hammer - Describe live_dedup_cache_size sysctl
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
3b998fa9 2 * Copyright (c) 2003-2010 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
60f60350 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
60f60350 20 *
8c10bfcf
MD
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
b37f18d6 47#include <sys/kinfo.h>
8ad65e08 48#include <sys/queue.h>
7d0bac62 49#include <sys/sysctl.h>
99df837e 50#include <sys/kthread.h>
f1d1c3fa 51#include <machine/cpu.h>
99df837e 52#include <sys/lock.h>
f6bf3af1 53#include <sys/caps.h>
9d265729 54#include <sys/spinlock.h>
57aa743c 55#include <sys/ktr.h>
9d265729
MD
56
57#include <sys/thread2.h>
58#include <sys/spinlock2.h>
684a93c4 59#include <sys/mplock2.h>
f1d1c3fa 60
8c72e3d5
AH
61#include <sys/dsched.h>
62
7d0bac62
MD
63#include <vm/vm.h>
64#include <vm/vm_param.h>
65#include <vm/vm_kern.h>
66#include <vm/vm_object.h>
67#include <vm/vm_page.h>
68#include <vm/vm_map.h>
69#include <vm/vm_pager.h>
70#include <vm/vm_extern.h>
7d0bac62 71
99df837e 72#include <machine/stdarg.h>
96728c05 73#include <machine/smp.h>
99df837e 74
d850923c
AE
75#if !defined(KTR_CTXSW)
76#define KTR_CTXSW KTR_ALL
77#endif
78KTR_INFO_MASTER(ctxsw);
a1f0fb66
AE
79KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
80 sizeof(int) + sizeof(struct thread *));
81KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
82 sizeof(int) + sizeof(struct thread *));
83KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
84 sizeof (struct thread *) + sizeof(char *));
85KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
1541028a 86
40aaf5fc
NT
87static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
88
0f7a3396
MD
89#ifdef INVARIANTS
90static int panic_on_cscount = 0;
91#endif
05220613
MD
92static __int64_t switch_count = 0;
93static __int64_t preempt_hit = 0;
94static __int64_t preempt_miss = 0;
95static __int64_t preempt_weird = 0;
f64b567c 96static __int64_t token_contention_count __debugvar = 0;
fb0f29c4 97static int lwkt_use_spin_port;
40aaf5fc 98static struct objcache *thread_cache;
05220613 99
88ebb169 100#ifdef SMP
e381e77c 101static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
88ebb169 102#endif
f9235b6d 103static void lwkt_fairq_accumulate(globaldata_t gd, thread_t td);
e381e77c 104
0855a2af
JG
105extern void cpu_heavy_restore(void);
106extern void cpu_lwkt_restore(void);
107extern void cpu_kthread_restore(void);
108extern void cpu_idle_restore(void);
109
fb0f29c4
MD
110/*
111 * We can make all thread ports use the spin backend instead of the thread
112 * backend. This should only be set to debug the spin backend.
113 */
114TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
115
0f7a3396 116#ifdef INVARIANTS
0c52fa62
SG
117SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
118 "Panic if attempting to switch lwkt's while mastering cpusync");
0f7a3396 119#endif
0c52fa62
SG
120SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
121 "Number of switched threads");
9733f757 122SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
0c52fa62 123 "Successful preemption events");
9733f757 124SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
0c52fa62
SG
125 "Failed preemption events");
126SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
127 "Number of preempted threads.");
38717797
HP
128#ifdef INVARIANTS
129SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
130 &token_contention_count, 0, "spinning due to token contention");
38717797 131#endif
f9235b6d 132static int fairq_enable = 1;
2a418930
MD
133SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
134 &fairq_enable, 0, "Turn on fairq priority accumulators");
135static int lwkt_spin_loops = 10;
136SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
137 &lwkt_spin_loops, 0, "");
138static int lwkt_spin_delay = 1;
139SYSCTL_INT(_lwkt, OID_AUTO, spin_delay, CTLFLAG_RW,
140 &lwkt_spin_delay, 0, "Scheduler spin delay in microseconds 0=auto");
141static int lwkt_spin_method = 1;
142SYSCTL_INT(_lwkt, OID_AUTO, spin_method, CTLFLAG_RW,
143 &lwkt_spin_method, 0, "LWKT scheduler behavior when contended");
fbc024e4 144static int preempt_enable = 1;
2a418930
MD
145SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
146 &preempt_enable, 0, "Enable preemption");
fbc024e4 147
2a418930
MD
148static __cachealign int lwkt_cseq_rindex;
149static __cachealign int lwkt_cseq_windex;
05220613 150
4b5f931b
MD
151/*
152 * These helper procedures handle the runq, they can only be called from
153 * within a critical section.
75cdbe6c
MD
154 *
155 * WARNING! Prior to SMP being brought up it is possible to enqueue and
156 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
157 * instead of 'mycpu' when referencing the globaldata structure. Once
158 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 159 */
f1d1c3fa
MD
160static __inline
161void
162_lwkt_dequeue(thread_t td)
163{
164 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 165 struct globaldata *gd = td->td_gd;
4b5f931b 166
f1d1c3fa 167 td->td_flags &= ~TDF_RUNQ;
f9235b6d
MD
168 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
169 gd->gd_fairq_total_pri -= td->td_pri;
170 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
2a418930 171 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
172 }
173}
174
f9235b6d
MD
175/*
176 * Priority enqueue.
177 *
178 * NOTE: There are a limited number of lwkt threads runnable since user
179 * processes only schedule one at a time per cpu.
180 */
f1d1c3fa
MD
181static __inline
182void
183_lwkt_enqueue(thread_t td)
184{
f9235b6d
MD
185 thread_t xtd;
186
7f5d7ed7 187 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 188 struct globaldata *gd = td->td_gd;
4b5f931b 189
f1d1c3fa 190 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
191 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
192 if (xtd == NULL) {
193 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
2a418930 194 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
f9235b6d
MD
195 } else {
196 while (xtd && xtd->td_pri > td->td_pri)
197 xtd = TAILQ_NEXT(xtd, td_threadq);
198 if (xtd)
199 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
200 else
201 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
202 }
203 gd->gd_fairq_total_pri += td->td_pri;
f1d1c3fa
MD
204 }
205}
8ad65e08 206
40aaf5fc
NT
207static __boolean_t
208_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
209{
210 struct thread *td = (struct thread *)obj;
211
212 td->td_kstack = NULL;
213 td->td_kstack_size = 0;
214 td->td_flags = TDF_ALLOCATED_THREAD;
215 return (1);
216}
217
218static void
219_lwkt_thread_dtor(void *obj, void *privdata)
220{
221 struct thread *td = (struct thread *)obj;
222
223 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
224 ("_lwkt_thread_dtor: not allocated from objcache"));
225 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
226 td->td_kstack_size > 0,
227 ("_lwkt_thread_dtor: corrupted stack"));
228 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
229}
230
231/*
232 * Initialize the lwkt s/system.
233 */
234void
235lwkt_init(void)
236{
237 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
238 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
239 NULL, CACHE_NTHREADS/2,
240 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
241}
242
37af14fe
MD
243/*
244 * Schedule a thread to run. As the current thread we can always safely
245 * schedule ourselves, and a shortcut procedure is provided for that
246 * function.
247 *
248 * (non-blocking, self contained on a per cpu basis)
249 */
250void
251lwkt_schedule_self(thread_t td)
252{
253 crit_enter_quick(td);
f9235b6d
MD
254 KASSERT(td != &td->td_gd->gd_idlethread,
255 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 256 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 257 _lwkt_enqueue(td);
37af14fe
MD
258 crit_exit_quick(td);
259}
260
261/*
262 * Deschedule a thread.
263 *
264 * (non-blocking, self contained on a per cpu basis)
265 */
266void
267lwkt_deschedule_self(thread_t td)
268{
269 crit_enter_quick(td);
37af14fe
MD
270 _lwkt_dequeue(td);
271 crit_exit_quick(td);
272}
273
8ad65e08
MD
274/*
275 * LWKTs operate on a per-cpu basis
276 *
73e4f7b9 277 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
278 */
279void
280lwkt_gdinit(struct globaldata *gd)
281{
f9235b6d 282 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 283 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
284}
285
7d0bac62
MD
286/*
287 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
288 * or LWKT start address before it can be scheduled. If the target cpu is
289 * -1 the thread will be created on the current cpu.
0cfcada1
MD
290 *
291 * If you intend to create a thread without a process context this function
292 * does everything except load the startup and switcher function.
7d0bac62
MD
293 */
294thread_t
d3d32139 295lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 296{
c070746a 297 globaldata_t gd = mycpu;
99df837e 298 void *stack;
7d0bac62 299
c070746a
MD
300 /*
301 * If static thread storage is not supplied allocate a thread. Reuse
302 * a cached free thread if possible. gd_freetd is used to keep an exiting
303 * thread intact through the exit.
304 */
ef0fdad1 305 if (td == NULL) {
cf709dd2
MD
306 crit_enter_gd(gd);
307 if ((td = gd->gd_freetd) != NULL) {
308 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
309 TDF_RUNQ)) == 0);
c070746a 310 gd->gd_freetd = NULL;
cf709dd2 311 } else {
c070746a 312 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
313 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
314 TDF_RUNQ)) == 0);
315 }
316 crit_exit_gd(gd);
40aaf5fc
NT
317 KASSERT((td->td_flags &
318 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
319 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
320 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 321 }
c070746a
MD
322
323 /*
324 * Try to reuse cached stack.
325 */
f470d0c8
MD
326 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
327 if (flags & TDF_ALLOCATED_STACK) {
e4846942 328 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
329 stack = NULL;
330 }
331 }
332 if (stack == NULL) {
e40cfbd7 333 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 334 flags |= TDF_ALLOCATED_STACK;
99df837e 335 }
75cdbe6c 336 if (cpu < 0)
c070746a 337 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 338 else
f470d0c8 339 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 340 return(td);
7d0bac62
MD
341}
342
343/*
344 * Initialize a preexisting thread structure. This function is used by
345 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
346 *
f8c3996b
MD
347 * All threads start out in a critical section at a priority of
348 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
349 * appropriate. This function may send an IPI message when the
350 * requested cpu is not the current cpu and consequently gd_tdallq may
351 * not be initialized synchronously from the point of view of the originating
352 * cpu.
353 *
354 * NOTE! we have to be careful in regards to creating threads for other cpus
355 * if SMP has not yet been activated.
7d0bac62 356 */
41a01a4d
MD
357#ifdef SMP
358
75cdbe6c
MD
359static void
360lwkt_init_thread_remote(void *arg)
361{
362 thread_t td = arg;
363
52eedfb5
MD
364 /*
365 * Protected by critical section held by IPI dispatch
366 */
75cdbe6c
MD
367 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
368}
369
41a01a4d
MD
370#endif
371
fdce8919
MD
372/*
373 * lwkt core thread structural initialization.
374 *
375 * NOTE: All threads are initialized as mpsafe threads.
376 */
7d0bac62 377void
f470d0c8
MD
378lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
379 struct globaldata *gd)
7d0bac62 380{
37af14fe
MD
381 globaldata_t mygd = mycpu;
382
99df837e
MD
383 bzero(td, sizeof(struct thread));
384 td->td_kstack = stack;
f470d0c8 385 td->td_kstack_size = stksize;
d3d32139 386 td->td_flags = flags;
26a0694b 387 td->td_gd = gd;
f9235b6d
MD
388 td->td_pri = TDPRI_KERN_DAEMON;
389 td->td_critcount = 1;
3b998fa9 390 td->td_toks_stop = &td->td_toks_base;
fb0f29c4
MD
391 if (lwkt_use_spin_port)
392 lwkt_initport_spin(&td->td_msgport);
393 else
394 lwkt_initport_thread(&td->td_msgport, td);
99df837e 395 pmap_init_thread(td);
0f7a3396 396#ifdef SMP
5d21b981
MD
397 /*
398 * Normally initializing a thread for a remote cpu requires sending an
399 * IPI. However, the idlethread is setup before the other cpus are
400 * activated so we have to treat it as a special case. XXX manipulation
401 * of gd_tdallq requires the BGL.
402 */
403 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 404 crit_enter_gd(mygd);
75cdbe6c 405 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 406 crit_exit_gd(mygd);
75cdbe6c 407 } else {
2db3b277 408 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 409 }
0f7a3396 410#else
37af14fe 411 crit_enter_gd(mygd);
0f7a3396 412 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 413 crit_exit_gd(mygd);
0f7a3396 414#endif
8c72e3d5
AH
415
416 dsched_new_thread(td);
73e4f7b9
MD
417}
418
419void
420lwkt_set_comm(thread_t td, const char *ctl, ...)
421{
e2565a42 422 __va_list va;
73e4f7b9 423
e2565a42 424 __va_start(va, ctl);
379210cb 425 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 426 __va_end(va);
e7c0dbba 427 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
428}
429
99df837e 430void
73e4f7b9 431lwkt_hold(thread_t td)
99df837e 432{
73e4f7b9
MD
433 ++td->td_refs;
434}
435
436void
437lwkt_rele(thread_t td)
438{
439 KKASSERT(td->td_refs > 0);
440 --td->td_refs;
441}
442
443void
444lwkt_wait_free(thread_t td)
445{
446 while (td->td_refs)
377d4740 447 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
448}
449
450void
451lwkt_free_thread(thread_t td)
452{
cf709dd2 453 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|TDF_RUNQ)) == 0);
40aaf5fc
NT
454 if (td->td_flags & TDF_ALLOCATED_THREAD) {
455 objcache_put(thread_cache, td);
456 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
457 /* client-allocated struct with internally allocated stack */
458 KASSERT(td->td_kstack && td->td_kstack_size > 0,
459 ("lwkt_free_thread: corrupted stack"));
460 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
461 td->td_kstack = NULL;
462 td->td_kstack_size = 0;
99df837e 463 }
e7c0dbba 464 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
465}
466
467
8ad65e08
MD
468/*
469 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
470 * switch to the idlethread. Switching must occur within a critical
471 * section to avoid races with the scheduling queue.
472 *
473 * We always have full control over our cpu's run queue. Other cpus
474 * that wish to manipulate our queue must use the cpu_*msg() calls to
475 * talk to our cpu, so a critical section is all that is needed and
476 * the result is very, very fast thread switching.
477 *
96728c05
MD
478 * The LWKT scheduler uses a fixed priority model and round-robins at
479 * each priority level. User process scheduling is a totally
480 * different beast and LWKT priorities should not be confused with
481 * user process priorities.
f1d1c3fa 482 *
3933a3ab
MD
483 * Note that the td_switch() function cannot do anything that requires
484 * the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
485 * the target thread (not the current thread). It's nice to have a scheduler
486 * that does not need the MP lock to work because it allows us to do some
487 * really cool high-performance MP lock optimizations.
69d78e99
MD
488 *
489 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
490 * is not called by the current thread in the preemption case, only when
491 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
492 */
493void
494lwkt_switch(void)
495{
37af14fe
MD
496 globaldata_t gd = mycpu;
497 thread_t td = gd->gd_curthread;
8ad65e08 498 thread_t ntd;
f9235b6d 499 thread_t xtd;
2a418930
MD
500 int spinning = lwkt_spin_loops; /* loops before HLTing */
501 int reqflags;
502 int cseq;
0f0466c0 503 int oseq;
8ad65e08 504
46a3f46d 505 /*
27e88a6e
MD
506 * Switching from within a 'fast' (non thread switched) interrupt or IPI
507 * is illegal. However, we may have to do it anyway if we hit a fatal
508 * kernel trap or we have paniced.
509 *
510 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 511 */
27e88a6e
MD
512 if (gd->gd_intr_nesting_level) {
513 int savegdnest;
514 int savegdtrap;
515
5fddbda2 516 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22
MD
517 panic("lwkt_switch: Attempt to switch from a "
518 "a fast interrupt, ipi, or hard code section, "
519 "td %p\n",
520 td);
27e88a6e
MD
521 } else {
522 savegdnest = gd->gd_intr_nesting_level;
523 savegdtrap = gd->gd_trap_nesting_level;
524 gd->gd_intr_nesting_level = 0;
525 gd->gd_trap_nesting_level = 0;
a7422615
MD
526 if ((td->td_flags & TDF_PANICWARN) == 0) {
527 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
528 kprintf("Warning: thread switch from interrupt, IPI, "
529 "or hard code section.\n"
a7422615 530 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 531 print_backtrace(-1);
a7422615 532 }
27e88a6e
MD
533 lwkt_switch();
534 gd->gd_intr_nesting_level = savegdnest;
535 gd->gd_trap_nesting_level = savegdtrap;
536 return;
537 }
96728c05 538 }
ef0fdad1 539
cb973d15
MD
540 /*
541 * Passive release (used to transition from user to kernel mode
542 * when we block or switch rather then when we enter the kernel).
543 * This function is NOT called if we are switching into a preemption
544 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
545 * our current process designation (if we have one) and become a true
546 * LWKT thread, and may also hand the current process designation to
547 * another process and schedule thread.
cb973d15
MD
548 */
549 if (td->td_release)
550 td->td_release(td);
551
37af14fe 552 crit_enter_gd(gd);
3b998fa9 553 if (TD_TOKS_HELD(td))
9d265729
MD
554 lwkt_relalltokens(td);
555
556 /*
b02926de
MD
557 * We had better not be holding any spin locks, but don't get into an
558 * endless panic loop.
9d265729 559 */
d666840a
MD
560 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
561 ("lwkt_switch: still holding %d exclusive spinlocks!",
562 gd->gd_spinlocks_wr));
9d265729 563
8a8d5d85
MD
564
565#ifdef SMP
0f7a3396
MD
566#ifdef INVARIANTS
567 if (td->td_cscount) {
6ea70f76 568 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
569 td);
570 if (panic_on_cscount)
571 panic("switching while mastering cpusync");
572 }
573#endif
8a8d5d85 574#endif
f9235b6d
MD
575
576 /*
577 * If we had preempted another thread on this cpu, resume the preempted
578 * thread. This occurs transparently, whether the preempted thread
579 * was scheduled or not (it may have been preempted after descheduling
580 * itself).
581 *
582 * We have to setup the MP lock for the original thread after backing
583 * out the adjustment that was made to curthread when the original
584 * was preempted.
585 */
99df837e 586 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
587 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
588 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
589
590 /*
b9eb1c19
MD
591 * The interrupt may have woken a thread up, we need to properly
592 * set the reschedule flag if the originally interrupted thread is
593 * at a lower priority.
8ec60c3f 594 */
f9235b6d
MD
595 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
596 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
8ec60c3f 597 need_lwkt_resched();
f9235b6d 598 }
8a8d5d85 599 /* YYY release mp lock on switchback if original doesn't need it */
f9235b6d
MD
600 goto havethread_preempted;
601 }
602
603 /*
604 * Implement round-robin fairq with priority insertion. The priority
605 * insertion is handled by _lwkt_enqueue()
606 *
607 * We have to adjust the MP lock for the target thread. If we
608 * need the MP lock and cannot obtain it we try to locate a
609 * thread that does not need the MP lock. If we cannot, we spin
610 * instead of HLT.
611 *
612 * A similar issue exists for the tokens held by the target thread.
613 * If we cannot obtain ownership of the tokens we cannot immediately
614 * schedule the thread.
615 */
616 for (;;) {
2a418930
MD
617 /*
618 * Clear RQF_AST_LWKT_RESCHED (we handle the reschedule request)
619 * and set RQF_WAKEUP (prevent unnecessary IPIs from being
620 * received).
621 */
622 for (;;) {
623 reqflags = gd->gd_reqflags;
624 if (atomic_cmpset_int(&gd->gd_reqflags, reqflags,
625 (reqflags & ~RQF_AST_LWKT_RESCHED) |
626 RQF_WAKEUP)) {
627 break;
628 }
629 }
f9235b6d 630
4b5f931b 631 /*
2a418930
MD
632 * Hotpath - pull the head of the run queue and attempt to schedule
633 * it. Fairq exhaustion moves the task to the end of the list. If
634 * no threads are runnable we switch to the idle thread.
41a01a4d 635 */
2a418930
MD
636 for (;;) {
637 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
638
639 if (ntd == NULL) {
640 /*
641 * Runq is empty, switch to idle and clear RQF_WAKEUP
642 * to allow it to halt.
643 */
644 ntd = &gd->gd_idlethread;
6f207a2c 645#ifdef SMP
2a418930 646 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
b5d16701 647 ASSERT_NO_TOKENS_HELD(ntd);
6f207a2c 648#endif
2a418930
MD
649 cpu_time.cp_msg[0] = 0;
650 cpu_time.cp_stallpc = 0;
651 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP);
652 goto haveidle;
653 }
654
655 if (ntd->td_fairq_accum >= 0)
656 break;
657
658 splz_check();
659 lwkt_fairq_accumulate(gd, ntd);
660 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
661 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
f9235b6d 662 }
41a01a4d 663
8ec60c3f 664 /*
2a418930
MD
665 * Hotpath - schedule ntd. Leaves RQF_WAKEUP set to prevent
666 * unwanted decontention IPIs.
6f207a2c
MD
667 *
668 * NOTE: For UP there is no mplock and lwkt_getalltokens()
669 * always succeeds.
8ec60c3f 670 */
2a418930 671 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
f9235b6d 672 goto havethread;
f9235b6d 673
f9235b6d 674 /*
2a418930
MD
675 * Coldpath (SMP only since tokens always succeed on UP)
676 *
677 * We had some contention on the thread we wanted to schedule.
678 * What we do now is try to find a thread that we can schedule
679 * in its stead until decontention reschedules on our cpu.
680 *
681 * The coldpath scan does NOT rearrange threads in the run list
682 * and it also ignores the accumulator.
683 *
684 * We do not immediately schedule a user priority thread, instead
685 * we record it in xtd and continue looking for kernel threads.
686 * A cpu can only have one user priority thread (normally) so just
687 * record the first one.
688 *
689 * NOTE: This scan will also include threads whos fairq's were
690 * accumulated in the first loop.
f9235b6d 691 */
2a418930
MD
692 ++token_contention_count;
693 xtd = NULL;
694 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
41a01a4d 695 /*
2a418930
MD
696 * Try to switch to this thread. If the thread is running at
697 * user priority we clear WAKEUP to allow decontention IPIs
698 * (since this thread is simply running until the one we wanted
699 * decontends), and we make sure that LWKT_RESCHED is not set.
df6b8ba0 700 *
2a418930
MD
701 * Otherwise for kernel threads we leave WAKEUP set to avoid
702 * unnecessary decontention IPIs.
41a01a4d 703 */
2a418930
MD
704 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
705 if (xtd == NULL)
706 xtd = ntd;
707 continue;
f9235b6d 708 }
a453459d 709
f9235b6d 710 /*
2a418930
MD
711 * Do not let the fairq get too negative. Even though we are
712 * ignoring it atm once the scheduler decontends a very negative
713 * thread will get moved to the end of the queue.
f9235b6d 714 */
2a418930
MD
715 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd)) {
716 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd))
717 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd);
718 goto havethread;
8a8d5d85 719 }
f9235b6d 720
df6b8ba0 721 /*
2a418930 722 * Well fubar, this thread is contended as well, loop
df6b8ba0 723 */
2a418930
MD
724 /* */
725 }
726
727 /*
728 * We exhausted the run list but we may have recorded a user
729 * thread to try. We have three choices based on
730 * lwkt.decontention_method.
731 *
732 * (0) Atomically clear RQF_WAKEUP in order to receive decontention
733 * IPIs (to interrupt the user process) and test
734 * RQF_AST_LWKT_RESCHED at the same time.
735 *
736 * This results in significant decontention IPI traffic but may
737 * be more responsive.
738 *
739 * (1) Leave RQF_WAKEUP set so we do not receive a decontention IPI.
740 * An automatic LWKT reschedule will occur on the next hardclock
741 * (typically 100hz).
742 *
743 * This results in no decontention IPI traffic but may be less
744 * responsive. This is the default.
745 *
746 * (2) Refuse to schedule the user process at this time.
747 *
748 * This is highly experimental and should not be used under
749 * normal circumstances. This can cause a user process to
750 * get starved out in situations where kernel threads are
751 * fighting each other for tokens.
752 */
753 if (xtd) {
754 ntd = xtd;
755
756 switch(lwkt_spin_method) {
757 case 0:
758 for (;;) {
759 reqflags = gd->gd_reqflags;
760 if (atomic_cmpset_int(&gd->gd_reqflags,
761 reqflags,
762 reqflags & ~RQF_WAKEUP)) {
763 break;
764 }
765 }
766 break;
767 case 1:
768 reqflags = gd->gd_reqflags;
769 break;
770 default:
771 goto skip;
772 break;
773 }
774 if ((reqflags & RQF_AST_LWKT_RESCHED) == 0 &&
b5d16701 775 (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
f9235b6d 776 ) {
2a418930
MD
777 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd))
778 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd);
779 goto havethread;
df6b8ba0 780 }
9ac1ee6e 781
2a418930 782skip:
9ac1ee6e 783 /*
2a418930
MD
784 * Make sure RQF_WAKEUP is set if we failed to schedule the
785 * user thread to prevent the idle thread from halting.
9ac1ee6e 786 */
2a418930
MD
787 atomic_set_int(&gd->gd_reqflags, RQF_WAKEUP);
788 }
789
790 /*
791 * We exhausted the run list, meaning that all runnable threads
792 * are contended.
793 */
794 cpu_pause();
795 ntd = &gd->gd_idlethread;
796#ifdef SMP
797 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
798 ASSERT_NO_TOKENS_HELD(ntd);
799 /* contention case, do not clear contention mask */
800#endif
801
802 /*
803 * Ok, we might want to spin a few times as some tokens are held for
804 * very short periods of time and IPI overhead is 1uS or worse
805 * (meaning it is usually better to spin). Regardless we have to
806 * call splz_check() to be sure to service any interrupts blocked
807 * by our critical section, otherwise we could livelock e.g. IPIs.
808 *
809 * The IPI mechanic is really a last resort. In nearly all other
810 * cases RQF_WAKEUP is left set to prevent decontention IPIs.
811 *
812 * When we decide not to spin we clear RQF_WAKEUP and switch to
813 * the idle thread. Clearing RQF_WEAKEUP allows the idle thread
814 * to halt and decontended tokens will issue an IPI to us. The
815 * idle thread will check for pending reschedules already set
816 * (RQF_AST_LWKT_RESCHED) before actually halting so we don't have
817 * to here.
818 */
819 if (spinning <= 0) {
820 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP);
821 goto haveidle;
4b5f931b 822 }
2a418930 823 --spinning;
c5724852
MD
824
825 /*
2a418930
MD
826 * When spinning a delay is required both to avoid livelocks from
827 * token order reversals (a thread may be trying to acquire multiple
828 * tokens), and also to reduce cpu cache management traffic.
829 *
830 * In order to scale to a large number of CPUs we use a time slot
831 * resequencer to force contending cpus into non-contending
832 * time-slots. The scheduler may still contend with the lock holder
833 * but will not (generally) contend with all the other cpus trying
834 * trying to get the same token.
835 *
836 * The resequencer uses a FIFO counter mechanic. The owner of the
837 * rindex at the head of the FIFO is allowed to pull itself off
838 * the FIFO and fetchadd is used to enter into the FIFO. This bit
839 * of code is VERY cache friendly and forces all spinning schedulers
840 * into their own time slots.
c5724852 841 *
2a418930
MD
842 * This code has been tested to 48-cpus and caps the cache
843 * contention load at ~1uS intervals regardless of the number of
844 * cpus. Scaling beyond 64 cpus might require additional smarts
845 * (such as separate FIFOs for specific token cases).
846 *
847 * WARNING! We can't call splz_check() or anything else here as
848 * it could cause a deadlock.
c5724852 849 */
2a418930 850 cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
0f0466c0
MD
851 while ((oseq = lwkt_cseq_rindex) != cseq) {
852 cpu_ccfence();
8b402283 853#if !defined(_KERNEL_VIRTUAL)
0f0466c0
MD
854 if (cpu_mi_feature & CPU_MI_MONITOR) {
855 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
8b402283
SW
856 } else
857#endif
858 {
0f0466c0
MD
859 DELAY(1);
860 cpu_lfence();
861 }
2a418930
MD
862 }
863 cseq = lwkt_spin_delay; /* don't trust the system operator */
864 cpu_ccfence();
865 if (cseq < 1)
866 cseq = 1;
867 if (cseq > 1000)
868 cseq = 1000;
869 DELAY(cseq);
870 atomic_add_int(&lwkt_cseq_rindex, 1);
c5724852 871 splz_check();
2a418930 872 /* highest level for(;;) loop */
f1d1c3fa 873 }
8a8d5d85 874
2a418930 875havethread:
8a8d5d85 876 /*
f9235b6d
MD
877 * We must always decrement td_fairq_accum on non-idle threads just
878 * in case a thread never gets a tick due to being in a continuous
2a418930 879 * critical section. The page-zeroing code does this, for example.
f9235b6d
MD
880 *
881 * If the thread we came up with is a higher or equal priority verses
882 * the thread at the head of the queue we move our thread to the
883 * front. This way we can always check the front of the queue.
be71787b
MD
884 *
885 * Clear gd_idle_repeat when doing a normal switch to a non-idle
886 * thread.
f9235b6d 887 */
f9235b6d
MD
888 ++gd->gd_cnt.v_swtch;
889 --ntd->td_fairq_accum;
9ac1ee6e 890 ntd->td_wmesg = NULL;
f9235b6d
MD
891 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
892 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
893 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
894 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
895 }
be71787b 896 gd->gd_idle_repeat = 0;
2a418930 897
f9235b6d 898havethread_preempted:
f9235b6d
MD
899 /*
900 * If the new target does not need the MP lock and we are holding it,
901 * release the MP lock. If the new target requires the MP lock we have
902 * already acquired it for the target.
8a8d5d85 903 */
2a418930 904 ;
f9235b6d
MD
905haveidle:
906 KASSERT(ntd->td_critcount,
b5d16701
MD
907 ("priority problem in lwkt_switch %d %d",
908 td->td_critcount, ntd->td_critcount));
909
94f6d86e
MD
910 if (td != ntd) {
911 ++switch_count;
a1f0fb66 912 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
f1d1c3fa 913 td->td_switch(ntd);
94f6d86e 914 }
37af14fe
MD
915 /* NOTE: current cpu may have changed after switch */
916 crit_exit_quick(td);
8ad65e08
MD
917}
918
b68b7282 919/*
96728c05
MD
920 * Request that the target thread preempt the current thread. Preemption
921 * only works under a specific set of conditions:
b68b7282 922 *
96728c05
MD
923 * - We are not preempting ourselves
924 * - The target thread is owned by the current cpu
925 * - We are not currently being preempted
926 * - The target is not currently being preempted
d3d1cbc8
MD
927 * - We are not holding any spin locks
928 * - The target thread is not holding any tokens
96728c05
MD
929 * - We are able to satisfy the target's MP lock requirements (if any).
930 *
931 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
932 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 933 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
934 * to determine whether preemption is possible (aka usually just the crit
935 * priority of lwkt_schedule() itself).
b68b7282 936 *
26a0694b
MD
937 * XXX at the moment we run the target thread in a critical section during
938 * the preemption in order to prevent the target from taking interrupts
939 * that *WE* can't. Preemption is strictly limited to interrupt threads
940 * and interrupt-like threads, outside of a critical section, and the
941 * preempted source thread will be resumed the instant the target blocks
942 * whether or not the source is scheduled (i.e. preemption is supposed to
943 * be as transparent as possible).
b68b7282
MD
944 */
945void
f9235b6d 946lwkt_preempt(thread_t ntd, int critcount)
b68b7282 947{
46a3f46d 948 struct globaldata *gd = mycpu;
0a3f9b47 949 thread_t td;
2d910aaf 950 int save_gd_intr_nesting_level;
b68b7282 951
26a0694b 952 /*
96728c05
MD
953 * The caller has put us in a critical section. We can only preempt
954 * if the caller of the caller was not in a critical section (basically
f9235b6d 955 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 956 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 957 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
958 *
959 * YYY The target thread must be in a critical section (else it must
960 * inherit our critical section? I dunno yet).
41a01a4d 961 *
0a3f9b47 962 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b 963 */
f9235b6d 964 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 965
fbc024e4
MD
966 if (preempt_enable == 0) {
967 ++preempt_miss;
968 return;
969 }
970
0a3f9b47 971 td = gd->gd_curthread;
f9235b6d 972 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
973 ++preempt_miss;
974 return;
975 }
f9235b6d 976 if (td->td_critcount > critcount) {
96728c05 977 ++preempt_miss;
8ec60c3f 978 need_lwkt_resched();
96728c05
MD
979 return;
980 }
981#ifdef SMP
46a3f46d 982 if (ntd->td_gd != gd) {
96728c05 983 ++preempt_miss;
8ec60c3f 984 need_lwkt_resched();
96728c05
MD
985 return;
986 }
987#endif
41a01a4d 988 /*
77912481
MD
989 * We don't have to check spinlocks here as they will also bump
990 * td_critcount.
d3d1cbc8
MD
991 *
992 * Do not try to preempt if the target thread is holding any tokens.
993 * We could try to acquire the tokens but this case is so rare there
994 * is no need to support it.
41a01a4d 995 */
77912481
MD
996 KKASSERT(gd->gd_spinlocks_wr == 0);
997
3b998fa9 998 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8
MD
999 ++preempt_miss;
1000 need_lwkt_resched();
1001 return;
1002 }
26a0694b
MD
1003 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1004 ++preempt_weird;
8ec60c3f 1005 need_lwkt_resched();
26a0694b
MD
1006 return;
1007 }
1008 if (ntd->td_preempted) {
4b5f931b 1009 ++preempt_hit;
8ec60c3f 1010 need_lwkt_resched();
26a0694b 1011 return;
b68b7282 1012 }
26a0694b 1013
8ec60c3f
MD
1014 /*
1015 * Since we are able to preempt the current thread, there is no need to
1016 * call need_lwkt_resched().
2d910aaf
MD
1017 *
1018 * We must temporarily clear gd_intr_nesting_level around the switch
1019 * since switchouts from the target thread are allowed (they will just
1020 * return to our thread), and since the target thread has its own stack.
8ec60c3f 1021 */
26a0694b
MD
1022 ++preempt_hit;
1023 ntd->td_preempted = td;
1024 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1025 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1026 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1027 gd->gd_intr_nesting_level = 0;
26a0694b 1028 td->td_switch(ntd);
2d910aaf 1029 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1030
26a0694b
MD
1031 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1032 ntd->td_preempted = NULL;
1033 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1034}
1035
f1d1c3fa 1036/*
faaeffac 1037 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1038 * This will work inside a critical section but not inside a hard code
1039 * section.
ef0fdad1 1040 *
f1d1c3fa
MD
1041 * (self contained on a per cpu basis)
1042 */
1043void
faaeffac 1044splz_check(void)
f1d1c3fa 1045{
7966cb69
MD
1046 globaldata_t gd = mycpu;
1047 thread_t td = gd->gd_curthread;
ef0fdad1 1048
4a28fe22
MD
1049 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1050 gd->gd_intr_nesting_level == 0 &&
1051 td->td_nest_count < 2)
1052 {
f1d1c3fa 1053 splz();
4a28fe22
MD
1054 }
1055}
1056
1057/*
1058 * This version is integrated into crit_exit, reqflags has already
1059 * been tested but td_critcount has not.
1060 *
1061 * We only want to execute the splz() on the 1->0 transition of
1062 * critcount and not in a hard code section or if too deeply nested.
1063 */
1064void
1065lwkt_maybe_splz(thread_t td)
1066{
1067 globaldata_t gd = td->td_gd;
1068
1069 if (td->td_critcount == 0 &&
1070 gd->gd_intr_nesting_level == 0 &&
1071 td->td_nest_count < 2)
1072 {
1073 splz();
1074 }
f1d1c3fa
MD
1075}
1076
8ad65e08 1077/*
f9235b6d
MD
1078 * This function is used to negotiate a passive release of the current
1079 * process/lwp designation with the user scheduler, allowing the user
1080 * scheduler to schedule another user thread. The related kernel thread
1081 * (curthread) continues running in the released state.
8ad65e08
MD
1082 */
1083void
f9235b6d 1084lwkt_passive_release(struct thread *td)
8ad65e08 1085{
f9235b6d
MD
1086 struct lwp *lp = td->td_lwp;
1087
1088 td->td_release = NULL;
1089 lwkt_setpri_self(TDPRI_KERN_USER);
1090 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1091}
1092
f9235b6d 1093
3824f392 1094/*
f9235b6d
MD
1095 * This implements a normal yield. This routine is virtually a nop if
1096 * there is nothing to yield to but it will always run any pending interrupts
1097 * if called from a critical section.
1098 *
1099 * This yield is designed for kernel threads without a user context.
1100 *
1101 * (self contained on a per cpu basis)
3824f392
MD
1102 */
1103void
f9235b6d 1104lwkt_yield(void)
3824f392 1105{
f9235b6d
MD
1106 globaldata_t gd = mycpu;
1107 thread_t td = gd->gd_curthread;
1108 thread_t xtd;
3824f392 1109
f9235b6d
MD
1110 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1111 splz();
1112 if (td->td_fairq_accum < 0) {
1113 lwkt_schedule_self(curthread);
1114 lwkt_switch();
1115 } else {
1116 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1117 if (xtd && xtd->td_pri > td->td_pri) {
1118 lwkt_schedule_self(curthread);
1119 lwkt_switch();
1120 }
1121 }
3824f392
MD
1122}
1123
1124/*
f9235b6d
MD
1125 * This yield is designed for kernel threads with a user context.
1126 *
1127 * The kernel acting on behalf of the user is potentially cpu-bound,
1128 * this function will efficiently allow other threads to run and also
1129 * switch to other processes by releasing.
3824f392
MD
1130 *
1131 * The lwkt_user_yield() function is designed to have very low overhead
1132 * if no yield is determined to be needed.
1133 */
1134void
1135lwkt_user_yield(void)
1136{
f9235b6d
MD
1137 globaldata_t gd = mycpu;
1138 thread_t td = gd->gd_curthread;
1139
1140 /*
1141 * Always run any pending interrupts in case we are in a critical
1142 * section.
1143 */
1144 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1145 splz();
3824f392 1146
3824f392 1147 /*
f9235b6d
MD
1148 * Switch (which forces a release) if another kernel thread needs
1149 * the cpu, if userland wants us to resched, or if our kernel
1150 * quantum has run out.
3824f392 1151 */
f9235b6d
MD
1152 if (lwkt_resched_wanted() ||
1153 user_resched_wanted() ||
1154 td->td_fairq_accum < 0)
1155 {
3824f392 1156 lwkt_switch();
3824f392
MD
1157 }
1158
f9235b6d 1159#if 0
3824f392 1160 /*
f9235b6d
MD
1161 * Reacquire the current process if we are released.
1162 *
1163 * XXX not implemented atm. The kernel may be holding locks and such,
1164 * so we want the thread to continue to receive cpu.
3824f392 1165 */
f9235b6d
MD
1166 if (td->td_release == NULL && lp) {
1167 lp->lwp_proc->p_usched->acquire_curproc(lp);
1168 td->td_release = lwkt_passive_release;
1169 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1170 }
f9235b6d 1171#endif
b9eb1c19
MD
1172}
1173
8ad65e08 1174/*
f1d1c3fa
MD
1175 * Generic schedule. Possibly schedule threads belonging to other cpus and
1176 * deal with threads that might be blocked on a wait queue.
1177 *
0a3f9b47
MD
1178 * We have a little helper inline function which does additional work after
1179 * the thread has been enqueued, including dealing with preemption and
1180 * setting need_lwkt_resched() (which prevents the kernel from returning
1181 * to userland until it has processed higher priority threads).
6330a558
MD
1182 *
1183 * It is possible for this routine to be called after a failed _enqueue
1184 * (due to the target thread migrating, sleeping, or otherwise blocked).
1185 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1186 *
1187 * reschedok is an optimized constant propagated from lwkt_schedule() or
1188 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1189 * reschedule to be requested if the target thread has a higher priority.
1190 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1191 * be 0, prevented undesired reschedules.
8ad65e08 1192 */
0a3f9b47
MD
1193static __inline
1194void
f9235b6d 1195_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
0a3f9b47 1196{
b9eb1c19 1197 thread_t otd;
c730be20 1198
6330a558 1199 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1200 if (ntd->td_preemptable && reschedok) {
f9235b6d 1201 ntd->td_preemptable(ntd, ccount); /* YYY +token */
361d01dd 1202 } else if (reschedok) {
b9eb1c19 1203 otd = curthread;
f9235b6d 1204 if (ntd->td_pri > otd->td_pri)
c730be20 1205 need_lwkt_resched();
6330a558 1206 }
f9235b6d
MD
1207
1208 /*
1209 * Give the thread a little fair share scheduler bump if it
1210 * has been asleep for a while. This is primarily to avoid
1211 * a degenerate case for interrupt threads where accumulator
1212 * crosses into negative territory unnecessarily.
1213 */
1214 if (ntd->td_fairq_lticks != ticks) {
1215 ntd->td_fairq_lticks = ticks;
1216 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1217 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1218 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1219 }
0a3f9b47
MD
1220 }
1221}
1222
361d01dd 1223static __inline
8ad65e08 1224void
361d01dd 1225_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1226{
37af14fe
MD
1227 globaldata_t mygd = mycpu;
1228
cf709dd2
MD
1229 KASSERT(td != &td->td_gd->gd_idlethread,
1230 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1231 crit_enter_gd(mygd);
9388413d 1232 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1233 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1234 _lwkt_enqueue(td);
1235 } else {
f1d1c3fa 1236 /*
7cd8d145
MD
1237 * If we own the thread, there is no race (since we are in a
1238 * critical section). If we do not own the thread there might
1239 * be a race but the target cpu will deal with it.
f1d1c3fa 1240 */
0f7a3396 1241#ifdef SMP
7cd8d145 1242 if (td->td_gd == mygd) {
9d265729 1243 _lwkt_enqueue(td);
f9235b6d 1244 _lwkt_schedule_post(mygd, td, 1, reschedok);
f1d1c3fa 1245 } else {
e381e77c 1246 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1247 }
0f7a3396 1248#else
7cd8d145 1249 _lwkt_enqueue(td);
f9235b6d 1250 _lwkt_schedule_post(mygd, td, 1, reschedok);
0f7a3396 1251#endif
8ad65e08 1252 }
37af14fe 1253 crit_exit_gd(mygd);
8ad65e08
MD
1254}
1255
361d01dd
MD
1256void
1257lwkt_schedule(thread_t td)
1258{
1259 _lwkt_schedule(td, 1);
1260}
1261
1262void
1263lwkt_schedule_noresched(thread_t td)
1264{
1265 _lwkt_schedule(td, 0);
1266}
1267
88ebb169
SW
1268#ifdef SMP
1269
e381e77c
MD
1270/*
1271 * When scheduled remotely if frame != NULL the IPIQ is being
1272 * run via doreti or an interrupt then preemption can be allowed.
1273 *
1274 * To allow preemption we have to drop the critical section so only
1275 * one is present in _lwkt_schedule_post.
1276 */
1277static void
1278lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1279{
1280 thread_t td = curthread;
1281 thread_t ntd = arg;
1282
1283 if (frame && ntd->td_preemptable) {
1284 crit_exit_noyield(td);
1285 _lwkt_schedule(ntd, 1);
1286 crit_enter_quick(td);
1287 } else {
1288 _lwkt_schedule(ntd, 1);
1289 }
1290}
1291
d9eea1a5 1292/*
52eedfb5
MD
1293 * Thread migration using a 'Pull' method. The thread may or may not be
1294 * the current thread. It MUST be descheduled and in a stable state.
1295 * lwkt_giveaway() must be called on the cpu owning the thread.
1296 *
1297 * At any point after lwkt_giveaway() is called, the target cpu may
1298 * 'pull' the thread by calling lwkt_acquire().
1299 *
ae8e83e6
MD
1300 * We have to make sure the thread is not sitting on a per-cpu tsleep
1301 * queue or it will blow up when it moves to another cpu.
1302 *
52eedfb5 1303 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1304 */
52eedfb5
MD
1305void
1306lwkt_giveaway(thread_t td)
1307{
3b4192fb 1308 globaldata_t gd = mycpu;
52eedfb5 1309
3b4192fb
MD
1310 crit_enter_gd(gd);
1311 if (td->td_flags & TDF_TSLEEPQ)
1312 tsleep_remove(td);
1313 KKASSERT(td->td_gd == gd);
1314 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1315 td->td_flags |= TDF_MIGRATING;
1316 crit_exit_gd(gd);
52eedfb5
MD
1317}
1318
a2a5ad0d
MD
1319void
1320lwkt_acquire(thread_t td)
1321{
37af14fe
MD
1322 globaldata_t gd;
1323 globaldata_t mygd;
a2a5ad0d 1324
52eedfb5 1325 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1326 gd = td->td_gd;
37af14fe 1327 mygd = mycpu;
52eedfb5 1328 if (gd != mycpu) {
35238fa5 1329 cpu_lfence();
52eedfb5 1330 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1331 crit_enter_gd(mygd);
df910c23
MD
1332 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1333#ifdef SMP
1334 lwkt_process_ipiq();
1335#endif
52eedfb5 1336 cpu_lfence();
df910c23 1337 }
562273ea 1338 cpu_mfence();
37af14fe 1339 td->td_gd = mygd;
52eedfb5
MD
1340 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1341 td->td_flags &= ~TDF_MIGRATING;
1342 crit_exit_gd(mygd);
1343 } else {
1344 crit_enter_gd(mygd);
1345 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1346 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1347 crit_exit_gd(mygd);
a2a5ad0d
MD
1348 }
1349}
1350
52eedfb5
MD
1351#endif
1352
f1d1c3fa
MD
1353/*
1354 * Generic deschedule. Descheduling threads other then your own should be
1355 * done only in carefully controlled circumstances. Descheduling is
1356 * asynchronous.
1357 *
1358 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1359 */
1360void
1361lwkt_deschedule(thread_t td)
1362{
f1d1c3fa 1363 crit_enter();
b8a98473 1364#ifdef SMP
f1d1c3fa
MD
1365 if (td == curthread) {
1366 _lwkt_dequeue(td);
1367 } else {
a72187e9 1368 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1369 _lwkt_dequeue(td);
1370 } else {
b8a98473 1371 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1372 }
1373 }
b8a98473
MD
1374#else
1375 _lwkt_dequeue(td);
1376#endif
f1d1c3fa
MD
1377 crit_exit();
1378}
1379
4b5f931b
MD
1380/*
1381 * Set the target thread's priority. This routine does not automatically
1382 * switch to a higher priority thread, LWKT threads are not designed for
1383 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1384 */
1385void
1386lwkt_setpri(thread_t td, int pri)
1387{
a72187e9 1388 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1389 if (td->td_pri != pri) {
1390 KKASSERT(pri >= 0);
1391 crit_enter();
1392 if (td->td_flags & TDF_RUNQ) {
1393 _lwkt_dequeue(td);
1394 td->td_pri = pri;
1395 _lwkt_enqueue(td);
1396 } else {
1397 td->td_pri = pri;
1398 }
1399 crit_exit();
26a0694b 1400 }
26a0694b
MD
1401}
1402
03bd0a5e
MD
1403/*
1404 * Set the initial priority for a thread prior to it being scheduled for
1405 * the first time. The thread MUST NOT be scheduled before or during
1406 * this call. The thread may be assigned to a cpu other then the current
1407 * cpu.
1408 *
1409 * Typically used after a thread has been created with TDF_STOPPREQ,
1410 * and before the thread is initially scheduled.
1411 */
1412void
1413lwkt_setpri_initial(thread_t td, int pri)
1414{
1415 KKASSERT(pri >= 0);
1416 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1417 td->td_pri = pri;
03bd0a5e
MD
1418}
1419
26a0694b
MD
1420void
1421lwkt_setpri_self(int pri)
1422{
1423 thread_t td = curthread;
1424
4b5f931b
MD
1425 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1426 crit_enter();
1427 if (td->td_flags & TDF_RUNQ) {
1428 _lwkt_dequeue(td);
f9235b6d 1429 td->td_pri = pri;
4b5f931b
MD
1430 _lwkt_enqueue(td);
1431 } else {
f9235b6d 1432 td->td_pri = pri;
4b5f931b
MD
1433 }
1434 crit_exit();
1435}
1436
f9235b6d
MD
1437/*
1438 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1439 *
1440 * Example: two competing threads, same priority N. decrement by (2*N)
1441 * increment by N*8, each thread will get 4 ticks.
1442 */
1443void
1444lwkt_fairq_schedulerclock(thread_t td)
1445{
2a418930
MD
1446 globaldata_t gd;
1447
f9235b6d
MD
1448 if (fairq_enable) {
1449 while (td) {
2a418930
MD
1450 gd = td->td_gd;
1451 if (td != &gd->gd_idlethread) {
1452 td->td_fairq_accum -= gd->gd_fairq_total_pri;
1453 if (td->td_fairq_accum < -TDFAIRQ_MAX(gd))
1454 td->td_fairq_accum = -TDFAIRQ_MAX(gd);
f9235b6d
MD
1455 if (td->td_fairq_accum < 0)
1456 need_lwkt_resched();
1457 td->td_fairq_lticks = ticks;
1458 }
1459 td = td->td_preempted;
1460 }
1461 }
1462}
1463
1464static void
1465lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1466{
1467 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1468 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1469 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1470}
1471
5d21b981 1472/*
52eedfb5
MD
1473 * Migrate the current thread to the specified cpu.
1474 *
1475 * This is accomplished by descheduling ourselves from the current cpu,
1476 * moving our thread to the tdallq of the target cpu, IPI messaging the
1477 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1478 * races while the thread is being migrated.
ae8e83e6
MD
1479 *
1480 * We must be sure to remove ourselves from the current cpu's tsleepq
1481 * before potentially moving to another queue. The thread can be on
1482 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1483 */
3d28ff59 1484#ifdef SMP
5d21b981 1485static void lwkt_setcpu_remote(void *arg);
3d28ff59 1486#endif
5d21b981
MD
1487
1488void
1489lwkt_setcpu_self(globaldata_t rgd)
1490{
1491#ifdef SMP
1492 thread_t td = curthread;
1493
1494 if (td->td_gd != rgd) {
1495 crit_enter_quick(td);
ae8e83e6 1496 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1497 tsleep_remove(td);
5d21b981
MD
1498 td->td_flags |= TDF_MIGRATING;
1499 lwkt_deschedule_self(td);
52eedfb5 1500 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1501 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1502 lwkt_switch();
1503 /* we are now on the target cpu */
52eedfb5 1504 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1505 crit_exit_quick(td);
1506 }
1507#endif
1508}
1509
ecdefdda
MD
1510void
1511lwkt_migratecpu(int cpuid)
1512{
1513#ifdef SMP
1514 globaldata_t rgd;
1515
1516 rgd = globaldata_find(cpuid);
1517 lwkt_setcpu_self(rgd);
1518#endif
1519}
1520
5d21b981
MD
1521/*
1522 * Remote IPI for cpu migration (called while in a critical section so we
1523 * do not have to enter another one). The thread has already been moved to
1524 * our cpu's allq, but we must wait for the thread to be completely switched
1525 * out on the originating cpu before we schedule it on ours or the stack
1526 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1527 * change to main memory.
1528 *
1529 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1530 * against wakeups. It is best if this interface is used only when there
1531 * are no pending events that might try to schedule the thread.
1532 */
3d28ff59 1533#ifdef SMP
5d21b981
MD
1534static void
1535lwkt_setcpu_remote(void *arg)
1536{
1537 thread_t td = arg;
1538 globaldata_t gd = mycpu;
1539
df910c23
MD
1540 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1541#ifdef SMP
1542 lwkt_process_ipiq();
1543#endif
35238fa5 1544 cpu_lfence();
562273ea 1545 cpu_pause();
df910c23 1546 }
5d21b981 1547 td->td_gd = gd;
562273ea 1548 cpu_mfence();
5d21b981 1549 td->td_flags &= ~TDF_MIGRATING;
9388413d 1550 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1551 _lwkt_enqueue(td);
1552}
3d28ff59 1553#endif
5d21b981 1554
553ea3c8 1555struct lwp *
4b5f931b
MD
1556lwkt_preempted_proc(void)
1557{
73e4f7b9 1558 thread_t td = curthread;
4b5f931b
MD
1559 while (td->td_preempted)
1560 td = td->td_preempted;
553ea3c8 1561 return(td->td_lwp);
4b5f931b
MD
1562}
1563
99df837e
MD
1564/*
1565 * Create a kernel process/thread/whatever. It shares it's address space
1566 * with proc0 - ie: kernel only.
1567 *
365fa13f
MD
1568 * NOTE! By default new threads are created with the MP lock held. A
1569 * thread which does not require the MP lock should release it by calling
1570 * rel_mplock() at the start of the new thread.
99df837e
MD
1571 */
1572int
c9e9fb21
MD
1573lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1574 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1575{
73e4f7b9 1576 thread_t td;
e2565a42 1577 __va_list ap;
99df837e 1578
d3d32139 1579 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1580 tdflags);
a2a5ad0d
MD
1581 if (tdp)
1582 *tdp = td;
709799ea 1583 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1584
1585 /*
1586 * Set up arg0 for 'ps' etc
1587 */
e2565a42 1588 __va_start(ap, fmt);
379210cb 1589 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1590 __va_end(ap);
99df837e
MD
1591
1592 /*
1593 * Schedule the thread to run
1594 */
ef0fdad1
MD
1595 if ((td->td_flags & TDF_STOPREQ) == 0)
1596 lwkt_schedule(td);
1597 else
1598 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1599 return 0;
1600}
1601
1602/*
1603 * Destroy an LWKT thread. Warning! This function is not called when
1604 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1605 * uses a different reaping mechanism.
1606 */
1607void
1608lwkt_exit(void)
1609{
1610 thread_t td = curthread;
c070746a 1611 thread_t std;
8826f33a 1612 globaldata_t gd;
99df837e 1613
2883d2d8
MD
1614 /*
1615 * Do any cleanup that might block here
1616 */
99df837e 1617 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1618 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1619 caps_exit(td);
2883d2d8
MD
1620 biosched_done(td);
1621 dsched_exit_thread(td);
c070746a
MD
1622
1623 /*
1624 * Get us into a critical section to interlock gd_freetd and loop
1625 * until we can get it freed.
1626 *
1627 * We have to cache the current td in gd_freetd because objcache_put()ing
1628 * it would rip it out from under us while our thread is still active.
1629 */
1630 gd = mycpu;
37af14fe 1631 crit_enter_quick(td);
c070746a 1632 while ((std = gd->gd_freetd) != NULL) {
cf709dd2 1633 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
c070746a
MD
1634 gd->gd_freetd = NULL;
1635 objcache_put(thread_cache, std);
1636 }
3b4192fb
MD
1637
1638 /*
1639 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1640 * the last time. We cannot block after this point or we may end
1641 * up with a stale td on the tsleepq.
3b4192fb
MD
1642 */
1643 if (td->td_flags & TDF_TSLEEPQ)
1644 tsleep_remove(td);
37af14fe 1645 lwkt_deschedule_self(td);
e56e4dea 1646 lwkt_remove_tdallq(td);
2883d2d8
MD
1647
1648 /*
1649 * Final cleanup
1650 */
1651 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1652 if (td->td_flags & TDF_ALLOCATED_THREAD)
1653 gd->gd_freetd = td;
99df837e
MD
1654 cpu_thread_exit();
1655}
1656
e56e4dea
MD
1657void
1658lwkt_remove_tdallq(thread_t td)
1659{
1660 KKASSERT(td->td_gd == mycpu);
1661 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1662}
1663
9cf43f91
MD
1664/*
1665 * Code reduction and branch prediction improvements. Call/return
1666 * overhead on modern cpus often degenerates into 0 cycles due to
1667 * the cpu's branch prediction hardware and return pc cache. We
1668 * can take advantage of this by not inlining medium-complexity
1669 * functions and we can also reduce the branch prediction impact
1670 * by collapsing perfectly predictable branches into a single
1671 * procedure instead of duplicating it.
1672 *
1673 * Is any of this noticeable? Probably not, so I'll take the
1674 * smaller code size.
1675 */
1676void
b6468f56 1677crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1678{
b6468f56 1679 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1680}
1681
2d93b37a
MD
1682void
1683crit_panic(void)
1684{
1685 thread_t td = curthread;
850634cc 1686 int lcrit = td->td_critcount;
2d93b37a 1687
850634cc
AH
1688 td->td_critcount = 0;
1689 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1690 /* NOT REACHED */
2d93b37a
MD
1691}
1692
d165e668
MD
1693#ifdef SMP
1694
bd8015ca
MD
1695/*
1696 * Called from debugger/panic on cpus which have been stopped. We must still
1697 * process the IPIQ while stopped, even if we were stopped while in a critical
1698 * section (XXX).
1699 *
1700 * If we are dumping also try to process any pending interrupts. This may
1701 * or may not work depending on the state of the cpu at the point it was
1702 * stopped.
1703 */
1704void
1705lwkt_smp_stopped(void)
1706{
1707 globaldata_t gd = mycpu;
1708
1709 crit_enter_gd(gd);
1710 if (dumping) {
1711 lwkt_process_ipiq();
1712 splz();
1713 } else {
1714 lwkt_process_ipiq();
1715 }
1716 crit_exit_gd(gd);
1717}
1718
d165e668 1719#endif