utilities - Print the cpu id for running and runnable threads
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
3b998fa9 2 * Copyright (c) 2003-2010 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
60f60350 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
60f60350 20 *
8c10bfcf
MD
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
b37f18d6 47#include <sys/kinfo.h>
8ad65e08 48#include <sys/queue.h>
7d0bac62 49#include <sys/sysctl.h>
99df837e 50#include <sys/kthread.h>
f1d1c3fa 51#include <machine/cpu.h>
99df837e 52#include <sys/lock.h>
f6bf3af1 53#include <sys/caps.h>
9d265729 54#include <sys/spinlock.h>
57aa743c 55#include <sys/ktr.h>
9d265729
MD
56
57#include <sys/thread2.h>
58#include <sys/spinlock2.h>
684a93c4 59#include <sys/mplock2.h>
f1d1c3fa 60
8c72e3d5
AH
61#include <sys/dsched.h>
62
7d0bac62
MD
63#include <vm/vm.h>
64#include <vm/vm_param.h>
65#include <vm/vm_kern.h>
66#include <vm/vm_object.h>
67#include <vm/vm_page.h>
68#include <vm/vm_map.h>
69#include <vm/vm_pager.h>
70#include <vm/vm_extern.h>
7d0bac62 71
99df837e 72#include <machine/stdarg.h>
96728c05 73#include <machine/smp.h>
99df837e 74
d850923c
AE
75#if !defined(KTR_CTXSW)
76#define KTR_CTXSW KTR_ALL
77#endif
78KTR_INFO_MASTER(ctxsw);
a1f0fb66
AE
79KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
80 sizeof(int) + sizeof(struct thread *));
81KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
82 sizeof(int) + sizeof(struct thread *));
83KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
84 sizeof (struct thread *) + sizeof(char *));
85KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
1541028a 86
40aaf5fc
NT
87static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
88
0f7a3396
MD
89#ifdef INVARIANTS
90static int panic_on_cscount = 0;
91#endif
05220613
MD
92static __int64_t switch_count = 0;
93static __int64_t preempt_hit = 0;
94static __int64_t preempt_miss = 0;
95static __int64_t preempt_weird = 0;
f64b567c 96static __int64_t token_contention_count __debugvar = 0;
fb0f29c4 97static int lwkt_use_spin_port;
40aaf5fc 98static struct objcache *thread_cache;
05220613 99
88ebb169 100#ifdef SMP
e381e77c 101static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
88ebb169 102#endif
f9235b6d 103static void lwkt_fairq_accumulate(globaldata_t gd, thread_t td);
e381e77c 104
0855a2af
JG
105extern void cpu_heavy_restore(void);
106extern void cpu_lwkt_restore(void);
107extern void cpu_kthread_restore(void);
108extern void cpu_idle_restore(void);
109
fb0f29c4
MD
110/*
111 * We can make all thread ports use the spin backend instead of the thread
112 * backend. This should only be set to debug the spin backend.
113 */
114TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
115
0f7a3396 116#ifdef INVARIANTS
0c52fa62
SG
117SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
118 "Panic if attempting to switch lwkt's while mastering cpusync");
0f7a3396 119#endif
0c52fa62
SG
120SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
121 "Number of switched threads");
9733f757 122SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
0c52fa62 123 "Successful preemption events");
9733f757 124SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
0c52fa62
SG
125 "Failed preemption events");
126SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
127 "Number of preempted threads.");
38717797
HP
128#ifdef INVARIANTS
129SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
130 &token_contention_count, 0, "spinning due to token contention");
38717797 131#endif
f9235b6d 132static int fairq_enable = 1;
0c52fa62
SG
133SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, &fairq_enable, 0,
134 "Turn on fairq priority accumulators");
cbdd23b1 135static int user_pri_sched = 1;
0c52fa62
SG
136SYSCTL_INT(_lwkt, OID_AUTO, user_pri_sched, CTLFLAG_RW, &user_pri_sched, 0,
137 "");
fbc024e4 138static int preempt_enable = 1;
0c52fa62
SG
139SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, &preempt_enable, 0,
140 "Enable preemption");
fbc024e4 141
05220613 142
4b5f931b
MD
143/*
144 * These helper procedures handle the runq, they can only be called from
145 * within a critical section.
75cdbe6c
MD
146 *
147 * WARNING! Prior to SMP being brought up it is possible to enqueue and
148 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
149 * instead of 'mycpu' when referencing the globaldata structure. Once
150 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 151 */
f1d1c3fa
MD
152static __inline
153void
154_lwkt_dequeue(thread_t td)
155{
156 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 157 struct globaldata *gd = td->td_gd;
4b5f931b 158
f1d1c3fa 159 td->td_flags &= ~TDF_RUNQ;
f9235b6d
MD
160 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
161 gd->gd_fairq_total_pri -= td->td_pri;
162 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
163 atomic_clear_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
164 }
165}
166
f9235b6d
MD
167/*
168 * Priority enqueue.
169 *
170 * NOTE: There are a limited number of lwkt threads runnable since user
171 * processes only schedule one at a time per cpu.
172 */
f1d1c3fa
MD
173static __inline
174void
175_lwkt_enqueue(thread_t td)
176{
f9235b6d
MD
177 thread_t xtd;
178
7f5d7ed7 179 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 180 struct globaldata *gd = td->td_gd;
4b5f931b 181
f1d1c3fa 182 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
183 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
184 if (xtd == NULL) {
185 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
cbdd23b1
MD
186 atomic_set_int_nonlocked(&gd->gd_reqflags,
187 RQF_RUNNING | RQF_WAKEUP);
f9235b6d 188 } else {
cbdd23b1 189 atomic_set_int_nonlocked(&gd->gd_reqflags, RQF_WAKEUP);
f9235b6d
MD
190 while (xtd && xtd->td_pri > td->td_pri)
191 xtd = TAILQ_NEXT(xtd, td_threadq);
192 if (xtd)
193 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
194 else
195 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
196 }
197 gd->gd_fairq_total_pri += td->td_pri;
f1d1c3fa
MD
198 }
199}
8ad65e08 200
40aaf5fc
NT
201static __boolean_t
202_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
203{
204 struct thread *td = (struct thread *)obj;
205
206 td->td_kstack = NULL;
207 td->td_kstack_size = 0;
208 td->td_flags = TDF_ALLOCATED_THREAD;
209 return (1);
210}
211
212static void
213_lwkt_thread_dtor(void *obj, void *privdata)
214{
215 struct thread *td = (struct thread *)obj;
216
217 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
218 ("_lwkt_thread_dtor: not allocated from objcache"));
219 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
220 td->td_kstack_size > 0,
221 ("_lwkt_thread_dtor: corrupted stack"));
222 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
223}
224
225/*
226 * Initialize the lwkt s/system.
227 */
228void
229lwkt_init(void)
230{
231 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
232 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
233 NULL, CACHE_NTHREADS/2,
234 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
235}
236
37af14fe
MD
237/*
238 * Schedule a thread to run. As the current thread we can always safely
239 * schedule ourselves, and a shortcut procedure is provided for that
240 * function.
241 *
242 * (non-blocking, self contained on a per cpu basis)
243 */
244void
245lwkt_schedule_self(thread_t td)
246{
247 crit_enter_quick(td);
f9235b6d
MD
248 KASSERT(td != &td->td_gd->gd_idlethread,
249 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 250 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 251 _lwkt_enqueue(td);
37af14fe
MD
252 crit_exit_quick(td);
253}
254
255/*
256 * Deschedule a thread.
257 *
258 * (non-blocking, self contained on a per cpu basis)
259 */
260void
261lwkt_deschedule_self(thread_t td)
262{
263 crit_enter_quick(td);
37af14fe
MD
264 _lwkt_dequeue(td);
265 crit_exit_quick(td);
266}
267
8ad65e08
MD
268/*
269 * LWKTs operate on a per-cpu basis
270 *
73e4f7b9 271 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
272 */
273void
274lwkt_gdinit(struct globaldata *gd)
275{
f9235b6d 276 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 277 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
278}
279
7d0bac62
MD
280/*
281 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
282 * or LWKT start address before it can be scheduled. If the target cpu is
283 * -1 the thread will be created on the current cpu.
0cfcada1
MD
284 *
285 * If you intend to create a thread without a process context this function
286 * does everything except load the startup and switcher function.
7d0bac62
MD
287 */
288thread_t
d3d32139 289lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 290{
c070746a 291 globaldata_t gd = mycpu;
99df837e 292 void *stack;
7d0bac62 293
c070746a
MD
294 /*
295 * If static thread storage is not supplied allocate a thread. Reuse
296 * a cached free thread if possible. gd_freetd is used to keep an exiting
297 * thread intact through the exit.
298 */
ef0fdad1 299 if (td == NULL) {
cf709dd2
MD
300 crit_enter_gd(gd);
301 if ((td = gd->gd_freetd) != NULL) {
302 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
303 TDF_RUNQ)) == 0);
c070746a 304 gd->gd_freetd = NULL;
cf709dd2 305 } else {
c070746a 306 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
307 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
308 TDF_RUNQ)) == 0);
309 }
310 crit_exit_gd(gd);
40aaf5fc
NT
311 KASSERT((td->td_flags &
312 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
313 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
314 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 315 }
c070746a
MD
316
317 /*
318 * Try to reuse cached stack.
319 */
f470d0c8
MD
320 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
321 if (flags & TDF_ALLOCATED_STACK) {
e4846942 322 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
323 stack = NULL;
324 }
325 }
326 if (stack == NULL) {
e40cfbd7 327 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 328 flags |= TDF_ALLOCATED_STACK;
99df837e 329 }
75cdbe6c 330 if (cpu < 0)
c070746a 331 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 332 else
f470d0c8 333 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 334 return(td);
7d0bac62
MD
335}
336
337/*
338 * Initialize a preexisting thread structure. This function is used by
339 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
340 *
f8c3996b
MD
341 * All threads start out in a critical section at a priority of
342 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
343 * appropriate. This function may send an IPI message when the
344 * requested cpu is not the current cpu and consequently gd_tdallq may
345 * not be initialized synchronously from the point of view of the originating
346 * cpu.
347 *
348 * NOTE! we have to be careful in regards to creating threads for other cpus
349 * if SMP has not yet been activated.
7d0bac62 350 */
41a01a4d
MD
351#ifdef SMP
352
75cdbe6c
MD
353static void
354lwkt_init_thread_remote(void *arg)
355{
356 thread_t td = arg;
357
52eedfb5
MD
358 /*
359 * Protected by critical section held by IPI dispatch
360 */
75cdbe6c
MD
361 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
362}
363
41a01a4d
MD
364#endif
365
fdce8919
MD
366/*
367 * lwkt core thread structural initialization.
368 *
369 * NOTE: All threads are initialized as mpsafe threads.
370 */
7d0bac62 371void
f470d0c8
MD
372lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
373 struct globaldata *gd)
7d0bac62 374{
37af14fe
MD
375 globaldata_t mygd = mycpu;
376
99df837e
MD
377 bzero(td, sizeof(struct thread));
378 td->td_kstack = stack;
f470d0c8 379 td->td_kstack_size = stksize;
d3d32139 380 td->td_flags = flags;
26a0694b 381 td->td_gd = gd;
f9235b6d
MD
382 td->td_pri = TDPRI_KERN_DAEMON;
383 td->td_critcount = 1;
3b998fa9 384 td->td_toks_stop = &td->td_toks_base;
fb0f29c4
MD
385 if (lwkt_use_spin_port)
386 lwkt_initport_spin(&td->td_msgport);
387 else
388 lwkt_initport_thread(&td->td_msgport, td);
99df837e 389 pmap_init_thread(td);
0f7a3396 390#ifdef SMP
5d21b981
MD
391 /*
392 * Normally initializing a thread for a remote cpu requires sending an
393 * IPI. However, the idlethread is setup before the other cpus are
394 * activated so we have to treat it as a special case. XXX manipulation
395 * of gd_tdallq requires the BGL.
396 */
397 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 398 crit_enter_gd(mygd);
75cdbe6c 399 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 400 crit_exit_gd(mygd);
75cdbe6c 401 } else {
2db3b277 402 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 403 }
0f7a3396 404#else
37af14fe 405 crit_enter_gd(mygd);
0f7a3396 406 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 407 crit_exit_gd(mygd);
0f7a3396 408#endif
8c72e3d5
AH
409
410 dsched_new_thread(td);
73e4f7b9
MD
411}
412
413void
414lwkt_set_comm(thread_t td, const char *ctl, ...)
415{
e2565a42 416 __va_list va;
73e4f7b9 417
e2565a42 418 __va_start(va, ctl);
379210cb 419 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 420 __va_end(va);
e7c0dbba 421 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
422}
423
99df837e 424void
73e4f7b9 425lwkt_hold(thread_t td)
99df837e 426{
73e4f7b9
MD
427 ++td->td_refs;
428}
429
430void
431lwkt_rele(thread_t td)
432{
433 KKASSERT(td->td_refs > 0);
434 --td->td_refs;
435}
436
437void
438lwkt_wait_free(thread_t td)
439{
440 while (td->td_refs)
377d4740 441 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
442}
443
444void
445lwkt_free_thread(thread_t td)
446{
cf709dd2 447 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|TDF_RUNQ)) == 0);
40aaf5fc
NT
448 if (td->td_flags & TDF_ALLOCATED_THREAD) {
449 objcache_put(thread_cache, td);
450 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
451 /* client-allocated struct with internally allocated stack */
452 KASSERT(td->td_kstack && td->td_kstack_size > 0,
453 ("lwkt_free_thread: corrupted stack"));
454 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
455 td->td_kstack = NULL;
456 td->td_kstack_size = 0;
99df837e 457 }
e7c0dbba 458 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
459}
460
461
8ad65e08
MD
462/*
463 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
464 * switch to the idlethread. Switching must occur within a critical
465 * section to avoid races with the scheduling queue.
466 *
467 * We always have full control over our cpu's run queue. Other cpus
468 * that wish to manipulate our queue must use the cpu_*msg() calls to
469 * talk to our cpu, so a critical section is all that is needed and
470 * the result is very, very fast thread switching.
471 *
96728c05
MD
472 * The LWKT scheduler uses a fixed priority model and round-robins at
473 * each priority level. User process scheduling is a totally
474 * different beast and LWKT priorities should not be confused with
475 * user process priorities.
f1d1c3fa 476 *
3933a3ab
MD
477 * Note that the td_switch() function cannot do anything that requires
478 * the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
479 * the target thread (not the current thread). It's nice to have a scheduler
480 * that does not need the MP lock to work because it allows us to do some
481 * really cool high-performance MP lock optimizations.
69d78e99
MD
482 *
483 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
484 * is not called by the current thread in the preemption case, only when
485 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
486 */
487void
488lwkt_switch(void)
489{
37af14fe
MD
490 globaldata_t gd = mycpu;
491 thread_t td = gd->gd_curthread;
8ad65e08 492 thread_t ntd;
f9235b6d
MD
493 thread_t xtd;
494 thread_t nlast;
f9235b6d 495 int didaccumulate;
8ad65e08 496
46a3f46d 497 /*
27e88a6e
MD
498 * Switching from within a 'fast' (non thread switched) interrupt or IPI
499 * is illegal. However, we may have to do it anyway if we hit a fatal
500 * kernel trap or we have paniced.
501 *
502 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 503 */
27e88a6e
MD
504 if (gd->gd_intr_nesting_level) {
505 int savegdnest;
506 int savegdtrap;
507
5fddbda2 508 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22
MD
509 panic("lwkt_switch: Attempt to switch from a "
510 "a fast interrupt, ipi, or hard code section, "
511 "td %p\n",
512 td);
27e88a6e
MD
513 } else {
514 savegdnest = gd->gd_intr_nesting_level;
515 savegdtrap = gd->gd_trap_nesting_level;
516 gd->gd_intr_nesting_level = 0;
517 gd->gd_trap_nesting_level = 0;
a7422615
MD
518 if ((td->td_flags & TDF_PANICWARN) == 0) {
519 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
520 kprintf("Warning: thread switch from interrupt, IPI, "
521 "or hard code section.\n"
a7422615 522 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 523 print_backtrace(-1);
a7422615 524 }
27e88a6e
MD
525 lwkt_switch();
526 gd->gd_intr_nesting_level = savegdnest;
527 gd->gd_trap_nesting_level = savegdtrap;
528 return;
529 }
96728c05 530 }
ef0fdad1 531
cb973d15
MD
532 /*
533 * Passive release (used to transition from user to kernel mode
534 * when we block or switch rather then when we enter the kernel).
535 * This function is NOT called if we are switching into a preemption
536 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
537 * our current process designation (if we have one) and become a true
538 * LWKT thread, and may also hand the current process designation to
539 * another process and schedule thread.
cb973d15
MD
540 */
541 if (td->td_release)
542 td->td_release(td);
543
37af14fe 544 crit_enter_gd(gd);
3b998fa9 545 if (TD_TOKS_HELD(td))
9d265729
MD
546 lwkt_relalltokens(td);
547
548 /*
b02926de
MD
549 * We had better not be holding any spin locks, but don't get into an
550 * endless panic loop.
9d265729 551 */
d666840a
MD
552 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
553 ("lwkt_switch: still holding %d exclusive spinlocks!",
554 gd->gd_spinlocks_wr));
9d265729 555
8a8d5d85
MD
556
557#ifdef SMP
0f7a3396
MD
558#ifdef INVARIANTS
559 if (td->td_cscount) {
6ea70f76 560 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
561 td);
562 if (panic_on_cscount)
563 panic("switching while mastering cpusync");
564 }
565#endif
8a8d5d85 566#endif
f9235b6d
MD
567
568 /*
569 * If we had preempted another thread on this cpu, resume the preempted
570 * thread. This occurs transparently, whether the preempted thread
571 * was scheduled or not (it may have been preempted after descheduling
572 * itself).
573 *
574 * We have to setup the MP lock for the original thread after backing
575 * out the adjustment that was made to curthread when the original
576 * was preempted.
577 */
99df837e 578 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
579 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
580 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
581
582 /*
b9eb1c19
MD
583 * The interrupt may have woken a thread up, we need to properly
584 * set the reschedule flag if the originally interrupted thread is
585 * at a lower priority.
8ec60c3f 586 */
f9235b6d
MD
587 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
588 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
8ec60c3f 589 need_lwkt_resched();
f9235b6d 590 }
8a8d5d85 591 /* YYY release mp lock on switchback if original doesn't need it */
f9235b6d
MD
592 goto havethread_preempted;
593 }
594
595 /*
596 * Implement round-robin fairq with priority insertion. The priority
597 * insertion is handled by _lwkt_enqueue()
598 *
599 * We have to adjust the MP lock for the target thread. If we
600 * need the MP lock and cannot obtain it we try to locate a
601 * thread that does not need the MP lock. If we cannot, we spin
602 * instead of HLT.
603 *
604 * A similar issue exists for the tokens held by the target thread.
605 * If we cannot obtain ownership of the tokens we cannot immediately
606 * schedule the thread.
607 */
608 for (;;) {
cbdd23b1 609 atomic_clear_int_nonlocked(&mycpu->gd_reqflags, RQF_WAKEUP);
f9235b6d
MD
610 clear_lwkt_resched();
611 didaccumulate = 0;
612 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
613
4b5f931b 614 /*
f9235b6d 615 * Hotpath if we can get all necessary resources.
41a01a4d 616 *
f9235b6d 617 * If nothing is runnable switch to the idle thread
41a01a4d 618 */
f9235b6d
MD
619 if (ntd == NULL) {
620 ntd = &gd->gd_idlethread;
6f207a2c 621#ifdef SMP
b5d16701
MD
622 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
623 ASSERT_NO_TOKENS_HELD(ntd);
6f207a2c 624#endif
b37f18d6
MD
625 cpu_time.cp_msg[0] = 0;
626 cpu_time.cp_stallpc = 0;
f9235b6d
MD
627 goto haveidle;
628 }
41a01a4d 629
8ec60c3f 630 /*
f9235b6d 631 * Hotpath schedule
6f207a2c
MD
632 *
633 * NOTE: For UP there is no mplock and lwkt_getalltokens()
634 * always succeeds.
8ec60c3f 635 */
f9235b6d 636 if (ntd->td_fairq_accum >= 0 &&
b5d16701 637 (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
f9235b6d 638 ) {
f9235b6d
MD
639 goto havethread;
640 }
641
f9235b6d
MD
642 /*
643 * Coldpath - unable to schedule ntd, continue looking for threads
cbdd23b1 644 * to schedule.
f9235b6d 645 */
f9235b6d
MD
646 nlast = NULL;
647
648 for (;;) {
41a01a4d 649 /*
f9235b6d
MD
650 * If the fair-share scheduler ran out ntd gets moved to the
651 * end and its accumulator will be bumped, if it didn't we
652 * maintain the same queue position.
df6b8ba0 653 *
f9235b6d 654 * nlast keeps track of the last element prior to any moves.
41a01a4d 655 */
f9235b6d 656 if (ntd->td_fairq_accum < 0) {
f9235b6d
MD
657 lwkt_fairq_accumulate(gd, ntd);
658 didaccumulate = 1;
c5724852
MD
659
660 /*
661 * Move to end
662 */
663 xtd = TAILQ_NEXT(ntd, td_threadq);
f9235b6d
MD
664 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
665 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
c5724852
MD
666
667 /*
668 * Set terminal element (nlast)
669 */
f9235b6d
MD
670 if (nlast == NULL) {
671 nlast = ntd;
672 if (xtd == NULL)
673 xtd = ntd;
674 }
675 ntd = xtd;
676 } else {
677 ntd = TAILQ_NEXT(ntd, td_threadq);
678 }
a453459d 679
f9235b6d
MD
680 /*
681 * If we exhausted the run list switch to the idle thread.
f9235b6d
MD
682 *
683 * NOTE: nlast can be NULL.
684 */
685 if (ntd == nlast) {
e0a90d3b 686 cpu_pause();
f9235b6d 687 ntd = &gd->gd_idlethread;
6f207a2c 688#ifdef SMP
b5d16701
MD
689 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
690 ASSERT_NO_TOKENS_HELD(ntd);
691 /* contention case, do not clear contention mask */
6f207a2c 692#endif
684a93c4
MD
693
694 /*
f9235b6d
MD
695 * If fairq accumulations occured we do not schedule the
696 * idle thread. This will cause us to try again from
697 * the (almost) top.
684a93c4 698 */
f9235b6d 699 if (didaccumulate)
b37f18d6 700 break; /* try again from the top, almost */
f9235b6d 701 goto haveidle;
8a8d5d85 702 }
f9235b6d 703
df6b8ba0 704 /*
f9235b6d 705 * Try to switch to this thread.
6f207a2c
MD
706 *
707 * NOTE: For UP there is no mplock and lwkt_getalltokens()
708 * always succeeds.
df6b8ba0 709 */
cbdd23b1 710 if (ntd->td_fairq_accum >= 0 &&
b5d16701 711 (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
f9235b6d 712 ) {
f9235b6d 713 goto havethread;
df6b8ba0 714 }
9ac1ee6e
MD
715
716 /*
717 * Thread was runnable but we were unable to get the required
cbdd23b1 718 * resources (tokens and/or mplock), continue the scan.
9ac1ee6e 719 */
cbdd23b1 720 /* */
4b5f931b 721 }
c5724852
MD
722
723 /*
724 * All threads exhausted but we can loop due to a negative
725 * accumulator.
726 *
727 * While we are looping in the scheduler be sure to service
728 * any interrupts which were made pending due to our critical
729 * section, otherwise we could livelock (e.g.) IPIs.
c5724852
MD
730 */
731 splz_check();
f1d1c3fa 732 }
8a8d5d85
MD
733
734 /*
f9235b6d
MD
735 * We must always decrement td_fairq_accum on non-idle threads just
736 * in case a thread never gets a tick due to being in a continuous
737 * critical section. The page-zeroing code does that.
738 *
739 * If the thread we came up with is a higher or equal priority verses
740 * the thread at the head of the queue we move our thread to the
741 * front. This way we can always check the front of the queue.
742 */
743havethread:
744 ++gd->gd_cnt.v_swtch;
745 --ntd->td_fairq_accum;
9ac1ee6e 746 ntd->td_wmesg = NULL;
f9235b6d
MD
747 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
748 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
749 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
750 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
751 }
752havethread_preempted:
753 ;
754 /*
755 * If the new target does not need the MP lock and we are holding it,
756 * release the MP lock. If the new target requires the MP lock we have
757 * already acquired it for the target.
8a8d5d85 758 */
f9235b6d
MD
759haveidle:
760 KASSERT(ntd->td_critcount,
b5d16701
MD
761 ("priority problem in lwkt_switch %d %d",
762 td->td_critcount, ntd->td_critcount));
763
94f6d86e
MD
764 if (td != ntd) {
765 ++switch_count;
a1f0fb66 766 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
f1d1c3fa 767 td->td_switch(ntd);
94f6d86e 768 }
37af14fe
MD
769 /* NOTE: current cpu may have changed after switch */
770 crit_exit_quick(td);
8ad65e08
MD
771}
772
b68b7282 773/*
96728c05
MD
774 * Request that the target thread preempt the current thread. Preemption
775 * only works under a specific set of conditions:
b68b7282 776 *
96728c05
MD
777 * - We are not preempting ourselves
778 * - The target thread is owned by the current cpu
779 * - We are not currently being preempted
780 * - The target is not currently being preempted
d3d1cbc8
MD
781 * - We are not holding any spin locks
782 * - The target thread is not holding any tokens
96728c05
MD
783 * - We are able to satisfy the target's MP lock requirements (if any).
784 *
785 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
786 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 787 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
788 * to determine whether preemption is possible (aka usually just the crit
789 * priority of lwkt_schedule() itself).
b68b7282 790 *
26a0694b
MD
791 * XXX at the moment we run the target thread in a critical section during
792 * the preemption in order to prevent the target from taking interrupts
793 * that *WE* can't. Preemption is strictly limited to interrupt threads
794 * and interrupt-like threads, outside of a critical section, and the
795 * preempted source thread will be resumed the instant the target blocks
796 * whether or not the source is scheduled (i.e. preemption is supposed to
797 * be as transparent as possible).
b68b7282
MD
798 */
799void
f9235b6d 800lwkt_preempt(thread_t ntd, int critcount)
b68b7282 801{
46a3f46d 802 struct globaldata *gd = mycpu;
0a3f9b47 803 thread_t td;
2d910aaf 804 int save_gd_intr_nesting_level;
b68b7282 805
26a0694b 806 /*
96728c05
MD
807 * The caller has put us in a critical section. We can only preempt
808 * if the caller of the caller was not in a critical section (basically
f9235b6d 809 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 810 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 811 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
812 *
813 * YYY The target thread must be in a critical section (else it must
814 * inherit our critical section? I dunno yet).
41a01a4d 815 *
0a3f9b47 816 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b 817 */
f9235b6d 818 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 819
fbc024e4
MD
820 if (preempt_enable == 0) {
821 ++preempt_miss;
822 return;
823 }
824
0a3f9b47 825 td = gd->gd_curthread;
f9235b6d 826 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
827 ++preempt_miss;
828 return;
829 }
f9235b6d 830 if (td->td_critcount > critcount) {
96728c05 831 ++preempt_miss;
8ec60c3f 832 need_lwkt_resched();
96728c05
MD
833 return;
834 }
835#ifdef SMP
46a3f46d 836 if (ntd->td_gd != gd) {
96728c05 837 ++preempt_miss;
8ec60c3f 838 need_lwkt_resched();
96728c05
MD
839 return;
840 }
841#endif
41a01a4d 842 /*
77912481
MD
843 * We don't have to check spinlocks here as they will also bump
844 * td_critcount.
d3d1cbc8
MD
845 *
846 * Do not try to preempt if the target thread is holding any tokens.
847 * We could try to acquire the tokens but this case is so rare there
848 * is no need to support it.
41a01a4d 849 */
77912481
MD
850 KKASSERT(gd->gd_spinlocks_wr == 0);
851
3b998fa9 852 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8
MD
853 ++preempt_miss;
854 need_lwkt_resched();
855 return;
856 }
26a0694b
MD
857 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
858 ++preempt_weird;
8ec60c3f 859 need_lwkt_resched();
26a0694b
MD
860 return;
861 }
862 if (ntd->td_preempted) {
4b5f931b 863 ++preempt_hit;
8ec60c3f 864 need_lwkt_resched();
26a0694b 865 return;
b68b7282 866 }
26a0694b 867
8ec60c3f
MD
868 /*
869 * Since we are able to preempt the current thread, there is no need to
870 * call need_lwkt_resched().
2d910aaf
MD
871 *
872 * We must temporarily clear gd_intr_nesting_level around the switch
873 * since switchouts from the target thread are allowed (they will just
874 * return to our thread), and since the target thread has its own stack.
8ec60c3f 875 */
26a0694b
MD
876 ++preempt_hit;
877 ntd->td_preempted = td;
878 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 879 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
880 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
881 gd->gd_intr_nesting_level = 0;
26a0694b 882 td->td_switch(ntd);
2d910aaf 883 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 884
26a0694b
MD
885 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
886 ntd->td_preempted = NULL;
887 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
888}
889
f1d1c3fa 890/*
faaeffac 891 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
892 * This will work inside a critical section but not inside a hard code
893 * section.
ef0fdad1 894 *
f1d1c3fa
MD
895 * (self contained on a per cpu basis)
896 */
897void
faaeffac 898splz_check(void)
f1d1c3fa 899{
7966cb69
MD
900 globaldata_t gd = mycpu;
901 thread_t td = gd->gd_curthread;
ef0fdad1 902
4a28fe22
MD
903 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
904 gd->gd_intr_nesting_level == 0 &&
905 td->td_nest_count < 2)
906 {
f1d1c3fa 907 splz();
4a28fe22
MD
908 }
909}
910
911/*
912 * This version is integrated into crit_exit, reqflags has already
913 * been tested but td_critcount has not.
914 *
915 * We only want to execute the splz() on the 1->0 transition of
916 * critcount and not in a hard code section or if too deeply nested.
917 */
918void
919lwkt_maybe_splz(thread_t td)
920{
921 globaldata_t gd = td->td_gd;
922
923 if (td->td_critcount == 0 &&
924 gd->gd_intr_nesting_level == 0 &&
925 td->td_nest_count < 2)
926 {
927 splz();
928 }
f1d1c3fa
MD
929}
930
8ad65e08 931/*
f9235b6d
MD
932 * This function is used to negotiate a passive release of the current
933 * process/lwp designation with the user scheduler, allowing the user
934 * scheduler to schedule another user thread. The related kernel thread
935 * (curthread) continues running in the released state.
8ad65e08
MD
936 */
937void
f9235b6d 938lwkt_passive_release(struct thread *td)
8ad65e08 939{
f9235b6d
MD
940 struct lwp *lp = td->td_lwp;
941
942 td->td_release = NULL;
943 lwkt_setpri_self(TDPRI_KERN_USER);
944 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
945}
946
f9235b6d 947
3824f392 948/*
f9235b6d
MD
949 * This implements a normal yield. This routine is virtually a nop if
950 * there is nothing to yield to but it will always run any pending interrupts
951 * if called from a critical section.
952 *
953 * This yield is designed for kernel threads without a user context.
954 *
955 * (self contained on a per cpu basis)
3824f392
MD
956 */
957void
f9235b6d 958lwkt_yield(void)
3824f392 959{
f9235b6d
MD
960 globaldata_t gd = mycpu;
961 thread_t td = gd->gd_curthread;
962 thread_t xtd;
3824f392 963
f9235b6d
MD
964 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
965 splz();
966 if (td->td_fairq_accum < 0) {
967 lwkt_schedule_self(curthread);
968 lwkt_switch();
969 } else {
970 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
971 if (xtd && xtd->td_pri > td->td_pri) {
972 lwkt_schedule_self(curthread);
973 lwkt_switch();
974 }
975 }
3824f392
MD
976}
977
978/*
f9235b6d
MD
979 * This yield is designed for kernel threads with a user context.
980 *
981 * The kernel acting on behalf of the user is potentially cpu-bound,
982 * this function will efficiently allow other threads to run and also
983 * switch to other processes by releasing.
3824f392
MD
984 *
985 * The lwkt_user_yield() function is designed to have very low overhead
986 * if no yield is determined to be needed.
987 */
988void
989lwkt_user_yield(void)
990{
f9235b6d
MD
991 globaldata_t gd = mycpu;
992 thread_t td = gd->gd_curthread;
993
994 /*
995 * Always run any pending interrupts in case we are in a critical
996 * section.
997 */
998 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
999 splz();
3824f392 1000
3824f392 1001 /*
f9235b6d
MD
1002 * Switch (which forces a release) if another kernel thread needs
1003 * the cpu, if userland wants us to resched, or if our kernel
1004 * quantum has run out.
3824f392 1005 */
f9235b6d
MD
1006 if (lwkt_resched_wanted() ||
1007 user_resched_wanted() ||
1008 td->td_fairq_accum < 0)
1009 {
3824f392 1010 lwkt_switch();
3824f392
MD
1011 }
1012
f9235b6d 1013#if 0
3824f392 1014 /*
f9235b6d
MD
1015 * Reacquire the current process if we are released.
1016 *
1017 * XXX not implemented atm. The kernel may be holding locks and such,
1018 * so we want the thread to continue to receive cpu.
3824f392 1019 */
f9235b6d
MD
1020 if (td->td_release == NULL && lp) {
1021 lp->lwp_proc->p_usched->acquire_curproc(lp);
1022 td->td_release = lwkt_passive_release;
1023 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1024 }
f9235b6d 1025#endif
b9eb1c19
MD
1026}
1027
8ad65e08 1028/*
f1d1c3fa
MD
1029 * Generic schedule. Possibly schedule threads belonging to other cpus and
1030 * deal with threads that might be blocked on a wait queue.
1031 *
0a3f9b47
MD
1032 * We have a little helper inline function which does additional work after
1033 * the thread has been enqueued, including dealing with preemption and
1034 * setting need_lwkt_resched() (which prevents the kernel from returning
1035 * to userland until it has processed higher priority threads).
6330a558
MD
1036 *
1037 * It is possible for this routine to be called after a failed _enqueue
1038 * (due to the target thread migrating, sleeping, or otherwise blocked).
1039 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1040 *
1041 * reschedok is an optimized constant propagated from lwkt_schedule() or
1042 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1043 * reschedule to be requested if the target thread has a higher priority.
1044 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1045 * be 0, prevented undesired reschedules.
8ad65e08 1046 */
0a3f9b47
MD
1047static __inline
1048void
f9235b6d 1049_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
0a3f9b47 1050{
b9eb1c19 1051 thread_t otd;
c730be20 1052
6330a558 1053 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1054 if (ntd->td_preemptable && reschedok) {
f9235b6d 1055 ntd->td_preemptable(ntd, ccount); /* YYY +token */
361d01dd 1056 } else if (reschedok) {
b9eb1c19 1057 otd = curthread;
f9235b6d 1058 if (ntd->td_pri > otd->td_pri)
c730be20 1059 need_lwkt_resched();
6330a558 1060 }
f9235b6d
MD
1061
1062 /*
1063 * Give the thread a little fair share scheduler bump if it
1064 * has been asleep for a while. This is primarily to avoid
1065 * a degenerate case for interrupt threads where accumulator
1066 * crosses into negative territory unnecessarily.
1067 */
1068 if (ntd->td_fairq_lticks != ticks) {
1069 ntd->td_fairq_lticks = ticks;
1070 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1071 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1072 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1073 }
0a3f9b47
MD
1074 }
1075}
1076
361d01dd 1077static __inline
8ad65e08 1078void
361d01dd 1079_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1080{
37af14fe
MD
1081 globaldata_t mygd = mycpu;
1082
cf709dd2
MD
1083 KASSERT(td != &td->td_gd->gd_idlethread,
1084 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1085 crit_enter_gd(mygd);
9388413d 1086 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1087 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1088 _lwkt_enqueue(td);
1089 } else {
f1d1c3fa 1090 /*
7cd8d145
MD
1091 * If we own the thread, there is no race (since we are in a
1092 * critical section). If we do not own the thread there might
1093 * be a race but the target cpu will deal with it.
f1d1c3fa 1094 */
0f7a3396 1095#ifdef SMP
7cd8d145 1096 if (td->td_gd == mygd) {
9d265729 1097 _lwkt_enqueue(td);
f9235b6d 1098 _lwkt_schedule_post(mygd, td, 1, reschedok);
f1d1c3fa 1099 } else {
e381e77c 1100 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1101 }
0f7a3396 1102#else
7cd8d145 1103 _lwkt_enqueue(td);
f9235b6d 1104 _lwkt_schedule_post(mygd, td, 1, reschedok);
0f7a3396 1105#endif
8ad65e08 1106 }
37af14fe 1107 crit_exit_gd(mygd);
8ad65e08
MD
1108}
1109
361d01dd
MD
1110void
1111lwkt_schedule(thread_t td)
1112{
1113 _lwkt_schedule(td, 1);
1114}
1115
1116void
1117lwkt_schedule_noresched(thread_t td)
1118{
1119 _lwkt_schedule(td, 0);
1120}
1121
88ebb169
SW
1122#ifdef SMP
1123
e381e77c
MD
1124/*
1125 * When scheduled remotely if frame != NULL the IPIQ is being
1126 * run via doreti or an interrupt then preemption can be allowed.
1127 *
1128 * To allow preemption we have to drop the critical section so only
1129 * one is present in _lwkt_schedule_post.
1130 */
1131static void
1132lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1133{
1134 thread_t td = curthread;
1135 thread_t ntd = arg;
1136
1137 if (frame && ntd->td_preemptable) {
1138 crit_exit_noyield(td);
1139 _lwkt_schedule(ntd, 1);
1140 crit_enter_quick(td);
1141 } else {
1142 _lwkt_schedule(ntd, 1);
1143 }
1144}
1145
d9eea1a5 1146/*
52eedfb5
MD
1147 * Thread migration using a 'Pull' method. The thread may or may not be
1148 * the current thread. It MUST be descheduled and in a stable state.
1149 * lwkt_giveaway() must be called on the cpu owning the thread.
1150 *
1151 * At any point after lwkt_giveaway() is called, the target cpu may
1152 * 'pull' the thread by calling lwkt_acquire().
1153 *
ae8e83e6
MD
1154 * We have to make sure the thread is not sitting on a per-cpu tsleep
1155 * queue or it will blow up when it moves to another cpu.
1156 *
52eedfb5 1157 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1158 */
52eedfb5
MD
1159void
1160lwkt_giveaway(thread_t td)
1161{
3b4192fb 1162 globaldata_t gd = mycpu;
52eedfb5 1163
3b4192fb
MD
1164 crit_enter_gd(gd);
1165 if (td->td_flags & TDF_TSLEEPQ)
1166 tsleep_remove(td);
1167 KKASSERT(td->td_gd == gd);
1168 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1169 td->td_flags |= TDF_MIGRATING;
1170 crit_exit_gd(gd);
52eedfb5
MD
1171}
1172
a2a5ad0d
MD
1173void
1174lwkt_acquire(thread_t td)
1175{
37af14fe
MD
1176 globaldata_t gd;
1177 globaldata_t mygd;
a2a5ad0d 1178
52eedfb5 1179 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1180 gd = td->td_gd;
37af14fe 1181 mygd = mycpu;
52eedfb5 1182 if (gd != mycpu) {
35238fa5 1183 cpu_lfence();
52eedfb5 1184 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1185 crit_enter_gd(mygd);
df910c23
MD
1186 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1187#ifdef SMP
1188 lwkt_process_ipiq();
1189#endif
52eedfb5 1190 cpu_lfence();
df910c23 1191 }
562273ea 1192 cpu_mfence();
37af14fe 1193 td->td_gd = mygd;
52eedfb5
MD
1194 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1195 td->td_flags &= ~TDF_MIGRATING;
1196 crit_exit_gd(mygd);
1197 } else {
1198 crit_enter_gd(mygd);
1199 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1200 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1201 crit_exit_gd(mygd);
a2a5ad0d
MD
1202 }
1203}
1204
52eedfb5
MD
1205#endif
1206
f1d1c3fa
MD
1207/*
1208 * Generic deschedule. Descheduling threads other then your own should be
1209 * done only in carefully controlled circumstances. Descheduling is
1210 * asynchronous.
1211 *
1212 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1213 */
1214void
1215lwkt_deschedule(thread_t td)
1216{
f1d1c3fa 1217 crit_enter();
b8a98473 1218#ifdef SMP
f1d1c3fa
MD
1219 if (td == curthread) {
1220 _lwkt_dequeue(td);
1221 } else {
a72187e9 1222 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1223 _lwkt_dequeue(td);
1224 } else {
b8a98473 1225 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1226 }
1227 }
b8a98473
MD
1228#else
1229 _lwkt_dequeue(td);
1230#endif
f1d1c3fa
MD
1231 crit_exit();
1232}
1233
4b5f931b
MD
1234/*
1235 * Set the target thread's priority. This routine does not automatically
1236 * switch to a higher priority thread, LWKT threads are not designed for
1237 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1238 */
1239void
1240lwkt_setpri(thread_t td, int pri)
1241{
a72187e9 1242 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1243 if (td->td_pri != pri) {
1244 KKASSERT(pri >= 0);
1245 crit_enter();
1246 if (td->td_flags & TDF_RUNQ) {
1247 _lwkt_dequeue(td);
1248 td->td_pri = pri;
1249 _lwkt_enqueue(td);
1250 } else {
1251 td->td_pri = pri;
1252 }
1253 crit_exit();
26a0694b 1254 }
26a0694b
MD
1255}
1256
03bd0a5e
MD
1257/*
1258 * Set the initial priority for a thread prior to it being scheduled for
1259 * the first time. The thread MUST NOT be scheduled before or during
1260 * this call. The thread may be assigned to a cpu other then the current
1261 * cpu.
1262 *
1263 * Typically used after a thread has been created with TDF_STOPPREQ,
1264 * and before the thread is initially scheduled.
1265 */
1266void
1267lwkt_setpri_initial(thread_t td, int pri)
1268{
1269 KKASSERT(pri >= 0);
1270 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1271 td->td_pri = pri;
03bd0a5e
MD
1272}
1273
26a0694b
MD
1274void
1275lwkt_setpri_self(int pri)
1276{
1277 thread_t td = curthread;
1278
4b5f931b
MD
1279 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1280 crit_enter();
1281 if (td->td_flags & TDF_RUNQ) {
1282 _lwkt_dequeue(td);
f9235b6d 1283 td->td_pri = pri;
4b5f931b
MD
1284 _lwkt_enqueue(td);
1285 } else {
f9235b6d 1286 td->td_pri = pri;
4b5f931b
MD
1287 }
1288 crit_exit();
1289}
1290
f9235b6d
MD
1291/*
1292 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1293 *
1294 * Example: two competing threads, same priority N. decrement by (2*N)
1295 * increment by N*8, each thread will get 4 ticks.
1296 */
1297void
1298lwkt_fairq_schedulerclock(thread_t td)
1299{
1300 if (fairq_enable) {
1301 while (td) {
1302 if (td != &td->td_gd->gd_idlethread) {
1303 td->td_fairq_accum -= td->td_gd->gd_fairq_total_pri;
1304 if (td->td_fairq_accum < -TDFAIRQ_MAX(td->td_gd))
1305 td->td_fairq_accum = -TDFAIRQ_MAX(td->td_gd);
1306 if (td->td_fairq_accum < 0)
1307 need_lwkt_resched();
1308 td->td_fairq_lticks = ticks;
1309 }
1310 td = td->td_preempted;
1311 }
1312 }
1313}
1314
1315static void
1316lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1317{
1318 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1319 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1320 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1321}
1322
5d21b981 1323/*
52eedfb5
MD
1324 * Migrate the current thread to the specified cpu.
1325 *
1326 * This is accomplished by descheduling ourselves from the current cpu,
1327 * moving our thread to the tdallq of the target cpu, IPI messaging the
1328 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1329 * races while the thread is being migrated.
ae8e83e6
MD
1330 *
1331 * We must be sure to remove ourselves from the current cpu's tsleepq
1332 * before potentially moving to another queue. The thread can be on
1333 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1334 */
3d28ff59 1335#ifdef SMP
5d21b981 1336static void lwkt_setcpu_remote(void *arg);
3d28ff59 1337#endif
5d21b981
MD
1338
1339void
1340lwkt_setcpu_self(globaldata_t rgd)
1341{
1342#ifdef SMP
1343 thread_t td = curthread;
1344
1345 if (td->td_gd != rgd) {
1346 crit_enter_quick(td);
ae8e83e6 1347 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1348 tsleep_remove(td);
5d21b981
MD
1349 td->td_flags |= TDF_MIGRATING;
1350 lwkt_deschedule_self(td);
52eedfb5 1351 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1352 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1353 lwkt_switch();
1354 /* we are now on the target cpu */
52eedfb5 1355 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1356 crit_exit_quick(td);
1357 }
1358#endif
1359}
1360
ecdefdda
MD
1361void
1362lwkt_migratecpu(int cpuid)
1363{
1364#ifdef SMP
1365 globaldata_t rgd;
1366
1367 rgd = globaldata_find(cpuid);
1368 lwkt_setcpu_self(rgd);
1369#endif
1370}
1371
5d21b981
MD
1372/*
1373 * Remote IPI for cpu migration (called while in a critical section so we
1374 * do not have to enter another one). The thread has already been moved to
1375 * our cpu's allq, but we must wait for the thread to be completely switched
1376 * out on the originating cpu before we schedule it on ours or the stack
1377 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1378 * change to main memory.
1379 *
1380 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1381 * against wakeups. It is best if this interface is used only when there
1382 * are no pending events that might try to schedule the thread.
1383 */
3d28ff59 1384#ifdef SMP
5d21b981
MD
1385static void
1386lwkt_setcpu_remote(void *arg)
1387{
1388 thread_t td = arg;
1389 globaldata_t gd = mycpu;
1390
df910c23
MD
1391 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1392#ifdef SMP
1393 lwkt_process_ipiq();
1394#endif
35238fa5 1395 cpu_lfence();
562273ea 1396 cpu_pause();
df910c23 1397 }
5d21b981 1398 td->td_gd = gd;
562273ea 1399 cpu_mfence();
5d21b981 1400 td->td_flags &= ~TDF_MIGRATING;
9388413d 1401 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1402 _lwkt_enqueue(td);
1403}
3d28ff59 1404#endif
5d21b981 1405
553ea3c8 1406struct lwp *
4b5f931b
MD
1407lwkt_preempted_proc(void)
1408{
73e4f7b9 1409 thread_t td = curthread;
4b5f931b
MD
1410 while (td->td_preempted)
1411 td = td->td_preempted;
553ea3c8 1412 return(td->td_lwp);
4b5f931b
MD
1413}
1414
99df837e
MD
1415/*
1416 * Create a kernel process/thread/whatever. It shares it's address space
1417 * with proc0 - ie: kernel only.
1418 *
365fa13f
MD
1419 * NOTE! By default new threads are created with the MP lock held. A
1420 * thread which does not require the MP lock should release it by calling
1421 * rel_mplock() at the start of the new thread.
99df837e
MD
1422 */
1423int
c9e9fb21
MD
1424lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1425 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1426{
73e4f7b9 1427 thread_t td;
e2565a42 1428 __va_list ap;
99df837e 1429
d3d32139 1430 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1431 tdflags);
a2a5ad0d
MD
1432 if (tdp)
1433 *tdp = td;
709799ea 1434 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1435
1436 /*
1437 * Set up arg0 for 'ps' etc
1438 */
e2565a42 1439 __va_start(ap, fmt);
379210cb 1440 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1441 __va_end(ap);
99df837e
MD
1442
1443 /*
1444 * Schedule the thread to run
1445 */
ef0fdad1
MD
1446 if ((td->td_flags & TDF_STOPREQ) == 0)
1447 lwkt_schedule(td);
1448 else
1449 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1450 return 0;
1451}
1452
1453/*
1454 * Destroy an LWKT thread. Warning! This function is not called when
1455 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1456 * uses a different reaping mechanism.
1457 */
1458void
1459lwkt_exit(void)
1460{
1461 thread_t td = curthread;
c070746a 1462 thread_t std;
8826f33a 1463 globaldata_t gd;
99df837e 1464
2883d2d8
MD
1465 /*
1466 * Do any cleanup that might block here
1467 */
99df837e 1468 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1469 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1470 caps_exit(td);
2883d2d8
MD
1471 biosched_done(td);
1472 dsched_exit_thread(td);
c070746a
MD
1473
1474 /*
1475 * Get us into a critical section to interlock gd_freetd and loop
1476 * until we can get it freed.
1477 *
1478 * We have to cache the current td in gd_freetd because objcache_put()ing
1479 * it would rip it out from under us while our thread is still active.
1480 */
1481 gd = mycpu;
37af14fe 1482 crit_enter_quick(td);
c070746a 1483 while ((std = gd->gd_freetd) != NULL) {
cf709dd2 1484 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
c070746a
MD
1485 gd->gd_freetd = NULL;
1486 objcache_put(thread_cache, std);
1487 }
3b4192fb
MD
1488
1489 /*
1490 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1491 * the last time. We cannot block after this point or we may end
1492 * up with a stale td on the tsleepq.
3b4192fb
MD
1493 */
1494 if (td->td_flags & TDF_TSLEEPQ)
1495 tsleep_remove(td);
37af14fe 1496 lwkt_deschedule_self(td);
e56e4dea 1497 lwkt_remove_tdallq(td);
2883d2d8
MD
1498
1499 /*
1500 * Final cleanup
1501 */
1502 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1503 if (td->td_flags & TDF_ALLOCATED_THREAD)
1504 gd->gd_freetd = td;
99df837e
MD
1505 cpu_thread_exit();
1506}
1507
e56e4dea
MD
1508void
1509lwkt_remove_tdallq(thread_t td)
1510{
1511 KKASSERT(td->td_gd == mycpu);
1512 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1513}
1514
9cf43f91
MD
1515/*
1516 * Code reduction and branch prediction improvements. Call/return
1517 * overhead on modern cpus often degenerates into 0 cycles due to
1518 * the cpu's branch prediction hardware and return pc cache. We
1519 * can take advantage of this by not inlining medium-complexity
1520 * functions and we can also reduce the branch prediction impact
1521 * by collapsing perfectly predictable branches into a single
1522 * procedure instead of duplicating it.
1523 *
1524 * Is any of this noticeable? Probably not, so I'll take the
1525 * smaller code size.
1526 */
1527void
b6468f56 1528crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1529{
b6468f56 1530 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1531}
1532
2d93b37a
MD
1533void
1534crit_panic(void)
1535{
1536 thread_t td = curthread;
850634cc 1537 int lcrit = td->td_critcount;
2d93b37a 1538
850634cc
AH
1539 td->td_critcount = 0;
1540 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1541 /* NOT REACHED */
2d93b37a
MD
1542}
1543
d165e668
MD
1544#ifdef SMP
1545
bd8015ca
MD
1546/*
1547 * Called from debugger/panic on cpus which have been stopped. We must still
1548 * process the IPIQ while stopped, even if we were stopped while in a critical
1549 * section (XXX).
1550 *
1551 * If we are dumping also try to process any pending interrupts. This may
1552 * or may not work depending on the state of the cpu at the point it was
1553 * stopped.
1554 */
1555void
1556lwkt_smp_stopped(void)
1557{
1558 globaldata_t gd = mycpu;
1559
1560 crit_enter_gd(gd);
1561 if (dumping) {
1562 lwkt_process_ipiq();
1563 splz();
1564 } else {
1565 lwkt_process_ipiq();
1566 }
1567 crit_exit_gd(gd);
1568}
1569
d165e668 1570#endif