buildworld - Adjust ps for recent commits
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
3b998fa9 2 * Copyright (c) 2003-2010 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
60f60350 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
60f60350 20 *
8c10bfcf
MD
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
b37f18d6 47#include <sys/kinfo.h>
8ad65e08 48#include <sys/queue.h>
7d0bac62 49#include <sys/sysctl.h>
99df837e 50#include <sys/kthread.h>
f1d1c3fa 51#include <machine/cpu.h>
99df837e 52#include <sys/lock.h>
f6bf3af1 53#include <sys/caps.h>
9d265729 54#include <sys/spinlock.h>
57aa743c 55#include <sys/ktr.h>
9d265729
MD
56
57#include <sys/thread2.h>
58#include <sys/spinlock2.h>
684a93c4 59#include <sys/mplock2.h>
f1d1c3fa 60
8c72e3d5
AH
61#include <sys/dsched.h>
62
7d0bac62
MD
63#include <vm/vm.h>
64#include <vm/vm_param.h>
65#include <vm/vm_kern.h>
66#include <vm/vm_object.h>
67#include <vm/vm_page.h>
68#include <vm/vm_map.h>
69#include <vm/vm_pager.h>
70#include <vm/vm_extern.h>
7d0bac62 71
99df837e 72#include <machine/stdarg.h>
96728c05 73#include <machine/smp.h>
99df837e 74
d850923c
AE
75#if !defined(KTR_CTXSW)
76#define KTR_CTXSW KTR_ALL
77#endif
78KTR_INFO_MASTER(ctxsw);
a1f0fb66
AE
79KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
80 sizeof(int) + sizeof(struct thread *));
81KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
82 sizeof(int) + sizeof(struct thread *));
83KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
84 sizeof (struct thread *) + sizeof(char *));
85KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
1541028a 86
40aaf5fc
NT
87static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
88
0f7a3396
MD
89#ifdef INVARIANTS
90static int panic_on_cscount = 0;
91#endif
05220613
MD
92static __int64_t switch_count = 0;
93static __int64_t preempt_hit = 0;
94static __int64_t preempt_miss = 0;
95static __int64_t preempt_weird = 0;
f64b567c 96static __int64_t token_contention_count __debugvar = 0;
fb0f29c4 97static int lwkt_use_spin_port;
40aaf5fc 98static struct objcache *thread_cache;
05220613 99
88ebb169 100#ifdef SMP
e381e77c 101static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
88ebb169 102#endif
f9235b6d 103static void lwkt_fairq_accumulate(globaldata_t gd, thread_t td);
e381e77c 104
0855a2af
JG
105extern void cpu_heavy_restore(void);
106extern void cpu_lwkt_restore(void);
107extern void cpu_kthread_restore(void);
108extern void cpu_idle_restore(void);
109
b2b3ffcd 110#ifdef __x86_64__
85514115
MD
111
112static int
0855a2af
JG
113jg_tos_ok(struct thread *td)
114{
85514115
MD
115 void *tos;
116 int tos_ok;
117
0855a2af
JG
118 if (td == NULL) {
119 return 1;
120 }
121 KKASSERT(td->td_sp != NULL);
85514115
MD
122 tos = ((void **)td->td_sp)[0];
123 tos_ok = 0;
124 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
125 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
0855a2af
JG
126 tos_ok = 1;
127 }
128 return tos_ok;
129}
130
85514115
MD
131#endif
132
fb0f29c4
MD
133/*
134 * We can make all thread ports use the spin backend instead of the thread
135 * backend. This should only be set to debug the spin backend.
136 */
137TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
138
0f7a3396
MD
139#ifdef INVARIANTS
140SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
141#endif
4b5f931b 142SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
9733f757
VS
143SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
144 "Successful preemption events");
145SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
146 "Failed preemption events");
26a0694b 147SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
148#ifdef INVARIANTS
149SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
150 &token_contention_count, 0, "spinning due to token contention");
38717797 151#endif
f9235b6d
MD
152static int fairq_enable = 1;
153SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, &fairq_enable, 0, "");
05220613 154
4b5f931b
MD
155/*
156 * These helper procedures handle the runq, they can only be called from
157 * within a critical section.
75cdbe6c
MD
158 *
159 * WARNING! Prior to SMP being brought up it is possible to enqueue and
160 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
161 * instead of 'mycpu' when referencing the globaldata structure. Once
162 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 163 */
f1d1c3fa
MD
164static __inline
165void
166_lwkt_dequeue(thread_t td)
167{
168 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 169 struct globaldata *gd = td->td_gd;
4b5f931b 170
f1d1c3fa 171 td->td_flags &= ~TDF_RUNQ;
f9235b6d
MD
172 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
173 gd->gd_fairq_total_pri -= td->td_pri;
174 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
175 atomic_clear_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
176 }
177}
178
f9235b6d
MD
179/*
180 * Priority enqueue.
181 *
182 * NOTE: There are a limited number of lwkt threads runnable since user
183 * processes only schedule one at a time per cpu.
184 */
f1d1c3fa
MD
185static __inline
186void
187_lwkt_enqueue(thread_t td)
188{
f9235b6d
MD
189 thread_t xtd;
190
7f5d7ed7 191 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 192 struct globaldata *gd = td->td_gd;
4b5f931b 193
f1d1c3fa 194 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
195 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
196 if (xtd == NULL) {
197 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
198 atomic_set_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
199 } else {
200 while (xtd && xtd->td_pri > td->td_pri)
201 xtd = TAILQ_NEXT(xtd, td_threadq);
202 if (xtd)
203 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
204 else
205 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
206 }
207 gd->gd_fairq_total_pri += td->td_pri;
f1d1c3fa
MD
208 }
209}
8ad65e08 210
40aaf5fc
NT
211static __boolean_t
212_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
213{
214 struct thread *td = (struct thread *)obj;
215
216 td->td_kstack = NULL;
217 td->td_kstack_size = 0;
218 td->td_flags = TDF_ALLOCATED_THREAD;
219 return (1);
220}
221
222static void
223_lwkt_thread_dtor(void *obj, void *privdata)
224{
225 struct thread *td = (struct thread *)obj;
226
227 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
228 ("_lwkt_thread_dtor: not allocated from objcache"));
229 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
230 td->td_kstack_size > 0,
231 ("_lwkt_thread_dtor: corrupted stack"));
232 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
233}
234
235/*
236 * Initialize the lwkt s/system.
237 */
238void
239lwkt_init(void)
240{
241 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
242 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
243 NULL, CACHE_NTHREADS/2,
244 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
245}
246
37af14fe
MD
247/*
248 * Schedule a thread to run. As the current thread we can always safely
249 * schedule ourselves, and a shortcut procedure is provided for that
250 * function.
251 *
252 * (non-blocking, self contained on a per cpu basis)
253 */
254void
255lwkt_schedule_self(thread_t td)
256{
257 crit_enter_quick(td);
f9235b6d
MD
258 KASSERT(td != &td->td_gd->gd_idlethread,
259 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 260 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 261 _lwkt_enqueue(td);
37af14fe
MD
262 crit_exit_quick(td);
263}
264
265/*
266 * Deschedule a thread.
267 *
268 * (non-blocking, self contained on a per cpu basis)
269 */
270void
271lwkt_deschedule_self(thread_t td)
272{
273 crit_enter_quick(td);
37af14fe
MD
274 _lwkt_dequeue(td);
275 crit_exit_quick(td);
276}
277
8ad65e08
MD
278/*
279 * LWKTs operate on a per-cpu basis
280 *
73e4f7b9 281 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
282 */
283void
284lwkt_gdinit(struct globaldata *gd)
285{
f9235b6d 286 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 287 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
288}
289
7d0bac62
MD
290/*
291 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
292 * or LWKT start address before it can be scheduled. If the target cpu is
293 * -1 the thread will be created on the current cpu.
0cfcada1
MD
294 *
295 * If you intend to create a thread without a process context this function
296 * does everything except load the startup and switcher function.
7d0bac62
MD
297 */
298thread_t
d3d32139 299lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 300{
c070746a 301 globaldata_t gd = mycpu;
99df837e 302 void *stack;
7d0bac62 303
c070746a
MD
304 /*
305 * If static thread storage is not supplied allocate a thread. Reuse
306 * a cached free thread if possible. gd_freetd is used to keep an exiting
307 * thread intact through the exit.
308 */
ef0fdad1 309 if (td == NULL) {
c070746a
MD
310 if ((td = gd->gd_freetd) != NULL)
311 gd->gd_freetd = NULL;
312 else
313 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
314 KASSERT((td->td_flags &
315 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
316 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
317 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 318 }
c070746a
MD
319
320 /*
321 * Try to reuse cached stack.
322 */
f470d0c8
MD
323 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
324 if (flags & TDF_ALLOCATED_STACK) {
e4846942 325 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
326 stack = NULL;
327 }
328 }
329 if (stack == NULL) {
e4846942 330 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 331 flags |= TDF_ALLOCATED_STACK;
99df837e 332 }
75cdbe6c 333 if (cpu < 0)
c070746a 334 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 335 else
f470d0c8 336 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 337 return(td);
7d0bac62
MD
338}
339
340/*
341 * Initialize a preexisting thread structure. This function is used by
342 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
343 *
f8c3996b
MD
344 * All threads start out in a critical section at a priority of
345 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
346 * appropriate. This function may send an IPI message when the
347 * requested cpu is not the current cpu and consequently gd_tdallq may
348 * not be initialized synchronously from the point of view of the originating
349 * cpu.
350 *
351 * NOTE! we have to be careful in regards to creating threads for other cpus
352 * if SMP has not yet been activated.
7d0bac62 353 */
41a01a4d
MD
354#ifdef SMP
355
75cdbe6c
MD
356static void
357lwkt_init_thread_remote(void *arg)
358{
359 thread_t td = arg;
360
52eedfb5
MD
361 /*
362 * Protected by critical section held by IPI dispatch
363 */
75cdbe6c
MD
364 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
365}
366
41a01a4d
MD
367#endif
368
7d0bac62 369void
f470d0c8
MD
370lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
371 struct globaldata *gd)
7d0bac62 372{
37af14fe
MD
373 globaldata_t mygd = mycpu;
374
99df837e
MD
375 bzero(td, sizeof(struct thread));
376 td->td_kstack = stack;
f470d0c8 377 td->td_kstack_size = stksize;
d3d32139 378 td->td_flags = flags;
26a0694b 379 td->td_gd = gd;
f9235b6d
MD
380 td->td_pri = TDPRI_KERN_DAEMON;
381 td->td_critcount = 1;
3b998fa9 382 td->td_toks_stop = &td->td_toks_base;
d3d32139
MD
383#ifdef SMP
384 if ((flags & TDF_MPSAFE) == 0)
385 td->td_mpcount = 1;
386#endif
fb0f29c4
MD
387 if (lwkt_use_spin_port)
388 lwkt_initport_spin(&td->td_msgport);
389 else
390 lwkt_initport_thread(&td->td_msgport, td);
99df837e 391 pmap_init_thread(td);
0f7a3396 392#ifdef SMP
5d21b981
MD
393 /*
394 * Normally initializing a thread for a remote cpu requires sending an
395 * IPI. However, the idlethread is setup before the other cpus are
396 * activated so we have to treat it as a special case. XXX manipulation
397 * of gd_tdallq requires the BGL.
398 */
399 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 400 crit_enter_gd(mygd);
75cdbe6c 401 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 402 crit_exit_gd(mygd);
75cdbe6c 403 } else {
2db3b277 404 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 405 }
0f7a3396 406#else
37af14fe 407 crit_enter_gd(mygd);
0f7a3396 408 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 409 crit_exit_gd(mygd);
0f7a3396 410#endif
8c72e3d5
AH
411
412 dsched_new_thread(td);
73e4f7b9
MD
413}
414
415void
416lwkt_set_comm(thread_t td, const char *ctl, ...)
417{
e2565a42 418 __va_list va;
73e4f7b9 419
e2565a42 420 __va_start(va, ctl);
379210cb 421 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 422 __va_end(va);
e7c0dbba 423 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
424}
425
99df837e 426void
73e4f7b9 427lwkt_hold(thread_t td)
99df837e 428{
73e4f7b9
MD
429 ++td->td_refs;
430}
431
432void
433lwkt_rele(thread_t td)
434{
435 KKASSERT(td->td_refs > 0);
436 --td->td_refs;
437}
438
439void
440lwkt_wait_free(thread_t td)
441{
442 while (td->td_refs)
377d4740 443 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
444}
445
446void
447lwkt_free_thread(thread_t td)
448{
d9eea1a5 449 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
450 ("lwkt_free_thread: did not exit! %p", td));
451
40aaf5fc
NT
452 if (td->td_flags & TDF_ALLOCATED_THREAD) {
453 objcache_put(thread_cache, td);
454 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
455 /* client-allocated struct with internally allocated stack */
456 KASSERT(td->td_kstack && td->td_kstack_size > 0,
457 ("lwkt_free_thread: corrupted stack"));
458 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
459 td->td_kstack = NULL;
460 td->td_kstack_size = 0;
99df837e 461 }
e7c0dbba 462 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
463}
464
465
8ad65e08
MD
466/*
467 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
468 * switch to the idlethread. Switching must occur within a critical
469 * section to avoid races with the scheduling queue.
470 *
471 * We always have full control over our cpu's run queue. Other cpus
472 * that wish to manipulate our queue must use the cpu_*msg() calls to
473 * talk to our cpu, so a critical section is all that is needed and
474 * the result is very, very fast thread switching.
475 *
96728c05
MD
476 * The LWKT scheduler uses a fixed priority model and round-robins at
477 * each priority level. User process scheduling is a totally
478 * different beast and LWKT priorities should not be confused with
479 * user process priorities.
f1d1c3fa 480 *
96728c05
MD
481 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
482 * cleans it up. Note that the td_switch() function cannot do anything that
483 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
484 * the target thread (not the current thread). It's nice to have a scheduler
485 * that does not need the MP lock to work because it allows us to do some
486 * really cool high-performance MP lock optimizations.
69d78e99
MD
487 *
488 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
489 * is not called by the current thread in the preemption case, only when
490 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
491 */
492void
493lwkt_switch(void)
494{
37af14fe
MD
495 globaldata_t gd = mycpu;
496 thread_t td = gd->gd_curthread;
8ad65e08 497 thread_t ntd;
f9235b6d
MD
498 thread_t xtd;
499 thread_t nlast;
8a8d5d85 500#ifdef SMP
f9235b6d 501 int nquserok;
8a8d5d85
MD
502 int mpheld;
503#endif
f9235b6d 504 int didaccumulate;
b37f18d6
MD
505 const char *lmsg; /* diagnostic - 'systat -pv 1' */
506 const void *laddr;
8ad65e08 507
46a3f46d 508 /*
27e88a6e
MD
509 * Switching from within a 'fast' (non thread switched) interrupt or IPI
510 * is illegal. However, we may have to do it anyway if we hit a fatal
511 * kernel trap or we have paniced.
512 *
513 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 514 */
27e88a6e
MD
515 if (gd->gd_intr_nesting_level) {
516 int savegdnest;
517 int savegdtrap;
518
519 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
520 panic("lwkt_switch: cannot switch from within "
521 "a fast interrupt, yet, td %p\n", td);
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;
6ea70f76 529 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615 530 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 531 print_backtrace(-1);
a7422615 532 }
27e88a6e
MD
533 lwkt_switch();
534 gd->gd_intr_nesting_level = savegdnest;
535 gd->gd_trap_nesting_level = savegdtrap;
536 return;
537 }
96728c05 538 }
ef0fdad1 539
cb973d15
MD
540 /*
541 * Passive release (used to transition from user to kernel mode
542 * when we block or switch rather then when we enter the kernel).
543 * This function is NOT called if we are switching into a preemption
544 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
545 * our current process designation (if we have one) and become a true
546 * LWKT thread, and may also hand the current process designation to
547 * another process and schedule thread.
cb973d15
MD
548 */
549 if (td->td_release)
550 td->td_release(td);
551
37af14fe 552 crit_enter_gd(gd);
3b998fa9 553 if (TD_TOKS_HELD(td))
9d265729
MD
554 lwkt_relalltokens(td);
555
556 /*
b02926de
MD
557 * We had better not be holding any spin locks, but don't get into an
558 * endless panic loop.
9d265729 559 */
bbb31c5d
MD
560 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
561 ("lwkt_switch: still holding a shared spinlock %p!",
562 gd->gd_spinlock_rd));
d666840a
MD
563 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
564 ("lwkt_switch: still holding %d exclusive spinlocks!",
565 gd->gd_spinlocks_wr));
9d265729 566
8a8d5d85
MD
567
568#ifdef SMP
569 /*
570 * td_mpcount cannot be used to determine if we currently hold the
571 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
572 * to get the lock, and switch out if it can't. Our ownership of
573 * the actual lock will remain stable while we are in a critical section
574 * (but, of course, another cpu may own or release the lock so the
575 * actual value of mp_lock is not stable).
8a8d5d85
MD
576 */
577 mpheld = MP_LOCK_HELD();
0f7a3396
MD
578#ifdef INVARIANTS
579 if (td->td_cscount) {
6ea70f76 580 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
581 td);
582 if (panic_on_cscount)
583 panic("switching while mastering cpusync");
584 }
585#endif
8a8d5d85 586#endif
f9235b6d
MD
587
588 /*
589 * If we had preempted another thread on this cpu, resume the preempted
590 * thread. This occurs transparently, whether the preempted thread
591 * was scheduled or not (it may have been preempted after descheduling
592 * itself).
593 *
594 * We have to setup the MP lock for the original thread after backing
595 * out the adjustment that was made to curthread when the original
596 * was preempted.
597 */
99df837e 598 if ((ntd = td->td_preempted) != NULL) {
26a0694b 599 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 600#ifdef SMP
96728c05 601 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 602 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
603 td, ntd, td->td_mpcount, ntd->td_mpcount);
604 }
8a8d5d85
MD
605 if (ntd->td_mpcount) {
606 td->td_mpcount -= ntd->td_mpcount;
607 KKASSERT(td->td_mpcount >= 0);
608 }
609#endif
26a0694b 610 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
611
612 /*
b9eb1c19
MD
613 * The interrupt may have woken a thread up, we need to properly
614 * set the reschedule flag if the originally interrupted thread is
615 * at a lower priority.
8ec60c3f 616 */
f9235b6d
MD
617 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
618 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
8ec60c3f 619 need_lwkt_resched();
f9235b6d 620 }
8a8d5d85 621 /* YYY release mp lock on switchback if original doesn't need it */
f9235b6d
MD
622 goto havethread_preempted;
623 }
624
625 /*
626 * Implement round-robin fairq with priority insertion. The priority
627 * insertion is handled by _lwkt_enqueue()
628 *
629 * We have to adjust the MP lock for the target thread. If we
630 * need the MP lock and cannot obtain it we try to locate a
631 * thread that does not need the MP lock. If we cannot, we spin
632 * instead of HLT.
633 *
634 * A similar issue exists for the tokens held by the target thread.
635 * If we cannot obtain ownership of the tokens we cannot immediately
636 * schedule the thread.
637 */
638 for (;;) {
639 clear_lwkt_resched();
640 didaccumulate = 0;
641 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
642
4b5f931b 643 /*
f9235b6d 644 * Hotpath if we can get all necessary resources.
41a01a4d 645 *
f9235b6d 646 * If nothing is runnable switch to the idle thread
41a01a4d 647 */
f9235b6d
MD
648 if (ntd == NULL) {
649 ntd = &gd->gd_idlethread;
650 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
651 ntd->td_flags |= TDF_IDLE_NOHLT;
652 if (ntd->td_mpcount) {
653 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
654 panic("Idle thread %p was holding the BGL!", ntd);
655 if (mpheld == 0) {
656 cpu_pause();
657 continue;
658 }
659 }
b37f18d6
MD
660 cpu_time.cp_msg[0] = 0;
661 cpu_time.cp_stallpc = 0;
f9235b6d
MD
662 goto haveidle;
663 }
41a01a4d 664
8ec60c3f 665 /*
f9235b6d 666 * Hotpath schedule
8ec60c3f 667 */
f9235b6d
MD
668 if (ntd->td_fairq_accum >= 0 &&
669#ifdef SMP
670 (ntd->td_mpcount == 0 || mpheld || cpu_try_mplock()) &&
671#endif
b37f18d6 672 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
f9235b6d 673 ) {
8a8d5d85 674#ifdef SMP
f9235b6d
MD
675 clr_mplock_contention_mask(gd);
676#endif
677 goto havethread;
678 }
679
b37f18d6
MD
680 lmsg = NULL;
681 laddr = NULL;
682
f9235b6d
MD
683#ifdef SMP
684 /* Reload mpheld (it become stale after mplock/token ops) */
685 mpheld = MP_LOCK_HELD();
b37f18d6
MD
686 if (ntd->td_mpcount && mpheld == 0) {
687 lmsg = "mplock";
688 laddr = ntd->td_mplock_stallpc;
689 }
f9235b6d
MD
690#endif
691
692 /*
693 * Coldpath - unable to schedule ntd, continue looking for threads
694 * to schedule. This is only allowed of the (presumably) kernel
695 * thread exhausted its fair share. A kernel thread stuck on
696 * resources does not currently allow a user thread to get in
697 * front of it.
698 */
699#ifdef SMP
700 nquserok = ((ntd->td_pri < TDPRI_KERN_LPSCHED) ||
701 (ntd->td_fairq_accum < 0));
702#endif
703 nlast = NULL;
704
705 for (;;) {
41a01a4d 706 /*
f9235b6d
MD
707 * If the fair-share scheduler ran out ntd gets moved to the
708 * end and its accumulator will be bumped, if it didn't we
709 * maintain the same queue position.
df6b8ba0 710 *
f9235b6d 711 * nlast keeps track of the last element prior to any moves.
41a01a4d 712 */
f9235b6d
MD
713 if (ntd->td_fairq_accum < 0) {
714 xtd = TAILQ_NEXT(ntd, td_threadq);
715 lwkt_fairq_accumulate(gd, ntd);
716 didaccumulate = 1;
717 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
718 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
719 if (nlast == NULL) {
720 nlast = ntd;
721 if (xtd == NULL)
722 xtd = ntd;
723 }
724 ntd = xtd;
725 } else {
726 ntd = TAILQ_NEXT(ntd, td_threadq);
727 }
a453459d 728
f9235b6d
MD
729 /*
730 * If we exhausted the run list switch to the idle thread.
731 * Since one or more threads had resource acquisition issues
732 * we do not allow the idle thread to halt.
733 *
734 * NOTE: nlast can be NULL.
735 */
736 if (ntd == nlast) {
e0a90d3b 737 cpu_pause();
f9235b6d
MD
738 ntd = &gd->gd_idlethread;
739 ntd->td_flags |= TDF_IDLE_NOHLT;
740 set_mplock_contention_mask(gd);
741 cpu_mplock_contested();
742 if (ntd->td_mpcount) {
743 mpheld = MP_LOCK_HELD();
744 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
745 panic("Idle thread %p was holding the BGL!", ntd);
746 if (mpheld == 0) {
747 cpu_pause();
748 break; /* try again from the top, almost */
b9eb1c19 749 }
8a8d5d85 750 }
684a93c4
MD
751
752 /*
f9235b6d
MD
753 * If fairq accumulations occured we do not schedule the
754 * idle thread. This will cause us to try again from
755 * the (almost) top.
684a93c4 756 */
f9235b6d 757 if (didaccumulate)
b37f18d6
MD
758 break; /* try again from the top, almost */
759 if (lmsg)
760 strlcpy(cpu_time.cp_msg, lmsg, sizeof(cpu_time.cp_msg));
761 cpu_time.cp_stallpc = (uintptr_t)laddr;
f9235b6d 762 goto haveidle;
8a8d5d85 763 }
f9235b6d 764
df6b8ba0 765 /*
f9235b6d 766 * Try to switch to this thread.
df6b8ba0 767 */
f9235b6d
MD
768 if ((ntd->td_pri >= TDPRI_KERN_LPSCHED || nquserok) &&
769 ntd->td_fairq_accum >= 0 &&
770#ifdef SMP
771 (ntd->td_mpcount == 0 || mpheld || cpu_try_mplock()) &&
8a8d5d85 772#endif
b37f18d6 773 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
f9235b6d 774 ) {
a453459d 775#ifdef SMP
f9235b6d
MD
776 clr_mplock_contention_mask(gd);
777#endif
778 goto havethread;
df6b8ba0 779 }
f9235b6d
MD
780#ifdef SMP
781 /* Reload mpheld (it become stale after mplock/token ops) */
782 mpheld = MP_LOCK_HELD();
b37f18d6
MD
783 if (ntd->td_mpcount && mpheld == 0) {
784 lmsg = "mplock";
785 laddr = ntd->td_mplock_stallpc;
786 }
787
f9235b6d
MD
788 if (ntd->td_pri >= TDPRI_KERN_LPSCHED && ntd->td_fairq_accum >= 0)
789 nquserok = 0;
a453459d 790#endif
4b5f931b 791 }
f1d1c3fa 792 }
8a8d5d85
MD
793
794 /*
f9235b6d
MD
795 * Do the actual switch. WARNING: mpheld is stale here.
796 *
797 * We must always decrement td_fairq_accum on non-idle threads just
798 * in case a thread never gets a tick due to being in a continuous
799 * critical section. The page-zeroing code does that.
800 *
801 * If the thread we came up with is a higher or equal priority verses
802 * the thread at the head of the queue we move our thread to the
803 * front. This way we can always check the front of the queue.
804 */
805havethread:
806 ++gd->gd_cnt.v_swtch;
807 --ntd->td_fairq_accum;
808 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
809 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
810 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
811 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
812 }
813havethread_preempted:
814 ;
815 /*
816 * If the new target does not need the MP lock and we are holding it,
817 * release the MP lock. If the new target requires the MP lock we have
818 * already acquired it for the target.
819 *
820 * WARNING: mpheld is stale here.
8a8d5d85 821 */
f9235b6d
MD
822haveidle:
823 KASSERT(ntd->td_critcount,
824 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
825#ifdef SMP
826 if (ntd->td_mpcount == 0 ) {
827 if (MP_LOCK_HELD())
828 cpu_rel_mplock();
829 } else {
a453459d 830 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
831 }
832#endif
94f6d86e
MD
833 if (td != ntd) {
834 ++switch_count;
b2b3ffcd 835#ifdef __x86_64__
f9235b6d
MD
836 {
837 int tos_ok __debugvar = jg_tos_ok(ntd);
838 KKASSERT(tos_ok);
839 }
85514115 840#endif
a1f0fb66 841 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
f1d1c3fa 842 td->td_switch(ntd);
94f6d86e 843 }
37af14fe
MD
844 /* NOTE: current cpu may have changed after switch */
845 crit_exit_quick(td);
8ad65e08
MD
846}
847
b68b7282 848/*
96728c05
MD
849 * Request that the target thread preempt the current thread. Preemption
850 * only works under a specific set of conditions:
b68b7282 851 *
96728c05
MD
852 * - We are not preempting ourselves
853 * - The target thread is owned by the current cpu
854 * - We are not currently being preempted
855 * - The target is not currently being preempted
d3d1cbc8
MD
856 * - We are not holding any spin locks
857 * - The target thread is not holding any tokens
96728c05
MD
858 * - We are able to satisfy the target's MP lock requirements (if any).
859 *
860 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
861 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 862 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
863 * to determine whether preemption is possible (aka usually just the crit
864 * priority of lwkt_schedule() itself).
b68b7282 865 *
26a0694b
MD
866 * XXX at the moment we run the target thread in a critical section during
867 * the preemption in order to prevent the target from taking interrupts
868 * that *WE* can't. Preemption is strictly limited to interrupt threads
869 * and interrupt-like threads, outside of a critical section, and the
870 * preempted source thread will be resumed the instant the target blocks
871 * whether or not the source is scheduled (i.e. preemption is supposed to
872 * be as transparent as possible).
4b5f931b 873 *
8a8d5d85
MD
874 * The target thread inherits our MP count (added to its own) for the
875 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
876 * MP lock during the preemption. Therefore, any preempting targets must be
877 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
878 * out of sync with the physical mp_lock, but we do not have to preserve
879 * the original ownership of the lock if it was out of synch (that is, we
880 * can leave it synchronized on return).
b68b7282
MD
881 */
882void
f9235b6d 883lwkt_preempt(thread_t ntd, int critcount)
b68b7282 884{
46a3f46d 885 struct globaldata *gd = mycpu;
0a3f9b47 886 thread_t td;
8a8d5d85
MD
887#ifdef SMP
888 int mpheld;
57c254db 889 int savecnt;
8a8d5d85 890#endif
b68b7282 891
26a0694b 892 /*
96728c05
MD
893 * The caller has put us in a critical section. We can only preempt
894 * if the caller of the caller was not in a critical section (basically
f9235b6d 895 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 896 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 897 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
898 *
899 * YYY The target thread must be in a critical section (else it must
900 * inherit our critical section? I dunno yet).
41a01a4d 901 *
0a3f9b47 902 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b 903 */
f9235b6d 904 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 905
0a3f9b47 906 td = gd->gd_curthread;
f9235b6d 907 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
908 ++preempt_miss;
909 return;
910 }
f9235b6d 911 if (td->td_critcount > critcount) {
96728c05 912 ++preempt_miss;
8ec60c3f 913 need_lwkt_resched();
96728c05
MD
914 return;
915 }
916#ifdef SMP
46a3f46d 917 if (ntd->td_gd != gd) {
96728c05 918 ++preempt_miss;
8ec60c3f 919 need_lwkt_resched();
96728c05
MD
920 return;
921 }
922#endif
41a01a4d 923 /*
d3d1cbc8 924 * Take the easy way out and do not preempt if we are holding
d666840a 925 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
926 * preempted interlock against the target thread's tokens and whether
927 * we can get all the target thread's tokens, but this situation
928 * should not occur very often so its easier to simply not preempt.
d666840a
MD
929 * Also, plain spinlocks are impossible to figure out at this point so
930 * just don't preempt.
d3d1cbc8
MD
931 *
932 * Do not try to preempt if the target thread is holding any tokens.
933 * We could try to acquire the tokens but this case is so rare there
934 * is no need to support it.
41a01a4d 935 */
bbb31c5d 936 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 937 ++preempt_miss;
8ec60c3f 938 need_lwkt_resched();
41a01a4d
MD
939 return;
940 }
3b998fa9 941 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8
MD
942 ++preempt_miss;
943 need_lwkt_resched();
944 return;
945 }
26a0694b
MD
946 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
947 ++preempt_weird;
8ec60c3f 948 need_lwkt_resched();
26a0694b
MD
949 return;
950 }
951 if (ntd->td_preempted) {
4b5f931b 952 ++preempt_hit;
8ec60c3f 953 need_lwkt_resched();
26a0694b 954 return;
b68b7282 955 }
8a8d5d85 956#ifdef SMP
a2a5ad0d
MD
957 /*
958 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
959 * to or from the MP lock. In this case td_mpcount will be pre-disposed
960 * (non-zero) but not actually synchronized with the actual state of the
961 * lock. We can use it to imply an MP lock requirement for the
962 * preemption but we cannot use it to test whether we hold the MP lock
963 * or not.
a2a5ad0d 964 */
96728c05 965 savecnt = td->td_mpcount;
71ef2f5c 966 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
967 ntd->td_mpcount += td->td_mpcount;
968 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
969 ntd->td_mpcount -= td->td_mpcount;
970 ++preempt_miss;
8ec60c3f 971 need_lwkt_resched();
8a8d5d85
MD
972 return;
973 }
974#endif
26a0694b 975
8ec60c3f
MD
976 /*
977 * Since we are able to preempt the current thread, there is no need to
978 * call need_lwkt_resched().
979 */
26a0694b
MD
980 ++preempt_hit;
981 ntd->td_preempted = td;
982 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 983 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
26a0694b 984 td->td_switch(ntd);
b9eb1c19 985
26a0694b 986 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
987#ifdef SMP
988 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
989 mpheld = MP_LOCK_HELD();
990 if (mpheld && td->td_mpcount == 0)
96728c05 991 cpu_rel_mplock();
71ef2f5c 992 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
993 panic("lwkt_preempt(): MP lock was not held through");
994#endif
26a0694b
MD
995 ntd->td_preempted = NULL;
996 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
997}
998
f1d1c3fa 999/*
faaeffac 1000 * Conditionally call splz() if gd_reqflags indicates work is pending.
f1d1c3fa 1001 *
faaeffac
MD
1002 * td_nest_count prevents deep nesting via splz() or doreti() which
1003 * might otherwise blow out the kernel stack. Note that except for
1004 * this special case, we MUST call splz() here to handle any
1005 * pending ints, particularly after we switch, or we might accidently
1006 * halt the cpu with interrupts pending.
ef0fdad1 1007 *
f1d1c3fa
MD
1008 * (self contained on a per cpu basis)
1009 */
1010void
faaeffac 1011splz_check(void)
f1d1c3fa 1012{
7966cb69
MD
1013 globaldata_t gd = mycpu;
1014 thread_t td = gd->gd_curthread;
ef0fdad1 1015
f9235b6d 1016 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
f1d1c3fa 1017 splz();
f1d1c3fa
MD
1018}
1019
8ad65e08 1020/*
f9235b6d
MD
1021 * This function is used to negotiate a passive release of the current
1022 * process/lwp designation with the user scheduler, allowing the user
1023 * scheduler to schedule another user thread. The related kernel thread
1024 * (curthread) continues running in the released state.
8ad65e08
MD
1025 */
1026void
f9235b6d 1027lwkt_passive_release(struct thread *td)
8ad65e08 1028{
f9235b6d
MD
1029 struct lwp *lp = td->td_lwp;
1030
1031 td->td_release = NULL;
1032 lwkt_setpri_self(TDPRI_KERN_USER);
1033 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1034}
1035
f9235b6d 1036
3824f392 1037/*
f9235b6d
MD
1038 * This implements a normal yield. This routine is virtually a nop if
1039 * there is nothing to yield to but it will always run any pending interrupts
1040 * if called from a critical section.
1041 *
1042 * This yield is designed for kernel threads without a user context.
1043 *
1044 * (self contained on a per cpu basis)
3824f392
MD
1045 */
1046void
f9235b6d 1047lwkt_yield(void)
3824f392 1048{
f9235b6d
MD
1049 globaldata_t gd = mycpu;
1050 thread_t td = gd->gd_curthread;
1051 thread_t xtd;
3824f392 1052
f9235b6d
MD
1053 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1054 splz();
1055 if (td->td_fairq_accum < 0) {
1056 lwkt_schedule_self(curthread);
1057 lwkt_switch();
1058 } else {
1059 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1060 if (xtd && xtd->td_pri > td->td_pri) {
1061 lwkt_schedule_self(curthread);
1062 lwkt_switch();
1063 }
1064 }
3824f392
MD
1065}
1066
1067/*
f9235b6d
MD
1068 * This yield is designed for kernel threads with a user context.
1069 *
1070 * The kernel acting on behalf of the user is potentially cpu-bound,
1071 * this function will efficiently allow other threads to run and also
1072 * switch to other processes by releasing.
3824f392
MD
1073 *
1074 * The lwkt_user_yield() function is designed to have very low overhead
1075 * if no yield is determined to be needed.
1076 */
1077void
1078lwkt_user_yield(void)
1079{
f9235b6d
MD
1080 globaldata_t gd = mycpu;
1081 thread_t td = gd->gd_curthread;
1082
1083 /*
1084 * Always run any pending interrupts in case we are in a critical
1085 * section.
1086 */
1087 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1088 splz();
3824f392
MD
1089
1090#ifdef SMP
1091 /*
1092 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1093 * kernel can prevent other cpus from servicing interrupt threads
1094 * which still require the MP lock (which is a lot of them). This
1095 * has a chaining effect since if the interrupt is blocked, so is
1096 * the event, so normal scheduling will not pick up on the problem.
1097 */
684a93c4
MD
1098 if (mp_lock_contention_mask && td->td_mpcount) {
1099 yield_mplock(td);
3824f392
MD
1100 }
1101#endif
1102
1103 /*
f9235b6d
MD
1104 * Switch (which forces a release) if another kernel thread needs
1105 * the cpu, if userland wants us to resched, or if our kernel
1106 * quantum has run out.
3824f392 1107 */
f9235b6d
MD
1108 if (lwkt_resched_wanted() ||
1109 user_resched_wanted() ||
1110 td->td_fairq_accum < 0)
1111 {
3824f392 1112 lwkt_switch();
3824f392
MD
1113 }
1114
f9235b6d 1115#if 0
3824f392 1116 /*
f9235b6d
MD
1117 * Reacquire the current process if we are released.
1118 *
1119 * XXX not implemented atm. The kernel may be holding locks and such,
1120 * so we want the thread to continue to receive cpu.
3824f392 1121 */
f9235b6d
MD
1122 if (td->td_release == NULL && lp) {
1123 lp->lwp_proc->p_usched->acquire_curproc(lp);
1124 td->td_release = lwkt_passive_release;
1125 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1126 }
f9235b6d 1127#endif
b9eb1c19
MD
1128}
1129
8ad65e08 1130/*
f1d1c3fa
MD
1131 * Generic schedule. Possibly schedule threads belonging to other cpus and
1132 * deal with threads that might be blocked on a wait queue.
1133 *
0a3f9b47
MD
1134 * We have a little helper inline function which does additional work after
1135 * the thread has been enqueued, including dealing with preemption and
1136 * setting need_lwkt_resched() (which prevents the kernel from returning
1137 * to userland until it has processed higher priority threads).
6330a558
MD
1138 *
1139 * It is possible for this routine to be called after a failed _enqueue
1140 * (due to the target thread migrating, sleeping, or otherwise blocked).
1141 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1142 *
1143 * reschedok is an optimized constant propagated from lwkt_schedule() or
1144 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1145 * reschedule to be requested if the target thread has a higher priority.
1146 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1147 * be 0, prevented undesired reschedules.
8ad65e08 1148 */
0a3f9b47
MD
1149static __inline
1150void
f9235b6d 1151_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
0a3f9b47 1152{
b9eb1c19 1153 thread_t otd;
c730be20 1154
6330a558 1155 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1156 if (ntd->td_preemptable && reschedok) {
f9235b6d 1157 ntd->td_preemptable(ntd, ccount); /* YYY +token */
361d01dd 1158 } else if (reschedok) {
b9eb1c19 1159 otd = curthread;
f9235b6d 1160 if (ntd->td_pri > otd->td_pri)
c730be20 1161 need_lwkt_resched();
6330a558 1162 }
f9235b6d
MD
1163
1164 /*
1165 * Give the thread a little fair share scheduler bump if it
1166 * has been asleep for a while. This is primarily to avoid
1167 * a degenerate case for interrupt threads where accumulator
1168 * crosses into negative territory unnecessarily.
1169 */
1170 if (ntd->td_fairq_lticks != ticks) {
1171 ntd->td_fairq_lticks = ticks;
1172 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1173 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1174 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1175 }
0a3f9b47
MD
1176 }
1177}
1178
361d01dd 1179static __inline
8ad65e08 1180void
361d01dd 1181_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1182{
37af14fe
MD
1183 globaldata_t mygd = mycpu;
1184
41a01a4d 1185 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1186 crit_enter_gd(mygd);
9388413d 1187 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1188 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1189 _lwkt_enqueue(td);
1190 } else {
f1d1c3fa 1191 /*
7cd8d145
MD
1192 * If we own the thread, there is no race (since we are in a
1193 * critical section). If we do not own the thread there might
1194 * be a race but the target cpu will deal with it.
f1d1c3fa 1195 */
0f7a3396 1196#ifdef SMP
7cd8d145 1197 if (td->td_gd == mygd) {
9d265729 1198 _lwkt_enqueue(td);
f9235b6d 1199 _lwkt_schedule_post(mygd, td, 1, reschedok);
f1d1c3fa 1200 } else {
e381e77c 1201 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1202 }
0f7a3396 1203#else
7cd8d145 1204 _lwkt_enqueue(td);
f9235b6d 1205 _lwkt_schedule_post(mygd, td, 1, reschedok);
0f7a3396 1206#endif
8ad65e08 1207 }
37af14fe 1208 crit_exit_gd(mygd);
8ad65e08
MD
1209}
1210
361d01dd
MD
1211void
1212lwkt_schedule(thread_t td)
1213{
1214 _lwkt_schedule(td, 1);
1215}
1216
1217void
1218lwkt_schedule_noresched(thread_t td)
1219{
1220 _lwkt_schedule(td, 0);
1221}
1222
88ebb169
SW
1223#ifdef SMP
1224
e381e77c
MD
1225/*
1226 * When scheduled remotely if frame != NULL the IPIQ is being
1227 * run via doreti or an interrupt then preemption can be allowed.
1228 *
1229 * To allow preemption we have to drop the critical section so only
1230 * one is present in _lwkt_schedule_post.
1231 */
1232static void
1233lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1234{
1235 thread_t td = curthread;
1236 thread_t ntd = arg;
1237
1238 if (frame && ntd->td_preemptable) {
1239 crit_exit_noyield(td);
1240 _lwkt_schedule(ntd, 1);
1241 crit_enter_quick(td);
1242 } else {
1243 _lwkt_schedule(ntd, 1);
1244 }
1245}
1246
d9eea1a5 1247/*
52eedfb5
MD
1248 * Thread migration using a 'Pull' method. The thread may or may not be
1249 * the current thread. It MUST be descheduled and in a stable state.
1250 * lwkt_giveaway() must be called on the cpu owning the thread.
1251 *
1252 * At any point after lwkt_giveaway() is called, the target cpu may
1253 * 'pull' the thread by calling lwkt_acquire().
1254 *
ae8e83e6
MD
1255 * We have to make sure the thread is not sitting on a per-cpu tsleep
1256 * queue or it will blow up when it moves to another cpu.
1257 *
52eedfb5 1258 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1259 */
52eedfb5
MD
1260void
1261lwkt_giveaway(thread_t td)
1262{
3b4192fb 1263 globaldata_t gd = mycpu;
52eedfb5 1264
3b4192fb
MD
1265 crit_enter_gd(gd);
1266 if (td->td_flags & TDF_TSLEEPQ)
1267 tsleep_remove(td);
1268 KKASSERT(td->td_gd == gd);
1269 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1270 td->td_flags |= TDF_MIGRATING;
1271 crit_exit_gd(gd);
52eedfb5
MD
1272}
1273
a2a5ad0d
MD
1274void
1275lwkt_acquire(thread_t td)
1276{
37af14fe
MD
1277 globaldata_t gd;
1278 globaldata_t mygd;
a2a5ad0d 1279
52eedfb5 1280 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1281 gd = td->td_gd;
37af14fe 1282 mygd = mycpu;
52eedfb5 1283 if (gd != mycpu) {
35238fa5 1284 cpu_lfence();
52eedfb5 1285 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1286 crit_enter_gd(mygd);
df910c23
MD
1287 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1288#ifdef SMP
1289 lwkt_process_ipiq();
1290#endif
52eedfb5 1291 cpu_lfence();
df910c23 1292 }
37af14fe 1293 td->td_gd = mygd;
52eedfb5
MD
1294 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1295 td->td_flags &= ~TDF_MIGRATING;
1296 crit_exit_gd(mygd);
1297 } else {
1298 crit_enter_gd(mygd);
1299 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1300 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1301 crit_exit_gd(mygd);
a2a5ad0d
MD
1302 }
1303}
1304
52eedfb5
MD
1305#endif
1306
f1d1c3fa
MD
1307/*
1308 * Generic deschedule. Descheduling threads other then your own should be
1309 * done only in carefully controlled circumstances. Descheduling is
1310 * asynchronous.
1311 *
1312 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1313 */
1314void
1315lwkt_deschedule(thread_t td)
1316{
f1d1c3fa 1317 crit_enter();
b8a98473 1318#ifdef SMP
f1d1c3fa
MD
1319 if (td == curthread) {
1320 _lwkt_dequeue(td);
1321 } else {
a72187e9 1322 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1323 _lwkt_dequeue(td);
1324 } else {
b8a98473 1325 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1326 }
1327 }
b8a98473
MD
1328#else
1329 _lwkt_dequeue(td);
1330#endif
f1d1c3fa
MD
1331 crit_exit();
1332}
1333
4b5f931b
MD
1334/*
1335 * Set the target thread's priority. This routine does not automatically
1336 * switch to a higher priority thread, LWKT threads are not designed for
1337 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1338 */
1339void
1340lwkt_setpri(thread_t td, int pri)
1341{
a72187e9 1342 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1343 if (td->td_pri != pri) {
1344 KKASSERT(pri >= 0);
1345 crit_enter();
1346 if (td->td_flags & TDF_RUNQ) {
1347 _lwkt_dequeue(td);
1348 td->td_pri = pri;
1349 _lwkt_enqueue(td);
1350 } else {
1351 td->td_pri = pri;
1352 }
1353 crit_exit();
26a0694b 1354 }
26a0694b
MD
1355}
1356
03bd0a5e
MD
1357/*
1358 * Set the initial priority for a thread prior to it being scheduled for
1359 * the first time. The thread MUST NOT be scheduled before or during
1360 * this call. The thread may be assigned to a cpu other then the current
1361 * cpu.
1362 *
1363 * Typically used after a thread has been created with TDF_STOPPREQ,
1364 * and before the thread is initially scheduled.
1365 */
1366void
1367lwkt_setpri_initial(thread_t td, int pri)
1368{
1369 KKASSERT(pri >= 0);
1370 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1371 td->td_pri = pri;
03bd0a5e
MD
1372}
1373
26a0694b
MD
1374void
1375lwkt_setpri_self(int pri)
1376{
1377 thread_t td = curthread;
1378
4b5f931b
MD
1379 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1380 crit_enter();
1381 if (td->td_flags & TDF_RUNQ) {
1382 _lwkt_dequeue(td);
f9235b6d 1383 td->td_pri = pri;
4b5f931b
MD
1384 _lwkt_enqueue(td);
1385 } else {
f9235b6d 1386 td->td_pri = pri;
4b5f931b
MD
1387 }
1388 crit_exit();
1389}
1390
f9235b6d
MD
1391/*
1392 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1393 *
1394 * Example: two competing threads, same priority N. decrement by (2*N)
1395 * increment by N*8, each thread will get 4 ticks.
1396 */
1397void
1398lwkt_fairq_schedulerclock(thread_t td)
1399{
1400 if (fairq_enable) {
1401 while (td) {
1402 if (td != &td->td_gd->gd_idlethread) {
1403 td->td_fairq_accum -= td->td_gd->gd_fairq_total_pri;
1404 if (td->td_fairq_accum < -TDFAIRQ_MAX(td->td_gd))
1405 td->td_fairq_accum = -TDFAIRQ_MAX(td->td_gd);
1406 if (td->td_fairq_accum < 0)
1407 need_lwkt_resched();
1408 td->td_fairq_lticks = ticks;
1409 }
1410 td = td->td_preempted;
1411 }
1412 }
1413}
1414
1415static void
1416lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1417{
1418 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1419 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1420 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1421}
1422
5d21b981 1423/*
52eedfb5
MD
1424 * Migrate the current thread to the specified cpu.
1425 *
1426 * This is accomplished by descheduling ourselves from the current cpu,
1427 * moving our thread to the tdallq of the target cpu, IPI messaging the
1428 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1429 * races while the thread is being migrated.
ae8e83e6
MD
1430 *
1431 * We must be sure to remove ourselves from the current cpu's tsleepq
1432 * before potentially moving to another queue. The thread can be on
1433 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1434 */
3d28ff59 1435#ifdef SMP
5d21b981 1436static void lwkt_setcpu_remote(void *arg);
3d28ff59 1437#endif
5d21b981
MD
1438
1439void
1440lwkt_setcpu_self(globaldata_t rgd)
1441{
1442#ifdef SMP
1443 thread_t td = curthread;
1444
1445 if (td->td_gd != rgd) {
1446 crit_enter_quick(td);
ae8e83e6 1447 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1448 tsleep_remove(td);
5d21b981
MD
1449 td->td_flags |= TDF_MIGRATING;
1450 lwkt_deschedule_self(td);
52eedfb5 1451 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1452 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1453 lwkt_switch();
1454 /* we are now on the target cpu */
52eedfb5 1455 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1456 crit_exit_quick(td);
1457 }
1458#endif
1459}
1460
ecdefdda
MD
1461void
1462lwkt_migratecpu(int cpuid)
1463{
1464#ifdef SMP
1465 globaldata_t rgd;
1466
1467 rgd = globaldata_find(cpuid);
1468 lwkt_setcpu_self(rgd);
1469#endif
1470}
1471
5d21b981
MD
1472/*
1473 * Remote IPI for cpu migration (called while in a critical section so we
1474 * do not have to enter another one). The thread has already been moved to
1475 * our cpu's allq, but we must wait for the thread to be completely switched
1476 * out on the originating cpu before we schedule it on ours or the stack
1477 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1478 * change to main memory.
1479 *
1480 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1481 * against wakeups. It is best if this interface is used only when there
1482 * are no pending events that might try to schedule the thread.
1483 */
3d28ff59 1484#ifdef SMP
5d21b981
MD
1485static void
1486lwkt_setcpu_remote(void *arg)
1487{
1488 thread_t td = arg;
1489 globaldata_t gd = mycpu;
1490
df910c23
MD
1491 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1492#ifdef SMP
1493 lwkt_process_ipiq();
1494#endif
35238fa5 1495 cpu_lfence();
df910c23 1496 }
5d21b981 1497 td->td_gd = gd;
35238fa5 1498 cpu_sfence();
5d21b981 1499 td->td_flags &= ~TDF_MIGRATING;
9388413d 1500 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1501 _lwkt_enqueue(td);
1502}
3d28ff59 1503#endif
5d21b981 1504
553ea3c8 1505struct lwp *
4b5f931b
MD
1506lwkt_preempted_proc(void)
1507{
73e4f7b9 1508 thread_t td = curthread;
4b5f931b
MD
1509 while (td->td_preempted)
1510 td = td->td_preempted;
553ea3c8 1511 return(td->td_lwp);
4b5f931b
MD
1512}
1513
99df837e
MD
1514/*
1515 * Create a kernel process/thread/whatever. It shares it's address space
1516 * with proc0 - ie: kernel only.
1517 *
365fa13f
MD
1518 * NOTE! By default new threads are created with the MP lock held. A
1519 * thread which does not require the MP lock should release it by calling
1520 * rel_mplock() at the start of the new thread.
99df837e
MD
1521 */
1522int
1523lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1524 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1525 const char *fmt, ...)
99df837e 1526{
73e4f7b9 1527 thread_t td;
e2565a42 1528 __va_list ap;
99df837e 1529
d3d32139 1530 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1531 tdflags);
a2a5ad0d
MD
1532 if (tdp)
1533 *tdp = td;
709799ea 1534 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1535
1536 /*
1537 * Set up arg0 for 'ps' etc
1538 */
e2565a42 1539 __va_start(ap, fmt);
379210cb 1540 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1541 __va_end(ap);
99df837e
MD
1542
1543 /*
1544 * Schedule the thread to run
1545 */
ef0fdad1
MD
1546 if ((td->td_flags & TDF_STOPREQ) == 0)
1547 lwkt_schedule(td);
1548 else
1549 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1550 return 0;
1551}
1552
1553/*
1554 * Destroy an LWKT thread. Warning! This function is not called when
1555 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1556 * uses a different reaping mechanism.
1557 */
1558void
1559lwkt_exit(void)
1560{
1561 thread_t td = curthread;
c070746a 1562 thread_t std;
8826f33a 1563 globaldata_t gd;
99df837e
MD
1564
1565 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1566 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1567 caps_exit(td);
c070746a
MD
1568
1569 /*
1570 * Get us into a critical section to interlock gd_freetd and loop
1571 * until we can get it freed.
1572 *
1573 * We have to cache the current td in gd_freetd because objcache_put()ing
1574 * it would rip it out from under us while our thread is still active.
1575 */
1576 gd = mycpu;
37af14fe 1577 crit_enter_quick(td);
c070746a
MD
1578 while ((std = gd->gd_freetd) != NULL) {
1579 gd->gd_freetd = NULL;
1580 objcache_put(thread_cache, std);
1581 }
3b4192fb
MD
1582
1583 /*
1584 * Remove thread resources from kernel lists and deschedule us for
1585 * the last time.
1586 */
1587 if (td->td_flags & TDF_TSLEEPQ)
1588 tsleep_remove(td);
79eae878 1589 biosched_done(td);
f8abf63c 1590 dsched_exit_thread(td);
37af14fe 1591 lwkt_deschedule_self(td);
e56e4dea 1592 lwkt_remove_tdallq(td);
c070746a
MD
1593 if (td->td_flags & TDF_ALLOCATED_THREAD)
1594 gd->gd_freetd = td;
99df837e
MD
1595 cpu_thread_exit();
1596}
1597
e56e4dea
MD
1598void
1599lwkt_remove_tdallq(thread_t td)
1600{
1601 KKASSERT(td->td_gd == mycpu);
1602 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1603}
1604
2d93b37a
MD
1605void
1606crit_panic(void)
1607{
1608 thread_t td = curthread;
1609 int lpri = td->td_pri;
1610
1611 td->td_pri = 0;
1612 panic("td_pri is/would-go negative! %p %d", td, lpri);
1613}
1614
d165e668
MD
1615#ifdef SMP
1616
bd8015ca
MD
1617/*
1618 * Called from debugger/panic on cpus which have been stopped. We must still
1619 * process the IPIQ while stopped, even if we were stopped while in a critical
1620 * section (XXX).
1621 *
1622 * If we are dumping also try to process any pending interrupts. This may
1623 * or may not work depending on the state of the cpu at the point it was
1624 * stopped.
1625 */
1626void
1627lwkt_smp_stopped(void)
1628{
1629 globaldata_t gd = mycpu;
1630
1631 crit_enter_gd(gd);
1632 if (dumping) {
1633 lwkt_process_ipiq();
1634 splz();
1635 } else {
1636 lwkt_process_ipiq();
1637 }
1638 crit_exit_gd(gd);
1639}
1640
d165e668 1641#endif