objcache: Cache align magazinedepot and percpu_objcache
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
b12defdc 2 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
60f60350 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
60f60350 20 *
8c10bfcf
MD
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
b37f18d6 47#include <sys/kinfo.h>
8ad65e08 48#include <sys/queue.h>
7d0bac62 49#include <sys/sysctl.h>
99df837e 50#include <sys/kthread.h>
f1d1c3fa 51#include <machine/cpu.h>
99df837e 52#include <sys/lock.h>
f6bf3af1 53#include <sys/caps.h>
9d265729 54#include <sys/spinlock.h>
57aa743c 55#include <sys/ktr.h>
9d265729
MD
56
57#include <sys/thread2.h>
58#include <sys/spinlock2.h>
684a93c4 59#include <sys/mplock2.h>
f1d1c3fa 60
8c72e3d5
AH
61#include <sys/dsched.h>
62
7d0bac62
MD
63#include <vm/vm.h>
64#include <vm/vm_param.h>
65#include <vm/vm_kern.h>
66#include <vm/vm_object.h>
67#include <vm/vm_page.h>
68#include <vm/vm_map.h>
69#include <vm/vm_pager.h>
70#include <vm/vm_extern.h>
7d0bac62 71
99df837e 72#include <machine/stdarg.h>
96728c05 73#include <machine/smp.h>
99df837e 74
d850923c
AE
75#if !defined(KTR_CTXSW)
76#define KTR_CTXSW KTR_ALL
77#endif
78KTR_INFO_MASTER(ctxsw);
5bf48697
AE
79KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td);
80KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td);
81KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm);
82KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td);
1541028a 83
40aaf5fc
NT
84static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
85
0f7a3396
MD
86#ifdef INVARIANTS
87static int panic_on_cscount = 0;
88#endif
05220613
MD
89static __int64_t switch_count = 0;
90static __int64_t preempt_hit = 0;
91static __int64_t preempt_miss = 0;
92static __int64_t preempt_weird = 0;
fb0f29c4 93static int lwkt_use_spin_port;
40aaf5fc 94static struct objcache *thread_cache;
05220613 95
88ebb169 96#ifdef SMP
e381e77c 97static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
cc9b6223 98static void lwkt_setcpu_remote(void *arg);
88ebb169 99#endif
e381e77c 100
0855a2af
JG
101extern void cpu_heavy_restore(void);
102extern void cpu_lwkt_restore(void);
103extern void cpu_kthread_restore(void);
104extern void cpu_idle_restore(void);
105
fb0f29c4
MD
106/*
107 * We can make all thread ports use the spin backend instead of the thread
108 * backend. This should only be set to debug the spin backend.
109 */
110TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
111
0f7a3396 112#ifdef INVARIANTS
0c52fa62
SG
113SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
114 "Panic if attempting to switch lwkt's while mastering cpusync");
0f7a3396 115#endif
0c52fa62
SG
116SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
117 "Number of switched threads");
9733f757 118SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
0c52fa62 119 "Successful preemption events");
9733f757 120SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
0c52fa62
SG
121 "Failed preemption events");
122SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
123 "Number of preempted threads.");
b12defdc 124static int fairq_enable = 0;
2a418930
MD
125SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
126 &fairq_enable, 0, "Turn on fairq priority accumulators");
85946b6c 127static int fairq_bypass = -1;
b12defdc
MD
128SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
129 &fairq_bypass, 0, "Allow fairq to bypass td on token failure");
130extern int lwkt_sched_debug;
131int lwkt_sched_debug = 0;
132SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
133 &lwkt_sched_debug, 0, "Scheduler debug");
2a418930
MD
134static int lwkt_spin_loops = 10;
135SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
b12defdc
MD
136 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
137static int lwkt_spin_reseq = 0;
138SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW,
139 &lwkt_spin_reseq, 0, "Scheduler resequencer enable");
140static int lwkt_spin_monitor = 0;
141SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW,
142 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait");
d5b2d319
MD
143static int lwkt_spin_fatal = 0; /* disabled */
144SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
145 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
fbc024e4 146static int preempt_enable = 1;
2a418930
MD
147SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
148 &preempt_enable, 0, "Enable preemption");
7b234d8c 149static int lwkt_cache_threads = 0;
765b1ae0
MD
150SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
151 &lwkt_cache_threads, 0, "thread+kstack cache");
fbc024e4 152
2a418930
MD
153static __cachealign int lwkt_cseq_rindex;
154static __cachealign int lwkt_cseq_windex;
05220613 155
4b5f931b
MD
156/*
157 * These helper procedures handle the runq, they can only be called from
158 * within a critical section.
75cdbe6c
MD
159 *
160 * WARNING! Prior to SMP being brought up it is possible to enqueue and
161 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
162 * instead of 'mycpu' when referencing the globaldata structure. Once
163 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 164 */
f1d1c3fa
MD
165static __inline
166void
167_lwkt_dequeue(thread_t td)
168{
169 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 170 struct globaldata *gd = td->td_gd;
4b5f931b 171
f1d1c3fa 172 td->td_flags &= ~TDF_RUNQ;
f9235b6d 173 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
de4d4cb0 174 --gd->gd_tdrunqcount;
f9235b6d 175 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
2a418930 176 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
177 }
178}
179
f9235b6d
MD
180/*
181 * Priority enqueue.
182 *
d992c377
MD
183 * There are a limited number of lwkt threads runnable since user
184 * processes only schedule one at a time per cpu. However, there can
185 * be many user processes in kernel mode exiting from a tsleep() which
e3e6be1f 186 * become runnable.
d992c377
MD
187 *
188 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and
189 * will ignore user priority. This is to ensure that user threads in
190 * kernel mode get cpu at some point regardless of what the user
191 * scheduler thinks.
f9235b6d 192 */
f1d1c3fa
MD
193static __inline
194void
195_lwkt_enqueue(thread_t td)
196{
f9235b6d
MD
197 thread_t xtd;
198
7f5d7ed7 199 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 200 struct globaldata *gd = td->td_gd;
4b5f931b 201
f1d1c3fa 202 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
203 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
204 if (xtd == NULL) {
85946b6c
MD
205 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
206 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
f9235b6d 207 } else {
e3e6be1f
MD
208 /*
209 * NOTE: td_upri - higher numbers more desireable, same sense
210 * as td_pri (typically reversed from lwp_upri).
211 *
212 * In the equal priority case we want the best selection
213 * at the beginning so the less desireable selections know
214 * that they have to setrunqueue/go-to-another-cpu, even
215 * though it means switching back to the 'best' selection.
216 * This also avoids degenerate situations when many threads
217 * are runnable or waking up at the same time.
218 *
219 * If upri matches exactly place at end/round-robin.
220 */
d992c377 221 while (xtd &&
e3e6be1f 222 (xtd->td_pri >= td->td_pri ||
d992c377 223 (xtd->td_pri == td->td_pri &&
e3e6be1f 224 xtd->td_upri >= td->td_upri))) {
85946b6c 225 xtd = TAILQ_NEXT(xtd, td_threadq);
d992c377 226 }
85946b6c
MD
227 if (xtd)
228 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
229 else
230 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
f9235b6d 231 }
de4d4cb0 232 ++gd->gd_tdrunqcount;
b12defdc
MD
233
234 /*
85946b6c 235 * Request a LWKT reschedule if we are now at the head of the queue.
b12defdc 236 */
85946b6c
MD
237 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
238 need_lwkt_resched();
f1d1c3fa
MD
239 }
240}
8ad65e08 241
40aaf5fc
NT
242static __boolean_t
243_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
244{
245 struct thread *td = (struct thread *)obj;
246
247 td->td_kstack = NULL;
248 td->td_kstack_size = 0;
249 td->td_flags = TDF_ALLOCATED_THREAD;
4643740a 250 td->td_mpflags = 0;
40aaf5fc
NT
251 return (1);
252}
253
254static void
255_lwkt_thread_dtor(void *obj, void *privdata)
256{
257 struct thread *td = (struct thread *)obj;
258
259 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
260 ("_lwkt_thread_dtor: not allocated from objcache"));
261 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
262 td->td_kstack_size > 0,
263 ("_lwkt_thread_dtor: corrupted stack"));
264 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
7b234d8c
MD
265 td->td_kstack = NULL;
266 td->td_flags = 0;
40aaf5fc
NT
267}
268
269/*
270 * Initialize the lwkt s/system.
765b1ae0 271 *
7b234d8c
MD
272 * Nominally cache up to 32 thread + kstack structures. Cache more on
273 * systems with a lot of cpu cores.
40aaf5fc
NT
274 */
275void
276lwkt_init(void)
277{
765b1ae0 278 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
7b234d8c
MD
279 if (lwkt_cache_threads == 0) {
280 lwkt_cache_threads = ncpus * 4;
281 if (lwkt_cache_threads < 32)
282 lwkt_cache_threads = 32;
283 }
765b1ae0
MD
284 thread_cache = objcache_create_mbacked(
285 M_THREAD, sizeof(struct thread),
286 NULL, lwkt_cache_threads,
287 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
288}
289
37af14fe
MD
290/*
291 * Schedule a thread to run. As the current thread we can always safely
292 * schedule ourselves, and a shortcut procedure is provided for that
293 * function.
294 *
295 * (non-blocking, self contained on a per cpu basis)
296 */
297void
298lwkt_schedule_self(thread_t td)
299{
cfaeae2a 300 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 301 crit_enter_quick(td);
f9235b6d
MD
302 KASSERT(td != &td->td_gd->gd_idlethread,
303 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
4643740a
MD
304 KKASSERT(td->td_lwp == NULL ||
305 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
37af14fe 306 _lwkt_enqueue(td);
37af14fe
MD
307 crit_exit_quick(td);
308}
309
310/*
311 * Deschedule a thread.
312 *
313 * (non-blocking, self contained on a per cpu basis)
314 */
315void
316lwkt_deschedule_self(thread_t td)
317{
318 crit_enter_quick(td);
37af14fe
MD
319 _lwkt_dequeue(td);
320 crit_exit_quick(td);
321}
322
8ad65e08
MD
323/*
324 * LWKTs operate on a per-cpu basis
325 *
73e4f7b9 326 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
327 */
328void
329lwkt_gdinit(struct globaldata *gd)
330{
f9235b6d 331 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 332 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
333}
334
7d0bac62
MD
335/*
336 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
337 * or LWKT start address before it can be scheduled. If the target cpu is
338 * -1 the thread will be created on the current cpu.
0cfcada1
MD
339 *
340 * If you intend to create a thread without a process context this function
341 * does everything except load the startup and switcher function.
7d0bac62
MD
342 */
343thread_t
d3d32139 344lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 345{
d2d8515b 346 static int cpu_rotator;
c070746a 347 globaldata_t gd = mycpu;
99df837e 348 void *stack;
7d0bac62 349
c070746a
MD
350 /*
351 * If static thread storage is not supplied allocate a thread. Reuse
352 * a cached free thread if possible. gd_freetd is used to keep an exiting
353 * thread intact through the exit.
354 */
ef0fdad1 355 if (td == NULL) {
cf709dd2
MD
356 crit_enter_gd(gd);
357 if ((td = gd->gd_freetd) != NULL) {
358 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
359 TDF_RUNQ)) == 0);
c070746a 360 gd->gd_freetd = NULL;
cf709dd2 361 } else {
c070746a 362 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
363 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
364 TDF_RUNQ)) == 0);
365 }
366 crit_exit_gd(gd);
40aaf5fc 367 KASSERT((td->td_flags &
2af9d75d
MD
368 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
369 TDF_ALLOCATED_THREAD,
40aaf5fc
NT
370 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
371 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 372 }
c070746a
MD
373
374 /*
375 * Try to reuse cached stack.
376 */
f470d0c8
MD
377 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
378 if (flags & TDF_ALLOCATED_STACK) {
e4846942 379 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
380 stack = NULL;
381 }
382 }
383 if (stack == NULL) {
e40cfbd7 384 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 385 flags |= TDF_ALLOCATED_STACK;
99df837e 386 }
d2d8515b
MD
387 if (cpu < 0) {
388 cpu = ++cpu_rotator;
389 cpu_ccfence();
390 cpu %= ncpus;
391 }
392 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 393 return(td);
7d0bac62
MD
394}
395
396/*
397 * Initialize a preexisting thread structure. This function is used by
398 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
399 *
f8c3996b
MD
400 * All threads start out in a critical section at a priority of
401 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
402 * appropriate. This function may send an IPI message when the
403 * requested cpu is not the current cpu and consequently gd_tdallq may
404 * not be initialized synchronously from the point of view of the originating
405 * cpu.
406 *
407 * NOTE! we have to be careful in regards to creating threads for other cpus
408 * if SMP has not yet been activated.
7d0bac62 409 */
41a01a4d
MD
410#ifdef SMP
411
75cdbe6c
MD
412static void
413lwkt_init_thread_remote(void *arg)
414{
415 thread_t td = arg;
416
52eedfb5
MD
417 /*
418 * Protected by critical section held by IPI dispatch
419 */
75cdbe6c
MD
420 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
421}
422
41a01a4d
MD
423#endif
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;
26a0694b 441 td->td_gd = gd;
f9235b6d
MD
442 td->td_pri = TDPRI_KERN_DAEMON;
443 td->td_critcount = 1;
54341a3b 444 td->td_toks_have = NULL;
3b998fa9 445 td->td_toks_stop = &td->td_toks_base;
392cd266 446 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
d636cd48 447 lwkt_initport_spin(&td->td_msgport, td);
fb0f29c4
MD
448 else
449 lwkt_initport_thread(&td->td_msgport, td);
99df837e 450 pmap_init_thread(td);
0f7a3396 451#ifdef SMP
5d21b981
MD
452 /*
453 * Normally initializing a thread for a remote cpu requires sending an
454 * IPI. However, the idlethread is setup before the other cpus are
455 * activated so we have to treat it as a special case. XXX manipulation
456 * of gd_tdallq requires the BGL.
457 */
458 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 459 crit_enter_gd(mygd);
75cdbe6c 460 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 461 crit_exit_gd(mygd);
75cdbe6c 462 } else {
2db3b277 463 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 464 }
0f7a3396 465#else
37af14fe 466 crit_enter_gd(mygd);
0f7a3396 467 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 468 crit_exit_gd(mygd);
0f7a3396 469#endif
8c72e3d5
AH
470
471 dsched_new_thread(td);
73e4f7b9
MD
472}
473
474void
475lwkt_set_comm(thread_t td, const char *ctl, ...)
476{
e2565a42 477 __va_list va;
73e4f7b9 478
e2565a42 479 __va_start(va, ctl);
379210cb 480 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 481 __va_end(va);
5bf48697 482 KTR_LOG(ctxsw_newtd, td, td->td_comm);
7d0bac62
MD
483}
484
eb2adbf5
MD
485/*
486 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
487 * this does not prevent the thread from migrating to another cpu so the
488 * gd_tdallq state is not protected by this.
489 */
99df837e 490void
73e4f7b9 491lwkt_hold(thread_t td)
99df837e 492{
74c9628e 493 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
494}
495
496void
497lwkt_rele(thread_t td)
498{
499 KKASSERT(td->td_refs > 0);
74c9628e 500 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
501}
502
73e4f7b9
MD
503void
504lwkt_free_thread(thread_t td)
505{
74c9628e 506 KKASSERT(td->td_refs == 0);
c17a6852
MD
507 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
508 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
40aaf5fc
NT
509 if (td->td_flags & TDF_ALLOCATED_THREAD) {
510 objcache_put(thread_cache, td);
511 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
512 /* client-allocated struct with internally allocated stack */
513 KASSERT(td->td_kstack && td->td_kstack_size > 0,
514 ("lwkt_free_thread: corrupted stack"));
515 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
516 td->td_kstack = NULL;
517 td->td_kstack_size = 0;
99df837e 518 }
e7c0dbba 519 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
520}
521
522
8ad65e08
MD
523/*
524 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
525 * switch to the idlethread. Switching must occur within a critical
526 * section to avoid races with the scheduling queue.
527 *
528 * We always have full control over our cpu's run queue. Other cpus
529 * that wish to manipulate our queue must use the cpu_*msg() calls to
530 * talk to our cpu, so a critical section is all that is needed and
531 * the result is very, very fast thread switching.
532 *
96728c05
MD
533 * The LWKT scheduler uses a fixed priority model and round-robins at
534 * each priority level. User process scheduling is a totally
535 * different beast and LWKT priorities should not be confused with
536 * user process priorities.
f1d1c3fa 537 *
69d78e99
MD
538 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
539 * is not called by the current thread in the preemption case, only when
540 * the preempting thread blocks (in order to return to the original thread).
cfaeae2a
MD
541 *
542 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
543 * migration and tsleep deschedule the current lwkt thread and call
544 * lwkt_switch(). In particular, the target cpu of the migration fully
545 * expects the thread to become non-runnable and can deadlock against
546 * cpusync operations if we run any IPIs prior to switching the thread out.
547 *
548 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
95858b91 549 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
8ad65e08
MD
550 */
551void
552lwkt_switch(void)
553{
37af14fe
MD
554 globaldata_t gd = mycpu;
555 thread_t td = gd->gd_curthread;
8ad65e08 556 thread_t ntd;
b12defdc 557 int spinning = 0;
8ad65e08 558
da0b0e8b 559 KKASSERT(gd->gd_processing_ipiq == 0);
121f93bc 560 KKASSERT(td->td_flags & TDF_RUNNING);
da0b0e8b 561
46a3f46d 562 /*
27e88a6e
MD
563 * Switching from within a 'fast' (non thread switched) interrupt or IPI
564 * is illegal. However, we may have to do it anyway if we hit a fatal
565 * kernel trap or we have paniced.
566 *
567 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 568 */
27e88a6e
MD
569 if (gd->gd_intr_nesting_level) {
570 int savegdnest;
571 int savegdtrap;
572
5fddbda2 573 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22 574 panic("lwkt_switch: Attempt to switch from a "
5a8df152 575 "fast interrupt, ipi, or hard code section, "
4a28fe22
MD
576 "td %p\n",
577 td);
27e88a6e
MD
578 } else {
579 savegdnest = gd->gd_intr_nesting_level;
580 savegdtrap = gd->gd_trap_nesting_level;
581 gd->gd_intr_nesting_level = 0;
582 gd->gd_trap_nesting_level = 0;
a7422615
MD
583 if ((td->td_flags & TDF_PANICWARN) == 0) {
584 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
585 kprintf("Warning: thread switch from interrupt, IPI, "
586 "or hard code section.\n"
a7422615 587 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 588 print_backtrace(-1);
a7422615 589 }
27e88a6e
MD
590 lwkt_switch();
591 gd->gd_intr_nesting_level = savegdnest;
592 gd->gd_trap_nesting_level = savegdtrap;
593 return;
594 }
96728c05 595 }
ef0fdad1 596
cb973d15 597 /*
85946b6c
MD
598 * Release our current user process designation if we are blocking
599 * or if a user reschedule was requested.
600 *
601 * NOTE: This function is NOT called if we are switching into or
602 * returning from a preemption.
603 *
604 * NOTE: Releasing our current user process designation may cause
605 * it to be assigned to another thread, which in turn will
606 * cause us to block in the usched acquire code when we attempt
607 * to return to userland.
608 *
609 * NOTE: On SMP systems this can be very nasty when heavy token
610 * contention is present so we want to be careful not to
611 * release the designation gratuitously.
cb973d15 612 */
85946b6c
MD
613 if (td->td_release &&
614 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
cb973d15 615 td->td_release(td);
85946b6c 616 }
cb973d15 617
85946b6c
MD
618 /*
619 * Release all tokens
620 */
37af14fe 621 crit_enter_gd(gd);
3b998fa9 622 if (TD_TOKS_HELD(td))
9d265729
MD
623 lwkt_relalltokens(td);
624
625 /*
b02926de
MD
626 * We had better not be holding any spin locks, but don't get into an
627 * endless panic loop.
9d265729 628 */
0846e4ce 629 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL,
d666840a 630 ("lwkt_switch: still holding %d exclusive spinlocks!",
0846e4ce 631 gd->gd_spinlocks));
9d265729 632
8a8d5d85
MD
633
634#ifdef SMP
0f7a3396
MD
635#ifdef INVARIANTS
636 if (td->td_cscount) {
6ea70f76 637 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
638 td);
639 if (panic_on_cscount)
640 panic("switching while mastering cpusync");
641 }
642#endif
8a8d5d85 643#endif
f9235b6d
MD
644
645 /*
646 * If we had preempted another thread on this cpu, resume the preempted
647 * thread. This occurs transparently, whether the preempted thread
648 * was scheduled or not (it may have been preempted after descheduling
649 * itself).
650 *
651 * We have to setup the MP lock for the original thread after backing
652 * out the adjustment that was made to curthread when the original
653 * was preempted.
654 */
99df837e 655 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
656 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
657 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
658
659 /*
b9eb1c19
MD
660 * The interrupt may have woken a thread up, we need to properly
661 * set the reschedule flag if the originally interrupted thread is
662 * at a lower priority.
85946b6c
MD
663 *
664 * The interrupt may not have descheduled.
8ec60c3f 665 */
85946b6c 666 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
8ec60c3f 667 need_lwkt_resched();
f9235b6d
MD
668 goto havethread_preempted;
669 }
670
b12defdc 671 /*
f9235b6d 672 * If we cannot obtain ownership of the tokens we cannot immediately
cfaeae2a
MD
673 * schedule the target thread.
674 *
675 * Reminder: Again, we cannot afford to run any IPIs in this path if
676 * the current thread has been descheduled.
f9235b6d
MD
677 */
678 for (;;) {
b12defdc 679 clear_lwkt_resched();
f9235b6d 680
4b5f931b 681 /*
2a418930 682 * Hotpath - pull the head of the run queue and attempt to schedule
85946b6c 683 * it.
41a01a4d 684 */
7c3dd309
NA
685 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
686
687 if (ntd == NULL) {
688 /*
689 * Runq is empty, switch to idle to allow it to halt.
690 */
691 ntd = &gd->gd_idlethread;
6f207a2c 692#ifdef SMP
7c3dd309
NA
693 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
694 ASSERT_NO_TOKENS_HELD(ntd);
6f207a2c 695#endif
7c3dd309
NA
696 cpu_time.cp_msg[0] = 0;
697 cpu_time.cp_stallpc = 0;
698 goto haveidle;
f9235b6d 699 }
41a01a4d 700
8ec60c3f 701 /*
b12defdc 702 * Hotpath - schedule ntd.
6f207a2c
MD
703 *
704 * NOTE: For UP there is no mplock and lwkt_getalltokens()
705 * always succeeds.
8ec60c3f 706 */
b12defdc
MD
707 if (TD_TOKS_NOT_HELD(ntd) ||
708 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
709 {
f9235b6d 710 goto havethread;
b12defdc 711 }
f9235b6d 712
f9235b6d 713 /*
2a418930
MD
714 * Coldpath (SMP only since tokens always succeed on UP)
715 *
716 * We had some contention on the thread we wanted to schedule.
717 * What we do now is try to find a thread that we can schedule
b12defdc 718 * in its stead.
2a418930 719 *
85946b6c
MD
720 * The coldpath scan does NOT rearrange threads in the run list.
721 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
722 * the next tick whenever the current head is not the current thread.
f9235b6d 723 */
b12defdc 724#ifdef INVARIANTS
85946b6c 725 ++ntd->td_contended;
b12defdc 726#endif
574339e6 727 ++gd->gd_cnt.v_token_colls;
b12defdc 728
85946b6c 729 if (fairq_bypass > 0)
b12defdc
MD
730 goto skip;
731
b12defdc 732 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
e3e6be1f 733#ifndef NO_LWKT_SPLIT_USERPRI
85946b6c
MD
734 /*
735 * Never schedule threads returning to userland or the
736 * user thread scheduler helper thread when higher priority
bb6860b5
NA
737 * threads are present. The runq is sorted by priority
738 * so we can give up traversing it when we find the first
739 * low priority thread.
85946b6c
MD
740 */
741 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
742 ntd = NULL;
743 break;
744 }
d992c377 745#endif
85946b6c
MD
746
747 /*
748 * Try this one.
749 */
b12defdc
MD
750 if (TD_TOKS_NOT_HELD(ntd) ||
751 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
752 goto havethread;
753 }
85946b6c 754#ifdef INVARIANTS
85946b6c 755 ++ntd->td_contended;
b12defdc 756#endif
574339e6 757 ++gd->gd_cnt.v_token_colls;
2a418930
MD
758 }
759
b12defdc 760skip:
2a418930
MD
761 /*
762 * We exhausted the run list, meaning that all runnable threads
b12defdc 763 * are contested.
2a418930
MD
764 */
765 cpu_pause();
766 ntd = &gd->gd_idlethread;
767#ifdef SMP
768 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
769 ASSERT_NO_TOKENS_HELD(ntd);
770 /* contention case, do not clear contention mask */
771#endif
772
773 /*
b12defdc
MD
774 * We are going to have to retry but if the current thread is not
775 * on the runq we instead switch through the idle thread to get away
776 * from the current thread. We have to flag for lwkt reschedule
777 * to prevent the idle thread from halting.
2a418930 778 *
b12defdc
MD
779 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
780 * instruct it to deal with the potential for deadlocks by
781 * ordering the tokens by address.
2a418930 782 */
b12defdc 783 if ((td->td_flags & TDF_RUNQ) == 0) {
85946b6c 784 need_lwkt_resched(); /* prevent hlt */
2a418930 785 goto haveidle;
4b5f931b 786 }
bd52bedf 787#if defined(INVARIANTS) && defined(__amd64__)
d5b2d319
MD
788 if ((read_rflags() & PSL_I) == 0) {
789 cpu_enable_intr();
790 panic("lwkt_switch() called with interrupts disabled");
791 }
792#endif
b12defdc
MD
793
794 /*
795 * Number iterations so far. After a certain point we switch to
796 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
797 */
798 if (spinning < 0x7FFFFFFF)
799 ++spinning;
800
801#ifdef SMP
802 /*
803 * lwkt_getalltokens() failed in sorted token mode, we can use
804 * monitor/mwait in this case.
805 */
806 if (spinning >= lwkt_spin_loops &&
807 (cpu_mi_feature & CPU_MI_MONITOR) &&
808 lwkt_spin_monitor)
809 {
810 cpu_mmw_pause_int(&gd->gd_reqflags,
811 (gd->gd_reqflags | RQF_SPINNING) &
812 ~RQF_IDLECHECK_WK_MASK);
813 }
814#endif
815
816 /*
817 * We already checked that td is still scheduled so this should be
818 * safe.
819 */
820 splz_check();
821
822 /*
823 * This experimental resequencer is used as a fall-back to reduce
824 * hw cache line contention by placing each core's scheduler into a
825 * time-domain-multplexed slot.
826 *
827 * The resequencer is disabled by default. It's functionality has
828 * largely been superceeded by the token algorithm which limits races
829 * to a subset of cores.
830 *
831 * The resequencer algorithm tends to break down when more than
832 * 20 cores are contending. What appears to happen is that new
833 * tokens can be obtained out of address-sorted order by new cores
834 * while existing cores languish in long delays between retries and
835 * wind up being starved-out of the token acquisition.
836 */
837 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
838 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
839 int oseq;
840
841 while ((oseq = lwkt_cseq_rindex) != cseq) {
842 cpu_ccfence();
843#if 1
844 if (cpu_mi_feature & CPU_MI_MONITOR) {
845 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
846 } else {
847#endif
848 cpu_pause();
849 cpu_lfence();
850#if 1
851 }
8b402283 852#endif
0f0466c0 853 }
b12defdc
MD
854 DELAY(1);
855 atomic_add_int(&lwkt_cseq_rindex, 1);
2a418930 856 }
2a418930 857 /* highest level for(;;) loop */
f1d1c3fa 858 }
8a8d5d85 859
2a418930 860havethread:
b12defdc 861 /*
be71787b
MD
862 * Clear gd_idle_repeat when doing a normal switch to a non-idle
863 * thread.
f9235b6d 864 */
9ac1ee6e 865 ntd->td_wmesg = NULL;
b12defdc 866 ++gd->gd_cnt.v_swtch;
be71787b 867 gd->gd_idle_repeat = 0;
2a418930 868
f9235b6d 869havethread_preempted:
f9235b6d
MD
870 /*
871 * If the new target does not need the MP lock and we are holding it,
872 * release the MP lock. If the new target requires the MP lock we have
873 * already acquired it for the target.
8a8d5d85 874 */
2a418930 875 ;
f9235b6d
MD
876haveidle:
877 KASSERT(ntd->td_critcount,
b5d16701
MD
878 ("priority problem in lwkt_switch %d %d",
879 td->td_critcount, ntd->td_critcount));
880
94f6d86e 881 if (td != ntd) {
cc9b6223
MD
882 /*
883 * Execute the actual thread switch operation. This function
884 * returns to the current thread and returns the previous thread
885 * (which may be different from the thread we switched to).
886 *
887 * We are responsible for marking ntd as TDF_RUNNING.
888 */
121f93bc 889 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
94f6d86e 890 ++switch_count;
a1f0fb66 891 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
892 ntd->td_flags |= TDF_RUNNING;
893 lwkt_switch_return(td->td_switch(ntd));
894 /* ntd invalid, td_switch() can return a different thread_t */
94f6d86e 895 }
b12defdc 896
b12defdc 897 /*
54341a3b 898 * catch-all. XXX is this strictly needed?
b12defdc
MD
899 */
900 splz_check();
54341a3b 901
37af14fe
MD
902 /* NOTE: current cpu may have changed after switch */
903 crit_exit_quick(td);
8ad65e08
MD
904}
905
cc9b6223
MD
906/*
907 * Called by assembly in the td_switch (thread restore path) for thread
908 * bootstrap cases which do not 'return' to lwkt_switch().
909 */
910void
911lwkt_switch_return(thread_t otd)
912{
913#ifdef SMP
914 globaldata_t rgd;
915
916 /*
917 * Check if otd was migrating. Now that we are on ntd we can finish
918 * up the migration. This is a bit messy but it is the only place
919 * where td is known to be fully descheduled.
920 *
921 * We can only activate the migration if otd was migrating but not
922 * held on the cpu due to a preemption chain. We still have to
923 * clear TDF_RUNNING on the old thread either way.
924 *
925 * We are responsible for clearing the previously running thread's
926 * TDF_RUNNING.
927 */
928 if ((rgd = otd->td_migrate_gd) != NULL &&
929 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
930 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
931 (TDF_MIGRATING | TDF_RUNNING));
932 otd->td_migrate_gd = NULL;
933 otd->td_flags &= ~TDF_RUNNING;
934 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
935 } else {
936 otd->td_flags &= ~TDF_RUNNING;
937 }
938#else
939 otd->td_flags &= ~TDF_RUNNING;
940#endif
2b07d9aa
MD
941
942 /*
943 * Final exit validations (see lwp_wait()). Note that otd becomes
944 * invalid the *instant* we set TDF_MP_EXITSIG.
945 */
946 while (otd->td_flags & TDF_EXITING) {
947 u_int mpflags;
948
949 mpflags = otd->td_mpflags;
950 cpu_ccfence();
951
952 if (mpflags & TDF_MP_EXITWAIT) {
953 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
954 mpflags | TDF_MP_EXITSIG)) {
955 wakeup(otd);
956 break;
957 }
958 } else {
959 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
960 mpflags | TDF_MP_EXITSIG)) {
961 wakeup(otd);
962 break;
963 }
964 }
965 }
cc9b6223
MD
966}
967
b68b7282 968/*
96728c05 969 * Request that the target thread preempt the current thread. Preemption
54341a3b
MD
970 * can only occur if our only critical section is the one that we were called
971 * with, the relative priority of the target thread is higher, and the target
972 * thread holds no tokens. This also only works if we are not holding any
973 * spinlocks (obviously).
96728c05
MD
974 *
975 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
976 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 977 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
978 * to determine whether preemption is possible (aka usually just the crit
979 * priority of lwkt_schedule() itself).
b68b7282 980 *
54341a3b
MD
981 * Preemption is typically limited to interrupt threads.
982 *
983 * Operation works in a fairly straight-forward manner. The normal
984 * scheduling code is bypassed and we switch directly to the target
985 * thread. When the target thread attempts to block or switch away
986 * code at the base of lwkt_switch() will switch directly back to our
987 * thread. Our thread is able to retain whatever tokens it holds and
988 * if the target needs one of them the target will switch back to us
989 * and reschedule itself normally.
b68b7282
MD
990 */
991void
f9235b6d 992lwkt_preempt(thread_t ntd, int critcount)
b68b7282 993{
46a3f46d 994 struct globaldata *gd = mycpu;
cc9b6223 995 thread_t xtd;
0a3f9b47 996 thread_t td;
2d910aaf 997 int save_gd_intr_nesting_level;
b68b7282 998
26a0694b 999 /*
96728c05
MD
1000 * The caller has put us in a critical section. We can only preempt
1001 * if the caller of the caller was not in a critical section (basically
f9235b6d 1002 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 1003 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 1004 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
1005 *
1006 * YYY The target thread must be in a critical section (else it must
1007 * inherit our critical section? I dunno yet).
26a0694b 1008 */
f9235b6d 1009 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 1010
b12defdc 1011 td = gd->gd_curthread;
fbc024e4
MD
1012 if (preempt_enable == 0) {
1013 ++preempt_miss;
1014 return;
1015 }
f9235b6d 1016 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
1017 ++preempt_miss;
1018 return;
1019 }
f9235b6d 1020 if (td->td_critcount > critcount) {
96728c05
MD
1021 ++preempt_miss;
1022 return;
1023 }
1024#ifdef SMP
121f93bc
MD
1025 if (td->td_cscount) {
1026 ++preempt_miss;
1027 return;
1028 }
46a3f46d 1029 if (ntd->td_gd != gd) {
96728c05
MD
1030 ++preempt_miss;
1031 return;
1032 }
1033#endif
41a01a4d 1034 /*
77912481
MD
1035 * We don't have to check spinlocks here as they will also bump
1036 * td_critcount.
d3d1cbc8
MD
1037 *
1038 * Do not try to preempt if the target thread is holding any tokens.
1039 * We could try to acquire the tokens but this case is so rare there
1040 * is no need to support it.
41a01a4d 1041 */
0846e4ce 1042 KKASSERT(gd->gd_spinlocks == 0);
77912481 1043
3b998fa9 1044 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8 1045 ++preempt_miss;
d3d1cbc8
MD
1046 return;
1047 }
26a0694b
MD
1048 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1049 ++preempt_weird;
1050 return;
1051 }
1052 if (ntd->td_preempted) {
4b5f931b 1053 ++preempt_hit;
26a0694b 1054 return;
b68b7282 1055 }
da0b0e8b 1056 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 1057
8ec60c3f
MD
1058 /*
1059 * Since we are able to preempt the current thread, there is no need to
1060 * call need_lwkt_resched().
2d910aaf
MD
1061 *
1062 * We must temporarily clear gd_intr_nesting_level around the switch
1063 * since switchouts from the target thread are allowed (they will just
1064 * return to our thread), and since the target thread has its own stack.
cc9b6223
MD
1065 *
1066 * A preemption must switch back to the original thread, assert the
1067 * case.
8ec60c3f 1068 */
26a0694b
MD
1069 ++preempt_hit;
1070 ntd->td_preempted = td;
1071 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1072 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1073 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1074 gd->gd_intr_nesting_level = 0;
121f93bc
MD
1075
1076 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
cc9b6223
MD
1077 ntd->td_flags |= TDF_RUNNING;
1078 xtd = td->td_switch(ntd);
1079 KKASSERT(xtd == ntd);
1080 lwkt_switch_return(xtd);
2d910aaf 1081 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1082
26a0694b
MD
1083 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1084 ntd->td_preempted = NULL;
1085 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1086}
1087
f1d1c3fa 1088/*
faaeffac 1089 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1090 * This will work inside a critical section but not inside a hard code
1091 * section.
ef0fdad1 1092 *
f1d1c3fa
MD
1093 * (self contained on a per cpu basis)
1094 */
1095void
faaeffac 1096splz_check(void)
f1d1c3fa 1097{
7966cb69
MD
1098 globaldata_t gd = mycpu;
1099 thread_t td = gd->gd_curthread;
ef0fdad1 1100
4a28fe22
MD
1101 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1102 gd->gd_intr_nesting_level == 0 &&
1103 td->td_nest_count < 2)
1104 {
f1d1c3fa 1105 splz();
4a28fe22
MD
1106 }
1107}
1108
1109/*
1110 * This version is integrated into crit_exit, reqflags has already
1111 * been tested but td_critcount has not.
1112 *
1113 * We only want to execute the splz() on the 1->0 transition of
1114 * critcount and not in a hard code section or if too deeply nested.
925040f2 1115 *
0846e4ce 1116 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
4a28fe22
MD
1117 */
1118void
1119lwkt_maybe_splz(thread_t td)
1120{
1121 globaldata_t gd = td->td_gd;
1122
1123 if (td->td_critcount == 0 &&
1124 gd->gd_intr_nesting_level == 0 &&
1125 td->td_nest_count < 2)
1126 {
1127 splz();
1128 }
f1d1c3fa
MD
1129}
1130
e6546af9
MD
1131/*
1132 * Drivers which set up processing co-threads can call this function to
1133 * run the co-thread at a higher priority and to allow it to preempt
1134 * normal threads.
1135 */
1136void
1137lwkt_set_interrupt_support_thread(void)
1138{
1139 thread_t td = curthread;
1140
1141 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1142 td->td_flags |= TDF_INTTHREAD;
1143 td->td_preemptable = lwkt_preempt;
1144}
1145
1146
8ad65e08 1147/*
f9235b6d
MD
1148 * This function is used to negotiate a passive release of the current
1149 * process/lwp designation with the user scheduler, allowing the user
1150 * scheduler to schedule another user thread. The related kernel thread
1151 * (curthread) continues running in the released state.
8ad65e08
MD
1152 */
1153void
f9235b6d 1154lwkt_passive_release(struct thread *td)
8ad65e08 1155{
f9235b6d
MD
1156 struct lwp *lp = td->td_lwp;
1157
e3e6be1f 1158#ifndef NO_LWKT_SPLIT_USERPRI
f9235b6d
MD
1159 td->td_release = NULL;
1160 lwkt_setpri_self(TDPRI_KERN_USER);
d992c377
MD
1161#endif
1162
f9235b6d 1163 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1164}
1165
f9235b6d 1166
3824f392 1167/*
d2d8515b
MD
1168 * This implements a LWKT yield, allowing a kernel thread to yield to other
1169 * kernel threads at the same or higher priority. This function can be
1170 * called in a tight loop and will typically only yield once per tick.
f9235b6d 1171 *
d2d8515b
MD
1172 * Most kernel threads run at the same priority in order to allow equal
1173 * sharing.
f9235b6d
MD
1174 *
1175 * (self contained on a per cpu basis)
3824f392
MD
1176 */
1177void
f9235b6d 1178lwkt_yield(void)
3824f392 1179{
f9235b6d
MD
1180 globaldata_t gd = mycpu;
1181 thread_t td = gd->gd_curthread;
3824f392 1182
f9235b6d
MD
1183 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1184 splz();
85946b6c 1185 if (lwkt_resched_wanted()) {
f9235b6d
MD
1186 lwkt_schedule_self(curthread);
1187 lwkt_switch();
f9235b6d 1188 }
3824f392
MD
1189}
1190
40504122
MD
1191/*
1192 * The quick version processes pending interrupts and higher-priority
1193 * LWKT threads but will not round-robin same-priority LWKT threads.
de4d4cb0
MD
1194 *
1195 * When called while attempting to return to userland the only same-pri
1196 * threads are the ones which have already tried to become the current
1197 * user process.
40504122
MD
1198 */
1199void
1200lwkt_yield_quick(void)
1201{
1202 globaldata_t gd = mycpu;
1203 thread_t td = gd->gd_curthread;
1204
1205 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1206 splz();
1207 if (lwkt_resched_wanted()) {
1208 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1209 clear_lwkt_resched();
1210 } else {
1211 lwkt_schedule_self(curthread);
1212 lwkt_switch();
1213 }
1214 }
1215}
1216
3824f392 1217/*
f9235b6d
MD
1218 * This yield is designed for kernel threads with a user context.
1219 *
1220 * The kernel acting on behalf of the user is potentially cpu-bound,
1221 * this function will efficiently allow other threads to run and also
1222 * switch to other processes by releasing.
3824f392
MD
1223 *
1224 * The lwkt_user_yield() function is designed to have very low overhead
1225 * if no yield is determined to be needed.
1226 */
1227void
1228lwkt_user_yield(void)
1229{
f9235b6d
MD
1230 globaldata_t gd = mycpu;
1231 thread_t td = gd->gd_curthread;
1232
1233 /*
1234 * Always run any pending interrupts in case we are in a critical
1235 * section.
1236 */
1237 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1238 splz();
3824f392 1239
3824f392 1240 /*
f9235b6d
MD
1241 * Switch (which forces a release) if another kernel thread needs
1242 * the cpu, if userland wants us to resched, or if our kernel
1243 * quantum has run out.
3824f392 1244 */
f9235b6d 1245 if (lwkt_resched_wanted() ||
85946b6c 1246 user_resched_wanted())
f9235b6d 1247 {
3824f392 1248 lwkt_switch();
3824f392
MD
1249 }
1250
f9235b6d 1251#if 0
3824f392 1252 /*
f9235b6d
MD
1253 * Reacquire the current process if we are released.
1254 *
1255 * XXX not implemented atm. The kernel may be holding locks and such,
1256 * so we want the thread to continue to receive cpu.
3824f392 1257 */
f9235b6d
MD
1258 if (td->td_release == NULL && lp) {
1259 lp->lwp_proc->p_usched->acquire_curproc(lp);
1260 td->td_release = lwkt_passive_release;
1261 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1262 }
f9235b6d 1263#endif
b9eb1c19
MD
1264}
1265
8ad65e08 1266/*
f1d1c3fa
MD
1267 * Generic schedule. Possibly schedule threads belonging to other cpus and
1268 * deal with threads that might be blocked on a wait queue.
1269 *
0a3f9b47
MD
1270 * We have a little helper inline function which does additional work after
1271 * the thread has been enqueued, including dealing with preemption and
1272 * setting need_lwkt_resched() (which prevents the kernel from returning
1273 * to userland until it has processed higher priority threads).
6330a558
MD
1274 *
1275 * It is possible for this routine to be called after a failed _enqueue
1276 * (due to the target thread migrating, sleeping, or otherwise blocked).
1277 * We have to check that the thread is actually on the run queue!
8ad65e08 1278 */
0a3f9b47
MD
1279static __inline
1280void
85946b6c 1281_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
0a3f9b47 1282{
6330a558 1283 if (ntd->td_flags & TDF_RUNQ) {
85946b6c 1284 if (ntd->td_preemptable) {
f9235b6d 1285 ntd->td_preemptable(ntd, ccount); /* YYY +token */
6330a558 1286 }
0a3f9b47
MD
1287 }
1288}
1289
361d01dd 1290static __inline
8ad65e08 1291void
85946b6c 1292_lwkt_schedule(thread_t td)
8ad65e08 1293{
37af14fe
MD
1294 globaldata_t mygd = mycpu;
1295
cf709dd2
MD
1296 KASSERT(td != &td->td_gd->gd_idlethread,
1297 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1298 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1299 crit_enter_gd(mygd);
4643740a
MD
1300 KKASSERT(td->td_lwp == NULL ||
1301 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1302
37af14fe 1303 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1304 _lwkt_enqueue(td);
1305 } else {
f1d1c3fa 1306 /*
7cd8d145
MD
1307 * If we own the thread, there is no race (since we are in a
1308 * critical section). If we do not own the thread there might
1309 * be a race but the target cpu will deal with it.
f1d1c3fa 1310 */
0f7a3396 1311#ifdef SMP
7cd8d145 1312 if (td->td_gd == mygd) {
9d265729 1313 _lwkt_enqueue(td);
85946b6c 1314 _lwkt_schedule_post(mygd, td, 1);
f1d1c3fa 1315 } else {
e381e77c 1316 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1317 }
0f7a3396 1318#else
7cd8d145 1319 _lwkt_enqueue(td);
85946b6c 1320 _lwkt_schedule_post(mygd, td, 1);
0f7a3396 1321#endif
8ad65e08 1322 }
37af14fe 1323 crit_exit_gd(mygd);
8ad65e08
MD
1324}
1325
361d01dd
MD
1326void
1327lwkt_schedule(thread_t td)
1328{
85946b6c 1329 _lwkt_schedule(td);
361d01dd
MD
1330}
1331
1332void
85946b6c 1333lwkt_schedule_noresched(thread_t td) /* XXX not impl */
361d01dd 1334{
85946b6c 1335 _lwkt_schedule(td);
361d01dd
MD
1336}
1337
88ebb169
SW
1338#ifdef SMP
1339
e381e77c
MD
1340/*
1341 * When scheduled remotely if frame != NULL the IPIQ is being
1342 * run via doreti or an interrupt then preemption can be allowed.
1343 *
1344 * To allow preemption we have to drop the critical section so only
1345 * one is present in _lwkt_schedule_post.
1346 */
1347static void
1348lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1349{
1350 thread_t td = curthread;
1351 thread_t ntd = arg;
1352
1353 if (frame && ntd->td_preemptable) {
1354 crit_exit_noyield(td);
85946b6c 1355 _lwkt_schedule(ntd);
e381e77c
MD
1356 crit_enter_quick(td);
1357 } else {
85946b6c 1358 _lwkt_schedule(ntd);
e381e77c
MD
1359 }
1360}
1361
d9eea1a5 1362/*
52eedfb5
MD
1363 * Thread migration using a 'Pull' method. The thread may or may not be
1364 * the current thread. It MUST be descheduled and in a stable state.
1365 * lwkt_giveaway() must be called on the cpu owning the thread.
1366 *
1367 * At any point after lwkt_giveaway() is called, the target cpu may
1368 * 'pull' the thread by calling lwkt_acquire().
1369 *
ae8e83e6
MD
1370 * We have to make sure the thread is not sitting on a per-cpu tsleep
1371 * queue or it will blow up when it moves to another cpu.
1372 *
52eedfb5 1373 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1374 */
52eedfb5
MD
1375void
1376lwkt_giveaway(thread_t td)
1377{
3b4192fb 1378 globaldata_t gd = mycpu;
52eedfb5 1379
3b4192fb
MD
1380 crit_enter_gd(gd);
1381 if (td->td_flags & TDF_TSLEEPQ)
1382 tsleep_remove(td);
1383 KKASSERT(td->td_gd == gd);
1384 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1385 td->td_flags |= TDF_MIGRATING;
1386 crit_exit_gd(gd);
52eedfb5
MD
1387}
1388
a2a5ad0d
MD
1389void
1390lwkt_acquire(thread_t td)
1391{
37af14fe
MD
1392 globaldata_t gd;
1393 globaldata_t mygd;
cc9b6223 1394 int retry = 10000000;
a2a5ad0d 1395
52eedfb5 1396 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1397 gd = td->td_gd;
37af14fe 1398 mygd = mycpu;
52eedfb5 1399 if (gd != mycpu) {
35238fa5 1400 cpu_lfence();
52eedfb5 1401 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1402 crit_enter_gd(mygd);
cfaeae2a 1403 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23
MD
1404 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1405#ifdef SMP
1406 lwkt_process_ipiq();
1407#endif
52eedfb5 1408 cpu_lfence();
cc9b6223
MD
1409 if (--retry == 0) {
1410 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1411 td, td->td_flags);
1412 retry = 10000000;
1413 }
df910c23 1414 }
cfaeae2a 1415 DEBUG_POP_INFO();
562273ea 1416 cpu_mfence();
37af14fe 1417 td->td_gd = mygd;
52eedfb5
MD
1418 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1419 td->td_flags &= ~TDF_MIGRATING;
1420 crit_exit_gd(mygd);
1421 } else {
1422 crit_enter_gd(mygd);
1423 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1424 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1425 crit_exit_gd(mygd);
a2a5ad0d
MD
1426 }
1427}
1428
52eedfb5
MD
1429#endif
1430
f1d1c3fa
MD
1431/*
1432 * Generic deschedule. Descheduling threads other then your own should be
1433 * done only in carefully controlled circumstances. Descheduling is
1434 * asynchronous.
1435 *
1436 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1437 */
1438void
1439lwkt_deschedule(thread_t td)
1440{
f1d1c3fa 1441 crit_enter();
b8a98473 1442#ifdef SMP
f1d1c3fa
MD
1443 if (td == curthread) {
1444 _lwkt_dequeue(td);
1445 } else {
a72187e9 1446 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1447 _lwkt_dequeue(td);
1448 } else {
b8a98473 1449 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1450 }
1451 }
b8a98473
MD
1452#else
1453 _lwkt_dequeue(td);
1454#endif
f1d1c3fa
MD
1455 crit_exit();
1456}
1457
4b5f931b
MD
1458/*
1459 * Set the target thread's priority. This routine does not automatically
1460 * switch to a higher priority thread, LWKT threads are not designed for
1461 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1462 */
1463void
1464lwkt_setpri(thread_t td, int pri)
1465{
f9235b6d
MD
1466 if (td->td_pri != pri) {
1467 KKASSERT(pri >= 0);
1468 crit_enter();
1469 if (td->td_flags & TDF_RUNQ) {
d2d8515b 1470 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1471 _lwkt_dequeue(td);
1472 td->td_pri = pri;
1473 _lwkt_enqueue(td);
1474 } else {
1475 td->td_pri = pri;
1476 }
1477 crit_exit();
26a0694b 1478 }
26a0694b
MD
1479}
1480
03bd0a5e
MD
1481/*
1482 * Set the initial priority for a thread prior to it being scheduled for
1483 * the first time. The thread MUST NOT be scheduled before or during
1484 * this call. The thread may be assigned to a cpu other then the current
1485 * cpu.
1486 *
1487 * Typically used after a thread has been created with TDF_STOPPREQ,
1488 * and before the thread is initially scheduled.
1489 */
1490void
1491lwkt_setpri_initial(thread_t td, int pri)
1492{
1493 KKASSERT(pri >= 0);
1494 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1495 td->td_pri = pri;
03bd0a5e
MD
1496}
1497
26a0694b
MD
1498void
1499lwkt_setpri_self(int pri)
1500{
1501 thread_t td = curthread;
1502
4b5f931b
MD
1503 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1504 crit_enter();
1505 if (td->td_flags & TDF_RUNQ) {
1506 _lwkt_dequeue(td);
f9235b6d 1507 td->td_pri = pri;
4b5f931b
MD
1508 _lwkt_enqueue(td);
1509 } else {
f9235b6d 1510 td->td_pri = pri;
4b5f931b
MD
1511 }
1512 crit_exit();
1513}
1514
f9235b6d 1515/*
85946b6c 1516 * hz tick scheduler clock for LWKT threads
f9235b6d
MD
1517 */
1518void
85946b6c 1519lwkt_schedulerclock(thread_t td)
f9235b6d 1520{
85946b6c
MD
1521 globaldata_t gd = td->td_gd;
1522 thread_t xtd;
2a418930 1523
85946b6c
MD
1524 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1525 /*
1526 * If the current thread is at the head of the runq shift it to the
1527 * end of any equal-priority threads and request a LWKT reschedule
1528 * if it moved.
d992c377
MD
1529 *
1530 * Ignore upri in this situation. There will only be one user thread
1531 * in user mode, all others will be user threads running in kernel
1532 * mode and we have to make sure they get some cpu.
85946b6c
MD
1533 */
1534 xtd = TAILQ_NEXT(td, td_threadq);
1535 if (xtd && xtd->td_pri == td->td_pri) {
1536 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1537 while (xtd && xtd->td_pri == td->td_pri)
1538 xtd = TAILQ_NEXT(xtd, td_threadq);
1539 if (xtd)
1540 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1541 else
1542 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1543 need_lwkt_resched();
f9235b6d 1544 }
85946b6c
MD
1545 } else {
1546 /*
1547 * If we scheduled a thread other than the one at the head of the
1548 * queue always request a reschedule every tick.
1549 */
1550 need_lwkt_resched();
f9235b6d
MD
1551 }
1552}
1553
5d21b981 1554/*
52eedfb5
MD
1555 * Migrate the current thread to the specified cpu.
1556 *
cc9b6223
MD
1557 * This is accomplished by descheduling ourselves from the current cpu
1558 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1559 * 'old' thread wants to migrate after it has been completely switched out
1560 * and will complete the migration.
1561 *
1562 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1563 *
1564 * We must be sure to release our current process designation (if a user
1565 * process) before clearing out any tsleepq we are on because the release
1566 * code may re-add us.
ae8e83e6
MD
1567 *
1568 * We must be sure to remove ourselves from the current cpu's tsleepq
1569 * before potentially moving to another queue. The thread can be on
1570 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1571 */
5d21b981
MD
1572
1573void
1574lwkt_setcpu_self(globaldata_t rgd)
1575{
1576#ifdef SMP
1577 thread_t td = curthread;
1578
1579 if (td->td_gd != rgd) {
1580 crit_enter_quick(td);
cc9b6223 1581
95858b91
MD
1582 if (td->td_release)
1583 td->td_release(td);
ae8e83e6 1584 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1585 tsleep_remove(td);
cc9b6223
MD
1586
1587 /*
1588 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1589 * trying to deschedule ourselves and switch away, then deschedule
1590 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1591 * call lwkt_switch() to complete the operation.
1592 */
5d21b981
MD
1593 td->td_flags |= TDF_MIGRATING;
1594 lwkt_deschedule_self(td);
52eedfb5 1595 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1596 td->td_migrate_gd = rgd;
5d21b981 1597 lwkt_switch();
cc9b6223
MD
1598
1599 /*
1600 * We are now on the target cpu
1601 */
1602 KKASSERT(rgd == mycpu);
52eedfb5 1603 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1604 crit_exit_quick(td);
1605 }
1606#endif
1607}
1608
ecdefdda
MD
1609void
1610lwkt_migratecpu(int cpuid)
1611{
1612#ifdef SMP
1613 globaldata_t rgd;
1614
1615 rgd = globaldata_find(cpuid);
1616 lwkt_setcpu_self(rgd);
1617#endif
1618}
1619
cc9b6223 1620#ifdef SMP
5d21b981
MD
1621/*
1622 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
1623 * do not have to enter another one).
1624 *
1625 * The thread (td) has already been completely descheduled from the
1626 * originating cpu and we can simply assert the case. The thread is
1627 * assigned to the new cpu and enqueued.
5d21b981 1628 *
cc9b6223 1629 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1630 */
1631static void
1632lwkt_setcpu_remote(void *arg)
1633{
1634 thread_t td = arg;
1635 globaldata_t gd = mycpu;
1636
cc9b6223 1637 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1638 td->td_gd = gd;
562273ea 1639 cpu_mfence();
5d21b981 1640 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1641 KKASSERT(td->td_migrate_gd == NULL);
4643740a
MD
1642 KKASSERT(td->td_lwp == NULL ||
1643 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
5d21b981
MD
1644 _lwkt_enqueue(td);
1645}
3d28ff59 1646#endif
5d21b981 1647
553ea3c8 1648struct lwp *
4b5f931b
MD
1649lwkt_preempted_proc(void)
1650{
73e4f7b9 1651 thread_t td = curthread;
4b5f931b
MD
1652 while (td->td_preempted)
1653 td = td->td_preempted;
553ea3c8 1654 return(td->td_lwp);
4b5f931b
MD
1655}
1656
99df837e
MD
1657/*
1658 * Create a kernel process/thread/whatever. It shares it's address space
1659 * with proc0 - ie: kernel only.
1660 *
d2d8515b
MD
1661 * If the cpu is not specified one will be selected. In the future
1662 * specifying a cpu of -1 will enable kernel thread migration between
1663 * cpus.
99df837e
MD
1664 */
1665int
c9e9fb21
MD
1666lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1667 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1668{
73e4f7b9 1669 thread_t td;
e2565a42 1670 __va_list ap;
99df837e 1671
d3d32139 1672 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1673 tdflags);
a2a5ad0d
MD
1674 if (tdp)
1675 *tdp = td;
709799ea 1676 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1677
1678 /*
1679 * Set up arg0 for 'ps' etc
1680 */
e2565a42 1681 __va_start(ap, fmt);
379210cb 1682 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1683 __va_end(ap);
99df837e
MD
1684
1685 /*
1686 * Schedule the thread to run
1687 */
4643740a
MD
1688 if (td->td_flags & TDF_NOSTART)
1689 td->td_flags &= ~TDF_NOSTART;
ef0fdad1 1690 else
4643740a 1691 lwkt_schedule(td);
99df837e
MD
1692 return 0;
1693}
1694
1695/*
1696 * Destroy an LWKT thread. Warning! This function is not called when
1697 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1698 * uses a different reaping mechanism.
1699 */
1700void
1701lwkt_exit(void)
1702{
1703 thread_t td = curthread;
c070746a 1704 thread_t std;
8826f33a 1705 globaldata_t gd;
99df837e 1706
2883d2d8
MD
1707 /*
1708 * Do any cleanup that might block here
1709 */
99df837e 1710 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1711 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1712 caps_exit(td);
2883d2d8
MD
1713 biosched_done(td);
1714 dsched_exit_thread(td);
c070746a
MD
1715
1716 /*
1717 * Get us into a critical section to interlock gd_freetd and loop
1718 * until we can get it freed.
1719 *
1720 * We have to cache the current td in gd_freetd because objcache_put()ing
1721 * it would rip it out from under us while our thread is still active.
2af9d75d
MD
1722 *
1723 * We are the current thread so of course our own TDF_RUNNING bit will
1724 * be set, so unlike the lwp reap code we don't wait for it to clear.
c070746a
MD
1725 */
1726 gd = mycpu;
37af14fe 1727 crit_enter_quick(td);
2af9d75d
MD
1728 for (;;) {
1729 if (td->td_refs) {
1730 tsleep(td, 0, "tdreap", 1);
1731 continue;
1732 }
1733 if ((std = gd->gd_freetd) != NULL) {
1734 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1735 gd->gd_freetd = NULL;
1736 objcache_put(thread_cache, std);
1737 continue;
1738 }
1739 break;
c070746a 1740 }
3b4192fb
MD
1741
1742 /*
1743 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1744 * the last time. We cannot block after this point or we may end
1745 * up with a stale td on the tsleepq.
eb2adbf5
MD
1746 *
1747 * None of this may block, the critical section is the only thing
1748 * protecting tdallq and the only thing preventing new lwkt_hold()
1749 * thread refs now.
3b4192fb
MD
1750 */
1751 if (td->td_flags & TDF_TSLEEPQ)
1752 tsleep_remove(td);
37af14fe 1753 lwkt_deschedule_self(td);
e56e4dea 1754 lwkt_remove_tdallq(td);
74c9628e 1755 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1756
1757 /*
1758 * Final cleanup
1759 */
1760 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1761 if (td->td_flags & TDF_ALLOCATED_THREAD)
1762 gd->gd_freetd = td;
99df837e
MD
1763 cpu_thread_exit();
1764}
1765
e56e4dea
MD
1766void
1767lwkt_remove_tdallq(thread_t td)
1768{
1769 KKASSERT(td->td_gd == mycpu);
1770 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1771}
1772
9cf43f91
MD
1773/*
1774 * Code reduction and branch prediction improvements. Call/return
1775 * overhead on modern cpus often degenerates into 0 cycles due to
1776 * the cpu's branch prediction hardware and return pc cache. We
1777 * can take advantage of this by not inlining medium-complexity
1778 * functions and we can also reduce the branch prediction impact
1779 * by collapsing perfectly predictable branches into a single
1780 * procedure instead of duplicating it.
1781 *
1782 * Is any of this noticeable? Probably not, so I'll take the
1783 * smaller code size.
1784 */
1785void
b6468f56 1786crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1787{
b6468f56 1788 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1789}
1790
2d93b37a
MD
1791void
1792crit_panic(void)
1793{
1794 thread_t td = curthread;
850634cc 1795 int lcrit = td->td_critcount;
2d93b37a 1796
850634cc
AH
1797 td->td_critcount = 0;
1798 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1799 /* NOT REACHED */
2d93b37a
MD
1800}
1801
d165e668
MD
1802#ifdef SMP
1803
bd8015ca
MD
1804/*
1805 * Called from debugger/panic on cpus which have been stopped. We must still
1806 * process the IPIQ while stopped, even if we were stopped while in a critical
1807 * section (XXX).
1808 *
1809 * If we are dumping also try to process any pending interrupts. This may
1810 * or may not work depending on the state of the cpu at the point it was
1811 * stopped.
1812 */
1813void
1814lwkt_smp_stopped(void)
1815{
1816 globaldata_t gd = mycpu;
1817
1818 crit_enter_gd(gd);
1819 if (dumping) {
1820 lwkt_process_ipiq();
1821 splz();
1822 } else {
1823 lwkt_process_ipiq();
1824 }
1825 crit_exit_gd(gd);
1826}
1827
d165e668 1828#endif