Support disablement of chflags in a jail, part 1/2.
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
8c10bfcf
MD
2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
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:
8c10bfcf 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.
20 *
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.
8c10bfcf 33 *
bd8015ca 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.80 2005/07/20 20:21:28 dillon Exp $
75cdbe6c
MD
35 */
36
37/*
38 * Each cpu in a system has its own self-contained light weight kernel
39 * thread scheduler, which means that generally speaking we only need
40 * to use a critical section to avoid problems. Foreign thread
41 * scheduling is queued via (async) IPIs.
8ad65e08
MD
42 */
43
05220613
MD
44#ifdef _KERNEL
45
8ad65e08
MD
46#include <sys/param.h>
47#include <sys/systm.h>
48#include <sys/kernel.h>
49#include <sys/proc.h>
50#include <sys/rtprio.h>
51#include <sys/queue.h>
f1d1c3fa 52#include <sys/thread2.h>
7d0bac62 53#include <sys/sysctl.h>
99df837e 54#include <sys/kthread.h>
f1d1c3fa 55#include <machine/cpu.h>
99df837e 56#include <sys/lock.h>
f6bf3af1 57#include <sys/caps.h>
f1d1c3fa 58
7d0bac62
MD
59#include <vm/vm.h>
60#include <vm/vm_param.h>
61#include <vm/vm_kern.h>
62#include <vm/vm_object.h>
63#include <vm/vm_page.h>
64#include <vm/vm_map.h>
65#include <vm/vm_pager.h>
66#include <vm/vm_extern.h>
67#include <vm/vm_zone.h>
68
99df837e 69#include <machine/stdarg.h>
57c254db 70#include <machine/ipl.h>
96728c05 71#include <machine/smp.h>
99df837e 72
05220613
MD
73#else
74
75#include <sys/stdint.h>
fb04f4fd 76#include <libcaps/thread.h>
05220613
MD
77#include <sys/thread.h>
78#include <sys/msgport.h>
79#include <sys/errno.h>
fb04f4fd 80#include <libcaps/globaldata.h>
7e8303ad 81#include <machine/cpufunc.h>
05220613
MD
82#include <sys/thread2.h>
83#include <sys/msgport2.h>
709799ea 84#include <stdio.h>
05220613 85#include <stdlib.h>
709799ea 86#include <string.h>
709799ea 87#include <machine/lock.h>
b950d4b8
EN
88#include <machine/atomic.h>
89#include <machine/cpu.h>
05220613
MD
90
91#endif
92
7d0bac62 93static int untimely_switch = 0;
0f7a3396
MD
94#ifdef INVARIANTS
95static int panic_on_cscount = 0;
96#endif
05220613
MD
97static __int64_t switch_count = 0;
98static __int64_t preempt_hit = 0;
99static __int64_t preempt_miss = 0;
100static __int64_t preempt_weird = 0;
38717797
HP
101static __int64_t token_contention_count = 0;
102static __int64_t mplock_contention_count = 0;
05220613
MD
103
104#ifdef _KERNEL
105
106SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
0f7a3396
MD
107#ifdef INVARIANTS
108SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
109#endif
4b5f931b 110SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 111SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 112SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 113SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
114#ifdef INVARIANTS
115SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
116 &token_contention_count, 0, "spinning due to token contention");
117SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
118 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
119#endif
05220613
MD
120#endif
121
4b5f931b
MD
122/*
123 * These helper procedures handle the runq, they can only be called from
124 * within a critical section.
75cdbe6c
MD
125 *
126 * WARNING! Prior to SMP being brought up it is possible to enqueue and
127 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
128 * instead of 'mycpu' when referencing the globaldata structure. Once
129 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 130 */
f1d1c3fa
MD
131static __inline
132void
133_lwkt_dequeue(thread_t td)
134{
135 if (td->td_flags & TDF_RUNQ) {
4b5f931b 136 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 137 struct globaldata *gd = td->td_gd;
4b5f931b 138
f1d1c3fa 139 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
140 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
141 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
142 }
143}
144
145static __inline
146void
147_lwkt_enqueue(thread_t td)
148{
5d21b981 149 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING)) == 0) {
4b5f931b 150 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 151 struct globaldata *gd = td->td_gd;
4b5f931b 152
f1d1c3fa 153 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
154 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
155 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
156 }
157}
8ad65e08 158
37af14fe
MD
159/*
160 * Schedule a thread to run. As the current thread we can always safely
161 * schedule ourselves, and a shortcut procedure is provided for that
162 * function.
163 *
164 * (non-blocking, self contained on a per cpu basis)
165 */
166void
167lwkt_schedule_self(thread_t td)
168{
169 crit_enter_quick(td);
170 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
171 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
172 _lwkt_enqueue(td);
173#ifdef _KERNEL
174 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
175 panic("SCHED SELF PANIC");
176#endif
177 crit_exit_quick(td);
178}
179
180/*
181 * Deschedule a thread.
182 *
183 * (non-blocking, self contained on a per cpu basis)
184 */
185void
186lwkt_deschedule_self(thread_t td)
187{
188 crit_enter_quick(td);
189 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
190 _lwkt_dequeue(td);
191 crit_exit_quick(td);
192}
193
2d93b37a
MD
194#ifdef _KERNEL
195
8ad65e08
MD
196/*
197 * LWKTs operate on a per-cpu basis
198 *
73e4f7b9 199 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
200 */
201void
202lwkt_gdinit(struct globaldata *gd)
203{
4b5f931b
MD
204 int i;
205
206 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
207 TAILQ_INIT(&gd->gd_tdrunq[i]);
208 gd->gd_runqmask = 0;
73e4f7b9 209 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
210}
211
2d93b37a
MD
212#endif /* _KERNEL */
213
7d0bac62
MD
214/*
215 * Initialize a thread wait structure prior to first use.
216 *
217 * NOTE! called from low level boot code, we cannot do anything fancy!
218 */
219void
41a01a4d 220lwkt_wait_init(lwkt_wait_t w)
7d0bac62 221{
41a01a4d 222 lwkt_token_init(&w->wa_token);
7d0bac62 223 TAILQ_INIT(&w->wa_waitq);
41a01a4d
MD
224 w->wa_gen = 0;
225 w->wa_count = 0;
7d0bac62
MD
226}
227
228/*
229 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
230 * or LWKT start address before it can be scheduled. If the target cpu is
231 * -1 the thread will be created on the current cpu.
0cfcada1
MD
232 *
233 * If you intend to create a thread without a process context this function
234 * does everything except load the startup and switcher function.
7d0bac62
MD
235 */
236thread_t
f470d0c8 237lwkt_alloc_thread(struct thread *td, int stksize, int cpu)
7d0bac62 238{
99df837e 239 void *stack;
ef0fdad1 240 int flags = 0;
37af14fe 241 globaldata_t gd = mycpu;
7d0bac62 242
ef0fdad1 243 if (td == NULL) {
37af14fe
MD
244 crit_enter_gd(gd);
245 if (gd->gd_tdfreecount > 0) {
246 --gd->gd_tdfreecount;
247 td = TAILQ_FIRST(&gd->gd_tdfreeq);
d9eea1a5 248 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1 249 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
37af14fe
MD
250 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq);
251 crit_exit_gd(gd);
ef0fdad1
MD
252 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
253 } else {
37af14fe 254 crit_exit_gd(gd);
05220613 255#ifdef _KERNEL
ef0fdad1 256 td = zalloc(thread_zone);
05220613
MD
257#else
258 td = malloc(sizeof(struct thread));
259#endif
ef0fdad1 260 td->td_kstack = NULL;
f470d0c8 261 td->td_kstack_size = 0;
ef0fdad1
MD
262 flags |= TDF_ALLOCATED_THREAD;
263 }
264 }
f470d0c8
MD
265 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
266 if (flags & TDF_ALLOCATED_STACK) {
a9e5ae82 267#ifdef _KERNEL
f470d0c8 268 kmem_free(kernel_map, (vm_offset_t)stack, td->td_kstack_size);
a9e5ae82
JS
269#else
270 libcaps_free_stack(stack, td->td_kstack_size);
271#endif
f470d0c8
MD
272 stack = NULL;
273 }
274 }
275 if (stack == NULL) {
05220613 276#ifdef _KERNEL
f470d0c8 277 stack = (void *)kmem_alloc(kernel_map, stksize);
05220613 278#else
f470d0c8 279 stack = libcaps_alloc_stack(stksize);
05220613 280#endif
ef0fdad1 281 flags |= TDF_ALLOCATED_STACK;
99df837e 282 }
75cdbe6c 283 if (cpu < 0)
f470d0c8 284 lwkt_init_thread(td, stack, stksize, flags, mycpu);
75cdbe6c 285 else
f470d0c8 286 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 287 return(td);
7d0bac62
MD
288}
289
709799ea
MD
290#ifdef _KERNEL
291
7d0bac62
MD
292/*
293 * Initialize a preexisting thread structure. This function is used by
294 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
295 *
f8c3996b
MD
296 * All threads start out in a critical section at a priority of
297 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
298 * appropriate. This function may send an IPI message when the
299 * requested cpu is not the current cpu and consequently gd_tdallq may
300 * not be initialized synchronously from the point of view of the originating
301 * cpu.
302 *
303 * NOTE! we have to be careful in regards to creating threads for other cpus
304 * if SMP has not yet been activated.
7d0bac62 305 */
41a01a4d
MD
306#ifdef SMP
307
75cdbe6c
MD
308static void
309lwkt_init_thread_remote(void *arg)
310{
311 thread_t td = arg;
312
313 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
314}
315
41a01a4d
MD
316#endif
317
7d0bac62 318void
f470d0c8
MD
319lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
320 struct globaldata *gd)
7d0bac62 321{
37af14fe
MD
322 globaldata_t mygd = mycpu;
323
99df837e
MD
324 bzero(td, sizeof(struct thread));
325 td->td_kstack = stack;
f470d0c8 326 td->td_kstack_size = stksize;
99df837e 327 td->td_flags |= flags;
26a0694b 328 td->td_gd = gd;
f8c3996b 329 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
c95cd171 330 lwkt_initport(&td->td_msgport, td);
99df837e 331 pmap_init_thread(td);
0f7a3396 332#ifdef SMP
5d21b981
MD
333 /*
334 * Normally initializing a thread for a remote cpu requires sending an
335 * IPI. However, the idlethread is setup before the other cpus are
336 * activated so we have to treat it as a special case. XXX manipulation
337 * of gd_tdallq requires the BGL.
338 */
339 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 340 crit_enter_gd(mygd);
75cdbe6c 341 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 342 crit_exit_gd(mygd);
75cdbe6c 343 } else {
2db3b277 344 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 345 }
0f7a3396 346#else
37af14fe 347 crit_enter_gd(mygd);
0f7a3396 348 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 349 crit_exit_gd(mygd);
0f7a3396 350#endif
73e4f7b9
MD
351}
352
2d93b37a
MD
353#endif /* _KERNEL */
354
73e4f7b9
MD
355void
356lwkt_set_comm(thread_t td, const char *ctl, ...)
357{
e2565a42 358 __va_list va;
73e4f7b9 359
e2565a42 360 __va_start(va, ctl);
73e4f7b9 361 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 362 __va_end(va);
7d0bac62
MD
363}
364
99df837e 365void
73e4f7b9 366lwkt_hold(thread_t td)
99df837e 367{
73e4f7b9
MD
368 ++td->td_refs;
369}
370
371void
372lwkt_rele(thread_t td)
373{
374 KKASSERT(td->td_refs > 0);
375 --td->td_refs;
376}
377
c95cd171
MD
378#ifdef _KERNEL
379
73e4f7b9
MD
380void
381lwkt_wait_free(thread_t td)
382{
383 while (td->td_refs)
377d4740 384 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
385}
386
c95cd171
MD
387#endif
388
73e4f7b9
MD
389void
390lwkt_free_thread(thread_t td)
391{
392 struct globaldata *gd = mycpu;
393
d9eea1a5 394 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
395 ("lwkt_free_thread: did not exit! %p", td));
396
37af14fe 397 crit_enter_gd(gd);
73e4f7b9
MD
398 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
399 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
400 (td->td_flags & TDF_ALLOCATED_THREAD)
401 ) {
73e4f7b9
MD
402 ++gd->gd_tdfreecount;
403 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
37af14fe 404 crit_exit_gd(gd);
99df837e 405 } else {
37af14fe 406 crit_exit_gd(gd);
99df837e 407 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613 408#ifdef _KERNEL
f470d0c8 409 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
05220613 410#else
f470d0c8 411 libcaps_free_stack(td->td_kstack, td->td_kstack_size);
05220613 412#endif
73e4f7b9 413 /* gd invalid */
99df837e 414 td->td_kstack = NULL;
f470d0c8 415 td->td_kstack_size = 0;
99df837e 416 }
05220613
MD
417 if (td->td_flags & TDF_ALLOCATED_THREAD) {
418#ifdef _KERNEL
99df837e 419 zfree(thread_zone, td);
05220613
MD
420#else
421 free(td);
422#endif
423 }
99df837e
MD
424 }
425}
426
427
8ad65e08
MD
428/*
429 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
430 * switch to the idlethread. Switching must occur within a critical
431 * section to avoid races with the scheduling queue.
432 *
433 * We always have full control over our cpu's run queue. Other cpus
434 * that wish to manipulate our queue must use the cpu_*msg() calls to
435 * talk to our cpu, so a critical section is all that is needed and
436 * the result is very, very fast thread switching.
437 *
96728c05
MD
438 * The LWKT scheduler uses a fixed priority model and round-robins at
439 * each priority level. User process scheduling is a totally
440 * different beast and LWKT priorities should not be confused with
441 * user process priorities.
f1d1c3fa 442 *
96728c05
MD
443 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
444 * cleans it up. Note that the td_switch() function cannot do anything that
445 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
446 * the target thread (not the current thread). It's nice to have a scheduler
447 * that does not need the MP lock to work because it allows us to do some
448 * really cool high-performance MP lock optimizations.
8ad65e08 449 */
96728c05 450
8ad65e08
MD
451void
452lwkt_switch(void)
453{
37af14fe
MD
454 globaldata_t gd = mycpu;
455 thread_t td = gd->gd_curthread;
8ad65e08 456 thread_t ntd;
8a8d5d85
MD
457#ifdef SMP
458 int mpheld;
459#endif
8ad65e08 460
46a3f46d 461 /*
27e88a6e
MD
462 * Switching from within a 'fast' (non thread switched) interrupt or IPI
463 * is illegal. However, we may have to do it anyway if we hit a fatal
464 * kernel trap or we have paniced.
465 *
466 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 467 */
27e88a6e
MD
468 if (gd->gd_intr_nesting_level) {
469 int savegdnest;
470 int savegdtrap;
471
472 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
473 panic("lwkt_switch: cannot switch from within "
474 "a fast interrupt, yet, td %p\n", td);
475 } else {
476 savegdnest = gd->gd_intr_nesting_level;
477 savegdtrap = gd->gd_trap_nesting_level;
478 gd->gd_intr_nesting_level = 0;
479 gd->gd_trap_nesting_level = 0;
a7422615
MD
480 if ((td->td_flags & TDF_PANICWARN) == 0) {
481 td->td_flags |= TDF_PANICWARN;
482 printf("Warning: thread switch from interrupt or IPI, "
483 "thread %p (%s)\n", td, td->td_comm);
484#ifdef DDB
485 db_print_backtrace();
486#endif
487 }
27e88a6e
MD
488 lwkt_switch();
489 gd->gd_intr_nesting_level = savegdnest;
490 gd->gd_trap_nesting_level = savegdtrap;
491 return;
492 }
96728c05 493 }
ef0fdad1 494
cb973d15
MD
495 /*
496 * Passive release (used to transition from user to kernel mode
497 * when we block or switch rather then when we enter the kernel).
498 * This function is NOT called if we are switching into a preemption
499 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
500 * our current process designation (if we have one) and become a true
501 * LWKT thread, and may also hand the current process designation to
502 * another process and schedule thread.
cb973d15
MD
503 */
504 if (td->td_release)
505 td->td_release(td);
506
37af14fe 507 crit_enter_gd(gd);
8a8d5d85
MD
508
509#ifdef SMP
510 /*
511 * td_mpcount cannot be used to determine if we currently hold the
512 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
513 * to get the lock, and switch out if it can't. Our ownership of
514 * the actual lock will remain stable while we are in a critical section
515 * (but, of course, another cpu may own or release the lock so the
516 * actual value of mp_lock is not stable).
8a8d5d85
MD
517 */
518 mpheld = MP_LOCK_HELD();
0f7a3396
MD
519#ifdef INVARIANTS
520 if (td->td_cscount) {
521 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
522 td);
523 if (panic_on_cscount)
524 panic("switching while mastering cpusync");
525 }
526#endif
8a8d5d85 527#endif
99df837e
MD
528 if ((ntd = td->td_preempted) != NULL) {
529 /*
530 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
531 * thread. This occurs transparently, whether the preempted thread
532 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
533 * itself).
534 *
535 * We have to setup the MP lock for the original thread after backing
536 * out the adjustment that was made to curthread when the original
537 * was preempted.
99df837e 538 */
26a0694b 539 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 540#ifdef SMP
96728c05 541 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 542 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
543 td, ntd, td->td_mpcount, ntd->td_mpcount);
544 }
8a8d5d85
MD
545 if (ntd->td_mpcount) {
546 td->td_mpcount -= ntd->td_mpcount;
547 KKASSERT(td->td_mpcount >= 0);
548 }
549#endif
26a0694b 550 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
551
552 /*
553 * XXX. The interrupt may have woken a thread up, we need to properly
554 * set the reschedule flag if the originally interrupted thread is at
555 * a lower priority.
556 */
557 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
558 need_lwkt_resched();
8a8d5d85 559 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 560 } else {
4b5f931b
MD
561 /*
562 * Priority queue / round-robin at each priority. Note that user
563 * processes run at a fixed, low priority and the user process
564 * scheduler deals with interactions between user processes
565 * by scheduling and descheduling them from the LWKT queue as
566 * necessary.
8a8d5d85
MD
567 *
568 * We have to adjust the MP lock for the target thread. If we
569 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
570 * thread that does not need the MP lock. If we cannot, we spin
571 * instead of HLT.
572 *
573 * A similar issue exists for the tokens held by the target thread.
574 * If we cannot obtain ownership of the tokens we cannot immediately
575 * schedule the thread.
576 */
577
578 /*
579 * We are switching threads. If there are any pending requests for
580 * tokens we can satisfy all of them here.
4b5f931b 581 */
41a01a4d
MD
582#ifdef SMP
583 if (gd->gd_tokreqbase)
584 lwkt_drain_token_requests();
585#endif
586
8ec60c3f
MD
587 /*
588 * If an LWKT reschedule was requested, well that is what we are
589 * doing now so clear it.
590 */
591 clear_lwkt_resched();
4b5f931b
MD
592again:
593 if (gd->gd_runqmask) {
594 int nq = bsrl(gd->gd_runqmask);
595 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
596 gd->gd_runqmask &= ~(1 << nq);
597 goto again;
598 }
8a8d5d85 599#ifdef SMP
41a01a4d
MD
600 /*
601 * If the target needs the MP lock and we couldn't get it,
602 * or if the target is holding tokens and we could not
603 * gain ownership of the tokens, continue looking for a
604 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
605 *
606 * NOTE: the mpheld variable invalid after this conditional, it
607 * can change due to both cpu_try_mplock() returning success
608 * AND interactions in lwkt_chktokens() due to the fact that
609 * we are trying to check the mpcount of a thread other then
610 * the current thread. Because of this, if the current thread
611 * is not holding td_mpcount, an IPI indirectly run via
612 * lwkt_chktokens() can obtain and release the MP lock and
613 * cause the core MP lock to be released.
41a01a4d
MD
614 */
615 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
616 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
617 ) {
8a8d5d85 618 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
619
620 mpheld = MP_LOCK_HELD();
621 ntd = NULL;
8a8d5d85
MD
622 while (rqmask) {
623 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 624 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 625 /* spinning due to MP lock being held */
38717797 626#ifdef INVARIANTS
a453459d 627 ++mplock_contention_count;
38717797 628#endif
a453459d 629 /* mplock still not held, 'mpheld' still valid */
41a01a4d 630 continue;
38717797 631 }
a453459d
MD
632
633 /*
634 * mpheld state invalid after chktokens call returns
635 * failure, but the variable is only needed for
636 * the loop.
637 */
38717797 638 if (ntd->td_toks && !lwkt_chktokens(ntd)) {
a453459d 639 /* spinning due to token contention */
38717797 640#ifdef INVARIANTS
a453459d 641 ++token_contention_count;
38717797 642#endif
a453459d 643 mpheld = MP_LOCK_HELD();
41a01a4d 644 continue;
38717797 645 }
41a01a4d 646 break;
8a8d5d85
MD
647 }
648 if (ntd)
649 break;
650 rqmask &= ~(1 << nq);
651 nq = bsrl(rqmask);
652 }
653 if (ntd == NULL) {
a2a5ad0d
MD
654 ntd = &gd->gd_idlethread;
655 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 656 KASSERT(ntd->td_mpcount == 0, ("Idlex thread %p was holding the BGL!", ntd));
8a8d5d85
MD
657 } else {
658 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
659 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
660 }
661 } else {
662 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
663 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
664 }
665#else
4b5f931b
MD
666 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
667 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 668#endif
4b5f931b 669 } else {
3c23a41a 670 /*
60f945af
MD
671 * We have nothing to run but only let the idle loop halt
672 * the cpu if there are no pending interrupts.
3c23a41a 673 */
a2a5ad0d 674 ntd = &gd->gd_idlethread;
60f945af 675 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 676 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d
MD
677#ifdef SMP
678 KASSERT(ntd->td_mpcount == 0, ("Idley thread %p was holding the BGL!", ntd));
679#endif
4b5f931b 680 }
f1d1c3fa 681 }
26a0694b
MD
682 KASSERT(ntd->td_pri >= TDPRI_CRIT,
683 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
684
685 /*
686 * Do the actual switch. If the new target does not need the MP lock
687 * and we are holding it, release the MP lock. If the new target requires
688 * the MP lock we have already acquired it for the target.
689 */
690#ifdef SMP
691 if (ntd->td_mpcount == 0 ) {
692 if (MP_LOCK_HELD())
693 cpu_rel_mplock();
694 } else {
a453459d 695 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
696 }
697#endif
94f6d86e
MD
698 if (td != ntd) {
699 ++switch_count;
f1d1c3fa 700 td->td_switch(ntd);
94f6d86e 701 }
37af14fe
MD
702 /* NOTE: current cpu may have changed after switch */
703 crit_exit_quick(td);
8ad65e08
MD
704}
705
b68b7282 706/*
96728c05
MD
707 * Request that the target thread preempt the current thread. Preemption
708 * only works under a specific set of conditions:
b68b7282 709 *
96728c05
MD
710 * - We are not preempting ourselves
711 * - The target thread is owned by the current cpu
712 * - We are not currently being preempted
713 * - The target is not currently being preempted
714 * - We are able to satisfy the target's MP lock requirements (if any).
715 *
716 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
717 * this is called via lwkt_schedule() through the td_preemptable callback.
718 * critpri is the managed critical priority that we should ignore in order
719 * to determine whether preemption is possible (aka usually just the crit
720 * priority of lwkt_schedule() itself).
b68b7282 721 *
26a0694b
MD
722 * XXX at the moment we run the target thread in a critical section during
723 * the preemption in order to prevent the target from taking interrupts
724 * that *WE* can't. Preemption is strictly limited to interrupt threads
725 * and interrupt-like threads, outside of a critical section, and the
726 * preempted source thread will be resumed the instant the target blocks
727 * whether or not the source is scheduled (i.e. preemption is supposed to
728 * be as transparent as possible).
4b5f931b 729 *
8a8d5d85
MD
730 * The target thread inherits our MP count (added to its own) for the
731 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
732 * MP lock during the preemption. Therefore, any preempting targets must be
733 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
734 * out of sync with the physical mp_lock, but we do not have to preserve
735 * the original ownership of the lock if it was out of synch (that is, we
736 * can leave it synchronized on return).
b68b7282
MD
737 */
738void
96728c05 739lwkt_preempt(thread_t ntd, int critpri)
b68b7282 740{
46a3f46d 741 struct globaldata *gd = mycpu;
0a3f9b47 742 thread_t td;
8a8d5d85
MD
743#ifdef SMP
744 int mpheld;
57c254db 745 int savecnt;
8a8d5d85 746#endif
b68b7282 747
26a0694b 748 /*
96728c05
MD
749 * The caller has put us in a critical section. We can only preempt
750 * if the caller of the caller was not in a critical section (basically
0a3f9b47 751 * a local interrupt), as determined by the 'critpri' parameter.
96728c05
MD
752 *
753 * YYY The target thread must be in a critical section (else it must
754 * inherit our critical section? I dunno yet).
41a01a4d
MD
755 *
756 * Any tokens held by the target may not be held by thread(s) being
757 * preempted. We take the easy way out and do not preempt if
758 * the target is holding tokens.
0a3f9b47
MD
759 *
760 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
761 */
762 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 763
0a3f9b47 764 td = gd->gd_curthread;
0a3f9b47 765 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
766 ++preempt_miss;
767 return;
768 }
96728c05
MD
769 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
770 ++preempt_miss;
8ec60c3f 771 need_lwkt_resched();
96728c05
MD
772 return;
773 }
774#ifdef SMP
46a3f46d 775 if (ntd->td_gd != gd) {
96728c05 776 ++preempt_miss;
8ec60c3f 777 need_lwkt_resched();
96728c05
MD
778 return;
779 }
780#endif
41a01a4d
MD
781 /*
782 * Take the easy way out and do not preempt if the target is holding
783 * one or more tokens. We could test whether the thread(s) being
784 * preempted interlock against the target thread's tokens and whether
785 * we can get all the target thread's tokens, but this situation
786 * should not occur very often so its easier to simply not preempt.
787 */
788 if (ntd->td_toks != NULL) {
789 ++preempt_miss;
8ec60c3f 790 need_lwkt_resched();
41a01a4d
MD
791 return;
792 }
26a0694b
MD
793 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
794 ++preempt_weird;
8ec60c3f 795 need_lwkt_resched();
26a0694b
MD
796 return;
797 }
798 if (ntd->td_preempted) {
4b5f931b 799 ++preempt_hit;
8ec60c3f 800 need_lwkt_resched();
26a0694b 801 return;
b68b7282 802 }
8a8d5d85 803#ifdef SMP
a2a5ad0d
MD
804 /*
805 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
806 * to or from the MP lock. In this case td_mpcount will be pre-disposed
807 * (non-zero) but not actually synchronized with the actual state of the
808 * lock. We can use it to imply an MP lock requirement for the
809 * preemption but we cannot use it to test whether we hold the MP lock
810 * or not.
a2a5ad0d 811 */
96728c05 812 savecnt = td->td_mpcount;
71ef2f5c 813 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
814 ntd->td_mpcount += td->td_mpcount;
815 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
816 ntd->td_mpcount -= td->td_mpcount;
817 ++preempt_miss;
8ec60c3f 818 need_lwkt_resched();
8a8d5d85
MD
819 return;
820 }
821#endif
26a0694b 822
8ec60c3f
MD
823 /*
824 * Since we are able to preempt the current thread, there is no need to
825 * call need_lwkt_resched().
826 */
26a0694b
MD
827 ++preempt_hit;
828 ntd->td_preempted = td;
829 td->td_flags |= TDF_PREEMPT_LOCK;
830 td->td_switch(ntd);
831 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
832#ifdef SMP
833 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
834 mpheld = MP_LOCK_HELD();
835 if (mpheld && td->td_mpcount == 0)
96728c05 836 cpu_rel_mplock();
71ef2f5c 837 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
838 panic("lwkt_preempt(): MP lock was not held through");
839#endif
26a0694b
MD
840 ntd->td_preempted = NULL;
841 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
842}
843
f1d1c3fa
MD
844/*
845 * Yield our thread while higher priority threads are pending. This is
846 * typically called when we leave a critical section but it can be safely
847 * called while we are in a critical section.
848 *
849 * This function will not generally yield to equal priority threads but it
850 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 851 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
852 * lwkt_yield_quick().
853 *
235957ed 854 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
855 * came along but was blocked and made pending.
856 *
f1d1c3fa
MD
857 * (self contained on a per cpu basis)
858 */
859void
860lwkt_yield_quick(void)
861{
7966cb69
MD
862 globaldata_t gd = mycpu;
863 thread_t td = gd->gd_curthread;
ef0fdad1 864
a2a5ad0d 865 /*
235957ed 866 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
867 * it with a non-zero cpl then we might not wind up calling splz after
868 * a task switch when the critical section is exited even though the
46a3f46d 869 * new task could accept the interrupt.
a2a5ad0d
MD
870 *
871 * XXX from crit_exit() only called after last crit section is released.
872 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
873 *
874 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
875 * except for this special case, we MUST call splz() here to handle any
876 * pending ints, particularly after we switch, or we might accidently
877 * halt the cpu with interrupts pending.
a2a5ad0d 878 */
46a3f46d 879 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 880 splz();
f1d1c3fa
MD
881
882 /*
883 * YYY enabling will cause wakeup() to task-switch, which really
884 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
885 * preemption and MP without actually doing preemption or MP, because a
886 * lot of code assumes that wakeup() does not block.
f1d1c3fa 887 */
46a3f46d
MD
888 if (untimely_switch && td->td_nest_count == 0 &&
889 gd->gd_intr_nesting_level == 0
890 ) {
37af14fe 891 crit_enter_quick(td);
f1d1c3fa
MD
892 /*
893 * YYY temporary hacks until we disassociate the userland scheduler
894 * from the LWKT scheduler.
895 */
896 if (td->td_flags & TDF_RUNQ) {
897 lwkt_switch(); /* will not reenter yield function */
898 } else {
37af14fe 899 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 900 lwkt_switch(); /* will not reenter yield function */
37af14fe 901 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 902 }
7966cb69 903 crit_exit_noyield(td);
f1d1c3fa 904 }
f1d1c3fa
MD
905}
906
8ad65e08 907/*
f1d1c3fa 908 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 909 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
910 * the crit_exit() call in lwkt_switch().
911 *
912 * (self contained on a per cpu basis)
8ad65e08
MD
913 */
914void
f1d1c3fa 915lwkt_yield(void)
8ad65e08 916{
37af14fe 917 lwkt_schedule_self(curthread);
f1d1c3fa
MD
918 lwkt_switch();
919}
920
8ad65e08 921/*
f1d1c3fa
MD
922 * Generic schedule. Possibly schedule threads belonging to other cpus and
923 * deal with threads that might be blocked on a wait queue.
924 *
0a3f9b47
MD
925 * We have a little helper inline function which does additional work after
926 * the thread has been enqueued, including dealing with preemption and
927 * setting need_lwkt_resched() (which prevents the kernel from returning
928 * to userland until it has processed higher priority threads).
8ad65e08 929 */
0a3f9b47
MD
930static __inline
931void
8ec60c3f 932_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
0a3f9b47
MD
933{
934 if (ntd->td_preemptable) {
935 ntd->td_preemptable(ntd, cpri); /* YYY +token */
8ec60c3f
MD
936 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
937 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
938 ) {
939 need_lwkt_resched();
0a3f9b47
MD
940 }
941}
942
8ad65e08
MD
943void
944lwkt_schedule(thread_t td)
945{
37af14fe
MD
946 globaldata_t mygd = mycpu;
947
96728c05 948#ifdef INVARIANTS
41a01a4d 949 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
26a0694b
MD
950 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
951 && td->td_proc->p_stat == SSLEEP
952 ) {
953 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
954 curthread,
955 curthread->td_proc ? curthread->td_proc->p_pid : -1,
956 curthread->td_proc ? curthread->td_proc->p_stat : -1,
957 td,
958 td->td_proc ? curthread->td_proc->p_pid : -1,
959 td->td_proc ? curthread->td_proc->p_stat : -1
960 );
961 panic("SCHED PANIC");
962 }
96728c05 963#endif
37af14fe
MD
964 crit_enter_gd(mygd);
965 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
966 _lwkt_enqueue(td);
967 } else {
968 lwkt_wait_t w;
969
970 /*
971 * If the thread is on a wait list we have to send our scheduling
972 * request to the owner of the wait structure. Otherwise we send
973 * the scheduling request to the cpu owning the thread. Races
974 * are ok, the target will forward the message as necessary (the
975 * message may chase the thread around before it finally gets
976 * acted upon).
977 *
978 * (remember, wait structures use stable storage)
0a3f9b47 979 *
0c453950
MD
980 * NOTE: we have to account for the number of critical sections
981 * under our control when calling _lwkt_schedule_post() so it
982 * can figure out whether preemption is allowed.
983 *
984 * NOTE: The wait structure algorithms are a mess and need to be
985 * rewritten.
986 *
987 * NOTE: We cannot safely acquire or release a token, even
988 * non-blocking, because this routine may be called in the context
989 * of a thread already holding the token and thus not provide any
990 * interlock protection. We cannot safely manipulate the td_toks
991 * list for the same reason. Instead we depend on our critical
992 * section if the token is owned by our cpu.
f1d1c3fa
MD
993 */
994 if ((w = td->td_wait) != NULL) {
0c453950 995 if (w->wa_token.t_cpu == mygd) {
f1d1c3fa
MD
996 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
997 --w->wa_count;
998 td->td_wait = NULL;
0f7a3396 999#ifdef SMP
8ec60c3f 1000 if (td->td_gd == mygd) {
f1d1c3fa 1001 _lwkt_enqueue(td);
8ec60c3f 1002 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 1003 } else {
2db3b277 1004 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 1005 }
0f7a3396
MD
1006#else
1007 _lwkt_enqueue(td);
8ec60c3f 1008 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 1009#endif
f1d1c3fa 1010 } else {
96728c05 1011 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
1012 }
1013 } else {
1014 /*
1015 * If the wait structure is NULL and we own the thread, there
1016 * is no race (since we are in a critical section). If we
1017 * do not own the thread there might be a race but the
1018 * target cpu will deal with it.
1019 */
0f7a3396 1020#ifdef SMP
37af14fe 1021 if (td->td_gd == mygd) {
f1d1c3fa 1022 _lwkt_enqueue(td);
8ec60c3f 1023 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 1024 } else {
2db3b277 1025 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 1026 }
0f7a3396
MD
1027#else
1028 _lwkt_enqueue(td);
8ec60c3f 1029 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 1030#endif
f1d1c3fa 1031 }
8ad65e08 1032 }
37af14fe 1033 crit_exit_gd(mygd);
8ad65e08
MD
1034}
1035
d9eea1a5
MD
1036/*
1037 * Managed acquisition. This code assumes that the MP lock is held for
1038 * the tdallq operation and that the thread has been descheduled from its
1039 * original cpu. We also have to wait for the thread to be entirely switched
1040 * out on its original cpu (this is usually fast enough that we never loop)
1041 * since the LWKT system does not have to hold the MP lock while switching
1042 * and the target may have released it before switching.
1043 */
a2a5ad0d
MD
1044void
1045lwkt_acquire(thread_t td)
1046{
37af14fe
MD
1047 globaldata_t gd;
1048 globaldata_t mygd;
a2a5ad0d
MD
1049
1050 gd = td->td_gd;
37af14fe 1051 mygd = mycpu;
35238fa5 1052 cpu_lfence();
a2a5ad0d 1053 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5 1054 while (td->td_flags & TDF_RUNNING) /* XXX spin */
35238fa5 1055 cpu_lfence();
37af14fe
MD
1056 if (gd != mygd) {
1057 crit_enter_gd(mygd);
a2a5ad0d 1058 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
37af14fe
MD
1059 td->td_gd = mygd;
1060 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
1061 crit_exit_gd(mygd);
a2a5ad0d
MD
1062 }
1063}
1064
f1d1c3fa
MD
1065/*
1066 * Generic deschedule. Descheduling threads other then your own should be
1067 * done only in carefully controlled circumstances. Descheduling is
1068 * asynchronous.
1069 *
1070 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1071 */
1072void
1073lwkt_deschedule(thread_t td)
1074{
f1d1c3fa
MD
1075 crit_enter();
1076 if (td == curthread) {
1077 _lwkt_dequeue(td);
1078 } else {
a72187e9 1079 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1080 _lwkt_dequeue(td);
1081 } else {
2db3b277 1082 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
1083 }
1084 }
1085 crit_exit();
1086}
1087
4b5f931b
MD
1088/*
1089 * Set the target thread's priority. This routine does not automatically
1090 * switch to a higher priority thread, LWKT threads are not designed for
1091 * continuous priority changes. Yield if you want to switch.
1092 *
1093 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1094 * of the td_pri field. The specified priority may also indicate zero or
1095 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1096 *
1097 * Note that we requeue the thread whether it winds up on a different runq
1098 * or not. uio_yield() depends on this and the routine is not normally
1099 * called with the same priority otherwise.
4b5f931b
MD
1100 */
1101void
1102lwkt_setpri(thread_t td, int pri)
1103{
26a0694b 1104 KKASSERT(pri >= 0);
a72187e9 1105 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1106 crit_enter();
1107 if (td->td_flags & TDF_RUNQ) {
1108 _lwkt_dequeue(td);
1109 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1110 _lwkt_enqueue(td);
1111 } else {
1112 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1113 }
1114 crit_exit();
1115}
1116
1117void
1118lwkt_setpri_self(int pri)
1119{
1120 thread_t td = curthread;
1121
4b5f931b
MD
1122 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1123 crit_enter();
1124 if (td->td_flags & TDF_RUNQ) {
1125 _lwkt_dequeue(td);
1126 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1127 _lwkt_enqueue(td);
1128 } else {
1129 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1130 }
1131 crit_exit();
1132}
1133
8ec60c3f
MD
1134/*
1135 * Determine if there is a runnable thread at a higher priority then
1136 * the current thread. lwkt_setpri() does not check this automatically.
1137 * Return 1 if there is, 0 if there isn't.
1138 *
1139 * Example: if bit 31 of runqmask is set and the current thread is priority
1140 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1141 *
1142 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1143 * up comparing against 0xffffffff, a comparison that will always be false.
1144 */
1145int
1146lwkt_checkpri_self(void)
1147{
1148 globaldata_t gd = mycpu;
1149 thread_t td = gd->gd_curthread;
1150 int nq = td->td_pri & TDPRI_MASK;
1151
1152 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1153 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1154 return(1);
1155 ++nq;
1156 }
1157 return(0);
1158}
1159
5d21b981
MD
1160/*
1161 * Migrate the current thread to the specified cpu. The BGL must be held
1162 * (for the gd_tdallq manipulation XXX). This is accomplished by
1163 * descheduling ourselves from the current cpu, moving our thread to the
1164 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1165 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1166 */
3d28ff59 1167#ifdef SMP
5d21b981 1168static void lwkt_setcpu_remote(void *arg);
3d28ff59 1169#endif
5d21b981
MD
1170
1171void
1172lwkt_setcpu_self(globaldata_t rgd)
1173{
1174#ifdef SMP
1175 thread_t td = curthread;
1176
1177 if (td->td_gd != rgd) {
1178 crit_enter_quick(td);
1179 td->td_flags |= TDF_MIGRATING;
1180 lwkt_deschedule_self(td);
1181 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1182 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
1183 lwkt_send_ipiq(rgd, (ipifunc_t)lwkt_setcpu_remote, td);
1184 lwkt_switch();
1185 /* we are now on the target cpu */
1186 crit_exit_quick(td);
1187 }
1188#endif
1189}
1190
1191/*
1192 * Remote IPI for cpu migration (called while in a critical section so we
1193 * do not have to enter another one). The thread has already been moved to
1194 * our cpu's allq, but we must wait for the thread to be completely switched
1195 * out on the originating cpu before we schedule it on ours or the stack
1196 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1197 * change to main memory.
1198 *
1199 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1200 * against wakeups. It is best if this interface is used only when there
1201 * are no pending events that might try to schedule the thread.
1202 */
3d28ff59 1203#ifdef SMP
5d21b981
MD
1204static void
1205lwkt_setcpu_remote(void *arg)
1206{
1207 thread_t td = arg;
1208 globaldata_t gd = mycpu;
1209
1210 while (td->td_flags & TDF_RUNNING)
35238fa5 1211 cpu_lfence();
5d21b981 1212 td->td_gd = gd;
35238fa5 1213 cpu_sfence();
5d21b981
MD
1214 td->td_flags &= ~TDF_MIGRATING;
1215 _lwkt_enqueue(td);
1216}
3d28ff59 1217#endif
5d21b981 1218
4b5f931b
MD
1219struct proc *
1220lwkt_preempted_proc(void)
1221{
73e4f7b9 1222 thread_t td = curthread;
4b5f931b
MD
1223 while (td->td_preempted)
1224 td = td->td_preempted;
1225 return(td->td_proc);
1226}
1227
f1d1c3fa 1228/*
41a01a4d
MD
1229 * Block on the specified wait queue until signaled. A generation number
1230 * must be supplied to interlock the wait queue. The function will
1231 * return immediately if the generation number does not match the wait
1232 * structure's generation number.
f1d1c3fa
MD
1233 */
1234void
ae8050a4 1235lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
1236{
1237 thread_t td = curthread;
41a01a4d 1238 lwkt_tokref ilock;
f1d1c3fa 1239
41a01a4d
MD
1240 lwkt_gettoken(&ilock, &w->wa_token);
1241 crit_enter();
ae8050a4 1242 if (w->wa_gen == *gen) {
f1d1c3fa
MD
1243 _lwkt_dequeue(td);
1244 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1245 ++w->wa_count;
1246 td->td_wait = w;
ae8050a4 1247 td->td_wmesg = wmesg;
41a01a4d 1248 again:
f1d1c3fa 1249 lwkt_switch();
ece04fd0
MD
1250 if (td->td_wmesg != NULL) {
1251 _lwkt_dequeue(td);
1252 goto again;
1253 }
8ad65e08 1254 }
41a01a4d 1255 crit_exit();
ae8050a4 1256 *gen = w->wa_gen;
41a01a4d 1257 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1258}
1259
1260/*
1261 * Signal a wait queue. We gain ownership of the wait queue in order to
1262 * signal it. Once a thread is removed from the wait queue we have to
1263 * deal with the cpu owning the thread.
1264 *
1265 * Note: alternatively we could message the target cpu owning the wait
1266 * queue. YYY implement as sysctl.
1267 */
1268void
ece04fd0 1269lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1270{
1271 thread_t td;
41a01a4d 1272 lwkt_tokref ilock;
f1d1c3fa 1273
41a01a4d 1274 lwkt_gettoken(&ilock, &w->wa_token);
f1d1c3fa 1275 ++w->wa_gen;
41a01a4d 1276 crit_enter();
ece04fd0
MD
1277 if (count < 0)
1278 count = w->wa_count;
f1d1c3fa
MD
1279 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1280 --count;
1281 --w->wa_count;
1282 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1283 td->td_wait = NULL;
ae8050a4 1284 td->td_wmesg = NULL;
a72187e9 1285 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1286 _lwkt_enqueue(td);
1287 } else {
2db3b277 1288 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 1289 }
f1d1c3fa 1290 }
41a01a4d
MD
1291 crit_exit();
1292 lwkt_reltoken(&ilock);
f1d1c3fa
MD
1293}
1294
99df837e
MD
1295/*
1296 * Create a kernel process/thread/whatever. It shares it's address space
1297 * with proc0 - ie: kernel only.
1298 *
365fa13f
MD
1299 * NOTE! By default new threads are created with the MP lock held. A
1300 * thread which does not require the MP lock should release it by calling
1301 * rel_mplock() at the start of the new thread.
99df837e
MD
1302 */
1303int
1304lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1305 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1306 const char *fmt, ...)
99df837e 1307{
73e4f7b9 1308 thread_t td;
e2565a42 1309 __va_list ap;
99df837e 1310
f470d0c8 1311 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu);
a2a5ad0d
MD
1312 if (tdp)
1313 *tdp = td;
709799ea 1314 cpu_set_thread_handler(td, lwkt_exit, func, arg);
ef0fdad1 1315 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1316#ifdef SMP
1317 td->td_mpcount = 1;
1318#endif
99df837e
MD
1319
1320 /*
1321 * Set up arg0 for 'ps' etc
1322 */
e2565a42 1323 __va_start(ap, fmt);
99df837e 1324 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1325 __va_end(ap);
99df837e
MD
1326
1327 /*
1328 * Schedule the thread to run
1329 */
ef0fdad1
MD
1330 if ((td->td_flags & TDF_STOPREQ) == 0)
1331 lwkt_schedule(td);
1332 else
1333 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1334 return 0;
1335}
1336
2d93b37a 1337/*
2d93b37a
MD
1338 * kthread_* is specific to the kernel and is not needed by userland.
1339 */
1340#ifdef _KERNEL
1341
99df837e
MD
1342/*
1343 * Destroy an LWKT thread. Warning! This function is not called when
1344 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1345 * uses a different reaping mechanism.
1346 */
1347void
1348lwkt_exit(void)
1349{
1350 thread_t td = curthread;
8826f33a 1351 globaldata_t gd;
99df837e
MD
1352
1353 if (td->td_flags & TDF_VERBOSE)
1354 printf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1355 caps_exit(td);
37af14fe
MD
1356 crit_enter_quick(td);
1357 lwkt_deschedule_self(td);
8826f33a
MD
1358 gd = mycpu;
1359 KKASSERT(gd == td->td_gd);
1360 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1361 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1362 ++gd->gd_tdfreecount;
1363 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1364 }
99df837e
MD
1365 cpu_thread_exit();
1366}
1367
2d93b37a
MD
1368#endif /* _KERNEL */
1369
1370void
1371crit_panic(void)
1372{
1373 thread_t td = curthread;
1374 int lpri = td->td_pri;
1375
1376 td->td_pri = 0;
1377 panic("td_pri is/would-go negative! %p %d", td, lpri);
1378}
1379
bd8015ca
MD
1380/*
1381 * Called from debugger/panic on cpus which have been stopped. We must still
1382 * process the IPIQ while stopped, even if we were stopped while in a critical
1383 * section (XXX).
1384 *
1385 * If we are dumping also try to process any pending interrupts. This may
1386 * or may not work depending on the state of the cpu at the point it was
1387 * stopped.
1388 */
1389void
1390lwkt_smp_stopped(void)
1391{
1392 globaldata_t gd = mycpu;
1393
1394 crit_enter_gd(gd);
1395 if (dumping) {
1396 lwkt_process_ipiq();
1397 splz();
1398 } else {
1399 lwkt_process_ipiq();
1400 }
1401 crit_exit_gd(gd);
1402}
1403