Misc mdoc(7) cleanup:
[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 *
dd55d707 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.95 2006/05/19 18:26:28 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>
7d0bac62 52#include <sys/sysctl.h>
99df837e 53#include <sys/kthread.h>
f1d1c3fa 54#include <machine/cpu.h>
99df837e 55#include <sys/lock.h>
f6bf3af1 56#include <sys/caps.h>
9d265729
MD
57#include <sys/spinlock.h>
58
59#include <sys/thread2.h>
60#include <sys/spinlock2.h>
f1d1c3fa 61
7d0bac62
MD
62#include <vm/vm.h>
63#include <vm/vm_param.h>
64#include <vm/vm_kern.h>
65#include <vm/vm_object.h>
66#include <vm/vm_page.h>
67#include <vm/vm_map.h>
68#include <vm/vm_pager.h>
69#include <vm/vm_extern.h>
70#include <vm/vm_zone.h>
71
99df837e 72#include <machine/stdarg.h>
57c254db 73#include <machine/ipl.h>
96728c05 74#include <machine/smp.h>
99df837e 75
05220613
MD
76#else
77
78#include <sys/stdint.h>
fb04f4fd 79#include <libcaps/thread.h>
05220613
MD
80#include <sys/thread.h>
81#include <sys/msgport.h>
82#include <sys/errno.h>
fb04f4fd 83#include <libcaps/globaldata.h>
7e8303ad 84#include <machine/cpufunc.h>
05220613
MD
85#include <sys/thread2.h>
86#include <sys/msgport2.h>
709799ea 87#include <stdio.h>
05220613 88#include <stdlib.h>
709799ea 89#include <string.h>
709799ea 90#include <machine/lock.h>
b950d4b8
EN
91#include <machine/atomic.h>
92#include <machine/cpu.h>
05220613
MD
93
94#endif
95
7d0bac62 96static int untimely_switch = 0;
0f7a3396
MD
97#ifdef INVARIANTS
98static int panic_on_cscount = 0;
99#endif
05220613
MD
100static __int64_t switch_count = 0;
101static __int64_t preempt_hit = 0;
102static __int64_t preempt_miss = 0;
103static __int64_t preempt_weird = 0;
38717797
HP
104static __int64_t token_contention_count = 0;
105static __int64_t mplock_contention_count = 0;
05220613
MD
106
107#ifdef _KERNEL
108
109SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
0f7a3396
MD
110#ifdef INVARIANTS
111SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
112#endif
4b5f931b 113SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 114SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 115SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 116SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
117#ifdef INVARIANTS
118SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
119 &token_contention_count, 0, "spinning due to token contention");
120SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
121 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
122#endif
05220613
MD
123#endif
124
4b5f931b
MD
125/*
126 * These helper procedures handle the runq, they can only be called from
127 * within a critical section.
75cdbe6c
MD
128 *
129 * WARNING! Prior to SMP being brought up it is possible to enqueue and
130 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
131 * instead of 'mycpu' when referencing the globaldata structure. Once
132 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 133 */
f1d1c3fa
MD
134static __inline
135void
136_lwkt_dequeue(thread_t td)
137{
138 if (td->td_flags & TDF_RUNQ) {
4b5f931b 139 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 140 struct globaldata *gd = td->td_gd;
4b5f931b 141
f1d1c3fa 142 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
143 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
144 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
145 }
146}
147
148static __inline
149void
150_lwkt_enqueue(thread_t td)
151{
344ad853 152 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
4b5f931b 153 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 154 struct globaldata *gd = td->td_gd;
4b5f931b 155
f1d1c3fa 156 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
157 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
158 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
159 }
160}
8ad65e08 161
37af14fe
MD
162/*
163 * Schedule a thread to run. As the current thread we can always safely
164 * schedule ourselves, and a shortcut procedure is provided for that
165 * function.
166 *
167 * (non-blocking, self contained on a per cpu basis)
168 */
169void
170lwkt_schedule_self(thread_t td)
171{
172 crit_enter_quick(td);
173 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
174 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
344ad853 175 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
37af14fe 176 _lwkt_enqueue(td);
37af14fe
MD
177 crit_exit_quick(td);
178}
179
180/*
181 * Deschedule a thread.
182 *
183 * (non-blocking, self contained on a per cpu basis)
184 */
185void
186lwkt_deschedule_self(thread_t td)
187{
188 crit_enter_quick(td);
189 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
190 _lwkt_dequeue(td);
191 crit_exit_quick(td);
192}
193
2d93b37a
MD
194#ifdef _KERNEL
195
8ad65e08
MD
196/*
197 * LWKTs operate on a per-cpu basis
198 *
73e4f7b9 199 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
200 */
201void
202lwkt_gdinit(struct globaldata *gd)
203{
4b5f931b
MD
204 int i;
205
206 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
207 TAILQ_INIT(&gd->gd_tdrunq[i]);
208 gd->gd_runqmask = 0;
73e4f7b9 209 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
210}
211
2d93b37a
MD
212#endif /* _KERNEL */
213
7d0bac62
MD
214/*
215 * Initialize a thread wait structure prior to first use.
216 *
217 * NOTE! called from low level boot code, we cannot do anything fancy!
218 */
219void
41a01a4d 220lwkt_wait_init(lwkt_wait_t w)
7d0bac62 221{
9d265729 222 spin_init(&w->wa_spinlock);
7d0bac62 223 TAILQ_INIT(&w->wa_waitq);
41a01a4d
MD
224 w->wa_gen = 0;
225 w->wa_count = 0;
7d0bac62
MD
226}
227
228/*
229 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
230 * or LWKT start address before it can be scheduled. If the target cpu is
231 * -1 the thread will be created on the current cpu.
0cfcada1
MD
232 *
233 * If you intend to create a thread without a process context this function
234 * does everything except load the startup and switcher function.
7d0bac62
MD
235 */
236thread_t
d3d32139 237lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 238{
99df837e 239 void *stack;
37af14fe 240 globaldata_t gd = mycpu;
7d0bac62 241
ef0fdad1 242 if (td == NULL) {
37af14fe
MD
243 crit_enter_gd(gd);
244 if (gd->gd_tdfreecount > 0) {
245 --gd->gd_tdfreecount;
246 td = TAILQ_FIRST(&gd->gd_tdfreeq);
d9eea1a5 247 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1 248 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
37af14fe
MD
249 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq);
250 crit_exit_gd(gd);
d3d32139 251 flags |= td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
ef0fdad1 252 } else {
37af14fe 253 crit_exit_gd(gd);
05220613 254#ifdef _KERNEL
ef0fdad1 255 td = zalloc(thread_zone);
05220613
MD
256#else
257 td = malloc(sizeof(struct thread));
258#endif
ef0fdad1 259 td->td_kstack = NULL;
f470d0c8 260 td->td_kstack_size = 0;
ef0fdad1
MD
261 flags |= TDF_ALLOCATED_THREAD;
262 }
263 }
f470d0c8
MD
264 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
265 if (flags & TDF_ALLOCATED_STACK) {
a9e5ae82 266#ifdef _KERNEL
f470d0c8 267 kmem_free(kernel_map, (vm_offset_t)stack, td->td_kstack_size);
a9e5ae82
JS
268#else
269 libcaps_free_stack(stack, td->td_kstack_size);
270#endif
f470d0c8
MD
271 stack = NULL;
272 }
273 }
274 if (stack == NULL) {
05220613 275#ifdef _KERNEL
f470d0c8 276 stack = (void *)kmem_alloc(kernel_map, stksize);
05220613 277#else
f470d0c8 278 stack = libcaps_alloc_stack(stksize);
05220613 279#endif
ef0fdad1 280 flags |= TDF_ALLOCATED_STACK;
99df837e 281 }
75cdbe6c 282 if (cpu < 0)
f470d0c8 283 lwkt_init_thread(td, stack, stksize, flags, mycpu);
75cdbe6c 284 else
f470d0c8 285 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 286 return(td);
7d0bac62
MD
287}
288
709799ea
MD
289#ifdef _KERNEL
290
7d0bac62
MD
291/*
292 * Initialize a preexisting thread structure. This function is used by
293 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
294 *
f8c3996b
MD
295 * All threads start out in a critical section at a priority of
296 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
297 * appropriate. This function may send an IPI message when the
298 * requested cpu is not the current cpu and consequently gd_tdallq may
299 * not be initialized synchronously from the point of view of the originating
300 * cpu.
301 *
302 * NOTE! we have to be careful in regards to creating threads for other cpus
303 * if SMP has not yet been activated.
7d0bac62 304 */
41a01a4d
MD
305#ifdef SMP
306
75cdbe6c
MD
307static void
308lwkt_init_thread_remote(void *arg)
309{
310 thread_t td = arg;
311
312 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
313}
314
41a01a4d
MD
315#endif
316
7d0bac62 317void
f470d0c8
MD
318lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
319 struct globaldata *gd)
7d0bac62 320{
37af14fe
MD
321 globaldata_t mygd = mycpu;
322
99df837e
MD
323 bzero(td, sizeof(struct thread));
324 td->td_kstack = stack;
f470d0c8 325 td->td_kstack_size = stksize;
d3d32139 326 td->td_flags = flags;
26a0694b 327 td->td_gd = gd;
f8c3996b 328 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
329#ifdef SMP
330 if ((flags & TDF_MPSAFE) == 0)
331 td->td_mpcount = 1;
332#endif
c95cd171 333 lwkt_initport(&td->td_msgport, td);
99df837e 334 pmap_init_thread(td);
0f7a3396 335#ifdef SMP
5d21b981
MD
336 /*
337 * Normally initializing a thread for a remote cpu requires sending an
338 * IPI. However, the idlethread is setup before the other cpus are
339 * activated so we have to treat it as a special case. XXX manipulation
340 * of gd_tdallq requires the BGL.
341 */
342 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 343 crit_enter_gd(mygd);
75cdbe6c 344 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 345 crit_exit_gd(mygd);
75cdbe6c 346 } else {
2db3b277 347 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 348 }
0f7a3396 349#else
37af14fe 350 crit_enter_gd(mygd);
0f7a3396 351 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 352 crit_exit_gd(mygd);
0f7a3396 353#endif
73e4f7b9
MD
354}
355
2d93b37a
MD
356#endif /* _KERNEL */
357
73e4f7b9
MD
358void
359lwkt_set_comm(thread_t td, const char *ctl, ...)
360{
e2565a42 361 __va_list va;
73e4f7b9 362
e2565a42 363 __va_start(va, ctl);
73e4f7b9 364 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 365 __va_end(va);
7d0bac62
MD
366}
367
99df837e 368void
73e4f7b9 369lwkt_hold(thread_t td)
99df837e 370{
73e4f7b9
MD
371 ++td->td_refs;
372}
373
374void
375lwkt_rele(thread_t td)
376{
377 KKASSERT(td->td_refs > 0);
378 --td->td_refs;
379}
380
c95cd171
MD
381#ifdef _KERNEL
382
73e4f7b9
MD
383void
384lwkt_wait_free(thread_t td)
385{
386 while (td->td_refs)
377d4740 387 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
388}
389
c95cd171
MD
390#endif
391
73e4f7b9
MD
392void
393lwkt_free_thread(thread_t td)
394{
395 struct globaldata *gd = mycpu;
396
d9eea1a5 397 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
398 ("lwkt_free_thread: did not exit! %p", td));
399
37af14fe 400 crit_enter_gd(gd);
1ec20b9d 401 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* Protected by BGL */
73e4f7b9 402 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
403 (td->td_flags & TDF_ALLOCATED_THREAD)
404 ) {
73e4f7b9
MD
405 ++gd->gd_tdfreecount;
406 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
37af14fe 407 crit_exit_gd(gd);
99df837e 408 } else {
37af14fe 409 crit_exit_gd(gd);
99df837e 410 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613 411#ifdef _KERNEL
f470d0c8 412 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
05220613 413#else
f470d0c8 414 libcaps_free_stack(td->td_kstack, td->td_kstack_size);
05220613 415#endif
73e4f7b9 416 /* gd invalid */
99df837e 417 td->td_kstack = NULL;
f470d0c8 418 td->td_kstack_size = 0;
99df837e 419 }
05220613
MD
420 if (td->td_flags & TDF_ALLOCATED_THREAD) {
421#ifdef _KERNEL
99df837e 422 zfree(thread_zone, td);
05220613
MD
423#else
424 free(td);
425#endif
426 }
99df837e
MD
427 }
428}
429
430
8ad65e08
MD
431/*
432 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
433 * switch to the idlethread. Switching must occur within a critical
434 * section to avoid races with the scheduling queue.
435 *
436 * We always have full control over our cpu's run queue. Other cpus
437 * that wish to manipulate our queue must use the cpu_*msg() calls to
438 * talk to our cpu, so a critical section is all that is needed and
439 * the result is very, very fast thread switching.
440 *
96728c05
MD
441 * The LWKT scheduler uses a fixed priority model and round-robins at
442 * each priority level. User process scheduling is a totally
443 * different beast and LWKT priorities should not be confused with
444 * user process priorities.
f1d1c3fa 445 *
96728c05
MD
446 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
447 * cleans it up. Note that the td_switch() function cannot do anything that
448 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
449 * the target thread (not the current thread). It's nice to have a scheduler
450 * that does not need the MP lock to work because it allows us to do some
451 * really cool high-performance MP lock optimizations.
69d78e99
MD
452 *
453 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
454 * is not called by the current thread in the preemption case, only when
455 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
456 */
457void
458lwkt_switch(void)
459{
37af14fe
MD
460 globaldata_t gd = mycpu;
461 thread_t td = gd->gd_curthread;
8ad65e08 462 thread_t ntd;
8a8d5d85
MD
463#ifdef SMP
464 int mpheld;
465#endif
8ad65e08 466
46a3f46d 467 /*
27e88a6e
MD
468 * Switching from within a 'fast' (non thread switched) interrupt or IPI
469 * is illegal. However, we may have to do it anyway if we hit a fatal
470 * kernel trap or we have paniced.
471 *
472 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 473 */
27e88a6e
MD
474 if (gd->gd_intr_nesting_level) {
475 int savegdnest;
476 int savegdtrap;
477
478 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
479 panic("lwkt_switch: cannot switch from within "
480 "a fast interrupt, yet, td %p\n", td);
481 } else {
482 savegdnest = gd->gd_intr_nesting_level;
483 savegdtrap = gd->gd_trap_nesting_level;
484 gd->gd_intr_nesting_level = 0;
485 gd->gd_trap_nesting_level = 0;
a7422615
MD
486 if ((td->td_flags & TDF_PANICWARN) == 0) {
487 td->td_flags |= TDF_PANICWARN;
488 printf("Warning: thread switch from interrupt or IPI, "
489 "thread %p (%s)\n", td, td->td_comm);
490#ifdef DDB
491 db_print_backtrace();
492#endif
493 }
27e88a6e
MD
494 lwkt_switch();
495 gd->gd_intr_nesting_level = savegdnest;
496 gd->gd_trap_nesting_level = savegdtrap;
497 return;
498 }
96728c05 499 }
ef0fdad1 500
cb973d15
MD
501 /*
502 * Passive release (used to transition from user to kernel mode
503 * when we block or switch rather then when we enter the kernel).
504 * This function is NOT called if we are switching into a preemption
505 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
506 * our current process designation (if we have one) and become a true
507 * LWKT thread, and may also hand the current process designation to
508 * another process and schedule thread.
cb973d15
MD
509 */
510 if (td->td_release)
511 td->td_release(td);
512
37af14fe 513 crit_enter_gd(gd);
dd55d707 514#ifdef SMP
9d265729
MD
515 if (td->td_toks)
516 lwkt_relalltokens(td);
dd55d707 517#endif
9d265729
MD
518
519 /*
b02926de
MD
520 * We had better not be holding any spin locks, but don't get into an
521 * endless panic loop.
9d265729 522 */
b02926de 523 KASSERT(td->td_spinlocks == 0 || panicstr != NULL,
9d265729
MD
524 ("lwkt_switch: still holding %d spinlocks!", td->td_spinlocks));
525
8a8d5d85
MD
526
527#ifdef SMP
528 /*
529 * td_mpcount cannot be used to determine if we currently hold the
530 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
531 * to get the lock, and switch out if it can't. Our ownership of
532 * the actual lock will remain stable while we are in a critical section
533 * (but, of course, another cpu may own or release the lock so the
534 * actual value of mp_lock is not stable).
8a8d5d85
MD
535 */
536 mpheld = MP_LOCK_HELD();
0f7a3396
MD
537#ifdef INVARIANTS
538 if (td->td_cscount) {
539 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
540 td);
541 if (panic_on_cscount)
542 panic("switching while mastering cpusync");
543 }
544#endif
8a8d5d85 545#endif
99df837e
MD
546 if ((ntd = td->td_preempted) != NULL) {
547 /*
548 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
549 * thread. This occurs transparently, whether the preempted thread
550 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
551 * itself).
552 *
553 * We have to setup the MP lock for the original thread after backing
554 * out the adjustment that was made to curthread when the original
555 * was preempted.
99df837e 556 */
26a0694b 557 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 558#ifdef SMP
96728c05 559 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 560 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
561 td, ntd, td->td_mpcount, ntd->td_mpcount);
562 }
8a8d5d85
MD
563 if (ntd->td_mpcount) {
564 td->td_mpcount -= ntd->td_mpcount;
565 KKASSERT(td->td_mpcount >= 0);
566 }
567#endif
26a0694b 568 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
569
570 /*
571 * XXX. The interrupt may have woken a thread up, we need to properly
572 * set the reschedule flag if the originally interrupted thread is at
573 * a lower priority.
574 */
575 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
576 need_lwkt_resched();
8a8d5d85 577 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 578 } else {
4b5f931b
MD
579 /*
580 * Priority queue / round-robin at each priority. Note that user
581 * processes run at a fixed, low priority and the user process
582 * scheduler deals with interactions between user processes
583 * by scheduling and descheduling them from the LWKT queue as
584 * necessary.
8a8d5d85
MD
585 *
586 * We have to adjust the MP lock for the target thread. If we
587 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
588 * thread that does not need the MP lock. If we cannot, we spin
589 * instead of HLT.
590 *
591 * A similar issue exists for the tokens held by the target thread.
592 * If we cannot obtain ownership of the tokens we cannot immediately
593 * schedule the thread.
594 */
595
8ec60c3f
MD
596 /*
597 * If an LWKT reschedule was requested, well that is what we are
598 * doing now so clear it.
599 */
600 clear_lwkt_resched();
4b5f931b
MD
601again:
602 if (gd->gd_runqmask) {
603 int nq = bsrl(gd->gd_runqmask);
604 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
605 gd->gd_runqmask &= ~(1 << nq);
606 goto again;
607 }
8a8d5d85 608#ifdef SMP
41a01a4d 609 /*
df6b8ba0
MD
610 * THREAD SELECTION FOR AN SMP MACHINE BUILD
611 *
41a01a4d
MD
612 * If the target needs the MP lock and we couldn't get it,
613 * or if the target is holding tokens and we could not
614 * gain ownership of the tokens, continue looking for a
615 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
616 *
617 * NOTE: the mpheld variable invalid after this conditional, it
618 * can change due to both cpu_try_mplock() returning success
9d265729 619 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
620 * we are trying to check the mpcount of a thread other then
621 * the current thread. Because of this, if the current thread
622 * is not holding td_mpcount, an IPI indirectly run via
9d265729 623 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 624 * cause the core MP lock to be released.
41a01a4d
MD
625 */
626 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
9d265729 627 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
41a01a4d 628 ) {
8a8d5d85 629 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
630
631 mpheld = MP_LOCK_HELD();
632 ntd = NULL;
8a8d5d85
MD
633 while (rqmask) {
634 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 635 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 636 /* spinning due to MP lock being held */
38717797 637#ifdef INVARIANTS
a453459d 638 ++mplock_contention_count;
38717797 639#endif
a453459d 640 /* mplock still not held, 'mpheld' still valid */
41a01a4d 641 continue;
38717797 642 }
a453459d
MD
643
644 /*
9d265729 645 * mpheld state invalid after getalltokens call returns
a453459d
MD
646 * failure, but the variable is only needed for
647 * the loop.
648 */
9d265729 649 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
a453459d 650 /* spinning due to token contention */
38717797 651#ifdef INVARIANTS
a453459d 652 ++token_contention_count;
38717797 653#endif
a453459d 654 mpheld = MP_LOCK_HELD();
41a01a4d 655 continue;
38717797 656 }
41a01a4d 657 break;
8a8d5d85
MD
658 }
659 if (ntd)
660 break;
661 rqmask &= ~(1 << nq);
662 nq = bsrl(rqmask);
663 }
664 if (ntd == NULL) {
a2a5ad0d
MD
665 ntd = &gd->gd_idlethread;
666 ntd->td_flags |= TDF_IDLE_NOHLT;
df6b8ba0 667 goto using_idle_thread;
8a8d5d85 668 } else {
344ad853 669 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
670 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
671 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
672 }
673 } else {
344ad853 674 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
675 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
676 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
677 }
678#else
df6b8ba0
MD
679 /*
680 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
681 * worry about tokens or the BGL.
682 */
344ad853 683 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
684 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
685 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 686#endif
4b5f931b 687 } else {
3c23a41a 688 /*
60f945af
MD
689 * We have nothing to run but only let the idle loop halt
690 * the cpu if there are no pending interrupts.
3c23a41a 691 */
a2a5ad0d 692 ntd = &gd->gd_idlethread;
60f945af 693 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 694 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 695#ifdef SMP
df6b8ba0
MD
696using_idle_thread:
697 /*
698 * The idle thread should not be holding the MP lock unless we
699 * are trapping in the kernel or in a panic. Since we select the
700 * idle thread unconditionally when no other thread is available,
701 * if the MP lock is desired during a panic or kernel trap, we
702 * have to loop in the scheduler until we get it.
703 */
704 if (ntd->td_mpcount) {
705 mpheld = MP_LOCK_HELD();
706 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
707 panic("Idle thread %p was holding the BGL!", ntd);
708 else if (mpheld == 0)
709 goto again;
710 }
a453459d 711#endif
4b5f931b 712 }
f1d1c3fa 713 }
26a0694b
MD
714 KASSERT(ntd->td_pri >= TDPRI_CRIT,
715 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
716
717 /*
718 * Do the actual switch. If the new target does not need the MP lock
719 * and we are holding it, release the MP lock. If the new target requires
720 * the MP lock we have already acquired it for the target.
721 */
722#ifdef SMP
723 if (ntd->td_mpcount == 0 ) {
724 if (MP_LOCK_HELD())
725 cpu_rel_mplock();
726 } else {
a453459d 727 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
728 }
729#endif
94f6d86e
MD
730 if (td != ntd) {
731 ++switch_count;
f1d1c3fa 732 td->td_switch(ntd);
94f6d86e 733 }
37af14fe
MD
734 /* NOTE: current cpu may have changed after switch */
735 crit_exit_quick(td);
8ad65e08
MD
736}
737
b68b7282 738/*
96728c05
MD
739 * Request that the target thread preempt the current thread. Preemption
740 * only works under a specific set of conditions:
b68b7282 741 *
96728c05
MD
742 * - We are not preempting ourselves
743 * - The target thread is owned by the current cpu
744 * - We are not currently being preempted
745 * - The target is not currently being preempted
746 * - We are able to satisfy the target's MP lock requirements (if any).
747 *
748 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
749 * this is called via lwkt_schedule() through the td_preemptable callback.
750 * critpri is the managed critical priority that we should ignore in order
751 * to determine whether preemption is possible (aka usually just the crit
752 * priority of lwkt_schedule() itself).
b68b7282 753 *
26a0694b
MD
754 * XXX at the moment we run the target thread in a critical section during
755 * the preemption in order to prevent the target from taking interrupts
756 * that *WE* can't. Preemption is strictly limited to interrupt threads
757 * and interrupt-like threads, outside of a critical section, and the
758 * preempted source thread will be resumed the instant the target blocks
759 * whether or not the source is scheduled (i.e. preemption is supposed to
760 * be as transparent as possible).
4b5f931b 761 *
8a8d5d85
MD
762 * The target thread inherits our MP count (added to its own) for the
763 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
764 * MP lock during the preemption. Therefore, any preempting targets must be
765 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
766 * out of sync with the physical mp_lock, but we do not have to preserve
767 * the original ownership of the lock if it was out of synch (that is, we
768 * can leave it synchronized on return).
b68b7282
MD
769 */
770void
96728c05 771lwkt_preempt(thread_t ntd, int critpri)
b68b7282 772{
46a3f46d 773 struct globaldata *gd = mycpu;
0a3f9b47 774 thread_t td;
8a8d5d85
MD
775#ifdef SMP
776 int mpheld;
57c254db 777 int savecnt;
8a8d5d85 778#endif
b68b7282 779
26a0694b 780 /*
96728c05
MD
781 * The caller has put us in a critical section. We can only preempt
782 * if the caller of the caller was not in a critical section (basically
0a3f9b47 783 * a local interrupt), as determined by the 'critpri' parameter.
96728c05
MD
784 *
785 * YYY The target thread must be in a critical section (else it must
786 * inherit our critical section? I dunno yet).
41a01a4d
MD
787 *
788 * Any tokens held by the target may not be held by thread(s) being
789 * preempted. We take the easy way out and do not preempt if
790 * the target is holding tokens.
0a3f9b47
MD
791 *
792 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
793 */
794 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 795
0a3f9b47 796 td = gd->gd_curthread;
0a3f9b47 797 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
798 ++preempt_miss;
799 return;
800 }
96728c05
MD
801 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
802 ++preempt_miss;
8ec60c3f 803 need_lwkt_resched();
96728c05
MD
804 return;
805 }
806#ifdef SMP
46a3f46d 807 if (ntd->td_gd != gd) {
96728c05 808 ++preempt_miss;
8ec60c3f 809 need_lwkt_resched();
96728c05
MD
810 return;
811 }
812#endif
41a01a4d
MD
813 /*
814 * Take the easy way out and do not preempt if the target is holding
815 * one or more tokens. We could test whether the thread(s) being
816 * preempted interlock against the target thread's tokens and whether
817 * we can get all the target thread's tokens, but this situation
818 * should not occur very often so its easier to simply not preempt.
819 */
820 if (ntd->td_toks != NULL) {
821 ++preempt_miss;
8ec60c3f 822 need_lwkt_resched();
41a01a4d
MD
823 return;
824 }
26a0694b
MD
825 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
826 ++preempt_weird;
8ec60c3f 827 need_lwkt_resched();
26a0694b
MD
828 return;
829 }
830 if (ntd->td_preempted) {
4b5f931b 831 ++preempt_hit;
8ec60c3f 832 need_lwkt_resched();
26a0694b 833 return;
b68b7282 834 }
8a8d5d85 835#ifdef SMP
a2a5ad0d
MD
836 /*
837 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
838 * to or from the MP lock. In this case td_mpcount will be pre-disposed
839 * (non-zero) but not actually synchronized with the actual state of the
840 * lock. We can use it to imply an MP lock requirement for the
841 * preemption but we cannot use it to test whether we hold the MP lock
842 * or not.
a2a5ad0d 843 */
96728c05 844 savecnt = td->td_mpcount;
71ef2f5c 845 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
846 ntd->td_mpcount += td->td_mpcount;
847 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
848 ntd->td_mpcount -= td->td_mpcount;
849 ++preempt_miss;
8ec60c3f 850 need_lwkt_resched();
8a8d5d85
MD
851 return;
852 }
853#endif
26a0694b 854
8ec60c3f
MD
855 /*
856 * Since we are able to preempt the current thread, there is no need to
857 * call need_lwkt_resched().
858 */
26a0694b
MD
859 ++preempt_hit;
860 ntd->td_preempted = td;
861 td->td_flags |= TDF_PREEMPT_LOCK;
862 td->td_switch(ntd);
863 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
864#ifdef SMP
865 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
866 mpheld = MP_LOCK_HELD();
867 if (mpheld && td->td_mpcount == 0)
96728c05 868 cpu_rel_mplock();
71ef2f5c 869 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
870 panic("lwkt_preempt(): MP lock was not held through");
871#endif
26a0694b
MD
872 ntd->td_preempted = NULL;
873 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
874}
875
f1d1c3fa
MD
876/*
877 * Yield our thread while higher priority threads are pending. This is
878 * typically called when we leave a critical section but it can be safely
879 * called while we are in a critical section.
880 *
881 * This function will not generally yield to equal priority threads but it
882 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 883 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
884 * lwkt_yield_quick().
885 *
235957ed 886 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
887 * came along but was blocked and made pending.
888 *
f1d1c3fa
MD
889 * (self contained on a per cpu basis)
890 */
891void
892lwkt_yield_quick(void)
893{
7966cb69
MD
894 globaldata_t gd = mycpu;
895 thread_t td = gd->gd_curthread;
ef0fdad1 896
a2a5ad0d 897 /*
235957ed 898 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
899 * it with a non-zero cpl then we might not wind up calling splz after
900 * a task switch when the critical section is exited even though the
46a3f46d 901 * new task could accept the interrupt.
a2a5ad0d
MD
902 *
903 * XXX from crit_exit() only called after last crit section is released.
904 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
905 *
906 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
907 * except for this special case, we MUST call splz() here to handle any
908 * pending ints, particularly after we switch, or we might accidently
909 * halt the cpu with interrupts pending.
a2a5ad0d 910 */
46a3f46d 911 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 912 splz();
f1d1c3fa
MD
913
914 /*
915 * YYY enabling will cause wakeup() to task-switch, which really
916 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
917 * preemption and MP without actually doing preemption or MP, because a
918 * lot of code assumes that wakeup() does not block.
f1d1c3fa 919 */
46a3f46d
MD
920 if (untimely_switch && td->td_nest_count == 0 &&
921 gd->gd_intr_nesting_level == 0
922 ) {
37af14fe 923 crit_enter_quick(td);
f1d1c3fa
MD
924 /*
925 * YYY temporary hacks until we disassociate the userland scheduler
926 * from the LWKT scheduler.
927 */
928 if (td->td_flags & TDF_RUNQ) {
929 lwkt_switch(); /* will not reenter yield function */
930 } else {
37af14fe 931 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 932 lwkt_switch(); /* will not reenter yield function */
37af14fe 933 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 934 }
7966cb69 935 crit_exit_noyield(td);
f1d1c3fa 936 }
f1d1c3fa
MD
937}
938
8ad65e08 939/*
f1d1c3fa 940 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 941 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
942 * the crit_exit() call in lwkt_switch().
943 *
944 * (self contained on a per cpu basis)
8ad65e08
MD
945 */
946void
f1d1c3fa 947lwkt_yield(void)
8ad65e08 948{
37af14fe 949 lwkt_schedule_self(curthread);
f1d1c3fa
MD
950 lwkt_switch();
951}
952
8ad65e08 953/*
f1d1c3fa
MD
954 * Generic schedule. Possibly schedule threads belonging to other cpus and
955 * deal with threads that might be blocked on a wait queue.
956 *
0a3f9b47
MD
957 * We have a little helper inline function which does additional work after
958 * the thread has been enqueued, including dealing with preemption and
959 * setting need_lwkt_resched() (which prevents the kernel from returning
960 * to userland until it has processed higher priority threads).
6330a558
MD
961 *
962 * It is possible for this routine to be called after a failed _enqueue
963 * (due to the target thread migrating, sleeping, or otherwise blocked).
964 * We have to check that the thread is actually on the run queue!
8ad65e08 965 */
0a3f9b47
MD
966static __inline
967void
8ec60c3f 968_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
0a3f9b47 969{
6330a558
MD
970 if (ntd->td_flags & TDF_RUNQ) {
971 if (ntd->td_preemptable) {
972 ntd->td_preemptable(ntd, cpri); /* YYY +token */
973 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
974 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
975 ) {
976 need_lwkt_resched();
977 }
0a3f9b47
MD
978 }
979}
980
8ad65e08
MD
981void
982lwkt_schedule(thread_t td)
983{
37af14fe
MD
984 globaldata_t mygd = mycpu;
985
41a01a4d 986 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 987 crit_enter_gd(mygd);
344ad853 988 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
37af14fe 989 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
990 _lwkt_enqueue(td);
991 } else {
992 lwkt_wait_t w;
993
994 /*
995 * If the thread is on a wait list we have to send our scheduling
996 * request to the owner of the wait structure. Otherwise we send
997 * the scheduling request to the cpu owning the thread. Races
998 * are ok, the target will forward the message as necessary (the
999 * message may chase the thread around before it finally gets
1000 * acted upon).
1001 *
1002 * (remember, wait structures use stable storage)
0a3f9b47 1003 *
0c453950
MD
1004 * NOTE: we have to account for the number of critical sections
1005 * under our control when calling _lwkt_schedule_post() so it
1006 * can figure out whether preemption is allowed.
1007 *
1008 * NOTE: The wait structure algorithms are a mess and need to be
1009 * rewritten.
1010 *
1011 * NOTE: We cannot safely acquire or release a token, even
1012 * non-blocking, because this routine may be called in the context
1013 * of a thread already holding the token and thus not provide any
1014 * interlock protection. We cannot safely manipulate the td_toks
1015 * list for the same reason. Instead we depend on our critical
1016 * section if the token is owned by our cpu.
f1d1c3fa
MD
1017 */
1018 if ((w = td->td_wait) != NULL) {
9d265729
MD
1019 spin_lock_quick(&w->wa_spinlock);
1020 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1021 --w->wa_count;
1022 td->td_wait = NULL;
1023 spin_unlock_quick(&w->wa_spinlock);
0f7a3396 1024#ifdef SMP
9d265729 1025 if (td->td_gd == mygd) {
0f7a3396 1026 _lwkt_enqueue(td);
8ec60c3f 1027 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 1028 } else {
9d265729
MD
1029 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1030 }
b8a98473 1031#else
9d265729
MD
1032 _lwkt_enqueue(td);
1033 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
b8a98473 1034#endif
f1d1c3fa
MD
1035 } else {
1036 /*
1037 * If the wait structure is NULL and we own the thread, there
1038 * is no race (since we are in a critical section). If we
1039 * do not own the thread there might be a race but the
1040 * target cpu will deal with it.
1041 */
0f7a3396 1042#ifdef SMP
37af14fe 1043 if (td->td_gd == mygd) {
f1d1c3fa 1044 _lwkt_enqueue(td);
8ec60c3f 1045 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 1046 } else {
b8a98473 1047 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
f1d1c3fa 1048 }
0f7a3396
MD
1049#else
1050 _lwkt_enqueue(td);
8ec60c3f 1051 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 1052#endif
f1d1c3fa 1053 }
8ad65e08 1054 }
37af14fe 1055 crit_exit_gd(mygd);
8ad65e08
MD
1056}
1057
d9eea1a5
MD
1058/*
1059 * Managed acquisition. This code assumes that the MP lock is held for
1060 * the tdallq operation and that the thread has been descheduled from its
1061 * original cpu. We also have to wait for the thread to be entirely switched
1062 * out on its original cpu (this is usually fast enough that we never loop)
1063 * since the LWKT system does not have to hold the MP lock while switching
1064 * and the target may have released it before switching.
1065 */
a2a5ad0d
MD
1066void
1067lwkt_acquire(thread_t td)
1068{
37af14fe
MD
1069 globaldata_t gd;
1070 globaldata_t mygd;
a2a5ad0d
MD
1071
1072 gd = td->td_gd;
37af14fe 1073 mygd = mycpu;
35238fa5 1074 cpu_lfence();
a2a5ad0d 1075 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
c1102e9f 1076 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) /* XXX spin */
35238fa5 1077 cpu_lfence();
37af14fe
MD
1078 if (gd != mygd) {
1079 crit_enter_gd(mygd);
a2a5ad0d 1080 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
37af14fe
MD
1081 td->td_gd = mygd;
1082 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
1083 crit_exit_gd(mygd);
a2a5ad0d
MD
1084 }
1085}
1086
f1d1c3fa
MD
1087/*
1088 * Generic deschedule. Descheduling threads other then your own should be
1089 * done only in carefully controlled circumstances. Descheduling is
1090 * asynchronous.
1091 *
1092 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1093 */
1094void
1095lwkt_deschedule(thread_t td)
1096{
f1d1c3fa 1097 crit_enter();
b8a98473 1098#ifdef SMP
f1d1c3fa
MD
1099 if (td == curthread) {
1100 _lwkt_dequeue(td);
1101 } else {
a72187e9 1102 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1103 _lwkt_dequeue(td);
1104 } else {
b8a98473 1105 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1106 }
1107 }
b8a98473
MD
1108#else
1109 _lwkt_dequeue(td);
1110#endif
f1d1c3fa
MD
1111 crit_exit();
1112}
1113
4b5f931b
MD
1114/*
1115 * Set the target thread's priority. This routine does not automatically
1116 * switch to a higher priority thread, LWKT threads are not designed for
1117 * continuous priority changes. Yield if you want to switch.
1118 *
1119 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1120 * of the td_pri field. The specified priority may also indicate zero or
1121 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1122 *
1123 * Note that we requeue the thread whether it winds up on a different runq
1124 * or not. uio_yield() depends on this and the routine is not normally
1125 * called with the same priority otherwise.
4b5f931b
MD
1126 */
1127void
1128lwkt_setpri(thread_t td, int pri)
1129{
26a0694b 1130 KKASSERT(pri >= 0);
a72187e9 1131 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1132 crit_enter();
1133 if (td->td_flags & TDF_RUNQ) {
1134 _lwkt_dequeue(td);
1135 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1136 _lwkt_enqueue(td);
1137 } else {
1138 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1139 }
1140 crit_exit();
1141}
1142
1143void
1144lwkt_setpri_self(int pri)
1145{
1146 thread_t td = curthread;
1147
4b5f931b
MD
1148 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1149 crit_enter();
1150 if (td->td_flags & TDF_RUNQ) {
1151 _lwkt_dequeue(td);
1152 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1153 _lwkt_enqueue(td);
1154 } else {
1155 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1156 }
1157 crit_exit();
1158}
1159
8ec60c3f
MD
1160/*
1161 * Determine if there is a runnable thread at a higher priority then
1162 * the current thread. lwkt_setpri() does not check this automatically.
1163 * Return 1 if there is, 0 if there isn't.
1164 *
1165 * Example: if bit 31 of runqmask is set and the current thread is priority
1166 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1167 *
1168 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1169 * up comparing against 0xffffffff, a comparison that will always be false.
1170 */
1171int
1172lwkt_checkpri_self(void)
1173{
1174 globaldata_t gd = mycpu;
1175 thread_t td = gd->gd_curthread;
1176 int nq = td->td_pri & TDPRI_MASK;
1177
1178 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1179 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1180 return(1);
1181 ++nq;
1182 }
1183 return(0);
1184}
1185
5d21b981
MD
1186/*
1187 * Migrate the current thread to the specified cpu. The BGL must be held
1188 * (for the gd_tdallq manipulation XXX). This is accomplished by
1189 * descheduling ourselves from the current cpu, moving our thread to the
1190 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1191 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1192 */
3d28ff59 1193#ifdef SMP
5d21b981 1194static void lwkt_setcpu_remote(void *arg);
3d28ff59 1195#endif
5d21b981
MD
1196
1197void
1198lwkt_setcpu_self(globaldata_t rgd)
1199{
1200#ifdef SMP
1201 thread_t td = curthread;
1202
1203 if (td->td_gd != rgd) {
1204 crit_enter_quick(td);
1205 td->td_flags |= TDF_MIGRATING;
1206 lwkt_deschedule_self(td);
1207 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1208 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
b8a98473 1209 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1210 lwkt_switch();
1211 /* we are now on the target cpu */
1212 crit_exit_quick(td);
1213 }
1214#endif
1215}
1216
ecdefdda
MD
1217void
1218lwkt_migratecpu(int cpuid)
1219{
1220#ifdef SMP
1221 globaldata_t rgd;
1222
1223 rgd = globaldata_find(cpuid);
1224 lwkt_setcpu_self(rgd);
1225#endif
1226}
1227
5d21b981
MD
1228/*
1229 * Remote IPI for cpu migration (called while in a critical section so we
1230 * do not have to enter another one). The thread has already been moved to
1231 * our cpu's allq, but we must wait for the thread to be completely switched
1232 * out on the originating cpu before we schedule it on ours or the stack
1233 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1234 * change to main memory.
1235 *
1236 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1237 * against wakeups. It is best if this interface is used only when there
1238 * are no pending events that might try to schedule the thread.
1239 */
3d28ff59 1240#ifdef SMP
5d21b981
MD
1241static void
1242lwkt_setcpu_remote(void *arg)
1243{
1244 thread_t td = arg;
1245 globaldata_t gd = mycpu;
1246
c1102e9f 1247 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK))
35238fa5 1248 cpu_lfence();
5d21b981 1249 td->td_gd = gd;
35238fa5 1250 cpu_sfence();
5d21b981 1251 td->td_flags &= ~TDF_MIGRATING;
344ad853 1252 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
5d21b981
MD
1253 _lwkt_enqueue(td);
1254}
3d28ff59 1255#endif
5d21b981 1256
553ea3c8 1257struct lwp *
4b5f931b
MD
1258lwkt_preempted_proc(void)
1259{
73e4f7b9 1260 thread_t td = curthread;
4b5f931b
MD
1261 while (td->td_preempted)
1262 td = td->td_preempted;
553ea3c8 1263 return(td->td_lwp);
4b5f931b
MD
1264}
1265
f1d1c3fa 1266/*
41a01a4d
MD
1267 * Block on the specified wait queue until signaled. A generation number
1268 * must be supplied to interlock the wait queue. The function will
1269 * return immediately if the generation number does not match the wait
1270 * structure's generation number.
f1d1c3fa
MD
1271 */
1272void
ae8050a4 1273lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
1274{
1275 thread_t td = curthread;
f1d1c3fa 1276
9d265729 1277 spin_lock(&w->wa_spinlock);
ae8050a4 1278 if (w->wa_gen == *gen) {
f1d1c3fa 1279 _lwkt_dequeue(td);
344ad853 1280 td->td_flags |= TDF_BLOCKQ;
f1d1c3fa
MD
1281 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1282 ++w->wa_count;
1283 td->td_wait = w;
ae8050a4 1284 td->td_wmesg = wmesg;
9d265729 1285 spin_unlock(&w->wa_spinlock);
f1d1c3fa 1286 lwkt_switch();
344ad853
MD
1287 KKASSERT((td->td_flags & TDF_BLOCKQ) == 0);
1288 td->td_wmesg = NULL;
9d265729
MD
1289 *gen = w->wa_gen;
1290 } else {
1291 *gen = w->wa_gen;
1292 spin_unlock(&w->wa_spinlock);
8ad65e08 1293 }
f1d1c3fa
MD
1294}
1295
1296/*
1297 * Signal a wait queue. We gain ownership of the wait queue in order to
1298 * signal it. Once a thread is removed from the wait queue we have to
1299 * deal with the cpu owning the thread.
1300 *
1301 * Note: alternatively we could message the target cpu owning the wait
1302 * queue. YYY implement as sysctl.
1303 */
1304void
ece04fd0 1305lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1306{
1307 thread_t td;
f1d1c3fa 1308
9d265729 1309 spin_lock(&w->wa_spinlock);
f1d1c3fa 1310 ++w->wa_gen;
ece04fd0
MD
1311 if (count < 0)
1312 count = w->wa_count;
f1d1c3fa
MD
1313 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1314 --count;
1315 --w->wa_count;
344ad853 1316 KKASSERT(td->td_flags & TDF_BLOCKQ);
f1d1c3fa 1317 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
344ad853 1318 td->td_flags &= ~TDF_BLOCKQ;
f1d1c3fa 1319 td->td_wait = NULL;
9d265729 1320 spin_unlock(&w->wa_spinlock);
344ad853 1321 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
b8a98473 1322#ifdef SMP
a72187e9 1323 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1324 _lwkt_enqueue(td);
1325 } else {
b8a98473 1326 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
f1d1c3fa 1327 }
b8a98473
MD
1328#else
1329 _lwkt_enqueue(td);
1330#endif
9d265729 1331 spin_lock(&w->wa_spinlock);
f1d1c3fa 1332 }
9d265729 1333 spin_unlock(&w->wa_spinlock);
f1d1c3fa
MD
1334}
1335
99df837e
MD
1336/*
1337 * Create a kernel process/thread/whatever. It shares it's address space
1338 * with proc0 - ie: kernel only.
1339 *
365fa13f
MD
1340 * NOTE! By default new threads are created with the MP lock held. A
1341 * thread which does not require the MP lock should release it by calling
1342 * rel_mplock() at the start of the new thread.
99df837e
MD
1343 */
1344int
1345lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1346 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1347 const char *fmt, ...)
99df837e 1348{
73e4f7b9 1349 thread_t td;
e2565a42 1350 __va_list ap;
99df837e 1351
d3d32139
MD
1352 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1353 tdflags | TDF_VERBOSE);
a2a5ad0d
MD
1354 if (tdp)
1355 *tdp = td;
709799ea 1356 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1357
1358 /*
1359 * Set up arg0 for 'ps' etc
1360 */
e2565a42 1361 __va_start(ap, fmt);
99df837e 1362 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1363 __va_end(ap);
99df837e
MD
1364
1365 /*
1366 * Schedule the thread to run
1367 */
ef0fdad1
MD
1368 if ((td->td_flags & TDF_STOPREQ) == 0)
1369 lwkt_schedule(td);
1370 else
1371 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1372 return 0;
1373}
1374
2d93b37a 1375/*
2d93b37a
MD
1376 * kthread_* is specific to the kernel and is not needed by userland.
1377 */
1378#ifdef _KERNEL
1379
99df837e
MD
1380/*
1381 * Destroy an LWKT thread. Warning! This function is not called when
1382 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1383 * uses a different reaping mechanism.
1384 */
1385void
1386lwkt_exit(void)
1387{
1388 thread_t td = curthread;
8826f33a 1389 globaldata_t gd;
99df837e
MD
1390
1391 if (td->td_flags & TDF_VERBOSE)
1392 printf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1393 caps_exit(td);
37af14fe
MD
1394 crit_enter_quick(td);
1395 lwkt_deschedule_self(td);
8826f33a
MD
1396 gd = mycpu;
1397 KKASSERT(gd == td->td_gd);
1398 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1399 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1400 ++gd->gd_tdfreecount;
1401 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1402 }
99df837e
MD
1403 cpu_thread_exit();
1404}
1405
2d93b37a
MD
1406#endif /* _KERNEL */
1407
1408void
1409crit_panic(void)
1410{
1411 thread_t td = curthread;
1412 int lpri = td->td_pri;
1413
1414 td->td_pri = 0;
1415 panic("td_pri is/would-go negative! %p %d", td, lpri);
1416}
1417
d165e668
MD
1418#ifdef SMP
1419
bd8015ca
MD
1420/*
1421 * Called from debugger/panic on cpus which have been stopped. We must still
1422 * process the IPIQ while stopped, even if we were stopped while in a critical
1423 * section (XXX).
1424 *
1425 * If we are dumping also try to process any pending interrupts. This may
1426 * or may not work depending on the state of the cpu at the point it was
1427 * stopped.
1428 */
1429void
1430lwkt_smp_stopped(void)
1431{
1432 globaldata_t gd = mycpu;
1433
1434 crit_enter_gd(gd);
1435 if (dumping) {
1436 lwkt_process_ipiq();
1437 splz();
1438 } else {
1439 lwkt_process_ipiq();
1440 }
1441 crit_exit_gd(gd);
1442}
1443
d165e668 1444#endif