Get rid -y/-Y (sort by interactive measure). The interactive measure has
[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 *
d666840a 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.96 2006/05/21 20:23:25 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
312 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
313}
314
41a01a4d
MD
315#endif
316
7d0bac62 317void
f470d0c8
MD
318lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
319 struct globaldata *gd)
7d0bac62 320{
37af14fe
MD
321 globaldata_t mygd = mycpu;
322
99df837e
MD
323 bzero(td, sizeof(struct thread));
324 td->td_kstack = stack;
f470d0c8 325 td->td_kstack_size = stksize;
d3d32139 326 td->td_flags = flags;
26a0694b 327 td->td_gd = gd;
f8c3996b 328 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
329#ifdef SMP
330 if ((flags & TDF_MPSAFE) == 0)
331 td->td_mpcount = 1;
332#endif
c95cd171 333 lwkt_initport(&td->td_msgport, td);
99df837e 334 pmap_init_thread(td);
0f7a3396 335#ifdef SMP
5d21b981
MD
336 /*
337 * Normally initializing a thread for a remote cpu requires sending an
338 * IPI. However, the idlethread is setup before the other cpus are
339 * activated so we have to treat it as a special case. XXX manipulation
340 * of gd_tdallq requires the BGL.
341 */
342 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 343 crit_enter_gd(mygd);
75cdbe6c 344 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 345 crit_exit_gd(mygd);
75cdbe6c 346 } else {
2db3b277 347 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 348 }
0f7a3396 349#else
37af14fe 350 crit_enter_gd(mygd);
0f7a3396 351 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 352 crit_exit_gd(mygd);
0f7a3396 353#endif
73e4f7b9
MD
354}
355
2d93b37a
MD
356#endif /* _KERNEL */
357
73e4f7b9
MD
358void
359lwkt_set_comm(thread_t td, const char *ctl, ...)
360{
e2565a42 361 __va_list va;
73e4f7b9 362
e2565a42 363 __va_start(va, ctl);
73e4f7b9 364 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 365 __va_end(va);
7d0bac62
MD
366}
367
99df837e 368void
73e4f7b9 369lwkt_hold(thread_t td)
99df837e 370{
73e4f7b9
MD
371 ++td->td_refs;
372}
373
374void
375lwkt_rele(thread_t td)
376{
377 KKASSERT(td->td_refs > 0);
378 --td->td_refs;
379}
380
c95cd171
MD
381#ifdef _KERNEL
382
73e4f7b9
MD
383void
384lwkt_wait_free(thread_t td)
385{
386 while (td->td_refs)
377d4740 387 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
388}
389
c95cd171
MD
390#endif
391
73e4f7b9
MD
392void
393lwkt_free_thread(thread_t td)
394{
395 struct globaldata *gd = mycpu;
396
d9eea1a5 397 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
398 ("lwkt_free_thread: did not exit! %p", td));
399
37af14fe 400 crit_enter_gd(gd);
1ec20b9d 401 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* Protected by BGL */
73e4f7b9 402 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
403 (td->td_flags & TDF_ALLOCATED_THREAD)
404 ) {
73e4f7b9
MD
405 ++gd->gd_tdfreecount;
406 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
37af14fe 407 crit_exit_gd(gd);
99df837e 408 } else {
37af14fe 409 crit_exit_gd(gd);
99df837e 410 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613 411#ifdef _KERNEL
f470d0c8 412 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
05220613 413#else
f470d0c8 414 libcaps_free_stack(td->td_kstack, td->td_kstack_size);
05220613 415#endif
73e4f7b9 416 /* gd invalid */
99df837e 417 td->td_kstack = NULL;
f470d0c8 418 td->td_kstack_size = 0;
99df837e 419 }
05220613
MD
420 if (td->td_flags & TDF_ALLOCATED_THREAD) {
421#ifdef _KERNEL
99df837e 422 zfree(thread_zone, td);
05220613
MD
423#else
424 free(td);
425#endif
426 }
99df837e
MD
427 }
428}
429
430
8ad65e08
MD
431/*
432 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
433 * switch to the idlethread. Switching must occur within a critical
434 * section to avoid races with the scheduling queue.
435 *
436 * We always have full control over our cpu's run queue. Other cpus
437 * that wish to manipulate our queue must use the cpu_*msg() calls to
438 * talk to our cpu, so a critical section is all that is needed and
439 * the result is very, very fast thread switching.
440 *
96728c05
MD
441 * The LWKT scheduler uses a fixed priority model and round-robins at
442 * each priority level. User process scheduling is a totally
443 * different beast and LWKT priorities should not be confused with
444 * user process priorities.
f1d1c3fa 445 *
96728c05
MD
446 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
447 * cleans it up. Note that the td_switch() function cannot do anything that
448 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
449 * the target thread (not the current thread). It's nice to have a scheduler
450 * that does not need the MP lock to work because it allows us to do some
451 * really cool high-performance MP lock optimizations.
69d78e99
MD
452 *
453 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
454 * is not called by the current thread in the preemption case, only when
455 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
456 */
457void
458lwkt_switch(void)
459{
37af14fe
MD
460 globaldata_t gd = mycpu;
461 thread_t td = gd->gd_curthread;
8ad65e08 462 thread_t ntd;
8a8d5d85
MD
463#ifdef SMP
464 int mpheld;
465#endif
8ad65e08 466
46a3f46d 467 /*
27e88a6e
MD
468 * Switching from within a 'fast' (non thread switched) interrupt or IPI
469 * is illegal. However, we may have to do it anyway if we hit a fatal
470 * kernel trap or we have paniced.
471 *
472 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 473 */
27e88a6e
MD
474 if (gd->gd_intr_nesting_level) {
475 int savegdnest;
476 int savegdtrap;
477
478 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
479 panic("lwkt_switch: cannot switch from within "
480 "a fast interrupt, yet, td %p\n", td);
481 } else {
482 savegdnest = gd->gd_intr_nesting_level;
483 savegdtrap = gd->gd_trap_nesting_level;
484 gd->gd_intr_nesting_level = 0;
485 gd->gd_trap_nesting_level = 0;
a7422615
MD
486 if ((td->td_flags & TDF_PANICWARN) == 0) {
487 td->td_flags |= TDF_PANICWARN;
488 printf("Warning: thread switch from interrupt or IPI, "
489 "thread %p (%s)\n", td, td->td_comm);
490#ifdef DDB
491 db_print_backtrace();
492#endif
493 }
27e88a6e
MD
494 lwkt_switch();
495 gd->gd_intr_nesting_level = savegdnest;
496 gd->gd_trap_nesting_level = savegdtrap;
497 return;
498 }
96728c05 499 }
ef0fdad1 500
cb973d15
MD
501 /*
502 * Passive release (used to transition from user to kernel mode
503 * when we block or switch rather then when we enter the kernel).
504 * This function is NOT called if we are switching into a preemption
505 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
506 * our current process designation (if we have one) and become a true
507 * LWKT thread, and may also hand the current process designation to
508 * another process and schedule thread.
cb973d15
MD
509 */
510 if (td->td_release)
511 td->td_release(td);
512
37af14fe 513 crit_enter_gd(gd);
dd55d707 514#ifdef SMP
9d265729
MD
515 if (td->td_toks)
516 lwkt_relalltokens(td);
dd55d707 517#endif
9d265729
MD
518
519 /*
b02926de
MD
520 * We had better not be holding any spin locks, but don't get into an
521 * endless panic loop.
9d265729 522 */
d666840a
MD
523 KASSERT(gd->gd_spinlocks_rd == 0 || panicstr != NULL,
524 ("lwkt_switch: still holding %d shared spinlocks!",
525 gd->gd_spinlocks_rd));
526 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
527 ("lwkt_switch: still holding %d exclusive spinlocks!",
528 gd->gd_spinlocks_wr));
9d265729 529
8a8d5d85
MD
530
531#ifdef SMP
532 /*
533 * td_mpcount cannot be used to determine if we currently hold the
534 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
535 * to get the lock, and switch out if it can't. Our ownership of
536 * the actual lock will remain stable while we are in a critical section
537 * (but, of course, another cpu may own or release the lock so the
538 * actual value of mp_lock is not stable).
8a8d5d85
MD
539 */
540 mpheld = MP_LOCK_HELD();
0f7a3396
MD
541#ifdef INVARIANTS
542 if (td->td_cscount) {
543 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
544 td);
545 if (panic_on_cscount)
546 panic("switching while mastering cpusync");
547 }
548#endif
8a8d5d85 549#endif
99df837e
MD
550 if ((ntd = td->td_preempted) != NULL) {
551 /*
552 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
553 * thread. This occurs transparently, whether the preempted thread
554 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
555 * itself).
556 *
557 * We have to setup the MP lock for the original thread after backing
558 * out the adjustment that was made to curthread when the original
559 * was preempted.
99df837e 560 */
26a0694b 561 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 562#ifdef SMP
96728c05 563 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 564 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
565 td, ntd, td->td_mpcount, ntd->td_mpcount);
566 }
8a8d5d85
MD
567 if (ntd->td_mpcount) {
568 td->td_mpcount -= ntd->td_mpcount;
569 KKASSERT(td->td_mpcount >= 0);
570 }
571#endif
26a0694b 572 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
573
574 /*
575 * XXX. The interrupt may have woken a thread up, we need to properly
576 * set the reschedule flag if the originally interrupted thread is at
577 * a lower priority.
578 */
579 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
580 need_lwkt_resched();
8a8d5d85 581 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 582 } else {
4b5f931b
MD
583 /*
584 * Priority queue / round-robin at each priority. Note that user
585 * processes run at a fixed, low priority and the user process
586 * scheduler deals with interactions between user processes
587 * by scheduling and descheduling them from the LWKT queue as
588 * necessary.
8a8d5d85
MD
589 *
590 * We have to adjust the MP lock for the target thread. If we
591 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
592 * thread that does not need the MP lock. If we cannot, we spin
593 * instead of HLT.
594 *
595 * A similar issue exists for the tokens held by the target thread.
596 * If we cannot obtain ownership of the tokens we cannot immediately
597 * schedule the thread.
598 */
599
8ec60c3f
MD
600 /*
601 * If an LWKT reschedule was requested, well that is what we are
602 * doing now so clear it.
603 */
604 clear_lwkt_resched();
4b5f931b
MD
605again:
606 if (gd->gd_runqmask) {
607 int nq = bsrl(gd->gd_runqmask);
608 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
609 gd->gd_runqmask &= ~(1 << nq);
610 goto again;
611 }
8a8d5d85 612#ifdef SMP
41a01a4d 613 /*
df6b8ba0
MD
614 * THREAD SELECTION FOR AN SMP MACHINE BUILD
615 *
41a01a4d
MD
616 * If the target needs the MP lock and we couldn't get it,
617 * or if the target is holding tokens and we could not
618 * gain ownership of the tokens, continue looking for a
619 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
620 *
621 * NOTE: the mpheld variable invalid after this conditional, it
622 * can change due to both cpu_try_mplock() returning success
9d265729 623 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
624 * we are trying to check the mpcount of a thread other then
625 * the current thread. Because of this, if the current thread
626 * is not holding td_mpcount, an IPI indirectly run via
9d265729 627 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 628 * cause the core MP lock to be released.
41a01a4d
MD
629 */
630 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
9d265729 631 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
41a01a4d 632 ) {
8a8d5d85 633 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
634
635 mpheld = MP_LOCK_HELD();
636 ntd = NULL;
8a8d5d85
MD
637 while (rqmask) {
638 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 639 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 640 /* spinning due to MP lock being held */
38717797 641#ifdef INVARIANTS
a453459d 642 ++mplock_contention_count;
38717797 643#endif
a453459d 644 /* mplock still not held, 'mpheld' still valid */
41a01a4d 645 continue;
38717797 646 }
a453459d
MD
647
648 /*
9d265729 649 * mpheld state invalid after getalltokens call returns
a453459d
MD
650 * failure, but the variable is only needed for
651 * the loop.
652 */
9d265729 653 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
a453459d 654 /* spinning due to token contention */
38717797 655#ifdef INVARIANTS
a453459d 656 ++token_contention_count;
38717797 657#endif
a453459d 658 mpheld = MP_LOCK_HELD();
41a01a4d 659 continue;
38717797 660 }
41a01a4d 661 break;
8a8d5d85
MD
662 }
663 if (ntd)
664 break;
665 rqmask &= ~(1 << nq);
666 nq = bsrl(rqmask);
667 }
668 if (ntd == NULL) {
a2a5ad0d
MD
669 ntd = &gd->gd_idlethread;
670 ntd->td_flags |= TDF_IDLE_NOHLT;
df6b8ba0 671 goto using_idle_thread;
8a8d5d85 672 } else {
344ad853 673 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
674 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
675 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
676 }
677 } else {
344ad853 678 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
679 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
680 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
681 }
682#else
df6b8ba0
MD
683 /*
684 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
685 * worry about tokens or the BGL.
686 */
344ad853 687 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
688 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
689 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 690#endif
4b5f931b 691 } else {
3c23a41a 692 /*
60f945af
MD
693 * We have nothing to run but only let the idle loop halt
694 * the cpu if there are no pending interrupts.
3c23a41a 695 */
a2a5ad0d 696 ntd = &gd->gd_idlethread;
60f945af 697 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 698 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 699#ifdef SMP
df6b8ba0
MD
700using_idle_thread:
701 /*
702 * The idle thread should not be holding the MP lock unless we
703 * are trapping in the kernel or in a panic. Since we select the
704 * idle thread unconditionally when no other thread is available,
705 * if the MP lock is desired during a panic or kernel trap, we
706 * have to loop in the scheduler until we get it.
707 */
708 if (ntd->td_mpcount) {
709 mpheld = MP_LOCK_HELD();
710 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
711 panic("Idle thread %p was holding the BGL!", ntd);
712 else if (mpheld == 0)
713 goto again;
714 }
a453459d 715#endif
4b5f931b 716 }
f1d1c3fa 717 }
26a0694b
MD
718 KASSERT(ntd->td_pri >= TDPRI_CRIT,
719 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
720
721 /*
722 * Do the actual switch. If the new target does not need the MP lock
723 * and we are holding it, release the MP lock. If the new target requires
724 * the MP lock we have already acquired it for the target.
725 */
726#ifdef SMP
727 if (ntd->td_mpcount == 0 ) {
728 if (MP_LOCK_HELD())
729 cpu_rel_mplock();
730 } else {
a453459d 731 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
732 }
733#endif
94f6d86e
MD
734 if (td != ntd) {
735 ++switch_count;
f1d1c3fa 736 td->td_switch(ntd);
94f6d86e 737 }
37af14fe
MD
738 /* NOTE: current cpu may have changed after switch */
739 crit_exit_quick(td);
8ad65e08
MD
740}
741
b68b7282 742/*
96728c05
MD
743 * Request that the target thread preempt the current thread. Preemption
744 * only works under a specific set of conditions:
b68b7282 745 *
96728c05
MD
746 * - We are not preempting ourselves
747 * - The target thread is owned by the current cpu
748 * - We are not currently being preempted
749 * - The target is not currently being preempted
750 * - We are able to satisfy the target's MP lock requirements (if any).
751 *
752 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
753 * this is called via lwkt_schedule() through the td_preemptable callback.
754 * critpri is the managed critical priority that we should ignore in order
755 * to determine whether preemption is possible (aka usually just the crit
756 * priority of lwkt_schedule() itself).
b68b7282 757 *
26a0694b
MD
758 * XXX at the moment we run the target thread in a critical section during
759 * the preemption in order to prevent the target from taking interrupts
760 * that *WE* can't. Preemption is strictly limited to interrupt threads
761 * and interrupt-like threads, outside of a critical section, and the
762 * preempted source thread will be resumed the instant the target blocks
763 * whether or not the source is scheduled (i.e. preemption is supposed to
764 * be as transparent as possible).
4b5f931b 765 *
8a8d5d85
MD
766 * The target thread inherits our MP count (added to its own) for the
767 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
768 * MP lock during the preemption. Therefore, any preempting targets must be
769 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
770 * out of sync with the physical mp_lock, but we do not have to preserve
771 * the original ownership of the lock if it was out of synch (that is, we
772 * can leave it synchronized on return).
b68b7282
MD
773 */
774void
96728c05 775lwkt_preempt(thread_t ntd, int critpri)
b68b7282 776{
46a3f46d 777 struct globaldata *gd = mycpu;
0a3f9b47 778 thread_t td;
8a8d5d85
MD
779#ifdef SMP
780 int mpheld;
57c254db 781 int savecnt;
8a8d5d85 782#endif
b68b7282 783
26a0694b 784 /*
96728c05
MD
785 * The caller has put us in a critical section. We can only preempt
786 * if the caller of the caller was not in a critical section (basically
d666840a
MD
787 * a local interrupt), as determined by the 'critpri' parameter. We
788 * also acn't preempt if the caller is holding any spinlocks (even if
789 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
790 *
791 * YYY The target thread must be in a critical section (else it must
792 * inherit our critical section? I dunno yet).
41a01a4d 793 *
0a3f9b47 794 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
795 */
796 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 797
0a3f9b47 798 td = gd->gd_curthread;
0a3f9b47 799 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
800 ++preempt_miss;
801 return;
802 }
96728c05
MD
803 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
804 ++preempt_miss;
8ec60c3f 805 need_lwkt_resched();
96728c05
MD
806 return;
807 }
808#ifdef SMP
46a3f46d 809 if (ntd->td_gd != gd) {
96728c05 810 ++preempt_miss;
8ec60c3f 811 need_lwkt_resched();
96728c05
MD
812 return;
813 }
814#endif
41a01a4d
MD
815 /*
816 * Take the easy way out and do not preempt if the target is holding
d666840a 817 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
818 * preempted interlock against the target thread's tokens and whether
819 * we can get all the target thread's tokens, but this situation
820 * should not occur very often so its easier to simply not preempt.
d666840a
MD
821 * Also, plain spinlocks are impossible to figure out at this point so
822 * just don't preempt.
41a01a4d 823 */
d666840a 824 if (gd->gd_spinlocks_rd + gd->gd_spinlocks_wr != 0) {
41a01a4d 825 ++preempt_miss;
8ec60c3f 826 need_lwkt_resched();
41a01a4d
MD
827 return;
828 }
26a0694b
MD
829 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
830 ++preempt_weird;
8ec60c3f 831 need_lwkt_resched();
26a0694b
MD
832 return;
833 }
834 if (ntd->td_preempted) {
4b5f931b 835 ++preempt_hit;
8ec60c3f 836 need_lwkt_resched();
26a0694b 837 return;
b68b7282 838 }
8a8d5d85 839#ifdef SMP
a2a5ad0d
MD
840 /*
841 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
842 * to or from the MP lock. In this case td_mpcount will be pre-disposed
843 * (non-zero) but not actually synchronized with the actual state of the
844 * lock. We can use it to imply an MP lock requirement for the
845 * preemption but we cannot use it to test whether we hold the MP lock
846 * or not.
a2a5ad0d 847 */
96728c05 848 savecnt = td->td_mpcount;
71ef2f5c 849 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
850 ntd->td_mpcount += td->td_mpcount;
851 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
852 ntd->td_mpcount -= td->td_mpcount;
853 ++preempt_miss;
8ec60c3f 854 need_lwkt_resched();
8a8d5d85
MD
855 return;
856 }
857#endif
26a0694b 858
8ec60c3f
MD
859 /*
860 * Since we are able to preempt the current thread, there is no need to
861 * call need_lwkt_resched().
862 */
26a0694b
MD
863 ++preempt_hit;
864 ntd->td_preempted = td;
865 td->td_flags |= TDF_PREEMPT_LOCK;
866 td->td_switch(ntd);
867 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
868#ifdef SMP
869 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
870 mpheld = MP_LOCK_HELD();
871 if (mpheld && td->td_mpcount == 0)
96728c05 872 cpu_rel_mplock();
71ef2f5c 873 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
874 panic("lwkt_preempt(): MP lock was not held through");
875#endif
26a0694b
MD
876 ntd->td_preempted = NULL;
877 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
878}
879
f1d1c3fa
MD
880/*
881 * Yield our thread while higher priority threads are pending. This is
882 * typically called when we leave a critical section but it can be safely
883 * called while we are in a critical section.
884 *
885 * This function will not generally yield to equal priority threads but it
886 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 887 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
888 * lwkt_yield_quick().
889 *
235957ed 890 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
891 * came along but was blocked and made pending.
892 *
f1d1c3fa
MD
893 * (self contained on a per cpu basis)
894 */
895void
896lwkt_yield_quick(void)
897{
7966cb69
MD
898 globaldata_t gd = mycpu;
899 thread_t td = gd->gd_curthread;
ef0fdad1 900
a2a5ad0d 901 /*
235957ed 902 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
903 * it with a non-zero cpl then we might not wind up calling splz after
904 * a task switch when the critical section is exited even though the
46a3f46d 905 * new task could accept the interrupt.
a2a5ad0d
MD
906 *
907 * XXX from crit_exit() only called after last crit section is released.
908 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
909 *
910 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
911 * except for this special case, we MUST call splz() here to handle any
912 * pending ints, particularly after we switch, or we might accidently
913 * halt the cpu with interrupts pending.
a2a5ad0d 914 */
46a3f46d 915 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 916 splz();
f1d1c3fa
MD
917
918 /*
919 * YYY enabling will cause wakeup() to task-switch, which really
920 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
921 * preemption and MP without actually doing preemption or MP, because a
922 * lot of code assumes that wakeup() does not block.
f1d1c3fa 923 */
46a3f46d
MD
924 if (untimely_switch && td->td_nest_count == 0 &&
925 gd->gd_intr_nesting_level == 0
926 ) {
37af14fe 927 crit_enter_quick(td);
f1d1c3fa
MD
928 /*
929 * YYY temporary hacks until we disassociate the userland scheduler
930 * from the LWKT scheduler.
931 */
932 if (td->td_flags & TDF_RUNQ) {
933 lwkt_switch(); /* will not reenter yield function */
934 } else {
37af14fe 935 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 936 lwkt_switch(); /* will not reenter yield function */
37af14fe 937 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 938 }
7966cb69 939 crit_exit_noyield(td);
f1d1c3fa 940 }
f1d1c3fa
MD
941}
942
8ad65e08 943/*
f1d1c3fa 944 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 945 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
946 * the crit_exit() call in lwkt_switch().
947 *
948 * (self contained on a per cpu basis)
8ad65e08
MD
949 */
950void
f1d1c3fa 951lwkt_yield(void)
8ad65e08 952{
37af14fe 953 lwkt_schedule_self(curthread);
f1d1c3fa
MD
954 lwkt_switch();
955}
956
8ad65e08 957/*
f1d1c3fa
MD
958 * Generic schedule. Possibly schedule threads belonging to other cpus and
959 * deal with threads that might be blocked on a wait queue.
960 *
0a3f9b47
MD
961 * We have a little helper inline function which does additional work after
962 * the thread has been enqueued, including dealing with preemption and
963 * setting need_lwkt_resched() (which prevents the kernel from returning
964 * to userland until it has processed higher priority threads).
6330a558
MD
965 *
966 * It is possible for this routine to be called after a failed _enqueue
967 * (due to the target thread migrating, sleeping, or otherwise blocked).
968 * We have to check that the thread is actually on the run queue!
8ad65e08 969 */
0a3f9b47
MD
970static __inline
971void
8ec60c3f 972_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
0a3f9b47 973{
6330a558
MD
974 if (ntd->td_flags & TDF_RUNQ) {
975 if (ntd->td_preemptable) {
976 ntd->td_preemptable(ntd, cpri); /* YYY +token */
977 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
978 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
979 ) {
980 need_lwkt_resched();
981 }
0a3f9b47
MD
982 }
983}
984
8ad65e08
MD
985void
986lwkt_schedule(thread_t td)
987{
37af14fe
MD
988 globaldata_t mygd = mycpu;
989
41a01a4d 990 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 991 crit_enter_gd(mygd);
344ad853 992 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
37af14fe 993 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
994 _lwkt_enqueue(td);
995 } else {
996 lwkt_wait_t w;
997
998 /*
999 * If the thread is on a wait list we have to send our scheduling
1000 * request to the owner of the wait structure. Otherwise we send
1001 * the scheduling request to the cpu owning the thread. Races
1002 * are ok, the target will forward the message as necessary (the
1003 * message may chase the thread around before it finally gets
1004 * acted upon).
1005 *
1006 * (remember, wait structures use stable storage)
0a3f9b47 1007 *
0c453950
MD
1008 * NOTE: we have to account for the number of critical sections
1009 * under our control when calling _lwkt_schedule_post() so it
1010 * can figure out whether preemption is allowed.
1011 *
1012 * NOTE: The wait structure algorithms are a mess and need to be
1013 * rewritten.
1014 *
1015 * NOTE: We cannot safely acquire or release a token, even
1016 * non-blocking, because this routine may be called in the context
1017 * of a thread already holding the token and thus not provide any
1018 * interlock protection. We cannot safely manipulate the td_toks
1019 * list for the same reason. Instead we depend on our critical
1020 * section if the token is owned by our cpu.
f1d1c3fa
MD
1021 */
1022 if ((w = td->td_wait) != NULL) {
d666840a 1023 spin_lock_wr(&w->wa_spinlock);
9d265729
MD
1024 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1025 --w->wa_count;
1026 td->td_wait = NULL;
d666840a 1027 spin_unlock_wr(&w->wa_spinlock);
0f7a3396 1028#ifdef SMP
9d265729 1029 if (td->td_gd == mygd) {
0f7a3396 1030 _lwkt_enqueue(td);
8ec60c3f 1031 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 1032 } else {
9d265729
MD
1033 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1034 }
b8a98473 1035#else
9d265729
MD
1036 _lwkt_enqueue(td);
1037 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
b8a98473 1038#endif
f1d1c3fa
MD
1039 } else {
1040 /*
1041 * If the wait structure is NULL and we own the thread, there
1042 * is no race (since we are in a critical section). If we
1043 * do not own the thread there might be a race but the
1044 * target cpu will deal with it.
1045 */
0f7a3396 1046#ifdef SMP
37af14fe 1047 if (td->td_gd == mygd) {
f1d1c3fa 1048 _lwkt_enqueue(td);
8ec60c3f 1049 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 1050 } else {
b8a98473 1051 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
f1d1c3fa 1052 }
0f7a3396
MD
1053#else
1054 _lwkt_enqueue(td);
8ec60c3f 1055 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 1056#endif
f1d1c3fa 1057 }
8ad65e08 1058 }
37af14fe 1059 crit_exit_gd(mygd);
8ad65e08
MD
1060}
1061
d9eea1a5
MD
1062/*
1063 * Managed acquisition. This code assumes that the MP lock is held for
1064 * the tdallq operation and that the thread has been descheduled from its
1065 * original cpu. We also have to wait for the thread to be entirely switched
1066 * out on its original cpu (this is usually fast enough that we never loop)
1067 * since the LWKT system does not have to hold the MP lock while switching
1068 * and the target may have released it before switching.
1069 */
a2a5ad0d
MD
1070void
1071lwkt_acquire(thread_t td)
1072{
37af14fe
MD
1073 globaldata_t gd;
1074 globaldata_t mygd;
a2a5ad0d
MD
1075
1076 gd = td->td_gd;
37af14fe 1077 mygd = mycpu;
35238fa5 1078 cpu_lfence();
a2a5ad0d 1079 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
c1102e9f 1080 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) /* XXX spin */
35238fa5 1081 cpu_lfence();
37af14fe
MD
1082 if (gd != mygd) {
1083 crit_enter_gd(mygd);
a2a5ad0d 1084 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
37af14fe
MD
1085 td->td_gd = mygd;
1086 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
1087 crit_exit_gd(mygd);
a2a5ad0d
MD
1088 }
1089}
1090
f1d1c3fa
MD
1091/*
1092 * Generic deschedule. Descheduling threads other then your own should be
1093 * done only in carefully controlled circumstances. Descheduling is
1094 * asynchronous.
1095 *
1096 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1097 */
1098void
1099lwkt_deschedule(thread_t td)
1100{
f1d1c3fa 1101 crit_enter();
b8a98473 1102#ifdef SMP
f1d1c3fa
MD
1103 if (td == curthread) {
1104 _lwkt_dequeue(td);
1105 } else {
a72187e9 1106 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1107 _lwkt_dequeue(td);
1108 } else {
b8a98473 1109 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1110 }
1111 }
b8a98473
MD
1112#else
1113 _lwkt_dequeue(td);
1114#endif
f1d1c3fa
MD
1115 crit_exit();
1116}
1117
4b5f931b
MD
1118/*
1119 * Set the target thread's priority. This routine does not automatically
1120 * switch to a higher priority thread, LWKT threads are not designed for
1121 * continuous priority changes. Yield if you want to switch.
1122 *
1123 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1124 * of the td_pri field. The specified priority may also indicate zero or
1125 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1126 *
1127 * Note that we requeue the thread whether it winds up on a different runq
1128 * or not. uio_yield() depends on this and the routine is not normally
1129 * called with the same priority otherwise.
4b5f931b
MD
1130 */
1131void
1132lwkt_setpri(thread_t td, int pri)
1133{
26a0694b 1134 KKASSERT(pri >= 0);
a72187e9 1135 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1136 crit_enter();
1137 if (td->td_flags & TDF_RUNQ) {
1138 _lwkt_dequeue(td);
1139 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1140 _lwkt_enqueue(td);
1141 } else {
1142 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1143 }
1144 crit_exit();
1145}
1146
1147void
1148lwkt_setpri_self(int pri)
1149{
1150 thread_t td = curthread;
1151
4b5f931b
MD
1152 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1153 crit_enter();
1154 if (td->td_flags & TDF_RUNQ) {
1155 _lwkt_dequeue(td);
1156 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1157 _lwkt_enqueue(td);
1158 } else {
1159 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1160 }
1161 crit_exit();
1162}
1163
8ec60c3f
MD
1164/*
1165 * Determine if there is a runnable thread at a higher priority then
1166 * the current thread. lwkt_setpri() does not check this automatically.
1167 * Return 1 if there is, 0 if there isn't.
1168 *
1169 * Example: if bit 31 of runqmask is set and the current thread is priority
1170 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1171 *
1172 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1173 * up comparing against 0xffffffff, a comparison that will always be false.
1174 */
1175int
1176lwkt_checkpri_self(void)
1177{
1178 globaldata_t gd = mycpu;
1179 thread_t td = gd->gd_curthread;
1180 int nq = td->td_pri & TDPRI_MASK;
1181
1182 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1183 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1184 return(1);
1185 ++nq;
1186 }
1187 return(0);
1188}
1189
5d21b981
MD
1190/*
1191 * Migrate the current thread to the specified cpu. The BGL must be held
1192 * (for the gd_tdallq manipulation XXX). This is accomplished by
1193 * descheduling ourselves from the current cpu, moving our thread to the
1194 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1195 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1196 */
3d28ff59 1197#ifdef SMP
5d21b981 1198static void lwkt_setcpu_remote(void *arg);
3d28ff59 1199#endif
5d21b981
MD
1200
1201void
1202lwkt_setcpu_self(globaldata_t rgd)
1203{
1204#ifdef SMP
1205 thread_t td = curthread;
1206
1207 if (td->td_gd != rgd) {
1208 crit_enter_quick(td);
1209 td->td_flags |= TDF_MIGRATING;
1210 lwkt_deschedule_self(td);
1211 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1212 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
b8a98473 1213 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1214 lwkt_switch();
1215 /* we are now on the target cpu */
1216 crit_exit_quick(td);
1217 }
1218#endif
1219}
1220
ecdefdda
MD
1221void
1222lwkt_migratecpu(int cpuid)
1223{
1224#ifdef SMP
1225 globaldata_t rgd;
1226
1227 rgd = globaldata_find(cpuid);
1228 lwkt_setcpu_self(rgd);
1229#endif
1230}
1231
5d21b981
MD
1232/*
1233 * Remote IPI for cpu migration (called while in a critical section so we
1234 * do not have to enter another one). The thread has already been moved to
1235 * our cpu's allq, but we must wait for the thread to be completely switched
1236 * out on the originating cpu before we schedule it on ours or the stack
1237 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1238 * change to main memory.
1239 *
1240 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1241 * against wakeups. It is best if this interface is used only when there
1242 * are no pending events that might try to schedule the thread.
1243 */
3d28ff59 1244#ifdef SMP
5d21b981
MD
1245static void
1246lwkt_setcpu_remote(void *arg)
1247{
1248 thread_t td = arg;
1249 globaldata_t gd = mycpu;
1250
c1102e9f 1251 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK))
35238fa5 1252 cpu_lfence();
5d21b981 1253 td->td_gd = gd;
35238fa5 1254 cpu_sfence();
5d21b981 1255 td->td_flags &= ~TDF_MIGRATING;
344ad853 1256 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
5d21b981
MD
1257 _lwkt_enqueue(td);
1258}
3d28ff59 1259#endif
5d21b981 1260
553ea3c8 1261struct lwp *
4b5f931b
MD
1262lwkt_preempted_proc(void)
1263{
73e4f7b9 1264 thread_t td = curthread;
4b5f931b
MD
1265 while (td->td_preempted)
1266 td = td->td_preempted;
553ea3c8 1267 return(td->td_lwp);
4b5f931b
MD
1268}
1269
f1d1c3fa 1270/*
41a01a4d
MD
1271 * Block on the specified wait queue until signaled. A generation number
1272 * must be supplied to interlock the wait queue. The function will
1273 * return immediately if the generation number does not match the wait
1274 * structure's generation number.
f1d1c3fa
MD
1275 */
1276void
ae8050a4 1277lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
1278{
1279 thread_t td = curthread;
f1d1c3fa 1280
d666840a 1281 spin_lock_wr(&w->wa_spinlock);
ae8050a4 1282 if (w->wa_gen == *gen) {
f1d1c3fa 1283 _lwkt_dequeue(td);
344ad853 1284 td->td_flags |= TDF_BLOCKQ;
f1d1c3fa
MD
1285 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1286 ++w->wa_count;
1287 td->td_wait = w;
ae8050a4 1288 td->td_wmesg = wmesg;
d666840a 1289 spin_unlock_wr(&w->wa_spinlock);
f1d1c3fa 1290 lwkt_switch();
344ad853
MD
1291 KKASSERT((td->td_flags & TDF_BLOCKQ) == 0);
1292 td->td_wmesg = NULL;
9d265729
MD
1293 *gen = w->wa_gen;
1294 } else {
1295 *gen = w->wa_gen;
d666840a 1296 spin_unlock_wr(&w->wa_spinlock);
8ad65e08 1297 }
f1d1c3fa
MD
1298}
1299
1300/*
1301 * Signal a wait queue. We gain ownership of the wait queue in order to
1302 * signal it. Once a thread is removed from the wait queue we have to
1303 * deal with the cpu owning the thread.
1304 *
1305 * Note: alternatively we could message the target cpu owning the wait
1306 * queue. YYY implement as sysctl.
1307 */
1308void
ece04fd0 1309lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1310{
1311 thread_t td;
f1d1c3fa 1312
d666840a 1313 spin_lock_wr(&w->wa_spinlock);
f1d1c3fa 1314 ++w->wa_gen;
ece04fd0
MD
1315 if (count < 0)
1316 count = w->wa_count;
f1d1c3fa
MD
1317 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1318 --count;
1319 --w->wa_count;
344ad853 1320 KKASSERT(td->td_flags & TDF_BLOCKQ);
f1d1c3fa 1321 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
344ad853 1322 td->td_flags &= ~TDF_BLOCKQ;
f1d1c3fa 1323 td->td_wait = NULL;
d666840a 1324 spin_unlock_wr(&w->wa_spinlock);
344ad853 1325 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
b8a98473 1326#ifdef SMP
a72187e9 1327 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1328 _lwkt_enqueue(td);
1329 } else {
b8a98473 1330 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
f1d1c3fa 1331 }
b8a98473
MD
1332#else
1333 _lwkt_enqueue(td);
1334#endif
d666840a 1335 spin_lock_wr(&w->wa_spinlock);
f1d1c3fa 1336 }
d666840a 1337 spin_unlock_wr(&w->wa_spinlock);
f1d1c3fa
MD
1338}
1339
99df837e
MD
1340/*
1341 * Create a kernel process/thread/whatever. It shares it's address space
1342 * with proc0 - ie: kernel only.
1343 *
365fa13f
MD
1344 * NOTE! By default new threads are created with the MP lock held. A
1345 * thread which does not require the MP lock should release it by calling
1346 * rel_mplock() at the start of the new thread.
99df837e
MD
1347 */
1348int
1349lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1350 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1351 const char *fmt, ...)
99df837e 1352{
73e4f7b9 1353 thread_t td;
e2565a42 1354 __va_list ap;
99df837e 1355
d3d32139
MD
1356 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1357 tdflags | TDF_VERBOSE);
a2a5ad0d
MD
1358 if (tdp)
1359 *tdp = td;
709799ea 1360 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1361
1362 /*
1363 * Set up arg0 for 'ps' etc
1364 */
e2565a42 1365 __va_start(ap, fmt);
99df837e 1366 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1367 __va_end(ap);
99df837e
MD
1368
1369 /*
1370 * Schedule the thread to run
1371 */
ef0fdad1
MD
1372 if ((td->td_flags & TDF_STOPREQ) == 0)
1373 lwkt_schedule(td);
1374 else
1375 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1376 return 0;
1377}
1378
2d93b37a 1379/*
2d93b37a
MD
1380 * kthread_* is specific to the kernel and is not needed by userland.
1381 */
1382#ifdef _KERNEL
1383
99df837e
MD
1384/*
1385 * Destroy an LWKT thread. Warning! This function is not called when
1386 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1387 * uses a different reaping mechanism.
1388 */
1389void
1390lwkt_exit(void)
1391{
1392 thread_t td = curthread;
8826f33a 1393 globaldata_t gd;
99df837e
MD
1394
1395 if (td->td_flags & TDF_VERBOSE)
1396 printf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1397 caps_exit(td);
37af14fe
MD
1398 crit_enter_quick(td);
1399 lwkt_deschedule_self(td);
8826f33a
MD
1400 gd = mycpu;
1401 KKASSERT(gd == td->td_gd);
1402 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1403 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1404 ++gd->gd_tdfreecount;
1405 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1406 }
99df837e
MD
1407 cpu_thread_exit();
1408}
1409
2d93b37a
MD
1410#endif /* _KERNEL */
1411
1412void
1413crit_panic(void)
1414{
1415 thread_t td = curthread;
1416 int lpri = td->td_pri;
1417
1418 td->td_pri = 0;
1419 panic("td_pri is/would-go negative! %p %d", td, lpri);
1420}
1421
d165e668
MD
1422#ifdef SMP
1423
bd8015ca
MD
1424/*
1425 * Called from debugger/panic on cpus which have been stopped. We must still
1426 * process the IPIQ while stopped, even if we were stopped while in a critical
1427 * section (XXX).
1428 *
1429 * If we are dumping also try to process any pending interrupts. This may
1430 * or may not work depending on the state of the cpu at the point it was
1431 * stopped.
1432 */
1433void
1434lwkt_smp_stopped(void)
1435{
1436 globaldata_t gd = mycpu;
1437
1438 crit_enter_gd(gd);
1439 if (dumping) {
1440 lwkt_process_ipiq();
1441 splz();
1442 } else {
1443 lwkt_process_ipiq();
1444 }
1445 crit_exit_gd(gd);
1446}
1447
d165e668 1448#endif