kernel - Remove CACHE_*MPLOCK* macros & sysctl
[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;
8ad65e08 501
46a3f46d 502 /*
27e88a6e
MD
503 * Switching from within a 'fast' (non thread switched) interrupt or IPI
504 * is illegal. However, we may have to do it anyway if we hit a fatal
505 * kernel trap or we have paniced.
506 *
507 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 508 */
27e88a6e
MD
509 if (gd->gd_intr_nesting_level) {
510 int savegdnest;
511 int savegdtrap;
512
5fddbda2 513 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22
MD
514 panic("lwkt_switch: Attempt to switch from a "
515 "a fast interrupt, ipi, or hard code section, "
516 "td %p\n",
517 td);
27e88a6e
MD
518 } else {
519 savegdnest = gd->gd_intr_nesting_level;
520 savegdtrap = gd->gd_trap_nesting_level;
521 gd->gd_intr_nesting_level = 0;
522 gd->gd_trap_nesting_level = 0;
a7422615
MD
523 if ((td->td_flags & TDF_PANICWARN) == 0) {
524 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
525 kprintf("Warning: thread switch from interrupt, IPI, "
526 "or hard code section.\n"
a7422615 527 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 528 print_backtrace(-1);
a7422615 529 }
27e88a6e
MD
530 lwkt_switch();
531 gd->gd_intr_nesting_level = savegdnest;
532 gd->gd_trap_nesting_level = savegdtrap;
533 return;
534 }
96728c05 535 }
ef0fdad1 536
cb973d15
MD
537 /*
538 * Passive release (used to transition from user to kernel mode
539 * when we block or switch rather then when we enter the kernel).
540 * This function is NOT called if we are switching into a preemption
541 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
542 * our current process designation (if we have one) and become a true
543 * LWKT thread, and may also hand the current process designation to
544 * another process and schedule thread.
cb973d15
MD
545 */
546 if (td->td_release)
547 td->td_release(td);
548
37af14fe 549 crit_enter_gd(gd);
3b998fa9 550 if (TD_TOKS_HELD(td))
9d265729
MD
551 lwkt_relalltokens(td);
552
553 /*
b02926de
MD
554 * We had better not be holding any spin locks, but don't get into an
555 * endless panic loop.
9d265729 556 */
d666840a
MD
557 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
558 ("lwkt_switch: still holding %d exclusive spinlocks!",
559 gd->gd_spinlocks_wr));
9d265729 560
8a8d5d85
MD
561
562#ifdef SMP
563 /*
3933a3ab
MD
564 * td_mpcount + td_xpcount cannot be used to determine if we currently
565 * hold the MP lock because get_mplock() will increment it prior to
566 * attempting to get the lock, and switch out if it can't. Our
567 * ownership of the actual lock will remain stable while we are
568 * in a critical section, and once we actually acquire the underlying
569 * lock as long as the count is greater than 0.
8a8d5d85 570 */
c5724852 571 mpheld = MP_LOCK_HELD(gd);
0f7a3396
MD
572#ifdef INVARIANTS
573 if (td->td_cscount) {
6ea70f76 574 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
575 td);
576 if (panic_on_cscount)
577 panic("switching while mastering cpusync");
578 }
579#endif
8a8d5d85 580#endif
f9235b6d
MD
581
582 /*
583 * If we had preempted another thread on this cpu, resume the preempted
584 * thread. This occurs transparently, whether the preempted thread
585 * was scheduled or not (it may have been preempted after descheduling
586 * itself).
587 *
588 * We have to setup the MP lock for the original thread after backing
589 * out the adjustment that was made to curthread when the original
590 * was preempted.
591 */
99df837e 592 if ((ntd = td->td_preempted) != NULL) {
26a0694b 593 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 594#ifdef SMP
3933a3ab 595 if (ntd->td_mpcount + ntd->td_xpcount && mpheld == 0) {
fc92d4aa 596 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
3933a3ab 597 td, ntd, td->td_mpcount, ntd->td_mpcount + ntd->td_xpcount);
8a8d5d85 598 }
3933a3ab 599 td->td_xpcount = 0;
8a8d5d85 600#endif
26a0694b 601 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
602
603 /*
b9eb1c19
MD
604 * The interrupt may have woken a thread up, we need to properly
605 * set the reschedule flag if the originally interrupted thread is
606 * at a lower priority.
8ec60c3f 607 */
f9235b6d
MD
608 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
609 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
8ec60c3f 610 need_lwkt_resched();
f9235b6d 611 }
8a8d5d85 612 /* YYY release mp lock on switchback if original doesn't need it */
f9235b6d
MD
613 goto havethread_preempted;
614 }
615
616 /*
617 * Implement round-robin fairq with priority insertion. The priority
618 * insertion is handled by _lwkt_enqueue()
619 *
620 * We have to adjust the MP lock for the target thread. If we
621 * need the MP lock and cannot obtain it we try to locate a
622 * thread that does not need the MP lock. If we cannot, we spin
623 * instead of HLT.
624 *
625 * A similar issue exists for the tokens held by the target thread.
626 * If we cannot obtain ownership of the tokens we cannot immediately
627 * schedule the thread.
628 */
629 for (;;) {
630 clear_lwkt_resched();
631 didaccumulate = 0;
632 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
633
4b5f931b 634 /*
f9235b6d 635 * Hotpath if we can get all necessary resources.
41a01a4d 636 *
f9235b6d 637 * If nothing is runnable switch to the idle thread
41a01a4d 638 */
f9235b6d
MD
639 if (ntd == NULL) {
640 ntd = &gd->gd_idlethread;
641 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
642 ntd->td_flags |= TDF_IDLE_NOHLT;
6f207a2c 643#ifdef SMP
3933a3ab 644 KKASSERT(ntd->td_xpcount == 0);
f9235b6d
MD
645 if (ntd->td_mpcount) {
646 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
647 panic("Idle thread %p was holding the BGL!", ntd);
648 if (mpheld == 0) {
c5724852
MD
649 set_cpu_contention_mask(gd);
650 handle_cpu_contention_mask();
651 cpu_try_mplock();
652 mpheld = MP_LOCK_HELD(gd);
f9235b6d
MD
653 cpu_pause();
654 continue;
655 }
656 }
c5724852 657 clr_cpu_contention_mask(gd);
6f207a2c 658#endif
b37f18d6
MD
659 cpu_time.cp_msg[0] = 0;
660 cpu_time.cp_stallpc = 0;
f9235b6d
MD
661 goto haveidle;
662 }
41a01a4d 663
8ec60c3f 664 /*
f9235b6d 665 * Hotpath schedule
6f207a2c
MD
666 *
667 * NOTE: For UP there is no mplock and lwkt_getalltokens()
668 * always succeeds.
8ec60c3f 669 */
f9235b6d
MD
670 if (ntd->td_fairq_accum >= 0 &&
671#ifdef SMP
3933a3ab 672 (ntd->td_mpcount + ntd->td_xpcount == 0 ||
2a9d4663 673 mpheld || cpu_try_mplock_msg(&ntd->td_wmesg)) &&
f9235b6d 674#endif
2a9d4663
MD
675 (!TD_TOKS_HELD(ntd) ||
676 lwkt_getalltokens(ntd))
f9235b6d 677 ) {
8a8d5d85 678#ifdef SMP
c5724852 679 clr_cpu_contention_mask(gd);
f9235b6d
MD
680#endif
681 goto havethread;
682 }
683
684#ifdef SMP
c5724852
MD
685 if (ntd->td_fairq_accum >= 0)
686 set_cpu_contention_mask(gd);
f9235b6d 687 /* Reload mpheld (it become stale after mplock/token ops) */
c5724852 688 mpheld = MP_LOCK_HELD(gd);
f9235b6d
MD
689#endif
690
691 /*
692 * Coldpath - unable to schedule ntd, continue looking for threads
693 * to schedule. This is only allowed of the (presumably) kernel
694 * thread exhausted its fair share. A kernel thread stuck on
695 * resources does not currently allow a user thread to get in
696 * front of it.
697 */
698#ifdef SMP
699 nquserok = ((ntd->td_pri < TDPRI_KERN_LPSCHED) ||
700 (ntd->td_fairq_accum < 0));
6f207a2c
MD
701#else
702 nquserok = 1;
f9235b6d
MD
703#endif
704 nlast = NULL;
705
706 for (;;) {
41a01a4d 707 /*
f9235b6d
MD
708 * If the fair-share scheduler ran out ntd gets moved to the
709 * end and its accumulator will be bumped, if it didn't we
710 * maintain the same queue position.
df6b8ba0 711 *
f9235b6d 712 * nlast keeps track of the last element prior to any moves.
41a01a4d 713 */
f9235b6d 714 if (ntd->td_fairq_accum < 0) {
f9235b6d
MD
715 lwkt_fairq_accumulate(gd, ntd);
716 didaccumulate = 1;
c5724852
MD
717
718 /*
719 * Move to end
720 */
721 xtd = TAILQ_NEXT(ntd, td_threadq);
f9235b6d
MD
722 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
723 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
c5724852
MD
724
725 /*
726 * Set terminal element (nlast)
727 */
f9235b6d
MD
728 if (nlast == NULL) {
729 nlast = ntd;
730 if (xtd == NULL)
731 xtd = ntd;
732 }
733 ntd = xtd;
734 } else {
735 ntd = TAILQ_NEXT(ntd, td_threadq);
736 }
a453459d 737
f9235b6d
MD
738 /*
739 * If we exhausted the run list switch to the idle thread.
740 * Since one or more threads had resource acquisition issues
741 * we do not allow the idle thread to halt.
742 *
743 * NOTE: nlast can be NULL.
744 */
745 if (ntd == nlast) {
e0a90d3b 746 cpu_pause();
f9235b6d
MD
747 ntd = &gd->gd_idlethread;
748 ntd->td_flags |= TDF_IDLE_NOHLT;
6f207a2c 749#ifdef SMP
3933a3ab 750 KKASSERT(ntd->td_xpcount == 0);
f9235b6d 751 if (ntd->td_mpcount) {
c5724852 752 mpheld = MP_LOCK_HELD(gd);
f9235b6d
MD
753 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
754 panic("Idle thread %p was holding the BGL!", ntd);
755 if (mpheld == 0) {
c5724852
MD
756 set_cpu_contention_mask(gd);
757 handle_cpu_contention_mask();
758 cpu_try_mplock();
759 mpheld = MP_LOCK_HELD(gd);
f9235b6d
MD
760 cpu_pause();
761 break; /* try again from the top, almost */
b9eb1c19 762 }
8a8d5d85 763 }
6f207a2c 764#endif
684a93c4
MD
765
766 /*
f9235b6d
MD
767 * If fairq accumulations occured we do not schedule the
768 * idle thread. This will cause us to try again from
769 * the (almost) top.
684a93c4 770 */
f9235b6d 771 if (didaccumulate)
b37f18d6 772 break; /* try again from the top, almost */
f9235b6d 773 goto haveidle;
8a8d5d85 774 }
f9235b6d 775
df6b8ba0 776 /*
f9235b6d 777 * Try to switch to this thread.
6f207a2c
MD
778 *
779 * NOTE: For UP there is no mplock and lwkt_getalltokens()
780 * always succeeds.
df6b8ba0 781 */
77912481
MD
782 if ((ntd->td_pri >= TDPRI_KERN_LPSCHED || nquserok ||
783 user_pri_sched) && ntd->td_fairq_accum >= 0 &&
f9235b6d 784#ifdef SMP
3933a3ab 785 (ntd->td_mpcount + ntd->td_xpcount == 0 ||
2a9d4663 786 mpheld || cpu_try_mplock_msg(&ntd->td_wmesg)) &&
8a8d5d85 787#endif
2a9d4663 788 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd))
f9235b6d 789 ) {
a453459d 790#ifdef SMP
c5724852 791 clr_cpu_contention_mask(gd);
f9235b6d
MD
792#endif
793 goto havethread;
df6b8ba0 794 }
9ac1ee6e
MD
795
796 /*
797 * Thread was runnable but we were unable to get the required
798 * resources (tokens and/or mplock).
799 */
f9235b6d 800#ifdef SMP
c5724852
MD
801 if (ntd->td_fairq_accum >= 0)
802 set_cpu_contention_mask(gd);
803 /*
804 * Reload mpheld (it become stale after mplock/token ops).
805 */
806 mpheld = MP_LOCK_HELD(gd);
f9235b6d
MD
807 if (ntd->td_pri >= TDPRI_KERN_LPSCHED && ntd->td_fairq_accum >= 0)
808 nquserok = 0;
a453459d 809#endif
4b5f931b 810 }
c5724852
MD
811
812 /*
813 * All threads exhausted but we can loop due to a negative
814 * accumulator.
815 *
816 * While we are looping in the scheduler be sure to service
817 * any interrupts which were made pending due to our critical
818 * section, otherwise we could livelock (e.g.) IPIs.
819 *
820 * NOTE: splz can enter and exit the mplock so mpheld is
821 * stale after this call.
822 */
823 splz_check();
824
825#ifdef SMP
826 /*
827 * Our mplock can be cached and cause other cpus to livelock
828 * if we loop due to e.g. not being able to acquire tokens.
829 */
830 if (MP_LOCK_HELD(gd))
831 cpu_rel_mplock(gd->gd_cpuid);
832 mpheld = 0;
833#endif
f1d1c3fa 834 }
8a8d5d85
MD
835
836 /*
f9235b6d
MD
837 * Do the actual switch. WARNING: mpheld is stale here.
838 *
839 * We must always decrement td_fairq_accum on non-idle threads just
840 * in case a thread never gets a tick due to being in a continuous
841 * critical section. The page-zeroing code does that.
842 *
843 * If the thread we came up with is a higher or equal priority verses
844 * the thread at the head of the queue we move our thread to the
845 * front. This way we can always check the front of the queue.
846 */
847havethread:
848 ++gd->gd_cnt.v_swtch;
849 --ntd->td_fairq_accum;
9ac1ee6e 850 ntd->td_wmesg = NULL;
f9235b6d
MD
851 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
852 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
853 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
854 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
855 }
856havethread_preempted:
857 ;
858 /*
859 * If the new target does not need the MP lock and we are holding it,
860 * release the MP lock. If the new target requires the MP lock we have
861 * already acquired it for the target.
862 *
863 * WARNING: mpheld is stale here.
8a8d5d85 864 */
f9235b6d
MD
865haveidle:
866 KASSERT(ntd->td_critcount,
867 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85 868#ifdef SMP
3933a3ab 869 if (ntd->td_mpcount + ntd->td_xpcount == 0 ) {
c5724852
MD
870 if (MP_LOCK_HELD(gd))
871 cpu_rel_mplock(gd->gd_cpuid);
8a8d5d85 872 } else {
a453459d 873 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
874 }
875#endif
94f6d86e
MD
876 if (td != ntd) {
877 ++switch_count;
a1f0fb66 878 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
f1d1c3fa 879 td->td_switch(ntd);
94f6d86e 880 }
37af14fe
MD
881 /* NOTE: current cpu may have changed after switch */
882 crit_exit_quick(td);
8ad65e08
MD
883}
884
b68b7282 885/*
96728c05
MD
886 * Request that the target thread preempt the current thread. Preemption
887 * only works under a specific set of conditions:
b68b7282 888 *
96728c05
MD
889 * - We are not preempting ourselves
890 * - The target thread is owned by the current cpu
891 * - We are not currently being preempted
892 * - The target is not currently being preempted
d3d1cbc8
MD
893 * - We are not holding any spin locks
894 * - The target thread is not holding any tokens
96728c05
MD
895 * - We are able to satisfy the target's MP lock requirements (if any).
896 *
897 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
898 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 899 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
900 * to determine whether preemption is possible (aka usually just the crit
901 * priority of lwkt_schedule() itself).
b68b7282 902 *
26a0694b
MD
903 * XXX at the moment we run the target thread in a critical section during
904 * the preemption in order to prevent the target from taking interrupts
905 * that *WE* can't. Preemption is strictly limited to interrupt threads
906 * and interrupt-like threads, outside of a critical section, and the
907 * preempted source thread will be resumed the instant the target blocks
908 * whether or not the source is scheduled (i.e. preemption is supposed to
909 * be as transparent as possible).
4b5f931b 910 *
8a8d5d85
MD
911 * The target thread inherits our MP count (added to its own) for the
912 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
913 * MP lock during the preemption. Therefore, any preempting targets must be
914 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
915 * out of sync with the physical mp_lock, but we do not have to preserve
916 * the original ownership of the lock if it was out of synch (that is, we
917 * can leave it synchronized on return).
b68b7282
MD
918 */
919void
f9235b6d 920lwkt_preempt(thread_t ntd, int critcount)
b68b7282 921{
46a3f46d 922 struct globaldata *gd = mycpu;
0a3f9b47 923 thread_t td;
8a8d5d85
MD
924#ifdef SMP
925 int mpheld;
57c254db 926 int savecnt;
8a8d5d85 927#endif
2d910aaf 928 int save_gd_intr_nesting_level;
b68b7282 929
26a0694b 930 /*
96728c05
MD
931 * The caller has put us in a critical section. We can only preempt
932 * if the caller of the caller was not in a critical section (basically
f9235b6d 933 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 934 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 935 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
936 *
937 * YYY The target thread must be in a critical section (else it must
938 * inherit our critical section? I dunno yet).
41a01a4d 939 *
0a3f9b47 940 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b 941 */
f9235b6d 942 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 943
fbc024e4
MD
944 if (preempt_enable == 0) {
945 ++preempt_miss;
946 return;
947 }
948
0a3f9b47 949 td = gd->gd_curthread;
f9235b6d 950 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
951 ++preempt_miss;
952 return;
953 }
f9235b6d 954 if (td->td_critcount > critcount) {
96728c05 955 ++preempt_miss;
8ec60c3f 956 need_lwkt_resched();
96728c05
MD
957 return;
958 }
959#ifdef SMP
46a3f46d 960 if (ntd->td_gd != gd) {
96728c05 961 ++preempt_miss;
8ec60c3f 962 need_lwkt_resched();
96728c05
MD
963 return;
964 }
965#endif
41a01a4d 966 /*
77912481
MD
967 * We don't have to check spinlocks here as they will also bump
968 * td_critcount.
d3d1cbc8
MD
969 *
970 * Do not try to preempt if the target thread is holding any tokens.
971 * We could try to acquire the tokens but this case is so rare there
972 * is no need to support it.
41a01a4d 973 */
77912481
MD
974 KKASSERT(gd->gd_spinlocks_wr == 0);
975
3b998fa9 976 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8
MD
977 ++preempt_miss;
978 need_lwkt_resched();
979 return;
980 }
26a0694b
MD
981 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
982 ++preempt_weird;
8ec60c3f 983 need_lwkt_resched();
26a0694b
MD
984 return;
985 }
986 if (ntd->td_preempted) {
4b5f931b 987 ++preempt_hit;
8ec60c3f 988 need_lwkt_resched();
26a0694b 989 return;
b68b7282 990 }
8a8d5d85 991#ifdef SMP
a2a5ad0d 992 /*
3933a3ab 993 * NOTE: An interrupt might have occured just as we were transitioning
71ef2f5c 994 * to or from the MP lock. In this case td_mpcount will be pre-disposed
3933a3ab
MD
995 * (non-zero) but not actually synchronized with the mp_lock itself.
996 * We can use it to imply an MP lock requirement for the preemption but
997 * we cannot use it to test whether we hold the MP lock or not.
a2a5ad0d 998 */
96728c05 999 savecnt = td->td_mpcount;
c5724852 1000 mpheld = MP_LOCK_HELD(gd);
6d9b99db 1001 ntd->td_xpcount = td->td_mpcount + td->td_xpcount;
3933a3ab
MD
1002 if (mpheld == 0 && ntd->td_mpcount + ntd->td_xpcount && !cpu_try_mplock()) {
1003 ntd->td_xpcount = 0;
8a8d5d85 1004 ++preempt_miss;
8ec60c3f 1005 need_lwkt_resched();
8a8d5d85
MD
1006 return;
1007 }
1008#endif
26a0694b 1009
8ec60c3f
MD
1010 /*
1011 * Since we are able to preempt the current thread, there is no need to
1012 * call need_lwkt_resched().
2d910aaf
MD
1013 *
1014 * We must temporarily clear gd_intr_nesting_level around the switch
1015 * since switchouts from the target thread are allowed (they will just
1016 * return to our thread), and since the target thread has its own stack.
8ec60c3f 1017 */
26a0694b
MD
1018 ++preempt_hit;
1019 ntd->td_preempted = td;
1020 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1021 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1022 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1023 gd->gd_intr_nesting_level = 0;
26a0694b 1024 td->td_switch(ntd);
2d910aaf 1025 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1026
26a0694b 1027 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
1028#ifdef SMP
1029 KKASSERT(savecnt == td->td_mpcount);
c5724852 1030 mpheld = MP_LOCK_HELD(gd);
71ef2f5c 1031 if (mpheld && td->td_mpcount == 0)
c5724852 1032 cpu_rel_mplock(gd->gd_cpuid);
3933a3ab 1033 else if (mpheld == 0 && td->td_mpcount + td->td_xpcount)
96728c05
MD
1034 panic("lwkt_preempt(): MP lock was not held through");
1035#endif
26a0694b
MD
1036 ntd->td_preempted = NULL;
1037 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1038}
1039
f1d1c3fa 1040/*
faaeffac 1041 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1042 * This will work inside a critical section but not inside a hard code
1043 * section.
ef0fdad1 1044 *
f1d1c3fa
MD
1045 * (self contained on a per cpu basis)
1046 */
1047void
faaeffac 1048splz_check(void)
f1d1c3fa 1049{
7966cb69
MD
1050 globaldata_t gd = mycpu;
1051 thread_t td = gd->gd_curthread;
ef0fdad1 1052
4a28fe22
MD
1053 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1054 gd->gd_intr_nesting_level == 0 &&
1055 td->td_nest_count < 2)
1056 {
f1d1c3fa 1057 splz();
4a28fe22
MD
1058 }
1059}
1060
1061/*
1062 * This version is integrated into crit_exit, reqflags has already
1063 * been tested but td_critcount has not.
1064 *
1065 * We only want to execute the splz() on the 1->0 transition of
1066 * critcount and not in a hard code section or if too deeply nested.
1067 */
1068void
1069lwkt_maybe_splz(thread_t td)
1070{
1071 globaldata_t gd = td->td_gd;
1072
1073 if (td->td_critcount == 0 &&
1074 gd->gd_intr_nesting_level == 0 &&
1075 td->td_nest_count < 2)
1076 {
1077 splz();
1078 }
f1d1c3fa
MD
1079}
1080
8ad65e08 1081/*
f9235b6d
MD
1082 * This function is used to negotiate a passive release of the current
1083 * process/lwp designation with the user scheduler, allowing the user
1084 * scheduler to schedule another user thread. The related kernel thread
1085 * (curthread) continues running in the released state.
8ad65e08
MD
1086 */
1087void
f9235b6d 1088lwkt_passive_release(struct thread *td)
8ad65e08 1089{
f9235b6d
MD
1090 struct lwp *lp = td->td_lwp;
1091
1092 td->td_release = NULL;
1093 lwkt_setpri_self(TDPRI_KERN_USER);
1094 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1095}
1096
f9235b6d 1097
3824f392 1098/*
f9235b6d
MD
1099 * This implements a normal yield. This routine is virtually a nop if
1100 * there is nothing to yield to but it will always run any pending interrupts
1101 * if called from a critical section.
1102 *
1103 * This yield is designed for kernel threads without a user context.
1104 *
1105 * (self contained on a per cpu basis)
3824f392
MD
1106 */
1107void
f9235b6d 1108lwkt_yield(void)
3824f392 1109{
f9235b6d
MD
1110 globaldata_t gd = mycpu;
1111 thread_t td = gd->gd_curthread;
1112 thread_t xtd;
3824f392 1113
f9235b6d
MD
1114 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1115 splz();
1116 if (td->td_fairq_accum < 0) {
1117 lwkt_schedule_self(curthread);
1118 lwkt_switch();
1119 } else {
1120 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1121 if (xtd && xtd->td_pri > td->td_pri) {
1122 lwkt_schedule_self(curthread);
1123 lwkt_switch();
1124 }
1125 }
3824f392
MD
1126}
1127
1128/*
f9235b6d
MD
1129 * This yield is designed for kernel threads with a user context.
1130 *
1131 * The kernel acting on behalf of the user is potentially cpu-bound,
1132 * this function will efficiently allow other threads to run and also
1133 * switch to other processes by releasing.
3824f392
MD
1134 *
1135 * The lwkt_user_yield() function is designed to have very low overhead
1136 * if no yield is determined to be needed.
1137 */
1138void
1139lwkt_user_yield(void)
1140{
f9235b6d
MD
1141 globaldata_t gd = mycpu;
1142 thread_t td = gd->gd_curthread;
1143
1144 /*
1145 * Always run any pending interrupts in case we are in a critical
1146 * section.
1147 */
1148 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1149 splz();
3824f392
MD
1150
1151#ifdef SMP
1152 /*
1153 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1154 * kernel can prevent other cpus from servicing interrupt threads
1155 * which still require the MP lock (which is a lot of them). This
1156 * has a chaining effect since if the interrupt is blocked, so is
1157 * the event, so normal scheduling will not pick up on the problem.
1158 */
3933a3ab 1159 if (cpu_contention_mask && td->td_mpcount + td->td_xpcount) {
684a93c4 1160 yield_mplock(td);
3824f392
MD
1161 }
1162#endif
1163
1164 /*
f9235b6d
MD
1165 * Switch (which forces a release) if another kernel thread needs
1166 * the cpu, if userland wants us to resched, or if our kernel
1167 * quantum has run out.
3824f392 1168 */
f9235b6d
MD
1169 if (lwkt_resched_wanted() ||
1170 user_resched_wanted() ||
1171 td->td_fairq_accum < 0)
1172 {
3824f392 1173 lwkt_switch();
3824f392
MD
1174 }
1175
f9235b6d 1176#if 0
3824f392 1177 /*
f9235b6d
MD
1178 * Reacquire the current process if we are released.
1179 *
1180 * XXX not implemented atm. The kernel may be holding locks and such,
1181 * so we want the thread to continue to receive cpu.
3824f392 1182 */
f9235b6d
MD
1183 if (td->td_release == NULL && lp) {
1184 lp->lwp_proc->p_usched->acquire_curproc(lp);
1185 td->td_release = lwkt_passive_release;
1186 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1187 }
f9235b6d 1188#endif
b9eb1c19
MD
1189}
1190
8ad65e08 1191/*
f1d1c3fa
MD
1192 * Generic schedule. Possibly schedule threads belonging to other cpus and
1193 * deal with threads that might be blocked on a wait queue.
1194 *
0a3f9b47
MD
1195 * We have a little helper inline function which does additional work after
1196 * the thread has been enqueued, including dealing with preemption and
1197 * setting need_lwkt_resched() (which prevents the kernel from returning
1198 * to userland until it has processed higher priority threads).
6330a558
MD
1199 *
1200 * It is possible for this routine to be called after a failed _enqueue
1201 * (due to the target thread migrating, sleeping, or otherwise blocked).
1202 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1203 *
1204 * reschedok is an optimized constant propagated from lwkt_schedule() or
1205 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1206 * reschedule to be requested if the target thread has a higher priority.
1207 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1208 * be 0, prevented undesired reschedules.
8ad65e08 1209 */
0a3f9b47
MD
1210static __inline
1211void
f9235b6d 1212_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
0a3f9b47 1213{
b9eb1c19 1214 thread_t otd;
c730be20 1215
6330a558 1216 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1217 if (ntd->td_preemptable && reschedok) {
f9235b6d 1218 ntd->td_preemptable(ntd, ccount); /* YYY +token */
361d01dd 1219 } else if (reschedok) {
b9eb1c19 1220 otd = curthread;
f9235b6d 1221 if (ntd->td_pri > otd->td_pri)
c730be20 1222 need_lwkt_resched();
6330a558 1223 }
f9235b6d
MD
1224
1225 /*
1226 * Give the thread a little fair share scheduler bump if it
1227 * has been asleep for a while. This is primarily to avoid
1228 * a degenerate case for interrupt threads where accumulator
1229 * crosses into negative territory unnecessarily.
1230 */
1231 if (ntd->td_fairq_lticks != ticks) {
1232 ntd->td_fairq_lticks = ticks;
1233 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1234 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1235 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1236 }
0a3f9b47
MD
1237 }
1238}
1239
361d01dd 1240static __inline
8ad65e08 1241void
361d01dd 1242_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1243{
37af14fe
MD
1244 globaldata_t mygd = mycpu;
1245
cf709dd2
MD
1246 KASSERT(td != &td->td_gd->gd_idlethread,
1247 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1248 crit_enter_gd(mygd);
9388413d 1249 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1250 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1251 _lwkt_enqueue(td);
1252 } else {
f1d1c3fa 1253 /*
7cd8d145
MD
1254 * If we own the thread, there is no race (since we are in a
1255 * critical section). If we do not own the thread there might
1256 * be a race but the target cpu will deal with it.
f1d1c3fa 1257 */
0f7a3396 1258#ifdef SMP
7cd8d145 1259 if (td->td_gd == mygd) {
9d265729 1260 _lwkt_enqueue(td);
f9235b6d 1261 _lwkt_schedule_post(mygd, td, 1, reschedok);
f1d1c3fa 1262 } else {
e381e77c 1263 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1264 }
0f7a3396 1265#else
7cd8d145 1266 _lwkt_enqueue(td);
f9235b6d 1267 _lwkt_schedule_post(mygd, td, 1, reschedok);
0f7a3396 1268#endif
8ad65e08 1269 }
37af14fe 1270 crit_exit_gd(mygd);
8ad65e08
MD
1271}
1272
361d01dd
MD
1273void
1274lwkt_schedule(thread_t td)
1275{
1276 _lwkt_schedule(td, 1);
1277}
1278
1279void
1280lwkt_schedule_noresched(thread_t td)
1281{
1282 _lwkt_schedule(td, 0);
1283}
1284
88ebb169
SW
1285#ifdef SMP
1286
e381e77c
MD
1287/*
1288 * When scheduled remotely if frame != NULL the IPIQ is being
1289 * run via doreti or an interrupt then preemption can be allowed.
1290 *
1291 * To allow preemption we have to drop the critical section so only
1292 * one is present in _lwkt_schedule_post.
1293 */
1294static void
1295lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1296{
1297 thread_t td = curthread;
1298 thread_t ntd = arg;
1299
1300 if (frame && ntd->td_preemptable) {
1301 crit_exit_noyield(td);
1302 _lwkt_schedule(ntd, 1);
1303 crit_enter_quick(td);
1304 } else {
1305 _lwkt_schedule(ntd, 1);
1306 }
1307}
1308
d9eea1a5 1309/*
52eedfb5
MD
1310 * Thread migration using a 'Pull' method. The thread may or may not be
1311 * the current thread. It MUST be descheduled and in a stable state.
1312 * lwkt_giveaway() must be called on the cpu owning the thread.
1313 *
1314 * At any point after lwkt_giveaway() is called, the target cpu may
1315 * 'pull' the thread by calling lwkt_acquire().
1316 *
ae8e83e6
MD
1317 * We have to make sure the thread is not sitting on a per-cpu tsleep
1318 * queue or it will blow up when it moves to another cpu.
1319 *
52eedfb5 1320 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1321 */
52eedfb5
MD
1322void
1323lwkt_giveaway(thread_t td)
1324{
3b4192fb 1325 globaldata_t gd = mycpu;
52eedfb5 1326
3b4192fb
MD
1327 crit_enter_gd(gd);
1328 if (td->td_flags & TDF_TSLEEPQ)
1329 tsleep_remove(td);
1330 KKASSERT(td->td_gd == gd);
1331 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1332 td->td_flags |= TDF_MIGRATING;
1333 crit_exit_gd(gd);
52eedfb5
MD
1334}
1335
a2a5ad0d
MD
1336void
1337lwkt_acquire(thread_t td)
1338{
37af14fe
MD
1339 globaldata_t gd;
1340 globaldata_t mygd;
a2a5ad0d 1341
52eedfb5 1342 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1343 gd = td->td_gd;
37af14fe 1344 mygd = mycpu;
52eedfb5 1345 if (gd != mycpu) {
35238fa5 1346 cpu_lfence();
52eedfb5 1347 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1348 crit_enter_gd(mygd);
df910c23
MD
1349 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1350#ifdef SMP
1351 lwkt_process_ipiq();
1352#endif
52eedfb5 1353 cpu_lfence();
df910c23 1354 }
562273ea 1355 cpu_mfence();
37af14fe 1356 td->td_gd = mygd;
52eedfb5
MD
1357 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1358 td->td_flags &= ~TDF_MIGRATING;
1359 crit_exit_gd(mygd);
1360 } else {
1361 crit_enter_gd(mygd);
1362 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1363 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1364 crit_exit_gd(mygd);
a2a5ad0d
MD
1365 }
1366}
1367
52eedfb5
MD
1368#endif
1369
f1d1c3fa
MD
1370/*
1371 * Generic deschedule. Descheduling threads other then your own should be
1372 * done only in carefully controlled circumstances. Descheduling is
1373 * asynchronous.
1374 *
1375 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1376 */
1377void
1378lwkt_deschedule(thread_t td)
1379{
f1d1c3fa 1380 crit_enter();
b8a98473 1381#ifdef SMP
f1d1c3fa
MD
1382 if (td == curthread) {
1383 _lwkt_dequeue(td);
1384 } else {
a72187e9 1385 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1386 _lwkt_dequeue(td);
1387 } else {
b8a98473 1388 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1389 }
1390 }
b8a98473
MD
1391#else
1392 _lwkt_dequeue(td);
1393#endif
f1d1c3fa
MD
1394 crit_exit();
1395}
1396
4b5f931b
MD
1397/*
1398 * Set the target thread's priority. This routine does not automatically
1399 * switch to a higher priority thread, LWKT threads are not designed for
1400 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1401 */
1402void
1403lwkt_setpri(thread_t td, int pri)
1404{
a72187e9 1405 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1406 if (td->td_pri != pri) {
1407 KKASSERT(pri >= 0);
1408 crit_enter();
1409 if (td->td_flags & TDF_RUNQ) {
1410 _lwkt_dequeue(td);
1411 td->td_pri = pri;
1412 _lwkt_enqueue(td);
1413 } else {
1414 td->td_pri = pri;
1415 }
1416 crit_exit();
26a0694b 1417 }
26a0694b
MD
1418}
1419
03bd0a5e
MD
1420/*
1421 * Set the initial priority for a thread prior to it being scheduled for
1422 * the first time. The thread MUST NOT be scheduled before or during
1423 * this call. The thread may be assigned to a cpu other then the current
1424 * cpu.
1425 *
1426 * Typically used after a thread has been created with TDF_STOPPREQ,
1427 * and before the thread is initially scheduled.
1428 */
1429void
1430lwkt_setpri_initial(thread_t td, int pri)
1431{
1432 KKASSERT(pri >= 0);
1433 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1434 td->td_pri = pri;
03bd0a5e
MD
1435}
1436
26a0694b
MD
1437void
1438lwkt_setpri_self(int pri)
1439{
1440 thread_t td = curthread;
1441
4b5f931b
MD
1442 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1443 crit_enter();
1444 if (td->td_flags & TDF_RUNQ) {
1445 _lwkt_dequeue(td);
f9235b6d 1446 td->td_pri = pri;
4b5f931b
MD
1447 _lwkt_enqueue(td);
1448 } else {
f9235b6d 1449 td->td_pri = pri;
4b5f931b
MD
1450 }
1451 crit_exit();
1452}
1453
f9235b6d
MD
1454/*
1455 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1456 *
1457 * Example: two competing threads, same priority N. decrement by (2*N)
1458 * increment by N*8, each thread will get 4 ticks.
1459 */
1460void
1461lwkt_fairq_schedulerclock(thread_t td)
1462{
1463 if (fairq_enable) {
1464 while (td) {
1465 if (td != &td->td_gd->gd_idlethread) {
1466 td->td_fairq_accum -= td->td_gd->gd_fairq_total_pri;
1467 if (td->td_fairq_accum < -TDFAIRQ_MAX(td->td_gd))
1468 td->td_fairq_accum = -TDFAIRQ_MAX(td->td_gd);
1469 if (td->td_fairq_accum < 0)
1470 need_lwkt_resched();
1471 td->td_fairq_lticks = ticks;
1472 }
1473 td = td->td_preempted;
1474 }
1475 }
1476}
1477
1478static void
1479lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1480{
1481 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1482 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1483 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1484}
1485
5d21b981 1486/*
52eedfb5
MD
1487 * Migrate the current thread to the specified cpu.
1488 *
1489 * This is accomplished by descheduling ourselves from the current cpu,
1490 * moving our thread to the tdallq of the target cpu, IPI messaging the
1491 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1492 * races while the thread is being migrated.
ae8e83e6
MD
1493 *
1494 * We must be sure to remove ourselves from the current cpu's tsleepq
1495 * before potentially moving to another queue. The thread can be on
1496 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1497 */
3d28ff59 1498#ifdef SMP
5d21b981 1499static void lwkt_setcpu_remote(void *arg);
3d28ff59 1500#endif
5d21b981
MD
1501
1502void
1503lwkt_setcpu_self(globaldata_t rgd)
1504{
1505#ifdef SMP
1506 thread_t td = curthread;
1507
1508 if (td->td_gd != rgd) {
1509 crit_enter_quick(td);
ae8e83e6 1510 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1511 tsleep_remove(td);
5d21b981
MD
1512 td->td_flags |= TDF_MIGRATING;
1513 lwkt_deschedule_self(td);
52eedfb5 1514 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1515 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1516 lwkt_switch();
1517 /* we are now on the target cpu */
52eedfb5 1518 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1519 crit_exit_quick(td);
1520 }
1521#endif
1522}
1523
ecdefdda
MD
1524void
1525lwkt_migratecpu(int cpuid)
1526{
1527#ifdef SMP
1528 globaldata_t rgd;
1529
1530 rgd = globaldata_find(cpuid);
1531 lwkt_setcpu_self(rgd);
1532#endif
1533}
1534
5d21b981
MD
1535/*
1536 * Remote IPI for cpu migration (called while in a critical section so we
1537 * do not have to enter another one). The thread has already been moved to
1538 * our cpu's allq, but we must wait for the thread to be completely switched
1539 * out on the originating cpu before we schedule it on ours or the stack
1540 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1541 * change to main memory.
1542 *
1543 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1544 * against wakeups. It is best if this interface is used only when there
1545 * are no pending events that might try to schedule the thread.
1546 */
3d28ff59 1547#ifdef SMP
5d21b981
MD
1548static void
1549lwkt_setcpu_remote(void *arg)
1550{
1551 thread_t td = arg;
1552 globaldata_t gd = mycpu;
1553
df910c23
MD
1554 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1555#ifdef SMP
1556 lwkt_process_ipiq();
1557#endif
35238fa5 1558 cpu_lfence();
562273ea 1559 cpu_pause();
df910c23 1560 }
5d21b981 1561 td->td_gd = gd;
562273ea 1562 cpu_mfence();
5d21b981 1563 td->td_flags &= ~TDF_MIGRATING;
9388413d 1564 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1565 _lwkt_enqueue(td);
1566}
3d28ff59 1567#endif
5d21b981 1568
553ea3c8 1569struct lwp *
4b5f931b
MD
1570lwkt_preempted_proc(void)
1571{
73e4f7b9 1572 thread_t td = curthread;
4b5f931b
MD
1573 while (td->td_preempted)
1574 td = td->td_preempted;
553ea3c8 1575 return(td->td_lwp);
4b5f931b
MD
1576}
1577
99df837e
MD
1578/*
1579 * Create a kernel process/thread/whatever. It shares it's address space
1580 * with proc0 - ie: kernel only.
1581 *
365fa13f
MD
1582 * NOTE! By default new threads are created with the MP lock held. A
1583 * thread which does not require the MP lock should release it by calling
1584 * rel_mplock() at the start of the new thread.
99df837e
MD
1585 */
1586int
c9e9fb21
MD
1587lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1588 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1589{
73e4f7b9 1590 thread_t td;
e2565a42 1591 __va_list ap;
99df837e 1592
d3d32139 1593 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1594 tdflags);
a2a5ad0d
MD
1595 if (tdp)
1596 *tdp = td;
709799ea 1597 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1598
1599 /*
1600 * Set up arg0 for 'ps' etc
1601 */
e2565a42 1602 __va_start(ap, fmt);
379210cb 1603 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1604 __va_end(ap);
99df837e
MD
1605
1606 /*
1607 * Schedule the thread to run
1608 */
ef0fdad1
MD
1609 if ((td->td_flags & TDF_STOPREQ) == 0)
1610 lwkt_schedule(td);
1611 else
1612 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1613 return 0;
1614}
1615
1616/*
1617 * Destroy an LWKT thread. Warning! This function is not called when
1618 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1619 * uses a different reaping mechanism.
1620 */
1621void
1622lwkt_exit(void)
1623{
1624 thread_t td = curthread;
c070746a 1625 thread_t std;
8826f33a 1626 globaldata_t gd;
99df837e 1627
2883d2d8
MD
1628 /*
1629 * Do any cleanup that might block here
1630 */
99df837e 1631 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1632 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1633 caps_exit(td);
2883d2d8
MD
1634 biosched_done(td);
1635 dsched_exit_thread(td);
c070746a
MD
1636
1637 /*
1638 * Get us into a critical section to interlock gd_freetd and loop
1639 * until we can get it freed.
1640 *
1641 * We have to cache the current td in gd_freetd because objcache_put()ing
1642 * it would rip it out from under us while our thread is still active.
1643 */
1644 gd = mycpu;
37af14fe 1645 crit_enter_quick(td);
c070746a 1646 while ((std = gd->gd_freetd) != NULL) {
cf709dd2 1647 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
c070746a
MD
1648 gd->gd_freetd = NULL;
1649 objcache_put(thread_cache, std);
1650 }
3b4192fb
MD
1651
1652 /*
1653 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1654 * the last time. We cannot block after this point or we may end
1655 * up with a stale td on the tsleepq.
3b4192fb
MD
1656 */
1657 if (td->td_flags & TDF_TSLEEPQ)
1658 tsleep_remove(td);
37af14fe 1659 lwkt_deschedule_self(td);
e56e4dea 1660 lwkt_remove_tdallq(td);
2883d2d8
MD
1661
1662 /*
1663 * Final cleanup
1664 */
1665 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1666 if (td->td_flags & TDF_ALLOCATED_THREAD)
1667 gd->gd_freetd = td;
99df837e
MD
1668 cpu_thread_exit();
1669}
1670
e56e4dea
MD
1671void
1672lwkt_remove_tdallq(thread_t td)
1673{
1674 KKASSERT(td->td_gd == mycpu);
1675 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1676}
1677
9cf43f91
MD
1678/*
1679 * Code reduction and branch prediction improvements. Call/return
1680 * overhead on modern cpus often degenerates into 0 cycles due to
1681 * the cpu's branch prediction hardware and return pc cache. We
1682 * can take advantage of this by not inlining medium-complexity
1683 * functions and we can also reduce the branch prediction impact
1684 * by collapsing perfectly predictable branches into a single
1685 * procedure instead of duplicating it.
1686 *
1687 * Is any of this noticeable? Probably not, so I'll take the
1688 * smaller code size.
1689 */
1690void
b6468f56 1691crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1692{
b6468f56 1693 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1694}
1695
2d93b37a
MD
1696void
1697crit_panic(void)
1698{
1699 thread_t td = curthread;
850634cc 1700 int lcrit = td->td_critcount;
2d93b37a 1701
850634cc
AH
1702 td->td_critcount = 0;
1703 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1704 /* NOT REACHED */
2d93b37a
MD
1705}
1706
d165e668
MD
1707#ifdef SMP
1708
bd8015ca
MD
1709/*
1710 * Called from debugger/panic on cpus which have been stopped. We must still
1711 * process the IPIQ while stopped, even if we were stopped while in a critical
1712 * section (XXX).
1713 *
1714 * If we are dumping also try to process any pending interrupts. This may
1715 * or may not work depending on the state of the cpu at the point it was
1716 * stopped.
1717 */
1718void
1719lwkt_smp_stopped(void)
1720{
1721 globaldata_t gd = mycpu;
1722
1723 crit_enter_gd(gd);
1724 if (dumping) {
1725 lwkt_process_ipiq();
1726 splz();
1727 } else {
1728 lwkt_process_ipiq();
1729 }
1730 crit_exit_gd(gd);
1731}
1732
d165e668 1733#endif