dsched - Periph.: call dsched_exit on thread exit
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
8c10bfcf 2 * Copyright (c) 2003,2004 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>
47#include <sys/queue.h>
7d0bac62 48#include <sys/sysctl.h>
99df837e 49#include <sys/kthread.h>
f1d1c3fa 50#include <machine/cpu.h>
99df837e 51#include <sys/lock.h>
f6bf3af1 52#include <sys/caps.h>
9d265729 53#include <sys/spinlock.h>
57aa743c 54#include <sys/ktr.h>
9d265729
MD
55
56#include <sys/thread2.h>
57#include <sys/spinlock2.h>
684a93c4 58#include <sys/mplock2.h>
f1d1c3fa 59
8c72e3d5
AH
60#include <sys/dsched.h>
61
7d0bac62
MD
62#include <vm/vm.h>
63#include <vm/vm_param.h>
64#include <vm/vm_kern.h>
65#include <vm/vm_object.h>
66#include <vm/vm_page.h>
67#include <vm/vm_map.h>
68#include <vm/vm_pager.h>
69#include <vm/vm_extern.h>
7d0bac62 70
99df837e 71#include <machine/stdarg.h>
96728c05 72#include <machine/smp.h>
99df837e 73
d850923c
AE
74#if !defined(KTR_CTXSW)
75#define KTR_CTXSW KTR_ALL
76#endif
77KTR_INFO_MASTER(ctxsw);
78KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "sw %p > %p", 2 * sizeof(struct thread *));
79KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "pre %p > %p", 2 * sizeof(struct thread *));
e7c0dbba
AE
80KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "new_td %p %s", sizeof (struct thread *) +
81 sizeof(char *));
82KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "dead_td %p", sizeof (struct thread *));
1541028a 83
40aaf5fc
NT
84static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
85
0f7a3396
MD
86#ifdef INVARIANTS
87static int panic_on_cscount = 0;
88#endif
05220613
MD
89static __int64_t switch_count = 0;
90static __int64_t preempt_hit = 0;
91static __int64_t preempt_miss = 0;
92static __int64_t preempt_weird = 0;
f64b567c 93static __int64_t token_contention_count __debugvar = 0;
fb0f29c4 94static int lwkt_use_spin_port;
40aaf5fc 95static struct objcache *thread_cache;
05220613 96
88ebb169 97#ifdef SMP
e381e77c 98static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
88ebb169 99#endif
e381e77c 100
0855a2af
JG
101extern void cpu_heavy_restore(void);
102extern void cpu_lwkt_restore(void);
103extern void cpu_kthread_restore(void);
104extern void cpu_idle_restore(void);
105
b2b3ffcd 106#ifdef __x86_64__
85514115
MD
107
108static int
0855a2af
JG
109jg_tos_ok(struct thread *td)
110{
85514115
MD
111 void *tos;
112 int tos_ok;
113
0855a2af
JG
114 if (td == NULL) {
115 return 1;
116 }
117 KKASSERT(td->td_sp != NULL);
85514115
MD
118 tos = ((void **)td->td_sp)[0];
119 tos_ok = 0;
120 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
121 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
0855a2af
JG
122 tos_ok = 1;
123 }
124 return tos_ok;
125}
126
85514115
MD
127#endif
128
fb0f29c4
MD
129/*
130 * We can make all thread ports use the spin backend instead of the thread
131 * backend. This should only be set to debug the spin backend.
132 */
133TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
134
0f7a3396
MD
135#ifdef INVARIANTS
136SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
137#endif
4b5f931b 138SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 139SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 140SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 141SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
142#ifdef INVARIANTS
143SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
144 &token_contention_count, 0, "spinning due to token contention");
38717797 145#endif
05220613 146
4b5f931b
MD
147/*
148 * These helper procedures handle the runq, they can only be called from
149 * within a critical section.
75cdbe6c
MD
150 *
151 * WARNING! Prior to SMP being brought up it is possible to enqueue and
152 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
153 * instead of 'mycpu' when referencing the globaldata structure. Once
154 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 155 */
f1d1c3fa
MD
156static __inline
157void
158_lwkt_dequeue(thread_t td)
159{
160 if (td->td_flags & TDF_RUNQ) {
4b5f931b 161 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 162 struct globaldata *gd = td->td_gd;
4b5f931b 163
f1d1c3fa 164 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
165 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
166 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
167 }
168}
169
170static __inline
171void
172_lwkt_enqueue(thread_t td)
173{
7f5d7ed7 174 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
4b5f931b 175 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 176 struct globaldata *gd = td->td_gd;
4b5f931b 177
f1d1c3fa 178 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
179 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
180 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
181 }
182}
8ad65e08 183
40aaf5fc
NT
184static __boolean_t
185_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
186{
187 struct thread *td = (struct thread *)obj;
188
189 td->td_kstack = NULL;
190 td->td_kstack_size = 0;
191 td->td_flags = TDF_ALLOCATED_THREAD;
192 return (1);
193}
194
195static void
196_lwkt_thread_dtor(void *obj, void *privdata)
197{
198 struct thread *td = (struct thread *)obj;
199
200 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
201 ("_lwkt_thread_dtor: not allocated from objcache"));
202 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
203 td->td_kstack_size > 0,
204 ("_lwkt_thread_dtor: corrupted stack"));
205 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
206}
207
208/*
209 * Initialize the lwkt s/system.
210 */
211void
212lwkt_init(void)
213{
214 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
215 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
216 NULL, CACHE_NTHREADS/2,
217 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
218}
219
37af14fe
MD
220/*
221 * Schedule a thread to run. As the current thread we can always safely
222 * schedule ourselves, and a shortcut procedure is provided for that
223 * function.
224 *
225 * (non-blocking, self contained on a per cpu basis)
226 */
227void
228lwkt_schedule_self(thread_t td)
229{
230 crit_enter_quick(td);
37af14fe 231 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 232 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 233 _lwkt_enqueue(td);
37af14fe
MD
234 crit_exit_quick(td);
235}
236
237/*
238 * Deschedule a thread.
239 *
240 * (non-blocking, self contained on a per cpu basis)
241 */
242void
243lwkt_deschedule_self(thread_t td)
244{
245 crit_enter_quick(td);
37af14fe
MD
246 _lwkt_dequeue(td);
247 crit_exit_quick(td);
248}
249
8ad65e08
MD
250/*
251 * LWKTs operate on a per-cpu basis
252 *
73e4f7b9 253 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
254 */
255void
256lwkt_gdinit(struct globaldata *gd)
257{
4b5f931b
MD
258 int i;
259
260 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
261 TAILQ_INIT(&gd->gd_tdrunq[i]);
262 gd->gd_runqmask = 0;
73e4f7b9 263 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
264}
265
266/*
7d0bac62 267 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
268 * or LWKT start address before it can be scheduled. If the target cpu is
269 * -1 the thread will be created on the current cpu.
0cfcada1
MD
270 *
271 * If you intend to create a thread without a process context this function
272 * does everything except load the startup and switcher function.
7d0bac62
MD
273 */
274thread_t
d3d32139 275lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 276{
c070746a 277 globaldata_t gd = mycpu;
99df837e 278 void *stack;
7d0bac62 279
c070746a
MD
280 /*
281 * If static thread storage is not supplied allocate a thread. Reuse
282 * a cached free thread if possible. gd_freetd is used to keep an exiting
283 * thread intact through the exit.
284 */
ef0fdad1 285 if (td == NULL) {
c070746a
MD
286 if ((td = gd->gd_freetd) != NULL)
287 gd->gd_freetd = NULL;
288 else
289 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
290 KASSERT((td->td_flags &
291 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
292 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
293 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 294 }
c070746a
MD
295
296 /*
297 * Try to reuse cached stack.
298 */
f470d0c8
MD
299 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
300 if (flags & TDF_ALLOCATED_STACK) {
e4846942 301 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
302 stack = NULL;
303 }
304 }
305 if (stack == NULL) {
e4846942 306 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 307 flags |= TDF_ALLOCATED_STACK;
99df837e 308 }
75cdbe6c 309 if (cpu < 0)
c070746a 310 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 311 else
f470d0c8 312 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 313 return(td);
7d0bac62
MD
314}
315
316/*
317 * Initialize a preexisting thread structure. This function is used by
318 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
319 *
f8c3996b
MD
320 * All threads start out in a critical section at a priority of
321 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
322 * appropriate. This function may send an IPI message when the
323 * requested cpu is not the current cpu and consequently gd_tdallq may
324 * not be initialized synchronously from the point of view of the originating
325 * cpu.
326 *
327 * NOTE! we have to be careful in regards to creating threads for other cpus
328 * if SMP has not yet been activated.
7d0bac62 329 */
41a01a4d
MD
330#ifdef SMP
331
75cdbe6c
MD
332static void
333lwkt_init_thread_remote(void *arg)
334{
335 thread_t td = arg;
336
52eedfb5
MD
337 /*
338 * Protected by critical section held by IPI dispatch
339 */
75cdbe6c
MD
340 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
341}
342
41a01a4d
MD
343#endif
344
7d0bac62 345void
f470d0c8
MD
346lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
347 struct globaldata *gd)
7d0bac62 348{
37af14fe
MD
349 globaldata_t mygd = mycpu;
350
99df837e
MD
351 bzero(td, sizeof(struct thread));
352 td->td_kstack = stack;
f470d0c8 353 td->td_kstack_size = stksize;
d3d32139 354 td->td_flags = flags;
26a0694b 355 td->td_gd = gd;
f8c3996b 356 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
357#ifdef SMP
358 if ((flags & TDF_MPSAFE) == 0)
359 td->td_mpcount = 1;
360#endif
fb0f29c4
MD
361 if (lwkt_use_spin_port)
362 lwkt_initport_spin(&td->td_msgport);
363 else
364 lwkt_initport_thread(&td->td_msgport, td);
99df837e 365 pmap_init_thread(td);
0f7a3396 366#ifdef SMP
5d21b981
MD
367 /*
368 * Normally initializing a thread for a remote cpu requires sending an
369 * IPI. However, the idlethread is setup before the other cpus are
370 * activated so we have to treat it as a special case. XXX manipulation
371 * of gd_tdallq requires the BGL.
372 */
373 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 374 crit_enter_gd(mygd);
75cdbe6c 375 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 376 crit_exit_gd(mygd);
75cdbe6c 377 } else {
2db3b277 378 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 379 }
0f7a3396 380#else
37af14fe 381 crit_enter_gd(mygd);
0f7a3396 382 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 383 crit_exit_gd(mygd);
0f7a3396 384#endif
8c72e3d5
AH
385
386 dsched_new_thread(td);
73e4f7b9
MD
387}
388
389void
390lwkt_set_comm(thread_t td, const char *ctl, ...)
391{
e2565a42 392 __va_list va;
73e4f7b9 393
e2565a42 394 __va_start(va, ctl);
379210cb 395 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 396 __va_end(va);
e7c0dbba 397 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
398}
399
99df837e 400void
73e4f7b9 401lwkt_hold(thread_t td)
99df837e 402{
73e4f7b9
MD
403 ++td->td_refs;
404}
405
406void
407lwkt_rele(thread_t td)
408{
409 KKASSERT(td->td_refs > 0);
410 --td->td_refs;
411}
412
413void
414lwkt_wait_free(thread_t td)
415{
416 while (td->td_refs)
377d4740 417 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
418}
419
420void
421lwkt_free_thread(thread_t td)
422{
d9eea1a5 423 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
424 ("lwkt_free_thread: did not exit! %p", td));
425
40aaf5fc
NT
426 if (td->td_flags & TDF_ALLOCATED_THREAD) {
427 objcache_put(thread_cache, td);
428 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
429 /* client-allocated struct with internally allocated stack */
430 KASSERT(td->td_kstack && td->td_kstack_size > 0,
431 ("lwkt_free_thread: corrupted stack"));
432 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
433 td->td_kstack = NULL;
434 td->td_kstack_size = 0;
99df837e 435 }
e7c0dbba 436 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
437}
438
439
7d0bac62 440/*
8ad65e08 441 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
442 * switch to the idlethread. Switching must occur within a critical
443 * section to avoid races with the scheduling queue.
444 *
445 * We always have full control over our cpu's run queue. Other cpus
446 * that wish to manipulate our queue must use the cpu_*msg() calls to
447 * talk to our cpu, so a critical section is all that is needed and
448 * the result is very, very fast thread switching.
449 *
96728c05
MD
450 * The LWKT scheduler uses a fixed priority model and round-robins at
451 * each priority level. User process scheduling is a totally
452 * different beast and LWKT priorities should not be confused with
453 * user process priorities.
f1d1c3fa 454 *
96728c05
MD
455 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
456 * cleans it up. Note that the td_switch() function cannot do anything that
457 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
458 * the target thread (not the current thread). It's nice to have a scheduler
459 * that does not need the MP lock to work because it allows us to do some
460 * really cool high-performance MP lock optimizations.
69d78e99
MD
461 *
462 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
463 * is not called by the current thread in the preemption case, only when
464 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
465 */
466void
467lwkt_switch(void)
468{
37af14fe
MD
469 globaldata_t gd = mycpu;
470 thread_t td = gd->gd_curthread;
8ad65e08 471 thread_t ntd;
8a8d5d85
MD
472#ifdef SMP
473 int mpheld;
474#endif
8ad65e08 475
46a3f46d 476 /*
27e88a6e
MD
477 * Switching from within a 'fast' (non thread switched) interrupt or IPI
478 * is illegal. However, we may have to do it anyway if we hit a fatal
479 * kernel trap or we have paniced.
480 *
481 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 482 */
27e88a6e
MD
483 if (gd->gd_intr_nesting_level) {
484 int savegdnest;
485 int savegdtrap;
486
487 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
488 panic("lwkt_switch: cannot switch from within "
489 "a fast interrupt, yet, td %p\n", td);
490 } else {
491 savegdnest = gd->gd_intr_nesting_level;
492 savegdtrap = gd->gd_trap_nesting_level;
493 gd->gd_intr_nesting_level = 0;
494 gd->gd_trap_nesting_level = 0;
a7422615
MD
495 if ((td->td_flags & TDF_PANICWARN) == 0) {
496 td->td_flags |= TDF_PANICWARN;
6ea70f76 497 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615 498 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 499 print_backtrace(-1);
a7422615 500 }
27e88a6e
MD
501 lwkt_switch();
502 gd->gd_intr_nesting_level = savegdnest;
503 gd->gd_trap_nesting_level = savegdtrap;
504 return;
505 }
96728c05 506 }
ef0fdad1 507
cb973d15
MD
508 /*
509 * Passive release (used to transition from user to kernel mode
510 * when we block or switch rather then when we enter the kernel).
511 * This function is NOT called if we are switching into a preemption
512 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
513 * our current process designation (if we have one) and become a true
514 * LWKT thread, and may also hand the current process designation to
515 * another process and schedule thread.
cb973d15
MD
516 */
517 if (td->td_release)
518 td->td_release(td);
519
37af14fe 520 crit_enter_gd(gd);
9d265729
MD
521 if (td->td_toks)
522 lwkt_relalltokens(td);
523
524 /*
b02926de
MD
525 * We had better not be holding any spin locks, but don't get into an
526 * endless panic loop.
9d265729 527 */
bbb31c5d
MD
528 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
529 ("lwkt_switch: still holding a shared spinlock %p!",
530 gd->gd_spinlock_rd));
d666840a
MD
531 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
532 ("lwkt_switch: still holding %d exclusive spinlocks!",
533 gd->gd_spinlocks_wr));
9d265729 534
8a8d5d85
MD
535
536#ifdef SMP
537 /*
538 * td_mpcount cannot be used to determine if we currently hold the
539 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
540 * to get the lock, and switch out if it can't. Our ownership of
541 * the actual lock will remain stable while we are in a critical section
542 * (but, of course, another cpu may own or release the lock so the
543 * actual value of mp_lock is not stable).
8a8d5d85
MD
544 */
545 mpheld = MP_LOCK_HELD();
0f7a3396
MD
546#ifdef INVARIANTS
547 if (td->td_cscount) {
6ea70f76 548 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
549 td);
550 if (panic_on_cscount)
551 panic("switching while mastering cpusync");
552 }
553#endif
8a8d5d85 554#endif
99df837e
MD
555 if ((ntd = td->td_preempted) != NULL) {
556 /*
557 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
558 * thread. This occurs transparently, whether the preempted thread
559 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
560 * itself).
561 *
562 * We have to setup the MP lock for the original thread after backing
563 * out the adjustment that was made to curthread when the original
564 * was preempted.
99df837e 565 */
26a0694b 566 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 567#ifdef SMP
96728c05 568 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 569 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
570 td, ntd, td->td_mpcount, ntd->td_mpcount);
571 }
8a8d5d85
MD
572 if (ntd->td_mpcount) {
573 td->td_mpcount -= ntd->td_mpcount;
574 KKASSERT(td->td_mpcount >= 0);
575 }
576#endif
26a0694b 577 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
578
579 /*
b9eb1c19
MD
580 * The interrupt may have woken a thread up, we need to properly
581 * set the reschedule flag if the originally interrupted thread is
582 * at a lower priority.
8ec60c3f
MD
583 */
584 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
585 need_lwkt_resched();
8a8d5d85 586 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 587 } else {
4b5f931b
MD
588 /*
589 * Priority queue / round-robin at each priority. Note that user
590 * processes run at a fixed, low priority and the user process
591 * scheduler deals with interactions between user processes
592 * by scheduling and descheduling them from the LWKT queue as
593 * necessary.
8a8d5d85
MD
594 *
595 * We have to adjust the MP lock for the target thread. If we
596 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
597 * thread that does not need the MP lock. If we cannot, we spin
598 * instead of HLT.
599 *
600 * A similar issue exists for the tokens held by the target thread.
601 * If we cannot obtain ownership of the tokens we cannot immediately
602 * schedule the thread.
603 */
604
605 /*
8ec60c3f
MD
606 * If an LWKT reschedule was requested, well that is what we are
607 * doing now so clear it.
608 */
609 clear_lwkt_resched();
4b5f931b
MD
610again:
611 if (gd->gd_runqmask) {
612 int nq = bsrl(gd->gd_runqmask);
613 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
614 gd->gd_runqmask &= ~(1 << nq);
615 goto again;
616 }
8a8d5d85 617#ifdef SMP
41a01a4d 618 /*
df6b8ba0
MD
619 * THREAD SELECTION FOR AN SMP MACHINE BUILD
620 *
41a01a4d
MD
621 * If the target needs the MP lock and we couldn't get it,
622 * or if the target is holding tokens and we could not
623 * gain ownership of the tokens, continue looking for a
624 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
625 *
626 * NOTE: the mpheld variable invalid after this conditional, it
627 * can change due to both cpu_try_mplock() returning success
9d265729 628 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
629 * we are trying to check the mpcount of a thread other then
630 * the current thread. Because of this, if the current thread
631 * is not holding td_mpcount, an IPI indirectly run via
9d265729 632 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 633 * cause the core MP lock to be released.
41a01a4d
MD
634 */
635 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
9d265729 636 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
41a01a4d 637 ) {
8a8d5d85 638 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
639
640 mpheld = MP_LOCK_HELD();
641 ntd = NULL;
8a8d5d85
MD
642 while (rqmask) {
643 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 644 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 645 /* spinning due to MP lock being held */
41a01a4d 646 continue;
38717797 647 }
a453459d
MD
648
649 /*
9d265729 650 * mpheld state invalid after getalltokens call returns
a453459d
MD
651 * failure, but the variable is only needed for
652 * the loop.
653 */
9d265729 654 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
a453459d 655 /* spinning due to token contention */
38717797 656#ifdef INVARIANTS
a453459d 657 ++token_contention_count;
38717797 658#endif
a453459d 659 mpheld = MP_LOCK_HELD();
41a01a4d 660 continue;
38717797 661 }
41a01a4d 662 break;
8a8d5d85
MD
663 }
664 if (ntd)
665 break;
666 rqmask &= ~(1 << nq);
667 nq = bsrl(rqmask);
b9eb1c19
MD
668
669 /*
670 * We have two choices. We can either refuse to run a
671 * user thread when a kernel thread needs the MP lock
672 * but could not get it, or we can allow it to run but
673 * then expect an IPI (hopefully) later on to force a
674 * reschedule when the MP lock might become available.
675 */
676 if (nq < TDPRI_KERN_LPSCHED) {
684a93c4
MD
677 break; /* for now refuse to run */
678#if 0
b9eb1c19
MD
679 if (chain_mplock == 0)
680 break;
b9eb1c19 681 /* continue loop, allow user threads to be scheduled */
684a93c4 682#endif
b9eb1c19 683 }
8a8d5d85 684 }
684a93c4
MD
685
686 /*
687 * Case where a (kernel) thread needed the MP lock and could
688 * not get one, and we may or may not have found another
689 * thread which does not need the MP lock to run while
690 * we wait (ntd).
691 */
8a8d5d85 692 if (ntd == NULL) {
a2a5ad0d
MD
693 ntd = &gd->gd_idlethread;
694 ntd->td_flags |= TDF_IDLE_NOHLT;
684a93c4
MD
695 set_mplock_contention_mask(gd);
696 cpu_mplock_contested();
df6b8ba0 697 goto using_idle_thread;
8a8d5d85 698 } else {
684a93c4 699 clr_mplock_contention_mask(gd);
344ad853 700 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
701 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
702 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
703 }
704 } else {
684a93c4 705 clr_mplock_contention_mask(gd);
344ad853 706 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
707 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
708 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
709 }
710#else
df6b8ba0
MD
711 /*
712 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
7eb611ef
MD
713 * worry about tokens or the BGL. However, we still have
714 * to call lwkt_getalltokens() in order to properly detect
715 * stale tokens. This call cannot fail for a UP build!
df6b8ba0 716 */
7eb611ef 717 lwkt_getalltokens(ntd);
344ad853 718 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
719 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
720 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 721#endif
4b5f931b 722 } else {
3c23a41a 723 /*
60f945af
MD
724 * We have nothing to run but only let the idle loop halt
725 * the cpu if there are no pending interrupts.
3c23a41a 726 */
a2a5ad0d 727 ntd = &gd->gd_idlethread;
60f945af 728 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 729 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 730#ifdef SMP
df6b8ba0
MD
731using_idle_thread:
732 /*
733 * The idle thread should not be holding the MP lock unless we
734 * are trapping in the kernel or in a panic. Since we select the
735 * idle thread unconditionally when no other thread is available,
736 * if the MP lock is desired during a panic or kernel trap, we
737 * have to loop in the scheduler until we get it.
738 */
739 if (ntd->td_mpcount) {
740 mpheld = MP_LOCK_HELD();
684a93c4 741 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
df6b8ba0 742 panic("Idle thread %p was holding the BGL!", ntd);
684a93c4 743 if (mpheld == 0)
df6b8ba0
MD
744 goto again;
745 }
a453459d 746#endif
4b5f931b 747 }
f1d1c3fa 748 }
26a0694b
MD
749 KASSERT(ntd->td_pri >= TDPRI_CRIT,
750 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
751
752 /*
753 * Do the actual switch. If the new target does not need the MP lock
754 * and we are holding it, release the MP lock. If the new target requires
755 * the MP lock we have already acquired it for the target.
756 */
757#ifdef SMP
758 if (ntd->td_mpcount == 0 ) {
759 if (MP_LOCK_HELD())
760 cpu_rel_mplock();
761 } else {
a453459d 762 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
763 }
764#endif
94f6d86e
MD
765 if (td != ntd) {
766 ++switch_count;
b2b3ffcd 767#ifdef __x86_64__
f64b567c
SW
768 {
769 int tos_ok __debugvar = jg_tos_ok(ntd);
770 KKASSERT(tos_ok);
771 }
85514115 772#endif
d850923c 773 KTR_LOG(ctxsw_sw, td, ntd);
f1d1c3fa 774 td->td_switch(ntd);
94f6d86e 775 }
37af14fe
MD
776 /* NOTE: current cpu may have changed after switch */
777 crit_exit_quick(td);
8ad65e08
MD
778}
779
f1d1c3fa 780/*
96728c05
MD
781 * Request that the target thread preempt the current thread. Preemption
782 * only works under a specific set of conditions:
b68b7282 783 *
96728c05
MD
784 * - We are not preempting ourselves
785 * - The target thread is owned by the current cpu
786 * - We are not currently being preempted
787 * - The target is not currently being preempted
d3d1cbc8
MD
788 * - We are not holding any spin locks
789 * - The target thread is not holding any tokens
96728c05
MD
790 * - We are able to satisfy the target's MP lock requirements (if any).
791 *
792 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
793 * this is called via lwkt_schedule() through the td_preemptable callback.
794 * critpri is the managed critical priority that we should ignore in order
795 * to determine whether preemption is possible (aka usually just the crit
796 * priority of lwkt_schedule() itself).
b68b7282 797 *
26a0694b
MD
798 * XXX at the moment we run the target thread in a critical section during
799 * the preemption in order to prevent the target from taking interrupts
800 * that *WE* can't. Preemption is strictly limited to interrupt threads
801 * and interrupt-like threads, outside of a critical section, and the
802 * preempted source thread will be resumed the instant the target blocks
803 * whether or not the source is scheduled (i.e. preemption is supposed to
804 * be as transparent as possible).
4b5f931b 805 *
8a8d5d85
MD
806 * The target thread inherits our MP count (added to its own) for the
807 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
808 * MP lock during the preemption. Therefore, any preempting targets must be
809 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
810 * out of sync with the physical mp_lock, but we do not have to preserve
811 * the original ownership of the lock if it was out of synch (that is, we
812 * can leave it synchronized on return).
b68b7282
MD
813 */
814void
96728c05 815lwkt_preempt(thread_t ntd, int critpri)
b68b7282 816{
46a3f46d 817 struct globaldata *gd = mycpu;
0a3f9b47 818 thread_t td;
8a8d5d85
MD
819#ifdef SMP
820 int mpheld;
57c254db 821 int savecnt;
8a8d5d85 822#endif
b68b7282 823
26a0694b 824 /*
96728c05
MD
825 * The caller has put us in a critical section. We can only preempt
826 * if the caller of the caller was not in a critical section (basically
d666840a 827 * a local interrupt), as determined by the 'critpri' parameter. We
47737962 828 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 829 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
830 *
831 * YYY The target thread must be in a critical section (else it must
832 * inherit our critical section? I dunno yet).
41a01a4d 833 *
0a3f9b47 834 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
835 */
836 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 837
0a3f9b47 838 td = gd->gd_curthread;
0a3f9b47 839 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
840 ++preempt_miss;
841 return;
842 }
96728c05
MD
843 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
844 ++preempt_miss;
8ec60c3f 845 need_lwkt_resched();
96728c05
MD
846 return;
847 }
848#ifdef SMP
46a3f46d 849 if (ntd->td_gd != gd) {
96728c05 850 ++preempt_miss;
8ec60c3f 851 need_lwkt_resched();
96728c05
MD
852 return;
853 }
854#endif
41a01a4d 855 /*
d3d1cbc8 856 * Take the easy way out and do not preempt if we are holding
d666840a 857 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
858 * preempted interlock against the target thread's tokens and whether
859 * we can get all the target thread's tokens, but this situation
860 * should not occur very often so its easier to simply not preempt.
d666840a
MD
861 * Also, plain spinlocks are impossible to figure out at this point so
862 * just don't preempt.
d3d1cbc8
MD
863 *
864 * Do not try to preempt if the target thread is holding any tokens.
865 * We could try to acquire the tokens but this case is so rare there
866 * is no need to support it.
41a01a4d 867 */
bbb31c5d 868 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 869 ++preempt_miss;
8ec60c3f 870 need_lwkt_resched();
41a01a4d
MD
871 return;
872 }
d3d1cbc8
MD
873 if (ntd->td_toks) {
874 ++preempt_miss;
875 need_lwkt_resched();
876 return;
877 }
26a0694b
MD
878 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
879 ++preempt_weird;
8ec60c3f 880 need_lwkt_resched();
26a0694b
MD
881 return;
882 }
883 if (ntd->td_preempted) {
4b5f931b 884 ++preempt_hit;
8ec60c3f 885 need_lwkt_resched();
26a0694b 886 return;
b68b7282 887 }
8a8d5d85 888#ifdef SMP
a2a5ad0d
MD
889 /*
890 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
891 * to or from the MP lock. In this case td_mpcount will be pre-disposed
892 * (non-zero) but not actually synchronized with the actual state of the
893 * lock. We can use it to imply an MP lock requirement for the
894 * preemption but we cannot use it to test whether we hold the MP lock
895 * or not.
a2a5ad0d 896 */
96728c05 897 savecnt = td->td_mpcount;
71ef2f5c 898 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
899 ntd->td_mpcount += td->td_mpcount;
900 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
901 ntd->td_mpcount -= td->td_mpcount;
902 ++preempt_miss;
8ec60c3f 903 need_lwkt_resched();
8a8d5d85
MD
904 return;
905 }
906#endif
26a0694b 907
8ec60c3f
MD
908 /*
909 * Since we are able to preempt the current thread, there is no need to
910 * call need_lwkt_resched().
911 */
26a0694b
MD
912 ++preempt_hit;
913 ntd->td_preempted = td;
914 td->td_flags |= TDF_PREEMPT_LOCK;
d850923c 915 KTR_LOG(ctxsw_pre, td, ntd);
26a0694b 916 td->td_switch(ntd);
b9eb1c19 917
26a0694b 918 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
919#ifdef SMP
920 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
921 mpheld = MP_LOCK_HELD();
922 if (mpheld && td->td_mpcount == 0)
96728c05 923 cpu_rel_mplock();
71ef2f5c 924 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
925 panic("lwkt_preempt(): MP lock was not held through");
926#endif
26a0694b
MD
927 ntd->td_preempted = NULL;
928 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
929}
930
931/*
faaeffac 932 * Conditionally call splz() if gd_reqflags indicates work is pending.
f1d1c3fa 933 *
faaeffac
MD
934 * td_nest_count prevents deep nesting via splz() or doreti() which
935 * might otherwise blow out the kernel stack. Note that except for
936 * this special case, we MUST call splz() here to handle any
937 * pending ints, particularly after we switch, or we might accidently
938 * halt the cpu with interrupts pending.
ef0fdad1 939 *
f1d1c3fa
MD
940 * (self contained on a per cpu basis)
941 */
942void
faaeffac 943splz_check(void)
f1d1c3fa 944{
7966cb69
MD
945 globaldata_t gd = mycpu;
946 thread_t td = gd->gd_curthread;
ef0fdad1 947
46a3f46d 948 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 949 splz();
f1d1c3fa
MD
950}
951
8ad65e08 952/*
faaeffac
MD
953 * This implements a normal yield which will yield to equal priority
954 * threads as well as higher priority threads. Note that gd_reqflags
955 * tests will be handled by the crit_exit() call in lwkt_switch().
f1d1c3fa
MD
956 *
957 * (self contained on a per cpu basis)
8ad65e08
MD
958 */
959void
f1d1c3fa 960lwkt_yield(void)
8ad65e08 961{
37af14fe 962 lwkt_schedule_self(curthread);
f1d1c3fa
MD
963 lwkt_switch();
964}
965
966/*
3824f392
MD
967 * This function is used along with the lwkt_passive_recover() inline
968 * by the trap code to negotiate a passive release of the current
969 * process/lwp designation with the user scheduler.
970 */
971void
972lwkt_passive_release(struct thread *td)
973{
974 struct lwp *lp = td->td_lwp;
975
976 td->td_release = NULL;
977 lwkt_setpri_self(TDPRI_KERN_USER);
978 lp->lwp_proc->p_usched->release_curproc(lp);
979}
980
981/*
982 * Make a kernel thread act as if it were in user mode with regards
983 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel
984 * loops which may be potentially cpu-bound can call lwkt_user_yield().
985 *
986 * The lwkt_user_yield() function is designed to have very low overhead
987 * if no yield is determined to be needed.
988 */
989void
990lwkt_user_yield(void)
991{
992 thread_t td = curthread;
993 struct lwp *lp = td->td_lwp;
994
995#ifdef SMP
996 /*
997 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
998 * kernel can prevent other cpus from servicing interrupt threads
999 * which still require the MP lock (which is a lot of them). This
1000 * has a chaining effect since if the interrupt is blocked, so is
1001 * the event, so normal scheduling will not pick up on the problem.
1002 */
684a93c4
MD
1003 if (mp_lock_contention_mask && td->td_mpcount) {
1004 yield_mplock(td);
3824f392
MD
1005 }
1006#endif
1007
1008 /*
1009 * Another kernel thread wants the cpu
1010 */
1011 if (lwkt_resched_wanted())
1012 lwkt_switch();
1013
1014 /*
1015 * If the user scheduler has asynchronously determined that the current
1016 * process (when running in user mode) needs to lose the cpu then make
1017 * sure we are released.
1018 */
1019 if (user_resched_wanted()) {
1020 if (td->td_release)
1021 td->td_release(td);
1022 }
1023
1024 /*
1025 * If we are released reduce our priority
1026 */
1027 if (td->td_release == NULL) {
1028 if (lwkt_check_resched(td) > 0)
1029 lwkt_switch();
e381e77c
MD
1030 if (lp) {
1031 lp->lwp_proc->p_usched->acquire_curproc(lp);
1032 td->td_release = lwkt_passive_release;
1033 lwkt_setpri_self(TDPRI_USER_NORM);
1034 }
3824f392
MD
1035 }
1036}
1037
1038/*
b9eb1c19
MD
1039 * Return 0 if no runnable threads are pending at the same or higher
1040 * priority as the passed thread.
1041 *
1042 * Return 1 if runnable threads are pending at the same priority.
1043 *
1044 * Return 2 if runnable threads are pending at a higher priority.
1045 */
1046int
1047lwkt_check_resched(thread_t td)
1048{
1049 int pri = td->td_pri & TDPRI_MASK;
1050
1051 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1052 return(2);
1053 if (TAILQ_NEXT(td, td_threadq))
1054 return(1);
1055 return(0);
1056}
1057
1058/*
f1d1c3fa
MD
1059 * Generic schedule. Possibly schedule threads belonging to other cpus and
1060 * deal with threads that might be blocked on a wait queue.
1061 *
0a3f9b47
MD
1062 * We have a little helper inline function which does additional work after
1063 * the thread has been enqueued, including dealing with preemption and
1064 * setting need_lwkt_resched() (which prevents the kernel from returning
1065 * to userland until it has processed higher priority threads).
6330a558
MD
1066 *
1067 * It is possible for this routine to be called after a failed _enqueue
1068 * (due to the target thread migrating, sleeping, or otherwise blocked).
1069 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1070 *
1071 * reschedok is an optimized constant propagated from lwkt_schedule() or
1072 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1073 * reschedule to be requested if the target thread has a higher priority.
1074 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1075 * be 0, prevented undesired reschedules.
8ad65e08 1076 */
0a3f9b47
MD
1077static __inline
1078void
361d01dd 1079_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
0a3f9b47 1080{
b9eb1c19 1081 thread_t otd;
c730be20 1082
6330a558 1083 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1084 if (ntd->td_preemptable && reschedok) {
6330a558 1085 ntd->td_preemptable(ntd, cpri); /* YYY +token */
361d01dd 1086 } else if (reschedok) {
b9eb1c19
MD
1087 otd = curthread;
1088 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
c730be20 1089 need_lwkt_resched();
6330a558 1090 }
0a3f9b47
MD
1091 }
1092}
1093
361d01dd 1094static __inline
8ad65e08 1095void
361d01dd 1096_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1097{
37af14fe
MD
1098 globaldata_t mygd = mycpu;
1099
41a01a4d 1100 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1101 crit_enter_gd(mygd);
9388413d 1102 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1103 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1104 _lwkt_enqueue(td);
1105 } else {
f1d1c3fa 1106 /*
7cd8d145
MD
1107 * If we own the thread, there is no race (since we are in a
1108 * critical section). If we do not own the thread there might
1109 * be a race but the target cpu will deal with it.
f1d1c3fa 1110 */
0f7a3396 1111#ifdef SMP
7cd8d145 1112 if (td->td_gd == mygd) {
9d265729 1113 _lwkt_enqueue(td);
361d01dd 1114 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
f1d1c3fa 1115 } else {
e381e77c 1116 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1117 }
0f7a3396 1118#else
7cd8d145 1119 _lwkt_enqueue(td);
361d01dd 1120 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
0f7a3396 1121#endif
8ad65e08 1122 }
37af14fe 1123 crit_exit_gd(mygd);
8ad65e08
MD
1124}
1125
361d01dd
MD
1126void
1127lwkt_schedule(thread_t td)
1128{
1129 _lwkt_schedule(td, 1);
1130}
1131
1132void
1133lwkt_schedule_noresched(thread_t td)
1134{
1135 _lwkt_schedule(td, 0);
1136}
1137
88ebb169
SW
1138#ifdef SMP
1139
e381e77c
MD
1140/*
1141 * When scheduled remotely if frame != NULL the IPIQ is being
1142 * run via doreti or an interrupt then preemption can be allowed.
1143 *
1144 * To allow preemption we have to drop the critical section so only
1145 * one is present in _lwkt_schedule_post.
1146 */
1147static void
1148lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1149{
1150 thread_t td = curthread;
1151 thread_t ntd = arg;
1152
1153 if (frame && ntd->td_preemptable) {
1154 crit_exit_noyield(td);
1155 _lwkt_schedule(ntd, 1);
1156 crit_enter_quick(td);
1157 } else {
1158 _lwkt_schedule(ntd, 1);
1159 }
1160}
1161
d9eea1a5 1162/*
52eedfb5
MD
1163 * Thread migration using a 'Pull' method. The thread may or may not be
1164 * the current thread. It MUST be descheduled and in a stable state.
1165 * lwkt_giveaway() must be called on the cpu owning the thread.
1166 *
1167 * At any point after lwkt_giveaway() is called, the target cpu may
1168 * 'pull' the thread by calling lwkt_acquire().
1169 *
ae8e83e6
MD
1170 * We have to make sure the thread is not sitting on a per-cpu tsleep
1171 * queue or it will blow up when it moves to another cpu.
1172 *
52eedfb5 1173 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1174 */
a2a5ad0d 1175void
52eedfb5
MD
1176lwkt_giveaway(thread_t td)
1177{
3b4192fb 1178 globaldata_t gd = mycpu;
52eedfb5 1179
3b4192fb
MD
1180 crit_enter_gd(gd);
1181 if (td->td_flags & TDF_TSLEEPQ)
1182 tsleep_remove(td);
1183 KKASSERT(td->td_gd == gd);
1184 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1185 td->td_flags |= TDF_MIGRATING;
1186 crit_exit_gd(gd);
52eedfb5
MD
1187}
1188
1189void
a2a5ad0d
MD
1190lwkt_acquire(thread_t td)
1191{
37af14fe
MD
1192 globaldata_t gd;
1193 globaldata_t mygd;
a2a5ad0d 1194
52eedfb5 1195 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1196 gd = td->td_gd;
37af14fe 1197 mygd = mycpu;
52eedfb5 1198 if (gd != mycpu) {
35238fa5 1199 cpu_lfence();
52eedfb5 1200 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1201 crit_enter_gd(mygd);
df910c23
MD
1202 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1203#ifdef SMP
1204 lwkt_process_ipiq();
1205#endif
52eedfb5 1206 cpu_lfence();
df910c23 1207 }
37af14fe 1208 td->td_gd = mygd;
52eedfb5
MD
1209 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1210 td->td_flags &= ~TDF_MIGRATING;
1211 crit_exit_gd(mygd);
1212 } else {
1213 crit_enter_gd(mygd);
1214 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1215 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1216 crit_exit_gd(mygd);
a2a5ad0d
MD
1217 }
1218}
1219
52eedfb5
MD
1220#endif
1221
8ad65e08 1222/*
f1d1c3fa
MD
1223 * Generic deschedule. Descheduling threads other then your own should be
1224 * done only in carefully controlled circumstances. Descheduling is
1225 * asynchronous.
1226 *
1227 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1228 */
1229void
1230lwkt_deschedule(thread_t td)
1231{
f1d1c3fa 1232 crit_enter();
b8a98473 1233#ifdef SMP
f1d1c3fa
MD
1234 if (td == curthread) {
1235 _lwkt_dequeue(td);
1236 } else {
a72187e9 1237 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1238 _lwkt_dequeue(td);
1239 } else {
b8a98473 1240 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1241 }
1242 }
b8a98473
MD
1243#else
1244 _lwkt_dequeue(td);
1245#endif
f1d1c3fa
MD
1246 crit_exit();
1247}
1248
1249/*
4b5f931b
MD
1250 * Set the target thread's priority. This routine does not automatically
1251 * switch to a higher priority thread, LWKT threads are not designed for
1252 * continuous priority changes. Yield if you want to switch.
1253 *
1254 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1255 * of the td_pri field. The specified priority may also indicate zero or
1256 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1257 *
1258 * Note that we requeue the thread whether it winds up on a different runq
1259 * or not. uio_yield() depends on this and the routine is not normally
1260 * called with the same priority otherwise.
4b5f931b
MD
1261 */
1262void
1263lwkt_setpri(thread_t td, int pri)
1264{
26a0694b 1265 KKASSERT(pri >= 0);
a72187e9 1266 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1267 crit_enter();
1268 if (td->td_flags & TDF_RUNQ) {
1269 _lwkt_dequeue(td);
1270 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1271 _lwkt_enqueue(td);
1272 } else {
1273 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1274 }
1275 crit_exit();
1276}
1277
03bd0a5e
MD
1278/*
1279 * Set the initial priority for a thread prior to it being scheduled for
1280 * the first time. The thread MUST NOT be scheduled before or during
1281 * this call. The thread may be assigned to a cpu other then the current
1282 * cpu.
1283 *
1284 * Typically used after a thread has been created with TDF_STOPPREQ,
1285 * and before the thread is initially scheduled.
1286 */
1287void
1288lwkt_setpri_initial(thread_t td, int pri)
1289{
1290 KKASSERT(pri >= 0);
1291 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1292 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1293}
1294
26a0694b
MD
1295void
1296lwkt_setpri_self(int pri)
1297{
1298 thread_t td = curthread;
1299
4b5f931b
MD
1300 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1301 crit_enter();
1302 if (td->td_flags & TDF_RUNQ) {
1303 _lwkt_dequeue(td);
1304 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1305 _lwkt_enqueue(td);
1306 } else {
1307 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1308 }
1309 crit_exit();
1310}
1311
5d21b981 1312/*
52eedfb5
MD
1313 * Migrate the current thread to the specified cpu.
1314 *
1315 * This is accomplished by descheduling ourselves from the current cpu,
1316 * moving our thread to the tdallq of the target cpu, IPI messaging the
1317 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1318 * races while the thread is being migrated.
ae8e83e6
MD
1319 *
1320 * We must be sure to remove ourselves from the current cpu's tsleepq
1321 * before potentially moving to another queue. The thread can be on
1322 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1323 */
3d28ff59 1324#ifdef SMP
5d21b981 1325static void lwkt_setcpu_remote(void *arg);
3d28ff59 1326#endif
5d21b981
MD
1327
1328void
1329lwkt_setcpu_self(globaldata_t rgd)
1330{
1331#ifdef SMP
1332 thread_t td = curthread;
1333
1334 if (td->td_gd != rgd) {
1335 crit_enter_quick(td);
ae8e83e6 1336 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1337 tsleep_remove(td);
5d21b981
MD
1338 td->td_flags |= TDF_MIGRATING;
1339 lwkt_deschedule_self(td);
52eedfb5 1340 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1341 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1342 lwkt_switch();
1343 /* we are now on the target cpu */
52eedfb5 1344 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1345 crit_exit_quick(td);
1346 }
1347#endif
1348}
1349
ecdefdda
MD
1350void
1351lwkt_migratecpu(int cpuid)
1352{
1353#ifdef SMP
1354 globaldata_t rgd;
1355
1356 rgd = globaldata_find(cpuid);
1357 lwkt_setcpu_self(rgd);
1358#endif
1359}
1360
5d21b981
MD
1361/*
1362 * Remote IPI for cpu migration (called while in a critical section so we
1363 * do not have to enter another one). The thread has already been moved to
1364 * our cpu's allq, but we must wait for the thread to be completely switched
1365 * out on the originating cpu before we schedule it on ours or the stack
1366 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1367 * change to main memory.
1368 *
1369 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1370 * against wakeups. It is best if this interface is used only when there
1371 * are no pending events that might try to schedule the thread.
1372 */
3d28ff59 1373#ifdef SMP
5d21b981
MD
1374static void
1375lwkt_setcpu_remote(void *arg)
1376{
1377 thread_t td = arg;
1378 globaldata_t gd = mycpu;
1379
df910c23
MD
1380 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1381#ifdef SMP
1382 lwkt_process_ipiq();
1383#endif
35238fa5 1384 cpu_lfence();
df910c23 1385 }
5d21b981 1386 td->td_gd = gd;
35238fa5 1387 cpu_sfence();
5d21b981 1388 td->td_flags &= ~TDF_MIGRATING;
9388413d 1389 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1390 _lwkt_enqueue(td);
1391}
3d28ff59 1392#endif
5d21b981 1393
553ea3c8 1394struct lwp *
4b5f931b
MD
1395lwkt_preempted_proc(void)
1396{
73e4f7b9 1397 thread_t td = curthread;
4b5f931b
MD
1398 while (td->td_preempted)
1399 td = td->td_preempted;
553ea3c8 1400 return(td->td_lwp);
4b5f931b
MD
1401}
1402
4b5f931b 1403/*
99df837e
MD
1404 * Create a kernel process/thread/whatever. It shares it's address space
1405 * with proc0 - ie: kernel only.
1406 *
365fa13f
MD
1407 * NOTE! By default new threads are created with the MP lock held. A
1408 * thread which does not require the MP lock should release it by calling
1409 * rel_mplock() at the start of the new thread.
99df837e
MD
1410 */
1411int
1412lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1413 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1414 const char *fmt, ...)
99df837e 1415{
73e4f7b9 1416 thread_t td;
e2565a42 1417 __va_list ap;
99df837e 1418
d3d32139 1419 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1420 tdflags);
a2a5ad0d
MD
1421 if (tdp)
1422 *tdp = td;
709799ea 1423 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1424
1425 /*
1426 * Set up arg0 for 'ps' etc
1427 */
e2565a42 1428 __va_start(ap, fmt);
379210cb 1429 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1430 __va_end(ap);
99df837e
MD
1431
1432 /*
1433 * Schedule the thread to run
1434 */
ef0fdad1
MD
1435 if ((td->td_flags & TDF_STOPREQ) == 0)
1436 lwkt_schedule(td);
1437 else
1438 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1439 return 0;
1440}
1441
1442/*
1443 * Destroy an LWKT thread. Warning! This function is not called when
1444 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1445 * uses a different reaping mechanism.
1446 */
1447void
1448lwkt_exit(void)
1449{
1450 thread_t td = curthread;
c070746a 1451 thread_t std;
8826f33a 1452 globaldata_t gd;
99df837e
MD
1453
1454 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1455 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1456 caps_exit(td);
c070746a
MD
1457
1458 /*
1459 * Get us into a critical section to interlock gd_freetd and loop
1460 * until we can get it freed.
1461 *
1462 * We have to cache the current td in gd_freetd because objcache_put()ing
1463 * it would rip it out from under us while our thread is still active.
1464 */
1465 gd = mycpu;
37af14fe 1466 crit_enter_quick(td);
c070746a
MD
1467 while ((std = gd->gd_freetd) != NULL) {
1468 gd->gd_freetd = NULL;
1469 objcache_put(thread_cache, std);
1470 }
3b4192fb
MD
1471
1472 /*
1473 * Remove thread resources from kernel lists and deschedule us for
1474 * the last time.
1475 */
1476 if (td->td_flags & TDF_TSLEEPQ)
1477 tsleep_remove(td);
79eae878 1478 biosched_done(td);
f8abf63c 1479 dsched_exit_thread(td);
37af14fe 1480 lwkt_deschedule_self(td);
e56e4dea 1481 lwkt_remove_tdallq(td);
c070746a
MD
1482 if (td->td_flags & TDF_ALLOCATED_THREAD)
1483 gd->gd_freetd = td;
99df837e
MD
1484 cpu_thread_exit();
1485}
1486
e56e4dea
MD
1487void
1488lwkt_remove_tdallq(thread_t td)
1489{
1490 KKASSERT(td->td_gd == mycpu);
1491 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1492}
1493
2d93b37a
MD
1494void
1495crit_panic(void)
1496{
1497 thread_t td = curthread;
1498 int lpri = td->td_pri;
1499
1500 td->td_pri = 0;
1501 panic("td_pri is/would-go negative! %p %d", td, lpri);
1502}
1503
d165e668
MD
1504#ifdef SMP
1505
bd8015ca
MD
1506/*
1507 * Called from debugger/panic on cpus which have been stopped. We must still
1508 * process the IPIQ while stopped, even if we were stopped while in a critical
1509 * section (XXX).
1510 *
1511 * If we are dumping also try to process any pending interrupts. This may
1512 * or may not work depending on the state of the cpu at the point it was
1513 * stopped.
1514 */
1515void
1516lwkt_smp_stopped(void)
1517{
1518 globaldata_t gd = mycpu;
1519
1520 crit_enter_gd(gd);
1521 if (dumping) {
1522 lwkt_process_ipiq();
1523 splz();
1524 } else {
1525 lwkt_process_ipiq();
1526 }
1527 crit_exit_gd(gd);
1528}
1529
d165e668 1530#endif