kernel: Make SMP support default (and non-optional).
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
b12defdc 2 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 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:
60f60350 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.
60f60350 20 *
8c10bfcf
MD
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.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
b37f18d6 47#include <sys/kinfo.h>
8ad65e08 48#include <sys/queue.h>
7d0bac62 49#include <sys/sysctl.h>
99df837e 50#include <sys/kthread.h>
f1d1c3fa 51#include <machine/cpu.h>
99df837e 52#include <sys/lock.h>
9d265729 53#include <sys/spinlock.h>
57aa743c 54#include <sys/ktr.h>
9d265729
MD
55
56#include <sys/thread2.h>
57#include <sys/spinlock2.h>
684a93c4 58#include <sys/mplock2.h>
f1d1c3fa 59
8c72e3d5
AH
60#include <sys/dsched.h>
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>
7d0bac62 70
99df837e 71#include <machine/stdarg.h>
96728c05 72#include <machine/smp.h>
99df837e 73
d850923c
AE
74#if !defined(KTR_CTXSW)
75#define KTR_CTXSW KTR_ALL
76#endif
77KTR_INFO_MASTER(ctxsw);
5bf48697
AE
78KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td);
79KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td);
80KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm);
81KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td);
1541028a 82
40aaf5fc
NT
83static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
84
0f7a3396
MD
85#ifdef INVARIANTS
86static int panic_on_cscount = 0;
87#endif
05220613
MD
88static __int64_t switch_count = 0;
89static __int64_t preempt_hit = 0;
90static __int64_t preempt_miss = 0;
91static __int64_t preempt_weird = 0;
fb0f29c4 92static int lwkt_use_spin_port;
40aaf5fc 93static struct objcache *thread_cache;
05220613 94
e381e77c 95static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
cc9b6223 96static void lwkt_setcpu_remote(void *arg);
e381e77c 97
0855a2af
JG
98extern void cpu_heavy_restore(void);
99extern void cpu_lwkt_restore(void);
100extern void cpu_kthread_restore(void);
101extern void cpu_idle_restore(void);
102
fb0f29c4
MD
103/*
104 * We can make all thread ports use the spin backend instead of the thread
105 * backend. This should only be set to debug the spin backend.
106 */
107TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
108
0f7a3396 109#ifdef INVARIANTS
0c52fa62
SG
110SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
111 "Panic if attempting to switch lwkt's while mastering cpusync");
0f7a3396 112#endif
0c52fa62
SG
113SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
114 "Number of switched threads");
9733f757 115SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
0c52fa62 116 "Successful preemption events");
9733f757 117SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
0c52fa62
SG
118 "Failed preemption events");
119SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
120 "Number of preempted threads.");
b12defdc 121static int fairq_enable = 0;
2a418930
MD
122SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
123 &fairq_enable, 0, "Turn on fairq priority accumulators");
85946b6c 124static int fairq_bypass = -1;
b12defdc
MD
125SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
126 &fairq_bypass, 0, "Allow fairq to bypass td on token failure");
127extern int lwkt_sched_debug;
128int lwkt_sched_debug = 0;
129SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
130 &lwkt_sched_debug, 0, "Scheduler debug");
2a418930
MD
131static int lwkt_spin_loops = 10;
132SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
b12defdc
MD
133 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
134static int lwkt_spin_reseq = 0;
135SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW,
136 &lwkt_spin_reseq, 0, "Scheduler resequencer enable");
137static int lwkt_spin_monitor = 0;
138SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW,
139 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait");
d5b2d319
MD
140static int lwkt_spin_fatal = 0; /* disabled */
141SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
142 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
fbc024e4 143static int preempt_enable = 1;
2a418930
MD
144SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
145 &preempt_enable, 0, "Enable preemption");
7b234d8c 146static int lwkt_cache_threads = 0;
765b1ae0
MD
147SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
148 &lwkt_cache_threads, 0, "thread+kstack cache");
fbc024e4 149
2a418930
MD
150static __cachealign int lwkt_cseq_rindex;
151static __cachealign int lwkt_cseq_windex;
05220613 152
4b5f931b
MD
153/*
154 * These helper procedures handle the runq, they can only be called from
155 * within a critical section.
75cdbe6c
MD
156 *
157 * WARNING! Prior to SMP being brought up it is possible to enqueue and
158 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
159 * instead of 'mycpu' when referencing the globaldata structure. Once
160 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 161 */
f1d1c3fa
MD
162static __inline
163void
164_lwkt_dequeue(thread_t td)
165{
166 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 167 struct globaldata *gd = td->td_gd;
4b5f931b 168
f1d1c3fa 169 td->td_flags &= ~TDF_RUNQ;
f9235b6d 170 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
de4d4cb0 171 --gd->gd_tdrunqcount;
f9235b6d 172 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
2a418930 173 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
174 }
175}
176
f9235b6d
MD
177/*
178 * Priority enqueue.
179 *
d992c377
MD
180 * There are a limited number of lwkt threads runnable since user
181 * processes only schedule one at a time per cpu. However, there can
182 * be many user processes in kernel mode exiting from a tsleep() which
e3e6be1f 183 * become runnable.
d992c377
MD
184 *
185 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and
186 * will ignore user priority. This is to ensure that user threads in
187 * kernel mode get cpu at some point regardless of what the user
188 * scheduler thinks.
f9235b6d 189 */
f1d1c3fa
MD
190static __inline
191void
192_lwkt_enqueue(thread_t td)
193{
f9235b6d
MD
194 thread_t xtd;
195
7f5d7ed7 196 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 197 struct globaldata *gd = td->td_gd;
4b5f931b 198
f1d1c3fa 199 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
200 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
201 if (xtd == NULL) {
85946b6c
MD
202 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
203 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
f9235b6d 204 } else {
e3e6be1f
MD
205 /*
206 * NOTE: td_upri - higher numbers more desireable, same sense
207 * as td_pri (typically reversed from lwp_upri).
208 *
209 * In the equal priority case we want the best selection
210 * at the beginning so the less desireable selections know
211 * that they have to setrunqueue/go-to-another-cpu, even
212 * though it means switching back to the 'best' selection.
213 * This also avoids degenerate situations when many threads
214 * are runnable or waking up at the same time.
215 *
216 * If upri matches exactly place at end/round-robin.
217 */
d992c377 218 while (xtd &&
e3e6be1f 219 (xtd->td_pri >= td->td_pri ||
d992c377 220 (xtd->td_pri == td->td_pri &&
e3e6be1f 221 xtd->td_upri >= td->td_upri))) {
85946b6c 222 xtd = TAILQ_NEXT(xtd, td_threadq);
d992c377 223 }
85946b6c
MD
224 if (xtd)
225 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
226 else
227 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
f9235b6d 228 }
de4d4cb0 229 ++gd->gd_tdrunqcount;
b12defdc
MD
230
231 /*
85946b6c 232 * Request a LWKT reschedule if we are now at the head of the queue.
b12defdc 233 */
85946b6c
MD
234 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
235 need_lwkt_resched();
f1d1c3fa
MD
236 }
237}
8ad65e08 238
40aaf5fc
NT
239static __boolean_t
240_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
241{
242 struct thread *td = (struct thread *)obj;
243
244 td->td_kstack = NULL;
245 td->td_kstack_size = 0;
246 td->td_flags = TDF_ALLOCATED_THREAD;
4643740a 247 td->td_mpflags = 0;
40aaf5fc
NT
248 return (1);
249}
250
251static void
252_lwkt_thread_dtor(void *obj, void *privdata)
253{
254 struct thread *td = (struct thread *)obj;
255
256 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
257 ("_lwkt_thread_dtor: not allocated from objcache"));
258 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
259 td->td_kstack_size > 0,
260 ("_lwkt_thread_dtor: corrupted stack"));
261 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
7b234d8c
MD
262 td->td_kstack = NULL;
263 td->td_flags = 0;
40aaf5fc
NT
264}
265
266/*
267 * Initialize the lwkt s/system.
765b1ae0 268 *
7b234d8c
MD
269 * Nominally cache up to 32 thread + kstack structures. Cache more on
270 * systems with a lot of cpu cores.
40aaf5fc
NT
271 */
272void
273lwkt_init(void)
274{
765b1ae0 275 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
7b234d8c
MD
276 if (lwkt_cache_threads == 0) {
277 lwkt_cache_threads = ncpus * 4;
278 if (lwkt_cache_threads < 32)
279 lwkt_cache_threads = 32;
280 }
765b1ae0
MD
281 thread_cache = objcache_create_mbacked(
282 M_THREAD, sizeof(struct thread),
2fce2579 283 0, lwkt_cache_threads,
765b1ae0 284 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
285}
286
37af14fe
MD
287/*
288 * Schedule a thread to run. As the current thread we can always safely
289 * schedule ourselves, and a shortcut procedure is provided for that
290 * function.
291 *
292 * (non-blocking, self contained on a per cpu basis)
293 */
294void
295lwkt_schedule_self(thread_t td)
296{
cfaeae2a 297 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 298 crit_enter_quick(td);
f9235b6d
MD
299 KASSERT(td != &td->td_gd->gd_idlethread,
300 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
4643740a
MD
301 KKASSERT(td->td_lwp == NULL ||
302 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
37af14fe 303 _lwkt_enqueue(td);
37af14fe
MD
304 crit_exit_quick(td);
305}
306
307/*
308 * Deschedule a thread.
309 *
310 * (non-blocking, self contained on a per cpu basis)
311 */
312void
313lwkt_deschedule_self(thread_t td)
314{
315 crit_enter_quick(td);
37af14fe
MD
316 _lwkt_dequeue(td);
317 crit_exit_quick(td);
318}
319
8ad65e08
MD
320/*
321 * LWKTs operate on a per-cpu basis
322 *
73e4f7b9 323 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
324 */
325void
326lwkt_gdinit(struct globaldata *gd)
327{
f9235b6d 328 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 329 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
330}
331
7d0bac62
MD
332/*
333 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
334 * or LWKT start address before it can be scheduled. If the target cpu is
335 * -1 the thread will be created on the current cpu.
0cfcada1
MD
336 *
337 * If you intend to create a thread without a process context this function
338 * does everything except load the startup and switcher function.
7d0bac62
MD
339 */
340thread_t
d3d32139 341lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 342{
d2d8515b 343 static int cpu_rotator;
c070746a 344 globaldata_t gd = mycpu;
99df837e 345 void *stack;
7d0bac62 346
c070746a
MD
347 /*
348 * If static thread storage is not supplied allocate a thread. Reuse
349 * a cached free thread if possible. gd_freetd is used to keep an exiting
350 * thread intact through the exit.
351 */
ef0fdad1 352 if (td == NULL) {
cf709dd2
MD
353 crit_enter_gd(gd);
354 if ((td = gd->gd_freetd) != NULL) {
355 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
356 TDF_RUNQ)) == 0);
c070746a 357 gd->gd_freetd = NULL;
cf709dd2 358 } else {
c070746a 359 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
360 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
361 TDF_RUNQ)) == 0);
362 }
363 crit_exit_gd(gd);
40aaf5fc 364 KASSERT((td->td_flags &
2af9d75d
MD
365 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
366 TDF_ALLOCATED_THREAD,
40aaf5fc
NT
367 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
368 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 369 }
c070746a
MD
370
371 /*
372 * Try to reuse cached stack.
373 */
f470d0c8
MD
374 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
375 if (flags & TDF_ALLOCATED_STACK) {
e4846942 376 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
377 stack = NULL;
378 }
379 }
380 if (stack == NULL) {
e40cfbd7 381 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 382 flags |= TDF_ALLOCATED_STACK;
99df837e 383 }
d2d8515b
MD
384 if (cpu < 0) {
385 cpu = ++cpu_rotator;
386 cpu_ccfence();
387 cpu %= ncpus;
388 }
389 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 390 return(td);
7d0bac62
MD
391}
392
393/*
394 * Initialize a preexisting thread structure. This function is used by
395 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
396 *
f8c3996b
MD
397 * All threads start out in a critical section at a priority of
398 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
399 * appropriate. This function may send an IPI message when the
400 * requested cpu is not the current cpu and consequently gd_tdallq may
401 * not be initialized synchronously from the point of view of the originating
402 * cpu.
403 *
404 * NOTE! we have to be careful in regards to creating threads for other cpus
405 * if SMP has not yet been activated.
7d0bac62 406 */
75cdbe6c
MD
407static void
408lwkt_init_thread_remote(void *arg)
409{
410 thread_t td = arg;
411
52eedfb5
MD
412 /*
413 * Protected by critical section held by IPI dispatch
414 */
75cdbe6c
MD
415 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
416}
417
fdce8919
MD
418/*
419 * lwkt core thread structural initialization.
420 *
421 * NOTE: All threads are initialized as mpsafe threads.
422 */
7d0bac62 423void
f470d0c8
MD
424lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
425 struct globaldata *gd)
7d0bac62 426{
37af14fe
MD
427 globaldata_t mygd = mycpu;
428
99df837e
MD
429 bzero(td, sizeof(struct thread));
430 td->td_kstack = stack;
f470d0c8 431 td->td_kstack_size = stksize;
d3d32139 432 td->td_flags = flags;
4643740a 433 td->td_mpflags = 0;
26a0694b 434 td->td_gd = gd;
f9235b6d
MD
435 td->td_pri = TDPRI_KERN_DAEMON;
436 td->td_critcount = 1;
54341a3b 437 td->td_toks_have = NULL;
3b998fa9 438 td->td_toks_stop = &td->td_toks_base;
392cd266 439 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
d636cd48 440 lwkt_initport_spin(&td->td_msgport, td);
fb0f29c4
MD
441 else
442 lwkt_initport_thread(&td->td_msgport, td);
99df837e 443 pmap_init_thread(td);
5d21b981
MD
444 /*
445 * Normally initializing a thread for a remote cpu requires sending an
446 * IPI. However, the idlethread is setup before the other cpus are
447 * activated so we have to treat it as a special case. XXX manipulation
448 * of gd_tdallq requires the BGL.
449 */
450 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 451 crit_enter_gd(mygd);
75cdbe6c 452 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 453 crit_exit_gd(mygd);
75cdbe6c 454 } else {
2db3b277 455 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 456 }
8c72e3d5 457 dsched_new_thread(td);
73e4f7b9
MD
458}
459
460void
461lwkt_set_comm(thread_t td, const char *ctl, ...)
462{
e2565a42 463 __va_list va;
73e4f7b9 464
e2565a42 465 __va_start(va, ctl);
379210cb 466 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 467 __va_end(va);
5bf48697 468 KTR_LOG(ctxsw_newtd, td, td->td_comm);
7d0bac62
MD
469}
470
eb2adbf5
MD
471/*
472 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
473 * this does not prevent the thread from migrating to another cpu so the
474 * gd_tdallq state is not protected by this.
475 */
99df837e 476void
73e4f7b9 477lwkt_hold(thread_t td)
99df837e 478{
74c9628e 479 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
480}
481
482void
483lwkt_rele(thread_t td)
484{
485 KKASSERT(td->td_refs > 0);
74c9628e 486 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
487}
488
73e4f7b9
MD
489void
490lwkt_free_thread(thread_t td)
491{
74c9628e 492 KKASSERT(td->td_refs == 0);
c17a6852
MD
493 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
494 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
40aaf5fc
NT
495 if (td->td_flags & TDF_ALLOCATED_THREAD) {
496 objcache_put(thread_cache, td);
497 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
498 /* client-allocated struct with internally allocated stack */
499 KASSERT(td->td_kstack && td->td_kstack_size > 0,
500 ("lwkt_free_thread: corrupted stack"));
501 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
502 td->td_kstack = NULL;
503 td->td_kstack_size = 0;
99df837e 504 }
e7c0dbba 505 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
506}
507
508
8ad65e08
MD
509/*
510 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
511 * switch to the idlethread. Switching must occur within a critical
512 * section to avoid races with the scheduling queue.
513 *
514 * We always have full control over our cpu's run queue. Other cpus
515 * that wish to manipulate our queue must use the cpu_*msg() calls to
516 * talk to our cpu, so a critical section is all that is needed and
517 * the result is very, very fast thread switching.
518 *
96728c05
MD
519 * The LWKT scheduler uses a fixed priority model and round-robins at
520 * each priority level. User process scheduling is a totally
521 * different beast and LWKT priorities should not be confused with
522 * user process priorities.
f1d1c3fa 523 *
69d78e99
MD
524 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
525 * is not called by the current thread in the preemption case, only when
526 * the preempting thread blocks (in order to return to the original thread).
cfaeae2a
MD
527 *
528 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
529 * migration and tsleep deschedule the current lwkt thread and call
530 * lwkt_switch(). In particular, the target cpu of the migration fully
531 * expects the thread to become non-runnable and can deadlock against
532 * cpusync operations if we run any IPIs prior to switching the thread out.
533 *
534 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
95858b91 535 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
8ad65e08
MD
536 */
537void
538lwkt_switch(void)
539{
37af14fe
MD
540 globaldata_t gd = mycpu;
541 thread_t td = gd->gd_curthread;
8ad65e08 542 thread_t ntd;
b12defdc 543 int spinning = 0;
8ad65e08 544
da0b0e8b 545 KKASSERT(gd->gd_processing_ipiq == 0);
121f93bc 546 KKASSERT(td->td_flags & TDF_RUNNING);
da0b0e8b 547
46a3f46d 548 /*
27e88a6e
MD
549 * Switching from within a 'fast' (non thread switched) interrupt or IPI
550 * is illegal. However, we may have to do it anyway if we hit a fatal
551 * kernel trap or we have paniced.
552 *
553 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 554 */
27e88a6e
MD
555 if (gd->gd_intr_nesting_level) {
556 int savegdnest;
557 int savegdtrap;
558
5fddbda2 559 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22 560 panic("lwkt_switch: Attempt to switch from a "
5a8df152 561 "fast interrupt, ipi, or hard code section, "
4a28fe22
MD
562 "td %p\n",
563 td);
27e88a6e
MD
564 } else {
565 savegdnest = gd->gd_intr_nesting_level;
566 savegdtrap = gd->gd_trap_nesting_level;
567 gd->gd_intr_nesting_level = 0;
568 gd->gd_trap_nesting_level = 0;
a7422615
MD
569 if ((td->td_flags & TDF_PANICWARN) == 0) {
570 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
571 kprintf("Warning: thread switch from interrupt, IPI, "
572 "or hard code section.\n"
a7422615 573 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 574 print_backtrace(-1);
a7422615 575 }
27e88a6e
MD
576 lwkt_switch();
577 gd->gd_intr_nesting_level = savegdnest;
578 gd->gd_trap_nesting_level = savegdtrap;
579 return;
580 }
96728c05 581 }
ef0fdad1 582
cb973d15 583 /*
85946b6c
MD
584 * Release our current user process designation if we are blocking
585 * or if a user reschedule was requested.
586 *
587 * NOTE: This function is NOT called if we are switching into or
588 * returning from a preemption.
589 *
590 * NOTE: Releasing our current user process designation may cause
591 * it to be assigned to another thread, which in turn will
592 * cause us to block in the usched acquire code when we attempt
593 * to return to userland.
594 *
595 * NOTE: On SMP systems this can be very nasty when heavy token
596 * contention is present so we want to be careful not to
597 * release the designation gratuitously.
cb973d15 598 */
85946b6c
MD
599 if (td->td_release &&
600 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
cb973d15 601 td->td_release(td);
85946b6c 602 }
cb973d15 603
85946b6c
MD
604 /*
605 * Release all tokens
606 */
37af14fe 607 crit_enter_gd(gd);
3b998fa9 608 if (TD_TOKS_HELD(td))
9d265729
MD
609 lwkt_relalltokens(td);
610
611 /*
b02926de
MD
612 * We had better not be holding any spin locks, but don't get into an
613 * endless panic loop.
9d265729 614 */
0846e4ce 615 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL,
d666840a 616 ("lwkt_switch: still holding %d exclusive spinlocks!",
0846e4ce 617 gd->gd_spinlocks));
9d265729 618
8a8d5d85 619
0f7a3396
MD
620#ifdef INVARIANTS
621 if (td->td_cscount) {
6ea70f76 622 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
623 td);
624 if (panic_on_cscount)
625 panic("switching while mastering cpusync");
626 }
8a8d5d85 627#endif
f9235b6d
MD
628
629 /*
630 * If we had preempted another thread on this cpu, resume the preempted
631 * thread. This occurs transparently, whether the preempted thread
632 * was scheduled or not (it may have been preempted after descheduling
633 * itself).
634 *
635 * We have to setup the MP lock for the original thread after backing
636 * out the adjustment that was made to curthread when the original
637 * was preempted.
638 */
99df837e 639 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
640 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
641 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
642
643 /*
b9eb1c19
MD
644 * The interrupt may have woken a thread up, we need to properly
645 * set the reschedule flag if the originally interrupted thread is
646 * at a lower priority.
85946b6c
MD
647 *
648 * The interrupt may not have descheduled.
8ec60c3f 649 */
85946b6c 650 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
8ec60c3f 651 need_lwkt_resched();
f9235b6d
MD
652 goto havethread_preempted;
653 }
654
b12defdc 655 /*
f9235b6d 656 * If we cannot obtain ownership of the tokens we cannot immediately
cfaeae2a
MD
657 * schedule the target thread.
658 *
659 * Reminder: Again, we cannot afford to run any IPIs in this path if
660 * the current thread has been descheduled.
f9235b6d
MD
661 */
662 for (;;) {
b12defdc 663 clear_lwkt_resched();
f9235b6d 664
4b5f931b 665 /*
2a418930 666 * Hotpath - pull the head of the run queue and attempt to schedule
85946b6c 667 * it.
41a01a4d 668 */
7c3dd309
NA
669 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
670
671 if (ntd == NULL) {
672 /*
673 * Runq is empty, switch to idle to allow it to halt.
674 */
675 ntd = &gd->gd_idlethread;
7c3dd309
NA
676 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
677 ASSERT_NO_TOKENS_HELD(ntd);
7c3dd309
NA
678 cpu_time.cp_msg[0] = 0;
679 cpu_time.cp_stallpc = 0;
680 goto haveidle;
f9235b6d 681 }
41a01a4d 682
8ec60c3f 683 /*
b12defdc 684 * Hotpath - schedule ntd.
6f207a2c
MD
685 *
686 * NOTE: For UP there is no mplock and lwkt_getalltokens()
687 * always succeeds.
8ec60c3f 688 */
b12defdc
MD
689 if (TD_TOKS_NOT_HELD(ntd) ||
690 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
691 {
f9235b6d 692 goto havethread;
b12defdc 693 }
f9235b6d 694
f9235b6d 695 /*
2a418930
MD
696 * Coldpath (SMP only since tokens always succeed on UP)
697 *
698 * We had some contention on the thread we wanted to schedule.
699 * What we do now is try to find a thread that we can schedule
b12defdc 700 * in its stead.
2a418930 701 *
85946b6c
MD
702 * The coldpath scan does NOT rearrange threads in the run list.
703 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
704 * the next tick whenever the current head is not the current thread.
f9235b6d 705 */
b12defdc 706#ifdef INVARIANTS
85946b6c 707 ++ntd->td_contended;
b12defdc 708#endif
574339e6 709 ++gd->gd_cnt.v_token_colls;
b12defdc 710
85946b6c 711 if (fairq_bypass > 0)
b12defdc
MD
712 goto skip;
713
b12defdc 714 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
e3e6be1f 715#ifndef NO_LWKT_SPLIT_USERPRI
85946b6c
MD
716 /*
717 * Never schedule threads returning to userland or the
718 * user thread scheduler helper thread when higher priority
bb6860b5
NA
719 * threads are present. The runq is sorted by priority
720 * so we can give up traversing it when we find the first
721 * low priority thread.
85946b6c
MD
722 */
723 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
724 ntd = NULL;
725 break;
726 }
d992c377 727#endif
85946b6c
MD
728
729 /*
730 * Try this one.
731 */
b12defdc
MD
732 if (TD_TOKS_NOT_HELD(ntd) ||
733 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
734 goto havethread;
735 }
85946b6c 736#ifdef INVARIANTS
85946b6c 737 ++ntd->td_contended;
b12defdc 738#endif
574339e6 739 ++gd->gd_cnt.v_token_colls;
2a418930
MD
740 }
741
b12defdc 742skip:
2a418930
MD
743 /*
744 * We exhausted the run list, meaning that all runnable threads
b12defdc 745 * are contested.
2a418930
MD
746 */
747 cpu_pause();
748 ntd = &gd->gd_idlethread;
2a418930
MD
749 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
750 ASSERT_NO_TOKENS_HELD(ntd);
751 /* contention case, do not clear contention mask */
2a418930
MD
752
753 /*
b12defdc
MD
754 * We are going to have to retry but if the current thread is not
755 * on the runq we instead switch through the idle thread to get away
756 * from the current thread. We have to flag for lwkt reschedule
757 * to prevent the idle thread from halting.
2a418930 758 *
b12defdc
MD
759 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
760 * instruct it to deal with the potential for deadlocks by
761 * ordering the tokens by address.
2a418930 762 */
b12defdc 763 if ((td->td_flags & TDF_RUNQ) == 0) {
85946b6c 764 need_lwkt_resched(); /* prevent hlt */
2a418930 765 goto haveidle;
4b5f931b 766 }
bd52bedf 767#if defined(INVARIANTS) && defined(__amd64__)
d5b2d319
MD
768 if ((read_rflags() & PSL_I) == 0) {
769 cpu_enable_intr();
770 panic("lwkt_switch() called with interrupts disabled");
771 }
772#endif
b12defdc
MD
773
774 /*
775 * Number iterations so far. After a certain point we switch to
776 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
777 */
778 if (spinning < 0x7FFFFFFF)
779 ++spinning;
780
b12defdc
MD
781 /*
782 * lwkt_getalltokens() failed in sorted token mode, we can use
783 * monitor/mwait in this case.
784 */
785 if (spinning >= lwkt_spin_loops &&
786 (cpu_mi_feature & CPU_MI_MONITOR) &&
787 lwkt_spin_monitor)
788 {
789 cpu_mmw_pause_int(&gd->gd_reqflags,
790 (gd->gd_reqflags | RQF_SPINNING) &
791 ~RQF_IDLECHECK_WK_MASK);
792 }
b12defdc
MD
793
794 /*
795 * We already checked that td is still scheduled so this should be
796 * safe.
797 */
798 splz_check();
799
800 /*
801 * This experimental resequencer is used as a fall-back to reduce
802 * hw cache line contention by placing each core's scheduler into a
803 * time-domain-multplexed slot.
804 *
805 * The resequencer is disabled by default. It's functionality has
806 * largely been superceeded by the token algorithm which limits races
807 * to a subset of cores.
808 *
809 * The resequencer algorithm tends to break down when more than
810 * 20 cores are contending. What appears to happen is that new
811 * tokens can be obtained out of address-sorted order by new cores
812 * while existing cores languish in long delays between retries and
813 * wind up being starved-out of the token acquisition.
814 */
815 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
816 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
817 int oseq;
818
819 while ((oseq = lwkt_cseq_rindex) != cseq) {
820 cpu_ccfence();
821#if 1
822 if (cpu_mi_feature & CPU_MI_MONITOR) {
823 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
824 } else {
825#endif
826 cpu_pause();
827 cpu_lfence();
828#if 1
829 }
8b402283 830#endif
0f0466c0 831 }
b12defdc
MD
832 DELAY(1);
833 atomic_add_int(&lwkt_cseq_rindex, 1);
2a418930 834 }
2a418930 835 /* highest level for(;;) loop */
f1d1c3fa 836 }
8a8d5d85 837
2a418930 838havethread:
b12defdc 839 /*
be71787b
MD
840 * Clear gd_idle_repeat when doing a normal switch to a non-idle
841 * thread.
f9235b6d 842 */
9ac1ee6e 843 ntd->td_wmesg = NULL;
b12defdc 844 ++gd->gd_cnt.v_swtch;
be71787b 845 gd->gd_idle_repeat = 0;
2a418930 846
f9235b6d 847havethread_preempted:
f9235b6d
MD
848 /*
849 * If the new target does not need the MP lock and we are holding it,
850 * release the MP lock. If the new target requires the MP lock we have
851 * already acquired it for the target.
8a8d5d85 852 */
2a418930 853 ;
f9235b6d
MD
854haveidle:
855 KASSERT(ntd->td_critcount,
b5d16701
MD
856 ("priority problem in lwkt_switch %d %d",
857 td->td_critcount, ntd->td_critcount));
858
94f6d86e 859 if (td != ntd) {
cc9b6223
MD
860 /*
861 * Execute the actual thread switch operation. This function
862 * returns to the current thread and returns the previous thread
863 * (which may be different from the thread we switched to).
864 *
865 * We are responsible for marking ntd as TDF_RUNNING.
866 */
121f93bc 867 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
94f6d86e 868 ++switch_count;
a1f0fb66 869 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
870 ntd->td_flags |= TDF_RUNNING;
871 lwkt_switch_return(td->td_switch(ntd));
872 /* ntd invalid, td_switch() can return a different thread_t */
94f6d86e 873 }
b12defdc 874
b12defdc 875 /*
54341a3b 876 * catch-all. XXX is this strictly needed?
b12defdc
MD
877 */
878 splz_check();
54341a3b 879
37af14fe
MD
880 /* NOTE: current cpu may have changed after switch */
881 crit_exit_quick(td);
8ad65e08
MD
882}
883
cc9b6223
MD
884/*
885 * Called by assembly in the td_switch (thread restore path) for thread
886 * bootstrap cases which do not 'return' to lwkt_switch().
887 */
888void
889lwkt_switch_return(thread_t otd)
890{
cc9b6223
MD
891 globaldata_t rgd;
892
893 /*
894 * Check if otd was migrating. Now that we are on ntd we can finish
895 * up the migration. This is a bit messy but it is the only place
896 * where td is known to be fully descheduled.
897 *
898 * We can only activate the migration if otd was migrating but not
899 * held on the cpu due to a preemption chain. We still have to
900 * clear TDF_RUNNING on the old thread either way.
901 *
902 * We are responsible for clearing the previously running thread's
903 * TDF_RUNNING.
904 */
905 if ((rgd = otd->td_migrate_gd) != NULL &&
906 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
907 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
908 (TDF_MIGRATING | TDF_RUNNING));
909 otd->td_migrate_gd = NULL;
910 otd->td_flags &= ~TDF_RUNNING;
911 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
912 } else {
913 otd->td_flags &= ~TDF_RUNNING;
914 }
2b07d9aa
MD
915
916 /*
917 * Final exit validations (see lwp_wait()). Note that otd becomes
918 * invalid the *instant* we set TDF_MP_EXITSIG.
919 */
920 while (otd->td_flags & TDF_EXITING) {
921 u_int mpflags;
922
923 mpflags = otd->td_mpflags;
924 cpu_ccfence();
925
926 if (mpflags & TDF_MP_EXITWAIT) {
927 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
928 mpflags | TDF_MP_EXITSIG)) {
929 wakeup(otd);
930 break;
931 }
932 } else {
933 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
934 mpflags | TDF_MP_EXITSIG)) {
935 wakeup(otd);
936 break;
937 }
938 }
939 }
cc9b6223
MD
940}
941
b68b7282 942/*
96728c05 943 * Request that the target thread preempt the current thread. Preemption
54341a3b
MD
944 * can only occur if our only critical section is the one that we were called
945 * with, the relative priority of the target thread is higher, and the target
946 * thread holds no tokens. This also only works if we are not holding any
947 * spinlocks (obviously).
96728c05
MD
948 *
949 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
950 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 951 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
952 * to determine whether preemption is possible (aka usually just the crit
953 * priority of lwkt_schedule() itself).
b68b7282 954 *
54341a3b
MD
955 * Preemption is typically limited to interrupt threads.
956 *
957 * Operation works in a fairly straight-forward manner. The normal
958 * scheduling code is bypassed and we switch directly to the target
959 * thread. When the target thread attempts to block or switch away
960 * code at the base of lwkt_switch() will switch directly back to our
961 * thread. Our thread is able to retain whatever tokens it holds and
962 * if the target needs one of them the target will switch back to us
963 * and reschedule itself normally.
b68b7282
MD
964 */
965void
f9235b6d 966lwkt_preempt(thread_t ntd, int critcount)
b68b7282 967{
46a3f46d 968 struct globaldata *gd = mycpu;
cc9b6223 969 thread_t xtd;
0a3f9b47 970 thread_t td;
2d910aaf 971 int save_gd_intr_nesting_level;
b68b7282 972
26a0694b 973 /*
96728c05
MD
974 * The caller has put us in a critical section. We can only preempt
975 * if the caller of the caller was not in a critical section (basically
f9235b6d 976 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 977 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 978 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
979 *
980 * YYY The target thread must be in a critical section (else it must
981 * inherit our critical section? I dunno yet).
26a0694b 982 */
f9235b6d 983 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 984
b12defdc 985 td = gd->gd_curthread;
fbc024e4
MD
986 if (preempt_enable == 0) {
987 ++preempt_miss;
988 return;
989 }
f9235b6d 990 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
991 ++preempt_miss;
992 return;
993 }
f9235b6d 994 if (td->td_critcount > critcount) {
96728c05
MD
995 ++preempt_miss;
996 return;
997 }
121f93bc
MD
998 if (td->td_cscount) {
999 ++preempt_miss;
1000 return;
1001 }
46a3f46d 1002 if (ntd->td_gd != gd) {
96728c05
MD
1003 ++preempt_miss;
1004 return;
1005 }
41a01a4d 1006 /*
77912481
MD
1007 * We don't have to check spinlocks here as they will also bump
1008 * td_critcount.
d3d1cbc8
MD
1009 *
1010 * Do not try to preempt if the target thread is holding any tokens.
1011 * We could try to acquire the tokens but this case is so rare there
1012 * is no need to support it.
41a01a4d 1013 */
0846e4ce 1014 KKASSERT(gd->gd_spinlocks == 0);
77912481 1015
3b998fa9 1016 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8 1017 ++preempt_miss;
d3d1cbc8
MD
1018 return;
1019 }
26a0694b
MD
1020 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1021 ++preempt_weird;
1022 return;
1023 }
1024 if (ntd->td_preempted) {
4b5f931b 1025 ++preempt_hit;
26a0694b 1026 return;
b68b7282 1027 }
da0b0e8b 1028 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 1029
8ec60c3f
MD
1030 /*
1031 * Since we are able to preempt the current thread, there is no need to
1032 * call need_lwkt_resched().
2d910aaf
MD
1033 *
1034 * We must temporarily clear gd_intr_nesting_level around the switch
1035 * since switchouts from the target thread are allowed (they will just
1036 * return to our thread), and since the target thread has its own stack.
cc9b6223
MD
1037 *
1038 * A preemption must switch back to the original thread, assert the
1039 * case.
8ec60c3f 1040 */
26a0694b
MD
1041 ++preempt_hit;
1042 ntd->td_preempted = td;
1043 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1044 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1045 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1046 gd->gd_intr_nesting_level = 0;
121f93bc
MD
1047
1048 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
cc9b6223
MD
1049 ntd->td_flags |= TDF_RUNNING;
1050 xtd = td->td_switch(ntd);
1051 KKASSERT(xtd == ntd);
1052 lwkt_switch_return(xtd);
2d910aaf 1053 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1054
26a0694b
MD
1055 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1056 ntd->td_preempted = NULL;
1057 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1058}
1059
f1d1c3fa 1060/*
faaeffac 1061 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1062 * This will work inside a critical section but not inside a hard code
1063 * section.
ef0fdad1 1064 *
f1d1c3fa
MD
1065 * (self contained on a per cpu basis)
1066 */
1067void
faaeffac 1068splz_check(void)
f1d1c3fa 1069{
7966cb69
MD
1070 globaldata_t gd = mycpu;
1071 thread_t td = gd->gd_curthread;
ef0fdad1 1072
4a28fe22
MD
1073 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1074 gd->gd_intr_nesting_level == 0 &&
1075 td->td_nest_count < 2)
1076 {
f1d1c3fa 1077 splz();
4a28fe22
MD
1078 }
1079}
1080
1081/*
1082 * This version is integrated into crit_exit, reqflags has already
1083 * been tested but td_critcount has not.
1084 *
1085 * We only want to execute the splz() on the 1->0 transition of
1086 * critcount and not in a hard code section or if too deeply nested.
925040f2 1087 *
0846e4ce 1088 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
4a28fe22
MD
1089 */
1090void
1091lwkt_maybe_splz(thread_t td)
1092{
1093 globaldata_t gd = td->td_gd;
1094
1095 if (td->td_critcount == 0 &&
1096 gd->gd_intr_nesting_level == 0 &&
1097 td->td_nest_count < 2)
1098 {
1099 splz();
1100 }
f1d1c3fa
MD
1101}
1102
e6546af9
MD
1103/*
1104 * Drivers which set up processing co-threads can call this function to
1105 * run the co-thread at a higher priority and to allow it to preempt
1106 * normal threads.
1107 */
1108void
1109lwkt_set_interrupt_support_thread(void)
1110{
1111 thread_t td = curthread;
1112
1113 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1114 td->td_flags |= TDF_INTTHREAD;
1115 td->td_preemptable = lwkt_preempt;
1116}
1117
1118
8ad65e08 1119/*
f9235b6d
MD
1120 * This function is used to negotiate a passive release of the current
1121 * process/lwp designation with the user scheduler, allowing the user
1122 * scheduler to schedule another user thread. The related kernel thread
1123 * (curthread) continues running in the released state.
8ad65e08
MD
1124 */
1125void
f9235b6d 1126lwkt_passive_release(struct thread *td)
8ad65e08 1127{
f9235b6d
MD
1128 struct lwp *lp = td->td_lwp;
1129
e3e6be1f 1130#ifndef NO_LWKT_SPLIT_USERPRI
f9235b6d
MD
1131 td->td_release = NULL;
1132 lwkt_setpri_self(TDPRI_KERN_USER);
d992c377
MD
1133#endif
1134
f9235b6d 1135 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1136}
1137
f9235b6d 1138
3824f392 1139/*
d2d8515b
MD
1140 * This implements a LWKT yield, allowing a kernel thread to yield to other
1141 * kernel threads at the same or higher priority. This function can be
1142 * called in a tight loop and will typically only yield once per tick.
f9235b6d 1143 *
d2d8515b
MD
1144 * Most kernel threads run at the same priority in order to allow equal
1145 * sharing.
f9235b6d
MD
1146 *
1147 * (self contained on a per cpu basis)
3824f392
MD
1148 */
1149void
f9235b6d 1150lwkt_yield(void)
3824f392 1151{
f9235b6d
MD
1152 globaldata_t gd = mycpu;
1153 thread_t td = gd->gd_curthread;
3824f392 1154
f9235b6d
MD
1155 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1156 splz();
85946b6c 1157 if (lwkt_resched_wanted()) {
f9235b6d
MD
1158 lwkt_schedule_self(curthread);
1159 lwkt_switch();
f9235b6d 1160 }
3824f392
MD
1161}
1162
40504122
MD
1163/*
1164 * The quick version processes pending interrupts and higher-priority
1165 * LWKT threads but will not round-robin same-priority LWKT threads.
de4d4cb0
MD
1166 *
1167 * When called while attempting to return to userland the only same-pri
1168 * threads are the ones which have already tried to become the current
1169 * user process.
40504122
MD
1170 */
1171void
1172lwkt_yield_quick(void)
1173{
1174 globaldata_t gd = mycpu;
1175 thread_t td = gd->gd_curthread;
1176
1177 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1178 splz();
1179 if (lwkt_resched_wanted()) {
1180 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1181 clear_lwkt_resched();
1182 } else {
1183 lwkt_schedule_self(curthread);
1184 lwkt_switch();
1185 }
1186 }
1187}
1188
3824f392 1189/*
f9235b6d
MD
1190 * This yield is designed for kernel threads with a user context.
1191 *
1192 * The kernel acting on behalf of the user is potentially cpu-bound,
1193 * this function will efficiently allow other threads to run and also
1194 * switch to other processes by releasing.
3824f392
MD
1195 *
1196 * The lwkt_user_yield() function is designed to have very low overhead
1197 * if no yield is determined to be needed.
1198 */
1199void
1200lwkt_user_yield(void)
1201{
f9235b6d
MD
1202 globaldata_t gd = mycpu;
1203 thread_t td = gd->gd_curthread;
1204
1205 /*
1206 * Always run any pending interrupts in case we are in a critical
1207 * section.
1208 */
1209 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1210 splz();
3824f392 1211
3824f392 1212 /*
f9235b6d
MD
1213 * Switch (which forces a release) if another kernel thread needs
1214 * the cpu, if userland wants us to resched, or if our kernel
1215 * quantum has run out.
3824f392 1216 */
f9235b6d 1217 if (lwkt_resched_wanted() ||
85946b6c 1218 user_resched_wanted())
f9235b6d 1219 {
3824f392 1220 lwkt_switch();
3824f392
MD
1221 }
1222
f9235b6d 1223#if 0
3824f392 1224 /*
f9235b6d
MD
1225 * Reacquire the current process if we are released.
1226 *
1227 * XXX not implemented atm. The kernel may be holding locks and such,
1228 * so we want the thread to continue to receive cpu.
3824f392 1229 */
f9235b6d
MD
1230 if (td->td_release == NULL && lp) {
1231 lp->lwp_proc->p_usched->acquire_curproc(lp);
1232 td->td_release = lwkt_passive_release;
1233 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1234 }
f9235b6d 1235#endif
b9eb1c19
MD
1236}
1237
8ad65e08 1238/*
f1d1c3fa
MD
1239 * Generic schedule. Possibly schedule threads belonging to other cpus and
1240 * deal with threads that might be blocked on a wait queue.
1241 *
0a3f9b47
MD
1242 * We have a little helper inline function which does additional work after
1243 * the thread has been enqueued, including dealing with preemption and
1244 * setting need_lwkt_resched() (which prevents the kernel from returning
1245 * to userland until it has processed higher priority threads).
6330a558
MD
1246 *
1247 * It is possible for this routine to be called after a failed _enqueue
1248 * (due to the target thread migrating, sleeping, or otherwise blocked).
1249 * We have to check that the thread is actually on the run queue!
8ad65e08 1250 */
0a3f9b47
MD
1251static __inline
1252void
85946b6c 1253_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
0a3f9b47 1254{
6330a558 1255 if (ntd->td_flags & TDF_RUNQ) {
85946b6c 1256 if (ntd->td_preemptable) {
f9235b6d 1257 ntd->td_preemptable(ntd, ccount); /* YYY +token */
6330a558 1258 }
0a3f9b47
MD
1259 }
1260}
1261
361d01dd 1262static __inline
8ad65e08 1263void
85946b6c 1264_lwkt_schedule(thread_t td)
8ad65e08 1265{
37af14fe
MD
1266 globaldata_t mygd = mycpu;
1267
cf709dd2
MD
1268 KASSERT(td != &td->td_gd->gd_idlethread,
1269 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1270 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1271 crit_enter_gd(mygd);
4643740a
MD
1272 KKASSERT(td->td_lwp == NULL ||
1273 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1274
37af14fe 1275 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1276 _lwkt_enqueue(td);
1277 } else {
f1d1c3fa 1278 /*
7cd8d145
MD
1279 * If we own the thread, there is no race (since we are in a
1280 * critical section). If we do not own the thread there might
1281 * be a race but the target cpu will deal with it.
f1d1c3fa 1282 */
7cd8d145 1283 if (td->td_gd == mygd) {
9d265729 1284 _lwkt_enqueue(td);
85946b6c 1285 _lwkt_schedule_post(mygd, td, 1);
f1d1c3fa 1286 } else {
e381e77c 1287 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1288 }
8ad65e08 1289 }
37af14fe 1290 crit_exit_gd(mygd);
8ad65e08
MD
1291}
1292
361d01dd
MD
1293void
1294lwkt_schedule(thread_t td)
1295{
85946b6c 1296 _lwkt_schedule(td);
361d01dd
MD
1297}
1298
1299void
85946b6c 1300lwkt_schedule_noresched(thread_t td) /* XXX not impl */
361d01dd 1301{
85946b6c 1302 _lwkt_schedule(td);
361d01dd
MD
1303}
1304
e381e77c
MD
1305/*
1306 * When scheduled remotely if frame != NULL the IPIQ is being
1307 * run via doreti or an interrupt then preemption can be allowed.
1308 *
1309 * To allow preemption we have to drop the critical section so only
1310 * one is present in _lwkt_schedule_post.
1311 */
1312static void
1313lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1314{
1315 thread_t td = curthread;
1316 thread_t ntd = arg;
1317
1318 if (frame && ntd->td_preemptable) {
1319 crit_exit_noyield(td);
85946b6c 1320 _lwkt_schedule(ntd);
e381e77c
MD
1321 crit_enter_quick(td);
1322 } else {
85946b6c 1323 _lwkt_schedule(ntd);
e381e77c
MD
1324 }
1325}
1326
d9eea1a5 1327/*
52eedfb5
MD
1328 * Thread migration using a 'Pull' method. The thread may or may not be
1329 * the current thread. It MUST be descheduled and in a stable state.
1330 * lwkt_giveaway() must be called on the cpu owning the thread.
1331 *
1332 * At any point after lwkt_giveaway() is called, the target cpu may
1333 * 'pull' the thread by calling lwkt_acquire().
1334 *
ae8e83e6
MD
1335 * We have to make sure the thread is not sitting on a per-cpu tsleep
1336 * queue or it will blow up when it moves to another cpu.
1337 *
52eedfb5 1338 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1339 */
52eedfb5
MD
1340void
1341lwkt_giveaway(thread_t td)
1342{
3b4192fb 1343 globaldata_t gd = mycpu;
52eedfb5 1344
3b4192fb
MD
1345 crit_enter_gd(gd);
1346 if (td->td_flags & TDF_TSLEEPQ)
1347 tsleep_remove(td);
1348 KKASSERT(td->td_gd == gd);
1349 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1350 td->td_flags |= TDF_MIGRATING;
1351 crit_exit_gd(gd);
52eedfb5
MD
1352}
1353
a2a5ad0d
MD
1354void
1355lwkt_acquire(thread_t td)
1356{
37af14fe
MD
1357 globaldata_t gd;
1358 globaldata_t mygd;
cc9b6223 1359 int retry = 10000000;
a2a5ad0d 1360
52eedfb5 1361 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1362 gd = td->td_gd;
37af14fe 1363 mygd = mycpu;
52eedfb5 1364 if (gd != mycpu) {
35238fa5 1365 cpu_lfence();
52eedfb5 1366 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1367 crit_enter_gd(mygd);
cfaeae2a 1368 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23 1369 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
df910c23 1370 lwkt_process_ipiq();
52eedfb5 1371 cpu_lfence();
cc9b6223
MD
1372 if (--retry == 0) {
1373 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1374 td, td->td_flags);
1375 retry = 10000000;
1376 }
df910c23 1377 }
cfaeae2a 1378 DEBUG_POP_INFO();
562273ea 1379 cpu_mfence();
37af14fe 1380 td->td_gd = mygd;
52eedfb5
MD
1381 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1382 td->td_flags &= ~TDF_MIGRATING;
1383 crit_exit_gd(mygd);
1384 } else {
1385 crit_enter_gd(mygd);
1386 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1387 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1388 crit_exit_gd(mygd);
a2a5ad0d
MD
1389 }
1390}
1391
f1d1c3fa
MD
1392/*
1393 * Generic deschedule. Descheduling threads other then your own should be
1394 * done only in carefully controlled circumstances. Descheduling is
1395 * asynchronous.
1396 *
1397 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1398 */
1399void
1400lwkt_deschedule(thread_t td)
1401{
f1d1c3fa
MD
1402 crit_enter();
1403 if (td == curthread) {
1404 _lwkt_dequeue(td);
1405 } else {
a72187e9 1406 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1407 _lwkt_dequeue(td);
1408 } else {
b8a98473 1409 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1410 }
1411 }
1412 crit_exit();
1413}
1414
4b5f931b
MD
1415/*
1416 * Set the target thread's priority. This routine does not automatically
1417 * switch to a higher priority thread, LWKT threads are not designed for
1418 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1419 */
1420void
1421lwkt_setpri(thread_t td, int pri)
1422{
f9235b6d
MD
1423 if (td->td_pri != pri) {
1424 KKASSERT(pri >= 0);
1425 crit_enter();
1426 if (td->td_flags & TDF_RUNQ) {
d2d8515b 1427 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1428 _lwkt_dequeue(td);
1429 td->td_pri = pri;
1430 _lwkt_enqueue(td);
1431 } else {
1432 td->td_pri = pri;
1433 }
1434 crit_exit();
26a0694b 1435 }
26a0694b
MD
1436}
1437
03bd0a5e
MD
1438/*
1439 * Set the initial priority for a thread prior to it being scheduled for
1440 * the first time. The thread MUST NOT be scheduled before or during
1441 * this call. The thread may be assigned to a cpu other then the current
1442 * cpu.
1443 *
1444 * Typically used after a thread has been created with TDF_STOPPREQ,
1445 * and before the thread is initially scheduled.
1446 */
1447void
1448lwkt_setpri_initial(thread_t td, int pri)
1449{
1450 KKASSERT(pri >= 0);
1451 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1452 td->td_pri = pri;
03bd0a5e
MD
1453}
1454
26a0694b
MD
1455void
1456lwkt_setpri_self(int pri)
1457{
1458 thread_t td = curthread;
1459
4b5f931b
MD
1460 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1461 crit_enter();
1462 if (td->td_flags & TDF_RUNQ) {
1463 _lwkt_dequeue(td);
f9235b6d 1464 td->td_pri = pri;
4b5f931b
MD
1465 _lwkt_enqueue(td);
1466 } else {
f9235b6d 1467 td->td_pri = pri;
4b5f931b
MD
1468 }
1469 crit_exit();
1470}
1471
f9235b6d 1472/*
85946b6c 1473 * hz tick scheduler clock for LWKT threads
f9235b6d
MD
1474 */
1475void
85946b6c 1476lwkt_schedulerclock(thread_t td)
f9235b6d 1477{
85946b6c
MD
1478 globaldata_t gd = td->td_gd;
1479 thread_t xtd;
2a418930 1480
85946b6c
MD
1481 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1482 /*
1483 * If the current thread is at the head of the runq shift it to the
1484 * end of any equal-priority threads and request a LWKT reschedule
1485 * if it moved.
d992c377
MD
1486 *
1487 * Ignore upri in this situation. There will only be one user thread
1488 * in user mode, all others will be user threads running in kernel
1489 * mode and we have to make sure they get some cpu.
85946b6c
MD
1490 */
1491 xtd = TAILQ_NEXT(td, td_threadq);
1492 if (xtd && xtd->td_pri == td->td_pri) {
1493 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1494 while (xtd && xtd->td_pri == td->td_pri)
1495 xtd = TAILQ_NEXT(xtd, td_threadq);
1496 if (xtd)
1497 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1498 else
1499 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1500 need_lwkt_resched();
f9235b6d 1501 }
85946b6c
MD
1502 } else {
1503 /*
1504 * If we scheduled a thread other than the one at the head of the
1505 * queue always request a reschedule every tick.
1506 */
1507 need_lwkt_resched();
f9235b6d
MD
1508 }
1509}
1510
5d21b981 1511/*
52eedfb5
MD
1512 * Migrate the current thread to the specified cpu.
1513 *
cc9b6223
MD
1514 * This is accomplished by descheduling ourselves from the current cpu
1515 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1516 * 'old' thread wants to migrate after it has been completely switched out
1517 * and will complete the migration.
1518 *
1519 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1520 *
1521 * We must be sure to release our current process designation (if a user
1522 * process) before clearing out any tsleepq we are on because the release
1523 * code may re-add us.
ae8e83e6
MD
1524 *
1525 * We must be sure to remove ourselves from the current cpu's tsleepq
1526 * before potentially moving to another queue. The thread can be on
1527 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1528 */
5d21b981
MD
1529
1530void
1531lwkt_setcpu_self(globaldata_t rgd)
1532{
5d21b981
MD
1533 thread_t td = curthread;
1534
1535 if (td->td_gd != rgd) {
1536 crit_enter_quick(td);
cc9b6223 1537
95858b91
MD
1538 if (td->td_release)
1539 td->td_release(td);
ae8e83e6 1540 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1541 tsleep_remove(td);
cc9b6223
MD
1542
1543 /*
1544 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1545 * trying to deschedule ourselves and switch away, then deschedule
1546 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1547 * call lwkt_switch() to complete the operation.
1548 */
5d21b981
MD
1549 td->td_flags |= TDF_MIGRATING;
1550 lwkt_deschedule_self(td);
52eedfb5 1551 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1552 td->td_migrate_gd = rgd;
5d21b981 1553 lwkt_switch();
cc9b6223
MD
1554
1555 /*
1556 * We are now on the target cpu
1557 */
1558 KKASSERT(rgd == mycpu);
52eedfb5 1559 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1560 crit_exit_quick(td);
1561 }
5d21b981
MD
1562}
1563
ecdefdda
MD
1564void
1565lwkt_migratecpu(int cpuid)
1566{
ecdefdda
MD
1567 globaldata_t rgd;
1568
1569 rgd = globaldata_find(cpuid);
1570 lwkt_setcpu_self(rgd);
ecdefdda
MD
1571}
1572
5d21b981
MD
1573/*
1574 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
1575 * do not have to enter another one).
1576 *
1577 * The thread (td) has already been completely descheduled from the
1578 * originating cpu and we can simply assert the case. The thread is
1579 * assigned to the new cpu and enqueued.
5d21b981 1580 *
cc9b6223 1581 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1582 */
1583static void
1584lwkt_setcpu_remote(void *arg)
1585{
1586 thread_t td = arg;
1587 globaldata_t gd = mycpu;
1588
cc9b6223 1589 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1590 td->td_gd = gd;
562273ea 1591 cpu_mfence();
5d21b981 1592 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1593 KKASSERT(td->td_migrate_gd == NULL);
4643740a
MD
1594 KKASSERT(td->td_lwp == NULL ||
1595 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
5d21b981
MD
1596 _lwkt_enqueue(td);
1597}
1598
553ea3c8 1599struct lwp *
4b5f931b
MD
1600lwkt_preempted_proc(void)
1601{
73e4f7b9 1602 thread_t td = curthread;
4b5f931b
MD
1603 while (td->td_preempted)
1604 td = td->td_preempted;
553ea3c8 1605 return(td->td_lwp);
4b5f931b
MD
1606}
1607
99df837e
MD
1608/*
1609 * Create a kernel process/thread/whatever. It shares it's address space
1610 * with proc0 - ie: kernel only.
1611 *
d2d8515b
MD
1612 * If the cpu is not specified one will be selected. In the future
1613 * specifying a cpu of -1 will enable kernel thread migration between
1614 * cpus.
99df837e
MD
1615 */
1616int
c9e9fb21
MD
1617lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1618 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1619{
73e4f7b9 1620 thread_t td;
e2565a42 1621 __va_list ap;
99df837e 1622
d3d32139 1623 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1624 tdflags);
a2a5ad0d
MD
1625 if (tdp)
1626 *tdp = td;
709799ea 1627 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1628
1629 /*
1630 * Set up arg0 for 'ps' etc
1631 */
e2565a42 1632 __va_start(ap, fmt);
379210cb 1633 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1634 __va_end(ap);
99df837e
MD
1635
1636 /*
1637 * Schedule the thread to run
1638 */
4643740a
MD
1639 if (td->td_flags & TDF_NOSTART)
1640 td->td_flags &= ~TDF_NOSTART;
ef0fdad1 1641 else
4643740a 1642 lwkt_schedule(td);
99df837e
MD
1643 return 0;
1644}
1645
1646/*
1647 * Destroy an LWKT thread. Warning! This function is not called when
1648 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1649 * uses a different reaping mechanism.
1650 */
1651void
1652lwkt_exit(void)
1653{
1654 thread_t td = curthread;
c070746a 1655 thread_t std;
8826f33a 1656 globaldata_t gd;
99df837e 1657
2883d2d8
MD
1658 /*
1659 * Do any cleanup that might block here
1660 */
99df837e 1661 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1662 kprintf("kthread %p %s has exited\n", td, td->td_comm);
2883d2d8
MD
1663 biosched_done(td);
1664 dsched_exit_thread(td);
c070746a
MD
1665
1666 /*
1667 * Get us into a critical section to interlock gd_freetd and loop
1668 * until we can get it freed.
1669 *
1670 * We have to cache the current td in gd_freetd because objcache_put()ing
1671 * it would rip it out from under us while our thread is still active.
2af9d75d
MD
1672 *
1673 * We are the current thread so of course our own TDF_RUNNING bit will
1674 * be set, so unlike the lwp reap code we don't wait for it to clear.
c070746a
MD
1675 */
1676 gd = mycpu;
37af14fe 1677 crit_enter_quick(td);
2af9d75d
MD
1678 for (;;) {
1679 if (td->td_refs) {
1680 tsleep(td, 0, "tdreap", 1);
1681 continue;
1682 }
1683 if ((std = gd->gd_freetd) != NULL) {
1684 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1685 gd->gd_freetd = NULL;
1686 objcache_put(thread_cache, std);
1687 continue;
1688 }
1689 break;
c070746a 1690 }
3b4192fb
MD
1691
1692 /*
1693 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1694 * the last time. We cannot block after this point or we may end
1695 * up with a stale td on the tsleepq.
eb2adbf5
MD
1696 *
1697 * None of this may block, the critical section is the only thing
1698 * protecting tdallq and the only thing preventing new lwkt_hold()
1699 * thread refs now.
3b4192fb
MD
1700 */
1701 if (td->td_flags & TDF_TSLEEPQ)
1702 tsleep_remove(td);
37af14fe 1703 lwkt_deschedule_self(td);
e56e4dea 1704 lwkt_remove_tdallq(td);
74c9628e 1705 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1706
1707 /*
1708 * Final cleanup
1709 */
1710 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1711 if (td->td_flags & TDF_ALLOCATED_THREAD)
1712 gd->gd_freetd = td;
99df837e
MD
1713 cpu_thread_exit();
1714}
1715
e56e4dea
MD
1716void
1717lwkt_remove_tdallq(thread_t td)
1718{
1719 KKASSERT(td->td_gd == mycpu);
1720 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1721}
1722
9cf43f91
MD
1723/*
1724 * Code reduction and branch prediction improvements. Call/return
1725 * overhead on modern cpus often degenerates into 0 cycles due to
1726 * the cpu's branch prediction hardware and return pc cache. We
1727 * can take advantage of this by not inlining medium-complexity
1728 * functions and we can also reduce the branch prediction impact
1729 * by collapsing perfectly predictable branches into a single
1730 * procedure instead of duplicating it.
1731 *
1732 * Is any of this noticeable? Probably not, so I'll take the
1733 * smaller code size.
1734 */
1735void
b6468f56 1736crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1737{
b6468f56 1738 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1739}
1740
2d93b37a
MD
1741void
1742crit_panic(void)
1743{
1744 thread_t td = curthread;
850634cc 1745 int lcrit = td->td_critcount;
2d93b37a 1746
850634cc
AH
1747 td->td_critcount = 0;
1748 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1749 /* NOT REACHED */
2d93b37a
MD
1750}
1751
bd8015ca
MD
1752/*
1753 * Called from debugger/panic on cpus which have been stopped. We must still
1754 * process the IPIQ while stopped, even if we were stopped while in a critical
1755 * section (XXX).
1756 *
1757 * If we are dumping also try to process any pending interrupts. This may
1758 * or may not work depending on the state of the cpu at the point it was
1759 * stopped.
1760 */
1761void
1762lwkt_smp_stopped(void)
1763{
1764 globaldata_t gd = mycpu;
1765
1766 crit_enter_gd(gd);
1767 if (dumping) {
1768 lwkt_process_ipiq();
1769 splz();
1770 } else {
1771 lwkt_process_ipiq();
1772 }
1773 crit_exit_gd(gd);
1774}