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