msgport: Merge several sendmsg functions
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
b12defdc 2 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
60f60350 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
60f60350 20 *
8c10bfcf
MD
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
b37f18d6 47#include <sys/kinfo.h>
8ad65e08 48#include <sys/queue.h>
7d0bac62 49#include <sys/sysctl.h>
99df837e 50#include <sys/kthread.h>
f1d1c3fa 51#include <machine/cpu.h>
99df837e 52#include <sys/lock.h>
9d265729 53#include <sys/spinlock.h>
57aa743c 54#include <sys/ktr.h>
9d265729
MD
55
56#include <sys/thread2.h>
57#include <sys/spinlock2.h>
684a93c4 58#include <sys/mplock2.h>
f1d1c3fa 59
8c72e3d5
AH
60#include <sys/dsched.h>
61
7d0bac62
MD
62#include <vm/vm.h>
63#include <vm/vm_param.h>
64#include <vm/vm_kern.h>
65#include <vm/vm_object.h>
66#include <vm/vm_page.h>
67#include <vm/vm_map.h>
68#include <vm/vm_pager.h>
69#include <vm/vm_extern.h>
7d0bac62 70
99df837e 71#include <machine/stdarg.h>
96728c05 72#include <machine/smp.h>
99df837e 73
8cee56f4
MD
74#ifdef _KERNEL_VIRTUAL
75#include <pthread.h>
76#endif
77
d850923c
AE
78#if !defined(KTR_CTXSW)
79#define KTR_CTXSW KTR_ALL
80#endif
81KTR_INFO_MASTER(ctxsw);
5bf48697
AE
82KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td);
83KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td);
84KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm);
85KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td);
1541028a 86
40aaf5fc
NT
87static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
88
0f7a3396
MD
89#ifdef INVARIANTS
90static int panic_on_cscount = 0;
91#endif
05220613
MD
92static __int64_t switch_count = 0;
93static __int64_t preempt_hit = 0;
94static __int64_t preempt_miss = 0;
95static __int64_t preempt_weird = 0;
fb0f29c4 96static int lwkt_use_spin_port;
40aaf5fc 97static struct objcache *thread_cache;
a46b4a23 98int cpu_mwait_spin = 0;
05220613 99
e381e77c 100static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
cc9b6223 101static void lwkt_setcpu_remote(void *arg);
e381e77c 102
0855a2af
JG
103extern void cpu_heavy_restore(void);
104extern void cpu_lwkt_restore(void);
105extern void cpu_kthread_restore(void);
106extern void cpu_idle_restore(void);
107
fb0f29c4
MD
108/*
109 * We can make all thread ports use the spin backend instead of the thread
110 * backend. This should only be set to debug the spin backend.
111 */
112TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
113
0f7a3396 114#ifdef INVARIANTS
0c52fa62
SG
115SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
116 "Panic if attempting to switch lwkt's while mastering cpusync");
0f7a3396 117#endif
a46b4a23
MD
118SYSCTL_INT(_hw, OID_AUTO, cpu_mwait_spin, CTLFLAG_RW, &cpu_mwait_spin, 0,
119 "monitor/mwait target state");
0c52fa62
SG
120SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
121 "Number of switched threads");
9733f757 122SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
0c52fa62 123 "Successful preemption events");
9733f757 124SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
0c52fa62
SG
125 "Failed preemption events");
126SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
127 "Number of preempted threads.");
b12defdc 128static int fairq_enable = 0;
2a418930
MD
129SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
130 &fairq_enable, 0, "Turn on fairq priority accumulators");
85946b6c 131static int fairq_bypass = -1;
b12defdc
MD
132SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
133 &fairq_bypass, 0, "Allow fairq to bypass td on token failure");
134extern int lwkt_sched_debug;
135int lwkt_sched_debug = 0;
136SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
137 &lwkt_sched_debug, 0, "Scheduler debug");
2a418930
MD
138static int lwkt_spin_loops = 10;
139SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
b12defdc
MD
140 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
141static int lwkt_spin_reseq = 0;
142SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW,
143 &lwkt_spin_reseq, 0, "Scheduler resequencer enable");
144static int lwkt_spin_monitor = 0;
145SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW,
146 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait");
d5b2d319
MD
147static int lwkt_spin_fatal = 0; /* disabled */
148SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
149 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
fbc024e4 150static int preempt_enable = 1;
2a418930
MD
151SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
152 &preempt_enable, 0, "Enable preemption");
7b234d8c 153static int lwkt_cache_threads = 0;
765b1ae0
MD
154SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
155 &lwkt_cache_threads, 0, "thread+kstack cache");
fbc024e4 156
8cee56f4 157#ifndef _KERNEL_VIRTUAL
2a418930
MD
158static __cachealign int lwkt_cseq_rindex;
159static __cachealign int lwkt_cseq_windex;
8cee56f4 160#endif
05220613 161
4b5f931b
MD
162/*
163 * These helper procedures handle the runq, they can only be called from
164 * within a critical section.
75cdbe6c
MD
165 *
166 * WARNING! Prior to SMP being brought up it is possible to enqueue and
167 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
168 * instead of 'mycpu' when referencing the globaldata structure. Once
169 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 170 */
f1d1c3fa
MD
171static __inline
172void
173_lwkt_dequeue(thread_t td)
174{
175 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 176 struct globaldata *gd = td->td_gd;
4b5f931b 177
f1d1c3fa 178 td->td_flags &= ~TDF_RUNQ;
f9235b6d 179 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
de4d4cb0 180 --gd->gd_tdrunqcount;
f9235b6d 181 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
2a418930 182 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
183 }
184}
185
f9235b6d
MD
186/*
187 * Priority enqueue.
188 *
d992c377
MD
189 * There are a limited number of lwkt threads runnable since user
190 * processes only schedule one at a time per cpu. However, there can
191 * be many user processes in kernel mode exiting from a tsleep() which
e3e6be1f 192 * become runnable.
d992c377
MD
193 *
194 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and
195 * will ignore user priority. This is to ensure that user threads in
196 * kernel mode get cpu at some point regardless of what the user
197 * scheduler thinks.
f9235b6d 198 */
f1d1c3fa
MD
199static __inline
200void
201_lwkt_enqueue(thread_t td)
202{
f9235b6d
MD
203 thread_t xtd;
204
7f5d7ed7 205 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 206 struct globaldata *gd = td->td_gd;
4b5f931b 207
f1d1c3fa 208 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
209 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
210 if (xtd == NULL) {
85946b6c
MD
211 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
212 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
f9235b6d 213 } else {
e3e6be1f
MD
214 /*
215 * NOTE: td_upri - higher numbers more desireable, same sense
216 * as td_pri (typically reversed from lwp_upri).
217 *
218 * In the equal priority case we want the best selection
219 * at the beginning so the less desireable selections know
220 * that they have to setrunqueue/go-to-another-cpu, even
221 * though it means switching back to the 'best' selection.
222 * This also avoids degenerate situations when many threads
223 * are runnable or waking up at the same time.
224 *
225 * If upri matches exactly place at end/round-robin.
226 */
d992c377 227 while (xtd &&
e3e6be1f 228 (xtd->td_pri >= td->td_pri ||
d992c377 229 (xtd->td_pri == td->td_pri &&
e3e6be1f 230 xtd->td_upri >= td->td_upri))) {
85946b6c 231 xtd = TAILQ_NEXT(xtd, td_threadq);
d992c377 232 }
85946b6c
MD
233 if (xtd)
234 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
235 else
236 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
f9235b6d 237 }
de4d4cb0 238 ++gd->gd_tdrunqcount;
b12defdc
MD
239
240 /*
85946b6c 241 * Request a LWKT reschedule if we are now at the head of the queue.
b12defdc 242 */
85946b6c
MD
243 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
244 need_lwkt_resched();
f1d1c3fa
MD
245 }
246}
8ad65e08 247
40aaf5fc
NT
248static __boolean_t
249_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
250{
251 struct thread *td = (struct thread *)obj;
252
253 td->td_kstack = NULL;
254 td->td_kstack_size = 0;
255 td->td_flags = TDF_ALLOCATED_THREAD;
4643740a 256 td->td_mpflags = 0;
40aaf5fc
NT
257 return (1);
258}
259
260static void
261_lwkt_thread_dtor(void *obj, void *privdata)
262{
263 struct thread *td = (struct thread *)obj;
264
265 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
266 ("_lwkt_thread_dtor: not allocated from objcache"));
267 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
268 td->td_kstack_size > 0,
269 ("_lwkt_thread_dtor: corrupted stack"));
270 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
7b234d8c
MD
271 td->td_kstack = NULL;
272 td->td_flags = 0;
40aaf5fc
NT
273}
274
275/*
276 * Initialize the lwkt s/system.
765b1ae0 277 *
7b234d8c
MD
278 * Nominally cache up to 32 thread + kstack structures. Cache more on
279 * systems with a lot of cpu cores.
40aaf5fc
NT
280 */
281void
282lwkt_init(void)
283{
765b1ae0 284 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
7b234d8c
MD
285 if (lwkt_cache_threads == 0) {
286 lwkt_cache_threads = ncpus * 4;
287 if (lwkt_cache_threads < 32)
288 lwkt_cache_threads = 32;
289 }
765b1ae0
MD
290 thread_cache = objcache_create_mbacked(
291 M_THREAD, sizeof(struct thread),
2fce2579 292 0, lwkt_cache_threads,
765b1ae0 293 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
294}
295
37af14fe
MD
296/*
297 * Schedule a thread to run. As the current thread we can always safely
298 * schedule ourselves, and a shortcut procedure is provided for that
299 * function.
300 *
301 * (non-blocking, self contained on a per cpu basis)
302 */
303void
304lwkt_schedule_self(thread_t td)
305{
cfaeae2a 306 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 307 crit_enter_quick(td);
f9235b6d
MD
308 KASSERT(td != &td->td_gd->gd_idlethread,
309 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
4643740a
MD
310 KKASSERT(td->td_lwp == NULL ||
311 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
37af14fe 312 _lwkt_enqueue(td);
37af14fe
MD
313 crit_exit_quick(td);
314}
315
316/*
317 * Deschedule a thread.
318 *
319 * (non-blocking, self contained on a per cpu basis)
320 */
321void
322lwkt_deschedule_self(thread_t td)
323{
324 crit_enter_quick(td);
37af14fe
MD
325 _lwkt_dequeue(td);
326 crit_exit_quick(td);
327}
328
8ad65e08
MD
329/*
330 * LWKTs operate on a per-cpu basis
331 *
73e4f7b9 332 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
333 */
334void
335lwkt_gdinit(struct globaldata *gd)
336{
f9235b6d 337 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 338 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
339}
340
7d0bac62
MD
341/*
342 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
343 * or LWKT start address before it can be scheduled. If the target cpu is
344 * -1 the thread will be created on the current cpu.
0cfcada1
MD
345 *
346 * If you intend to create a thread without a process context this function
347 * does everything except load the startup and switcher function.
7d0bac62
MD
348 */
349thread_t
d3d32139 350lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 351{
d2d8515b 352 static int cpu_rotator;
c070746a 353 globaldata_t gd = mycpu;
99df837e 354 void *stack;
7d0bac62 355
c070746a
MD
356 /*
357 * If static thread storage is not supplied allocate a thread. Reuse
358 * a cached free thread if possible. gd_freetd is used to keep an exiting
359 * thread intact through the exit.
360 */
ef0fdad1 361 if (td == NULL) {
cf709dd2
MD
362 crit_enter_gd(gd);
363 if ((td = gd->gd_freetd) != NULL) {
364 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
365 TDF_RUNQ)) == 0);
c070746a 366 gd->gd_freetd = NULL;
cf709dd2 367 } else {
c070746a 368 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
369 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
370 TDF_RUNQ)) == 0);
371 }
372 crit_exit_gd(gd);
40aaf5fc 373 KASSERT((td->td_flags &
2af9d75d
MD
374 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
375 TDF_ALLOCATED_THREAD,
40aaf5fc
NT
376 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
377 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 378 }
c070746a
MD
379
380 /*
381 * Try to reuse cached stack.
382 */
f470d0c8
MD
383 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
384 if (flags & TDF_ALLOCATED_STACK) {
e4846942 385 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
386 stack = NULL;
387 }
388 }
389 if (stack == NULL) {
e40cfbd7 390 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 391 flags |= TDF_ALLOCATED_STACK;
99df837e 392 }
d2d8515b
MD
393 if (cpu < 0) {
394 cpu = ++cpu_rotator;
395 cpu_ccfence();
396 cpu %= ncpus;
397 }
398 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 399 return(td);
7d0bac62
MD
400}
401
402/*
403 * Initialize a preexisting thread structure. This function is used by
404 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
405 *
f8c3996b
MD
406 * All threads start out in a critical section at a priority of
407 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
408 * appropriate. This function may send an IPI message when the
409 * requested cpu is not the current cpu and consequently gd_tdallq may
410 * not be initialized synchronously from the point of view of the originating
411 * cpu.
412 *
413 * NOTE! we have to be careful in regards to creating threads for other cpus
414 * if SMP has not yet been activated.
7d0bac62 415 */
75cdbe6c
MD
416static void
417lwkt_init_thread_remote(void *arg)
418{
419 thread_t td = arg;
420
52eedfb5
MD
421 /*
422 * Protected by critical section held by IPI dispatch
423 */
75cdbe6c
MD
424 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
425}
426
fdce8919
MD
427/*
428 * lwkt core thread structural initialization.
429 *
430 * NOTE: All threads are initialized as mpsafe threads.
431 */
7d0bac62 432void
f470d0c8
MD
433lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
434 struct globaldata *gd)
7d0bac62 435{
37af14fe
MD
436 globaldata_t mygd = mycpu;
437
99df837e
MD
438 bzero(td, sizeof(struct thread));
439 td->td_kstack = stack;
f470d0c8 440 td->td_kstack_size = stksize;
d3d32139 441 td->td_flags = flags;
4643740a 442 td->td_mpflags = 0;
f256b6c0 443 td->td_type = TD_TYPE_GENERIC;
26a0694b 444 td->td_gd = gd;
f9235b6d
MD
445 td->td_pri = TDPRI_KERN_DAEMON;
446 td->td_critcount = 1;
54341a3b 447 td->td_toks_have = NULL;
3b998fa9 448 td->td_toks_stop = &td->td_toks_base;
392cd266 449 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
d636cd48 450 lwkt_initport_spin(&td->td_msgport, td);
fb0f29c4
MD
451 else
452 lwkt_initport_thread(&td->td_msgport, td);
99df837e 453 pmap_init_thread(td);
5d21b981
MD
454 /*
455 * Normally initializing a thread for a remote cpu requires sending an
456 * IPI. However, the idlethread is setup before the other cpus are
457 * activated so we have to treat it as a special case. XXX manipulation
458 * of gd_tdallq requires the BGL.
459 */
460 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 461 crit_enter_gd(mygd);
75cdbe6c 462 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 463 crit_exit_gd(mygd);
75cdbe6c 464 } else {
2db3b277 465 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 466 }
8c72e3d5 467 dsched_new_thread(td);
73e4f7b9
MD
468}
469
470void
471lwkt_set_comm(thread_t td, const char *ctl, ...)
472{
e2565a42 473 __va_list va;
73e4f7b9 474
e2565a42 475 __va_start(va, ctl);
379210cb 476 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 477 __va_end(va);
5bf48697 478 KTR_LOG(ctxsw_newtd, td, td->td_comm);
7d0bac62
MD
479}
480
eb2adbf5
MD
481/*
482 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
483 * this does not prevent the thread from migrating to another cpu so the
484 * gd_tdallq state is not protected by this.
485 */
99df837e 486void
73e4f7b9 487lwkt_hold(thread_t td)
99df837e 488{
74c9628e 489 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
490}
491
492void
493lwkt_rele(thread_t td)
494{
495 KKASSERT(td->td_refs > 0);
74c9628e 496 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
497}
498
73e4f7b9
MD
499void
500lwkt_free_thread(thread_t td)
501{
74c9628e 502 KKASSERT(td->td_refs == 0);
c17a6852
MD
503 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
504 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
40aaf5fc
NT
505 if (td->td_flags & TDF_ALLOCATED_THREAD) {
506 objcache_put(thread_cache, td);
507 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
508 /* client-allocated struct with internally allocated stack */
509 KASSERT(td->td_kstack && td->td_kstack_size > 0,
510 ("lwkt_free_thread: corrupted stack"));
511 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
512 td->td_kstack = NULL;
513 td->td_kstack_size = 0;
99df837e 514 }
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
574339e6 720 ++gd->gd_cnt.v_token_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
574339e6 750 ++gd->gd_cnt.v_token_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
MD
806 ~RQF_IDLECHECK_WK_MASK,
807 cpu_mwait_spin);
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) {
a46b4a23 841 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq, cpu_mwait_spin);
b12defdc
MD
842 } else {
843#endif
844 cpu_pause();
845 cpu_lfence();
846#if 1
847 }
8b402283 848#endif
0f0466c0 849 }
b12defdc
MD
850 DELAY(1);
851 atomic_add_int(&lwkt_cseq_rindex, 1);
2a418930 852 }
8cee56f4 853#endif
2a418930 854 /* highest level for(;;) loop */
f1d1c3fa 855 }
8a8d5d85 856
2a418930 857havethread:
b12defdc 858 /*
be71787b
MD
859 * Clear gd_idle_repeat when doing a normal switch to a non-idle
860 * thread.
f9235b6d 861 */
9ac1ee6e 862 ntd->td_wmesg = NULL;
b12defdc 863 ++gd->gd_cnt.v_swtch;
be71787b 864 gd->gd_idle_repeat = 0;
2a418930 865
f9235b6d 866havethread_preempted:
f9235b6d
MD
867 /*
868 * If the new target does not need the MP lock and we are holding it,
869 * release the MP lock. If the new target requires the MP lock we have
870 * already acquired it for the target.
8a8d5d85 871 */
2a418930 872 ;
f9235b6d
MD
873haveidle:
874 KASSERT(ntd->td_critcount,
b5d16701
MD
875 ("priority problem in lwkt_switch %d %d",
876 td->td_critcount, ntd->td_critcount));
877
94f6d86e 878 if (td != ntd) {
cc9b6223
MD
879 /*
880 * Execute the actual thread switch operation. This function
881 * returns to the current thread and returns the previous thread
882 * (which may be different from the thread we switched to).
883 *
884 * We are responsible for marking ntd as TDF_RUNNING.
885 */
121f93bc 886 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
94f6d86e 887 ++switch_count;
a1f0fb66 888 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
889 ntd->td_flags |= TDF_RUNNING;
890 lwkt_switch_return(td->td_switch(ntd));
891 /* ntd invalid, td_switch() can return a different thread_t */
94f6d86e 892 }
b12defdc 893
b12defdc 894 /*
54341a3b 895 * catch-all. XXX is this strictly needed?
b12defdc
MD
896 */
897 splz_check();
54341a3b 898
37af14fe
MD
899 /* NOTE: current cpu may have changed after switch */
900 crit_exit_quick(td);
8ad65e08
MD
901}
902
cc9b6223
MD
903/*
904 * Called by assembly in the td_switch (thread restore path) for thread
905 * bootstrap cases which do not 'return' to lwkt_switch().
906 */
907void
908lwkt_switch_return(thread_t otd)
909{
cc9b6223
MD
910 globaldata_t rgd;
911
912 /*
913 * Check if otd was migrating. Now that we are on ntd we can finish
914 * up the migration. This is a bit messy but it is the only place
915 * where td is known to be fully descheduled.
916 *
917 * We can only activate the migration if otd was migrating but not
918 * held on the cpu due to a preemption chain. We still have to
919 * clear TDF_RUNNING on the old thread either way.
920 *
921 * We are responsible for clearing the previously running thread's
922 * TDF_RUNNING.
923 */
924 if ((rgd = otd->td_migrate_gd) != NULL &&
925 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
926 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
927 (TDF_MIGRATING | TDF_RUNNING));
928 otd->td_migrate_gd = NULL;
929 otd->td_flags &= ~TDF_RUNNING;
930 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
931 } else {
932 otd->td_flags &= ~TDF_RUNNING;
933 }
2b07d9aa
MD
934
935 /*
936 * Final exit validations (see lwp_wait()). Note that otd becomes
937 * invalid the *instant* we set TDF_MP_EXITSIG.
938 */
939 while (otd->td_flags & TDF_EXITING) {
940 u_int mpflags;
941
942 mpflags = otd->td_mpflags;
943 cpu_ccfence();
944
945 if (mpflags & TDF_MP_EXITWAIT) {
946 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
947 mpflags | TDF_MP_EXITSIG)) {
948 wakeup(otd);
949 break;
950 }
951 } else {
952 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
953 mpflags | TDF_MP_EXITSIG)) {
954 wakeup(otd);
955 break;
956 }
957 }
958 }
cc9b6223
MD
959}
960
b68b7282 961/*
96728c05 962 * Request that the target thread preempt the current thread. Preemption
54341a3b
MD
963 * can only occur if our only critical section is the one that we were called
964 * with, the relative priority of the target thread is higher, and the target
965 * thread holds no tokens. This also only works if we are not holding any
966 * spinlocks (obviously).
96728c05
MD
967 *
968 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
969 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 970 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
971 * to determine whether preemption is possible (aka usually just the crit
972 * priority of lwkt_schedule() itself).
b68b7282 973 *
54341a3b
MD
974 * Preemption is typically limited to interrupt threads.
975 *
976 * Operation works in a fairly straight-forward manner. The normal
977 * scheduling code is bypassed and we switch directly to the target
978 * thread. When the target thread attempts to block or switch away
979 * code at the base of lwkt_switch() will switch directly back to our
980 * thread. Our thread is able to retain whatever tokens it holds and
981 * if the target needs one of them the target will switch back to us
982 * and reschedule itself normally.
b68b7282
MD
983 */
984void
f9235b6d 985lwkt_preempt(thread_t ntd, int critcount)
b68b7282 986{
46a3f46d 987 struct globaldata *gd = mycpu;
cc9b6223 988 thread_t xtd;
0a3f9b47 989 thread_t td;
2d910aaf 990 int save_gd_intr_nesting_level;
b68b7282 991
26a0694b 992 /*
96728c05
MD
993 * The caller has put us in a critical section. We can only preempt
994 * if the caller of the caller was not in a critical section (basically
f9235b6d 995 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 996 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 997 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
998 *
999 * YYY The target thread must be in a critical section (else it must
1000 * inherit our critical section? I dunno yet).
26a0694b 1001 */
f9235b6d 1002 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 1003
b12defdc 1004 td = gd->gd_curthread;
fbc024e4
MD
1005 if (preempt_enable == 0) {
1006 ++preempt_miss;
1007 return;
1008 }
f9235b6d 1009 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
1010 ++preempt_miss;
1011 return;
1012 }
f9235b6d 1013 if (td->td_critcount > critcount) {
96728c05
MD
1014 ++preempt_miss;
1015 return;
1016 }
121f93bc
MD
1017 if (td->td_cscount) {
1018 ++preempt_miss;
1019 return;
1020 }
46a3f46d 1021 if (ntd->td_gd != gd) {
96728c05
MD
1022 ++preempt_miss;
1023 return;
1024 }
41a01a4d 1025 /*
77912481
MD
1026 * We don't have to check spinlocks here as they will also bump
1027 * td_critcount.
d3d1cbc8
MD
1028 *
1029 * Do not try to preempt if the target thread is holding any tokens.
1030 * We could try to acquire the tokens but this case is so rare there
1031 * is no need to support it.
41a01a4d 1032 */
0846e4ce 1033 KKASSERT(gd->gd_spinlocks == 0);
77912481 1034
3b998fa9 1035 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8 1036 ++preempt_miss;
d3d1cbc8
MD
1037 return;
1038 }
26a0694b
MD
1039 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1040 ++preempt_weird;
1041 return;
1042 }
1043 if (ntd->td_preempted) {
4b5f931b 1044 ++preempt_hit;
26a0694b 1045 return;
b68b7282 1046 }
da0b0e8b 1047 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 1048
8ec60c3f
MD
1049 /*
1050 * Since we are able to preempt the current thread, there is no need to
1051 * call need_lwkt_resched().
2d910aaf
MD
1052 *
1053 * We must temporarily clear gd_intr_nesting_level around the switch
1054 * since switchouts from the target thread are allowed (they will just
1055 * return to our thread), and since the target thread has its own stack.
cc9b6223
MD
1056 *
1057 * A preemption must switch back to the original thread, assert the
1058 * case.
8ec60c3f 1059 */
26a0694b
MD
1060 ++preempt_hit;
1061 ntd->td_preempted = td;
1062 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1063 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1064 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1065 gd->gd_intr_nesting_level = 0;
121f93bc
MD
1066
1067 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
cc9b6223
MD
1068 ntd->td_flags |= TDF_RUNNING;
1069 xtd = td->td_switch(ntd);
1070 KKASSERT(xtd == ntd);
1071 lwkt_switch_return(xtd);
2d910aaf 1072 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1073
26a0694b
MD
1074 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1075 ntd->td_preempted = NULL;
1076 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1077}
1078
f1d1c3fa 1079/*
faaeffac 1080 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1081 * This will work inside a critical section but not inside a hard code
1082 * section.
ef0fdad1 1083 *
f1d1c3fa
MD
1084 * (self contained on a per cpu basis)
1085 */
1086void
faaeffac 1087splz_check(void)
f1d1c3fa 1088{
7966cb69
MD
1089 globaldata_t gd = mycpu;
1090 thread_t td = gd->gd_curthread;
ef0fdad1 1091
4a28fe22
MD
1092 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1093 gd->gd_intr_nesting_level == 0 &&
1094 td->td_nest_count < 2)
1095 {
f1d1c3fa 1096 splz();
4a28fe22
MD
1097 }
1098}
1099
1100/*
1101 * This version is integrated into crit_exit, reqflags has already
1102 * been tested but td_critcount has not.
1103 *
1104 * We only want to execute the splz() on the 1->0 transition of
1105 * critcount and not in a hard code section or if too deeply nested.
925040f2 1106 *
0846e4ce 1107 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
4a28fe22
MD
1108 */
1109void
1110lwkt_maybe_splz(thread_t td)
1111{
1112 globaldata_t gd = td->td_gd;
1113
1114 if (td->td_critcount == 0 &&
1115 gd->gd_intr_nesting_level == 0 &&
1116 td->td_nest_count < 2)
1117 {
1118 splz();
1119 }
f1d1c3fa
MD
1120}
1121
e6546af9
MD
1122/*
1123 * Drivers which set up processing co-threads can call this function to
1124 * run the co-thread at a higher priority and to allow it to preempt
1125 * normal threads.
1126 */
1127void
1128lwkt_set_interrupt_support_thread(void)
1129{
1130 thread_t td = curthread;
1131
1132 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1133 td->td_flags |= TDF_INTTHREAD;
1134 td->td_preemptable = lwkt_preempt;
1135}
1136
1137
8ad65e08 1138/*
f9235b6d
MD
1139 * This function is used to negotiate a passive release of the current
1140 * process/lwp designation with the user scheduler, allowing the user
1141 * scheduler to schedule another user thread. The related kernel thread
1142 * (curthread) continues running in the released state.
8ad65e08
MD
1143 */
1144void
f9235b6d 1145lwkt_passive_release(struct thread *td)
8ad65e08 1146{
f9235b6d
MD
1147 struct lwp *lp = td->td_lwp;
1148
e3e6be1f 1149#ifndef NO_LWKT_SPLIT_USERPRI
f9235b6d
MD
1150 td->td_release = NULL;
1151 lwkt_setpri_self(TDPRI_KERN_USER);
d992c377
MD
1152#endif
1153
f9235b6d 1154 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1155}
1156
f9235b6d 1157
3824f392 1158/*
d2d8515b
MD
1159 * This implements a LWKT yield, allowing a kernel thread to yield to other
1160 * kernel threads at the same or higher priority. This function can be
1161 * called in a tight loop and will typically only yield once per tick.
f9235b6d 1162 *
d2d8515b
MD
1163 * Most kernel threads run at the same priority in order to allow equal
1164 * sharing.
f9235b6d
MD
1165 *
1166 * (self contained on a per cpu basis)
3824f392
MD
1167 */
1168void
f9235b6d 1169lwkt_yield(void)
3824f392 1170{
f9235b6d
MD
1171 globaldata_t gd = mycpu;
1172 thread_t td = gd->gd_curthread;
3824f392 1173
f9235b6d
MD
1174 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1175 splz();
85946b6c 1176 if (lwkt_resched_wanted()) {
f9235b6d
MD
1177 lwkt_schedule_self(curthread);
1178 lwkt_switch();
f9235b6d 1179 }
3824f392
MD
1180}
1181
40504122
MD
1182/*
1183 * The quick version processes pending interrupts and higher-priority
1184 * LWKT threads but will not round-robin same-priority LWKT threads.
de4d4cb0
MD
1185 *
1186 * When called while attempting to return to userland the only same-pri
1187 * threads are the ones which have already tried to become the current
1188 * user process.
40504122
MD
1189 */
1190void
1191lwkt_yield_quick(void)
1192{
1193 globaldata_t gd = mycpu;
1194 thread_t td = gd->gd_curthread;
1195
1196 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1197 splz();
1198 if (lwkt_resched_wanted()) {
9c99cb33 1199 crit_enter();
40504122
MD
1200 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1201 clear_lwkt_resched();
1202 } else {
1203 lwkt_schedule_self(curthread);
1204 lwkt_switch();
1205 }
9c99cb33 1206 crit_exit();
40504122
MD
1207 }
1208}
1209
3824f392 1210/*
f9235b6d
MD
1211 * This yield is designed for kernel threads with a user context.
1212 *
1213 * The kernel acting on behalf of the user is potentially cpu-bound,
1214 * this function will efficiently allow other threads to run and also
1215 * switch to other processes by releasing.
3824f392
MD
1216 *
1217 * The lwkt_user_yield() function is designed to have very low overhead
1218 * if no yield is determined to be needed.
1219 */
1220void
1221lwkt_user_yield(void)
1222{
f9235b6d
MD
1223 globaldata_t gd = mycpu;
1224 thread_t td = gd->gd_curthread;
1225
1226 /*
1227 * Always run any pending interrupts in case we are in a critical
1228 * section.
1229 */
1230 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1231 splz();
3824f392 1232
3824f392 1233 /*
f9235b6d
MD
1234 * Switch (which forces a release) if another kernel thread needs
1235 * the cpu, if userland wants us to resched, or if our kernel
1236 * quantum has run out.
3824f392 1237 */
f9235b6d 1238 if (lwkt_resched_wanted() ||
85946b6c 1239 user_resched_wanted())
f9235b6d 1240 {
3824f392 1241 lwkt_switch();
3824f392
MD
1242 }
1243
f9235b6d 1244#if 0
3824f392 1245 /*
f9235b6d
MD
1246 * Reacquire the current process if we are released.
1247 *
1248 * XXX not implemented atm. The kernel may be holding locks and such,
1249 * so we want the thread to continue to receive cpu.
3824f392 1250 */
f9235b6d
MD
1251 if (td->td_release == NULL && lp) {
1252 lp->lwp_proc->p_usched->acquire_curproc(lp);
1253 td->td_release = lwkt_passive_release;
1254 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1255 }
f9235b6d 1256#endif
b9eb1c19
MD
1257}
1258
8ad65e08 1259/*
f1d1c3fa
MD
1260 * Generic schedule. Possibly schedule threads belonging to other cpus and
1261 * deal with threads that might be blocked on a wait queue.
1262 *
0a3f9b47
MD
1263 * We have a little helper inline function which does additional work after
1264 * the thread has been enqueued, including dealing with preemption and
1265 * setting need_lwkt_resched() (which prevents the kernel from returning
1266 * to userland until it has processed higher priority threads).
6330a558
MD
1267 *
1268 * It is possible for this routine to be called after a failed _enqueue
1269 * (due to the target thread migrating, sleeping, or otherwise blocked).
1270 * We have to check that the thread is actually on the run queue!
8ad65e08 1271 */
0a3f9b47
MD
1272static __inline
1273void
85946b6c 1274_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
0a3f9b47 1275{
6330a558 1276 if (ntd->td_flags & TDF_RUNQ) {
85946b6c 1277 if (ntd->td_preemptable) {
f9235b6d 1278 ntd->td_preemptable(ntd, ccount); /* YYY +token */
6330a558 1279 }
0a3f9b47
MD
1280 }
1281}
1282
361d01dd 1283static __inline
8ad65e08 1284void
85946b6c 1285_lwkt_schedule(thread_t td)
8ad65e08 1286{
37af14fe
MD
1287 globaldata_t mygd = mycpu;
1288
cf709dd2
MD
1289 KASSERT(td != &td->td_gd->gd_idlethread,
1290 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1291 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1292 crit_enter_gd(mygd);
4643740a
MD
1293 KKASSERT(td->td_lwp == NULL ||
1294 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1295
37af14fe 1296 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1297 _lwkt_enqueue(td);
1298 } else {
f1d1c3fa 1299 /*
7cd8d145
MD
1300 * If we own the thread, there is no race (since we are in a
1301 * critical section). If we do not own the thread there might
1302 * be a race but the target cpu will deal with it.
f1d1c3fa 1303 */
7cd8d145 1304 if (td->td_gd == mygd) {
9d265729 1305 _lwkt_enqueue(td);
85946b6c 1306 _lwkt_schedule_post(mygd, td, 1);
f1d1c3fa 1307 } else {
e381e77c 1308 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1309 }
8ad65e08 1310 }
37af14fe 1311 crit_exit_gd(mygd);
8ad65e08
MD
1312}
1313
361d01dd
MD
1314void
1315lwkt_schedule(thread_t td)
1316{
85946b6c 1317 _lwkt_schedule(td);
361d01dd
MD
1318}
1319
1320void
85946b6c 1321lwkt_schedule_noresched(thread_t td) /* XXX not impl */
361d01dd 1322{
85946b6c 1323 _lwkt_schedule(td);
361d01dd
MD
1324}
1325
e381e77c
MD
1326/*
1327 * When scheduled remotely if frame != NULL the IPIQ is being
1328 * run via doreti or an interrupt then preemption can be allowed.
1329 *
1330 * To allow preemption we have to drop the critical section so only
1331 * one is present in _lwkt_schedule_post.
1332 */
1333static void
1334lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1335{
1336 thread_t td = curthread;
1337 thread_t ntd = arg;
1338
1339 if (frame && ntd->td_preemptable) {
1340 crit_exit_noyield(td);
85946b6c 1341 _lwkt_schedule(ntd);
e381e77c
MD
1342 crit_enter_quick(td);
1343 } else {
85946b6c 1344 _lwkt_schedule(ntd);
e381e77c
MD
1345 }
1346}
1347
d9eea1a5 1348/*
52eedfb5
MD
1349 * Thread migration using a 'Pull' method. The thread may or may not be
1350 * the current thread. It MUST be descheduled and in a stable state.
1351 * lwkt_giveaway() must be called on the cpu owning the thread.
1352 *
1353 * At any point after lwkt_giveaway() is called, the target cpu may
1354 * 'pull' the thread by calling lwkt_acquire().
1355 *
ae8e83e6
MD
1356 * We have to make sure the thread is not sitting on a per-cpu tsleep
1357 * queue or it will blow up when it moves to another cpu.
1358 *
52eedfb5 1359 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1360 */
52eedfb5
MD
1361void
1362lwkt_giveaway(thread_t td)
1363{
3b4192fb 1364 globaldata_t gd = mycpu;
52eedfb5 1365
3b4192fb
MD
1366 crit_enter_gd(gd);
1367 if (td->td_flags & TDF_TSLEEPQ)
1368 tsleep_remove(td);
1369 KKASSERT(td->td_gd == gd);
1370 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1371 td->td_flags |= TDF_MIGRATING;
1372 crit_exit_gd(gd);
52eedfb5
MD
1373}
1374
a2a5ad0d
MD
1375void
1376lwkt_acquire(thread_t td)
1377{
37af14fe
MD
1378 globaldata_t gd;
1379 globaldata_t mygd;
cc9b6223 1380 int retry = 10000000;
a2a5ad0d 1381
52eedfb5 1382 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1383 gd = td->td_gd;
37af14fe 1384 mygd = mycpu;
52eedfb5 1385 if (gd != mycpu) {
35238fa5 1386 cpu_lfence();
52eedfb5 1387 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1388 crit_enter_gd(mygd);
cfaeae2a 1389 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23 1390 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
df910c23 1391 lwkt_process_ipiq();
52eedfb5 1392 cpu_lfence();
cc9b6223
MD
1393 if (--retry == 0) {
1394 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1395 td, td->td_flags);
1396 retry = 10000000;
1397 }
a86ce0cd
MD
1398#ifdef _KERNEL_VIRTUAL
1399 pthread_yield();
1400#endif
df910c23 1401 }
cfaeae2a 1402 DEBUG_POP_INFO();
562273ea 1403 cpu_mfence();
37af14fe 1404 td->td_gd = mygd;
52eedfb5
MD
1405 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1406 td->td_flags &= ~TDF_MIGRATING;
1407 crit_exit_gd(mygd);
1408 } else {
1409 crit_enter_gd(mygd);
1410 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1411 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1412 crit_exit_gd(mygd);
a2a5ad0d
MD
1413 }
1414}
1415
f1d1c3fa
MD
1416/*
1417 * Generic deschedule. Descheduling threads other then your own should be
1418 * done only in carefully controlled circumstances. Descheduling is
1419 * asynchronous.
1420 *
1421 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1422 */
1423void
1424lwkt_deschedule(thread_t td)
1425{
f1d1c3fa
MD
1426 crit_enter();
1427 if (td == curthread) {
1428 _lwkt_dequeue(td);
1429 } else {
a72187e9 1430 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1431 _lwkt_dequeue(td);
1432 } else {
b8a98473 1433 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1434 }
1435 }
1436 crit_exit();
1437}
1438
4b5f931b
MD
1439/*
1440 * Set the target thread's priority. This routine does not automatically
1441 * switch to a higher priority thread, LWKT threads are not designed for
1442 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1443 */
1444void
1445lwkt_setpri(thread_t td, int pri)
1446{
f9235b6d
MD
1447 if (td->td_pri != pri) {
1448 KKASSERT(pri >= 0);
1449 crit_enter();
1450 if (td->td_flags & TDF_RUNQ) {
d2d8515b 1451 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1452 _lwkt_dequeue(td);
1453 td->td_pri = pri;
1454 _lwkt_enqueue(td);
1455 } else {
1456 td->td_pri = pri;
1457 }
1458 crit_exit();
26a0694b 1459 }
26a0694b
MD
1460}
1461
03bd0a5e
MD
1462/*
1463 * Set the initial priority for a thread prior to it being scheduled for
1464 * the first time. The thread MUST NOT be scheduled before or during
1465 * this call. The thread may be assigned to a cpu other then the current
1466 * cpu.
1467 *
1468 * Typically used after a thread has been created with TDF_STOPPREQ,
1469 * and before the thread is initially scheduled.
1470 */
1471void
1472lwkt_setpri_initial(thread_t td, int pri)
1473{
1474 KKASSERT(pri >= 0);
1475 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1476 td->td_pri = pri;
03bd0a5e
MD
1477}
1478
26a0694b
MD
1479void
1480lwkt_setpri_self(int pri)
1481{
1482 thread_t td = curthread;
1483
4b5f931b
MD
1484 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1485 crit_enter();
1486 if (td->td_flags & TDF_RUNQ) {
1487 _lwkt_dequeue(td);
f9235b6d 1488 td->td_pri = pri;
4b5f931b
MD
1489 _lwkt_enqueue(td);
1490 } else {
f9235b6d 1491 td->td_pri = pri;
4b5f931b
MD
1492 }
1493 crit_exit();
1494}
1495
f9235b6d 1496/*
85946b6c 1497 * hz tick scheduler clock for LWKT threads
f9235b6d
MD
1498 */
1499void
85946b6c 1500lwkt_schedulerclock(thread_t td)
f9235b6d 1501{
85946b6c
MD
1502 globaldata_t gd = td->td_gd;
1503 thread_t xtd;
2a418930 1504
85946b6c
MD
1505 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1506 /*
1507 * If the current thread is at the head of the runq shift it to the
1508 * end of any equal-priority threads and request a LWKT reschedule
1509 * if it moved.
d992c377
MD
1510 *
1511 * Ignore upri in this situation. There will only be one user thread
1512 * in user mode, all others will be user threads running in kernel
1513 * mode and we have to make sure they get some cpu.
85946b6c
MD
1514 */
1515 xtd = TAILQ_NEXT(td, td_threadq);
1516 if (xtd && xtd->td_pri == td->td_pri) {
1517 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1518 while (xtd && xtd->td_pri == td->td_pri)
1519 xtd = TAILQ_NEXT(xtd, td_threadq);
1520 if (xtd)
1521 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1522 else
1523 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1524 need_lwkt_resched();
f9235b6d 1525 }
85946b6c
MD
1526 } else {
1527 /*
1528 * If we scheduled a thread other than the one at the head of the
1529 * queue always request a reschedule every tick.
1530 */
1531 need_lwkt_resched();
f9235b6d
MD
1532 }
1533}
1534
5d21b981 1535/*
52eedfb5
MD
1536 * Migrate the current thread to the specified cpu.
1537 *
cc9b6223
MD
1538 * This is accomplished by descheduling ourselves from the current cpu
1539 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1540 * 'old' thread wants to migrate after it has been completely switched out
1541 * and will complete the migration.
1542 *
1543 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1544 *
1545 * We must be sure to release our current process designation (if a user
1546 * process) before clearing out any tsleepq we are on because the release
1547 * code may re-add us.
ae8e83e6
MD
1548 *
1549 * We must be sure to remove ourselves from the current cpu's tsleepq
1550 * before potentially moving to another queue. The thread can be on
1551 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1552 */
5d21b981
MD
1553
1554void
1555lwkt_setcpu_self(globaldata_t rgd)
1556{
5d21b981
MD
1557 thread_t td = curthread;
1558
1559 if (td->td_gd != rgd) {
1560 crit_enter_quick(td);
cc9b6223 1561
95858b91
MD
1562 if (td->td_release)
1563 td->td_release(td);
ae8e83e6 1564 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1565 tsleep_remove(td);
cc9b6223
MD
1566
1567 /*
1568 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1569 * trying to deschedule ourselves and switch away, then deschedule
1570 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1571 * call lwkt_switch() to complete the operation.
1572 */
5d21b981
MD
1573 td->td_flags |= TDF_MIGRATING;
1574 lwkt_deschedule_self(td);
52eedfb5 1575 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1576 td->td_migrate_gd = rgd;
5d21b981 1577 lwkt_switch();
cc9b6223
MD
1578
1579 /*
1580 * We are now on the target cpu
1581 */
1582 KKASSERT(rgd == mycpu);
52eedfb5 1583 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1584 crit_exit_quick(td);
1585 }
5d21b981
MD
1586}
1587
ecdefdda
MD
1588void
1589lwkt_migratecpu(int cpuid)
1590{
ecdefdda
MD
1591 globaldata_t rgd;
1592
1593 rgd = globaldata_find(cpuid);
1594 lwkt_setcpu_self(rgd);
ecdefdda
MD
1595}
1596
5d21b981
MD
1597/*
1598 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
1599 * do not have to enter another one).
1600 *
1601 * The thread (td) has already been completely descheduled from the
1602 * originating cpu and we can simply assert the case. The thread is
1603 * assigned to the new cpu and enqueued.
5d21b981 1604 *
cc9b6223 1605 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1606 */
1607static void
1608lwkt_setcpu_remote(void *arg)
1609{
1610 thread_t td = arg;
1611 globaldata_t gd = mycpu;
1612
cc9b6223 1613 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1614 td->td_gd = gd;
562273ea 1615 cpu_mfence();
5d21b981 1616 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1617 KKASSERT(td->td_migrate_gd == NULL);
4643740a
MD
1618 KKASSERT(td->td_lwp == NULL ||
1619 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
5d21b981
MD
1620 _lwkt_enqueue(td);
1621}
1622
553ea3c8 1623struct lwp *
4b5f931b
MD
1624lwkt_preempted_proc(void)
1625{
73e4f7b9 1626 thread_t td = curthread;
4b5f931b
MD
1627 while (td->td_preempted)
1628 td = td->td_preempted;
553ea3c8 1629 return(td->td_lwp);
4b5f931b
MD
1630}
1631
99df837e
MD
1632/*
1633 * Create a kernel process/thread/whatever. It shares it's address space
1634 * with proc0 - ie: kernel only.
1635 *
d2d8515b
MD
1636 * If the cpu is not specified one will be selected. In the future
1637 * specifying a cpu of -1 will enable kernel thread migration between
1638 * cpus.
99df837e
MD
1639 */
1640int
c9e9fb21
MD
1641lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1642 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1643{
73e4f7b9 1644 thread_t td;
e2565a42 1645 __va_list ap;
99df837e 1646
d3d32139 1647 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1648 tdflags);
a2a5ad0d
MD
1649 if (tdp)
1650 *tdp = td;
709799ea 1651 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1652
1653 /*
1654 * Set up arg0 for 'ps' etc
1655 */
e2565a42 1656 __va_start(ap, fmt);
379210cb 1657 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1658 __va_end(ap);
99df837e
MD
1659
1660 /*
1661 * Schedule the thread to run
1662 */
4643740a
MD
1663 if (td->td_flags & TDF_NOSTART)
1664 td->td_flags &= ~TDF_NOSTART;
ef0fdad1 1665 else
4643740a 1666 lwkt_schedule(td);
99df837e
MD
1667 return 0;
1668}
1669
1670/*
1671 * Destroy an LWKT thread. Warning! This function is not called when
1672 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1673 * uses a different reaping mechanism.
1674 */
1675void
1676lwkt_exit(void)
1677{
1678 thread_t td = curthread;
c070746a 1679 thread_t std;
8826f33a 1680 globaldata_t gd;
99df837e 1681
2883d2d8
MD
1682 /*
1683 * Do any cleanup that might block here
1684 */
99df837e 1685 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1686 kprintf("kthread %p %s has exited\n", td, td->td_comm);
2883d2d8
MD
1687 biosched_done(td);
1688 dsched_exit_thread(td);
c070746a
MD
1689
1690 /*
1691 * Get us into a critical section to interlock gd_freetd and loop
1692 * until we can get it freed.
1693 *
1694 * We have to cache the current td in gd_freetd because objcache_put()ing
1695 * it would rip it out from under us while our thread is still active.
2af9d75d
MD
1696 *
1697 * We are the current thread so of course our own TDF_RUNNING bit will
1698 * be set, so unlike the lwp reap code we don't wait for it to clear.
c070746a
MD
1699 */
1700 gd = mycpu;
37af14fe 1701 crit_enter_quick(td);
2af9d75d
MD
1702 for (;;) {
1703 if (td->td_refs) {
1704 tsleep(td, 0, "tdreap", 1);
1705 continue;
1706 }
1707 if ((std = gd->gd_freetd) != NULL) {
1708 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1709 gd->gd_freetd = NULL;
1710 objcache_put(thread_cache, std);
1711 continue;
1712 }
1713 break;
c070746a 1714 }
3b4192fb
MD
1715
1716 /*
1717 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1718 * the last time. We cannot block after this point or we may end
1719 * up with a stale td on the tsleepq.
eb2adbf5
MD
1720 *
1721 * None of this may block, the critical section is the only thing
1722 * protecting tdallq and the only thing preventing new lwkt_hold()
1723 * thread refs now.
3b4192fb
MD
1724 */
1725 if (td->td_flags & TDF_TSLEEPQ)
1726 tsleep_remove(td);
37af14fe 1727 lwkt_deschedule_self(td);
e56e4dea 1728 lwkt_remove_tdallq(td);
74c9628e 1729 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1730
1731 /*
1732 * Final cleanup
1733 */
1734 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1735 if (td->td_flags & TDF_ALLOCATED_THREAD)
1736 gd->gd_freetd = td;
99df837e
MD
1737 cpu_thread_exit();
1738}
1739
e56e4dea
MD
1740void
1741lwkt_remove_tdallq(thread_t td)
1742{
1743 KKASSERT(td->td_gd == mycpu);
1744 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1745}
1746
9cf43f91
MD
1747/*
1748 * Code reduction and branch prediction improvements. Call/return
1749 * overhead on modern cpus often degenerates into 0 cycles due to
1750 * the cpu's branch prediction hardware and return pc cache. We
1751 * can take advantage of this by not inlining medium-complexity
1752 * functions and we can also reduce the branch prediction impact
1753 * by collapsing perfectly predictable branches into a single
1754 * procedure instead of duplicating it.
1755 *
1756 * Is any of this noticeable? Probably not, so I'll take the
1757 * smaller code size.
1758 */
1759void
b6468f56 1760crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1761{
b6468f56 1762 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1763}
1764
2d93b37a
MD
1765void
1766crit_panic(void)
1767{
1768 thread_t td = curthread;
850634cc 1769 int lcrit = td->td_critcount;
2d93b37a 1770
850634cc
AH
1771 td->td_critcount = 0;
1772 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1773 /* NOT REACHED */
2d93b37a
MD
1774}
1775
bd8015ca
MD
1776/*
1777 * Called from debugger/panic on cpus which have been stopped. We must still
1778 * process the IPIQ while stopped, even if we were stopped while in a critical
1779 * section (XXX).
1780 *
1781 * If we are dumping also try to process any pending interrupts. This may
1782 * or may not work depending on the state of the cpu at the point it was
1783 * stopped.
1784 */
1785void
1786lwkt_smp_stopped(void)
1787{
1788 globaldata_t gd = mycpu;
1789
1790 crit_enter_gd(gd);
1791 if (dumping) {
1792 lwkt_process_ipiq();
1793 splz();
1794 } else {
1795 lwkt_process_ipiq();
1796 }
1797 crit_exit_gd(gd);
1798}