dhclient - Do not remove default route unless setting a default route
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
3b998fa9 2 * Copyright (c) 2003-2010 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
60f60350 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
60f60350 20 *
8c10bfcf
MD
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
b37f18d6 47#include <sys/kinfo.h>
8ad65e08 48#include <sys/queue.h>
7d0bac62 49#include <sys/sysctl.h>
99df837e 50#include <sys/kthread.h>
f1d1c3fa 51#include <machine/cpu.h>
99df837e 52#include <sys/lock.h>
f6bf3af1 53#include <sys/caps.h>
9d265729 54#include <sys/spinlock.h>
57aa743c 55#include <sys/ktr.h>
9d265729
MD
56
57#include <sys/thread2.h>
58#include <sys/spinlock2.h>
684a93c4 59#include <sys/mplock2.h>
f1d1c3fa 60
8c72e3d5
AH
61#include <sys/dsched.h>
62
7d0bac62
MD
63#include <vm/vm.h>
64#include <vm/vm_param.h>
65#include <vm/vm_kern.h>
66#include <vm/vm_object.h>
67#include <vm/vm_page.h>
68#include <vm/vm_map.h>
69#include <vm/vm_pager.h>
70#include <vm/vm_extern.h>
7d0bac62 71
99df837e 72#include <machine/stdarg.h>
96728c05 73#include <machine/smp.h>
99df837e 74
d850923c
AE
75#if !defined(KTR_CTXSW)
76#define KTR_CTXSW KTR_ALL
77#endif
78KTR_INFO_MASTER(ctxsw);
a1f0fb66
AE
79KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
80 sizeof(int) + sizeof(struct thread *));
81KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
82 sizeof(int) + sizeof(struct thread *));
83KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
84 sizeof (struct thread *) + sizeof(char *));
85KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
1541028a 86
40aaf5fc
NT
87static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
88
0f7a3396
MD
89#ifdef INVARIANTS
90static int panic_on_cscount = 0;
91#endif
05220613
MD
92static __int64_t switch_count = 0;
93static __int64_t preempt_hit = 0;
94static __int64_t preempt_miss = 0;
95static __int64_t preempt_weird = 0;
f64b567c 96static __int64_t token_contention_count __debugvar = 0;
fb0f29c4 97static int lwkt_use_spin_port;
40aaf5fc 98static struct objcache *thread_cache;
05220613 99
88ebb169 100#ifdef SMP
e381e77c 101static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
88ebb169 102#endif
f9235b6d 103static void lwkt_fairq_accumulate(globaldata_t gd, thread_t td);
e381e77c 104
0855a2af
JG
105extern void cpu_heavy_restore(void);
106extern void cpu_lwkt_restore(void);
107extern void cpu_kthread_restore(void);
108extern void cpu_idle_restore(void);
109
b2b3ffcd 110#ifdef __x86_64__
85514115
MD
111
112static int
0855a2af
JG
113jg_tos_ok(struct thread *td)
114{
85514115
MD
115 void *tos;
116 int tos_ok;
117
0855a2af
JG
118 if (td == NULL) {
119 return 1;
120 }
121 KKASSERT(td->td_sp != NULL);
85514115
MD
122 tos = ((void **)td->td_sp)[0];
123 tos_ok = 0;
124 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
125 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
0855a2af
JG
126 tos_ok = 1;
127 }
128 return tos_ok;
129}
130
85514115
MD
131#endif
132
fb0f29c4
MD
133/*
134 * We can make all thread ports use the spin backend instead of the thread
135 * backend. This should only be set to debug the spin backend.
136 */
137TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
138
0f7a3396
MD
139#ifdef INVARIANTS
140SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
141#endif
4b5f931b 142SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
9733f757
VS
143SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
144 "Successful preemption events");
145SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
146 "Failed preemption events");
26a0694b 147SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
148#ifdef INVARIANTS
149SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
150 &token_contention_count, 0, "spinning due to token contention");
38717797 151#endif
f9235b6d
MD
152static int fairq_enable = 1;
153SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, &fairq_enable, 0, "");
77912481
MD
154static int user_pri_sched = 0;
155SYSCTL_INT(_lwkt, OID_AUTO, user_pri_sched, CTLFLAG_RW, &user_pri_sched, 0, "");
fbc024e4
MD
156static int preempt_enable = 1;
157SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, &preempt_enable, 0, "");
158
05220613 159
4b5f931b
MD
160/*
161 * These helper procedures handle the runq, they can only be called from
162 * within a critical section.
75cdbe6c
MD
163 *
164 * WARNING! Prior to SMP being brought up it is possible to enqueue and
165 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
166 * instead of 'mycpu' when referencing the globaldata structure. Once
167 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 168 */
f1d1c3fa
MD
169static __inline
170void
171_lwkt_dequeue(thread_t td)
172{
173 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 174 struct globaldata *gd = td->td_gd;
4b5f931b 175
f1d1c3fa 176 td->td_flags &= ~TDF_RUNQ;
f9235b6d
MD
177 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
178 gd->gd_fairq_total_pri -= td->td_pri;
179 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
180 atomic_clear_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
181 }
182}
183
f9235b6d
MD
184/*
185 * Priority enqueue.
186 *
187 * NOTE: There are a limited number of lwkt threads runnable since user
188 * processes only schedule one at a time per cpu.
189 */
f1d1c3fa
MD
190static __inline
191void
192_lwkt_enqueue(thread_t td)
193{
f9235b6d
MD
194 thread_t xtd;
195
7f5d7ed7 196 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 197 struct globaldata *gd = td->td_gd;
4b5f931b 198
f1d1c3fa 199 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
200 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
201 if (xtd == NULL) {
202 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
203 atomic_set_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
204 } else {
205 while (xtd && xtd->td_pri > td->td_pri)
206 xtd = TAILQ_NEXT(xtd, td_threadq);
207 if (xtd)
208 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
209 else
210 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
211 }
212 gd->gd_fairq_total_pri += td->td_pri;
f1d1c3fa
MD
213 }
214}
8ad65e08 215
40aaf5fc
NT
216static __boolean_t
217_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
218{
219 struct thread *td = (struct thread *)obj;
220
221 td->td_kstack = NULL;
222 td->td_kstack_size = 0;
223 td->td_flags = TDF_ALLOCATED_THREAD;
224 return (1);
225}
226
227static void
228_lwkt_thread_dtor(void *obj, void *privdata)
229{
230 struct thread *td = (struct thread *)obj;
231
232 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
233 ("_lwkt_thread_dtor: not allocated from objcache"));
234 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
235 td->td_kstack_size > 0,
236 ("_lwkt_thread_dtor: corrupted stack"));
237 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
238}
239
240/*
241 * Initialize the lwkt s/system.
242 */
243void
244lwkt_init(void)
245{
246 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
247 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
248 NULL, CACHE_NTHREADS/2,
249 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
250}
251
37af14fe
MD
252/*
253 * Schedule a thread to run. As the current thread we can always safely
254 * schedule ourselves, and a shortcut procedure is provided for that
255 * function.
256 *
257 * (non-blocking, self contained on a per cpu basis)
258 */
259void
260lwkt_schedule_self(thread_t td)
261{
262 crit_enter_quick(td);
f9235b6d
MD
263 KASSERT(td != &td->td_gd->gd_idlethread,
264 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 265 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 266 _lwkt_enqueue(td);
37af14fe
MD
267 crit_exit_quick(td);
268}
269
270/*
271 * Deschedule a thread.
272 *
273 * (non-blocking, self contained on a per cpu basis)
274 */
275void
276lwkt_deschedule_self(thread_t td)
277{
278 crit_enter_quick(td);
37af14fe
MD
279 _lwkt_dequeue(td);
280 crit_exit_quick(td);
281}
282
8ad65e08
MD
283/*
284 * LWKTs operate on a per-cpu basis
285 *
73e4f7b9 286 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
287 */
288void
289lwkt_gdinit(struct globaldata *gd)
290{
f9235b6d 291 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 292 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
293}
294
7d0bac62
MD
295/*
296 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
297 * or LWKT start address before it can be scheduled. If the target cpu is
298 * -1 the thread will be created on the current cpu.
0cfcada1
MD
299 *
300 * If you intend to create a thread without a process context this function
301 * does everything except load the startup and switcher function.
7d0bac62
MD
302 */
303thread_t
d3d32139 304lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 305{
c070746a 306 globaldata_t gd = mycpu;
99df837e 307 void *stack;
7d0bac62 308
c070746a
MD
309 /*
310 * If static thread storage is not supplied allocate a thread. Reuse
311 * a cached free thread if possible. gd_freetd is used to keep an exiting
312 * thread intact through the exit.
313 */
ef0fdad1 314 if (td == NULL) {
c070746a
MD
315 if ((td = gd->gd_freetd) != NULL)
316 gd->gd_freetd = NULL;
317 else
318 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
319 KASSERT((td->td_flags &
320 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
321 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
322 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 323 }
c070746a
MD
324
325 /*
326 * Try to reuse cached stack.
327 */
f470d0c8
MD
328 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
329 if (flags & TDF_ALLOCATED_STACK) {
e4846942 330 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
331 stack = NULL;
332 }
333 }
334 if (stack == NULL) {
e4846942 335 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 336 flags |= TDF_ALLOCATED_STACK;
99df837e 337 }
75cdbe6c 338 if (cpu < 0)
c070746a 339 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 340 else
f470d0c8 341 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 342 return(td);
7d0bac62
MD
343}
344
345/*
346 * Initialize a preexisting thread structure. This function is used by
347 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
348 *
f8c3996b
MD
349 * All threads start out in a critical section at a priority of
350 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
351 * appropriate. This function may send an IPI message when the
352 * requested cpu is not the current cpu and consequently gd_tdallq may
353 * not be initialized synchronously from the point of view of the originating
354 * cpu.
355 *
356 * NOTE! we have to be careful in regards to creating threads for other cpus
357 * if SMP has not yet been activated.
7d0bac62 358 */
41a01a4d
MD
359#ifdef SMP
360
75cdbe6c
MD
361static void
362lwkt_init_thread_remote(void *arg)
363{
364 thread_t td = arg;
365
52eedfb5
MD
366 /*
367 * Protected by critical section held by IPI dispatch
368 */
75cdbe6c
MD
369 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
370}
371
41a01a4d
MD
372#endif
373
fdce8919
MD
374/*
375 * lwkt core thread structural initialization.
376 *
377 * NOTE: All threads are initialized as mpsafe threads.
378 */
7d0bac62 379void
f470d0c8
MD
380lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
381 struct globaldata *gd)
7d0bac62 382{
37af14fe
MD
383 globaldata_t mygd = mycpu;
384
99df837e
MD
385 bzero(td, sizeof(struct thread));
386 td->td_kstack = stack;
f470d0c8 387 td->td_kstack_size = stksize;
d3d32139 388 td->td_flags = flags;
26a0694b 389 td->td_gd = gd;
f9235b6d
MD
390 td->td_pri = TDPRI_KERN_DAEMON;
391 td->td_critcount = 1;
3b998fa9 392 td->td_toks_stop = &td->td_toks_base;
fb0f29c4
MD
393 if (lwkt_use_spin_port)
394 lwkt_initport_spin(&td->td_msgport);
395 else
396 lwkt_initport_thread(&td->td_msgport, td);
99df837e 397 pmap_init_thread(td);
0f7a3396 398#ifdef SMP
5d21b981
MD
399 /*
400 * Normally initializing a thread for a remote cpu requires sending an
401 * IPI. However, the idlethread is setup before the other cpus are
402 * activated so we have to treat it as a special case. XXX manipulation
403 * of gd_tdallq requires the BGL.
404 */
405 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 406 crit_enter_gd(mygd);
75cdbe6c 407 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 408 crit_exit_gd(mygd);
75cdbe6c 409 } else {
2db3b277 410 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 411 }
0f7a3396 412#else
37af14fe 413 crit_enter_gd(mygd);
0f7a3396 414 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 415 crit_exit_gd(mygd);
0f7a3396 416#endif
8c72e3d5
AH
417
418 dsched_new_thread(td);
73e4f7b9
MD
419}
420
421void
422lwkt_set_comm(thread_t td, const char *ctl, ...)
423{
e2565a42 424 __va_list va;
73e4f7b9 425
e2565a42 426 __va_start(va, ctl);
379210cb 427 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 428 __va_end(va);
e7c0dbba 429 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
430}
431
99df837e 432void
73e4f7b9 433lwkt_hold(thread_t td)
99df837e 434{
73e4f7b9
MD
435 ++td->td_refs;
436}
437
438void
439lwkt_rele(thread_t td)
440{
441 KKASSERT(td->td_refs > 0);
442 --td->td_refs;
443}
444
445void
446lwkt_wait_free(thread_t td)
447{
448 while (td->td_refs)
377d4740 449 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
450}
451
452void
453lwkt_free_thread(thread_t td)
454{
d9eea1a5 455 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
456 ("lwkt_free_thread: did not exit! %p", td));
457
40aaf5fc
NT
458 if (td->td_flags & TDF_ALLOCATED_THREAD) {
459 objcache_put(thread_cache, td);
460 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
461 /* client-allocated struct with internally allocated stack */
462 KASSERT(td->td_kstack && td->td_kstack_size > 0,
463 ("lwkt_free_thread: corrupted stack"));
464 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
465 td->td_kstack = NULL;
466 td->td_kstack_size = 0;
99df837e 467 }
e7c0dbba 468 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
469}
470
471
8ad65e08
MD
472/*
473 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
474 * switch to the idlethread. Switching must occur within a critical
475 * section to avoid races with the scheduling queue.
476 *
477 * We always have full control over our cpu's run queue. Other cpus
478 * that wish to manipulate our queue must use the cpu_*msg() calls to
479 * talk to our cpu, so a critical section is all that is needed and
480 * the result is very, very fast thread switching.
481 *
96728c05
MD
482 * The LWKT scheduler uses a fixed priority model and round-robins at
483 * each priority level. User process scheduling is a totally
484 * different beast and LWKT priorities should not be confused with
485 * user process priorities.
f1d1c3fa 486 *
3933a3ab
MD
487 * The MP lock may be out of sync with the thread's td_mpcount + td_xpcount.
488 * lwkt_switch() cleans it up.
489 *
490 * Note that the td_switch() function cannot do anything that requires
491 * the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
492 * the target thread (not the current thread). It's nice to have a scheduler
493 * that does not need the MP lock to work because it allows us to do some
494 * really cool high-performance MP lock optimizations.
69d78e99
MD
495 *
496 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
497 * is not called by the current thread in the preemption case, only when
498 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
499 */
500void
501lwkt_switch(void)
502{
37af14fe
MD
503 globaldata_t gd = mycpu;
504 thread_t td = gd->gd_curthread;
8ad65e08 505 thread_t ntd;
f9235b6d
MD
506 thread_t xtd;
507 thread_t nlast;
f9235b6d 508 int nquserok;
6f207a2c 509#ifdef SMP
8a8d5d85
MD
510 int mpheld;
511#endif
f9235b6d 512 int didaccumulate;
b37f18d6
MD
513 const char *lmsg; /* diagnostic - 'systat -pv 1' */
514 const void *laddr;
8ad65e08 515
46a3f46d 516 /*
27e88a6e
MD
517 * Switching from within a 'fast' (non thread switched) interrupt or IPI
518 * is illegal. However, we may have to do it anyway if we hit a fatal
519 * kernel trap or we have paniced.
520 *
521 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 522 */
27e88a6e
MD
523 if (gd->gd_intr_nesting_level) {
524 int savegdnest;
525 int savegdtrap;
526
5fddbda2 527 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22
MD
528 panic("lwkt_switch: Attempt to switch from a "
529 "a fast interrupt, ipi, or hard code section, "
530 "td %p\n",
531 td);
27e88a6e
MD
532 } else {
533 savegdnest = gd->gd_intr_nesting_level;
534 savegdtrap = gd->gd_trap_nesting_level;
535 gd->gd_intr_nesting_level = 0;
536 gd->gd_trap_nesting_level = 0;
a7422615
MD
537 if ((td->td_flags & TDF_PANICWARN) == 0) {
538 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
539 kprintf("Warning: thread switch from interrupt, IPI, "
540 "or hard code section.\n"
a7422615 541 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 542 print_backtrace(-1);
a7422615 543 }
27e88a6e
MD
544 lwkt_switch();
545 gd->gd_intr_nesting_level = savegdnest;
546 gd->gd_trap_nesting_level = savegdtrap;
547 return;
548 }
96728c05 549 }
ef0fdad1 550
cb973d15
MD
551 /*
552 * Passive release (used to transition from user to kernel mode
553 * when we block or switch rather then when we enter the kernel).
554 * This function is NOT called if we are switching into a preemption
555 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
556 * our current process designation (if we have one) and become a true
557 * LWKT thread, and may also hand the current process designation to
558 * another process and schedule thread.
cb973d15
MD
559 */
560 if (td->td_release)
561 td->td_release(td);
562
37af14fe 563 crit_enter_gd(gd);
3b998fa9 564 if (TD_TOKS_HELD(td))
9d265729
MD
565 lwkt_relalltokens(td);
566
567 /*
b02926de
MD
568 * We had better not be holding any spin locks, but don't get into an
569 * endless panic loop.
9d265729 570 */
d666840a
MD
571 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
572 ("lwkt_switch: still holding %d exclusive spinlocks!",
573 gd->gd_spinlocks_wr));
9d265729 574
8a8d5d85
MD
575
576#ifdef SMP
577 /*
3933a3ab
MD
578 * td_mpcount + td_xpcount cannot be used to determine if we currently
579 * hold the MP lock because get_mplock() will increment it prior to
580 * attempting to get the lock, and switch out if it can't. Our
581 * ownership of the actual lock will remain stable while we are
582 * in a critical section, and once we actually acquire the underlying
583 * lock as long as the count is greater than 0.
8a8d5d85 584 */
c5724852 585 mpheld = MP_LOCK_HELD(gd);
0f7a3396
MD
586#ifdef INVARIANTS
587 if (td->td_cscount) {
6ea70f76 588 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
589 td);
590 if (panic_on_cscount)
591 panic("switching while mastering cpusync");
592 }
593#endif
8a8d5d85 594#endif
f9235b6d
MD
595
596 /*
597 * If we had preempted another thread on this cpu, resume the preempted
598 * thread. This occurs transparently, whether the preempted thread
599 * was scheduled or not (it may have been preempted after descheduling
600 * itself).
601 *
602 * We have to setup the MP lock for the original thread after backing
603 * out the adjustment that was made to curthread when the original
604 * was preempted.
605 */
99df837e 606 if ((ntd = td->td_preempted) != NULL) {
26a0694b 607 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 608#ifdef SMP
3933a3ab 609 if (ntd->td_mpcount + ntd->td_xpcount && mpheld == 0) {
fc92d4aa 610 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
3933a3ab 611 td, ntd, td->td_mpcount, ntd->td_mpcount + ntd->td_xpcount);
8a8d5d85 612 }
3933a3ab 613 td->td_xpcount = 0;
8a8d5d85 614#endif
26a0694b 615 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
616
617 /*
b9eb1c19
MD
618 * The interrupt may have woken a thread up, we need to properly
619 * set the reschedule flag if the originally interrupted thread is
620 * at a lower priority.
8ec60c3f 621 */
f9235b6d
MD
622 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
623 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
8ec60c3f 624 need_lwkt_resched();
f9235b6d 625 }
8a8d5d85 626 /* YYY release mp lock on switchback if original doesn't need it */
f9235b6d
MD
627 goto havethread_preempted;
628 }
629
630 /*
631 * Implement round-robin fairq with priority insertion. The priority
632 * insertion is handled by _lwkt_enqueue()
633 *
634 * We have to adjust the MP lock for the target thread. If we
635 * need the MP lock and cannot obtain it we try to locate a
636 * thread that does not need the MP lock. If we cannot, we spin
637 * instead of HLT.
638 *
639 * A similar issue exists for the tokens held by the target thread.
640 * If we cannot obtain ownership of the tokens we cannot immediately
641 * schedule the thread.
642 */
643 for (;;) {
644 clear_lwkt_resched();
645 didaccumulate = 0;
646 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
647
4b5f931b 648 /*
f9235b6d 649 * Hotpath if we can get all necessary resources.
41a01a4d 650 *
f9235b6d 651 * If nothing is runnable switch to the idle thread
41a01a4d 652 */
f9235b6d
MD
653 if (ntd == NULL) {
654 ntd = &gd->gd_idlethread;
655 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
656 ntd->td_flags |= TDF_IDLE_NOHLT;
6f207a2c 657#ifdef SMP
3933a3ab 658 KKASSERT(ntd->td_xpcount == 0);
f9235b6d
MD
659 if (ntd->td_mpcount) {
660 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
661 panic("Idle thread %p was holding the BGL!", ntd);
662 if (mpheld == 0) {
c5724852
MD
663 set_cpu_contention_mask(gd);
664 handle_cpu_contention_mask();
665 cpu_try_mplock();
666 mpheld = MP_LOCK_HELD(gd);
f9235b6d
MD
667 cpu_pause();
668 continue;
669 }
670 }
c5724852 671 clr_cpu_contention_mask(gd);
6f207a2c 672#endif
b37f18d6
MD
673 cpu_time.cp_msg[0] = 0;
674 cpu_time.cp_stallpc = 0;
f9235b6d
MD
675 goto haveidle;
676 }
41a01a4d 677
8ec60c3f 678 /*
f9235b6d 679 * Hotpath schedule
6f207a2c
MD
680 *
681 * NOTE: For UP there is no mplock and lwkt_getalltokens()
682 * always succeeds.
8ec60c3f 683 */
f9235b6d
MD
684 if (ntd->td_fairq_accum >= 0 &&
685#ifdef SMP
3933a3ab
MD
686 (ntd->td_mpcount + ntd->td_xpcount == 0 ||
687 mpheld || cpu_try_mplock()) &&
f9235b6d 688#endif
b37f18d6 689 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
f9235b6d 690 ) {
8a8d5d85 691#ifdef SMP
c5724852 692 clr_cpu_contention_mask(gd);
f9235b6d
MD
693#endif
694 goto havethread;
695 }
696
b37f18d6
MD
697 lmsg = NULL;
698 laddr = NULL;
699
f9235b6d 700#ifdef SMP
c5724852
MD
701 if (ntd->td_fairq_accum >= 0)
702 set_cpu_contention_mask(gd);
f9235b6d 703 /* Reload mpheld (it become stale after mplock/token ops) */
c5724852 704 mpheld = MP_LOCK_HELD(gd);
3933a3ab 705 if (ntd->td_mpcount + ntd->td_xpcount && mpheld == 0) {
b37f18d6
MD
706 lmsg = "mplock";
707 laddr = ntd->td_mplock_stallpc;
708 }
f9235b6d
MD
709#endif
710
711 /*
712 * Coldpath - unable to schedule ntd, continue looking for threads
713 * to schedule. This is only allowed of the (presumably) kernel
714 * thread exhausted its fair share. A kernel thread stuck on
715 * resources does not currently allow a user thread to get in
716 * front of it.
717 */
718#ifdef SMP
719 nquserok = ((ntd->td_pri < TDPRI_KERN_LPSCHED) ||
720 (ntd->td_fairq_accum < 0));
6f207a2c
MD
721#else
722 nquserok = 1;
f9235b6d
MD
723#endif
724 nlast = NULL;
725
726 for (;;) {
41a01a4d 727 /*
f9235b6d
MD
728 * If the fair-share scheduler ran out ntd gets moved to the
729 * end and its accumulator will be bumped, if it didn't we
730 * maintain the same queue position.
df6b8ba0 731 *
f9235b6d 732 * nlast keeps track of the last element prior to any moves.
41a01a4d 733 */
f9235b6d 734 if (ntd->td_fairq_accum < 0) {
f9235b6d
MD
735 lwkt_fairq_accumulate(gd, ntd);
736 didaccumulate = 1;
c5724852
MD
737
738 /*
739 * Move to end
740 */
741 xtd = TAILQ_NEXT(ntd, td_threadq);
f9235b6d
MD
742 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
743 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
c5724852
MD
744
745 /*
746 * Set terminal element (nlast)
747 */
f9235b6d
MD
748 if (nlast == NULL) {
749 nlast = ntd;
750 if (xtd == NULL)
751 xtd = ntd;
752 }
753 ntd = xtd;
754 } else {
755 ntd = TAILQ_NEXT(ntd, td_threadq);
756 }
a453459d 757
f9235b6d
MD
758 /*
759 * If we exhausted the run list switch to the idle thread.
760 * Since one or more threads had resource acquisition issues
761 * we do not allow the idle thread to halt.
762 *
763 * NOTE: nlast can be NULL.
764 */
765 if (ntd == nlast) {
e0a90d3b 766 cpu_pause();
f9235b6d
MD
767 ntd = &gd->gd_idlethread;
768 ntd->td_flags |= TDF_IDLE_NOHLT;
6f207a2c 769#ifdef SMP
3933a3ab 770 KKASSERT(ntd->td_xpcount == 0);
f9235b6d 771 if (ntd->td_mpcount) {
c5724852 772 mpheld = MP_LOCK_HELD(gd);
f9235b6d
MD
773 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
774 panic("Idle thread %p was holding the BGL!", ntd);
775 if (mpheld == 0) {
c5724852
MD
776 set_cpu_contention_mask(gd);
777 handle_cpu_contention_mask();
778 cpu_try_mplock();
779 mpheld = MP_LOCK_HELD(gd);
f9235b6d
MD
780 cpu_pause();
781 break; /* try again from the top, almost */
b9eb1c19 782 }
8a8d5d85 783 }
6f207a2c 784#endif
684a93c4
MD
785
786 /*
f9235b6d
MD
787 * If fairq accumulations occured we do not schedule the
788 * idle thread. This will cause us to try again from
789 * the (almost) top.
684a93c4 790 */
f9235b6d 791 if (didaccumulate)
b37f18d6
MD
792 break; /* try again from the top, almost */
793 if (lmsg)
794 strlcpy(cpu_time.cp_msg, lmsg, sizeof(cpu_time.cp_msg));
795 cpu_time.cp_stallpc = (uintptr_t)laddr;
f9235b6d 796 goto haveidle;
8a8d5d85 797 }
f9235b6d 798
df6b8ba0 799 /*
f9235b6d 800 * Try to switch to this thread.
6f207a2c
MD
801 *
802 * NOTE: For UP there is no mplock and lwkt_getalltokens()
803 * always succeeds.
df6b8ba0 804 */
77912481
MD
805 if ((ntd->td_pri >= TDPRI_KERN_LPSCHED || nquserok ||
806 user_pri_sched) && ntd->td_fairq_accum >= 0 &&
f9235b6d 807#ifdef SMP
3933a3ab
MD
808 (ntd->td_mpcount + ntd->td_xpcount == 0 ||
809 mpheld || cpu_try_mplock()) &&
8a8d5d85 810#endif
b37f18d6 811 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
f9235b6d 812 ) {
a453459d 813#ifdef SMP
c5724852 814 clr_cpu_contention_mask(gd);
f9235b6d
MD
815#endif
816 goto havethread;
df6b8ba0 817 }
f9235b6d 818#ifdef SMP
c5724852
MD
819 if (ntd->td_fairq_accum >= 0)
820 set_cpu_contention_mask(gd);
821 /*
822 * Reload mpheld (it become stale after mplock/token ops).
823 */
824 mpheld = MP_LOCK_HELD(gd);
3933a3ab 825 if (ntd->td_mpcount + ntd->td_xpcount && mpheld == 0) {
b37f18d6
MD
826 lmsg = "mplock";
827 laddr = ntd->td_mplock_stallpc;
828 }
f9235b6d
MD
829 if (ntd->td_pri >= TDPRI_KERN_LPSCHED && ntd->td_fairq_accum >= 0)
830 nquserok = 0;
a453459d 831#endif
4b5f931b 832 }
c5724852
MD
833
834 /*
835 * All threads exhausted but we can loop due to a negative
836 * accumulator.
837 *
838 * While we are looping in the scheduler be sure to service
839 * any interrupts which were made pending due to our critical
840 * section, otherwise we could livelock (e.g.) IPIs.
841 *
842 * NOTE: splz can enter and exit the mplock so mpheld is
843 * stale after this call.
844 */
845 splz_check();
846
847#ifdef SMP
848 /*
849 * Our mplock can be cached and cause other cpus to livelock
850 * if we loop due to e.g. not being able to acquire tokens.
851 */
852 if (MP_LOCK_HELD(gd))
853 cpu_rel_mplock(gd->gd_cpuid);
854 mpheld = 0;
855#endif
f1d1c3fa 856 }
8a8d5d85
MD
857
858 /*
f9235b6d
MD
859 * Do the actual switch. WARNING: mpheld is stale here.
860 *
861 * We must always decrement td_fairq_accum on non-idle threads just
862 * in case a thread never gets a tick due to being in a continuous
863 * critical section. The page-zeroing code does that.
864 *
865 * If the thread we came up with is a higher or equal priority verses
866 * the thread at the head of the queue we move our thread to the
867 * front. This way we can always check the front of the queue.
868 */
869havethread:
870 ++gd->gd_cnt.v_swtch;
871 --ntd->td_fairq_accum;
872 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
873 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
874 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
875 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
876 }
877havethread_preempted:
878 ;
879 /*
880 * If the new target does not need the MP lock and we are holding it,
881 * release the MP lock. If the new target requires the MP lock we have
882 * already acquired it for the target.
883 *
884 * WARNING: mpheld is stale here.
8a8d5d85 885 */
f9235b6d
MD
886haveidle:
887 KASSERT(ntd->td_critcount,
888 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85 889#ifdef SMP
3933a3ab 890 if (ntd->td_mpcount + ntd->td_xpcount == 0 ) {
c5724852
MD
891 if (MP_LOCK_HELD(gd))
892 cpu_rel_mplock(gd->gd_cpuid);
8a8d5d85 893 } else {
a453459d 894 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
895 }
896#endif
94f6d86e
MD
897 if (td != ntd) {
898 ++switch_count;
b2b3ffcd 899#ifdef __x86_64__
f9235b6d
MD
900 {
901 int tos_ok __debugvar = jg_tos_ok(ntd);
902 KKASSERT(tos_ok);
903 }
85514115 904#endif
a1f0fb66 905 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
f1d1c3fa 906 td->td_switch(ntd);
94f6d86e 907 }
37af14fe
MD
908 /* NOTE: current cpu may have changed after switch */
909 crit_exit_quick(td);
8ad65e08
MD
910}
911
b68b7282 912/*
96728c05
MD
913 * Request that the target thread preempt the current thread. Preemption
914 * only works under a specific set of conditions:
b68b7282 915 *
96728c05
MD
916 * - We are not preempting ourselves
917 * - The target thread is owned by the current cpu
918 * - We are not currently being preempted
919 * - The target is not currently being preempted
d3d1cbc8
MD
920 * - We are not holding any spin locks
921 * - The target thread is not holding any tokens
96728c05
MD
922 * - We are able to satisfy the target's MP lock requirements (if any).
923 *
924 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
925 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 926 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
927 * to determine whether preemption is possible (aka usually just the crit
928 * priority of lwkt_schedule() itself).
b68b7282 929 *
26a0694b
MD
930 * XXX at the moment we run the target thread in a critical section during
931 * the preemption in order to prevent the target from taking interrupts
932 * that *WE* can't. Preemption is strictly limited to interrupt threads
933 * and interrupt-like threads, outside of a critical section, and the
934 * preempted source thread will be resumed the instant the target blocks
935 * whether or not the source is scheduled (i.e. preemption is supposed to
936 * be as transparent as possible).
4b5f931b 937 *
8a8d5d85
MD
938 * The target thread inherits our MP count (added to its own) for the
939 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
940 * MP lock during the preemption. Therefore, any preempting targets must be
941 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
942 * out of sync with the physical mp_lock, but we do not have to preserve
943 * the original ownership of the lock if it was out of synch (that is, we
944 * can leave it synchronized on return).
b68b7282
MD
945 */
946void
f9235b6d 947lwkt_preempt(thread_t ntd, int critcount)
b68b7282 948{
46a3f46d 949 struct globaldata *gd = mycpu;
0a3f9b47 950 thread_t td;
8a8d5d85
MD
951#ifdef SMP
952 int mpheld;
57c254db 953 int savecnt;
8a8d5d85 954#endif
b68b7282 955
26a0694b 956 /*
96728c05
MD
957 * The caller has put us in a critical section. We can only preempt
958 * if the caller of the caller was not in a critical section (basically
f9235b6d 959 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 960 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 961 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
962 *
963 * YYY The target thread must be in a critical section (else it must
964 * inherit our critical section? I dunno yet).
41a01a4d 965 *
0a3f9b47 966 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b 967 */
f9235b6d 968 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 969
fbc024e4
MD
970 if (preempt_enable == 0) {
971 ++preempt_miss;
972 return;
973 }
974
0a3f9b47 975 td = gd->gd_curthread;
f9235b6d 976 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
977 ++preempt_miss;
978 return;
979 }
f9235b6d 980 if (td->td_critcount > critcount) {
96728c05 981 ++preempt_miss;
8ec60c3f 982 need_lwkt_resched();
96728c05
MD
983 return;
984 }
985#ifdef SMP
46a3f46d 986 if (ntd->td_gd != gd) {
96728c05 987 ++preempt_miss;
8ec60c3f 988 need_lwkt_resched();
96728c05
MD
989 return;
990 }
991#endif
41a01a4d 992 /*
77912481
MD
993 * We don't have to check spinlocks here as they will also bump
994 * td_critcount.
d3d1cbc8
MD
995 *
996 * Do not try to preempt if the target thread is holding any tokens.
997 * We could try to acquire the tokens but this case is so rare there
998 * is no need to support it.
41a01a4d 999 */
77912481
MD
1000 KKASSERT(gd->gd_spinlocks_wr == 0);
1001
3b998fa9 1002 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8
MD
1003 ++preempt_miss;
1004 need_lwkt_resched();
1005 return;
1006 }
26a0694b
MD
1007 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1008 ++preempt_weird;
8ec60c3f 1009 need_lwkt_resched();
26a0694b
MD
1010 return;
1011 }
1012 if (ntd->td_preempted) {
4b5f931b 1013 ++preempt_hit;
8ec60c3f 1014 need_lwkt_resched();
26a0694b 1015 return;
b68b7282 1016 }
8a8d5d85 1017#ifdef SMP
a2a5ad0d 1018 /*
3933a3ab 1019 * NOTE: An interrupt might have occured just as we were transitioning
71ef2f5c 1020 * to or from the MP lock. In this case td_mpcount will be pre-disposed
3933a3ab
MD
1021 * (non-zero) but not actually synchronized with the mp_lock itself.
1022 * We can use it to imply an MP lock requirement for the preemption but
1023 * we cannot use it to test whether we hold the MP lock or not.
a2a5ad0d 1024 */
96728c05 1025 savecnt = td->td_mpcount;
c5724852 1026 mpheld = MP_LOCK_HELD(gd);
6d9b99db 1027 ntd->td_xpcount = td->td_mpcount + td->td_xpcount;
3933a3ab
MD
1028 if (mpheld == 0 && ntd->td_mpcount + ntd->td_xpcount && !cpu_try_mplock()) {
1029 ntd->td_xpcount = 0;
8a8d5d85 1030 ++preempt_miss;
8ec60c3f 1031 need_lwkt_resched();
8a8d5d85
MD
1032 return;
1033 }
1034#endif
26a0694b 1035
8ec60c3f
MD
1036 /*
1037 * Since we are able to preempt the current thread, there is no need to
1038 * call need_lwkt_resched().
1039 */
26a0694b
MD
1040 ++preempt_hit;
1041 ntd->td_preempted = td;
1042 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1043 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
26a0694b 1044 td->td_switch(ntd);
b9eb1c19 1045
26a0694b 1046 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
1047#ifdef SMP
1048 KKASSERT(savecnt == td->td_mpcount);
c5724852 1049 mpheld = MP_LOCK_HELD(gd);
71ef2f5c 1050 if (mpheld && td->td_mpcount == 0)
c5724852 1051 cpu_rel_mplock(gd->gd_cpuid);
3933a3ab 1052 else if (mpheld == 0 && td->td_mpcount + td->td_xpcount)
96728c05
MD
1053 panic("lwkt_preempt(): MP lock was not held through");
1054#endif
26a0694b
MD
1055 ntd->td_preempted = NULL;
1056 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1057}
1058
f1d1c3fa 1059/*
faaeffac 1060 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1061 * This will work inside a critical section but not inside a hard code
1062 * section.
ef0fdad1 1063 *
f1d1c3fa
MD
1064 * (self contained on a per cpu basis)
1065 */
1066void
faaeffac 1067splz_check(void)
f1d1c3fa 1068{
7966cb69
MD
1069 globaldata_t gd = mycpu;
1070 thread_t td = gd->gd_curthread;
ef0fdad1 1071
4a28fe22
MD
1072 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1073 gd->gd_intr_nesting_level == 0 &&
1074 td->td_nest_count < 2)
1075 {
f1d1c3fa 1076 splz();
4a28fe22
MD
1077 }
1078}
1079
1080/*
1081 * This version is integrated into crit_exit, reqflags has already
1082 * been tested but td_critcount has not.
1083 *
1084 * We only want to execute the splz() on the 1->0 transition of
1085 * critcount and not in a hard code section or if too deeply nested.
1086 */
1087void
1088lwkt_maybe_splz(thread_t td)
1089{
1090 globaldata_t gd = td->td_gd;
1091
1092 if (td->td_critcount == 0 &&
1093 gd->gd_intr_nesting_level == 0 &&
1094 td->td_nest_count < 2)
1095 {
1096 splz();
1097 }
f1d1c3fa
MD
1098}
1099
8ad65e08 1100/*
f9235b6d
MD
1101 * This function is used to negotiate a passive release of the current
1102 * process/lwp designation with the user scheduler, allowing the user
1103 * scheduler to schedule another user thread. The related kernel thread
1104 * (curthread) continues running in the released state.
8ad65e08
MD
1105 */
1106void
f9235b6d 1107lwkt_passive_release(struct thread *td)
8ad65e08 1108{
f9235b6d
MD
1109 struct lwp *lp = td->td_lwp;
1110
1111 td->td_release = NULL;
1112 lwkt_setpri_self(TDPRI_KERN_USER);
1113 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1114}
1115
f9235b6d 1116
3824f392 1117/*
f9235b6d
MD
1118 * This implements a normal yield. This routine is virtually a nop if
1119 * there is nothing to yield to but it will always run any pending interrupts
1120 * if called from a critical section.
1121 *
1122 * This yield is designed for kernel threads without a user context.
1123 *
1124 * (self contained on a per cpu basis)
3824f392
MD
1125 */
1126void
f9235b6d 1127lwkt_yield(void)
3824f392 1128{
f9235b6d
MD
1129 globaldata_t gd = mycpu;
1130 thread_t td = gd->gd_curthread;
1131 thread_t xtd;
3824f392 1132
f9235b6d
MD
1133 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1134 splz();
1135 if (td->td_fairq_accum < 0) {
1136 lwkt_schedule_self(curthread);
1137 lwkt_switch();
1138 } else {
1139 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1140 if (xtd && xtd->td_pri > td->td_pri) {
1141 lwkt_schedule_self(curthread);
1142 lwkt_switch();
1143 }
1144 }
3824f392
MD
1145}
1146
1147/*
f9235b6d
MD
1148 * This yield is designed for kernel threads with a user context.
1149 *
1150 * The kernel acting on behalf of the user is potentially cpu-bound,
1151 * this function will efficiently allow other threads to run and also
1152 * switch to other processes by releasing.
3824f392
MD
1153 *
1154 * The lwkt_user_yield() function is designed to have very low overhead
1155 * if no yield is determined to be needed.
1156 */
1157void
1158lwkt_user_yield(void)
1159{
f9235b6d
MD
1160 globaldata_t gd = mycpu;
1161 thread_t td = gd->gd_curthread;
1162
1163 /*
1164 * Always run any pending interrupts in case we are in a critical
1165 * section.
1166 */
1167 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1168 splz();
3824f392
MD
1169
1170#ifdef SMP
1171 /*
1172 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1173 * kernel can prevent other cpus from servicing interrupt threads
1174 * which still require the MP lock (which is a lot of them). This
1175 * has a chaining effect since if the interrupt is blocked, so is
1176 * the event, so normal scheduling will not pick up on the problem.
1177 */
3933a3ab 1178 if (cpu_contention_mask && td->td_mpcount + td->td_xpcount) {
684a93c4 1179 yield_mplock(td);
3824f392
MD
1180 }
1181#endif
1182
1183 /*
f9235b6d
MD
1184 * Switch (which forces a release) if another kernel thread needs
1185 * the cpu, if userland wants us to resched, or if our kernel
1186 * quantum has run out.
3824f392 1187 */
f9235b6d
MD
1188 if (lwkt_resched_wanted() ||
1189 user_resched_wanted() ||
1190 td->td_fairq_accum < 0)
1191 {
3824f392 1192 lwkt_switch();
3824f392
MD
1193 }
1194
f9235b6d 1195#if 0
3824f392 1196 /*
f9235b6d
MD
1197 * Reacquire the current process if we are released.
1198 *
1199 * XXX not implemented atm. The kernel may be holding locks and such,
1200 * so we want the thread to continue to receive cpu.
3824f392 1201 */
f9235b6d
MD
1202 if (td->td_release == NULL && lp) {
1203 lp->lwp_proc->p_usched->acquire_curproc(lp);
1204 td->td_release = lwkt_passive_release;
1205 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1206 }
f9235b6d 1207#endif
b9eb1c19
MD
1208}
1209
8ad65e08 1210/*
f1d1c3fa
MD
1211 * Generic schedule. Possibly schedule threads belonging to other cpus and
1212 * deal with threads that might be blocked on a wait queue.
1213 *
0a3f9b47
MD
1214 * We have a little helper inline function which does additional work after
1215 * the thread has been enqueued, including dealing with preemption and
1216 * setting need_lwkt_resched() (which prevents the kernel from returning
1217 * to userland until it has processed higher priority threads).
6330a558
MD
1218 *
1219 * It is possible for this routine to be called after a failed _enqueue
1220 * (due to the target thread migrating, sleeping, or otherwise blocked).
1221 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1222 *
1223 * reschedok is an optimized constant propagated from lwkt_schedule() or
1224 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1225 * reschedule to be requested if the target thread has a higher priority.
1226 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1227 * be 0, prevented undesired reschedules.
8ad65e08 1228 */
0a3f9b47
MD
1229static __inline
1230void
f9235b6d 1231_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
0a3f9b47 1232{
b9eb1c19 1233 thread_t otd;
c730be20 1234
6330a558 1235 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1236 if (ntd->td_preemptable && reschedok) {
f9235b6d 1237 ntd->td_preemptable(ntd, ccount); /* YYY +token */
361d01dd 1238 } else if (reschedok) {
b9eb1c19 1239 otd = curthread;
f9235b6d 1240 if (ntd->td_pri > otd->td_pri)
c730be20 1241 need_lwkt_resched();
6330a558 1242 }
f9235b6d
MD
1243
1244 /*
1245 * Give the thread a little fair share scheduler bump if it
1246 * has been asleep for a while. This is primarily to avoid
1247 * a degenerate case for interrupt threads where accumulator
1248 * crosses into negative territory unnecessarily.
1249 */
1250 if (ntd->td_fairq_lticks != ticks) {
1251 ntd->td_fairq_lticks = ticks;
1252 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1253 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1254 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1255 }
0a3f9b47
MD
1256 }
1257}
1258
361d01dd 1259static __inline
8ad65e08 1260void
361d01dd 1261_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1262{
37af14fe
MD
1263 globaldata_t mygd = mycpu;
1264
41a01a4d 1265 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1266 crit_enter_gd(mygd);
9388413d 1267 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1268 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1269 _lwkt_enqueue(td);
1270 } else {
f1d1c3fa 1271 /*
7cd8d145
MD
1272 * If we own the thread, there is no race (since we are in a
1273 * critical section). If we do not own the thread there might
1274 * be a race but the target cpu will deal with it.
f1d1c3fa 1275 */
0f7a3396 1276#ifdef SMP
7cd8d145 1277 if (td->td_gd == mygd) {
9d265729 1278 _lwkt_enqueue(td);
f9235b6d 1279 _lwkt_schedule_post(mygd, td, 1, reschedok);
f1d1c3fa 1280 } else {
e381e77c 1281 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1282 }
0f7a3396 1283#else
7cd8d145 1284 _lwkt_enqueue(td);
f9235b6d 1285 _lwkt_schedule_post(mygd, td, 1, reschedok);
0f7a3396 1286#endif
8ad65e08 1287 }
37af14fe 1288 crit_exit_gd(mygd);
8ad65e08
MD
1289}
1290
361d01dd
MD
1291void
1292lwkt_schedule(thread_t td)
1293{
1294 _lwkt_schedule(td, 1);
1295}
1296
1297void
1298lwkt_schedule_noresched(thread_t td)
1299{
1300 _lwkt_schedule(td, 0);
1301}
1302
88ebb169
SW
1303#ifdef SMP
1304
e381e77c
MD
1305/*
1306 * When scheduled remotely if frame != NULL the IPIQ is being
1307 * run via doreti or an interrupt then preemption can be allowed.
1308 *
1309 * To allow preemption we have to drop the critical section so only
1310 * one is present in _lwkt_schedule_post.
1311 */
1312static void
1313lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1314{
1315 thread_t td = curthread;
1316 thread_t ntd = arg;
1317
1318 if (frame && ntd->td_preemptable) {
1319 crit_exit_noyield(td);
1320 _lwkt_schedule(ntd, 1);
1321 crit_enter_quick(td);
1322 } else {
1323 _lwkt_schedule(ntd, 1);
1324 }
1325}
1326
d9eea1a5 1327/*
52eedfb5
MD
1328 * Thread migration using a 'Pull' method. The thread may or may not be
1329 * the current thread. It MUST be descheduled and in a stable state.
1330 * lwkt_giveaway() must be called on the cpu owning the thread.
1331 *
1332 * At any point after lwkt_giveaway() is called, the target cpu may
1333 * 'pull' the thread by calling lwkt_acquire().
1334 *
ae8e83e6
MD
1335 * We have to make sure the thread is not sitting on a per-cpu tsleep
1336 * queue or it will blow up when it moves to another cpu.
1337 *
52eedfb5 1338 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1339 */
52eedfb5
MD
1340void
1341lwkt_giveaway(thread_t td)
1342{
3b4192fb 1343 globaldata_t gd = mycpu;
52eedfb5 1344
3b4192fb
MD
1345 crit_enter_gd(gd);
1346 if (td->td_flags & TDF_TSLEEPQ)
1347 tsleep_remove(td);
1348 KKASSERT(td->td_gd == gd);
1349 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1350 td->td_flags |= TDF_MIGRATING;
1351 crit_exit_gd(gd);
52eedfb5
MD
1352}
1353
a2a5ad0d
MD
1354void
1355lwkt_acquire(thread_t td)
1356{
37af14fe
MD
1357 globaldata_t gd;
1358 globaldata_t mygd;
a2a5ad0d 1359
52eedfb5 1360 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1361 gd = td->td_gd;
37af14fe 1362 mygd = mycpu;
52eedfb5 1363 if (gd != mycpu) {
35238fa5 1364 cpu_lfence();
52eedfb5 1365 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1366 crit_enter_gd(mygd);
df910c23
MD
1367 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1368#ifdef SMP
1369 lwkt_process_ipiq();
1370#endif
52eedfb5 1371 cpu_lfence();
df910c23 1372 }
562273ea 1373 cpu_mfence();
37af14fe 1374 td->td_gd = mygd;
52eedfb5
MD
1375 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1376 td->td_flags &= ~TDF_MIGRATING;
1377 crit_exit_gd(mygd);
1378 } else {
1379 crit_enter_gd(mygd);
1380 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1381 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1382 crit_exit_gd(mygd);
a2a5ad0d
MD
1383 }
1384}
1385
52eedfb5
MD
1386#endif
1387
f1d1c3fa
MD
1388/*
1389 * Generic deschedule. Descheduling threads other then your own should be
1390 * done only in carefully controlled circumstances. Descheduling is
1391 * asynchronous.
1392 *
1393 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1394 */
1395void
1396lwkt_deschedule(thread_t td)
1397{
f1d1c3fa 1398 crit_enter();
b8a98473 1399#ifdef SMP
f1d1c3fa
MD
1400 if (td == curthread) {
1401 _lwkt_dequeue(td);
1402 } else {
a72187e9 1403 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1404 _lwkt_dequeue(td);
1405 } else {
b8a98473 1406 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1407 }
1408 }
b8a98473
MD
1409#else
1410 _lwkt_dequeue(td);
1411#endif
f1d1c3fa
MD
1412 crit_exit();
1413}
1414
4b5f931b
MD
1415/*
1416 * Set the target thread's priority. This routine does not automatically
1417 * switch to a higher priority thread, LWKT threads are not designed for
1418 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1419 */
1420void
1421lwkt_setpri(thread_t td, int pri)
1422{
a72187e9 1423 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1424 if (td->td_pri != pri) {
1425 KKASSERT(pri >= 0);
1426 crit_enter();
1427 if (td->td_flags & TDF_RUNQ) {
1428 _lwkt_dequeue(td);
1429 td->td_pri = pri;
1430 _lwkt_enqueue(td);
1431 } else {
1432 td->td_pri = pri;
1433 }
1434 crit_exit();
26a0694b 1435 }
26a0694b
MD
1436}
1437
03bd0a5e
MD
1438/*
1439 * Set the initial priority for a thread prior to it being scheduled for
1440 * the first time. The thread MUST NOT be scheduled before or during
1441 * this call. The thread may be assigned to a cpu other then the current
1442 * cpu.
1443 *
1444 * Typically used after a thread has been created with TDF_STOPPREQ,
1445 * and before the thread is initially scheduled.
1446 */
1447void
1448lwkt_setpri_initial(thread_t td, int pri)
1449{
1450 KKASSERT(pri >= 0);
1451 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1452 td->td_pri = pri;
03bd0a5e
MD
1453}
1454
26a0694b
MD
1455void
1456lwkt_setpri_self(int pri)
1457{
1458 thread_t td = curthread;
1459
4b5f931b
MD
1460 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1461 crit_enter();
1462 if (td->td_flags & TDF_RUNQ) {
1463 _lwkt_dequeue(td);
f9235b6d 1464 td->td_pri = pri;
4b5f931b
MD
1465 _lwkt_enqueue(td);
1466 } else {
f9235b6d 1467 td->td_pri = pri;
4b5f931b
MD
1468 }
1469 crit_exit();
1470}
1471
f9235b6d
MD
1472/*
1473 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1474 *
1475 * Example: two competing threads, same priority N. decrement by (2*N)
1476 * increment by N*8, each thread will get 4 ticks.
1477 */
1478void
1479lwkt_fairq_schedulerclock(thread_t td)
1480{
1481 if (fairq_enable) {
1482 while (td) {
1483 if (td != &td->td_gd->gd_idlethread) {
1484 td->td_fairq_accum -= td->td_gd->gd_fairq_total_pri;
1485 if (td->td_fairq_accum < -TDFAIRQ_MAX(td->td_gd))
1486 td->td_fairq_accum = -TDFAIRQ_MAX(td->td_gd);
1487 if (td->td_fairq_accum < 0)
1488 need_lwkt_resched();
1489 td->td_fairq_lticks = ticks;
1490 }
1491 td = td->td_preempted;
1492 }
1493 }
1494}
1495
1496static void
1497lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1498{
1499 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1500 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1501 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1502}
1503
5d21b981 1504/*
52eedfb5
MD
1505 * Migrate the current thread to the specified cpu.
1506 *
1507 * This is accomplished by descheduling ourselves from the current cpu,
1508 * moving our thread to the tdallq of the target cpu, IPI messaging the
1509 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1510 * races while the thread is being migrated.
ae8e83e6
MD
1511 *
1512 * We must be sure to remove ourselves from the current cpu's tsleepq
1513 * before potentially moving to another queue. The thread can be on
1514 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1515 */
3d28ff59 1516#ifdef SMP
5d21b981 1517static void lwkt_setcpu_remote(void *arg);
3d28ff59 1518#endif
5d21b981
MD
1519
1520void
1521lwkt_setcpu_self(globaldata_t rgd)
1522{
1523#ifdef SMP
1524 thread_t td = curthread;
1525
1526 if (td->td_gd != rgd) {
1527 crit_enter_quick(td);
ae8e83e6 1528 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1529 tsleep_remove(td);
5d21b981
MD
1530 td->td_flags |= TDF_MIGRATING;
1531 lwkt_deschedule_self(td);
52eedfb5 1532 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1533 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1534 lwkt_switch();
1535 /* we are now on the target cpu */
52eedfb5 1536 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1537 crit_exit_quick(td);
1538 }
1539#endif
1540}
1541
ecdefdda
MD
1542void
1543lwkt_migratecpu(int cpuid)
1544{
1545#ifdef SMP
1546 globaldata_t rgd;
1547
1548 rgd = globaldata_find(cpuid);
1549 lwkt_setcpu_self(rgd);
1550#endif
1551}
1552
5d21b981
MD
1553/*
1554 * Remote IPI for cpu migration (called while in a critical section so we
1555 * do not have to enter another one). The thread has already been moved to
1556 * our cpu's allq, but we must wait for the thread to be completely switched
1557 * out on the originating cpu before we schedule it on ours or the stack
1558 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1559 * change to main memory.
1560 *
1561 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1562 * against wakeups. It is best if this interface is used only when there
1563 * are no pending events that might try to schedule the thread.
1564 */
3d28ff59 1565#ifdef SMP
5d21b981
MD
1566static void
1567lwkt_setcpu_remote(void *arg)
1568{
1569 thread_t td = arg;
1570 globaldata_t gd = mycpu;
1571
df910c23
MD
1572 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1573#ifdef SMP
1574 lwkt_process_ipiq();
1575#endif
35238fa5 1576 cpu_lfence();
562273ea 1577 cpu_pause();
df910c23 1578 }
5d21b981 1579 td->td_gd = gd;
562273ea 1580 cpu_mfence();
5d21b981 1581 td->td_flags &= ~TDF_MIGRATING;
9388413d 1582 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1583 _lwkt_enqueue(td);
1584}
3d28ff59 1585#endif
5d21b981 1586
553ea3c8 1587struct lwp *
4b5f931b
MD
1588lwkt_preempted_proc(void)
1589{
73e4f7b9 1590 thread_t td = curthread;
4b5f931b
MD
1591 while (td->td_preempted)
1592 td = td->td_preempted;
553ea3c8 1593 return(td->td_lwp);
4b5f931b
MD
1594}
1595
99df837e
MD
1596/*
1597 * Create a kernel process/thread/whatever. It shares it's address space
1598 * with proc0 - ie: kernel only.
1599 *
365fa13f
MD
1600 * NOTE! By default new threads are created with the MP lock held. A
1601 * thread which does not require the MP lock should release it by calling
1602 * rel_mplock() at the start of the new thread.
99df837e
MD
1603 */
1604int
c9e9fb21
MD
1605lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1606 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1607{
73e4f7b9 1608 thread_t td;
e2565a42 1609 __va_list ap;
99df837e 1610
d3d32139 1611 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1612 tdflags);
a2a5ad0d
MD
1613 if (tdp)
1614 *tdp = td;
709799ea 1615 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1616
1617 /*
1618 * Set up arg0 for 'ps' etc
1619 */
e2565a42 1620 __va_start(ap, fmt);
379210cb 1621 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1622 __va_end(ap);
99df837e
MD
1623
1624 /*
1625 * Schedule the thread to run
1626 */
ef0fdad1
MD
1627 if ((td->td_flags & TDF_STOPREQ) == 0)
1628 lwkt_schedule(td);
1629 else
1630 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1631 return 0;
1632}
1633
1634/*
1635 * Destroy an LWKT thread. Warning! This function is not called when
1636 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1637 * uses a different reaping mechanism.
1638 */
1639void
1640lwkt_exit(void)
1641{
1642 thread_t td = curthread;
c070746a 1643 thread_t std;
8826f33a 1644 globaldata_t gd;
99df837e
MD
1645
1646 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1647 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1648 caps_exit(td);
c070746a
MD
1649
1650 /*
1651 * Get us into a critical section to interlock gd_freetd and loop
1652 * until we can get it freed.
1653 *
1654 * We have to cache the current td in gd_freetd because objcache_put()ing
1655 * it would rip it out from under us while our thread is still active.
1656 */
1657 gd = mycpu;
37af14fe 1658 crit_enter_quick(td);
c070746a
MD
1659 while ((std = gd->gd_freetd) != NULL) {
1660 gd->gd_freetd = NULL;
1661 objcache_put(thread_cache, std);
1662 }
3b4192fb
MD
1663
1664 /*
1665 * Remove thread resources from kernel lists and deschedule us for
1666 * the last time.
1667 */
1668 if (td->td_flags & TDF_TSLEEPQ)
1669 tsleep_remove(td);
79eae878 1670 biosched_done(td);
f8abf63c 1671 dsched_exit_thread(td);
37af14fe 1672 lwkt_deschedule_self(td);
e56e4dea 1673 lwkt_remove_tdallq(td);
c070746a
MD
1674 if (td->td_flags & TDF_ALLOCATED_THREAD)
1675 gd->gd_freetd = td;
99df837e
MD
1676 cpu_thread_exit();
1677}
1678
e56e4dea
MD
1679void
1680lwkt_remove_tdallq(thread_t td)
1681{
1682 KKASSERT(td->td_gd == mycpu);
1683 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1684}
1685
9cf43f91
MD
1686/*
1687 * Code reduction and branch prediction improvements. Call/return
1688 * overhead on modern cpus often degenerates into 0 cycles due to
1689 * the cpu's branch prediction hardware and return pc cache. We
1690 * can take advantage of this by not inlining medium-complexity
1691 * functions and we can also reduce the branch prediction impact
1692 * by collapsing perfectly predictable branches into a single
1693 * procedure instead of duplicating it.
1694 *
1695 * Is any of this noticeable? Probably not, so I'll take the
1696 * smaller code size.
1697 */
1698void
b6468f56 1699crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1700{
b6468f56 1701 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1702}
1703
2d93b37a
MD
1704void
1705crit_panic(void)
1706{
1707 thread_t td = curthread;
850634cc 1708 int lcrit = td->td_critcount;
2d93b37a 1709
850634cc
AH
1710 td->td_critcount = 0;
1711 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1712 /* NOT REACHED */
2d93b37a
MD
1713}
1714
d165e668
MD
1715#ifdef SMP
1716
bd8015ca
MD
1717/*
1718 * Called from debugger/panic on cpus which have been stopped. We must still
1719 * process the IPIQ while stopped, even if we were stopped while in a critical
1720 * section (XXX).
1721 *
1722 * If we are dumping also try to process any pending interrupts. This may
1723 * or may not work depending on the state of the cpu at the point it was
1724 * stopped.
1725 */
1726void
1727lwkt_smp_stopped(void)
1728{
1729 globaldata_t gd = mycpu;
1730
1731 crit_enter_gd(gd);
1732 if (dumping) {
1733 lwkt_process_ipiq();
1734 splz();
1735 } else {
1736 lwkt_process_ipiq();
1737 }
1738 crit_exit_gd(gd);
1739}
1740
d165e668 1741#endif