make.conf.5: Add some words about WANT_NETGRAPH7.
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
b12defdc 2 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
60f60350 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
60f60350 20 *
8c10bfcf
MD
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
b37f18d6 47#include <sys/kinfo.h>
8ad65e08 48#include <sys/queue.h>
7d0bac62 49#include <sys/sysctl.h>
99df837e 50#include <sys/kthread.h>
f1d1c3fa 51#include <machine/cpu.h>
99df837e 52#include <sys/lock.h>
f6bf3af1 53#include <sys/caps.h>
9d265729 54#include <sys/spinlock.h>
57aa743c 55#include <sys/ktr.h>
9d265729
MD
56
57#include <sys/thread2.h>
58#include <sys/spinlock2.h>
684a93c4 59#include <sys/mplock2.h>
f1d1c3fa 60
8c72e3d5
AH
61#include <sys/dsched.h>
62
7d0bac62
MD
63#include <vm/vm.h>
64#include <vm/vm_param.h>
65#include <vm/vm_kern.h>
66#include <vm/vm_object.h>
67#include <vm/vm_page.h>
68#include <vm/vm_map.h>
69#include <vm/vm_pager.h>
70#include <vm/vm_extern.h>
7d0bac62 71
99df837e 72#include <machine/stdarg.h>
96728c05 73#include <machine/smp.h>
99df837e 74
d850923c
AE
75#if !defined(KTR_CTXSW)
76#define KTR_CTXSW KTR_ALL
77#endif
78KTR_INFO_MASTER(ctxsw);
5bf48697
AE
79KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td);
80KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td);
81KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm);
82KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td);
1541028a 83
40aaf5fc
NT
84static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
85
0f7a3396
MD
86#ifdef INVARIANTS
87static int panic_on_cscount = 0;
88#endif
05220613
MD
89static __int64_t switch_count = 0;
90static __int64_t preempt_hit = 0;
91static __int64_t preempt_miss = 0;
92static __int64_t preempt_weird = 0;
fb0f29c4 93static int lwkt_use_spin_port;
40aaf5fc 94static struct objcache *thread_cache;
05220613 95
88ebb169 96#ifdef SMP
e381e77c 97static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
cc9b6223 98static void lwkt_setcpu_remote(void *arg);
88ebb169 99#endif
e381e77c 100
0855a2af
JG
101extern void cpu_heavy_restore(void);
102extern void cpu_lwkt_restore(void);
103extern void cpu_kthread_restore(void);
104extern void cpu_idle_restore(void);
105
fb0f29c4
MD
106/*
107 * We can make all thread ports use the spin backend instead of the thread
108 * backend. This should only be set to debug the spin backend.
109 */
110TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
111
0f7a3396 112#ifdef INVARIANTS
0c52fa62
SG
113SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
114 "Panic if attempting to switch lwkt's while mastering cpusync");
0f7a3396 115#endif
0c52fa62
SG
116SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
117 "Number of switched threads");
9733f757 118SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
0c52fa62 119 "Successful preemption events");
9733f757 120SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
0c52fa62
SG
121 "Failed preemption events");
122SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
123 "Number of preempted threads.");
b12defdc 124static int fairq_enable = 0;
2a418930
MD
125SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
126 &fairq_enable, 0, "Turn on fairq priority accumulators");
85946b6c 127static int fairq_bypass = -1;
b12defdc
MD
128SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
129 &fairq_bypass, 0, "Allow fairq to bypass td on token failure");
130extern int lwkt_sched_debug;
131int lwkt_sched_debug = 0;
132SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
133 &lwkt_sched_debug, 0, "Scheduler debug");
2a418930
MD
134static int lwkt_spin_loops = 10;
135SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
b12defdc
MD
136 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
137static int lwkt_spin_reseq = 0;
138SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW,
139 &lwkt_spin_reseq, 0, "Scheduler resequencer enable");
140static int lwkt_spin_monitor = 0;
141SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW,
142 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait");
d5b2d319
MD
143static int lwkt_spin_fatal = 0; /* disabled */
144SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
145 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
fbc024e4 146static int preempt_enable = 1;
2a418930
MD
147SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
148 &preempt_enable, 0, "Enable preemption");
7b234d8c 149static int lwkt_cache_threads = 0;
765b1ae0
MD
150SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
151 &lwkt_cache_threads, 0, "thread+kstack cache");
fbc024e4 152
2a418930
MD
153static __cachealign int lwkt_cseq_rindex;
154static __cachealign int lwkt_cseq_windex;
05220613 155
4b5f931b
MD
156/*
157 * These helper procedures handle the runq, they can only be called from
158 * within a critical section.
75cdbe6c
MD
159 *
160 * WARNING! Prior to SMP being brought up it is possible to enqueue and
161 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
162 * instead of 'mycpu' when referencing the globaldata structure. Once
163 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 164 */
f1d1c3fa
MD
165static __inline
166void
167_lwkt_dequeue(thread_t td)
168{
169 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 170 struct globaldata *gd = td->td_gd;
4b5f931b 171
f1d1c3fa 172 td->td_flags &= ~TDF_RUNQ;
f9235b6d 173 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
de4d4cb0 174 --gd->gd_tdrunqcount;
f9235b6d 175 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
2a418930 176 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
177 }
178}
179
f9235b6d
MD
180/*
181 * Priority enqueue.
182 *
d992c377
MD
183 * There are a limited number of lwkt threads runnable since user
184 * processes only schedule one at a time per cpu. However, there can
185 * be many user processes in kernel mode exiting from a tsleep() which
186 * become runnable. We do a secondary comparison using td_upri to try
187 * to order these in the situation where several wake up at the same time
188 * to avoid excessive switching.
189 *
190 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and
191 * will ignore user priority. This is to ensure that user threads in
192 * kernel mode get cpu at some point regardless of what the user
193 * scheduler thinks.
f9235b6d 194 */
f1d1c3fa
MD
195static __inline
196void
197_lwkt_enqueue(thread_t td)
198{
f9235b6d
MD
199 thread_t xtd;
200
7f5d7ed7 201 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 202 struct globaldata *gd = td->td_gd;
4b5f931b 203
f1d1c3fa 204 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
205 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
206 if (xtd == NULL) {
85946b6c
MD
207 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
208 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
f9235b6d 209 } else {
d992c377
MD
210 while (xtd &&
211 (xtd->td_pri > td->td_pri ||
212 (xtd->td_pri == td->td_pri &&
213 xtd->td_upri >= td->td_pri))) {
85946b6c 214 xtd = TAILQ_NEXT(xtd, td_threadq);
d992c377 215 }
85946b6c
MD
216 if (xtd)
217 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
218 else
219 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
f9235b6d 220 }
de4d4cb0 221 ++gd->gd_tdrunqcount;
b12defdc
MD
222
223 /*
85946b6c 224 * Request a LWKT reschedule if we are now at the head of the queue.
b12defdc 225 */
85946b6c
MD
226 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
227 need_lwkt_resched();
f1d1c3fa
MD
228 }
229}
8ad65e08 230
40aaf5fc
NT
231static __boolean_t
232_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
233{
234 struct thread *td = (struct thread *)obj;
235
236 td->td_kstack = NULL;
237 td->td_kstack_size = 0;
238 td->td_flags = TDF_ALLOCATED_THREAD;
4643740a 239 td->td_mpflags = 0;
40aaf5fc
NT
240 return (1);
241}
242
243static void
244_lwkt_thread_dtor(void *obj, void *privdata)
245{
246 struct thread *td = (struct thread *)obj;
247
248 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
249 ("_lwkt_thread_dtor: not allocated from objcache"));
250 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
251 td->td_kstack_size > 0,
252 ("_lwkt_thread_dtor: corrupted stack"));
253 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
7b234d8c
MD
254 td->td_kstack = NULL;
255 td->td_flags = 0;
40aaf5fc
NT
256}
257
258/*
259 * Initialize the lwkt s/system.
765b1ae0 260 *
7b234d8c
MD
261 * Nominally cache up to 32 thread + kstack structures. Cache more on
262 * systems with a lot of cpu cores.
40aaf5fc
NT
263 */
264void
265lwkt_init(void)
266{
765b1ae0 267 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
7b234d8c
MD
268 if (lwkt_cache_threads == 0) {
269 lwkt_cache_threads = ncpus * 4;
270 if (lwkt_cache_threads < 32)
271 lwkt_cache_threads = 32;
272 }
765b1ae0
MD
273 thread_cache = objcache_create_mbacked(
274 M_THREAD, sizeof(struct thread),
275 NULL, lwkt_cache_threads,
276 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
277}
278
37af14fe
MD
279/*
280 * Schedule a thread to run. As the current thread we can always safely
281 * schedule ourselves, and a shortcut procedure is provided for that
282 * function.
283 *
284 * (non-blocking, self contained on a per cpu basis)
285 */
286void
287lwkt_schedule_self(thread_t td)
288{
cfaeae2a 289 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 290 crit_enter_quick(td);
f9235b6d
MD
291 KASSERT(td != &td->td_gd->gd_idlethread,
292 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
4643740a
MD
293 KKASSERT(td->td_lwp == NULL ||
294 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
37af14fe 295 _lwkt_enqueue(td);
37af14fe
MD
296 crit_exit_quick(td);
297}
298
299/*
300 * Deschedule a thread.
301 *
302 * (non-blocking, self contained on a per cpu basis)
303 */
304void
305lwkt_deschedule_self(thread_t td)
306{
307 crit_enter_quick(td);
37af14fe
MD
308 _lwkt_dequeue(td);
309 crit_exit_quick(td);
310}
311
8ad65e08
MD
312/*
313 * LWKTs operate on a per-cpu basis
314 *
73e4f7b9 315 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
316 */
317void
318lwkt_gdinit(struct globaldata *gd)
319{
f9235b6d 320 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 321 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
322}
323
7d0bac62
MD
324/*
325 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
326 * or LWKT start address before it can be scheduled. If the target cpu is
327 * -1 the thread will be created on the current cpu.
0cfcada1
MD
328 *
329 * If you intend to create a thread without a process context this function
330 * does everything except load the startup and switcher function.
7d0bac62
MD
331 */
332thread_t
d3d32139 333lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 334{
d2d8515b 335 static int cpu_rotator;
c070746a 336 globaldata_t gd = mycpu;
99df837e 337 void *stack;
7d0bac62 338
c070746a
MD
339 /*
340 * If static thread storage is not supplied allocate a thread. Reuse
341 * a cached free thread if possible. gd_freetd is used to keep an exiting
342 * thread intact through the exit.
343 */
ef0fdad1 344 if (td == NULL) {
cf709dd2
MD
345 crit_enter_gd(gd);
346 if ((td = gd->gd_freetd) != NULL) {
347 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
348 TDF_RUNQ)) == 0);
c070746a 349 gd->gd_freetd = NULL;
cf709dd2 350 } else {
c070746a 351 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
352 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
353 TDF_RUNQ)) == 0);
354 }
355 crit_exit_gd(gd);
40aaf5fc 356 KASSERT((td->td_flags &
2af9d75d
MD
357 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
358 TDF_ALLOCATED_THREAD,
40aaf5fc
NT
359 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
360 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 361 }
c070746a
MD
362
363 /*
364 * Try to reuse cached stack.
365 */
f470d0c8
MD
366 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
367 if (flags & TDF_ALLOCATED_STACK) {
e4846942 368 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
369 stack = NULL;
370 }
371 }
372 if (stack == NULL) {
e40cfbd7 373 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 374 flags |= TDF_ALLOCATED_STACK;
99df837e 375 }
d2d8515b
MD
376 if (cpu < 0) {
377 cpu = ++cpu_rotator;
378 cpu_ccfence();
379 cpu %= ncpus;
380 }
381 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 382 return(td);
7d0bac62
MD
383}
384
385/*
386 * Initialize a preexisting thread structure. This function is used by
387 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
388 *
f8c3996b
MD
389 * All threads start out in a critical section at a priority of
390 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
391 * appropriate. This function may send an IPI message when the
392 * requested cpu is not the current cpu and consequently gd_tdallq may
393 * not be initialized synchronously from the point of view of the originating
394 * cpu.
395 *
396 * NOTE! we have to be careful in regards to creating threads for other cpus
397 * if SMP has not yet been activated.
7d0bac62 398 */
41a01a4d
MD
399#ifdef SMP
400
75cdbe6c
MD
401static void
402lwkt_init_thread_remote(void *arg)
403{
404 thread_t td = arg;
405
52eedfb5
MD
406 /*
407 * Protected by critical section held by IPI dispatch
408 */
75cdbe6c
MD
409 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
410}
411
41a01a4d
MD
412#endif
413
fdce8919
MD
414/*
415 * lwkt core thread structural initialization.
416 *
417 * NOTE: All threads are initialized as mpsafe threads.
418 */
7d0bac62 419void
f470d0c8
MD
420lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
421 struct globaldata *gd)
7d0bac62 422{
37af14fe
MD
423 globaldata_t mygd = mycpu;
424
99df837e
MD
425 bzero(td, sizeof(struct thread));
426 td->td_kstack = stack;
f470d0c8 427 td->td_kstack_size = stksize;
d3d32139 428 td->td_flags = flags;
4643740a 429 td->td_mpflags = 0;
26a0694b 430 td->td_gd = gd;
f9235b6d
MD
431 td->td_pri = TDPRI_KERN_DAEMON;
432 td->td_critcount = 1;
54341a3b 433 td->td_toks_have = NULL;
3b998fa9 434 td->td_toks_stop = &td->td_toks_base;
392cd266 435 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
d636cd48 436 lwkt_initport_spin(&td->td_msgport, td);
fb0f29c4
MD
437 else
438 lwkt_initport_thread(&td->td_msgport, td);
99df837e 439 pmap_init_thread(td);
0f7a3396 440#ifdef SMP
5d21b981
MD
441 /*
442 * Normally initializing a thread for a remote cpu requires sending an
443 * IPI. However, the idlethread is setup before the other cpus are
444 * activated so we have to treat it as a special case. XXX manipulation
445 * of gd_tdallq requires the BGL.
446 */
447 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 448 crit_enter_gd(mygd);
75cdbe6c 449 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 450 crit_exit_gd(mygd);
75cdbe6c 451 } else {
2db3b277 452 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 453 }
0f7a3396 454#else
37af14fe 455 crit_enter_gd(mygd);
0f7a3396 456 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 457 crit_exit_gd(mygd);
0f7a3396 458#endif
8c72e3d5
AH
459
460 dsched_new_thread(td);
73e4f7b9
MD
461}
462
463void
464lwkt_set_comm(thread_t td, const char *ctl, ...)
465{
e2565a42 466 __va_list va;
73e4f7b9 467
e2565a42 468 __va_start(va, ctl);
379210cb 469 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 470 __va_end(va);
5bf48697 471 KTR_LOG(ctxsw_newtd, td, td->td_comm);
7d0bac62
MD
472}
473
eb2adbf5
MD
474/*
475 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
476 * this does not prevent the thread from migrating to another cpu so the
477 * gd_tdallq state is not protected by this.
478 */
99df837e 479void
73e4f7b9 480lwkt_hold(thread_t td)
99df837e 481{
74c9628e 482 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
483}
484
485void
486lwkt_rele(thread_t td)
487{
488 KKASSERT(td->td_refs > 0);
74c9628e 489 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
490}
491
73e4f7b9
MD
492void
493lwkt_free_thread(thread_t td)
494{
74c9628e 495 KKASSERT(td->td_refs == 0);
c17a6852
MD
496 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
497 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
40aaf5fc
NT
498 if (td->td_flags & TDF_ALLOCATED_THREAD) {
499 objcache_put(thread_cache, td);
500 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
501 /* client-allocated struct with internally allocated stack */
502 KASSERT(td->td_kstack && td->td_kstack_size > 0,
503 ("lwkt_free_thread: corrupted stack"));
504 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
505 td->td_kstack = NULL;
506 td->td_kstack_size = 0;
99df837e 507 }
e7c0dbba 508 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
509}
510
511
8ad65e08
MD
512/*
513 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
514 * switch to the idlethread. Switching must occur within a critical
515 * section to avoid races with the scheduling queue.
516 *
517 * We always have full control over our cpu's run queue. Other cpus
518 * that wish to manipulate our queue must use the cpu_*msg() calls to
519 * talk to our cpu, so a critical section is all that is needed and
520 * the result is very, very fast thread switching.
521 *
96728c05
MD
522 * The LWKT scheduler uses a fixed priority model and round-robins at
523 * each priority level. User process scheduling is a totally
524 * different beast and LWKT priorities should not be confused with
525 * user process priorities.
f1d1c3fa 526 *
69d78e99
MD
527 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
528 * is not called by the current thread in the preemption case, only when
529 * the preempting thread blocks (in order to return to the original thread).
cfaeae2a
MD
530 *
531 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
532 * migration and tsleep deschedule the current lwkt thread and call
533 * lwkt_switch(). In particular, the target cpu of the migration fully
534 * expects the thread to become non-runnable and can deadlock against
535 * cpusync operations if we run any IPIs prior to switching the thread out.
536 *
537 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
95858b91 538 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
8ad65e08
MD
539 */
540void
541lwkt_switch(void)
542{
37af14fe
MD
543 globaldata_t gd = mycpu;
544 thread_t td = gd->gd_curthread;
8ad65e08 545 thread_t ntd;
b12defdc 546 int spinning = 0;
8ad65e08 547
da0b0e8b 548 KKASSERT(gd->gd_processing_ipiq == 0);
121f93bc 549 KKASSERT(td->td_flags & TDF_RUNNING);
da0b0e8b 550
46a3f46d 551 /*
27e88a6e
MD
552 * Switching from within a 'fast' (non thread switched) interrupt or IPI
553 * is illegal. However, we may have to do it anyway if we hit a fatal
554 * kernel trap or we have paniced.
555 *
556 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 557 */
27e88a6e
MD
558 if (gd->gd_intr_nesting_level) {
559 int savegdnest;
560 int savegdtrap;
561
5fddbda2 562 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22 563 panic("lwkt_switch: Attempt to switch from a "
5a8df152 564 "fast interrupt, ipi, or hard code section, "
4a28fe22
MD
565 "td %p\n",
566 td);
27e88a6e
MD
567 } else {
568 savegdnest = gd->gd_intr_nesting_level;
569 savegdtrap = gd->gd_trap_nesting_level;
570 gd->gd_intr_nesting_level = 0;
571 gd->gd_trap_nesting_level = 0;
a7422615
MD
572 if ((td->td_flags & TDF_PANICWARN) == 0) {
573 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
574 kprintf("Warning: thread switch from interrupt, IPI, "
575 "or hard code section.\n"
a7422615 576 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 577 print_backtrace(-1);
a7422615 578 }
27e88a6e
MD
579 lwkt_switch();
580 gd->gd_intr_nesting_level = savegdnest;
581 gd->gd_trap_nesting_level = savegdtrap;
582 return;
583 }
96728c05 584 }
ef0fdad1 585
cb973d15 586 /*
85946b6c
MD
587 * Release our current user process designation if we are blocking
588 * or if a user reschedule was requested.
589 *
590 * NOTE: This function is NOT called if we are switching into or
591 * returning from a preemption.
592 *
593 * NOTE: Releasing our current user process designation may cause
594 * it to be assigned to another thread, which in turn will
595 * cause us to block in the usched acquire code when we attempt
596 * to return to userland.
597 *
598 * NOTE: On SMP systems this can be very nasty when heavy token
599 * contention is present so we want to be careful not to
600 * release the designation gratuitously.
cb973d15 601 */
85946b6c
MD
602 if (td->td_release &&
603 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
cb973d15 604 td->td_release(td);
85946b6c 605 }
cb973d15 606
85946b6c
MD
607 /*
608 * Release all tokens
609 */
37af14fe 610 crit_enter_gd(gd);
3b998fa9 611 if (TD_TOKS_HELD(td))
9d265729
MD
612 lwkt_relalltokens(td);
613
614 /*
b02926de
MD
615 * We had better not be holding any spin locks, but don't get into an
616 * endless panic loop.
9d265729 617 */
0846e4ce 618 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL,
d666840a 619 ("lwkt_switch: still holding %d exclusive spinlocks!",
0846e4ce 620 gd->gd_spinlocks));
9d265729 621
8a8d5d85
MD
622
623#ifdef SMP
0f7a3396
MD
624#ifdef INVARIANTS
625 if (td->td_cscount) {
6ea70f76 626 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
627 td);
628 if (panic_on_cscount)
629 panic("switching while mastering cpusync");
630 }
631#endif
8a8d5d85 632#endif
f9235b6d
MD
633
634 /*
635 * If we had preempted another thread on this cpu, resume the preempted
636 * thread. This occurs transparently, whether the preempted thread
637 * was scheduled or not (it may have been preempted after descheduling
638 * itself).
639 *
640 * We have to setup the MP lock for the original thread after backing
641 * out the adjustment that was made to curthread when the original
642 * was preempted.
643 */
99df837e 644 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
645 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
646 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
647
648 /*
b9eb1c19
MD
649 * The interrupt may have woken a thread up, we need to properly
650 * set the reschedule flag if the originally interrupted thread is
651 * at a lower priority.
85946b6c
MD
652 *
653 * The interrupt may not have descheduled.
8ec60c3f 654 */
85946b6c 655 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
8ec60c3f 656 need_lwkt_resched();
f9235b6d
MD
657 goto havethread_preempted;
658 }
659
b12defdc 660 /*
f9235b6d 661 * If we cannot obtain ownership of the tokens we cannot immediately
cfaeae2a
MD
662 * schedule the target thread.
663 *
664 * Reminder: Again, we cannot afford to run any IPIs in this path if
665 * the current thread has been descheduled.
f9235b6d
MD
666 */
667 for (;;) {
b12defdc 668 clear_lwkt_resched();
f9235b6d 669
4b5f931b 670 /*
2a418930 671 * Hotpath - pull the head of the run queue and attempt to schedule
85946b6c 672 * it.
41a01a4d 673 */
7c3dd309
NA
674 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
675
676 if (ntd == NULL) {
677 /*
678 * Runq is empty, switch to idle to allow it to halt.
679 */
680 ntd = &gd->gd_idlethread;
6f207a2c 681#ifdef SMP
7c3dd309
NA
682 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
683 ASSERT_NO_TOKENS_HELD(ntd);
6f207a2c 684#endif
7c3dd309
NA
685 cpu_time.cp_msg[0] = 0;
686 cpu_time.cp_stallpc = 0;
687 goto haveidle;
f9235b6d 688 }
41a01a4d 689
8ec60c3f 690 /*
b12defdc 691 * Hotpath - schedule ntd.
6f207a2c
MD
692 *
693 * NOTE: For UP there is no mplock and lwkt_getalltokens()
694 * always succeeds.
8ec60c3f 695 */
b12defdc
MD
696 if (TD_TOKS_NOT_HELD(ntd) ||
697 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
698 {
f9235b6d 699 goto havethread;
b12defdc 700 }
f9235b6d 701
f9235b6d 702 /*
2a418930
MD
703 * Coldpath (SMP only since tokens always succeed on UP)
704 *
705 * We had some contention on the thread we wanted to schedule.
706 * What we do now is try to find a thread that we can schedule
b12defdc 707 * in its stead.
2a418930 708 *
85946b6c
MD
709 * The coldpath scan does NOT rearrange threads in the run list.
710 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
711 * the next tick whenever the current head is not the current thread.
f9235b6d 712 */
b12defdc 713#ifdef INVARIANTS
85946b6c 714 ++ntd->td_contended;
b12defdc 715#endif
574339e6 716 ++gd->gd_cnt.v_token_colls;
b12defdc 717
85946b6c 718 if (fairq_bypass > 0)
b12defdc
MD
719 goto skip;
720
b12defdc 721 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
d992c377 722#ifdef LWKT_SPLIT_USERPRI
85946b6c
MD
723 /*
724 * Never schedule threads returning to userland or the
725 * user thread scheduler helper thread when higher priority
bb6860b5
NA
726 * threads are present. The runq is sorted by priority
727 * so we can give up traversing it when we find the first
728 * low priority thread.
85946b6c
MD
729 */
730 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
731 ntd = NULL;
732 break;
733 }
d992c377 734#endif
85946b6c
MD
735
736 /*
737 * Try this one.
738 */
b12defdc
MD
739 if (TD_TOKS_NOT_HELD(ntd) ||
740 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
741 goto havethread;
742 }
85946b6c 743#ifdef INVARIANTS
85946b6c 744 ++ntd->td_contended;
b12defdc 745#endif
574339e6 746 ++gd->gd_cnt.v_token_colls;
2a418930
MD
747 }
748
b12defdc 749skip:
2a418930
MD
750 /*
751 * We exhausted the run list, meaning that all runnable threads
b12defdc 752 * are contested.
2a418930
MD
753 */
754 cpu_pause();
755 ntd = &gd->gd_idlethread;
756#ifdef SMP
757 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
758 ASSERT_NO_TOKENS_HELD(ntd);
759 /* contention case, do not clear contention mask */
760#endif
761
762 /*
b12defdc
MD
763 * We are going to have to retry but if the current thread is not
764 * on the runq we instead switch through the idle thread to get away
765 * from the current thread. We have to flag for lwkt reschedule
766 * to prevent the idle thread from halting.
2a418930 767 *
b12defdc
MD
768 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
769 * instruct it to deal with the potential for deadlocks by
770 * ordering the tokens by address.
2a418930 771 */
b12defdc 772 if ((td->td_flags & TDF_RUNQ) == 0) {
85946b6c 773 need_lwkt_resched(); /* prevent hlt */
2a418930 774 goto haveidle;
4b5f931b 775 }
bd52bedf 776#if defined(INVARIANTS) && defined(__amd64__)
d5b2d319
MD
777 if ((read_rflags() & PSL_I) == 0) {
778 cpu_enable_intr();
779 panic("lwkt_switch() called with interrupts disabled");
780 }
781#endif
b12defdc
MD
782
783 /*
784 * Number iterations so far. After a certain point we switch to
785 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
786 */
787 if (spinning < 0x7FFFFFFF)
788 ++spinning;
789
790#ifdef SMP
791 /*
792 * lwkt_getalltokens() failed in sorted token mode, we can use
793 * monitor/mwait in this case.
794 */
795 if (spinning >= lwkt_spin_loops &&
796 (cpu_mi_feature & CPU_MI_MONITOR) &&
797 lwkt_spin_monitor)
798 {
799 cpu_mmw_pause_int(&gd->gd_reqflags,
800 (gd->gd_reqflags | RQF_SPINNING) &
801 ~RQF_IDLECHECK_WK_MASK);
802 }
803#endif
804
805 /*
806 * We already checked that td is still scheduled so this should be
807 * safe.
808 */
809 splz_check();
810
811 /*
812 * This experimental resequencer is used as a fall-back to reduce
813 * hw cache line contention by placing each core's scheduler into a
814 * time-domain-multplexed slot.
815 *
816 * The resequencer is disabled by default. It's functionality has
817 * largely been superceeded by the token algorithm which limits races
818 * to a subset of cores.
819 *
820 * The resequencer algorithm tends to break down when more than
821 * 20 cores are contending. What appears to happen is that new
822 * tokens can be obtained out of address-sorted order by new cores
823 * while existing cores languish in long delays between retries and
824 * wind up being starved-out of the token acquisition.
825 */
826 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
827 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
828 int oseq;
829
830 while ((oseq = lwkt_cseq_rindex) != cseq) {
831 cpu_ccfence();
832#if 1
833 if (cpu_mi_feature & CPU_MI_MONITOR) {
834 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
835 } else {
836#endif
837 cpu_pause();
838 cpu_lfence();
839#if 1
840 }
8b402283 841#endif
0f0466c0 842 }
b12defdc
MD
843 DELAY(1);
844 atomic_add_int(&lwkt_cseq_rindex, 1);
2a418930 845 }
2a418930 846 /* highest level for(;;) loop */
f1d1c3fa 847 }
8a8d5d85 848
2a418930 849havethread:
b12defdc 850 /*
be71787b
MD
851 * Clear gd_idle_repeat when doing a normal switch to a non-idle
852 * thread.
f9235b6d 853 */
9ac1ee6e 854 ntd->td_wmesg = NULL;
b12defdc 855 ++gd->gd_cnt.v_swtch;
be71787b 856 gd->gd_idle_repeat = 0;
2a418930 857
f9235b6d 858havethread_preempted:
f9235b6d
MD
859 /*
860 * If the new target does not need the MP lock and we are holding it,
861 * release the MP lock. If the new target requires the MP lock we have
862 * already acquired it for the target.
8a8d5d85 863 */
2a418930 864 ;
f9235b6d
MD
865haveidle:
866 KASSERT(ntd->td_critcount,
b5d16701
MD
867 ("priority problem in lwkt_switch %d %d",
868 td->td_critcount, ntd->td_critcount));
869
94f6d86e 870 if (td != ntd) {
cc9b6223
MD
871 /*
872 * Execute the actual thread switch operation. This function
873 * returns to the current thread and returns the previous thread
874 * (which may be different from the thread we switched to).
875 *
876 * We are responsible for marking ntd as TDF_RUNNING.
877 */
121f93bc 878 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
94f6d86e 879 ++switch_count;
a1f0fb66 880 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
881 ntd->td_flags |= TDF_RUNNING;
882 lwkt_switch_return(td->td_switch(ntd));
883 /* ntd invalid, td_switch() can return a different thread_t */
94f6d86e 884 }
b12defdc 885
b12defdc 886 /*
54341a3b 887 * catch-all. XXX is this strictly needed?
b12defdc
MD
888 */
889 splz_check();
54341a3b 890
37af14fe
MD
891 /* NOTE: current cpu may have changed after switch */
892 crit_exit_quick(td);
8ad65e08
MD
893}
894
cc9b6223
MD
895/*
896 * Called by assembly in the td_switch (thread restore path) for thread
897 * bootstrap cases which do not 'return' to lwkt_switch().
898 */
899void
900lwkt_switch_return(thread_t otd)
901{
902#ifdef SMP
903 globaldata_t rgd;
904
905 /*
906 * Check if otd was migrating. Now that we are on ntd we can finish
907 * up the migration. This is a bit messy but it is the only place
908 * where td is known to be fully descheduled.
909 *
910 * We can only activate the migration if otd was migrating but not
911 * held on the cpu due to a preemption chain. We still have to
912 * clear TDF_RUNNING on the old thread either way.
913 *
914 * We are responsible for clearing the previously running thread's
915 * TDF_RUNNING.
916 */
917 if ((rgd = otd->td_migrate_gd) != NULL &&
918 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
919 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
920 (TDF_MIGRATING | TDF_RUNNING));
921 otd->td_migrate_gd = NULL;
922 otd->td_flags &= ~TDF_RUNNING;
923 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
924 } else {
925 otd->td_flags &= ~TDF_RUNNING;
926 }
927#else
928 otd->td_flags &= ~TDF_RUNNING;
929#endif
2b07d9aa
MD
930
931 /*
932 * Final exit validations (see lwp_wait()). Note that otd becomes
933 * invalid the *instant* we set TDF_MP_EXITSIG.
934 */
935 while (otd->td_flags & TDF_EXITING) {
936 u_int mpflags;
937
938 mpflags = otd->td_mpflags;
939 cpu_ccfence();
940
941 if (mpflags & TDF_MP_EXITWAIT) {
942 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
943 mpflags | TDF_MP_EXITSIG)) {
944 wakeup(otd);
945 break;
946 }
947 } else {
948 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
949 mpflags | TDF_MP_EXITSIG)) {
950 wakeup(otd);
951 break;
952 }
953 }
954 }
cc9b6223
MD
955}
956
b68b7282 957/*
96728c05 958 * Request that the target thread preempt the current thread. Preemption
54341a3b
MD
959 * can only occur if our only critical section is the one that we were called
960 * with, the relative priority of the target thread is higher, and the target
961 * thread holds no tokens. This also only works if we are not holding any
962 * spinlocks (obviously).
96728c05
MD
963 *
964 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
965 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 966 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
967 * to determine whether preemption is possible (aka usually just the crit
968 * priority of lwkt_schedule() itself).
b68b7282 969 *
54341a3b
MD
970 * Preemption is typically limited to interrupt threads.
971 *
972 * Operation works in a fairly straight-forward manner. The normal
973 * scheduling code is bypassed and we switch directly to the target
974 * thread. When the target thread attempts to block or switch away
975 * code at the base of lwkt_switch() will switch directly back to our
976 * thread. Our thread is able to retain whatever tokens it holds and
977 * if the target needs one of them the target will switch back to us
978 * and reschedule itself normally.
b68b7282
MD
979 */
980void
f9235b6d 981lwkt_preempt(thread_t ntd, int critcount)
b68b7282 982{
46a3f46d 983 struct globaldata *gd = mycpu;
cc9b6223 984 thread_t xtd;
0a3f9b47 985 thread_t td;
2d910aaf 986 int save_gd_intr_nesting_level;
b68b7282 987
26a0694b 988 /*
96728c05
MD
989 * The caller has put us in a critical section. We can only preempt
990 * if the caller of the caller was not in a critical section (basically
f9235b6d 991 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 992 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 993 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
994 *
995 * YYY The target thread must be in a critical section (else it must
996 * inherit our critical section? I dunno yet).
26a0694b 997 */
f9235b6d 998 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 999
b12defdc 1000 td = gd->gd_curthread;
fbc024e4
MD
1001 if (preempt_enable == 0) {
1002 ++preempt_miss;
1003 return;
1004 }
f9235b6d 1005 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
1006 ++preempt_miss;
1007 return;
1008 }
f9235b6d 1009 if (td->td_critcount > critcount) {
96728c05
MD
1010 ++preempt_miss;
1011 return;
1012 }
1013#ifdef SMP
121f93bc
MD
1014 if (td->td_cscount) {
1015 ++preempt_miss;
1016 return;
1017 }
46a3f46d 1018 if (ntd->td_gd != gd) {
96728c05
MD
1019 ++preempt_miss;
1020 return;
1021 }
1022#endif
41a01a4d 1023 /*
77912481
MD
1024 * We don't have to check spinlocks here as they will also bump
1025 * td_critcount.
d3d1cbc8
MD
1026 *
1027 * Do not try to preempt if the target thread is holding any tokens.
1028 * We could try to acquire the tokens but this case is so rare there
1029 * is no need to support it.
41a01a4d 1030 */
0846e4ce 1031 KKASSERT(gd->gd_spinlocks == 0);
77912481 1032
3b998fa9 1033 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8 1034 ++preempt_miss;
d3d1cbc8
MD
1035 return;
1036 }
26a0694b
MD
1037 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1038 ++preempt_weird;
1039 return;
1040 }
1041 if (ntd->td_preempted) {
4b5f931b 1042 ++preempt_hit;
26a0694b 1043 return;
b68b7282 1044 }
da0b0e8b 1045 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 1046
8ec60c3f
MD
1047 /*
1048 * Since we are able to preempt the current thread, there is no need to
1049 * call need_lwkt_resched().
2d910aaf
MD
1050 *
1051 * We must temporarily clear gd_intr_nesting_level around the switch
1052 * since switchouts from the target thread are allowed (they will just
1053 * return to our thread), and since the target thread has its own stack.
cc9b6223
MD
1054 *
1055 * A preemption must switch back to the original thread, assert the
1056 * case.
8ec60c3f 1057 */
26a0694b
MD
1058 ++preempt_hit;
1059 ntd->td_preempted = td;
1060 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1061 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1062 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1063 gd->gd_intr_nesting_level = 0;
121f93bc
MD
1064
1065 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
cc9b6223
MD
1066 ntd->td_flags |= TDF_RUNNING;
1067 xtd = td->td_switch(ntd);
1068 KKASSERT(xtd == ntd);
1069 lwkt_switch_return(xtd);
2d910aaf 1070 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1071
26a0694b
MD
1072 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1073 ntd->td_preempted = NULL;
1074 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1075}
1076
f1d1c3fa 1077/*
faaeffac 1078 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1079 * This will work inside a critical section but not inside a hard code
1080 * section.
ef0fdad1 1081 *
f1d1c3fa
MD
1082 * (self contained on a per cpu basis)
1083 */
1084void
faaeffac 1085splz_check(void)
f1d1c3fa 1086{
7966cb69
MD
1087 globaldata_t gd = mycpu;
1088 thread_t td = gd->gd_curthread;
ef0fdad1 1089
4a28fe22
MD
1090 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1091 gd->gd_intr_nesting_level == 0 &&
1092 td->td_nest_count < 2)
1093 {
f1d1c3fa 1094 splz();
4a28fe22
MD
1095 }
1096}
1097
1098/*
1099 * This version is integrated into crit_exit, reqflags has already
1100 * been tested but td_critcount has not.
1101 *
1102 * We only want to execute the splz() on the 1->0 transition of
1103 * critcount and not in a hard code section or if too deeply nested.
925040f2 1104 *
0846e4ce 1105 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
4a28fe22
MD
1106 */
1107void
1108lwkt_maybe_splz(thread_t td)
1109{
1110 globaldata_t gd = td->td_gd;
1111
1112 if (td->td_critcount == 0 &&
1113 gd->gd_intr_nesting_level == 0 &&
1114 td->td_nest_count < 2)
1115 {
1116 splz();
1117 }
f1d1c3fa
MD
1118}
1119
e6546af9
MD
1120/*
1121 * Drivers which set up processing co-threads can call this function to
1122 * run the co-thread at a higher priority and to allow it to preempt
1123 * normal threads.
1124 */
1125void
1126lwkt_set_interrupt_support_thread(void)
1127{
1128 thread_t td = curthread;
1129
1130 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1131 td->td_flags |= TDF_INTTHREAD;
1132 td->td_preemptable = lwkt_preempt;
1133}
1134
1135
8ad65e08 1136/*
f9235b6d
MD
1137 * This function is used to negotiate a passive release of the current
1138 * process/lwp designation with the user scheduler, allowing the user
1139 * scheduler to schedule another user thread. The related kernel thread
1140 * (curthread) continues running in the released state.
8ad65e08
MD
1141 */
1142void
f9235b6d 1143lwkt_passive_release(struct thread *td)
8ad65e08 1144{
f9235b6d
MD
1145 struct lwp *lp = td->td_lwp;
1146
d992c377 1147#ifdef LWKT_SPLIT_USERPRI
f9235b6d
MD
1148 td->td_release = NULL;
1149 lwkt_setpri_self(TDPRI_KERN_USER);
d992c377
MD
1150#endif
1151
f9235b6d 1152 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1153}
1154
f9235b6d 1155
3824f392 1156/*
d2d8515b
MD
1157 * This implements a LWKT yield, allowing a kernel thread to yield to other
1158 * kernel threads at the same or higher priority. This function can be
1159 * called in a tight loop and will typically only yield once per tick.
f9235b6d 1160 *
d2d8515b
MD
1161 * Most kernel threads run at the same priority in order to allow equal
1162 * sharing.
f9235b6d
MD
1163 *
1164 * (self contained on a per cpu basis)
3824f392
MD
1165 */
1166void
f9235b6d 1167lwkt_yield(void)
3824f392 1168{
f9235b6d
MD
1169 globaldata_t gd = mycpu;
1170 thread_t td = gd->gd_curthread;
3824f392 1171
f9235b6d
MD
1172 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1173 splz();
85946b6c 1174 if (lwkt_resched_wanted()) {
f9235b6d
MD
1175 lwkt_schedule_self(curthread);
1176 lwkt_switch();
f9235b6d 1177 }
3824f392
MD
1178}
1179
40504122
MD
1180/*
1181 * The quick version processes pending interrupts and higher-priority
1182 * LWKT threads but will not round-robin same-priority LWKT threads.
de4d4cb0
MD
1183 *
1184 * When called while attempting to return to userland the only same-pri
1185 * threads are the ones which have already tried to become the current
1186 * user process.
40504122
MD
1187 */
1188void
1189lwkt_yield_quick(void)
1190{
1191 globaldata_t gd = mycpu;
1192 thread_t td = gd->gd_curthread;
1193
1194 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1195 splz();
1196 if (lwkt_resched_wanted()) {
1197 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1198 clear_lwkt_resched();
1199 } else {
1200 lwkt_schedule_self(curthread);
1201 lwkt_switch();
1202 }
1203 }
1204}
1205
3824f392 1206/*
f9235b6d
MD
1207 * This yield is designed for kernel threads with a user context.
1208 *
1209 * The kernel acting on behalf of the user is potentially cpu-bound,
1210 * this function will efficiently allow other threads to run and also
1211 * switch to other processes by releasing.
3824f392
MD
1212 *
1213 * The lwkt_user_yield() function is designed to have very low overhead
1214 * if no yield is determined to be needed.
1215 */
1216void
1217lwkt_user_yield(void)
1218{
f9235b6d
MD
1219 globaldata_t gd = mycpu;
1220 thread_t td = gd->gd_curthread;
1221
1222 /*
1223 * Always run any pending interrupts in case we are in a critical
1224 * section.
1225 */
1226 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1227 splz();
3824f392 1228
3824f392 1229 /*
f9235b6d
MD
1230 * Switch (which forces a release) if another kernel thread needs
1231 * the cpu, if userland wants us to resched, or if our kernel
1232 * quantum has run out.
3824f392 1233 */
f9235b6d 1234 if (lwkt_resched_wanted() ||
85946b6c 1235 user_resched_wanted())
f9235b6d 1236 {
3824f392 1237 lwkt_switch();
3824f392
MD
1238 }
1239
f9235b6d 1240#if 0
3824f392 1241 /*
f9235b6d
MD
1242 * Reacquire the current process if we are released.
1243 *
1244 * XXX not implemented atm. The kernel may be holding locks and such,
1245 * so we want the thread to continue to receive cpu.
3824f392 1246 */
f9235b6d
MD
1247 if (td->td_release == NULL && lp) {
1248 lp->lwp_proc->p_usched->acquire_curproc(lp);
1249 td->td_release = lwkt_passive_release;
1250 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1251 }
f9235b6d 1252#endif
b9eb1c19
MD
1253}
1254
8ad65e08 1255/*
f1d1c3fa
MD
1256 * Generic schedule. Possibly schedule threads belonging to other cpus and
1257 * deal with threads that might be blocked on a wait queue.
1258 *
0a3f9b47
MD
1259 * We have a little helper inline function which does additional work after
1260 * the thread has been enqueued, including dealing with preemption and
1261 * setting need_lwkt_resched() (which prevents the kernel from returning
1262 * to userland until it has processed higher priority threads).
6330a558
MD
1263 *
1264 * It is possible for this routine to be called after a failed _enqueue
1265 * (due to the target thread migrating, sleeping, or otherwise blocked).
1266 * We have to check that the thread is actually on the run queue!
8ad65e08 1267 */
0a3f9b47
MD
1268static __inline
1269void
85946b6c 1270_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
0a3f9b47 1271{
6330a558 1272 if (ntd->td_flags & TDF_RUNQ) {
85946b6c 1273 if (ntd->td_preemptable) {
f9235b6d 1274 ntd->td_preemptable(ntd, ccount); /* YYY +token */
6330a558 1275 }
0a3f9b47
MD
1276 }
1277}
1278
361d01dd 1279static __inline
8ad65e08 1280void
85946b6c 1281_lwkt_schedule(thread_t td)
8ad65e08 1282{
37af14fe
MD
1283 globaldata_t mygd = mycpu;
1284
cf709dd2
MD
1285 KASSERT(td != &td->td_gd->gd_idlethread,
1286 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1287 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1288 crit_enter_gd(mygd);
4643740a
MD
1289 KKASSERT(td->td_lwp == NULL ||
1290 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1291
37af14fe 1292 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1293 _lwkt_enqueue(td);
1294 } else {
f1d1c3fa 1295 /*
7cd8d145
MD
1296 * If we own the thread, there is no race (since we are in a
1297 * critical section). If we do not own the thread there might
1298 * be a race but the target cpu will deal with it.
f1d1c3fa 1299 */
0f7a3396 1300#ifdef SMP
7cd8d145 1301 if (td->td_gd == mygd) {
9d265729 1302 _lwkt_enqueue(td);
85946b6c 1303 _lwkt_schedule_post(mygd, td, 1);
f1d1c3fa 1304 } else {
e381e77c 1305 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1306 }
0f7a3396 1307#else
7cd8d145 1308 _lwkt_enqueue(td);
85946b6c 1309 _lwkt_schedule_post(mygd, td, 1);
0f7a3396 1310#endif
8ad65e08 1311 }
37af14fe 1312 crit_exit_gd(mygd);
8ad65e08
MD
1313}
1314
361d01dd
MD
1315void
1316lwkt_schedule(thread_t td)
1317{
85946b6c 1318 _lwkt_schedule(td);
361d01dd
MD
1319}
1320
1321void
85946b6c 1322lwkt_schedule_noresched(thread_t td) /* XXX not impl */
361d01dd 1323{
85946b6c 1324 _lwkt_schedule(td);
361d01dd
MD
1325}
1326
88ebb169
SW
1327#ifdef SMP
1328
e381e77c
MD
1329/*
1330 * When scheduled remotely if frame != NULL the IPIQ is being
1331 * run via doreti or an interrupt then preemption can be allowed.
1332 *
1333 * To allow preemption we have to drop the critical section so only
1334 * one is present in _lwkt_schedule_post.
1335 */
1336static void
1337lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1338{
1339 thread_t td = curthread;
1340 thread_t ntd = arg;
1341
1342 if (frame && ntd->td_preemptable) {
1343 crit_exit_noyield(td);
85946b6c 1344 _lwkt_schedule(ntd);
e381e77c
MD
1345 crit_enter_quick(td);
1346 } else {
85946b6c 1347 _lwkt_schedule(ntd);
e381e77c
MD
1348 }
1349}
1350
d9eea1a5 1351/*
52eedfb5
MD
1352 * Thread migration using a 'Pull' method. The thread may or may not be
1353 * the current thread. It MUST be descheduled and in a stable state.
1354 * lwkt_giveaway() must be called on the cpu owning the thread.
1355 *
1356 * At any point after lwkt_giveaway() is called, the target cpu may
1357 * 'pull' the thread by calling lwkt_acquire().
1358 *
ae8e83e6
MD
1359 * We have to make sure the thread is not sitting on a per-cpu tsleep
1360 * queue or it will blow up when it moves to another cpu.
1361 *
52eedfb5 1362 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1363 */
52eedfb5
MD
1364void
1365lwkt_giveaway(thread_t td)
1366{
3b4192fb 1367 globaldata_t gd = mycpu;
52eedfb5 1368
3b4192fb
MD
1369 crit_enter_gd(gd);
1370 if (td->td_flags & TDF_TSLEEPQ)
1371 tsleep_remove(td);
1372 KKASSERT(td->td_gd == gd);
1373 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1374 td->td_flags |= TDF_MIGRATING;
1375 crit_exit_gd(gd);
52eedfb5
MD
1376}
1377
a2a5ad0d
MD
1378void
1379lwkt_acquire(thread_t td)
1380{
37af14fe
MD
1381 globaldata_t gd;
1382 globaldata_t mygd;
cc9b6223 1383 int retry = 10000000;
a2a5ad0d 1384
52eedfb5 1385 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1386 gd = td->td_gd;
37af14fe 1387 mygd = mycpu;
52eedfb5 1388 if (gd != mycpu) {
35238fa5 1389 cpu_lfence();
52eedfb5 1390 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1391 crit_enter_gd(mygd);
cfaeae2a 1392 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23
MD
1393 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1394#ifdef SMP
1395 lwkt_process_ipiq();
1396#endif
52eedfb5 1397 cpu_lfence();
cc9b6223
MD
1398 if (--retry == 0) {
1399 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1400 td, td->td_flags);
1401 retry = 10000000;
1402 }
df910c23 1403 }
cfaeae2a 1404 DEBUG_POP_INFO();
562273ea 1405 cpu_mfence();
37af14fe 1406 td->td_gd = mygd;
52eedfb5
MD
1407 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1408 td->td_flags &= ~TDF_MIGRATING;
1409 crit_exit_gd(mygd);
1410 } else {
1411 crit_enter_gd(mygd);
1412 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1413 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1414 crit_exit_gd(mygd);
a2a5ad0d
MD
1415 }
1416}
1417
52eedfb5
MD
1418#endif
1419
f1d1c3fa
MD
1420/*
1421 * Generic deschedule. Descheduling threads other then your own should be
1422 * done only in carefully controlled circumstances. Descheduling is
1423 * asynchronous.
1424 *
1425 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1426 */
1427void
1428lwkt_deschedule(thread_t td)
1429{
f1d1c3fa 1430 crit_enter();
b8a98473 1431#ifdef SMP
f1d1c3fa
MD
1432 if (td == curthread) {
1433 _lwkt_dequeue(td);
1434 } else {
a72187e9 1435 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1436 _lwkt_dequeue(td);
1437 } else {
b8a98473 1438 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1439 }
1440 }
b8a98473
MD
1441#else
1442 _lwkt_dequeue(td);
1443#endif
f1d1c3fa
MD
1444 crit_exit();
1445}
1446
4b5f931b
MD
1447/*
1448 * Set the target thread's priority. This routine does not automatically
1449 * switch to a higher priority thread, LWKT threads are not designed for
1450 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1451 */
1452void
1453lwkt_setpri(thread_t td, int pri)
1454{
f9235b6d
MD
1455 if (td->td_pri != pri) {
1456 KKASSERT(pri >= 0);
1457 crit_enter();
1458 if (td->td_flags & TDF_RUNQ) {
d2d8515b 1459 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1460 _lwkt_dequeue(td);
1461 td->td_pri = pri;
1462 _lwkt_enqueue(td);
1463 } else {
1464 td->td_pri = pri;
1465 }
1466 crit_exit();
26a0694b 1467 }
26a0694b
MD
1468}
1469
03bd0a5e
MD
1470/*
1471 * Set the initial priority for a thread prior to it being scheduled for
1472 * the first time. The thread MUST NOT be scheduled before or during
1473 * this call. The thread may be assigned to a cpu other then the current
1474 * cpu.
1475 *
1476 * Typically used after a thread has been created with TDF_STOPPREQ,
1477 * and before the thread is initially scheduled.
1478 */
1479void
1480lwkt_setpri_initial(thread_t td, int pri)
1481{
1482 KKASSERT(pri >= 0);
1483 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1484 td->td_pri = pri;
03bd0a5e
MD
1485}
1486
26a0694b
MD
1487void
1488lwkt_setpri_self(int pri)
1489{
1490 thread_t td = curthread;
1491
4b5f931b
MD
1492 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1493 crit_enter();
1494 if (td->td_flags & TDF_RUNQ) {
1495 _lwkt_dequeue(td);
f9235b6d 1496 td->td_pri = pri;
4b5f931b
MD
1497 _lwkt_enqueue(td);
1498 } else {
f9235b6d 1499 td->td_pri = pri;
4b5f931b
MD
1500 }
1501 crit_exit();
1502}
1503
f9235b6d 1504/*
85946b6c 1505 * hz tick scheduler clock for LWKT threads
f9235b6d
MD
1506 */
1507void
85946b6c 1508lwkt_schedulerclock(thread_t td)
f9235b6d 1509{
85946b6c
MD
1510 globaldata_t gd = td->td_gd;
1511 thread_t xtd;
2a418930 1512
85946b6c
MD
1513 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1514 /*
1515 * If the current thread is at the head of the runq shift it to the
1516 * end of any equal-priority threads and request a LWKT reschedule
1517 * if it moved.
d992c377
MD
1518 *
1519 * Ignore upri in this situation. There will only be one user thread
1520 * in user mode, all others will be user threads running in kernel
1521 * mode and we have to make sure they get some cpu.
85946b6c
MD
1522 */
1523 xtd = TAILQ_NEXT(td, td_threadq);
1524 if (xtd && xtd->td_pri == td->td_pri) {
1525 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1526 while (xtd && xtd->td_pri == td->td_pri)
1527 xtd = TAILQ_NEXT(xtd, td_threadq);
1528 if (xtd)
1529 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1530 else
1531 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1532 need_lwkt_resched();
f9235b6d 1533 }
85946b6c
MD
1534 } else {
1535 /*
1536 * If we scheduled a thread other than the one at the head of the
1537 * queue always request a reschedule every tick.
1538 */
1539 need_lwkt_resched();
f9235b6d
MD
1540 }
1541}
1542
5d21b981 1543/*
52eedfb5
MD
1544 * Migrate the current thread to the specified cpu.
1545 *
cc9b6223
MD
1546 * This is accomplished by descheduling ourselves from the current cpu
1547 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1548 * 'old' thread wants to migrate after it has been completely switched out
1549 * and will complete the migration.
1550 *
1551 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1552 *
1553 * We must be sure to release our current process designation (if a user
1554 * process) before clearing out any tsleepq we are on because the release
1555 * code may re-add us.
ae8e83e6
MD
1556 *
1557 * We must be sure to remove ourselves from the current cpu's tsleepq
1558 * before potentially moving to another queue. The thread can be on
1559 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1560 */
5d21b981
MD
1561
1562void
1563lwkt_setcpu_self(globaldata_t rgd)
1564{
1565#ifdef SMP
1566 thread_t td = curthread;
1567
1568 if (td->td_gd != rgd) {
1569 crit_enter_quick(td);
cc9b6223 1570
95858b91
MD
1571 if (td->td_release)
1572 td->td_release(td);
ae8e83e6 1573 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1574 tsleep_remove(td);
cc9b6223
MD
1575
1576 /*
1577 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1578 * trying to deschedule ourselves and switch away, then deschedule
1579 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1580 * call lwkt_switch() to complete the operation.
1581 */
5d21b981
MD
1582 td->td_flags |= TDF_MIGRATING;
1583 lwkt_deschedule_self(td);
52eedfb5 1584 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1585 td->td_migrate_gd = rgd;
5d21b981 1586 lwkt_switch();
cc9b6223
MD
1587
1588 /*
1589 * We are now on the target cpu
1590 */
1591 KKASSERT(rgd == mycpu);
52eedfb5 1592 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1593 crit_exit_quick(td);
1594 }
1595#endif
1596}
1597
ecdefdda
MD
1598void
1599lwkt_migratecpu(int cpuid)
1600{
1601#ifdef SMP
1602 globaldata_t rgd;
1603
1604 rgd = globaldata_find(cpuid);
1605 lwkt_setcpu_self(rgd);
1606#endif
1607}
1608
cc9b6223 1609#ifdef SMP
5d21b981
MD
1610/*
1611 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
1612 * do not have to enter another one).
1613 *
1614 * The thread (td) has already been completely descheduled from the
1615 * originating cpu and we can simply assert the case. The thread is
1616 * assigned to the new cpu and enqueued.
5d21b981 1617 *
cc9b6223 1618 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1619 */
1620static void
1621lwkt_setcpu_remote(void *arg)
1622{
1623 thread_t td = arg;
1624 globaldata_t gd = mycpu;
1625
cc9b6223 1626 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1627 td->td_gd = gd;
562273ea 1628 cpu_mfence();
5d21b981 1629 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1630 KKASSERT(td->td_migrate_gd == NULL);
4643740a
MD
1631 KKASSERT(td->td_lwp == NULL ||
1632 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
5d21b981
MD
1633 _lwkt_enqueue(td);
1634}
3d28ff59 1635#endif
5d21b981 1636
553ea3c8 1637struct lwp *
4b5f931b
MD
1638lwkt_preempted_proc(void)
1639{
73e4f7b9 1640 thread_t td = curthread;
4b5f931b
MD
1641 while (td->td_preempted)
1642 td = td->td_preempted;
553ea3c8 1643 return(td->td_lwp);
4b5f931b
MD
1644}
1645
99df837e
MD
1646/*
1647 * Create a kernel process/thread/whatever. It shares it's address space
1648 * with proc0 - ie: kernel only.
1649 *
d2d8515b
MD
1650 * If the cpu is not specified one will be selected. In the future
1651 * specifying a cpu of -1 will enable kernel thread migration between
1652 * cpus.
99df837e
MD
1653 */
1654int
c9e9fb21
MD
1655lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1656 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1657{
73e4f7b9 1658 thread_t td;
e2565a42 1659 __va_list ap;
99df837e 1660
d3d32139 1661 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1662 tdflags);
a2a5ad0d
MD
1663 if (tdp)
1664 *tdp = td;
709799ea 1665 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1666
1667 /*
1668 * Set up arg0 for 'ps' etc
1669 */
e2565a42 1670 __va_start(ap, fmt);
379210cb 1671 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1672 __va_end(ap);
99df837e
MD
1673
1674 /*
1675 * Schedule the thread to run
1676 */
4643740a
MD
1677 if (td->td_flags & TDF_NOSTART)
1678 td->td_flags &= ~TDF_NOSTART;
ef0fdad1 1679 else
4643740a 1680 lwkt_schedule(td);
99df837e
MD
1681 return 0;
1682}
1683
1684/*
1685 * Destroy an LWKT thread. Warning! This function is not called when
1686 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1687 * uses a different reaping mechanism.
1688 */
1689void
1690lwkt_exit(void)
1691{
1692 thread_t td = curthread;
c070746a 1693 thread_t std;
8826f33a 1694 globaldata_t gd;
99df837e 1695
2883d2d8
MD
1696 /*
1697 * Do any cleanup that might block here
1698 */
99df837e 1699 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1700 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1701 caps_exit(td);
2883d2d8
MD
1702 biosched_done(td);
1703 dsched_exit_thread(td);
c070746a
MD
1704
1705 /*
1706 * Get us into a critical section to interlock gd_freetd and loop
1707 * until we can get it freed.
1708 *
1709 * We have to cache the current td in gd_freetd because objcache_put()ing
1710 * it would rip it out from under us while our thread is still active.
2af9d75d
MD
1711 *
1712 * We are the current thread so of course our own TDF_RUNNING bit will
1713 * be set, so unlike the lwp reap code we don't wait for it to clear.
c070746a
MD
1714 */
1715 gd = mycpu;
37af14fe 1716 crit_enter_quick(td);
2af9d75d
MD
1717 for (;;) {
1718 if (td->td_refs) {
1719 tsleep(td, 0, "tdreap", 1);
1720 continue;
1721 }
1722 if ((std = gd->gd_freetd) != NULL) {
1723 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1724 gd->gd_freetd = NULL;
1725 objcache_put(thread_cache, std);
1726 continue;
1727 }
1728 break;
c070746a 1729 }
3b4192fb
MD
1730
1731 /*
1732 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1733 * the last time. We cannot block after this point or we may end
1734 * up with a stale td on the tsleepq.
eb2adbf5
MD
1735 *
1736 * None of this may block, the critical section is the only thing
1737 * protecting tdallq and the only thing preventing new lwkt_hold()
1738 * thread refs now.
3b4192fb
MD
1739 */
1740 if (td->td_flags & TDF_TSLEEPQ)
1741 tsleep_remove(td);
37af14fe 1742 lwkt_deschedule_self(td);
e56e4dea 1743 lwkt_remove_tdallq(td);
74c9628e 1744 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1745
1746 /*
1747 * Final cleanup
1748 */
1749 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1750 if (td->td_flags & TDF_ALLOCATED_THREAD)
1751 gd->gd_freetd = td;
99df837e
MD
1752 cpu_thread_exit();
1753}
1754
e56e4dea
MD
1755void
1756lwkt_remove_tdallq(thread_t td)
1757{
1758 KKASSERT(td->td_gd == mycpu);
1759 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1760}
1761
9cf43f91
MD
1762/*
1763 * Code reduction and branch prediction improvements. Call/return
1764 * overhead on modern cpus often degenerates into 0 cycles due to
1765 * the cpu's branch prediction hardware and return pc cache. We
1766 * can take advantage of this by not inlining medium-complexity
1767 * functions and we can also reduce the branch prediction impact
1768 * by collapsing perfectly predictable branches into a single
1769 * procedure instead of duplicating it.
1770 *
1771 * Is any of this noticeable? Probably not, so I'll take the
1772 * smaller code size.
1773 */
1774void
b6468f56 1775crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1776{
b6468f56 1777 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1778}
1779
2d93b37a
MD
1780void
1781crit_panic(void)
1782{
1783 thread_t td = curthread;
850634cc 1784 int lcrit = td->td_critcount;
2d93b37a 1785
850634cc
AH
1786 td->td_critcount = 0;
1787 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1788 /* NOT REACHED */
2d93b37a
MD
1789}
1790
d165e668
MD
1791#ifdef SMP
1792
bd8015ca
MD
1793/*
1794 * Called from debugger/panic on cpus which have been stopped. We must still
1795 * process the IPIQ while stopped, even if we were stopped while in a critical
1796 * section (XXX).
1797 *
1798 * If we are dumping also try to process any pending interrupts. This may
1799 * or may not work depending on the state of the cpu at the point it was
1800 * stopped.
1801 */
1802void
1803lwkt_smp_stopped(void)
1804{
1805 globaldata_t gd = mycpu;
1806
1807 crit_enter_gd(gd);
1808 if (dumping) {
1809 lwkt_process_ipiq();
1810 splz();
1811 } else {
1812 lwkt_process_ipiq();
1813 }
1814 crit_exit_gd(gd);
1815}
1816
d165e668 1817#endif