
46a3f46d 749 if (ntd->td_gd != gd) {
96728c05 750 ++preempt_miss;
8ec60c3f 751 need_lwkt_resched();
96728c05
MD
752 return;
753 }
754#endif
41a01a4d
MD
755 /*
756 * Take the easy way out and do not preempt if the target is holding
757 * one or more tokens. We could test whether the thread(s) being
758 * preempted interlock against the target thread's tokens and whether
759 * we can get all the target thread's tokens, but this situation
760 * should not occur very often so its easier to simply not preempt.
761 */
762 if (ntd->td_toks != NULL) {
763 ++preempt_miss;
8ec60c3f 764 need_lwkt_resched();
41a01a4d
MD
765 return;
766 }
26a0694b
MD
767 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
768 ++preempt_weird;
8ec60c3f 769 need_lwkt_resched();
26a0694b
MD
770 return;
771 }
772 if (ntd->td_preempted) {
4b5f931b 773 ++preempt_hit;
8ec60c3f 774 need_lwkt_resched();
26a0694b 775 return;
b68b7282 776 }
8a8d5d85 777#ifdef SMP
a2a5ad0d
MD
778 /*
779 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
780 * to or from the MP lock. In this case td_mpcount will be pre-disposed
781 * (non-zero) but not actually synchronized with the actual state of the
782 * lock. We can use it to imply an MP lock requirement for the
783 * preemption but we cannot use it to test whether we hold the MP lock
784 * or not.
a2a5ad0d 785 */
96728c05 786 savecnt = td->td_mpcount;
71ef2f5c 787 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
788 ntd->td_mpcount += td->td_mpcount;
789 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
790 ntd->td_mpcount -= td->td_mpcount;
791 ++preempt_miss;
8ec60c3f 792 need_lwkt_resched();
8a8d5d85
MD
793 return;
794 }
795#endif
26a0694b 796
8ec60c3f
MD
797 /*
798 * Since we are able to preempt the current thread, there is no need to
799 * call need_lwkt_resched().
800 */
26a0694b
MD
801 ++preempt_hit;
802 ntd->td_preempted = td;
803 td->td_flags |= TDF_PREEMPT_LOCK;
804 td->td_switch(ntd);
805 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
806#ifdef SMP
807 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
808 mpheld = MP_LOCK_HELD();
809 if (mpheld && td->td_mpcount == 0)
96728c05 810 cpu_rel_mplock();
71ef2f5c 811 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
812 panic("lwkt_preempt(): MP lock was not held through");
813#endif
26a0694b
MD
814 ntd->td_preempted = NULL;
815 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
816}
817
f1d1c3fa
MD
818/*
819 * Yield our thread while higher priority threads are pending. This is
820 * typically called when we leave a critical section but it can be safely
821 * called while we are in a critical section.
822 *
823 * This function will not generally yield to equal priority threads but it
824 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 825 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
826 * lwkt_yield_quick().
827 *
235957ed 828 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
829 * came along but was blocked and made pending.
830 *
f1d1c3fa
MD
831 * (self contained on a per cpu basis)
832 */
833void
834lwkt_yield_quick(void)
835{
7966cb69
MD
836 globaldata_t gd = mycpu;
837 thread_t td = gd->gd_curthread;
ef0fdad1 838
a2a5ad0d 839 /*
235957ed 840 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
841 * it with a non-zero cpl then we might not wind up calling splz after
842 * a task switch when the critical section is exited even though the
46a3f46d 843 * new task could accept the interrupt.
a2a5ad0d
MD
844 *
845 * XXX from crit_exit() only called after last crit section is released.
846 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
847 *
848 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
849 * except for this special case, we MUST call splz() here to handle any
850 * pending ints, particularly after we switch, or we might accidently
851 * halt the cpu with interrupts pending.
a2a5ad0d 852 */
46a3f46d 853 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 854 splz();
f1d1c3fa
MD
855
856 /*
857 * YYY enabling will cause wakeup() to task-switch, which really
858 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
859 * preemption and MP without actually doing preemption or MP, because a
860 * lot of code assumes that wakeup() does not block.
f1d1c3fa 861 */
46a3f46d
MD
862 if (untimely_switch && td->td_nest_count == 0 &&
863 gd->gd_intr_nesting_level == 0
864 ) {
37af14fe 865 crit_enter_quick(td);
f1d1c3fa
MD
866 /*
867 * YYY temporary hacks until we disassociate the userland scheduler
868 * from the LWKT scheduler.
869 */
870 if (td->td_flags & TDF_RUNQ) {
871 lwkt_switch(); /* will not reenter yield function */
872 } else {
37af14fe 873 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 874 lwkt_switch(); /* will not reenter yield function */
37af14fe 875 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 876 }
7966cb69 877 crit_exit_noyield(td);
f1d1c3fa 878 }
f1d1c3fa
MD
879}
880
8ad65e08 881/*
f1d1c3fa 882 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 883 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
884 * the crit_exit() call in lwkt_switch().
885 *
886 * (self contained on a per cpu basis)
8ad65e08
MD
887 */
888void
f1d1c3fa 889lwkt_yield(void)
8ad65e08 890{
37af14fe 891 lwkt_schedule_self(curthread);
f1d1c3fa
MD
892 lwkt_switch();
893}
894
8ad65e08 895/*
f1d1c3fa
MD
896 * Generic schedule. Possibly schedule threads belonging to other cpus and
897 * deal with threads that might be blocked on a wait queue.
898 *
0a3f9b47
MD
899 * We have a little helper inline function which does additional work after
900 * the thread has been enqueued, including dealing with preemption and
901 * setting need_lwkt_resched() (which prevents the kernel from returning
902 * to userland until it has processed higher priority threads).
8ad65e08 903 */
0a3f9b47
MD
904static __inline
905void
8ec60c3f 906_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
0a3f9b47
MD
907{
908 if (ntd->td_preemptable) {
909 ntd->td_preemptable(ntd, cpri); /* YYY +token */
8ec60c3f
MD
910 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
911 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
912 ) {
913 need_lwkt_resched();
0a3f9b47
MD
914 }
915}
916
8ad65e08
MD
917void
918lwkt_schedule(thread_t td)
919{
37af14fe
MD
920 globaldata_t mygd = mycpu;
921
96728c05 922#ifdef INVARIANTS
41a01a4d 923 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
26a0694b
MD
924 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
925 && td->td_proc->p_stat == SSLEEP
926 ) {
927 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
928 curthread,
929 curthread->td_proc ? curthread->td_proc->p_pid : -1,
930 curthread->td_proc ? curthread->td_proc->p_stat : -1,
931 td,
932 td->td_proc ? curthread->td_proc->p_pid : -1,
933 td->td_proc ? curthread->td_proc->p_stat : -1
934 );
935 panic("SCHED PANIC");
936 }
96728c05 937#endif
37af14fe
MD
938 crit_enter_gd(mygd);
939 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
940 _lwkt_enqueue(td);
941 } else {
942 lwkt_wait_t w;
943
944 /*
945 * If the thread is on a wait list we have to send our scheduling
946 * request to the owner of the wait structure. Otherwise we send
947 * the scheduling request to the cpu owning the thread. Races
948 * are ok, the target will forward the message as necessary (the
949 * message may chase the thread around before it finally gets
950 * acted upon).
951 *
952 * (remember, wait structures use stable storage)
0a3f9b47 953 *
0c453950
MD
954 * NOTE: we have to account for the number of critical sections
955 * under our control when calling _lwkt_schedule_post() so it
956 * can figure out whether preemption is allowed.
957 *
958 * NOTE: The wait structure algorithms are a mess and need to be
959 * rewritten.
960 *
961 * NOTE: We cannot safely acquire or release a token, even
962 * non-blocking, because this routine may be called in the context
963 * of a thread already holding the token and thus not provide any
964 * interlock protection. We cannot safely manipulate the td_toks
965 * list for the same reason. Instead we depend on our critical
966 * section if the token is owned by our cpu.
f1d1c3fa
MD
967 */
968 if ((w = td->td_wait) != NULL) {
0c453950 969 if (w->wa_token.t_cpu == mygd) {
f1d1c3fa
MD
970 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
971 --w->wa_count;
972 td->td_wait = NULL;
0f7a3396 973#ifdef SMP
8ec60c3f 974 if (td->td_gd == mygd) {
f1d1c3fa 975 _lwkt_enqueue(td);
8ec60c3f 976 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 977 } else {
2db3b277 978 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 979 }
0f7a3396
MD
980#else
981 _lwkt_enqueue(td);
8ec60c3f 982 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 983#endif
f1d1c3fa 984 } else {
96728c05 985 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
986 }
987 } else {
988 /*
989 * If the wait structure is NULL and we own the thread, there
990 * is no race (since we are in a critical section). If we
991 * do not own the thread there might be a race but the
992 * target cpu will deal with it.
993 */
0f7a3396 994#ifdef SMP
37af14fe 995 if (td->td_gd == mygd) {
f1d1c3fa 996 _lwkt_enqueue(td);
8ec60c3f 997 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 998 } else {
2db3b277 999 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 1000 }
0f7a3396
MD
1001#else
1002 _lwkt_enqueue(td);
8ec60c3f 1003 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 1004#endif
f1d1c3fa 1005 }
8ad65e08 1006 }
37af14fe 1007 crit_exit_gd(mygd);
8ad65e08
MD
1008}
1009
d9eea1a5
MD
1010/*
1011 * Managed acquisition. This code assumes that the MP lock is held for
1012 * the tdallq operation and that the thread has been descheduled from its
1013 * original cpu. We also have to wait for the thread to be entirely switched
1014 * out on its original cpu (this is usually fast enough that we never loop)
1015 * since the LWKT system does not have to hold the MP lock while switching
1016 * and the target may have released it before switching.
1017 */
a2a5ad0d
MD
1018void
1019lwkt_acquire(thread_t td)
1020{
37af14fe
MD
1021 globaldata_t gd;
1022 globaldata_t mygd;
a2a5ad0d
MD
1023
1024 gd = td->td_gd;
37af14fe 1025 mygd = mycpu;
35238fa5 1026 cpu_lfence();
a2a5ad0d 1027 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5 1028 while (td->td_flags & TDF_RUNNING) /* XXX spin */
35238fa5 1029 cpu_lfence();
37af14fe
MD
1030 if (gd != mygd) {
1031 crit_enter_gd(mygd);
a2a5ad0d 1032 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
37af14fe
MD
1033 td->td_gd = mygd;
1034 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
1035 crit_exit_gd(mygd);
a2a5ad0d
MD
1036 }
1037}
1038
f1d1c3fa
MD
1039/*
1040 * Generic deschedule. Descheduling threads other then your own should be
1041 * done only in carefully controlled circumstances. Descheduling is
1042 * asynchronous.
1043 *
1044 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1045 */
1046void
1047lwkt_deschedule(thread_t td)
1048{
f1d1c3fa
MD
1049 crit_enter();
1050 if (td == curthread) {
1051 _lwkt_dequeue(td);
1052 } else {
a72187e9 1053 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1054 _lwkt_dequeue(td);
1055 } else {
2db3b277 1056 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
1057 }
1058 }
1059 crit_exit();
1060}
1061
4b5f931b
MD
1062/*
1063 * Set the target thread's priority. This routine does not automatically
1064 * switch to a higher priority thread, LWKT threads are not designed for
1065 * continuous priority changes. Yield if you want to switch.
1066 *
1067 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1068 * of the td_pri field. The specified priority may also indicate zero or
1069 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1070 *
1071 * Note that we requeue the thread whether it winds up on a different runq
1072 * or not. uio_yield() depends on this and the routine is not normally
1073 * called with the same priority otherwise.
4b5f931b
MD
1074 */
1075void
1076lwkt_setpri(thread_t td, int pri)
1077{
26a0694b 1078 KKASSERT(pri >= 0);
a72187e9 1079 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1080 crit_enter();
1081 if (td->td_flags & TDF_RUNQ) {
1082 _lwkt_dequeue(td);
1083 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1084 _lwkt_enqueue(td);
1085 } else {
1086 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1087 }
1088 crit_exit();
1089}
1090
1091void
1092lwkt_setpri_self(int pri)
1093{
1094 thread_t td = curthread;
1095
4b5f931b
MD
1096 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1097 crit_enter();
1098 if (td->td_flags & TDF_RUNQ) {
1099 _lwkt_dequeue(td);
1100 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1101 _lwkt_enqueue(td);
1102 } else {
1103 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1104 }
1105 crit_exit();
1106}
1107
8ec60c3f
MD
1108/*
1109 * Determine if there is a runnable thread at a higher priority then
1110 * the current thread. lwkt_setpri() does not check this automatically.
1111 * Return 1 if there is, 0 if there isn't.
1112 *
1113 * Example: if bit 31 of runqmask is set and the current thread is priority
1114 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1115 *
1116 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1117 * up comparing against 0xffffffff, a comparison that will always be false.
1118 */
1119int
1120lwkt_checkpri_self(void)
1121{
1122 globaldata_t gd = mycpu;
1123 thread_t td = gd->gd_curthread;
1124 int nq = td->td_pri & TDPRI_MASK;
1125
1126 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1127 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1128 return(1);
1129 ++nq;
1130 }
1131 return(0);
1132}
1133
5d21b981
MD
1134/*
1135 * Migrate the current thread to the specified cpu. The BGL must be held
1136 * (for the gd_tdallq manipulation XXX). This is accomplished by
1137 * descheduling ourselves from the current cpu, moving our thread to the
1138 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1139 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1140 */
3d28ff59 1141#ifdef SMP
5d21b981 1142static void lwkt_setcpu_remote(void *arg);
3d28ff59 1143#endif
5d21b981
MD
1144
1145void
1146lwkt_setcpu_self(globaldata_t rgd)
1147{
1148#ifdef SMP
1149 thread_t td = curthread;
1150
1151 if (td->td_gd != rgd) {
1152 crit_enter_quick(td);
1153 td->td_flags |= TDF_MIGRATING;
1154 lwkt_deschedule_self(td);
1155 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1156 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
1157 lwkt_send_ipiq(rgd, (ipifunc_t)lwkt_setcpu_remote, td);
1158 lwkt_switch();
1159 /* we are now on the target cpu */
1160 crit_exit_quick(td);
1161 }
1162#endif
1163}
1164
1165/*
1166 * Remote IPI for cpu migration (called while in a critical section so we
1167 * do not have to enter another one). The thread has already been moved to
1168 * our cpu's allq, but we must wait for the thread to be completely switched
1169 * out on the originating cpu before we schedule it on ours or the stack
1170 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1171 * change to main memory.
1172 *
1173 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1174 * against wakeups. It is best if this interface is used only when there
1175 * are no pending events that might try to schedule the thread.
1176 */
3d28ff59 1177#ifdef SMP
5d21b981
MD
1178static void
1179lwkt_setcpu_remote(void *arg)
1180{
1181 thread_t td = arg;
1182 globaldata_t gd = mycpu;
1183
1184 while (td->td_flags & TDF_RUNNING)
35238fa5 1185 cpu_lfence();
5d21b981 1186 td->td_gd = gd;
35238fa5 1187 cpu_sfence();
5d21b981
MD
1188 td->td_flags &= ~TDF_MIGRATING;
1189 _lwkt_enqueue(td);
1190}
3d28ff59 1191#endif
5d21b981 1192
4b5f931b
MD
1193struct proc *
1194lwkt_preempted_proc(void)
1195{
73e4f7b9 1196 thread_t td = curthread;
4b5f931b
MD
1197 while (td->td_preempted)
1198 td = td->td_preempted;
1199 return(td->td_proc);
1200}
1201
f1d1c3fa 1202/*
41a01a4d
MD
1203 * Block on the specified wait queue until signaled. A generation number
1204 * must be supplied to interlock the wait queue. The function will
1205 * return immediately if the generation number does not match the wait
1206 * structure's generation number.
f1d1c3fa
MD
1207 */
1208void
ae8050a4 1209lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
1210{
1211 thread_t td = curthread;
41a01a4d 1212 lwkt_tokref ilock;
f1d1c3fa 1213
41a01a4d
MD
1214 lwkt_gettoken(&ilock, &w->wa_token);
1215 crit_enter();
ae8050a4 1216 if (w->wa_gen == *gen) {
f1d1c3fa
MD
1217 _lwkt_dequeue(td);
1218 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1219 ++w->wa_count;
1220 td->td_wait = w;
ae8050a4 1221 td->td_wmesg = wmesg;
41a01a4d 1222 again:
f1d1c3fa 1223 lwkt_switch();
ece04fd0
MD
1224 if (td->td_wmesg != NULL) {
1225 _lwkt_dequeue(td);
1226 goto again;
1227 }
8ad65e08 1228 }
41a01a4d 1229 crit_exit();
ae8050a4 1230 *gen = w->wa_gen;
41a01a4d 1231 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1232}
1233
1234/*
1235 * Signal a wait queue. We gain ownership of the wait queue in order to
1236 * signal it. Once a thread is removed from the wait queue we have to
1237 * deal with the cpu owning the thread.
1238 *
1239 * Note: alternatively we could message the target cpu owning the wait
1240 * queue. YYY implement as sysctl.
1241 */
1242void
ece04fd0 1243lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1244{
1245 thread_t td;
41a01a4d 1246 lwkt_tokref ilock;
f1d1c3fa 1247
41a01a4d 1248 lwkt_gettoken(&ilock, &w->wa_token);
f1d1c3fa 1249 ++w->wa_gen;
41a01a4d 1250 crit_enter();
ece04fd0
MD
1251 if (count < 0)
1252 count = w->wa_count;
f1d1c3fa
MD
1253 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1254 --count;
1255 --w->wa_count;
1256 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1257 td->td_wait = NULL;
ae8050a4 1258 td->td_wmesg = NULL;
a72187e9 1259 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1260 _lwkt_enqueue(td);
1261 } else {
2db3b277 1262 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 1263 }
f1d1c3fa 1264 }
41a01a4d
MD
1265 crit_exit();
1266 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1267}
1268
99df837e
MD
1269/*
1270 * Create a kernel process/thread/whatever. It shares it's address space
1271 * with proc0 - ie: kernel only.
1272 *
365fa13f
MD
1273 * NOTE! By default new threads are created with the MP lock held. A
1274 * thread which does not require the MP lock should release it by calling
1275 * rel_mplock() at the start of the new thread.
99df837e
MD
1276 */
1277int
1278lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1279 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1280 const char *fmt, ...)
99df837e 1281{
73e4f7b9 1282 thread_t td;
e2565a42 1283 __va_list ap;
99df837e 1284
f470d0c8 1285 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu);
a2a5ad0d
MD
1286 if (tdp)
1287 *tdp = td;
709799ea 1288 cpu_set_thread_handler(td, lwkt_exit, func, arg);
ef0fdad1 1289 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1290#ifdef SMP
1291 td->td_mpcount = 1;
1292#endif
99df837e
MD
1293
1294 /*
1295 * Set up arg0 for 'ps' etc
1296 */
e2565a42 1297 __va_start(ap, fmt);
99df837e 1298 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1299 __va_end(ap);
99df837e
MD
1300
1301 /*
1302 * Schedule the thread to run
1303 */
ef0fdad1
MD
1304 if ((td->td_flags & TDF_STOPREQ) == 0)
1305 lwkt_schedule(td);
1306 else
1307 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1308 return 0;
1309}
1310
2d93b37a 1311/*
2d93b37a
MD
1312 * kthread_* is specific to the kernel and is not needed by userland.
1313 */
1314#ifdef _KERNEL
1315
99df837e
MD
1316/*
1317 * Destroy an LWKT thread. Warning! This function is not called when
1318 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1319 * uses a different reaping mechanism.
1320 */
1321void
1322lwkt_exit(void)
1323{
1324 thread_t td = curthread;
8826f33a 1325 globaldata_t gd;
99df837e
MD
1326
1327 if (td->td_flags & TDF_VERBOSE)
1328 printf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1329 caps_exit(td);
37af14fe
MD
1330 crit_enter_quick(td);
1331 lwkt_deschedule_self(td);
8826f33a
MD
1332 gd = mycpu;
1333 KKASSERT(gd == td->td_gd);
1334 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1335 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1336 ++gd->gd_tdfreecount;
1337 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1338 }
99df837e
MD
1339 cpu_thread_exit();
1340}
1341
2d93b37a
MD
1342#endif /* _KERNEL */
1343
1344void
1345crit_panic(void)
1346{
1347 thread_t td = curthread;
1348 int lpri = td->td_pri;
1349
1350 td->td_pri = 0;
1351 panic("td_pri is/would-go negative! %p %d", td, lpri);
1352}
1353