inet6: Make non-prefix and directly reachable inet6 routes work.
[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
MD
724 * We generally don't want to do this because it represents a
725 * priority inversion. Do not allow the case if the thread
726 * is returning to userland (not a kernel thread) AND the thread
727 * has a lower upri.
f9235b6d 728 */
b12defdc 729 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
5411d8f1
MD
730 if (ntd->td_pri < TDPRI_KERN_LPSCHED && upri > ntd->td_upri)
731 break;
732 upri = ntd->td_upri;
b12defdc 733
5411d8f1
MD
734 /*
735 * Try this one.
736 */
737 if (TD_TOKS_NOT_HELD(ntd) ||
738 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) {
739 goto havethread;
740 }
7fb451cb 741 ++ntd->td_contended; /* overflow ok */
b12defdc 742 }
b12defdc
MD
743
744 /*
5411d8f1
MD
745 * Fall through, switch to idle thread to get us out of the current
746 * context. Since we were contended, prevent HLT by flagging a
747 * LWKT reschedule.
b12defdc 748 */
5411d8f1 749 need_lwkt_resched();
f1d1c3fa 750 }
8a8d5d85 751
5411d8f1
MD
752 /*
753 * We either contended on ntd or the runq is empty. We must switch
754 * through the idle thread to get out of the current context.
755 */
756 ntd = &gd->gd_idlethread;
757 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
758 ASSERT_NO_TOKENS_HELD(ntd);
759 cpu_time.cp_msg[0] = 0;
5411d8f1
MD
760 goto haveidle;
761
2a418930 762havethread:
b12defdc 763 /*
be71787b
MD
764 * Clear gd_idle_repeat when doing a normal switch to a non-idle
765 * thread.
f9235b6d 766 */
9ac1ee6e 767 ntd->td_wmesg = NULL;
7fb451cb 768 ntd->td_contended = 0; /* reset once scheduled */
b12defdc 769 ++gd->gd_cnt.v_swtch;
be71787b 770 gd->gd_idle_repeat = 0;
2a418930 771
5b49787b
MD
772 /*
773 * If we were busy waiting record final disposition
774 */
b1793cc6
MD
775 if (gd->gd_indefinite.type)
776 indefinite_done(&gd->gd_indefinite);
5b49787b 777
f9235b6d 778havethread_preempted:
f9235b6d
MD
779 /*
780 * If the new target does not need the MP lock and we are holding it,
781 * release the MP lock. If the new target requires the MP lock we have
782 * already acquired it for the target.
8a8d5d85 783 */
2a418930 784 ;
f9235b6d
MD
785haveidle:
786 KASSERT(ntd->td_critcount,
b5d16701
MD
787 ("priority problem in lwkt_switch %d %d",
788 td->td_critcount, ntd->td_critcount));
789
94f6d86e 790 if (td != ntd) {
cc9b6223
MD
791 /*
792 * Execute the actual thread switch operation. This function
793 * returns to the current thread and returns the previous thread
794 * (which may be different from the thread we switched to).
795 *
796 * We are responsible for marking ntd as TDF_RUNNING.
797 */
121f93bc 798 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
94f6d86e 799 ++switch_count;
a1f0fb66 800 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
801 ntd->td_flags |= TDF_RUNNING;
802 lwkt_switch_return(td->td_switch(ntd));
803 /* ntd invalid, td_switch() can return a different thread_t */
94f6d86e 804 }
b12defdc 805
b12defdc 806 /*
54341a3b 807 * catch-all. XXX is this strictly needed?
b12defdc
MD
808 */
809 splz_check();
54341a3b 810
37af14fe
MD
811 /* NOTE: current cpu may have changed after switch */
812 crit_exit_quick(td);
8ad65e08
MD
813}
814
cc9b6223
MD
815/*
816 * Called by assembly in the td_switch (thread restore path) for thread
817 * bootstrap cases which do not 'return' to lwkt_switch().
818 */
819void
820lwkt_switch_return(thread_t otd)
821{
cc9b6223 822 globaldata_t rgd;
3a06728e
MD
823#ifdef LOOPMASK
824 uint64_t tsc_base = rdtsc();
825#endif
826 int exiting;
827
828 exiting = otd->td_flags & TDF_EXITING;
829 cpu_ccfence();
cc9b6223
MD
830
831 /*
832 * Check if otd was migrating. Now that we are on ntd we can finish
833 * up the migration. This is a bit messy but it is the only place
834 * where td is known to be fully descheduled.
835 *
836 * We can only activate the migration if otd was migrating but not
837 * held on the cpu due to a preemption chain. We still have to
838 * clear TDF_RUNNING on the old thread either way.
839 *
840 * We are responsible for clearing the previously running thread's
841 * TDF_RUNNING.
842 */
843 if ((rgd = otd->td_migrate_gd) != NULL &&
844 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
845 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
846 (TDF_MIGRATING | TDF_RUNNING));
847 otd->td_migrate_gd = NULL;
848 otd->td_flags &= ~TDF_RUNNING;
849 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
850 } else {
851 otd->td_flags &= ~TDF_RUNNING;
852 }
2b07d9aa
MD
853
854 /*
855 * Final exit validations (see lwp_wait()). Note that otd becomes
856 * invalid the *instant* we set TDF_MP_EXITSIG.
3a06728e
MD
857 *
858 * Use the EXITING status loaded from before we clear TDF_RUNNING,
859 * because if it is not set otd becomes invalid the instant we clear
860 * TDF_RUNNING on it (otherwise, if the system is fast enough, we
861 * might 'steal' TDF_EXITING from another switch-return!).
2b07d9aa 862 */
3a06728e 863 while (exiting) {
2b07d9aa
MD
864 u_int mpflags;
865
866 mpflags = otd->td_mpflags;
867 cpu_ccfence();
868
869 if (mpflags & TDF_MP_EXITWAIT) {
870 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
871 mpflags | TDF_MP_EXITSIG)) {
872 wakeup(otd);
873 break;
874 }
875 } else {
876 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
877 mpflags | TDF_MP_EXITSIG)) {
878 wakeup(otd);
879 break;
880 }
881 }
3a06728e
MD
882
883#ifdef LOOPMASK
884 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
885 kprintf("lwkt_switch_return: excessive TDF_EXITING "
886 "thread %p\n", otd);
887 tsc_base = rdtsc();
888 }
889#endif
2b07d9aa 890 }
cc9b6223
MD
891}
892
b68b7282 893/*
96728c05 894 * Request that the target thread preempt the current thread. Preemption
203592a0
MD
895 * can only occur only:
896 *
897 * - If our critical section is the one that we were called with
898 * - The relative priority of the target thread is higher
899 * - The target is not excessively interrupt-nested via td_nest_count
900 * - The target thread holds no tokens.
901 * - The target thread is not already scheduled and belongs to the
902 * current cpu.
903 * - The current thread is not holding any spin-locks.
96728c05
MD
904 *
905 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
906 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 907 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
908 * to determine whether preemption is possible (aka usually just the crit
909 * priority of lwkt_schedule() itself).
b68b7282 910 *
54341a3b
MD
911 * Preemption is typically limited to interrupt threads.
912 *
913 * Operation works in a fairly straight-forward manner. The normal
914 * scheduling code is bypassed and we switch directly to the target
915 * thread. When the target thread attempts to block or switch away
916 * code at the base of lwkt_switch() will switch directly back to our
917 * thread. Our thread is able to retain whatever tokens it holds and
918 * if the target needs one of them the target will switch back to us
919 * and reschedule itself normally.
b68b7282
MD
920 */
921void
f9235b6d 922lwkt_preempt(thread_t ntd, int critcount)
b68b7282 923{
46a3f46d 924 struct globaldata *gd = mycpu;
cc9b6223 925 thread_t xtd;
0a3f9b47 926 thread_t td;
2d910aaf 927 int save_gd_intr_nesting_level;
b68b7282 928
26a0694b 929 /*
96728c05
MD
930 * The caller has put us in a critical section. We can only preempt
931 * if the caller of the caller was not in a critical section (basically
f9235b6d 932 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 933 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 934 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
935 *
936 * YYY The target thread must be in a critical section (else it must
937 * inherit our critical section? I dunno yet).
26a0694b 938 */
f9235b6d 939 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 940
b12defdc 941 td = gd->gd_curthread;
fbc024e4
MD
942 if (preempt_enable == 0) {
943 ++preempt_miss;
944 return;
945 }
f9235b6d 946 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
947 ++preempt_miss;
948 return;
949 }
f9235b6d 950 if (td->td_critcount > critcount) {
96728c05
MD
951 ++preempt_miss;
952 return;
953 }
203592a0
MD
954 if (td->td_nest_count >= 2) {
955 ++preempt_miss;
956 return;
957 }
121f93bc
MD
958 if (td->td_cscount) {
959 ++preempt_miss;
960 return;
961 }
46a3f46d 962 if (ntd->td_gd != gd) {
96728c05
MD
963 ++preempt_miss;
964 return;
965 }
ee89e80b 966
41a01a4d 967 /*
77912481
MD
968 * We don't have to check spinlocks here as they will also bump
969 * td_critcount.
d3d1cbc8
MD
970 *
971 * Do not try to preempt if the target thread is holding any tokens.
972 * We could try to acquire the tokens but this case is so rare there
973 * is no need to support it.
41a01a4d 974 */
0846e4ce 975 KKASSERT(gd->gd_spinlocks == 0);
77912481 976
3b998fa9 977 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8 978 ++preempt_miss;
d3d1cbc8
MD
979 return;
980 }
26a0694b
MD
981 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
982 ++preempt_weird;
983 return;
984 }
985 if (ntd->td_preempted) {
4b5f931b 986 ++preempt_hit;
26a0694b 987 return;
b68b7282 988 }
da0b0e8b 989 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 990
8ec60c3f
MD
991 /*
992 * Since we are able to preempt the current thread, there is no need to
993 * call need_lwkt_resched().
2d910aaf
MD
994 *
995 * We must temporarily clear gd_intr_nesting_level around the switch
996 * since switchouts from the target thread are allowed (they will just
997 * return to our thread), and since the target thread has its own stack.
cc9b6223
MD
998 *
999 * A preemption must switch back to the original thread, assert the
1000 * case.
8ec60c3f 1001 */
26a0694b
MD
1002 ++preempt_hit;
1003 ntd->td_preempted = td;
1004 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1005 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1006 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1007 gd->gd_intr_nesting_level = 0;
121f93bc
MD
1008
1009 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
cc9b6223
MD
1010 ntd->td_flags |= TDF_RUNNING;
1011 xtd = td->td_switch(ntd);
1012 KKASSERT(xtd == ntd);
1013 lwkt_switch_return(xtd);
2d910aaf 1014 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1015
26a0694b
MD
1016 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1017 ntd->td_preempted = NULL;
1018 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1019}
1020
f1d1c3fa 1021/*
faaeffac 1022 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1023 * This will work inside a critical section but not inside a hard code
1024 * section.
ef0fdad1 1025 *
f1d1c3fa
MD
1026 * (self contained on a per cpu basis)
1027 */
1028void
faaeffac 1029splz_check(void)
f1d1c3fa 1030{
7966cb69
MD
1031 globaldata_t gd = mycpu;
1032 thread_t td = gd->gd_curthread;
ef0fdad1 1033
4a28fe22
MD
1034 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1035 gd->gd_intr_nesting_level == 0 &&
1036 td->td_nest_count < 2)
1037 {
f1d1c3fa 1038 splz();
4a28fe22
MD
1039 }
1040}
1041
1042/*
1043 * This version is integrated into crit_exit, reqflags has already
1044 * been tested but td_critcount has not.
1045 *
1046 * We only want to execute the splz() on the 1->0 transition of
1047 * critcount and not in a hard code section or if too deeply nested.
925040f2 1048 *
0846e4ce 1049 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
4a28fe22
MD
1050 */
1051void
1052lwkt_maybe_splz(thread_t td)
1053{
1054 globaldata_t gd = td->td_gd;
1055
1056 if (td->td_critcount == 0 &&
1057 gd->gd_intr_nesting_level == 0 &&
1058 td->td_nest_count < 2)
1059 {
1060 splz();
1061 }
f1d1c3fa
MD
1062}
1063
e6546af9
MD
1064/*
1065 * Drivers which set up processing co-threads can call this function to
1066 * run the co-thread at a higher priority and to allow it to preempt
1067 * normal threads.
1068 */
1069void
1070lwkt_set_interrupt_support_thread(void)
1071{
1072 thread_t td = curthread;
1073
1074 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1075 td->td_flags |= TDF_INTTHREAD;
1076 td->td_preemptable = lwkt_preempt;
1077}
1078
1079
8ad65e08 1080/*
f9235b6d
MD
1081 * This function is used to negotiate a passive release of the current
1082 * process/lwp designation with the user scheduler, allowing the user
1083 * scheduler to schedule another user thread. The related kernel thread
1084 * (curthread) continues running in the released state.
8ad65e08
MD
1085 */
1086void
f9235b6d 1087lwkt_passive_release(struct thread *td)
8ad65e08 1088{
f9235b6d
MD
1089 struct lwp *lp = td->td_lwp;
1090
1091 td->td_release = NULL;
1092 lwkt_setpri_self(TDPRI_KERN_USER);
d992c377 1093
f9235b6d 1094 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1095}
1096
f9235b6d 1097
3824f392 1098/*
d2d8515b
MD
1099 * This implements a LWKT yield, allowing a kernel thread to yield to other
1100 * kernel threads at the same or higher priority. This function can be
1101 * called in a tight loop and will typically only yield once per tick.
f9235b6d 1102 *
d2d8515b
MD
1103 * Most kernel threads run at the same priority in order to allow equal
1104 * sharing.
f9235b6d
MD
1105 *
1106 * (self contained on a per cpu basis)
3824f392
MD
1107 */
1108void
f9235b6d 1109lwkt_yield(void)
3824f392 1110{
f9235b6d
MD
1111 globaldata_t gd = mycpu;
1112 thread_t td = gd->gd_curthread;
3824f392 1113
fbe96076
MD
1114 /*
1115 * Should never be called with spinlocks held but there is a path
1116 * via ACPI where it might happen.
1117 */
1118 if (gd->gd_spinlocks)
1119 return;
1120
1121 /*
1122 * Safe to call splz if we are not too-heavily nested.
1123 */
f9235b6d
MD
1124 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1125 splz();
fbe96076
MD
1126
1127 /*
1128 * Caller allows switching
1129 */
85946b6c 1130 if (lwkt_resched_wanted()) {
f9235b6d
MD
1131 lwkt_schedule_self(curthread);
1132 lwkt_switch();
f9235b6d 1133 }
3824f392
MD
1134}
1135
40504122
MD
1136/*
1137 * The quick version processes pending interrupts and higher-priority
1138 * LWKT threads but will not round-robin same-priority LWKT threads.
de4d4cb0
MD
1139 *
1140 * When called while attempting to return to userland the only same-pri
1141 * threads are the ones which have already tried to become the current
1142 * user process.
40504122
MD
1143 */
1144void
1145lwkt_yield_quick(void)
1146{
1147 globaldata_t gd = mycpu;
1148 thread_t td = gd->gd_curthread;
1149
1150 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1151 splz();
1152 if (lwkt_resched_wanted()) {
9c99cb33 1153 crit_enter();
40504122
MD
1154 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1155 clear_lwkt_resched();
1156 } else {
1157 lwkt_schedule_self(curthread);
1158 lwkt_switch();
1159 }
9c99cb33 1160 crit_exit();
40504122
MD
1161 }
1162}
1163
3824f392 1164/*
f9235b6d
MD
1165 * This yield is designed for kernel threads with a user context.
1166 *
1167 * The kernel acting on behalf of the user is potentially cpu-bound,
1168 * this function will efficiently allow other threads to run and also
1169 * switch to other processes by releasing.
3824f392
MD
1170 *
1171 * The lwkt_user_yield() function is designed to have very low overhead
1172 * if no yield is determined to be needed.
1173 */
1174void
1175lwkt_user_yield(void)
1176{
f9235b6d
MD
1177 globaldata_t gd = mycpu;
1178 thread_t td = gd->gd_curthread;
1179
fbe96076
MD
1180 /*
1181 * Should never be called with spinlocks held but there is a path
1182 * via ACPI where it might happen.
1183 */
1184 if (gd->gd_spinlocks)
1185 return;
1186
f9235b6d
MD
1187 /*
1188 * Always run any pending interrupts in case we are in a critical
1189 * section.
1190 */
1191 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1192 splz();
3824f392 1193
3824f392 1194 /*
f9235b6d
MD
1195 * Switch (which forces a release) if another kernel thread needs
1196 * the cpu, if userland wants us to resched, or if our kernel
1197 * quantum has run out.
3824f392 1198 */
f9235b6d 1199 if (lwkt_resched_wanted() ||
85946b6c 1200 user_resched_wanted())
f9235b6d 1201 {
3824f392 1202 lwkt_switch();
3824f392
MD
1203 }
1204
f9235b6d 1205#if 0
3824f392 1206 /*
f9235b6d
MD
1207 * Reacquire the current process if we are released.
1208 *
1209 * XXX not implemented atm. The kernel may be holding locks and such,
1210 * so we want the thread to continue to receive cpu.
3824f392 1211 */
f9235b6d
MD
1212 if (td->td_release == NULL && lp) {
1213 lp->lwp_proc->p_usched->acquire_curproc(lp);
1214 td->td_release = lwkt_passive_release;
1215 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1216 }
f9235b6d 1217#endif
b9eb1c19
MD
1218}
1219
8ad65e08 1220/*
f1d1c3fa
MD
1221 * Generic schedule. Possibly schedule threads belonging to other cpus and
1222 * deal with threads that might be blocked on a wait queue.
1223 *
0a3f9b47
MD
1224 * We have a little helper inline function which does additional work after
1225 * the thread has been enqueued, including dealing with preemption and
1226 * setting need_lwkt_resched() (which prevents the kernel from returning
1227 * to userland until it has processed higher priority threads).
6330a558
MD
1228 *
1229 * It is possible for this routine to be called after a failed _enqueue
1230 * (due to the target thread migrating, sleeping, or otherwise blocked).
1231 * We have to check that the thread is actually on the run queue!
8ad65e08 1232 */
0a3f9b47
MD
1233static __inline
1234void
85946b6c 1235_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
0a3f9b47 1236{
6330a558 1237 if (ntd->td_flags & TDF_RUNQ) {
85946b6c 1238 if (ntd->td_preemptable) {
f9235b6d 1239 ntd->td_preemptable(ntd, ccount); /* YYY +token */
6330a558 1240 }
0a3f9b47
MD
1241 }
1242}
1243
361d01dd 1244static __inline
8ad65e08 1245void
85946b6c 1246_lwkt_schedule(thread_t td)
8ad65e08 1247{
37af14fe
MD
1248 globaldata_t mygd = mycpu;
1249
cf709dd2
MD
1250 KASSERT(td != &td->td_gd->gd_idlethread,
1251 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1252 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1253 crit_enter_gd(mygd);
4643740a
MD
1254 KKASSERT(td->td_lwp == NULL ||
1255 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1256
37af14fe 1257 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1258 _lwkt_enqueue(td);
1259 } else {
f1d1c3fa 1260 /*
7cd8d145
MD
1261 * If we own the thread, there is no race (since we are in a
1262 * critical section). If we do not own the thread there might
1263 * be a race but the target cpu will deal with it.
f1d1c3fa 1264 */
7cd8d145 1265 if (td->td_gd == mygd) {
9d265729 1266 _lwkt_enqueue(td);
85946b6c 1267 _lwkt_schedule_post(mygd, td, 1);
f1d1c3fa 1268 } else {
e381e77c 1269 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1270 }
8ad65e08 1271 }
37af14fe 1272 crit_exit_gd(mygd);
8ad65e08
MD
1273}
1274
361d01dd
MD
1275void
1276lwkt_schedule(thread_t td)
1277{
85946b6c 1278 _lwkt_schedule(td);
361d01dd
MD
1279}
1280
1281void
85946b6c 1282lwkt_schedule_noresched(thread_t td) /* XXX not impl */
361d01dd 1283{
85946b6c 1284 _lwkt_schedule(td);
361d01dd
MD
1285}
1286
e381e77c
MD
1287/*
1288 * When scheduled remotely if frame != NULL the IPIQ is being
1289 * run via doreti or an interrupt then preemption can be allowed.
1290 *
1291 * To allow preemption we have to drop the critical section so only
1292 * one is present in _lwkt_schedule_post.
1293 */
1294static void
1295lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1296{
1297 thread_t td = curthread;
1298 thread_t ntd = arg;
1299
1300 if (frame && ntd->td_preemptable) {
1301 crit_exit_noyield(td);
85946b6c 1302 _lwkt_schedule(ntd);
e381e77c
MD
1303 crit_enter_quick(td);
1304 } else {
85946b6c 1305 _lwkt_schedule(ntd);
e381e77c
MD
1306 }
1307}
1308
d9eea1a5 1309/*
52eedfb5
MD
1310 * Thread migration using a 'Pull' method. The thread may or may not be
1311 * the current thread. It MUST be descheduled and in a stable state.
1312 * lwkt_giveaway() must be called on the cpu owning the thread.
1313 *
1314 * At any point after lwkt_giveaway() is called, the target cpu may
1315 * 'pull' the thread by calling lwkt_acquire().
1316 *
ae8e83e6
MD
1317 * We have to make sure the thread is not sitting on a per-cpu tsleep
1318 * queue or it will blow up when it moves to another cpu.
1319 *
52eedfb5 1320 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1321 */
52eedfb5
MD
1322void
1323lwkt_giveaway(thread_t td)
1324{
3b4192fb 1325 globaldata_t gd = mycpu;
52eedfb5 1326
3b4192fb
MD
1327 crit_enter_gd(gd);
1328 if (td->td_flags & TDF_TSLEEPQ)
1329 tsleep_remove(td);
1330 KKASSERT(td->td_gd == gd);
1331 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1332 td->td_flags |= TDF_MIGRATING;
1333 crit_exit_gd(gd);
52eedfb5
MD
1334}
1335
a2a5ad0d
MD
1336void
1337lwkt_acquire(thread_t td)
1338{
37af14fe
MD
1339 globaldata_t gd;
1340 globaldata_t mygd;
a2a5ad0d 1341
52eedfb5 1342 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1343 gd = td->td_gd;
37af14fe 1344 mygd = mycpu;
52eedfb5 1345 if (gd != mycpu) {
3a06728e
MD
1346#ifdef LOOPMASK
1347 uint64_t tsc_base = rdtsc();
1348#endif
35238fa5 1349 cpu_lfence();
52eedfb5 1350 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1351 crit_enter_gd(mygd);
cfaeae2a 1352 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23 1353 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
df910c23 1354 lwkt_process_ipiq();
52eedfb5 1355 cpu_lfence();
a86ce0cd
MD
1356#ifdef _KERNEL_VIRTUAL
1357 pthread_yield();
3a06728e
MD
1358#endif
1359#ifdef LOOPMASK
1360 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
1361 kprintf("lwkt_acquire: stuck td %p td->td_flags %08x\n",
1362 td, td->td_flags);
1363 tsc_base = rdtsc();
1364 }
a86ce0cd 1365#endif
df910c23 1366 }
cfaeae2a 1367 DEBUG_POP_INFO();
562273ea 1368 cpu_mfence();
37af14fe 1369 td->td_gd = mygd;
52eedfb5
MD
1370 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1371 td->td_flags &= ~TDF_MIGRATING;
1372 crit_exit_gd(mygd);
1373 } else {
1374 crit_enter_gd(mygd);
1375 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1376 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1377 crit_exit_gd(mygd);
a2a5ad0d
MD
1378 }
1379}
1380
f1d1c3fa
MD
1381/*
1382 * Generic deschedule. Descheduling threads other then your own should be
1383 * done only in carefully controlled circumstances. Descheduling is
1384 * asynchronous.
1385 *
1386 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1387 */
1388void
1389lwkt_deschedule(thread_t td)
1390{
f1d1c3fa
MD
1391 crit_enter();
1392 if (td == curthread) {
1393 _lwkt_dequeue(td);
1394 } else {
a72187e9 1395 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1396 _lwkt_dequeue(td);
1397 } else {
b8a98473 1398 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1399 }
1400 }
1401 crit_exit();
1402}
1403
4b5f931b
MD
1404/*
1405 * Set the target thread's priority. This routine does not automatically
1406 * switch to a higher priority thread, LWKT threads are not designed for
1407 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1408 */
1409void
1410lwkt_setpri(thread_t td, int pri)
1411{
f9235b6d
MD
1412 if (td->td_pri != pri) {
1413 KKASSERT(pri >= 0);
1414 crit_enter();
1415 if (td->td_flags & TDF_RUNQ) {
d2d8515b 1416 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1417 _lwkt_dequeue(td);
1418 td->td_pri = pri;
1419 _lwkt_enqueue(td);
1420 } else {
1421 td->td_pri = pri;
1422 }
1423 crit_exit();
26a0694b 1424 }
26a0694b
MD
1425}
1426
03bd0a5e
MD
1427/*
1428 * Set the initial priority for a thread prior to it being scheduled for
1429 * the first time. The thread MUST NOT be scheduled before or during
1430 * this call. The thread may be assigned to a cpu other then the current
1431 * cpu.
1432 *
1433 * Typically used after a thread has been created with TDF_STOPPREQ,
1434 * and before the thread is initially scheduled.
1435 */
1436void
1437lwkt_setpri_initial(thread_t td, int pri)
1438{
1439 KKASSERT(pri >= 0);
1440 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1441 td->td_pri = pri;
03bd0a5e
MD
1442}
1443
26a0694b
MD
1444void
1445lwkt_setpri_self(int pri)
1446{
1447 thread_t td = curthread;
1448
4b5f931b
MD
1449 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1450 crit_enter();
1451 if (td->td_flags & TDF_RUNQ) {
1452 _lwkt_dequeue(td);
f9235b6d 1453 td->td_pri = pri;
4b5f931b
MD
1454 _lwkt_enqueue(td);
1455 } else {
f9235b6d 1456 td->td_pri = pri;
4b5f931b
MD
1457 }
1458 crit_exit();
1459}
1460
f9235b6d 1461/*
85946b6c 1462 * hz tick scheduler clock for LWKT threads
f9235b6d
MD
1463 */
1464void
85946b6c 1465lwkt_schedulerclock(thread_t td)
f9235b6d 1466{
85946b6c
MD
1467 globaldata_t gd = td->td_gd;
1468 thread_t xtd;
2a418930 1469
c6a766f4
MD
1470 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1471 if (xtd == td) {
85946b6c
MD
1472 /*
1473 * If the current thread is at the head of the runq shift it to the
1474 * end of any equal-priority threads and request a LWKT reschedule
1475 * if it moved.
d992c377
MD
1476 *
1477 * Ignore upri in this situation. There will only be one user thread
1478 * in user mode, all others will be user threads running in kernel
1479 * mode and we have to make sure they get some cpu.
85946b6c
MD
1480 */
1481 xtd = TAILQ_NEXT(td, td_threadq);
1482 if (xtd && xtd->td_pri == td->td_pri) {
1483 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1484 while (xtd && xtd->td_pri == td->td_pri)
1485 xtd = TAILQ_NEXT(xtd, td_threadq);
1486 if (xtd)
1487 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1488 else
1489 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1490 need_lwkt_resched();
f9235b6d 1491 }
c6a766f4 1492 } else if (xtd) {
85946b6c
MD
1493 /*
1494 * If we scheduled a thread other than the one at the head of the
1495 * queue always request a reschedule every tick.
1496 */
1497 need_lwkt_resched();
f9235b6d 1498 }
c6a766f4 1499 /* else curthread probably the idle thread, no need to reschedule */
f9235b6d
MD
1500}
1501
5d21b981 1502/*
52eedfb5
MD
1503 * Migrate the current thread to the specified cpu.
1504 *
cc9b6223
MD
1505 * This is accomplished by descheduling ourselves from the current cpu
1506 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1507 * 'old' thread wants to migrate after it has been completely switched out
1508 * and will complete the migration.
1509 *
1510 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1511 *
1512 * We must be sure to release our current process designation (if a user
1513 * process) before clearing out any tsleepq we are on because the release
1514 * code may re-add us.
ae8e83e6
MD
1515 *
1516 * We must be sure to remove ourselves from the current cpu's tsleepq
1517 * before potentially moving to another queue. The thread can be on
1518 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1519 */
5d21b981
MD
1520
1521void
1522lwkt_setcpu_self(globaldata_t rgd)
1523{
5d21b981
MD
1524 thread_t td = curthread;
1525
1526 if (td->td_gd != rgd) {
1527 crit_enter_quick(td);
cc9b6223 1528
95858b91
MD
1529 if (td->td_release)
1530 td->td_release(td);
ae8e83e6 1531 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1532 tsleep_remove(td);
cc9b6223
MD
1533
1534 /*
1535 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1536 * trying to deschedule ourselves and switch away, then deschedule
1537 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1538 * call lwkt_switch() to complete the operation.
1539 */
5d21b981
MD
1540 td->td_flags |= TDF_MIGRATING;
1541 lwkt_deschedule_self(td);
52eedfb5 1542 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1543 td->td_migrate_gd = rgd;
5d21b981 1544 lwkt_switch();
cc9b6223
MD
1545
1546 /*
1547 * We are now on the target cpu
1548 */
1549 KKASSERT(rgd == mycpu);
52eedfb5 1550 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1551 crit_exit_quick(td);
1552 }
5d21b981
MD
1553}
1554
ecdefdda
MD
1555void
1556lwkt_migratecpu(int cpuid)
1557{
ecdefdda
MD
1558 globaldata_t rgd;
1559
1560 rgd = globaldata_find(cpuid);
1561 lwkt_setcpu_self(rgd);
ecdefdda
MD
1562}
1563
5d21b981
MD
1564/*
1565 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
1566 * do not have to enter another one).
1567 *
1568 * The thread (td) has already been completely descheduled from the
1569 * originating cpu and we can simply assert the case. The thread is
1570 * assigned to the new cpu and enqueued.
5d21b981 1571 *
cc9b6223 1572 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1573 */
1574static void
1575lwkt_setcpu_remote(void *arg)
1576{
1577 thread_t td = arg;
1578 globaldata_t gd = mycpu;
1579
cc9b6223 1580 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1581 td->td_gd = gd;
562273ea 1582 cpu_mfence();
5d21b981 1583 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1584 KKASSERT(td->td_migrate_gd == NULL);
4643740a
MD
1585 KKASSERT(td->td_lwp == NULL ||
1586 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
5d21b981
MD
1587 _lwkt_enqueue(td);
1588}
1589
553ea3c8 1590struct lwp *
4b5f931b
MD
1591lwkt_preempted_proc(void)
1592{
73e4f7b9 1593 thread_t td = curthread;
4b5f931b
MD
1594 while (td->td_preempted)
1595 td = td->td_preempted;
553ea3c8 1596 return(td->td_lwp);
4b5f931b
MD
1597}
1598
99df837e
MD
1599/*
1600 * Create a kernel process/thread/whatever. It shares it's address space
1601 * with proc0 - ie: kernel only.
1602 *
d2d8515b
MD
1603 * If the cpu is not specified one will be selected. In the future
1604 * specifying a cpu of -1 will enable kernel thread migration between
1605 * cpus.
99df837e
MD
1606 */
1607int
c9e9fb21
MD
1608lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1609 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1610{
73e4f7b9 1611 thread_t td;
e2565a42 1612 __va_list ap;
99df837e 1613
d3d32139 1614 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1615 tdflags);
a2a5ad0d
MD
1616 if (tdp)
1617 *tdp = td;
709799ea 1618 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1619
1620 /*
1621 * Set up arg0 for 'ps' etc
1622 */
e2565a42 1623 __va_start(ap, fmt);
379210cb 1624 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1625 __va_end(ap);
99df837e
MD
1626
1627 /*
1628 * Schedule the thread to run
1629 */
4643740a
MD
1630 if (td->td_flags & TDF_NOSTART)
1631 td->td_flags &= ~TDF_NOSTART;
ef0fdad1 1632 else
4643740a 1633 lwkt_schedule(td);
99df837e
MD
1634 return 0;
1635}
1636
1637/*
1638 * Destroy an LWKT thread. Warning! This function is not called when
1639 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1640 * uses a different reaping mechanism.
1641 */
1642void
1643lwkt_exit(void)
1644{
1645 thread_t td = curthread;
c070746a 1646 thread_t std;
8826f33a 1647 globaldata_t gd;
99df837e 1648
2883d2d8
MD
1649 /*
1650 * Do any cleanup that might block here
1651 */
99df837e 1652 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1653 kprintf("kthread %p %s has exited\n", td, td->td_comm);
2883d2d8
MD
1654 biosched_done(td);
1655 dsched_exit_thread(td);
c070746a
MD
1656
1657 /*
1658 * Get us into a critical section to interlock gd_freetd and loop
1659 * until we can get it freed.
1660 *
1661 * We have to cache the current td in gd_freetd because objcache_put()ing
1662 * it would rip it out from under us while our thread is still active.
2af9d75d
MD
1663 *
1664 * We are the current thread so of course our own TDF_RUNNING bit will
1665 * be set, so unlike the lwp reap code we don't wait for it to clear.
c070746a
MD
1666 */
1667 gd = mycpu;
37af14fe 1668 crit_enter_quick(td);
2af9d75d
MD
1669 for (;;) {
1670 if (td->td_refs) {
1671 tsleep(td, 0, "tdreap", 1);
1672 continue;
1673 }
1674 if ((std = gd->gd_freetd) != NULL) {
1675 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1676 gd->gd_freetd = NULL;
1677 objcache_put(thread_cache, std);
1678 continue;
1679 }
1680 break;
c070746a 1681 }
3b4192fb
MD
1682
1683 /*
1684 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1685 * the last time. We cannot block after this point or we may end
1686 * up with a stale td on the tsleepq.
eb2adbf5
MD
1687 *
1688 * None of this may block, the critical section is the only thing
1689 * protecting tdallq and the only thing preventing new lwkt_hold()
1690 * thread refs now.
3b4192fb
MD
1691 */
1692 if (td->td_flags & TDF_TSLEEPQ)
1693 tsleep_remove(td);
37af14fe 1694 lwkt_deschedule_self(td);
e56e4dea 1695 lwkt_remove_tdallq(td);
74c9628e 1696 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1697
1698 /*
1699 * Final cleanup
1700 */
1701 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1702 if (td->td_flags & TDF_ALLOCATED_THREAD)
1703 gd->gd_freetd = td;
99df837e
MD
1704 cpu_thread_exit();
1705}
1706
e56e4dea
MD
1707void
1708lwkt_remove_tdallq(thread_t td)
1709{
1710 KKASSERT(td->td_gd == mycpu);
1711 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1712}
1713
9cf43f91
MD
1714/*
1715 * Code reduction and branch prediction improvements. Call/return
1716 * overhead on modern cpus often degenerates into 0 cycles due to
1717 * the cpu's branch prediction hardware and return pc cache. We
1718 * can take advantage of this by not inlining medium-complexity
1719 * functions and we can also reduce the branch prediction impact
1720 * by collapsing perfectly predictable branches into a single
1721 * procedure instead of duplicating it.
1722 *
1723 * Is any of this noticeable? Probably not, so I'll take the
1724 * smaller code size.
1725 */
1726void
b6468f56 1727crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1728{
b6468f56 1729 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1730}
1731
2d93b37a
MD
1732void
1733crit_panic(void)
1734{
1735 thread_t td = curthread;
850634cc 1736 int lcrit = td->td_critcount;
2d93b37a 1737
850634cc 1738 td->td_critcount = 0;
a4d95680 1739 cpu_ccfence();
850634cc 1740 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1741 /* NOT REACHED */
2d93b37a
MD
1742}
1743
bd8015ca
MD
1744/*
1745 * Called from debugger/panic on cpus which have been stopped. We must still
b19f40a4 1746 * process the IPIQ while stopped.
bd8015ca
MD
1747 *
1748 * If we are dumping also try to process any pending interrupts. This may
1749 * or may not work depending on the state of the cpu at the point it was
1750 * stopped.
1751 */
1752void
1753lwkt_smp_stopped(void)
1754{
1755 globaldata_t gd = mycpu;
1756
bd8015ca
MD
1757 if (dumping) {
1758 lwkt_process_ipiq();
b19f40a4 1759 --gd->gd_intr_nesting_level;
bd8015ca 1760 splz();
b19f40a4 1761 ++gd->gd_intr_nesting_level;
bd8015ca
MD
1762 } else {
1763 lwkt_process_ipiq();
1764 }
63cff036 1765 cpu_smp_stopped();
bd8015ca 1766}