dc(1): Sync with OpenBSD
[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>
9d265729
MD
55
56#include <sys/thread2.h>
57#include <sys/spinlock2.h>
684a93c4 58#include <sys/mplock2.h>
f1d1c3fa 59
8c72e3d5
AH
60#include <sys/dsched.h>
61
7d0bac62
MD
62#include <vm/vm.h>
63#include <vm/vm_param.h>
64#include <vm/vm_kern.h>
65#include <vm/vm_object.h>
66#include <vm/vm_page.h>
67#include <vm/vm_map.h>
68#include <vm/vm_pager.h>
69#include <vm/vm_extern.h>
7d0bac62 70
99df837e 71#include <machine/stdarg.h>
96728c05 72#include <machine/smp.h>
99df837e 73
8cee56f4
MD
74#ifdef _KERNEL_VIRTUAL
75#include <pthread.h>
76#endif
77
d850923c
AE
78#if !defined(KTR_CTXSW)
79#define KTR_CTXSW KTR_ALL
80#endif
81KTR_INFO_MASTER(ctxsw);
5bf48697
AE
82KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td);
83KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td);
84KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm);
85KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td);
1541028a 86
40aaf5fc
NT
87static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
88
0f7a3396
MD
89#ifdef INVARIANTS
90static int panic_on_cscount = 0;
91#endif
05220613
MD
92static __int64_t switch_count = 0;
93static __int64_t preempt_hit = 0;
94static __int64_t preempt_miss = 0;
95static __int64_t preempt_weird = 0;
fb0f29c4 96static int lwkt_use_spin_port;
40aaf5fc 97static struct objcache *thread_cache;
a46b4a23 98int cpu_mwait_spin = 0;
05220613 99
e381e77c 100static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
cc9b6223 101static void lwkt_setcpu_remote(void *arg);
e381e77c 102
0855a2af
JG
103extern void cpu_heavy_restore(void);
104extern void cpu_lwkt_restore(void);
105extern void cpu_kthread_restore(void);
106extern void cpu_idle_restore(void);
107
fb0f29c4
MD
108/*
109 * We can make all thread ports use the spin backend instead of the thread
110 * backend. This should only be set to debug the spin backend.
111 */
112TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
113
0f7a3396 114#ifdef INVARIANTS
0c52fa62
SG
115SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
116 "Panic if attempting to switch lwkt's while mastering cpusync");
0f7a3396 117#endif
a46b4a23
MD
118SYSCTL_INT(_hw, OID_AUTO, cpu_mwait_spin, CTLFLAG_RW, &cpu_mwait_spin, 0,
119 "monitor/mwait target state");
0c52fa62
SG
120SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
121 "Number of switched threads");
9733f757 122SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
0c52fa62 123 "Successful preemption events");
9733f757 124SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
0c52fa62
SG
125 "Failed preemption events");
126SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
127 "Number of preempted threads.");
b12defdc 128static int fairq_enable = 0;
2a418930
MD
129SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
130 &fairq_enable, 0, "Turn on fairq priority accumulators");
85946b6c 131static int fairq_bypass = -1;
b12defdc
MD
132SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
133 &fairq_bypass, 0, "Allow fairq to bypass td on token failure");
134extern int lwkt_sched_debug;
135int lwkt_sched_debug = 0;
136SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
137 &lwkt_sched_debug, 0, "Scheduler debug");
2a418930
MD
138static int lwkt_spin_loops = 10;
139SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
b12defdc
MD
140 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
141static int lwkt_spin_reseq = 0;
142SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW,
143 &lwkt_spin_reseq, 0, "Scheduler resequencer enable");
144static int lwkt_spin_monitor = 0;
145SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW,
146 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait");
d5b2d319
MD
147static int lwkt_spin_fatal = 0; /* disabled */
148SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
149 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
fbc024e4 150static int preempt_enable = 1;
2a418930
MD
151SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
152 &preempt_enable, 0, "Enable preemption");
7b234d8c 153static int lwkt_cache_threads = 0;
765b1ae0
MD
154SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
155 &lwkt_cache_threads, 0, "thread+kstack cache");
fbc024e4 156
8cee56f4 157#ifndef _KERNEL_VIRTUAL
2a418930
MD
158static __cachealign int lwkt_cseq_rindex;
159static __cachealign int lwkt_cseq_windex;
8cee56f4 160#endif
05220613 161
4b5f931b
MD
162/*
163 * These helper procedures handle the runq, they can only be called from
164 * within a critical section.
75cdbe6c
MD
165 *
166 * WARNING! Prior to SMP being brought up it is possible to enqueue and
167 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
168 * instead of 'mycpu' when referencing the globaldata structure. Once
169 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 170 */
f1d1c3fa
MD
171static __inline
172void
173_lwkt_dequeue(thread_t td)
174{
175 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 176 struct globaldata *gd = td->td_gd;
4b5f931b 177
f1d1c3fa 178 td->td_flags &= ~TDF_RUNQ;
f9235b6d 179 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
de4d4cb0 180 --gd->gd_tdrunqcount;
f9235b6d 181 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
2a418930 182 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
183 }
184}
185
f9235b6d
MD
186/*
187 * Priority enqueue.
188 *
d992c377
MD
189 * There are a limited number of lwkt threads runnable since user
190 * processes only schedule one at a time per cpu. However, there can
191 * be many user processes in kernel mode exiting from a tsleep() which
e3e6be1f 192 * become runnable.
d992c377
MD
193 *
194 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and
195 * will ignore user priority. This is to ensure that user threads in
196 * kernel mode get cpu at some point regardless of what the user
197 * scheduler thinks.
f9235b6d 198 */
f1d1c3fa
MD
199static __inline
200void
201_lwkt_enqueue(thread_t td)
202{
f9235b6d
MD
203 thread_t xtd;
204
7f5d7ed7 205 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 206 struct globaldata *gd = td->td_gd;
4b5f931b 207
f1d1c3fa 208 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
209 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
210 if (xtd == NULL) {
85946b6c
MD
211 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
212 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
f9235b6d 213 } else {
e3e6be1f
MD
214 /*
215 * NOTE: td_upri - higher numbers more desireable, same sense
216 * as td_pri (typically reversed from lwp_upri).
217 *
218 * In the equal priority case we want the best selection
219 * at the beginning so the less desireable selections know
220 * that they have to setrunqueue/go-to-another-cpu, even
221 * though it means switching back to the 'best' selection.
222 * This also avoids degenerate situations when many threads
223 * are runnable or waking up at the same time.
224 *
225 * If upri matches exactly place at end/round-robin.
226 */
d992c377 227 while (xtd &&
e3e6be1f 228 (xtd->td_pri >= td->td_pri ||
d992c377 229 (xtd->td_pri == td->td_pri &&
e3e6be1f 230 xtd->td_upri >= td->td_upri))) {
85946b6c 231 xtd = TAILQ_NEXT(xtd, td_threadq);
d992c377 232 }
85946b6c
MD
233 if (xtd)
234 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
235 else
236 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
f9235b6d 237 }
de4d4cb0 238 ++gd->gd_tdrunqcount;
b12defdc
MD
239
240 /*
85946b6c 241 * Request a LWKT reschedule if we are now at the head of the queue.
b12defdc 242 */
85946b6c
MD
243 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
244 need_lwkt_resched();
f1d1c3fa
MD
245 }
246}
8ad65e08 247
40aaf5fc
NT
248static __boolean_t
249_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
250{
251 struct thread *td = (struct thread *)obj;
252
253 td->td_kstack = NULL;
254 td->td_kstack_size = 0;
255 td->td_flags = TDF_ALLOCATED_THREAD;
4643740a 256 td->td_mpflags = 0;
40aaf5fc
NT
257 return (1);
258}
259
260static void
261_lwkt_thread_dtor(void *obj, void *privdata)
262{
263 struct thread *td = (struct thread *)obj;
264
265 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
266 ("_lwkt_thread_dtor: not allocated from objcache"));
267 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
268 td->td_kstack_size > 0,
269 ("_lwkt_thread_dtor: corrupted stack"));
270 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
7b234d8c
MD
271 td->td_kstack = NULL;
272 td->td_flags = 0;
40aaf5fc
NT
273}
274
275/*
276 * Initialize the lwkt s/system.
765b1ae0 277 *
7b234d8c
MD
278 * Nominally cache up to 32 thread + kstack structures. Cache more on
279 * systems with a lot of cpu cores.
40aaf5fc
NT
280 */
281void
282lwkt_init(void)
283{
765b1ae0 284 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
7b234d8c
MD
285 if (lwkt_cache_threads == 0) {
286 lwkt_cache_threads = ncpus * 4;
287 if (lwkt_cache_threads < 32)
288 lwkt_cache_threads = 32;
289 }
765b1ae0
MD
290 thread_cache = objcache_create_mbacked(
291 M_THREAD, sizeof(struct thread),
2fce2579 292 0, lwkt_cache_threads,
765b1ae0 293 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
294}
295
37af14fe
MD
296/*
297 * Schedule a thread to run. As the current thread we can always safely
298 * schedule ourselves, and a shortcut procedure is provided for that
299 * function.
300 *
301 * (non-blocking, self contained on a per cpu basis)
302 */
303void
304lwkt_schedule_self(thread_t td)
305{
cfaeae2a 306 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 307 crit_enter_quick(td);
f9235b6d
MD
308 KASSERT(td != &td->td_gd->gd_idlethread,
309 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
4643740a
MD
310 KKASSERT(td->td_lwp == NULL ||
311 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
37af14fe 312 _lwkt_enqueue(td);
37af14fe
MD
313 crit_exit_quick(td);
314}
315
316/*
317 * Deschedule a thread.
318 *
319 * (non-blocking, self contained on a per cpu basis)
320 */
321void
322lwkt_deschedule_self(thread_t td)
323{
324 crit_enter_quick(td);
37af14fe
MD
325 _lwkt_dequeue(td);
326 crit_exit_quick(td);
327}
328
8ad65e08
MD
329/*
330 * LWKTs operate on a per-cpu basis
331 *
73e4f7b9 332 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
333 */
334void
335lwkt_gdinit(struct globaldata *gd)
336{
f9235b6d 337 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 338 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
339}
340
7d0bac62
MD
341/*
342 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
343 * or LWKT start address before it can be scheduled. If the target cpu is
344 * -1 the thread will be created on the current cpu.
0cfcada1
MD
345 *
346 * If you intend to create a thread without a process context this function
347 * does everything except load the startup and switcher function.
7d0bac62
MD
348 */
349thread_t
d3d32139 350lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 351{
d2d8515b 352 static int cpu_rotator;
c070746a 353 globaldata_t gd = mycpu;
99df837e 354 void *stack;
7d0bac62 355
c070746a
MD
356 /*
357 * If static thread storage is not supplied allocate a thread. Reuse
358 * a cached free thread if possible. gd_freetd is used to keep an exiting
359 * thread intact through the exit.
360 */
ef0fdad1 361 if (td == NULL) {
cf709dd2
MD
362 crit_enter_gd(gd);
363 if ((td = gd->gd_freetd) != NULL) {
364 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
365 TDF_RUNQ)) == 0);
c070746a 366 gd->gd_freetd = NULL;
cf709dd2 367 } else {
c070746a 368 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
369 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
370 TDF_RUNQ)) == 0);
371 }
372 crit_exit_gd(gd);
40aaf5fc 373 KASSERT((td->td_flags &
2af9d75d
MD
374 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
375 TDF_ALLOCATED_THREAD,
40aaf5fc
NT
376 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
377 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 378 }
c070746a
MD
379
380 /*
381 * Try to reuse cached stack.
382 */
f470d0c8
MD
383 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
384 if (flags & TDF_ALLOCATED_STACK) {
e4846942 385 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
386 stack = NULL;
387 }
388 }
389 if (stack == NULL) {
e40cfbd7 390 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 391 flags |= TDF_ALLOCATED_STACK;
99df837e 392 }
d2d8515b
MD
393 if (cpu < 0) {
394 cpu = ++cpu_rotator;
395 cpu_ccfence();
396 cpu %= ncpus;
397 }
398 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 399 return(td);
7d0bac62
MD
400}
401
402/*
403 * Initialize a preexisting thread structure. This function is used by
404 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
405 *
f8c3996b
MD
406 * All threads start out in a critical section at a priority of
407 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
408 * appropriate. This function may send an IPI message when the
409 * requested cpu is not the current cpu and consequently gd_tdallq may
410 * not be initialized synchronously from the point of view of the originating
411 * cpu.
412 *
413 * NOTE! we have to be careful in regards to creating threads for other cpus
414 * if SMP has not yet been activated.
7d0bac62 415 */
75cdbe6c
MD
416static void
417lwkt_init_thread_remote(void *arg)
418{
419 thread_t td = arg;
420
52eedfb5
MD
421 /*
422 * Protected by critical section held by IPI dispatch
423 */
75cdbe6c
MD
424 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
425}
426
fdce8919
MD
427/*
428 * lwkt core thread structural initialization.
429 *
430 * NOTE: All threads are initialized as mpsafe threads.
431 */
7d0bac62 432void
f470d0c8
MD
433lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
434 struct globaldata *gd)
7d0bac62 435{
37af14fe
MD
436 globaldata_t mygd = mycpu;
437
99df837e
MD
438 bzero(td, sizeof(struct thread));
439 td->td_kstack = stack;
f470d0c8 440 td->td_kstack_size = stksize;
d3d32139 441 td->td_flags = flags;
4643740a 442 td->td_mpflags = 0;
f256b6c0 443 td->td_type = TD_TYPE_GENERIC;
26a0694b 444 td->td_gd = gd;
f9235b6d
MD
445 td->td_pri = TDPRI_KERN_DAEMON;
446 td->td_critcount = 1;
54341a3b 447 td->td_toks_have = NULL;
3b998fa9 448 td->td_toks_stop = &td->td_toks_base;
392cd266 449 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
d636cd48 450 lwkt_initport_spin(&td->td_msgport, td);
fb0f29c4
MD
451 else
452 lwkt_initport_thread(&td->td_msgport, td);
99df837e 453 pmap_init_thread(td);
5d21b981
MD
454 /*
455 * Normally initializing a thread for a remote cpu requires sending an
456 * IPI. However, the idlethread is setup before the other cpus are
457 * activated so we have to treat it as a special case. XXX manipulation
458 * of gd_tdallq requires the BGL.
459 */
460 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 461 crit_enter_gd(mygd);
75cdbe6c 462 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 463 crit_exit_gd(mygd);
75cdbe6c 464 } else {
2db3b277 465 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 466 }
8c72e3d5 467 dsched_new_thread(td);
73e4f7b9
MD
468}
469
470void
471lwkt_set_comm(thread_t td, const char *ctl, ...)
472{
e2565a42 473 __va_list va;
73e4f7b9 474
e2565a42 475 __va_start(va, ctl);
379210cb 476 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 477 __va_end(va);
5bf48697 478 KTR_LOG(ctxsw_newtd, td, td->td_comm);
7d0bac62
MD
479}
480
eb2adbf5
MD
481/*
482 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
483 * this does not prevent the thread from migrating to another cpu so the
484 * gd_tdallq state is not protected by this.
485 */
99df837e 486void
73e4f7b9 487lwkt_hold(thread_t td)
99df837e 488{
74c9628e 489 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
490}
491
492void
493lwkt_rele(thread_t td)
494{
495 KKASSERT(td->td_refs > 0);
74c9628e 496 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
497}
498
73e4f7b9
MD
499void
500lwkt_free_thread(thread_t td)
501{
74c9628e 502 KKASSERT(td->td_refs == 0);
c17a6852
MD
503 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
504 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
40aaf5fc
NT
505 if (td->td_flags & TDF_ALLOCATED_THREAD) {
506 objcache_put(thread_cache, td);
507 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
508 /* client-allocated struct with internally allocated stack */
509 KASSERT(td->td_kstack && td->td_kstack_size > 0,
510 ("lwkt_free_thread: corrupted stack"));
511 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
512 td->td_kstack = NULL;
513 td->td_kstack_size = 0;
99df837e 514 }
e7c0dbba 515 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
516}
517
518
8ad65e08
MD
519/*
520 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
521 * switch to the idlethread. Switching must occur within a critical
522 * section to avoid races with the scheduling queue.
523 *
524 * We always have full control over our cpu's run queue. Other cpus
525 * that wish to manipulate our queue must use the cpu_*msg() calls to
526 * talk to our cpu, so a critical section is all that is needed and
527 * the result is very, very fast thread switching.
528 *
96728c05
MD
529 * The LWKT scheduler uses a fixed priority model and round-robins at
530 * each priority level. User process scheduling is a totally
531 * different beast and LWKT priorities should not be confused with
532 * user process priorities.
f1d1c3fa 533 *
69d78e99
MD
534 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
535 * is not called by the current thread in the preemption case, only when
536 * the preempting thread blocks (in order to return to the original thread).
cfaeae2a
MD
537 *
538 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
539 * migration and tsleep deschedule the current lwkt thread and call
540 * lwkt_switch(). In particular, the target cpu of the migration fully
541 * expects the thread to become non-runnable and can deadlock against
542 * cpusync operations if we run any IPIs prior to switching the thread out.
543 *
544 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
95858b91 545 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
8ad65e08
MD
546 */
547void
548lwkt_switch(void)
549{
37af14fe
MD
550 globaldata_t gd = mycpu;
551 thread_t td = gd->gd_curthread;
8ad65e08 552 thread_t ntd;
b12defdc 553 int spinning = 0;
8ad65e08 554
da0b0e8b 555 KKASSERT(gd->gd_processing_ipiq == 0);
121f93bc 556 KKASSERT(td->td_flags & TDF_RUNNING);
da0b0e8b 557
46a3f46d 558 /*
27e88a6e
MD
559 * Switching from within a 'fast' (non thread switched) interrupt or IPI
560 * is illegal. However, we may have to do it anyway if we hit a fatal
561 * kernel trap or we have paniced.
562 *
563 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 564 */
27e88a6e
MD
565 if (gd->gd_intr_nesting_level) {
566 int savegdnest;
567 int savegdtrap;
568
5fddbda2 569 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22 570 panic("lwkt_switch: Attempt to switch from a "
5a8df152 571 "fast interrupt, ipi, or hard code section, "
4a28fe22
MD
572 "td %p\n",
573 td);
27e88a6e
MD
574 } else {
575 savegdnest = gd->gd_intr_nesting_level;
576 savegdtrap = gd->gd_trap_nesting_level;
577 gd->gd_intr_nesting_level = 0;
578 gd->gd_trap_nesting_level = 0;
a7422615
MD
579 if ((td->td_flags & TDF_PANICWARN) == 0) {
580 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
581 kprintf("Warning: thread switch from interrupt, IPI, "
582 "or hard code section.\n"
a7422615 583 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 584 print_backtrace(-1);
a7422615 585 }
27e88a6e
MD
586 lwkt_switch();
587 gd->gd_intr_nesting_level = savegdnest;
588 gd->gd_trap_nesting_level = savegdtrap;
589 return;
590 }
96728c05 591 }
ef0fdad1 592
cb973d15 593 /*
85946b6c
MD
594 * Release our current user process designation if we are blocking
595 * or if a user reschedule was requested.
596 *
597 * NOTE: This function is NOT called if we are switching into or
598 * returning from a preemption.
599 *
600 * NOTE: Releasing our current user process designation may cause
601 * it to be assigned to another thread, which in turn will
602 * cause us to block in the usched acquire code when we attempt
603 * to return to userland.
604 *
605 * NOTE: On SMP systems this can be very nasty when heavy token
606 * contention is present so we want to be careful not to
607 * release the designation gratuitously.
cb973d15 608 */
85946b6c
MD
609 if (td->td_release &&
610 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
cb973d15 611 td->td_release(td);
85946b6c 612 }
cb973d15 613
85946b6c
MD
614 /*
615 * Release all tokens
616 */
37af14fe 617 crit_enter_gd(gd);
3b998fa9 618 if (TD_TOKS_HELD(td))
9d265729
MD
619 lwkt_relalltokens(td);
620
621 /*
b02926de
MD
622 * We had better not be holding any spin locks, but don't get into an
623 * endless panic loop.
9d265729 624 */
0846e4ce 625 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL,
d666840a 626 ("lwkt_switch: still holding %d exclusive spinlocks!",
0846e4ce 627 gd->gd_spinlocks));
9d265729 628
8a8d5d85 629
0f7a3396
MD
630#ifdef INVARIANTS
631 if (td->td_cscount) {
6ea70f76 632 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
633 td);
634 if (panic_on_cscount)
635 panic("switching while mastering cpusync");
636 }
8a8d5d85 637#endif
f9235b6d
MD
638
639 /*
640 * If we had preempted another thread on this cpu, resume the preempted
641 * thread. This occurs transparently, whether the preempted thread
642 * was scheduled or not (it may have been preempted after descheduling
643 * itself).
644 *
645 * We have to setup the MP lock for the original thread after backing
646 * out the adjustment that was made to curthread when the original
647 * was preempted.
648 */
99df837e 649 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
650 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
651 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
652
653 /*
b9eb1c19
MD
654 * The interrupt may have woken a thread up, we need to properly
655 * set the reschedule flag if the originally interrupted thread is
656 * at a lower priority.
85946b6c
MD
657 *
658 * The interrupt may not have descheduled.
8ec60c3f 659 */
85946b6c 660 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
8ec60c3f 661 need_lwkt_resched();
f9235b6d
MD
662 goto havethread_preempted;
663 }
664
b12defdc 665 /*
f9235b6d 666 * If we cannot obtain ownership of the tokens we cannot immediately
cfaeae2a
MD
667 * schedule the target thread.
668 *
669 * Reminder: Again, we cannot afford to run any IPIs in this path if
670 * the current thread has been descheduled.
f9235b6d
MD
671 */
672 for (;;) {
b12defdc 673 clear_lwkt_resched();
f9235b6d 674
4b5f931b 675 /*
2a418930 676 * Hotpath - pull the head of the run queue and attempt to schedule
85946b6c 677 * it.
41a01a4d 678 */
7c3dd309
NA
679 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
680
681 if (ntd == NULL) {
682 /*
683 * Runq is empty, switch to idle to allow it to halt.
684 */
685 ntd = &gd->gd_idlethread;
7c3dd309
NA
686 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
687 ASSERT_NO_TOKENS_HELD(ntd);
7c3dd309
NA
688 cpu_time.cp_msg[0] = 0;
689 cpu_time.cp_stallpc = 0;
690 goto haveidle;
f9235b6d 691 }
41a01a4d 692
8ec60c3f 693 /*
b12defdc 694 * Hotpath - schedule ntd.
6f207a2c
MD
695 *
696 * NOTE: For UP there is no mplock and lwkt_getalltokens()
697 * always succeeds.
8ec60c3f 698 */
b12defdc
MD
699 if (TD_TOKS_NOT_HELD(ntd) ||
700 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
701 {
f9235b6d 702 goto havethread;
b12defdc 703 }
f9235b6d 704
f9235b6d 705 /*
2a418930
MD
706 * Coldpath (SMP only since tokens always succeed on UP)
707 *
708 * We had some contention on the thread we wanted to schedule.
709 * What we do now is try to find a thread that we can schedule
b12defdc 710 * in its stead.
2a418930 711 *
85946b6c
MD
712 * The coldpath scan does NOT rearrange threads in the run list.
713 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
714 * the next tick whenever the current head is not the current thread.
f9235b6d 715 */
b12defdc 716#ifdef INVARIANTS
85946b6c 717 ++ntd->td_contended;
b12defdc 718#endif
574339e6 719 ++gd->gd_cnt.v_token_colls;
b12defdc 720
85946b6c 721 if (fairq_bypass > 0)
b12defdc
MD
722 goto skip;
723
b12defdc 724 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
e3e6be1f 725#ifndef NO_LWKT_SPLIT_USERPRI
85946b6c
MD
726 /*
727 * Never schedule threads returning to userland or the
728 * user thread scheduler helper thread when higher priority
bb6860b5
NA
729 * threads are present. The runq is sorted by priority
730 * so we can give up traversing it when we find the first
731 * low priority thread.
85946b6c
MD
732 */
733 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
734 ntd = NULL;
735 break;
736 }
d992c377 737#endif
85946b6c
MD
738
739 /*
740 * Try this one.
741 */
b12defdc
MD
742 if (TD_TOKS_NOT_HELD(ntd) ||
743 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
744 goto havethread;
745 }
85946b6c 746#ifdef INVARIANTS
85946b6c 747 ++ntd->td_contended;
b12defdc 748#endif
574339e6 749 ++gd->gd_cnt.v_token_colls;
2a418930
MD
750 }
751
b12defdc 752skip:
2a418930
MD
753 /*
754 * We exhausted the run list, meaning that all runnable threads
b12defdc 755 * are contested.
2a418930
MD
756 */
757 cpu_pause();
8cee56f4
MD
758#ifdef _KERNEL_VIRTUAL
759 pthread_yield();
760#endif
2a418930 761 ntd = &gd->gd_idlethread;
2a418930
MD
762 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
763 ASSERT_NO_TOKENS_HELD(ntd);
764 /* contention case, do not clear contention mask */
2a418930
MD
765
766 /*
b12defdc
MD
767 * We are going to have to retry but if the current thread is not
768 * on the runq we instead switch through the idle thread to get away
769 * from the current thread. We have to flag for lwkt reschedule
770 * to prevent the idle thread from halting.
2a418930 771 *
b12defdc
MD
772 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
773 * instruct it to deal with the potential for deadlocks by
774 * ordering the tokens by address.
2a418930 775 */
b12defdc 776 if ((td->td_flags & TDF_RUNQ) == 0) {
85946b6c 777 need_lwkt_resched(); /* prevent hlt */
2a418930 778 goto haveidle;
4b5f931b 779 }
dbefba87 780#if defined(INVARIANTS) && defined(__x86_64__)
d5b2d319
MD
781 if ((read_rflags() & PSL_I) == 0) {
782 cpu_enable_intr();
783 panic("lwkt_switch() called with interrupts disabled");
784 }
785#endif
b12defdc
MD
786
787 /*
788 * Number iterations so far. After a certain point we switch to
789 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
790 */
791 if (spinning < 0x7FFFFFFF)
792 ++spinning;
793
8cee56f4 794#ifndef _KERNEL_VIRTUAL
b12defdc
MD
795 /*
796 * lwkt_getalltokens() failed in sorted token mode, we can use
797 * monitor/mwait in this case.
798 */
799 if (spinning >= lwkt_spin_loops &&
800 (cpu_mi_feature & CPU_MI_MONITOR) &&
801 lwkt_spin_monitor)
802 {
803 cpu_mmw_pause_int(&gd->gd_reqflags,
804 (gd->gd_reqflags | RQF_SPINNING) &
a46b4a23
MD
805 ~RQF_IDLECHECK_WK_MASK,
806 cpu_mwait_spin);
b12defdc 807 }
8cee56f4 808#endif
b12defdc
MD
809
810 /*
811 * We already checked that td is still scheduled so this should be
812 * safe.
813 */
814 splz_check();
815
8cee56f4 816#ifndef _KERNEL_VIRTUAL
b12defdc
MD
817 /*
818 * This experimental resequencer is used as a fall-back to reduce
819 * hw cache line contention by placing each core's scheduler into a
820 * time-domain-multplexed slot.
821 *
822 * The resequencer is disabled by default. It's functionality has
823 * largely been superceeded by the token algorithm which limits races
824 * to a subset of cores.
825 *
826 * The resequencer algorithm tends to break down when more than
827 * 20 cores are contending. What appears to happen is that new
828 * tokens can be obtained out of address-sorted order by new cores
829 * while existing cores languish in long delays between retries and
830 * wind up being starved-out of the token acquisition.
831 */
832 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
833 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
834 int oseq;
835
836 while ((oseq = lwkt_cseq_rindex) != cseq) {
837 cpu_ccfence();
838#if 1
839 if (cpu_mi_feature & CPU_MI_MONITOR) {
a46b4a23 840 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq, cpu_mwait_spin);
b12defdc
MD
841 } else {
842#endif
843 cpu_pause();
844 cpu_lfence();
845#if 1
846 }
8b402283 847#endif
0f0466c0 848 }
b12defdc
MD
849 DELAY(1);
850 atomic_add_int(&lwkt_cseq_rindex, 1);
2a418930 851 }
8cee56f4 852#endif
2a418930 853 /* highest level for(;;) loop */
f1d1c3fa 854 }
8a8d5d85 855
2a418930 856havethread:
b12defdc 857 /*
be71787b
MD
858 * Clear gd_idle_repeat when doing a normal switch to a non-idle
859 * thread.
f9235b6d 860 */
9ac1ee6e 861 ntd->td_wmesg = NULL;
b12defdc 862 ++gd->gd_cnt.v_swtch;
be71787b 863 gd->gd_idle_repeat = 0;
2a418930 864
f9235b6d 865havethread_preempted:
f9235b6d
MD
866 /*
867 * If the new target does not need the MP lock and we are holding it,
868 * release the MP lock. If the new target requires the MP lock we have
869 * already acquired it for the target.
8a8d5d85 870 */
2a418930 871 ;
f9235b6d
MD
872haveidle:
873 KASSERT(ntd->td_critcount,
b5d16701
MD
874 ("priority problem in lwkt_switch %d %d",
875 td->td_critcount, ntd->td_critcount));
876
94f6d86e 877 if (td != ntd) {
cc9b6223
MD
878 /*
879 * Execute the actual thread switch operation. This function
880 * returns to the current thread and returns the previous thread
881 * (which may be different from the thread we switched to).
882 *
883 * We are responsible for marking ntd as TDF_RUNNING.
884 */
121f93bc 885 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
94f6d86e 886 ++switch_count;
a1f0fb66 887 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
888 ntd->td_flags |= TDF_RUNNING;
889 lwkt_switch_return(td->td_switch(ntd));
890 /* ntd invalid, td_switch() can return a different thread_t */
94f6d86e 891 }
b12defdc 892
b12defdc 893 /*
54341a3b 894 * catch-all. XXX is this strictly needed?
b12defdc
MD
895 */
896 splz_check();
54341a3b 897
37af14fe
MD
898 /* NOTE: current cpu may have changed after switch */
899 crit_exit_quick(td);
8ad65e08
MD
900}
901
cc9b6223
MD
902/*
903 * Called by assembly in the td_switch (thread restore path) for thread
904 * bootstrap cases which do not 'return' to lwkt_switch().
905 */
906void
907lwkt_switch_return(thread_t otd)
908{
cc9b6223
MD
909 globaldata_t rgd;
910
911 /*
912 * Check if otd was migrating. Now that we are on ntd we can finish
913 * up the migration. This is a bit messy but it is the only place
914 * where td is known to be fully descheduled.
915 *
916 * We can only activate the migration if otd was migrating but not
917 * held on the cpu due to a preemption chain. We still have to
918 * clear TDF_RUNNING on the old thread either way.
919 *
920 * We are responsible for clearing the previously running thread's
921 * TDF_RUNNING.
922 */
923 if ((rgd = otd->td_migrate_gd) != NULL &&
924 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
925 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
926 (TDF_MIGRATING | TDF_RUNNING));
927 otd->td_migrate_gd = NULL;
928 otd->td_flags &= ~TDF_RUNNING;
929 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
930 } else {
931 otd->td_flags &= ~TDF_RUNNING;
932 }
2b07d9aa
MD
933
934 /*
935 * Final exit validations (see lwp_wait()). Note that otd becomes
936 * invalid the *instant* we set TDF_MP_EXITSIG.
937 */
938 while (otd->td_flags & TDF_EXITING) {
939 u_int mpflags;
940
941 mpflags = otd->td_mpflags;
942 cpu_ccfence();
943
944 if (mpflags & TDF_MP_EXITWAIT) {
945 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
946 mpflags | TDF_MP_EXITSIG)) {
947 wakeup(otd);
948 break;
949 }
950 } else {
951 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
952 mpflags | TDF_MP_EXITSIG)) {
953 wakeup(otd);
954 break;
955 }
956 }
957 }
cc9b6223
MD
958}
959
b68b7282 960/*
96728c05 961 * Request that the target thread preempt the current thread. Preemption
54341a3b
MD
962 * can only occur if our only critical section is the one that we were called
963 * with, the relative priority of the target thread is higher, and the target
964 * thread holds no tokens. This also only works if we are not holding any
965 * spinlocks (obviously).
96728c05
MD
966 *
967 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
968 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 969 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
970 * to determine whether preemption is possible (aka usually just the crit
971 * priority of lwkt_schedule() itself).
b68b7282 972 *
54341a3b
MD
973 * Preemption is typically limited to interrupt threads.
974 *
975 * Operation works in a fairly straight-forward manner. The normal
976 * scheduling code is bypassed and we switch directly to the target
977 * thread. When the target thread attempts to block or switch away
978 * code at the base of lwkt_switch() will switch directly back to our
979 * thread. Our thread is able to retain whatever tokens it holds and
980 * if the target needs one of them the target will switch back to us
981 * and reschedule itself normally.
b68b7282
MD
982 */
983void
f9235b6d 984lwkt_preempt(thread_t ntd, int critcount)
b68b7282 985{
46a3f46d 986 struct globaldata *gd = mycpu;
cc9b6223 987 thread_t xtd;
0a3f9b47 988 thread_t td;
2d910aaf 989 int save_gd_intr_nesting_level;
b68b7282 990
26a0694b 991 /*
96728c05
MD
992 * The caller has put us in a critical section. We can only preempt
993 * if the caller of the caller was not in a critical section (basically
f9235b6d 994 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 995 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 996 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
997 *
998 * YYY The target thread must be in a critical section (else it must
999 * inherit our critical section? I dunno yet).
26a0694b 1000 */
f9235b6d 1001 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 1002
b12defdc 1003 td = gd->gd_curthread;
fbc024e4
MD
1004 if (preempt_enable == 0) {
1005 ++preempt_miss;
1006 return;
1007 }
f9235b6d 1008 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
1009 ++preempt_miss;
1010 return;
1011 }
f9235b6d 1012 if (td->td_critcount > critcount) {
96728c05
MD
1013 ++preempt_miss;
1014 return;
1015 }
121f93bc
MD
1016 if (td->td_cscount) {
1017 ++preempt_miss;
1018 return;
1019 }
46a3f46d 1020 if (ntd->td_gd != gd) {
96728c05
MD
1021 ++preempt_miss;
1022 return;
1023 }
41a01a4d 1024 /*
77912481
MD
1025 * We don't have to check spinlocks here as they will also bump
1026 * td_critcount.
d3d1cbc8
MD
1027 *
1028 * Do not try to preempt if the target thread is holding any tokens.
1029 * We could try to acquire the tokens but this case is so rare there
1030 * is no need to support it.
41a01a4d 1031 */
0846e4ce 1032 KKASSERT(gd->gd_spinlocks == 0);
77912481 1033
3b998fa9 1034 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8 1035 ++preempt_miss;
d3d1cbc8
MD
1036 return;
1037 }
26a0694b
MD
1038 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1039 ++preempt_weird;
1040 return;
1041 }
1042 if (ntd->td_preempted) {
4b5f931b 1043 ++preempt_hit;
26a0694b 1044 return;
b68b7282 1045 }
da0b0e8b 1046 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 1047
8ec60c3f
MD
1048 /*
1049 * Since we are able to preempt the current thread, there is no need to
1050 * call need_lwkt_resched().
2d910aaf
MD
1051 *
1052 * We must temporarily clear gd_intr_nesting_level around the switch
1053 * since switchouts from the target thread are allowed (they will just
1054 * return to our thread), and since the target thread has its own stack.
cc9b6223
MD
1055 *
1056 * A preemption must switch back to the original thread, assert the
1057 * case.
8ec60c3f 1058 */
26a0694b
MD
1059 ++preempt_hit;
1060 ntd->td_preempted = td;
1061 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1062 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1063 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1064 gd->gd_intr_nesting_level = 0;
121f93bc
MD
1065
1066 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
cc9b6223
MD
1067 ntd->td_flags |= TDF_RUNNING;
1068 xtd = td->td_switch(ntd);
1069 KKASSERT(xtd == ntd);
1070 lwkt_switch_return(xtd);
2d910aaf 1071 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1072
26a0694b
MD
1073 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1074 ntd->td_preempted = NULL;
1075 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1076}
1077
f1d1c3fa 1078/*
faaeffac 1079 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1080 * This will work inside a critical section but not inside a hard code
1081 * section.
ef0fdad1 1082 *
f1d1c3fa
MD
1083 * (self contained on a per cpu basis)
1084 */
1085void
faaeffac 1086splz_check(void)
f1d1c3fa 1087{
7966cb69
MD
1088 globaldata_t gd = mycpu;
1089 thread_t td = gd->gd_curthread;
ef0fdad1 1090
4a28fe22
MD
1091 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1092 gd->gd_intr_nesting_level == 0 &&
1093 td->td_nest_count < 2)
1094 {
f1d1c3fa 1095 splz();
4a28fe22
MD
1096 }
1097}
1098
1099/*
1100 * This version is integrated into crit_exit, reqflags has already
1101 * been tested but td_critcount has not.
1102 *
1103 * We only want to execute the splz() on the 1->0 transition of
1104 * critcount and not in a hard code section or if too deeply nested.
925040f2 1105 *
0846e4ce 1106 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
4a28fe22
MD
1107 */
1108void
1109lwkt_maybe_splz(thread_t td)
1110{
1111 globaldata_t gd = td->td_gd;
1112
1113 if (td->td_critcount == 0 &&
1114 gd->gd_intr_nesting_level == 0 &&
1115 td->td_nest_count < 2)
1116 {
1117 splz();
1118 }
f1d1c3fa
MD
1119}
1120
e6546af9
MD
1121/*
1122 * Drivers which set up processing co-threads can call this function to
1123 * run the co-thread at a higher priority and to allow it to preempt
1124 * normal threads.
1125 */
1126void
1127lwkt_set_interrupt_support_thread(void)
1128{
1129 thread_t td = curthread;
1130
1131 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1132 td->td_flags |= TDF_INTTHREAD;
1133 td->td_preemptable = lwkt_preempt;
1134}
1135
1136
8ad65e08 1137/*
f9235b6d
MD
1138 * This function is used to negotiate a passive release of the current
1139 * process/lwp designation with the user scheduler, allowing the user
1140 * scheduler to schedule another user thread. The related kernel thread
1141 * (curthread) continues running in the released state.
8ad65e08
MD
1142 */
1143void
f9235b6d 1144lwkt_passive_release(struct thread *td)
8ad65e08 1145{
f9235b6d
MD
1146 struct lwp *lp = td->td_lwp;
1147
e3e6be1f 1148#ifndef NO_LWKT_SPLIT_USERPRI
f9235b6d
MD
1149 td->td_release = NULL;
1150 lwkt_setpri_self(TDPRI_KERN_USER);
d992c377
MD
1151#endif
1152
f9235b6d 1153 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1154}
1155
f9235b6d 1156
3824f392 1157/*
d2d8515b
MD
1158 * This implements a LWKT yield, allowing a kernel thread to yield to other
1159 * kernel threads at the same or higher priority. This function can be
1160 * called in a tight loop and will typically only yield once per tick.
f9235b6d 1161 *
d2d8515b
MD
1162 * Most kernel threads run at the same priority in order to allow equal
1163 * sharing.
f9235b6d
MD
1164 *
1165 * (self contained on a per cpu basis)
3824f392
MD
1166 */
1167void
f9235b6d 1168lwkt_yield(void)
3824f392 1169{
f9235b6d
MD
1170 globaldata_t gd = mycpu;
1171 thread_t td = gd->gd_curthread;
3824f392 1172
f9235b6d
MD
1173 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1174 splz();
85946b6c 1175 if (lwkt_resched_wanted()) {
f9235b6d
MD
1176 lwkt_schedule_self(curthread);
1177 lwkt_switch();
f9235b6d 1178 }
3824f392
MD
1179}
1180
40504122
MD
1181/*
1182 * The quick version processes pending interrupts and higher-priority
1183 * LWKT threads but will not round-robin same-priority LWKT threads.
de4d4cb0
MD
1184 *
1185 * When called while attempting to return to userland the only same-pri
1186 * threads are the ones which have already tried to become the current
1187 * user process.
40504122
MD
1188 */
1189void
1190lwkt_yield_quick(void)
1191{
1192 globaldata_t gd = mycpu;
1193 thread_t td = gd->gd_curthread;
1194
1195 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1196 splz();
1197 if (lwkt_resched_wanted()) {
9c99cb33 1198 crit_enter();
40504122
MD
1199 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1200 clear_lwkt_resched();
1201 } else {
1202 lwkt_schedule_self(curthread);
1203 lwkt_switch();
1204 }
9c99cb33 1205 crit_exit();
40504122
MD
1206 }
1207}
1208
3824f392 1209/*
f9235b6d
MD
1210 * This yield is designed for kernel threads with a user context.
1211 *
1212 * The kernel acting on behalf of the user is potentially cpu-bound,
1213 * this function will efficiently allow other threads to run and also
1214 * switch to other processes by releasing.
3824f392
MD
1215 *
1216 * The lwkt_user_yield() function is designed to have very low overhead
1217 * if no yield is determined to be needed.
1218 */
1219void
1220lwkt_user_yield(void)
1221{
f9235b6d
MD
1222 globaldata_t gd = mycpu;
1223 thread_t td = gd->gd_curthread;
1224
1225 /*
1226 * Always run any pending interrupts in case we are in a critical
1227 * section.
1228 */
1229 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1230 splz();
3824f392 1231
3824f392 1232 /*
f9235b6d
MD
1233 * Switch (which forces a release) if another kernel thread needs
1234 * the cpu, if userland wants us to resched, or if our kernel
1235 * quantum has run out.
3824f392 1236 */
f9235b6d 1237 if (lwkt_resched_wanted() ||
85946b6c 1238 user_resched_wanted())
f9235b6d 1239 {
3824f392 1240 lwkt_switch();
3824f392
MD
1241 }
1242
f9235b6d 1243#if 0
3824f392 1244 /*
f9235b6d
MD
1245 * Reacquire the current process if we are released.
1246 *
1247 * XXX not implemented atm. The kernel may be holding locks and such,
1248 * so we want the thread to continue to receive cpu.
3824f392 1249 */
f9235b6d
MD
1250 if (td->td_release == NULL && lp) {
1251 lp->lwp_proc->p_usched->acquire_curproc(lp);
1252 td->td_release = lwkt_passive_release;
1253 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1254 }
f9235b6d 1255#endif
b9eb1c19
MD
1256}
1257
8ad65e08 1258/*
f1d1c3fa
MD
1259 * Generic schedule. Possibly schedule threads belonging to other cpus and
1260 * deal with threads that might be blocked on a wait queue.
1261 *
0a3f9b47
MD
1262 * We have a little helper inline function which does additional work after
1263 * the thread has been enqueued, including dealing with preemption and
1264 * setting need_lwkt_resched() (which prevents the kernel from returning
1265 * to userland until it has processed higher priority threads).
6330a558
MD
1266 *
1267 * It is possible for this routine to be called after a failed _enqueue
1268 * (due to the target thread migrating, sleeping, or otherwise blocked).
1269 * We have to check that the thread is actually on the run queue!
8ad65e08 1270 */
0a3f9b47
MD
1271static __inline
1272void
85946b6c 1273_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
0a3f9b47 1274{
6330a558 1275 if (ntd->td_flags & TDF_RUNQ) {
85946b6c 1276 if (ntd->td_preemptable) {
f9235b6d 1277 ntd->td_preemptable(ntd, ccount); /* YYY +token */
6330a558 1278 }
0a3f9b47
MD
1279 }
1280}
1281
361d01dd 1282static __inline
8ad65e08 1283void
85946b6c 1284_lwkt_schedule(thread_t td)
8ad65e08 1285{
37af14fe
MD
1286 globaldata_t mygd = mycpu;
1287
cf709dd2
MD
1288 KASSERT(td != &td->td_gd->gd_idlethread,
1289 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1290 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1291 crit_enter_gd(mygd);
4643740a
MD
1292 KKASSERT(td->td_lwp == NULL ||
1293 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1294
37af14fe 1295 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1296 _lwkt_enqueue(td);
1297 } else {
f1d1c3fa 1298 /*
7cd8d145
MD
1299 * If we own the thread, there is no race (since we are in a
1300 * critical section). If we do not own the thread there might
1301 * be a race but the target cpu will deal with it.
f1d1c3fa 1302 */
7cd8d145 1303 if (td->td_gd == mygd) {
9d265729 1304 _lwkt_enqueue(td);
85946b6c 1305 _lwkt_schedule_post(mygd, td, 1);
f1d1c3fa 1306 } else {
e381e77c 1307 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1308 }
8ad65e08 1309 }
37af14fe 1310 crit_exit_gd(mygd);
8ad65e08
MD
1311}
1312
361d01dd
MD
1313void
1314lwkt_schedule(thread_t td)
1315{
85946b6c 1316 _lwkt_schedule(td);
361d01dd
MD
1317}
1318
1319void
85946b6c 1320lwkt_schedule_noresched(thread_t td) /* XXX not impl */
361d01dd 1321{
85946b6c 1322 _lwkt_schedule(td);
361d01dd
MD
1323}
1324
e381e77c
MD
1325/*
1326 * When scheduled remotely if frame != NULL the IPIQ is being
1327 * run via doreti or an interrupt then preemption can be allowed.
1328 *
1329 * To allow preemption we have to drop the critical section so only
1330 * one is present in _lwkt_schedule_post.
1331 */
1332static void
1333lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1334{
1335 thread_t td = curthread;
1336 thread_t ntd = arg;
1337
1338 if (frame && ntd->td_preemptable) {
1339 crit_exit_noyield(td);
85946b6c 1340 _lwkt_schedule(ntd);
e381e77c
MD
1341 crit_enter_quick(td);
1342 } else {
85946b6c 1343 _lwkt_schedule(ntd);
e381e77c
MD
1344 }
1345}
1346
d9eea1a5 1347/*
52eedfb5
MD
1348 * Thread migration using a 'Pull' method. The thread may or may not be
1349 * the current thread. It MUST be descheduled and in a stable state.
1350 * lwkt_giveaway() must be called on the cpu owning the thread.
1351 *
1352 * At any point after lwkt_giveaway() is called, the target cpu may
1353 * 'pull' the thread by calling lwkt_acquire().
1354 *
ae8e83e6
MD
1355 * We have to make sure the thread is not sitting on a per-cpu tsleep
1356 * queue or it will blow up when it moves to another cpu.
1357 *
52eedfb5 1358 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1359 */
52eedfb5
MD
1360void
1361lwkt_giveaway(thread_t td)
1362{
3b4192fb 1363 globaldata_t gd = mycpu;
52eedfb5 1364
3b4192fb
MD
1365 crit_enter_gd(gd);
1366 if (td->td_flags & TDF_TSLEEPQ)
1367 tsleep_remove(td);
1368 KKASSERT(td->td_gd == gd);
1369 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1370 td->td_flags |= TDF_MIGRATING;
1371 crit_exit_gd(gd);
52eedfb5
MD
1372}
1373
a2a5ad0d
MD
1374void
1375lwkt_acquire(thread_t td)
1376{
37af14fe
MD
1377 globaldata_t gd;
1378 globaldata_t mygd;
cc9b6223 1379 int retry = 10000000;
a2a5ad0d 1380
52eedfb5 1381 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1382 gd = td->td_gd;
37af14fe 1383 mygd = mycpu;
52eedfb5 1384 if (gd != mycpu) {
35238fa5 1385 cpu_lfence();
52eedfb5 1386 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1387 crit_enter_gd(mygd);
cfaeae2a 1388 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23 1389 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
df910c23 1390 lwkt_process_ipiq();
52eedfb5 1391 cpu_lfence();
cc9b6223
MD
1392 if (--retry == 0) {
1393 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1394 td, td->td_flags);
1395 retry = 10000000;
1396 }
df910c23 1397 }
cfaeae2a 1398 DEBUG_POP_INFO();
562273ea 1399 cpu_mfence();
37af14fe 1400 td->td_gd = mygd;
52eedfb5
MD
1401 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1402 td->td_flags &= ~TDF_MIGRATING;
1403 crit_exit_gd(mygd);
1404 } else {
1405 crit_enter_gd(mygd);
1406 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1407 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1408 crit_exit_gd(mygd);
a2a5ad0d
MD
1409 }
1410}
1411
f1d1c3fa
MD
1412/*
1413 * Generic deschedule. Descheduling threads other then your own should be
1414 * done only in carefully controlled circumstances. Descheduling is
1415 * asynchronous.
1416 *
1417 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1418 */
1419void
1420lwkt_deschedule(thread_t td)
1421{
f1d1c3fa
MD
1422 crit_enter();
1423 if (td == curthread) {
1424 _lwkt_dequeue(td);
1425 } else {
a72187e9 1426 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1427 _lwkt_dequeue(td);
1428 } else {
b8a98473 1429 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1430 }
1431 }
1432 crit_exit();
1433}
1434
4b5f931b
MD
1435/*
1436 * Set the target thread's priority. This routine does not automatically
1437 * switch to a higher priority thread, LWKT threads are not designed for
1438 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1439 */
1440void
1441lwkt_setpri(thread_t td, int pri)
1442{
f9235b6d
MD
1443 if (td->td_pri != pri) {
1444 KKASSERT(pri >= 0);
1445 crit_enter();
1446 if (td->td_flags & TDF_RUNQ) {
d2d8515b 1447 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1448 _lwkt_dequeue(td);
1449 td->td_pri = pri;
1450 _lwkt_enqueue(td);
1451 } else {
1452 td->td_pri = pri;
1453 }
1454 crit_exit();
26a0694b 1455 }
26a0694b
MD
1456}
1457
03bd0a5e
MD
1458/*
1459 * Set the initial priority for a thread prior to it being scheduled for
1460 * the first time. The thread MUST NOT be scheduled before or during
1461 * this call. The thread may be assigned to a cpu other then the current
1462 * cpu.
1463 *
1464 * Typically used after a thread has been created with TDF_STOPPREQ,
1465 * and before the thread is initially scheduled.
1466 */
1467void
1468lwkt_setpri_initial(thread_t td, int pri)
1469{
1470 KKASSERT(pri >= 0);
1471 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1472 td->td_pri = pri;
03bd0a5e
MD
1473}
1474
26a0694b
MD
1475void
1476lwkt_setpri_self(int pri)
1477{
1478 thread_t td = curthread;
1479
4b5f931b
MD
1480 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1481 crit_enter();
1482 if (td->td_flags & TDF_RUNQ) {
1483 _lwkt_dequeue(td);
f9235b6d 1484 td->td_pri = pri;
4b5f931b
MD
1485 _lwkt_enqueue(td);
1486 } else {
f9235b6d 1487 td->td_pri = pri;
4b5f931b
MD
1488 }
1489 crit_exit();
1490}
1491
f9235b6d 1492/*
85946b6c 1493 * hz tick scheduler clock for LWKT threads
f9235b6d
MD
1494 */
1495void
85946b6c 1496lwkt_schedulerclock(thread_t td)
f9235b6d 1497{
85946b6c
MD
1498 globaldata_t gd = td->td_gd;
1499 thread_t xtd;
2a418930 1500
85946b6c
MD
1501 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1502 /*
1503 * If the current thread is at the head of the runq shift it to the
1504 * end of any equal-priority threads and request a LWKT reschedule
1505 * if it moved.
d992c377
MD
1506 *
1507 * Ignore upri in this situation. There will only be one user thread
1508 * in user mode, all others will be user threads running in kernel
1509 * mode and we have to make sure they get some cpu.
85946b6c
MD
1510 */
1511 xtd = TAILQ_NEXT(td, td_threadq);
1512 if (xtd && xtd->td_pri == td->td_pri) {
1513 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1514 while (xtd && xtd->td_pri == td->td_pri)
1515 xtd = TAILQ_NEXT(xtd, td_threadq);
1516 if (xtd)
1517 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1518 else
1519 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1520 need_lwkt_resched();
f9235b6d 1521 }
85946b6c
MD
1522 } else {
1523 /*
1524 * If we scheduled a thread other than the one at the head of the
1525 * queue always request a reschedule every tick.
1526 */
1527 need_lwkt_resched();
f9235b6d
MD
1528 }
1529}
1530
5d21b981 1531/*
52eedfb5
MD
1532 * Migrate the current thread to the specified cpu.
1533 *
cc9b6223
MD
1534 * This is accomplished by descheduling ourselves from the current cpu
1535 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1536 * 'old' thread wants to migrate after it has been completely switched out
1537 * and will complete the migration.
1538 *
1539 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1540 *
1541 * We must be sure to release our current process designation (if a user
1542 * process) before clearing out any tsleepq we are on because the release
1543 * code may re-add us.
ae8e83e6
MD
1544 *
1545 * We must be sure to remove ourselves from the current cpu's tsleepq
1546 * before potentially moving to another queue. The thread can be on
1547 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1548 */
5d21b981
MD
1549
1550void
1551lwkt_setcpu_self(globaldata_t rgd)
1552{
5d21b981
MD
1553 thread_t td = curthread;
1554
1555 if (td->td_gd != rgd) {
1556 crit_enter_quick(td);
cc9b6223 1557
95858b91
MD
1558 if (td->td_release)
1559 td->td_release(td);
ae8e83e6 1560 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1561 tsleep_remove(td);
cc9b6223
MD
1562
1563 /*
1564 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1565 * trying to deschedule ourselves and switch away, then deschedule
1566 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1567 * call lwkt_switch() to complete the operation.
1568 */
5d21b981
MD
1569 td->td_flags |= TDF_MIGRATING;
1570 lwkt_deschedule_self(td);
52eedfb5 1571 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1572 td->td_migrate_gd = rgd;
5d21b981 1573 lwkt_switch();
cc9b6223
MD
1574
1575 /*
1576 * We are now on the target cpu
1577 */
1578 KKASSERT(rgd == mycpu);
52eedfb5 1579 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1580 crit_exit_quick(td);
1581 }
5d21b981
MD
1582}
1583
ecdefdda
MD
1584void
1585lwkt_migratecpu(int cpuid)
1586{
ecdefdda
MD
1587 globaldata_t rgd;
1588
1589 rgd = globaldata_find(cpuid);
1590 lwkt_setcpu_self(rgd);
ecdefdda
MD
1591}
1592
5d21b981
MD
1593/*
1594 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
1595 * do not have to enter another one).
1596 *
1597 * The thread (td) has already been completely descheduled from the
1598 * originating cpu and we can simply assert the case. The thread is
1599 * assigned to the new cpu and enqueued.
5d21b981 1600 *
cc9b6223 1601 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1602 */
1603static void
1604lwkt_setcpu_remote(void *arg)
1605{
1606 thread_t td = arg;
1607 globaldata_t gd = mycpu;
1608
cc9b6223 1609 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1610 td->td_gd = gd;
562273ea 1611 cpu_mfence();
5d21b981 1612 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1613 KKASSERT(td->td_migrate_gd == NULL);
4643740a
MD
1614 KKASSERT(td->td_lwp == NULL ||
1615 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
5d21b981
MD
1616 _lwkt_enqueue(td);
1617}
1618
553ea3c8 1619struct lwp *
4b5f931b
MD
1620lwkt_preempted_proc(void)
1621{
73e4f7b9 1622 thread_t td = curthread;
4b5f931b
MD
1623 while (td->td_preempted)
1624 td = td->td_preempted;
553ea3c8 1625 return(td->td_lwp);
4b5f931b
MD
1626}
1627
99df837e
MD
1628/*
1629 * Create a kernel process/thread/whatever. It shares it's address space
1630 * with proc0 - ie: kernel only.
1631 *
d2d8515b
MD
1632 * If the cpu is not specified one will be selected. In the future
1633 * specifying a cpu of -1 will enable kernel thread migration between
1634 * cpus.
99df837e
MD
1635 */
1636int
c9e9fb21
MD
1637lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1638 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1639{
73e4f7b9 1640 thread_t td;
e2565a42 1641 __va_list ap;
99df837e 1642
d3d32139 1643 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1644 tdflags);
a2a5ad0d
MD
1645 if (tdp)
1646 *tdp = td;
709799ea 1647 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1648
1649 /*
1650 * Set up arg0 for 'ps' etc
1651 */
e2565a42 1652 __va_start(ap, fmt);
379210cb 1653 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1654 __va_end(ap);
99df837e
MD
1655
1656 /*
1657 * Schedule the thread to run
1658 */
4643740a
MD
1659 if (td->td_flags & TDF_NOSTART)
1660 td->td_flags &= ~TDF_NOSTART;
ef0fdad1 1661 else
4643740a 1662 lwkt_schedule(td);
99df837e
MD
1663 return 0;
1664}
1665
1666/*
1667 * Destroy an LWKT thread. Warning! This function is not called when
1668 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1669 * uses a different reaping mechanism.
1670 */
1671void
1672lwkt_exit(void)
1673{
1674 thread_t td = curthread;
c070746a 1675 thread_t std;
8826f33a 1676 globaldata_t gd;
99df837e 1677
2883d2d8
MD
1678 /*
1679 * Do any cleanup that might block here
1680 */
99df837e 1681 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1682 kprintf("kthread %p %s has exited\n", td, td->td_comm);
2883d2d8
MD
1683 biosched_done(td);
1684 dsched_exit_thread(td);
c070746a
MD
1685
1686 /*
1687 * Get us into a critical section to interlock gd_freetd and loop
1688 * until we can get it freed.
1689 *
1690 * We have to cache the current td in gd_freetd because objcache_put()ing
1691 * it would rip it out from under us while our thread is still active.
2af9d75d
MD
1692 *
1693 * We are the current thread so of course our own TDF_RUNNING bit will
1694 * be set, so unlike the lwp reap code we don't wait for it to clear.
c070746a
MD
1695 */
1696 gd = mycpu;
37af14fe 1697 crit_enter_quick(td);
2af9d75d
MD
1698 for (;;) {
1699 if (td->td_refs) {
1700 tsleep(td, 0, "tdreap", 1);
1701 continue;
1702 }
1703 if ((std = gd->gd_freetd) != NULL) {
1704 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1705 gd->gd_freetd = NULL;
1706 objcache_put(thread_cache, std);
1707 continue;
1708 }
1709 break;
c070746a 1710 }
3b4192fb
MD
1711
1712 /*
1713 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1714 * the last time. We cannot block after this point or we may end
1715 * up with a stale td on the tsleepq.
eb2adbf5
MD
1716 *
1717 * None of this may block, the critical section is the only thing
1718 * protecting tdallq and the only thing preventing new lwkt_hold()
1719 * thread refs now.
3b4192fb
MD
1720 */
1721 if (td->td_flags & TDF_TSLEEPQ)
1722 tsleep_remove(td);
37af14fe 1723 lwkt_deschedule_self(td);
e56e4dea 1724 lwkt_remove_tdallq(td);
74c9628e 1725 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1726
1727 /*
1728 * Final cleanup
1729 */
1730 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1731 if (td->td_flags & TDF_ALLOCATED_THREAD)
1732 gd->gd_freetd = td;
99df837e
MD
1733 cpu_thread_exit();
1734}
1735
e56e4dea
MD
1736void
1737lwkt_remove_tdallq(thread_t td)
1738{
1739 KKASSERT(td->td_gd == mycpu);
1740 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1741}
1742
9cf43f91
MD
1743/*
1744 * Code reduction and branch prediction improvements. Call/return
1745 * overhead on modern cpus often degenerates into 0 cycles due to
1746 * the cpu's branch prediction hardware and return pc cache. We
1747 * can take advantage of this by not inlining medium-complexity
1748 * functions and we can also reduce the branch prediction impact
1749 * by collapsing perfectly predictable branches into a single
1750 * procedure instead of duplicating it.
1751 *
1752 * Is any of this noticeable? Probably not, so I'll take the
1753 * smaller code size.
1754 */
1755void
b6468f56 1756crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1757{
b6468f56 1758 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1759}
1760
2d93b37a
MD
1761void
1762crit_panic(void)
1763{
1764 thread_t td = curthread;
850634cc 1765 int lcrit = td->td_critcount;
2d93b37a 1766
850634cc
AH
1767 td->td_critcount = 0;
1768 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1769 /* NOT REACHED */
2d93b37a
MD
1770}
1771
bd8015ca
MD
1772/*
1773 * Called from debugger/panic on cpus which have been stopped. We must still
1774 * process the IPIQ while stopped, even if we were stopped while in a critical
1775 * section (XXX).
1776 *
1777 * If we are dumping also try to process any pending interrupts. This may
1778 * or may not work depending on the state of the cpu at the point it was
1779 * stopped.
1780 */
1781void
1782lwkt_smp_stopped(void)
1783{
1784 globaldata_t gd = mycpu;
1785
1786 crit_enter_gd(gd);
1787 if (dumping) {
1788 lwkt_process_ipiq();
1789 splz();
1790 } else {
1791 lwkt_process_ipiq();
1792 }
1793 crit_exit_gd(gd);
1794}