vkernel - Fix corrupt tailq (vkernel64 only)
[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
286/*
7d0bac62 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{
74c9628e 433 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
434}
435
436void
437lwkt_rele(thread_t td)
438{
439 KKASSERT(td->td_refs > 0);
74c9628e 440 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
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{
74c9628e 453 KKASSERT(td->td_refs == 0);
cf709dd2 454 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|TDF_RUNQ)) == 0);
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
7d0bac62 469/*
8ad65e08 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 *
3933a3ab
MD
484 * Note that the td_switch() function cannot do anything that requires
485 * the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
486 * the target thread (not the current thread). It's nice to have a scheduler
487 * that does not need the MP lock to work because it allows us to do some
488 * really cool high-performance MP lock optimizations.
69d78e99
MD
489 *
490 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
491 * is not called by the current thread in the preemption case, only when
492 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
493 */
494void
495lwkt_switch(void)
496{
37af14fe
MD
497 globaldata_t gd = mycpu;
498 thread_t td = gd->gd_curthread;
8ad65e08 499 thread_t ntd;
f9235b6d 500 thread_t xtd;
2a418930
MD
501 int spinning = lwkt_spin_loops; /* loops before HLTing */
502 int reqflags;
503 int cseq;
0f0466c0 504 int oseq;
8ad65e08 505
46a3f46d 506 /*
27e88a6e
MD
507 * Switching from within a 'fast' (non thread switched) interrupt or IPI
508 * is illegal. However, we may have to do it anyway if we hit a fatal
509 * kernel trap or we have paniced.
510 *
511 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 512 */
27e88a6e
MD
513 if (gd->gd_intr_nesting_level) {
514 int savegdnest;
515 int savegdtrap;
516
5fddbda2 517 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22
MD
518 panic("lwkt_switch: Attempt to switch from a "
519 "a fast interrupt, ipi, or hard code section, "
520 "td %p\n",
521 td);
27e88a6e
MD
522 } else {
523 savegdnest = gd->gd_intr_nesting_level;
524 savegdtrap = gd->gd_trap_nesting_level;
525 gd->gd_intr_nesting_level = 0;
526 gd->gd_trap_nesting_level = 0;
a7422615
MD
527 if ((td->td_flags & TDF_PANICWARN) == 0) {
528 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
529 kprintf("Warning: thread switch from interrupt, IPI, "
530 "or hard code section.\n"
a7422615 531 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 532 print_backtrace(-1);
a7422615 533 }
27e88a6e
MD
534 lwkt_switch();
535 gd->gd_intr_nesting_level = savegdnest;
536 gd->gd_trap_nesting_level = savegdtrap;
537 return;
538 }
96728c05 539 }
ef0fdad1 540
cb973d15
MD
541 /*
542 * Passive release (used to transition from user to kernel mode
543 * when we block or switch rather then when we enter the kernel).
544 * This function is NOT called if we are switching into a preemption
545 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
546 * our current process designation (if we have one) and become a true
547 * LWKT thread, and may also hand the current process designation to
548 * another process and schedule thread.
cb973d15
MD
549 */
550 if (td->td_release)
551 td->td_release(td);
552
37af14fe 553 crit_enter_gd(gd);
3b998fa9 554 if (TD_TOKS_HELD(td))
9d265729
MD
555 lwkt_relalltokens(td);
556
557 /*
b02926de
MD
558 * We had better not be holding any spin locks, but don't get into an
559 * endless panic loop.
9d265729 560 */
d666840a
MD
561 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
562 ("lwkt_switch: still holding %d exclusive spinlocks!",
563 gd->gd_spinlocks_wr));
9d265729 564
8a8d5d85
MD
565
566#ifdef SMP
0f7a3396
MD
567#ifdef INVARIANTS
568 if (td->td_cscount) {
6ea70f76 569 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
570 td);
571 if (panic_on_cscount)
572 panic("switching while mastering cpusync");
573 }
574#endif
8a8d5d85 575#endif
f9235b6d
MD
576
577 /*
578 * If we had preempted another thread on this cpu, resume the preempted
579 * thread. This occurs transparently, whether the preempted thread
580 * was scheduled or not (it may have been preempted after descheduling
581 * itself).
582 *
583 * We have to setup the MP lock for the original thread after backing
584 * out the adjustment that was made to curthread when the original
585 * was preempted.
586 */
99df837e 587 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
588 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
589 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
590
591 /*
b9eb1c19
MD
592 * The interrupt may have woken a thread up, we need to properly
593 * set the reschedule flag if the originally interrupted thread is
594 * at a lower priority.
8ec60c3f 595 */
f9235b6d
MD
596 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
597 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
8ec60c3f 598 need_lwkt_resched();
f9235b6d 599 }
8a8d5d85 600 /* YYY release mp lock on switchback if original doesn't need it */
f9235b6d
MD
601 goto havethread_preempted;
602 }
603
604 /*
605 * Implement round-robin fairq with priority insertion. The priority
606 * insertion is handled by _lwkt_enqueue()
607 *
608 * We have to adjust the MP lock for the target thread. If we
609 * need the MP lock and cannot obtain it we try to locate a
610 * thread that does not need the MP lock. If we cannot, we spin
611 * instead of HLT.
612 *
613 * A similar issue exists for the tokens held by the target thread.
614 * If we cannot obtain ownership of the tokens we cannot immediately
615 * schedule the thread.
616 */
617 for (;;) {
2a418930
MD
618 /*
619 * Clear RQF_AST_LWKT_RESCHED (we handle the reschedule request)
620 * and set RQF_WAKEUP (prevent unnecessary IPIs from being
621 * received).
622 */
623 for (;;) {
624 reqflags = gd->gd_reqflags;
625 if (atomic_cmpset_int(&gd->gd_reqflags, reqflags,
626 (reqflags & ~RQF_AST_LWKT_RESCHED) |
627 RQF_WAKEUP)) {
628 break;
629 }
630 }
f9235b6d 631
4b5f931b 632 /*
2a418930
MD
633 * Hotpath - pull the head of the run queue and attempt to schedule
634 * it. Fairq exhaustion moves the task to the end of the list. If
635 * no threads are runnable we switch to the idle thread.
41a01a4d 636 */
2a418930
MD
637 for (;;) {
638 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
639
640 if (ntd == NULL) {
641 /*
642 * Runq is empty, switch to idle and clear RQF_WAKEUP
643 * to allow it to halt.
644 */
645 ntd = &gd->gd_idlethread;
6f207a2c 646#ifdef SMP
2a418930 647 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
b5d16701 648 ASSERT_NO_TOKENS_HELD(ntd);
6f207a2c 649#endif
2a418930
MD
650 cpu_time.cp_msg[0] = 0;
651 cpu_time.cp_stallpc = 0;
652 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP);
653 goto haveidle;
654 }
655
656 if (ntd->td_fairq_accum >= 0)
657 break;
658
659 splz_check();
660 lwkt_fairq_accumulate(gd, ntd);
661 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
662 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
f9235b6d 663 }
41a01a4d
MD
664
665 /*
2a418930
MD
666 * Hotpath - schedule ntd. Leaves RQF_WAKEUP set to prevent
667 * unwanted decontention IPIs.
6f207a2c
MD
668 *
669 * NOTE: For UP there is no mplock and lwkt_getalltokens()
670 * always succeeds.
8ec60c3f 671 */
2a418930 672 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
f9235b6d 673 goto havethread;
f9235b6d 674
f9235b6d 675 /*
2a418930
MD
676 * Coldpath (SMP only since tokens always succeed on UP)
677 *
678 * We had some contention on the thread we wanted to schedule.
679 * What we do now is try to find a thread that we can schedule
680 * in its stead until decontention reschedules on our cpu.
681 *
682 * The coldpath scan does NOT rearrange threads in the run list
683 * and it also ignores the accumulator.
684 *
685 * We do not immediately schedule a user priority thread, instead
686 * we record it in xtd and continue looking for kernel threads.
687 * A cpu can only have one user priority thread (normally) so just
688 * record the first one.
689 *
690 * NOTE: This scan will also include threads whos fairq's were
691 * accumulated in the first loop.
f9235b6d 692 */
2a418930
MD
693 ++token_contention_count;
694 xtd = NULL;
695 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
41a01a4d 696 /*
2a418930
MD
697 * Try to switch to this thread. If the thread is running at
698 * user priority we clear WAKEUP to allow decontention IPIs
699 * (since this thread is simply running until the one we wanted
700 * decontends), and we make sure that LWKT_RESCHED is not set.
df6b8ba0 701 *
2a418930
MD
702 * Otherwise for kernel threads we leave WAKEUP set to avoid
703 * unnecessary decontention IPIs.
41a01a4d 704 */
2a418930
MD
705 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
706 if (xtd == NULL)
707 xtd = ntd;
708 continue;
f9235b6d 709 }
a453459d 710
f9235b6d 711 /*
2a418930
MD
712 * Do not let the fairq get too negative. Even though we are
713 * ignoring it atm once the scheduler decontends a very negative
714 * thread will get moved to the end of the queue.
f9235b6d 715 */
2a418930
MD
716 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd)) {
717 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd))
718 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd);
719 goto havethread;
8a8d5d85 720 }
f9235b6d 721
df6b8ba0 722 /*
2a418930 723 * Well fubar, this thread is contended as well, loop
df6b8ba0 724 */
2a418930
MD
725 /* */
726 }
727
728 /*
729 * We exhausted the run list but we may have recorded a user
730 * thread to try. We have three choices based on
731 * lwkt.decontention_method.
732 *
733 * (0) Atomically clear RQF_WAKEUP in order to receive decontention
734 * IPIs (to interrupt the user process) and test
735 * RQF_AST_LWKT_RESCHED at the same time.
736 *
737 * This results in significant decontention IPI traffic but may
738 * be more responsive.
739 *
740 * (1) Leave RQF_WAKEUP set so we do not receive a decontention IPI.
741 * An automatic LWKT reschedule will occur on the next hardclock
742 * (typically 100hz).
743 *
744 * This results in no decontention IPI traffic but may be less
745 * responsive. This is the default.
746 *
747 * (2) Refuse to schedule the user process at this time.
748 *
749 * This is highly experimental and should not be used under
750 * normal circumstances. This can cause a user process to
751 * get starved out in situations where kernel threads are
752 * fighting each other for tokens.
753 */
754 if (xtd) {
755 ntd = xtd;
756
757 switch(lwkt_spin_method) {
758 case 0:
759 for (;;) {
760 reqflags = gd->gd_reqflags;
761 if (atomic_cmpset_int(&gd->gd_reqflags,
762 reqflags,
763 reqflags & ~RQF_WAKEUP)) {
764 break;
765 }
766 }
767 break;
768 case 1:
769 reqflags = gd->gd_reqflags;
770 break;
771 default:
772 goto skip;
773 break;
774 }
775 if ((reqflags & RQF_AST_LWKT_RESCHED) == 0 &&
b5d16701 776 (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
f9235b6d 777 ) {
2a418930
MD
778 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd))
779 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd);
780 goto havethread;
df6b8ba0 781 }
9ac1ee6e 782
2a418930 783skip:
9ac1ee6e 784 /*
2a418930
MD
785 * Make sure RQF_WAKEUP is set if we failed to schedule the
786 * user thread to prevent the idle thread from halting.
9ac1ee6e 787 */
2a418930
MD
788 atomic_set_int(&gd->gd_reqflags, RQF_WAKEUP);
789 }
790
791 /*
792 * We exhausted the run list, meaning that all runnable threads
793 * are contended.
794 */
795 cpu_pause();
796 ntd = &gd->gd_idlethread;
797#ifdef SMP
798 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
799 ASSERT_NO_TOKENS_HELD(ntd);
800 /* contention case, do not clear contention mask */
801#endif
802
803 /*
804 * Ok, we might want to spin a few times as some tokens are held for
805 * very short periods of time and IPI overhead is 1uS or worse
806 * (meaning it is usually better to spin). Regardless we have to
807 * call splz_check() to be sure to service any interrupts blocked
808 * by our critical section, otherwise we could livelock e.g. IPIs.
809 *
810 * The IPI mechanic is really a last resort. In nearly all other
811 * cases RQF_WAKEUP is left set to prevent decontention IPIs.
812 *
813 * When we decide not to spin we clear RQF_WAKEUP and switch to
814 * the idle thread. Clearing RQF_WEAKEUP allows the idle thread
815 * to halt and decontended tokens will issue an IPI to us. The
816 * idle thread will check for pending reschedules already set
817 * (RQF_AST_LWKT_RESCHED) before actually halting so we don't have
818 * to here.
819 */
820 if (spinning <= 0) {
821 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP);
822 goto haveidle;
4b5f931b 823 }
2a418930 824 --spinning;
c5724852
MD
825
826 /*
2a418930
MD
827 * When spinning a delay is required both to avoid livelocks from
828 * token order reversals (a thread may be trying to acquire multiple
829 * tokens), and also to reduce cpu cache management traffic.
830 *
831 * In order to scale to a large number of CPUs we use a time slot
832 * resequencer to force contending cpus into non-contending
833 * time-slots. The scheduler may still contend with the lock holder
834 * but will not (generally) contend with all the other cpus trying
835 * trying to get the same token.
836 *
837 * The resequencer uses a FIFO counter mechanic. The owner of the
838 * rindex at the head of the FIFO is allowed to pull itself off
839 * the FIFO and fetchadd is used to enter into the FIFO. This bit
840 * of code is VERY cache friendly and forces all spinning schedulers
841 * into their own time slots.
c5724852 842 *
2a418930
MD
843 * This code has been tested to 48-cpus and caps the cache
844 * contention load at ~1uS intervals regardless of the number of
845 * cpus. Scaling beyond 64 cpus might require additional smarts
846 * (such as separate FIFOs for specific token cases).
847 *
848 * WARNING! We can't call splz_check() or anything else here as
849 * it could cause a deadlock.
c5724852 850 */
2a418930 851 cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
0f0466c0
MD
852 while ((oseq = lwkt_cseq_rindex) != cseq) {
853 cpu_ccfence();
8b402283 854#if !defined(_KERNEL_VIRTUAL)
0f0466c0
MD
855 if (cpu_mi_feature & CPU_MI_MONITOR) {
856 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
8b402283
SW
857 } else
858#endif
859 {
0f0466c0
MD
860 DELAY(1);
861 cpu_lfence();
862 }
2a418930
MD
863 }
864 cseq = lwkt_spin_delay; /* don't trust the system operator */
865 cpu_ccfence();
866 if (cseq < 1)
867 cseq = 1;
868 if (cseq > 1000)
869 cseq = 1000;
870 DELAY(cseq);
871 atomic_add_int(&lwkt_cseq_rindex, 1);
c5724852 872 splz_check();
2a418930 873 /* highest level for(;;) loop */
f1d1c3fa 874 }
8a8d5d85 875
2a418930 876havethread:
8a8d5d85 877 /*
f9235b6d
MD
878 * We must always decrement td_fairq_accum on non-idle threads just
879 * in case a thread never gets a tick due to being in a continuous
2a418930 880 * critical section. The page-zeroing code does this, for example.
f9235b6d
MD
881 *
882 * If the thread we came up with is a higher or equal priority verses
883 * the thread at the head of the queue we move our thread to the
884 * front. This way we can always check the front of the queue.
be71787b
MD
885 *
886 * Clear gd_idle_repeat when doing a normal switch to a non-idle
887 * thread.
f9235b6d 888 */
f9235b6d
MD
889 ++gd->gd_cnt.v_swtch;
890 --ntd->td_fairq_accum;
9ac1ee6e 891 ntd->td_wmesg = NULL;
f9235b6d
MD
892 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
893 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
894 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
895 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
896 }
be71787b 897 gd->gd_idle_repeat = 0;
2a418930 898
f9235b6d 899havethread_preempted:
f9235b6d
MD
900 /*
901 * If the new target does not need the MP lock and we are holding it,
902 * release the MP lock. If the new target requires the MP lock we have
903 * already acquired it for the target.
8a8d5d85 904 */
2a418930 905 ;
f9235b6d
MD
906haveidle:
907 KASSERT(ntd->td_critcount,
b5d16701
MD
908 ("priority problem in lwkt_switch %d %d",
909 td->td_critcount, ntd->td_critcount));
910
94f6d86e
MD
911 if (td != ntd) {
912 ++switch_count;
a1f0fb66 913 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
f1d1c3fa 914 td->td_switch(ntd);
94f6d86e 915 }
37af14fe
MD
916 /* NOTE: current cpu may have changed after switch */
917 crit_exit_quick(td);
8ad65e08
MD
918}
919
f1d1c3fa 920/*
96728c05
MD
921 * Request that the target thread preempt the current thread. Preemption
922 * only works under a specific set of conditions:
b68b7282 923 *
96728c05
MD
924 * - We are not preempting ourselves
925 * - The target thread is owned by the current cpu
926 * - We are not currently being preempted
927 * - The target is not currently being preempted
d3d1cbc8
MD
928 * - We are not holding any spin locks
929 * - The target thread is not holding any tokens
96728c05
MD
930 * - We are able to satisfy the target's MP lock requirements (if any).
931 *
932 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
933 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 934 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
935 * to determine whether preemption is possible (aka usually just the crit
936 * priority of lwkt_schedule() itself).
b68b7282 937 *
26a0694b
MD
938 * XXX at the moment we run the target thread in a critical section during
939 * the preemption in order to prevent the target from taking interrupts
940 * that *WE* can't. Preemption is strictly limited to interrupt threads
941 * and interrupt-like threads, outside of a critical section, and the
942 * preempted source thread will be resumed the instant the target blocks
943 * whether or not the source is scheduled (i.e. preemption is supposed to
944 * be as transparent as possible).
b68b7282
MD
945 */
946void
f9235b6d 947lwkt_preempt(thread_t ntd, int critcount)
b68b7282 948{
46a3f46d 949 struct globaldata *gd = mycpu;
0a3f9b47 950 thread_t td;
2d910aaf 951 int save_gd_intr_nesting_level;
b68b7282 952
26a0694b 953 /*
96728c05
MD
954 * The caller has put us in a critical section. We can only preempt
955 * if the caller of the caller was not in a critical section (basically
f9235b6d 956 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 957 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 958 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
959 *
960 * YYY The target thread must be in a critical section (else it must
961 * inherit our critical section? I dunno yet).
41a01a4d 962 *
0a3f9b47 963 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b 964 */
f9235b6d 965 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 966
fbc024e4
MD
967 if (preempt_enable == 0) {
968 ++preempt_miss;
969 return;
970 }
971
0a3f9b47 972 td = gd->gd_curthread;
f9235b6d 973 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
974 ++preempt_miss;
975 return;
976 }
f9235b6d 977 if (td->td_critcount > critcount) {
96728c05 978 ++preempt_miss;
8ec60c3f 979 need_lwkt_resched();
96728c05
MD
980 return;
981 }
982#ifdef SMP
46a3f46d 983 if (ntd->td_gd != gd) {
96728c05 984 ++preempt_miss;
8ec60c3f 985 need_lwkt_resched();
96728c05
MD
986 return;
987 }
988#endif
41a01a4d 989 /*
77912481
MD
990 * We don't have to check spinlocks here as they will also bump
991 * td_critcount.
d3d1cbc8
MD
992 *
993 * Do not try to preempt if the target thread is holding any tokens.
994 * We could try to acquire the tokens but this case is so rare there
995 * is no need to support it.
41a01a4d 996 */
77912481
MD
997 KKASSERT(gd->gd_spinlocks_wr == 0);
998
3b998fa9 999 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8
MD
1000 ++preempt_miss;
1001 need_lwkt_resched();
1002 return;
1003 }
26a0694b
MD
1004 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1005 ++preempt_weird;
8ec60c3f 1006 need_lwkt_resched();
26a0694b
MD
1007 return;
1008 }
1009 if (ntd->td_preempted) {
4b5f931b 1010 ++preempt_hit;
8ec60c3f 1011 need_lwkt_resched();
26a0694b 1012 return;
b68b7282 1013 }
26a0694b 1014
8ec60c3f
MD
1015 /*
1016 * Since we are able to preempt the current thread, there is no need to
1017 * call need_lwkt_resched().
2d910aaf
MD
1018 *
1019 * We must temporarily clear gd_intr_nesting_level around the switch
1020 * since switchouts from the target thread are allowed (they will just
1021 * return to our thread), and since the target thread has its own stack.
8ec60c3f 1022 */
26a0694b
MD
1023 ++preempt_hit;
1024 ntd->td_preempted = td;
1025 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1026 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1027 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1028 gd->gd_intr_nesting_level = 0;
26a0694b 1029 td->td_switch(ntd);
2d910aaf 1030 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1031
26a0694b
MD
1032 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1033 ntd->td_preempted = NULL;
1034 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1035}
1036
1037/*
faaeffac 1038 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1039 * This will work inside a critical section but not inside a hard code
1040 * section.
ef0fdad1 1041 *
f1d1c3fa
MD
1042 * (self contained on a per cpu basis)
1043 */
1044void
faaeffac 1045splz_check(void)
f1d1c3fa 1046{
7966cb69
MD
1047 globaldata_t gd = mycpu;
1048 thread_t td = gd->gd_curthread;
ef0fdad1 1049
4a28fe22
MD
1050 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1051 gd->gd_intr_nesting_level == 0 &&
1052 td->td_nest_count < 2)
1053 {
f1d1c3fa 1054 splz();
4a28fe22
MD
1055 }
1056}
1057
1058/*
1059 * This version is integrated into crit_exit, reqflags has already
1060 * been tested but td_critcount has not.
1061 *
1062 * We only want to execute the splz() on the 1->0 transition of
1063 * critcount and not in a hard code section or if too deeply nested.
1064 */
1065void
1066lwkt_maybe_splz(thread_t td)
1067{
1068 globaldata_t gd = td->td_gd;
1069
1070 if (td->td_critcount == 0 &&
1071 gd->gd_intr_nesting_level == 0 &&
1072 td->td_nest_count < 2)
1073 {
1074 splz();
1075 }
f1d1c3fa
MD
1076}
1077
8ad65e08 1078/*
f9235b6d
MD
1079 * This function is used to negotiate a passive release of the current
1080 * process/lwp designation with the user scheduler, allowing the user
1081 * scheduler to schedule another user thread. The related kernel thread
1082 * (curthread) continues running in the released state.
8ad65e08
MD
1083 */
1084void
f9235b6d 1085lwkt_passive_release(struct thread *td)
8ad65e08 1086{
f9235b6d
MD
1087 struct lwp *lp = td->td_lwp;
1088
1089 td->td_release = NULL;
1090 lwkt_setpri_self(TDPRI_KERN_USER);
1091 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1092}
1093
f9235b6d 1094
f1d1c3fa 1095/*
f9235b6d
MD
1096 * This implements a normal yield. This routine is virtually a nop if
1097 * there is nothing to yield to but it will always run any pending interrupts
1098 * if called from a critical section.
1099 *
1100 * This yield is designed for kernel threads without a user context.
1101 *
1102 * (self contained on a per cpu basis)
3824f392
MD
1103 */
1104void
f9235b6d 1105lwkt_yield(void)
3824f392 1106{
f9235b6d
MD
1107 globaldata_t gd = mycpu;
1108 thread_t td = gd->gd_curthread;
1109 thread_t xtd;
3824f392 1110
f9235b6d
MD
1111 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1112 splz();
1113 if (td->td_fairq_accum < 0) {
1114 lwkt_schedule_self(curthread);
1115 lwkt_switch();
1116 } else {
1117 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1118 if (xtd && xtd->td_pri > td->td_pri) {
1119 lwkt_schedule_self(curthread);
1120 lwkt_switch();
1121 }
1122 }
3824f392
MD
1123}
1124
1125/*
f9235b6d
MD
1126 * This yield is designed for kernel threads with a user context.
1127 *
1128 * The kernel acting on behalf of the user is potentially cpu-bound,
1129 * this function will efficiently allow other threads to run and also
1130 * switch to other processes by releasing.
3824f392
MD
1131 *
1132 * The lwkt_user_yield() function is designed to have very low overhead
1133 * if no yield is determined to be needed.
1134 */
1135void
1136lwkt_user_yield(void)
1137{
f9235b6d
MD
1138 globaldata_t gd = mycpu;
1139 thread_t td = gd->gd_curthread;
1140
1141 /*
1142 * Always run any pending interrupts in case we are in a critical
1143 * section.
1144 */
1145 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1146 splz();
3824f392 1147
3824f392 1148 /*
f9235b6d
MD
1149 * Switch (which forces a release) if another kernel thread needs
1150 * the cpu, if userland wants us to resched, or if our kernel
1151 * quantum has run out.
3824f392 1152 */
f9235b6d
MD
1153 if (lwkt_resched_wanted() ||
1154 user_resched_wanted() ||
1155 td->td_fairq_accum < 0)
1156 {
3824f392 1157 lwkt_switch();
3824f392
MD
1158 }
1159
f9235b6d 1160#if 0
3824f392 1161 /*
f9235b6d
MD
1162 * Reacquire the current process if we are released.
1163 *
1164 * XXX not implemented atm. The kernel may be holding locks and such,
1165 * so we want the thread to continue to receive cpu.
3824f392 1166 */
f9235b6d
MD
1167 if (td->td_release == NULL && lp) {
1168 lp->lwp_proc->p_usched->acquire_curproc(lp);
1169 td->td_release = lwkt_passive_release;
1170 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1171 }
f9235b6d 1172#endif
b9eb1c19
MD
1173}
1174
1175/*
f1d1c3fa
MD
1176 * Generic schedule. Possibly schedule threads belonging to other cpus and
1177 * deal with threads that might be blocked on a wait queue.
1178 *
0a3f9b47
MD
1179 * We have a little helper inline function which does additional work after
1180 * the thread has been enqueued, including dealing with preemption and
1181 * setting need_lwkt_resched() (which prevents the kernel from returning
1182 * to userland until it has processed higher priority threads).
6330a558
MD
1183 *
1184 * It is possible for this routine to be called after a failed _enqueue
1185 * (due to the target thread migrating, sleeping, or otherwise blocked).
1186 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1187 *
1188 * reschedok is an optimized constant propagated from lwkt_schedule() or
1189 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1190 * reschedule to be requested if the target thread has a higher priority.
1191 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1192 * be 0, prevented undesired reschedules.
8ad65e08 1193 */
0a3f9b47
MD
1194static __inline
1195void
f9235b6d 1196_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
0a3f9b47 1197{
b9eb1c19 1198 thread_t otd;
c730be20 1199
6330a558 1200 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1201 if (ntd->td_preemptable && reschedok) {
f9235b6d 1202 ntd->td_preemptable(ntd, ccount); /* YYY +token */
361d01dd 1203 } else if (reschedok) {
b9eb1c19 1204 otd = curthread;
f9235b6d 1205 if (ntd->td_pri > otd->td_pri)
c730be20 1206 need_lwkt_resched();
6330a558 1207 }
f9235b6d
MD
1208
1209 /*
1210 * Give the thread a little fair share scheduler bump if it
1211 * has been asleep for a while. This is primarily to avoid
1212 * a degenerate case for interrupt threads where accumulator
1213 * crosses into negative territory unnecessarily.
1214 */
1215 if (ntd->td_fairq_lticks != ticks) {
1216 ntd->td_fairq_lticks = ticks;
1217 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1218 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1219 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1220 }
0a3f9b47
MD
1221 }
1222}
1223
361d01dd 1224static __inline
8ad65e08 1225void
361d01dd 1226_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1227{
37af14fe
MD
1228 globaldata_t mygd = mycpu;
1229
cf709dd2
MD
1230 KASSERT(td != &td->td_gd->gd_idlethread,
1231 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1232 crit_enter_gd(mygd);
9388413d 1233 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1234 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1235 _lwkt_enqueue(td);
1236 } else {
f1d1c3fa 1237 /*
7cd8d145
MD
1238 * If we own the thread, there is no race (since we are in a
1239 * critical section). If we do not own the thread there might
1240 * be a race but the target cpu will deal with it.
f1d1c3fa 1241 */
0f7a3396 1242#ifdef SMP
7cd8d145 1243 if (td->td_gd == mygd) {
9d265729 1244 _lwkt_enqueue(td);
f9235b6d 1245 _lwkt_schedule_post(mygd, td, 1, reschedok);
f1d1c3fa 1246 } else {
e381e77c 1247 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1248 }
0f7a3396 1249#else
7cd8d145 1250 _lwkt_enqueue(td);
f9235b6d 1251 _lwkt_schedule_post(mygd, td, 1, reschedok);
0f7a3396 1252#endif
8ad65e08 1253 }
37af14fe 1254 crit_exit_gd(mygd);
8ad65e08
MD
1255}
1256
361d01dd
MD
1257void
1258lwkt_schedule(thread_t td)
1259{
1260 _lwkt_schedule(td, 1);
1261}
1262
1263void
1264lwkt_schedule_noresched(thread_t td)
1265{
1266 _lwkt_schedule(td, 0);
1267}
1268
88ebb169
SW
1269#ifdef SMP
1270
e381e77c
MD
1271/*
1272 * When scheduled remotely if frame != NULL the IPIQ is being
1273 * run via doreti or an interrupt then preemption can be allowed.
1274 *
1275 * To allow preemption we have to drop the critical section so only
1276 * one is present in _lwkt_schedule_post.
1277 */
1278static void
1279lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1280{
1281 thread_t td = curthread;
1282 thread_t ntd = arg;
1283
1284 if (frame && ntd->td_preemptable) {
1285 crit_exit_noyield(td);
1286 _lwkt_schedule(ntd, 1);
1287 crit_enter_quick(td);
1288 } else {
1289 _lwkt_schedule(ntd, 1);
1290 }
1291}
1292
d9eea1a5 1293/*
52eedfb5
MD
1294 * Thread migration using a 'Pull' method. The thread may or may not be
1295 * the current thread. It MUST be descheduled and in a stable state.
1296 * lwkt_giveaway() must be called on the cpu owning the thread.
1297 *
1298 * At any point after lwkt_giveaway() is called, the target cpu may
1299 * 'pull' the thread by calling lwkt_acquire().
1300 *
ae8e83e6
MD
1301 * We have to make sure the thread is not sitting on a per-cpu tsleep
1302 * queue or it will blow up when it moves to another cpu.
1303 *
52eedfb5 1304 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1305 */
a2a5ad0d 1306void
52eedfb5
MD
1307lwkt_giveaway(thread_t td)
1308{
3b4192fb 1309 globaldata_t gd = mycpu;
52eedfb5 1310
3b4192fb
MD
1311 crit_enter_gd(gd);
1312 if (td->td_flags & TDF_TSLEEPQ)
1313 tsleep_remove(td);
1314 KKASSERT(td->td_gd == gd);
1315 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1316 td->td_flags |= TDF_MIGRATING;
1317 crit_exit_gd(gd);
52eedfb5
MD
1318}
1319
1320void
a2a5ad0d
MD
1321lwkt_acquire(thread_t td)
1322{
37af14fe
MD
1323 globaldata_t gd;
1324 globaldata_t mygd;
a2a5ad0d 1325
52eedfb5 1326 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1327 gd = td->td_gd;
37af14fe 1328 mygd = mycpu;
52eedfb5 1329 if (gd != mycpu) {
35238fa5 1330 cpu_lfence();
52eedfb5 1331 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1332 crit_enter_gd(mygd);
df910c23
MD
1333 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1334#ifdef SMP
1335 lwkt_process_ipiq();
1336#endif
52eedfb5 1337 cpu_lfence();
df910c23 1338 }
562273ea 1339 cpu_mfence();
37af14fe 1340 td->td_gd = mygd;
52eedfb5
MD
1341 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1342 td->td_flags &= ~TDF_MIGRATING;
1343 crit_exit_gd(mygd);
1344 } else {
1345 crit_enter_gd(mygd);
1346 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1347 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1348 crit_exit_gd(mygd);
a2a5ad0d
MD
1349 }
1350}
1351
52eedfb5
MD
1352#endif
1353
8ad65e08 1354/*
f1d1c3fa
MD
1355 * Generic deschedule. Descheduling threads other then your own should be
1356 * done only in carefully controlled circumstances. Descheduling is
1357 * asynchronous.
1358 *
1359 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1360 */
1361void
1362lwkt_deschedule(thread_t td)
1363{
f1d1c3fa 1364 crit_enter();
b8a98473 1365#ifdef SMP
f1d1c3fa
MD
1366 if (td == curthread) {
1367 _lwkt_dequeue(td);
1368 } else {
a72187e9 1369 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1370 _lwkt_dequeue(td);
1371 } else {
b8a98473 1372 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1373 }
1374 }
b8a98473
MD
1375#else
1376 _lwkt_dequeue(td);
1377#endif
f1d1c3fa
MD
1378 crit_exit();
1379}
1380
1381/*
4b5f931b
MD
1382 * Set the target thread's priority. This routine does not automatically
1383 * switch to a higher priority thread, LWKT threads are not designed for
1384 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1385 */
1386void
1387lwkt_setpri(thread_t td, int pri)
1388{
a72187e9 1389 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1390 if (td->td_pri != pri) {
1391 KKASSERT(pri >= 0);
1392 crit_enter();
1393 if (td->td_flags & TDF_RUNQ) {
1394 _lwkt_dequeue(td);
1395 td->td_pri = pri;
1396 _lwkt_enqueue(td);
1397 } else {
1398 td->td_pri = pri;
1399 }
1400 crit_exit();
26a0694b 1401 }
26a0694b
MD
1402}
1403
03bd0a5e
MD
1404/*
1405 * Set the initial priority for a thread prior to it being scheduled for
1406 * the first time. The thread MUST NOT be scheduled before or during
1407 * this call. The thread may be assigned to a cpu other then the current
1408 * cpu.
1409 *
1410 * Typically used after a thread has been created with TDF_STOPPREQ,
1411 * and before the thread is initially scheduled.
1412 */
1413void
1414lwkt_setpri_initial(thread_t td, int pri)
1415{
1416 KKASSERT(pri >= 0);
1417 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1418 td->td_pri = pri;
03bd0a5e
MD
1419}
1420
26a0694b
MD
1421void
1422lwkt_setpri_self(int pri)
1423{
1424 thread_t td = curthread;
1425
4b5f931b
MD
1426 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1427 crit_enter();
1428 if (td->td_flags & TDF_RUNQ) {
1429 _lwkt_dequeue(td);
f9235b6d 1430 td->td_pri = pri;
4b5f931b
MD
1431 _lwkt_enqueue(td);
1432 } else {
f9235b6d 1433 td->td_pri = pri;
4b5f931b
MD
1434 }
1435 crit_exit();
1436}
1437
5d21b981 1438/*
f9235b6d
MD
1439 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1440 *
1441 * Example: two competing threads, same priority N. decrement by (2*N)
1442 * increment by N*8, each thread will get 4 ticks.
1443 */
1444void
1445lwkt_fairq_schedulerclock(thread_t td)
1446{
2a418930
MD
1447 globaldata_t gd;
1448
f9235b6d
MD
1449 if (fairq_enable) {
1450 while (td) {
2a418930
MD
1451 gd = td->td_gd;
1452 if (td != &gd->gd_idlethread) {
1453 td->td_fairq_accum -= gd->gd_fairq_total_pri;
1454 if (td->td_fairq_accum < -TDFAIRQ_MAX(gd))
1455 td->td_fairq_accum = -TDFAIRQ_MAX(gd);
f9235b6d
MD
1456 if (td->td_fairq_accum < 0)
1457 need_lwkt_resched();
1458 td->td_fairq_lticks = ticks;
1459 }
1460 td = td->td_preempted;
1461 }
1462 }
1463}
1464
1465static void
1466lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1467{
1468 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1469 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1470 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1471}
1472
1473/*
52eedfb5
MD
1474 * Migrate the current thread to the specified cpu.
1475 *
1476 * This is accomplished by descheduling ourselves from the current cpu,
1477 * moving our thread to the tdallq of the target cpu, IPI messaging the
1478 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1479 * races while the thread is being migrated.
ae8e83e6
MD
1480 *
1481 * We must be sure to remove ourselves from the current cpu's tsleepq
1482 * before potentially moving to another queue. The thread can be on
1483 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1484 */
3d28ff59 1485#ifdef SMP
5d21b981 1486static void lwkt_setcpu_remote(void *arg);
3d28ff59 1487#endif
5d21b981
MD
1488
1489void
1490lwkt_setcpu_self(globaldata_t rgd)
1491{
1492#ifdef SMP
1493 thread_t td = curthread;
1494
1495 if (td->td_gd != rgd) {
1496 crit_enter_quick(td);
ae8e83e6 1497 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1498 tsleep_remove(td);
5d21b981
MD
1499 td->td_flags |= TDF_MIGRATING;
1500 lwkt_deschedule_self(td);
52eedfb5 1501 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1502 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1503 lwkt_switch();
1504 /* we are now on the target cpu */
52eedfb5 1505 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1506 crit_exit_quick(td);
1507 }
1508#endif
1509}
1510
ecdefdda
MD
1511void
1512lwkt_migratecpu(int cpuid)
1513{
1514#ifdef SMP
1515 globaldata_t rgd;
1516
1517 rgd = globaldata_find(cpuid);
1518 lwkt_setcpu_self(rgd);
1519#endif
1520}
1521
5d21b981
MD
1522/*
1523 * Remote IPI for cpu migration (called while in a critical section so we
1524 * do not have to enter another one). The thread has already been moved to
1525 * our cpu's allq, but we must wait for the thread to be completely switched
1526 * out on the originating cpu before we schedule it on ours or the stack
1527 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1528 * change to main memory.
1529 *
1530 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1531 * against wakeups. It is best if this interface is used only when there
1532 * are no pending events that might try to schedule the thread.
1533 */
3d28ff59 1534#ifdef SMP
5d21b981
MD
1535static void
1536lwkt_setcpu_remote(void *arg)
1537{
1538 thread_t td = arg;
1539 globaldata_t gd = mycpu;
1540
df910c23
MD
1541 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1542#ifdef SMP
1543 lwkt_process_ipiq();
1544#endif
35238fa5 1545 cpu_lfence();
562273ea 1546 cpu_pause();
df910c23 1547 }
5d21b981 1548 td->td_gd = gd;
562273ea 1549 cpu_mfence();
5d21b981 1550 td->td_flags &= ~TDF_MIGRATING;
9388413d 1551 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1552 _lwkt_enqueue(td);
1553}
3d28ff59 1554#endif
5d21b981 1555
553ea3c8 1556struct lwp *
4b5f931b
MD
1557lwkt_preempted_proc(void)
1558{
73e4f7b9 1559 thread_t td = curthread;
4b5f931b
MD
1560 while (td->td_preempted)
1561 td = td->td_preempted;
553ea3c8 1562 return(td->td_lwp);
4b5f931b
MD
1563}
1564
4b5f931b 1565/*
99df837e
MD
1566 * Create a kernel process/thread/whatever. It shares it's address space
1567 * with proc0 - ie: kernel only.
1568 *
365fa13f
MD
1569 * NOTE! By default new threads are created with the MP lock held. A
1570 * thread which does not require the MP lock should release it by calling
1571 * rel_mplock() at the start of the new thread.
99df837e
MD
1572 */
1573int
c9e9fb21
MD
1574lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1575 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1576{
73e4f7b9 1577 thread_t td;
e2565a42 1578 __va_list ap;
99df837e 1579
d3d32139 1580 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1581 tdflags);
a2a5ad0d
MD
1582 if (tdp)
1583 *tdp = td;
709799ea 1584 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1585
1586 /*
1587 * Set up arg0 for 'ps' etc
1588 */
e2565a42 1589 __va_start(ap, fmt);
379210cb 1590 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1591 __va_end(ap);
99df837e
MD
1592
1593 /*
1594 * Schedule the thread to run
1595 */
ef0fdad1
MD
1596 if ((td->td_flags & TDF_STOPREQ) == 0)
1597 lwkt_schedule(td);
1598 else
1599 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1600 return 0;
1601}
1602
1603/*
1604 * Destroy an LWKT thread. Warning! This function is not called when
1605 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1606 * uses a different reaping mechanism.
1607 */
1608void
1609lwkt_exit(void)
1610{
1611 thread_t td = curthread;
c070746a 1612 thread_t std;
8826f33a 1613 globaldata_t gd;
99df837e 1614
2883d2d8
MD
1615 /*
1616 * Do any cleanup that might block here
1617 */
99df837e 1618 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1619 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1620 caps_exit(td);
2883d2d8
MD
1621 biosched_done(td);
1622 dsched_exit_thread(td);
c070746a
MD
1623
1624 /*
1625 * Get us into a critical section to interlock gd_freetd and loop
1626 * until we can get it freed.
1627 *
1628 * We have to cache the current td in gd_freetd because objcache_put()ing
1629 * it would rip it out from under us while our thread is still active.
1630 */
1631 gd = mycpu;
37af14fe 1632 crit_enter_quick(td);
c070746a 1633 while ((std = gd->gd_freetd) != NULL) {
cf709dd2 1634 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
c070746a
MD
1635 gd->gd_freetd = NULL;
1636 objcache_put(thread_cache, std);
1637 }
3b4192fb
MD
1638
1639 /*
1640 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1641 * the last time. We cannot block after this point or we may end
1642 * up with a stale td on the tsleepq.
3b4192fb
MD
1643 */
1644 if (td->td_flags & TDF_TSLEEPQ)
1645 tsleep_remove(td);
37af14fe 1646 lwkt_deschedule_self(td);
e56e4dea 1647 lwkt_remove_tdallq(td);
74c9628e 1648 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1649
1650 /*
1651 * Final cleanup
1652 */
1653 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1654 if (td->td_flags & TDF_ALLOCATED_THREAD)
1655 gd->gd_freetd = td;
99df837e
MD
1656 cpu_thread_exit();
1657}
1658
e56e4dea
MD
1659void
1660lwkt_remove_tdallq(thread_t td)
1661{
1662 KKASSERT(td->td_gd == mycpu);
1663 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1664}
1665
9cf43f91
MD
1666/*
1667 * Code reduction and branch prediction improvements. Call/return
1668 * overhead on modern cpus often degenerates into 0 cycles due to
1669 * the cpu's branch prediction hardware and return pc cache. We
1670 * can take advantage of this by not inlining medium-complexity
1671 * functions and we can also reduce the branch prediction impact
1672 * by collapsing perfectly predictable branches into a single
1673 * procedure instead of duplicating it.
1674 *
1675 * Is any of this noticeable? Probably not, so I'll take the
1676 * smaller code size.
1677 */
1678void
b6468f56 1679crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1680{
b6468f56 1681 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1682}
1683
2d93b37a
MD
1684void
1685crit_panic(void)
1686{
1687 thread_t td = curthread;
850634cc 1688 int lcrit = td->td_critcount;
2d93b37a 1689
850634cc
AH
1690 td->td_critcount = 0;
1691 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1692 /* NOT REACHED */
2d93b37a
MD
1693}
1694
d165e668
MD
1695#ifdef SMP
1696
bd8015ca
MD
1697/*
1698 * Called from debugger/panic on cpus which have been stopped. We must still
1699 * process the IPIQ while stopped, even if we were stopped while in a critical
1700 * section (XXX).
1701 *
1702 * If we are dumping also try to process any pending interrupts. This may
1703 * or may not work depending on the state of the cpu at the point it was
1704 * stopped.
1705 */
1706void
1707lwkt_smp_stopped(void)
1708{
1709 globaldata_t gd = mycpu;
1710
1711 crit_enter_gd(gd);
1712 if (dumping) {
1713 lwkt_process_ipiq();
1714 splz();
1715 } else {
1716 lwkt_process_ipiq();
1717 }
1718 crit_exit_gd(gd);
1719}
1720
d165e668 1721#endif