network - Fix nasty bug in udp6_send()
[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
b2b3ffcd 110#ifdef __x86_64__
85514115
MD
111
112static int
0855a2af
JG
113jg_tos_ok(struct thread *td)
114{
85514115
MD
115 void *tos;
116 int tos_ok;
117
0855a2af
JG
118 if (td == NULL) {
119 return 1;
120 }
121 KKASSERT(td->td_sp != NULL);
85514115
MD
122 tos = ((void **)td->td_sp)[0];
123 tos_ok = 0;
124 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
125 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
0855a2af
JG
126 tos_ok = 1;
127 }
128 return tos_ok;
129}
130
85514115
MD
131#endif
132
fb0f29c4
MD
133/*
134 * We can make all thread ports use the spin backend instead of the thread
135 * backend. This should only be set to debug the spin backend.
136 */
137TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
138
0f7a3396
MD
139#ifdef INVARIANTS
140SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
141#endif
4b5f931b 142SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
9733f757
VS
143SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
144 "Successful preemption events");
145SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
146 "Failed preemption events");
26a0694b 147SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
148#ifdef INVARIANTS
149SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
150 &token_contention_count, 0, "spinning due to token contention");
38717797 151#endif
f9235b6d
MD
152static int fairq_enable = 1;
153SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, &fairq_enable, 0, "");
77912481
MD
154static int user_pri_sched = 0;
155SYSCTL_INT(_lwkt, OID_AUTO, user_pri_sched, CTLFLAG_RW, &user_pri_sched, 0, "");
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)
177 atomic_clear_int_nonlocked(&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);
200 atomic_set_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
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.
239 */
240void
241lwkt_init(void)
242{
243 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
244 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
245 NULL, CACHE_NTHREADS/2,
246 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
247}
248
37af14fe
MD
249/*
250 * Schedule a thread to run. As the current thread we can always safely
251 * schedule ourselves, and a shortcut procedure is provided for that
252 * function.
253 *
254 * (non-blocking, self contained on a per cpu basis)
255 */
256void
257lwkt_schedule_self(thread_t td)
258{
259 crit_enter_quick(td);
f9235b6d
MD
260 KASSERT(td != &td->td_gd->gd_idlethread,
261 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 262 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 263 _lwkt_enqueue(td);
37af14fe
MD
264 crit_exit_quick(td);
265}
266
267/*
268 * Deschedule a thread.
269 *
270 * (non-blocking, self contained on a per cpu basis)
271 */
272void
273lwkt_deschedule_self(thread_t td)
274{
275 crit_enter_quick(td);
37af14fe
MD
276 _lwkt_dequeue(td);
277 crit_exit_quick(td);
278}
279
8ad65e08
MD
280/*
281 * LWKTs operate on a per-cpu basis
282 *
73e4f7b9 283 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
284 */
285void
286lwkt_gdinit(struct globaldata *gd)
287{
f9235b6d 288 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 289 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
290}
291
7d0bac62
MD
292/*
293 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
294 * or LWKT start address before it can be scheduled. If the target cpu is
295 * -1 the thread will be created on the current cpu.
0cfcada1
MD
296 *
297 * If you intend to create a thread without a process context this function
298 * does everything except load the startup and switcher function.
7d0bac62
MD
299 */
300thread_t
d3d32139 301lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 302{
c070746a 303 globaldata_t gd = mycpu;
99df837e 304 void *stack;
7d0bac62 305
c070746a
MD
306 /*
307 * If static thread storage is not supplied allocate a thread. Reuse
308 * a cached free thread if possible. gd_freetd is used to keep an exiting
309 * thread intact through the exit.
310 */
ef0fdad1 311 if (td == NULL) {
c070746a
MD
312 if ((td = gd->gd_freetd) != NULL)
313 gd->gd_freetd = NULL;
314 else
315 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
316 KASSERT((td->td_flags &
317 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
318 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
319 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 320 }
c070746a
MD
321
322 /*
323 * Try to reuse cached stack.
324 */
f470d0c8
MD
325 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
326 if (flags & TDF_ALLOCATED_STACK) {
e4846942 327 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
328 stack = NULL;
329 }
330 }
331 if (stack == NULL) {
e4846942 332 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 333 flags |= TDF_ALLOCATED_STACK;
99df837e 334 }
75cdbe6c 335 if (cpu < 0)
c070746a 336 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 337 else
f470d0c8 338 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 339 return(td);
7d0bac62
MD
340}
341
342/*
343 * Initialize a preexisting thread structure. This function is used by
344 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
345 *
f8c3996b
MD
346 * All threads start out in a critical section at a priority of
347 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
348 * appropriate. This function may send an IPI message when the
349 * requested cpu is not the current cpu and consequently gd_tdallq may
350 * not be initialized synchronously from the point of view of the originating
351 * cpu.
352 *
353 * NOTE! we have to be careful in regards to creating threads for other cpus
354 * if SMP has not yet been activated.
7d0bac62 355 */
41a01a4d
MD
356#ifdef SMP
357
75cdbe6c
MD
358static void
359lwkt_init_thread_remote(void *arg)
360{
361 thread_t td = arg;
362
52eedfb5
MD
363 /*
364 * Protected by critical section held by IPI dispatch
365 */
75cdbe6c
MD
366 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
367}
368
41a01a4d
MD
369#endif
370
fdce8919
MD
371/*
372 * lwkt core thread structural initialization.
373 *
374 * NOTE: All threads are initialized as mpsafe threads.
375 */
7d0bac62 376void
f470d0c8
MD
377lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
378 struct globaldata *gd)
7d0bac62 379{
37af14fe
MD
380 globaldata_t mygd = mycpu;
381
99df837e
MD
382 bzero(td, sizeof(struct thread));
383 td->td_kstack = stack;
f470d0c8 384 td->td_kstack_size = stksize;
d3d32139 385 td->td_flags = flags;
26a0694b 386 td->td_gd = gd;
f9235b6d
MD
387 td->td_pri = TDPRI_KERN_DAEMON;
388 td->td_critcount = 1;
3b998fa9 389 td->td_toks_stop = &td->td_toks_base;
fb0f29c4
MD
390 if (lwkt_use_spin_port)
391 lwkt_initport_spin(&td->td_msgport);
392 else
393 lwkt_initport_thread(&td->td_msgport, td);
99df837e 394 pmap_init_thread(td);
0f7a3396 395#ifdef SMP
5d21b981
MD
396 /*
397 * Normally initializing a thread for a remote cpu requires sending an
398 * IPI. However, the idlethread is setup before the other cpus are
399 * activated so we have to treat it as a special case. XXX manipulation
400 * of gd_tdallq requires the BGL.
401 */
402 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 403 crit_enter_gd(mygd);
75cdbe6c 404 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 405 crit_exit_gd(mygd);
75cdbe6c 406 } else {
2db3b277 407 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 408 }
0f7a3396 409#else
37af14fe 410 crit_enter_gd(mygd);
0f7a3396 411 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 412 crit_exit_gd(mygd);
0f7a3396 413#endif
8c72e3d5
AH
414
415 dsched_new_thread(td);
73e4f7b9
MD
416}
417
418void
419lwkt_set_comm(thread_t td, const char *ctl, ...)
420{
e2565a42 421 __va_list va;
73e4f7b9 422
e2565a42 423 __va_start(va, ctl);
379210cb 424 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 425 __va_end(va);
e7c0dbba 426 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
427}
428
99df837e 429void
73e4f7b9 430lwkt_hold(thread_t td)
99df837e 431{
73e4f7b9
MD
432 ++td->td_refs;
433}
434
435void
436lwkt_rele(thread_t td)
437{
438 KKASSERT(td->td_refs > 0);
439 --td->td_refs;
440}
441
442void
443lwkt_wait_free(thread_t td)
444{
445 while (td->td_refs)
377d4740 446 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
447}
448
449void
450lwkt_free_thread(thread_t td)
451{
d9eea1a5 452 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
453 ("lwkt_free_thread: did not exit! %p", td));
454
40aaf5fc
NT
455 if (td->td_flags & TDF_ALLOCATED_THREAD) {
456 objcache_put(thread_cache, td);
457 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
458 /* client-allocated struct with internally allocated stack */
459 KASSERT(td->td_kstack && td->td_kstack_size > 0,
460 ("lwkt_free_thread: corrupted stack"));
461 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
462 td->td_kstack = NULL;
463 td->td_kstack_size = 0;
99df837e 464 }
e7c0dbba 465 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
466}
467
468
8ad65e08
MD
469/*
470 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
471 * switch to the idlethread. Switching must occur within a critical
472 * section to avoid races with the scheduling queue.
473 *
474 * We always have full control over our cpu's run queue. Other cpus
475 * that wish to manipulate our queue must use the cpu_*msg() calls to
476 * talk to our cpu, so a critical section is all that is needed and
477 * the result is very, very fast thread switching.
478 *
96728c05
MD
479 * The LWKT scheduler uses a fixed priority model and round-robins at
480 * each priority level. User process scheduling is a totally
481 * different beast and LWKT priorities should not be confused with
482 * user process priorities.
f1d1c3fa 483 *
3933a3ab
MD
484 * The MP lock may be out of sync with the thread's td_mpcount + td_xpcount.
485 * lwkt_switch() cleans it up.
486 *
487 * Note that the td_switch() function cannot do anything that requires
488 * the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
489 * the target thread (not the current thread). It's nice to have a scheduler
490 * that does not need the MP lock to work because it allows us to do some
491 * really cool high-performance MP lock optimizations.
69d78e99
MD
492 *
493 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
494 * is not called by the current thread in the preemption case, only when
495 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
496 */
497void
498lwkt_switch(void)
499{
37af14fe
MD
500 globaldata_t gd = mycpu;
501 thread_t td = gd->gd_curthread;
8ad65e08 502 thread_t ntd;
f9235b6d
MD
503 thread_t xtd;
504 thread_t nlast;
f9235b6d 505 int nquserok;
6f207a2c 506#ifdef SMP
8a8d5d85
MD
507 int mpheld;
508#endif
f9235b6d 509 int didaccumulate;
b37f18d6
MD
510 const char *lmsg; /* diagnostic - 'systat -pv 1' */
511 const void *laddr;
8ad65e08 512
46a3f46d 513 /*
27e88a6e
MD
514 * Switching from within a 'fast' (non thread switched) interrupt or IPI
515 * is illegal. However, we may have to do it anyway if we hit a fatal
516 * kernel trap or we have paniced.
517 *
518 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 519 */
27e88a6e
MD
520 if (gd->gd_intr_nesting_level) {
521 int savegdnest;
522 int savegdtrap;
523
5fddbda2 524 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22
MD
525 panic("lwkt_switch: Attempt to switch from a "
526 "a fast interrupt, ipi, or hard code section, "
527 "td %p\n",
528 td);
27e88a6e
MD
529 } else {
530 savegdnest = gd->gd_intr_nesting_level;
531 savegdtrap = gd->gd_trap_nesting_level;
532 gd->gd_intr_nesting_level = 0;
533 gd->gd_trap_nesting_level = 0;
a7422615
MD
534 if ((td->td_flags & TDF_PANICWARN) == 0) {
535 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
536 kprintf("Warning: thread switch from interrupt, IPI, "
537 "or hard code section.\n"
a7422615 538 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 539 print_backtrace(-1);
a7422615 540 }
27e88a6e
MD
541 lwkt_switch();
542 gd->gd_intr_nesting_level = savegdnest;
543 gd->gd_trap_nesting_level = savegdtrap;
544 return;
545 }
96728c05 546 }
ef0fdad1 547
cb973d15
MD
548 /*
549 * Passive release (used to transition from user to kernel mode
550 * when we block or switch rather then when we enter the kernel).
551 * This function is NOT called if we are switching into a preemption
552 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
553 * our current process designation (if we have one) and become a true
554 * LWKT thread, and may also hand the current process designation to
555 * another process and schedule thread.
cb973d15
MD
556 */
557 if (td->td_release)
558 td->td_release(td);
559
37af14fe 560 crit_enter_gd(gd);
3b998fa9 561 if (TD_TOKS_HELD(td))
9d265729
MD
562 lwkt_relalltokens(td);
563
564 /*
b02926de
MD
565 * We had better not be holding any spin locks, but don't get into an
566 * endless panic loop.
9d265729 567 */
d666840a
MD
568 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
569 ("lwkt_switch: still holding %d exclusive spinlocks!",
570 gd->gd_spinlocks_wr));
9d265729 571
8a8d5d85
MD
572
573#ifdef SMP
574 /*
3933a3ab
MD
575 * td_mpcount + td_xpcount cannot be used to determine if we currently
576 * hold the MP lock because get_mplock() will increment it prior to
577 * attempting to get the lock, and switch out if it can't. Our
578 * ownership of the actual lock will remain stable while we are
579 * in a critical section, and once we actually acquire the underlying
580 * lock as long as the count is greater than 0.
8a8d5d85 581 */
c5724852 582 mpheld = MP_LOCK_HELD(gd);
0f7a3396
MD
583#ifdef INVARIANTS
584 if (td->td_cscount) {
6ea70f76 585 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
586 td);
587 if (panic_on_cscount)
588 panic("switching while mastering cpusync");
589 }
590#endif
8a8d5d85 591#endif
f9235b6d
MD
592
593 /*
594 * If we had preempted another thread on this cpu, resume the preempted
595 * thread. This occurs transparently, whether the preempted thread
596 * was scheduled or not (it may have been preempted after descheduling
597 * itself).
598 *
599 * We have to setup the MP lock for the original thread after backing
600 * out the adjustment that was made to curthread when the original
601 * was preempted.
602 */
99df837e 603 if ((ntd = td->td_preempted) != NULL) {
26a0694b 604 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 605#ifdef SMP
3933a3ab 606 if (ntd->td_mpcount + ntd->td_xpcount && mpheld == 0) {
fc92d4aa 607 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
3933a3ab 608 td, ntd, td->td_mpcount, ntd->td_mpcount + ntd->td_xpcount);
8a8d5d85 609 }
3933a3ab 610 td->td_xpcount = 0;
8a8d5d85 611#endif
26a0694b 612 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
613
614 /*
b9eb1c19
MD
615 * The interrupt may have woken a thread up, we need to properly
616 * set the reschedule flag if the originally interrupted thread is
617 * at a lower priority.
8ec60c3f 618 */
f9235b6d
MD
619 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
620 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
8ec60c3f 621 need_lwkt_resched();
f9235b6d 622 }
8a8d5d85 623 /* YYY release mp lock on switchback if original doesn't need it */
f9235b6d
MD
624 goto havethread_preempted;
625 }
626
627 /*
628 * Implement round-robin fairq with priority insertion. The priority
629 * insertion is handled by _lwkt_enqueue()
630 *
631 * We have to adjust the MP lock for the target thread. If we
632 * need the MP lock and cannot obtain it we try to locate a
633 * thread that does not need the MP lock. If we cannot, we spin
634 * instead of HLT.
635 *
636 * A similar issue exists for the tokens held by the target thread.
637 * If we cannot obtain ownership of the tokens we cannot immediately
638 * schedule the thread.
639 */
640 for (;;) {
641 clear_lwkt_resched();
642 didaccumulate = 0;
643 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
644
4b5f931b 645 /*
f9235b6d 646 * Hotpath if we can get all necessary resources.
41a01a4d 647 *
f9235b6d 648 * If nothing is runnable switch to the idle thread
41a01a4d 649 */
f9235b6d
MD
650 if (ntd == NULL) {
651 ntd = &gd->gd_idlethread;
652 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
653 ntd->td_flags |= TDF_IDLE_NOHLT;
6f207a2c 654#ifdef SMP
3933a3ab 655 KKASSERT(ntd->td_xpcount == 0);
f9235b6d
MD
656 if (ntd->td_mpcount) {
657 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
658 panic("Idle thread %p was holding the BGL!", ntd);
659 if (mpheld == 0) {
c5724852
MD
660 set_cpu_contention_mask(gd);
661 handle_cpu_contention_mask();
662 cpu_try_mplock();
663 mpheld = MP_LOCK_HELD(gd);
f9235b6d
MD
664 cpu_pause();
665 continue;
666 }
667 }
c5724852 668 clr_cpu_contention_mask(gd);
6f207a2c 669#endif
b37f18d6
MD
670 cpu_time.cp_msg[0] = 0;
671 cpu_time.cp_stallpc = 0;
f9235b6d
MD
672 goto haveidle;
673 }
41a01a4d 674
8ec60c3f 675 /*
f9235b6d 676 * Hotpath schedule
6f207a2c
MD
677 *
678 * NOTE: For UP there is no mplock and lwkt_getalltokens()
679 * always succeeds.
8ec60c3f 680 */
f9235b6d
MD
681 if (ntd->td_fairq_accum >= 0 &&
682#ifdef SMP
3933a3ab
MD
683 (ntd->td_mpcount + ntd->td_xpcount == 0 ||
684 mpheld || cpu_try_mplock()) &&
f9235b6d 685#endif
b37f18d6 686 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
f9235b6d 687 ) {
8a8d5d85 688#ifdef SMP
c5724852 689 clr_cpu_contention_mask(gd);
f9235b6d
MD
690#endif
691 goto havethread;
692 }
693
b37f18d6
MD
694 lmsg = NULL;
695 laddr = NULL;
696
f9235b6d 697#ifdef SMP
c5724852
MD
698 if (ntd->td_fairq_accum >= 0)
699 set_cpu_contention_mask(gd);
f9235b6d 700 /* Reload mpheld (it become stale after mplock/token ops) */
c5724852 701 mpheld = MP_LOCK_HELD(gd);
3933a3ab 702 if (ntd->td_mpcount + ntd->td_xpcount && mpheld == 0) {
b37f18d6
MD
703 lmsg = "mplock";
704 laddr = ntd->td_mplock_stallpc;
705 }
f9235b6d
MD
706#endif
707
708 /*
709 * Coldpath - unable to schedule ntd, continue looking for threads
710 * to schedule. This is only allowed of the (presumably) kernel
711 * thread exhausted its fair share. A kernel thread stuck on
712 * resources does not currently allow a user thread to get in
713 * front of it.
714 */
715#ifdef SMP
716 nquserok = ((ntd->td_pri < TDPRI_KERN_LPSCHED) ||
717 (ntd->td_fairq_accum < 0));
6f207a2c
MD
718#else
719 nquserok = 1;
f9235b6d
MD
720#endif
721 nlast = NULL;
722
723 for (;;) {
41a01a4d 724 /*
f9235b6d
MD
725 * If the fair-share scheduler ran out ntd gets moved to the
726 * end and its accumulator will be bumped, if it didn't we
727 * maintain the same queue position.
df6b8ba0 728 *
f9235b6d 729 * nlast keeps track of the last element prior to any moves.
41a01a4d 730 */
f9235b6d 731 if (ntd->td_fairq_accum < 0) {
f9235b6d
MD
732 lwkt_fairq_accumulate(gd, ntd);
733 didaccumulate = 1;
c5724852
MD
734
735 /*
736 * Move to end
737 */
738 xtd = TAILQ_NEXT(ntd, td_threadq);
f9235b6d
MD
739 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
740 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
c5724852
MD
741
742 /*
743 * Set terminal element (nlast)
744 */
f9235b6d
MD
745 if (nlast == NULL) {
746 nlast = ntd;
747 if (xtd == NULL)
748 xtd = ntd;
749 }
750 ntd = xtd;
751 } else {
752 ntd = TAILQ_NEXT(ntd, td_threadq);
753 }
a453459d 754
f9235b6d
MD
755 /*
756 * If we exhausted the run list switch to the idle thread.
757 * Since one or more threads had resource acquisition issues
758 * we do not allow the idle thread to halt.
759 *
760 * NOTE: nlast can be NULL.
761 */
762 if (ntd == nlast) {
e0a90d3b 763 cpu_pause();
f9235b6d
MD
764 ntd = &gd->gd_idlethread;
765 ntd->td_flags |= TDF_IDLE_NOHLT;
6f207a2c 766#ifdef SMP
3933a3ab 767 KKASSERT(ntd->td_xpcount == 0);
f9235b6d 768 if (ntd->td_mpcount) {
c5724852 769 mpheld = MP_LOCK_HELD(gd);
f9235b6d
MD
770 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
771 panic("Idle thread %p was holding the BGL!", ntd);
772 if (mpheld == 0) {
c5724852
MD
773 set_cpu_contention_mask(gd);
774 handle_cpu_contention_mask();
775 cpu_try_mplock();
776 mpheld = MP_LOCK_HELD(gd);
f9235b6d
MD
777 cpu_pause();
778 break; /* try again from the top, almost */
b9eb1c19 779 }
8a8d5d85 780 }
6f207a2c 781#endif
684a93c4
MD
782
783 /*
f9235b6d
MD
784 * If fairq accumulations occured we do not schedule the
785 * idle thread. This will cause us to try again from
786 * the (almost) top.
684a93c4 787 */
f9235b6d 788 if (didaccumulate)
b37f18d6
MD
789 break; /* try again from the top, almost */
790 if (lmsg)
791 strlcpy(cpu_time.cp_msg, lmsg, sizeof(cpu_time.cp_msg));
792 cpu_time.cp_stallpc = (uintptr_t)laddr;
f9235b6d 793 goto haveidle;
8a8d5d85 794 }
f9235b6d 795
df6b8ba0 796 /*
f9235b6d 797 * Try to switch to this thread.
6f207a2c
MD
798 *
799 * NOTE: For UP there is no mplock and lwkt_getalltokens()
800 * always succeeds.
df6b8ba0 801 */
77912481
MD
802 if ((ntd->td_pri >= TDPRI_KERN_LPSCHED || nquserok ||
803 user_pri_sched) && ntd->td_fairq_accum >= 0 &&
f9235b6d 804#ifdef SMP
3933a3ab
MD
805 (ntd->td_mpcount + ntd->td_xpcount == 0 ||
806 mpheld || cpu_try_mplock()) &&
8a8d5d85 807#endif
b37f18d6 808 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
f9235b6d 809 ) {
a453459d 810#ifdef SMP
c5724852 811 clr_cpu_contention_mask(gd);
f9235b6d
MD
812#endif
813 goto havethread;
df6b8ba0 814 }
f9235b6d 815#ifdef SMP
c5724852
MD
816 if (ntd->td_fairq_accum >= 0)
817 set_cpu_contention_mask(gd);
818 /*
819 * Reload mpheld (it become stale after mplock/token ops).
820 */
821 mpheld = MP_LOCK_HELD(gd);
3933a3ab 822 if (ntd->td_mpcount + ntd->td_xpcount && mpheld == 0) {
b37f18d6
MD
823 lmsg = "mplock";
824 laddr = ntd->td_mplock_stallpc;
825 }
f9235b6d
MD
826 if (ntd->td_pri >= TDPRI_KERN_LPSCHED && ntd->td_fairq_accum >= 0)
827 nquserok = 0;
a453459d 828#endif
4b5f931b 829 }
c5724852
MD
830
831 /*
832 * All threads exhausted but we can loop due to a negative
833 * accumulator.
834 *
835 * While we are looping in the scheduler be sure to service
836 * any interrupts which were made pending due to our critical
837 * section, otherwise we could livelock (e.g.) IPIs.
838 *
839 * NOTE: splz can enter and exit the mplock so mpheld is
840 * stale after this call.
841 */
842 splz_check();
843
844#ifdef SMP
845 /*
846 * Our mplock can be cached and cause other cpus to livelock
847 * if we loop due to e.g. not being able to acquire tokens.
848 */
849 if (MP_LOCK_HELD(gd))
850 cpu_rel_mplock(gd->gd_cpuid);
851 mpheld = 0;
852#endif
f1d1c3fa 853 }
8a8d5d85
MD
854
855 /*
f9235b6d
MD
856 * Do the actual switch. WARNING: mpheld is stale here.
857 *
858 * We must always decrement td_fairq_accum on non-idle threads just
859 * in case a thread never gets a tick due to being in a continuous
860 * critical section. The page-zeroing code does that.
861 *
862 * If the thread we came up with is a higher or equal priority verses
863 * the thread at the head of the queue we move our thread to the
864 * front. This way we can always check the front of the queue.
865 */
866havethread:
867 ++gd->gd_cnt.v_swtch;
868 --ntd->td_fairq_accum;
869 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
870 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
871 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
872 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
873 }
874havethread_preempted:
875 ;
876 /*
877 * If the new target does not need the MP lock and we are holding it,
878 * release the MP lock. If the new target requires the MP lock we have
879 * already acquired it for the target.
880 *
881 * WARNING: mpheld is stale here.
8a8d5d85 882 */
f9235b6d
MD
883haveidle:
884 KASSERT(ntd->td_critcount,
885 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85 886#ifdef SMP
3933a3ab 887 if (ntd->td_mpcount + ntd->td_xpcount == 0 ) {
c5724852
MD
888 if (MP_LOCK_HELD(gd))
889 cpu_rel_mplock(gd->gd_cpuid);
8a8d5d85 890 } else {
a453459d 891 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
892 }
893#endif
94f6d86e
MD
894 if (td != ntd) {
895 ++switch_count;
b2b3ffcd 896#ifdef __x86_64__
f9235b6d
MD
897 {
898 int tos_ok __debugvar = jg_tos_ok(ntd);
899 KKASSERT(tos_ok);
900 }
85514115 901#endif
a1f0fb66 902 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
f1d1c3fa 903 td->td_switch(ntd);
94f6d86e 904 }
37af14fe
MD
905 /* NOTE: current cpu may have changed after switch */
906 crit_exit_quick(td);
8ad65e08
MD
907}
908
b68b7282 909/*
96728c05
MD
910 * Request that the target thread preempt the current thread. Preemption
911 * only works under a specific set of conditions:
b68b7282 912 *
96728c05
MD
913 * - We are not preempting ourselves
914 * - The target thread is owned by the current cpu
915 * - We are not currently being preempted
916 * - The target is not currently being preempted
d3d1cbc8
MD
917 * - We are not holding any spin locks
918 * - The target thread is not holding any tokens
96728c05
MD
919 * - We are able to satisfy the target's MP lock requirements (if any).
920 *
921 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
922 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 923 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
924 * to determine whether preemption is possible (aka usually just the crit
925 * priority of lwkt_schedule() itself).
b68b7282 926 *
26a0694b
MD
927 * XXX at the moment we run the target thread in a critical section during
928 * the preemption in order to prevent the target from taking interrupts
929 * that *WE* can't. Preemption is strictly limited to interrupt threads
930 * and interrupt-like threads, outside of a critical section, and the
931 * preempted source thread will be resumed the instant the target blocks
932 * whether or not the source is scheduled (i.e. preemption is supposed to
933 * be as transparent as possible).
4b5f931b 934 *
8a8d5d85
MD
935 * The target thread inherits our MP count (added to its own) for the
936 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
937 * MP lock during the preemption. Therefore, any preempting targets must be
938 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
939 * out of sync with the physical mp_lock, but we do not have to preserve
940 * the original ownership of the lock if it was out of synch (that is, we
941 * can leave it synchronized on return).
b68b7282
MD
942 */
943void
f9235b6d 944lwkt_preempt(thread_t ntd, int critcount)
b68b7282 945{
46a3f46d 946 struct globaldata *gd = mycpu;
0a3f9b47 947 thread_t td;
8a8d5d85
MD
948#ifdef SMP
949 int mpheld;
57c254db 950 int savecnt;
8a8d5d85 951#endif
b68b7282 952
26a0694b 953 /*
96728c05
MD
954 * The caller has put us in a critical section. We can only preempt
955 * if the caller of the caller was not in a critical section (basically
f9235b6d 956 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 957 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 958 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
959 *
960 * YYY The target thread must be in a critical section (else it must
961 * inherit our critical section? I dunno yet).
41a01a4d 962 *
0a3f9b47 963 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b 964 */
f9235b6d 965 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 966
0a3f9b47 967 td = gd->gd_curthread;
f9235b6d 968 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
969 ++preempt_miss;
970 return;
971 }
f9235b6d 972 if (td->td_critcount > critcount) {
96728c05 973 ++preempt_miss;
8ec60c3f 974 need_lwkt_resched();
96728c05
MD
975 return;
976 }
977#ifdef SMP
46a3f46d 978 if (ntd->td_gd != gd) {
96728c05 979 ++preempt_miss;
8ec60c3f 980 need_lwkt_resched();
96728c05
MD
981 return;
982 }
983#endif
41a01a4d 984 /*
77912481
MD
985 * We don't have to check spinlocks here as they will also bump
986 * td_critcount.
d3d1cbc8
MD
987 *
988 * Do not try to preempt if the target thread is holding any tokens.
989 * We could try to acquire the tokens but this case is so rare there
990 * is no need to support it.
41a01a4d 991 */
77912481
MD
992 KKASSERT(gd->gd_spinlocks_wr == 0);
993
3b998fa9 994 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8
MD
995 ++preempt_miss;
996 need_lwkt_resched();
997 return;
998 }
26a0694b
MD
999 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1000 ++preempt_weird;
8ec60c3f 1001 need_lwkt_resched();
26a0694b
MD
1002 return;
1003 }
1004 if (ntd->td_preempted) {
4b5f931b 1005 ++preempt_hit;
8ec60c3f 1006 need_lwkt_resched();
26a0694b 1007 return;
b68b7282 1008 }
8a8d5d85 1009#ifdef SMP
a2a5ad0d 1010 /*
3933a3ab 1011 * NOTE: An interrupt might have occured just as we were transitioning
71ef2f5c 1012 * to or from the MP lock. In this case td_mpcount will be pre-disposed
3933a3ab
MD
1013 * (non-zero) but not actually synchronized with the mp_lock itself.
1014 * We can use it to imply an MP lock requirement for the preemption but
1015 * we cannot use it to test whether we hold the MP lock or not.
a2a5ad0d 1016 */
96728c05 1017 savecnt = td->td_mpcount;
c5724852 1018 mpheld = MP_LOCK_HELD(gd);
6d9b99db 1019 ntd->td_xpcount = td->td_mpcount + td->td_xpcount;
3933a3ab
MD
1020 if (mpheld == 0 && ntd->td_mpcount + ntd->td_xpcount && !cpu_try_mplock()) {
1021 ntd->td_xpcount = 0;
8a8d5d85 1022 ++preempt_miss;
8ec60c3f 1023 need_lwkt_resched();
8a8d5d85
MD
1024 return;
1025 }
1026#endif
26a0694b 1027
8ec60c3f
MD
1028 /*
1029 * Since we are able to preempt the current thread, there is no need to
1030 * call need_lwkt_resched().
1031 */
26a0694b
MD
1032 ++preempt_hit;
1033 ntd->td_preempted = td;
1034 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1035 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
26a0694b 1036 td->td_switch(ntd);
b9eb1c19 1037
26a0694b 1038 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
1039#ifdef SMP
1040 KKASSERT(savecnt == td->td_mpcount);
c5724852 1041 mpheld = MP_LOCK_HELD(gd);
71ef2f5c 1042 if (mpheld && td->td_mpcount == 0)
c5724852 1043 cpu_rel_mplock(gd->gd_cpuid);
3933a3ab 1044 else if (mpheld == 0 && td->td_mpcount + td->td_xpcount)
96728c05
MD
1045 panic("lwkt_preempt(): MP lock was not held through");
1046#endif
26a0694b
MD
1047 ntd->td_preempted = NULL;
1048 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1049}
1050
f1d1c3fa 1051/*
faaeffac 1052 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1053 * This will work inside a critical section but not inside a hard code
1054 * section.
ef0fdad1 1055 *
f1d1c3fa
MD
1056 * (self contained on a per cpu basis)
1057 */
1058void
faaeffac 1059splz_check(void)
f1d1c3fa 1060{
7966cb69
MD
1061 globaldata_t gd = mycpu;
1062 thread_t td = gd->gd_curthread;
ef0fdad1 1063
4a28fe22
MD
1064 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1065 gd->gd_intr_nesting_level == 0 &&
1066 td->td_nest_count < 2)
1067 {
f1d1c3fa 1068 splz();
4a28fe22
MD
1069 }
1070}
1071
1072/*
1073 * This version is integrated into crit_exit, reqflags has already
1074 * been tested but td_critcount has not.
1075 *
1076 * We only want to execute the splz() on the 1->0 transition of
1077 * critcount and not in a hard code section or if too deeply nested.
1078 */
1079void
1080lwkt_maybe_splz(thread_t td)
1081{
1082 globaldata_t gd = td->td_gd;
1083
1084 if (td->td_critcount == 0 &&
1085 gd->gd_intr_nesting_level == 0 &&
1086 td->td_nest_count < 2)
1087 {
1088 splz();
1089 }
f1d1c3fa
MD
1090}
1091
8ad65e08 1092/*
f9235b6d
MD
1093 * This function is used to negotiate a passive release of the current
1094 * process/lwp designation with the user scheduler, allowing the user
1095 * scheduler to schedule another user thread. The related kernel thread
1096 * (curthread) continues running in the released state.
8ad65e08
MD
1097 */
1098void
f9235b6d 1099lwkt_passive_release(struct thread *td)
8ad65e08 1100{
f9235b6d
MD
1101 struct lwp *lp = td->td_lwp;
1102
1103 td->td_release = NULL;
1104 lwkt_setpri_self(TDPRI_KERN_USER);
1105 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1106}
1107
f9235b6d 1108
3824f392 1109/*
f9235b6d
MD
1110 * This implements a normal yield. This routine is virtually a nop if
1111 * there is nothing to yield to but it will always run any pending interrupts
1112 * if called from a critical section.
1113 *
1114 * This yield is designed for kernel threads without a user context.
1115 *
1116 * (self contained on a per cpu basis)
3824f392
MD
1117 */
1118void
f9235b6d 1119lwkt_yield(void)
3824f392 1120{
f9235b6d
MD
1121 globaldata_t gd = mycpu;
1122 thread_t td = gd->gd_curthread;
1123 thread_t xtd;
3824f392 1124
f9235b6d
MD
1125 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1126 splz();
1127 if (td->td_fairq_accum < 0) {
1128 lwkt_schedule_self(curthread);
1129 lwkt_switch();
1130 } else {
1131 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1132 if (xtd && xtd->td_pri > td->td_pri) {
1133 lwkt_schedule_self(curthread);
1134 lwkt_switch();
1135 }
1136 }
3824f392
MD
1137}
1138
1139/*
f9235b6d
MD
1140 * This yield is designed for kernel threads with a user context.
1141 *
1142 * The kernel acting on behalf of the user is potentially cpu-bound,
1143 * this function will efficiently allow other threads to run and also
1144 * switch to other processes by releasing.
3824f392
MD
1145 *
1146 * The lwkt_user_yield() function is designed to have very low overhead
1147 * if no yield is determined to be needed.
1148 */
1149void
1150lwkt_user_yield(void)
1151{
f9235b6d
MD
1152 globaldata_t gd = mycpu;
1153 thread_t td = gd->gd_curthread;
1154
1155 /*
1156 * Always run any pending interrupts in case we are in a critical
1157 * section.
1158 */
1159 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1160 splz();
3824f392
MD
1161
1162#ifdef SMP
1163 /*
1164 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1165 * kernel can prevent other cpus from servicing interrupt threads
1166 * which still require the MP lock (which is a lot of them). This
1167 * has a chaining effect since if the interrupt is blocked, so is
1168 * the event, so normal scheduling will not pick up on the problem.
1169 */
3933a3ab 1170 if (cpu_contention_mask && td->td_mpcount + td->td_xpcount) {
684a93c4 1171 yield_mplock(td);
3824f392
MD
1172 }
1173#endif
1174
1175 /*
f9235b6d
MD
1176 * Switch (which forces a release) if another kernel thread needs
1177 * the cpu, if userland wants us to resched, or if our kernel
1178 * quantum has run out.
3824f392 1179 */
f9235b6d
MD
1180 if (lwkt_resched_wanted() ||
1181 user_resched_wanted() ||
1182 td->td_fairq_accum < 0)
1183 {
3824f392 1184 lwkt_switch();
3824f392
MD
1185 }
1186
f9235b6d 1187#if 0
3824f392 1188 /*
f9235b6d
MD
1189 * Reacquire the current process if we are released.
1190 *
1191 * XXX not implemented atm. The kernel may be holding locks and such,
1192 * so we want the thread to continue to receive cpu.
3824f392 1193 */
f9235b6d
MD
1194 if (td->td_release == NULL && lp) {
1195 lp->lwp_proc->p_usched->acquire_curproc(lp);
1196 td->td_release = lwkt_passive_release;
1197 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1198 }
f9235b6d 1199#endif
b9eb1c19
MD
1200}
1201
8ad65e08 1202/*
f1d1c3fa
MD
1203 * Generic schedule. Possibly schedule threads belonging to other cpus and
1204 * deal with threads that might be blocked on a wait queue.
1205 *
0a3f9b47
MD
1206 * We have a little helper inline function which does additional work after
1207 * the thread has been enqueued, including dealing with preemption and
1208 * setting need_lwkt_resched() (which prevents the kernel from returning
1209 * to userland until it has processed higher priority threads).
6330a558
MD
1210 *
1211 * It is possible for this routine to be called after a failed _enqueue
1212 * (due to the target thread migrating, sleeping, or otherwise blocked).
1213 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1214 *
1215 * reschedok is an optimized constant propagated from lwkt_schedule() or
1216 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1217 * reschedule to be requested if the target thread has a higher priority.
1218 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1219 * be 0, prevented undesired reschedules.
8ad65e08 1220 */
0a3f9b47
MD
1221static __inline
1222void
f9235b6d 1223_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
0a3f9b47 1224{
b9eb1c19 1225 thread_t otd;
c730be20 1226
6330a558 1227 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1228 if (ntd->td_preemptable && reschedok) {
f9235b6d 1229 ntd->td_preemptable(ntd, ccount); /* YYY +token */
361d01dd 1230 } else if (reschedok) {
b9eb1c19 1231 otd = curthread;
f9235b6d 1232 if (ntd->td_pri > otd->td_pri)
c730be20 1233 need_lwkt_resched();
6330a558 1234 }
f9235b6d
MD
1235
1236 /*
1237 * Give the thread a little fair share scheduler bump if it
1238 * has been asleep for a while. This is primarily to avoid
1239 * a degenerate case for interrupt threads where accumulator
1240 * crosses into negative territory unnecessarily.
1241 */
1242 if (ntd->td_fairq_lticks != ticks) {
1243 ntd->td_fairq_lticks = ticks;
1244 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1245 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1246 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1247 }
0a3f9b47
MD
1248 }
1249}
1250
361d01dd 1251static __inline
8ad65e08 1252void
361d01dd 1253_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1254{
37af14fe
MD
1255 globaldata_t mygd = mycpu;
1256
41a01a4d 1257 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1258 crit_enter_gd(mygd);
9388413d 1259 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1260 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1261 _lwkt_enqueue(td);
1262 } else {
f1d1c3fa 1263 /*
7cd8d145
MD
1264 * If we own the thread, there is no race (since we are in a
1265 * critical section). If we do not own the thread there might
1266 * be a race but the target cpu will deal with it.
f1d1c3fa 1267 */
0f7a3396 1268#ifdef SMP
7cd8d145 1269 if (td->td_gd == mygd) {
9d265729 1270 _lwkt_enqueue(td);
f9235b6d 1271 _lwkt_schedule_post(mygd, td, 1, reschedok);
f1d1c3fa 1272 } else {
e381e77c 1273 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1274 }
0f7a3396 1275#else
7cd8d145 1276 _lwkt_enqueue(td);
f9235b6d 1277 _lwkt_schedule_post(mygd, td, 1, reschedok);
0f7a3396 1278#endif
8ad65e08 1279 }
37af14fe 1280 crit_exit_gd(mygd);
8ad65e08
MD
1281}
1282
361d01dd
MD
1283void
1284lwkt_schedule(thread_t td)
1285{
1286 _lwkt_schedule(td, 1);
1287}
1288
1289void
1290lwkt_schedule_noresched(thread_t td)
1291{
1292 _lwkt_schedule(td, 0);
1293}
1294
88ebb169
SW
1295#ifdef SMP
1296
e381e77c
MD
1297/*
1298 * When scheduled remotely if frame != NULL the IPIQ is being
1299 * run via doreti or an interrupt then preemption can be allowed.
1300 *
1301 * To allow preemption we have to drop the critical section so only
1302 * one is present in _lwkt_schedule_post.
1303 */
1304static void
1305lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1306{
1307 thread_t td = curthread;
1308 thread_t ntd = arg;
1309
1310 if (frame && ntd->td_preemptable) {
1311 crit_exit_noyield(td);
1312 _lwkt_schedule(ntd, 1);
1313 crit_enter_quick(td);
1314 } else {
1315 _lwkt_schedule(ntd, 1);
1316 }
1317}
1318
d9eea1a5 1319/*
52eedfb5
MD
1320 * Thread migration using a 'Pull' method. The thread may or may not be
1321 * the current thread. It MUST be descheduled and in a stable state.
1322 * lwkt_giveaway() must be called on the cpu owning the thread.
1323 *
1324 * At any point after lwkt_giveaway() is called, the target cpu may
1325 * 'pull' the thread by calling lwkt_acquire().
1326 *
ae8e83e6
MD
1327 * We have to make sure the thread is not sitting on a per-cpu tsleep
1328 * queue or it will blow up when it moves to another cpu.
1329 *
52eedfb5 1330 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1331 */
52eedfb5
MD
1332void
1333lwkt_giveaway(thread_t td)
1334{
3b4192fb 1335 globaldata_t gd = mycpu;
52eedfb5 1336
3b4192fb
MD
1337 crit_enter_gd(gd);
1338 if (td->td_flags & TDF_TSLEEPQ)
1339 tsleep_remove(td);
1340 KKASSERT(td->td_gd == gd);
1341 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1342 td->td_flags |= TDF_MIGRATING;
1343 crit_exit_gd(gd);
52eedfb5
MD
1344}
1345
a2a5ad0d
MD
1346void
1347lwkt_acquire(thread_t td)
1348{
37af14fe
MD
1349 globaldata_t gd;
1350 globaldata_t mygd;
a2a5ad0d 1351
52eedfb5 1352 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1353 gd = td->td_gd;
37af14fe 1354 mygd = mycpu;
52eedfb5 1355 if (gd != mycpu) {
35238fa5 1356 cpu_lfence();
52eedfb5 1357 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1358 crit_enter_gd(mygd);
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 }
37af14fe 1365 td->td_gd = mygd;
52eedfb5
MD
1366 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1367 td->td_flags &= ~TDF_MIGRATING;
1368 crit_exit_gd(mygd);
1369 } else {
1370 crit_enter_gd(mygd);
1371 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1372 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1373 crit_exit_gd(mygd);
a2a5ad0d
MD
1374 }
1375}
1376
52eedfb5
MD
1377#endif
1378
f1d1c3fa
MD
1379/*
1380 * Generic deschedule. Descheduling threads other then your own should be
1381 * done only in carefully controlled circumstances. Descheduling is
1382 * asynchronous.
1383 *
1384 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1385 */
1386void
1387lwkt_deschedule(thread_t td)
1388{
f1d1c3fa 1389 crit_enter();
b8a98473 1390#ifdef SMP
f1d1c3fa
MD
1391 if (td == curthread) {
1392 _lwkt_dequeue(td);
1393 } else {
a72187e9 1394 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1395 _lwkt_dequeue(td);
1396 } else {
b8a98473 1397 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1398 }
1399 }
b8a98473
MD
1400#else
1401 _lwkt_dequeue(td);
1402#endif
f1d1c3fa
MD
1403 crit_exit();
1404}
1405
4b5f931b
MD
1406/*
1407 * Set the target thread's priority. This routine does not automatically
1408 * switch to a higher priority thread, LWKT threads are not designed for
1409 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1410 */
1411void
1412lwkt_setpri(thread_t td, int pri)
1413{
a72187e9 1414 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1415 if (td->td_pri != pri) {
1416 KKASSERT(pri >= 0);
1417 crit_enter();
1418 if (td->td_flags & TDF_RUNQ) {
1419 _lwkt_dequeue(td);
1420 td->td_pri = pri;
1421 _lwkt_enqueue(td);
1422 } else {
1423 td->td_pri = pri;
1424 }
1425 crit_exit();
26a0694b 1426 }
26a0694b
MD
1427}
1428
03bd0a5e
MD
1429/*
1430 * Set the initial priority for a thread prior to it being scheduled for
1431 * the first time. The thread MUST NOT be scheduled before or during
1432 * this call. The thread may be assigned to a cpu other then the current
1433 * cpu.
1434 *
1435 * Typically used after a thread has been created with TDF_STOPPREQ,
1436 * and before the thread is initially scheduled.
1437 */
1438void
1439lwkt_setpri_initial(thread_t td, int pri)
1440{
1441 KKASSERT(pri >= 0);
1442 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1443 td->td_pri = pri;
03bd0a5e
MD
1444}
1445
26a0694b
MD
1446void
1447lwkt_setpri_self(int pri)
1448{
1449 thread_t td = curthread;
1450
4b5f931b
MD
1451 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1452 crit_enter();
1453 if (td->td_flags & TDF_RUNQ) {
1454 _lwkt_dequeue(td);
f9235b6d 1455 td->td_pri = pri;
4b5f931b
MD
1456 _lwkt_enqueue(td);
1457 } else {
f9235b6d 1458 td->td_pri = pri;
4b5f931b
MD
1459 }
1460 crit_exit();
1461}
1462
f9235b6d
MD
1463/*
1464 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1465 *
1466 * Example: two competing threads, same priority N. decrement by (2*N)
1467 * increment by N*8, each thread will get 4 ticks.
1468 */
1469void
1470lwkt_fairq_schedulerclock(thread_t td)
1471{
1472 if (fairq_enable) {
1473 while (td) {
1474 if (td != &td->td_gd->gd_idlethread) {
1475 td->td_fairq_accum -= td->td_gd->gd_fairq_total_pri;
1476 if (td->td_fairq_accum < -TDFAIRQ_MAX(td->td_gd))
1477 td->td_fairq_accum = -TDFAIRQ_MAX(td->td_gd);
1478 if (td->td_fairq_accum < 0)
1479 need_lwkt_resched();
1480 td->td_fairq_lticks = ticks;
1481 }
1482 td = td->td_preempted;
1483 }
1484 }
1485}
1486
1487static void
1488lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1489{
1490 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1491 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1492 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1493}
1494
5d21b981 1495/*
52eedfb5
MD
1496 * Migrate the current thread to the specified cpu.
1497 *
1498 * This is accomplished by descheduling ourselves from the current cpu,
1499 * moving our thread to the tdallq of the target cpu, IPI messaging the
1500 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1501 * races while the thread is being migrated.
ae8e83e6
MD
1502 *
1503 * We must be sure to remove ourselves from the current cpu's tsleepq
1504 * before potentially moving to another queue. The thread can be on
1505 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1506 */
3d28ff59 1507#ifdef SMP
5d21b981 1508static void lwkt_setcpu_remote(void *arg);
3d28ff59 1509#endif
5d21b981
MD
1510
1511void
1512lwkt_setcpu_self(globaldata_t rgd)
1513{
1514#ifdef SMP
1515 thread_t td = curthread;
1516
1517 if (td->td_gd != rgd) {
1518 crit_enter_quick(td);
ae8e83e6 1519 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1520 tsleep_remove(td);
5d21b981
MD
1521 td->td_flags |= TDF_MIGRATING;
1522 lwkt_deschedule_self(td);
52eedfb5 1523 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1524 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1525 lwkt_switch();
1526 /* we are now on the target cpu */
52eedfb5 1527 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1528 crit_exit_quick(td);
1529 }
1530#endif
1531}
1532
ecdefdda
MD
1533void
1534lwkt_migratecpu(int cpuid)
1535{
1536#ifdef SMP
1537 globaldata_t rgd;
1538
1539 rgd = globaldata_find(cpuid);
1540 lwkt_setcpu_self(rgd);
1541#endif
1542}
1543
5d21b981
MD
1544/*
1545 * Remote IPI for cpu migration (called while in a critical section so we
1546 * do not have to enter another one). The thread has already been moved to
1547 * our cpu's allq, but we must wait for the thread to be completely switched
1548 * out on the originating cpu before we schedule it on ours or the stack
1549 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1550 * change to main memory.
1551 *
1552 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1553 * against wakeups. It is best if this interface is used only when there
1554 * are no pending events that might try to schedule the thread.
1555 */
3d28ff59 1556#ifdef SMP
5d21b981
MD
1557static void
1558lwkt_setcpu_remote(void *arg)
1559{
1560 thread_t td = arg;
1561 globaldata_t gd = mycpu;
1562
df910c23
MD
1563 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1564#ifdef SMP
1565 lwkt_process_ipiq();
1566#endif
35238fa5 1567 cpu_lfence();
df910c23 1568 }
5d21b981 1569 td->td_gd = gd;
35238fa5 1570 cpu_sfence();
5d21b981 1571 td->td_flags &= ~TDF_MIGRATING;
9388413d 1572 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1573 _lwkt_enqueue(td);
1574}
3d28ff59 1575#endif
5d21b981 1576
553ea3c8 1577struct lwp *
4b5f931b
MD
1578lwkt_preempted_proc(void)
1579{
73e4f7b9 1580 thread_t td = curthread;
4b5f931b
MD
1581 while (td->td_preempted)
1582 td = td->td_preempted;
553ea3c8 1583 return(td->td_lwp);
4b5f931b
MD
1584}
1585
99df837e
MD
1586/*
1587 * Create a kernel process/thread/whatever. It shares it's address space
1588 * with proc0 - ie: kernel only.
1589 *
365fa13f
MD
1590 * NOTE! By default new threads are created with the MP lock held. A
1591 * thread which does not require the MP lock should release it by calling
1592 * rel_mplock() at the start of the new thread.
99df837e
MD
1593 */
1594int
c9e9fb21
MD
1595lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1596 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1597{
73e4f7b9 1598 thread_t td;
e2565a42 1599 __va_list ap;
99df837e 1600
d3d32139 1601 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1602 tdflags);
a2a5ad0d
MD
1603 if (tdp)
1604 *tdp = td;
709799ea 1605 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1606
1607 /*
1608 * Set up arg0 for 'ps' etc
1609 */
e2565a42 1610 __va_start(ap, fmt);
379210cb 1611 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1612 __va_end(ap);
99df837e
MD
1613
1614 /*
1615 * Schedule the thread to run
1616 */
ef0fdad1
MD
1617 if ((td->td_flags & TDF_STOPREQ) == 0)
1618 lwkt_schedule(td);
1619 else
1620 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1621 return 0;
1622}
1623
1624/*
1625 * Destroy an LWKT thread. Warning! This function is not called when
1626 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1627 * uses a different reaping mechanism.
1628 */
1629void
1630lwkt_exit(void)
1631{
1632 thread_t td = curthread;
c070746a 1633 thread_t std;
8826f33a 1634 globaldata_t gd;
99df837e
MD
1635
1636 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1637 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1638 caps_exit(td);
c070746a
MD
1639
1640 /*
1641 * Get us into a critical section to interlock gd_freetd and loop
1642 * until we can get it freed.
1643 *
1644 * We have to cache the current td in gd_freetd because objcache_put()ing
1645 * it would rip it out from under us while our thread is still active.
1646 */
1647 gd = mycpu;
37af14fe 1648 crit_enter_quick(td);
c070746a
MD
1649 while ((std = gd->gd_freetd) != NULL) {
1650 gd->gd_freetd = NULL;
1651 objcache_put(thread_cache, std);
1652 }
3b4192fb
MD
1653
1654 /*
1655 * Remove thread resources from kernel lists and deschedule us for
1656 * the last time.
1657 */
1658 if (td->td_flags & TDF_TSLEEPQ)
1659 tsleep_remove(td);
79eae878 1660 biosched_done(td);
f8abf63c 1661 dsched_exit_thread(td);
37af14fe 1662 lwkt_deschedule_self(td);
e56e4dea 1663 lwkt_remove_tdallq(td);
c070746a
MD
1664 if (td->td_flags & TDF_ALLOCATED_THREAD)
1665 gd->gd_freetd = td;
99df837e
MD
1666 cpu_thread_exit();
1667}
1668
e56e4dea
MD
1669void
1670lwkt_remove_tdallq(thread_t td)
1671{
1672 KKASSERT(td->td_gd == mycpu);
1673 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1674}
1675
9cf43f91
MD
1676/*
1677 * Code reduction and branch prediction improvements. Call/return
1678 * overhead on modern cpus often degenerates into 0 cycles due to
1679 * the cpu's branch prediction hardware and return pc cache. We
1680 * can take advantage of this by not inlining medium-complexity
1681 * functions and we can also reduce the branch prediction impact
1682 * by collapsing perfectly predictable branches into a single
1683 * procedure instead of duplicating it.
1684 *
1685 * Is any of this noticeable? Probably not, so I'll take the
1686 * smaller code size.
1687 */
1688void
b6468f56 1689crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1690{
b6468f56 1691 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1692}
1693
2d93b37a
MD
1694void
1695crit_panic(void)
1696{
1697 thread_t td = curthread;
850634cc 1698 int lcrit = td->td_critcount;
2d93b37a 1699
850634cc
AH
1700 td->td_critcount = 0;
1701 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1702 /* NOT REACHED */
2d93b37a
MD
1703}
1704
d165e668
MD
1705#ifdef SMP
1706
bd8015ca
MD
1707/*
1708 * Called from debugger/panic on cpus which have been stopped. We must still
1709 * process the IPIQ while stopped, even if we were stopped while in a critical
1710 * section (XXX).
1711 *
1712 * If we are dumping also try to process any pending interrupts. This may
1713 * or may not work depending on the state of the cpu at the point it was
1714 * stopped.
1715 */
1716void
1717lwkt_smp_stopped(void)
1718{
1719 globaldata_t gd = mycpu;
1720
1721 crit_enter_gd(gd);
1722 if (dumping) {
1723 lwkt_process_ipiq();
1724 splz();
1725 } else {
1726 lwkt_process_ipiq();
1727 }
1728 crit_exit_gd(gd);
1729}
1730
d165e668 1731#endif