kernel - Optimize shared -> excl spinlock contention
[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>
5b49787b 55#include <sys/indefinite.h>
9d265729
MD
56
57#include <sys/thread2.h>
58#include <sys/spinlock2.h>
5b49787b 59#include <sys/indefinite2.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>
3a06728e 74#include <machine/clock.h>
99df837e 75
8cee56f4
MD
76#ifdef _KERNEL_VIRTUAL
77#include <pthread.h>
78#endif
79
3a06728e
MD
80#define LOOPMASK
81
d850923c
AE
82#if !defined(KTR_CTXSW)
83#define KTR_CTXSW KTR_ALL
84#endif
85KTR_INFO_MASTER(ctxsw);
5bf48697
AE
86KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td);
87KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td);
88KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm);
89KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td);
1541028a 90
40aaf5fc
NT
91static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
92
0f7a3396
MD
93#ifdef INVARIANTS
94static int panic_on_cscount = 0;
95#endif
e28c8ef4
SW
96static int64_t switch_count = 0;
97static int64_t preempt_hit = 0;
98static int64_t preempt_miss = 0;
99static int64_t preempt_weird = 0;
fb0f29c4 100static int lwkt_use_spin_port;
40aaf5fc 101static struct objcache *thread_cache;
a46b4a23 102int cpu_mwait_spin = 0;
05220613 103
e381e77c 104static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
cc9b6223 105static void lwkt_setcpu_remote(void *arg);
e381e77c 106
fb0f29c4
MD
107/*
108 * We can make all thread ports use the spin backend instead of the thread
109 * backend. This should only be set to debug the spin backend.
110 */
111TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
112
0f7a3396 113#ifdef INVARIANTS
0c52fa62
SG
114SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
115 "Panic if attempting to switch lwkt's while mastering cpusync");
0f7a3396 116#endif
0c52fa62
SG
117SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
118 "Number of switched threads");
9733f757 119SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
0c52fa62 120 "Successful preemption events");
9733f757 121SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
0c52fa62
SG
122 "Failed preemption events");
123SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
124 "Number of preempted threads.");
b12defdc 125static int fairq_enable = 0;
2a418930
MD
126SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
127 &fairq_enable, 0, "Turn on fairq priority accumulators");
85946b6c 128static int fairq_bypass = -1;
b12defdc
MD
129SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
130 &fairq_bypass, 0, "Allow fairq to bypass td on token failure");
131extern int lwkt_sched_debug;
132int lwkt_sched_debug = 0;
133SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
134 &lwkt_sched_debug, 0, "Scheduler debug");
7fb451cb
MD
135static u_int lwkt_spin_loops = 10;
136SYSCTL_UINT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
b12defdc 137 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
fbc024e4 138static int preempt_enable = 1;
2a418930
MD
139SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
140 &preempt_enable, 0, "Enable preemption");
7b234d8c 141static int lwkt_cache_threads = 0;
765b1ae0
MD
142SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
143 &lwkt_cache_threads, 0, "thread+kstack cache");
fbc024e4 144
4b5f931b
MD
145/*
146 * These helper procedures handle the runq, they can only be called from
147 * within a critical section.
75cdbe6c
MD
148 *
149 * WARNING! Prior to SMP being brought up it is possible to enqueue and
150 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
151 * instead of 'mycpu' when referencing the globaldata structure. Once
152 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 153 */
f1d1c3fa
MD
154static __inline
155void
156_lwkt_dequeue(thread_t td)
157{
158 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 159 struct globaldata *gd = td->td_gd;
4b5f931b 160
f1d1c3fa 161 td->td_flags &= ~TDF_RUNQ;
f9235b6d 162 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
de4d4cb0 163 --gd->gd_tdrunqcount;
f9235b6d 164 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
2a418930 165 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
166 }
167}
168
f9235b6d
MD
169/*
170 * Priority enqueue.
171 *
d992c377
MD
172 * There are a limited number of lwkt threads runnable since user
173 * processes only schedule one at a time per cpu. However, there can
174 * be many user processes in kernel mode exiting from a tsleep() which
e3e6be1f 175 * become runnable.
d992c377 176 *
e6b81333
MD
177 * We scan the queue in both directions to help deal with degenerate
178 * situations when hundreds or thousands (or more) threads are runnable.
179 *
d992c377
MD
180 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and
181 * will ignore user priority. This is to ensure that user threads in
182 * kernel mode get cpu at some point regardless of what the user
183 * scheduler thinks.
f9235b6d 184 */
f1d1c3fa
MD
185static __inline
186void
187_lwkt_enqueue(thread_t td)
188{
e6b81333
MD
189 thread_t xtd; /* forward scan */
190 thread_t rtd; /* reverse scan */
f9235b6d 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 {
e3e6be1f
MD
201 /*
202 * NOTE: td_upri - higher numbers more desireable, same sense
203 * as td_pri (typically reversed from lwp_upri).
204 *
205 * In the equal priority case we want the best selection
206 * at the beginning so the less desireable selections know
207 * that they have to setrunqueue/go-to-another-cpu, even
208 * though it means switching back to the 'best' selection.
209 * This also avoids degenerate situations when many threads
210 * are runnable or waking up at the same time.
211 *
212 * If upri matches exactly place at end/round-robin.
213 */
e6b81333
MD
214 rtd = TAILQ_LAST(&gd->gd_tdrunq, lwkt_queue);
215
d992c377 216 while (xtd &&
e6b81333 217 (xtd->td_pri > td->td_pri ||
d992c377 218 (xtd->td_pri == td->td_pri &&
e3e6be1f 219 xtd->td_upri >= td->td_upri))) {
85946b6c 220 xtd = TAILQ_NEXT(xtd, td_threadq);
e6b81333
MD
221
222 /*
223 * Doing a reverse scan at the same time is an optimization
224 * for the insert-closer-to-tail case that avoids having to
225 * scan the entire list. This situation can occur when
226 * thousands of threads are woken up at the same time.
227 */
228 if (rtd->td_pri > td->td_pri ||
229 (rtd->td_pri == td->td_pri &&
230 rtd->td_upri >= td->td_upri)) {
231 TAILQ_INSERT_AFTER(&gd->gd_tdrunq, rtd, td, td_threadq);
232 goto skip;
233 }
234 rtd = TAILQ_PREV(rtd, lwkt_queue, td_threadq);
d992c377 235 }
85946b6c
MD
236 if (xtd)
237 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
238 else
239 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
f9235b6d 240 }
e6b81333 241skip:
de4d4cb0 242 ++gd->gd_tdrunqcount;
b12defdc
MD
243
244 /*
85946b6c 245 * Request a LWKT reschedule if we are now at the head of the queue.
b12defdc 246 */
85946b6c
MD
247 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
248 need_lwkt_resched();
f1d1c3fa
MD
249 }
250}
8ad65e08 251
e28c8ef4 252static boolean_t
40aaf5fc
NT
253_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
254{
255 struct thread *td = (struct thread *)obj;
256
257 td->td_kstack = NULL;
258 td->td_kstack_size = 0;
259 td->td_flags = TDF_ALLOCATED_THREAD;
4643740a 260 td->td_mpflags = 0;
40aaf5fc
NT
261 return (1);
262}
263
264static void
265_lwkt_thread_dtor(void *obj, void *privdata)
266{
267 struct thread *td = (struct thread *)obj;
268
269 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
270 ("_lwkt_thread_dtor: not allocated from objcache"));
271 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
272 td->td_kstack_size > 0,
273 ("_lwkt_thread_dtor: corrupted stack"));
274 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
7b234d8c
MD
275 td->td_kstack = NULL;
276 td->td_flags = 0;
40aaf5fc
NT
277}
278
279/*
280 * Initialize the lwkt s/system.
765b1ae0 281 *
7b234d8c
MD
282 * Nominally cache up to 32 thread + kstack structures. Cache more on
283 * systems with a lot of cpu cores.
40aaf5fc 284 */
ced589cb 285static void
40aaf5fc
NT
286lwkt_init(void)
287{
765b1ae0 288 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
7b234d8c
MD
289 if (lwkt_cache_threads == 0) {
290 lwkt_cache_threads = ncpus * 4;
291 if (lwkt_cache_threads < 32)
292 lwkt_cache_threads = 32;
293 }
765b1ae0
MD
294 thread_cache = objcache_create_mbacked(
295 M_THREAD, sizeof(struct thread),
2fce2579 296 0, lwkt_cache_threads,
765b1ae0 297 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc 298}
ced589cb 299SYSINIT(lwkt_init, SI_BOOT2_LWKT_INIT, SI_ORDER_FIRST, lwkt_init, NULL);
40aaf5fc 300
37af14fe
MD
301/*
302 * Schedule a thread to run. As the current thread we can always safely
303 * schedule ourselves, and a shortcut procedure is provided for that
304 * function.
305 *
306 * (non-blocking, self contained on a per cpu basis)
307 */
308void
309lwkt_schedule_self(thread_t td)
310{
cfaeae2a 311 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 312 crit_enter_quick(td);
f9235b6d
MD
313 KASSERT(td != &td->td_gd->gd_idlethread,
314 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
4643740a
MD
315 KKASSERT(td->td_lwp == NULL ||
316 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
37af14fe 317 _lwkt_enqueue(td);
37af14fe
MD
318 crit_exit_quick(td);
319}
320
321/*
322 * Deschedule a thread.
323 *
324 * (non-blocking, self contained on a per cpu basis)
325 */
326void
327lwkt_deschedule_self(thread_t td)
328{
329 crit_enter_quick(td);
37af14fe
MD
330 _lwkt_dequeue(td);
331 crit_exit_quick(td);
332}
333
8ad65e08
MD
334/*
335 * LWKTs operate on a per-cpu basis
336 *
73e4f7b9 337 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
338 */
339void
340lwkt_gdinit(struct globaldata *gd)
341{
f9235b6d 342 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 343 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
344}
345
7d0bac62
MD
346/*
347 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
348 * or LWKT start address before it can be scheduled. If the target cpu is
349 * -1 the thread will be created on the current cpu.
0cfcada1
MD
350 *
351 * If you intend to create a thread without a process context this function
352 * does everything except load the startup and switcher function.
7d0bac62
MD
353 */
354thread_t
d3d32139 355lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 356{
d2d8515b 357 static int cpu_rotator;
c070746a 358 globaldata_t gd = mycpu;
99df837e 359 void *stack;
7d0bac62 360
c070746a
MD
361 /*
362 * If static thread storage is not supplied allocate a thread. Reuse
363 * a cached free thread if possible. gd_freetd is used to keep an exiting
364 * thread intact through the exit.
365 */
ef0fdad1 366 if (td == NULL) {
cf709dd2
MD
367 crit_enter_gd(gd);
368 if ((td = gd->gd_freetd) != NULL) {
369 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
370 TDF_RUNQ)) == 0);
c070746a 371 gd->gd_freetd = NULL;
cf709dd2 372 } else {
c070746a 373 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
374 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
375 TDF_RUNQ)) == 0);
376 }
377 crit_exit_gd(gd);
40aaf5fc 378 KASSERT((td->td_flags &
2af9d75d
MD
379 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
380 TDF_ALLOCATED_THREAD,
40aaf5fc
NT
381 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
382 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 383 }
c070746a
MD
384
385 /*
386 * Try to reuse cached stack.
387 */
f470d0c8
MD
388 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
389 if (flags & TDF_ALLOCATED_STACK) {
e4846942 390 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
391 stack = NULL;
392 }
393 }
394 if (stack == NULL) {
070a58b3
MD
395 if (cpu < 0)
396 stack = (void *)kmem_alloc_stack(&kernel_map, stksize, 0);
397 else
398 stack = (void *)kmem_alloc_stack(&kernel_map, stksize,
399 KM_CPU(cpu));
ef0fdad1 400 flags |= TDF_ALLOCATED_STACK;
99df837e 401 }
d2d8515b
MD
402 if (cpu < 0) {
403 cpu = ++cpu_rotator;
404 cpu_ccfence();
5dd85b08 405 cpu = (uint32_t)cpu % (uint32_t)ncpus;
d2d8515b
MD
406 }
407 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 408 return(td);
7d0bac62
MD
409}
410
411/*
412 * Initialize a preexisting thread structure. This function is used by
413 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
414 *
f8c3996b
MD
415 * All threads start out in a critical section at a priority of
416 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
417 * appropriate. This function may send an IPI message when the
418 * requested cpu is not the current cpu and consequently gd_tdallq may
419 * not be initialized synchronously from the point of view of the originating
420 * cpu.
421 *
422 * NOTE! we have to be careful in regards to creating threads for other cpus
423 * if SMP has not yet been activated.
7d0bac62 424 */
75cdbe6c
MD
425static void
426lwkt_init_thread_remote(void *arg)
427{
428 thread_t td = arg;
429
52eedfb5
MD
430 /*
431 * Protected by critical section held by IPI dispatch
432 */
75cdbe6c
MD
433 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
434}
435
fdce8919
MD
436/*
437 * lwkt core thread structural initialization.
438 *
439 * NOTE: All threads are initialized as mpsafe threads.
440 */
7d0bac62 441void
f470d0c8
MD
442lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
443 struct globaldata *gd)
7d0bac62 444{
37af14fe
MD
445 globaldata_t mygd = mycpu;
446
99df837e
MD
447 bzero(td, sizeof(struct thread));
448 td->td_kstack = stack;
f470d0c8 449 td->td_kstack_size = stksize;
d3d32139 450 td->td_flags = flags;
4643740a 451 td->td_mpflags = 0;
f256b6c0 452 td->td_type = TD_TYPE_GENERIC;
26a0694b 453 td->td_gd = gd;
f9235b6d
MD
454 td->td_pri = TDPRI_KERN_DAEMON;
455 td->td_critcount = 1;
54341a3b 456 td->td_toks_have = NULL;
3b998fa9 457 td->td_toks_stop = &td->td_toks_base;
c068fb59
SZ
458 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) {
459 lwkt_initport_spin(&td->td_msgport, td,
460 (flags & TDF_FIXEDCPU) ? TRUE : FALSE);
461 } else {
fb0f29c4 462 lwkt_initport_thread(&td->td_msgport, td);
c068fb59 463 }
99df837e 464 pmap_init_thread(td);
5d21b981
MD
465 /*
466 * Normally initializing a thread for a remote cpu requires sending an
467 * IPI. However, the idlethread is setup before the other cpus are
468 * activated so we have to treat it as a special case. XXX manipulation
469 * of gd_tdallq requires the BGL.
470 */
471 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 472 crit_enter_gd(mygd);
75cdbe6c 473 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 474 crit_exit_gd(mygd);
75cdbe6c 475 } else {
2db3b277 476 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 477 }
3573cf7b 478 dsched_enter_thread(td);
73e4f7b9
MD
479}
480
481void
482lwkt_set_comm(thread_t td, const char *ctl, ...)
483{
e2565a42 484 __va_list va;
73e4f7b9 485
e2565a42 486 __va_start(va, ctl);
379210cb 487 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 488 __va_end(va);
5bf48697 489 KTR_LOG(ctxsw_newtd, td, td->td_comm);
7d0bac62
MD
490}
491
eb2adbf5
MD
492/*
493 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
494 * this does not prevent the thread from migrating to another cpu so the
495 * gd_tdallq state is not protected by this.
496 */
99df837e 497void
73e4f7b9 498lwkt_hold(thread_t td)
99df837e 499{
74c9628e 500 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
501}
502
503void
504lwkt_rele(thread_t td)
505{
506 KKASSERT(td->td_refs > 0);
74c9628e 507 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
508}
509
73e4f7b9
MD
510void
511lwkt_free_thread(thread_t td)
512{
74c9628e 513 KKASSERT(td->td_refs == 0);
c17a6852
MD
514 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
515 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
40aaf5fc
NT
516 if (td->td_flags & TDF_ALLOCATED_THREAD) {
517 objcache_put(thread_cache, td);
518 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
519 /* client-allocated struct with internally allocated stack */
520 KASSERT(td->td_kstack && td->td_kstack_size > 0,
521 ("lwkt_free_thread: corrupted stack"));
522 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
523 td->td_kstack = NULL;
524 td->td_kstack_size = 0;
99df837e 525 }
a86ce0cd 526
e7c0dbba 527 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
528}
529
530
8ad65e08
MD
531/*
532 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
533 * switch to the idlethread. Switching must occur within a critical
534 * section to avoid races with the scheduling queue.
535 *
536 * We always have full control over our cpu's run queue. Other cpus
537 * that wish to manipulate our queue must use the cpu_*msg() calls to
538 * talk to our cpu, so a critical section is all that is needed and
539 * the result is very, very fast thread switching.
540 *
96728c05
MD
541 * The LWKT scheduler uses a fixed priority model and round-robins at
542 * each priority level. User process scheduling is a totally
543 * different beast and LWKT priorities should not be confused with
544 * user process priorities.
f1d1c3fa 545 *
69d78e99
MD
546 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
547 * is not called by the current thread in the preemption case, only when
548 * the preempting thread blocks (in order to return to the original thread).
cfaeae2a
MD
549 *
550 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
551 * migration and tsleep deschedule the current lwkt thread and call
552 * lwkt_switch(). In particular, the target cpu of the migration fully
553 * expects the thread to become non-runnable and can deadlock against
554 * cpusync operations if we run any IPIs prior to switching the thread out.
555 *
556 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
95858b91 557 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
8ad65e08
MD
558 */
559void
560lwkt_switch(void)
561{
37af14fe
MD
562 globaldata_t gd = mycpu;
563 thread_t td = gd->gd_curthread;
8ad65e08 564 thread_t ntd;
c6a766f4 565 thread_t xtd;
5411d8f1 566 int upri;
3a06728e
MD
567#ifdef LOOPMASK
568 uint64_t tsc_base = rdtsc();
569#endif
8ad65e08 570
da0b0e8b 571 KKASSERT(gd->gd_processing_ipiq == 0);
121f93bc 572 KKASSERT(td->td_flags & TDF_RUNNING);
da0b0e8b 573
46a3f46d 574 /*
27e88a6e
MD
575 * Switching from within a 'fast' (non thread switched) interrupt or IPI
576 * is illegal. However, we may have to do it anyway if we hit a fatal
577 * kernel trap or we have paniced.
578 *
579 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 580 */
27e88a6e
MD
581 if (gd->gd_intr_nesting_level) {
582 int savegdnest;
583 int savegdtrap;
584
5fddbda2 585 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22 586 panic("lwkt_switch: Attempt to switch from a "
5a8df152 587 "fast interrupt, ipi, or hard code section, "
4a28fe22
MD
588 "td %p\n",
589 td);
27e88a6e
MD
590 } else {
591 savegdnest = gd->gd_intr_nesting_level;
592 savegdtrap = gd->gd_trap_nesting_level;
593 gd->gd_intr_nesting_level = 0;
594 gd->gd_trap_nesting_level = 0;
a7422615
MD
595 if ((td->td_flags & TDF_PANICWARN) == 0) {
596 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
597 kprintf("Warning: thread switch from interrupt, IPI, "
598 "or hard code section.\n"
a7422615 599 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 600 print_backtrace(-1);
a7422615 601 }
27e88a6e
MD
602 lwkt_switch();
603 gd->gd_intr_nesting_level = savegdnest;
604 gd->gd_trap_nesting_level = savegdtrap;
605 return;
606 }
96728c05 607 }
ef0fdad1 608
cb973d15 609 /*
85946b6c
MD
610 * Release our current user process designation if we are blocking
611 * or if a user reschedule was requested.
612 *
613 * NOTE: This function is NOT called if we are switching into or
614 * returning from a preemption.
615 *
616 * NOTE: Releasing our current user process designation may cause
617 * it to be assigned to another thread, which in turn will
618 * cause us to block in the usched acquire code when we attempt
619 * to return to userland.
620 *
621 * NOTE: On SMP systems this can be very nasty when heavy token
622 * contention is present so we want to be careful not to
623 * release the designation gratuitously.
cb973d15 624 */
85946b6c
MD
625 if (td->td_release &&
626 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
cb973d15 627 td->td_release(td);
85946b6c 628 }
cb973d15 629
85946b6c 630 /*
5411d8f1
MD
631 * Release all tokens. Once we do this we must remain in the critical
632 * section and cannot run IPIs or other interrupts until we switch away
633 * because they may implode if they try to get a token using our thread
634 * context.
85946b6c 635 */
37af14fe 636 crit_enter_gd(gd);
3b998fa9 637 if (TD_TOKS_HELD(td))
9d265729
MD
638 lwkt_relalltokens(td);
639
640 /*
b02926de
MD
641 * We had better not be holding any spin locks, but don't get into an
642 * endless panic loop.
9d265729 643 */
0846e4ce 644 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL,
d666840a 645 ("lwkt_switch: still holding %d exclusive spinlocks!",
0846e4ce 646 gd->gd_spinlocks));
9d265729 647
0f7a3396
MD
648#ifdef INVARIANTS
649 if (td->td_cscount) {
6ea70f76 650 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
651 td);
652 if (panic_on_cscount)
653 panic("switching while mastering cpusync");
654 }
8a8d5d85 655#endif
f9235b6d
MD
656
657 /*
658 * If we had preempted another thread on this cpu, resume the preempted
659 * thread. This occurs transparently, whether the preempted thread
660 * was scheduled or not (it may have been preempted after descheduling
661 * itself).
662 *
663 * We have to setup the MP lock for the original thread after backing
664 * out the adjustment that was made to curthread when the original
665 * was preempted.
666 */
99df837e 667 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
668 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
669 ntd->td_flags |= TDF_PREEMPT_DONE;
7fb451cb 670 ntd->td_contended = 0; /* reset contended */
8ec60c3f
MD
671
672 /*
b9eb1c19
MD
673 * The interrupt may have woken a thread up, we need to properly
674 * set the reschedule flag if the originally interrupted thread is
675 * at a lower priority.
85946b6c 676 *
c6a766f4
MD
677 * NOTE: The interrupt may not have descheduled ntd.
678 *
679 * NOTE: We do not reschedule if there are no threads on the runq.
680 * (ntd could be the idlethread).
8ec60c3f 681 */
c6a766f4
MD
682 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
683 if (xtd && xtd != ntd)
8ec60c3f 684 need_lwkt_resched();
f9235b6d
MD
685 goto havethread_preempted;
686 }
687
b12defdc 688 /*
5411d8f1
MD
689 * Figure out switch target. If we cannot switch to our desired target
690 * look for a thread that we can switch to.
cfaeae2a 691 *
5411d8f1
MD
692 * NOTE! The limited spin loop and related parameters are extremely
693 * important for system performance, particularly for pipes and
694 * concurrent conflicting VM faults.
f9235b6d 695 */
5411d8f1
MD
696 clear_lwkt_resched();
697 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
698
699 if (ntd) {
700 do {
701 if (TD_TOKS_NOT_HELD(ntd) ||
702 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops)))
703 {
704 goto havethread;
705 }
7fb451cb 706 ++ntd->td_contended; /* overflow ok */
3a06728e
MD
707#ifdef LOOPMASK
708 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
709 kprintf("lwkt_switch: excessive contended %d "
710 "thread %p\n", ntd->td_contended, ntd);
711 tsc_base = rdtsc();
712 }
713#endif
5411d8f1
MD
714 } while (ntd->td_contended < (lwkt_spin_loops >> 1));
715 upri = ntd->td_upri;
f9235b6d 716
f9235b6d 717 /*
5411d8f1
MD
718 * Bleh, the thread we wanted to switch to has a contended token.
719 * See if we can switch to another thread.
2a418930 720 *
5411d8f1
MD
721 * We generally don't want to do this because it represents a
722 * priority inversion. Do not allow the case if the thread
723 * is returning to userland (not a kernel thread) AND the thread
724 * has a lower upri.
f9235b6d 725 */
b12defdc 726 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
5411d8f1
MD
727 if (ntd->td_pri < TDPRI_KERN_LPSCHED && upri > ntd->td_upri)
728 break;
729 upri = ntd->td_upri;
b12defdc 730
5411d8f1
MD
731 /*
732 * Try this one.
733 */
734 if (TD_TOKS_NOT_HELD(ntd) ||
735 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) {
736 goto havethread;
737 }
7fb451cb 738 ++ntd->td_contended; /* overflow ok */
b12defdc 739 }
b12defdc
MD
740
741 /*
5411d8f1
MD
742 * Fall through, switch to idle thread to get us out of the current
743 * context. Since we were contended, prevent HLT by flagging a
744 * LWKT reschedule.
b12defdc 745 */
5411d8f1 746 need_lwkt_resched();
f1d1c3fa 747 }
8a8d5d85 748
5411d8f1
MD
749 /*
750 * We either contended on ntd or the runq is empty. We must switch
751 * through the idle thread to get out of the current context.
752 */
753 ntd = &gd->gd_idlethread;
754 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
755 ASSERT_NO_TOKENS_HELD(ntd);
756 cpu_time.cp_msg[0] = 0;
5411d8f1
MD
757 goto haveidle;
758
2a418930 759havethread:
b12defdc 760 /*
be71787b
MD
761 * Clear gd_idle_repeat when doing a normal switch to a non-idle
762 * thread.
f9235b6d 763 */
9ac1ee6e 764 ntd->td_wmesg = NULL;
7fb451cb 765 ntd->td_contended = 0; /* reset once scheduled */
b12defdc 766 ++gd->gd_cnt.v_swtch;
be71787b 767 gd->gd_idle_repeat = 0;
2a418930 768
5b49787b
MD
769 /*
770 * If we were busy waiting record final disposition
771 */
772 if (ntd->td_indefinite.type)
773 indefinite_done(&ntd->td_indefinite);
774
f9235b6d 775havethread_preempted:
f9235b6d
MD
776 /*
777 * If the new target does not need the MP lock and we are holding it,
778 * release the MP lock. If the new target requires the MP lock we have
779 * already acquired it for the target.
8a8d5d85 780 */
2a418930 781 ;
f9235b6d
MD
782haveidle:
783 KASSERT(ntd->td_critcount,
b5d16701
MD
784 ("priority problem in lwkt_switch %d %d",
785 td->td_critcount, ntd->td_critcount));
786
94f6d86e 787 if (td != ntd) {
cc9b6223
MD
788 /*
789 * Execute the actual thread switch operation. This function
790 * returns to the current thread and returns the previous thread
791 * (which may be different from the thread we switched to).
792 *
793 * We are responsible for marking ntd as TDF_RUNNING.
794 */
121f93bc 795 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
94f6d86e 796 ++switch_count;
a1f0fb66 797 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
798 ntd->td_flags |= TDF_RUNNING;
799 lwkt_switch_return(td->td_switch(ntd));
800 /* ntd invalid, td_switch() can return a different thread_t */
94f6d86e 801 }
b12defdc 802
b12defdc 803 /*
54341a3b 804 * catch-all. XXX is this strictly needed?
b12defdc
MD
805 */
806 splz_check();
54341a3b 807
37af14fe
MD
808 /* NOTE: current cpu may have changed after switch */
809 crit_exit_quick(td);
8ad65e08
MD
810}
811
cc9b6223
MD
812/*
813 * Called by assembly in the td_switch (thread restore path) for thread
814 * bootstrap cases which do not 'return' to lwkt_switch().
815 */
816void
817lwkt_switch_return(thread_t otd)
818{
cc9b6223 819 globaldata_t rgd;
3a06728e
MD
820#ifdef LOOPMASK
821 uint64_t tsc_base = rdtsc();
822#endif
823 int exiting;
824
825 exiting = otd->td_flags & TDF_EXITING;
826 cpu_ccfence();
cc9b6223
MD
827
828 /*
829 * Check if otd was migrating. Now that we are on ntd we can finish
830 * up the migration. This is a bit messy but it is the only place
831 * where td is known to be fully descheduled.
832 *
833 * We can only activate the migration if otd was migrating but not
834 * held on the cpu due to a preemption chain. We still have to
835 * clear TDF_RUNNING on the old thread either way.
836 *
837 * We are responsible for clearing the previously running thread's
838 * TDF_RUNNING.
839 */
840 if ((rgd = otd->td_migrate_gd) != NULL &&
841 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
842 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
843 (TDF_MIGRATING | TDF_RUNNING));
844 otd->td_migrate_gd = NULL;
845 otd->td_flags &= ~TDF_RUNNING;
846 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
847 } else {
848 otd->td_flags &= ~TDF_RUNNING;
849 }
2b07d9aa
MD
850
851 /*
852 * Final exit validations (see lwp_wait()). Note that otd becomes
853 * invalid the *instant* we set TDF_MP_EXITSIG.
3a06728e
MD
854 *
855 * Use the EXITING status loaded from before we clear TDF_RUNNING,
856 * because if it is not set otd becomes invalid the instant we clear
857 * TDF_RUNNING on it (otherwise, if the system is fast enough, we
858 * might 'steal' TDF_EXITING from another switch-return!).
2b07d9aa 859 */
3a06728e 860 while (exiting) {
2b07d9aa
MD
861 u_int mpflags;
862
863 mpflags = otd->td_mpflags;
864 cpu_ccfence();
865
866 if (mpflags & TDF_MP_EXITWAIT) {
867 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
868 mpflags | TDF_MP_EXITSIG)) {
869 wakeup(otd);
870 break;
871 }
872 } else {
873 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
874 mpflags | TDF_MP_EXITSIG)) {
875 wakeup(otd);
876 break;
877 }
878 }
3a06728e
MD
879
880#ifdef LOOPMASK
881 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
882 kprintf("lwkt_switch_return: excessive TDF_EXITING "
883 "thread %p\n", otd);
884 tsc_base = rdtsc();
885 }
886#endif
2b07d9aa 887 }
cc9b6223
MD
888}
889
b68b7282 890/*
96728c05 891 * Request that the target thread preempt the current thread. Preemption
203592a0
MD
892 * can only occur only:
893 *
894 * - If our critical section is the one that we were called with
895 * - The relative priority of the target thread is higher
896 * - The target is not excessively interrupt-nested via td_nest_count
897 * - The target thread holds no tokens.
898 * - The target thread is not already scheduled and belongs to the
899 * current cpu.
900 * - The current thread is not holding any spin-locks.
96728c05
MD
901 *
902 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
903 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 904 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
905 * to determine whether preemption is possible (aka usually just the crit
906 * priority of lwkt_schedule() itself).
b68b7282 907 *
54341a3b
MD
908 * Preemption is typically limited to interrupt threads.
909 *
910 * Operation works in a fairly straight-forward manner. The normal
911 * scheduling code is bypassed and we switch directly to the target
912 * thread. When the target thread attempts to block or switch away
913 * code at the base of lwkt_switch() will switch directly back to our
914 * thread. Our thread is able to retain whatever tokens it holds and
915 * if the target needs one of them the target will switch back to us
916 * and reschedule itself normally.
b68b7282
MD
917 */
918void
f9235b6d 919lwkt_preempt(thread_t ntd, int critcount)
b68b7282 920{
46a3f46d 921 struct globaldata *gd = mycpu;
cc9b6223 922 thread_t xtd;
0a3f9b47 923 thread_t td;
2d910aaf 924 int save_gd_intr_nesting_level;
b68b7282 925
26a0694b 926 /*
96728c05
MD
927 * The caller has put us in a critical section. We can only preempt
928 * if the caller of the caller was not in a critical section (basically
f9235b6d 929 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 930 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 931 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
932 *
933 * YYY The target thread must be in a critical section (else it must
934 * inherit our critical section? I dunno yet).
26a0694b 935 */
f9235b6d 936 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 937
b12defdc 938 td = gd->gd_curthread;
fbc024e4
MD
939 if (preempt_enable == 0) {
940 ++preempt_miss;
941 return;
942 }
f9235b6d 943 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
944 ++preempt_miss;
945 return;
946 }
f9235b6d 947 if (td->td_critcount > critcount) {
96728c05
MD
948 ++preempt_miss;
949 return;
950 }
203592a0
MD
951 if (td->td_nest_count >= 2) {
952 ++preempt_miss;
953 return;
954 }
121f93bc
MD
955 if (td->td_cscount) {
956 ++preempt_miss;
957 return;
958 }
46a3f46d 959 if (ntd->td_gd != gd) {
96728c05
MD
960 ++preempt_miss;
961 return;
962 }
ee89e80b 963
41a01a4d 964 /*
77912481
MD
965 * We don't have to check spinlocks here as they will also bump
966 * td_critcount.
d3d1cbc8
MD
967 *
968 * Do not try to preempt if the target thread is holding any tokens.
969 * We could try to acquire the tokens but this case is so rare there
970 * is no need to support it.
41a01a4d 971 */
0846e4ce 972 KKASSERT(gd->gd_spinlocks == 0);
77912481 973
3b998fa9 974 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8 975 ++preempt_miss;
d3d1cbc8
MD
976 return;
977 }
26a0694b
MD
978 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
979 ++preempt_weird;
980 return;
981 }
982 if (ntd->td_preempted) {
4b5f931b 983 ++preempt_hit;
26a0694b 984 return;
b68b7282 985 }
da0b0e8b 986 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 987
8ec60c3f
MD
988 /*
989 * Since we are able to preempt the current thread, there is no need to
990 * call need_lwkt_resched().
2d910aaf
MD
991 *
992 * We must temporarily clear gd_intr_nesting_level around the switch
993 * since switchouts from the target thread are allowed (they will just
994 * return to our thread), and since the target thread has its own stack.
cc9b6223
MD
995 *
996 * A preemption must switch back to the original thread, assert the
997 * case.
8ec60c3f 998 */
26a0694b
MD
999 ++preempt_hit;
1000 ntd->td_preempted = td;
1001 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1002 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1003 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1004 gd->gd_intr_nesting_level = 0;
121f93bc
MD
1005
1006 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
cc9b6223
MD
1007 ntd->td_flags |= TDF_RUNNING;
1008 xtd = td->td_switch(ntd);
1009 KKASSERT(xtd == ntd);
1010 lwkt_switch_return(xtd);
2d910aaf 1011 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1012
26a0694b
MD
1013 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1014 ntd->td_preempted = NULL;
1015 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1016}
1017
f1d1c3fa 1018/*
faaeffac 1019 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1020 * This will work inside a critical section but not inside a hard code
1021 * section.
ef0fdad1 1022 *
f1d1c3fa
MD
1023 * (self contained on a per cpu basis)
1024 */
1025void
faaeffac 1026splz_check(void)
f1d1c3fa 1027{
7966cb69
MD
1028 globaldata_t gd = mycpu;
1029 thread_t td = gd->gd_curthread;
ef0fdad1 1030
4a28fe22
MD
1031 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1032 gd->gd_intr_nesting_level == 0 &&
1033 td->td_nest_count < 2)
1034 {
f1d1c3fa 1035 splz();
4a28fe22
MD
1036 }
1037}
1038
1039/*
1040 * This version is integrated into crit_exit, reqflags has already
1041 * been tested but td_critcount has not.
1042 *
1043 * We only want to execute the splz() on the 1->0 transition of
1044 * critcount and not in a hard code section or if too deeply nested.
925040f2 1045 *
0846e4ce 1046 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
4a28fe22
MD
1047 */
1048void
1049lwkt_maybe_splz(thread_t td)
1050{
1051 globaldata_t gd = td->td_gd;
1052
1053 if (td->td_critcount == 0 &&
1054 gd->gd_intr_nesting_level == 0 &&
1055 td->td_nest_count < 2)
1056 {
1057 splz();
1058 }
f1d1c3fa
MD
1059}
1060
e6546af9
MD
1061/*
1062 * Drivers which set up processing co-threads can call this function to
1063 * run the co-thread at a higher priority and to allow it to preempt
1064 * normal threads.
1065 */
1066void
1067lwkt_set_interrupt_support_thread(void)
1068{
1069 thread_t td = curthread;
1070
1071 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1072 td->td_flags |= TDF_INTTHREAD;
1073 td->td_preemptable = lwkt_preempt;
1074}
1075
1076
8ad65e08 1077/*
f9235b6d
MD
1078 * This function is used to negotiate a passive release of the current
1079 * process/lwp designation with the user scheduler, allowing the user
1080 * scheduler to schedule another user thread. The related kernel thread
1081 * (curthread) continues running in the released state.
8ad65e08
MD
1082 */
1083void
f9235b6d 1084lwkt_passive_release(struct thread *td)
8ad65e08 1085{
f9235b6d
MD
1086 struct lwp *lp = td->td_lwp;
1087
1088 td->td_release = NULL;
1089 lwkt_setpri_self(TDPRI_KERN_USER);
d992c377 1090
f9235b6d 1091 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1092}
1093
f9235b6d 1094
3824f392 1095/*
d2d8515b
MD
1096 * This implements a LWKT yield, allowing a kernel thread to yield to other
1097 * kernel threads at the same or higher priority. This function can be
1098 * called in a tight loop and will typically only yield once per tick.
f9235b6d 1099 *
d2d8515b
MD
1100 * Most kernel threads run at the same priority in order to allow equal
1101 * sharing.
f9235b6d
MD
1102 *
1103 * (self contained on a per cpu basis)
3824f392
MD
1104 */
1105void
f9235b6d 1106lwkt_yield(void)
3824f392 1107{
f9235b6d
MD
1108 globaldata_t gd = mycpu;
1109 thread_t td = gd->gd_curthread;
3824f392 1110
fbe96076
MD
1111 /*
1112 * Should never be called with spinlocks held but there is a path
1113 * via ACPI where it might happen.
1114 */
1115 if (gd->gd_spinlocks)
1116 return;
1117
1118 /*
1119 * Safe to call splz if we are not too-heavily nested.
1120 */
f9235b6d
MD
1121 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1122 splz();
fbe96076
MD
1123
1124 /*
1125 * Caller allows switching
1126 */
85946b6c 1127 if (lwkt_resched_wanted()) {
f9235b6d
MD
1128 lwkt_schedule_self(curthread);
1129 lwkt_switch();
f9235b6d 1130 }
3824f392
MD
1131}
1132
40504122
MD
1133/*
1134 * The quick version processes pending interrupts and higher-priority
1135 * LWKT threads but will not round-robin same-priority LWKT threads.
de4d4cb0
MD
1136 *
1137 * When called while attempting to return to userland the only same-pri
1138 * threads are the ones which have already tried to become the current
1139 * user process.
40504122
MD
1140 */
1141void
1142lwkt_yield_quick(void)
1143{
1144 globaldata_t gd = mycpu;
1145 thread_t td = gd->gd_curthread;
1146
1147 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1148 splz();
1149 if (lwkt_resched_wanted()) {
9c99cb33 1150 crit_enter();
40504122
MD
1151 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1152 clear_lwkt_resched();
1153 } else {
1154 lwkt_schedule_self(curthread);
1155 lwkt_switch();
1156 }
9c99cb33 1157 crit_exit();
40504122
MD
1158 }
1159}
1160
3824f392 1161/*
f9235b6d
MD
1162 * This yield is designed for kernel threads with a user context.
1163 *
1164 * The kernel acting on behalf of the user is potentially cpu-bound,
1165 * this function will efficiently allow other threads to run and also
1166 * switch to other processes by releasing.
3824f392
MD
1167 *
1168 * The lwkt_user_yield() function is designed to have very low overhead
1169 * if no yield is determined to be needed.
1170 */
1171void
1172lwkt_user_yield(void)
1173{
f9235b6d
MD
1174 globaldata_t gd = mycpu;
1175 thread_t td = gd->gd_curthread;
1176
fbe96076
MD
1177 /*
1178 * Should never be called with spinlocks held but there is a path
1179 * via ACPI where it might happen.
1180 */
1181 if (gd->gd_spinlocks)
1182 return;
1183
f9235b6d
MD
1184 /*
1185 * Always run any pending interrupts in case we are in a critical
1186 * section.
1187 */
1188 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1189 splz();
3824f392 1190
3824f392 1191 /*
f9235b6d
MD
1192 * Switch (which forces a release) if another kernel thread needs
1193 * the cpu, if userland wants us to resched, or if our kernel
1194 * quantum has run out.
3824f392 1195 */
f9235b6d 1196 if (lwkt_resched_wanted() ||
85946b6c 1197 user_resched_wanted())
f9235b6d 1198 {
3824f392 1199 lwkt_switch();
3824f392
MD
1200 }
1201
f9235b6d 1202#if 0
3824f392 1203 /*
f9235b6d
MD
1204 * Reacquire the current process if we are released.
1205 *
1206 * XXX not implemented atm. The kernel may be holding locks and such,
1207 * so we want the thread to continue to receive cpu.
3824f392 1208 */
f9235b6d
MD
1209 if (td->td_release == NULL && lp) {
1210 lp->lwp_proc->p_usched->acquire_curproc(lp);
1211 td->td_release = lwkt_passive_release;
1212 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1213 }
f9235b6d 1214#endif
b9eb1c19
MD
1215}
1216
8ad65e08 1217/*
f1d1c3fa
MD
1218 * Generic schedule. Possibly schedule threads belonging to other cpus and
1219 * deal with threads that might be blocked on a wait queue.
1220 *
0a3f9b47
MD
1221 * We have a little helper inline function which does additional work after
1222 * the thread has been enqueued, including dealing with preemption and
1223 * setting need_lwkt_resched() (which prevents the kernel from returning
1224 * to userland until it has processed higher priority threads).
6330a558
MD
1225 *
1226 * It is possible for this routine to be called after a failed _enqueue
1227 * (due to the target thread migrating, sleeping, or otherwise blocked).
1228 * We have to check that the thread is actually on the run queue!
8ad65e08 1229 */
0a3f9b47
MD
1230static __inline
1231void
85946b6c 1232_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
0a3f9b47 1233{
6330a558 1234 if (ntd->td_flags & TDF_RUNQ) {
85946b6c 1235 if (ntd->td_preemptable) {
f9235b6d 1236 ntd->td_preemptable(ntd, ccount); /* YYY +token */
6330a558 1237 }
0a3f9b47
MD
1238 }
1239}
1240
361d01dd 1241static __inline
8ad65e08 1242void
85946b6c 1243_lwkt_schedule(thread_t td)
8ad65e08 1244{
37af14fe
MD
1245 globaldata_t mygd = mycpu;
1246
cf709dd2
MD
1247 KASSERT(td != &td->td_gd->gd_idlethread,
1248 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1249 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1250 crit_enter_gd(mygd);
4643740a
MD
1251 KKASSERT(td->td_lwp == NULL ||
1252 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1253
37af14fe 1254 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1255 _lwkt_enqueue(td);
1256 } else {
f1d1c3fa 1257 /*
7cd8d145
MD
1258 * If we own the thread, there is no race (since we are in a
1259 * critical section). If we do not own the thread there might
1260 * be a race but the target cpu will deal with it.
f1d1c3fa 1261 */
7cd8d145 1262 if (td->td_gd == mygd) {
9d265729 1263 _lwkt_enqueue(td);
85946b6c 1264 _lwkt_schedule_post(mygd, td, 1);
f1d1c3fa 1265 } else {
e381e77c 1266 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1267 }
8ad65e08 1268 }
37af14fe 1269 crit_exit_gd(mygd);
8ad65e08
MD
1270}
1271
361d01dd
MD
1272void
1273lwkt_schedule(thread_t td)
1274{
85946b6c 1275 _lwkt_schedule(td);
361d01dd
MD
1276}
1277
1278void
85946b6c 1279lwkt_schedule_noresched(thread_t td) /* XXX not impl */
361d01dd 1280{
85946b6c 1281 _lwkt_schedule(td);
361d01dd
MD
1282}
1283
e381e77c
MD
1284/*
1285 * When scheduled remotely if frame != NULL the IPIQ is being
1286 * run via doreti or an interrupt then preemption can be allowed.
1287 *
1288 * To allow preemption we have to drop the critical section so only
1289 * one is present in _lwkt_schedule_post.
1290 */
1291static void
1292lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1293{
1294 thread_t td = curthread;
1295 thread_t ntd = arg;
1296
1297 if (frame && ntd->td_preemptable) {
1298 crit_exit_noyield(td);
85946b6c 1299 _lwkt_schedule(ntd);
e381e77c
MD
1300 crit_enter_quick(td);
1301 } else {
85946b6c 1302 _lwkt_schedule(ntd);
e381e77c
MD
1303 }
1304}
1305
d9eea1a5 1306/*
52eedfb5
MD
1307 * Thread migration using a 'Pull' method. The thread may or may not be
1308 * the current thread. It MUST be descheduled and in a stable state.
1309 * lwkt_giveaway() must be called on the cpu owning the thread.
1310 *
1311 * At any point after lwkt_giveaway() is called, the target cpu may
1312 * 'pull' the thread by calling lwkt_acquire().
1313 *
ae8e83e6
MD
1314 * We have to make sure the thread is not sitting on a per-cpu tsleep
1315 * queue or it will blow up when it moves to another cpu.
1316 *
52eedfb5 1317 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1318 */
52eedfb5
MD
1319void
1320lwkt_giveaway(thread_t td)
1321{
3b4192fb 1322 globaldata_t gd = mycpu;
52eedfb5 1323
3b4192fb
MD
1324 crit_enter_gd(gd);
1325 if (td->td_flags & TDF_TSLEEPQ)
1326 tsleep_remove(td);
1327 KKASSERT(td->td_gd == gd);
1328 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1329 td->td_flags |= TDF_MIGRATING;
1330 crit_exit_gd(gd);
52eedfb5
MD
1331}
1332
a2a5ad0d
MD
1333void
1334lwkt_acquire(thread_t td)
1335{
37af14fe
MD
1336 globaldata_t gd;
1337 globaldata_t mygd;
a2a5ad0d 1338
52eedfb5 1339 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1340 gd = td->td_gd;
37af14fe 1341 mygd = mycpu;
52eedfb5 1342 if (gd != mycpu) {
3a06728e
MD
1343#ifdef LOOPMASK
1344 uint64_t tsc_base = rdtsc();
1345#endif
35238fa5 1346 cpu_lfence();
52eedfb5 1347 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1348 crit_enter_gd(mygd);
cfaeae2a 1349 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23 1350 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
df910c23 1351 lwkt_process_ipiq();
52eedfb5 1352 cpu_lfence();
a86ce0cd
MD
1353#ifdef _KERNEL_VIRTUAL
1354 pthread_yield();
3a06728e
MD
1355#endif
1356#ifdef LOOPMASK
1357 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
1358 kprintf("lwkt_acquire: stuck td %p td->td_flags %08x\n",
1359 td, td->td_flags);
1360 tsc_base = rdtsc();
1361 }
a86ce0cd 1362#endif
df910c23 1363 }
cfaeae2a 1364 DEBUG_POP_INFO();
562273ea 1365 cpu_mfence();
37af14fe 1366 td->td_gd = mygd;
52eedfb5
MD
1367 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1368 td->td_flags &= ~TDF_MIGRATING;
1369 crit_exit_gd(mygd);
1370 } else {
1371 crit_enter_gd(mygd);
1372 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1373 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1374 crit_exit_gd(mygd);
a2a5ad0d
MD
1375 }
1376}
1377
f1d1c3fa
MD
1378/*
1379 * Generic deschedule. Descheduling threads other then your own should be
1380 * done only in carefully controlled circumstances. Descheduling is
1381 * asynchronous.
1382 *
1383 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1384 */
1385void
1386lwkt_deschedule(thread_t td)
1387{
f1d1c3fa
MD
1388 crit_enter();
1389 if (td == curthread) {
1390 _lwkt_dequeue(td);
1391 } else {
a72187e9 1392 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1393 _lwkt_dequeue(td);
1394 } else {
b8a98473 1395 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1396 }
1397 }
1398 crit_exit();
1399}
1400
4b5f931b
MD
1401/*
1402 * Set the target thread's priority. This routine does not automatically
1403 * switch to a higher priority thread, LWKT threads are not designed for
1404 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1405 */
1406void
1407lwkt_setpri(thread_t td, int pri)
1408{
f9235b6d
MD
1409 if (td->td_pri != pri) {
1410 KKASSERT(pri >= 0);
1411 crit_enter();
1412 if (td->td_flags & TDF_RUNQ) {
d2d8515b 1413 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1414 _lwkt_dequeue(td);
1415 td->td_pri = pri;
1416 _lwkt_enqueue(td);
1417 } else {
1418 td->td_pri = pri;
1419 }
1420 crit_exit();
26a0694b 1421 }
26a0694b
MD
1422}
1423
03bd0a5e
MD
1424/*
1425 * Set the initial priority for a thread prior to it being scheduled for
1426 * the first time. The thread MUST NOT be scheduled before or during
1427 * this call. The thread may be assigned to a cpu other then the current
1428 * cpu.
1429 *
1430 * Typically used after a thread has been created with TDF_STOPPREQ,
1431 * and before the thread is initially scheduled.
1432 */
1433void
1434lwkt_setpri_initial(thread_t td, int pri)
1435{
1436 KKASSERT(pri >= 0);
1437 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1438 td->td_pri = pri;
03bd0a5e
MD
1439}
1440
26a0694b
MD
1441void
1442lwkt_setpri_self(int pri)
1443{
1444 thread_t td = curthread;
1445
4b5f931b
MD
1446 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1447 crit_enter();
1448 if (td->td_flags & TDF_RUNQ) {
1449 _lwkt_dequeue(td);
f9235b6d 1450 td->td_pri = pri;
4b5f931b
MD
1451 _lwkt_enqueue(td);
1452 } else {
f9235b6d 1453 td->td_pri = pri;
4b5f931b
MD
1454 }
1455 crit_exit();
1456}
1457
f9235b6d 1458/*
85946b6c 1459 * hz tick scheduler clock for LWKT threads
f9235b6d
MD
1460 */
1461void
85946b6c 1462lwkt_schedulerclock(thread_t td)
f9235b6d 1463{
85946b6c
MD
1464 globaldata_t gd = td->td_gd;
1465 thread_t xtd;
2a418930 1466
c6a766f4
MD
1467 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1468 if (xtd == td) {
85946b6c
MD
1469 /*
1470 * If the current thread is at the head of the runq shift it to the
1471 * end of any equal-priority threads and request a LWKT reschedule
1472 * if it moved.
d992c377
MD
1473 *
1474 * Ignore upri in this situation. There will only be one user thread
1475 * in user mode, all others will be user threads running in kernel
1476 * mode and we have to make sure they get some cpu.
85946b6c
MD
1477 */
1478 xtd = TAILQ_NEXT(td, td_threadq);
1479 if (xtd && xtd->td_pri == td->td_pri) {
1480 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1481 while (xtd && xtd->td_pri == td->td_pri)
1482 xtd = TAILQ_NEXT(xtd, td_threadq);
1483 if (xtd)
1484 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1485 else
1486 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1487 need_lwkt_resched();
f9235b6d 1488 }
c6a766f4 1489 } else if (xtd) {
85946b6c
MD
1490 /*
1491 * If we scheduled a thread other than the one at the head of the
1492 * queue always request a reschedule every tick.
1493 */
1494 need_lwkt_resched();
f9235b6d 1495 }
c6a766f4 1496 /* else curthread probably the idle thread, no need to reschedule */
f9235b6d
MD
1497}
1498
5d21b981 1499/*
52eedfb5
MD
1500 * Migrate the current thread to the specified cpu.
1501 *
cc9b6223
MD
1502 * This is accomplished by descheduling ourselves from the current cpu
1503 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1504 * 'old' thread wants to migrate after it has been completely switched out
1505 * and will complete the migration.
1506 *
1507 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1508 *
1509 * We must be sure to release our current process designation (if a user
1510 * process) before clearing out any tsleepq we are on because the release
1511 * code may re-add us.
ae8e83e6
MD
1512 *
1513 * We must be sure to remove ourselves from the current cpu's tsleepq
1514 * before potentially moving to another queue. The thread can be on
1515 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1516 */
5d21b981
MD
1517
1518void
1519lwkt_setcpu_self(globaldata_t rgd)
1520{
5d21b981
MD
1521 thread_t td = curthread;
1522
1523 if (td->td_gd != rgd) {
1524 crit_enter_quick(td);
cc9b6223 1525
95858b91
MD
1526 if (td->td_release)
1527 td->td_release(td);
ae8e83e6 1528 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1529 tsleep_remove(td);
cc9b6223
MD
1530
1531 /*
1532 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1533 * trying to deschedule ourselves and switch away, then deschedule
1534 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1535 * call lwkt_switch() to complete the operation.
1536 */
5d21b981
MD
1537 td->td_flags |= TDF_MIGRATING;
1538 lwkt_deschedule_self(td);
52eedfb5 1539 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1540 td->td_migrate_gd = rgd;
5d21b981 1541 lwkt_switch();
cc9b6223
MD
1542
1543 /*
1544 * We are now on the target cpu
1545 */
1546 KKASSERT(rgd == mycpu);
52eedfb5 1547 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1548 crit_exit_quick(td);
1549 }
5d21b981
MD
1550}
1551
ecdefdda
MD
1552void
1553lwkt_migratecpu(int cpuid)
1554{
ecdefdda
MD
1555 globaldata_t rgd;
1556
1557 rgd = globaldata_find(cpuid);
1558 lwkt_setcpu_self(rgd);
ecdefdda
MD
1559}
1560
5d21b981
MD
1561/*
1562 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
1563 * do not have to enter another one).
1564 *
1565 * The thread (td) has already been completely descheduled from the
1566 * originating cpu and we can simply assert the case. The thread is
1567 * assigned to the new cpu and enqueued.
5d21b981 1568 *
cc9b6223 1569 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1570 */
1571static void
1572lwkt_setcpu_remote(void *arg)
1573{
1574 thread_t td = arg;
1575 globaldata_t gd = mycpu;
1576
cc9b6223 1577 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1578 td->td_gd = gd;
562273ea 1579 cpu_mfence();
5d21b981 1580 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1581 KKASSERT(td->td_migrate_gd == NULL);
4643740a
MD
1582 KKASSERT(td->td_lwp == NULL ||
1583 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
5d21b981
MD
1584 _lwkt_enqueue(td);
1585}
1586
553ea3c8 1587struct lwp *
4b5f931b
MD
1588lwkt_preempted_proc(void)
1589{
73e4f7b9 1590 thread_t td = curthread;
4b5f931b
MD
1591 while (td->td_preempted)
1592 td = td->td_preempted;
553ea3c8 1593 return(td->td_lwp);
4b5f931b
MD
1594}
1595
99df837e
MD
1596/*
1597 * Create a kernel process/thread/whatever. It shares it's address space
1598 * with proc0 - ie: kernel only.
1599 *
d2d8515b
MD
1600 * If the cpu is not specified one will be selected. In the future
1601 * specifying a cpu of -1 will enable kernel thread migration between
1602 * cpus.
99df837e
MD
1603 */
1604int
c9e9fb21
MD
1605lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1606 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1607{
73e4f7b9 1608 thread_t td;
e2565a42 1609 __va_list ap;
99df837e 1610
d3d32139 1611 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1612 tdflags);
a2a5ad0d
MD
1613 if (tdp)
1614 *tdp = td;
709799ea 1615 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1616
1617 /*
1618 * Set up arg0 for 'ps' etc
1619 */
e2565a42 1620 __va_start(ap, fmt);
379210cb 1621 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1622 __va_end(ap);
99df837e
MD
1623
1624 /*
1625 * Schedule the thread to run
1626 */
4643740a
MD
1627 if (td->td_flags & TDF_NOSTART)
1628 td->td_flags &= ~TDF_NOSTART;
ef0fdad1 1629 else
4643740a 1630 lwkt_schedule(td);
99df837e
MD
1631 return 0;
1632}
1633
1634/*
1635 * Destroy an LWKT thread. Warning! This function is not called when
1636 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1637 * uses a different reaping mechanism.
1638 */
1639void
1640lwkt_exit(void)
1641{
1642 thread_t td = curthread;
c070746a 1643 thread_t std;
8826f33a 1644 globaldata_t gd;
99df837e 1645
2883d2d8
MD
1646 /*
1647 * Do any cleanup that might block here
1648 */
99df837e 1649 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1650 kprintf("kthread %p %s has exited\n", td, td->td_comm);
2883d2d8
MD
1651 biosched_done(td);
1652 dsched_exit_thread(td);
c070746a
MD
1653
1654 /*
1655 * Get us into a critical section to interlock gd_freetd and loop
1656 * until we can get it freed.
1657 *
1658 * We have to cache the current td in gd_freetd because objcache_put()ing
1659 * it would rip it out from under us while our thread is still active.
2af9d75d
MD
1660 *
1661 * We are the current thread so of course our own TDF_RUNNING bit will
1662 * be set, so unlike the lwp reap code we don't wait for it to clear.
c070746a
MD
1663 */
1664 gd = mycpu;
37af14fe 1665 crit_enter_quick(td);
2af9d75d
MD
1666 for (;;) {
1667 if (td->td_refs) {
1668 tsleep(td, 0, "tdreap", 1);
1669 continue;
1670 }
1671 if ((std = gd->gd_freetd) != NULL) {
1672 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1673 gd->gd_freetd = NULL;
1674 objcache_put(thread_cache, std);
1675 continue;
1676 }
1677 break;
c070746a 1678 }
3b4192fb
MD
1679
1680 /*
1681 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1682 * the last time. We cannot block after this point or we may end
1683 * up with a stale td on the tsleepq.
eb2adbf5
MD
1684 *
1685 * None of this may block, the critical section is the only thing
1686 * protecting tdallq and the only thing preventing new lwkt_hold()
1687 * thread refs now.
3b4192fb
MD
1688 */
1689 if (td->td_flags & TDF_TSLEEPQ)
1690 tsleep_remove(td);
37af14fe 1691 lwkt_deschedule_self(td);
e56e4dea 1692 lwkt_remove_tdallq(td);
74c9628e 1693 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1694
1695 /*
1696 * Final cleanup
1697 */
1698 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1699 if (td->td_flags & TDF_ALLOCATED_THREAD)
1700 gd->gd_freetd = td;
99df837e
MD
1701 cpu_thread_exit();
1702}
1703
e56e4dea
MD
1704void
1705lwkt_remove_tdallq(thread_t td)
1706{
1707 KKASSERT(td->td_gd == mycpu);
1708 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1709}
1710
9cf43f91
MD
1711/*
1712 * Code reduction and branch prediction improvements. Call/return
1713 * overhead on modern cpus often degenerates into 0 cycles due to
1714 * the cpu's branch prediction hardware and return pc cache. We
1715 * can take advantage of this by not inlining medium-complexity
1716 * functions and we can also reduce the branch prediction impact
1717 * by collapsing perfectly predictable branches into a single
1718 * procedure instead of duplicating it.
1719 *
1720 * Is any of this noticeable? Probably not, so I'll take the
1721 * smaller code size.
1722 */
1723void
b6468f56 1724crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1725{
b6468f56 1726 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1727}
1728
2d93b37a
MD
1729void
1730crit_panic(void)
1731{
1732 thread_t td = curthread;
850634cc 1733 int lcrit = td->td_critcount;
2d93b37a 1734
850634cc 1735 td->td_critcount = 0;
a4d95680 1736 cpu_ccfence();
850634cc 1737 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1738 /* NOT REACHED */
2d93b37a
MD
1739}
1740
bd8015ca
MD
1741/*
1742 * Called from debugger/panic on cpus which have been stopped. We must still
b19f40a4 1743 * process the IPIQ while stopped.
bd8015ca
MD
1744 *
1745 * If we are dumping also try to process any pending interrupts. This may
1746 * or may not work depending on the state of the cpu at the point it was
1747 * stopped.
1748 */
1749void
1750lwkt_smp_stopped(void)
1751{
1752 globaldata_t gd = mycpu;
1753
bd8015ca
MD
1754 if (dumping) {
1755 lwkt_process_ipiq();
b19f40a4 1756 --gd->gd_intr_nesting_level;
bd8015ca 1757 splz();
b19f40a4 1758 ++gd->gd_intr_nesting_level;
bd8015ca
MD
1759 } else {
1760 lwkt_process_ipiq();
1761 }
63cff036 1762 cpu_smp_stopped();
bd8015ca 1763}