kernel - Quick pass add __read_frequently
[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);
e8e93b5f 344 lockinit(&gd->gd_sysctllock, "sysctl", 0, LK_CANRECURSE);
8ad65e08
MD
345}
346
7d0bac62
MD
347/*
348 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
349 * or LWKT start address before it can be scheduled. If the target cpu is
350 * -1 the thread will be created on the current cpu.
0cfcada1
MD
351 *
352 * If you intend to create a thread without a process context this function
353 * does everything except load the startup and switcher function.
7d0bac62
MD
354 */
355thread_t
d3d32139 356lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 357{
d2d8515b 358 static int cpu_rotator;
c070746a 359 globaldata_t gd = mycpu;
99df837e 360 void *stack;
7d0bac62 361
c070746a
MD
362 /*
363 * If static thread storage is not supplied allocate a thread. Reuse
364 * a cached free thread if possible. gd_freetd is used to keep an exiting
365 * thread intact through the exit.
366 */
ef0fdad1 367 if (td == NULL) {
cf709dd2
MD
368 crit_enter_gd(gd);
369 if ((td = gd->gd_freetd) != NULL) {
370 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
371 TDF_RUNQ)) == 0);
c070746a 372 gd->gd_freetd = NULL;
cf709dd2 373 } else {
c070746a 374 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
375 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
376 TDF_RUNQ)) == 0);
377 }
378 crit_exit_gd(gd);
40aaf5fc 379 KASSERT((td->td_flags &
2af9d75d
MD
380 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
381 TDF_ALLOCATED_THREAD,
40aaf5fc
NT
382 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
383 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 384 }
c070746a
MD
385
386 /*
387 * Try to reuse cached stack.
388 */
f470d0c8
MD
389 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
390 if (flags & TDF_ALLOCATED_STACK) {
e4846942 391 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
392 stack = NULL;
393 }
394 }
395 if (stack == NULL) {
070a58b3
MD
396 if (cpu < 0)
397 stack = (void *)kmem_alloc_stack(&kernel_map, stksize, 0);
398 else
399 stack = (void *)kmem_alloc_stack(&kernel_map, stksize,
400 KM_CPU(cpu));
ef0fdad1 401 flags |= TDF_ALLOCATED_STACK;
99df837e 402 }
d2d8515b
MD
403 if (cpu < 0) {
404 cpu = ++cpu_rotator;
405 cpu_ccfence();
5dd85b08 406 cpu = (uint32_t)cpu % (uint32_t)ncpus;
d2d8515b
MD
407 }
408 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 409 return(td);
7d0bac62
MD
410}
411
412/*
413 * Initialize a preexisting thread structure. This function is used by
414 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
415 *
f8c3996b
MD
416 * All threads start out in a critical section at a priority of
417 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
418 * appropriate. This function may send an IPI message when the
419 * requested cpu is not the current cpu and consequently gd_tdallq may
420 * not be initialized synchronously from the point of view of the originating
421 * cpu.
422 *
423 * NOTE! we have to be careful in regards to creating threads for other cpus
424 * if SMP has not yet been activated.
7d0bac62 425 */
75cdbe6c
MD
426static void
427lwkt_init_thread_remote(void *arg)
428{
429 thread_t td = arg;
430
52eedfb5
MD
431 /*
432 * Protected by critical section held by IPI dispatch
433 */
75cdbe6c
MD
434 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
435}
436
fdce8919
MD
437/*
438 * lwkt core thread structural initialization.
439 *
440 * NOTE: All threads are initialized as mpsafe threads.
441 */
7d0bac62 442void
f470d0c8
MD
443lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
444 struct globaldata *gd)
7d0bac62 445{
37af14fe
MD
446 globaldata_t mygd = mycpu;
447
99df837e
MD
448 bzero(td, sizeof(struct thread));
449 td->td_kstack = stack;
f470d0c8 450 td->td_kstack_size = stksize;
d3d32139 451 td->td_flags = flags;
4643740a 452 td->td_mpflags = 0;
f256b6c0 453 td->td_type = TD_TYPE_GENERIC;
26a0694b 454 td->td_gd = gd;
f9235b6d
MD
455 td->td_pri = TDPRI_KERN_DAEMON;
456 td->td_critcount = 1;
54341a3b 457 td->td_toks_have = NULL;
3b998fa9 458 td->td_toks_stop = &td->td_toks_base;
c068fb59
SZ
459 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) {
460 lwkt_initport_spin(&td->td_msgport, td,
461 (flags & TDF_FIXEDCPU) ? TRUE : FALSE);
462 } else {
fb0f29c4 463 lwkt_initport_thread(&td->td_msgport, td);
c068fb59 464 }
99df837e 465 pmap_init_thread(td);
5d21b981
MD
466 /*
467 * Normally initializing a thread for a remote cpu requires sending an
468 * IPI. However, the idlethread is setup before the other cpus are
469 * activated so we have to treat it as a special case. XXX manipulation
470 * of gd_tdallq requires the BGL.
471 */
472 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 473 crit_enter_gd(mygd);
75cdbe6c 474 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 475 crit_exit_gd(mygd);
75cdbe6c 476 } else {
2db3b277 477 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 478 }
3573cf7b 479 dsched_enter_thread(td);
73e4f7b9
MD
480}
481
482void
483lwkt_set_comm(thread_t td, const char *ctl, ...)
484{
e2565a42 485 __va_list va;
73e4f7b9 486
e2565a42 487 __va_start(va, ctl);
379210cb 488 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 489 __va_end(va);
5bf48697 490 KTR_LOG(ctxsw_newtd, td, td->td_comm);
7d0bac62
MD
491}
492
eb2adbf5
MD
493/*
494 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
495 * this does not prevent the thread from migrating to another cpu so the
496 * gd_tdallq state is not protected by this.
497 */
99df837e 498void
73e4f7b9 499lwkt_hold(thread_t td)
99df837e 500{
74c9628e 501 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
502}
503
504void
505lwkt_rele(thread_t td)
506{
507 KKASSERT(td->td_refs > 0);
74c9628e 508 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
509}
510
73e4f7b9
MD
511void
512lwkt_free_thread(thread_t td)
513{
74c9628e 514 KKASSERT(td->td_refs == 0);
c17a6852
MD
515 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
516 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
40aaf5fc
NT
517 if (td->td_flags & TDF_ALLOCATED_THREAD) {
518 objcache_put(thread_cache, td);
519 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
520 /* client-allocated struct with internally allocated stack */
521 KASSERT(td->td_kstack && td->td_kstack_size > 0,
522 ("lwkt_free_thread: corrupted stack"));
523 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
524 td->td_kstack = NULL;
525 td->td_kstack_size = 0;
99df837e 526 }
a86ce0cd 527
e7c0dbba 528 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
529}
530
531
8ad65e08
MD
532/*
533 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
534 * switch to the idlethread. Switching must occur within a critical
535 * section to avoid races with the scheduling queue.
536 *
537 * We always have full control over our cpu's run queue. Other cpus
538 * that wish to manipulate our queue must use the cpu_*msg() calls to
539 * talk to our cpu, so a critical section is all that is needed and
540 * the result is very, very fast thread switching.
541 *
96728c05
MD
542 * The LWKT scheduler uses a fixed priority model and round-robins at
543 * each priority level. User process scheduling is a totally
544 * different beast and LWKT priorities should not be confused with
545 * user process priorities.
f1d1c3fa 546 *
69d78e99
MD
547 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
548 * is not called by the current thread in the preemption case, only when
549 * the preempting thread blocks (in order to return to the original thread).
cfaeae2a
MD
550 *
551 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
552 * migration and tsleep deschedule the current lwkt thread and call
553 * lwkt_switch(). In particular, the target cpu of the migration fully
554 * expects the thread to become non-runnable and can deadlock against
555 * cpusync operations if we run any IPIs prior to switching the thread out.
556 *
557 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
95858b91 558 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
8ad65e08
MD
559 */
560void
561lwkt_switch(void)
562{
37af14fe
MD
563 globaldata_t gd = mycpu;
564 thread_t td = gd->gd_curthread;
8ad65e08 565 thread_t ntd;
c6a766f4 566 thread_t xtd;
5411d8f1 567 int upri;
3a06728e
MD
568#ifdef LOOPMASK
569 uint64_t tsc_base = rdtsc();
570#endif
8ad65e08 571
da0b0e8b 572 KKASSERT(gd->gd_processing_ipiq == 0);
121f93bc 573 KKASSERT(td->td_flags & TDF_RUNNING);
da0b0e8b 574
46a3f46d 575 /*
27e88a6e
MD
576 * Switching from within a 'fast' (non thread switched) interrupt or IPI
577 * is illegal. However, we may have to do it anyway if we hit a fatal
578 * kernel trap or we have paniced.
579 *
580 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 581 */
27e88a6e
MD
582 if (gd->gd_intr_nesting_level) {
583 int savegdnest;
584 int savegdtrap;
585
5fddbda2 586 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22 587 panic("lwkt_switch: Attempt to switch from a "
5a8df152 588 "fast interrupt, ipi, or hard code section, "
4a28fe22
MD
589 "td %p\n",
590 td);
27e88a6e
MD
591 } else {
592 savegdnest = gd->gd_intr_nesting_level;
593 savegdtrap = gd->gd_trap_nesting_level;
594 gd->gd_intr_nesting_level = 0;
595 gd->gd_trap_nesting_level = 0;
a7422615
MD
596 if ((td->td_flags & TDF_PANICWARN) == 0) {
597 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
598 kprintf("Warning: thread switch from interrupt, IPI, "
599 "or hard code section.\n"
a7422615 600 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 601 print_backtrace(-1);
a7422615 602 }
27e88a6e
MD
603 lwkt_switch();
604 gd->gd_intr_nesting_level = savegdnest;
605 gd->gd_trap_nesting_level = savegdtrap;
606 return;
607 }
96728c05 608 }
ef0fdad1 609
cb973d15 610 /*
85946b6c
MD
611 * Release our current user process designation if we are blocking
612 * or if a user reschedule was requested.
613 *
614 * NOTE: This function is NOT called if we are switching into or
615 * returning from a preemption.
616 *
617 * NOTE: Releasing our current user process designation may cause
618 * it to be assigned to another thread, which in turn will
619 * cause us to block in the usched acquire code when we attempt
620 * to return to userland.
621 *
622 * NOTE: On SMP systems this can be very nasty when heavy token
623 * contention is present so we want to be careful not to
624 * release the designation gratuitously.
cb973d15 625 */
85946b6c
MD
626 if (td->td_release &&
627 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
cb973d15 628 td->td_release(td);
85946b6c 629 }
cb973d15 630
85946b6c 631 /*
5411d8f1
MD
632 * Release all tokens. Once we do this we must remain in the critical
633 * section and cannot run IPIs or other interrupts until we switch away
634 * because they may implode if they try to get a token using our thread
635 * context.
85946b6c 636 */
37af14fe 637 crit_enter_gd(gd);
3b998fa9 638 if (TD_TOKS_HELD(td))
9d265729
MD
639 lwkt_relalltokens(td);
640
641 /*
b02926de
MD
642 * We had better not be holding any spin locks, but don't get into an
643 * endless panic loop.
9d265729 644 */
0846e4ce 645 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL,
d666840a 646 ("lwkt_switch: still holding %d exclusive spinlocks!",
0846e4ce 647 gd->gd_spinlocks));
9d265729 648
0f7a3396
MD
649#ifdef INVARIANTS
650 if (td->td_cscount) {
6ea70f76 651 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
652 td);
653 if (panic_on_cscount)
654 panic("switching while mastering cpusync");
655 }
8a8d5d85 656#endif
f9235b6d
MD
657
658 /*
659 * If we had preempted another thread on this cpu, resume the preempted
660 * thread. This occurs transparently, whether the preempted thread
661 * was scheduled or not (it may have been preempted after descheduling
662 * itself).
663 *
664 * We have to setup the MP lock for the original thread after backing
665 * out the adjustment that was made to curthread when the original
666 * was preempted.
667 */
99df837e 668 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
669 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
670 ntd->td_flags |= TDF_PREEMPT_DONE;
7fb451cb 671 ntd->td_contended = 0; /* reset contended */
8ec60c3f
MD
672
673 /*
b9eb1c19
MD
674 * The interrupt may have woken a thread up, we need to properly
675 * set the reschedule flag if the originally interrupted thread is
676 * at a lower priority.
85946b6c 677 *
c6a766f4
MD
678 * NOTE: The interrupt may not have descheduled ntd.
679 *
680 * NOTE: We do not reschedule if there are no threads on the runq.
681 * (ntd could be the idlethread).
8ec60c3f 682 */
c6a766f4
MD
683 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
684 if (xtd && xtd != ntd)
8ec60c3f 685 need_lwkt_resched();
f9235b6d
MD
686 goto havethread_preempted;
687 }
688
b12defdc 689 /*
5411d8f1
MD
690 * Figure out switch target. If we cannot switch to our desired target
691 * look for a thread that we can switch to.
cfaeae2a 692 *
5411d8f1
MD
693 * NOTE! The limited spin loop and related parameters are extremely
694 * important for system performance, particularly for pipes and
695 * concurrent conflicting VM faults.
f9235b6d 696 */
5411d8f1
MD
697 clear_lwkt_resched();
698 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
699
700 if (ntd) {
701 do {
702 if (TD_TOKS_NOT_HELD(ntd) ||
703 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops)))
704 {
705 goto havethread;
706 }
7fb451cb 707 ++ntd->td_contended; /* overflow ok */
b1793cc6
MD
708 if (gd->gd_indefinite.type == 0)
709 indefinite_init(&gd->gd_indefinite, NULL, 0, 't');
3a06728e
MD
710#ifdef LOOPMASK
711 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
712 kprintf("lwkt_switch: excessive contended %d "
713 "thread %p\n", ntd->td_contended, ntd);
714 tsc_base = rdtsc();
715 }
716#endif
5411d8f1
MD
717 } while (ntd->td_contended < (lwkt_spin_loops >> 1));
718 upri = ntd->td_upri;
f9235b6d 719
f9235b6d 720 /*
5411d8f1
MD
721 * Bleh, the thread we wanted to switch to has a contended token.
722 * See if we can switch to another thread.
2a418930 723 *
5411d8f1 724 * We generally don't want to do this because it represents a
5042f74d
MD
725 * priority inversion, but contending tokens on the same cpu can
726 * cause real problems if we don't now that we have an exclusive
727 * priority mechanism over shared for tokens.
728 *
729 * The solution is to allow threads with pending tokens to compete
730 * for them (a lower priority thread will get less cpu once it
731 * returns from the kernel anyway). If a thread does not have
732 * any contending tokens, we go by td_pri and upri.
f9235b6d 733 */
b12defdc 734 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
5042f74d
MD
735 if (TD_TOKS_NOT_HELD(ntd) &&
736 ntd->td_pri < TDPRI_KERN_LPSCHED && upri > ntd->td_upri) {
737 continue;
738 }
739 if (upri < ntd->td_upri)
740 upri = ntd->td_upri;
b12defdc 741
5411d8f1
MD
742 /*
743 * Try this one.
744 */
745 if (TD_TOKS_NOT_HELD(ntd) ||
746 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) {
747 goto havethread;
748 }
7fb451cb 749 ++ntd->td_contended; /* overflow ok */
b12defdc 750 }
b12defdc
MD
751
752 /*
5411d8f1
MD
753 * Fall through, switch to idle thread to get us out of the current
754 * context. Since we were contended, prevent HLT by flagging a
755 * LWKT reschedule.
b12defdc 756 */
5411d8f1 757 need_lwkt_resched();
f1d1c3fa 758 }
8a8d5d85 759
5411d8f1
MD
760 /*
761 * We either contended on ntd or the runq is empty. We must switch
762 * through the idle thread to get out of the current context.
763 */
764 ntd = &gd->gd_idlethread;
765 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
766 ASSERT_NO_TOKENS_HELD(ntd);
767 cpu_time.cp_msg[0] = 0;
5411d8f1
MD
768 goto haveidle;
769
2a418930 770havethread:
b12defdc 771 /*
be71787b
MD
772 * Clear gd_idle_repeat when doing a normal switch to a non-idle
773 * thread.
f9235b6d 774 */
9ac1ee6e 775 ntd->td_wmesg = NULL;
7fb451cb 776 ntd->td_contended = 0; /* reset once scheduled */
b12defdc 777 ++gd->gd_cnt.v_swtch;
be71787b 778 gd->gd_idle_repeat = 0;
2a418930 779
5b49787b
MD
780 /*
781 * If we were busy waiting record final disposition
782 */
b1793cc6
MD
783 if (gd->gd_indefinite.type)
784 indefinite_done(&gd->gd_indefinite);
5b49787b 785
f9235b6d 786havethread_preempted:
f9235b6d
MD
787 /*
788 * If the new target does not need the MP lock and we are holding it,
789 * release the MP lock. If the new target requires the MP lock we have
790 * already acquired it for the target.
8a8d5d85 791 */
2a418930 792 ;
f9235b6d
MD
793haveidle:
794 KASSERT(ntd->td_critcount,
b5d16701
MD
795 ("priority problem in lwkt_switch %d %d",
796 td->td_critcount, ntd->td_critcount));
797
94f6d86e 798 if (td != ntd) {
cc9b6223
MD
799 /*
800 * Execute the actual thread switch operation. This function
801 * returns to the current thread and returns the previous thread
802 * (which may be different from the thread we switched to).
803 *
804 * We are responsible for marking ntd as TDF_RUNNING.
805 */
121f93bc 806 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
94f6d86e 807 ++switch_count;
a1f0fb66 808 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
809 ntd->td_flags |= TDF_RUNNING;
810 lwkt_switch_return(td->td_switch(ntd));
811 /* ntd invalid, td_switch() can return a different thread_t */
94f6d86e 812 }
b12defdc 813
b12defdc 814 /*
54341a3b 815 * catch-all. XXX is this strictly needed?
b12defdc
MD
816 */
817 splz_check();
54341a3b 818
37af14fe
MD
819 /* NOTE: current cpu may have changed after switch */
820 crit_exit_quick(td);
8ad65e08
MD
821}
822
cc9b6223
MD
823/*
824 * Called by assembly in the td_switch (thread restore path) for thread
825 * bootstrap cases which do not 'return' to lwkt_switch().
826 */
827void
828lwkt_switch_return(thread_t otd)
829{
cc9b6223 830 globaldata_t rgd;
3a06728e
MD
831#ifdef LOOPMASK
832 uint64_t tsc_base = rdtsc();
833#endif
834 int exiting;
835
836 exiting = otd->td_flags & TDF_EXITING;
837 cpu_ccfence();
cc9b6223
MD
838
839 /*
840 * Check if otd was migrating. Now that we are on ntd we can finish
841 * up the migration. This is a bit messy but it is the only place
842 * where td is known to be fully descheduled.
843 *
844 * We can only activate the migration if otd was migrating but not
845 * held on the cpu due to a preemption chain. We still have to
846 * clear TDF_RUNNING on the old thread either way.
847 *
848 * We are responsible for clearing the previously running thread's
849 * TDF_RUNNING.
850 */
851 if ((rgd = otd->td_migrate_gd) != NULL &&
852 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
853 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
854 (TDF_MIGRATING | TDF_RUNNING));
855 otd->td_migrate_gd = NULL;
856 otd->td_flags &= ~TDF_RUNNING;
857 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
858 } else {
859 otd->td_flags &= ~TDF_RUNNING;
860 }
2b07d9aa
MD
861
862 /*
863 * Final exit validations (see lwp_wait()). Note that otd becomes
864 * invalid the *instant* we set TDF_MP_EXITSIG.
3a06728e
MD
865 *
866 * Use the EXITING status loaded from before we clear TDF_RUNNING,
867 * because if it is not set otd becomes invalid the instant we clear
868 * TDF_RUNNING on it (otherwise, if the system is fast enough, we
869 * might 'steal' TDF_EXITING from another switch-return!).
2b07d9aa 870 */
3a06728e 871 while (exiting) {
2b07d9aa
MD
872 u_int mpflags;
873
874 mpflags = otd->td_mpflags;
875 cpu_ccfence();
876
877 if (mpflags & TDF_MP_EXITWAIT) {
878 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
879 mpflags | TDF_MP_EXITSIG)) {
880 wakeup(otd);
881 break;
882 }
883 } else {
884 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
885 mpflags | TDF_MP_EXITSIG)) {
886 wakeup(otd);
887 break;
888 }
889 }
3a06728e
MD
890
891#ifdef LOOPMASK
892 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
893 kprintf("lwkt_switch_return: excessive TDF_EXITING "
894 "thread %p\n", otd);
895 tsc_base = rdtsc();
896 }
897#endif
2b07d9aa 898 }
cc9b6223
MD
899}
900
b68b7282 901/*
96728c05 902 * Request that the target thread preempt the current thread. Preemption
203592a0
MD
903 * can only occur only:
904 *
905 * - If our critical section is the one that we were called with
906 * - The relative priority of the target thread is higher
907 * - The target is not excessively interrupt-nested via td_nest_count
908 * - The target thread holds no tokens.
909 * - The target thread is not already scheduled and belongs to the
910 * current cpu.
911 * - The current thread is not holding any spin-locks.
96728c05
MD
912 *
913 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
914 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 915 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
916 * to determine whether preemption is possible (aka usually just the crit
917 * priority of lwkt_schedule() itself).
b68b7282 918 *
54341a3b
MD
919 * Preemption is typically limited to interrupt threads.
920 *
921 * Operation works in a fairly straight-forward manner. The normal
922 * scheduling code is bypassed and we switch directly to the target
923 * thread. When the target thread attempts to block or switch away
924 * code at the base of lwkt_switch() will switch directly back to our
925 * thread. Our thread is able to retain whatever tokens it holds and
926 * if the target needs one of them the target will switch back to us
927 * and reschedule itself normally.
b68b7282
MD
928 */
929void
f9235b6d 930lwkt_preempt(thread_t ntd, int critcount)
b68b7282 931{
46a3f46d 932 struct globaldata *gd = mycpu;
cc9b6223 933 thread_t xtd;
0a3f9b47 934 thread_t td;
2d910aaf 935 int save_gd_intr_nesting_level;
b68b7282 936
26a0694b 937 /*
96728c05
MD
938 * The caller has put us in a critical section. We can only preempt
939 * if the caller of the caller was not in a critical section (basically
f9235b6d 940 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 941 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 942 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
943 *
944 * YYY The target thread must be in a critical section (else it must
945 * inherit our critical section? I dunno yet).
26a0694b 946 */
f9235b6d 947 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 948
b12defdc 949 td = gd->gd_curthread;
fbc024e4
MD
950 if (preempt_enable == 0) {
951 ++preempt_miss;
952 return;
953 }
f9235b6d 954 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
955 ++preempt_miss;
956 return;
957 }
f9235b6d 958 if (td->td_critcount > critcount) {
96728c05
MD
959 ++preempt_miss;
960 return;
961 }
203592a0
MD
962 if (td->td_nest_count >= 2) {
963 ++preempt_miss;
964 return;
965 }
121f93bc
MD
966 if (td->td_cscount) {
967 ++preempt_miss;
968 return;
969 }
46a3f46d 970 if (ntd->td_gd != gd) {
96728c05
MD
971 ++preempt_miss;
972 return;
973 }
ee89e80b 974
41a01a4d 975 /*
77912481
MD
976 * We don't have to check spinlocks here as they will also bump
977 * td_critcount.
d3d1cbc8
MD
978 *
979 * Do not try to preempt if the target thread is holding any tokens.
980 * We could try to acquire the tokens but this case is so rare there
981 * is no need to support it.
41a01a4d 982 */
0846e4ce 983 KKASSERT(gd->gd_spinlocks == 0);
77912481 984
3b998fa9 985 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8 986 ++preempt_miss;
d3d1cbc8
MD
987 return;
988 }
26a0694b
MD
989 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
990 ++preempt_weird;
991 return;
992 }
993 if (ntd->td_preempted) {
4b5f931b 994 ++preempt_hit;
26a0694b 995 return;
b68b7282 996 }
da0b0e8b 997 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 998
8ec60c3f
MD
999 /*
1000 * Since we are able to preempt the current thread, there is no need to
1001 * call need_lwkt_resched().
2d910aaf
MD
1002 *
1003 * We must temporarily clear gd_intr_nesting_level around the switch
1004 * since switchouts from the target thread are allowed (they will just
1005 * return to our thread), and since the target thread has its own stack.
cc9b6223
MD
1006 *
1007 * A preemption must switch back to the original thread, assert the
1008 * case.
8ec60c3f 1009 */
26a0694b
MD
1010 ++preempt_hit;
1011 ntd->td_preempted = td;
1012 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1013 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1014 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1015 gd->gd_intr_nesting_level = 0;
121f93bc
MD
1016
1017 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
cc9b6223
MD
1018 ntd->td_flags |= TDF_RUNNING;
1019 xtd = td->td_switch(ntd);
1020 KKASSERT(xtd == ntd);
1021 lwkt_switch_return(xtd);
2d910aaf 1022 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1023
26a0694b
MD
1024 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1025 ntd->td_preempted = NULL;
1026 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1027}
1028
f1d1c3fa 1029/*
faaeffac 1030 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1031 * This will work inside a critical section but not inside a hard code
1032 * section.
ef0fdad1 1033 *
f1d1c3fa
MD
1034 * (self contained on a per cpu basis)
1035 */
1036void
faaeffac 1037splz_check(void)
f1d1c3fa 1038{
7966cb69
MD
1039 globaldata_t gd = mycpu;
1040 thread_t td = gd->gd_curthread;
ef0fdad1 1041
4a28fe22
MD
1042 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1043 gd->gd_intr_nesting_level == 0 &&
1044 td->td_nest_count < 2)
1045 {
f1d1c3fa 1046 splz();
4a28fe22
MD
1047 }
1048}
1049
1050/*
1051 * This version is integrated into crit_exit, reqflags has already
1052 * been tested but td_critcount has not.
1053 *
1054 * We only want to execute the splz() on the 1->0 transition of
1055 * critcount and not in a hard code section or if too deeply nested.
925040f2 1056 *
0846e4ce 1057 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
4a28fe22
MD
1058 */
1059void
1060lwkt_maybe_splz(thread_t td)
1061{
1062 globaldata_t gd = td->td_gd;
1063
1064 if (td->td_critcount == 0 &&
1065 gd->gd_intr_nesting_level == 0 &&
1066 td->td_nest_count < 2)
1067 {
1068 splz();
1069 }
f1d1c3fa
MD
1070}
1071
e6546af9
MD
1072/*
1073 * Drivers which set up processing co-threads can call this function to
1074 * run the co-thread at a higher priority and to allow it to preempt
1075 * normal threads.
1076 */
1077void
1078lwkt_set_interrupt_support_thread(void)
1079{
1080 thread_t td = curthread;
1081
1082 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1083 td->td_flags |= TDF_INTTHREAD;
1084 td->td_preemptable = lwkt_preempt;
1085}
1086
1087
8ad65e08 1088/*
f9235b6d
MD
1089 * This function is used to negotiate a passive release of the current
1090 * process/lwp designation with the user scheduler, allowing the user
1091 * scheduler to schedule another user thread. The related kernel thread
1092 * (curthread) continues running in the released state.
8ad65e08
MD
1093 */
1094void
f9235b6d 1095lwkt_passive_release(struct thread *td)
8ad65e08 1096{
f9235b6d
MD
1097 struct lwp *lp = td->td_lwp;
1098
1099 td->td_release = NULL;
1100 lwkt_setpri_self(TDPRI_KERN_USER);
d992c377 1101
f9235b6d 1102 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1103}
1104
f9235b6d 1105
3824f392 1106/*
d2d8515b
MD
1107 * This implements a LWKT yield, allowing a kernel thread to yield to other
1108 * kernel threads at the same or higher priority. This function can be
1109 * called in a tight loop and will typically only yield once per tick.
f9235b6d 1110 *
d2d8515b
MD
1111 * Most kernel threads run at the same priority in order to allow equal
1112 * sharing.
f9235b6d
MD
1113 *
1114 * (self contained on a per cpu basis)
3824f392
MD
1115 */
1116void
f9235b6d 1117lwkt_yield(void)
3824f392 1118{
f9235b6d
MD
1119 globaldata_t gd = mycpu;
1120 thread_t td = gd->gd_curthread;
3824f392 1121
fbe96076
MD
1122 /*
1123 * Should never be called with spinlocks held but there is a path
1124 * via ACPI where it might happen.
1125 */
1126 if (gd->gd_spinlocks)
1127 return;
1128
1129 /*
1130 * Safe to call splz if we are not too-heavily nested.
1131 */
f9235b6d
MD
1132 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1133 splz();
fbe96076
MD
1134
1135 /*
1136 * Caller allows switching
1137 */
85946b6c 1138 if (lwkt_resched_wanted()) {
dd8be70a
MD
1139 atomic_set_int(&td->td_mpflags, TDF_MP_DIDYIELD);
1140 lwkt_schedule_self(td);
f9235b6d 1141 lwkt_switch();
f9235b6d 1142 }
3824f392
MD
1143}
1144
40504122
MD
1145/*
1146 * The quick version processes pending interrupts and higher-priority
1147 * LWKT threads but will not round-robin same-priority LWKT threads.
de4d4cb0
MD
1148 *
1149 * When called while attempting to return to userland the only same-pri
1150 * threads are the ones which have already tried to become the current
1151 * user process.
40504122
MD
1152 */
1153void
1154lwkt_yield_quick(void)
1155{
1156 globaldata_t gd = mycpu;
1157 thread_t td = gd->gd_curthread;
1158
1159 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1160 splz();
1161 if (lwkt_resched_wanted()) {
9c99cb33 1162 crit_enter();
40504122
MD
1163 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1164 clear_lwkt_resched();
1165 } else {
dd8be70a 1166 atomic_set_int(&td->td_mpflags, TDF_MP_DIDYIELD);
40504122
MD
1167 lwkt_schedule_self(curthread);
1168 lwkt_switch();
1169 }
9c99cb33 1170 crit_exit();
40504122
MD
1171 }
1172}
1173
3824f392 1174/*
f9235b6d
MD
1175 * This yield is designed for kernel threads with a user context.
1176 *
1177 * The kernel acting on behalf of the user is potentially cpu-bound,
1178 * this function will efficiently allow other threads to run and also
1179 * switch to other processes by releasing.
3824f392
MD
1180 *
1181 * The lwkt_user_yield() function is designed to have very low overhead
1182 * if no yield is determined to be needed.
1183 */
1184void
1185lwkt_user_yield(void)
1186{
f9235b6d
MD
1187 globaldata_t gd = mycpu;
1188 thread_t td = gd->gd_curthread;
1189
fbe96076
MD
1190 /*
1191 * Should never be called with spinlocks held but there is a path
1192 * via ACPI where it might happen.
1193 */
1194 if (gd->gd_spinlocks)
1195 return;
1196
f9235b6d
MD
1197 /*
1198 * Always run any pending interrupts in case we are in a critical
1199 * section.
1200 */
1201 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1202 splz();
3824f392 1203
3824f392 1204 /*
f9235b6d
MD
1205 * Switch (which forces a release) if another kernel thread needs
1206 * the cpu, if userland wants us to resched, or if our kernel
1207 * quantum has run out.
3824f392 1208 */
f9235b6d 1209 if (lwkt_resched_wanted() ||
85946b6c 1210 user_resched_wanted())
f9235b6d 1211 {
3824f392 1212 lwkt_switch();
3824f392
MD
1213 }
1214
f9235b6d 1215#if 0
3824f392 1216 /*
f9235b6d
MD
1217 * Reacquire the current process if we are released.
1218 *
1219 * XXX not implemented atm. The kernel may be holding locks and such,
1220 * so we want the thread to continue to receive cpu.
3824f392 1221 */
f9235b6d
MD
1222 if (td->td_release == NULL && lp) {
1223 lp->lwp_proc->p_usched->acquire_curproc(lp);
1224 td->td_release = lwkt_passive_release;
1225 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1226 }
f9235b6d 1227#endif
b9eb1c19
MD
1228}
1229
8ad65e08 1230/*
f1d1c3fa
MD
1231 * Generic schedule. Possibly schedule threads belonging to other cpus and
1232 * deal with threads that might be blocked on a wait queue.
1233 *
0a3f9b47
MD
1234 * We have a little helper inline function which does additional work after
1235 * the thread has been enqueued, including dealing with preemption and
1236 * setting need_lwkt_resched() (which prevents the kernel from returning
1237 * to userland until it has processed higher priority threads).
6330a558
MD
1238 *
1239 * It is possible for this routine to be called after a failed _enqueue
1240 * (due to the target thread migrating, sleeping, or otherwise blocked).
1241 * We have to check that the thread is actually on the run queue!
8ad65e08 1242 */
0a3f9b47
MD
1243static __inline
1244void
85946b6c 1245_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
0a3f9b47 1246{
6330a558 1247 if (ntd->td_flags & TDF_RUNQ) {
85946b6c 1248 if (ntd->td_preemptable) {
f9235b6d 1249 ntd->td_preemptable(ntd, ccount); /* YYY +token */
6330a558 1250 }
0a3f9b47
MD
1251 }
1252}
1253
361d01dd 1254static __inline
8ad65e08 1255void
85946b6c 1256_lwkt_schedule(thread_t td)
8ad65e08 1257{
37af14fe
MD
1258 globaldata_t mygd = mycpu;
1259
cf709dd2
MD
1260 KASSERT(td != &td->td_gd->gd_idlethread,
1261 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1262 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1263 crit_enter_gd(mygd);
4643740a
MD
1264 KKASSERT(td->td_lwp == NULL ||
1265 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1266
37af14fe 1267 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1268 _lwkt_enqueue(td);
1269 } else {
f1d1c3fa 1270 /*
7cd8d145
MD
1271 * If we own the thread, there is no race (since we are in a
1272 * critical section). If we do not own the thread there might
1273 * be a race but the target cpu will deal with it.
f1d1c3fa 1274 */
7cd8d145 1275 if (td->td_gd == mygd) {
9d265729 1276 _lwkt_enqueue(td);
85946b6c 1277 _lwkt_schedule_post(mygd, td, 1);
f1d1c3fa 1278 } else {
e381e77c 1279 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1280 }
8ad65e08 1281 }
37af14fe 1282 crit_exit_gd(mygd);
8ad65e08
MD
1283}
1284
361d01dd
MD
1285void
1286lwkt_schedule(thread_t td)
1287{
85946b6c 1288 _lwkt_schedule(td);
361d01dd
MD
1289}
1290
1291void
85946b6c 1292lwkt_schedule_noresched(thread_t td) /* XXX not impl */
361d01dd 1293{
85946b6c 1294 _lwkt_schedule(td);
361d01dd
MD
1295}
1296
e381e77c
MD
1297/*
1298 * When scheduled remotely if frame != NULL the IPIQ is being
1299 * run via doreti or an interrupt then preemption can be allowed.
1300 *
1301 * To allow preemption we have to drop the critical section so only
1302 * one is present in _lwkt_schedule_post.
1303 */
1304static void
1305lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1306{
1307 thread_t td = curthread;
1308 thread_t ntd = arg;
1309
1310 if (frame && ntd->td_preemptable) {
1311 crit_exit_noyield(td);
85946b6c 1312 _lwkt_schedule(ntd);
e381e77c
MD
1313 crit_enter_quick(td);
1314 } else {
85946b6c 1315 _lwkt_schedule(ntd);
e381e77c
MD
1316 }
1317}
1318
d9eea1a5 1319/*
52eedfb5
MD
1320 * Thread migration using a 'Pull' method. The thread may or may not be
1321 * the current thread. It MUST be descheduled and in a stable state.
1322 * lwkt_giveaway() must be called on the cpu owning the thread.
1323 *
1324 * At any point after lwkt_giveaway() is called, the target cpu may
1325 * 'pull' the thread by calling lwkt_acquire().
1326 *
ae8e83e6
MD
1327 * We have to make sure the thread is not sitting on a per-cpu tsleep
1328 * queue or it will blow up when it moves to another cpu.
1329 *
52eedfb5 1330 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1331 */
52eedfb5
MD
1332void
1333lwkt_giveaway(thread_t td)
1334{
3b4192fb 1335 globaldata_t gd = mycpu;
52eedfb5 1336
3b4192fb
MD
1337 crit_enter_gd(gd);
1338 if (td->td_flags & TDF_TSLEEPQ)
1339 tsleep_remove(td);
1340 KKASSERT(td->td_gd == gd);
1341 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1342 td->td_flags |= TDF_MIGRATING;
1343 crit_exit_gd(gd);
52eedfb5
MD
1344}
1345
a2a5ad0d
MD
1346void
1347lwkt_acquire(thread_t td)
1348{
37af14fe
MD
1349 globaldata_t gd;
1350 globaldata_t mygd;
a2a5ad0d 1351
52eedfb5 1352 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1353 gd = td->td_gd;
37af14fe 1354 mygd = mycpu;
52eedfb5 1355 if (gd != mycpu) {
3a06728e
MD
1356#ifdef LOOPMASK
1357 uint64_t tsc_base = rdtsc();
1358#endif
35238fa5 1359 cpu_lfence();
52eedfb5 1360 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1361 crit_enter_gd(mygd);
cfaeae2a 1362 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23 1363 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
df910c23 1364 lwkt_process_ipiq();
52eedfb5 1365 cpu_lfence();
a86ce0cd
MD
1366#ifdef _KERNEL_VIRTUAL
1367 pthread_yield();
3a06728e
MD
1368#endif
1369#ifdef LOOPMASK
1370 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
1371 kprintf("lwkt_acquire: stuck td %p td->td_flags %08x\n",
1372 td, td->td_flags);
1373 tsc_base = rdtsc();
1374 }
a86ce0cd 1375#endif
df910c23 1376 }
cfaeae2a 1377 DEBUG_POP_INFO();
562273ea 1378 cpu_mfence();
37af14fe 1379 td->td_gd = mygd;
52eedfb5
MD
1380 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1381 td->td_flags &= ~TDF_MIGRATING;
1382 crit_exit_gd(mygd);
1383 } else {
1384 crit_enter_gd(mygd);
1385 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1386 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1387 crit_exit_gd(mygd);
a2a5ad0d
MD
1388 }
1389}
1390
f1d1c3fa
MD
1391/*
1392 * Generic deschedule. Descheduling threads other then your own should be
1393 * done only in carefully controlled circumstances. Descheduling is
1394 * asynchronous.
1395 *
1396 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1397 */
1398void
1399lwkt_deschedule(thread_t td)
1400{
f1d1c3fa
MD
1401 crit_enter();
1402 if (td == curthread) {
1403 _lwkt_dequeue(td);
1404 } else {
a72187e9 1405 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1406 _lwkt_dequeue(td);
1407 } else {
b8a98473 1408 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1409 }
1410 }
1411 crit_exit();
1412}
1413
4b5f931b
MD
1414/*
1415 * Set the target thread's priority. This routine does not automatically
1416 * switch to a higher priority thread, LWKT threads are not designed for
1417 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1418 */
1419void
1420lwkt_setpri(thread_t td, int pri)
1421{
f9235b6d
MD
1422 if (td->td_pri != pri) {
1423 KKASSERT(pri >= 0);
1424 crit_enter();
1425 if (td->td_flags & TDF_RUNQ) {
d2d8515b 1426 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1427 _lwkt_dequeue(td);
1428 td->td_pri = pri;
1429 _lwkt_enqueue(td);
1430 } else {
1431 td->td_pri = pri;
1432 }
1433 crit_exit();
26a0694b 1434 }
26a0694b
MD
1435}
1436
03bd0a5e
MD
1437/*
1438 * Set the initial priority for a thread prior to it being scheduled for
1439 * the first time. The thread MUST NOT be scheduled before or during
1440 * this call. The thread may be assigned to a cpu other then the current
1441 * cpu.
1442 *
1443 * Typically used after a thread has been created with TDF_STOPPREQ,
1444 * and before the thread is initially scheduled.
1445 */
1446void
1447lwkt_setpri_initial(thread_t td, int pri)
1448{
1449 KKASSERT(pri >= 0);
1450 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1451 td->td_pri = pri;
03bd0a5e
MD
1452}
1453
26a0694b
MD
1454void
1455lwkt_setpri_self(int pri)
1456{
1457 thread_t td = curthread;
1458
4b5f931b
MD
1459 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1460 crit_enter();
1461 if (td->td_flags & TDF_RUNQ) {
1462 _lwkt_dequeue(td);
f9235b6d 1463 td->td_pri = pri;
4b5f931b
MD
1464 _lwkt_enqueue(td);
1465 } else {
f9235b6d 1466 td->td_pri = pri;
4b5f931b
MD
1467 }
1468 crit_exit();
1469}
1470
f9235b6d 1471/*
85946b6c 1472 * hz tick scheduler clock for LWKT threads
f9235b6d
MD
1473 */
1474void
85946b6c 1475lwkt_schedulerclock(thread_t td)
f9235b6d 1476{
85946b6c
MD
1477 globaldata_t gd = td->td_gd;
1478 thread_t xtd;
2a418930 1479
c6a766f4
MD
1480 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1481 if (xtd == td) {
85946b6c
MD
1482 /*
1483 * If the current thread is at the head of the runq shift it to the
1484 * end of any equal-priority threads and request a LWKT reschedule
1485 * if it moved.
d992c377
MD
1486 *
1487 * Ignore upri in this situation. There will only be one user thread
1488 * in user mode, all others will be user threads running in kernel
1489 * mode and we have to make sure they get some cpu.
85946b6c
MD
1490 */
1491 xtd = TAILQ_NEXT(td, td_threadq);
1492 if (xtd && xtd->td_pri == td->td_pri) {
1493 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1494 while (xtd && xtd->td_pri == td->td_pri)
1495 xtd = TAILQ_NEXT(xtd, td_threadq);
1496 if (xtd)
1497 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1498 else
1499 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1500 need_lwkt_resched();
f9235b6d 1501 }
c6a766f4 1502 } else if (xtd) {
85946b6c
MD
1503 /*
1504 * If we scheduled a thread other than the one at the head of the
1505 * queue always request a reschedule every tick.
1506 */
1507 need_lwkt_resched();
f9235b6d 1508 }
c6a766f4 1509 /* else curthread probably the idle thread, no need to reschedule */
f9235b6d
MD
1510}
1511
5d21b981 1512/*
52eedfb5
MD
1513 * Migrate the current thread to the specified cpu.
1514 *
cc9b6223
MD
1515 * This is accomplished by descheduling ourselves from the current cpu
1516 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1517 * 'old' thread wants to migrate after it has been completely switched out
1518 * and will complete the migration.
1519 *
1520 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1521 *
1522 * We must be sure to release our current process designation (if a user
1523 * process) before clearing out any tsleepq we are on because the release
1524 * code may re-add us.
ae8e83e6
MD
1525 *
1526 * We must be sure to remove ourselves from the current cpu's tsleepq
1527 * before potentially moving to another queue. The thread can be on
1528 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1529 */
5d21b981
MD
1530
1531void
1532lwkt_setcpu_self(globaldata_t rgd)
1533{
5d21b981
MD
1534 thread_t td = curthread;
1535
1536 if (td->td_gd != rgd) {
1537 crit_enter_quick(td);
cc9b6223 1538
95858b91
MD
1539 if (td->td_release)
1540 td->td_release(td);
ae8e83e6 1541 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1542 tsleep_remove(td);
cc9b6223
MD
1543
1544 /*
1545 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1546 * trying to deschedule ourselves and switch away, then deschedule
1547 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1548 * call lwkt_switch() to complete the operation.
1549 */
5d21b981
MD
1550 td->td_flags |= TDF_MIGRATING;
1551 lwkt_deschedule_self(td);
52eedfb5 1552 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1553 td->td_migrate_gd = rgd;
5d21b981 1554 lwkt_switch();
cc9b6223
MD
1555
1556 /*
1557 * We are now on the target cpu
1558 */
1559 KKASSERT(rgd == mycpu);
52eedfb5 1560 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1561 crit_exit_quick(td);
1562 }
5d21b981
MD
1563}
1564
ecdefdda
MD
1565void
1566lwkt_migratecpu(int cpuid)
1567{
ecdefdda
MD
1568 globaldata_t rgd;
1569
1570 rgd = globaldata_find(cpuid);
1571 lwkt_setcpu_self(rgd);
ecdefdda
MD
1572}
1573
5d21b981
MD
1574/*
1575 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
1576 * do not have to enter another one).
1577 *
1578 * The thread (td) has already been completely descheduled from the
1579 * originating cpu and we can simply assert the case. The thread is
1580 * assigned to the new cpu and enqueued.
5d21b981 1581 *
cc9b6223 1582 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1583 */
1584static void
1585lwkt_setcpu_remote(void *arg)
1586{
1587 thread_t td = arg;
1588 globaldata_t gd = mycpu;
1589
cc9b6223 1590 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1591 td->td_gd = gd;
562273ea 1592 cpu_mfence();
5d21b981 1593 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1594 KKASSERT(td->td_migrate_gd == NULL);
4643740a
MD
1595 KKASSERT(td->td_lwp == NULL ||
1596 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
5d21b981
MD
1597 _lwkt_enqueue(td);
1598}
1599
553ea3c8 1600struct lwp *
4b5f931b
MD
1601lwkt_preempted_proc(void)
1602{
73e4f7b9 1603 thread_t td = curthread;
4b5f931b
MD
1604 while (td->td_preempted)
1605 td = td->td_preempted;
553ea3c8 1606 return(td->td_lwp);
4b5f931b
MD
1607}
1608
99df837e
MD
1609/*
1610 * Create a kernel process/thread/whatever. It shares it's address space
1611 * with proc0 - ie: kernel only.
1612 *
d2d8515b
MD
1613 * If the cpu is not specified one will be selected. In the future
1614 * specifying a cpu of -1 will enable kernel thread migration between
1615 * cpus.
99df837e
MD
1616 */
1617int
c9e9fb21
MD
1618lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1619 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1620{
73e4f7b9 1621 thread_t td;
e2565a42 1622 __va_list ap;
99df837e 1623
d3d32139 1624 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1625 tdflags);
a2a5ad0d
MD
1626 if (tdp)
1627 *tdp = td;
709799ea 1628 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1629
1630 /*
1631 * Set up arg0 for 'ps' etc
1632 */
e2565a42 1633 __va_start(ap, fmt);
379210cb 1634 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1635 __va_end(ap);
99df837e
MD
1636
1637 /*
1638 * Schedule the thread to run
1639 */
4643740a
MD
1640 if (td->td_flags & TDF_NOSTART)
1641 td->td_flags &= ~TDF_NOSTART;
ef0fdad1 1642 else
4643740a 1643 lwkt_schedule(td);
99df837e
MD
1644 return 0;
1645}
1646
1647/*
1648 * Destroy an LWKT thread. Warning! This function is not called when
1649 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1650 * uses a different reaping mechanism.
1651 */
1652void
1653lwkt_exit(void)
1654{
1655 thread_t td = curthread;
c070746a 1656 thread_t std;
8826f33a 1657 globaldata_t gd;
99df837e 1658
2883d2d8
MD
1659 /*
1660 * Do any cleanup that might block here
1661 */
2883d2d8
MD
1662 biosched_done(td);
1663 dsched_exit_thread(td);
c070746a
MD
1664
1665 /*
1666 * Get us into a critical section to interlock gd_freetd and loop
1667 * until we can get it freed.
1668 *
1669 * We have to cache the current td in gd_freetd because objcache_put()ing
1670 * it would rip it out from under us while our thread is still active.
2af9d75d
MD
1671 *
1672 * We are the current thread so of course our own TDF_RUNNING bit will
1673 * be set, so unlike the lwp reap code we don't wait for it to clear.
c070746a
MD
1674 */
1675 gd = mycpu;
37af14fe 1676 crit_enter_quick(td);
2af9d75d
MD
1677 for (;;) {
1678 if (td->td_refs) {
1679 tsleep(td, 0, "tdreap", 1);
1680 continue;
1681 }
1682 if ((std = gd->gd_freetd) != NULL) {
1683 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1684 gd->gd_freetd = NULL;
1685 objcache_put(thread_cache, std);
1686 continue;
1687 }
1688 break;
c070746a 1689 }
3b4192fb
MD
1690
1691 /*
1692 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1693 * the last time. We cannot block after this point or we may end
1694 * up with a stale td on the tsleepq.
eb2adbf5
MD
1695 *
1696 * None of this may block, the critical section is the only thing
1697 * protecting tdallq and the only thing preventing new lwkt_hold()
1698 * thread refs now.
3b4192fb
MD
1699 */
1700 if (td->td_flags & TDF_TSLEEPQ)
1701 tsleep_remove(td);
37af14fe 1702 lwkt_deschedule_self(td);
e56e4dea 1703 lwkt_remove_tdallq(td);
74c9628e 1704 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1705
1706 /*
1707 * Final cleanup
1708 */
1709 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1710 if (td->td_flags & TDF_ALLOCATED_THREAD)
1711 gd->gd_freetd = td;
99df837e
MD
1712 cpu_thread_exit();
1713}
1714
e56e4dea
MD
1715void
1716lwkt_remove_tdallq(thread_t td)
1717{
1718 KKASSERT(td->td_gd == mycpu);
1719 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1720}
1721
9cf43f91
MD
1722/*
1723 * Code reduction and branch prediction improvements. Call/return
1724 * overhead on modern cpus often degenerates into 0 cycles due to
1725 * the cpu's branch prediction hardware and return pc cache. We
1726 * can take advantage of this by not inlining medium-complexity
1727 * functions and we can also reduce the branch prediction impact
1728 * by collapsing perfectly predictable branches into a single
1729 * procedure instead of duplicating it.
1730 *
1731 * Is any of this noticeable? Probably not, so I'll take the
1732 * smaller code size.
1733 */
1734void
b6468f56 1735crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1736{
b6468f56 1737 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1738}
1739
2d93b37a
MD
1740void
1741crit_panic(void)
1742{
1743 thread_t td = curthread;
850634cc 1744 int lcrit = td->td_critcount;
2d93b37a 1745
850634cc 1746 td->td_critcount = 0;
a4d95680 1747 cpu_ccfence();
850634cc 1748 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1749 /* NOT REACHED */
2d93b37a
MD
1750}
1751
bd8015ca
MD
1752/*
1753 * Called from debugger/panic on cpus which have been stopped. We must still
b19f40a4 1754 * process the IPIQ while stopped.
bd8015ca
MD
1755 *
1756 * If we are dumping also try to process any pending interrupts. This may
1757 * or may not work depending on the state of the cpu at the point it was
1758 * stopped.
1759 */
1760void
1761lwkt_smp_stopped(void)
1762{
1763 globaldata_t gd = mycpu;
1764
bd8015ca
MD
1765 if (dumping) {
1766 lwkt_process_ipiq();
b19f40a4 1767 --gd->gd_intr_nesting_level;
bd8015ca 1768 splz();
b19f40a4 1769 ++gd->gd_intr_nesting_level;
bd8015ca
MD
1770 } else {
1771 lwkt_process_ipiq();
1772 }
63cff036 1773 cpu_smp_stopped();
bd8015ca 1774}