Optimize lwkt_send_ipiq() - the IPI based inter-cpu messaging routine.
[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 *
b950d4b8 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.70 2004/10/13 18:42:34 eirikn 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) {
fc92d4aa 459 panic("lwkt_switch: cannot switch from within a fast interrupt, yet");
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);
4b5f931b 475 ++switch_count;
8a8d5d85
MD
476
477#ifdef SMP
478 /*
479 * td_mpcount cannot be used to determine if we currently hold the
480 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
481 * to get the lock, and switch out if it can't. Our ownership of
482 * the actual lock will remain stable while we are in a critical section
483 * (but, of course, another cpu may own or release the lock so the
484 * actual value of mp_lock is not stable).
8a8d5d85
MD
485 */
486 mpheld = MP_LOCK_HELD();
0f7a3396
MD
487#ifdef INVARIANTS
488 if (td->td_cscount) {
489 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
490 td);
491 if (panic_on_cscount)
492 panic("switching while mastering cpusync");
493 }
494#endif
8a8d5d85 495#endif
99df837e
MD
496 if ((ntd = td->td_preempted) != NULL) {
497 /*
498 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
499 * thread. This occurs transparently, whether the preempted thread
500 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
501 * itself).
502 *
503 * We have to setup the MP lock for the original thread after backing
504 * out the adjustment that was made to curthread when the original
505 * was preempted.
99df837e 506 */
26a0694b 507 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 508#ifdef SMP
96728c05 509 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 510 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
511 td, ntd, td->td_mpcount, ntd->td_mpcount);
512 }
8a8d5d85
MD
513 if (ntd->td_mpcount) {
514 td->td_mpcount -= ntd->td_mpcount;
515 KKASSERT(td->td_mpcount >= 0);
516 }
517#endif
26a0694b 518 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
519
520 /*
521 * XXX. The interrupt may have woken a thread up, we need to properly
522 * set the reschedule flag if the originally interrupted thread is at
523 * a lower priority.
524 */
525 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
526 need_lwkt_resched();
8a8d5d85 527 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 528 } else {
4b5f931b
MD
529 /*
530 * Priority queue / round-robin at each priority. Note that user
531 * processes run at a fixed, low priority and the user process
532 * scheduler deals with interactions between user processes
533 * by scheduling and descheduling them from the LWKT queue as
534 * necessary.
8a8d5d85
MD
535 *
536 * We have to adjust the MP lock for the target thread. If we
537 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
538 * thread that does not need the MP lock. If we cannot, we spin
539 * instead of HLT.
540 *
541 * A similar issue exists for the tokens held by the target thread.
542 * If we cannot obtain ownership of the tokens we cannot immediately
543 * schedule the thread.
544 */
545
546 /*
547 * We are switching threads. If there are any pending requests for
548 * tokens we can satisfy all of them here.
4b5f931b 549 */
41a01a4d
MD
550#ifdef SMP
551 if (gd->gd_tokreqbase)
552 lwkt_drain_token_requests();
553#endif
554
8ec60c3f
MD
555 /*
556 * If an LWKT reschedule was requested, well that is what we are
557 * doing now so clear it.
558 */
559 clear_lwkt_resched();
4b5f931b
MD
560again:
561 if (gd->gd_runqmask) {
562 int nq = bsrl(gd->gd_runqmask);
563 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
564 gd->gd_runqmask &= ~(1 << nq);
565 goto again;
566 }
8a8d5d85 567#ifdef SMP
41a01a4d
MD
568 /*
569 * If the target needs the MP lock and we couldn't get it,
570 * or if the target is holding tokens and we could not
571 * gain ownership of the tokens, continue looking for a
572 * thread to schedule and spin instead of HLT if we can't.
573 */
574 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
575 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
576 ) {
8a8d5d85
MD
577 u_int32_t rqmask = gd->gd_runqmask;
578 while (rqmask) {
579 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
41a01a4d
MD
580 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock())
581 continue;
582 mpheld = MP_LOCK_HELD();
583 if (ntd->td_toks && !lwkt_chktokens(ntd))
584 continue;
585 break;
8a8d5d85
MD
586 }
587 if (ntd)
588 break;
589 rqmask &= ~(1 << nq);
590 nq = bsrl(rqmask);
591 }
592 if (ntd == NULL) {
a2a5ad0d
MD
593 ntd = &gd->gd_idlethread;
594 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
595 } else {
596 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
597 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
598 }
599 } else {
600 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
601 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
602 }
603#else
4b5f931b
MD
604 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
605 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 606#endif
4b5f931b 607 } else {
3c23a41a 608 /*
60f945af
MD
609 * We have nothing to run but only let the idle loop halt
610 * the cpu if there are no pending interrupts.
3c23a41a 611 */
a2a5ad0d 612 ntd = &gd->gd_idlethread;
60f945af 613 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 614 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 615 }
f1d1c3fa 616 }
26a0694b
MD
617 KASSERT(ntd->td_pri >= TDPRI_CRIT,
618 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
619
620 /*
621 * Do the actual switch. If the new target does not need the MP lock
622 * and we are holding it, release the MP lock. If the new target requires
623 * the MP lock we have already acquired it for the target.
624 */
625#ifdef SMP
626 if (ntd->td_mpcount == 0 ) {
627 if (MP_LOCK_HELD())
628 cpu_rel_mplock();
629 } else {
630 ASSERT_MP_LOCK_HELD();
631 }
632#endif
37af14fe 633 if (td != ntd)
f1d1c3fa 634 td->td_switch(ntd);
37af14fe
MD
635 /* NOTE: current cpu may have changed after switch */
636 crit_exit_quick(td);
8ad65e08
MD
637}
638
b68b7282 639/*
96728c05
MD
640 * Request that the target thread preempt the current thread. Preemption
641 * only works under a specific set of conditions:
b68b7282 642 *
96728c05
MD
643 * - We are not preempting ourselves
644 * - The target thread is owned by the current cpu
645 * - We are not currently being preempted
646 * - The target is not currently being preempted
647 * - We are able to satisfy the target's MP lock requirements (if any).
648 *
649 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
650 * this is called via lwkt_schedule() through the td_preemptable callback.
651 * critpri is the managed critical priority that we should ignore in order
652 * to determine whether preemption is possible (aka usually just the crit
653 * priority of lwkt_schedule() itself).
b68b7282 654 *
26a0694b
MD
655 * XXX at the moment we run the target thread in a critical section during
656 * the preemption in order to prevent the target from taking interrupts
657 * that *WE* can't. Preemption is strictly limited to interrupt threads
658 * and interrupt-like threads, outside of a critical section, and the
659 * preempted source thread will be resumed the instant the target blocks
660 * whether or not the source is scheduled (i.e. preemption is supposed to
661 * be as transparent as possible).
4b5f931b 662 *
8a8d5d85
MD
663 * The target thread inherits our MP count (added to its own) for the
664 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
665 * MP lock during the preemption. Therefore, any preempting targets must be
666 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
667 * out of sync with the physical mp_lock, but we do not have to preserve
668 * the original ownership of the lock if it was out of synch (that is, we
669 * can leave it synchronized on return).
b68b7282
MD
670 */
671void
96728c05 672lwkt_preempt(thread_t ntd, int critpri)
b68b7282 673{
46a3f46d 674 struct globaldata *gd = mycpu;
0a3f9b47 675 thread_t td;
8a8d5d85
MD
676#ifdef SMP
677 int mpheld;
57c254db 678 int savecnt;
8a8d5d85 679#endif
b68b7282 680
26a0694b 681 /*
96728c05
MD
682 * The caller has put us in a critical section. We can only preempt
683 * if the caller of the caller was not in a critical section (basically
0a3f9b47 684 * a local interrupt), as determined by the 'critpri' parameter.
96728c05
MD
685 *
686 * YYY The target thread must be in a critical section (else it must
687 * inherit our critical section? I dunno yet).
41a01a4d
MD
688 *
689 * Any tokens held by the target may not be held by thread(s) being
690 * preempted. We take the easy way out and do not preempt if
691 * the target is holding tokens.
0a3f9b47
MD
692 *
693 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
694 */
695 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 696
0a3f9b47 697 td = gd->gd_curthread;
0a3f9b47 698 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
699 ++preempt_miss;
700 return;
701 }
96728c05
MD
702 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
703 ++preempt_miss;
8ec60c3f 704 need_lwkt_resched();
96728c05
MD
705 return;
706 }
707#ifdef SMP
46a3f46d 708 if (ntd->td_gd != gd) {
96728c05 709 ++preempt_miss;
8ec60c3f 710 need_lwkt_resched();
96728c05
MD
711 return;
712 }
713#endif
41a01a4d
MD
714 /*
715 * Take the easy way out and do not preempt if the target is holding
716 * one or more tokens. We could test whether the thread(s) being
717 * preempted interlock against the target thread's tokens and whether
718 * we can get all the target thread's tokens, but this situation
719 * should not occur very often so its easier to simply not preempt.
720 */
721 if (ntd->td_toks != NULL) {
722 ++preempt_miss;
8ec60c3f 723 need_lwkt_resched();
41a01a4d
MD
724 return;
725 }
26a0694b
MD
726 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
727 ++preempt_weird;
8ec60c3f 728 need_lwkt_resched();
26a0694b
MD
729 return;
730 }
731 if (ntd->td_preempted) {
4b5f931b 732 ++preempt_hit;
8ec60c3f 733 need_lwkt_resched();
26a0694b 734 return;
b68b7282 735 }
8a8d5d85 736#ifdef SMP
a2a5ad0d
MD
737 /*
738 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
739 * to or from the MP lock. In this case td_mpcount will be pre-disposed
740 * (non-zero) but not actually synchronized with the actual state of the
741 * lock. We can use it to imply an MP lock requirement for the
742 * preemption but we cannot use it to test whether we hold the MP lock
743 * or not.
a2a5ad0d 744 */
96728c05 745 savecnt = td->td_mpcount;
71ef2f5c 746 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
747 ntd->td_mpcount += td->td_mpcount;
748 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
749 ntd->td_mpcount -= td->td_mpcount;
750 ++preempt_miss;
8ec60c3f 751 need_lwkt_resched();
8a8d5d85
MD
752 return;
753 }
754#endif
26a0694b 755
8ec60c3f
MD
756 /*
757 * Since we are able to preempt the current thread, there is no need to
758 * call need_lwkt_resched().
759 */
26a0694b
MD
760 ++preempt_hit;
761 ntd->td_preempted = td;
762 td->td_flags |= TDF_PREEMPT_LOCK;
763 td->td_switch(ntd);
764 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
765#ifdef SMP
766 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
767 mpheld = MP_LOCK_HELD();
768 if (mpheld && td->td_mpcount == 0)
96728c05 769 cpu_rel_mplock();
71ef2f5c 770 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
771 panic("lwkt_preempt(): MP lock was not held through");
772#endif
26a0694b
MD
773 ntd->td_preempted = NULL;
774 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
775}
776
f1d1c3fa
MD
777/*
778 * Yield our thread while higher priority threads are pending. This is
779 * typically called when we leave a critical section but it can be safely
780 * called while we are in a critical section.
781 *
782 * This function will not generally yield to equal priority threads but it
783 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 784 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
785 * lwkt_yield_quick().
786 *
235957ed 787 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
788 * came along but was blocked and made pending.
789 *
f1d1c3fa
MD
790 * (self contained on a per cpu basis)
791 */
792void
793lwkt_yield_quick(void)
794{
7966cb69
MD
795 globaldata_t gd = mycpu;
796 thread_t td = gd->gd_curthread;
ef0fdad1 797
a2a5ad0d 798 /*
235957ed 799 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
800 * it with a non-zero cpl then we might not wind up calling splz after
801 * a task switch when the critical section is exited even though the
46a3f46d 802 * new task could accept the interrupt.
a2a5ad0d
MD
803 *
804 * XXX from crit_exit() only called after last crit section is released.
805 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
806 *
807 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
808 * except for this special case, we MUST call splz() here to handle any
809 * pending ints, particularly after we switch, or we might accidently
810 * halt the cpu with interrupts pending.
a2a5ad0d 811 */
46a3f46d 812 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 813 splz();
f1d1c3fa
MD
814
815 /*
816 * YYY enabling will cause wakeup() to task-switch, which really
817 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
818 * preemption and MP without actually doing preemption or MP, because a
819 * lot of code assumes that wakeup() does not block.
f1d1c3fa 820 */
46a3f46d
MD
821 if (untimely_switch && td->td_nest_count == 0 &&
822 gd->gd_intr_nesting_level == 0
823 ) {
37af14fe 824 crit_enter_quick(td);
f1d1c3fa
MD
825 /*
826 * YYY temporary hacks until we disassociate the userland scheduler
827 * from the LWKT scheduler.
828 */
829 if (td->td_flags & TDF_RUNQ) {
830 lwkt_switch(); /* will not reenter yield function */
831 } else {
37af14fe 832 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 833 lwkt_switch(); /* will not reenter yield function */
37af14fe 834 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 835 }
7966cb69 836 crit_exit_noyield(td);
f1d1c3fa 837 }
f1d1c3fa
MD
838}
839
8ad65e08 840/*
f1d1c3fa 841 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 842 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
843 * the crit_exit() call in lwkt_switch().
844 *
845 * (self contained on a per cpu basis)
8ad65e08
MD
846 */
847void
f1d1c3fa 848lwkt_yield(void)
8ad65e08 849{
37af14fe 850 lwkt_schedule_self(curthread);
f1d1c3fa
MD
851 lwkt_switch();
852}
853
8ad65e08 854/*
f1d1c3fa
MD
855 * Generic schedule. Possibly schedule threads belonging to other cpus and
856 * deal with threads that might be blocked on a wait queue.
857 *
0a3f9b47
MD
858 * We have a little helper inline function which does additional work after
859 * the thread has been enqueued, including dealing with preemption and
860 * setting need_lwkt_resched() (which prevents the kernel from returning
861 * to userland until it has processed higher priority threads).
8ad65e08 862 */
0a3f9b47
MD
863static __inline
864void
8ec60c3f 865_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
0a3f9b47
MD
866{
867 if (ntd->td_preemptable) {
868 ntd->td_preemptable(ntd, cpri); /* YYY +token */
8ec60c3f
MD
869 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
870 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
871 ) {
872 need_lwkt_resched();
0a3f9b47
MD
873 }
874}
875
8ad65e08
MD
876void
877lwkt_schedule(thread_t td)
878{
37af14fe
MD
879 globaldata_t mygd = mycpu;
880
96728c05 881#ifdef INVARIANTS
41a01a4d 882 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
26a0694b
MD
883 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
884 && td->td_proc->p_stat == SSLEEP
885 ) {
886 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
887 curthread,
888 curthread->td_proc ? curthread->td_proc->p_pid : -1,
889 curthread->td_proc ? curthread->td_proc->p_stat : -1,
890 td,
891 td->td_proc ? curthread->td_proc->p_pid : -1,
892 td->td_proc ? curthread->td_proc->p_stat : -1
893 );
894 panic("SCHED PANIC");
895 }
96728c05 896#endif
37af14fe
MD
897 crit_enter_gd(mygd);
898 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
899 _lwkt_enqueue(td);
900 } else {
901 lwkt_wait_t w;
902
903 /*
904 * If the thread is on a wait list we have to send our scheduling
905 * request to the owner of the wait structure. Otherwise we send
906 * the scheduling request to the cpu owning the thread. Races
907 * are ok, the target will forward the message as necessary (the
908 * message may chase the thread around before it finally gets
909 * acted upon).
910 *
911 * (remember, wait structures use stable storage)
0a3f9b47
MD
912 *
913 * NOTE: tokens no longer enter a critical section, so we only need
914 * to account for the crit_enter() above when calling
915 * _lwkt_schedule_post().
f1d1c3fa
MD
916 */
917 if ((w = td->td_wait) != NULL) {
41a01a4d
MD
918 lwkt_tokref wref;
919
920 if (lwkt_trytoken(&wref, &w->wa_token)) {
f1d1c3fa
MD
921 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
922 --w->wa_count;
923 td->td_wait = NULL;
0f7a3396 924#ifdef SMP
8ec60c3f 925 if (td->td_gd == mygd) {
f1d1c3fa 926 _lwkt_enqueue(td);
8ec60c3f 927 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 928 } else {
2db3b277 929 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 930 }
0f7a3396
MD
931#else
932 _lwkt_enqueue(td);
8ec60c3f 933 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 934#endif
41a01a4d 935 lwkt_reltoken(&wref);
f1d1c3fa 936 } else {
96728c05 937 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
938 }
939 } else {
940 /*
941 * If the wait structure is NULL and we own the thread, there
942 * is no race (since we are in a critical section). If we
943 * do not own the thread there might be a race but the
944 * target cpu will deal with it.
945 */
0f7a3396 946#ifdef SMP
37af14fe 947 if (td->td_gd == mygd) {
f1d1c3fa 948 _lwkt_enqueue(td);
8ec60c3f 949 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 950 } else {
2db3b277 951 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 952 }
0f7a3396
MD
953#else
954 _lwkt_enqueue(td);
8ec60c3f 955 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 956#endif
f1d1c3fa 957 }
8ad65e08 958 }
37af14fe 959 crit_exit_gd(mygd);
8ad65e08
MD
960}
961
d9eea1a5
MD
962/*
963 * Managed acquisition. This code assumes that the MP lock is held for
964 * the tdallq operation and that the thread has been descheduled from its
965 * original cpu. We also have to wait for the thread to be entirely switched
966 * out on its original cpu (this is usually fast enough that we never loop)
967 * since the LWKT system does not have to hold the MP lock while switching
968 * and the target may have released it before switching.
969 */
a2a5ad0d
MD
970void
971lwkt_acquire(thread_t td)
972{
37af14fe
MD
973 globaldata_t gd;
974 globaldata_t mygd;
a2a5ad0d
MD
975
976 gd = td->td_gd;
37af14fe 977 mygd = mycpu;
a2a5ad0d 978 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5 979 while (td->td_flags & TDF_RUNNING) /* XXX spin */
5d21b981 980 cpu_mb1();
37af14fe
MD
981 if (gd != mygd) {
982 crit_enter_gd(mygd);
a2a5ad0d 983 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
37af14fe
MD
984 td->td_gd = mygd;
985 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
986 crit_exit_gd(mygd);
a2a5ad0d
MD
987 }
988}
989
f1d1c3fa
MD
990/*
991 * Generic deschedule. Descheduling threads other then your own should be
992 * done only in carefully controlled circumstances. Descheduling is
993 * asynchronous.
994 *
995 * This function may block if the cpu has run out of messages.
8ad65e08
MD
996 */
997void
998lwkt_deschedule(thread_t td)
999{
f1d1c3fa
MD
1000 crit_enter();
1001 if (td == curthread) {
1002 _lwkt_dequeue(td);
1003 } else {
a72187e9 1004 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1005 _lwkt_dequeue(td);
1006 } else {
2db3b277 1007 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
1008 }
1009 }
1010 crit_exit();
1011}
1012
4b5f931b
MD
1013/*
1014 * Set the target thread's priority. This routine does not automatically
1015 * switch to a higher priority thread, LWKT threads are not designed for
1016 * continuous priority changes. Yield if you want to switch.
1017 *
1018 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1019 * of the td_pri field. The specified priority may also indicate zero or
1020 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1021 *
1022 * Note that we requeue the thread whether it winds up on a different runq
1023 * or not. uio_yield() depends on this and the routine is not normally
1024 * called with the same priority otherwise.
4b5f931b
MD
1025 */
1026void
1027lwkt_setpri(thread_t td, int pri)
1028{
26a0694b 1029 KKASSERT(pri >= 0);
a72187e9 1030 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1031 crit_enter();
1032 if (td->td_flags & TDF_RUNQ) {
1033 _lwkt_dequeue(td);
1034 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1035 _lwkt_enqueue(td);
1036 } else {
1037 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1038 }
1039 crit_exit();
1040}
1041
1042void
1043lwkt_setpri_self(int pri)
1044{
1045 thread_t td = curthread;
1046
4b5f931b
MD
1047 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1048 crit_enter();
1049 if (td->td_flags & TDF_RUNQ) {
1050 _lwkt_dequeue(td);
1051 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1052 _lwkt_enqueue(td);
1053 } else {
1054 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1055 }
1056 crit_exit();
1057}
1058
8ec60c3f
MD
1059/*
1060 * Determine if there is a runnable thread at a higher priority then
1061 * the current thread. lwkt_setpri() does not check this automatically.
1062 * Return 1 if there is, 0 if there isn't.
1063 *
1064 * Example: if bit 31 of runqmask is set and the current thread is priority
1065 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1066 *
1067 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1068 * up comparing against 0xffffffff, a comparison that will always be false.
1069 */
1070int
1071lwkt_checkpri_self(void)
1072{
1073 globaldata_t gd = mycpu;
1074 thread_t td = gd->gd_curthread;
1075 int nq = td->td_pri & TDPRI_MASK;
1076
1077 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1078 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1079 return(1);
1080 ++nq;
1081 }
1082 return(0);
1083}
1084
5d21b981
MD
1085/*
1086 * Migrate the current thread to the specified cpu. The BGL must be held
1087 * (for the gd_tdallq manipulation XXX). This is accomplished by
1088 * descheduling ourselves from the current cpu, moving our thread to the
1089 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1090 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1091 */
3d28ff59 1092#ifdef SMP
5d21b981 1093static void lwkt_setcpu_remote(void *arg);
3d28ff59 1094#endif
5d21b981
MD
1095
1096void
1097lwkt_setcpu_self(globaldata_t rgd)
1098{
1099#ifdef SMP
1100 thread_t td = curthread;
1101
1102 if (td->td_gd != rgd) {
1103 crit_enter_quick(td);
1104 td->td_flags |= TDF_MIGRATING;
1105 lwkt_deschedule_self(td);
1106 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1107 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
1108 lwkt_send_ipiq(rgd, (ipifunc_t)lwkt_setcpu_remote, td);
1109 lwkt_switch();
1110 /* we are now on the target cpu */
1111 crit_exit_quick(td);
1112 }
1113#endif
1114}
1115
1116/*
1117 * Remote IPI for cpu migration (called while in a critical section so we
1118 * do not have to enter another one). The thread has already been moved to
1119 * our cpu's allq, but we must wait for the thread to be completely switched
1120 * out on the originating cpu before we schedule it on ours or the stack
1121 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1122 * change to main memory.
1123 *
1124 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1125 * against wakeups. It is best if this interface is used only when there
1126 * are no pending events that might try to schedule the thread.
1127 */
3d28ff59 1128#ifdef SMP
5d21b981
MD
1129static void
1130lwkt_setcpu_remote(void *arg)
1131{
1132 thread_t td = arg;
1133 globaldata_t gd = mycpu;
1134
1135 while (td->td_flags & TDF_RUNNING)
1136 cpu_mb1();
1137 td->td_gd = gd;
1138 cpu_mb2();
1139 td->td_flags &= ~TDF_MIGRATING;
1140 _lwkt_enqueue(td);
1141}
3d28ff59 1142#endif
5d21b981 1143
4b5f931b
MD
1144struct proc *
1145lwkt_preempted_proc(void)
1146{
73e4f7b9 1147 thread_t td = curthread;
4b5f931b
MD
1148 while (td->td_preempted)
1149 td = td->td_preempted;
1150 return(td->td_proc);
1151}
1152
f1d1c3fa 1153/*
41a01a4d
MD
1154 * Block on the specified wait queue until signaled. A generation number
1155 * must be supplied to interlock the wait queue. The function will
1156 * return immediately if the generation number does not match the wait
1157 * structure's generation number.
f1d1c3fa
MD
1158 */
1159void
ae8050a4 1160lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
1161{
1162 thread_t td = curthread;
41a01a4d 1163 lwkt_tokref ilock;
f1d1c3fa 1164
41a01a4d
MD
1165 lwkt_gettoken(&ilock, &w->wa_token);
1166 crit_enter();
ae8050a4 1167 if (w->wa_gen == *gen) {
f1d1c3fa
MD
1168 _lwkt_dequeue(td);
1169 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1170 ++w->wa_count;
1171 td->td_wait = w;
ae8050a4 1172 td->td_wmesg = wmesg;
41a01a4d 1173 again:
f1d1c3fa 1174 lwkt_switch();
ece04fd0
MD
1175 if (td->td_wmesg != NULL) {
1176 _lwkt_dequeue(td);
1177 goto again;
1178 }
8ad65e08 1179 }
41a01a4d 1180 crit_exit();
ae8050a4 1181 *gen = w->wa_gen;
41a01a4d 1182 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1183}
1184
1185/*
1186 * Signal a wait queue. We gain ownership of the wait queue in order to
1187 * signal it. Once a thread is removed from the wait queue we have to
1188 * deal with the cpu owning the thread.
1189 *
1190 * Note: alternatively we could message the target cpu owning the wait
1191 * queue. YYY implement as sysctl.
1192 */
1193void
ece04fd0 1194lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1195{
1196 thread_t td;
41a01a4d 1197 lwkt_tokref ilock;
f1d1c3fa 1198
41a01a4d 1199 lwkt_gettoken(&ilock, &w->wa_token);
f1d1c3fa 1200 ++w->wa_gen;
41a01a4d 1201 crit_enter();
ece04fd0
MD
1202 if (count < 0)
1203 count = w->wa_count;
f1d1c3fa
MD
1204 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1205 --count;
1206 --w->wa_count;
1207 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1208 td->td_wait = NULL;
ae8050a4 1209 td->td_wmesg = NULL;
a72187e9 1210 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1211 _lwkt_enqueue(td);
1212 } else {
2db3b277 1213 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 1214 }
f1d1c3fa 1215 }
41a01a4d
MD
1216 crit_exit();
1217 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1218}
1219
99df837e
MD
1220/*
1221 * Create a kernel process/thread/whatever. It shares it's address space
1222 * with proc0 - ie: kernel only.
1223 *
365fa13f
MD
1224 * NOTE! By default new threads are created with the MP lock held. A
1225 * thread which does not require the MP lock should release it by calling
1226 * rel_mplock() at the start of the new thread.
99df837e
MD
1227 */
1228int
1229lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1230 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1231 const char *fmt, ...)
99df837e 1232{
73e4f7b9 1233 thread_t td;
e2565a42 1234 __va_list ap;
99df837e 1235
f470d0c8 1236 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu);
a2a5ad0d
MD
1237 if (tdp)
1238 *tdp = td;
709799ea 1239 cpu_set_thread_handler(td, lwkt_exit, func, arg);
ef0fdad1 1240 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1241#ifdef SMP
1242 td->td_mpcount = 1;
1243#endif
99df837e
MD
1244
1245 /*
1246 * Set up arg0 for 'ps' etc
1247 */
e2565a42 1248 __va_start(ap, fmt);
99df837e 1249 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1250 __va_end(ap);
99df837e
MD
1251
1252 /*
1253 * Schedule the thread to run
1254 */
ef0fdad1
MD
1255 if ((td->td_flags & TDF_STOPREQ) == 0)
1256 lwkt_schedule(td);
1257 else
1258 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1259 return 0;
1260}
1261
2d93b37a 1262/*
2d93b37a
MD
1263 * kthread_* is specific to the kernel and is not needed by userland.
1264 */
1265#ifdef _KERNEL
1266
99df837e
MD
1267/*
1268 * Destroy an LWKT thread. Warning! This function is not called when
1269 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1270 * uses a different reaping mechanism.
1271 */
1272void
1273lwkt_exit(void)
1274{
1275 thread_t td = curthread;
8826f33a 1276 globaldata_t gd;
99df837e
MD
1277
1278 if (td->td_flags & TDF_VERBOSE)
1279 printf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1280 caps_exit(td);
37af14fe
MD
1281 crit_enter_quick(td);
1282 lwkt_deschedule_self(td);
8826f33a
MD
1283 gd = mycpu;
1284 KKASSERT(gd == td->td_gd);
1285 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1286 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1287 ++gd->gd_tdfreecount;
1288 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1289 }
99df837e
MD
1290 cpu_thread_exit();
1291}
1292
2d93b37a
MD
1293#endif /* _KERNEL */
1294
1295void
1296crit_panic(void)
1297{
1298 thread_t td = curthread;
1299 int lpri = td->td_pri;
1300
1301 td->td_pri = 0;
1302 panic("td_pri is/would-go negative! %p %d", td, lpri);
1303}
1304