Remove conditional memory allocation based on KTR_ALL. Allocate memory
[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 *
52eedfb5 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.97 2006/05/29 03:57:20 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>
7d0bac62 52#include <sys/sysctl.h>
99df837e 53#include <sys/kthread.h>
f1d1c3fa 54#include <machine/cpu.h>
99df837e 55#include <sys/lock.h>
f6bf3af1 56#include <sys/caps.h>
9d265729
MD
57#include <sys/spinlock.h>
58
59#include <sys/thread2.h>
60#include <sys/spinlock2.h>
f1d1c3fa 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>
70#include <vm/vm_zone.h>
71
99df837e 72#include <machine/stdarg.h>
57c254db 73#include <machine/ipl.h>
96728c05 74#include <machine/smp.h>
99df837e 75
05220613
MD
76#else
77
78#include <sys/stdint.h>
fb04f4fd 79#include <libcaps/thread.h>
05220613
MD
80#include <sys/thread.h>
81#include <sys/msgport.h>
82#include <sys/errno.h>
fb04f4fd 83#include <libcaps/globaldata.h>
7e8303ad 84#include <machine/cpufunc.h>
05220613
MD
85#include <sys/thread2.h>
86#include <sys/msgport2.h>
709799ea 87#include <stdio.h>
05220613 88#include <stdlib.h>
709799ea 89#include <string.h>
709799ea 90#include <machine/lock.h>
b950d4b8
EN
91#include <machine/atomic.h>
92#include <machine/cpu.h>
05220613
MD
93
94#endif
95
7d0bac62 96static int untimely_switch = 0;
0f7a3396
MD
97#ifdef INVARIANTS
98static int panic_on_cscount = 0;
99#endif
05220613
MD
100static __int64_t switch_count = 0;
101static __int64_t preempt_hit = 0;
102static __int64_t preempt_miss = 0;
103static __int64_t preempt_weird = 0;
38717797
HP
104static __int64_t token_contention_count = 0;
105static __int64_t mplock_contention_count = 0;
05220613
MD
106
107#ifdef _KERNEL
108
109SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
0f7a3396
MD
110#ifdef INVARIANTS
111SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
112#endif
4b5f931b 113SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 114SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 115SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 116SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
117#ifdef INVARIANTS
118SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
119 &token_contention_count, 0, "spinning due to token contention");
120SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
121 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
122#endif
05220613
MD
123#endif
124
4b5f931b
MD
125/*
126 * These helper procedures handle the runq, they can only be called from
127 * within a critical section.
75cdbe6c
MD
128 *
129 * WARNING! Prior to SMP being brought up it is possible to enqueue and
130 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
131 * instead of 'mycpu' when referencing the globaldata structure. Once
132 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 133 */
f1d1c3fa
MD
134static __inline
135void
136_lwkt_dequeue(thread_t td)
137{
138 if (td->td_flags & TDF_RUNQ) {
4b5f931b 139 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 140 struct globaldata *gd = td->td_gd;
4b5f931b 141
f1d1c3fa 142 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
143 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
144 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
145 }
146}
147
148static __inline
149void
150_lwkt_enqueue(thread_t td)
151{
344ad853 152 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
4b5f931b 153 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 154 struct globaldata *gd = td->td_gd;
4b5f931b 155
f1d1c3fa 156 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
157 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
158 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
159 }
160}
8ad65e08 161
37af14fe
MD
162/*
163 * Schedule a thread to run. As the current thread we can always safely
164 * schedule ourselves, and a shortcut procedure is provided for that
165 * function.
166 *
167 * (non-blocking, self contained on a per cpu basis)
168 */
169void
170lwkt_schedule_self(thread_t td)
171{
172 crit_enter_quick(td);
173 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
174 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
344ad853 175 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
37af14fe 176 _lwkt_enqueue(td);
37af14fe
MD
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{
9d265729 222 spin_init(&w->wa_spinlock);
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
d3d32139 237lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 238{
99df837e 239 void *stack;
37af14fe 240 globaldata_t gd = mycpu;
7d0bac62 241
ef0fdad1 242 if (td == NULL) {
37af14fe
MD
243 crit_enter_gd(gd);
244 if (gd->gd_tdfreecount > 0) {
245 --gd->gd_tdfreecount;
246 td = TAILQ_FIRST(&gd->gd_tdfreeq);
d9eea1a5 247 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1 248 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
37af14fe
MD
249 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq);
250 crit_exit_gd(gd);
d3d32139 251 flags |= td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
ef0fdad1 252 } else {
37af14fe 253 crit_exit_gd(gd);
05220613 254#ifdef _KERNEL
ef0fdad1 255 td = zalloc(thread_zone);
05220613
MD
256#else
257 td = malloc(sizeof(struct thread));
258#endif
ef0fdad1 259 td->td_kstack = NULL;
f470d0c8 260 td->td_kstack_size = 0;
ef0fdad1
MD
261 flags |= TDF_ALLOCATED_THREAD;
262 }
263 }
f470d0c8
MD
264 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
265 if (flags & TDF_ALLOCATED_STACK) {
a9e5ae82 266#ifdef _KERNEL
f470d0c8 267 kmem_free(kernel_map, (vm_offset_t)stack, td->td_kstack_size);
a9e5ae82
JS
268#else
269 libcaps_free_stack(stack, td->td_kstack_size);
270#endif
f470d0c8
MD
271 stack = NULL;
272 }
273 }
274 if (stack == NULL) {
05220613 275#ifdef _KERNEL
f470d0c8 276 stack = (void *)kmem_alloc(kernel_map, stksize);
05220613 277#else
f470d0c8 278 stack = libcaps_alloc_stack(stksize);
05220613 279#endif
ef0fdad1 280 flags |= TDF_ALLOCATED_STACK;
99df837e 281 }
75cdbe6c 282 if (cpu < 0)
f470d0c8 283 lwkt_init_thread(td, stack, stksize, flags, mycpu);
75cdbe6c 284 else
f470d0c8 285 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 286 return(td);
7d0bac62
MD
287}
288
709799ea
MD
289#ifdef _KERNEL
290
7d0bac62
MD
291/*
292 * Initialize a preexisting thread structure. This function is used by
293 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
294 *
f8c3996b
MD
295 * All threads start out in a critical section at a priority of
296 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
297 * appropriate. This function may send an IPI message when the
298 * requested cpu is not the current cpu and consequently gd_tdallq may
299 * not be initialized synchronously from the point of view of the originating
300 * cpu.
301 *
302 * NOTE! we have to be careful in regards to creating threads for other cpus
303 * if SMP has not yet been activated.
7d0bac62 304 */
41a01a4d
MD
305#ifdef SMP
306
75cdbe6c
MD
307static void
308lwkt_init_thread_remote(void *arg)
309{
310 thread_t td = arg;
311
52eedfb5
MD
312 /*
313 * Protected by critical section held by IPI dispatch
314 */
75cdbe6c
MD
315 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
316}
317
41a01a4d
MD
318#endif
319
7d0bac62 320void
f470d0c8
MD
321lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
322 struct globaldata *gd)
7d0bac62 323{
37af14fe
MD
324 globaldata_t mygd = mycpu;
325
99df837e
MD
326 bzero(td, sizeof(struct thread));
327 td->td_kstack = stack;
f470d0c8 328 td->td_kstack_size = stksize;
d3d32139 329 td->td_flags = flags;
26a0694b 330 td->td_gd = gd;
f8c3996b 331 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
332#ifdef SMP
333 if ((flags & TDF_MPSAFE) == 0)
334 td->td_mpcount = 1;
335#endif
c95cd171 336 lwkt_initport(&td->td_msgport, td);
99df837e 337 pmap_init_thread(td);
0f7a3396 338#ifdef SMP
5d21b981
MD
339 /*
340 * Normally initializing a thread for a remote cpu requires sending an
341 * IPI. However, the idlethread is setup before the other cpus are
342 * activated so we have to treat it as a special case. XXX manipulation
343 * of gd_tdallq requires the BGL.
344 */
345 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 346 crit_enter_gd(mygd);
75cdbe6c 347 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 348 crit_exit_gd(mygd);
75cdbe6c 349 } else {
2db3b277 350 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 351 }
0f7a3396 352#else
37af14fe 353 crit_enter_gd(mygd);
0f7a3396 354 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 355 crit_exit_gd(mygd);
0f7a3396 356#endif
73e4f7b9
MD
357}
358
2d93b37a
MD
359#endif /* _KERNEL */
360
73e4f7b9
MD
361void
362lwkt_set_comm(thread_t td, const char *ctl, ...)
363{
e2565a42 364 __va_list va;
73e4f7b9 365
e2565a42 366 __va_start(va, ctl);
73e4f7b9 367 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 368 __va_end(va);
7d0bac62
MD
369}
370
99df837e 371void
73e4f7b9 372lwkt_hold(thread_t td)
99df837e 373{
73e4f7b9
MD
374 ++td->td_refs;
375}
376
377void
378lwkt_rele(thread_t td)
379{
380 KKASSERT(td->td_refs > 0);
381 --td->td_refs;
382}
383
c95cd171
MD
384#ifdef _KERNEL
385
73e4f7b9
MD
386void
387lwkt_wait_free(thread_t td)
388{
389 while (td->td_refs)
377d4740 390 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
391}
392
c95cd171
MD
393#endif
394
73e4f7b9
MD
395void
396lwkt_free_thread(thread_t td)
397{
398 struct globaldata *gd = mycpu;
399
d9eea1a5 400 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
401 ("lwkt_free_thread: did not exit! %p", td));
402
37af14fe 403 crit_enter_gd(gd);
1ec20b9d 404 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* Protected by BGL */
73e4f7b9 405 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
406 (td->td_flags & TDF_ALLOCATED_THREAD)
407 ) {
73e4f7b9
MD
408 ++gd->gd_tdfreecount;
409 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
37af14fe 410 crit_exit_gd(gd);
99df837e 411 } else {
37af14fe 412 crit_exit_gd(gd);
99df837e 413 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613 414#ifdef _KERNEL
f470d0c8 415 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
05220613 416#else
f470d0c8 417 libcaps_free_stack(td->td_kstack, td->td_kstack_size);
05220613 418#endif
73e4f7b9 419 /* gd invalid */
99df837e 420 td->td_kstack = NULL;
f470d0c8 421 td->td_kstack_size = 0;
99df837e 422 }
05220613
MD
423 if (td->td_flags & TDF_ALLOCATED_THREAD) {
424#ifdef _KERNEL
99df837e 425 zfree(thread_zone, td);
05220613
MD
426#else
427 free(td);
428#endif
429 }
99df837e
MD
430 }
431}
432
433
8ad65e08
MD
434/*
435 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
436 * switch to the idlethread. Switching must occur within a critical
437 * section to avoid races with the scheduling queue.
438 *
439 * We always have full control over our cpu's run queue. Other cpus
440 * that wish to manipulate our queue must use the cpu_*msg() calls to
441 * talk to our cpu, so a critical section is all that is needed and
442 * the result is very, very fast thread switching.
443 *
96728c05
MD
444 * The LWKT scheduler uses a fixed priority model and round-robins at
445 * each priority level. User process scheduling is a totally
446 * different beast and LWKT priorities should not be confused with
447 * user process priorities.
f1d1c3fa 448 *
96728c05
MD
449 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
450 * cleans it up. Note that the td_switch() function cannot do anything that
451 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
452 * the target thread (not the current thread). It's nice to have a scheduler
453 * that does not need the MP lock to work because it allows us to do some
454 * really cool high-performance MP lock optimizations.
69d78e99
MD
455 *
456 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
457 * is not called by the current thread in the preemption case, only when
458 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
459 */
460void
461lwkt_switch(void)
462{
37af14fe
MD
463 globaldata_t gd = mycpu;
464 thread_t td = gd->gd_curthread;
8ad65e08 465 thread_t ntd;
8a8d5d85
MD
466#ifdef SMP
467 int mpheld;
468#endif
8ad65e08 469
46a3f46d 470 /*
27e88a6e
MD
471 * Switching from within a 'fast' (non thread switched) interrupt or IPI
472 * is illegal. However, we may have to do it anyway if we hit a fatal
473 * kernel trap or we have paniced.
474 *
475 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 476 */
27e88a6e
MD
477 if (gd->gd_intr_nesting_level) {
478 int savegdnest;
479 int savegdtrap;
480
481 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
482 panic("lwkt_switch: cannot switch from within "
483 "a fast interrupt, yet, td %p\n", td);
484 } else {
485 savegdnest = gd->gd_intr_nesting_level;
486 savegdtrap = gd->gd_trap_nesting_level;
487 gd->gd_intr_nesting_level = 0;
488 gd->gd_trap_nesting_level = 0;
a7422615
MD
489 if ((td->td_flags & TDF_PANICWARN) == 0) {
490 td->td_flags |= TDF_PANICWARN;
491 printf("Warning: thread switch from interrupt or IPI, "
492 "thread %p (%s)\n", td, td->td_comm);
493#ifdef DDB
494 db_print_backtrace();
495#endif
496 }
27e88a6e
MD
497 lwkt_switch();
498 gd->gd_intr_nesting_level = savegdnest;
499 gd->gd_trap_nesting_level = savegdtrap;
500 return;
501 }
96728c05 502 }
ef0fdad1 503
cb973d15
MD
504 /*
505 * Passive release (used to transition from user to kernel mode
506 * when we block or switch rather then when we enter the kernel).
507 * This function is NOT called if we are switching into a preemption
508 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
509 * our current process designation (if we have one) and become a true
510 * LWKT thread, and may also hand the current process designation to
511 * another process and schedule thread.
cb973d15
MD
512 */
513 if (td->td_release)
514 td->td_release(td);
515
37af14fe 516 crit_enter_gd(gd);
dd55d707 517#ifdef SMP
9d265729
MD
518 if (td->td_toks)
519 lwkt_relalltokens(td);
dd55d707 520#endif
9d265729
MD
521
522 /*
b02926de
MD
523 * We had better not be holding any spin locks, but don't get into an
524 * endless panic loop.
9d265729 525 */
d666840a
MD
526 KASSERT(gd->gd_spinlocks_rd == 0 || panicstr != NULL,
527 ("lwkt_switch: still holding %d shared spinlocks!",
528 gd->gd_spinlocks_rd));
529 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
530 ("lwkt_switch: still holding %d exclusive spinlocks!",
531 gd->gd_spinlocks_wr));
9d265729 532
8a8d5d85
MD
533
534#ifdef SMP
535 /*
536 * td_mpcount cannot be used to determine if we currently hold the
537 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
538 * to get the lock, and switch out if it can't. Our ownership of
539 * the actual lock will remain stable while we are in a critical section
540 * (but, of course, another cpu may own or release the lock so the
541 * actual value of mp_lock is not stable).
8a8d5d85
MD
542 */
543 mpheld = MP_LOCK_HELD();
0f7a3396
MD
544#ifdef INVARIANTS
545 if (td->td_cscount) {
546 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
547 td);
548 if (panic_on_cscount)
549 panic("switching while mastering cpusync");
550 }
551#endif
8a8d5d85 552#endif
99df837e
MD
553 if ((ntd = td->td_preempted) != NULL) {
554 /*
555 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
556 * thread. This occurs transparently, whether the preempted thread
557 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
558 * itself).
559 *
560 * We have to setup the MP lock for the original thread after backing
561 * out the adjustment that was made to curthread when the original
562 * was preempted.
99df837e 563 */
26a0694b 564 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 565#ifdef SMP
96728c05 566 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 567 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
568 td, ntd, td->td_mpcount, ntd->td_mpcount);
569 }
8a8d5d85
MD
570 if (ntd->td_mpcount) {
571 td->td_mpcount -= ntd->td_mpcount;
572 KKASSERT(td->td_mpcount >= 0);
573 }
574#endif
26a0694b 575 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
576
577 /*
578 * XXX. The interrupt may have woken a thread up, we need to properly
579 * set the reschedule flag if the originally interrupted thread is at
580 * a lower priority.
581 */
582 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
583 need_lwkt_resched();
8a8d5d85 584 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 585 } else {
4b5f931b
MD
586 /*
587 * Priority queue / round-robin at each priority. Note that user
588 * processes run at a fixed, low priority and the user process
589 * scheduler deals with interactions between user processes
590 * by scheduling and descheduling them from the LWKT queue as
591 * necessary.
8a8d5d85
MD
592 *
593 * We have to adjust the MP lock for the target thread. If we
594 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
595 * thread that does not need the MP lock. If we cannot, we spin
596 * instead of HLT.
597 *
598 * A similar issue exists for the tokens held by the target thread.
599 * If we cannot obtain ownership of the tokens we cannot immediately
600 * schedule the thread.
601 */
602
8ec60c3f
MD
603 /*
604 * If an LWKT reschedule was requested, well that is what we are
605 * doing now so clear it.
606 */
607 clear_lwkt_resched();
4b5f931b
MD
608again:
609 if (gd->gd_runqmask) {
610 int nq = bsrl(gd->gd_runqmask);
611 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
612 gd->gd_runqmask &= ~(1 << nq);
613 goto again;
614 }
8a8d5d85 615#ifdef SMP
41a01a4d 616 /*
df6b8ba0
MD
617 * THREAD SELECTION FOR AN SMP MACHINE BUILD
618 *
41a01a4d
MD
619 * If the target needs the MP lock and we couldn't get it,
620 * or if the target is holding tokens and we could not
621 * gain ownership of the tokens, continue looking for a
622 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
623 *
624 * NOTE: the mpheld variable invalid after this conditional, it
625 * can change due to both cpu_try_mplock() returning success
9d265729 626 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
627 * we are trying to check the mpcount of a thread other then
628 * the current thread. Because of this, if the current thread
629 * is not holding td_mpcount, an IPI indirectly run via
9d265729 630 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 631 * cause the core MP lock to be released.
41a01a4d
MD
632 */
633 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
9d265729 634 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
41a01a4d 635 ) {
8a8d5d85 636 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
637
638 mpheld = MP_LOCK_HELD();
639 ntd = NULL;
8a8d5d85
MD
640 while (rqmask) {
641 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 642 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 643 /* spinning due to MP lock being held */
38717797 644#ifdef INVARIANTS
a453459d 645 ++mplock_contention_count;
38717797 646#endif
a453459d 647 /* mplock still not held, 'mpheld' still valid */
41a01a4d 648 continue;
38717797 649 }
a453459d
MD
650
651 /*
9d265729 652 * mpheld state invalid after getalltokens call returns
a453459d
MD
653 * failure, but the variable is only needed for
654 * the loop.
655 */
9d265729 656 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
a453459d 657 /* spinning due to token contention */
38717797 658#ifdef INVARIANTS
a453459d 659 ++token_contention_count;
38717797 660#endif
a453459d 661 mpheld = MP_LOCK_HELD();
41a01a4d 662 continue;
38717797 663 }
41a01a4d 664 break;
8a8d5d85
MD
665 }
666 if (ntd)
667 break;
668 rqmask &= ~(1 << nq);
669 nq = bsrl(rqmask);
670 }
671 if (ntd == NULL) {
a2a5ad0d
MD
672 ntd = &gd->gd_idlethread;
673 ntd->td_flags |= TDF_IDLE_NOHLT;
df6b8ba0 674 goto using_idle_thread;
8a8d5d85 675 } else {
344ad853 676 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
677 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
678 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
679 }
680 } else {
344ad853 681 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
682 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
683 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
684 }
685#else
df6b8ba0
MD
686 /*
687 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
688 * worry about tokens or the BGL.
689 */
344ad853 690 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
691 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
692 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 693#endif
4b5f931b 694 } else {
3c23a41a 695 /*
60f945af
MD
696 * We have nothing to run but only let the idle loop halt
697 * the cpu if there are no pending interrupts.
3c23a41a 698 */
a2a5ad0d 699 ntd = &gd->gd_idlethread;
60f945af 700 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 701 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 702#ifdef SMP
df6b8ba0
MD
703using_idle_thread:
704 /*
705 * The idle thread should not be holding the MP lock unless we
706 * are trapping in the kernel or in a panic. Since we select the
707 * idle thread unconditionally when no other thread is available,
708 * if the MP lock is desired during a panic or kernel trap, we
709 * have to loop in the scheduler until we get it.
710 */
711 if (ntd->td_mpcount) {
712 mpheld = MP_LOCK_HELD();
713 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
714 panic("Idle thread %p was holding the BGL!", ntd);
715 else if (mpheld == 0)
716 goto again;
717 }
a453459d 718#endif
4b5f931b 719 }
f1d1c3fa 720 }
26a0694b
MD
721 KASSERT(ntd->td_pri >= TDPRI_CRIT,
722 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
723
724 /*
725 * Do the actual switch. If the new target does not need the MP lock
726 * and we are holding it, release the MP lock. If the new target requires
727 * the MP lock we have already acquired it for the target.
728 */
729#ifdef SMP
730 if (ntd->td_mpcount == 0 ) {
731 if (MP_LOCK_HELD())
732 cpu_rel_mplock();
733 } else {
a453459d 734 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
735 }
736#endif
94f6d86e
MD
737 if (td != ntd) {
738 ++switch_count;
f1d1c3fa 739 td->td_switch(ntd);
94f6d86e 740 }
37af14fe
MD
741 /* NOTE: current cpu may have changed after switch */
742 crit_exit_quick(td);
8ad65e08
MD
743}
744
b68b7282 745/*
96728c05
MD
746 * Request that the target thread preempt the current thread. Preemption
747 * only works under a specific set of conditions:
b68b7282 748 *
96728c05
MD
749 * - We are not preempting ourselves
750 * - The target thread is owned by the current cpu
751 * - We are not currently being preempted
752 * - The target is not currently being preempted
753 * - We are able to satisfy the target's MP lock requirements (if any).
754 *
755 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
756 * this is called via lwkt_schedule() through the td_preemptable callback.
757 * critpri is the managed critical priority that we should ignore in order
758 * to determine whether preemption is possible (aka usually just the crit
759 * priority of lwkt_schedule() itself).
b68b7282 760 *
26a0694b
MD
761 * XXX at the moment we run the target thread in a critical section during
762 * the preemption in order to prevent the target from taking interrupts
763 * that *WE* can't. Preemption is strictly limited to interrupt threads
764 * and interrupt-like threads, outside of a critical section, and the
765 * preempted source thread will be resumed the instant the target blocks
766 * whether or not the source is scheduled (i.e. preemption is supposed to
767 * be as transparent as possible).
4b5f931b 768 *
8a8d5d85
MD
769 * The target thread inherits our MP count (added to its own) for the
770 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
771 * MP lock during the preemption. Therefore, any preempting targets must be
772 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
773 * out of sync with the physical mp_lock, but we do not have to preserve
774 * the original ownership of the lock if it was out of synch (that is, we
775 * can leave it synchronized on return).
b68b7282
MD
776 */
777void
96728c05 778lwkt_preempt(thread_t ntd, int critpri)
b68b7282 779{
46a3f46d 780 struct globaldata *gd = mycpu;
0a3f9b47 781 thread_t td;
8a8d5d85
MD
782#ifdef SMP
783 int mpheld;
57c254db 784 int savecnt;
8a8d5d85 785#endif
b68b7282 786
26a0694b 787 /*
96728c05
MD
788 * The caller has put us in a critical section. We can only preempt
789 * if the caller of the caller was not in a critical section (basically
d666840a
MD
790 * a local interrupt), as determined by the 'critpri' parameter. We
791 * also acn't preempt if the caller is holding any spinlocks (even if
792 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
793 *
794 * YYY The target thread must be in a critical section (else it must
795 * inherit our critical section? I dunno yet).
41a01a4d 796 *
0a3f9b47 797 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
798 */
799 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 800
0a3f9b47 801 td = gd->gd_curthread;
0a3f9b47 802 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
803 ++preempt_miss;
804 return;
805 }
96728c05
MD
806 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
807 ++preempt_miss;
8ec60c3f 808 need_lwkt_resched();
96728c05
MD
809 return;
810 }
811#ifdef SMP
46a3f46d 812 if (ntd->td_gd != gd) {
96728c05 813 ++preempt_miss;
8ec60c3f 814 need_lwkt_resched();
96728c05
MD
815 return;
816 }
817#endif
41a01a4d
MD
818 /*
819 * Take the easy way out and do not preempt if the target is holding
d666840a 820 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
821 * preempted interlock against the target thread's tokens and whether
822 * we can get all the target thread's tokens, but this situation
823 * should not occur very often so its easier to simply not preempt.
d666840a
MD
824 * Also, plain spinlocks are impossible to figure out at this point so
825 * just don't preempt.
41a01a4d 826 */
d666840a 827 if (gd->gd_spinlocks_rd + gd->gd_spinlocks_wr != 0) {
41a01a4d 828 ++preempt_miss;
8ec60c3f 829 need_lwkt_resched();
41a01a4d
MD
830 return;
831 }
26a0694b
MD
832 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
833 ++preempt_weird;
8ec60c3f 834 need_lwkt_resched();
26a0694b
MD
835 return;
836 }
837 if (ntd->td_preempted) {
4b5f931b 838 ++preempt_hit;
8ec60c3f 839 need_lwkt_resched();
26a0694b 840 return;
b68b7282 841 }
8a8d5d85 842#ifdef SMP
a2a5ad0d
MD
843 /*
844 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
845 * to or from the MP lock. In this case td_mpcount will be pre-disposed
846 * (non-zero) but not actually synchronized with the actual state of the
847 * lock. We can use it to imply an MP lock requirement for the
848 * preemption but we cannot use it to test whether we hold the MP lock
849 * or not.
a2a5ad0d 850 */
96728c05 851 savecnt = td->td_mpcount;
71ef2f5c 852 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
853 ntd->td_mpcount += td->td_mpcount;
854 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
855 ntd->td_mpcount -= td->td_mpcount;
856 ++preempt_miss;
8ec60c3f 857 need_lwkt_resched();
8a8d5d85
MD
858 return;
859 }
860#endif
26a0694b 861
8ec60c3f
MD
862 /*
863 * Since we are able to preempt the current thread, there is no need to
864 * call need_lwkt_resched().
865 */
26a0694b
MD
866 ++preempt_hit;
867 ntd->td_preempted = td;
868 td->td_flags |= TDF_PREEMPT_LOCK;
869 td->td_switch(ntd);
870 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
871#ifdef SMP
872 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
873 mpheld = MP_LOCK_HELD();
874 if (mpheld && td->td_mpcount == 0)
96728c05 875 cpu_rel_mplock();
71ef2f5c 876 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
877 panic("lwkt_preempt(): MP lock was not held through");
878#endif
26a0694b
MD
879 ntd->td_preempted = NULL;
880 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
881}
882
f1d1c3fa
MD
883/*
884 * Yield our thread while higher priority threads are pending. This is
885 * typically called when we leave a critical section but it can be safely
886 * called while we are in a critical section.
887 *
888 * This function will not generally yield to equal priority threads but it
889 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 890 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
891 * lwkt_yield_quick().
892 *
235957ed 893 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
894 * came along but was blocked and made pending.
895 *
f1d1c3fa
MD
896 * (self contained on a per cpu basis)
897 */
898void
899lwkt_yield_quick(void)
900{
7966cb69
MD
901 globaldata_t gd = mycpu;
902 thread_t td = gd->gd_curthread;
ef0fdad1 903
a2a5ad0d 904 /*
235957ed 905 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
906 * it with a non-zero cpl then we might not wind up calling splz after
907 * a task switch when the critical section is exited even though the
46a3f46d 908 * new task could accept the interrupt.
a2a5ad0d
MD
909 *
910 * XXX from crit_exit() only called after last crit section is released.
911 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
912 *
913 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
914 * except for this special case, we MUST call splz() here to handle any
915 * pending ints, particularly after we switch, or we might accidently
916 * halt the cpu with interrupts pending.
a2a5ad0d 917 */
46a3f46d 918 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 919 splz();
f1d1c3fa
MD
920
921 /*
922 * YYY enabling will cause wakeup() to task-switch, which really
923 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
924 * preemption and MP without actually doing preemption or MP, because a
925 * lot of code assumes that wakeup() does not block.
f1d1c3fa 926 */
46a3f46d
MD
927 if (untimely_switch && td->td_nest_count == 0 &&
928 gd->gd_intr_nesting_level == 0
929 ) {
37af14fe 930 crit_enter_quick(td);
f1d1c3fa
MD
931 /*
932 * YYY temporary hacks until we disassociate the userland scheduler
933 * from the LWKT scheduler.
934 */
935 if (td->td_flags & TDF_RUNQ) {
936 lwkt_switch(); /* will not reenter yield function */
937 } else {
37af14fe 938 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 939 lwkt_switch(); /* will not reenter yield function */
37af14fe 940 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 941 }
7966cb69 942 crit_exit_noyield(td);
f1d1c3fa 943 }
f1d1c3fa
MD
944}
945
8ad65e08 946/*
f1d1c3fa 947 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 948 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
949 * the crit_exit() call in lwkt_switch().
950 *
951 * (self contained on a per cpu basis)
8ad65e08
MD
952 */
953void
f1d1c3fa 954lwkt_yield(void)
8ad65e08 955{
37af14fe 956 lwkt_schedule_self(curthread);
f1d1c3fa
MD
957 lwkt_switch();
958}
959
8ad65e08 960/*
f1d1c3fa
MD
961 * Generic schedule. Possibly schedule threads belonging to other cpus and
962 * deal with threads that might be blocked on a wait queue.
963 *
0a3f9b47
MD
964 * We have a little helper inline function which does additional work after
965 * the thread has been enqueued, including dealing with preemption and
966 * setting need_lwkt_resched() (which prevents the kernel from returning
967 * to userland until it has processed higher priority threads).
6330a558
MD
968 *
969 * It is possible for this routine to be called after a failed _enqueue
970 * (due to the target thread migrating, sleeping, or otherwise blocked).
971 * We have to check that the thread is actually on the run queue!
8ad65e08 972 */
0a3f9b47
MD
973static __inline
974void
8ec60c3f 975_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
0a3f9b47 976{
6330a558
MD
977 if (ntd->td_flags & TDF_RUNQ) {
978 if (ntd->td_preemptable) {
979 ntd->td_preemptable(ntd, cpri); /* YYY +token */
980 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
981 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
982 ) {
983 need_lwkt_resched();
984 }
0a3f9b47
MD
985 }
986}
987
8ad65e08
MD
988void
989lwkt_schedule(thread_t td)
990{
37af14fe
MD
991 globaldata_t mygd = mycpu;
992
41a01a4d 993 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 994 crit_enter_gd(mygd);
344ad853 995 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
37af14fe 996 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
997 _lwkt_enqueue(td);
998 } else {
999 lwkt_wait_t w;
1000
1001 /*
1002 * If the thread is on a wait list we have to send our scheduling
1003 * request to the owner of the wait structure. Otherwise we send
1004 * the scheduling request to the cpu owning the thread. Races
1005 * are ok, the target will forward the message as necessary (the
1006 * message may chase the thread around before it finally gets
1007 * acted upon).
1008 *
1009 * (remember, wait structures use stable storage)
0a3f9b47 1010 *
0c453950
MD
1011 * NOTE: we have to account for the number of critical sections
1012 * under our control when calling _lwkt_schedule_post() so it
1013 * can figure out whether preemption is allowed.
1014 *
1015 * NOTE: The wait structure algorithms are a mess and need to be
1016 * rewritten.
1017 *
1018 * NOTE: We cannot safely acquire or release a token, even
1019 * non-blocking, because this routine may be called in the context
1020 * of a thread already holding the token and thus not provide any
1021 * interlock protection. We cannot safely manipulate the td_toks
1022 * list for the same reason. Instead we depend on our critical
1023 * section if the token is owned by our cpu.
f1d1c3fa
MD
1024 */
1025 if ((w = td->td_wait) != NULL) {
d666840a 1026 spin_lock_wr(&w->wa_spinlock);
9d265729
MD
1027 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1028 --w->wa_count;
1029 td->td_wait = NULL;
d666840a 1030 spin_unlock_wr(&w->wa_spinlock);
0f7a3396 1031#ifdef SMP
9d265729 1032 if (td->td_gd == mygd) {
0f7a3396 1033 _lwkt_enqueue(td);
8ec60c3f 1034 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 1035 } else {
9d265729
MD
1036 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1037 }
b8a98473 1038#else
9d265729
MD
1039 _lwkt_enqueue(td);
1040 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
b8a98473 1041#endif
f1d1c3fa
MD
1042 } else {
1043 /*
1044 * If the wait structure is NULL and we own the thread, there
1045 * is no race (since we are in a critical section). If we
1046 * do not own the thread there might be a race but the
1047 * target cpu will deal with it.
1048 */
0f7a3396 1049#ifdef SMP
37af14fe 1050 if (td->td_gd == mygd) {
f1d1c3fa 1051 _lwkt_enqueue(td);
8ec60c3f 1052 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 1053 } else {
b8a98473 1054 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
f1d1c3fa 1055 }
0f7a3396
MD
1056#else
1057 _lwkt_enqueue(td);
8ec60c3f 1058 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 1059#endif
f1d1c3fa 1060 }
8ad65e08 1061 }
37af14fe 1062 crit_exit_gd(mygd);
8ad65e08
MD
1063}
1064
52eedfb5
MD
1065#ifdef SMP
1066
d9eea1a5 1067/*
52eedfb5
MD
1068 * Thread migration using a 'Pull' method. The thread may or may not be
1069 * the current thread. It MUST be descheduled and in a stable state.
1070 * lwkt_giveaway() must be called on the cpu owning the thread.
1071 *
1072 * At any point after lwkt_giveaway() is called, the target cpu may
1073 * 'pull' the thread by calling lwkt_acquire().
1074 *
1075 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1076 */
52eedfb5
MD
1077void
1078lwkt_giveaway(thread_t td)
1079{
1080 globaldata_t gd = mycpu;
1081
1082 crit_enter_gd(gd);
1083 KKASSERT(td->td_gd == gd);
1084 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1085 td->td_flags |= TDF_MIGRATING;
1086 crit_exit_gd(gd);
1087}
1088
a2a5ad0d
MD
1089void
1090lwkt_acquire(thread_t td)
1091{
37af14fe
MD
1092 globaldata_t gd;
1093 globaldata_t mygd;
a2a5ad0d 1094
52eedfb5 1095 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1096 gd = td->td_gd;
37af14fe 1097 mygd = mycpu;
52eedfb5 1098 if (gd != mycpu) {
35238fa5 1099 cpu_lfence();
52eedfb5 1100 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1101 crit_enter_gd(mygd);
52eedfb5
MD
1102 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK))
1103 cpu_lfence();
37af14fe 1104 td->td_gd = mygd;
52eedfb5
MD
1105 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1106 td->td_flags &= ~TDF_MIGRATING;
1107 crit_exit_gd(mygd);
1108 } else {
1109 crit_enter_gd(mygd);
1110 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1111 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1112 crit_exit_gd(mygd);
a2a5ad0d
MD
1113 }
1114}
1115
52eedfb5
MD
1116#endif
1117
f1d1c3fa
MD
1118/*
1119 * Generic deschedule. Descheduling threads other then your own should be
1120 * done only in carefully controlled circumstances. Descheduling is
1121 * asynchronous.
1122 *
1123 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1124 */
1125void
1126lwkt_deschedule(thread_t td)
1127{
f1d1c3fa 1128 crit_enter();
b8a98473 1129#ifdef SMP
f1d1c3fa
MD
1130 if (td == curthread) {
1131 _lwkt_dequeue(td);
1132 } else {
a72187e9 1133 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1134 _lwkt_dequeue(td);
1135 } else {
b8a98473 1136 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1137 }
1138 }
b8a98473
MD
1139#else
1140 _lwkt_dequeue(td);
1141#endif
f1d1c3fa
MD
1142 crit_exit();
1143}
1144
4b5f931b
MD
1145/*
1146 * Set the target thread's priority. This routine does not automatically
1147 * switch to a higher priority thread, LWKT threads are not designed for
1148 * continuous priority changes. Yield if you want to switch.
1149 *
1150 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1151 * of the td_pri field. The specified priority may also indicate zero or
1152 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1153 *
1154 * Note that we requeue the thread whether it winds up on a different runq
1155 * or not. uio_yield() depends on this and the routine is not normally
1156 * called with the same priority otherwise.
4b5f931b
MD
1157 */
1158void
1159lwkt_setpri(thread_t td, int pri)
1160{
26a0694b 1161 KKASSERT(pri >= 0);
a72187e9 1162 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1163 crit_enter();
1164 if (td->td_flags & TDF_RUNQ) {
1165 _lwkt_dequeue(td);
1166 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1167 _lwkt_enqueue(td);
1168 } else {
1169 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1170 }
1171 crit_exit();
1172}
1173
1174void
1175lwkt_setpri_self(int pri)
1176{
1177 thread_t td = curthread;
1178
4b5f931b
MD
1179 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1180 crit_enter();
1181 if (td->td_flags & TDF_RUNQ) {
1182 _lwkt_dequeue(td);
1183 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1184 _lwkt_enqueue(td);
1185 } else {
1186 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1187 }
1188 crit_exit();
1189}
1190
8ec60c3f
MD
1191/*
1192 * Determine if there is a runnable thread at a higher priority then
1193 * the current thread. lwkt_setpri() does not check this automatically.
1194 * Return 1 if there is, 0 if there isn't.
1195 *
1196 * Example: if bit 31 of runqmask is set and the current thread is priority
1197 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1198 *
1199 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1200 * up comparing against 0xffffffff, a comparison that will always be false.
1201 */
1202int
1203lwkt_checkpri_self(void)
1204{
1205 globaldata_t gd = mycpu;
1206 thread_t td = gd->gd_curthread;
1207 int nq = td->td_pri & TDPRI_MASK;
1208
1209 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1210 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1211 return(1);
1212 ++nq;
1213 }
1214 return(0);
1215}
1216
5d21b981 1217/*
52eedfb5
MD
1218 * Migrate the current thread to the specified cpu.
1219 *
1220 * This is accomplished by descheduling ourselves from the current cpu,
1221 * moving our thread to the tdallq of the target cpu, IPI messaging the
1222 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1223 * races while the thread is being migrated.
5d21b981 1224 */
3d28ff59 1225#ifdef SMP
5d21b981 1226static void lwkt_setcpu_remote(void *arg);
3d28ff59 1227#endif
5d21b981
MD
1228
1229void
1230lwkt_setcpu_self(globaldata_t rgd)
1231{
1232#ifdef SMP
1233 thread_t td = curthread;
1234
1235 if (td->td_gd != rgd) {
1236 crit_enter_quick(td);
1237 td->td_flags |= TDF_MIGRATING;
1238 lwkt_deschedule_self(td);
52eedfb5 1239 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1240 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1241 lwkt_switch();
1242 /* we are now on the target cpu */
52eedfb5 1243 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1244 crit_exit_quick(td);
1245 }
1246#endif
1247}
1248
ecdefdda
MD
1249void
1250lwkt_migratecpu(int cpuid)
1251{
1252#ifdef SMP
1253 globaldata_t rgd;
1254
1255 rgd = globaldata_find(cpuid);
1256 lwkt_setcpu_self(rgd);
1257#endif
1258}
1259
5d21b981
MD
1260/*
1261 * Remote IPI for cpu migration (called while in a critical section so we
1262 * do not have to enter another one). The thread has already been moved to
1263 * our cpu's allq, but we must wait for the thread to be completely switched
1264 * out on the originating cpu before we schedule it on ours or the stack
1265 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1266 * change to main memory.
1267 *
1268 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1269 * against wakeups. It is best if this interface is used only when there
1270 * are no pending events that might try to schedule the thread.
1271 */
3d28ff59 1272#ifdef SMP
5d21b981
MD
1273static void
1274lwkt_setcpu_remote(void *arg)
1275{
1276 thread_t td = arg;
1277 globaldata_t gd = mycpu;
1278
c1102e9f 1279 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK))
35238fa5 1280 cpu_lfence();
5d21b981 1281 td->td_gd = gd;
35238fa5 1282 cpu_sfence();
5d21b981 1283 td->td_flags &= ~TDF_MIGRATING;
344ad853 1284 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
5d21b981
MD
1285 _lwkt_enqueue(td);
1286}
3d28ff59 1287#endif
5d21b981 1288
553ea3c8 1289struct lwp *
4b5f931b
MD
1290lwkt_preempted_proc(void)
1291{
73e4f7b9 1292 thread_t td = curthread;
4b5f931b
MD
1293 while (td->td_preempted)
1294 td = td->td_preempted;
553ea3c8 1295 return(td->td_lwp);
4b5f931b
MD
1296}
1297
f1d1c3fa 1298/*
41a01a4d
MD
1299 * Block on the specified wait queue until signaled. A generation number
1300 * must be supplied to interlock the wait queue. The function will
1301 * return immediately if the generation number does not match the wait
1302 * structure's generation number.
f1d1c3fa
MD
1303 */
1304void
ae8050a4 1305lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
1306{
1307 thread_t td = curthread;
f1d1c3fa 1308
d666840a 1309 spin_lock_wr(&w->wa_spinlock);
ae8050a4 1310 if (w->wa_gen == *gen) {
f1d1c3fa 1311 _lwkt_dequeue(td);
344ad853 1312 td->td_flags |= TDF_BLOCKQ;
f1d1c3fa
MD
1313 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1314 ++w->wa_count;
1315 td->td_wait = w;
ae8050a4 1316 td->td_wmesg = wmesg;
d666840a 1317 spin_unlock_wr(&w->wa_spinlock);
f1d1c3fa 1318 lwkt_switch();
344ad853
MD
1319 KKASSERT((td->td_flags & TDF_BLOCKQ) == 0);
1320 td->td_wmesg = NULL;
9d265729
MD
1321 *gen = w->wa_gen;
1322 } else {
1323 *gen = w->wa_gen;
d666840a 1324 spin_unlock_wr(&w->wa_spinlock);
8ad65e08 1325 }
f1d1c3fa
MD
1326}
1327
1328/*
1329 * Signal a wait queue. We gain ownership of the wait queue in order to
1330 * signal it. Once a thread is removed from the wait queue we have to
1331 * deal with the cpu owning the thread.
1332 *
1333 * Note: alternatively we could message the target cpu owning the wait
1334 * queue. YYY implement as sysctl.
1335 */
1336void
ece04fd0 1337lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1338{
1339 thread_t td;
f1d1c3fa 1340
d666840a 1341 spin_lock_wr(&w->wa_spinlock);
f1d1c3fa 1342 ++w->wa_gen;
ece04fd0
MD
1343 if (count < 0)
1344 count = w->wa_count;
f1d1c3fa
MD
1345 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1346 --count;
1347 --w->wa_count;
344ad853 1348 KKASSERT(td->td_flags & TDF_BLOCKQ);
f1d1c3fa 1349 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
344ad853 1350 td->td_flags &= ~TDF_BLOCKQ;
f1d1c3fa 1351 td->td_wait = NULL;
d666840a 1352 spin_unlock_wr(&w->wa_spinlock);
344ad853 1353 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
b8a98473 1354#ifdef SMP
a72187e9 1355 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1356 _lwkt_enqueue(td);
1357 } else {
b8a98473 1358 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
f1d1c3fa 1359 }
b8a98473
MD
1360#else
1361 _lwkt_enqueue(td);
1362#endif
d666840a 1363 spin_lock_wr(&w->wa_spinlock);
f1d1c3fa 1364 }
d666840a 1365 spin_unlock_wr(&w->wa_spinlock);
f1d1c3fa
MD
1366}
1367
99df837e
MD
1368/*
1369 * Create a kernel process/thread/whatever. It shares it's address space
1370 * with proc0 - ie: kernel only.
1371 *
365fa13f
MD
1372 * NOTE! By default new threads are created with the MP lock held. A
1373 * thread which does not require the MP lock should release it by calling
1374 * rel_mplock() at the start of the new thread.
99df837e
MD
1375 */
1376int
1377lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1378 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1379 const char *fmt, ...)
99df837e 1380{
73e4f7b9 1381 thread_t td;
e2565a42 1382 __va_list ap;
99df837e 1383
d3d32139
MD
1384 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1385 tdflags | TDF_VERBOSE);
a2a5ad0d
MD
1386 if (tdp)
1387 *tdp = td;
709799ea 1388 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1389
1390 /*
1391 * Set up arg0 for 'ps' etc
1392 */
e2565a42 1393 __va_start(ap, fmt);
99df837e 1394 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1395 __va_end(ap);
99df837e
MD
1396
1397 /*
1398 * Schedule the thread to run
1399 */
ef0fdad1
MD
1400 if ((td->td_flags & TDF_STOPREQ) == 0)
1401 lwkt_schedule(td);
1402 else
1403 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1404 return 0;
1405}
1406
2d93b37a 1407/*
2d93b37a
MD
1408 * kthread_* is specific to the kernel and is not needed by userland.
1409 */
1410#ifdef _KERNEL
1411
99df837e
MD
1412/*
1413 * Destroy an LWKT thread. Warning! This function is not called when
1414 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1415 * uses a different reaping mechanism.
1416 */
1417void
1418lwkt_exit(void)
1419{
1420 thread_t td = curthread;
8826f33a 1421 globaldata_t gd;
99df837e
MD
1422
1423 if (td->td_flags & TDF_VERBOSE)
1424 printf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1425 caps_exit(td);
37af14fe
MD
1426 crit_enter_quick(td);
1427 lwkt_deschedule_self(td);
8826f33a
MD
1428 gd = mycpu;
1429 KKASSERT(gd == td->td_gd);
1430 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1431 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1432 ++gd->gd_tdfreecount;
1433 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1434 }
99df837e
MD
1435 cpu_thread_exit();
1436}
1437
2d93b37a
MD
1438#endif /* _KERNEL */
1439
1440void
1441crit_panic(void)
1442{
1443 thread_t td = curthread;
1444 int lpri = td->td_pri;
1445
1446 td->td_pri = 0;
1447 panic("td_pri is/would-go negative! %p %d", td, lpri);
1448}
1449
d165e668
MD
1450#ifdef SMP
1451
bd8015ca
MD
1452/*
1453 * Called from debugger/panic on cpus which have been stopped. We must still
1454 * process the IPIQ while stopped, even if we were stopped while in a critical
1455 * section (XXX).
1456 *
1457 * If we are dumping also try to process any pending interrupts. This may
1458 * or may not work depending on the state of the cpu at the point it was
1459 * stopped.
1460 */
1461void
1462lwkt_smp_stopped(void)
1463{
1464 globaldata_t gd = mycpu;
1465
1466 crit_enter_gd(gd);
1467 if (dumping) {
1468 lwkt_process_ipiq();
1469 splz();
1470 } else {
1471 lwkt_process_ipiq();
1472 }
1473 crit_exit_gd(gd);
1474}
1475
d165e668 1476#endif