These files have been moved to bus/usb.
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08
MD
1/*
2 * Copyright (c) 2003 Matthew Dillon <dillon@backplane.com>
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 *
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24 * SUCH DAMAGE.
25 *
709799ea 26 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.46 2003/12/04 20:09:33 dillon Exp $
75cdbe6c
MD
27 */
28
29/*
30 * Each cpu in a system has its own self-contained light weight kernel
31 * thread scheduler, which means that generally speaking we only need
32 * to use a critical section to avoid problems. Foreign thread
33 * scheduling is queued via (async) IPIs.
f1d1c3fa 34 *
75cdbe6c
MD
35 * NOTE: on UP machines smp_active is defined to be 0. On SMP machines
36 * smp_active is 0 prior to SMP activation, then it is 1. The LWKT module
37 * uses smp_active to optimize UP builds and to avoid sending IPIs during
38 * early boot (primarily interrupt and network thread initialization).
8ad65e08
MD
39 */
40
05220613
MD
41#ifdef _KERNEL
42
8ad65e08
MD
43#include <sys/param.h>
44#include <sys/systm.h>
45#include <sys/kernel.h>
46#include <sys/proc.h>
47#include <sys/rtprio.h>
48#include <sys/queue.h>
f1d1c3fa 49#include <sys/thread2.h>
7d0bac62 50#include <sys/sysctl.h>
99df837e 51#include <sys/kthread.h>
f1d1c3fa 52#include <machine/cpu.h>
99df837e 53#include <sys/lock.h>
f1d1c3fa 54
7d0bac62
MD
55#include <vm/vm.h>
56#include <vm/vm_param.h>
57#include <vm/vm_kern.h>
58#include <vm/vm_object.h>
59#include <vm/vm_page.h>
60#include <vm/vm_map.h>
61#include <vm/vm_pager.h>
62#include <vm/vm_extern.h>
63#include <vm/vm_zone.h>
64
99df837e 65#include <machine/stdarg.h>
57c254db 66#include <machine/ipl.h>
96728c05 67#include <machine/smp.h>
99df837e 68
05220613
MD
69#define THREAD_STACK (UPAGES * PAGE_SIZE)
70
71#else
72
73#include <sys/stdint.h>
fb04f4fd 74#include <libcaps/thread.h>
05220613
MD
75#include <sys/thread.h>
76#include <sys/msgport.h>
77#include <sys/errno.h>
fb04f4fd 78#include <libcaps/globaldata.h>
05220613
MD
79#include <sys/thread2.h>
80#include <sys/msgport2.h>
709799ea 81#include <stdio.h>
05220613 82#include <stdlib.h>
709799ea 83#include <string.h>
c95cd171 84#include <machine/cpufunc.h>
709799ea 85#include <machine/lock.h>
05220613
MD
86
87#endif
88
7d0bac62 89static int untimely_switch = 0;
57c254db
MD
90#ifdef INVARIANTS
91static int token_debug = 0;
05220613
MD
92#endif
93static __int64_t switch_count = 0;
94static __int64_t preempt_hit = 0;
95static __int64_t preempt_miss = 0;
96static __int64_t preempt_weird = 0;
709799ea 97#ifdef SMP
05220613
MD
98static __int64_t ipiq_count = 0;
99static __int64_t ipiq_fifofull = 0;
709799ea 100#endif
05220613
MD
101
102#ifdef _KERNEL
103
104SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
105#ifdef INVARIANTS
57c254db
MD
106SYSCTL_INT(_lwkt, OID_AUTO, token_debug, CTLFLAG_RW, &token_debug, 0, "");
107#endif
4b5f931b 108SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 109SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 110SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 111SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
709799ea 112#ifdef SMP
96728c05 113SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
96728c05 114SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
709799ea 115#endif
7d0bac62 116
05220613
MD
117#endif
118
4b5f931b
MD
119/*
120 * These helper procedures handle the runq, they can only be called from
121 * within a critical section.
75cdbe6c
MD
122 *
123 * WARNING! Prior to SMP being brought up it is possible to enqueue and
124 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
125 * instead of 'mycpu' when referencing the globaldata structure. Once
126 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 127 */
f1d1c3fa
MD
128static __inline
129void
130_lwkt_dequeue(thread_t td)
131{
132 if (td->td_flags & TDF_RUNQ) {
4b5f931b 133 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 134 struct globaldata *gd = td->td_gd;
4b5f931b 135
f1d1c3fa 136 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
137 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
138 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
139 }
140}
141
142static __inline
143void
144_lwkt_enqueue(thread_t td)
145{
146 if ((td->td_flags & TDF_RUNQ) == 0) {
4b5f931b 147 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 148 struct globaldata *gd = td->td_gd;
4b5f931b 149
f1d1c3fa 150 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
151 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
152 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
153 }
154}
8ad65e08 155
57c254db
MD
156static __inline
157int
158_lwkt_wantresched(thread_t ntd, thread_t cur)
159{
160 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
161}
162
2d93b37a
MD
163#ifdef _KERNEL
164
8ad65e08
MD
165/*
166 * LWKTs operate on a per-cpu basis
167 *
73e4f7b9 168 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
169 */
170void
171lwkt_gdinit(struct globaldata *gd)
172{
4b5f931b
MD
173 int i;
174
175 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
176 TAILQ_INIT(&gd->gd_tdrunq[i]);
177 gd->gd_runqmask = 0;
73e4f7b9 178 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
179}
180
2d93b37a
MD
181#endif /* _KERNEL */
182
7d0bac62
MD
183/*
184 * Initialize a thread wait structure prior to first use.
185 *
186 * NOTE! called from low level boot code, we cannot do anything fancy!
187 */
188void
189lwkt_init_wait(lwkt_wait_t w)
190{
191 TAILQ_INIT(&w->wa_waitq);
192}
193
194/*
195 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
196 * or LWKT start address before it can be scheduled. If the target cpu is
197 * -1 the thread will be created on the current cpu.
0cfcada1
MD
198 *
199 * If you intend to create a thread without a process context this function
200 * does everything except load the startup and switcher function.
7d0bac62
MD
201 */
202thread_t
75cdbe6c 203lwkt_alloc_thread(struct thread *td, int cpu)
7d0bac62 204{
99df837e 205 void *stack;
ef0fdad1 206 int flags = 0;
7d0bac62 207
ef0fdad1 208 if (td == NULL) {
26a0694b 209 crit_enter();
ef0fdad1
MD
210 if (mycpu->gd_tdfreecount > 0) {
211 --mycpu->gd_tdfreecount;
212 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
d9eea1a5 213 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1
MD
214 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
215 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
216 crit_exit();
217 stack = td->td_kstack;
218 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
219 } else {
220 crit_exit();
05220613 221#ifdef _KERNEL
ef0fdad1 222 td = zalloc(thread_zone);
05220613
MD
223#else
224 td = malloc(sizeof(struct thread));
225#endif
ef0fdad1
MD
226 td->td_kstack = NULL;
227 flags |= TDF_ALLOCATED_THREAD;
228 }
229 }
230 if ((stack = td->td_kstack) == NULL) {
05220613
MD
231#ifdef _KERNEL
232 stack = (void *)kmem_alloc(kernel_map, THREAD_STACK);
233#else
fb04f4fd 234 stack = libcaps_alloc_stack(THREAD_STACK);
05220613 235#endif
ef0fdad1 236 flags |= TDF_ALLOCATED_STACK;
99df837e 237 }
75cdbe6c
MD
238 if (cpu < 0)
239 lwkt_init_thread(td, stack, flags, mycpu);
240 else
241 lwkt_init_thread(td, stack, flags, globaldata_find(cpu));
99df837e 242 return(td);
7d0bac62
MD
243}
244
709799ea
MD
245#ifdef _KERNEL
246
7d0bac62
MD
247/*
248 * Initialize a preexisting thread structure. This function is used by
249 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
250 *
f8c3996b
MD
251 * All threads start out in a critical section at a priority of
252 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
253 * appropriate. This function may send an IPI message when the
254 * requested cpu is not the current cpu and consequently gd_tdallq may
255 * not be initialized synchronously from the point of view of the originating
256 * cpu.
257 *
258 * NOTE! we have to be careful in regards to creating threads for other cpus
259 * if SMP has not yet been activated.
7d0bac62 260 */
75cdbe6c
MD
261static void
262lwkt_init_thread_remote(void *arg)
263{
264 thread_t td = arg;
265
266 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
267}
268
7d0bac62 269void
26a0694b 270lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
7d0bac62 271{
99df837e
MD
272 bzero(td, sizeof(struct thread));
273 td->td_kstack = stack;
274 td->td_flags |= flags;
26a0694b 275 td->td_gd = gd;
f8c3996b 276 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
c95cd171 277 lwkt_initport(&td->td_msgport, td);
99df837e 278 pmap_init_thread(td);
75cdbe6c
MD
279 if (smp_active == 0 || gd == mycpu) {
280 crit_enter();
281 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
282 crit_exit();
283 } else {
284 lwkt_send_ipiq(gd->gd_cpuid, lwkt_init_thread_remote, td);
285 }
73e4f7b9
MD
286}
287
2d93b37a
MD
288#endif /* _KERNEL */
289
73e4f7b9
MD
290void
291lwkt_set_comm(thread_t td, const char *ctl, ...)
292{
e2565a42 293 __va_list va;
73e4f7b9 294
e2565a42 295 __va_start(va, ctl);
73e4f7b9 296 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 297 __va_end(va);
7d0bac62
MD
298}
299
99df837e 300void
73e4f7b9 301lwkt_hold(thread_t td)
99df837e 302{
73e4f7b9
MD
303 ++td->td_refs;
304}
305
306void
307lwkt_rele(thread_t td)
308{
309 KKASSERT(td->td_refs > 0);
310 --td->td_refs;
311}
312
c95cd171
MD
313#ifdef _KERNEL
314
73e4f7b9
MD
315void
316lwkt_wait_free(thread_t td)
317{
318 while (td->td_refs)
377d4740 319 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
320}
321
c95cd171
MD
322#endif
323
73e4f7b9
MD
324void
325lwkt_free_thread(thread_t td)
326{
327 struct globaldata *gd = mycpu;
328
d9eea1a5 329 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
330 ("lwkt_free_thread: did not exit! %p", td));
331
332 crit_enter();
73e4f7b9
MD
333 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
334 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
335 (td->td_flags & TDF_ALLOCATED_THREAD)
336 ) {
73e4f7b9
MD
337 ++gd->gd_tdfreecount;
338 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
99df837e
MD
339 crit_exit();
340 } else {
341 crit_exit();
342 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613
MD
343#ifdef _KERNEL
344 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK);
345#else
fb04f4fd 346 libcaps_free_stack(td->td_kstack, THREAD_STACK);
05220613 347#endif
73e4f7b9 348 /* gd invalid */
99df837e
MD
349 td->td_kstack = NULL;
350 }
05220613
MD
351 if (td->td_flags & TDF_ALLOCATED_THREAD) {
352#ifdef _KERNEL
99df837e 353 zfree(thread_zone, td);
05220613
MD
354#else
355 free(td);
356#endif
357 }
99df837e
MD
358 }
359}
360
361
8ad65e08
MD
362/*
363 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
364 * switch to the idlethread. Switching must occur within a critical
365 * section to avoid races with the scheduling queue.
366 *
367 * We always have full control over our cpu's run queue. Other cpus
368 * that wish to manipulate our queue must use the cpu_*msg() calls to
369 * talk to our cpu, so a critical section is all that is needed and
370 * the result is very, very fast thread switching.
371 *
96728c05
MD
372 * The LWKT scheduler uses a fixed priority model and round-robins at
373 * each priority level. User process scheduling is a totally
374 * different beast and LWKT priorities should not be confused with
375 * user process priorities.
f1d1c3fa 376 *
96728c05
MD
377 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
378 * cleans it up. Note that the td_switch() function cannot do anything that
379 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
380 * the target thread (not the current thread). It's nice to have a scheduler
381 * that does not need the MP lock to work because it allows us to do some
382 * really cool high-performance MP lock optimizations.
8ad65e08 383 */
96728c05 384
8ad65e08
MD
385void
386lwkt_switch(void)
387{
4b5f931b 388 struct globaldata *gd;
f1d1c3fa 389 thread_t td = curthread;
8ad65e08 390 thread_t ntd;
8a8d5d85
MD
391#ifdef SMP
392 int mpheld;
393#endif
8ad65e08 394
46a3f46d
MD
395 /*
396 * Switching from within a 'fast' (non thread switched) interrupt is
397 * illegal.
398 */
399 if (mycpu->gd_intr_nesting_level && panicstr == NULL) {
03aa8d99 400 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
96728c05 401 }
ef0fdad1 402
cb973d15
MD
403 /*
404 * Passive release (used to transition from user to kernel mode
405 * when we block or switch rather then when we enter the kernel).
406 * This function is NOT called if we are switching into a preemption
407 * or returning from a preemption. Typically this causes us to lose
408 * our P_CURPROC designation (if we have one) and become a true LWKT
409 * thread, and may also hand P_CURPROC to another process and schedule
410 * its thread.
411 */
412 if (td->td_release)
413 td->td_release(td);
414
f1d1c3fa 415 crit_enter();
4b5f931b 416 ++switch_count;
8a8d5d85
MD
417
418#ifdef SMP
419 /*
420 * td_mpcount cannot be used to determine if we currently hold the
421 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
422 * to get the lock, and switch out if it can't. Our ownership of
423 * the actual lock will remain stable while we are in a critical section
424 * (but, of course, another cpu may own or release the lock so the
425 * actual value of mp_lock is not stable).
8a8d5d85
MD
426 */
427 mpheld = MP_LOCK_HELD();
428#endif
99df837e
MD
429 if ((ntd = td->td_preempted) != NULL) {
430 /*
431 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
432 * thread. This occurs transparently, whether the preempted thread
433 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
434 * itself).
435 *
436 * We have to setup the MP lock for the original thread after backing
437 * out the adjustment that was made to curthread when the original
438 * was preempted.
99df837e 439 */
26a0694b 440 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 441#ifdef SMP
96728c05
MD
442 if (ntd->td_mpcount && mpheld == 0) {
443 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
444 td, ntd, td->td_mpcount, ntd->td_mpcount);
445 }
8a8d5d85
MD
446 if (ntd->td_mpcount) {
447 td->td_mpcount -= ntd->td_mpcount;
448 KKASSERT(td->td_mpcount >= 0);
449 }
450#endif
26a0694b 451 ntd->td_flags |= TDF_PREEMPT_DONE;
8a8d5d85 452 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 453 } else {
4b5f931b
MD
454 /*
455 * Priority queue / round-robin at each priority. Note that user
456 * processes run at a fixed, low priority and the user process
457 * scheduler deals with interactions between user processes
458 * by scheduling and descheduling them from the LWKT queue as
459 * necessary.
8a8d5d85
MD
460 *
461 * We have to adjust the MP lock for the target thread. If we
462 * need the MP lock and cannot obtain it we try to locate a
463 * thread that does not need the MP lock.
4b5f931b
MD
464 */
465 gd = mycpu;
4b5f931b
MD
466again:
467 if (gd->gd_runqmask) {
468 int nq = bsrl(gd->gd_runqmask);
469 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
470 gd->gd_runqmask &= ~(1 << nq);
471 goto again;
472 }
8a8d5d85
MD
473#ifdef SMP
474 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
475 /*
96728c05
MD
476 * Target needs MP lock and we couldn't get it, try
477 * to locate a thread which does not need the MP lock
3c23a41a 478 * to run. If we cannot locate a thread spin in idle.
8a8d5d85
MD
479 */
480 u_int32_t rqmask = gd->gd_runqmask;
481 while (rqmask) {
482 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
483 if (ntd->td_mpcount == 0)
484 break;
485 }
486 if (ntd)
487 break;
488 rqmask &= ~(1 << nq);
489 nq = bsrl(rqmask);
490 }
491 if (ntd == NULL) {
a2a5ad0d
MD
492 ntd = &gd->gd_idlethread;
493 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
494 } else {
495 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
496 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
497 }
498 } else {
499 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
500 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
501 }
502#else
4b5f931b
MD
503 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
504 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 505#endif
4b5f931b 506 } else {
3c23a41a
MD
507 /*
508 * Nothing to run but we may still need the BGL to deal with
509 * pending interrupts, spin in idle if so.
510 */
a2a5ad0d 511 ntd = &gd->gd_idlethread;
235957ed 512 if (gd->gd_reqflags)
3c23a41a 513 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 514 }
f1d1c3fa 515 }
26a0694b
MD
516 KASSERT(ntd->td_pri >= TDPRI_CRIT,
517 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
518
519 /*
520 * Do the actual switch. If the new target does not need the MP lock
521 * and we are holding it, release the MP lock. If the new target requires
522 * the MP lock we have already acquired it for the target.
523 */
524#ifdef SMP
525 if (ntd->td_mpcount == 0 ) {
526 if (MP_LOCK_HELD())
527 cpu_rel_mplock();
528 } else {
529 ASSERT_MP_LOCK_HELD();
530 }
531#endif
8a8d5d85 532 if (td != ntd) {
f1d1c3fa 533 td->td_switch(ntd);
8a8d5d85 534 }
96728c05 535
f1d1c3fa 536 crit_exit();
8ad65e08
MD
537}
538
cb973d15
MD
539/*
540 * Switch if another thread has a higher priority. Do not switch to other
541 * threads at the same priority.
542 */
543void
544lwkt_maybe_switch()
545{
546 struct globaldata *gd = mycpu;
547 struct thread *td = gd->gd_curthread;
548
549 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
550 lwkt_switch();
551 }
552}
553
b68b7282 554/*
96728c05
MD
555 * Request that the target thread preempt the current thread. Preemption
556 * only works under a specific set of conditions:
b68b7282 557 *
96728c05
MD
558 * - We are not preempting ourselves
559 * - The target thread is owned by the current cpu
560 * - We are not currently being preempted
561 * - The target is not currently being preempted
562 * - We are able to satisfy the target's MP lock requirements (if any).
563 *
564 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
565 * this is called via lwkt_schedule() through the td_preemptable callback.
566 * critpri is the managed critical priority that we should ignore in order
567 * to determine whether preemption is possible (aka usually just the crit
568 * priority of lwkt_schedule() itself).
b68b7282 569 *
26a0694b
MD
570 * XXX at the moment we run the target thread in a critical section during
571 * the preemption in order to prevent the target from taking interrupts
572 * that *WE* can't. Preemption is strictly limited to interrupt threads
573 * and interrupt-like threads, outside of a critical section, and the
574 * preempted source thread will be resumed the instant the target blocks
575 * whether or not the source is scheduled (i.e. preemption is supposed to
576 * be as transparent as possible).
4b5f931b 577 *
8a8d5d85
MD
578 * The target thread inherits our MP count (added to its own) for the
579 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
580 * MP lock during the preemption. Therefore, any preempting targets must be
581 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
582 * out of sync with the physical mp_lock, but we do not have to preserve
583 * the original ownership of the lock if it was out of synch (that is, we
584 * can leave it synchronized on return).
b68b7282
MD
585 */
586void
96728c05 587lwkt_preempt(thread_t ntd, int critpri)
b68b7282 588{
46a3f46d
MD
589 struct globaldata *gd = mycpu;
590 thread_t td = gd->gd_curthread;
8a8d5d85
MD
591#ifdef SMP
592 int mpheld;
57c254db 593 int savecnt;
8a8d5d85 594#endif
b68b7282 595
26a0694b 596 /*
96728c05
MD
597 * The caller has put us in a critical section. We can only preempt
598 * if the caller of the caller was not in a critical section (basically
57c254db
MD
599 * a local interrupt), as determined by the 'critpri' parameter. If
600 * we are unable to preempt
96728c05
MD
601 *
602 * YYY The target thread must be in a critical section (else it must
603 * inherit our critical section? I dunno yet).
26a0694b
MD
604 */
605 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 606
cb973d15 607 need_resched();
57c254db
MD
608 if (!_lwkt_wantresched(ntd, td)) {
609 ++preempt_miss;
610 return;
611 }
96728c05
MD
612 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
613 ++preempt_miss;
614 return;
615 }
616#ifdef SMP
46a3f46d 617 if (ntd->td_gd != gd) {
96728c05
MD
618 ++preempt_miss;
619 return;
620 }
621#endif
26a0694b
MD
622 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
623 ++preempt_weird;
624 return;
625 }
626 if (ntd->td_preempted) {
4b5f931b 627 ++preempt_hit;
26a0694b 628 return;
b68b7282 629 }
8a8d5d85 630#ifdef SMP
a2a5ad0d
MD
631 /*
632 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
633 * to or from the MP lock. In this case td_mpcount will be pre-disposed
634 * (non-zero) but not actually synchronized with the actual state of the
635 * lock. We can use it to imply an MP lock requirement for the
636 * preemption but we cannot use it to test whether we hold the MP lock
637 * or not.
a2a5ad0d 638 */
96728c05 639 savecnt = td->td_mpcount;
71ef2f5c 640 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
641 ntd->td_mpcount += td->td_mpcount;
642 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
643 ntd->td_mpcount -= td->td_mpcount;
644 ++preempt_miss;
645 return;
646 }
647#endif
26a0694b
MD
648
649 ++preempt_hit;
650 ntd->td_preempted = td;
651 td->td_flags |= TDF_PREEMPT_LOCK;
652 td->td_switch(ntd);
653 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
654#ifdef SMP
655 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
656 mpheld = MP_LOCK_HELD();
657 if (mpheld && td->td_mpcount == 0)
96728c05 658 cpu_rel_mplock();
71ef2f5c 659 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
660 panic("lwkt_preempt(): MP lock was not held through");
661#endif
26a0694b
MD
662 ntd->td_preempted = NULL;
663 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
664}
665
f1d1c3fa
MD
666/*
667 * Yield our thread while higher priority threads are pending. This is
668 * typically called when we leave a critical section but it can be safely
669 * called while we are in a critical section.
670 *
671 * This function will not generally yield to equal priority threads but it
672 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 673 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
674 * lwkt_yield_quick().
675 *
235957ed 676 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
677 * came along but was blocked and made pending.
678 *
f1d1c3fa
MD
679 * (self contained on a per cpu basis)
680 */
681void
682lwkt_yield_quick(void)
683{
7966cb69
MD
684 globaldata_t gd = mycpu;
685 thread_t td = gd->gd_curthread;
ef0fdad1 686
a2a5ad0d 687 /*
235957ed 688 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
689 * it with a non-zero cpl then we might not wind up calling splz after
690 * a task switch when the critical section is exited even though the
46a3f46d 691 * new task could accept the interrupt.
a2a5ad0d
MD
692 *
693 * XXX from crit_exit() only called after last crit section is released.
694 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
695 *
696 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
697 * except for this special case, we MUST call splz() here to handle any
698 * pending ints, particularly after we switch, or we might accidently
699 * halt the cpu with interrupts pending.
a2a5ad0d 700 */
46a3f46d 701 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 702 splz();
f1d1c3fa
MD
703
704 /*
705 * YYY enabling will cause wakeup() to task-switch, which really
706 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
707 * preemption and MP without actually doing preemption or MP, because a
708 * lot of code assumes that wakeup() does not block.
f1d1c3fa 709 */
46a3f46d
MD
710 if (untimely_switch && td->td_nest_count == 0 &&
711 gd->gd_intr_nesting_level == 0
712 ) {
f1d1c3fa
MD
713 crit_enter();
714 /*
715 * YYY temporary hacks until we disassociate the userland scheduler
716 * from the LWKT scheduler.
717 */
718 if (td->td_flags & TDF_RUNQ) {
719 lwkt_switch(); /* will not reenter yield function */
720 } else {
721 lwkt_schedule_self(); /* make sure we are scheduled */
722 lwkt_switch(); /* will not reenter yield function */
723 lwkt_deschedule_self(); /* make sure we are descheduled */
724 }
7966cb69 725 crit_exit_noyield(td);
f1d1c3fa 726 }
f1d1c3fa
MD
727}
728
8ad65e08 729/*
f1d1c3fa 730 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 731 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
732 * the crit_exit() call in lwkt_switch().
733 *
734 * (self contained on a per cpu basis)
8ad65e08
MD
735 */
736void
f1d1c3fa 737lwkt_yield(void)
8ad65e08 738{
f1d1c3fa
MD
739 lwkt_schedule_self();
740 lwkt_switch();
741}
742
743/*
744 * Schedule a thread to run. As the current thread we can always safely
745 * schedule ourselves, and a shortcut procedure is provided for that
746 * function.
747 *
748 * (non-blocking, self contained on a per cpu basis)
749 */
750void
751lwkt_schedule_self(void)
752{
753 thread_t td = curthread;
754
755 crit_enter();
756 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa 757 _lwkt_enqueue(td);
05220613 758#ifdef _KERNEL
26a0694b
MD
759 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
760 panic("SCHED SELF PANIC");
05220613 761#endif
f1d1c3fa 762 crit_exit();
8ad65e08 763}
8ad65e08
MD
764
765/*
f1d1c3fa
MD
766 * Generic schedule. Possibly schedule threads belonging to other cpus and
767 * deal with threads that might be blocked on a wait queue.
768 *
96728c05 769 * YYY this is one of the best places to implement load balancing code.
f1d1c3fa
MD
770 * Load balancing can be accomplished by requesting other sorts of actions
771 * for the thread in question.
8ad65e08
MD
772 */
773void
774lwkt_schedule(thread_t td)
775{
96728c05 776#ifdef INVARIANTS
26a0694b
MD
777 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
778 && td->td_proc->p_stat == SSLEEP
779 ) {
780 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
781 curthread,
782 curthread->td_proc ? curthread->td_proc->p_pid : -1,
783 curthread->td_proc ? curthread->td_proc->p_stat : -1,
784 td,
785 td->td_proc ? curthread->td_proc->p_pid : -1,
786 td->td_proc ? curthread->td_proc->p_stat : -1
787 );
788 panic("SCHED PANIC");
789 }
96728c05 790#endif
f1d1c3fa
MD
791 crit_enter();
792 if (td == curthread) {
793 _lwkt_enqueue(td);
794 } else {
795 lwkt_wait_t w;
796
797 /*
798 * If the thread is on a wait list we have to send our scheduling
799 * request to the owner of the wait structure. Otherwise we send
800 * the scheduling request to the cpu owning the thread. Races
801 * are ok, the target will forward the message as necessary (the
802 * message may chase the thread around before it finally gets
803 * acted upon).
804 *
805 * (remember, wait structures use stable storage)
806 */
807 if ((w = td->td_wait) != NULL) {
96728c05 808 if (lwkt_trytoken(&w->wa_token)) {
f1d1c3fa
MD
809 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
810 --w->wa_count;
811 td->td_wait = NULL;
75cdbe6c 812 if (smp_active == 0 || td->td_gd == mycpu) {
f1d1c3fa 813 _lwkt_enqueue(td);
57c254db 814 if (td->td_preemptable) {
96728c05 815 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
57c254db
MD
816 } else if (_lwkt_wantresched(td, curthread)) {
817 need_resched();
818 }
f1d1c3fa 819 } else {
a72187e9 820 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 821 }
96728c05 822 lwkt_reltoken(&w->wa_token);
f1d1c3fa 823 } else {
96728c05 824 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
825 }
826 } else {
827 /*
828 * If the wait structure is NULL and we own the thread, there
829 * is no race (since we are in a critical section). If we
830 * do not own the thread there might be a race but the
831 * target cpu will deal with it.
832 */
75cdbe6c 833 if (smp_active == 0 || td->td_gd == mycpu) {
f1d1c3fa 834 _lwkt_enqueue(td);
57c254db 835 if (td->td_preemptable) {
96728c05 836 td->td_preemptable(td, TDPRI_CRIT);
57c254db
MD
837 } else if (_lwkt_wantresched(td, curthread)) {
838 need_resched();
839 }
f1d1c3fa 840 } else {
a72187e9 841 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
842 }
843 }
8ad65e08 844 }
f1d1c3fa 845 crit_exit();
8ad65e08
MD
846}
847
d9eea1a5
MD
848/*
849 * Managed acquisition. This code assumes that the MP lock is held for
850 * the tdallq operation and that the thread has been descheduled from its
851 * original cpu. We also have to wait for the thread to be entirely switched
852 * out on its original cpu (this is usually fast enough that we never loop)
853 * since the LWKT system does not have to hold the MP lock while switching
854 * and the target may have released it before switching.
855 */
a2a5ad0d
MD
856void
857lwkt_acquire(thread_t td)
858{
859 struct globaldata *gd;
860
861 gd = td->td_gd;
862 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5
MD
863 while (td->td_flags & TDF_RUNNING) /* XXX spin */
864 ;
a2a5ad0d
MD
865 if (gd != mycpu) {
866 crit_enter();
867 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
868 gd = mycpu;
869 td->td_gd = gd;
a2a5ad0d
MD
870 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
871 crit_exit();
872 }
873}
874
8ad65e08 875/*
f1d1c3fa
MD
876 * Deschedule a thread.
877 *
878 * (non-blocking, self contained on a per cpu basis)
879 */
880void
881lwkt_deschedule_self(void)
882{
883 thread_t td = curthread;
884
885 crit_enter();
886 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa
MD
887 _lwkt_dequeue(td);
888 crit_exit();
889}
890
891/*
892 * Generic deschedule. Descheduling threads other then your own should be
893 * done only in carefully controlled circumstances. Descheduling is
894 * asynchronous.
895 *
896 * This function may block if the cpu has run out of messages.
8ad65e08
MD
897 */
898void
899lwkt_deschedule(thread_t td)
900{
f1d1c3fa
MD
901 crit_enter();
902 if (td == curthread) {
903 _lwkt_dequeue(td);
904 } else {
a72187e9 905 if (td->td_gd == mycpu) {
f1d1c3fa
MD
906 _lwkt_dequeue(td);
907 } else {
a72187e9 908 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
909 }
910 }
911 crit_exit();
912}
913
4b5f931b
MD
914/*
915 * Set the target thread's priority. This routine does not automatically
916 * switch to a higher priority thread, LWKT threads are not designed for
917 * continuous priority changes. Yield if you want to switch.
918 *
919 * We have to retain the critical section count which uses the high bits
26a0694b
MD
920 * of the td_pri field. The specified priority may also indicate zero or
921 * more critical sections by adding TDPRI_CRIT*N.
4b5f931b
MD
922 */
923void
924lwkt_setpri(thread_t td, int pri)
925{
26a0694b 926 KKASSERT(pri >= 0);
a72187e9 927 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
928 crit_enter();
929 if (td->td_flags & TDF_RUNQ) {
930 _lwkt_dequeue(td);
931 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
932 _lwkt_enqueue(td);
933 } else {
934 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
935 }
936 crit_exit();
937}
938
939void
940lwkt_setpri_self(int pri)
941{
942 thread_t td = curthread;
943
4b5f931b
MD
944 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
945 crit_enter();
946 if (td->td_flags & TDF_RUNQ) {
947 _lwkt_dequeue(td);
948 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
949 _lwkt_enqueue(td);
950 } else {
951 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
952 }
953 crit_exit();
954}
955
956struct proc *
957lwkt_preempted_proc(void)
958{
73e4f7b9 959 thread_t td = curthread;
4b5f931b
MD
960 while (td->td_preempted)
961 td = td->td_preempted;
962 return(td->td_proc);
963}
964
ece04fd0
MD
965typedef struct lwkt_gettoken_req {
966 lwkt_token_t tok;
967 int cpu;
968} lwkt_gettoken_req;
969
970#if 0
4b5f931b 971
f1d1c3fa
MD
972/*
973 * This function deschedules the current thread and blocks on the specified
974 * wait queue. We obtain ownership of the wait queue in order to block
975 * on it. A generation number is used to interlock the wait queue in case
976 * it gets signalled while we are blocked waiting on the token.
977 *
978 * Note: alternatively we could dequeue our thread and then message the
979 * target cpu owning the wait queue. YYY implement as sysctl.
980 *
981 * Note: wait queue signals normally ping-pong the cpu as an optimization.
982 */
96728c05 983
f1d1c3fa 984void
ae8050a4 985lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
986{
987 thread_t td = curthread;
f1d1c3fa 988
f1d1c3fa 989 lwkt_gettoken(&w->wa_token);
ae8050a4 990 if (w->wa_gen == *gen) {
f1d1c3fa
MD
991 _lwkt_dequeue(td);
992 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
993 ++w->wa_count;
994 td->td_wait = w;
ae8050a4 995 td->td_wmesg = wmesg;
ece04fd0 996again:
f1d1c3fa 997 lwkt_switch();
ece04fd0
MD
998 lwkt_regettoken(&w->wa_token);
999 if (td->td_wmesg != NULL) {
1000 _lwkt_dequeue(td);
1001 goto again;
1002 }
8ad65e08 1003 }
ae8050a4
MD
1004 /* token might be lost, doesn't matter for gen update */
1005 *gen = w->wa_gen;
f1d1c3fa
MD
1006 lwkt_reltoken(&w->wa_token);
1007}
1008
1009/*
1010 * Signal a wait queue. We gain ownership of the wait queue in order to
1011 * signal it. Once a thread is removed from the wait queue we have to
1012 * deal with the cpu owning the thread.
1013 *
1014 * Note: alternatively we could message the target cpu owning the wait
1015 * queue. YYY implement as sysctl.
1016 */
1017void
ece04fd0 1018lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1019{
1020 thread_t td;
1021 int count;
1022
1023 lwkt_gettoken(&w->wa_token);
1024 ++w->wa_gen;
ece04fd0
MD
1025 if (count < 0)
1026 count = w->wa_count;
f1d1c3fa
MD
1027 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1028 --count;
1029 --w->wa_count;
1030 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1031 td->td_wait = NULL;
ae8050a4 1032 td->td_wmesg = NULL;
a72187e9 1033 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1034 _lwkt_enqueue(td);
1035 } else {
a72187e9 1036 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
1037 }
1038 lwkt_regettoken(&w->wa_token);
1039 }
1040 lwkt_reltoken(&w->wa_token);
1041}
1042
ece04fd0
MD
1043#endif
1044
f1d1c3fa 1045/*
96728c05 1046 * Acquire ownership of a token
f1d1c3fa 1047 *
96728c05 1048 * Acquire ownership of a token. The token may have spl and/or critical
f1d1c3fa
MD
1049 * section side effects, depending on its purpose. These side effects
1050 * guarentee that you will maintain ownership of the token as long as you
1051 * do not block. If you block you may lose access to the token (but you
1052 * must still release it even if you lose your access to it).
1053 *
96728c05 1054 * YYY for now we use a critical section to prevent IPIs from taking away
a2a5ad0d 1055 * a token, but do we really only need to disable IPIs ?
96728c05
MD
1056 *
1057 * YYY certain tokens could be made to act like mutexes when performance
1058 * would be better (e.g. t_cpu == -1). This is not yet implemented.
1059 *
a2a5ad0d
MD
1060 * YYY the tokens replace 4.x's simplelocks for the most part, but this
1061 * means that 4.x does not expect a switch so for now we cannot switch
1062 * when waiting for an IPI to be returned.
1063 *
1064 * YYY If the token is owned by another cpu we may have to send an IPI to
96728c05
MD
1065 * it and then block. The IPI causes the token to be given away to the
1066 * requesting cpu, unless it has already changed hands. Since only the
1067 * current cpu can give away a token it owns we do not need a memory barrier.
a2a5ad0d 1068 * This needs serious optimization.
f1d1c3fa 1069 */
57c254db
MD
1070
1071#ifdef SMP
1072
96728c05
MD
1073static
1074void
1075lwkt_gettoken_remote(void *arg)
1076{
1077 lwkt_gettoken_req *req = arg;
1078 if (req->tok->t_cpu == mycpu->gd_cpuid) {
634081ff 1079#ifdef INVARIANTS
a2a5ad0d
MD
1080 if (token_debug)
1081 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
634081ff 1082#endif
96728c05 1083 req->tok->t_cpu = req->cpu;
a2a5ad0d
MD
1084 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
1085 /* else set reqcpu to point to current cpu for release */
96728c05
MD
1086 }
1087}
1088
57c254db
MD
1089#endif
1090
8a8d5d85 1091int
f1d1c3fa
MD
1092lwkt_gettoken(lwkt_token_t tok)
1093{
1094 /*
1095 * Prevent preemption so the token can't be taken away from us once
1096 * we gain ownership of it. Use a synchronous request which might
1097 * block. The request will be forwarded as necessary playing catchup
1098 * to the token.
1099 */
96728c05 1100
f1d1c3fa 1101 crit_enter();
57c254db 1102#ifdef INVARIANTS
1d4c9574
MD
1103 if (curthread->td_pri > 1800) {
1104 printf("lwkt_gettoken: %p called from %p: crit sect nesting warning\n",
1105 tok, ((int **)&tok)[-1]);
1106 }
a2a5ad0d
MD
1107 if (curthread->td_pri > 2000) {
1108 curthread->td_pri = 1000;
1109 panic("too HIGH!");
57c254db
MD
1110 }
1111#endif
96728c05 1112#ifdef SMP
d0e06f83 1113 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1114 struct lwkt_gettoken_req req;
1115 int seq;
96728c05
MD
1116 int dcpu;
1117
1118 req.cpu = mycpu->gd_cpuid;
1119 req.tok = tok;
1120 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1121 KKASSERT(dcpu >= 0 && dcpu < ncpus);
634081ff 1122#ifdef INVARIANTS
a2a5ad0d
MD
1123 if (token_debug)
1124 printf("REQT%d ", dcpu);
634081ff 1125#endif
96728c05
MD
1126 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1127 lwkt_wait_ipiq(dcpu, seq);
634081ff 1128#ifdef INVARIANTS
a2a5ad0d
MD
1129 if (token_debug)
1130 printf("REQR%d ", tok->t_cpu);
634081ff 1131#endif
f1d1c3fa
MD
1132 }
1133#endif
1134 /*
1135 * leave us in a critical section on return. This will be undone
8a8d5d85 1136 * by lwkt_reltoken(). Bump the generation number.
f1d1c3fa 1137 */
8a8d5d85 1138 return(++tok->t_gen);
f1d1c3fa
MD
1139}
1140
96728c05
MD
1141/*
1142 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1143 * failure.
1144 */
1145int
1146lwkt_trytoken(lwkt_token_t tok)
1147{
1148 crit_enter();
1149#ifdef SMP
1150 if (tok->t_cpu != mycpu->gd_cpuid) {
a015262c 1151 crit_exit();
96728c05
MD
1152 return(0);
1153 }
1154#endif
1155 /* leave us in the critical section */
1156 ++tok->t_gen;
1157 return(1);
1158}
1159
f1d1c3fa
MD
1160/*
1161 * Release your ownership of a token. Releases must occur in reverse
1162 * order to aquisitions, eventually so priorities can be unwound properly
1163 * like SPLs. At the moment the actual implemention doesn't care.
1164 *
1165 * We can safely hand a token that we own to another cpu without notifying
1166 * it, but once we do we can't get it back without requesting it (unless
1167 * the other cpu hands it back to us before we check).
1168 *
1169 * We might have lost the token, so check that.
7ba9c17c
MD
1170 *
1171 * Return the token's generation number. The number is useful to callers
1172 * who may want to know if the token was stolen during potential blockages.
f1d1c3fa 1173 */
7ba9c17c 1174int
f1d1c3fa
MD
1175lwkt_reltoken(lwkt_token_t tok)
1176{
7ba9c17c
MD
1177 int gen;
1178
d0e06f83 1179 if (tok->t_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
1180 tok->t_cpu = tok->t_reqcpu;
1181 }
7ba9c17c 1182 gen = tok->t_gen;
f1d1c3fa 1183 crit_exit();
7ba9c17c 1184 return(gen);
f1d1c3fa
MD
1185}
1186
1187/*
7ba9c17c
MD
1188 * Reacquire a token that might have been lost. 0 is returned if the
1189 * generation has not changed (nobody stole the token from us), -1 is
1190 * returned otherwise. The token is reacquired regardless but the
1191 * generation number is not bumped further if we already own the token.
8a8d5d85 1192 *
7ba9c17c
MD
1193 * For efficiency we inline the best-case situation for lwkt_regettoken()
1194 * (i.e .we still own the token).
8a8d5d85
MD
1195 */
1196int
1197lwkt_gentoken(lwkt_token_t tok, int *gen)
1198{
7ba9c17c 1199 if (tok->t_cpu == mycpu->gd_cpuid && tok->t_gen == *gen)
8a8d5d85 1200 return(0);
7ba9c17c
MD
1201 *gen = lwkt_regettoken(tok);
1202 return(-1);
8a8d5d85
MD
1203}
1204
8a8d5d85 1205/*
435ff993
MD
1206 * Re-acquire a token that might have been lost. The generation number
1207 * is bumped and returned regardless of whether the token had been lost
1208 * or not (because we only have cpu granularity we have to bump the token
1209 * either way).
f1d1c3fa
MD
1210 */
1211int
1212lwkt_regettoken(lwkt_token_t tok)
1213{
96728c05 1214 /* assert we are in a critical section */
d0e06f83 1215 if (tok->t_cpu != mycpu->gd_cpuid) {
96728c05 1216#ifdef SMP
d0e06f83 1217 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1218 struct lwkt_gettoken_req req;
1219 int seq;
96728c05 1220 int dcpu;
57c254db 1221
96728c05
MD
1222 req.cpu = mycpu->gd_cpuid;
1223 req.tok = tok;
1224 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1225 KKASSERT(dcpu >= 0 && dcpu < ncpus);
634081ff 1226#ifdef INVARIANTS
cb973d15
MD
1227 if (token_debug)
1228 printf("REQT%d ", dcpu);
634081ff 1229#endif
96728c05
MD
1230 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1231 lwkt_wait_ipiq(dcpu, seq);
634081ff 1232#ifdef INVARIATNS
cb973d15
MD
1233 if (token_debug)
1234 printf("REQR%d ", tok->t_cpu);
634081ff 1235#endif
f1d1c3fa 1236 }
f1d1c3fa 1237#endif
96728c05 1238 }
435ff993 1239 ++tok->t_gen;
8a8d5d85 1240 return(tok->t_gen);
8ad65e08
MD
1241}
1242
72740893
MD
1243void
1244lwkt_inittoken(lwkt_token_t tok)
1245{
1246 /*
1247 * Zero structure and set cpu owner and reqcpu to cpu 0.
1248 */
1249 bzero(tok, sizeof(*tok));
1250}
1251
99df837e
MD
1252/*
1253 * Create a kernel process/thread/whatever. It shares it's address space
1254 * with proc0 - ie: kernel only.
1255 *
365fa13f
MD
1256 * NOTE! By default new threads are created with the MP lock held. A
1257 * thread which does not require the MP lock should release it by calling
1258 * rel_mplock() at the start of the new thread.
99df837e
MD
1259 */
1260int
1261lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1262 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1263 const char *fmt, ...)
99df837e 1264{
73e4f7b9 1265 thread_t td;
e2565a42 1266 __va_list ap;
99df837e 1267
75cdbe6c 1268 td = lwkt_alloc_thread(template, cpu);
a2a5ad0d
MD
1269 if (tdp)
1270 *tdp = td;
709799ea 1271 cpu_set_thread_handler(td, lwkt_exit, func, arg);
ef0fdad1 1272 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1273#ifdef SMP
1274 td->td_mpcount = 1;
1275#endif
99df837e
MD
1276
1277 /*
1278 * Set up arg0 for 'ps' etc
1279 */
e2565a42 1280 __va_start(ap, fmt);
99df837e 1281 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1282 __va_end(ap);
99df837e
MD
1283
1284 /*
1285 * Schedule the thread to run
1286 */
ef0fdad1
MD
1287 if ((td->td_flags & TDF_STOPREQ) == 0)
1288 lwkt_schedule(td);
1289 else
1290 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1291 return 0;
1292}
1293
2d93b37a 1294/*
2d93b37a
MD
1295 * kthread_* is specific to the kernel and is not needed by userland.
1296 */
1297#ifdef _KERNEL
1298
99df837e
MD
1299/*
1300 * Destroy an LWKT thread. Warning! This function is not called when
1301 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1302 * uses a different reaping mechanism.
1303 */
1304void
1305lwkt_exit(void)
1306{
1307 thread_t td = curthread;
1308
1309 if (td->td_flags & TDF_VERBOSE)
1310 printf("kthread %p %s has exited\n", td, td->td_comm);
1311 crit_enter();
1312 lwkt_deschedule_self();
1313 ++mycpu->gd_tdfreecount;
1314 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1315 cpu_thread_exit();
1316}
1317
1318/*
1319 * Create a kernel process/thread/whatever. It shares it's address space
ef0fdad1 1320 * with proc0 - ie: kernel only. 5.x compatible.
365fa13f
MD
1321 *
1322 * NOTE! By default kthreads are created with the MP lock held. A
1323 * thread which does not require the MP lock should release it by calling
1324 * rel_mplock() at the start of the new thread.
99df837e
MD
1325 */
1326int
1327kthread_create(void (*func)(void *), void *arg,
1328 struct thread **tdp, const char *fmt, ...)
1329{
73e4f7b9 1330 thread_t td;
e2565a42 1331 __va_list ap;
99df837e 1332
75cdbe6c 1333 td = lwkt_alloc_thread(NULL, -1);
a2a5ad0d
MD
1334 if (tdp)
1335 *tdp = td;
99df837e
MD
1336 cpu_set_thread_handler(td, kthread_exit, func, arg);
1337 td->td_flags |= TDF_VERBOSE;
8a8d5d85
MD
1338#ifdef SMP
1339 td->td_mpcount = 1;
1340#endif
99df837e
MD
1341
1342 /*
1343 * Set up arg0 for 'ps' etc
1344 */
e2565a42 1345 __va_start(ap, fmt);
99df837e 1346 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1347 __va_end(ap);
99df837e
MD
1348
1349 /*
1350 * Schedule the thread to run
1351 */
1352 lwkt_schedule(td);
1353 return 0;
1354}
1355
1356/*
1357 * Destroy an LWKT thread. Warning! This function is not called when
1358 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1359 * uses a different reaping mechanism.
1360 *
1361 * XXX duplicates lwkt_exit()
1362 */
1363void
1364kthread_exit(void)
1365{
1366 lwkt_exit();
1367}
1368
2d93b37a
MD
1369#endif /* _KERNEL */
1370
1371void
1372crit_panic(void)
1373{
1374 thread_t td = curthread;
1375 int lpri = td->td_pri;
1376
1377 td->td_pri = 0;
1378 panic("td_pri is/would-go negative! %p %d", td, lpri);
1379}
1380
96728c05
MD
1381#ifdef SMP
1382
1383/*
1384 * Send a function execution request to another cpu. The request is queued
1385 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1386 * possible target cpu. The FIFO can be written.
1387 *
1388 * YYY If the FIFO fills up we have to enable interrupts and process the
1389 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1390 * Create a CPU_*() function to do this!
1391 *
46a3f46d
MD
1392 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
1393 * end will take care of any pending interrupts.
1394 *
96728c05
MD
1395 * Must be called from a critical section.
1396 */
1397int
1398lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1399{
1400 lwkt_ipiq_t ip;
1401 int windex;
a2a5ad0d 1402 struct globaldata *gd = mycpu;
96728c05 1403
a2a5ad0d 1404 if (dcpu == gd->gd_cpuid) {
96728c05
MD
1405 func(arg);
1406 return(0);
1407 }
cb973d15 1408 crit_enter();
a2a5ad0d
MD
1409 ++gd->gd_intr_nesting_level;
1410#ifdef INVARIANTS
1411 if (gd->gd_intr_nesting_level > 20)
1412 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1413#endif
96728c05
MD
1414 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1415 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1416 ++ipiq_count;
a2a5ad0d 1417 ip = &gd->gd_ipiq[dcpu];
cb973d15
MD
1418
1419 /*
1420 * We always drain before the FIFO becomes full so it should never
1421 * become full. We need to leave enough entries to deal with
1422 * reentrancy.
1423 */
1424 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1425 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1426 ip->ip_func[windex] = func;
1427 ip->ip_arg[windex] = arg;
1428 /* YYY memory barrier */
1429 ++ip->ip_windex;
96728c05
MD
1430 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1431 unsigned int eflags = read_eflags();
1432 cpu_enable_intr();
1433 ++ipiq_fifofull;
cb973d15 1434 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
96728c05
MD
1435 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1436 lwkt_process_ipiq();
1437 }
1438 write_eflags(eflags);
1439 }
a2a5ad0d 1440 --gd->gd_intr_nesting_level;
96728c05 1441 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
cb973d15 1442 crit_exit();
96728c05
MD
1443 return(ip->ip_windex);
1444}
1445
cb973d15
MD
1446/*
1447 * Send a message to several target cpus. Typically used for scheduling.
435ff993 1448 * The message will not be sent to stopped cpus.
cb973d15
MD
1449 */
1450void
1451lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1452{
1453 int cpuid;
1454
435ff993 1455 mask &= ~stopped_cpus;
cb973d15
MD
1456 while (mask) {
1457 cpuid = bsfl(mask);
1458 lwkt_send_ipiq(cpuid, func, arg);
1459 mask &= ~(1 << cpuid);
1460 }
1461}
1462
96728c05
MD
1463/*
1464 * Wait for the remote cpu to finish processing a function.
1465 *
1466 * YYY we have to enable interrupts and process the IPIQ while waiting
1467 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1468 * function to do this! YYY we really should 'block' here.
1469 *
1470 * Must be called from a critical section. Thsi routine may be called
1471 * from an interrupt (for example, if an interrupt wakes a foreign thread
1472 * up).
1473 */
1474void
1475lwkt_wait_ipiq(int dcpu, int seq)
1476{
1477 lwkt_ipiq_t ip;
a2a5ad0d 1478 int maxc = 100000000;
96728c05
MD
1479
1480 if (dcpu != mycpu->gd_cpuid) {
1481 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1482 ip = &mycpu->gd_ipiq[dcpu];
cb973d15 1483 if ((int)(ip->ip_xindex - seq) < 0) {
96728c05
MD
1484 unsigned int eflags = read_eflags();
1485 cpu_enable_intr();
cb973d15 1486 while ((int)(ip->ip_xindex - seq) < 0) {
96728c05 1487 lwkt_process_ipiq();
a2a5ad0d 1488 if (--maxc == 0)
cb973d15 1489 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
a2a5ad0d
MD
1490 if (maxc < -1000000)
1491 panic("LWKT_WAIT_IPIQ");
96728c05
MD
1492 }
1493 write_eflags(eflags);
1494 }
1495 }
1496}
1497
1498/*
1499 * Called from IPI interrupt (like a fast interrupt), which has placed
1500 * us in a critical section. The MP lock may or may not be held.
cb973d15
MD
1501 * May also be called from doreti or splz, or be reentrantly called
1502 * indirectly through the ip_func[] we run.
96728c05
MD
1503 */
1504void
1505lwkt_process_ipiq(void)
1506{
1507 int n;
1508 int cpuid = mycpu->gd_cpuid;
1509
1510 for (n = 0; n < ncpus; ++n) {
1511 lwkt_ipiq_t ip;
1512 int ri;
1513
1514 if (n == cpuid)
1515 continue;
1516 ip = globaldata_find(n)->gd_ipiq;
1517 if (ip == NULL)
1518 continue;
1519 ip = &ip[cpuid];
cb973d15
MD
1520
1521 /*
1522 * Note: xindex is only updated after we are sure the function has
1523 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1524 * function may send an IPI which may block/drain.
1525 */
96728c05
MD
1526 while (ip->ip_rindex != ip->ip_windex) {
1527 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
96728c05 1528 ++ip->ip_rindex;
cb973d15
MD
1529 ip->ip_func[ri](ip->ip_arg[ri]);
1530 /* YYY memory barrier */
1531 ip->ip_xindex = ip->ip_rindex;
96728c05
MD
1532 }
1533 }
1534}
1535
1536#else
1537
1538int
1539lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1540{
1541 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1542 return(0); /* NOT REACHED */
1543}
1544
1545void
1546lwkt_wait_ipiq(int dcpu, int seq)
1547{
1548 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);
1549}
1550
1551#endif