kernel - Silence warnings from older gcc
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
3b998fa9 2 * Copyright (c) 2003-2010 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);
a1f0fb66
AE
79KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
80 sizeof(int) + sizeof(struct thread *));
81KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
82 sizeof(int) + sizeof(struct thread *));
83KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
84 sizeof (struct thread *) + sizeof(char *));
85KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
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;
f64b567c 96static __int64_t token_contention_count __debugvar = 0;
fb0f29c4 97static int lwkt_use_spin_port;
40aaf5fc 98static struct objcache *thread_cache;
05220613 99
88ebb169 100#ifdef SMP
e381e77c 101static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
88ebb169 102#endif
f9235b6d 103static void lwkt_fairq_accumulate(globaldata_t gd, thread_t td);
e381e77c 104
0855a2af
JG
105extern void cpu_heavy_restore(void);
106extern void cpu_lwkt_restore(void);
107extern void cpu_kthread_restore(void);
108extern void cpu_idle_restore(void);
109
fb0f29c4
MD
110/*
111 * We can make all thread ports use the spin backend instead of the thread
112 * backend. This should only be set to debug the spin backend.
113 */
114TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
115
0f7a3396 116#ifdef INVARIANTS
0c52fa62
SG
117SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
118 "Panic if attempting to switch lwkt's while mastering cpusync");
0f7a3396 119#endif
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.");
38717797
HP
128#ifdef INVARIANTS
129SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
130 &token_contention_count, 0, "spinning due to token contention");
38717797 131#endif
f9235b6d 132static int fairq_enable = 1;
2a418930
MD
133SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
134 &fairq_enable, 0, "Turn on fairq priority accumulators");
135static int lwkt_spin_loops = 10;
136SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
137 &lwkt_spin_loops, 0, "");
138static int lwkt_spin_delay = 1;
139SYSCTL_INT(_lwkt, OID_AUTO, spin_delay, CTLFLAG_RW,
140 &lwkt_spin_delay, 0, "Scheduler spin delay in microseconds 0=auto");
141static int lwkt_spin_method = 1;
142SYSCTL_INT(_lwkt, OID_AUTO, spin_method, CTLFLAG_RW,
143 &lwkt_spin_method, 0, "LWKT scheduler behavior when contended");
d5b2d319
MD
144static int lwkt_spin_fatal = 0; /* disabled */
145SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
146 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
fbc024e4 147static int preempt_enable = 1;
2a418930
MD
148SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
149 &preempt_enable, 0, "Enable preemption");
765b1ae0
MD
150static int lwkt_cache_threads = 32;
151SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
152 &lwkt_cache_threads, 0, "thread+kstack cache");
fbc024e4 153
2a418930
MD
154static __cachealign int lwkt_cseq_rindex;
155static __cachealign int lwkt_cseq_windex;
05220613 156
4b5f931b
MD
157/*
158 * These helper procedures handle the runq, they can only be called from
159 * within a critical section.
75cdbe6c
MD
160 *
161 * WARNING! Prior to SMP being brought up it is possible to enqueue and
162 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
163 * instead of 'mycpu' when referencing the globaldata structure. Once
164 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 165 */
f1d1c3fa
MD
166static __inline
167void
168_lwkt_dequeue(thread_t td)
169{
170 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 171 struct globaldata *gd = td->td_gd;
4b5f931b 172
f1d1c3fa 173 td->td_flags &= ~TDF_RUNQ;
f9235b6d
MD
174 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
175 gd->gd_fairq_total_pri -= td->td_pri;
176 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
2a418930 177 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
178 }
179}
180
f9235b6d
MD
181/*
182 * Priority enqueue.
183 *
184 * NOTE: There are a limited number of lwkt threads runnable since user
185 * processes only schedule one at a time per cpu.
186 */
f1d1c3fa
MD
187static __inline
188void
189_lwkt_enqueue(thread_t td)
190{
f9235b6d
MD
191 thread_t xtd;
192
7f5d7ed7 193 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 194 struct globaldata *gd = td->td_gd;
4b5f931b 195
f1d1c3fa 196 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
197 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
198 if (xtd == NULL) {
199 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
2a418930 200 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
f9235b6d
MD
201 } else {
202 while (xtd && xtd->td_pri > td->td_pri)
203 xtd = TAILQ_NEXT(xtd, td_threadq);
204 if (xtd)
205 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
206 else
207 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
208 }
209 gd->gd_fairq_total_pri += td->td_pri;
f1d1c3fa
MD
210 }
211}
8ad65e08 212
40aaf5fc
NT
213static __boolean_t
214_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
215{
216 struct thread *td = (struct thread *)obj;
217
218 td->td_kstack = NULL;
219 td->td_kstack_size = 0;
220 td->td_flags = TDF_ALLOCATED_THREAD;
221 return (1);
222}
223
224static void
225_lwkt_thread_dtor(void *obj, void *privdata)
226{
227 struct thread *td = (struct thread *)obj;
228
229 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
230 ("_lwkt_thread_dtor: not allocated from objcache"));
231 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
232 td->td_kstack_size > 0,
233 ("_lwkt_thread_dtor: corrupted stack"));
234 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
235}
236
237/*
238 * Initialize the lwkt s/system.
765b1ae0
MD
239 *
240 * Nominally cache up to 32 thread + kstack structures.
40aaf5fc
NT
241 */
242void
243lwkt_init(void)
244{
765b1ae0
MD
245 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
246 thread_cache = objcache_create_mbacked(
247 M_THREAD, sizeof(struct thread),
248 NULL, lwkt_cache_threads,
249 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
250}
251
37af14fe
MD
252/*
253 * Schedule a thread to run. As the current thread we can always safely
254 * schedule ourselves, and a shortcut procedure is provided for that
255 * function.
256 *
257 * (non-blocking, self contained on a per cpu basis)
258 */
259void
260lwkt_schedule_self(thread_t td)
261{
cfaeae2a 262 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 263 crit_enter_quick(td);
f9235b6d
MD
264 KASSERT(td != &td->td_gd->gd_idlethread,
265 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 266 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 267 _lwkt_enqueue(td);
37af14fe
MD
268 crit_exit_quick(td);
269}
270
271/*
272 * Deschedule a thread.
273 *
274 * (non-blocking, self contained on a per cpu basis)
275 */
276void
277lwkt_deschedule_self(thread_t td)
278{
279 crit_enter_quick(td);
37af14fe
MD
280 _lwkt_dequeue(td);
281 crit_exit_quick(td);
282}
283
8ad65e08
MD
284/*
285 * LWKTs operate on a per-cpu basis
286 *
73e4f7b9 287 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
288 */
289void
290lwkt_gdinit(struct globaldata *gd)
291{
f9235b6d 292 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 293 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
294}
295
7d0bac62
MD
296/*
297 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
298 * or LWKT start address before it can be scheduled. If the target cpu is
299 * -1 the thread will be created on the current cpu.
0cfcada1
MD
300 *
301 * If you intend to create a thread without a process context this function
302 * does everything except load the startup and switcher function.
7d0bac62
MD
303 */
304thread_t
d3d32139 305lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 306{
c070746a 307 globaldata_t gd = mycpu;
99df837e 308 void *stack;
7d0bac62 309
c070746a
MD
310 /*
311 * If static thread storage is not supplied allocate a thread. Reuse
312 * a cached free thread if possible. gd_freetd is used to keep an exiting
313 * thread intact through the exit.
314 */
ef0fdad1 315 if (td == NULL) {
cf709dd2
MD
316 crit_enter_gd(gd);
317 if ((td = gd->gd_freetd) != NULL) {
318 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
319 TDF_RUNQ)) == 0);
c070746a 320 gd->gd_freetd = NULL;
cf709dd2 321 } else {
c070746a 322 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
323 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
324 TDF_RUNQ)) == 0);
325 }
326 crit_exit_gd(gd);
40aaf5fc
NT
327 KASSERT((td->td_flags &
328 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
329 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
330 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 331 }
c070746a
MD
332
333 /*
334 * Try to reuse cached stack.
335 */
f470d0c8
MD
336 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
337 if (flags & TDF_ALLOCATED_STACK) {
e4846942 338 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
339 stack = NULL;
340 }
341 }
342 if (stack == NULL) {
e40cfbd7 343 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 344 flags |= TDF_ALLOCATED_STACK;
99df837e 345 }
75cdbe6c 346 if (cpu < 0)
c070746a 347 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 348 else
f470d0c8 349 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 350 return(td);
7d0bac62
MD
351}
352
353/*
354 * Initialize a preexisting thread structure. This function is used by
355 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
356 *
f8c3996b
MD
357 * All threads start out in a critical section at a priority of
358 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
359 * appropriate. This function may send an IPI message when the
360 * requested cpu is not the current cpu and consequently gd_tdallq may
361 * not be initialized synchronously from the point of view of the originating
362 * cpu.
363 *
364 * NOTE! we have to be careful in regards to creating threads for other cpus
365 * if SMP has not yet been activated.
7d0bac62 366 */
41a01a4d
MD
367#ifdef SMP
368
75cdbe6c
MD
369static void
370lwkt_init_thread_remote(void *arg)
371{
372 thread_t td = arg;
373
52eedfb5
MD
374 /*
375 * Protected by critical section held by IPI dispatch
376 */
75cdbe6c
MD
377 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
378}
379
41a01a4d
MD
380#endif
381
fdce8919
MD
382/*
383 * lwkt core thread structural initialization.
384 *
385 * NOTE: All threads are initialized as mpsafe threads.
386 */
7d0bac62 387void
f470d0c8
MD
388lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
389 struct globaldata *gd)
7d0bac62 390{
37af14fe
MD
391 globaldata_t mygd = mycpu;
392
99df837e
MD
393 bzero(td, sizeof(struct thread));
394 td->td_kstack = stack;
f470d0c8 395 td->td_kstack_size = stksize;
d3d32139 396 td->td_flags = flags;
26a0694b 397 td->td_gd = gd;
f9235b6d
MD
398 td->td_pri = TDPRI_KERN_DAEMON;
399 td->td_critcount = 1;
3b998fa9 400 td->td_toks_stop = &td->td_toks_base;
fb0f29c4
MD
401 if (lwkt_use_spin_port)
402 lwkt_initport_spin(&td->td_msgport);
403 else
404 lwkt_initport_thread(&td->td_msgport, td);
99df837e 405 pmap_init_thread(td);
0f7a3396 406#ifdef SMP
5d21b981
MD
407 /*
408 * Normally initializing a thread for a remote cpu requires sending an
409 * IPI. However, the idlethread is setup before the other cpus are
410 * activated so we have to treat it as a special case. XXX manipulation
411 * of gd_tdallq requires the BGL.
412 */
413 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 414 crit_enter_gd(mygd);
75cdbe6c 415 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 416 crit_exit_gd(mygd);
75cdbe6c 417 } else {
2db3b277 418 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 419 }
0f7a3396 420#else
37af14fe 421 crit_enter_gd(mygd);
0f7a3396 422 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 423 crit_exit_gd(mygd);
0f7a3396 424#endif
8c72e3d5
AH
425
426 dsched_new_thread(td);
73e4f7b9
MD
427}
428
429void
430lwkt_set_comm(thread_t td, const char *ctl, ...)
431{
e2565a42 432 __va_list va;
73e4f7b9 433
e2565a42 434 __va_start(va, ctl);
379210cb 435 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 436 __va_end(va);
e7c0dbba 437 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
438}
439
99df837e 440void
73e4f7b9 441lwkt_hold(thread_t td)
99df837e 442{
74c9628e 443 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
444}
445
446void
447lwkt_rele(thread_t td)
448{
449 KKASSERT(td->td_refs > 0);
74c9628e 450 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
451}
452
453void
454lwkt_wait_free(thread_t td)
455{
456 while (td->td_refs)
377d4740 457 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
458}
459
460void
461lwkt_free_thread(thread_t td)
462{
74c9628e 463 KKASSERT(td->td_refs == 0);
cf709dd2 464 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|TDF_RUNQ)) == 0);
40aaf5fc
NT
465 if (td->td_flags & TDF_ALLOCATED_THREAD) {
466 objcache_put(thread_cache, td);
467 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
468 /* client-allocated struct with internally allocated stack */
469 KASSERT(td->td_kstack && td->td_kstack_size > 0,
470 ("lwkt_free_thread: corrupted stack"));
471 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
472 td->td_kstack = NULL;
473 td->td_kstack_size = 0;
99df837e 474 }
e7c0dbba 475 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
476}
477
478
8ad65e08
MD
479/*
480 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
481 * switch to the idlethread. Switching must occur within a critical
482 * section to avoid races with the scheduling queue.
483 *
484 * We always have full control over our cpu's run queue. Other cpus
485 * that wish to manipulate our queue must use the cpu_*msg() calls to
486 * talk to our cpu, so a critical section is all that is needed and
487 * the result is very, very fast thread switching.
488 *
96728c05
MD
489 * The LWKT scheduler uses a fixed priority model and round-robins at
490 * each priority level. User process scheduling is a totally
491 * different beast and LWKT priorities should not be confused with
492 * user process priorities.
f1d1c3fa 493 *
69d78e99
MD
494 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
495 * is not called by the current thread in the preemption case, only when
496 * the preempting thread blocks (in order to return to the original thread).
cfaeae2a
MD
497 *
498 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
499 * migration and tsleep deschedule the current lwkt thread and call
500 * lwkt_switch(). In particular, the target cpu of the migration fully
501 * expects the thread to become non-runnable and can deadlock against
502 * cpusync operations if we run any IPIs prior to switching the thread out.
503 *
504 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
95858b91 505 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
8ad65e08
MD
506 */
507void
508lwkt_switch(void)
509{
37af14fe
MD
510 globaldata_t gd = mycpu;
511 thread_t td = gd->gd_curthread;
8ad65e08 512 thread_t ntd;
f9235b6d 513 thread_t xtd;
2a418930
MD
514 int spinning = lwkt_spin_loops; /* loops before HLTing */
515 int reqflags;
516 int cseq;
0f0466c0 517 int oseq;
d5b2d319 518 int fatal_count;
8ad65e08 519
46a3f46d 520 /*
27e88a6e
MD
521 * Switching from within a 'fast' (non thread switched) interrupt or IPI
522 * is illegal. However, we may have to do it anyway if we hit a fatal
523 * kernel trap or we have paniced.
524 *
525 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 526 */
27e88a6e
MD
527 if (gd->gd_intr_nesting_level) {
528 int savegdnest;
529 int savegdtrap;
530
5fddbda2 531 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22
MD
532 panic("lwkt_switch: Attempt to switch from a "
533 "a fast interrupt, ipi, or hard code section, "
534 "td %p\n",
535 td);
27e88a6e
MD
536 } else {
537 savegdnest = gd->gd_intr_nesting_level;
538 savegdtrap = gd->gd_trap_nesting_level;
539 gd->gd_intr_nesting_level = 0;
540 gd->gd_trap_nesting_level = 0;
a7422615
MD
541 if ((td->td_flags & TDF_PANICWARN) == 0) {
542 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
543 kprintf("Warning: thread switch from interrupt, IPI, "
544 "or hard code section.\n"
a7422615 545 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 546 print_backtrace(-1);
a7422615 547 }
27e88a6e
MD
548 lwkt_switch();
549 gd->gd_intr_nesting_level = savegdnest;
550 gd->gd_trap_nesting_level = savegdtrap;
551 return;
552 }
96728c05 553 }
ef0fdad1 554
cb973d15
MD
555 /*
556 * Passive release (used to transition from user to kernel mode
557 * when we block or switch rather then when we enter the kernel).
558 * This function is NOT called if we are switching into a preemption
559 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
560 * our current process designation (if we have one) and become a true
561 * LWKT thread, and may also hand the current process designation to
562 * another process and schedule thread.
cb973d15
MD
563 */
564 if (td->td_release)
565 td->td_release(td);
566
37af14fe 567 crit_enter_gd(gd);
3b998fa9 568 if (TD_TOKS_HELD(td))
9d265729
MD
569 lwkt_relalltokens(td);
570
571 /*
b02926de
MD
572 * We had better not be holding any spin locks, but don't get into an
573 * endless panic loop.
9d265729 574 */
d666840a
MD
575 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
576 ("lwkt_switch: still holding %d exclusive spinlocks!",
577 gd->gd_spinlocks_wr));
9d265729 578
8a8d5d85
MD
579
580#ifdef SMP
0f7a3396
MD
581#ifdef INVARIANTS
582 if (td->td_cscount) {
6ea70f76 583 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
584 td);
585 if (panic_on_cscount)
586 panic("switching while mastering cpusync");
587 }
588#endif
8a8d5d85 589#endif
f9235b6d
MD
590
591 /*
592 * If we had preempted another thread on this cpu, resume the preempted
593 * thread. This occurs transparently, whether the preempted thread
594 * was scheduled or not (it may have been preempted after descheduling
595 * itself).
596 *
597 * We have to setup the MP lock for the original thread after backing
598 * out the adjustment that was made to curthread when the original
599 * was preempted.
600 */
99df837e 601 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
602 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
603 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
604
605 /*
b9eb1c19
MD
606 * The interrupt may have woken a thread up, we need to properly
607 * set the reschedule flag if the originally interrupted thread is
608 * at a lower priority.
8ec60c3f 609 */
f9235b6d
MD
610 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
611 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
8ec60c3f 612 need_lwkt_resched();
f9235b6d 613 }
8a8d5d85 614 /* YYY release mp lock on switchback if original doesn't need it */
f9235b6d
MD
615 goto havethread_preempted;
616 }
617
618 /*
619 * Implement round-robin fairq with priority insertion. The priority
620 * insertion is handled by _lwkt_enqueue()
621 *
f9235b6d 622 * If we cannot obtain ownership of the tokens we cannot immediately
cfaeae2a
MD
623 * schedule the target thread.
624 *
625 * Reminder: Again, we cannot afford to run any IPIs in this path if
626 * the current thread has been descheduled.
f9235b6d
MD
627 */
628 for (;;) {
2a418930
MD
629 /*
630 * Clear RQF_AST_LWKT_RESCHED (we handle the reschedule request)
631 * and set RQF_WAKEUP (prevent unnecessary IPIs from being
632 * received).
633 */
634 for (;;) {
635 reqflags = gd->gd_reqflags;
636 if (atomic_cmpset_int(&gd->gd_reqflags, reqflags,
637 (reqflags & ~RQF_AST_LWKT_RESCHED) |
638 RQF_WAKEUP)) {
639 break;
640 }
641 }
f9235b6d 642
4b5f931b 643 /*
2a418930
MD
644 * Hotpath - pull the head of the run queue and attempt to schedule
645 * it. Fairq exhaustion moves the task to the end of the list. If
646 * no threads are runnable we switch to the idle thread.
41a01a4d 647 */
2a418930
MD
648 for (;;) {
649 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
650
651 if (ntd == NULL) {
652 /*
653 * Runq is empty, switch to idle and clear RQF_WAKEUP
654 * to allow it to halt.
655 */
656 ntd = &gd->gd_idlethread;
6f207a2c 657#ifdef SMP
2a418930 658 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
b5d16701 659 ASSERT_NO_TOKENS_HELD(ntd);
6f207a2c 660#endif
2a418930
MD
661 cpu_time.cp_msg[0] = 0;
662 cpu_time.cp_stallpc = 0;
663 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP);
664 goto haveidle;
665 }
666
667 if (ntd->td_fairq_accum >= 0)
668 break;
669
cfaeae2a 670 /*splz_check(); cannot do this here, see above */
2a418930
MD
671 lwkt_fairq_accumulate(gd, ntd);
672 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
673 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
f9235b6d 674 }
41a01a4d 675
8ec60c3f 676 /*
2a418930
MD
677 * Hotpath - schedule ntd. Leaves RQF_WAKEUP set to prevent
678 * unwanted decontention IPIs.
6f207a2c
MD
679 *
680 * NOTE: For UP there is no mplock and lwkt_getalltokens()
681 * always succeeds.
8ec60c3f 682 */
2a418930 683 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
f9235b6d 684 goto havethread;
f9235b6d 685
f9235b6d 686 /*
2a418930
MD
687 * Coldpath (SMP only since tokens always succeed on UP)
688 *
689 * We had some contention on the thread we wanted to schedule.
690 * What we do now is try to find a thread that we can schedule
691 * in its stead until decontention reschedules on our cpu.
692 *
693 * The coldpath scan does NOT rearrange threads in the run list
694 * and it also ignores the accumulator.
695 *
696 * We do not immediately schedule a user priority thread, instead
697 * we record it in xtd and continue looking for kernel threads.
698 * A cpu can only have one user priority thread (normally) so just
699 * record the first one.
700 *
701 * NOTE: This scan will also include threads whos fairq's were
702 * accumulated in the first loop.
f9235b6d 703 */
2a418930
MD
704 ++token_contention_count;
705 xtd = NULL;
706 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
41a01a4d 707 /*
2a418930
MD
708 * Try to switch to this thread. If the thread is running at
709 * user priority we clear WAKEUP to allow decontention IPIs
710 * (since this thread is simply running until the one we wanted
711 * decontends), and we make sure that LWKT_RESCHED is not set.
df6b8ba0 712 *
2a418930
MD
713 * Otherwise for kernel threads we leave WAKEUP set to avoid
714 * unnecessary decontention IPIs.
41a01a4d 715 */
2a418930
MD
716 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
717 if (xtd == NULL)
718 xtd = ntd;
719 continue;
f9235b6d 720 }
a453459d 721
f9235b6d 722 /*
2a418930
MD
723 * Do not let the fairq get too negative. Even though we are
724 * ignoring it atm once the scheduler decontends a very negative
725 * thread will get moved to the end of the queue.
f9235b6d 726 */
2a418930
MD
727 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd)) {
728 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd))
729 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd);
730 goto havethread;
8a8d5d85 731 }
f9235b6d 732
df6b8ba0 733 /*
2a418930 734 * Well fubar, this thread is contended as well, loop
df6b8ba0 735 */
2a418930
MD
736 /* */
737 }
738
739 /*
740 * We exhausted the run list but we may have recorded a user
741 * thread to try. We have three choices based on
742 * lwkt.decontention_method.
743 *
744 * (0) Atomically clear RQF_WAKEUP in order to receive decontention
745 * IPIs (to interrupt the user process) and test
746 * RQF_AST_LWKT_RESCHED at the same time.
747 *
748 * This results in significant decontention IPI traffic but may
749 * be more responsive.
750 *
751 * (1) Leave RQF_WAKEUP set so we do not receive a decontention IPI.
752 * An automatic LWKT reschedule will occur on the next hardclock
753 * (typically 100hz).
754 *
755 * This results in no decontention IPI traffic but may be less
756 * responsive. This is the default.
757 *
758 * (2) Refuse to schedule the user process at this time.
759 *
760 * This is highly experimental and should not be used under
761 * normal circumstances. This can cause a user process to
762 * get starved out in situations where kernel threads are
763 * fighting each other for tokens.
764 */
765 if (xtd) {
766 ntd = xtd;
767
768 switch(lwkt_spin_method) {
769 case 0:
770 for (;;) {
771 reqflags = gd->gd_reqflags;
772 if (atomic_cmpset_int(&gd->gd_reqflags,
773 reqflags,
774 reqflags & ~RQF_WAKEUP)) {
775 break;
776 }
777 }
778 break;
779 case 1:
780 reqflags = gd->gd_reqflags;
781 break;
782 default:
783 goto skip;
784 break;
785 }
786 if ((reqflags & RQF_AST_LWKT_RESCHED) == 0 &&
b5d16701 787 (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
f9235b6d 788 ) {
2a418930
MD
789 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd))
790 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd);
791 goto havethread;
df6b8ba0 792 }
9ac1ee6e 793
2a418930 794skip:
9ac1ee6e 795 /*
2a418930
MD
796 * Make sure RQF_WAKEUP is set if we failed to schedule the
797 * user thread to prevent the idle thread from halting.
9ac1ee6e 798 */
2a418930
MD
799 atomic_set_int(&gd->gd_reqflags, RQF_WAKEUP);
800 }
801
802 /*
803 * We exhausted the run list, meaning that all runnable threads
804 * are contended.
805 */
806 cpu_pause();
807 ntd = &gd->gd_idlethread;
808#ifdef SMP
809 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
810 ASSERT_NO_TOKENS_HELD(ntd);
811 /* contention case, do not clear contention mask */
812#endif
813
814 /*
815 * Ok, we might want to spin a few times as some tokens are held for
816 * very short periods of time and IPI overhead is 1uS or worse
817 * (meaning it is usually better to spin). Regardless we have to
818 * call splz_check() to be sure to service any interrupts blocked
819 * by our critical section, otherwise we could livelock e.g. IPIs.
820 *
821 * The IPI mechanic is really a last resort. In nearly all other
822 * cases RQF_WAKEUP is left set to prevent decontention IPIs.
823 *
824 * When we decide not to spin we clear RQF_WAKEUP and switch to
825 * the idle thread. Clearing RQF_WEAKEUP allows the idle thread
826 * to halt and decontended tokens will issue an IPI to us. The
827 * idle thread will check for pending reschedules already set
828 * (RQF_AST_LWKT_RESCHED) before actually halting so we don't have
829 * to here.
cfaeae2a
MD
830 *
831 * Also, if TDF_RUNQ is not set the current thread is trying to
832 * deschedule, possibly in an atomic fashion. We cannot afford to
833 * stay here.
2a418930 834 */
cfaeae2a 835 if (spinning <= 0 || (td->td_flags & TDF_RUNQ) == 0) {
2a418930
MD
836 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP);
837 goto haveidle;
4b5f931b 838 }
2a418930 839 --spinning;
c5724852
MD
840
841 /*
2a418930
MD
842 * When spinning a delay is required both to avoid livelocks from
843 * token order reversals (a thread may be trying to acquire multiple
844 * tokens), and also to reduce cpu cache management traffic.
845 *
846 * In order to scale to a large number of CPUs we use a time slot
847 * resequencer to force contending cpus into non-contending
848 * time-slots. The scheduler may still contend with the lock holder
849 * but will not (generally) contend with all the other cpus trying
850 * trying to get the same token.
851 *
852 * The resequencer uses a FIFO counter mechanic. The owner of the
853 * rindex at the head of the FIFO is allowed to pull itself off
854 * the FIFO and fetchadd is used to enter into the FIFO. This bit
855 * of code is VERY cache friendly and forces all spinning schedulers
856 * into their own time slots.
c5724852 857 *
2a418930
MD
858 * This code has been tested to 48-cpus and caps the cache
859 * contention load at ~1uS intervals regardless of the number of
860 * cpus. Scaling beyond 64 cpus might require additional smarts
861 * (such as separate FIFOs for specific token cases).
862 *
863 * WARNING! We can't call splz_check() or anything else here as
864 * it could cause a deadlock.
c5724852 865 */
bd52bedf 866#if defined(INVARIANTS) && defined(__amd64__)
d5b2d319
MD
867 if ((read_rflags() & PSL_I) == 0) {
868 cpu_enable_intr();
869 panic("lwkt_switch() called with interrupts disabled");
870 }
871#endif
2a418930 872 cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
d5b2d319 873 fatal_count = lwkt_spin_fatal;
0f0466c0
MD
874 while ((oseq = lwkt_cseq_rindex) != cseq) {
875 cpu_ccfence();
8b402283 876#if !defined(_KERNEL_VIRTUAL)
0f0466c0
MD
877 if (cpu_mi_feature & CPU_MI_MONITOR) {
878 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
8b402283
SW
879 } else
880#endif
881 {
0f0466c0
MD
882 DELAY(1);
883 cpu_lfence();
884 }
d5b2d319
MD
885 if (fatal_count && --fatal_count == 0)
886 panic("lwkt_switch: fatal spin wait");
2a418930
MD
887 }
888 cseq = lwkt_spin_delay; /* don't trust the system operator */
889 cpu_ccfence();
890 if (cseq < 1)
891 cseq = 1;
892 if (cseq > 1000)
893 cseq = 1000;
894 DELAY(cseq);
895 atomic_add_int(&lwkt_cseq_rindex, 1);
cfaeae2a 896 splz_check(); /* ok, we already checked that td is still scheduled */
2a418930 897 /* highest level for(;;) loop */
f1d1c3fa 898 }
8a8d5d85 899
2a418930 900havethread:
8a8d5d85 901 /*
f9235b6d
MD
902 * We must always decrement td_fairq_accum on non-idle threads just
903 * in case a thread never gets a tick due to being in a continuous
2a418930 904 * critical section. The page-zeroing code does this, for example.
f9235b6d
MD
905 *
906 * If the thread we came up with is a higher or equal priority verses
907 * the thread at the head of the queue we move our thread to the
908 * front. This way we can always check the front of the queue.
be71787b
MD
909 *
910 * Clear gd_idle_repeat when doing a normal switch to a non-idle
911 * thread.
f9235b6d 912 */
f9235b6d
MD
913 ++gd->gd_cnt.v_swtch;
914 --ntd->td_fairq_accum;
9ac1ee6e 915 ntd->td_wmesg = NULL;
f9235b6d
MD
916 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
917 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
918 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
919 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
920 }
be71787b 921 gd->gd_idle_repeat = 0;
2a418930 922
f9235b6d 923havethread_preempted:
f9235b6d
MD
924 /*
925 * If the new target does not need the MP lock and we are holding it,
926 * release the MP lock. If the new target requires the MP lock we have
927 * already acquired it for the target.
8a8d5d85 928 */
2a418930 929 ;
f9235b6d
MD
930haveidle:
931 KASSERT(ntd->td_critcount,
b5d16701
MD
932 ("priority problem in lwkt_switch %d %d",
933 td->td_critcount, ntd->td_critcount));
934
94f6d86e
MD
935 if (td != ntd) {
936 ++switch_count;
a1f0fb66 937 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
f1d1c3fa 938 td->td_switch(ntd);
94f6d86e 939 }
37af14fe
MD
940 /* NOTE: current cpu may have changed after switch */
941 crit_exit_quick(td);
8ad65e08
MD
942}
943
b68b7282 944/*
96728c05
MD
945 * Request that the target thread preempt the current thread. Preemption
946 * only works under a specific set of conditions:
b68b7282 947 *
96728c05
MD
948 * - We are not preempting ourselves
949 * - The target thread is owned by the current cpu
950 * - We are not currently being preempted
951 * - The target is not currently being preempted
d3d1cbc8
MD
952 * - We are not holding any spin locks
953 * - The target thread is not holding any tokens
96728c05
MD
954 * - We are able to satisfy the target's MP lock requirements (if any).
955 *
956 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
957 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 958 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
959 * to determine whether preemption is possible (aka usually just the crit
960 * priority of lwkt_schedule() itself).
b68b7282 961 *
26a0694b
MD
962 * XXX at the moment we run the target thread in a critical section during
963 * the preemption in order to prevent the target from taking interrupts
964 * that *WE* can't. Preemption is strictly limited to interrupt threads
965 * and interrupt-like threads, outside of a critical section, and the
966 * preempted source thread will be resumed the instant the target blocks
967 * whether or not the source is scheduled (i.e. preemption is supposed to
968 * be as transparent as possible).
b68b7282
MD
969 */
970void
f9235b6d 971lwkt_preempt(thread_t ntd, int critcount)
b68b7282 972{
46a3f46d 973 struct globaldata *gd = mycpu;
0a3f9b47 974 thread_t td;
2d910aaf 975 int save_gd_intr_nesting_level;
b68b7282 976
26a0694b 977 /*
96728c05
MD
978 * The caller has put us in a critical section. We can only preempt
979 * if the caller of the caller was not in a critical section (basically
f9235b6d 980 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 981 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 982 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
983 *
984 * YYY The target thread must be in a critical section (else it must
985 * inherit our critical section? I dunno yet).
41a01a4d 986 *
0a3f9b47 987 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b 988 */
f9235b6d 989 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 990
fbc024e4
MD
991 if (preempt_enable == 0) {
992 ++preempt_miss;
993 return;
994 }
995
0a3f9b47 996 td = gd->gd_curthread;
f9235b6d 997 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
998 ++preempt_miss;
999 return;
1000 }
f9235b6d 1001 if (td->td_critcount > critcount) {
96728c05 1002 ++preempt_miss;
8ec60c3f 1003 need_lwkt_resched();
96728c05
MD
1004 return;
1005 }
1006#ifdef SMP
46a3f46d 1007 if (ntd->td_gd != gd) {
96728c05 1008 ++preempt_miss;
8ec60c3f 1009 need_lwkt_resched();
96728c05
MD
1010 return;
1011 }
1012#endif
41a01a4d 1013 /*
77912481
MD
1014 * We don't have to check spinlocks here as they will also bump
1015 * td_critcount.
d3d1cbc8
MD
1016 *
1017 * Do not try to preempt if the target thread is holding any tokens.
1018 * We could try to acquire the tokens but this case is so rare there
1019 * is no need to support it.
41a01a4d 1020 */
77912481
MD
1021 KKASSERT(gd->gd_spinlocks_wr == 0);
1022
3b998fa9 1023 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8
MD
1024 ++preempt_miss;
1025 need_lwkt_resched();
1026 return;
1027 }
26a0694b
MD
1028 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1029 ++preempt_weird;
8ec60c3f 1030 need_lwkt_resched();
26a0694b
MD
1031 return;
1032 }
1033 if (ntd->td_preempted) {
4b5f931b 1034 ++preempt_hit;
8ec60c3f 1035 need_lwkt_resched();
26a0694b 1036 return;
b68b7282 1037 }
26a0694b 1038
8ec60c3f
MD
1039 /*
1040 * Since we are able to preempt the current thread, there is no need to
1041 * call need_lwkt_resched().
2d910aaf
MD
1042 *
1043 * We must temporarily clear gd_intr_nesting_level around the switch
1044 * since switchouts from the target thread are allowed (they will just
1045 * return to our thread), and since the target thread has its own stack.
8ec60c3f 1046 */
26a0694b
MD
1047 ++preempt_hit;
1048 ntd->td_preempted = td;
1049 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1050 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1051 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1052 gd->gd_intr_nesting_level = 0;
26a0694b 1053 td->td_switch(ntd);
2d910aaf 1054 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1055
26a0694b
MD
1056 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1057 ntd->td_preempted = NULL;
1058 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1059}
1060
f1d1c3fa 1061/*
faaeffac 1062 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1063 * This will work inside a critical section but not inside a hard code
1064 * section.
ef0fdad1 1065 *
f1d1c3fa
MD
1066 * (self contained on a per cpu basis)
1067 */
1068void
faaeffac 1069splz_check(void)
f1d1c3fa 1070{
7966cb69
MD
1071 globaldata_t gd = mycpu;
1072 thread_t td = gd->gd_curthread;
ef0fdad1 1073
4a28fe22
MD
1074 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1075 gd->gd_intr_nesting_level == 0 &&
1076 td->td_nest_count < 2)
1077 {
f1d1c3fa 1078 splz();
4a28fe22
MD
1079 }
1080}
1081
1082/*
1083 * This version is integrated into crit_exit, reqflags has already
1084 * been tested but td_critcount has not.
1085 *
1086 * We only want to execute the splz() on the 1->0 transition of
1087 * critcount and not in a hard code section or if too deeply nested.
1088 */
1089void
1090lwkt_maybe_splz(thread_t td)
1091{
1092 globaldata_t gd = td->td_gd;
1093
1094 if (td->td_critcount == 0 &&
1095 gd->gd_intr_nesting_level == 0 &&
1096 td->td_nest_count < 2)
1097 {
1098 splz();
1099 }
f1d1c3fa
MD
1100}
1101
8ad65e08 1102/*
f9235b6d
MD
1103 * This function is used to negotiate a passive release of the current
1104 * process/lwp designation with the user scheduler, allowing the user
1105 * scheduler to schedule another user thread. The related kernel thread
1106 * (curthread) continues running in the released state.
8ad65e08
MD
1107 */
1108void
f9235b6d 1109lwkt_passive_release(struct thread *td)
8ad65e08 1110{
f9235b6d
MD
1111 struct lwp *lp = td->td_lwp;
1112
1113 td->td_release = NULL;
1114 lwkt_setpri_self(TDPRI_KERN_USER);
1115 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1116}
1117
f9235b6d 1118
3824f392 1119/*
f9235b6d
MD
1120 * This implements a normal yield. This routine is virtually a nop if
1121 * there is nothing to yield to but it will always run any pending interrupts
1122 * if called from a critical section.
1123 *
1124 * This yield is designed for kernel threads without a user context.
1125 *
1126 * (self contained on a per cpu basis)
3824f392
MD
1127 */
1128void
f9235b6d 1129lwkt_yield(void)
3824f392 1130{
f9235b6d
MD
1131 globaldata_t gd = mycpu;
1132 thread_t td = gd->gd_curthread;
1133 thread_t xtd;
3824f392 1134
f9235b6d
MD
1135 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1136 splz();
1137 if (td->td_fairq_accum < 0) {
1138 lwkt_schedule_self(curthread);
1139 lwkt_switch();
1140 } else {
1141 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1142 if (xtd && xtd->td_pri > td->td_pri) {
1143 lwkt_schedule_self(curthread);
1144 lwkt_switch();
1145 }
1146 }
3824f392
MD
1147}
1148
1149/*
f9235b6d
MD
1150 * This yield is designed for kernel threads with a user context.
1151 *
1152 * The kernel acting on behalf of the user is potentially cpu-bound,
1153 * this function will efficiently allow other threads to run and also
1154 * switch to other processes by releasing.
3824f392
MD
1155 *
1156 * The lwkt_user_yield() function is designed to have very low overhead
1157 * if no yield is determined to be needed.
1158 */
1159void
1160lwkt_user_yield(void)
1161{
f9235b6d
MD
1162 globaldata_t gd = mycpu;
1163 thread_t td = gd->gd_curthread;
1164
1165 /*
1166 * Always run any pending interrupts in case we are in a critical
1167 * section.
1168 */
1169 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1170 splz();
3824f392 1171
3824f392 1172 /*
f9235b6d
MD
1173 * Switch (which forces a release) if another kernel thread needs
1174 * the cpu, if userland wants us to resched, or if our kernel
1175 * quantum has run out.
3824f392 1176 */
f9235b6d
MD
1177 if (lwkt_resched_wanted() ||
1178 user_resched_wanted() ||
1179 td->td_fairq_accum < 0)
1180 {
3824f392 1181 lwkt_switch();
3824f392
MD
1182 }
1183
f9235b6d 1184#if 0
3824f392 1185 /*
f9235b6d
MD
1186 * Reacquire the current process if we are released.
1187 *
1188 * XXX not implemented atm. The kernel may be holding locks and such,
1189 * so we want the thread to continue to receive cpu.
3824f392 1190 */
f9235b6d
MD
1191 if (td->td_release == NULL && lp) {
1192 lp->lwp_proc->p_usched->acquire_curproc(lp);
1193 td->td_release = lwkt_passive_release;
1194 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1195 }
f9235b6d 1196#endif
b9eb1c19
MD
1197}
1198
8ad65e08 1199/*
f1d1c3fa
MD
1200 * Generic schedule. Possibly schedule threads belonging to other cpus and
1201 * deal with threads that might be blocked on a wait queue.
1202 *
0a3f9b47
MD
1203 * We have a little helper inline function which does additional work after
1204 * the thread has been enqueued, including dealing with preemption and
1205 * setting need_lwkt_resched() (which prevents the kernel from returning
1206 * to userland until it has processed higher priority threads).
6330a558
MD
1207 *
1208 * It is possible for this routine to be called after a failed _enqueue
1209 * (due to the target thread migrating, sleeping, or otherwise blocked).
1210 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1211 *
1212 * reschedok is an optimized constant propagated from lwkt_schedule() or
1213 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1214 * reschedule to be requested if the target thread has a higher priority.
1215 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1216 * be 0, prevented undesired reschedules.
8ad65e08 1217 */
0a3f9b47
MD
1218static __inline
1219void
f9235b6d 1220_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
0a3f9b47 1221{
b9eb1c19 1222 thread_t otd;
c730be20 1223
6330a558 1224 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1225 if (ntd->td_preemptable && reschedok) {
f9235b6d 1226 ntd->td_preemptable(ntd, ccount); /* YYY +token */
361d01dd 1227 } else if (reschedok) {
b9eb1c19 1228 otd = curthread;
f9235b6d 1229 if (ntd->td_pri > otd->td_pri)
c730be20 1230 need_lwkt_resched();
6330a558 1231 }
f9235b6d
MD
1232
1233 /*
1234 * Give the thread a little fair share scheduler bump if it
1235 * has been asleep for a while. This is primarily to avoid
1236 * a degenerate case for interrupt threads where accumulator
1237 * crosses into negative territory unnecessarily.
1238 */
1239 if (ntd->td_fairq_lticks != ticks) {
1240 ntd->td_fairq_lticks = ticks;
1241 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1242 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1243 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1244 }
0a3f9b47
MD
1245 }
1246}
1247
361d01dd 1248static __inline
8ad65e08 1249void
361d01dd 1250_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1251{
37af14fe
MD
1252 globaldata_t mygd = mycpu;
1253
cf709dd2
MD
1254 KASSERT(td != &td->td_gd->gd_idlethread,
1255 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1256 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1257 crit_enter_gd(mygd);
9388413d 1258 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1259 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1260 _lwkt_enqueue(td);
1261 } else {
f1d1c3fa 1262 /*
7cd8d145
MD
1263 * If we own the thread, there is no race (since we are in a
1264 * critical section). If we do not own the thread there might
1265 * be a race but the target cpu will deal with it.
f1d1c3fa 1266 */
0f7a3396 1267#ifdef SMP
7cd8d145 1268 if (td->td_gd == mygd) {
9d265729 1269 _lwkt_enqueue(td);
f9235b6d 1270 _lwkt_schedule_post(mygd, td, 1, reschedok);
f1d1c3fa 1271 } else {
e381e77c 1272 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1273 }
0f7a3396 1274#else
7cd8d145 1275 _lwkt_enqueue(td);
f9235b6d 1276 _lwkt_schedule_post(mygd, td, 1, reschedok);
0f7a3396 1277#endif
8ad65e08 1278 }
37af14fe 1279 crit_exit_gd(mygd);
8ad65e08
MD
1280}
1281
361d01dd
MD
1282void
1283lwkt_schedule(thread_t td)
1284{
1285 _lwkt_schedule(td, 1);
1286}
1287
1288void
1289lwkt_schedule_noresched(thread_t td)
1290{
1291 _lwkt_schedule(td, 0);
1292}
1293
88ebb169
SW
1294#ifdef SMP
1295
e381e77c
MD
1296/*
1297 * When scheduled remotely if frame != NULL the IPIQ is being
1298 * run via doreti or an interrupt then preemption can be allowed.
1299 *
1300 * To allow preemption we have to drop the critical section so only
1301 * one is present in _lwkt_schedule_post.
1302 */
1303static void
1304lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1305{
1306 thread_t td = curthread;
1307 thread_t ntd = arg;
1308
1309 if (frame && ntd->td_preemptable) {
1310 crit_exit_noyield(td);
1311 _lwkt_schedule(ntd, 1);
1312 crit_enter_quick(td);
1313 } else {
1314 _lwkt_schedule(ntd, 1);
1315 }
1316}
1317
d9eea1a5 1318/*
52eedfb5
MD
1319 * Thread migration using a 'Pull' method. The thread may or may not be
1320 * the current thread. It MUST be descheduled and in a stable state.
1321 * lwkt_giveaway() must be called on the cpu owning the thread.
1322 *
1323 * At any point after lwkt_giveaway() is called, the target cpu may
1324 * 'pull' the thread by calling lwkt_acquire().
1325 *
ae8e83e6
MD
1326 * We have to make sure the thread is not sitting on a per-cpu tsleep
1327 * queue or it will blow up when it moves to another cpu.
1328 *
52eedfb5 1329 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1330 */
52eedfb5
MD
1331void
1332lwkt_giveaway(thread_t td)
1333{
3b4192fb 1334 globaldata_t gd = mycpu;
52eedfb5 1335
3b4192fb
MD
1336 crit_enter_gd(gd);
1337 if (td->td_flags & TDF_TSLEEPQ)
1338 tsleep_remove(td);
1339 KKASSERT(td->td_gd == gd);
1340 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1341 td->td_flags |= TDF_MIGRATING;
1342 crit_exit_gd(gd);
52eedfb5
MD
1343}
1344
a2a5ad0d
MD
1345void
1346lwkt_acquire(thread_t td)
1347{
37af14fe
MD
1348 globaldata_t gd;
1349 globaldata_t mygd;
a2a5ad0d 1350
52eedfb5 1351 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1352 gd = td->td_gd;
37af14fe 1353 mygd = mycpu;
52eedfb5 1354 if (gd != mycpu) {
35238fa5 1355 cpu_lfence();
52eedfb5 1356 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1357 crit_enter_gd(mygd);
cfaeae2a 1358 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23
MD
1359 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1360#ifdef SMP
1361 lwkt_process_ipiq();
1362#endif
52eedfb5 1363 cpu_lfence();
df910c23 1364 }
cfaeae2a 1365 DEBUG_POP_INFO();
562273ea 1366 cpu_mfence();
37af14fe 1367 td->td_gd = mygd;
52eedfb5
MD
1368 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1369 td->td_flags &= ~TDF_MIGRATING;
1370 crit_exit_gd(mygd);
1371 } else {
1372 crit_enter_gd(mygd);
1373 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1374 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1375 crit_exit_gd(mygd);
a2a5ad0d
MD
1376 }
1377}
1378
52eedfb5
MD
1379#endif
1380
f1d1c3fa
MD
1381/*
1382 * Generic deschedule. Descheduling threads other then your own should be
1383 * done only in carefully controlled circumstances. Descheduling is
1384 * asynchronous.
1385 *
1386 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1387 */
1388void
1389lwkt_deschedule(thread_t td)
1390{
f1d1c3fa 1391 crit_enter();
b8a98473 1392#ifdef SMP
f1d1c3fa
MD
1393 if (td == curthread) {
1394 _lwkt_dequeue(td);
1395 } else {
a72187e9 1396 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1397 _lwkt_dequeue(td);
1398 } else {
b8a98473 1399 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1400 }
1401 }
b8a98473
MD
1402#else
1403 _lwkt_dequeue(td);
1404#endif
f1d1c3fa
MD
1405 crit_exit();
1406}
1407
4b5f931b
MD
1408/*
1409 * Set the target thread's priority. This routine does not automatically
1410 * switch to a higher priority thread, LWKT threads are not designed for
1411 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1412 */
1413void
1414lwkt_setpri(thread_t td, int pri)
1415{
a72187e9 1416 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1417 if (td->td_pri != pri) {
1418 KKASSERT(pri >= 0);
1419 crit_enter();
1420 if (td->td_flags & TDF_RUNQ) {
1421 _lwkt_dequeue(td);
1422 td->td_pri = pri;
1423 _lwkt_enqueue(td);
1424 } else {
1425 td->td_pri = pri;
1426 }
1427 crit_exit();
26a0694b 1428 }
26a0694b
MD
1429}
1430
03bd0a5e
MD
1431/*
1432 * Set the initial priority for a thread prior to it being scheduled for
1433 * the first time. The thread MUST NOT be scheduled before or during
1434 * this call. The thread may be assigned to a cpu other then the current
1435 * cpu.
1436 *
1437 * Typically used after a thread has been created with TDF_STOPPREQ,
1438 * and before the thread is initially scheduled.
1439 */
1440void
1441lwkt_setpri_initial(thread_t td, int pri)
1442{
1443 KKASSERT(pri >= 0);
1444 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1445 td->td_pri = pri;
03bd0a5e
MD
1446}
1447
26a0694b
MD
1448void
1449lwkt_setpri_self(int pri)
1450{
1451 thread_t td = curthread;
1452
4b5f931b
MD
1453 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1454 crit_enter();
1455 if (td->td_flags & TDF_RUNQ) {
1456 _lwkt_dequeue(td);
f9235b6d 1457 td->td_pri = pri;
4b5f931b
MD
1458 _lwkt_enqueue(td);
1459 } else {
f9235b6d 1460 td->td_pri = pri;
4b5f931b
MD
1461 }
1462 crit_exit();
1463}
1464
f9235b6d
MD
1465/*
1466 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1467 *
1468 * Example: two competing threads, same priority N. decrement by (2*N)
1469 * increment by N*8, each thread will get 4 ticks.
1470 */
1471void
1472lwkt_fairq_schedulerclock(thread_t td)
1473{
2a418930
MD
1474 globaldata_t gd;
1475
f9235b6d
MD
1476 if (fairq_enable) {
1477 while (td) {
2a418930
MD
1478 gd = td->td_gd;
1479 if (td != &gd->gd_idlethread) {
1480 td->td_fairq_accum -= gd->gd_fairq_total_pri;
1481 if (td->td_fairq_accum < -TDFAIRQ_MAX(gd))
1482 td->td_fairq_accum = -TDFAIRQ_MAX(gd);
f9235b6d
MD
1483 if (td->td_fairq_accum < 0)
1484 need_lwkt_resched();
1485 td->td_fairq_lticks = ticks;
1486 }
1487 td = td->td_preempted;
1488 }
1489 }
1490}
1491
1492static void
1493lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1494{
1495 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1496 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1497 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1498}
1499
5d21b981 1500/*
52eedfb5
MD
1501 * Migrate the current thread to the specified cpu.
1502 *
1503 * This is accomplished by descheduling ourselves from the current cpu,
1504 * moving our thread to the tdallq of the target cpu, IPI messaging the
1505 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1506 * races while the thread is being migrated.
ae8e83e6
MD
1507 *
1508 * We must be sure to remove ourselves from the current cpu's tsleepq
1509 * before potentially moving to another queue. The thread can be on
1510 * a tsleepq due to a left-over tsleep_interlock().
95858b91
MD
1511 *
1512 * We also have to make sure that the switch code doesn't allow an IPI
1513 * processing operation to leak in between our send and our switch, or
1514 * any other potential livelock such that might occur when we release the
1515 * current process designation, so do that first.
5d21b981 1516 */
3d28ff59 1517#ifdef SMP
5d21b981 1518static void lwkt_setcpu_remote(void *arg);
3d28ff59 1519#endif
5d21b981
MD
1520
1521void
1522lwkt_setcpu_self(globaldata_t rgd)
1523{
1524#ifdef SMP
1525 thread_t td = curthread;
1526
1527 if (td->td_gd != rgd) {
1528 crit_enter_quick(td);
95858b91
MD
1529 if (td->td_release)
1530 td->td_release(td);
ae8e83e6 1531 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1532 tsleep_remove(td);
5d21b981
MD
1533 td->td_flags |= TDF_MIGRATING;
1534 lwkt_deschedule_self(td);
52eedfb5 1535 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1536 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1537 lwkt_switch();
1538 /* we are now on the target cpu */
52eedfb5 1539 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1540 crit_exit_quick(td);
1541 }
1542#endif
1543}
1544
ecdefdda
MD
1545void
1546lwkt_migratecpu(int cpuid)
1547{
1548#ifdef SMP
1549 globaldata_t rgd;
1550
1551 rgd = globaldata_find(cpuid);
1552 lwkt_setcpu_self(rgd);
1553#endif
1554}
1555
5d21b981
MD
1556/*
1557 * Remote IPI for cpu migration (called while in a critical section so we
1558 * do not have to enter another one). The thread has already been moved to
1559 * our cpu's allq, but we must wait for the thread to be completely switched
1560 * out on the originating cpu before we schedule it on ours or the stack
1561 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1562 * change to main memory.
1563 *
1564 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1565 * against wakeups. It is best if this interface is used only when there
1566 * are no pending events that might try to schedule the thread.
1567 */
3d28ff59 1568#ifdef SMP
5d21b981
MD
1569static void
1570lwkt_setcpu_remote(void *arg)
1571{
1572 thread_t td = arg;
1573 globaldata_t gd = mycpu;
cfaeae2a 1574 int retry = 10000000;
5d21b981 1575
cfaeae2a 1576 DEBUG_PUSH_INFO("lwkt_setcpu_remote");
df910c23
MD
1577 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1578#ifdef SMP
1579 lwkt_process_ipiq();
1580#endif
35238fa5 1581 cpu_lfence();
562273ea 1582 cpu_pause();
cfaeae2a
MD
1583 if (--retry == 0) {
1584 kprintf("lwkt_setcpu_remote: td->td_flags %08x\n",
1585 td->td_flags);
1586 retry = 10000000;
1587 }
df910c23 1588 }
cfaeae2a 1589 DEBUG_POP_INFO();
5d21b981 1590 td->td_gd = gd;
562273ea 1591 cpu_mfence();
5d21b981 1592 td->td_flags &= ~TDF_MIGRATING;
9388413d 1593 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1594 _lwkt_enqueue(td);
1595}
3d28ff59 1596#endif
5d21b981 1597
553ea3c8 1598struct lwp *
4b5f931b
MD
1599lwkt_preempted_proc(void)
1600{
73e4f7b9 1601 thread_t td = curthread;
4b5f931b
MD
1602 while (td->td_preempted)
1603 td = td->td_preempted;
553ea3c8 1604 return(td->td_lwp);
4b5f931b
MD
1605}
1606
99df837e
MD
1607/*
1608 * Create a kernel process/thread/whatever. It shares it's address space
1609 * with proc0 - ie: kernel only.
1610 *
365fa13f
MD
1611 * NOTE! By default new threads are created with the MP lock held. A
1612 * thread which does not require the MP lock should release it by calling
1613 * rel_mplock() at the start of the new thread.
99df837e
MD
1614 */
1615int
c9e9fb21
MD
1616lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1617 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1618{
73e4f7b9 1619 thread_t td;
e2565a42 1620 __va_list ap;
99df837e 1621
d3d32139 1622 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1623 tdflags);
a2a5ad0d
MD
1624 if (tdp)
1625 *tdp = td;
709799ea 1626 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1627
1628 /*
1629 * Set up arg0 for 'ps' etc
1630 */
e2565a42 1631 __va_start(ap, fmt);
379210cb 1632 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1633 __va_end(ap);
99df837e
MD
1634
1635 /*
1636 * Schedule the thread to run
1637 */
ef0fdad1
MD
1638 if ((td->td_flags & TDF_STOPREQ) == 0)
1639 lwkt_schedule(td);
1640 else
1641 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1642 return 0;
1643}
1644
1645/*
1646 * Destroy an LWKT thread. Warning! This function is not called when
1647 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1648 * uses a different reaping mechanism.
1649 */
1650void
1651lwkt_exit(void)
1652{
1653 thread_t td = curthread;
c070746a 1654 thread_t std;
8826f33a 1655 globaldata_t gd;
99df837e 1656
2883d2d8
MD
1657 /*
1658 * Do any cleanup that might block here
1659 */
99df837e 1660 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1661 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1662 caps_exit(td);
2883d2d8
MD
1663 biosched_done(td);
1664 dsched_exit_thread(td);
c070746a
MD
1665
1666 /*
1667 * Get us into a critical section to interlock gd_freetd and loop
1668 * until we can get it freed.
1669 *
1670 * We have to cache the current td in gd_freetd because objcache_put()ing
1671 * it would rip it out from under us while our thread is still active.
1672 */
1673 gd = mycpu;
37af14fe 1674 crit_enter_quick(td);
c070746a 1675 while ((std = gd->gd_freetd) != NULL) {
cf709dd2 1676 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
c070746a
MD
1677 gd->gd_freetd = NULL;
1678 objcache_put(thread_cache, std);
1679 }
3b4192fb
MD
1680
1681 /*
1682 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1683 * the last time. We cannot block after this point or we may end
1684 * up with a stale td on the tsleepq.
3b4192fb
MD
1685 */
1686 if (td->td_flags & TDF_TSLEEPQ)
1687 tsleep_remove(td);
37af14fe 1688 lwkt_deschedule_self(td);
e56e4dea 1689 lwkt_remove_tdallq(td);
74c9628e 1690 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1691
1692 /*
1693 * Final cleanup
1694 */
1695 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1696 if (td->td_flags & TDF_ALLOCATED_THREAD)
1697 gd->gd_freetd = td;
99df837e
MD
1698 cpu_thread_exit();
1699}
1700
e56e4dea
MD
1701void
1702lwkt_remove_tdallq(thread_t td)
1703{
1704 KKASSERT(td->td_gd == mycpu);
1705 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1706}
1707
9cf43f91
MD
1708/*
1709 * Code reduction and branch prediction improvements. Call/return
1710 * overhead on modern cpus often degenerates into 0 cycles due to
1711 * the cpu's branch prediction hardware and return pc cache. We
1712 * can take advantage of this by not inlining medium-complexity
1713 * functions and we can also reduce the branch prediction impact
1714 * by collapsing perfectly predictable branches into a single
1715 * procedure instead of duplicating it.
1716 *
1717 * Is any of this noticeable? Probably not, so I'll take the
1718 * smaller code size.
1719 */
1720void
b6468f56 1721crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1722{
b6468f56 1723 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1724}
1725
2d93b37a
MD
1726void
1727crit_panic(void)
1728{
1729 thread_t td = curthread;
850634cc 1730 int lcrit = td->td_critcount;
2d93b37a 1731
850634cc
AH
1732 td->td_critcount = 0;
1733 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1734 /* NOT REACHED */
2d93b37a
MD
1735}
1736
d165e668
MD
1737#ifdef SMP
1738
bd8015ca
MD
1739/*
1740 * Called from debugger/panic on cpus which have been stopped. We must still
1741 * process the IPIQ while stopped, even if we were stopped while in a critical
1742 * section (XXX).
1743 *
1744 * If we are dumping also try to process any pending interrupts. This may
1745 * or may not work depending on the state of the cpu at the point it was
1746 * stopped.
1747 */
1748void
1749lwkt_smp_stopped(void)
1750{
1751 globaldata_t gd = mycpu;
1752
1753 crit_enter_gd(gd);
1754 if (dumping) {
1755 lwkt_process_ipiq();
1756 splz();
1757 } else {
1758 lwkt_process_ipiq();
1759 }
1760 crit_exit_gd(gd);
1761}
1762
d165e668 1763#endif