kernel - Add debugging features
[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;
0c52fa62
SG
133SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, &fairq_enable, 0,
134 "Turn on fairq priority accumulators");
77912481 135static int user_pri_sched = 0;
0c52fa62
SG
136SYSCTL_INT(_lwkt, OID_AUTO, user_pri_sched, CTLFLAG_RW, &user_pri_sched, 0,
137 "");
fbc024e4 138static int preempt_enable = 1;
0c52fa62
SG
139SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, &preempt_enable, 0,
140 "Enable preemption");
fbc024e4 141
05220613 142
4b5f931b
MD
143/*
144 * These helper procedures handle the runq, they can only be called from
145 * within a critical section.
75cdbe6c
MD
146 *
147 * WARNING! Prior to SMP being brought up it is possible to enqueue and
148 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
149 * instead of 'mycpu' when referencing the globaldata structure. Once
150 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 151 */
f1d1c3fa
MD
152static __inline
153void
154_lwkt_dequeue(thread_t td)
155{
156 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 157 struct globaldata *gd = td->td_gd;
4b5f931b 158
f1d1c3fa 159 td->td_flags &= ~TDF_RUNQ;
f9235b6d
MD
160 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
161 gd->gd_fairq_total_pri -= td->td_pri;
162 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
163 atomic_clear_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
164 }
165}
166
f9235b6d
MD
167/*
168 * Priority enqueue.
169 *
170 * NOTE: There are a limited number of lwkt threads runnable since user
171 * processes only schedule one at a time per cpu.
172 */
f1d1c3fa
MD
173static __inline
174void
175_lwkt_enqueue(thread_t td)
176{
f9235b6d
MD
177 thread_t xtd;
178
7f5d7ed7 179 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 180 struct globaldata *gd = td->td_gd;
4b5f931b 181
f1d1c3fa 182 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
183 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
184 if (xtd == NULL) {
185 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
186 atomic_set_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
187 } else {
188 while (xtd && xtd->td_pri > td->td_pri)
189 xtd = TAILQ_NEXT(xtd, td_threadq);
190 if (xtd)
191 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
192 else
193 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
194 }
195 gd->gd_fairq_total_pri += td->td_pri;
f1d1c3fa
MD
196 }
197}
8ad65e08 198
40aaf5fc
NT
199static __boolean_t
200_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
201{
202 struct thread *td = (struct thread *)obj;
203
204 td->td_kstack = NULL;
205 td->td_kstack_size = 0;
206 td->td_flags = TDF_ALLOCATED_THREAD;
207 return (1);
208}
209
210static void
211_lwkt_thread_dtor(void *obj, void *privdata)
212{
213 struct thread *td = (struct thread *)obj;
214
215 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
216 ("_lwkt_thread_dtor: not allocated from objcache"));
217 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
218 td->td_kstack_size > 0,
219 ("_lwkt_thread_dtor: corrupted stack"));
220 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
221}
222
223/*
224 * Initialize the lwkt s/system.
225 */
226void
227lwkt_init(void)
228{
229 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
230 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
231 NULL, CACHE_NTHREADS/2,
232 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
233}
234
37af14fe
MD
235/*
236 * Schedule a thread to run. As the current thread we can always safely
237 * schedule ourselves, and a shortcut procedure is provided for that
238 * function.
239 *
240 * (non-blocking, self contained on a per cpu basis)
241 */
242void
243lwkt_schedule_self(thread_t td)
244{
245 crit_enter_quick(td);
f9235b6d
MD
246 KASSERT(td != &td->td_gd->gd_idlethread,
247 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 248 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 249 _lwkt_enqueue(td);
37af14fe
MD
250 crit_exit_quick(td);
251}
252
253/*
254 * Deschedule a thread.
255 *
256 * (non-blocking, self contained on a per cpu basis)
257 */
258void
259lwkt_deschedule_self(thread_t td)
260{
261 crit_enter_quick(td);
37af14fe
MD
262 _lwkt_dequeue(td);
263 crit_exit_quick(td);
264}
265
8ad65e08
MD
266/*
267 * LWKTs operate on a per-cpu basis
268 *
73e4f7b9 269 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
270 */
271void
272lwkt_gdinit(struct globaldata *gd)
273{
f9235b6d 274 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 275 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
276}
277
7d0bac62
MD
278/*
279 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
280 * or LWKT start address before it can be scheduled. If the target cpu is
281 * -1 the thread will be created on the current cpu.
0cfcada1
MD
282 *
283 * If you intend to create a thread without a process context this function
284 * does everything except load the startup and switcher function.
7d0bac62
MD
285 */
286thread_t
d3d32139 287lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 288{
c070746a 289 globaldata_t gd = mycpu;
99df837e 290 void *stack;
7d0bac62 291
c070746a
MD
292 /*
293 * If static thread storage is not supplied allocate a thread. Reuse
294 * a cached free thread if possible. gd_freetd is used to keep an exiting
295 * thread intact through the exit.
296 */
ef0fdad1 297 if (td == NULL) {
cf709dd2
MD
298 crit_enter_gd(gd);
299 if ((td = gd->gd_freetd) != NULL) {
300 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
301 TDF_RUNQ)) == 0);
c070746a 302 gd->gd_freetd = NULL;
cf709dd2 303 } else {
c070746a 304 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
305 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
306 TDF_RUNQ)) == 0);
307 }
308 crit_exit_gd(gd);
40aaf5fc
NT
309 KASSERT((td->td_flags &
310 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
311 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
312 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 313 }
c070746a
MD
314
315 /*
316 * Try to reuse cached stack.
317 */
f470d0c8
MD
318 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
319 if (flags & TDF_ALLOCATED_STACK) {
e4846942 320 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
321 stack = NULL;
322 }
323 }
324 if (stack == NULL) {
e40cfbd7 325 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 326 flags |= TDF_ALLOCATED_STACK;
99df837e 327 }
75cdbe6c 328 if (cpu < 0)
c070746a 329 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 330 else
f470d0c8 331 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 332 return(td);
7d0bac62
MD
333}
334
335/*
336 * Initialize a preexisting thread structure. This function is used by
337 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
338 *
f8c3996b
MD
339 * All threads start out in a critical section at a priority of
340 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
341 * appropriate. This function may send an IPI message when the
342 * requested cpu is not the current cpu and consequently gd_tdallq may
343 * not be initialized synchronously from the point of view of the originating
344 * cpu.
345 *
346 * NOTE! we have to be careful in regards to creating threads for other cpus
347 * if SMP has not yet been activated.
7d0bac62 348 */
41a01a4d
MD
349#ifdef SMP
350
75cdbe6c
MD
351static void
352lwkt_init_thread_remote(void *arg)
353{
354 thread_t td = arg;
355
52eedfb5
MD
356 /*
357 * Protected by critical section held by IPI dispatch
358 */
75cdbe6c
MD
359 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
360}
361
41a01a4d
MD
362#endif
363
fdce8919
MD
364/*
365 * lwkt core thread structural initialization.
366 *
367 * NOTE: All threads are initialized as mpsafe threads.
368 */
7d0bac62 369void
f470d0c8
MD
370lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
371 struct globaldata *gd)
7d0bac62 372{
37af14fe
MD
373 globaldata_t mygd = mycpu;
374
99df837e
MD
375 bzero(td, sizeof(struct thread));
376 td->td_kstack = stack;
f470d0c8 377 td->td_kstack_size = stksize;
d3d32139 378 td->td_flags = flags;
26a0694b 379 td->td_gd = gd;
f9235b6d
MD
380 td->td_pri = TDPRI_KERN_DAEMON;
381 td->td_critcount = 1;
3b998fa9 382 td->td_toks_stop = &td->td_toks_base;
fb0f29c4
MD
383 if (lwkt_use_spin_port)
384 lwkt_initport_spin(&td->td_msgport);
385 else
386 lwkt_initport_thread(&td->td_msgport, td);
99df837e 387 pmap_init_thread(td);
0f7a3396 388#ifdef SMP
5d21b981
MD
389 /*
390 * Normally initializing a thread for a remote cpu requires sending an
391 * IPI. However, the idlethread is setup before the other cpus are
392 * activated so we have to treat it as a special case. XXX manipulation
393 * of gd_tdallq requires the BGL.
394 */
395 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 396 crit_enter_gd(mygd);
75cdbe6c 397 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 398 crit_exit_gd(mygd);
75cdbe6c 399 } else {
2db3b277 400 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 401 }
0f7a3396 402#else
37af14fe 403 crit_enter_gd(mygd);
0f7a3396 404 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 405 crit_exit_gd(mygd);
0f7a3396 406#endif
8c72e3d5
AH
407
408 dsched_new_thread(td);
73e4f7b9
MD
409}
410
411void
412lwkt_set_comm(thread_t td, const char *ctl, ...)
413{
e2565a42 414 __va_list va;
73e4f7b9 415
e2565a42 416 __va_start(va, ctl);
379210cb 417 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 418 __va_end(va);
e7c0dbba 419 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
420}
421
99df837e 422void
73e4f7b9 423lwkt_hold(thread_t td)
99df837e 424{
73e4f7b9
MD
425 ++td->td_refs;
426}
427
428void
429lwkt_rele(thread_t td)
430{
431 KKASSERT(td->td_refs > 0);
432 --td->td_refs;
433}
434
435void
436lwkt_wait_free(thread_t td)
437{
438 while (td->td_refs)
377d4740 439 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
440}
441
442void
443lwkt_free_thread(thread_t td)
444{
cf709dd2 445 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|TDF_RUNQ)) == 0);
40aaf5fc
NT
446 if (td->td_flags & TDF_ALLOCATED_THREAD) {
447 objcache_put(thread_cache, td);
448 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
449 /* client-allocated struct with internally allocated stack */
450 KASSERT(td->td_kstack && td->td_kstack_size > 0,
451 ("lwkt_free_thread: corrupted stack"));
452 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
453 td->td_kstack = NULL;
454 td->td_kstack_size = 0;
99df837e 455 }
e7c0dbba 456 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
457}
458
459
8ad65e08
MD
460/*
461 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
462 * switch to the idlethread. Switching must occur within a critical
463 * section to avoid races with the scheduling queue.
464 *
465 * We always have full control over our cpu's run queue. Other cpus
466 * that wish to manipulate our queue must use the cpu_*msg() calls to
467 * talk to our cpu, so a critical section is all that is needed and
468 * the result is very, very fast thread switching.
469 *
96728c05
MD
470 * The LWKT scheduler uses a fixed priority model and round-robins at
471 * each priority level. User process scheduling is a totally
472 * different beast and LWKT priorities should not be confused with
473 * user process priorities.
f1d1c3fa 474 *
3933a3ab
MD
475 * The MP lock may be out of sync with the thread's td_mpcount + td_xpcount.
476 * lwkt_switch() cleans it up.
477 *
478 * Note that the td_switch() function cannot do anything that requires
479 * the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
480 * the target thread (not the current thread). It's nice to have a scheduler
481 * that does not need the MP lock to work because it allows us to do some
482 * really cool high-performance MP lock optimizations.
69d78e99
MD
483 *
484 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
485 * is not called by the current thread in the preemption case, only when
486 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
487 */
488void
489lwkt_switch(void)
490{
37af14fe
MD
491 globaldata_t gd = mycpu;
492 thread_t td = gd->gd_curthread;
8ad65e08 493 thread_t ntd;
f9235b6d
MD
494 thread_t xtd;
495 thread_t nlast;
f9235b6d 496 int nquserok;
6f207a2c 497#ifdef SMP
8a8d5d85
MD
498 int mpheld;
499#endif
f9235b6d 500 int didaccumulate;
b37f18d6
MD
501 const char *lmsg; /* diagnostic - 'systat -pv 1' */
502 const void *laddr;
8ad65e08 503
46a3f46d 504 /*
27e88a6e
MD
505 * Switching from within a 'fast' (non thread switched) interrupt or IPI
506 * is illegal. However, we may have to do it anyway if we hit a fatal
507 * kernel trap or we have paniced.
508 *
509 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 510 */
27e88a6e
MD
511 if (gd->gd_intr_nesting_level) {
512 int savegdnest;
513 int savegdtrap;
514
5fddbda2 515 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22
MD
516 panic("lwkt_switch: Attempt to switch from a "
517 "a fast interrupt, ipi, or hard code section, "
518 "td %p\n",
519 td);
27e88a6e
MD
520 } else {
521 savegdnest = gd->gd_intr_nesting_level;
522 savegdtrap = gd->gd_trap_nesting_level;
523 gd->gd_intr_nesting_level = 0;
524 gd->gd_trap_nesting_level = 0;
a7422615
MD
525 if ((td->td_flags & TDF_PANICWARN) == 0) {
526 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
527 kprintf("Warning: thread switch from interrupt, IPI, "
528 "or hard code section.\n"
a7422615 529 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 530 print_backtrace(-1);
a7422615 531 }
27e88a6e
MD
532 lwkt_switch();
533 gd->gd_intr_nesting_level = savegdnest;
534 gd->gd_trap_nesting_level = savegdtrap;
535 return;
536 }
96728c05 537 }
ef0fdad1 538
cb973d15
MD
539 /*
540 * Passive release (used to transition from user to kernel mode
541 * when we block or switch rather then when we enter the kernel).
542 * This function is NOT called if we are switching into a preemption
543 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
544 * our current process designation (if we have one) and become a true
545 * LWKT thread, and may also hand the current process designation to
546 * another process and schedule thread.
cb973d15
MD
547 */
548 if (td->td_release)
549 td->td_release(td);
550
37af14fe 551 crit_enter_gd(gd);
3b998fa9 552 if (TD_TOKS_HELD(td))
9d265729
MD
553 lwkt_relalltokens(td);
554
555 /*
b02926de
MD
556 * We had better not be holding any spin locks, but don't get into an
557 * endless panic loop.
9d265729 558 */
d666840a
MD
559 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
560 ("lwkt_switch: still holding %d exclusive spinlocks!",
561 gd->gd_spinlocks_wr));
9d265729 562
8a8d5d85
MD
563
564#ifdef SMP
565 /*
3933a3ab
MD
566 * td_mpcount + td_xpcount cannot be used to determine if we currently
567 * hold the MP lock because get_mplock() will increment it prior to
568 * attempting to get the lock, and switch out if it can't. Our
569 * ownership of the actual lock will remain stable while we are
570 * in a critical section, and once we actually acquire the underlying
571 * lock as long as the count is greater than 0.
8a8d5d85 572 */
c5724852 573 mpheld = MP_LOCK_HELD(gd);
0f7a3396
MD
574#ifdef INVARIANTS
575 if (td->td_cscount) {
6ea70f76 576 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
577 td);
578 if (panic_on_cscount)
579 panic("switching while mastering cpusync");
580 }
581#endif
8a8d5d85 582#endif
f9235b6d
MD
583
584 /*
585 * If we had preempted another thread on this cpu, resume the preempted
586 * thread. This occurs transparently, whether the preempted thread
587 * was scheduled or not (it may have been preempted after descheduling
588 * itself).
589 *
590 * We have to setup the MP lock for the original thread after backing
591 * out the adjustment that was made to curthread when the original
592 * was preempted.
593 */
99df837e 594 if ((ntd = td->td_preempted) != NULL) {
26a0694b 595 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 596#ifdef SMP
3933a3ab 597 if (ntd->td_mpcount + ntd->td_xpcount && mpheld == 0) {
fc92d4aa 598 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
3933a3ab 599 td, ntd, td->td_mpcount, ntd->td_mpcount + ntd->td_xpcount);
8a8d5d85 600 }
3933a3ab 601 td->td_xpcount = 0;
8a8d5d85 602#endif
26a0694b 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 *
622 * We have to adjust the MP lock for the target thread. If we
623 * need the MP lock and cannot obtain it we try to locate a
624 * thread that does not need the MP lock. If we cannot, we spin
625 * instead of HLT.
626 *
627 * A similar issue exists for the tokens held by the target thread.
628 * If we cannot obtain ownership of the tokens we cannot immediately
629 * schedule the thread.
630 */
631 for (;;) {
632 clear_lwkt_resched();
633 didaccumulate = 0;
634 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
635
4b5f931b 636 /*
f9235b6d 637 * Hotpath if we can get all necessary resources.
41a01a4d 638 *
f9235b6d 639 * If nothing is runnable switch to the idle thread
41a01a4d 640 */
f9235b6d
MD
641 if (ntd == NULL) {
642 ntd = &gd->gd_idlethread;
643 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
644 ntd->td_flags |= TDF_IDLE_NOHLT;
6f207a2c 645#ifdef SMP
3933a3ab 646 KKASSERT(ntd->td_xpcount == 0);
f9235b6d
MD
647 if (ntd->td_mpcount) {
648 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
649 panic("Idle thread %p was holding the BGL!", ntd);
650 if (mpheld == 0) {
c5724852
MD
651 set_cpu_contention_mask(gd);
652 handle_cpu_contention_mask();
653 cpu_try_mplock();
654 mpheld = MP_LOCK_HELD(gd);
f9235b6d
MD
655 cpu_pause();
656 continue;
657 }
658 }
c5724852 659 clr_cpu_contention_mask(gd);
6f207a2c 660#endif
b37f18d6
MD
661 cpu_time.cp_msg[0] = 0;
662 cpu_time.cp_stallpc = 0;
f9235b6d
MD
663 goto haveidle;
664 }
41a01a4d 665
8ec60c3f 666 /*
f9235b6d 667 * Hotpath schedule
6f207a2c
MD
668 *
669 * NOTE: For UP there is no mplock and lwkt_getalltokens()
670 * always succeeds.
8ec60c3f 671 */
f9235b6d
MD
672 if (ntd->td_fairq_accum >= 0 &&
673#ifdef SMP
3933a3ab
MD
674 (ntd->td_mpcount + ntd->td_xpcount == 0 ||
675 mpheld || cpu_try_mplock()) &&
f9235b6d 676#endif
b37f18d6 677 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
f9235b6d 678 ) {
8a8d5d85 679#ifdef SMP
c5724852 680 clr_cpu_contention_mask(gd);
f9235b6d
MD
681#endif
682 goto havethread;
683 }
684
b37f18d6
MD
685 lmsg = NULL;
686 laddr = NULL;
687
f9235b6d 688#ifdef SMP
c5724852
MD
689 if (ntd->td_fairq_accum >= 0)
690 set_cpu_contention_mask(gd);
f9235b6d 691 /* Reload mpheld (it become stale after mplock/token ops) */
c5724852 692 mpheld = MP_LOCK_HELD(gd);
3933a3ab 693 if (ntd->td_mpcount + ntd->td_xpcount && mpheld == 0) {
b37f18d6
MD
694 lmsg = "mplock";
695 laddr = ntd->td_mplock_stallpc;
696 }
f9235b6d
MD
697#endif
698
699 /*
700 * Coldpath - unable to schedule ntd, continue looking for threads
701 * to schedule. This is only allowed of the (presumably) kernel
702 * thread exhausted its fair share. A kernel thread stuck on
703 * resources does not currently allow a user thread to get in
704 * front of it.
705 */
706#ifdef SMP
707 nquserok = ((ntd->td_pri < TDPRI_KERN_LPSCHED) ||
708 (ntd->td_fairq_accum < 0));
6f207a2c
MD
709#else
710 nquserok = 1;
f9235b6d
MD
711#endif
712 nlast = NULL;
713
714 for (;;) {
41a01a4d 715 /*
f9235b6d
MD
716 * If the fair-share scheduler ran out ntd gets moved to the
717 * end and its accumulator will be bumped, if it didn't we
718 * maintain the same queue position.
df6b8ba0 719 *
f9235b6d 720 * nlast keeps track of the last element prior to any moves.
41a01a4d 721 */
f9235b6d 722 if (ntd->td_fairq_accum < 0) {
f9235b6d
MD
723 lwkt_fairq_accumulate(gd, ntd);
724 didaccumulate = 1;
c5724852
MD
725
726 /*
727 * Move to end
728 */
729 xtd = TAILQ_NEXT(ntd, td_threadq);
f9235b6d
MD
730 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
731 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
c5724852
MD
732
733 /*
734 * Set terminal element (nlast)
735 */
f9235b6d
MD
736 if (nlast == NULL) {
737 nlast = ntd;
738 if (xtd == NULL)
739 xtd = ntd;
740 }
741 ntd = xtd;
742 } else {
743 ntd = TAILQ_NEXT(ntd, td_threadq);
744 }
a453459d 745
f9235b6d
MD
746 /*
747 * If we exhausted the run list switch to the idle thread.
748 * Since one or more threads had resource acquisition issues
749 * we do not allow the idle thread to halt.
750 *
751 * NOTE: nlast can be NULL.
752 */
753 if (ntd == nlast) {
e0a90d3b 754 cpu_pause();
f9235b6d
MD
755 ntd = &gd->gd_idlethread;
756 ntd->td_flags |= TDF_IDLE_NOHLT;
6f207a2c 757#ifdef SMP
3933a3ab 758 KKASSERT(ntd->td_xpcount == 0);
f9235b6d 759 if (ntd->td_mpcount) {
c5724852 760 mpheld = MP_LOCK_HELD(gd);
f9235b6d
MD
761 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
762 panic("Idle thread %p was holding the BGL!", ntd);
763 if (mpheld == 0) {
c5724852
MD
764 set_cpu_contention_mask(gd);
765 handle_cpu_contention_mask();
766 cpu_try_mplock();
767 mpheld = MP_LOCK_HELD(gd);
f9235b6d
MD
768 cpu_pause();
769 break; /* try again from the top, almost */
b9eb1c19 770 }
8a8d5d85 771 }
6f207a2c 772#endif
684a93c4
MD
773
774 /*
f9235b6d
MD
775 * If fairq accumulations occured we do not schedule the
776 * idle thread. This will cause us to try again from
777 * the (almost) top.
684a93c4 778 */
f9235b6d 779 if (didaccumulate)
b37f18d6
MD
780 break; /* try again from the top, almost */
781 if (lmsg)
782 strlcpy(cpu_time.cp_msg, lmsg, sizeof(cpu_time.cp_msg));
783 cpu_time.cp_stallpc = (uintptr_t)laddr;
f9235b6d 784 goto haveidle;
8a8d5d85 785 }
f9235b6d 786
df6b8ba0 787 /*
f9235b6d 788 * Try to switch to this thread.
6f207a2c
MD
789 *
790 * NOTE: For UP there is no mplock and lwkt_getalltokens()
791 * always succeeds.
df6b8ba0 792 */
77912481
MD
793 if ((ntd->td_pri >= TDPRI_KERN_LPSCHED || nquserok ||
794 user_pri_sched) && ntd->td_fairq_accum >= 0 &&
f9235b6d 795#ifdef SMP
3933a3ab
MD
796 (ntd->td_mpcount + ntd->td_xpcount == 0 ||
797 mpheld || cpu_try_mplock()) &&
8a8d5d85 798#endif
b37f18d6 799 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
f9235b6d 800 ) {
a453459d 801#ifdef SMP
c5724852 802 clr_cpu_contention_mask(gd);
f9235b6d
MD
803#endif
804 goto havethread;
df6b8ba0 805 }
9ac1ee6e
MD
806
807 /*
808 * Thread was runnable but we were unable to get the required
809 * resources (tokens and/or mplock).
810 */
f9235b6d 811#ifdef SMP
9ac1ee6e 812 ntd->td_wmesg = lmsg;
c5724852
MD
813 if (ntd->td_fairq_accum >= 0)
814 set_cpu_contention_mask(gd);
815 /*
816 * Reload mpheld (it become stale after mplock/token ops).
817 */
818 mpheld = MP_LOCK_HELD(gd);
3933a3ab 819 if (ntd->td_mpcount + ntd->td_xpcount && mpheld == 0) {
b37f18d6
MD
820 lmsg = "mplock";
821 laddr = ntd->td_mplock_stallpc;
822 }
f9235b6d
MD
823 if (ntd->td_pri >= TDPRI_KERN_LPSCHED && ntd->td_fairq_accum >= 0)
824 nquserok = 0;
a453459d 825#endif
4b5f931b 826 }
c5724852
MD
827
828 /*
829 * All threads exhausted but we can loop due to a negative
830 * accumulator.
831 *
832 * While we are looping in the scheduler be sure to service
833 * any interrupts which were made pending due to our critical
834 * section, otherwise we could livelock (e.g.) IPIs.
835 *
836 * NOTE: splz can enter and exit the mplock so mpheld is
837 * stale after this call.
838 */
839 splz_check();
840
841#ifdef SMP
842 /*
843 * Our mplock can be cached and cause other cpus to livelock
844 * if we loop due to e.g. not being able to acquire tokens.
845 */
846 if (MP_LOCK_HELD(gd))
847 cpu_rel_mplock(gd->gd_cpuid);
848 mpheld = 0;
849#endif
f1d1c3fa 850 }
8a8d5d85
MD
851
852 /*
f9235b6d
MD
853 * Do the actual switch. WARNING: mpheld is stale here.
854 *
855 * We must always decrement td_fairq_accum on non-idle threads just
856 * in case a thread never gets a tick due to being in a continuous
857 * critical section. The page-zeroing code does that.
858 *
859 * If the thread we came up with is a higher or equal priority verses
860 * the thread at the head of the queue we move our thread to the
861 * front. This way we can always check the front of the queue.
862 */
863havethread:
864 ++gd->gd_cnt.v_swtch;
865 --ntd->td_fairq_accum;
9ac1ee6e 866 ntd->td_wmesg = NULL;
f9235b6d
MD
867 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
868 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
869 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
870 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
871 }
872havethread_preempted:
873 ;
874 /*
875 * If the new target does not need the MP lock and we are holding it,
876 * release the MP lock. If the new target requires the MP lock we have
877 * already acquired it for the target.
878 *
879 * WARNING: mpheld is stale here.
8a8d5d85 880 */
f9235b6d
MD
881haveidle:
882 KASSERT(ntd->td_critcount,
883 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85 884#ifdef SMP
3933a3ab 885 if (ntd->td_mpcount + ntd->td_xpcount == 0 ) {
c5724852
MD
886 if (MP_LOCK_HELD(gd))
887 cpu_rel_mplock(gd->gd_cpuid);
8a8d5d85 888 } else {
a453459d 889 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
890 }
891#endif
94f6d86e
MD
892 if (td != ntd) {
893 ++switch_count;
a1f0fb66 894 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
f1d1c3fa 895 td->td_switch(ntd);
94f6d86e 896 }
37af14fe
MD
897 /* NOTE: current cpu may have changed after switch */
898 crit_exit_quick(td);
8ad65e08
MD
899}
900
b68b7282 901/*
96728c05
MD
902 * Request that the target thread preempt the current thread. Preemption
903 * only works under a specific set of conditions:
b68b7282 904 *
96728c05
MD
905 * - We are not preempting ourselves
906 * - The target thread is owned by the current cpu
907 * - We are not currently being preempted
908 * - The target is not currently being preempted
d3d1cbc8
MD
909 * - We are not holding any spin locks
910 * - The target thread is not holding any tokens
96728c05
MD
911 * - We are able to satisfy the target's MP lock requirements (if any).
912 *
913 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
914 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 915 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
916 * to determine whether preemption is possible (aka usually just the crit
917 * priority of lwkt_schedule() itself).
b68b7282 918 *
26a0694b
MD
919 * XXX at the moment we run the target thread in a critical section during
920 * the preemption in order to prevent the target from taking interrupts
921 * that *WE* can't. Preemption is strictly limited to interrupt threads
922 * and interrupt-like threads, outside of a critical section, and the
923 * preempted source thread will be resumed the instant the target blocks
924 * whether or not the source is scheduled (i.e. preemption is supposed to
925 * be as transparent as possible).
4b5f931b 926 *
8a8d5d85
MD
927 * The target thread inherits our MP count (added to its own) for the
928 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
929 * MP lock during the preemption. Therefore, any preempting targets must be
930 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
931 * out of sync with the physical mp_lock, but we do not have to preserve
932 * the original ownership of the lock if it was out of synch (that is, we
933 * can leave it synchronized on return).
b68b7282
MD
934 */
935void
f9235b6d 936lwkt_preempt(thread_t ntd, int critcount)
b68b7282 937{
46a3f46d 938 struct globaldata *gd = mycpu;
0a3f9b47 939 thread_t td;
8a8d5d85
MD
940#ifdef SMP
941 int mpheld;
57c254db 942 int savecnt;
8a8d5d85 943#endif
2d910aaf 944 int save_gd_intr_nesting_level;
b68b7282 945
26a0694b 946 /*
96728c05
MD
947 * The caller has put us in a critical section. We can only preempt
948 * if the caller of the caller was not in a critical section (basically
f9235b6d 949 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 950 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 951 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
952 *
953 * YYY The target thread must be in a critical section (else it must
954 * inherit our critical section? I dunno yet).
41a01a4d 955 *
0a3f9b47 956 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b 957 */
f9235b6d 958 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 959
fbc024e4
MD
960 if (preempt_enable == 0) {
961 ++preempt_miss;
962 return;
963 }
964
0a3f9b47 965 td = gd->gd_curthread;
f9235b6d 966 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
967 ++preempt_miss;
968 return;
969 }
f9235b6d 970 if (td->td_critcount > critcount) {
96728c05 971 ++preempt_miss;
8ec60c3f 972 need_lwkt_resched();
96728c05
MD
973 return;
974 }
975#ifdef SMP
46a3f46d 976 if (ntd->td_gd != gd) {
96728c05 977 ++preempt_miss;
8ec60c3f 978 need_lwkt_resched();
96728c05
MD
979 return;
980 }
981#endif
41a01a4d 982 /*
77912481
MD
983 * We don't have to check spinlocks here as they will also bump
984 * td_critcount.
d3d1cbc8
MD
985 *
986 * Do not try to preempt if the target thread is holding any tokens.
987 * We could try to acquire the tokens but this case is so rare there
988 * is no need to support it.
41a01a4d 989 */
77912481
MD
990 KKASSERT(gd->gd_spinlocks_wr == 0);
991
3b998fa9 992 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8
MD
993 ++preempt_miss;
994 need_lwkt_resched();
995 return;
996 }
26a0694b
MD
997 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
998 ++preempt_weird;
8ec60c3f 999 need_lwkt_resched();
26a0694b
MD
1000 return;
1001 }
1002 if (ntd->td_preempted) {
4b5f931b 1003 ++preempt_hit;
8ec60c3f 1004 need_lwkt_resched();
26a0694b 1005 return;
b68b7282 1006 }
8a8d5d85 1007#ifdef SMP
a2a5ad0d 1008 /*
3933a3ab 1009 * NOTE: An interrupt might have occured just as we were transitioning
71ef2f5c 1010 * to or from the MP lock. In this case td_mpcount will be pre-disposed
3933a3ab
MD
1011 * (non-zero) but not actually synchronized with the mp_lock itself.
1012 * We can use it to imply an MP lock requirement for the preemption but
1013 * we cannot use it to test whether we hold the MP lock or not.
a2a5ad0d 1014 */
96728c05 1015 savecnt = td->td_mpcount;
c5724852 1016 mpheld = MP_LOCK_HELD(gd);
6d9b99db 1017 ntd->td_xpcount = td->td_mpcount + td->td_xpcount;
3933a3ab
MD
1018 if (mpheld == 0 && ntd->td_mpcount + ntd->td_xpcount && !cpu_try_mplock()) {
1019 ntd->td_xpcount = 0;
8a8d5d85 1020 ++preempt_miss;
8ec60c3f 1021 need_lwkt_resched();
8a8d5d85
MD
1022 return;
1023 }
1024#endif
26a0694b 1025
8ec60c3f
MD
1026 /*
1027 * Since we are able to preempt the current thread, there is no need to
1028 * call need_lwkt_resched().
2d910aaf
MD
1029 *
1030 * We must temporarily clear gd_intr_nesting_level around the switch
1031 * since switchouts from the target thread are allowed (they will just
1032 * return to our thread), and since the target thread has its own stack.
8ec60c3f 1033 */
26a0694b
MD
1034 ++preempt_hit;
1035 ntd->td_preempted = td;
1036 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1037 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1038 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1039 gd->gd_intr_nesting_level = 0;
26a0694b 1040 td->td_switch(ntd);
2d910aaf 1041 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1042
26a0694b 1043 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
1044#ifdef SMP
1045 KKASSERT(savecnt == td->td_mpcount);
c5724852 1046 mpheld = MP_LOCK_HELD(gd);
71ef2f5c 1047 if (mpheld && td->td_mpcount == 0)
c5724852 1048 cpu_rel_mplock(gd->gd_cpuid);
3933a3ab 1049 else if (mpheld == 0 && td->td_mpcount + td->td_xpcount)
96728c05
MD
1050 panic("lwkt_preempt(): MP lock was not held through");
1051#endif
26a0694b
MD
1052 ntd->td_preempted = NULL;
1053 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1054}
1055
f1d1c3fa 1056/*
faaeffac 1057 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1058 * This will work inside a critical section but not inside a hard code
1059 * section.
ef0fdad1 1060 *
f1d1c3fa
MD
1061 * (self contained on a per cpu basis)
1062 */
1063void
faaeffac 1064splz_check(void)
f1d1c3fa 1065{
7966cb69
MD
1066 globaldata_t gd = mycpu;
1067 thread_t td = gd->gd_curthread;
ef0fdad1 1068
4a28fe22
MD
1069 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1070 gd->gd_intr_nesting_level == 0 &&
1071 td->td_nest_count < 2)
1072 {
f1d1c3fa 1073 splz();
4a28fe22
MD
1074 }
1075}
1076
1077/*
1078 * This version is integrated into crit_exit, reqflags has already
1079 * been tested but td_critcount has not.
1080 *
1081 * We only want to execute the splz() on the 1->0 transition of
1082 * critcount and not in a hard code section or if too deeply nested.
1083 */
1084void
1085lwkt_maybe_splz(thread_t td)
1086{
1087 globaldata_t gd = td->td_gd;
1088
1089 if (td->td_critcount == 0 &&
1090 gd->gd_intr_nesting_level == 0 &&
1091 td->td_nest_count < 2)
1092 {
1093 splz();
1094 }
f1d1c3fa
MD
1095}
1096
8ad65e08 1097/*
f9235b6d
MD
1098 * This function is used to negotiate a passive release of the current
1099 * process/lwp designation with the user scheduler, allowing the user
1100 * scheduler to schedule another user thread. The related kernel thread
1101 * (curthread) continues running in the released state.
8ad65e08
MD
1102 */
1103void
f9235b6d 1104lwkt_passive_release(struct thread *td)
8ad65e08 1105{
f9235b6d
MD
1106 struct lwp *lp = td->td_lwp;
1107
1108 td->td_release = NULL;
1109 lwkt_setpri_self(TDPRI_KERN_USER);
1110 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1111}
1112
f9235b6d 1113
3824f392 1114/*
f9235b6d
MD
1115 * This implements a normal yield. This routine is virtually a nop if
1116 * there is nothing to yield to but it will always run any pending interrupts
1117 * if called from a critical section.
1118 *
1119 * This yield is designed for kernel threads without a user context.
1120 *
1121 * (self contained on a per cpu basis)
3824f392
MD
1122 */
1123void
f9235b6d 1124lwkt_yield(void)
3824f392 1125{
f9235b6d
MD
1126 globaldata_t gd = mycpu;
1127 thread_t td = gd->gd_curthread;
1128 thread_t xtd;
3824f392 1129
f9235b6d
MD
1130 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1131 splz();
1132 if (td->td_fairq_accum < 0) {
1133 lwkt_schedule_self(curthread);
1134 lwkt_switch();
1135 } else {
1136 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1137 if (xtd && xtd->td_pri > td->td_pri) {
1138 lwkt_schedule_self(curthread);
1139 lwkt_switch();
1140 }
1141 }
3824f392
MD
1142}
1143
1144/*
f9235b6d
MD
1145 * This yield is designed for kernel threads with a user context.
1146 *
1147 * The kernel acting on behalf of the user is potentially cpu-bound,
1148 * this function will efficiently allow other threads to run and also
1149 * switch to other processes by releasing.
3824f392
MD
1150 *
1151 * The lwkt_user_yield() function is designed to have very low overhead
1152 * if no yield is determined to be needed.
1153 */
1154void
1155lwkt_user_yield(void)
1156{
f9235b6d
MD
1157 globaldata_t gd = mycpu;
1158 thread_t td = gd->gd_curthread;
1159
1160 /*
1161 * Always run any pending interrupts in case we are in a critical
1162 * section.
1163 */
1164 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1165 splz();
3824f392
MD
1166
1167#ifdef SMP
1168 /*
1169 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1170 * kernel can prevent other cpus from servicing interrupt threads
1171 * which still require the MP lock (which is a lot of them). This
1172 * has a chaining effect since if the interrupt is blocked, so is
1173 * the event, so normal scheduling will not pick up on the problem.
1174 */
3933a3ab 1175 if (cpu_contention_mask && td->td_mpcount + td->td_xpcount) {
684a93c4 1176 yield_mplock(td);
3824f392
MD
1177 }
1178#endif
1179
1180 /*
f9235b6d
MD
1181 * Switch (which forces a release) if another kernel thread needs
1182 * the cpu, if userland wants us to resched, or if our kernel
1183 * quantum has run out.
3824f392 1184 */
f9235b6d
MD
1185 if (lwkt_resched_wanted() ||
1186 user_resched_wanted() ||
1187 td->td_fairq_accum < 0)
1188 {
3824f392 1189 lwkt_switch();
3824f392
MD
1190 }
1191
f9235b6d 1192#if 0
3824f392 1193 /*
f9235b6d
MD
1194 * Reacquire the current process if we are released.
1195 *
1196 * XXX not implemented atm. The kernel may be holding locks and such,
1197 * so we want the thread to continue to receive cpu.
3824f392 1198 */
f9235b6d
MD
1199 if (td->td_release == NULL && lp) {
1200 lp->lwp_proc->p_usched->acquire_curproc(lp);
1201 td->td_release = lwkt_passive_release;
1202 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1203 }
f9235b6d 1204#endif
b9eb1c19
MD
1205}
1206
8ad65e08 1207/*
f1d1c3fa
MD
1208 * Generic schedule. Possibly schedule threads belonging to other cpus and
1209 * deal with threads that might be blocked on a wait queue.
1210 *
0a3f9b47
MD
1211 * We have a little helper inline function which does additional work after
1212 * the thread has been enqueued, including dealing with preemption and
1213 * setting need_lwkt_resched() (which prevents the kernel from returning
1214 * to userland until it has processed higher priority threads).
6330a558
MD
1215 *
1216 * It is possible for this routine to be called after a failed _enqueue
1217 * (due to the target thread migrating, sleeping, or otherwise blocked).
1218 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1219 *
1220 * reschedok is an optimized constant propagated from lwkt_schedule() or
1221 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1222 * reschedule to be requested if the target thread has a higher priority.
1223 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1224 * be 0, prevented undesired reschedules.
8ad65e08 1225 */
0a3f9b47
MD
1226static __inline
1227void
f9235b6d 1228_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
0a3f9b47 1229{
b9eb1c19 1230 thread_t otd;
c730be20 1231
6330a558 1232 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1233 if (ntd->td_preemptable && reschedok) {
f9235b6d 1234 ntd->td_preemptable(ntd, ccount); /* YYY +token */
361d01dd 1235 } else if (reschedok) {
b9eb1c19 1236 otd = curthread;
f9235b6d 1237 if (ntd->td_pri > otd->td_pri)
c730be20 1238 need_lwkt_resched();
6330a558 1239 }
f9235b6d
MD
1240
1241 /*
1242 * Give the thread a little fair share scheduler bump if it
1243 * has been asleep for a while. This is primarily to avoid
1244 * a degenerate case for interrupt threads where accumulator
1245 * crosses into negative territory unnecessarily.
1246 */
1247 if (ntd->td_fairq_lticks != ticks) {
1248 ntd->td_fairq_lticks = ticks;
1249 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1250 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1251 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1252 }
0a3f9b47
MD
1253 }
1254}
1255
361d01dd 1256static __inline
8ad65e08 1257void
361d01dd 1258_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1259{
37af14fe
MD
1260 globaldata_t mygd = mycpu;
1261
cf709dd2
MD
1262 KASSERT(td != &td->td_gd->gd_idlethread,
1263 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1264 crit_enter_gd(mygd);
9388413d 1265 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1266 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1267 _lwkt_enqueue(td);
1268 } else {
f1d1c3fa 1269 /*
7cd8d145
MD
1270 * If we own the thread, there is no race (since we are in a
1271 * critical section). If we do not own the thread there might
1272 * be a race but the target cpu will deal with it.
f1d1c3fa 1273 */
0f7a3396 1274#ifdef SMP
7cd8d145 1275 if (td->td_gd == mygd) {
9d265729 1276 _lwkt_enqueue(td);
f9235b6d 1277 _lwkt_schedule_post(mygd, td, 1, reschedok);
f1d1c3fa 1278 } else {
e381e77c 1279 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1280 }
0f7a3396 1281#else
7cd8d145 1282 _lwkt_enqueue(td);
f9235b6d 1283 _lwkt_schedule_post(mygd, td, 1, reschedok);
0f7a3396 1284#endif
8ad65e08 1285 }
37af14fe 1286 crit_exit_gd(mygd);
8ad65e08
MD
1287}
1288
361d01dd
MD
1289void
1290lwkt_schedule(thread_t td)
1291{
1292 _lwkt_schedule(td, 1);
1293}
1294
1295void
1296lwkt_schedule_noresched(thread_t td)
1297{
1298 _lwkt_schedule(td, 0);
1299}
1300
88ebb169
SW
1301#ifdef SMP
1302
e381e77c
MD
1303/*
1304 * When scheduled remotely if frame != NULL the IPIQ is being
1305 * run via doreti or an interrupt then preemption can be allowed.
1306 *
1307 * To allow preemption we have to drop the critical section so only
1308 * one is present in _lwkt_schedule_post.
1309 */
1310static void
1311lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1312{
1313 thread_t td = curthread;
1314 thread_t ntd = arg;
1315
1316 if (frame && ntd->td_preemptable) {
1317 crit_exit_noyield(td);
1318 _lwkt_schedule(ntd, 1);
1319 crit_enter_quick(td);
1320 } else {
1321 _lwkt_schedule(ntd, 1);
1322 }
1323}
1324
d9eea1a5 1325/*
52eedfb5
MD
1326 * Thread migration using a 'Pull' method. The thread may or may not be
1327 * the current thread. It MUST be descheduled and in a stable state.
1328 * lwkt_giveaway() must be called on the cpu owning the thread.
1329 *
1330 * At any point after lwkt_giveaway() is called, the target cpu may
1331 * 'pull' the thread by calling lwkt_acquire().
1332 *
ae8e83e6
MD
1333 * We have to make sure the thread is not sitting on a per-cpu tsleep
1334 * queue or it will blow up when it moves to another cpu.
1335 *
52eedfb5 1336 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1337 */
52eedfb5
MD
1338void
1339lwkt_giveaway(thread_t td)
1340{
3b4192fb 1341 globaldata_t gd = mycpu;
52eedfb5 1342
3b4192fb
MD
1343 crit_enter_gd(gd);
1344 if (td->td_flags & TDF_TSLEEPQ)
1345 tsleep_remove(td);
1346 KKASSERT(td->td_gd == gd);
1347 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1348 td->td_flags |= TDF_MIGRATING;
1349 crit_exit_gd(gd);
52eedfb5
MD
1350}
1351
a2a5ad0d
MD
1352void
1353lwkt_acquire(thread_t td)
1354{
37af14fe
MD
1355 globaldata_t gd;
1356 globaldata_t mygd;
a2a5ad0d 1357
52eedfb5 1358 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1359 gd = td->td_gd;
37af14fe 1360 mygd = mycpu;
52eedfb5 1361 if (gd != mycpu) {
35238fa5 1362 cpu_lfence();
52eedfb5 1363 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1364 crit_enter_gd(mygd);
df910c23
MD
1365 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1366#ifdef SMP
1367 lwkt_process_ipiq();
1368#endif
52eedfb5 1369 cpu_lfence();
df910c23 1370 }
562273ea 1371 cpu_mfence();
37af14fe 1372 td->td_gd = mygd;
52eedfb5
MD
1373 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1374 td->td_flags &= ~TDF_MIGRATING;
1375 crit_exit_gd(mygd);
1376 } else {
1377 crit_enter_gd(mygd);
1378 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1379 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1380 crit_exit_gd(mygd);
a2a5ad0d
MD
1381 }
1382}
1383
52eedfb5
MD
1384#endif
1385
f1d1c3fa
MD
1386/*
1387 * Generic deschedule. Descheduling threads other then your own should be
1388 * done only in carefully controlled circumstances. Descheduling is
1389 * asynchronous.
1390 *
1391 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1392 */
1393void
1394lwkt_deschedule(thread_t td)
1395{
f1d1c3fa 1396 crit_enter();
b8a98473 1397#ifdef SMP
f1d1c3fa
MD
1398 if (td == curthread) {
1399 _lwkt_dequeue(td);
1400 } else {
a72187e9 1401 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1402 _lwkt_dequeue(td);
1403 } else {
b8a98473 1404 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1405 }
1406 }
b8a98473
MD
1407#else
1408 _lwkt_dequeue(td);
1409#endif
f1d1c3fa
MD
1410 crit_exit();
1411}
1412
4b5f931b
MD
1413/*
1414 * Set the target thread's priority. This routine does not automatically
1415 * switch to a higher priority thread, LWKT threads are not designed for
1416 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1417 */
1418void
1419lwkt_setpri(thread_t td, int pri)
1420{
a72187e9 1421 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1422 if (td->td_pri != pri) {
1423 KKASSERT(pri >= 0);
1424 crit_enter();
1425 if (td->td_flags & TDF_RUNQ) {
1426 _lwkt_dequeue(td);
1427 td->td_pri = pri;
1428 _lwkt_enqueue(td);
1429 } else {
1430 td->td_pri = pri;
1431 }
1432 crit_exit();
26a0694b 1433 }
26a0694b
MD
1434}
1435
03bd0a5e
MD
1436/*
1437 * Set the initial priority for a thread prior to it being scheduled for
1438 * the first time. The thread MUST NOT be scheduled before or during
1439 * this call. The thread may be assigned to a cpu other then the current
1440 * cpu.
1441 *
1442 * Typically used after a thread has been created with TDF_STOPPREQ,
1443 * and before the thread is initially scheduled.
1444 */
1445void
1446lwkt_setpri_initial(thread_t td, int pri)
1447{
1448 KKASSERT(pri >= 0);
1449 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1450 td->td_pri = pri;
03bd0a5e
MD
1451}
1452
26a0694b
MD
1453void
1454lwkt_setpri_self(int pri)
1455{
1456 thread_t td = curthread;
1457
4b5f931b
MD
1458 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1459 crit_enter();
1460 if (td->td_flags & TDF_RUNQ) {
1461 _lwkt_dequeue(td);
f9235b6d 1462 td->td_pri = pri;
4b5f931b
MD
1463 _lwkt_enqueue(td);
1464 } else {
f9235b6d 1465 td->td_pri = pri;
4b5f931b
MD
1466 }
1467 crit_exit();
1468}
1469
f9235b6d
MD
1470/*
1471 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1472 *
1473 * Example: two competing threads, same priority N. decrement by (2*N)
1474 * increment by N*8, each thread will get 4 ticks.
1475 */
1476void
1477lwkt_fairq_schedulerclock(thread_t td)
1478{
1479 if (fairq_enable) {
1480 while (td) {
1481 if (td != &td->td_gd->gd_idlethread) {
1482 td->td_fairq_accum -= td->td_gd->gd_fairq_total_pri;
1483 if (td->td_fairq_accum < -TDFAIRQ_MAX(td->td_gd))
1484 td->td_fairq_accum = -TDFAIRQ_MAX(td->td_gd);
1485 if (td->td_fairq_accum < 0)
1486 need_lwkt_resched();
1487 td->td_fairq_lticks = ticks;
1488 }
1489 td = td->td_preempted;
1490 }
1491 }
1492}
1493
1494static void
1495lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1496{
1497 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1498 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1499 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1500}
1501
5d21b981 1502/*
52eedfb5
MD
1503 * Migrate the current thread to the specified cpu.
1504 *
1505 * This is accomplished by descheduling ourselves from the current cpu,
1506 * moving our thread to the tdallq of the target cpu, IPI messaging the
1507 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1508 * races while the thread is being migrated.
ae8e83e6
MD
1509 *
1510 * We must be sure to remove ourselves from the current cpu's tsleepq
1511 * before potentially moving to another queue. The thread can be on
1512 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1513 */
3d28ff59 1514#ifdef SMP
5d21b981 1515static void lwkt_setcpu_remote(void *arg);
3d28ff59 1516#endif
5d21b981
MD
1517
1518void
1519lwkt_setcpu_self(globaldata_t rgd)
1520{
1521#ifdef SMP
1522 thread_t td = curthread;
1523
1524 if (td->td_gd != rgd) {
1525 crit_enter_quick(td);
ae8e83e6 1526 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1527 tsleep_remove(td);
5d21b981
MD
1528 td->td_flags |= TDF_MIGRATING;
1529 lwkt_deschedule_self(td);
52eedfb5 1530 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1531 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1532 lwkt_switch();
1533 /* we are now on the target cpu */
52eedfb5 1534 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1535 crit_exit_quick(td);
1536 }
1537#endif
1538}
1539
ecdefdda
MD
1540void
1541lwkt_migratecpu(int cpuid)
1542{
1543#ifdef SMP
1544 globaldata_t rgd;
1545
1546 rgd = globaldata_find(cpuid);
1547 lwkt_setcpu_self(rgd);
1548#endif
1549}
1550
5d21b981
MD
1551/*
1552 * Remote IPI for cpu migration (called while in a critical section so we
1553 * do not have to enter another one). The thread has already been moved to
1554 * our cpu's allq, but we must wait for the thread to be completely switched
1555 * out on the originating cpu before we schedule it on ours or the stack
1556 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1557 * change to main memory.
1558 *
1559 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1560 * against wakeups. It is best if this interface is used only when there
1561 * are no pending events that might try to schedule the thread.
1562 */
3d28ff59 1563#ifdef SMP
5d21b981
MD
1564static void
1565lwkt_setcpu_remote(void *arg)
1566{
1567 thread_t td = arg;
1568 globaldata_t gd = mycpu;
1569
df910c23
MD
1570 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1571#ifdef SMP
1572 lwkt_process_ipiq();
1573#endif
35238fa5 1574 cpu_lfence();
562273ea 1575 cpu_pause();
df910c23 1576 }
5d21b981 1577 td->td_gd = gd;
562273ea 1578 cpu_mfence();
5d21b981 1579 td->td_flags &= ~TDF_MIGRATING;
9388413d 1580 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1581 _lwkt_enqueue(td);
1582}
3d28ff59 1583#endif
5d21b981 1584
553ea3c8 1585struct lwp *
4b5f931b
MD
1586lwkt_preempted_proc(void)
1587{
73e4f7b9 1588 thread_t td = curthread;
4b5f931b
MD
1589 while (td->td_preempted)
1590 td = td->td_preempted;
553ea3c8 1591 return(td->td_lwp);
4b5f931b
MD
1592}
1593
99df837e
MD
1594/*
1595 * Create a kernel process/thread/whatever. It shares it's address space
1596 * with proc0 - ie: kernel only.
1597 *
365fa13f
MD
1598 * NOTE! By default new threads are created with the MP lock held. A
1599 * thread which does not require the MP lock should release it by calling
1600 * rel_mplock() at the start of the new thread.
99df837e
MD
1601 */
1602int
c9e9fb21
MD
1603lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1604 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1605{
73e4f7b9 1606 thread_t td;
e2565a42 1607 __va_list ap;
99df837e 1608
d3d32139 1609 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1610 tdflags);
a2a5ad0d
MD
1611 if (tdp)
1612 *tdp = td;
709799ea 1613 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1614
1615 /*
1616 * Set up arg0 for 'ps' etc
1617 */
e2565a42 1618 __va_start(ap, fmt);
379210cb 1619 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1620 __va_end(ap);
99df837e
MD
1621
1622 /*
1623 * Schedule the thread to run
1624 */
ef0fdad1
MD
1625 if ((td->td_flags & TDF_STOPREQ) == 0)
1626 lwkt_schedule(td);
1627 else
1628 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1629 return 0;
1630}
1631
1632/*
1633 * Destroy an LWKT thread. Warning! This function is not called when
1634 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1635 * uses a different reaping mechanism.
1636 */
1637void
1638lwkt_exit(void)
1639{
1640 thread_t td = curthread;
c070746a 1641 thread_t std;
8826f33a 1642 globaldata_t gd;
99df837e 1643
2883d2d8
MD
1644 /*
1645 * Do any cleanup that might block here
1646 */
99df837e 1647 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1648 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1649 caps_exit(td);
2883d2d8
MD
1650 biosched_done(td);
1651 dsched_exit_thread(td);
c070746a
MD
1652
1653 /*
1654 * Get us into a critical section to interlock gd_freetd and loop
1655 * until we can get it freed.
1656 *
1657 * We have to cache the current td in gd_freetd because objcache_put()ing
1658 * it would rip it out from under us while our thread is still active.
1659 */
1660 gd = mycpu;
37af14fe 1661 crit_enter_quick(td);
c070746a 1662 while ((std = gd->gd_freetd) != NULL) {
cf709dd2 1663 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
c070746a
MD
1664 gd->gd_freetd = NULL;
1665 objcache_put(thread_cache, std);
1666 }
3b4192fb
MD
1667
1668 /*
1669 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1670 * the last time. We cannot block after this point or we may end
1671 * up with a stale td on the tsleepq.
3b4192fb
MD
1672 */
1673 if (td->td_flags & TDF_TSLEEPQ)
1674 tsleep_remove(td);
37af14fe 1675 lwkt_deschedule_self(td);
e56e4dea 1676 lwkt_remove_tdallq(td);
2883d2d8
MD
1677
1678 /*
1679 * Final cleanup
1680 */
1681 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1682 if (td->td_flags & TDF_ALLOCATED_THREAD)
1683 gd->gd_freetd = td;
99df837e
MD
1684 cpu_thread_exit();
1685}
1686
e56e4dea
MD
1687void
1688lwkt_remove_tdallq(thread_t td)
1689{
1690 KKASSERT(td->td_gd == mycpu);
1691 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1692}
1693
9cf43f91
MD
1694/*
1695 * Code reduction and branch prediction improvements. Call/return
1696 * overhead on modern cpus often degenerates into 0 cycles due to
1697 * the cpu's branch prediction hardware and return pc cache. We
1698 * can take advantage of this by not inlining medium-complexity
1699 * functions and we can also reduce the branch prediction impact
1700 * by collapsing perfectly predictable branches into a single
1701 * procedure instead of duplicating it.
1702 *
1703 * Is any of this noticeable? Probably not, so I'll take the
1704 * smaller code size.
1705 */
1706void
b6468f56 1707crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1708{
b6468f56 1709 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1710}
1711
2d93b37a
MD
1712void
1713crit_panic(void)
1714{
1715 thread_t td = curthread;
850634cc 1716 int lcrit = td->td_critcount;
2d93b37a 1717
850634cc
AH
1718 td->td_critcount = 0;
1719 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1720 /* NOT REACHED */
2d93b37a
MD
1721}
1722
d165e668
MD
1723#ifdef SMP
1724
bd8015ca
MD
1725/*
1726 * Called from debugger/panic on cpus which have been stopped. We must still
1727 * process the IPIQ while stopped, even if we were stopped while in a critical
1728 * section (XXX).
1729 *
1730 * If we are dumping also try to process any pending interrupts. This may
1731 * or may not work depending on the state of the cpu at the point it was
1732 * stopped.
1733 */
1734void
1735lwkt_smp_stopped(void)
1736{
1737 globaldata_t gd = mycpu;
1738
1739 crit_enter_gd(gd);
1740 if (dumping) {
1741 lwkt_process_ipiq();
1742 splz();
1743 } else {
1744 lwkt_process_ipiq();
1745 }
1746 crit_exit_gd(gd);
1747}
1748
d165e668 1749#endif