hammer2 - slave sync work
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
b12defdc 2 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
60f60350 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
60f60350 20 *
8c10bfcf
MD
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
b37f18d6 47#include <sys/kinfo.h>
8ad65e08 48#include <sys/queue.h>
7d0bac62 49#include <sys/sysctl.h>
99df837e 50#include <sys/kthread.h>
f1d1c3fa 51#include <machine/cpu.h>
99df837e 52#include <sys/lock.h>
9d265729 53#include <sys/spinlock.h>
57aa743c 54#include <sys/ktr.h>
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
0c52fa62
SG
118SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
119 "Number of switched threads");
9733f757 120SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
0c52fa62 121 "Successful preemption events");
9733f757 122SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
0c52fa62
SG
123 "Failed preemption events");
124SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
125 "Number of preempted threads.");
b12defdc 126static int fairq_enable = 0;
2a418930
MD
127SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
128 &fairq_enable, 0, "Turn on fairq priority accumulators");
85946b6c 129static int fairq_bypass = -1;
b12defdc
MD
130SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
131 &fairq_bypass, 0, "Allow fairq to bypass td on token failure");
132extern int lwkt_sched_debug;
133int lwkt_sched_debug = 0;
134SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
135 &lwkt_sched_debug, 0, "Scheduler debug");
2a418930
MD
136static int lwkt_spin_loops = 10;
137SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
b12defdc
MD
138 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
139static int lwkt_spin_reseq = 0;
140SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW,
141 &lwkt_spin_reseq, 0, "Scheduler resequencer enable");
142static int lwkt_spin_monitor = 0;
143SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW,
144 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait");
d5b2d319
MD
145static int lwkt_spin_fatal = 0; /* disabled */
146SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
147 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
fbc024e4 148static int preempt_enable = 1;
2a418930
MD
149SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
150 &preempt_enable, 0, "Enable preemption");
7b234d8c 151static int lwkt_cache_threads = 0;
765b1ae0
MD
152SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
153 &lwkt_cache_threads, 0, "thread+kstack cache");
fbc024e4 154
8cee56f4 155#ifndef _KERNEL_VIRTUAL
2a418930
MD
156static __cachealign int lwkt_cseq_rindex;
157static __cachealign int lwkt_cseq_windex;
8cee56f4 158#endif
05220613 159
4b5f931b
MD
160/*
161 * These helper procedures handle the runq, they can only be called from
162 * within a critical section.
75cdbe6c
MD
163 *
164 * WARNING! Prior to SMP being brought up it is possible to enqueue and
165 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
166 * instead of 'mycpu' when referencing the globaldata structure. Once
167 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 168 */
f1d1c3fa
MD
169static __inline
170void
171_lwkt_dequeue(thread_t td)
172{
173 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 174 struct globaldata *gd = td->td_gd;
4b5f931b 175
f1d1c3fa 176 td->td_flags &= ~TDF_RUNQ;
f9235b6d 177 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
de4d4cb0 178 --gd->gd_tdrunqcount;
f9235b6d 179 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
2a418930 180 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
181 }
182}
183
f9235b6d
MD
184/*
185 * Priority enqueue.
186 *
d992c377
MD
187 * There are a limited number of lwkt threads runnable since user
188 * processes only schedule one at a time per cpu. However, there can
189 * be many user processes in kernel mode exiting from a tsleep() which
e3e6be1f 190 * become runnable.
d992c377
MD
191 *
192 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and
193 * will ignore user priority. This is to ensure that user threads in
194 * kernel mode get cpu at some point regardless of what the user
195 * scheduler thinks.
f9235b6d 196 */
f1d1c3fa
MD
197static __inline
198void
199_lwkt_enqueue(thread_t td)
200{
f9235b6d
MD
201 thread_t xtd;
202
7f5d7ed7 203 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 204 struct globaldata *gd = td->td_gd;
4b5f931b 205
f1d1c3fa 206 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
207 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
208 if (xtd == NULL) {
85946b6c
MD
209 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
210 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
f9235b6d 211 } else {
e3e6be1f
MD
212 /*
213 * NOTE: td_upri - higher numbers more desireable, same sense
214 * as td_pri (typically reversed from lwp_upri).
215 *
216 * In the equal priority case we want the best selection
217 * at the beginning so the less desireable selections know
218 * that they have to setrunqueue/go-to-another-cpu, even
219 * though it means switching back to the 'best' selection.
220 * This also avoids degenerate situations when many threads
221 * are runnable or waking up at the same time.
222 *
223 * If upri matches exactly place at end/round-robin.
224 */
d992c377 225 while (xtd &&
e3e6be1f 226 (xtd->td_pri >= td->td_pri ||
d992c377 227 (xtd->td_pri == td->td_pri &&
e3e6be1f 228 xtd->td_upri >= td->td_upri))) {
85946b6c 229 xtd = TAILQ_NEXT(xtd, td_threadq);
d992c377 230 }
85946b6c
MD
231 if (xtd)
232 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
233 else
234 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
f9235b6d 235 }
de4d4cb0 236 ++gd->gd_tdrunqcount;
b12defdc
MD
237
238 /*
85946b6c 239 * Request a LWKT reschedule if we are now at the head of the queue.
b12defdc 240 */
85946b6c
MD
241 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
242 need_lwkt_resched();
f1d1c3fa
MD
243 }
244}
8ad65e08 245
40aaf5fc
NT
246static __boolean_t
247_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
248{
249 struct thread *td = (struct thread *)obj;
250
251 td->td_kstack = NULL;
252 td->td_kstack_size = 0;
253 td->td_flags = TDF_ALLOCATED_THREAD;
4643740a 254 td->td_mpflags = 0;
40aaf5fc
NT
255 return (1);
256}
257
258static void
259_lwkt_thread_dtor(void *obj, void *privdata)
260{
261 struct thread *td = (struct thread *)obj;
262
263 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
264 ("_lwkt_thread_dtor: not allocated from objcache"));
265 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
266 td->td_kstack_size > 0,
267 ("_lwkt_thread_dtor: corrupted stack"));
268 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
7b234d8c
MD
269 td->td_kstack = NULL;
270 td->td_flags = 0;
40aaf5fc
NT
271}
272
273/*
274 * Initialize the lwkt s/system.
765b1ae0 275 *
7b234d8c
MD
276 * Nominally cache up to 32 thread + kstack structures. Cache more on
277 * systems with a lot of cpu cores.
40aaf5fc
NT
278 */
279void
280lwkt_init(void)
281{
765b1ae0 282 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
7b234d8c
MD
283 if (lwkt_cache_threads == 0) {
284 lwkt_cache_threads = ncpus * 4;
285 if (lwkt_cache_threads < 32)
286 lwkt_cache_threads = 32;
287 }
765b1ae0
MD
288 thread_cache = objcache_create_mbacked(
289 M_THREAD, sizeof(struct thread),
2fce2579 290 0, lwkt_cache_threads,
765b1ae0 291 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
292}
293
37af14fe
MD
294/*
295 * Schedule a thread to run. As the current thread we can always safely
296 * schedule ourselves, and a shortcut procedure is provided for that
297 * function.
298 *
299 * (non-blocking, self contained on a per cpu basis)
300 */
301void
302lwkt_schedule_self(thread_t td)
303{
cfaeae2a 304 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 305 crit_enter_quick(td);
f9235b6d
MD
306 KASSERT(td != &td->td_gd->gd_idlethread,
307 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
4643740a
MD
308 KKASSERT(td->td_lwp == NULL ||
309 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
37af14fe 310 _lwkt_enqueue(td);
37af14fe
MD
311 crit_exit_quick(td);
312}
313
314/*
315 * Deschedule a thread.
316 *
317 * (non-blocking, self contained on a per cpu basis)
318 */
319void
320lwkt_deschedule_self(thread_t td)
321{
322 crit_enter_quick(td);
37af14fe
MD
323 _lwkt_dequeue(td);
324 crit_exit_quick(td);
325}
326
8ad65e08
MD
327/*
328 * LWKTs operate on a per-cpu basis
329 *
73e4f7b9 330 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
331 */
332void
333lwkt_gdinit(struct globaldata *gd)
334{
f9235b6d 335 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 336 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
337}
338
7d0bac62
MD
339/*
340 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
341 * or LWKT start address before it can be scheduled. If the target cpu is
342 * -1 the thread will be created on the current cpu.
0cfcada1
MD
343 *
344 * If you intend to create a thread without a process context this function
345 * does everything except load the startup and switcher function.
7d0bac62
MD
346 */
347thread_t
d3d32139 348lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 349{
d2d8515b 350 static int cpu_rotator;
c070746a 351 globaldata_t gd = mycpu;
99df837e 352 void *stack;
7d0bac62 353
c070746a
MD
354 /*
355 * If static thread storage is not supplied allocate a thread. Reuse
356 * a cached free thread if possible. gd_freetd is used to keep an exiting
357 * thread intact through the exit.
358 */
ef0fdad1 359 if (td == NULL) {
cf709dd2
MD
360 crit_enter_gd(gd);
361 if ((td = gd->gd_freetd) != NULL) {
362 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
363 TDF_RUNQ)) == 0);
c070746a 364 gd->gd_freetd = NULL;
cf709dd2 365 } else {
c070746a 366 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
367 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
368 TDF_RUNQ)) == 0);
369 }
370 crit_exit_gd(gd);
40aaf5fc 371 KASSERT((td->td_flags &
2af9d75d
MD
372 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
373 TDF_ALLOCATED_THREAD,
40aaf5fc
NT
374 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
375 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 376 }
c070746a
MD
377
378 /*
379 * Try to reuse cached stack.
380 */
f470d0c8
MD
381 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
382 if (flags & TDF_ALLOCATED_STACK) {
e4846942 383 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
384 stack = NULL;
385 }
386 }
387 if (stack == NULL) {
e40cfbd7 388 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 389 flags |= TDF_ALLOCATED_STACK;
99df837e 390 }
d2d8515b
MD
391 if (cpu < 0) {
392 cpu = ++cpu_rotator;
393 cpu_ccfence();
394 cpu %= ncpus;
395 }
396 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 397 return(td);
7d0bac62
MD
398}
399
400/*
401 * Initialize a preexisting thread structure. This function is used by
402 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
403 *
f8c3996b
MD
404 * All threads start out in a critical section at a priority of
405 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
406 * appropriate. This function may send an IPI message when the
407 * requested cpu is not the current cpu and consequently gd_tdallq may
408 * not be initialized synchronously from the point of view of the originating
409 * cpu.
410 *
411 * NOTE! we have to be careful in regards to creating threads for other cpus
412 * if SMP has not yet been activated.
7d0bac62 413 */
75cdbe6c
MD
414static void
415lwkt_init_thread_remote(void *arg)
416{
417 thread_t td = arg;
418
52eedfb5
MD
419 /*
420 * Protected by critical section held by IPI dispatch
421 */
75cdbe6c
MD
422 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
423}
424
fdce8919
MD
425/*
426 * lwkt core thread structural initialization.
427 *
428 * NOTE: All threads are initialized as mpsafe threads.
429 */
7d0bac62 430void
f470d0c8
MD
431lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
432 struct globaldata *gd)
7d0bac62 433{
37af14fe
MD
434 globaldata_t mygd = mycpu;
435
99df837e
MD
436 bzero(td, sizeof(struct thread));
437 td->td_kstack = stack;
f470d0c8 438 td->td_kstack_size = stksize;
d3d32139 439 td->td_flags = flags;
4643740a 440 td->td_mpflags = 0;
f256b6c0 441 td->td_type = TD_TYPE_GENERIC;
26a0694b 442 td->td_gd = gd;
f9235b6d
MD
443 td->td_pri = TDPRI_KERN_DAEMON;
444 td->td_critcount = 1;
54341a3b 445 td->td_toks_have = NULL;
3b998fa9 446 td->td_toks_stop = &td->td_toks_base;
c068fb59
SZ
447 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) {
448 lwkt_initport_spin(&td->td_msgport, td,
449 (flags & TDF_FIXEDCPU) ? TRUE : FALSE);
450 } else {
fb0f29c4 451 lwkt_initport_thread(&td->td_msgport, td);
c068fb59 452 }
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 }
a86ce0cd 515
e7c0dbba 516 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
517}
518
519
8ad65e08
MD
520/*
521 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
522 * switch to the idlethread. Switching must occur within a critical
523 * section to avoid races with the scheduling queue.
524 *
525 * We always have full control over our cpu's run queue. Other cpus
526 * that wish to manipulate our queue must use the cpu_*msg() calls to
527 * talk to our cpu, so a critical section is all that is needed and
528 * the result is very, very fast thread switching.
529 *
96728c05
MD
530 * The LWKT scheduler uses a fixed priority model and round-robins at
531 * each priority level. User process scheduling is a totally
532 * different beast and LWKT priorities should not be confused with
533 * user process priorities.
f1d1c3fa 534 *
69d78e99
MD
535 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
536 * is not called by the current thread in the preemption case, only when
537 * the preempting thread blocks (in order to return to the original thread).
cfaeae2a
MD
538 *
539 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
540 * migration and tsleep deschedule the current lwkt thread and call
541 * lwkt_switch(). In particular, the target cpu of the migration fully
542 * expects the thread to become non-runnable and can deadlock against
543 * cpusync operations if we run any IPIs prior to switching the thread out.
544 *
545 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
95858b91 546 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
8ad65e08
MD
547 */
548void
549lwkt_switch(void)
550{
37af14fe
MD
551 globaldata_t gd = mycpu;
552 thread_t td = gd->gd_curthread;
8ad65e08 553 thread_t ntd;
b12defdc 554 int spinning = 0;
8ad65e08 555
da0b0e8b 556 KKASSERT(gd->gd_processing_ipiq == 0);
121f93bc 557 KKASSERT(td->td_flags & TDF_RUNNING);
da0b0e8b 558
46a3f46d 559 /*
27e88a6e
MD
560 * Switching from within a 'fast' (non thread switched) interrupt or IPI
561 * is illegal. However, we may have to do it anyway if we hit a fatal
562 * kernel trap or we have paniced.
563 *
564 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 565 */
27e88a6e
MD
566 if (gd->gd_intr_nesting_level) {
567 int savegdnest;
568 int savegdtrap;
569
5fddbda2 570 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22 571 panic("lwkt_switch: Attempt to switch from a "
5a8df152 572 "fast interrupt, ipi, or hard code section, "
4a28fe22
MD
573 "td %p\n",
574 td);
27e88a6e
MD
575 } else {
576 savegdnest = gd->gd_intr_nesting_level;
577 savegdtrap = gd->gd_trap_nesting_level;
578 gd->gd_intr_nesting_level = 0;
579 gd->gd_trap_nesting_level = 0;
a7422615
MD
580 if ((td->td_flags & TDF_PANICWARN) == 0) {
581 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
582 kprintf("Warning: thread switch from interrupt, IPI, "
583 "or hard code section.\n"
a7422615 584 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 585 print_backtrace(-1);
a7422615 586 }
27e88a6e
MD
587 lwkt_switch();
588 gd->gd_intr_nesting_level = savegdnest;
589 gd->gd_trap_nesting_level = savegdtrap;
590 return;
591 }
96728c05 592 }
ef0fdad1 593
cb973d15 594 /*
85946b6c
MD
595 * Release our current user process designation if we are blocking
596 * or if a user reschedule was requested.
597 *
598 * NOTE: This function is NOT called if we are switching into or
599 * returning from a preemption.
600 *
601 * NOTE: Releasing our current user process designation may cause
602 * it to be assigned to another thread, which in turn will
603 * cause us to block in the usched acquire code when we attempt
604 * to return to userland.
605 *
606 * NOTE: On SMP systems this can be very nasty when heavy token
607 * contention is present so we want to be careful not to
608 * release the designation gratuitously.
cb973d15 609 */
85946b6c
MD
610 if (td->td_release &&
611 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
cb973d15 612 td->td_release(td);
85946b6c 613 }
cb973d15 614
85946b6c
MD
615 /*
616 * Release all tokens
617 */
37af14fe 618 crit_enter_gd(gd);
3b998fa9 619 if (TD_TOKS_HELD(td))
9d265729
MD
620 lwkt_relalltokens(td);
621
622 /*
b02926de
MD
623 * We had better not be holding any spin locks, but don't get into an
624 * endless panic loop.
9d265729 625 */
0846e4ce 626 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL,
d666840a 627 ("lwkt_switch: still holding %d exclusive spinlocks!",
0846e4ce 628 gd->gd_spinlocks));
9d265729 629
8a8d5d85 630
0f7a3396
MD
631#ifdef INVARIANTS
632 if (td->td_cscount) {
6ea70f76 633 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
634 td);
635 if (panic_on_cscount)
636 panic("switching while mastering cpusync");
637 }
8a8d5d85 638#endif
f9235b6d
MD
639
640 /*
641 * If we had preempted another thread on this cpu, resume the preempted
642 * thread. This occurs transparently, whether the preempted thread
643 * was scheduled or not (it may have been preempted after descheduling
644 * itself).
645 *
646 * We have to setup the MP lock for the original thread after backing
647 * out the adjustment that was made to curthread when the original
648 * was preempted.
649 */
99df837e 650 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
651 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
652 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
653
654 /*
b9eb1c19
MD
655 * The interrupt may have woken a thread up, we need to properly
656 * set the reschedule flag if the originally interrupted thread is
657 * at a lower priority.
85946b6c
MD
658 *
659 * The interrupt may not have descheduled.
8ec60c3f 660 */
85946b6c 661 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
8ec60c3f 662 need_lwkt_resched();
f9235b6d
MD
663 goto havethread_preempted;
664 }
665
b12defdc 666 /*
f9235b6d 667 * If we cannot obtain ownership of the tokens we cannot immediately
cfaeae2a
MD
668 * schedule the target thread.
669 *
670 * Reminder: Again, we cannot afford to run any IPIs in this path if
671 * the current thread has been descheduled.
f9235b6d
MD
672 */
673 for (;;) {
b12defdc 674 clear_lwkt_resched();
f9235b6d 675
4b5f931b 676 /*
2a418930 677 * Hotpath - pull the head of the run queue and attempt to schedule
85946b6c 678 * it.
41a01a4d 679 */
7c3dd309
NA
680 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
681
682 if (ntd == NULL) {
683 /*
684 * Runq is empty, switch to idle to allow it to halt.
685 */
686 ntd = &gd->gd_idlethread;
7c3dd309
NA
687 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
688 ASSERT_NO_TOKENS_HELD(ntd);
7c3dd309
NA
689 cpu_time.cp_msg[0] = 0;
690 cpu_time.cp_stallpc = 0;
691 goto haveidle;
f9235b6d 692 }
41a01a4d 693
8ec60c3f 694 /*
b12defdc 695 * Hotpath - schedule ntd.
6f207a2c
MD
696 *
697 * NOTE: For UP there is no mplock and lwkt_getalltokens()
698 * always succeeds.
8ec60c3f 699 */
b12defdc
MD
700 if (TD_TOKS_NOT_HELD(ntd) ||
701 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
702 {
f9235b6d 703 goto havethread;
b12defdc 704 }
f9235b6d 705
f9235b6d 706 /*
2a418930
MD
707 * Coldpath (SMP only since tokens always succeed on UP)
708 *
709 * We had some contention on the thread we wanted to schedule.
710 * What we do now is try to find a thread that we can schedule
b12defdc 711 * in its stead.
2a418930 712 *
85946b6c
MD
713 * The coldpath scan does NOT rearrange threads in the run list.
714 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
715 * the next tick whenever the current head is not the current thread.
f9235b6d 716 */
b12defdc 717#ifdef INVARIANTS
85946b6c 718 ++ntd->td_contended;
b12defdc 719#endif
050032ec 720 ++gd->gd_cnt.v_lock_colls;
b12defdc 721
85946b6c 722 if (fairq_bypass > 0)
b12defdc
MD
723 goto skip;
724
b12defdc 725 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
e3e6be1f 726#ifndef NO_LWKT_SPLIT_USERPRI
85946b6c
MD
727 /*
728 * Never schedule threads returning to userland or the
729 * user thread scheduler helper thread when higher priority
bb6860b5
NA
730 * threads are present. The runq is sorted by priority
731 * so we can give up traversing it when we find the first
732 * low priority thread.
85946b6c
MD
733 */
734 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
735 ntd = NULL;
736 break;
737 }
d992c377 738#endif
85946b6c
MD
739
740 /*
741 * Try this one.
742 */
b12defdc
MD
743 if (TD_TOKS_NOT_HELD(ntd) ||
744 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
745 goto havethread;
746 }
85946b6c 747#ifdef INVARIANTS
85946b6c 748 ++ntd->td_contended;
b12defdc 749#endif
050032ec 750 ++gd->gd_cnt.v_lock_colls;
2a418930
MD
751 }
752
b12defdc 753skip:
2a418930
MD
754 /*
755 * We exhausted the run list, meaning that all runnable threads
b12defdc 756 * are contested.
2a418930
MD
757 */
758 cpu_pause();
8cee56f4
MD
759#ifdef _KERNEL_VIRTUAL
760 pthread_yield();
761#endif
2a418930 762 ntd = &gd->gd_idlethread;
2a418930
MD
763 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
764 ASSERT_NO_TOKENS_HELD(ntd);
765 /* contention case, do not clear contention mask */
2a418930
MD
766
767 /*
b12defdc
MD
768 * We are going to have to retry but if the current thread is not
769 * on the runq we instead switch through the idle thread to get away
770 * from the current thread. We have to flag for lwkt reschedule
771 * to prevent the idle thread from halting.
2a418930 772 *
b12defdc
MD
773 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
774 * instruct it to deal with the potential for deadlocks by
775 * ordering the tokens by address.
2a418930 776 */
b12defdc 777 if ((td->td_flags & TDF_RUNQ) == 0) {
85946b6c 778 need_lwkt_resched(); /* prevent hlt */
2a418930 779 goto haveidle;
4b5f931b 780 }
dbefba87 781#if defined(INVARIANTS) && defined(__x86_64__)
d5b2d319
MD
782 if ((read_rflags() & PSL_I) == 0) {
783 cpu_enable_intr();
784 panic("lwkt_switch() called with interrupts disabled");
785 }
786#endif
b12defdc
MD
787
788 /*
789 * Number iterations so far. After a certain point we switch to
790 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
791 */
792 if (spinning < 0x7FFFFFFF)
793 ++spinning;
794
8cee56f4 795#ifndef _KERNEL_VIRTUAL
b12defdc
MD
796 /*
797 * lwkt_getalltokens() failed in sorted token mode, we can use
798 * monitor/mwait in this case.
799 */
800 if (spinning >= lwkt_spin_loops &&
801 (cpu_mi_feature & CPU_MI_MONITOR) &&
802 lwkt_spin_monitor)
803 {
804 cpu_mmw_pause_int(&gd->gd_reqflags,
805 (gd->gd_reqflags | RQF_SPINNING) &
a46b4a23 806 ~RQF_IDLECHECK_WK_MASK,
e5c52dd7 807 cpu_mwait_spin, 0);
b12defdc 808 }
8cee56f4 809#endif
b12defdc
MD
810
811 /*
812 * We already checked that td is still scheduled so this should be
813 * safe.
814 */
815 splz_check();
816
8cee56f4 817#ifndef _KERNEL_VIRTUAL
b12defdc
MD
818 /*
819 * This experimental resequencer is used as a fall-back to reduce
820 * hw cache line contention by placing each core's scheduler into a
821 * time-domain-multplexed slot.
822 *
823 * The resequencer is disabled by default. It's functionality has
824 * largely been superceeded by the token algorithm which limits races
825 * to a subset of cores.
826 *
827 * The resequencer algorithm tends to break down when more than
828 * 20 cores are contending. What appears to happen is that new
829 * tokens can be obtained out of address-sorted order by new cores
830 * while existing cores languish in long delays between retries and
831 * wind up being starved-out of the token acquisition.
832 */
833 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
834 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
835 int oseq;
836
837 while ((oseq = lwkt_cseq_rindex) != cseq) {
838 cpu_ccfence();
839#if 1
840 if (cpu_mi_feature & CPU_MI_MONITOR) {
e5c52dd7
SZ
841 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq,
842 cpu_mwait_spin, 0);
b12defdc
MD
843 } else {
844#endif
845 cpu_pause();
846 cpu_lfence();
847#if 1
848 }
8b402283 849#endif
0f0466c0 850 }
b12defdc
MD
851 DELAY(1);
852 atomic_add_int(&lwkt_cseq_rindex, 1);
2a418930 853 }
8cee56f4 854#endif
2a418930 855 /* highest level for(;;) loop */
f1d1c3fa 856 }
8a8d5d85 857
2a418930 858havethread:
b12defdc 859 /*
be71787b
MD
860 * Clear gd_idle_repeat when doing a normal switch to a non-idle
861 * thread.
f9235b6d 862 */
9ac1ee6e 863 ntd->td_wmesg = NULL;
b12defdc 864 ++gd->gd_cnt.v_swtch;
be71787b 865 gd->gd_idle_repeat = 0;
2a418930 866
f9235b6d 867havethread_preempted:
f9235b6d
MD
868 /*
869 * If the new target does not need the MP lock and we are holding it,
870 * release the MP lock. If the new target requires the MP lock we have
871 * already acquired it for the target.
8a8d5d85 872 */
2a418930 873 ;
f9235b6d
MD
874haveidle:
875 KASSERT(ntd->td_critcount,
b5d16701
MD
876 ("priority problem in lwkt_switch %d %d",
877 td->td_critcount, ntd->td_critcount));
878
94f6d86e 879 if (td != ntd) {
cc9b6223
MD
880 /*
881 * Execute the actual thread switch operation. This function
882 * returns to the current thread and returns the previous thread
883 * (which may be different from the thread we switched to).
884 *
885 * We are responsible for marking ntd as TDF_RUNNING.
886 */
121f93bc 887 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
94f6d86e 888 ++switch_count;
a1f0fb66 889 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
890 ntd->td_flags |= TDF_RUNNING;
891 lwkt_switch_return(td->td_switch(ntd));
892 /* ntd invalid, td_switch() can return a different thread_t */
94f6d86e 893 }
b12defdc 894
b12defdc 895 /*
54341a3b 896 * catch-all. XXX is this strictly needed?
b12defdc
MD
897 */
898 splz_check();
54341a3b 899
37af14fe
MD
900 /* NOTE: current cpu may have changed after switch */
901 crit_exit_quick(td);
8ad65e08
MD
902}
903
cc9b6223
MD
904/*
905 * Called by assembly in the td_switch (thread restore path) for thread
906 * bootstrap cases which do not 'return' to lwkt_switch().
907 */
908void
909lwkt_switch_return(thread_t otd)
910{
cc9b6223
MD
911 globaldata_t rgd;
912
913 /*
914 * Check if otd was migrating. Now that we are on ntd we can finish
915 * up the migration. This is a bit messy but it is the only place
916 * where td is known to be fully descheduled.
917 *
918 * We can only activate the migration if otd was migrating but not
919 * held on the cpu due to a preemption chain. We still have to
920 * clear TDF_RUNNING on the old thread either way.
921 *
922 * We are responsible for clearing the previously running thread's
923 * TDF_RUNNING.
924 */
925 if ((rgd = otd->td_migrate_gd) != NULL &&
926 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
927 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
928 (TDF_MIGRATING | TDF_RUNNING));
929 otd->td_migrate_gd = NULL;
930 otd->td_flags &= ~TDF_RUNNING;
931 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
932 } else {
933 otd->td_flags &= ~TDF_RUNNING;
934 }
2b07d9aa
MD
935
936 /*
937 * Final exit validations (see lwp_wait()). Note that otd becomes
938 * invalid the *instant* we set TDF_MP_EXITSIG.
939 */
940 while (otd->td_flags & TDF_EXITING) {
941 u_int mpflags;
942
943 mpflags = otd->td_mpflags;
944 cpu_ccfence();
945
946 if (mpflags & TDF_MP_EXITWAIT) {
947 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
948 mpflags | TDF_MP_EXITSIG)) {
949 wakeup(otd);
950 break;
951 }
952 } else {
953 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
954 mpflags | TDF_MP_EXITSIG)) {
955 wakeup(otd);
956 break;
957 }
958 }
959 }
cc9b6223
MD
960}
961
b68b7282 962/*
96728c05 963 * Request that the target thread preempt the current thread. Preemption
54341a3b
MD
964 * can only occur if our only critical section is the one that we were called
965 * with, the relative priority of the target thread is higher, and the target
966 * thread holds no tokens. This also only works if we are not holding any
967 * spinlocks (obviously).
96728c05
MD
968 *
969 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
970 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 971 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
972 * to determine whether preemption is possible (aka usually just the crit
973 * priority of lwkt_schedule() itself).
b68b7282 974 *
54341a3b
MD
975 * Preemption is typically limited to interrupt threads.
976 *
977 * Operation works in a fairly straight-forward manner. The normal
978 * scheduling code is bypassed and we switch directly to the target
979 * thread. When the target thread attempts to block or switch away
980 * code at the base of lwkt_switch() will switch directly back to our
981 * thread. Our thread is able to retain whatever tokens it holds and
982 * if the target needs one of them the target will switch back to us
983 * and reschedule itself normally.
b68b7282
MD
984 */
985void
f9235b6d 986lwkt_preempt(thread_t ntd, int critcount)
b68b7282 987{
46a3f46d 988 struct globaldata *gd = mycpu;
cc9b6223 989 thread_t xtd;
0a3f9b47 990 thread_t td;
2d910aaf 991 int save_gd_intr_nesting_level;
b68b7282 992
26a0694b 993 /*
96728c05
MD
994 * The caller has put us in a critical section. We can only preempt
995 * if the caller of the caller was not in a critical section (basically
f9235b6d 996 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 997 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 998 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
999 *
1000 * YYY The target thread must be in a critical section (else it must
1001 * inherit our critical section? I dunno yet).
26a0694b 1002 */
f9235b6d 1003 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 1004
b12defdc 1005 td = gd->gd_curthread;
fbc024e4
MD
1006 if (preempt_enable == 0) {
1007 ++preempt_miss;
1008 return;
1009 }
f9235b6d 1010 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
1011 ++preempt_miss;
1012 return;
1013 }
f9235b6d 1014 if (td->td_critcount > critcount) {
96728c05
MD
1015 ++preempt_miss;
1016 return;
1017 }
121f93bc
MD
1018 if (td->td_cscount) {
1019 ++preempt_miss;
1020 return;
1021 }
46a3f46d 1022 if (ntd->td_gd != gd) {
96728c05
MD
1023 ++preempt_miss;
1024 return;
1025 }
41a01a4d 1026 /*
77912481
MD
1027 * We don't have to check spinlocks here as they will also bump
1028 * td_critcount.
d3d1cbc8
MD
1029 *
1030 * Do not try to preempt if the target thread is holding any tokens.
1031 * We could try to acquire the tokens but this case is so rare there
1032 * is no need to support it.
41a01a4d 1033 */
0846e4ce 1034 KKASSERT(gd->gd_spinlocks == 0);
77912481 1035
3b998fa9 1036 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8 1037 ++preempt_miss;
d3d1cbc8
MD
1038 return;
1039 }
26a0694b
MD
1040 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1041 ++preempt_weird;
1042 return;
1043 }
1044 if (ntd->td_preempted) {
4b5f931b 1045 ++preempt_hit;
26a0694b 1046 return;
b68b7282 1047 }
da0b0e8b 1048 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 1049
8ec60c3f
MD
1050 /*
1051 * Since we are able to preempt the current thread, there is no need to
1052 * call need_lwkt_resched().
2d910aaf
MD
1053 *
1054 * We must temporarily clear gd_intr_nesting_level around the switch
1055 * since switchouts from the target thread are allowed (they will just
1056 * return to our thread), and since the target thread has its own stack.
cc9b6223
MD
1057 *
1058 * A preemption must switch back to the original thread, assert the
1059 * case.
8ec60c3f 1060 */
26a0694b
MD
1061 ++preempt_hit;
1062 ntd->td_preempted = td;
1063 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1064 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1065 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1066 gd->gd_intr_nesting_level = 0;
121f93bc
MD
1067
1068 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
cc9b6223
MD
1069 ntd->td_flags |= TDF_RUNNING;
1070 xtd = td->td_switch(ntd);
1071 KKASSERT(xtd == ntd);
1072 lwkt_switch_return(xtd);
2d910aaf 1073 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1074
26a0694b
MD
1075 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1076 ntd->td_preempted = NULL;
1077 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1078}
1079
f1d1c3fa 1080/*
faaeffac 1081 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1082 * This will work inside a critical section but not inside a hard code
1083 * section.
ef0fdad1 1084 *
f1d1c3fa
MD
1085 * (self contained on a per cpu basis)
1086 */
1087void
faaeffac 1088splz_check(void)
f1d1c3fa 1089{
7966cb69
MD
1090 globaldata_t gd = mycpu;
1091 thread_t td = gd->gd_curthread;
ef0fdad1 1092
4a28fe22
MD
1093 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1094 gd->gd_intr_nesting_level == 0 &&
1095 td->td_nest_count < 2)
1096 {
f1d1c3fa 1097 splz();
4a28fe22
MD
1098 }
1099}
1100
1101/*
1102 * This version is integrated into crit_exit, reqflags has already
1103 * been tested but td_critcount has not.
1104 *
1105 * We only want to execute the splz() on the 1->0 transition of
1106 * critcount and not in a hard code section or if too deeply nested.
925040f2 1107 *
0846e4ce 1108 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
4a28fe22
MD
1109 */
1110void
1111lwkt_maybe_splz(thread_t td)
1112{
1113 globaldata_t gd = td->td_gd;
1114
1115 if (td->td_critcount == 0 &&
1116 gd->gd_intr_nesting_level == 0 &&
1117 td->td_nest_count < 2)
1118 {
1119 splz();
1120 }
f1d1c3fa
MD
1121}
1122
e6546af9
MD
1123/*
1124 * Drivers which set up processing co-threads can call this function to
1125 * run the co-thread at a higher priority and to allow it to preempt
1126 * normal threads.
1127 */
1128void
1129lwkt_set_interrupt_support_thread(void)
1130{
1131 thread_t td = curthread;
1132
1133 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1134 td->td_flags |= TDF_INTTHREAD;
1135 td->td_preemptable = lwkt_preempt;
1136}
1137
1138
8ad65e08 1139/*
f9235b6d
MD
1140 * This function is used to negotiate a passive release of the current
1141 * process/lwp designation with the user scheduler, allowing the user
1142 * scheduler to schedule another user thread. The related kernel thread
1143 * (curthread) continues running in the released state.
8ad65e08
MD
1144 */
1145void
f9235b6d 1146lwkt_passive_release(struct thread *td)
8ad65e08 1147{
f9235b6d
MD
1148 struct lwp *lp = td->td_lwp;
1149
e3e6be1f 1150#ifndef NO_LWKT_SPLIT_USERPRI
f9235b6d
MD
1151 td->td_release = NULL;
1152 lwkt_setpri_self(TDPRI_KERN_USER);
d992c377
MD
1153#endif
1154
f9235b6d 1155 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1156}
1157
f9235b6d 1158
3824f392 1159/*
d2d8515b
MD
1160 * This implements a LWKT yield, allowing a kernel thread to yield to other
1161 * kernel threads at the same or higher priority. This function can be
1162 * called in a tight loop and will typically only yield once per tick.
f9235b6d 1163 *
d2d8515b
MD
1164 * Most kernel threads run at the same priority in order to allow equal
1165 * sharing.
f9235b6d
MD
1166 *
1167 * (self contained on a per cpu basis)
3824f392
MD
1168 */
1169void
f9235b6d 1170lwkt_yield(void)
3824f392 1171{
f9235b6d
MD
1172 globaldata_t gd = mycpu;
1173 thread_t td = gd->gd_curthread;
3824f392 1174
f9235b6d
MD
1175 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1176 splz();
85946b6c 1177 if (lwkt_resched_wanted()) {
f9235b6d
MD
1178 lwkt_schedule_self(curthread);
1179 lwkt_switch();
f9235b6d 1180 }
3824f392
MD
1181}
1182
40504122
MD
1183/*
1184 * The quick version processes pending interrupts and higher-priority
1185 * LWKT threads but will not round-robin same-priority LWKT threads.
de4d4cb0
MD
1186 *
1187 * When called while attempting to return to userland the only same-pri
1188 * threads are the ones which have already tried to become the current
1189 * user process.
40504122
MD
1190 */
1191void
1192lwkt_yield_quick(void)
1193{
1194 globaldata_t gd = mycpu;
1195 thread_t td = gd->gd_curthread;
1196
1197 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1198 splz();
1199 if (lwkt_resched_wanted()) {
9c99cb33 1200 crit_enter();
40504122
MD
1201 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1202 clear_lwkt_resched();
1203 } else {
1204 lwkt_schedule_self(curthread);
1205 lwkt_switch();
1206 }
9c99cb33 1207 crit_exit();
40504122
MD
1208 }
1209}
1210
3824f392 1211/*
f9235b6d
MD
1212 * This yield is designed for kernel threads with a user context.
1213 *
1214 * The kernel acting on behalf of the user is potentially cpu-bound,
1215 * this function will efficiently allow other threads to run and also
1216 * switch to other processes by releasing.
3824f392
MD
1217 *
1218 * The lwkt_user_yield() function is designed to have very low overhead
1219 * if no yield is determined to be needed.
1220 */
1221void
1222lwkt_user_yield(void)
1223{
f9235b6d
MD
1224 globaldata_t gd = mycpu;
1225 thread_t td = gd->gd_curthread;
1226
1227 /*
1228 * Always run any pending interrupts in case we are in a critical
1229 * section.
1230 */
1231 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1232 splz();
3824f392 1233
3824f392 1234 /*
f9235b6d
MD
1235 * Switch (which forces a release) if another kernel thread needs
1236 * the cpu, if userland wants us to resched, or if our kernel
1237 * quantum has run out.
3824f392 1238 */
f9235b6d 1239 if (lwkt_resched_wanted() ||
85946b6c 1240 user_resched_wanted())
f9235b6d 1241 {
3824f392 1242 lwkt_switch();
3824f392
MD
1243 }
1244
f9235b6d 1245#if 0
3824f392 1246 /*
f9235b6d
MD
1247 * Reacquire the current process if we are released.
1248 *
1249 * XXX not implemented atm. The kernel may be holding locks and such,
1250 * so we want the thread to continue to receive cpu.
3824f392 1251 */
f9235b6d
MD
1252 if (td->td_release == NULL && lp) {
1253 lp->lwp_proc->p_usched->acquire_curproc(lp);
1254 td->td_release = lwkt_passive_release;
1255 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1256 }
f9235b6d 1257#endif
b9eb1c19
MD
1258}
1259
8ad65e08 1260/*
f1d1c3fa
MD
1261 * Generic schedule. Possibly schedule threads belonging to other cpus and
1262 * deal with threads that might be blocked on a wait queue.
1263 *
0a3f9b47
MD
1264 * We have a little helper inline function which does additional work after
1265 * the thread has been enqueued, including dealing with preemption and
1266 * setting need_lwkt_resched() (which prevents the kernel from returning
1267 * to userland until it has processed higher priority threads).
6330a558
MD
1268 *
1269 * It is possible for this routine to be called after a failed _enqueue
1270 * (due to the target thread migrating, sleeping, or otherwise blocked).
1271 * We have to check that the thread is actually on the run queue!
8ad65e08 1272 */
0a3f9b47
MD
1273static __inline
1274void
85946b6c 1275_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
0a3f9b47 1276{
6330a558 1277 if (ntd->td_flags & TDF_RUNQ) {
85946b6c 1278 if (ntd->td_preemptable) {
f9235b6d 1279 ntd->td_preemptable(ntd, ccount); /* YYY +token */
6330a558 1280 }
0a3f9b47
MD
1281 }
1282}
1283
361d01dd 1284static __inline
8ad65e08 1285void
85946b6c 1286_lwkt_schedule(thread_t td)
8ad65e08 1287{
37af14fe
MD
1288 globaldata_t mygd = mycpu;
1289
cf709dd2
MD
1290 KASSERT(td != &td->td_gd->gd_idlethread,
1291 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1292 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1293 crit_enter_gd(mygd);
4643740a
MD
1294 KKASSERT(td->td_lwp == NULL ||
1295 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1296
37af14fe 1297 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1298 _lwkt_enqueue(td);
1299 } else {
f1d1c3fa 1300 /*
7cd8d145
MD
1301 * If we own the thread, there is no race (since we are in a
1302 * critical section). If we do not own the thread there might
1303 * be a race but the target cpu will deal with it.
f1d1c3fa 1304 */
7cd8d145 1305 if (td->td_gd == mygd) {
9d265729 1306 _lwkt_enqueue(td);
85946b6c 1307 _lwkt_schedule_post(mygd, td, 1);
f1d1c3fa 1308 } else {
e381e77c 1309 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1310 }
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
e381e77c
MD
1327/*
1328 * When scheduled remotely if frame != NULL the IPIQ is being
1329 * run via doreti or an interrupt then preemption can be allowed.
1330 *
1331 * To allow preemption we have to drop the critical section so only
1332 * one is present in _lwkt_schedule_post.
1333 */
1334static void
1335lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1336{
1337 thread_t td = curthread;
1338 thread_t ntd = arg;
1339
1340 if (frame && ntd->td_preemptable) {
1341 crit_exit_noyield(td);
85946b6c 1342 _lwkt_schedule(ntd);
e381e77c
MD
1343 crit_enter_quick(td);
1344 } else {
85946b6c 1345 _lwkt_schedule(ntd);
e381e77c
MD
1346 }
1347}
1348
d9eea1a5 1349/*
52eedfb5
MD
1350 * Thread migration using a 'Pull' method. The thread may or may not be
1351 * the current thread. It MUST be descheduled and in a stable state.
1352 * lwkt_giveaway() must be called on the cpu owning the thread.
1353 *
1354 * At any point after lwkt_giveaway() is called, the target cpu may
1355 * 'pull' the thread by calling lwkt_acquire().
1356 *
ae8e83e6
MD
1357 * We have to make sure the thread is not sitting on a per-cpu tsleep
1358 * queue or it will blow up when it moves to another cpu.
1359 *
52eedfb5 1360 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1361 */
52eedfb5
MD
1362void
1363lwkt_giveaway(thread_t td)
1364{
3b4192fb 1365 globaldata_t gd = mycpu;
52eedfb5 1366
3b4192fb
MD
1367 crit_enter_gd(gd);
1368 if (td->td_flags & TDF_TSLEEPQ)
1369 tsleep_remove(td);
1370 KKASSERT(td->td_gd == gd);
1371 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1372 td->td_flags |= TDF_MIGRATING;
1373 crit_exit_gd(gd);
52eedfb5
MD
1374}
1375
a2a5ad0d
MD
1376void
1377lwkt_acquire(thread_t td)
1378{
37af14fe
MD
1379 globaldata_t gd;
1380 globaldata_t mygd;
cc9b6223 1381 int retry = 10000000;
a2a5ad0d 1382
52eedfb5 1383 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1384 gd = td->td_gd;
37af14fe 1385 mygd = mycpu;
52eedfb5 1386 if (gd != mycpu) {
35238fa5 1387 cpu_lfence();
52eedfb5 1388 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1389 crit_enter_gd(mygd);
cfaeae2a 1390 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23 1391 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
df910c23 1392 lwkt_process_ipiq();
52eedfb5 1393 cpu_lfence();
cc9b6223
MD
1394 if (--retry == 0) {
1395 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1396 td, td->td_flags);
1397 retry = 10000000;
1398 }
a86ce0cd
MD
1399#ifdef _KERNEL_VIRTUAL
1400 pthread_yield();
1401#endif
df910c23 1402 }
cfaeae2a 1403 DEBUG_POP_INFO();
562273ea 1404 cpu_mfence();
37af14fe 1405 td->td_gd = mygd;
52eedfb5
MD
1406 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1407 td->td_flags &= ~TDF_MIGRATING;
1408 crit_exit_gd(mygd);
1409 } else {
1410 crit_enter_gd(mygd);
1411 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1412 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1413 crit_exit_gd(mygd);
a2a5ad0d
MD
1414 }
1415}
1416
f1d1c3fa
MD
1417/*
1418 * Generic deschedule. Descheduling threads other then your own should be
1419 * done only in carefully controlled circumstances. Descheduling is
1420 * asynchronous.
1421 *
1422 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1423 */
1424void
1425lwkt_deschedule(thread_t td)
1426{
f1d1c3fa
MD
1427 crit_enter();
1428 if (td == curthread) {
1429 _lwkt_dequeue(td);
1430 } else {
a72187e9 1431 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1432 _lwkt_dequeue(td);
1433 } else {
b8a98473 1434 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1435 }
1436 }
1437 crit_exit();
1438}
1439
4b5f931b
MD
1440/*
1441 * Set the target thread's priority. This routine does not automatically
1442 * switch to a higher priority thread, LWKT threads are not designed for
1443 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1444 */
1445void
1446lwkt_setpri(thread_t td, int pri)
1447{
f9235b6d
MD
1448 if (td->td_pri != pri) {
1449 KKASSERT(pri >= 0);
1450 crit_enter();
1451 if (td->td_flags & TDF_RUNQ) {
d2d8515b 1452 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1453 _lwkt_dequeue(td);
1454 td->td_pri = pri;
1455 _lwkt_enqueue(td);
1456 } else {
1457 td->td_pri = pri;
1458 }
1459 crit_exit();
26a0694b 1460 }
26a0694b
MD
1461}
1462
03bd0a5e
MD
1463/*
1464 * Set the initial priority for a thread prior to it being scheduled for
1465 * the first time. The thread MUST NOT be scheduled before or during
1466 * this call. The thread may be assigned to a cpu other then the current
1467 * cpu.
1468 *
1469 * Typically used after a thread has been created with TDF_STOPPREQ,
1470 * and before the thread is initially scheduled.
1471 */
1472void
1473lwkt_setpri_initial(thread_t td, int pri)
1474{
1475 KKASSERT(pri >= 0);
1476 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1477 td->td_pri = pri;
03bd0a5e
MD
1478}
1479
26a0694b
MD
1480void
1481lwkt_setpri_self(int pri)
1482{
1483 thread_t td = curthread;
1484
4b5f931b
MD
1485 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1486 crit_enter();
1487 if (td->td_flags & TDF_RUNQ) {
1488 _lwkt_dequeue(td);
f9235b6d 1489 td->td_pri = pri;
4b5f931b
MD
1490 _lwkt_enqueue(td);
1491 } else {
f9235b6d 1492 td->td_pri = pri;
4b5f931b
MD
1493 }
1494 crit_exit();
1495}
1496
f9235b6d 1497/*
85946b6c 1498 * hz tick scheduler clock for LWKT threads
f9235b6d
MD
1499 */
1500void
85946b6c 1501lwkt_schedulerclock(thread_t td)
f9235b6d 1502{
85946b6c
MD
1503 globaldata_t gd = td->td_gd;
1504 thread_t xtd;
2a418930 1505
85946b6c
MD
1506 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1507 /*
1508 * If the current thread is at the head of the runq shift it to the
1509 * end of any equal-priority threads and request a LWKT reschedule
1510 * if it moved.
d992c377
MD
1511 *
1512 * Ignore upri in this situation. There will only be one user thread
1513 * in user mode, all others will be user threads running in kernel
1514 * mode and we have to make sure they get some cpu.
85946b6c
MD
1515 */
1516 xtd = TAILQ_NEXT(td, td_threadq);
1517 if (xtd && xtd->td_pri == td->td_pri) {
1518 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1519 while (xtd && xtd->td_pri == td->td_pri)
1520 xtd = TAILQ_NEXT(xtd, td_threadq);
1521 if (xtd)
1522 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1523 else
1524 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1525 need_lwkt_resched();
f9235b6d 1526 }
85946b6c
MD
1527 } else {
1528 /*
1529 * If we scheduled a thread other than the one at the head of the
1530 * queue always request a reschedule every tick.
1531 */
1532 need_lwkt_resched();
f9235b6d
MD
1533 }
1534}
1535
5d21b981 1536/*
52eedfb5
MD
1537 * Migrate the current thread to the specified cpu.
1538 *
cc9b6223
MD
1539 * This is accomplished by descheduling ourselves from the current cpu
1540 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1541 * 'old' thread wants to migrate after it has been completely switched out
1542 * and will complete the migration.
1543 *
1544 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1545 *
1546 * We must be sure to release our current process designation (if a user
1547 * process) before clearing out any tsleepq we are on because the release
1548 * code may re-add us.
ae8e83e6
MD
1549 *
1550 * We must be sure to remove ourselves from the current cpu's tsleepq
1551 * before potentially moving to another queue. The thread can be on
1552 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1553 */
5d21b981
MD
1554
1555void
1556lwkt_setcpu_self(globaldata_t rgd)
1557{
5d21b981
MD
1558 thread_t td = curthread;
1559
1560 if (td->td_gd != rgd) {
1561 crit_enter_quick(td);
cc9b6223 1562
95858b91
MD
1563 if (td->td_release)
1564 td->td_release(td);
ae8e83e6 1565 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1566 tsleep_remove(td);
cc9b6223
MD
1567
1568 /*
1569 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1570 * trying to deschedule ourselves and switch away, then deschedule
1571 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1572 * call lwkt_switch() to complete the operation.
1573 */
5d21b981
MD
1574 td->td_flags |= TDF_MIGRATING;
1575 lwkt_deschedule_self(td);
52eedfb5 1576 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1577 td->td_migrate_gd = rgd;
5d21b981 1578 lwkt_switch();
cc9b6223
MD
1579
1580 /*
1581 * We are now on the target cpu
1582 */
1583 KKASSERT(rgd == mycpu);
52eedfb5 1584 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1585 crit_exit_quick(td);
1586 }
5d21b981
MD
1587}
1588
ecdefdda
MD
1589void
1590lwkt_migratecpu(int cpuid)
1591{
ecdefdda
MD
1592 globaldata_t rgd;
1593
1594 rgd = globaldata_find(cpuid);
1595 lwkt_setcpu_self(rgd);
ecdefdda
MD
1596}
1597
5d21b981
MD
1598/*
1599 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
1600 * do not have to enter another one).
1601 *
1602 * The thread (td) has already been completely descheduled from the
1603 * originating cpu and we can simply assert the case. The thread is
1604 * assigned to the new cpu and enqueued.
5d21b981 1605 *
cc9b6223 1606 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1607 */
1608static void
1609lwkt_setcpu_remote(void *arg)
1610{
1611 thread_t td = arg;
1612 globaldata_t gd = mycpu;
1613
cc9b6223 1614 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1615 td->td_gd = gd;
562273ea 1616 cpu_mfence();
5d21b981 1617 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1618 KKASSERT(td->td_migrate_gd == NULL);
4643740a
MD
1619 KKASSERT(td->td_lwp == NULL ||
1620 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
5d21b981
MD
1621 _lwkt_enqueue(td);
1622}
1623
553ea3c8 1624struct lwp *
4b5f931b
MD
1625lwkt_preempted_proc(void)
1626{
73e4f7b9 1627 thread_t td = curthread;
4b5f931b
MD
1628 while (td->td_preempted)
1629 td = td->td_preempted;
553ea3c8 1630 return(td->td_lwp);
4b5f931b
MD
1631}
1632
99df837e
MD
1633/*
1634 * Create a kernel process/thread/whatever. It shares it's address space
1635 * with proc0 - ie: kernel only.
1636 *
d2d8515b
MD
1637 * If the cpu is not specified one will be selected. In the future
1638 * specifying a cpu of -1 will enable kernel thread migration between
1639 * cpus.
99df837e
MD
1640 */
1641int
c9e9fb21
MD
1642lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1643 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1644{
73e4f7b9 1645 thread_t td;
e2565a42 1646 __va_list ap;
99df837e 1647
d3d32139 1648 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1649 tdflags);
a2a5ad0d
MD
1650 if (tdp)
1651 *tdp = td;
709799ea 1652 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1653
1654 /*
1655 * Set up arg0 for 'ps' etc
1656 */
e2565a42 1657 __va_start(ap, fmt);
379210cb 1658 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1659 __va_end(ap);
99df837e
MD
1660
1661 /*
1662 * Schedule the thread to run
1663 */
4643740a
MD
1664 if (td->td_flags & TDF_NOSTART)
1665 td->td_flags &= ~TDF_NOSTART;
ef0fdad1 1666 else
4643740a 1667 lwkt_schedule(td);
99df837e
MD
1668 return 0;
1669}
1670
1671/*
1672 * Destroy an LWKT thread. Warning! This function is not called when
1673 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1674 * uses a different reaping mechanism.
1675 */
1676void
1677lwkt_exit(void)
1678{
1679 thread_t td = curthread;
c070746a 1680 thread_t std;
8826f33a 1681 globaldata_t gd;
99df837e 1682
2883d2d8
MD
1683 /*
1684 * Do any cleanup that might block here
1685 */
99df837e 1686 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1687 kprintf("kthread %p %s has exited\n", td, td->td_comm);
2883d2d8
MD
1688 biosched_done(td);
1689 dsched_exit_thread(td);
c070746a
MD
1690
1691 /*
1692 * Get us into a critical section to interlock gd_freetd and loop
1693 * until we can get it freed.
1694 *
1695 * We have to cache the current td in gd_freetd because objcache_put()ing
1696 * it would rip it out from under us while our thread is still active.
2af9d75d
MD
1697 *
1698 * We are the current thread so of course our own TDF_RUNNING bit will
1699 * be set, so unlike the lwp reap code we don't wait for it to clear.
c070746a
MD
1700 */
1701 gd = mycpu;
37af14fe 1702 crit_enter_quick(td);
2af9d75d
MD
1703 for (;;) {
1704 if (td->td_refs) {
1705 tsleep(td, 0, "tdreap", 1);
1706 continue;
1707 }
1708 if ((std = gd->gd_freetd) != NULL) {
1709 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1710 gd->gd_freetd = NULL;
1711 objcache_put(thread_cache, std);
1712 continue;
1713 }
1714 break;
c070746a 1715 }
3b4192fb
MD
1716
1717 /*
1718 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1719 * the last time. We cannot block after this point or we may end
1720 * up with a stale td on the tsleepq.
eb2adbf5
MD
1721 *
1722 * None of this may block, the critical section is the only thing
1723 * protecting tdallq and the only thing preventing new lwkt_hold()
1724 * thread refs now.
3b4192fb
MD
1725 */
1726 if (td->td_flags & TDF_TSLEEPQ)
1727 tsleep_remove(td);
37af14fe 1728 lwkt_deschedule_self(td);
e56e4dea 1729 lwkt_remove_tdallq(td);
74c9628e 1730 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1731
1732 /*
1733 * Final cleanup
1734 */
1735 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1736 if (td->td_flags & TDF_ALLOCATED_THREAD)
1737 gd->gd_freetd = td;
99df837e
MD
1738 cpu_thread_exit();
1739}
1740
e56e4dea
MD
1741void
1742lwkt_remove_tdallq(thread_t td)
1743{
1744 KKASSERT(td->td_gd == mycpu);
1745 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1746}
1747
9cf43f91
MD
1748/*
1749 * Code reduction and branch prediction improvements. Call/return
1750 * overhead on modern cpus often degenerates into 0 cycles due to
1751 * the cpu's branch prediction hardware and return pc cache. We
1752 * can take advantage of this by not inlining medium-complexity
1753 * functions and we can also reduce the branch prediction impact
1754 * by collapsing perfectly predictable branches into a single
1755 * procedure instead of duplicating it.
1756 *
1757 * Is any of this noticeable? Probably not, so I'll take the
1758 * smaller code size.
1759 */
1760void
b6468f56 1761crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1762{
b6468f56 1763 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1764}
1765
2d93b37a
MD
1766void
1767crit_panic(void)
1768{
1769 thread_t td = curthread;
850634cc 1770 int lcrit = td->td_critcount;
2d93b37a 1771
850634cc
AH
1772 td->td_critcount = 0;
1773 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1774 /* NOT REACHED */
2d93b37a
MD
1775}
1776
bd8015ca
MD
1777/*
1778 * Called from debugger/panic on cpus which have been stopped. We must still
1779 * process the IPIQ while stopped, even if we were stopped while in a critical
1780 * section (XXX).
1781 *
1782 * If we are dumping also try to process any pending interrupts. This may
1783 * or may not work depending on the state of the cpu at the point it was
1784 * stopped.
1785 */
1786void
1787lwkt_smp_stopped(void)
1788{
1789 globaldata_t gd = mycpu;
1790
1791 crit_enter_gd(gd);
1792 if (dumping) {
1793 lwkt_process_ipiq();
1794 splz();
1795 } else {
1796 lwkt_process_ipiq();
1797 }
1798 crit_exit_gd(gd);
1799}