Don't display the file and line by default.
[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 *
c1102e9f 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.89 2005/12/02 22:02:17 dillon Exp $
75cdbe6c
MD
35 */
36
37/*
38 * Each cpu in a system has its own self-contained light weight kernel
39 * thread scheduler, which means that generally speaking we only need
40 * to use a critical section to avoid problems. Foreign thread
41 * scheduling is queued via (async) IPIs.
8ad65e08
MD
42 */
43
05220613
MD
44#ifdef _KERNEL
45
8ad65e08
MD
46#include <sys/param.h>
47#include <sys/systm.h>
48#include <sys/kernel.h>
49#include <sys/proc.h>
50#include <sys/rtprio.h>
51#include <sys/queue.h>
f1d1c3fa 52#include <sys/thread2.h>
7d0bac62 53#include <sys/sysctl.h>
99df837e 54#include <sys/kthread.h>
f1d1c3fa 55#include <machine/cpu.h>
99df837e 56#include <sys/lock.h>
f6bf3af1 57#include <sys/caps.h>
f1d1c3fa 58
7d0bac62
MD
59#include <vm/vm.h>
60#include <vm/vm_param.h>
61#include <vm/vm_kern.h>
62#include <vm/vm_object.h>
63#include <vm/vm_page.h>
64#include <vm/vm_map.h>
65#include <vm/vm_pager.h>
66#include <vm/vm_extern.h>
67#include <vm/vm_zone.h>
68
99df837e 69#include <machine/stdarg.h>
57c254db 70#include <machine/ipl.h>
96728c05 71#include <machine/smp.h>
99df837e 72
05220613
MD
73#else
74
75#include <sys/stdint.h>
fb04f4fd 76#include <libcaps/thread.h>
05220613
MD
77#include <sys/thread.h>
78#include <sys/msgport.h>
79#include <sys/errno.h>
fb04f4fd 80#include <libcaps/globaldata.h>
7e8303ad 81#include <machine/cpufunc.h>
05220613
MD
82#include <sys/thread2.h>
83#include <sys/msgport2.h>
709799ea 84#include <stdio.h>
05220613 85#include <stdlib.h>
709799ea 86#include <string.h>
709799ea 87#include <machine/lock.h>
b950d4b8
EN
88#include <machine/atomic.h>
89#include <machine/cpu.h>
05220613
MD
90
91#endif
92
7d0bac62 93static int untimely_switch = 0;
0f7a3396
MD
94#ifdef INVARIANTS
95static int panic_on_cscount = 0;
96#endif
05220613
MD
97static __int64_t switch_count = 0;
98static __int64_t preempt_hit = 0;
99static __int64_t preempt_miss = 0;
100static __int64_t preempt_weird = 0;
38717797
HP
101static __int64_t token_contention_count = 0;
102static __int64_t mplock_contention_count = 0;
05220613
MD
103
104#ifdef _KERNEL
105
106SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
0f7a3396
MD
107#ifdef INVARIANTS
108SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
109#endif
4b5f931b 110SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 111SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 112SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 113SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
114#ifdef INVARIANTS
115SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
116 &token_contention_count, 0, "spinning due to token contention");
117SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
118 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
119#endif
05220613
MD
120#endif
121
4b5f931b
MD
122/*
123 * These helper procedures handle the runq, they can only be called from
124 * within a critical section.
75cdbe6c
MD
125 *
126 * WARNING! Prior to SMP being brought up it is possible to enqueue and
127 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
128 * instead of 'mycpu' when referencing the globaldata structure. Once
129 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 130 */
f1d1c3fa
MD
131static __inline
132void
133_lwkt_dequeue(thread_t td)
134{
135 if (td->td_flags & TDF_RUNQ) {
4b5f931b 136 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 137 struct globaldata *gd = td->td_gd;
4b5f931b 138
f1d1c3fa 139 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
140 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
141 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
142 }
143}
144
145static __inline
146void
147_lwkt_enqueue(thread_t td)
148{
344ad853 149 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
4b5f931b 150 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 151 struct globaldata *gd = td->td_gd;
4b5f931b 152
f1d1c3fa 153 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
154 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
155 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
156 }
157}
8ad65e08 158
37af14fe
MD
159/*
160 * Schedule a thread to run. As the current thread we can always safely
161 * schedule ourselves, and a shortcut procedure is provided for that
162 * function.
163 *
164 * (non-blocking, self contained on a per cpu basis)
165 */
166void
167lwkt_schedule_self(thread_t td)
168{
169 crit_enter_quick(td);
170 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
171 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
344ad853 172 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
37af14fe 173 _lwkt_enqueue(td);
37af14fe
MD
174 crit_exit_quick(td);
175}
176
177/*
178 * Deschedule a thread.
179 *
180 * (non-blocking, self contained on a per cpu basis)
181 */
182void
183lwkt_deschedule_self(thread_t td)
184{
185 crit_enter_quick(td);
186 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
187 _lwkt_dequeue(td);
188 crit_exit_quick(td);
189}
190
2d93b37a
MD
191#ifdef _KERNEL
192
8ad65e08
MD
193/*
194 * LWKTs operate on a per-cpu basis
195 *
73e4f7b9 196 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
197 */
198void
199lwkt_gdinit(struct globaldata *gd)
200{
4b5f931b
MD
201 int i;
202
203 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
204 TAILQ_INIT(&gd->gd_tdrunq[i]);
205 gd->gd_runqmask = 0;
73e4f7b9 206 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
207}
208
2d93b37a
MD
209#endif /* _KERNEL */
210
7d0bac62
MD
211/*
212 * Initialize a thread wait structure prior to first use.
213 *
214 * NOTE! called from low level boot code, we cannot do anything fancy!
215 */
216void
41a01a4d 217lwkt_wait_init(lwkt_wait_t w)
7d0bac62 218{
41a01a4d 219 lwkt_token_init(&w->wa_token);
7d0bac62 220 TAILQ_INIT(&w->wa_waitq);
41a01a4d
MD
221 w->wa_gen = 0;
222 w->wa_count = 0;
7d0bac62
MD
223}
224
225/*
226 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
227 * or LWKT start address before it can be scheduled. If the target cpu is
228 * -1 the thread will be created on the current cpu.
0cfcada1
MD
229 *
230 * If you intend to create a thread without a process context this function
231 * does everything except load the startup and switcher function.
7d0bac62
MD
232 */
233thread_t
d3d32139 234lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 235{
99df837e 236 void *stack;
37af14fe 237 globaldata_t gd = mycpu;
7d0bac62 238
ef0fdad1 239 if (td == NULL) {
37af14fe
MD
240 crit_enter_gd(gd);
241 if (gd->gd_tdfreecount > 0) {
242 --gd->gd_tdfreecount;
243 td = TAILQ_FIRST(&gd->gd_tdfreeq);
d9eea1a5 244 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1 245 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
37af14fe
MD
246 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq);
247 crit_exit_gd(gd);
d3d32139 248 flags |= td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
ef0fdad1 249 } else {
37af14fe 250 crit_exit_gd(gd);
05220613 251#ifdef _KERNEL
ef0fdad1 252 td = zalloc(thread_zone);
05220613
MD
253#else
254 td = malloc(sizeof(struct thread));
255#endif
ef0fdad1 256 td->td_kstack = NULL;
f470d0c8 257 td->td_kstack_size = 0;
ef0fdad1
MD
258 flags |= TDF_ALLOCATED_THREAD;
259 }
260 }
f470d0c8
MD
261 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
262 if (flags & TDF_ALLOCATED_STACK) {
a9e5ae82 263#ifdef _KERNEL
f470d0c8 264 kmem_free(kernel_map, (vm_offset_t)stack, td->td_kstack_size);
a9e5ae82
JS
265#else
266 libcaps_free_stack(stack, td->td_kstack_size);
267#endif
f470d0c8
MD
268 stack = NULL;
269 }
270 }
271 if (stack == NULL) {
05220613 272#ifdef _KERNEL
f470d0c8 273 stack = (void *)kmem_alloc(kernel_map, stksize);
05220613 274#else
f470d0c8 275 stack = libcaps_alloc_stack(stksize);
05220613 276#endif
ef0fdad1 277 flags |= TDF_ALLOCATED_STACK;
99df837e 278 }
75cdbe6c 279 if (cpu < 0)
f470d0c8 280 lwkt_init_thread(td, stack, stksize, flags, mycpu);
75cdbe6c 281 else
f470d0c8 282 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 283 return(td);
7d0bac62
MD
284}
285
709799ea
MD
286#ifdef _KERNEL
287
7d0bac62
MD
288/*
289 * Initialize a preexisting thread structure. This function is used by
290 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
291 *
f8c3996b
MD
292 * All threads start out in a critical section at a priority of
293 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
294 * appropriate. This function may send an IPI message when the
295 * requested cpu is not the current cpu and consequently gd_tdallq may
296 * not be initialized synchronously from the point of view of the originating
297 * cpu.
298 *
299 * NOTE! we have to be careful in regards to creating threads for other cpus
300 * if SMP has not yet been activated.
7d0bac62 301 */
41a01a4d
MD
302#ifdef SMP
303
75cdbe6c
MD
304static void
305lwkt_init_thread_remote(void *arg)
306{
307 thread_t td = arg;
308
309 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
310}
311
41a01a4d
MD
312#endif
313
7d0bac62 314void
f470d0c8
MD
315lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
316 struct globaldata *gd)
7d0bac62 317{
37af14fe
MD
318 globaldata_t mygd = mycpu;
319
99df837e
MD
320 bzero(td, sizeof(struct thread));
321 td->td_kstack = stack;
f470d0c8 322 td->td_kstack_size = stksize;
d3d32139 323 td->td_flags = flags;
26a0694b 324 td->td_gd = gd;
f8c3996b 325 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
326#ifdef SMP
327 if ((flags & TDF_MPSAFE) == 0)
328 td->td_mpcount = 1;
329#endif
c95cd171 330 lwkt_initport(&td->td_msgport, td);
99df837e 331 pmap_init_thread(td);
0f7a3396 332#ifdef SMP
5d21b981
MD
333 /*
334 * Normally initializing a thread for a remote cpu requires sending an
335 * IPI. However, the idlethread is setup before the other cpus are
336 * activated so we have to treat it as a special case. XXX manipulation
337 * of gd_tdallq requires the BGL.
338 */
339 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 340 crit_enter_gd(mygd);
75cdbe6c 341 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 342 crit_exit_gd(mygd);
75cdbe6c 343 } else {
2db3b277 344 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 345 }
0f7a3396 346#else
37af14fe 347 crit_enter_gd(mygd);
0f7a3396 348 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 349 crit_exit_gd(mygd);
0f7a3396 350#endif
73e4f7b9
MD
351}
352
2d93b37a
MD
353#endif /* _KERNEL */
354
73e4f7b9
MD
355void
356lwkt_set_comm(thread_t td, const char *ctl, ...)
357{
e2565a42 358 __va_list va;
73e4f7b9 359
e2565a42 360 __va_start(va, ctl);
73e4f7b9 361 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 362 __va_end(va);
7d0bac62
MD
363}
364
99df837e 365void
73e4f7b9 366lwkt_hold(thread_t td)
99df837e 367{
73e4f7b9
MD
368 ++td->td_refs;
369}
370
371void
372lwkt_rele(thread_t td)
373{
374 KKASSERT(td->td_refs > 0);
375 --td->td_refs;
376}
377
c95cd171
MD
378#ifdef _KERNEL
379
73e4f7b9
MD
380void
381lwkt_wait_free(thread_t td)
382{
383 while (td->td_refs)
377d4740 384 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
385}
386
c95cd171
MD
387#endif
388
73e4f7b9
MD
389void
390lwkt_free_thread(thread_t td)
391{
392 struct globaldata *gd = mycpu;
393
d9eea1a5 394 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
395 ("lwkt_free_thread: did not exit! %p", td));
396
37af14fe 397 crit_enter_gd(gd);
73e4f7b9
MD
398 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
399 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
400 (td->td_flags & TDF_ALLOCATED_THREAD)
401 ) {
73e4f7b9
MD
402 ++gd->gd_tdfreecount;
403 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
37af14fe 404 crit_exit_gd(gd);
99df837e 405 } else {
37af14fe 406 crit_exit_gd(gd);
99df837e 407 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613 408#ifdef _KERNEL
f470d0c8 409 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
05220613 410#else
f470d0c8 411 libcaps_free_stack(td->td_kstack, td->td_kstack_size);
05220613 412#endif
73e4f7b9 413 /* gd invalid */
99df837e 414 td->td_kstack = NULL;
f470d0c8 415 td->td_kstack_size = 0;
99df837e 416 }
05220613
MD
417 if (td->td_flags & TDF_ALLOCATED_THREAD) {
418#ifdef _KERNEL
99df837e 419 zfree(thread_zone, td);
05220613
MD
420#else
421 free(td);
422#endif
423 }
99df837e
MD
424 }
425}
426
427
8ad65e08
MD
428/*
429 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
430 * switch to the idlethread. Switching must occur within a critical
431 * section to avoid races with the scheduling queue.
432 *
433 * We always have full control over our cpu's run queue. Other cpus
434 * that wish to manipulate our queue must use the cpu_*msg() calls to
435 * talk to our cpu, so a critical section is all that is needed and
436 * the result is very, very fast thread switching.
437 *
96728c05
MD
438 * The LWKT scheduler uses a fixed priority model and round-robins at
439 * each priority level. User process scheduling is a totally
440 * different beast and LWKT priorities should not be confused with
441 * user process priorities.
f1d1c3fa 442 *
96728c05
MD
443 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
444 * cleans it up. Note that the td_switch() function cannot do anything that
445 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
446 * the target thread (not the current thread). It's nice to have a scheduler
447 * that does not need the MP lock to work because it allows us to do some
448 * really cool high-performance MP lock optimizations.
69d78e99
MD
449 *
450 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
451 * is not called by the current thread in the preemption case, only when
452 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
453 */
454void
455lwkt_switch(void)
456{
37af14fe
MD
457 globaldata_t gd = mycpu;
458 thread_t td = gd->gd_curthread;
8ad65e08 459 thread_t ntd;
8a8d5d85
MD
460#ifdef SMP
461 int mpheld;
462#endif
8ad65e08 463
69d78e99
MD
464 /*
465 * We had better not be holding any spin locks.
466 */
467 KKASSERT(td->td_spinlocks == 0);
468
46a3f46d 469 /*
27e88a6e
MD
470 * Switching from within a 'fast' (non thread switched) interrupt or IPI
471 * is illegal. However, we may have to do it anyway if we hit a fatal
472 * kernel trap or we have paniced.
473 *
474 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 475 */
27e88a6e
MD
476 if (gd->gd_intr_nesting_level) {
477 int savegdnest;
478 int savegdtrap;
479
480 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
481 panic("lwkt_switch: cannot switch from within "
482 "a fast interrupt, yet, td %p\n", td);
483 } else {
484 savegdnest = gd->gd_intr_nesting_level;
485 savegdtrap = gd->gd_trap_nesting_level;
486 gd->gd_intr_nesting_level = 0;
487 gd->gd_trap_nesting_level = 0;
a7422615
MD
488 if ((td->td_flags & TDF_PANICWARN) == 0) {
489 td->td_flags |= TDF_PANICWARN;
490 printf("Warning: thread switch from interrupt or IPI, "
491 "thread %p (%s)\n", td, td->td_comm);
492#ifdef DDB
493 db_print_backtrace();
494#endif
495 }
27e88a6e
MD
496 lwkt_switch();
497 gd->gd_intr_nesting_level = savegdnest;
498 gd->gd_trap_nesting_level = savegdtrap;
499 return;
500 }
96728c05 501 }
ef0fdad1 502
cb973d15
MD
503 /*
504 * Passive release (used to transition from user to kernel mode
505 * when we block or switch rather then when we enter the kernel).
506 * This function is NOT called if we are switching into a preemption
507 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
508 * our current process designation (if we have one) and become a true
509 * LWKT thread, and may also hand the current process designation to
510 * another process and schedule thread.
cb973d15
MD
511 */
512 if (td->td_release)
513 td->td_release(td);
514
37af14fe 515 crit_enter_gd(gd);
8a8d5d85
MD
516
517#ifdef SMP
518 /*
519 * td_mpcount cannot be used to determine if we currently hold the
520 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
521 * to get the lock, and switch out if it can't. Our ownership of
522 * the actual lock will remain stable while we are in a critical section
523 * (but, of course, another cpu may own or release the lock so the
524 * actual value of mp_lock is not stable).
8a8d5d85
MD
525 */
526 mpheld = MP_LOCK_HELD();
0f7a3396
MD
527#ifdef INVARIANTS
528 if (td->td_cscount) {
529 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
530 td);
531 if (panic_on_cscount)
532 panic("switching while mastering cpusync");
533 }
534#endif
8a8d5d85 535#endif
99df837e
MD
536 if ((ntd = td->td_preempted) != NULL) {
537 /*
538 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
539 * thread. This occurs transparently, whether the preempted thread
540 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
541 * itself).
542 *
543 * We have to setup the MP lock for the original thread after backing
544 * out the adjustment that was made to curthread when the original
545 * was preempted.
99df837e 546 */
26a0694b 547 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 548#ifdef SMP
96728c05 549 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 550 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
551 td, ntd, td->td_mpcount, ntd->td_mpcount);
552 }
8a8d5d85
MD
553 if (ntd->td_mpcount) {
554 td->td_mpcount -= ntd->td_mpcount;
555 KKASSERT(td->td_mpcount >= 0);
556 }
557#endif
26a0694b 558 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
559
560 /*
561 * XXX. The interrupt may have woken a thread up, we need to properly
562 * set the reschedule flag if the originally interrupted thread is at
563 * a lower priority.
564 */
565 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
566 need_lwkt_resched();
8a8d5d85 567 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 568 } else {
4b5f931b
MD
569 /*
570 * Priority queue / round-robin at each priority. Note that user
571 * processes run at a fixed, low priority and the user process
572 * scheduler deals with interactions between user processes
573 * by scheduling and descheduling them from the LWKT queue as
574 * necessary.
8a8d5d85
MD
575 *
576 * We have to adjust the MP lock for the target thread. If we
577 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
578 * thread that does not need the MP lock. If we cannot, we spin
579 * instead of HLT.
580 *
581 * A similar issue exists for the tokens held by the target thread.
582 * If we cannot obtain ownership of the tokens we cannot immediately
583 * schedule the thread.
584 */
585
586 /*
587 * We are switching threads. If there are any pending requests for
588 * tokens we can satisfy all of them here.
4b5f931b 589 */
41a01a4d
MD
590#ifdef SMP
591 if (gd->gd_tokreqbase)
592 lwkt_drain_token_requests();
593#endif
594
8ec60c3f
MD
595 /*
596 * If an LWKT reschedule was requested, well that is what we are
597 * doing now so clear it.
598 */
599 clear_lwkt_resched();
4b5f931b
MD
600again:
601 if (gd->gd_runqmask) {
602 int nq = bsrl(gd->gd_runqmask);
603 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
604 gd->gd_runqmask &= ~(1 << nq);
605 goto again;
606 }
8a8d5d85 607#ifdef SMP
41a01a4d 608 /*
df6b8ba0
MD
609 * THREAD SELECTION FOR AN SMP MACHINE BUILD
610 *
41a01a4d
MD
611 * If the target needs the MP lock and we couldn't get it,
612 * or if the target is holding tokens and we could not
613 * gain ownership of the tokens, continue looking for a
614 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
615 *
616 * NOTE: the mpheld variable invalid after this conditional, it
617 * can change due to both cpu_try_mplock() returning success
618 * AND interactions in lwkt_chktokens() due to the fact that
619 * we are trying to check the mpcount of a thread other then
620 * the current thread. Because of this, if the current thread
621 * is not holding td_mpcount, an IPI indirectly run via
622 * lwkt_chktokens() can obtain and release the MP lock and
623 * cause the core MP lock to be released.
41a01a4d
MD
624 */
625 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
626 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
627 ) {
8a8d5d85 628 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
629
630 mpheld = MP_LOCK_HELD();
631 ntd = NULL;
8a8d5d85
MD
632 while (rqmask) {
633 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 634 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 635 /* spinning due to MP lock being held */
38717797 636#ifdef INVARIANTS
a453459d 637 ++mplock_contention_count;
38717797 638#endif
a453459d 639 /* mplock still not held, 'mpheld' still valid */
41a01a4d 640 continue;
38717797 641 }
a453459d
MD
642
643 /*
644 * mpheld state invalid after chktokens call returns
645 * failure, but the variable is only needed for
646 * the loop.
647 */
38717797 648 if (ntd->td_toks && !lwkt_chktokens(ntd)) {
a453459d 649 /* spinning due to token contention */
38717797 650#ifdef INVARIANTS
a453459d 651 ++token_contention_count;
38717797 652#endif
a453459d 653 mpheld = MP_LOCK_HELD();
41a01a4d 654 continue;
38717797 655 }
41a01a4d 656 break;
8a8d5d85
MD
657 }
658 if (ntd)
659 break;
660 rqmask &= ~(1 << nq);
661 nq = bsrl(rqmask);
662 }
663 if (ntd == NULL) {
a2a5ad0d
MD
664 ntd = &gd->gd_idlethread;
665 ntd->td_flags |= TDF_IDLE_NOHLT;
df6b8ba0 666 goto using_idle_thread;
8a8d5d85 667 } else {
344ad853 668 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
669 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
670 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
671 }
672 } else {
344ad853 673 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
674 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
675 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
676 }
677#else
df6b8ba0
MD
678 /*
679 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
680 * worry about tokens or the BGL.
681 */
344ad853 682 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
683 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
684 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 685#endif
4b5f931b 686 } else {
3c23a41a 687 /*
60f945af
MD
688 * We have nothing to run but only let the idle loop halt
689 * the cpu if there are no pending interrupts.
3c23a41a 690 */
a2a5ad0d 691 ntd = &gd->gd_idlethread;
60f945af 692 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 693 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 694#ifdef SMP
df6b8ba0
MD
695using_idle_thread:
696 /*
697 * The idle thread should not be holding the MP lock unless we
698 * are trapping in the kernel or in a panic. Since we select the
699 * idle thread unconditionally when no other thread is available,
700 * if the MP lock is desired during a panic or kernel trap, we
701 * have to loop in the scheduler until we get it.
702 */
703 if (ntd->td_mpcount) {
704 mpheld = MP_LOCK_HELD();
705 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
706 panic("Idle thread %p was holding the BGL!", ntd);
707 else if (mpheld == 0)
708 goto again;
709 }
a453459d 710#endif
4b5f931b 711 }
f1d1c3fa 712 }
26a0694b
MD
713 KASSERT(ntd->td_pri >= TDPRI_CRIT,
714 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
715
716 /*
717 * Do the actual switch. If the new target does not need the MP lock
718 * and we are holding it, release the MP lock. If the new target requires
719 * the MP lock we have already acquired it for the target.
720 */
721#ifdef SMP
722 if (ntd->td_mpcount == 0 ) {
723 if (MP_LOCK_HELD())
724 cpu_rel_mplock();
725 } else {
a453459d 726 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
727 }
728#endif
94f6d86e
MD
729 if (td != ntd) {
730 ++switch_count;
f1d1c3fa 731 td->td_switch(ntd);
94f6d86e 732 }
37af14fe
MD
733 /* NOTE: current cpu may have changed after switch */
734 crit_exit_quick(td);
8ad65e08
MD
735}
736
b68b7282 737/*
96728c05
MD
738 * Request that the target thread preempt the current thread. Preemption
739 * only works under a specific set of conditions:
b68b7282 740 *
96728c05
MD
741 * - We are not preempting ourselves
742 * - The target thread is owned by the current cpu
743 * - We are not currently being preempted
744 * - The target is not currently being preempted
745 * - We are able to satisfy the target's MP lock requirements (if any).
746 *
747 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
748 * this is called via lwkt_schedule() through the td_preemptable callback.
749 * critpri is the managed critical priority that we should ignore in order
750 * to determine whether preemption is possible (aka usually just the crit
751 * priority of lwkt_schedule() itself).
b68b7282 752 *
26a0694b
MD
753 * XXX at the moment we run the target thread in a critical section during
754 * the preemption in order to prevent the target from taking interrupts
755 * that *WE* can't. Preemption is strictly limited to interrupt threads
756 * and interrupt-like threads, outside of a critical section, and the
757 * preempted source thread will be resumed the instant the target blocks
758 * whether or not the source is scheduled (i.e. preemption is supposed to
759 * be as transparent as possible).
4b5f931b 760 *
8a8d5d85
MD
761 * The target thread inherits our MP count (added to its own) for the
762 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
763 * MP lock during the preemption. Therefore, any preempting targets must be
764 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
765 * out of sync with the physical mp_lock, but we do not have to preserve
766 * the original ownership of the lock if it was out of synch (that is, we
767 * can leave it synchronized on return).
b68b7282
MD
768 */
769void
96728c05 770lwkt_preempt(thread_t ntd, int critpri)
b68b7282 771{
46a3f46d 772 struct globaldata *gd = mycpu;
0a3f9b47 773 thread_t td;
8a8d5d85
MD
774#ifdef SMP
775 int mpheld;
57c254db 776 int savecnt;
8a8d5d85 777#endif
b68b7282 778
26a0694b 779 /*
96728c05
MD
780 * The caller has put us in a critical section. We can only preempt
781 * if the caller of the caller was not in a critical section (basically
0a3f9b47 782 * a local interrupt), as determined by the 'critpri' parameter.
96728c05
MD
783 *
784 * YYY The target thread must be in a critical section (else it must
785 * inherit our critical section? I dunno yet).
41a01a4d
MD
786 *
787 * Any tokens held by the target may not be held by thread(s) being
788 * preempted. We take the easy way out and do not preempt if
789 * the target is holding tokens.
0a3f9b47
MD
790 *
791 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
792 */
793 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 794
0a3f9b47 795 td = gd->gd_curthread;
0a3f9b47 796 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
797 ++preempt_miss;
798 return;
799 }
96728c05
MD
800 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
801 ++preempt_miss;
8ec60c3f 802 need_lwkt_resched();
96728c05
MD
803 return;
804 }
805#ifdef SMP
46a3f46d 806 if (ntd->td_gd != gd) {
96728c05 807 ++preempt_miss;
8ec60c3f 808 need_lwkt_resched();
96728c05
MD
809 return;
810 }
811#endif
41a01a4d
MD
812 /*
813 * Take the easy way out and do not preempt if the target is holding
814 * one or more tokens. We could test whether the thread(s) being
815 * preempted interlock against the target thread's tokens and whether
816 * we can get all the target thread's tokens, but this situation
817 * should not occur very often so its easier to simply not preempt.
818 */
819 if (ntd->td_toks != NULL) {
820 ++preempt_miss;
8ec60c3f 821 need_lwkt_resched();
41a01a4d
MD
822 return;
823 }
26a0694b
MD
824 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
825 ++preempt_weird;
8ec60c3f 826 need_lwkt_resched();
26a0694b
MD
827 return;
828 }
829 if (ntd->td_preempted) {
4b5f931b 830 ++preempt_hit;
8ec60c3f 831 need_lwkt_resched();
26a0694b 832 return;
b68b7282 833 }
8a8d5d85 834#ifdef SMP
a2a5ad0d
MD
835 /*
836 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
837 * to or from the MP lock. In this case td_mpcount will be pre-disposed
838 * (non-zero) but not actually synchronized with the actual state of the
839 * lock. We can use it to imply an MP lock requirement for the
840 * preemption but we cannot use it to test whether we hold the MP lock
841 * or not.
a2a5ad0d 842 */
96728c05 843 savecnt = td->td_mpcount;
71ef2f5c 844 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
845 ntd->td_mpcount += td->td_mpcount;
846 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
847 ntd->td_mpcount -= td->td_mpcount;
848 ++preempt_miss;
8ec60c3f 849 need_lwkt_resched();
8a8d5d85
MD
850 return;
851 }
852#endif
26a0694b 853
8ec60c3f
MD
854 /*
855 * Since we are able to preempt the current thread, there is no need to
856 * call need_lwkt_resched().
857 */
26a0694b
MD
858 ++preempt_hit;
859 ntd->td_preempted = td;
860 td->td_flags |= TDF_PREEMPT_LOCK;
861 td->td_switch(ntd);
862 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
863#ifdef SMP
864 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
865 mpheld = MP_LOCK_HELD();
866 if (mpheld && td->td_mpcount == 0)
96728c05 867 cpu_rel_mplock();
71ef2f5c 868 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
869 panic("lwkt_preempt(): MP lock was not held through");
870#endif
26a0694b
MD
871 ntd->td_preempted = NULL;
872 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
873}
874
f1d1c3fa
MD
875/*
876 * Yield our thread while higher priority threads are pending. This is
877 * typically called when we leave a critical section but it can be safely
878 * called while we are in a critical section.
879 *
880 * This function will not generally yield to equal priority threads but it
881 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 882 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
883 * lwkt_yield_quick().
884 *
235957ed 885 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
886 * came along but was blocked and made pending.
887 *
f1d1c3fa
MD
888 * (self contained on a per cpu basis)
889 */
890void
891lwkt_yield_quick(void)
892{
7966cb69
MD
893 globaldata_t gd = mycpu;
894 thread_t td = gd->gd_curthread;
ef0fdad1 895
a2a5ad0d 896 /*
235957ed 897 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
898 * it with a non-zero cpl then we might not wind up calling splz after
899 * a task switch when the critical section is exited even though the
46a3f46d 900 * new task could accept the interrupt.
a2a5ad0d
MD
901 *
902 * XXX from crit_exit() only called after last crit section is released.
903 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
904 *
905 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
906 * except for this special case, we MUST call splz() here to handle any
907 * pending ints, particularly after we switch, or we might accidently
908 * halt the cpu with interrupts pending.
a2a5ad0d 909 */
46a3f46d 910 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 911 splz();
f1d1c3fa
MD
912
913 /*
914 * YYY enabling will cause wakeup() to task-switch, which really
915 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
916 * preemption and MP without actually doing preemption or MP, because a
917 * lot of code assumes that wakeup() does not block.
f1d1c3fa 918 */
46a3f46d
MD
919 if (untimely_switch && td->td_nest_count == 0 &&
920 gd->gd_intr_nesting_level == 0
921 ) {
37af14fe 922 crit_enter_quick(td);
f1d1c3fa
MD
923 /*
924 * YYY temporary hacks until we disassociate the userland scheduler
925 * from the LWKT scheduler.
926 */
927 if (td->td_flags & TDF_RUNQ) {
928 lwkt_switch(); /* will not reenter yield function */
929 } else {
37af14fe 930 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 931 lwkt_switch(); /* will not reenter yield function */
37af14fe 932 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 933 }
7966cb69 934 crit_exit_noyield(td);
f1d1c3fa 935 }
f1d1c3fa
MD
936}
937
8ad65e08 938/*
f1d1c3fa 939 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 940 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
941 * the crit_exit() call in lwkt_switch().
942 *
943 * (self contained on a per cpu basis)
8ad65e08
MD
944 */
945void
f1d1c3fa 946lwkt_yield(void)
8ad65e08 947{
37af14fe 948 lwkt_schedule_self(curthread);
f1d1c3fa
MD
949 lwkt_switch();
950}
951
8ad65e08 952/*
f1d1c3fa
MD
953 * Generic schedule. Possibly schedule threads belonging to other cpus and
954 * deal with threads that might be blocked on a wait queue.
955 *
0a3f9b47
MD
956 * We have a little helper inline function which does additional work after
957 * the thread has been enqueued, including dealing with preemption and
958 * setting need_lwkt_resched() (which prevents the kernel from returning
959 * to userland until it has processed higher priority threads).
8ad65e08 960 */
0a3f9b47
MD
961static __inline
962void
8ec60c3f 963_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
0a3f9b47
MD
964{
965 if (ntd->td_preemptable) {
966 ntd->td_preemptable(ntd, cpri); /* YYY +token */
8ec60c3f
MD
967 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
968 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
969 ) {
970 need_lwkt_resched();
0a3f9b47
MD
971 }
972}
973
8ad65e08
MD
974void
975lwkt_schedule(thread_t td)
976{
37af14fe
MD
977 globaldata_t mygd = mycpu;
978
41a01a4d 979 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 980 crit_enter_gd(mygd);
344ad853 981 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
37af14fe 982 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
983 _lwkt_enqueue(td);
984 } else {
985 lwkt_wait_t w;
986
987 /*
988 * If the thread is on a wait list we have to send our scheduling
989 * request to the owner of the wait structure. Otherwise we send
990 * the scheduling request to the cpu owning the thread. Races
991 * are ok, the target will forward the message as necessary (the
992 * message may chase the thread around before it finally gets
993 * acted upon).
994 *
995 * (remember, wait structures use stable storage)
0a3f9b47 996 *
0c453950
MD
997 * NOTE: we have to account for the number of critical sections
998 * under our control when calling _lwkt_schedule_post() so it
999 * can figure out whether preemption is allowed.
1000 *
1001 * NOTE: The wait structure algorithms are a mess and need to be
1002 * rewritten.
1003 *
1004 * NOTE: We cannot safely acquire or release a token, even
1005 * non-blocking, because this routine may be called in the context
1006 * of a thread already holding the token and thus not provide any
1007 * interlock protection. We cannot safely manipulate the td_toks
1008 * list for the same reason. Instead we depend on our critical
1009 * section if the token is owned by our cpu.
f1d1c3fa
MD
1010 */
1011 if ((w = td->td_wait) != NULL) {
0c453950 1012 if (w->wa_token.t_cpu == mygd) {
f1d1c3fa
MD
1013 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1014 --w->wa_count;
1015 td->td_wait = NULL;
0f7a3396 1016#ifdef SMP
8ec60c3f 1017 if (td->td_gd == mygd) {
f1d1c3fa 1018 _lwkt_enqueue(td);
8ec60c3f 1019 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 1020 } else {
b8a98473 1021 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
f1d1c3fa 1022 }
0f7a3396
MD
1023#else
1024 _lwkt_enqueue(td);
8ec60c3f 1025 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 1026#endif
f1d1c3fa 1027 } else {
b8a98473
MD
1028#ifdef SMP
1029 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc1_t)lwkt_schedule, td);
1030#else
1031 panic("bad token %p", &w->wa_token);
1032#endif
f1d1c3fa
MD
1033 }
1034 } else {
1035 /*
1036 * If the wait structure is NULL and we own the thread, there
1037 * is no race (since we are in a critical section). If we
1038 * do not own the thread there might be a race but the
1039 * target cpu will deal with it.
1040 */
0f7a3396 1041#ifdef SMP
37af14fe 1042 if (td->td_gd == mygd) {
f1d1c3fa 1043 _lwkt_enqueue(td);
8ec60c3f 1044 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 1045 } else {
b8a98473 1046 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
f1d1c3fa 1047 }
0f7a3396
MD
1048#else
1049 _lwkt_enqueue(td);
8ec60c3f 1050 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 1051#endif
f1d1c3fa 1052 }
8ad65e08 1053 }
37af14fe 1054 crit_exit_gd(mygd);
8ad65e08
MD
1055}
1056
d9eea1a5
MD
1057/*
1058 * Managed acquisition. This code assumes that the MP lock is held for
1059 * the tdallq operation and that the thread has been descheduled from its
1060 * original cpu. We also have to wait for the thread to be entirely switched
1061 * out on its original cpu (this is usually fast enough that we never loop)
1062 * since the LWKT system does not have to hold the MP lock while switching
1063 * and the target may have released it before switching.
1064 */
a2a5ad0d
MD
1065void
1066lwkt_acquire(thread_t td)
1067{
37af14fe
MD
1068 globaldata_t gd;
1069 globaldata_t mygd;
a2a5ad0d
MD
1070
1071 gd = td->td_gd;
37af14fe 1072 mygd = mycpu;
35238fa5 1073 cpu_lfence();
a2a5ad0d 1074 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
c1102e9f 1075 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) /* XXX spin */
35238fa5 1076 cpu_lfence();
37af14fe
MD
1077 if (gd != mygd) {
1078 crit_enter_gd(mygd);
a2a5ad0d 1079 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
37af14fe
MD
1080 td->td_gd = mygd;
1081 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
1082 crit_exit_gd(mygd);
a2a5ad0d
MD
1083 }
1084}
1085
f1d1c3fa
MD
1086/*
1087 * Generic deschedule. Descheduling threads other then your own should be
1088 * done only in carefully controlled circumstances. Descheduling is
1089 * asynchronous.
1090 *
1091 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1092 */
1093void
1094lwkt_deschedule(thread_t td)
1095{
f1d1c3fa 1096 crit_enter();
b8a98473 1097#ifdef SMP
f1d1c3fa
MD
1098 if (td == curthread) {
1099 _lwkt_dequeue(td);
1100 } else {
a72187e9 1101 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1102 _lwkt_dequeue(td);
1103 } else {
b8a98473 1104 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1105 }
1106 }
b8a98473
MD
1107#else
1108 _lwkt_dequeue(td);
1109#endif
f1d1c3fa
MD
1110 crit_exit();
1111}
1112
4b5f931b
MD
1113/*
1114 * Set the target thread's priority. This routine does not automatically
1115 * switch to a higher priority thread, LWKT threads are not designed for
1116 * continuous priority changes. Yield if you want to switch.
1117 *
1118 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1119 * of the td_pri field. The specified priority may also indicate zero or
1120 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1121 *
1122 * Note that we requeue the thread whether it winds up on a different runq
1123 * or not. uio_yield() depends on this and the routine is not normally
1124 * called with the same priority otherwise.
4b5f931b
MD
1125 */
1126void
1127lwkt_setpri(thread_t td, int pri)
1128{
26a0694b 1129 KKASSERT(pri >= 0);
a72187e9 1130 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1131 crit_enter();
1132 if (td->td_flags & TDF_RUNQ) {
1133 _lwkt_dequeue(td);
1134 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1135 _lwkt_enqueue(td);
1136 } else {
1137 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1138 }
1139 crit_exit();
1140}
1141
1142void
1143lwkt_setpri_self(int pri)
1144{
1145 thread_t td = curthread;
1146
4b5f931b
MD
1147 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1148 crit_enter();
1149 if (td->td_flags & TDF_RUNQ) {
1150 _lwkt_dequeue(td);
1151 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1152 _lwkt_enqueue(td);
1153 } else {
1154 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1155 }
1156 crit_exit();
1157}
1158
8ec60c3f
MD
1159/*
1160 * Determine if there is a runnable thread at a higher priority then
1161 * the current thread. lwkt_setpri() does not check this automatically.
1162 * Return 1 if there is, 0 if there isn't.
1163 *
1164 * Example: if bit 31 of runqmask is set and the current thread is priority
1165 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1166 *
1167 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1168 * up comparing against 0xffffffff, a comparison that will always be false.
1169 */
1170int
1171lwkt_checkpri_self(void)
1172{
1173 globaldata_t gd = mycpu;
1174 thread_t td = gd->gd_curthread;
1175 int nq = td->td_pri & TDPRI_MASK;
1176
1177 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1178 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1179 return(1);
1180 ++nq;
1181 }
1182 return(0);
1183}
1184
5d21b981
MD
1185/*
1186 * Migrate the current thread to the specified cpu. The BGL must be held
1187 * (for the gd_tdallq manipulation XXX). This is accomplished by
1188 * descheduling ourselves from the current cpu, moving our thread to the
1189 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1190 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1191 */
3d28ff59 1192#ifdef SMP
5d21b981 1193static void lwkt_setcpu_remote(void *arg);
3d28ff59 1194#endif
5d21b981
MD
1195
1196void
1197lwkt_setcpu_self(globaldata_t rgd)
1198{
1199#ifdef SMP
1200 thread_t td = curthread;
1201
1202 if (td->td_gd != rgd) {
1203 crit_enter_quick(td);
1204 td->td_flags |= TDF_MIGRATING;
1205 lwkt_deschedule_self(td);
1206 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1207 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
b8a98473 1208 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1209 lwkt_switch();
1210 /* we are now on the target cpu */
1211 crit_exit_quick(td);
1212 }
1213#endif
1214}
1215
1216/*
1217 * Remote IPI for cpu migration (called while in a critical section so we
1218 * do not have to enter another one). The thread has already been moved to
1219 * our cpu's allq, but we must wait for the thread to be completely switched
1220 * out on the originating cpu before we schedule it on ours or the stack
1221 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1222 * change to main memory.
1223 *
1224 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1225 * against wakeups. It is best if this interface is used only when there
1226 * are no pending events that might try to schedule the thread.
1227 */
3d28ff59 1228#ifdef SMP
5d21b981
MD
1229static void
1230lwkt_setcpu_remote(void *arg)
1231{
1232 thread_t td = arg;
1233 globaldata_t gd = mycpu;
1234
c1102e9f 1235 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK))
35238fa5 1236 cpu_lfence();
5d21b981 1237 td->td_gd = gd;
35238fa5 1238 cpu_sfence();
5d21b981 1239 td->td_flags &= ~TDF_MIGRATING;
344ad853 1240 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
5d21b981
MD
1241 _lwkt_enqueue(td);
1242}
3d28ff59 1243#endif
5d21b981 1244
553ea3c8 1245struct lwp *
4b5f931b
MD
1246lwkt_preempted_proc(void)
1247{
73e4f7b9 1248 thread_t td = curthread;
4b5f931b
MD
1249 while (td->td_preempted)
1250 td = td->td_preempted;
553ea3c8 1251 return(td->td_lwp);
4b5f931b
MD
1252}
1253
f1d1c3fa 1254/*
41a01a4d
MD
1255 * Block on the specified wait queue until signaled. A generation number
1256 * must be supplied to interlock the wait queue. The function will
1257 * return immediately if the generation number does not match the wait
1258 * structure's generation number.
f1d1c3fa
MD
1259 */
1260void
ae8050a4 1261lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
1262{
1263 thread_t td = curthread;
41a01a4d 1264 lwkt_tokref ilock;
f1d1c3fa 1265
41a01a4d
MD
1266 lwkt_gettoken(&ilock, &w->wa_token);
1267 crit_enter();
ae8050a4 1268 if (w->wa_gen == *gen) {
f1d1c3fa 1269 _lwkt_dequeue(td);
344ad853 1270 td->td_flags |= TDF_BLOCKQ;
f1d1c3fa
MD
1271 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1272 ++w->wa_count;
1273 td->td_wait = w;
ae8050a4 1274 td->td_wmesg = wmesg;
f1d1c3fa 1275 lwkt_switch();
344ad853
MD
1276 KKASSERT((td->td_flags & TDF_BLOCKQ) == 0);
1277 td->td_wmesg = NULL;
8ad65e08 1278 }
41a01a4d 1279 crit_exit();
ae8050a4 1280 *gen = w->wa_gen;
41a01a4d 1281 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1282}
1283
1284/*
1285 * Signal a wait queue. We gain ownership of the wait queue in order to
1286 * signal it. Once a thread is removed from the wait queue we have to
1287 * deal with the cpu owning the thread.
1288 *
1289 * Note: alternatively we could message the target cpu owning the wait
1290 * queue. YYY implement as sysctl.
1291 */
1292void
ece04fd0 1293lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1294{
1295 thread_t td;
41a01a4d 1296 lwkt_tokref ilock;
f1d1c3fa 1297
41a01a4d 1298 lwkt_gettoken(&ilock, &w->wa_token);
f1d1c3fa 1299 ++w->wa_gen;
41a01a4d 1300 crit_enter();
ece04fd0
MD
1301 if (count < 0)
1302 count = w->wa_count;
f1d1c3fa
MD
1303 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1304 --count;
1305 --w->wa_count;
344ad853 1306 KKASSERT(td->td_flags & TDF_BLOCKQ);
f1d1c3fa 1307 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
344ad853 1308 td->td_flags &= ~TDF_BLOCKQ;
f1d1c3fa 1309 td->td_wait = NULL;
344ad853 1310 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
b8a98473 1311#ifdef SMP
a72187e9 1312 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1313 _lwkt_enqueue(td);
1314 } else {
b8a98473 1315 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
f1d1c3fa 1316 }
b8a98473
MD
1317#else
1318 _lwkt_enqueue(td);
1319#endif
f1d1c3fa 1320 }
41a01a4d
MD
1321 crit_exit();
1322 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1323}
1324
99df837e
MD
1325/*
1326 * Create a kernel process/thread/whatever. It shares it's address space
1327 * with proc0 - ie: kernel only.
1328 *
365fa13f
MD
1329 * NOTE! By default new threads are created with the MP lock held. A
1330 * thread which does not require the MP lock should release it by calling
1331 * rel_mplock() at the start of the new thread.
99df837e
MD
1332 */
1333int
1334lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1335 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1336 const char *fmt, ...)
99df837e 1337{
73e4f7b9 1338 thread_t td;
e2565a42 1339 __va_list ap;
99df837e 1340
d3d32139
MD
1341 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1342 tdflags | TDF_VERBOSE);
a2a5ad0d
MD
1343 if (tdp)
1344 *tdp = td;
709799ea 1345 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1346
1347 /*
1348 * Set up arg0 for 'ps' etc
1349 */
e2565a42 1350 __va_start(ap, fmt);
99df837e 1351 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1352 __va_end(ap);
99df837e
MD
1353
1354 /*
1355 * Schedule the thread to run
1356 */
ef0fdad1
MD
1357 if ((td->td_flags & TDF_STOPREQ) == 0)
1358 lwkt_schedule(td);
1359 else
1360 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1361 return 0;
1362}
1363
2d93b37a 1364/*
2d93b37a
MD
1365 * kthread_* is specific to the kernel and is not needed by userland.
1366 */
1367#ifdef _KERNEL
1368
99df837e
MD
1369/*
1370 * Destroy an LWKT thread. Warning! This function is not called when
1371 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1372 * uses a different reaping mechanism.
1373 */
1374void
1375lwkt_exit(void)
1376{
1377 thread_t td = curthread;
8826f33a 1378 globaldata_t gd;
99df837e
MD
1379
1380 if (td->td_flags & TDF_VERBOSE)
1381 printf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1382 caps_exit(td);
37af14fe
MD
1383 crit_enter_quick(td);
1384 lwkt_deschedule_self(td);
8826f33a
MD
1385 gd = mycpu;
1386 KKASSERT(gd == td->td_gd);
1387 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1388 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1389 ++gd->gd_tdfreecount;
1390 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1391 }
99df837e
MD
1392 cpu_thread_exit();
1393}
1394
2d93b37a
MD
1395#endif /* _KERNEL */
1396
1397void
1398crit_panic(void)
1399{
1400 thread_t td = curthread;
1401 int lpri = td->td_pri;
1402
1403 td->td_pri = 0;
1404 panic("td_pri is/would-go negative! %p %d", td, lpri);
1405}
1406
d165e668
MD
1407#ifdef SMP
1408
bd8015ca
MD
1409/*
1410 * Called from debugger/panic on cpus which have been stopped. We must still
1411 * process the IPIQ while stopped, even if we were stopped while in a critical
1412 * section (XXX).
1413 *
1414 * If we are dumping also try to process any pending interrupts. This may
1415 * or may not work depending on the state of the cpu at the point it was
1416 * stopped.
1417 */
1418void
1419lwkt_smp_stopped(void)
1420{
1421 globaldata_t gd = mycpu;
1422
1423 crit_enter_gd(gd);
1424 if (dumping) {
1425 lwkt_process_ipiq();
1426 splz();
1427 } else {
1428 lwkt_process_ipiq();
1429 }
1430 crit_exit_gd(gd);
1431}
1432
d165e668 1433#endif