kernel - Improve pmap hinting, improve performance
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
b12defdc 2 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
60f60350 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
60f60350 20 *
8c10bfcf
MD
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
b37f18d6 47#include <sys/kinfo.h>
8ad65e08 48#include <sys/queue.h>
7d0bac62 49#include <sys/sysctl.h>
99df837e 50#include <sys/kthread.h>
f1d1c3fa 51#include <machine/cpu.h>
99df837e 52#include <sys/lock.h>
9d265729 53#include <sys/spinlock.h>
57aa743c 54#include <sys/ktr.h>
5b49787b 55#include <sys/indefinite.h>
9d265729
MD
56
57#include <sys/thread2.h>
58#include <sys/spinlock2.h>
5b49787b 59#include <sys/indefinite2.h>
f1d1c3fa 60
8c72e3d5
AH
61#include <sys/dsched.h>
62
7d0bac62
MD
63#include <vm/vm.h>
64#include <vm/vm_param.h>
65#include <vm/vm_kern.h>
66#include <vm/vm_object.h>
67#include <vm/vm_page.h>
68#include <vm/vm_map.h>
69#include <vm/vm_pager.h>
70#include <vm/vm_extern.h>
7d0bac62 71
99df837e 72#include <machine/stdarg.h>
96728c05 73#include <machine/smp.h>
3a06728e 74#include <machine/clock.h>
99df837e 75
8cee56f4
MD
76#ifdef _KERNEL_VIRTUAL
77#include <pthread.h>
78#endif
79
3a06728e
MD
80#define LOOPMASK
81
d850923c
AE
82#if !defined(KTR_CTXSW)
83#define KTR_CTXSW KTR_ALL
84#endif
85KTR_INFO_MASTER(ctxsw);
5bf48697
AE
86KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td);
87KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td);
88KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm);
89KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td);
1541028a 90
40aaf5fc
NT
91static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
92
0f7a3396
MD
93#ifdef INVARIANTS
94static int panic_on_cscount = 0;
95#endif
e28c8ef4
SW
96static int64_t switch_count = 0;
97static int64_t preempt_hit = 0;
98static int64_t preempt_miss = 0;
99static int64_t preempt_weird = 0;
fb0f29c4 100static int lwkt_use_spin_port;
40aaf5fc 101static struct objcache *thread_cache;
a46b4a23 102int cpu_mwait_spin = 0;
05220613 103
e381e77c 104static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
cc9b6223 105static void lwkt_setcpu_remote(void *arg);
e381e77c 106
fb0f29c4
MD
107/*
108 * We can make all thread ports use the spin backend instead of the thread
109 * backend. This should only be set to debug the spin backend.
110 */
111TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
112
0f7a3396 113#ifdef INVARIANTS
0c52fa62
SG
114SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
115 "Panic if attempting to switch lwkt's while mastering cpusync");
0f7a3396 116#endif
0c52fa62
SG
117SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
118 "Number of switched threads");
9733f757 119SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
0c52fa62 120 "Successful preemption events");
9733f757 121SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
0c52fa62
SG
122 "Failed preemption events");
123SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
124 "Number of preempted threads.");
b12defdc 125static int fairq_enable = 0;
2a418930
MD
126SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
127 &fairq_enable, 0, "Turn on fairq priority accumulators");
85946b6c 128static int fairq_bypass = -1;
b12defdc
MD
129SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
130 &fairq_bypass, 0, "Allow fairq to bypass td on token failure");
131extern int lwkt_sched_debug;
132int lwkt_sched_debug = 0;
133SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
134 &lwkt_sched_debug, 0, "Scheduler debug");
7fb451cb
MD
135static u_int lwkt_spin_loops = 10;
136SYSCTL_UINT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
b12defdc 137 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
fbc024e4 138static int preempt_enable = 1;
2a418930
MD
139SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
140 &preempt_enable, 0, "Enable preemption");
7b234d8c 141static int lwkt_cache_threads = 0;
765b1ae0
MD
142SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
143 &lwkt_cache_threads, 0, "thread+kstack cache");
fbc024e4 144
4b5f931b
MD
145/*
146 * These helper procedures handle the runq, they can only be called from
147 * within a critical section.
75cdbe6c
MD
148 *
149 * WARNING! Prior to SMP being brought up it is possible to enqueue and
150 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
151 * instead of 'mycpu' when referencing the globaldata structure. Once
152 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 153 */
f1d1c3fa
MD
154static __inline
155void
156_lwkt_dequeue(thread_t td)
157{
158 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 159 struct globaldata *gd = td->td_gd;
4b5f931b 160
f1d1c3fa 161 td->td_flags &= ~TDF_RUNQ;
f9235b6d 162 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
de4d4cb0 163 --gd->gd_tdrunqcount;
f9235b6d 164 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
2a418930 165 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
166 }
167}
168
f9235b6d
MD
169/*
170 * Priority enqueue.
171 *
d992c377
MD
172 * There are a limited number of lwkt threads runnable since user
173 * processes only schedule one at a time per cpu. However, there can
174 * be many user processes in kernel mode exiting from a tsleep() which
e3e6be1f 175 * become runnable.
d992c377 176 *
e6b81333
MD
177 * We scan the queue in both directions to help deal with degenerate
178 * situations when hundreds or thousands (or more) threads are runnable.
179 *
d992c377
MD
180 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and
181 * will ignore user priority. This is to ensure that user threads in
182 * kernel mode get cpu at some point regardless of what the user
183 * scheduler thinks.
f9235b6d 184 */
f1d1c3fa
MD
185static __inline
186void
187_lwkt_enqueue(thread_t td)
188{
e6b81333
MD
189 thread_t xtd; /* forward scan */
190 thread_t rtd; /* reverse scan */
f9235b6d 191
7f5d7ed7 192 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 193 struct globaldata *gd = td->td_gd;
4b5f931b 194
f1d1c3fa 195 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
196 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
197 if (xtd == NULL) {
85946b6c
MD
198 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
199 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
f9235b6d 200 } else {
e3e6be1f
MD
201 /*
202 * NOTE: td_upri - higher numbers more desireable, same sense
203 * as td_pri (typically reversed from lwp_upri).
204 *
205 * In the equal priority case we want the best selection
206 * at the beginning so the less desireable selections know
207 * that they have to setrunqueue/go-to-another-cpu, even
208 * though it means switching back to the 'best' selection.
209 * This also avoids degenerate situations when many threads
210 * are runnable or waking up at the same time.
211 *
212 * If upri matches exactly place at end/round-robin.
213 */
e6b81333
MD
214 rtd = TAILQ_LAST(&gd->gd_tdrunq, lwkt_queue);
215
d992c377 216 while (xtd &&
e6b81333 217 (xtd->td_pri > td->td_pri ||
d992c377 218 (xtd->td_pri == td->td_pri &&
e3e6be1f 219 xtd->td_upri >= td->td_upri))) {
85946b6c 220 xtd = TAILQ_NEXT(xtd, td_threadq);
e6b81333
MD
221
222 /*
223 * Doing a reverse scan at the same time is an optimization
224 * for the insert-closer-to-tail case that avoids having to
225 * scan the entire list. This situation can occur when
226 * thousands of threads are woken up at the same time.
227 */
228 if (rtd->td_pri > td->td_pri ||
229 (rtd->td_pri == td->td_pri &&
230 rtd->td_upri >= td->td_upri)) {
231 TAILQ_INSERT_AFTER(&gd->gd_tdrunq, rtd, td, td_threadq);
232 goto skip;
233 }
234 rtd = TAILQ_PREV(rtd, lwkt_queue, td_threadq);
d992c377 235 }
85946b6c
MD
236 if (xtd)
237 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
238 else
239 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
f9235b6d 240 }
e6b81333 241skip:
de4d4cb0 242 ++gd->gd_tdrunqcount;
b12defdc
MD
243
244 /*
85946b6c 245 * Request a LWKT reschedule if we are now at the head of the queue.
b12defdc 246 */
85946b6c
MD
247 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
248 need_lwkt_resched();
f1d1c3fa
MD
249 }
250}
8ad65e08 251
e28c8ef4 252static boolean_t
40aaf5fc
NT
253_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
254{
255 struct thread *td = (struct thread *)obj;
256
257 td->td_kstack = NULL;
258 td->td_kstack_size = 0;
259 td->td_flags = TDF_ALLOCATED_THREAD;
4643740a 260 td->td_mpflags = 0;
40aaf5fc
NT
261 return (1);
262}
263
264static void
265_lwkt_thread_dtor(void *obj, void *privdata)
266{
267 struct thread *td = (struct thread *)obj;
268
269 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
270 ("_lwkt_thread_dtor: not allocated from objcache"));
271 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
272 td->td_kstack_size > 0,
273 ("_lwkt_thread_dtor: corrupted stack"));
274 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
7b234d8c
MD
275 td->td_kstack = NULL;
276 td->td_flags = 0;
40aaf5fc
NT
277}
278
279/*
280 * Initialize the lwkt s/system.
765b1ae0 281 *
7b234d8c
MD
282 * Nominally cache up to 32 thread + kstack structures. Cache more on
283 * systems with a lot of cpu cores.
40aaf5fc 284 */
ced589cb 285static void
40aaf5fc
NT
286lwkt_init(void)
287{
765b1ae0 288 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
7b234d8c
MD
289 if (lwkt_cache_threads == 0) {
290 lwkt_cache_threads = ncpus * 4;
291 if (lwkt_cache_threads < 32)
292 lwkt_cache_threads = 32;
293 }
765b1ae0
MD
294 thread_cache = objcache_create_mbacked(
295 M_THREAD, sizeof(struct thread),
2fce2579 296 0, lwkt_cache_threads,
765b1ae0 297 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc 298}
ced589cb 299SYSINIT(lwkt_init, SI_BOOT2_LWKT_INIT, SI_ORDER_FIRST, lwkt_init, NULL);
40aaf5fc 300
37af14fe
MD
301/*
302 * Schedule a thread to run. As the current thread we can always safely
303 * schedule ourselves, and a shortcut procedure is provided for that
304 * function.
305 *
306 * (non-blocking, self contained on a per cpu basis)
307 */
308void
309lwkt_schedule_self(thread_t td)
310{
cfaeae2a 311 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 312 crit_enter_quick(td);
f9235b6d
MD
313 KASSERT(td != &td->td_gd->gd_idlethread,
314 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
4643740a
MD
315 KKASSERT(td->td_lwp == NULL ||
316 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
37af14fe 317 _lwkt_enqueue(td);
37af14fe
MD
318 crit_exit_quick(td);
319}
320
321/*
322 * Deschedule a thread.
323 *
324 * (non-blocking, self contained on a per cpu basis)
325 */
326void
327lwkt_deschedule_self(thread_t td)
328{
329 crit_enter_quick(td);
37af14fe
MD
330 _lwkt_dequeue(td);
331 crit_exit_quick(td);
332}
333
8ad65e08
MD
334/*
335 * LWKTs operate on a per-cpu basis
336 *
73e4f7b9 337 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
338 */
339void
340lwkt_gdinit(struct globaldata *gd)
341{
f9235b6d 342 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 343 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
344}
345
7d0bac62
MD
346/*
347 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
348 * or LWKT start address before it can be scheduled. If the target cpu is
349 * -1 the thread will be created on the current cpu.
0cfcada1
MD
350 *
351 * If you intend to create a thread without a process context this function
352 * does everything except load the startup and switcher function.
7d0bac62
MD
353 */
354thread_t
d3d32139 355lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 356{
d2d8515b 357 static int cpu_rotator;
c070746a 358 globaldata_t gd = mycpu;
99df837e 359 void *stack;
7d0bac62 360
c070746a
MD
361 /*
362 * If static thread storage is not supplied allocate a thread. Reuse
363 * a cached free thread if possible. gd_freetd is used to keep an exiting
364 * thread intact through the exit.
365 */
ef0fdad1 366 if (td == NULL) {
cf709dd2
MD
367 crit_enter_gd(gd);
368 if ((td = gd->gd_freetd) != NULL) {
369 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
370 TDF_RUNQ)) == 0);
c070746a 371 gd->gd_freetd = NULL;
cf709dd2 372 } else {
c070746a 373 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
374 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
375 TDF_RUNQ)) == 0);
376 }
377 crit_exit_gd(gd);
40aaf5fc 378 KASSERT((td->td_flags &
2af9d75d
MD
379 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
380 TDF_ALLOCATED_THREAD,
40aaf5fc
NT
381 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
382 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 383 }
c070746a
MD
384
385 /*
386 * Try to reuse cached stack.
387 */
f470d0c8
MD
388 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
389 if (flags & TDF_ALLOCATED_STACK) {
e4846942 390 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
391 stack = NULL;
392 }
393 }
394 if (stack == NULL) {
070a58b3
MD
395 if (cpu < 0)
396 stack = (void *)kmem_alloc_stack(&kernel_map, stksize, 0);
397 else
398 stack = (void *)kmem_alloc_stack(&kernel_map, stksize,
399 KM_CPU(cpu));
ef0fdad1 400 flags |= TDF_ALLOCATED_STACK;
99df837e 401 }
d2d8515b
MD
402 if (cpu < 0) {
403 cpu = ++cpu_rotator;
404 cpu_ccfence();
5dd85b08 405 cpu = (uint32_t)cpu % (uint32_t)ncpus;
d2d8515b
MD
406 }
407 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 408 return(td);
7d0bac62
MD
409}
410
411/*
412 * Initialize a preexisting thread structure. This function is used by
413 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
414 *
f8c3996b
MD
415 * All threads start out in a critical section at a priority of
416 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
417 * appropriate. This function may send an IPI message when the
418 * requested cpu is not the current cpu and consequently gd_tdallq may
419 * not be initialized synchronously from the point of view of the originating
420 * cpu.
421 *
422 * NOTE! we have to be careful in regards to creating threads for other cpus
423 * if SMP has not yet been activated.
7d0bac62 424 */
75cdbe6c
MD
425static void
426lwkt_init_thread_remote(void *arg)
427{
428 thread_t td = arg;
429
52eedfb5
MD
430 /*
431 * Protected by critical section held by IPI dispatch
432 */
75cdbe6c
MD
433 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
434}
435
fdce8919
MD
436/*
437 * lwkt core thread structural initialization.
438 *
439 * NOTE: All threads are initialized as mpsafe threads.
440 */
7d0bac62 441void
f470d0c8
MD
442lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
443 struct globaldata *gd)
7d0bac62 444{
37af14fe
MD
445 globaldata_t mygd = mycpu;
446
99df837e
MD
447 bzero(td, sizeof(struct thread));
448 td->td_kstack = stack;
f470d0c8 449 td->td_kstack_size = stksize;
d3d32139 450 td->td_flags = flags;
4643740a 451 td->td_mpflags = 0;
f256b6c0 452 td->td_type = TD_TYPE_GENERIC;
26a0694b 453 td->td_gd = gd;
f9235b6d
MD
454 td->td_pri = TDPRI_KERN_DAEMON;
455 td->td_critcount = 1;
54341a3b 456 td->td_toks_have = NULL;
3b998fa9 457 td->td_toks_stop = &td->td_toks_base;
c068fb59
SZ
458 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) {
459 lwkt_initport_spin(&td->td_msgport, td,
460 (flags & TDF_FIXEDCPU) ? TRUE : FALSE);
461 } else {
fb0f29c4 462 lwkt_initport_thread(&td->td_msgport, td);
c068fb59 463 }
99df837e 464 pmap_init_thread(td);
5d21b981
MD
465 /*
466 * Normally initializing a thread for a remote cpu requires sending an
467 * IPI. However, the idlethread is setup before the other cpus are
468 * activated so we have to treat it as a special case. XXX manipulation
469 * of gd_tdallq requires the BGL.
470 */
471 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 472 crit_enter_gd(mygd);
75cdbe6c 473 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 474 crit_exit_gd(mygd);
75cdbe6c 475 } else {
2db3b277 476 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 477 }
3573cf7b 478 dsched_enter_thread(td);
73e4f7b9
MD
479}
480
481void
482lwkt_set_comm(thread_t td, const char *ctl, ...)
483{
e2565a42 484 __va_list va;
73e4f7b9 485
e2565a42 486 __va_start(va, ctl);
379210cb 487 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 488 __va_end(va);
5bf48697 489 KTR_LOG(ctxsw_newtd, td, td->td_comm);
7d0bac62
MD
490}
491
eb2adbf5
MD
492/*
493 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
494 * this does not prevent the thread from migrating to another cpu so the
495 * gd_tdallq state is not protected by this.
496 */
99df837e 497void
73e4f7b9 498lwkt_hold(thread_t td)
99df837e 499{
74c9628e 500 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
501}
502
503void
504lwkt_rele(thread_t td)
505{
506 KKASSERT(td->td_refs > 0);
74c9628e 507 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
508}
509
73e4f7b9
MD
510void
511lwkt_free_thread(thread_t td)
512{
74c9628e 513 KKASSERT(td->td_refs == 0);
c17a6852
MD
514 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
515 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
40aaf5fc
NT
516 if (td->td_flags & TDF_ALLOCATED_THREAD) {
517 objcache_put(thread_cache, td);
518 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
519 /* client-allocated struct with internally allocated stack */
520 KASSERT(td->td_kstack && td->td_kstack_size > 0,
521 ("lwkt_free_thread: corrupted stack"));
522 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
523 td->td_kstack = NULL;
524 td->td_kstack_size = 0;
99df837e 525 }
a86ce0cd 526
e7c0dbba 527 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
528}
529
530
8ad65e08
MD
531/*
532 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
533 * switch to the idlethread. Switching must occur within a critical
534 * section to avoid races with the scheduling queue.
535 *
536 * We always have full control over our cpu's run queue. Other cpus
537 * that wish to manipulate our queue must use the cpu_*msg() calls to
538 * talk to our cpu, so a critical section is all that is needed and
539 * the result is very, very fast thread switching.
540 *
96728c05
MD
541 * The LWKT scheduler uses a fixed priority model and round-robins at
542 * each priority level. User process scheduling is a totally
543 * different beast and LWKT priorities should not be confused with
544 * user process priorities.
f1d1c3fa 545 *
69d78e99
MD
546 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
547 * is not called by the current thread in the preemption case, only when
548 * the preempting thread blocks (in order to return to the original thread).
cfaeae2a
MD
549 *
550 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
551 * migration and tsleep deschedule the current lwkt thread and call
552 * lwkt_switch(). In particular, the target cpu of the migration fully
553 * expects the thread to become non-runnable and can deadlock against
554 * cpusync operations if we run any IPIs prior to switching the thread out.
555 *
556 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
95858b91 557 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
8ad65e08
MD
558 */
559void
560lwkt_switch(void)
561{
37af14fe
MD
562 globaldata_t gd = mycpu;
563 thread_t td = gd->gd_curthread;
8ad65e08 564 thread_t ntd;
c6a766f4 565 thread_t xtd;
5411d8f1 566 int upri;
3a06728e
MD
567#ifdef LOOPMASK
568 uint64_t tsc_base = rdtsc();
569#endif
8ad65e08 570
da0b0e8b 571 KKASSERT(gd->gd_processing_ipiq == 0);
121f93bc 572 KKASSERT(td->td_flags & TDF_RUNNING);
da0b0e8b 573
46a3f46d 574 /*
27e88a6e
MD
575 * Switching from within a 'fast' (non thread switched) interrupt or IPI
576 * is illegal. However, we may have to do it anyway if we hit a fatal
577 * kernel trap or we have paniced.
578 *
579 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 580 */
27e88a6e
MD
581 if (gd->gd_intr_nesting_level) {
582 int savegdnest;
583 int savegdtrap;
584
5fddbda2 585 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22 586 panic("lwkt_switch: Attempt to switch from a "
5a8df152 587 "fast interrupt, ipi, or hard code section, "
4a28fe22
MD
588 "td %p\n",
589 td);
27e88a6e
MD
590 } else {
591 savegdnest = gd->gd_intr_nesting_level;
592 savegdtrap = gd->gd_trap_nesting_level;
593 gd->gd_intr_nesting_level = 0;
594 gd->gd_trap_nesting_level = 0;
a7422615
MD
595 if ((td->td_flags & TDF_PANICWARN) == 0) {
596 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
597 kprintf("Warning: thread switch from interrupt, IPI, "
598 "or hard code section.\n"
a7422615 599 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 600 print_backtrace(-1);
a7422615 601 }
27e88a6e
MD
602 lwkt_switch();
603 gd->gd_intr_nesting_level = savegdnest;
604 gd->gd_trap_nesting_level = savegdtrap;
605 return;
606 }
96728c05 607 }
ef0fdad1 608
cb973d15 609 /*
85946b6c
MD
610 * Release our current user process designation if we are blocking
611 * or if a user reschedule was requested.
612 *
613 * NOTE: This function is NOT called if we are switching into or
614 * returning from a preemption.
615 *
616 * NOTE: Releasing our current user process designation may cause
617 * it to be assigned to another thread, which in turn will
618 * cause us to block in the usched acquire code when we attempt
619 * to return to userland.
620 *
621 * NOTE: On SMP systems this can be very nasty when heavy token
622 * contention is present so we want to be careful not to
623 * release the designation gratuitously.
cb973d15 624 */
85946b6c
MD
625 if (td->td_release &&
626 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
cb973d15 627 td->td_release(td);
85946b6c 628 }
cb973d15 629
85946b6c 630 /*
5411d8f1
MD
631 * Release all tokens. Once we do this we must remain in the critical
632 * section and cannot run IPIs or other interrupts until we switch away
633 * because they may implode if they try to get a token using our thread
634 * context.
85946b6c 635 */
37af14fe 636 crit_enter_gd(gd);
3b998fa9 637 if (TD_TOKS_HELD(td))
9d265729
MD
638 lwkt_relalltokens(td);
639
640 /*
b02926de
MD
641 * We had better not be holding any spin locks, but don't get into an
642 * endless panic loop.
9d265729 643 */
0846e4ce 644 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL,
d666840a 645 ("lwkt_switch: still holding %d exclusive spinlocks!",
0846e4ce 646 gd->gd_spinlocks));
9d265729 647
0f7a3396
MD
648#ifdef INVARIANTS
649 if (td->td_cscount) {
6ea70f76 650 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
651 td);
652 if (panic_on_cscount)
653 panic("switching while mastering cpusync");
654 }
8a8d5d85 655#endif
f9235b6d
MD
656
657 /*
658 * If we had preempted another thread on this cpu, resume the preempted
659 * thread. This occurs transparently, whether the preempted thread
660 * was scheduled or not (it may have been preempted after descheduling
661 * itself).
662 *
663 * We have to setup the MP lock for the original thread after backing
664 * out the adjustment that was made to curthread when the original
665 * was preempted.
666 */
99df837e 667 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
668 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
669 ntd->td_flags |= TDF_PREEMPT_DONE;
7fb451cb 670 ntd->td_contended = 0; /* reset contended */
8ec60c3f
MD
671
672 /*
b9eb1c19
MD
673 * The interrupt may have woken a thread up, we need to properly
674 * set the reschedule flag if the originally interrupted thread is
675 * at a lower priority.
85946b6c 676 *
c6a766f4
MD
677 * NOTE: The interrupt may not have descheduled ntd.
678 *
679 * NOTE: We do not reschedule if there are no threads on the runq.
680 * (ntd could be the idlethread).
8ec60c3f 681 */
c6a766f4
MD
682 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
683 if (xtd && xtd != ntd)
8ec60c3f 684 need_lwkt_resched();
f9235b6d
MD
685 goto havethread_preempted;
686 }
687
b12defdc 688 /*
5411d8f1
MD
689 * Figure out switch target. If we cannot switch to our desired target
690 * look for a thread that we can switch to.
cfaeae2a 691 *
5411d8f1
MD
692 * NOTE! The limited spin loop and related parameters are extremely
693 * important for system performance, particularly for pipes and
694 * concurrent conflicting VM faults.
f9235b6d 695 */
5411d8f1
MD
696 clear_lwkt_resched();
697 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
698
699 if (ntd) {
700 do {
701 if (TD_TOKS_NOT_HELD(ntd) ||
702 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops)))
703 {
704 goto havethread;
705 }
7fb451cb 706 ++ntd->td_contended; /* overflow ok */
b1793cc6
MD
707 if (gd->gd_indefinite.type == 0)
708 indefinite_init(&gd->gd_indefinite, NULL, 0, 't');
3a06728e
MD
709#ifdef LOOPMASK
710 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
711 kprintf("lwkt_switch: excessive contended %d "
712 "thread %p\n", ntd->td_contended, ntd);
713 tsc_base = rdtsc();
714 }
715#endif
5411d8f1
MD
716 } while (ntd->td_contended < (lwkt_spin_loops >> 1));
717 upri = ntd->td_upri;
f9235b6d 718
f9235b6d 719 /*
5411d8f1
MD
720 * Bleh, the thread we wanted to switch to has a contended token.
721 * See if we can switch to another thread.
2a418930 722 *
5411d8f1
MD
723 * We generally don't want to do this because it represents a
724 * priority inversion. Do not allow the case if the thread
725 * is returning to userland (not a kernel thread) AND the thread
726 * has a lower upri.
f9235b6d 727 */
b12defdc 728 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
5411d8f1
MD
729 if (ntd->td_pri < TDPRI_KERN_LPSCHED && upri > ntd->td_upri)
730 break;
731 upri = ntd->td_upri;
b12defdc 732
5411d8f1
MD
733 /*
734 * Try this one.
735 */
736 if (TD_TOKS_NOT_HELD(ntd) ||
737 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) {
738 goto havethread;
739 }
7fb451cb 740 ++ntd->td_contended; /* overflow ok */
b12defdc 741 }
b12defdc
MD
742
743 /*
5411d8f1
MD
744 * Fall through, switch to idle thread to get us out of the current
745 * context. Since we were contended, prevent HLT by flagging a
746 * LWKT reschedule.
b12defdc 747 */
5411d8f1 748 need_lwkt_resched();
f1d1c3fa 749 }
8a8d5d85 750
5411d8f1
MD
751 /*
752 * We either contended on ntd or the runq is empty. We must switch
753 * through the idle thread to get out of the current context.
754 */
755 ntd = &gd->gd_idlethread;
756 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
757 ASSERT_NO_TOKENS_HELD(ntd);
758 cpu_time.cp_msg[0] = 0;
5411d8f1
MD
759 goto haveidle;
760
2a418930 761havethread:
b12defdc 762 /*
be71787b
MD
763 * Clear gd_idle_repeat when doing a normal switch to a non-idle
764 * thread.
f9235b6d 765 */
9ac1ee6e 766 ntd->td_wmesg = NULL;
7fb451cb 767 ntd->td_contended = 0; /* reset once scheduled */
b12defdc 768 ++gd->gd_cnt.v_swtch;
be71787b 769 gd->gd_idle_repeat = 0;
2a418930 770
5b49787b
MD
771 /*
772 * If we were busy waiting record final disposition
773 */
b1793cc6
MD
774 if (gd->gd_indefinite.type)
775 indefinite_done(&gd->gd_indefinite);
5b49787b 776
f9235b6d 777havethread_preempted:
f9235b6d
MD
778 /*
779 * If the new target does not need the MP lock and we are holding it,
780 * release the MP lock. If the new target requires the MP lock we have
781 * already acquired it for the target.
8a8d5d85 782 */
2a418930 783 ;
f9235b6d
MD
784haveidle:
785 KASSERT(ntd->td_critcount,
b5d16701
MD
786 ("priority problem in lwkt_switch %d %d",
787 td->td_critcount, ntd->td_critcount));
788
94f6d86e 789 if (td != ntd) {
cc9b6223
MD
790 /*
791 * Execute the actual thread switch operation. This function
792 * returns to the current thread and returns the previous thread
793 * (which may be different from the thread we switched to).
794 *
795 * We are responsible for marking ntd as TDF_RUNNING.
796 */
121f93bc 797 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
94f6d86e 798 ++switch_count;
a1f0fb66 799 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
800 ntd->td_flags |= TDF_RUNNING;
801 lwkt_switch_return(td->td_switch(ntd));
802 /* ntd invalid, td_switch() can return a different thread_t */
94f6d86e 803 }
b12defdc 804
b12defdc 805 /*
54341a3b 806 * catch-all. XXX is this strictly needed?
b12defdc
MD
807 */
808 splz_check();
54341a3b 809
37af14fe
MD
810 /* NOTE: current cpu may have changed after switch */
811 crit_exit_quick(td);
8ad65e08
MD
812}
813
cc9b6223
MD
814/*
815 * Called by assembly in the td_switch (thread restore path) for thread
816 * bootstrap cases which do not 'return' to lwkt_switch().
817 */
818void
819lwkt_switch_return(thread_t otd)
820{
cc9b6223 821 globaldata_t rgd;
3a06728e
MD
822#ifdef LOOPMASK
823 uint64_t tsc_base = rdtsc();
824#endif
825 int exiting;
826
827 exiting = otd->td_flags & TDF_EXITING;
828 cpu_ccfence();
cc9b6223
MD
829
830 /*
831 * Check if otd was migrating. Now that we are on ntd we can finish
832 * up the migration. This is a bit messy but it is the only place
833 * where td is known to be fully descheduled.
834 *
835 * We can only activate the migration if otd was migrating but not
836 * held on the cpu due to a preemption chain. We still have to
837 * clear TDF_RUNNING on the old thread either way.
838 *
839 * We are responsible for clearing the previously running thread's
840 * TDF_RUNNING.
841 */
842 if ((rgd = otd->td_migrate_gd) != NULL &&
843 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
844 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
845 (TDF_MIGRATING | TDF_RUNNING));
846 otd->td_migrate_gd = NULL;
847 otd->td_flags &= ~TDF_RUNNING;
848 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
849 } else {
850 otd->td_flags &= ~TDF_RUNNING;
851 }
2b07d9aa
MD
852
853 /*
854 * Final exit validations (see lwp_wait()). Note that otd becomes
855 * invalid the *instant* we set TDF_MP_EXITSIG.
3a06728e
MD
856 *
857 * Use the EXITING status loaded from before we clear TDF_RUNNING,
858 * because if it is not set otd becomes invalid the instant we clear
859 * TDF_RUNNING on it (otherwise, if the system is fast enough, we
860 * might 'steal' TDF_EXITING from another switch-return!).
2b07d9aa 861 */
3a06728e 862 while (exiting) {
2b07d9aa
MD
863 u_int mpflags;
864
865 mpflags = otd->td_mpflags;
866 cpu_ccfence();
867
868 if (mpflags & TDF_MP_EXITWAIT) {
869 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
870 mpflags | TDF_MP_EXITSIG)) {
871 wakeup(otd);
872 break;
873 }
874 } else {
875 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
876 mpflags | TDF_MP_EXITSIG)) {
877 wakeup(otd);
878 break;
879 }
880 }
3a06728e
MD
881
882#ifdef LOOPMASK
883 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
884 kprintf("lwkt_switch_return: excessive TDF_EXITING "
885 "thread %p\n", otd);
886 tsc_base = rdtsc();
887 }
888#endif
2b07d9aa 889 }
cc9b6223
MD
890}
891
b68b7282 892/*
96728c05 893 * Request that the target thread preempt the current thread. Preemption
203592a0
MD
894 * can only occur only:
895 *
896 * - If our critical section is the one that we were called with
897 * - The relative priority of the target thread is higher
898 * - The target is not excessively interrupt-nested via td_nest_count
899 * - The target thread holds no tokens.
900 * - The target thread is not already scheduled and belongs to the
901 * current cpu.
902 * - The current thread is not holding any spin-locks.
96728c05
MD
903 *
904 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
905 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 906 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
907 * to determine whether preemption is possible (aka usually just the crit
908 * priority of lwkt_schedule() itself).
b68b7282 909 *
54341a3b
MD
910 * Preemption is typically limited to interrupt threads.
911 *
912 * Operation works in a fairly straight-forward manner. The normal
913 * scheduling code is bypassed and we switch directly to the target
914 * thread. When the target thread attempts to block or switch away
915 * code at the base of lwkt_switch() will switch directly back to our
916 * thread. Our thread is able to retain whatever tokens it holds and
917 * if the target needs one of them the target will switch back to us
918 * and reschedule itself normally.
b68b7282
MD
919 */
920void
f9235b6d 921lwkt_preempt(thread_t ntd, int critcount)
b68b7282 922{
46a3f46d 923 struct globaldata *gd = mycpu;
cc9b6223 924 thread_t xtd;
0a3f9b47 925 thread_t td;
2d910aaf 926 int save_gd_intr_nesting_level;
b68b7282 927
26a0694b 928 /*
96728c05
MD
929 * The caller has put us in a critical section. We can only preempt
930 * if the caller of the caller was not in a critical section (basically
f9235b6d 931 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 932 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 933 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
934 *
935 * YYY The target thread must be in a critical section (else it must
936 * inherit our critical section? I dunno yet).
26a0694b 937 */
f9235b6d 938 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 939
b12defdc 940 td = gd->gd_curthread;
fbc024e4
MD
941 if (preempt_enable == 0) {
942 ++preempt_miss;
943 return;
944 }
f9235b6d 945 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
946 ++preempt_miss;
947 return;
948 }
f9235b6d 949 if (td->td_critcount > critcount) {
96728c05
MD
950 ++preempt_miss;
951 return;
952 }
203592a0
MD
953 if (td->td_nest_count >= 2) {
954 ++preempt_miss;
955 return;
956 }
121f93bc
MD
957 if (td->td_cscount) {
958 ++preempt_miss;
959 return;
960 }
46a3f46d 961 if (ntd->td_gd != gd) {
96728c05
MD
962 ++preempt_miss;
963 return;
964 }
ee89e80b 965
41a01a4d 966 /*
77912481
MD
967 * We don't have to check spinlocks here as they will also bump
968 * td_critcount.
d3d1cbc8
MD
969 *
970 * Do not try to preempt if the target thread is holding any tokens.
971 * We could try to acquire the tokens but this case is so rare there
972 * is no need to support it.
41a01a4d 973 */
0846e4ce 974 KKASSERT(gd->gd_spinlocks == 0);
77912481 975
3b998fa9 976 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8 977 ++preempt_miss;
d3d1cbc8
MD
978 return;
979 }
26a0694b
MD
980 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
981 ++preempt_weird;
982 return;
983 }
984 if (ntd->td_preempted) {
4b5f931b 985 ++preempt_hit;
26a0694b 986 return;
b68b7282 987 }
da0b0e8b 988 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 989
8ec60c3f
MD
990 /*
991 * Since we are able to preempt the current thread, there is no need to
992 * call need_lwkt_resched().
2d910aaf
MD
993 *
994 * We must temporarily clear gd_intr_nesting_level around the switch
995 * since switchouts from the target thread are allowed (they will just
996 * return to our thread), and since the target thread has its own stack.
cc9b6223
MD
997 *
998 * A preemption must switch back to the original thread, assert the
999 * case.
8ec60c3f 1000 */
26a0694b
MD
1001 ++preempt_hit;
1002 ntd->td_preempted = td;
1003 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1004 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1005 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1006 gd->gd_intr_nesting_level = 0;
121f93bc
MD
1007
1008 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
cc9b6223
MD
1009 ntd->td_flags |= TDF_RUNNING;
1010 xtd = td->td_switch(ntd);
1011 KKASSERT(xtd == ntd);
1012 lwkt_switch_return(xtd);
2d910aaf 1013 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1014
26a0694b
MD
1015 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1016 ntd->td_preempted = NULL;
1017 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1018}
1019
f1d1c3fa 1020/*
faaeffac 1021 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1022 * This will work inside a critical section but not inside a hard code
1023 * section.
ef0fdad1 1024 *
f1d1c3fa
MD
1025 * (self contained on a per cpu basis)
1026 */
1027void
faaeffac 1028splz_check(void)
f1d1c3fa 1029{
7966cb69
MD
1030 globaldata_t gd = mycpu;
1031 thread_t td = gd->gd_curthread;
ef0fdad1 1032
4a28fe22
MD
1033 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1034 gd->gd_intr_nesting_level == 0 &&
1035 td->td_nest_count < 2)
1036 {
f1d1c3fa 1037 splz();
4a28fe22
MD
1038 }
1039}
1040
1041/*
1042 * This version is integrated into crit_exit, reqflags has already
1043 * been tested but td_critcount has not.
1044 *
1045 * We only want to execute the splz() on the 1->0 transition of
1046 * critcount and not in a hard code section or if too deeply nested.
925040f2 1047 *
0846e4ce 1048 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
4a28fe22
MD
1049 */
1050void
1051lwkt_maybe_splz(thread_t td)
1052{
1053 globaldata_t gd = td->td_gd;
1054
1055 if (td->td_critcount == 0 &&
1056 gd->gd_intr_nesting_level == 0 &&
1057 td->td_nest_count < 2)
1058 {
1059 splz();
1060 }
f1d1c3fa
MD
1061}
1062
e6546af9
MD
1063/*
1064 * Drivers which set up processing co-threads can call this function to
1065 * run the co-thread at a higher priority and to allow it to preempt
1066 * normal threads.
1067 */
1068void
1069lwkt_set_interrupt_support_thread(void)
1070{
1071 thread_t td = curthread;
1072
1073 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1074 td->td_flags |= TDF_INTTHREAD;
1075 td->td_preemptable = lwkt_preempt;
1076}
1077
1078
8ad65e08 1079/*
f9235b6d
MD
1080 * This function is used to negotiate a passive release of the current
1081 * process/lwp designation with the user scheduler, allowing the user
1082 * scheduler to schedule another user thread. The related kernel thread
1083 * (curthread) continues running in the released state.
8ad65e08
MD
1084 */
1085void
f9235b6d 1086lwkt_passive_release(struct thread *td)
8ad65e08 1087{
f9235b6d
MD
1088 struct lwp *lp = td->td_lwp;
1089
1090 td->td_release = NULL;
1091 lwkt_setpri_self(TDPRI_KERN_USER);
d992c377 1092
f9235b6d 1093 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1094}
1095
f9235b6d 1096
3824f392 1097/*
d2d8515b
MD
1098 * This implements a LWKT yield, allowing a kernel thread to yield to other
1099 * kernel threads at the same or higher priority. This function can be
1100 * called in a tight loop and will typically only yield once per tick.
f9235b6d 1101 *
d2d8515b
MD
1102 * Most kernel threads run at the same priority in order to allow equal
1103 * sharing.
f9235b6d
MD
1104 *
1105 * (self contained on a per cpu basis)
3824f392
MD
1106 */
1107void
f9235b6d 1108lwkt_yield(void)
3824f392 1109{
f9235b6d
MD
1110 globaldata_t gd = mycpu;
1111 thread_t td = gd->gd_curthread;
3824f392 1112
fbe96076
MD
1113 /*
1114 * Should never be called with spinlocks held but there is a path
1115 * via ACPI where it might happen.
1116 */
1117 if (gd->gd_spinlocks)
1118 return;
1119
1120 /*
1121 * Safe to call splz if we are not too-heavily nested.
1122 */
f9235b6d
MD
1123 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1124 splz();
fbe96076
MD
1125
1126 /*
1127 * Caller allows switching
1128 */
85946b6c 1129 if (lwkt_resched_wanted()) {
f9235b6d
MD
1130 lwkt_schedule_self(curthread);
1131 lwkt_switch();
f9235b6d 1132 }
3824f392
MD
1133}
1134
40504122
MD
1135/*
1136 * The quick version processes pending interrupts and higher-priority
1137 * LWKT threads but will not round-robin same-priority LWKT threads.
de4d4cb0
MD
1138 *
1139 * When called while attempting to return to userland the only same-pri
1140 * threads are the ones which have already tried to become the current
1141 * user process.
40504122
MD
1142 */
1143void
1144lwkt_yield_quick(void)
1145{
1146 globaldata_t gd = mycpu;
1147 thread_t td = gd->gd_curthread;
1148
1149 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1150 splz();
1151 if (lwkt_resched_wanted()) {
9c99cb33 1152 crit_enter();
40504122
MD
1153 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1154 clear_lwkt_resched();
1155 } else {
1156 lwkt_schedule_self(curthread);
1157 lwkt_switch();
1158 }
9c99cb33 1159 crit_exit();
40504122
MD
1160 }
1161}
1162
3824f392 1163/*
f9235b6d
MD
1164 * This yield is designed for kernel threads with a user context.
1165 *
1166 * The kernel acting on behalf of the user is potentially cpu-bound,
1167 * this function will efficiently allow other threads to run and also
1168 * switch to other processes by releasing.
3824f392
MD
1169 *
1170 * The lwkt_user_yield() function is designed to have very low overhead
1171 * if no yield is determined to be needed.
1172 */
1173void
1174lwkt_user_yield(void)
1175{
f9235b6d
MD
1176 globaldata_t gd = mycpu;
1177 thread_t td = gd->gd_curthread;
1178
fbe96076
MD
1179 /*
1180 * Should never be called with spinlocks held but there is a path
1181 * via ACPI where it might happen.
1182 */
1183 if (gd->gd_spinlocks)
1184 return;
1185
f9235b6d
MD
1186 /*
1187 * Always run any pending interrupts in case we are in a critical
1188 * section.
1189 */
1190 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1191 splz();
3824f392 1192
3824f392 1193 /*
f9235b6d
MD
1194 * Switch (which forces a release) if another kernel thread needs
1195 * the cpu, if userland wants us to resched, or if our kernel
1196 * quantum has run out.
3824f392 1197 */
f9235b6d 1198 if (lwkt_resched_wanted() ||
85946b6c 1199 user_resched_wanted())
f9235b6d 1200 {
3824f392 1201 lwkt_switch();
3824f392
MD
1202 }
1203
f9235b6d 1204#if 0
3824f392 1205 /*
f9235b6d
MD
1206 * Reacquire the current process if we are released.
1207 *
1208 * XXX not implemented atm. The kernel may be holding locks and such,
1209 * so we want the thread to continue to receive cpu.
3824f392 1210 */
f9235b6d
MD
1211 if (td->td_release == NULL && lp) {
1212 lp->lwp_proc->p_usched->acquire_curproc(lp);
1213 td->td_release = lwkt_passive_release;
1214 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1215 }
f9235b6d 1216#endif
b9eb1c19
MD
1217}
1218
8ad65e08 1219/*
f1d1c3fa
MD
1220 * Generic schedule. Possibly schedule threads belonging to other cpus and
1221 * deal with threads that might be blocked on a wait queue.
1222 *
0a3f9b47
MD
1223 * We have a little helper inline function which does additional work after
1224 * the thread has been enqueued, including dealing with preemption and
1225 * setting need_lwkt_resched() (which prevents the kernel from returning
1226 * to userland until it has processed higher priority threads).
6330a558
MD
1227 *
1228 * It is possible for this routine to be called after a failed _enqueue
1229 * (due to the target thread migrating, sleeping, or otherwise blocked).
1230 * We have to check that the thread is actually on the run queue!
8ad65e08 1231 */
0a3f9b47
MD
1232static __inline
1233void
85946b6c 1234_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
0a3f9b47 1235{
6330a558 1236 if (ntd->td_flags & TDF_RUNQ) {
85946b6c 1237 if (ntd->td_preemptable) {
f9235b6d 1238 ntd->td_preemptable(ntd, ccount); /* YYY +token */
6330a558 1239 }
0a3f9b47
MD
1240 }
1241}
1242
361d01dd 1243static __inline
8ad65e08 1244void
85946b6c 1245_lwkt_schedule(thread_t td)
8ad65e08 1246{
37af14fe
MD
1247 globaldata_t mygd = mycpu;
1248
cf709dd2
MD
1249 KASSERT(td != &td->td_gd->gd_idlethread,
1250 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1251 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1252 crit_enter_gd(mygd);
4643740a
MD
1253 KKASSERT(td->td_lwp == NULL ||
1254 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1255
37af14fe 1256 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1257 _lwkt_enqueue(td);
1258 } else {
f1d1c3fa 1259 /*
7cd8d145
MD
1260 * If we own the thread, there is no race (since we are in a
1261 * critical section). If we do not own the thread there might
1262 * be a race but the target cpu will deal with it.
f1d1c3fa 1263 */
7cd8d145 1264 if (td->td_gd == mygd) {
9d265729 1265 _lwkt_enqueue(td);
85946b6c 1266 _lwkt_schedule_post(mygd, td, 1);
f1d1c3fa 1267 } else {
e381e77c 1268 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1269 }
8ad65e08 1270 }
37af14fe 1271 crit_exit_gd(mygd);
8ad65e08
MD
1272}
1273
361d01dd
MD
1274void
1275lwkt_schedule(thread_t td)
1276{
85946b6c 1277 _lwkt_schedule(td);
361d01dd
MD
1278}
1279
1280void
85946b6c 1281lwkt_schedule_noresched(thread_t td) /* XXX not impl */
361d01dd 1282{
85946b6c 1283 _lwkt_schedule(td);
361d01dd
MD
1284}
1285
e381e77c
MD
1286/*
1287 * When scheduled remotely if frame != NULL the IPIQ is being
1288 * run via doreti or an interrupt then preemption can be allowed.
1289 *
1290 * To allow preemption we have to drop the critical section so only
1291 * one is present in _lwkt_schedule_post.
1292 */
1293static void
1294lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1295{
1296 thread_t td = curthread;
1297 thread_t ntd = arg;
1298
1299 if (frame && ntd->td_preemptable) {
1300 crit_exit_noyield(td);
85946b6c 1301 _lwkt_schedule(ntd);
e381e77c
MD
1302 crit_enter_quick(td);
1303 } else {
85946b6c 1304 _lwkt_schedule(ntd);
e381e77c
MD
1305 }
1306}
1307
d9eea1a5 1308/*
52eedfb5
MD
1309 * Thread migration using a 'Pull' method. The thread may or may not be
1310 * the current thread. It MUST be descheduled and in a stable state.
1311 * lwkt_giveaway() must be called on the cpu owning the thread.
1312 *
1313 * At any point after lwkt_giveaway() is called, the target cpu may
1314 * 'pull' the thread by calling lwkt_acquire().
1315 *
ae8e83e6
MD
1316 * We have to make sure the thread is not sitting on a per-cpu tsleep
1317 * queue or it will blow up when it moves to another cpu.
1318 *
52eedfb5 1319 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1320 */
52eedfb5
MD
1321void
1322lwkt_giveaway(thread_t td)
1323{
3b4192fb 1324 globaldata_t gd = mycpu;
52eedfb5 1325
3b4192fb
MD
1326 crit_enter_gd(gd);
1327 if (td->td_flags & TDF_TSLEEPQ)
1328 tsleep_remove(td);
1329 KKASSERT(td->td_gd == gd);
1330 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1331 td->td_flags |= TDF_MIGRATING;
1332 crit_exit_gd(gd);
52eedfb5
MD
1333}
1334
a2a5ad0d
MD
1335void
1336lwkt_acquire(thread_t td)
1337{
37af14fe
MD
1338 globaldata_t gd;
1339 globaldata_t mygd;
a2a5ad0d 1340
52eedfb5 1341 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1342 gd = td->td_gd;
37af14fe 1343 mygd = mycpu;
52eedfb5 1344 if (gd != mycpu) {
3a06728e
MD
1345#ifdef LOOPMASK
1346 uint64_t tsc_base = rdtsc();
1347#endif
35238fa5 1348 cpu_lfence();
52eedfb5 1349 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1350 crit_enter_gd(mygd);
cfaeae2a 1351 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23 1352 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
df910c23 1353 lwkt_process_ipiq();
52eedfb5 1354 cpu_lfence();
a86ce0cd
MD
1355#ifdef _KERNEL_VIRTUAL
1356 pthread_yield();
3a06728e
MD
1357#endif
1358#ifdef LOOPMASK
1359 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
1360 kprintf("lwkt_acquire: stuck td %p td->td_flags %08x\n",
1361 td, td->td_flags);
1362 tsc_base = rdtsc();
1363 }
a86ce0cd 1364#endif
df910c23 1365 }
cfaeae2a 1366 DEBUG_POP_INFO();
562273ea 1367 cpu_mfence();
37af14fe 1368 td->td_gd = mygd;
52eedfb5
MD
1369 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1370 td->td_flags &= ~TDF_MIGRATING;
1371 crit_exit_gd(mygd);
1372 } else {
1373 crit_enter_gd(mygd);
1374 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1375 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1376 crit_exit_gd(mygd);
a2a5ad0d
MD
1377 }
1378}
1379
f1d1c3fa
MD
1380/*
1381 * Generic deschedule. Descheduling threads other then your own should be
1382 * done only in carefully controlled circumstances. Descheduling is
1383 * asynchronous.
1384 *
1385 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1386 */
1387void
1388lwkt_deschedule(thread_t td)
1389{
f1d1c3fa
MD
1390 crit_enter();
1391 if (td == curthread) {
1392 _lwkt_dequeue(td);
1393 } else {
a72187e9 1394 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1395 _lwkt_dequeue(td);
1396 } else {
b8a98473 1397 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1398 }
1399 }
1400 crit_exit();
1401}
1402
4b5f931b
MD
1403/*
1404 * Set the target thread's priority. This routine does not automatically
1405 * switch to a higher priority thread, LWKT threads are not designed for
1406 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1407 */
1408void
1409lwkt_setpri(thread_t td, int pri)
1410{
f9235b6d
MD
1411 if (td->td_pri != pri) {
1412 KKASSERT(pri >= 0);
1413 crit_enter();
1414 if (td->td_flags & TDF_RUNQ) {
d2d8515b 1415 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1416 _lwkt_dequeue(td);
1417 td->td_pri = pri;
1418 _lwkt_enqueue(td);
1419 } else {
1420 td->td_pri = pri;
1421 }
1422 crit_exit();
26a0694b 1423 }
26a0694b
MD
1424}
1425
03bd0a5e
MD
1426/*
1427 * Set the initial priority for a thread prior to it being scheduled for
1428 * the first time. The thread MUST NOT be scheduled before or during
1429 * this call. The thread may be assigned to a cpu other then the current
1430 * cpu.
1431 *
1432 * Typically used after a thread has been created with TDF_STOPPREQ,
1433 * and before the thread is initially scheduled.
1434 */
1435void
1436lwkt_setpri_initial(thread_t td, int pri)
1437{
1438 KKASSERT(pri >= 0);
1439 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1440 td->td_pri = pri;
03bd0a5e
MD
1441}
1442
26a0694b
MD
1443void
1444lwkt_setpri_self(int pri)
1445{
1446 thread_t td = curthread;
1447
4b5f931b
MD
1448 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1449 crit_enter();
1450 if (td->td_flags & TDF_RUNQ) {
1451 _lwkt_dequeue(td);
f9235b6d 1452 td->td_pri = pri;
4b5f931b
MD
1453 _lwkt_enqueue(td);
1454 } else {
f9235b6d 1455 td->td_pri = pri;
4b5f931b
MD
1456 }
1457 crit_exit();
1458}
1459
f9235b6d 1460/*
85946b6c 1461 * hz tick scheduler clock for LWKT threads
f9235b6d
MD
1462 */
1463void
85946b6c 1464lwkt_schedulerclock(thread_t td)
f9235b6d 1465{
85946b6c
MD
1466 globaldata_t gd = td->td_gd;
1467 thread_t xtd;
2a418930 1468
c6a766f4
MD
1469 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1470 if (xtd == td) {
85946b6c
MD
1471 /*
1472 * If the current thread is at the head of the runq shift it to the
1473 * end of any equal-priority threads and request a LWKT reschedule
1474 * if it moved.
d992c377
MD
1475 *
1476 * Ignore upri in this situation. There will only be one user thread
1477 * in user mode, all others will be user threads running in kernel
1478 * mode and we have to make sure they get some cpu.
85946b6c
MD
1479 */
1480 xtd = TAILQ_NEXT(td, td_threadq);
1481 if (xtd && xtd->td_pri == td->td_pri) {
1482 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1483 while (xtd && xtd->td_pri == td->td_pri)
1484 xtd = TAILQ_NEXT(xtd, td_threadq);
1485 if (xtd)
1486 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1487 else
1488 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1489 need_lwkt_resched();
f9235b6d 1490 }
c6a766f4 1491 } else if (xtd) {
85946b6c
MD
1492 /*
1493 * If we scheduled a thread other than the one at the head of the
1494 * queue always request a reschedule every tick.
1495 */
1496 need_lwkt_resched();
f9235b6d 1497 }
c6a766f4 1498 /* else curthread probably the idle thread, no need to reschedule */
f9235b6d
MD
1499}
1500
5d21b981 1501/*
52eedfb5
MD
1502 * Migrate the current thread to the specified cpu.
1503 *
cc9b6223
MD
1504 * This is accomplished by descheduling ourselves from the current cpu
1505 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1506 * 'old' thread wants to migrate after it has been completely switched out
1507 * and will complete the migration.
1508 *
1509 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1510 *
1511 * We must be sure to release our current process designation (if a user
1512 * process) before clearing out any tsleepq we are on because the release
1513 * code may re-add us.
ae8e83e6
MD
1514 *
1515 * We must be sure to remove ourselves from the current cpu's tsleepq
1516 * before potentially moving to another queue. The thread can be on
1517 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1518 */
5d21b981
MD
1519
1520void
1521lwkt_setcpu_self(globaldata_t rgd)
1522{
5d21b981
MD
1523 thread_t td = curthread;
1524
1525 if (td->td_gd != rgd) {
1526 crit_enter_quick(td);
cc9b6223 1527
95858b91
MD
1528 if (td->td_release)
1529 td->td_release(td);
ae8e83e6 1530 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1531 tsleep_remove(td);
cc9b6223
MD
1532
1533 /*
1534 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1535 * trying to deschedule ourselves and switch away, then deschedule
1536 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1537 * call lwkt_switch() to complete the operation.
1538 */
5d21b981
MD
1539 td->td_flags |= TDF_MIGRATING;
1540 lwkt_deschedule_self(td);
52eedfb5 1541 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1542 td->td_migrate_gd = rgd;
5d21b981 1543 lwkt_switch();
cc9b6223
MD
1544
1545 /*
1546 * We are now on the target cpu
1547 */
1548 KKASSERT(rgd == mycpu);
52eedfb5 1549 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1550 crit_exit_quick(td);
1551 }
5d21b981
MD
1552}
1553
ecdefdda
MD
1554void
1555lwkt_migratecpu(int cpuid)
1556{
ecdefdda
MD
1557 globaldata_t rgd;
1558
1559 rgd = globaldata_find(cpuid);
1560 lwkt_setcpu_self(rgd);
ecdefdda
MD
1561}
1562
5d21b981
MD
1563/*
1564 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
1565 * do not have to enter another one).
1566 *
1567 * The thread (td) has already been completely descheduled from the
1568 * originating cpu and we can simply assert the case. The thread is
1569 * assigned to the new cpu and enqueued.
5d21b981 1570 *
cc9b6223 1571 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1572 */
1573static void
1574lwkt_setcpu_remote(void *arg)
1575{
1576 thread_t td = arg;
1577 globaldata_t gd = mycpu;
1578
cc9b6223 1579 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1580 td->td_gd = gd;
562273ea 1581 cpu_mfence();
5d21b981 1582 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1583 KKASSERT(td->td_migrate_gd == NULL);
4643740a
MD
1584 KKASSERT(td->td_lwp == NULL ||
1585 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
5d21b981
MD
1586 _lwkt_enqueue(td);
1587}
1588
553ea3c8 1589struct lwp *
4b5f931b
MD
1590lwkt_preempted_proc(void)
1591{
73e4f7b9 1592 thread_t td = curthread;
4b5f931b
MD
1593 while (td->td_preempted)
1594 td = td->td_preempted;
553ea3c8 1595 return(td->td_lwp);
4b5f931b
MD
1596}
1597
99df837e
MD
1598/*
1599 * Create a kernel process/thread/whatever. It shares it's address space
1600 * with proc0 - ie: kernel only.
1601 *
d2d8515b
MD
1602 * If the cpu is not specified one will be selected. In the future
1603 * specifying a cpu of -1 will enable kernel thread migration between
1604 * cpus.
99df837e
MD
1605 */
1606int
c9e9fb21
MD
1607lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1608 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1609{
73e4f7b9 1610 thread_t td;
e2565a42 1611 __va_list ap;
99df837e 1612
d3d32139 1613 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1614 tdflags);
a2a5ad0d
MD
1615 if (tdp)
1616 *tdp = td;
709799ea 1617 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1618
1619 /*
1620 * Set up arg0 for 'ps' etc
1621 */
e2565a42 1622 __va_start(ap, fmt);
379210cb 1623 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1624 __va_end(ap);
99df837e
MD
1625
1626 /*
1627 * Schedule the thread to run
1628 */
4643740a
MD
1629 if (td->td_flags & TDF_NOSTART)
1630 td->td_flags &= ~TDF_NOSTART;
ef0fdad1 1631 else
4643740a 1632 lwkt_schedule(td);
99df837e
MD
1633 return 0;
1634}
1635
1636/*
1637 * Destroy an LWKT thread. Warning! This function is not called when
1638 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1639 * uses a different reaping mechanism.
1640 */
1641void
1642lwkt_exit(void)
1643{
1644 thread_t td = curthread;
c070746a 1645 thread_t std;
8826f33a 1646 globaldata_t gd;
99df837e 1647
2883d2d8
MD
1648 /*
1649 * Do any cleanup that might block here
1650 */
99df837e 1651 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1652 kprintf("kthread %p %s has exited\n", td, td->td_comm);
2883d2d8
MD
1653 biosched_done(td);
1654 dsched_exit_thread(td);
c070746a
MD
1655
1656 /*
1657 * Get us into a critical section to interlock gd_freetd and loop
1658 * until we can get it freed.
1659 *
1660 * We have to cache the current td in gd_freetd because objcache_put()ing
1661 * it would rip it out from under us while our thread is still active.
2af9d75d
MD
1662 *
1663 * We are the current thread so of course our own TDF_RUNNING bit will
1664 * be set, so unlike the lwp reap code we don't wait for it to clear.
c070746a
MD
1665 */
1666 gd = mycpu;
37af14fe 1667 crit_enter_quick(td);
2af9d75d
MD
1668 for (;;) {
1669 if (td->td_refs) {
1670 tsleep(td, 0, "tdreap", 1);
1671 continue;
1672 }
1673 if ((std = gd->gd_freetd) != NULL) {
1674 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1675 gd->gd_freetd = NULL;
1676 objcache_put(thread_cache, std);
1677 continue;
1678 }
1679 break;
c070746a 1680 }
3b4192fb
MD
1681
1682 /*
1683 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1684 * the last time. We cannot block after this point or we may end
1685 * up with a stale td on the tsleepq.
eb2adbf5
MD
1686 *
1687 * None of this may block, the critical section is the only thing
1688 * protecting tdallq and the only thing preventing new lwkt_hold()
1689 * thread refs now.
3b4192fb
MD
1690 */
1691 if (td->td_flags & TDF_TSLEEPQ)
1692 tsleep_remove(td);
37af14fe 1693 lwkt_deschedule_self(td);
e56e4dea 1694 lwkt_remove_tdallq(td);
74c9628e 1695 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1696
1697 /*
1698 * Final cleanup
1699 */
1700 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1701 if (td->td_flags & TDF_ALLOCATED_THREAD)
1702 gd->gd_freetd = td;
99df837e
MD
1703 cpu_thread_exit();
1704}
1705
e56e4dea
MD
1706void
1707lwkt_remove_tdallq(thread_t td)
1708{
1709 KKASSERT(td->td_gd == mycpu);
1710 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1711}
1712
9cf43f91
MD
1713/*
1714 * Code reduction and branch prediction improvements. Call/return
1715 * overhead on modern cpus often degenerates into 0 cycles due to
1716 * the cpu's branch prediction hardware and return pc cache. We
1717 * can take advantage of this by not inlining medium-complexity
1718 * functions and we can also reduce the branch prediction impact
1719 * by collapsing perfectly predictable branches into a single
1720 * procedure instead of duplicating it.
1721 *
1722 * Is any of this noticeable? Probably not, so I'll take the
1723 * smaller code size.
1724 */
1725void
b6468f56 1726crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1727{
b6468f56 1728 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1729}
1730
2d93b37a
MD
1731void
1732crit_panic(void)
1733{
1734 thread_t td = curthread;
850634cc 1735 int lcrit = td->td_critcount;
2d93b37a 1736
850634cc 1737 td->td_critcount = 0;
a4d95680 1738 cpu_ccfence();
850634cc 1739 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1740 /* NOT REACHED */
2d93b37a
MD
1741}
1742
bd8015ca
MD
1743/*
1744 * Called from debugger/panic on cpus which have been stopped. We must still
b19f40a4 1745 * process the IPIQ while stopped.
bd8015ca
MD
1746 *
1747 * If we are dumping also try to process any pending interrupts. This may
1748 * or may not work depending on the state of the cpu at the point it was
1749 * stopped.
1750 */
1751void
1752lwkt_smp_stopped(void)
1753{
1754 globaldata_t gd = mycpu;
1755
bd8015ca
MD
1756 if (dumping) {
1757 lwkt_process_ipiq();
b19f40a4 1758 --gd->gd_intr_nesting_level;
bd8015ca 1759 splz();
b19f40a4 1760 ++gd->gd_intr_nesting_level;
bd8015ca
MD
1761 } else {
1762 lwkt_process_ipiq();
1763 }
63cff036 1764 cpu_smp_stopped();
bd8015ca 1765}