Add a missing crit_exit(). The code path in question only got executed
[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 *
35238fa5 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.74 2005/06/03 23:57:32 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;
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
f470d0c8 230lwkt_alloc_thread(struct thread *td, int stksize, 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 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
246 } else {
37af14fe 247 crit_exit_gd(gd);
05220613 248#ifdef _KERNEL
ef0fdad1 249 td = zalloc(thread_zone);
05220613
MD
250#else
251 td = malloc(sizeof(struct thread));
252#endif
ef0fdad1 253 td->td_kstack = NULL;
f470d0c8 254 td->td_kstack_size = 0;
ef0fdad1
MD
255 flags |= TDF_ALLOCATED_THREAD;
256 }
257 }
f470d0c8
MD
258 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
259 if (flags & TDF_ALLOCATED_STACK) {
a9e5ae82 260#ifdef _KERNEL
f470d0c8 261 kmem_free(kernel_map, (vm_offset_t)stack, td->td_kstack_size);
a9e5ae82
JS
262#else
263 libcaps_free_stack(stack, td->td_kstack_size);
264#endif
f470d0c8
MD
265 stack = NULL;
266 }
267 }
268 if (stack == NULL) {
05220613 269#ifdef _KERNEL
f470d0c8 270 stack = (void *)kmem_alloc(kernel_map, stksize);
05220613 271#else
f470d0c8 272 stack = libcaps_alloc_stack(stksize);
05220613 273#endif
ef0fdad1 274 flags |= TDF_ALLOCATED_STACK;
99df837e 275 }
75cdbe6c 276 if (cpu < 0)
f470d0c8 277 lwkt_init_thread(td, stack, stksize, flags, mycpu);
75cdbe6c 278 else
f470d0c8 279 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 280 return(td);
7d0bac62
MD
281}
282
709799ea
MD
283#ifdef _KERNEL
284
7d0bac62
MD
285/*
286 * Initialize a preexisting thread structure. This function is used by
287 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
288 *
f8c3996b
MD
289 * All threads start out in a critical section at a priority of
290 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
291 * appropriate. This function may send an IPI message when the
292 * requested cpu is not the current cpu and consequently gd_tdallq may
293 * not be initialized synchronously from the point of view of the originating
294 * cpu.
295 *
296 * NOTE! we have to be careful in regards to creating threads for other cpus
297 * if SMP has not yet been activated.
7d0bac62 298 */
41a01a4d
MD
299#ifdef SMP
300
75cdbe6c
MD
301static void
302lwkt_init_thread_remote(void *arg)
303{
304 thread_t td = arg;
305
306 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
307}
308
41a01a4d
MD
309#endif
310
7d0bac62 311void
f470d0c8
MD
312lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
313 struct globaldata *gd)
7d0bac62 314{
37af14fe
MD
315 globaldata_t mygd = mycpu;
316
99df837e
MD
317 bzero(td, sizeof(struct thread));
318 td->td_kstack = stack;
f470d0c8 319 td->td_kstack_size = stksize;
99df837e 320 td->td_flags |= flags;
26a0694b 321 td->td_gd = gd;
f8c3996b 322 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
c95cd171 323 lwkt_initport(&td->td_msgport, td);
99df837e 324 pmap_init_thread(td);
0f7a3396 325#ifdef SMP
5d21b981
MD
326 /*
327 * Normally initializing a thread for a remote cpu requires sending an
328 * IPI. However, the idlethread is setup before the other cpus are
329 * activated so we have to treat it as a special case. XXX manipulation
330 * of gd_tdallq requires the BGL.
331 */
332 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 333 crit_enter_gd(mygd);
75cdbe6c 334 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 335 crit_exit_gd(mygd);
75cdbe6c 336 } else {
2db3b277 337 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 338 }
0f7a3396 339#else
37af14fe 340 crit_enter_gd(mygd);
0f7a3396 341 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 342 crit_exit_gd(mygd);
0f7a3396 343#endif
73e4f7b9
MD
344}
345
2d93b37a
MD
346#endif /* _KERNEL */
347
73e4f7b9
MD
348void
349lwkt_set_comm(thread_t td, const char *ctl, ...)
350{
e2565a42 351 __va_list va;
73e4f7b9 352
e2565a42 353 __va_start(va, ctl);
73e4f7b9 354 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 355 __va_end(va);
7d0bac62
MD
356}
357
99df837e 358void
73e4f7b9 359lwkt_hold(thread_t td)
99df837e 360{
73e4f7b9
MD
361 ++td->td_refs;
362}
363
364void
365lwkt_rele(thread_t td)
366{
367 KKASSERT(td->td_refs > 0);
368 --td->td_refs;
369}
370
c95cd171
MD
371#ifdef _KERNEL
372
73e4f7b9
MD
373void
374lwkt_wait_free(thread_t td)
375{
376 while (td->td_refs)
377d4740 377 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
378}
379
c95cd171
MD
380#endif
381
73e4f7b9
MD
382void
383lwkt_free_thread(thread_t td)
384{
385 struct globaldata *gd = mycpu;
386
d9eea1a5 387 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
388 ("lwkt_free_thread: did not exit! %p", td));
389
37af14fe 390 crit_enter_gd(gd);
73e4f7b9
MD
391 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
392 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
393 (td->td_flags & TDF_ALLOCATED_THREAD)
394 ) {
73e4f7b9
MD
395 ++gd->gd_tdfreecount;
396 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
37af14fe 397 crit_exit_gd(gd);
99df837e 398 } else {
37af14fe 399 crit_exit_gd(gd);
99df837e 400 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613 401#ifdef _KERNEL
f470d0c8 402 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
05220613 403#else
f470d0c8 404 libcaps_free_stack(td->td_kstack, td->td_kstack_size);
05220613 405#endif
73e4f7b9 406 /* gd invalid */
99df837e 407 td->td_kstack = NULL;
f470d0c8 408 td->td_kstack_size = 0;
99df837e 409 }
05220613
MD
410 if (td->td_flags & TDF_ALLOCATED_THREAD) {
411#ifdef _KERNEL
99df837e 412 zfree(thread_zone, td);
05220613
MD
413#else
414 free(td);
415#endif
416 }
99df837e
MD
417 }
418}
419
420
8ad65e08
MD
421/*
422 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
423 * switch to the idlethread. Switching must occur within a critical
424 * section to avoid races with the scheduling queue.
425 *
426 * We always have full control over our cpu's run queue. Other cpus
427 * that wish to manipulate our queue must use the cpu_*msg() calls to
428 * talk to our cpu, so a critical section is all that is needed and
429 * the result is very, very fast thread switching.
430 *
96728c05
MD
431 * The LWKT scheduler uses a fixed priority model and round-robins at
432 * each priority level. User process scheduling is a totally
433 * different beast and LWKT priorities should not be confused with
434 * user process priorities.
f1d1c3fa 435 *
96728c05
MD
436 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
437 * cleans it up. Note that the td_switch() function cannot do anything that
438 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
439 * the target thread (not the current thread). It's nice to have a scheduler
440 * that does not need the MP lock to work because it allows us to do some
441 * really cool high-performance MP lock optimizations.
8ad65e08 442 */
96728c05 443
8ad65e08
MD
444void
445lwkt_switch(void)
446{
37af14fe
MD
447 globaldata_t gd = mycpu;
448 thread_t td = gd->gd_curthread;
8ad65e08 449 thread_t ntd;
8a8d5d85
MD
450#ifdef SMP
451 int mpheld;
452#endif
8ad65e08 453
46a3f46d
MD
454 /*
455 * Switching from within a 'fast' (non thread switched) interrupt is
456 * illegal.
457 */
37af14fe 458 if (gd->gd_intr_nesting_level && panicstr == NULL) {
2cb56a3e 459 panic("lwkt_switch: cannot switch from within a fast interrupt, yet, td %p\n", td);
96728c05 460 }
ef0fdad1 461
cb973d15
MD
462 /*
463 * Passive release (used to transition from user to kernel mode
464 * when we block or switch rather then when we enter the kernel).
465 * This function is NOT called if we are switching into a preemption
466 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
467 * our current process designation (if we have one) and become a true
468 * LWKT thread, and may also hand the current process designation to
469 * another process and schedule thread.
cb973d15
MD
470 */
471 if (td->td_release)
472 td->td_release(td);
473
37af14fe 474 crit_enter_gd(gd);
8a8d5d85
MD
475
476#ifdef SMP
477 /*
478 * td_mpcount cannot be used to determine if we currently hold the
479 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
480 * to get the lock, and switch out if it can't. Our ownership of
481 * the actual lock will remain stable while we are in a critical section
482 * (but, of course, another cpu may own or release the lock so the
483 * actual value of mp_lock is not stable).
8a8d5d85
MD
484 */
485 mpheld = MP_LOCK_HELD();
0f7a3396
MD
486#ifdef INVARIANTS
487 if (td->td_cscount) {
488 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
489 td);
490 if (panic_on_cscount)
491 panic("switching while mastering cpusync");
492 }
493#endif
8a8d5d85 494#endif
99df837e
MD
495 if ((ntd = td->td_preempted) != NULL) {
496 /*
497 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
498 * thread. This occurs transparently, whether the preempted thread
499 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
500 * itself).
501 *
502 * We have to setup the MP lock for the original thread after backing
503 * out the adjustment that was made to curthread when the original
504 * was preempted.
99df837e 505 */
26a0694b 506 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 507#ifdef SMP
96728c05 508 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 509 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
510 td, ntd, td->td_mpcount, ntd->td_mpcount);
511 }
8a8d5d85
MD
512 if (ntd->td_mpcount) {
513 td->td_mpcount -= ntd->td_mpcount;
514 KKASSERT(td->td_mpcount >= 0);
515 }
516#endif
26a0694b 517 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
518
519 /*
520 * XXX. The interrupt may have woken a thread up, we need to properly
521 * set the reschedule flag if the originally interrupted thread is at
522 * a lower priority.
523 */
524 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
525 need_lwkt_resched();
8a8d5d85 526 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 527 } else {
4b5f931b
MD
528 /*
529 * Priority queue / round-robin at each priority. Note that user
530 * processes run at a fixed, low priority and the user process
531 * scheduler deals with interactions between user processes
532 * by scheduling and descheduling them from the LWKT queue as
533 * necessary.
8a8d5d85
MD
534 *
535 * We have to adjust the MP lock for the target thread. If we
536 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
537 * thread that does not need the MP lock. If we cannot, we spin
538 * instead of HLT.
539 *
540 * A similar issue exists for the tokens held by the target thread.
541 * If we cannot obtain ownership of the tokens we cannot immediately
542 * schedule the thread.
543 */
544
545 /*
546 * We are switching threads. If there are any pending requests for
547 * tokens we can satisfy all of them here.
4b5f931b 548 */
41a01a4d
MD
549#ifdef SMP
550 if (gd->gd_tokreqbase)
551 lwkt_drain_token_requests();
552#endif
553
8ec60c3f
MD
554 /*
555 * If an LWKT reschedule was requested, well that is what we are
556 * doing now so clear it.
557 */
558 clear_lwkt_resched();
4b5f931b
MD
559again:
560 if (gd->gd_runqmask) {
561 int nq = bsrl(gd->gd_runqmask);
562 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
563 gd->gd_runqmask &= ~(1 << nq);
564 goto again;
565 }
8a8d5d85 566#ifdef SMP
41a01a4d
MD
567 /*
568 * If the target needs the MP lock and we couldn't get it,
569 * or if the target is holding tokens and we could not
570 * gain ownership of the tokens, continue looking for a
571 * thread to schedule and spin instead of HLT if we can't.
572 */
573 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
574 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
575 ) {
8a8d5d85
MD
576 u_int32_t rqmask = gd->gd_runqmask;
577 while (rqmask) {
578 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
41a01a4d
MD
579 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock())
580 continue;
581 mpheld = MP_LOCK_HELD();
582 if (ntd->td_toks && !lwkt_chktokens(ntd))
583 continue;
584 break;
8a8d5d85
MD
585 }
586 if (ntd)
587 break;
588 rqmask &= ~(1 << nq);
589 nq = bsrl(rqmask);
590 }
591 if (ntd == NULL) {
a2a5ad0d
MD
592 ntd = &gd->gd_idlethread;
593 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
594 } else {
595 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
596 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
597 }
598 } else {
599 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
600 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
601 }
602#else
4b5f931b
MD
603 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
604 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 605#endif
4b5f931b 606 } else {
3c23a41a 607 /*
60f945af
MD
608 * We have nothing to run but only let the idle loop halt
609 * the cpu if there are no pending interrupts.
3c23a41a 610 */
a2a5ad0d 611 ntd = &gd->gd_idlethread;
60f945af 612 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 613 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 614 }
f1d1c3fa 615 }
26a0694b
MD
616 KASSERT(ntd->td_pri >= TDPRI_CRIT,
617 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
618
619 /*
620 * Do the actual switch. If the new target does not need the MP lock
621 * and we are holding it, release the MP lock. If the new target requires
622 * the MP lock we have already acquired it for the target.
623 */
624#ifdef SMP
625 if (ntd->td_mpcount == 0 ) {
626 if (MP_LOCK_HELD())
627 cpu_rel_mplock();
628 } else {
629 ASSERT_MP_LOCK_HELD();
630 }
631#endif
94f6d86e
MD
632 if (td != ntd) {
633 ++switch_count;
f1d1c3fa 634 td->td_switch(ntd);
94f6d86e 635 }
37af14fe
MD
636 /* NOTE: current cpu may have changed after switch */
637 crit_exit_quick(td);
8ad65e08
MD
638}
639
b68b7282 640/*
96728c05
MD
641 * Request that the target thread preempt the current thread. Preemption
642 * only works under a specific set of conditions:
b68b7282 643 *
96728c05
MD
644 * - We are not preempting ourselves
645 * - The target thread is owned by the current cpu
646 * - We are not currently being preempted
647 * - The target is not currently being preempted
648 * - We are able to satisfy the target's MP lock requirements (if any).
649 *
650 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
651 * this is called via lwkt_schedule() through the td_preemptable callback.
652 * critpri is the managed critical priority that we should ignore in order
653 * to determine whether preemption is possible (aka usually just the crit
654 * priority of lwkt_schedule() itself).
b68b7282 655 *
26a0694b
MD
656 * XXX at the moment we run the target thread in a critical section during
657 * the preemption in order to prevent the target from taking interrupts
658 * that *WE* can't. Preemption is strictly limited to interrupt threads
659 * and interrupt-like threads, outside of a critical section, and the
660 * preempted source thread will be resumed the instant the target blocks
661 * whether or not the source is scheduled (i.e. preemption is supposed to
662 * be as transparent as possible).
4b5f931b 663 *
8a8d5d85
MD
664 * The target thread inherits our MP count (added to its own) for the
665 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
666 * MP lock during the preemption. Therefore, any preempting targets must be
667 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
668 * out of sync with the physical mp_lock, but we do not have to preserve
669 * the original ownership of the lock if it was out of synch (that is, we
670 * can leave it synchronized on return).
b68b7282
MD
671 */
672void
96728c05 673lwkt_preempt(thread_t ntd, int critpri)
b68b7282 674{
46a3f46d 675 struct globaldata *gd = mycpu;
0a3f9b47 676 thread_t td;
8a8d5d85
MD
677#ifdef SMP
678 int mpheld;
57c254db 679 int savecnt;
8a8d5d85 680#endif
b68b7282 681
26a0694b 682 /*
96728c05
MD
683 * The caller has put us in a critical section. We can only preempt
684 * if the caller of the caller was not in a critical section (basically
0a3f9b47 685 * a local interrupt), as determined by the 'critpri' parameter.
96728c05
MD
686 *
687 * YYY The target thread must be in a critical section (else it must
688 * inherit our critical section? I dunno yet).
41a01a4d
MD
689 *
690 * Any tokens held by the target may not be held by thread(s) being
691 * preempted. We take the easy way out and do not preempt if
692 * the target is holding tokens.
0a3f9b47
MD
693 *
694 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
695 */
696 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 697
0a3f9b47 698 td = gd->gd_curthread;
0a3f9b47 699 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
700 ++preempt_miss;
701 return;
702 }
96728c05
MD
703 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
704 ++preempt_miss;
8ec60c3f 705 need_lwkt_resched();
96728c05
MD
706 return;
707 }
708#ifdef SMP
46a3f46d 709 if (ntd->td_gd != gd) {
96728c05 710 ++preempt_miss;
8ec60c3f 711 need_lwkt_resched();
96728c05
MD
712 return;
713 }
714#endif
41a01a4d
MD
715 /*
716 * Take the easy way out and do not preempt if the target is holding
717 * one or more tokens. We could test whether the thread(s) being
718 * preempted interlock against the target thread's tokens and whether
719 * we can get all the target thread's tokens, but this situation
720 * should not occur very often so its easier to simply not preempt.
721 */
722 if (ntd->td_toks != NULL) {
723 ++preempt_miss;
8ec60c3f 724 need_lwkt_resched();
41a01a4d
MD
725 return;
726 }
26a0694b
MD
727 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
728 ++preempt_weird;
8ec60c3f 729 need_lwkt_resched();
26a0694b
MD
730 return;
731 }
732 if (ntd->td_preempted) {
4b5f931b 733 ++preempt_hit;
8ec60c3f 734 need_lwkt_resched();
26a0694b 735 return;
b68b7282 736 }
8a8d5d85 737#ifdef SMP
a2a5ad0d
MD
738 /*
739 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
740 * to or from the MP lock. In this case td_mpcount will be pre-disposed
741 * (non-zero) but not actually synchronized with the actual state of the
742 * lock. We can use it to imply an MP lock requirement for the
743 * preemption but we cannot use it to test whether we hold the MP lock
744 * or not.
a2a5ad0d 745 */
96728c05 746 savecnt = td->td_mpcount;
71ef2f5c 747 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
748 ntd->td_mpcount += td->td_mpcount;
749 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
750 ntd->td_mpcount -= td->td_mpcount;
751 ++preempt_miss;
8ec60c3f 752 need_lwkt_resched();
8a8d5d85
MD
753 return;
754 }
755#endif
26a0694b 756
8ec60c3f
MD
757 /*
758 * Since we are able to preempt the current thread, there is no need to
759 * call need_lwkt_resched().
760 */
26a0694b
MD
761 ++preempt_hit;
762 ntd->td_preempted = td;
763 td->td_flags |= TDF_PREEMPT_LOCK;
764 td->td_switch(ntd);
765 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
766#ifdef SMP
767 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
768 mpheld = MP_LOCK_HELD();
769 if (mpheld && td->td_mpcount == 0)
96728c05 770 cpu_rel_mplock();
71ef2f5c 771 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
772 panic("lwkt_preempt(): MP lock was not held through");
773#endif
26a0694b
MD
774 ntd->td_preempted = NULL;
775 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
776}
777
f1d1c3fa
MD
778/*
779 * Yield our thread while higher priority threads are pending. This is
780 * typically called when we leave a critical section but it can be safely
781 * called while we are in a critical section.
782 *
783 * This function will not generally yield to equal priority threads but it
784 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 785 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
786 * lwkt_yield_quick().
787 *
235957ed 788 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
789 * came along but was blocked and made pending.
790 *
f1d1c3fa
MD
791 * (self contained on a per cpu basis)
792 */
793void
794lwkt_yield_quick(void)
795{
7966cb69
MD
796 globaldata_t gd = mycpu;
797 thread_t td = gd->gd_curthread;
ef0fdad1 798
a2a5ad0d 799 /*
235957ed 800 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
801 * it with a non-zero cpl then we might not wind up calling splz after
802 * a task switch when the critical section is exited even though the
46a3f46d 803 * new task could accept the interrupt.
a2a5ad0d
MD
804 *
805 * XXX from crit_exit() only called after last crit section is released.
806 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
807 *
808 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
809 * except for this special case, we MUST call splz() here to handle any
810 * pending ints, particularly after we switch, or we might accidently
811 * halt the cpu with interrupts pending.
a2a5ad0d 812 */
46a3f46d 813 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 814 splz();
f1d1c3fa
MD
815
816 /*
817 * YYY enabling will cause wakeup() to task-switch, which really
818 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
819 * preemption and MP without actually doing preemption or MP, because a
820 * lot of code assumes that wakeup() does not block.
f1d1c3fa 821 */
46a3f46d
MD
822 if (untimely_switch && td->td_nest_count == 0 &&
823 gd->gd_intr_nesting_level == 0
824 ) {
37af14fe 825 crit_enter_quick(td);
f1d1c3fa
MD
826 /*
827 * YYY temporary hacks until we disassociate the userland scheduler
828 * from the LWKT scheduler.
829 */
830 if (td->td_flags & TDF_RUNQ) {
831 lwkt_switch(); /* will not reenter yield function */
832 } else {
37af14fe 833 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 834 lwkt_switch(); /* will not reenter yield function */
37af14fe 835 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 836 }
7966cb69 837 crit_exit_noyield(td);
f1d1c3fa 838 }
f1d1c3fa
MD
839}
840
8ad65e08 841/*
f1d1c3fa 842 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 843 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
844 * the crit_exit() call in lwkt_switch().
845 *
846 * (self contained on a per cpu basis)
8ad65e08
MD
847 */
848void
f1d1c3fa 849lwkt_yield(void)
8ad65e08 850{
37af14fe 851 lwkt_schedule_self(curthread);
f1d1c3fa
MD
852 lwkt_switch();
853}
854
8ad65e08 855/*
f1d1c3fa
MD
856 * Generic schedule. Possibly schedule threads belonging to other cpus and
857 * deal with threads that might be blocked on a wait queue.
858 *
0a3f9b47
MD
859 * We have a little helper inline function which does additional work after
860 * the thread has been enqueued, including dealing with preemption and
861 * setting need_lwkt_resched() (which prevents the kernel from returning
862 * to userland until it has processed higher priority threads).
8ad65e08 863 */
0a3f9b47
MD
864static __inline
865void
8ec60c3f 866_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
0a3f9b47
MD
867{
868 if (ntd->td_preemptable) {
869 ntd->td_preemptable(ntd, cpri); /* YYY +token */
8ec60c3f
MD
870 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
871 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
872 ) {
873 need_lwkt_resched();
0a3f9b47
MD
874 }
875}
876
8ad65e08
MD
877void
878lwkt_schedule(thread_t td)
879{
37af14fe
MD
880 globaldata_t mygd = mycpu;
881
96728c05 882#ifdef INVARIANTS
41a01a4d 883 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
26a0694b
MD
884 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
885 && td->td_proc->p_stat == SSLEEP
886 ) {
887 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
888 curthread,
889 curthread->td_proc ? curthread->td_proc->p_pid : -1,
890 curthread->td_proc ? curthread->td_proc->p_stat : -1,
891 td,
892 td->td_proc ? curthread->td_proc->p_pid : -1,
893 td->td_proc ? curthread->td_proc->p_stat : -1
894 );
895 panic("SCHED PANIC");
896 }
96728c05 897#endif
37af14fe
MD
898 crit_enter_gd(mygd);
899 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
900 _lwkt_enqueue(td);
901 } else {
902 lwkt_wait_t w;
903
904 /*
905 * If the thread is on a wait list we have to send our scheduling
906 * request to the owner of the wait structure. Otherwise we send
907 * the scheduling request to the cpu owning the thread. Races
908 * are ok, the target will forward the message as necessary (the
909 * message may chase the thread around before it finally gets
910 * acted upon).
911 *
912 * (remember, wait structures use stable storage)
0a3f9b47
MD
913 *
914 * NOTE: tokens no longer enter a critical section, so we only need
915 * to account for the crit_enter() above when calling
916 * _lwkt_schedule_post().
f1d1c3fa
MD
917 */
918 if ((w = td->td_wait) != NULL) {
41a01a4d
MD
919 lwkt_tokref wref;
920
921 if (lwkt_trytoken(&wref, &w->wa_token)) {
f1d1c3fa
MD
922 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
923 --w->wa_count;
924 td->td_wait = NULL;
0f7a3396 925#ifdef SMP
8ec60c3f 926 if (td->td_gd == mygd) {
f1d1c3fa 927 _lwkt_enqueue(td);
8ec60c3f 928 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 929 } else {
2db3b277 930 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 931 }
0f7a3396
MD
932#else
933 _lwkt_enqueue(td);
8ec60c3f 934 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 935#endif
41a01a4d 936 lwkt_reltoken(&wref);
f1d1c3fa 937 } else {
96728c05 938 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
939 }
940 } else {
941 /*
942 * If the wait structure is NULL and we own the thread, there
943 * is no race (since we are in a critical section). If we
944 * do not own the thread there might be a race but the
945 * target cpu will deal with it.
946 */
0f7a3396 947#ifdef SMP
37af14fe 948 if (td->td_gd == mygd) {
f1d1c3fa 949 _lwkt_enqueue(td);
8ec60c3f 950 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 951 } else {
2db3b277 952 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 953 }
0f7a3396
MD
954#else
955 _lwkt_enqueue(td);
8ec60c3f 956 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 957#endif
f1d1c3fa 958 }
8ad65e08 959 }
37af14fe 960 crit_exit_gd(mygd);
8ad65e08
MD
961}
962
d9eea1a5
MD
963/*
964 * Managed acquisition. This code assumes that the MP lock is held for
965 * the tdallq operation and that the thread has been descheduled from its
966 * original cpu. We also have to wait for the thread to be entirely switched
967 * out on its original cpu (this is usually fast enough that we never loop)
968 * since the LWKT system does not have to hold the MP lock while switching
969 * and the target may have released it before switching.
970 */
a2a5ad0d
MD
971void
972lwkt_acquire(thread_t td)
973{
37af14fe
MD
974 globaldata_t gd;
975 globaldata_t mygd;
a2a5ad0d
MD
976
977 gd = td->td_gd;
37af14fe 978 mygd = mycpu;
35238fa5 979 cpu_lfence();
a2a5ad0d 980 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5 981 while (td->td_flags & TDF_RUNNING) /* XXX spin */
35238fa5 982 cpu_lfence();
37af14fe
MD
983 if (gd != mygd) {
984 crit_enter_gd(mygd);
a2a5ad0d 985 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
37af14fe
MD
986 td->td_gd = mygd;
987 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
988 crit_exit_gd(mygd);
a2a5ad0d
MD
989 }
990}
991
f1d1c3fa
MD
992/*
993 * Generic deschedule. Descheduling threads other then your own should be
994 * done only in carefully controlled circumstances. Descheduling is
995 * asynchronous.
996 *
997 * This function may block if the cpu has run out of messages.
8ad65e08
MD
998 */
999void
1000lwkt_deschedule(thread_t td)
1001{
f1d1c3fa
MD
1002 crit_enter();
1003 if (td == curthread) {
1004 _lwkt_dequeue(td);
1005 } else {
a72187e9 1006 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1007 _lwkt_dequeue(td);
1008 } else {
2db3b277 1009 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
1010 }
1011 }
1012 crit_exit();
1013}
1014
4b5f931b
MD
1015/*
1016 * Set the target thread's priority. This routine does not automatically
1017 * switch to a higher priority thread, LWKT threads are not designed for
1018 * continuous priority changes. Yield if you want to switch.
1019 *
1020 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1021 * of the td_pri field. The specified priority may also indicate zero or
1022 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1023 *
1024 * Note that we requeue the thread whether it winds up on a different runq
1025 * or not. uio_yield() depends on this and the routine is not normally
1026 * called with the same priority otherwise.
4b5f931b
MD
1027 */
1028void
1029lwkt_setpri(thread_t td, int pri)
1030{
26a0694b 1031 KKASSERT(pri >= 0);
a72187e9 1032 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1033 crit_enter();
1034 if (td->td_flags & TDF_RUNQ) {
1035 _lwkt_dequeue(td);
1036 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1037 _lwkt_enqueue(td);
1038 } else {
1039 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1040 }
1041 crit_exit();
1042}
1043
1044void
1045lwkt_setpri_self(int pri)
1046{
1047 thread_t td = curthread;
1048
4b5f931b
MD
1049 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1050 crit_enter();
1051 if (td->td_flags & TDF_RUNQ) {
1052 _lwkt_dequeue(td);
1053 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1054 _lwkt_enqueue(td);
1055 } else {
1056 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1057 }
1058 crit_exit();
1059}
1060
8ec60c3f
MD
1061/*
1062 * Determine if there is a runnable thread at a higher priority then
1063 * the current thread. lwkt_setpri() does not check this automatically.
1064 * Return 1 if there is, 0 if there isn't.
1065 *
1066 * Example: if bit 31 of runqmask is set and the current thread is priority
1067 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1068 *
1069 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1070 * up comparing against 0xffffffff, a comparison that will always be false.
1071 */
1072int
1073lwkt_checkpri_self(void)
1074{
1075 globaldata_t gd = mycpu;
1076 thread_t td = gd->gd_curthread;
1077 int nq = td->td_pri & TDPRI_MASK;
1078
1079 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1080 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1081 return(1);
1082 ++nq;
1083 }
1084 return(0);
1085}
1086
5d21b981
MD
1087/*
1088 * Migrate the current thread to the specified cpu. The BGL must be held
1089 * (for the gd_tdallq manipulation XXX). This is accomplished by
1090 * descheduling ourselves from the current cpu, moving our thread to the
1091 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1092 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1093 */
3d28ff59 1094#ifdef SMP
5d21b981 1095static void lwkt_setcpu_remote(void *arg);
3d28ff59 1096#endif
5d21b981
MD
1097
1098void
1099lwkt_setcpu_self(globaldata_t rgd)
1100{
1101#ifdef SMP
1102 thread_t td = curthread;
1103
1104 if (td->td_gd != rgd) {
1105 crit_enter_quick(td);
1106 td->td_flags |= TDF_MIGRATING;
1107 lwkt_deschedule_self(td);
1108 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1109 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
1110 lwkt_send_ipiq(rgd, (ipifunc_t)lwkt_setcpu_remote, td);
1111 lwkt_switch();
1112 /* we are now on the target cpu */
1113 crit_exit_quick(td);
1114 }
1115#endif
1116}
1117
1118/*
1119 * Remote IPI for cpu migration (called while in a critical section so we
1120 * do not have to enter another one). The thread has already been moved to
1121 * our cpu's allq, but we must wait for the thread to be completely switched
1122 * out on the originating cpu before we schedule it on ours or the stack
1123 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1124 * change to main memory.
1125 *
1126 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1127 * against wakeups. It is best if this interface is used only when there
1128 * are no pending events that might try to schedule the thread.
1129 */
3d28ff59 1130#ifdef SMP
5d21b981
MD
1131static void
1132lwkt_setcpu_remote(void *arg)
1133{
1134 thread_t td = arg;
1135 globaldata_t gd = mycpu;
1136
1137 while (td->td_flags & TDF_RUNNING)
35238fa5 1138 cpu_lfence();
5d21b981 1139 td->td_gd = gd;
35238fa5 1140 cpu_sfence();
5d21b981
MD
1141 td->td_flags &= ~TDF_MIGRATING;
1142 _lwkt_enqueue(td);
1143}
3d28ff59 1144#endif
5d21b981 1145
4b5f931b
MD
1146struct proc *
1147lwkt_preempted_proc(void)
1148{
73e4f7b9 1149 thread_t td = curthread;
4b5f931b
MD
1150 while (td->td_preempted)
1151 td = td->td_preempted;
1152 return(td->td_proc);
1153}
1154
f1d1c3fa 1155/*
41a01a4d
MD
1156 * Block on the specified wait queue until signaled. A generation number
1157 * must be supplied to interlock the wait queue. The function will
1158 * return immediately if the generation number does not match the wait
1159 * structure's generation number.
f1d1c3fa
MD
1160 */
1161void
ae8050a4 1162lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
1163{
1164 thread_t td = curthread;
41a01a4d 1165 lwkt_tokref ilock;
f1d1c3fa 1166
41a01a4d
MD
1167 lwkt_gettoken(&ilock, &w->wa_token);
1168 crit_enter();
ae8050a4 1169 if (w->wa_gen == *gen) {
f1d1c3fa
MD
1170 _lwkt_dequeue(td);
1171 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1172 ++w->wa_count;
1173 td->td_wait = w;
ae8050a4 1174 td->td_wmesg = wmesg;
41a01a4d 1175 again:
f1d1c3fa 1176 lwkt_switch();
ece04fd0
MD
1177 if (td->td_wmesg != NULL) {
1178 _lwkt_dequeue(td);
1179 goto again;
1180 }
8ad65e08 1181 }
41a01a4d 1182 crit_exit();
ae8050a4 1183 *gen = w->wa_gen;
41a01a4d 1184 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1185}
1186
1187/*
1188 * Signal a wait queue. We gain ownership of the wait queue in order to
1189 * signal it. Once a thread is removed from the wait queue we have to
1190 * deal with the cpu owning the thread.
1191 *
1192 * Note: alternatively we could message the target cpu owning the wait
1193 * queue. YYY implement as sysctl.
1194 */
1195void
ece04fd0 1196lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1197{
1198 thread_t td;
41a01a4d 1199 lwkt_tokref ilock;
f1d1c3fa 1200
41a01a4d 1201 lwkt_gettoken(&ilock, &w->wa_token);
f1d1c3fa 1202 ++w->wa_gen;
41a01a4d 1203 crit_enter();
ece04fd0
MD
1204 if (count < 0)
1205 count = w->wa_count;
f1d1c3fa
MD
1206 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1207 --count;
1208 --w->wa_count;
1209 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1210 td->td_wait = NULL;
ae8050a4 1211 td->td_wmesg = NULL;
a72187e9 1212 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1213 _lwkt_enqueue(td);
1214 } else {
2db3b277 1215 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 1216 }
f1d1c3fa 1217 }
41a01a4d
MD
1218 crit_exit();
1219 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1220}
1221
99df837e
MD
1222/*
1223 * Create a kernel process/thread/whatever. It shares it's address space
1224 * with proc0 - ie: kernel only.
1225 *
365fa13f
MD
1226 * NOTE! By default new threads are created with the MP lock held. A
1227 * thread which does not require the MP lock should release it by calling
1228 * rel_mplock() at the start of the new thread.
99df837e
MD
1229 */
1230int
1231lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1232 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1233 const char *fmt, ...)
99df837e 1234{
73e4f7b9 1235 thread_t td;
e2565a42 1236 __va_list ap;
99df837e 1237
f470d0c8 1238 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu);
a2a5ad0d
MD
1239 if (tdp)
1240 *tdp = td;
709799ea 1241 cpu_set_thread_handler(td, lwkt_exit, func, arg);
ef0fdad1 1242 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1243#ifdef SMP
1244 td->td_mpcount = 1;
1245#endif
99df837e
MD
1246
1247 /*
1248 * Set up arg0 for 'ps' etc
1249 */
e2565a42 1250 __va_start(ap, fmt);
99df837e 1251 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1252 __va_end(ap);
99df837e
MD
1253
1254 /*
1255 * Schedule the thread to run
1256 */
ef0fdad1
MD
1257 if ((td->td_flags & TDF_STOPREQ) == 0)
1258 lwkt_schedule(td);
1259 else
1260 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1261 return 0;
1262}
1263
2d93b37a 1264/*
2d93b37a
MD
1265 * kthread_* is specific to the kernel and is not needed by userland.
1266 */
1267#ifdef _KERNEL
1268
99df837e
MD
1269/*
1270 * Destroy an LWKT thread. Warning! This function is not called when
1271 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1272 * uses a different reaping mechanism.
1273 */
1274void
1275lwkt_exit(void)
1276{
1277 thread_t td = curthread;
8826f33a 1278 globaldata_t gd;
99df837e
MD
1279
1280 if (td->td_flags & TDF_VERBOSE)
1281 printf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1282 caps_exit(td);
37af14fe
MD
1283 crit_enter_quick(td);
1284 lwkt_deschedule_self(td);
8826f33a
MD
1285 gd = mycpu;
1286 KKASSERT(gd == td->td_gd);
1287 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1288 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1289 ++gd->gd_tdfreecount;
1290 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1291 }
99df837e
MD
1292 cpu_thread_exit();
1293}
1294
2d93b37a
MD
1295#endif /* _KERNEL */
1296
1297void
1298crit_panic(void)
1299{
1300 thread_t td = curthread;
1301 int lpri = td->td_pri;
1302
1303 td->td_pri = 0;
1304 panic("td_pri is/would-go negative! %p %d", td, lpri);
1305}
1306