systat - Ensure vmmeter output separates fields by at least one space.
[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>
f6bf3af1 53#include <sys/caps.h>
9d265729 54#include <sys/spinlock.h>
57aa743c 55#include <sys/ktr.h>
9d265729
MD
56
57#include <sys/thread2.h>
58#include <sys/spinlock2.h>
684a93c4 59#include <sys/mplock2.h>
f1d1c3fa 60
8c72e3d5
AH
61#include <sys/dsched.h>
62
7d0bac62
MD
63#include <vm/vm.h>
64#include <vm/vm_param.h>
65#include <vm/vm_kern.h>
66#include <vm/vm_object.h>
67#include <vm/vm_page.h>
68#include <vm/vm_map.h>
69#include <vm/vm_pager.h>
70#include <vm/vm_extern.h>
7d0bac62 71
99df837e 72#include <machine/stdarg.h>
96728c05 73#include <machine/smp.h>
99df837e 74
d850923c
AE
75#if !defined(KTR_CTXSW)
76#define KTR_CTXSW KTR_ALL
77#endif
78KTR_INFO_MASTER(ctxsw);
5bf48697
AE
79KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td);
80KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td);
81KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm);
82KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td);
1541028a 83
40aaf5fc
NT
84static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
85
0f7a3396
MD
86#ifdef INVARIANTS
87static int panic_on_cscount = 0;
88#endif
05220613
MD
89static __int64_t switch_count = 0;
90static __int64_t preempt_hit = 0;
91static __int64_t preempt_miss = 0;
92static __int64_t preempt_weird = 0;
fb0f29c4 93static int lwkt_use_spin_port;
40aaf5fc 94static struct objcache *thread_cache;
05220613 95
88ebb169 96#ifdef SMP
e381e77c 97static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
cc9b6223 98static void lwkt_setcpu_remote(void *arg);
88ebb169 99#endif
e381e77c 100
0855a2af
JG
101extern void cpu_heavy_restore(void);
102extern void cpu_lwkt_restore(void);
103extern void cpu_kthread_restore(void);
104extern void cpu_idle_restore(void);
105
fb0f29c4
MD
106/*
107 * We can make all thread ports use the spin backend instead of the thread
108 * backend. This should only be set to debug the spin backend.
109 */
110TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
111
0f7a3396 112#ifdef INVARIANTS
0c52fa62
SG
113SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
114 "Panic if attempting to switch lwkt's while mastering cpusync");
0f7a3396 115#endif
0c52fa62
SG
116SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
117 "Number of switched threads");
9733f757 118SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
0c52fa62 119 "Successful preemption events");
9733f757 120SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
0c52fa62
SG
121 "Failed preemption events");
122SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
123 "Number of preempted threads.");
b12defdc 124static int fairq_enable = 0;
2a418930
MD
125SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
126 &fairq_enable, 0, "Turn on fairq priority accumulators");
85946b6c 127static int fairq_bypass = -1;
b12defdc
MD
128SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
129 &fairq_bypass, 0, "Allow fairq to bypass td on token failure");
130extern int lwkt_sched_debug;
131int lwkt_sched_debug = 0;
132SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
133 &lwkt_sched_debug, 0, "Scheduler debug");
2a418930
MD
134static int lwkt_spin_loops = 10;
135SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
b12defdc
MD
136 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
137static int lwkt_spin_reseq = 0;
138SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW,
139 &lwkt_spin_reseq, 0, "Scheduler resequencer enable");
140static int lwkt_spin_monitor = 0;
141SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW,
142 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait");
d5b2d319
MD
143static int lwkt_spin_fatal = 0; /* disabled */
144SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
145 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
fbc024e4 146static int preempt_enable = 1;
2a418930
MD
147SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
148 &preempt_enable, 0, "Enable preemption");
7b234d8c 149static int lwkt_cache_threads = 0;
765b1ae0
MD
150SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
151 &lwkt_cache_threads, 0, "thread+kstack cache");
fbc024e4 152
2a418930
MD
153static __cachealign int lwkt_cseq_rindex;
154static __cachealign int lwkt_cseq_windex;
05220613 155
4b5f931b
MD
156/*
157 * These helper procedures handle the runq, they can only be called from
158 * within a critical section.
75cdbe6c
MD
159 *
160 * WARNING! Prior to SMP being brought up it is possible to enqueue and
161 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
162 * instead of 'mycpu' when referencing the globaldata structure. Once
163 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 164 */
f1d1c3fa
MD
165static __inline
166void
167_lwkt_dequeue(thread_t td)
168{
169 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 170 struct globaldata *gd = td->td_gd;
4b5f931b 171
f1d1c3fa 172 td->td_flags &= ~TDF_RUNQ;
f9235b6d 173 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
de4d4cb0 174 --gd->gd_tdrunqcount;
f9235b6d 175 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
2a418930 176 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
177 }
178}
179
f9235b6d
MD
180/*
181 * Priority enqueue.
182 *
183 * NOTE: There are a limited number of lwkt threads runnable since user
184 * processes only schedule one at a time per cpu.
185 */
f1d1c3fa
MD
186static __inline
187void
188_lwkt_enqueue(thread_t td)
189{
f9235b6d
MD
190 thread_t xtd;
191
7f5d7ed7 192 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 193 struct globaldata *gd = td->td_gd;
4b5f931b 194
f1d1c3fa 195 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
196 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
197 if (xtd == NULL) {
85946b6c
MD
198 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
199 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
f9235b6d 200 } else {
85946b6c
MD
201 while (xtd && xtd->td_pri >= td->td_pri)
202 xtd = TAILQ_NEXT(xtd, td_threadq);
203 if (xtd)
204 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
205 else
206 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
f9235b6d 207 }
de4d4cb0 208 ++gd->gd_tdrunqcount;
b12defdc
MD
209
210 /*
85946b6c 211 * Request a LWKT reschedule if we are now at the head of the queue.
b12defdc 212 */
85946b6c
MD
213 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
214 need_lwkt_resched();
f1d1c3fa
MD
215 }
216}
8ad65e08 217
40aaf5fc
NT
218static __boolean_t
219_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
220{
221 struct thread *td = (struct thread *)obj;
222
223 td->td_kstack = NULL;
224 td->td_kstack_size = 0;
225 td->td_flags = TDF_ALLOCATED_THREAD;
4643740a 226 td->td_mpflags = 0;
40aaf5fc
NT
227 return (1);
228}
229
230static void
231_lwkt_thread_dtor(void *obj, void *privdata)
232{
233 struct thread *td = (struct thread *)obj;
234
235 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
236 ("_lwkt_thread_dtor: not allocated from objcache"));
237 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
238 td->td_kstack_size > 0,
239 ("_lwkt_thread_dtor: corrupted stack"));
240 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
7b234d8c
MD
241 td->td_kstack = NULL;
242 td->td_flags = 0;
40aaf5fc
NT
243}
244
245/*
246 * Initialize the lwkt s/system.
765b1ae0 247 *
7b234d8c
MD
248 * Nominally cache up to 32 thread + kstack structures. Cache more on
249 * systems with a lot of cpu cores.
40aaf5fc
NT
250 */
251void
252lwkt_init(void)
253{
765b1ae0 254 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
7b234d8c
MD
255 if (lwkt_cache_threads == 0) {
256 lwkt_cache_threads = ncpus * 4;
257 if (lwkt_cache_threads < 32)
258 lwkt_cache_threads = 32;
259 }
765b1ae0
MD
260 thread_cache = objcache_create_mbacked(
261 M_THREAD, sizeof(struct thread),
262 NULL, lwkt_cache_threads,
263 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
264}
265
37af14fe
MD
266/*
267 * Schedule a thread to run. As the current thread we can always safely
268 * schedule ourselves, and a shortcut procedure is provided for that
269 * function.
270 *
271 * (non-blocking, self contained on a per cpu basis)
272 */
273void
274lwkt_schedule_self(thread_t td)
275{
cfaeae2a 276 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 277 crit_enter_quick(td);
f9235b6d
MD
278 KASSERT(td != &td->td_gd->gd_idlethread,
279 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
4643740a
MD
280 KKASSERT(td->td_lwp == NULL ||
281 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
37af14fe 282 _lwkt_enqueue(td);
37af14fe
MD
283 crit_exit_quick(td);
284}
285
286/*
287 * Deschedule a thread.
288 *
289 * (non-blocking, self contained on a per cpu basis)
290 */
291void
292lwkt_deschedule_self(thread_t td)
293{
294 crit_enter_quick(td);
37af14fe
MD
295 _lwkt_dequeue(td);
296 crit_exit_quick(td);
297}
298
8ad65e08
MD
299/*
300 * LWKTs operate on a per-cpu basis
301 *
73e4f7b9 302 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
303 */
304void
305lwkt_gdinit(struct globaldata *gd)
306{
f9235b6d 307 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 308 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
309}
310
7d0bac62
MD
311/*
312 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
313 * or LWKT start address before it can be scheduled. If the target cpu is
314 * -1 the thread will be created on the current cpu.
0cfcada1
MD
315 *
316 * If you intend to create a thread without a process context this function
317 * does everything except load the startup and switcher function.
7d0bac62
MD
318 */
319thread_t
d3d32139 320lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 321{
d2d8515b 322 static int cpu_rotator;
c070746a 323 globaldata_t gd = mycpu;
99df837e 324 void *stack;
7d0bac62 325
c070746a
MD
326 /*
327 * If static thread storage is not supplied allocate a thread. Reuse
328 * a cached free thread if possible. gd_freetd is used to keep an exiting
329 * thread intact through the exit.
330 */
ef0fdad1 331 if (td == NULL) {
cf709dd2
MD
332 crit_enter_gd(gd);
333 if ((td = gd->gd_freetd) != NULL) {
334 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
335 TDF_RUNQ)) == 0);
c070746a 336 gd->gd_freetd = NULL;
cf709dd2 337 } else {
c070746a 338 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
339 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
340 TDF_RUNQ)) == 0);
341 }
342 crit_exit_gd(gd);
40aaf5fc 343 KASSERT((td->td_flags &
2af9d75d
MD
344 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
345 TDF_ALLOCATED_THREAD,
40aaf5fc
NT
346 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
347 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 348 }
c070746a
MD
349
350 /*
351 * Try to reuse cached stack.
352 */
f470d0c8
MD
353 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
354 if (flags & TDF_ALLOCATED_STACK) {
e4846942 355 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
356 stack = NULL;
357 }
358 }
359 if (stack == NULL) {
e40cfbd7 360 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 361 flags |= TDF_ALLOCATED_STACK;
99df837e 362 }
d2d8515b
MD
363 if (cpu < 0) {
364 cpu = ++cpu_rotator;
365 cpu_ccfence();
366 cpu %= ncpus;
367 }
368 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 369 return(td);
7d0bac62
MD
370}
371
372/*
373 * Initialize a preexisting thread structure. This function is used by
374 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
375 *
f8c3996b
MD
376 * All threads start out in a critical section at a priority of
377 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
378 * appropriate. This function may send an IPI message when the
379 * requested cpu is not the current cpu and consequently gd_tdallq may
380 * not be initialized synchronously from the point of view of the originating
381 * cpu.
382 *
383 * NOTE! we have to be careful in regards to creating threads for other cpus
384 * if SMP has not yet been activated.
7d0bac62 385 */
41a01a4d
MD
386#ifdef SMP
387
75cdbe6c
MD
388static void
389lwkt_init_thread_remote(void *arg)
390{
391 thread_t td = arg;
392
52eedfb5
MD
393 /*
394 * Protected by critical section held by IPI dispatch
395 */
75cdbe6c
MD
396 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
397}
398
41a01a4d
MD
399#endif
400
fdce8919
MD
401/*
402 * lwkt core thread structural initialization.
403 *
404 * NOTE: All threads are initialized as mpsafe threads.
405 */
7d0bac62 406void
f470d0c8
MD
407lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
408 struct globaldata *gd)
7d0bac62 409{
37af14fe
MD
410 globaldata_t mygd = mycpu;
411
99df837e
MD
412 bzero(td, sizeof(struct thread));
413 td->td_kstack = stack;
f470d0c8 414 td->td_kstack_size = stksize;
d3d32139 415 td->td_flags = flags;
4643740a 416 td->td_mpflags = 0;
26a0694b 417 td->td_gd = gd;
f9235b6d
MD
418 td->td_pri = TDPRI_KERN_DAEMON;
419 td->td_critcount = 1;
54341a3b 420 td->td_toks_have = NULL;
3b998fa9 421 td->td_toks_stop = &td->td_toks_base;
392cd266 422 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
d636cd48 423 lwkt_initport_spin(&td->td_msgport, td);
fb0f29c4
MD
424 else
425 lwkt_initport_thread(&td->td_msgport, td);
99df837e 426 pmap_init_thread(td);
0f7a3396 427#ifdef SMP
5d21b981
MD
428 /*
429 * Normally initializing a thread for a remote cpu requires sending an
430 * IPI. However, the idlethread is setup before the other cpus are
431 * activated so we have to treat it as a special case. XXX manipulation
432 * of gd_tdallq requires the BGL.
433 */
434 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 435 crit_enter_gd(mygd);
75cdbe6c 436 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 437 crit_exit_gd(mygd);
75cdbe6c 438 } else {
2db3b277 439 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 440 }
0f7a3396 441#else
37af14fe 442 crit_enter_gd(mygd);
0f7a3396 443 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 444 crit_exit_gd(mygd);
0f7a3396 445#endif
8c72e3d5
AH
446
447 dsched_new_thread(td);
73e4f7b9
MD
448}
449
450void
451lwkt_set_comm(thread_t td, const char *ctl, ...)
452{
e2565a42 453 __va_list va;
73e4f7b9 454
e2565a42 455 __va_start(va, ctl);
379210cb 456 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 457 __va_end(va);
5bf48697 458 KTR_LOG(ctxsw_newtd, td, td->td_comm);
7d0bac62
MD
459}
460
eb2adbf5
MD
461/*
462 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
463 * this does not prevent the thread from migrating to another cpu so the
464 * gd_tdallq state is not protected by this.
465 */
99df837e 466void
73e4f7b9 467lwkt_hold(thread_t td)
99df837e 468{
74c9628e 469 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
470}
471
472void
473lwkt_rele(thread_t td)
474{
475 KKASSERT(td->td_refs > 0);
74c9628e 476 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
477}
478
73e4f7b9
MD
479void
480lwkt_free_thread(thread_t td)
481{
74c9628e 482 KKASSERT(td->td_refs == 0);
c17a6852
MD
483 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
484 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
40aaf5fc
NT
485 if (td->td_flags & TDF_ALLOCATED_THREAD) {
486 objcache_put(thread_cache, td);
487 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
488 /* client-allocated struct with internally allocated stack */
489 KASSERT(td->td_kstack && td->td_kstack_size > 0,
490 ("lwkt_free_thread: corrupted stack"));
491 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
492 td->td_kstack = NULL;
493 td->td_kstack_size = 0;
99df837e 494 }
e7c0dbba 495 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
496}
497
498
8ad65e08
MD
499/*
500 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
501 * switch to the idlethread. Switching must occur within a critical
502 * section to avoid races with the scheduling queue.
503 *
504 * We always have full control over our cpu's run queue. Other cpus
505 * that wish to manipulate our queue must use the cpu_*msg() calls to
506 * talk to our cpu, so a critical section is all that is needed and
507 * the result is very, very fast thread switching.
508 *
96728c05
MD
509 * The LWKT scheduler uses a fixed priority model and round-robins at
510 * each priority level. User process scheduling is a totally
511 * different beast and LWKT priorities should not be confused with
512 * user process priorities.
f1d1c3fa 513 *
69d78e99
MD
514 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
515 * is not called by the current thread in the preemption case, only when
516 * the preempting thread blocks (in order to return to the original thread).
cfaeae2a
MD
517 *
518 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
519 * migration and tsleep deschedule the current lwkt thread and call
520 * lwkt_switch(). In particular, the target cpu of the migration fully
521 * expects the thread to become non-runnable and can deadlock against
522 * cpusync operations if we run any IPIs prior to switching the thread out.
523 *
524 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
95858b91 525 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
8ad65e08
MD
526 */
527void
528lwkt_switch(void)
529{
37af14fe
MD
530 globaldata_t gd = mycpu;
531 thread_t td = gd->gd_curthread;
8ad65e08 532 thread_t ntd;
b12defdc 533 int spinning = 0;
8ad65e08 534
da0b0e8b 535 KKASSERT(gd->gd_processing_ipiq == 0);
121f93bc 536 KKASSERT(td->td_flags & TDF_RUNNING);
da0b0e8b 537
46a3f46d 538 /*
27e88a6e
MD
539 * Switching from within a 'fast' (non thread switched) interrupt or IPI
540 * is illegal. However, we may have to do it anyway if we hit a fatal
541 * kernel trap or we have paniced.
542 *
543 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 544 */
27e88a6e
MD
545 if (gd->gd_intr_nesting_level) {
546 int savegdnest;
547 int savegdtrap;
548
5fddbda2 549 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22 550 panic("lwkt_switch: Attempt to switch from a "
5a8df152 551 "fast interrupt, ipi, or hard code section, "
4a28fe22
MD
552 "td %p\n",
553 td);
27e88a6e
MD
554 } else {
555 savegdnest = gd->gd_intr_nesting_level;
556 savegdtrap = gd->gd_trap_nesting_level;
557 gd->gd_intr_nesting_level = 0;
558 gd->gd_trap_nesting_level = 0;
a7422615
MD
559 if ((td->td_flags & TDF_PANICWARN) == 0) {
560 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
561 kprintf("Warning: thread switch from interrupt, IPI, "
562 "or hard code section.\n"
a7422615 563 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 564 print_backtrace(-1);
a7422615 565 }
27e88a6e
MD
566 lwkt_switch();
567 gd->gd_intr_nesting_level = savegdnest;
568 gd->gd_trap_nesting_level = savegdtrap;
569 return;
570 }
96728c05 571 }
ef0fdad1 572
cb973d15 573 /*
85946b6c
MD
574 * Release our current user process designation if we are blocking
575 * or if a user reschedule was requested.
576 *
577 * NOTE: This function is NOT called if we are switching into or
578 * returning from a preemption.
579 *
580 * NOTE: Releasing our current user process designation may cause
581 * it to be assigned to another thread, which in turn will
582 * cause us to block in the usched acquire code when we attempt
583 * to return to userland.
584 *
585 * NOTE: On SMP systems this can be very nasty when heavy token
586 * contention is present so we want to be careful not to
587 * release the designation gratuitously.
cb973d15 588 */
85946b6c
MD
589 if (td->td_release &&
590 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
cb973d15 591 td->td_release(td);
85946b6c 592 }
cb973d15 593
85946b6c
MD
594 /*
595 * Release all tokens
596 */
37af14fe 597 crit_enter_gd(gd);
3b998fa9 598 if (TD_TOKS_HELD(td))
9d265729
MD
599 lwkt_relalltokens(td);
600
601 /*
b02926de
MD
602 * We had better not be holding any spin locks, but don't get into an
603 * endless panic loop.
9d265729 604 */
d666840a
MD
605 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
606 ("lwkt_switch: still holding %d exclusive spinlocks!",
607 gd->gd_spinlocks_wr));
9d265729 608
8a8d5d85
MD
609
610#ifdef SMP
0f7a3396
MD
611#ifdef INVARIANTS
612 if (td->td_cscount) {
6ea70f76 613 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
614 td);
615 if (panic_on_cscount)
616 panic("switching while mastering cpusync");
617 }
618#endif
8a8d5d85 619#endif
f9235b6d
MD
620
621 /*
622 * If we had preempted another thread on this cpu, resume the preempted
623 * thread. This occurs transparently, whether the preempted thread
624 * was scheduled or not (it may have been preempted after descheduling
625 * itself).
626 *
627 * We have to setup the MP lock for the original thread after backing
628 * out the adjustment that was made to curthread when the original
629 * was preempted.
630 */
99df837e 631 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
632 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
633 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
634
635 /*
b9eb1c19
MD
636 * The interrupt may have woken a thread up, we need to properly
637 * set the reschedule flag if the originally interrupted thread is
638 * at a lower priority.
85946b6c
MD
639 *
640 * The interrupt may not have descheduled.
8ec60c3f 641 */
85946b6c 642 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
8ec60c3f 643 need_lwkt_resched();
f9235b6d
MD
644 goto havethread_preempted;
645 }
646
b12defdc 647 /*
f9235b6d 648 * If we cannot obtain ownership of the tokens we cannot immediately
cfaeae2a
MD
649 * schedule the target thread.
650 *
651 * Reminder: Again, we cannot afford to run any IPIs in this path if
652 * the current thread has been descheduled.
f9235b6d
MD
653 */
654 for (;;) {
b12defdc 655 clear_lwkt_resched();
f9235b6d 656
4b5f931b 657 /*
2a418930 658 * Hotpath - pull the head of the run queue and attempt to schedule
85946b6c 659 * it.
41a01a4d 660 */
7c3dd309
NA
661 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
662
663 if (ntd == NULL) {
664 /*
665 * Runq is empty, switch to idle to allow it to halt.
666 */
667 ntd = &gd->gd_idlethread;
6f207a2c 668#ifdef SMP
7c3dd309
NA
669 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
670 ASSERT_NO_TOKENS_HELD(ntd);
6f207a2c 671#endif
7c3dd309
NA
672 cpu_time.cp_msg[0] = 0;
673 cpu_time.cp_stallpc = 0;
674 goto haveidle;
f9235b6d 675 }
41a01a4d 676
8ec60c3f 677 /*
b12defdc 678 * Hotpath - schedule ntd.
6f207a2c
MD
679 *
680 * NOTE: For UP there is no mplock and lwkt_getalltokens()
681 * always succeeds.
8ec60c3f 682 */
b12defdc
MD
683 if (TD_TOKS_NOT_HELD(ntd) ||
684 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
685 {
f9235b6d 686 goto havethread;
b12defdc 687 }
f9235b6d 688
f9235b6d 689 /*
2a418930
MD
690 * Coldpath (SMP only since tokens always succeed on UP)
691 *
692 * We had some contention on the thread we wanted to schedule.
693 * What we do now is try to find a thread that we can schedule
b12defdc 694 * in its stead.
2a418930 695 *
85946b6c
MD
696 * The coldpath scan does NOT rearrange threads in the run list.
697 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
698 * the next tick whenever the current head is not the current thread.
f9235b6d 699 */
b12defdc 700#ifdef INVARIANTS
85946b6c 701 ++ntd->td_contended;
b12defdc 702#endif
574339e6 703 ++gd->gd_cnt.v_token_colls;
b12defdc 704
85946b6c 705 if (fairq_bypass > 0)
b12defdc
MD
706 goto skip;
707
b12defdc 708 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
85946b6c
MD
709 /*
710 * Never schedule threads returning to userland or the
711 * user thread scheduler helper thread when higher priority
bb6860b5
NA
712 * threads are present. The runq is sorted by priority
713 * so we can give up traversing it when we find the first
714 * low priority thread.
85946b6c
MD
715 */
716 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
717 ntd = NULL;
718 break;
719 }
720
721 /*
722 * Try this one.
723 */
b12defdc
MD
724 if (TD_TOKS_NOT_HELD(ntd) ||
725 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
726 goto havethread;
727 }
85946b6c 728#ifdef INVARIANTS
85946b6c 729 ++ntd->td_contended;
b12defdc 730#endif
574339e6 731 ++gd->gd_cnt.v_token_colls;
2a418930
MD
732 }
733
b12defdc 734skip:
2a418930
MD
735 /*
736 * We exhausted the run list, meaning that all runnable threads
b12defdc 737 * are contested.
2a418930
MD
738 */
739 cpu_pause();
740 ntd = &gd->gd_idlethread;
741#ifdef SMP
742 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
743 ASSERT_NO_TOKENS_HELD(ntd);
744 /* contention case, do not clear contention mask */
745#endif
746
747 /*
b12defdc
MD
748 * We are going to have to retry but if the current thread is not
749 * on the runq we instead switch through the idle thread to get away
750 * from the current thread. We have to flag for lwkt reschedule
751 * to prevent the idle thread from halting.
2a418930 752 *
b12defdc
MD
753 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
754 * instruct it to deal with the potential for deadlocks by
755 * ordering the tokens by address.
2a418930 756 */
b12defdc 757 if ((td->td_flags & TDF_RUNQ) == 0) {
85946b6c 758 need_lwkt_resched(); /* prevent hlt */
2a418930 759 goto haveidle;
4b5f931b 760 }
bd52bedf 761#if defined(INVARIANTS) && defined(__amd64__)
d5b2d319
MD
762 if ((read_rflags() & PSL_I) == 0) {
763 cpu_enable_intr();
764 panic("lwkt_switch() called with interrupts disabled");
765 }
766#endif
b12defdc
MD
767
768 /*
769 * Number iterations so far. After a certain point we switch to
770 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
771 */
772 if (spinning < 0x7FFFFFFF)
773 ++spinning;
774
775#ifdef SMP
776 /*
777 * lwkt_getalltokens() failed in sorted token mode, we can use
778 * monitor/mwait in this case.
779 */
780 if (spinning >= lwkt_spin_loops &&
781 (cpu_mi_feature & CPU_MI_MONITOR) &&
782 lwkt_spin_monitor)
783 {
784 cpu_mmw_pause_int(&gd->gd_reqflags,
785 (gd->gd_reqflags | RQF_SPINNING) &
786 ~RQF_IDLECHECK_WK_MASK);
787 }
788#endif
789
790 /*
791 * We already checked that td is still scheduled so this should be
792 * safe.
793 */
794 splz_check();
795
796 /*
797 * This experimental resequencer is used as a fall-back to reduce
798 * hw cache line contention by placing each core's scheduler into a
799 * time-domain-multplexed slot.
800 *
801 * The resequencer is disabled by default. It's functionality has
802 * largely been superceeded by the token algorithm which limits races
803 * to a subset of cores.
804 *
805 * The resequencer algorithm tends to break down when more than
806 * 20 cores are contending. What appears to happen is that new
807 * tokens can be obtained out of address-sorted order by new cores
808 * while existing cores languish in long delays between retries and
809 * wind up being starved-out of the token acquisition.
810 */
811 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
812 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
813 int oseq;
814
815 while ((oseq = lwkt_cseq_rindex) != cseq) {
816 cpu_ccfence();
817#if 1
818 if (cpu_mi_feature & CPU_MI_MONITOR) {
819 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
820 } else {
821#endif
822 cpu_pause();
823 cpu_lfence();
824#if 1
825 }
8b402283 826#endif
0f0466c0 827 }
b12defdc
MD
828 DELAY(1);
829 atomic_add_int(&lwkt_cseq_rindex, 1);
2a418930 830 }
2a418930 831 /* highest level for(;;) loop */
f1d1c3fa 832 }
8a8d5d85 833
2a418930 834havethread:
b12defdc 835 /*
be71787b
MD
836 * Clear gd_idle_repeat when doing a normal switch to a non-idle
837 * thread.
f9235b6d 838 */
9ac1ee6e 839 ntd->td_wmesg = NULL;
b12defdc 840 ++gd->gd_cnt.v_swtch;
be71787b 841 gd->gd_idle_repeat = 0;
2a418930 842
f9235b6d 843havethread_preempted:
f9235b6d
MD
844 /*
845 * If the new target does not need the MP lock and we are holding it,
846 * release the MP lock. If the new target requires the MP lock we have
847 * already acquired it for the target.
8a8d5d85 848 */
2a418930 849 ;
f9235b6d
MD
850haveidle:
851 KASSERT(ntd->td_critcount,
b5d16701
MD
852 ("priority problem in lwkt_switch %d %d",
853 td->td_critcount, ntd->td_critcount));
854
94f6d86e 855 if (td != ntd) {
cc9b6223
MD
856 /*
857 * Execute the actual thread switch operation. This function
858 * returns to the current thread and returns the previous thread
859 * (which may be different from the thread we switched to).
860 *
861 * We are responsible for marking ntd as TDF_RUNNING.
862 */
121f93bc 863 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
94f6d86e 864 ++switch_count;
a1f0fb66 865 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
866 ntd->td_flags |= TDF_RUNNING;
867 lwkt_switch_return(td->td_switch(ntd));
868 /* ntd invalid, td_switch() can return a different thread_t */
94f6d86e 869 }
b12defdc 870
b12defdc 871 /*
54341a3b 872 * catch-all. XXX is this strictly needed?
b12defdc
MD
873 */
874 splz_check();
54341a3b 875
37af14fe
MD
876 /* NOTE: current cpu may have changed after switch */
877 crit_exit_quick(td);
8ad65e08
MD
878}
879
cc9b6223
MD
880/*
881 * Called by assembly in the td_switch (thread restore path) for thread
882 * bootstrap cases which do not 'return' to lwkt_switch().
883 */
884void
885lwkt_switch_return(thread_t otd)
886{
887#ifdef SMP
888 globaldata_t rgd;
889
890 /*
891 * Check if otd was migrating. Now that we are on ntd we can finish
892 * up the migration. This is a bit messy but it is the only place
893 * where td is known to be fully descheduled.
894 *
895 * We can only activate the migration if otd was migrating but not
896 * held on the cpu due to a preemption chain. We still have to
897 * clear TDF_RUNNING on the old thread either way.
898 *
899 * We are responsible for clearing the previously running thread's
900 * TDF_RUNNING.
901 */
902 if ((rgd = otd->td_migrate_gd) != NULL &&
903 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
904 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
905 (TDF_MIGRATING | TDF_RUNNING));
906 otd->td_migrate_gd = NULL;
907 otd->td_flags &= ~TDF_RUNNING;
908 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
909 } else {
910 otd->td_flags &= ~TDF_RUNNING;
911 }
912#else
913 otd->td_flags &= ~TDF_RUNNING;
914#endif
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 }
998#ifdef SMP
121f93bc
MD
999 if (td->td_cscount) {
1000 ++preempt_miss;
1001 return;
1002 }
46a3f46d 1003 if (ntd->td_gd != gd) {
96728c05
MD
1004 ++preempt_miss;
1005 return;
1006 }
1007#endif
41a01a4d 1008 /*
77912481
MD
1009 * We don't have to check spinlocks here as they will also bump
1010 * td_critcount.
d3d1cbc8
MD
1011 *
1012 * Do not try to preempt if the target thread is holding any tokens.
1013 * We could try to acquire the tokens but this case is so rare there
1014 * is no need to support it.
41a01a4d 1015 */
77912481
MD
1016 KKASSERT(gd->gd_spinlocks_wr == 0);
1017
3b998fa9 1018 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8 1019 ++preempt_miss;
d3d1cbc8
MD
1020 return;
1021 }
26a0694b
MD
1022 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1023 ++preempt_weird;
1024 return;
1025 }
1026 if (ntd->td_preempted) {
4b5f931b 1027 ++preempt_hit;
26a0694b 1028 return;
b68b7282 1029 }
da0b0e8b 1030 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 1031
8ec60c3f
MD
1032 /*
1033 * Since we are able to preempt the current thread, there is no need to
1034 * call need_lwkt_resched().
2d910aaf
MD
1035 *
1036 * We must temporarily clear gd_intr_nesting_level around the switch
1037 * since switchouts from the target thread are allowed (they will just
1038 * return to our thread), and since the target thread has its own stack.
cc9b6223
MD
1039 *
1040 * A preemption must switch back to the original thread, assert the
1041 * case.
8ec60c3f 1042 */
26a0694b
MD
1043 ++preempt_hit;
1044 ntd->td_preempted = td;
1045 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1046 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1047 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1048 gd->gd_intr_nesting_level = 0;
121f93bc
MD
1049
1050 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
cc9b6223
MD
1051 ntd->td_flags |= TDF_RUNNING;
1052 xtd = td->td_switch(ntd);
1053 KKASSERT(xtd == ntd);
1054 lwkt_switch_return(xtd);
2d910aaf 1055 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1056
26a0694b
MD
1057 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1058 ntd->td_preempted = NULL;
1059 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1060}
1061
f1d1c3fa 1062/*
faaeffac 1063 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1064 * This will work inside a critical section but not inside a hard code
1065 * section.
ef0fdad1 1066 *
f1d1c3fa
MD
1067 * (self contained on a per cpu basis)
1068 */
1069void
faaeffac 1070splz_check(void)
f1d1c3fa 1071{
7966cb69
MD
1072 globaldata_t gd = mycpu;
1073 thread_t td = gd->gd_curthread;
ef0fdad1 1074
4a28fe22
MD
1075 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1076 gd->gd_intr_nesting_level == 0 &&
1077 td->td_nest_count < 2)
1078 {
f1d1c3fa 1079 splz();
4a28fe22
MD
1080 }
1081}
1082
1083/*
1084 * This version is integrated into crit_exit, reqflags has already
1085 * been tested but td_critcount has not.
1086 *
1087 * We only want to execute the splz() on the 1->0 transition of
1088 * critcount and not in a hard code section or if too deeply nested.
925040f2
MD
1089 *
1090 * NOTE: gd->gd_spinlocks_wr is implied to be 0 when td_critcount is 0.
4a28fe22
MD
1091 */
1092void
1093lwkt_maybe_splz(thread_t td)
1094{
1095 globaldata_t gd = td->td_gd;
1096
1097 if (td->td_critcount == 0 &&
1098 gd->gd_intr_nesting_level == 0 &&
1099 td->td_nest_count < 2)
1100 {
1101 splz();
1102 }
f1d1c3fa
MD
1103}
1104
e6546af9
MD
1105/*
1106 * Drivers which set up processing co-threads can call this function to
1107 * run the co-thread at a higher priority and to allow it to preempt
1108 * normal threads.
1109 */
1110void
1111lwkt_set_interrupt_support_thread(void)
1112{
1113 thread_t td = curthread;
1114
1115 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1116 td->td_flags |= TDF_INTTHREAD;
1117 td->td_preemptable = lwkt_preempt;
1118}
1119
1120
8ad65e08 1121/*
f9235b6d
MD
1122 * This function is used to negotiate a passive release of the current
1123 * process/lwp designation with the user scheduler, allowing the user
1124 * scheduler to schedule another user thread. The related kernel thread
1125 * (curthread) continues running in the released state.
8ad65e08
MD
1126 */
1127void
f9235b6d 1128lwkt_passive_release(struct thread *td)
8ad65e08 1129{
f9235b6d
MD
1130 struct lwp *lp = td->td_lwp;
1131
1132 td->td_release = NULL;
1133 lwkt_setpri_self(TDPRI_KERN_USER);
1134 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1135}
1136
f9235b6d 1137
3824f392 1138/*
d2d8515b
MD
1139 * This implements a LWKT yield, allowing a kernel thread to yield to other
1140 * kernel threads at the same or higher priority. This function can be
1141 * called in a tight loop and will typically only yield once per tick.
f9235b6d 1142 *
d2d8515b
MD
1143 * Most kernel threads run at the same priority in order to allow equal
1144 * sharing.
f9235b6d
MD
1145 *
1146 * (self contained on a per cpu basis)
3824f392
MD
1147 */
1148void
f9235b6d 1149lwkt_yield(void)
3824f392 1150{
f9235b6d
MD
1151 globaldata_t gd = mycpu;
1152 thread_t td = gd->gd_curthread;
3824f392 1153
f9235b6d
MD
1154 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1155 splz();
85946b6c 1156 if (lwkt_resched_wanted()) {
f9235b6d
MD
1157 lwkt_schedule_self(curthread);
1158 lwkt_switch();
f9235b6d 1159 }
3824f392
MD
1160}
1161
40504122
MD
1162/*
1163 * The quick version processes pending interrupts and higher-priority
1164 * LWKT threads but will not round-robin same-priority LWKT threads.
de4d4cb0
MD
1165 *
1166 * When called while attempting to return to userland the only same-pri
1167 * threads are the ones which have already tried to become the current
1168 * user process.
40504122
MD
1169 */
1170void
1171lwkt_yield_quick(void)
1172{
1173 globaldata_t gd = mycpu;
1174 thread_t td = gd->gd_curthread;
1175
1176 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1177 splz();
1178 if (lwkt_resched_wanted()) {
1179 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1180 clear_lwkt_resched();
1181 } else {
1182 lwkt_schedule_self(curthread);
1183 lwkt_switch();
1184 }
1185 }
1186}
1187
3824f392 1188/*
f9235b6d
MD
1189 * This yield is designed for kernel threads with a user context.
1190 *
1191 * The kernel acting on behalf of the user is potentially cpu-bound,
1192 * this function will efficiently allow other threads to run and also
1193 * switch to other processes by releasing.
3824f392
MD
1194 *
1195 * The lwkt_user_yield() function is designed to have very low overhead
1196 * if no yield is determined to be needed.
1197 */
1198void
1199lwkt_user_yield(void)
1200{
f9235b6d
MD
1201 globaldata_t gd = mycpu;
1202 thread_t td = gd->gd_curthread;
1203
1204 /*
1205 * Always run any pending interrupts in case we are in a critical
1206 * section.
1207 */
1208 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1209 splz();
3824f392 1210
3824f392 1211 /*
f9235b6d
MD
1212 * Switch (which forces a release) if another kernel thread needs
1213 * the cpu, if userland wants us to resched, or if our kernel
1214 * quantum has run out.
3824f392 1215 */
f9235b6d 1216 if (lwkt_resched_wanted() ||
85946b6c 1217 user_resched_wanted())
f9235b6d 1218 {
3824f392 1219 lwkt_switch();
3824f392
MD
1220 }
1221
f9235b6d 1222#if 0
3824f392 1223 /*
f9235b6d
MD
1224 * Reacquire the current process if we are released.
1225 *
1226 * XXX not implemented atm. The kernel may be holding locks and such,
1227 * so we want the thread to continue to receive cpu.
3824f392 1228 */
f9235b6d
MD
1229 if (td->td_release == NULL && lp) {
1230 lp->lwp_proc->p_usched->acquire_curproc(lp);
1231 td->td_release = lwkt_passive_release;
1232 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1233 }
f9235b6d 1234#endif
b9eb1c19
MD
1235}
1236
8ad65e08 1237/*
f1d1c3fa
MD
1238 * Generic schedule. Possibly schedule threads belonging to other cpus and
1239 * deal with threads that might be blocked on a wait queue.
1240 *
0a3f9b47
MD
1241 * We have a little helper inline function which does additional work after
1242 * the thread has been enqueued, including dealing with preemption and
1243 * setting need_lwkt_resched() (which prevents the kernel from returning
1244 * to userland until it has processed higher priority threads).
6330a558
MD
1245 *
1246 * It is possible for this routine to be called after a failed _enqueue
1247 * (due to the target thread migrating, sleeping, or otherwise blocked).
1248 * We have to check that the thread is actually on the run queue!
8ad65e08 1249 */
0a3f9b47
MD
1250static __inline
1251void
85946b6c 1252_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
0a3f9b47 1253{
6330a558 1254 if (ntd->td_flags & TDF_RUNQ) {
85946b6c 1255 if (ntd->td_preemptable) {
f9235b6d 1256 ntd->td_preemptable(ntd, ccount); /* YYY +token */
6330a558 1257 }
0a3f9b47
MD
1258 }
1259}
1260
361d01dd 1261static __inline
8ad65e08 1262void
85946b6c 1263_lwkt_schedule(thread_t td)
8ad65e08 1264{
37af14fe
MD
1265 globaldata_t mygd = mycpu;
1266
cf709dd2
MD
1267 KASSERT(td != &td->td_gd->gd_idlethread,
1268 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1269 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1270 crit_enter_gd(mygd);
4643740a
MD
1271 KKASSERT(td->td_lwp == NULL ||
1272 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1273
37af14fe 1274 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1275 _lwkt_enqueue(td);
1276 } else {
f1d1c3fa 1277 /*
7cd8d145
MD
1278 * If we own the thread, there is no race (since we are in a
1279 * critical section). If we do not own the thread there might
1280 * be a race but the target cpu will deal with it.
f1d1c3fa 1281 */
0f7a3396 1282#ifdef SMP
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 }
0f7a3396 1289#else
7cd8d145 1290 _lwkt_enqueue(td);
85946b6c 1291 _lwkt_schedule_post(mygd, td, 1);
0f7a3396 1292#endif
8ad65e08 1293 }
37af14fe 1294 crit_exit_gd(mygd);
8ad65e08
MD
1295}
1296
361d01dd
MD
1297void
1298lwkt_schedule(thread_t td)
1299{
85946b6c 1300 _lwkt_schedule(td);
361d01dd
MD
1301}
1302
1303void
85946b6c 1304lwkt_schedule_noresched(thread_t td) /* XXX not impl */
361d01dd 1305{
85946b6c 1306 _lwkt_schedule(td);
361d01dd
MD
1307}
1308
88ebb169
SW
1309#ifdef SMP
1310
e381e77c
MD
1311/*
1312 * When scheduled remotely if frame != NULL the IPIQ is being
1313 * run via doreti or an interrupt then preemption can be allowed.
1314 *
1315 * To allow preemption we have to drop the critical section so only
1316 * one is present in _lwkt_schedule_post.
1317 */
1318static void
1319lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1320{
1321 thread_t td = curthread;
1322 thread_t ntd = arg;
1323
1324 if (frame && ntd->td_preemptable) {
1325 crit_exit_noyield(td);
85946b6c 1326 _lwkt_schedule(ntd);
e381e77c
MD
1327 crit_enter_quick(td);
1328 } else {
85946b6c 1329 _lwkt_schedule(ntd);
e381e77c
MD
1330 }
1331}
1332
d9eea1a5 1333/*
52eedfb5
MD
1334 * Thread migration using a 'Pull' method. The thread may or may not be
1335 * the current thread. It MUST be descheduled and in a stable state.
1336 * lwkt_giveaway() must be called on the cpu owning the thread.
1337 *
1338 * At any point after lwkt_giveaway() is called, the target cpu may
1339 * 'pull' the thread by calling lwkt_acquire().
1340 *
ae8e83e6
MD
1341 * We have to make sure the thread is not sitting on a per-cpu tsleep
1342 * queue or it will blow up when it moves to another cpu.
1343 *
52eedfb5 1344 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1345 */
52eedfb5
MD
1346void
1347lwkt_giveaway(thread_t td)
1348{
3b4192fb 1349 globaldata_t gd = mycpu;
52eedfb5 1350
3b4192fb
MD
1351 crit_enter_gd(gd);
1352 if (td->td_flags & TDF_TSLEEPQ)
1353 tsleep_remove(td);
1354 KKASSERT(td->td_gd == gd);
1355 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1356 td->td_flags |= TDF_MIGRATING;
1357 crit_exit_gd(gd);
52eedfb5
MD
1358}
1359
a2a5ad0d
MD
1360void
1361lwkt_acquire(thread_t td)
1362{
37af14fe
MD
1363 globaldata_t gd;
1364 globaldata_t mygd;
cc9b6223 1365 int retry = 10000000;
a2a5ad0d 1366
52eedfb5 1367 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1368 gd = td->td_gd;
37af14fe 1369 mygd = mycpu;
52eedfb5 1370 if (gd != mycpu) {
35238fa5 1371 cpu_lfence();
52eedfb5 1372 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1373 crit_enter_gd(mygd);
cfaeae2a 1374 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23
MD
1375 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1376#ifdef SMP
1377 lwkt_process_ipiq();
1378#endif
52eedfb5 1379 cpu_lfence();
cc9b6223
MD
1380 if (--retry == 0) {
1381 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1382 td, td->td_flags);
1383 retry = 10000000;
1384 }
df910c23 1385 }
cfaeae2a 1386 DEBUG_POP_INFO();
562273ea 1387 cpu_mfence();
37af14fe 1388 td->td_gd = mygd;
52eedfb5
MD
1389 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1390 td->td_flags &= ~TDF_MIGRATING;
1391 crit_exit_gd(mygd);
1392 } else {
1393 crit_enter_gd(mygd);
1394 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1395 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1396 crit_exit_gd(mygd);
a2a5ad0d
MD
1397 }
1398}
1399
52eedfb5
MD
1400#endif
1401
f1d1c3fa
MD
1402/*
1403 * Generic deschedule. Descheduling threads other then your own should be
1404 * done only in carefully controlled circumstances. Descheduling is
1405 * asynchronous.
1406 *
1407 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1408 */
1409void
1410lwkt_deschedule(thread_t td)
1411{
f1d1c3fa 1412 crit_enter();
b8a98473 1413#ifdef SMP
f1d1c3fa
MD
1414 if (td == curthread) {
1415 _lwkt_dequeue(td);
1416 } else {
a72187e9 1417 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1418 _lwkt_dequeue(td);
1419 } else {
b8a98473 1420 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1421 }
1422 }
b8a98473
MD
1423#else
1424 _lwkt_dequeue(td);
1425#endif
f1d1c3fa
MD
1426 crit_exit();
1427}
1428
4b5f931b
MD
1429/*
1430 * Set the target thread's priority. This routine does not automatically
1431 * switch to a higher priority thread, LWKT threads are not designed for
1432 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1433 */
1434void
1435lwkt_setpri(thread_t td, int pri)
1436{
f9235b6d
MD
1437 if (td->td_pri != pri) {
1438 KKASSERT(pri >= 0);
1439 crit_enter();
1440 if (td->td_flags & TDF_RUNQ) {
d2d8515b 1441 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1442 _lwkt_dequeue(td);
1443 td->td_pri = pri;
1444 _lwkt_enqueue(td);
1445 } else {
1446 td->td_pri = pri;
1447 }
1448 crit_exit();
26a0694b 1449 }
26a0694b
MD
1450}
1451
03bd0a5e
MD
1452/*
1453 * Set the initial priority for a thread prior to it being scheduled for
1454 * the first time. The thread MUST NOT be scheduled before or during
1455 * this call. The thread may be assigned to a cpu other then the current
1456 * cpu.
1457 *
1458 * Typically used after a thread has been created with TDF_STOPPREQ,
1459 * and before the thread is initially scheduled.
1460 */
1461void
1462lwkt_setpri_initial(thread_t td, int pri)
1463{
1464 KKASSERT(pri >= 0);
1465 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1466 td->td_pri = pri;
03bd0a5e
MD
1467}
1468
26a0694b
MD
1469void
1470lwkt_setpri_self(int pri)
1471{
1472 thread_t td = curthread;
1473
4b5f931b
MD
1474 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1475 crit_enter();
1476 if (td->td_flags & TDF_RUNQ) {
1477 _lwkt_dequeue(td);
f9235b6d 1478 td->td_pri = pri;
4b5f931b
MD
1479 _lwkt_enqueue(td);
1480 } else {
f9235b6d 1481 td->td_pri = pri;
4b5f931b
MD
1482 }
1483 crit_exit();
1484}
1485
f9235b6d 1486/*
85946b6c 1487 * hz tick scheduler clock for LWKT threads
f9235b6d
MD
1488 */
1489void
85946b6c 1490lwkt_schedulerclock(thread_t td)
f9235b6d 1491{
85946b6c
MD
1492 globaldata_t gd = td->td_gd;
1493 thread_t xtd;
2a418930 1494
85946b6c
MD
1495 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1496 /*
1497 * If the current thread is at the head of the runq shift it to the
1498 * end of any equal-priority threads and request a LWKT reschedule
1499 * if it moved.
1500 */
1501 xtd = TAILQ_NEXT(td, td_threadq);
1502 if (xtd && xtd->td_pri == td->td_pri) {
1503 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1504 while (xtd && xtd->td_pri == td->td_pri)
1505 xtd = TAILQ_NEXT(xtd, td_threadq);
1506 if (xtd)
1507 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1508 else
1509 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1510 need_lwkt_resched();
f9235b6d 1511 }
85946b6c
MD
1512 } else {
1513 /*
1514 * If we scheduled a thread other than the one at the head of the
1515 * queue always request a reschedule every tick.
1516 */
1517 need_lwkt_resched();
f9235b6d
MD
1518 }
1519}
1520
5d21b981 1521/*
52eedfb5
MD
1522 * Migrate the current thread to the specified cpu.
1523 *
cc9b6223
MD
1524 * This is accomplished by descheduling ourselves from the current cpu
1525 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1526 * 'old' thread wants to migrate after it has been completely switched out
1527 * and will complete the migration.
1528 *
1529 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1530 *
1531 * We must be sure to release our current process designation (if a user
1532 * process) before clearing out any tsleepq we are on because the release
1533 * code may re-add us.
ae8e83e6
MD
1534 *
1535 * We must be sure to remove ourselves from the current cpu's tsleepq
1536 * before potentially moving to another queue. The thread can be on
1537 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1538 */
5d21b981
MD
1539
1540void
1541lwkt_setcpu_self(globaldata_t rgd)
1542{
1543#ifdef SMP
1544 thread_t td = curthread;
1545
1546 if (td->td_gd != rgd) {
1547 crit_enter_quick(td);
cc9b6223 1548
95858b91
MD
1549 if (td->td_release)
1550 td->td_release(td);
ae8e83e6 1551 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1552 tsleep_remove(td);
cc9b6223
MD
1553
1554 /*
1555 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1556 * trying to deschedule ourselves and switch away, then deschedule
1557 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1558 * call lwkt_switch() to complete the operation.
1559 */
5d21b981
MD
1560 td->td_flags |= TDF_MIGRATING;
1561 lwkt_deschedule_self(td);
52eedfb5 1562 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1563 td->td_migrate_gd = rgd;
5d21b981 1564 lwkt_switch();
cc9b6223
MD
1565
1566 /*
1567 * We are now on the target cpu
1568 */
1569 KKASSERT(rgd == mycpu);
52eedfb5 1570 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1571 crit_exit_quick(td);
1572 }
1573#endif
1574}
1575
ecdefdda
MD
1576void
1577lwkt_migratecpu(int cpuid)
1578{
1579#ifdef SMP
1580 globaldata_t rgd;
1581
1582 rgd = globaldata_find(cpuid);
1583 lwkt_setcpu_self(rgd);
1584#endif
1585}
1586
cc9b6223 1587#ifdef SMP
5d21b981
MD
1588/*
1589 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
1590 * do not have to enter another one).
1591 *
1592 * The thread (td) has already been completely descheduled from the
1593 * originating cpu and we can simply assert the case. The thread is
1594 * assigned to the new cpu and enqueued.
5d21b981 1595 *
cc9b6223 1596 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1597 */
1598static void
1599lwkt_setcpu_remote(void *arg)
1600{
1601 thread_t td = arg;
1602 globaldata_t gd = mycpu;
1603
cc9b6223 1604 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1605 td->td_gd = gd;
562273ea 1606 cpu_mfence();
5d21b981 1607 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1608 KKASSERT(td->td_migrate_gd == NULL);
4643740a
MD
1609 KKASSERT(td->td_lwp == NULL ||
1610 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
5d21b981
MD
1611 _lwkt_enqueue(td);
1612}
3d28ff59 1613#endif
5d21b981 1614
553ea3c8 1615struct lwp *
4b5f931b
MD
1616lwkt_preempted_proc(void)
1617{
73e4f7b9 1618 thread_t td = curthread;
4b5f931b
MD
1619 while (td->td_preempted)
1620 td = td->td_preempted;
553ea3c8 1621 return(td->td_lwp);
4b5f931b
MD
1622}
1623
99df837e
MD
1624/*
1625 * Create a kernel process/thread/whatever. It shares it's address space
1626 * with proc0 - ie: kernel only.
1627 *
d2d8515b
MD
1628 * If the cpu is not specified one will be selected. In the future
1629 * specifying a cpu of -1 will enable kernel thread migration between
1630 * cpus.
99df837e
MD
1631 */
1632int
c9e9fb21
MD
1633lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1634 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1635{
73e4f7b9 1636 thread_t td;
e2565a42 1637 __va_list ap;
99df837e 1638
d3d32139 1639 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1640 tdflags);
a2a5ad0d
MD
1641 if (tdp)
1642 *tdp = td;
709799ea 1643 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1644
1645 /*
1646 * Set up arg0 for 'ps' etc
1647 */
e2565a42 1648 __va_start(ap, fmt);
379210cb 1649 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1650 __va_end(ap);
99df837e
MD
1651
1652 /*
1653 * Schedule the thread to run
1654 */
4643740a
MD
1655 if (td->td_flags & TDF_NOSTART)
1656 td->td_flags &= ~TDF_NOSTART;
ef0fdad1 1657 else
4643740a 1658 lwkt_schedule(td);
99df837e
MD
1659 return 0;
1660}
1661
1662/*
1663 * Destroy an LWKT thread. Warning! This function is not called when
1664 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1665 * uses a different reaping mechanism.
1666 */
1667void
1668lwkt_exit(void)
1669{
1670 thread_t td = curthread;
c070746a 1671 thread_t std;
8826f33a 1672 globaldata_t gd;
99df837e 1673
2883d2d8
MD
1674 /*
1675 * Do any cleanup that might block here
1676 */
99df837e 1677 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1678 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1679 caps_exit(td);
2883d2d8
MD
1680 biosched_done(td);
1681 dsched_exit_thread(td);
c070746a
MD
1682
1683 /*
1684 * Get us into a critical section to interlock gd_freetd and loop
1685 * until we can get it freed.
1686 *
1687 * We have to cache the current td in gd_freetd because objcache_put()ing
1688 * it would rip it out from under us while our thread is still active.
2af9d75d
MD
1689 *
1690 * We are the current thread so of course our own TDF_RUNNING bit will
1691 * be set, so unlike the lwp reap code we don't wait for it to clear.
c070746a
MD
1692 */
1693 gd = mycpu;
37af14fe 1694 crit_enter_quick(td);
2af9d75d
MD
1695 for (;;) {
1696 if (td->td_refs) {
1697 tsleep(td, 0, "tdreap", 1);
1698 continue;
1699 }
1700 if ((std = gd->gd_freetd) != NULL) {
1701 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1702 gd->gd_freetd = NULL;
1703 objcache_put(thread_cache, std);
1704 continue;
1705 }
1706 break;
c070746a 1707 }
3b4192fb
MD
1708
1709 /*
1710 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1711 * the last time. We cannot block after this point or we may end
1712 * up with a stale td on the tsleepq.
eb2adbf5
MD
1713 *
1714 * None of this may block, the critical section is the only thing
1715 * protecting tdallq and the only thing preventing new lwkt_hold()
1716 * thread refs now.
3b4192fb
MD
1717 */
1718 if (td->td_flags & TDF_TSLEEPQ)
1719 tsleep_remove(td);
37af14fe 1720 lwkt_deschedule_self(td);
e56e4dea 1721 lwkt_remove_tdallq(td);
74c9628e 1722 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1723
1724 /*
1725 * Final cleanup
1726 */
1727 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1728 if (td->td_flags & TDF_ALLOCATED_THREAD)
1729 gd->gd_freetd = td;
99df837e
MD
1730 cpu_thread_exit();
1731}
1732
e56e4dea
MD
1733void
1734lwkt_remove_tdallq(thread_t td)
1735{
1736 KKASSERT(td->td_gd == mycpu);
1737 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1738}
1739
9cf43f91
MD
1740/*
1741 * Code reduction and branch prediction improvements. Call/return
1742 * overhead on modern cpus often degenerates into 0 cycles due to
1743 * the cpu's branch prediction hardware and return pc cache. We
1744 * can take advantage of this by not inlining medium-complexity
1745 * functions and we can also reduce the branch prediction impact
1746 * by collapsing perfectly predictable branches into a single
1747 * procedure instead of duplicating it.
1748 *
1749 * Is any of this noticeable? Probably not, so I'll take the
1750 * smaller code size.
1751 */
1752void
b6468f56 1753crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1754{
b6468f56 1755 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1756}
1757
2d93b37a
MD
1758void
1759crit_panic(void)
1760{
1761 thread_t td = curthread;
850634cc 1762 int lcrit = td->td_critcount;
2d93b37a 1763
850634cc
AH
1764 td->td_critcount = 0;
1765 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1766 /* NOT REACHED */
2d93b37a
MD
1767}
1768
d165e668
MD
1769#ifdef SMP
1770
bd8015ca
MD
1771/*
1772 * Called from debugger/panic on cpus which have been stopped. We must still
1773 * process the IPIQ while stopped, even if we were stopped while in a critical
1774 * section (XXX).
1775 *
1776 * If we are dumping also try to process any pending interrupts. This may
1777 * or may not work depending on the state of the cpu at the point it was
1778 * stopped.
1779 */
1780void
1781lwkt_smp_stopped(void)
1782{
1783 globaldata_t gd = mycpu;
1784
1785 crit_enter_gd(gd);
1786 if (dumping) {
1787 lwkt_process_ipiq();
1788 splz();
1789 } else {
1790 lwkt_process_ipiq();
1791 }
1792 crit_exit_gd(gd);
1793}
1794
d165e668 1795#endif