zfreei->zfree (there is no zfreei anymore)
[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 *
f1d1c3fa
MD
26 * Each cpu in a system has its own self-contained light weight kernel
27 * thread scheduler, which means that generally speaking we only need
96728c05
MD
28 * to use a critical section to avoid problems. Foreign thread
29 * scheduling is queued via (async) IPIs.
f1d1c3fa 30 *
377d4740 31 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.24 2003/07/19 21:14:38 dillon Exp $
8ad65e08
MD
32 */
33
34#include <sys/param.h>
35#include <sys/systm.h>
36#include <sys/kernel.h>
37#include <sys/proc.h>
38#include <sys/rtprio.h>
39#include <sys/queue.h>
f1d1c3fa 40#include <sys/thread2.h>
7d0bac62 41#include <sys/sysctl.h>
99df837e 42#include <sys/kthread.h>
f1d1c3fa 43#include <machine/cpu.h>
99df837e 44#include <sys/lock.h>
f1d1c3fa 45
7d0bac62
MD
46#include <vm/vm.h>
47#include <vm/vm_param.h>
48#include <vm/vm_kern.h>
49#include <vm/vm_object.h>
50#include <vm/vm_page.h>
51#include <vm/vm_map.h>
52#include <vm/vm_pager.h>
53#include <vm/vm_extern.h>
54#include <vm/vm_zone.h>
55
99df837e 56#include <machine/stdarg.h>
57c254db 57#include <machine/ipl.h>
96728c05
MD
58#ifdef SMP
59#include <machine/smp.h>
60#endif
99df837e 61
7d0bac62 62static int untimely_switch = 0;
4b5f931b 63SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
57c254db
MD
64#ifdef INVARIANTS
65static int token_debug = 0;
66SYSCTL_INT(_lwkt, OID_AUTO, token_debug, CTLFLAG_RW, &token_debug, 0, "");
67#endif
4b5f931b
MD
68static quad_t switch_count = 0;
69SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
70static quad_t preempt_hit = 0;
71SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
72static quad_t preempt_miss = 0;
73SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b
MD
74static quad_t preempt_weird = 0;
75SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
96728c05
MD
76static quad_t ipiq_count = 0;
77SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
78static quad_t ipiq_fifofull = 0;
79SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
7d0bac62 80
4b5f931b
MD
81/*
82 * These helper procedures handle the runq, they can only be called from
83 * within a critical section.
84 */
f1d1c3fa
MD
85static __inline
86void
87_lwkt_dequeue(thread_t td)
88{
89 if (td->td_flags & TDF_RUNQ) {
4b5f931b
MD
90 int nq = td->td_pri & TDPRI_MASK;
91 struct globaldata *gd = mycpu;
92
f1d1c3fa 93 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
94 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
95 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
96 }
97}
98
99static __inline
100void
101_lwkt_enqueue(thread_t td)
102{
103 if ((td->td_flags & TDF_RUNQ) == 0) {
4b5f931b
MD
104 int nq = td->td_pri & TDPRI_MASK;
105 struct globaldata *gd = mycpu;
106
f1d1c3fa 107 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
108 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
109 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
110 }
111}
8ad65e08 112
57c254db
MD
113static __inline
114int
115_lwkt_wantresched(thread_t ntd, thread_t cur)
116{
117 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
118}
119
8ad65e08
MD
120/*
121 * LWKTs operate on a per-cpu basis
122 *
73e4f7b9 123 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
124 */
125void
126lwkt_gdinit(struct globaldata *gd)
127{
4b5f931b
MD
128 int i;
129
130 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
131 TAILQ_INIT(&gd->gd_tdrunq[i]);
132 gd->gd_runqmask = 0;
73e4f7b9 133 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
134}
135
7d0bac62
MD
136/*
137 * Initialize a thread wait structure prior to first use.
138 *
139 * NOTE! called from low level boot code, we cannot do anything fancy!
140 */
141void
142lwkt_init_wait(lwkt_wait_t w)
143{
144 TAILQ_INIT(&w->wa_waitq);
145}
146
147/*
148 * Create a new thread. The thread must be associated with a process context
149 * or LWKT start address before it can be scheduled.
0cfcada1
MD
150 *
151 * If you intend to create a thread without a process context this function
152 * does everything except load the startup and switcher function.
7d0bac62
MD
153 */
154thread_t
ef0fdad1 155lwkt_alloc_thread(struct thread *td)
7d0bac62 156{
99df837e 157 void *stack;
ef0fdad1 158 int flags = 0;
7d0bac62 159
ef0fdad1 160 if (td == NULL) {
26a0694b 161 crit_enter();
ef0fdad1
MD
162 if (mycpu->gd_tdfreecount > 0) {
163 --mycpu->gd_tdfreecount;
164 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
d9eea1a5 165 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1
MD
166 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
167 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
168 crit_exit();
169 stack = td->td_kstack;
170 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
171 } else {
172 crit_exit();
173 td = zalloc(thread_zone);
174 td->td_kstack = NULL;
175 flags |= TDF_ALLOCATED_THREAD;
176 }
177 }
178 if ((stack = td->td_kstack) == NULL) {
99df837e 179 stack = (void *)kmem_alloc(kernel_map, UPAGES * PAGE_SIZE);
ef0fdad1 180 flags |= TDF_ALLOCATED_STACK;
99df837e 181 }
26a0694b 182 lwkt_init_thread(td, stack, flags, mycpu);
99df837e 183 return(td);
7d0bac62
MD
184}
185
186/*
187 * Initialize a preexisting thread structure. This function is used by
188 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
189 *
190 * NOTE! called from low level boot code, we cannot do anything fancy!
191 */
192void
26a0694b 193lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
7d0bac62 194{
99df837e
MD
195 bzero(td, sizeof(struct thread));
196 td->td_kstack = stack;
197 td->td_flags |= flags;
26a0694b
MD
198 td->td_gd = gd;
199 td->td_pri = TDPRI_CRIT;
d9eea1a5 200 td->td_cpu = gd->gd_cpuid; /* YYY don't really need this if have td_gd */
99df837e 201 pmap_init_thread(td);
73e4f7b9
MD
202 crit_enter();
203 TAILQ_INSERT_TAIL(&mycpu->gd_tdallq, td, td_allq);
204 crit_exit();
205}
206
207void
208lwkt_set_comm(thread_t td, const char *ctl, ...)
209{
210 va_list va;
211
212 va_start(va, ctl);
213 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
214 va_end(va);
7d0bac62
MD
215}
216
99df837e 217void
73e4f7b9 218lwkt_hold(thread_t td)
99df837e 219{
73e4f7b9
MD
220 ++td->td_refs;
221}
222
223void
224lwkt_rele(thread_t td)
225{
226 KKASSERT(td->td_refs > 0);
227 --td->td_refs;
228}
229
230void
231lwkt_wait_free(thread_t td)
232{
233 while (td->td_refs)
377d4740 234 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
235}
236
237void
238lwkt_free_thread(thread_t td)
239{
240 struct globaldata *gd = mycpu;
241
d9eea1a5 242 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
243 ("lwkt_free_thread: did not exit! %p", td));
244
245 crit_enter();
73e4f7b9
MD
246 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
247 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
248 (td->td_flags & TDF_ALLOCATED_THREAD)
249 ) {
73e4f7b9
MD
250 ++gd->gd_tdfreecount;
251 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
99df837e
MD
252 crit_exit();
253 } else {
254 crit_exit();
255 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
256 kmem_free(kernel_map,
257 (vm_offset_t)td->td_kstack, UPAGES * PAGE_SIZE);
73e4f7b9 258 /* gd invalid */
99df837e
MD
259 td->td_kstack = NULL;
260 }
261 if (td->td_flags & TDF_ALLOCATED_THREAD)
262 zfree(thread_zone, td);
263 }
264}
265
266
8ad65e08
MD
267/*
268 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
269 * switch to the idlethread. Switching must occur within a critical
270 * section to avoid races with the scheduling queue.
271 *
272 * We always have full control over our cpu's run queue. Other cpus
273 * that wish to manipulate our queue must use the cpu_*msg() calls to
274 * talk to our cpu, so a critical section is all that is needed and
275 * the result is very, very fast thread switching.
276 *
96728c05
MD
277 * The LWKT scheduler uses a fixed priority model and round-robins at
278 * each priority level. User process scheduling is a totally
279 * different beast and LWKT priorities should not be confused with
280 * user process priorities.
f1d1c3fa 281 *
96728c05
MD
282 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
283 * cleans it up. Note that the td_switch() function cannot do anything that
284 * requires the MP lock since the MP lock will have already been setup for
285 * the target thread (not the current thread).
8ad65e08 286 */
96728c05 287
8ad65e08
MD
288void
289lwkt_switch(void)
290{
4b5f931b 291 struct globaldata *gd;
f1d1c3fa 292 thread_t td = curthread;
8ad65e08 293 thread_t ntd;
8a8d5d85
MD
294#ifdef SMP
295 int mpheld;
296#endif
8ad65e08 297
96728c05
MD
298 if (mycpu->gd_intr_nesting_level &&
299 td->td_preempted == NULL && panicstr == NULL
300 ) {
26a0694b 301 panic("lwkt_switch: cannot switch from within an interrupt, yet\n");
96728c05 302 }
ef0fdad1 303
cb973d15
MD
304 /*
305 * Passive release (used to transition from user to kernel mode
306 * when we block or switch rather then when we enter the kernel).
307 * This function is NOT called if we are switching into a preemption
308 * or returning from a preemption. Typically this causes us to lose
309 * our P_CURPROC designation (if we have one) and become a true LWKT
310 * thread, and may also hand P_CURPROC to another process and schedule
311 * its thread.
312 */
313 if (td->td_release)
314 td->td_release(td);
315
f1d1c3fa 316 crit_enter();
4b5f931b 317 ++switch_count;
8a8d5d85
MD
318
319#ifdef SMP
320 /*
321 * td_mpcount cannot be used to determine if we currently hold the
322 * MP lock because get_mplock() will increment it prior to attempting
323 * to get the lock, and switch out if it can't. Look at the actual lock.
324 */
325 mpheld = MP_LOCK_HELD();
326#endif
99df837e
MD
327 if ((ntd = td->td_preempted) != NULL) {
328 /*
329 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
330 * thread. This occurs transparently, whether the preempted thread
331 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
332 * itself).
333 *
334 * We have to setup the MP lock for the original thread after backing
335 * out the adjustment that was made to curthread when the original
336 * was preempted.
99df837e 337 */
26a0694b 338 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 339#ifdef SMP
96728c05
MD
340 if (ntd->td_mpcount && mpheld == 0) {
341 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
342 td, ntd, td->td_mpcount, ntd->td_mpcount);
343 }
8a8d5d85
MD
344 if (ntd->td_mpcount) {
345 td->td_mpcount -= ntd->td_mpcount;
346 KKASSERT(td->td_mpcount >= 0);
347 }
348#endif
26a0694b 349 ntd->td_flags |= TDF_PREEMPT_DONE;
8a8d5d85 350 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 351 } else {
4b5f931b
MD
352 /*
353 * Priority queue / round-robin at each priority. Note that user
354 * processes run at a fixed, low priority and the user process
355 * scheduler deals with interactions between user processes
356 * by scheduling and descheduling them from the LWKT queue as
357 * necessary.
8a8d5d85
MD
358 *
359 * We have to adjust the MP lock for the target thread. If we
360 * need the MP lock and cannot obtain it we try to locate a
361 * thread that does not need the MP lock.
4b5f931b
MD
362 */
363 gd = mycpu;
4b5f931b
MD
364again:
365 if (gd->gd_runqmask) {
366 int nq = bsrl(gd->gd_runqmask);
367 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
368 gd->gd_runqmask &= ~(1 << nq);
369 goto again;
370 }
8a8d5d85
MD
371#ifdef SMP
372 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
373 /*
96728c05
MD
374 * Target needs MP lock and we couldn't get it, try
375 * to locate a thread which does not need the MP lock
3c23a41a 376 * to run. If we cannot locate a thread spin in idle.
8a8d5d85
MD
377 */
378 u_int32_t rqmask = gd->gd_runqmask;
379 while (rqmask) {
380 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
381 if (ntd->td_mpcount == 0)
382 break;
383 }
384 if (ntd)
385 break;
386 rqmask &= ~(1 << nq);
387 nq = bsrl(rqmask);
388 }
389 if (ntd == NULL) {
a2a5ad0d
MD
390 ntd = &gd->gd_idlethread;
391 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
392 } else {
393 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
394 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
395 }
396 } else {
397 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
398 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
399 }
400#else
4b5f931b
MD
401 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
402 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 403#endif
4b5f931b 404 } else {
3c23a41a
MD
405 /*
406 * Nothing to run but we may still need the BGL to deal with
407 * pending interrupts, spin in idle if so.
408 */
a2a5ad0d 409 ntd = &gd->gd_idlethread;
235957ed 410 if (gd->gd_reqflags)
3c23a41a 411 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 412 }
f1d1c3fa 413 }
26a0694b
MD
414 KASSERT(ntd->td_pri >= TDPRI_CRIT,
415 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
416
417 /*
418 * Do the actual switch. If the new target does not need the MP lock
419 * and we are holding it, release the MP lock. If the new target requires
420 * the MP lock we have already acquired it for the target.
421 */
422#ifdef SMP
423 if (ntd->td_mpcount == 0 ) {
424 if (MP_LOCK_HELD())
425 cpu_rel_mplock();
426 } else {
427 ASSERT_MP_LOCK_HELD();
428 }
429#endif
8a8d5d85 430 if (td != ntd) {
f1d1c3fa 431 td->td_switch(ntd);
8a8d5d85 432 }
96728c05 433
f1d1c3fa 434 crit_exit();
8ad65e08
MD
435}
436
cb973d15
MD
437/*
438 * Switch if another thread has a higher priority. Do not switch to other
439 * threads at the same priority.
440 */
441void
442lwkt_maybe_switch()
443{
444 struct globaldata *gd = mycpu;
445 struct thread *td = gd->gd_curthread;
446
447 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
448 lwkt_switch();
449 }
450}
451
b68b7282 452/*
96728c05
MD
453 * Request that the target thread preempt the current thread. Preemption
454 * only works under a specific set of conditions:
b68b7282 455 *
96728c05
MD
456 * - We are not preempting ourselves
457 * - The target thread is owned by the current cpu
458 * - We are not currently being preempted
459 * - The target is not currently being preempted
460 * - We are able to satisfy the target's MP lock requirements (if any).
461 *
462 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
463 * this is called via lwkt_schedule() through the td_preemptable callback.
464 * critpri is the managed critical priority that we should ignore in order
465 * to determine whether preemption is possible (aka usually just the crit
466 * priority of lwkt_schedule() itself).
b68b7282 467 *
26a0694b
MD
468 * XXX at the moment we run the target thread in a critical section during
469 * the preemption in order to prevent the target from taking interrupts
470 * that *WE* can't. Preemption is strictly limited to interrupt threads
471 * and interrupt-like threads, outside of a critical section, and the
472 * preempted source thread will be resumed the instant the target blocks
473 * whether or not the source is scheduled (i.e. preemption is supposed to
474 * be as transparent as possible).
4b5f931b 475 *
8a8d5d85
MD
476 * The target thread inherits our MP count (added to its own) for the
477 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
478 * MP lock during the preemption. Therefore, any preempting targets must be
479 * careful in regards to MP assertions. Note that the MP count may be
480 * out of sync with the physical mp_lock. If we preempt we have to preserve
481 * the expected situation.
b68b7282
MD
482 */
483void
96728c05 484lwkt_preempt(thread_t ntd, int critpri)
b68b7282 485{
73e4f7b9 486 thread_t td = curthread;
8a8d5d85
MD
487#ifdef SMP
488 int mpheld;
57c254db 489 int savecnt;
8a8d5d85 490#endif
b68b7282 491
26a0694b 492 /*
96728c05
MD
493 * The caller has put us in a critical section. We can only preempt
494 * if the caller of the caller was not in a critical section (basically
57c254db
MD
495 * a local interrupt), as determined by the 'critpri' parameter. If
496 * we are unable to preempt
96728c05
MD
497 *
498 * YYY The target thread must be in a critical section (else it must
499 * inherit our critical section? I dunno yet).
26a0694b
MD
500 */
501 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 502
cb973d15 503 need_resched();
57c254db
MD
504 if (!_lwkt_wantresched(ntd, td)) {
505 ++preempt_miss;
506 return;
507 }
96728c05
MD
508 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
509 ++preempt_miss;
510 return;
511 }
512#ifdef SMP
513 if (ntd->td_cpu != mycpu->gd_cpuid) {
514 ++preempt_miss;
515 return;
516 }
517#endif
26a0694b
MD
518 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
519 ++preempt_weird;
520 return;
521 }
522 if (ntd->td_preempted) {
4b5f931b 523 ++preempt_hit;
26a0694b 524 return;
b68b7282 525 }
8a8d5d85 526#ifdef SMP
a2a5ad0d
MD
527 /*
528 * note: an interrupt might have occured just as we were transitioning
a5934754
MD
529 * to the MP lock. In this case td_mpcount will be pre-disposed but
530 * not actually synchronized with the actual state of the lock. We
531 * can use it to imply an MP lock requirement for the preemption but
532 * we cannot use it to test whether we hold the MP lock or not.
a2a5ad0d 533 */
a5934754 534 mpheld = MP_LOCK_HELD();
96728c05
MD
535 if (mpheld && td->td_mpcount == 0)
536 panic("lwkt_preempt(): held and no count");
537 savecnt = td->td_mpcount;
8a8d5d85
MD
538 ntd->td_mpcount += td->td_mpcount;
539 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
540 ntd->td_mpcount -= td->td_mpcount;
541 ++preempt_miss;
542 return;
543 }
544#endif
26a0694b
MD
545
546 ++preempt_hit;
547 ntd->td_preempted = td;
548 td->td_flags |= TDF_PREEMPT_LOCK;
549 td->td_switch(ntd);
550 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
551#ifdef SMP
552 KKASSERT(savecnt == td->td_mpcount);
553 if (mpheld == 0 && MP_LOCK_HELD())
554 cpu_rel_mplock();
555 else if (mpheld && !MP_LOCK_HELD())
556 panic("lwkt_preempt(): MP lock was not held through");
557#endif
26a0694b
MD
558 ntd->td_preempted = NULL;
559 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
560}
561
f1d1c3fa
MD
562/*
563 * Yield our thread while higher priority threads are pending. This is
564 * typically called when we leave a critical section but it can be safely
565 * called while we are in a critical section.
566 *
567 * This function will not generally yield to equal priority threads but it
568 * can occur as a side effect. Note that lwkt_switch() is called from
569 * inside the critical section to pervent its own crit_exit() from reentering
570 * lwkt_yield_quick().
571 *
235957ed 572 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
573 * came along but was blocked and made pending.
574 *
f1d1c3fa
MD
575 * (self contained on a per cpu basis)
576 */
577void
578lwkt_yield_quick(void)
579{
580 thread_t td = curthread;
ef0fdad1 581
a2a5ad0d 582 /*
235957ed 583 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
584 * it with a non-zero cpl then we might not wind up calling splz after
585 * a task switch when the critical section is exited even though the
586 * new task could accept the interrupt. YYY alternative is to have
587 * lwkt_switch() just call splz unconditionally.
588 *
589 * XXX from crit_exit() only called after last crit section is released.
590 * If called directly will run splz() even if in a critical section.
591 */
235957ed 592 if (mycpu->gd_reqflags)
f1d1c3fa 593 splz();
f1d1c3fa
MD
594
595 /*
596 * YYY enabling will cause wakeup() to task-switch, which really
597 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
598 * preemption and MP without actually doing preemption or MP, because a
599 * lot of code assumes that wakeup() does not block.
f1d1c3fa 600 */
ef0fdad1 601 if (untimely_switch && mycpu->gd_intr_nesting_level == 0) {
f1d1c3fa
MD
602 crit_enter();
603 /*
604 * YYY temporary hacks until we disassociate the userland scheduler
605 * from the LWKT scheduler.
606 */
607 if (td->td_flags & TDF_RUNQ) {
608 lwkt_switch(); /* will not reenter yield function */
609 } else {
610 lwkt_schedule_self(); /* make sure we are scheduled */
611 lwkt_switch(); /* will not reenter yield function */
612 lwkt_deschedule_self(); /* make sure we are descheduled */
613 }
614 crit_exit_noyield();
615 }
f1d1c3fa
MD
616}
617
8ad65e08 618/*
f1d1c3fa 619 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 620 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
621 * the crit_exit() call in lwkt_switch().
622 *
623 * (self contained on a per cpu basis)
8ad65e08
MD
624 */
625void
f1d1c3fa 626lwkt_yield(void)
8ad65e08 627{
f1d1c3fa
MD
628 lwkt_schedule_self();
629 lwkt_switch();
630}
631
632/*
633 * Schedule a thread to run. As the current thread we can always safely
634 * schedule ourselves, and a shortcut procedure is provided for that
635 * function.
636 *
637 * (non-blocking, self contained on a per cpu basis)
638 */
639void
640lwkt_schedule_self(void)
641{
642 thread_t td = curthread;
643
644 crit_enter();
645 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa 646 _lwkt_enqueue(td);
26a0694b
MD
647 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
648 panic("SCHED SELF PANIC");
f1d1c3fa 649 crit_exit();
8ad65e08 650}
8ad65e08
MD
651
652/*
f1d1c3fa
MD
653 * Generic schedule. Possibly schedule threads belonging to other cpus and
654 * deal with threads that might be blocked on a wait queue.
655 *
96728c05 656 * YYY this is one of the best places to implement load balancing code.
f1d1c3fa
MD
657 * Load balancing can be accomplished by requesting other sorts of actions
658 * for the thread in question.
8ad65e08
MD
659 */
660void
661lwkt_schedule(thread_t td)
662{
96728c05 663#ifdef INVARIANTS
26a0694b
MD
664 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
665 && td->td_proc->p_stat == SSLEEP
666 ) {
667 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
668 curthread,
669 curthread->td_proc ? curthread->td_proc->p_pid : -1,
670 curthread->td_proc ? curthread->td_proc->p_stat : -1,
671 td,
672 td->td_proc ? curthread->td_proc->p_pid : -1,
673 td->td_proc ? curthread->td_proc->p_stat : -1
674 );
675 panic("SCHED PANIC");
676 }
96728c05 677#endif
f1d1c3fa
MD
678 crit_enter();
679 if (td == curthread) {
680 _lwkt_enqueue(td);
681 } else {
682 lwkt_wait_t w;
683
684 /*
685 * If the thread is on a wait list we have to send our scheduling
686 * request to the owner of the wait structure. Otherwise we send
687 * the scheduling request to the cpu owning the thread. Races
688 * are ok, the target will forward the message as necessary (the
689 * message may chase the thread around before it finally gets
690 * acted upon).
691 *
692 * (remember, wait structures use stable storage)
693 */
694 if ((w = td->td_wait) != NULL) {
96728c05 695 if (lwkt_trytoken(&w->wa_token)) {
f1d1c3fa
MD
696 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
697 --w->wa_count;
698 td->td_wait = NULL;
d0e06f83 699 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa 700 _lwkt_enqueue(td);
57c254db 701 if (td->td_preemptable) {
96728c05 702 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
57c254db
MD
703 } else if (_lwkt_wantresched(td, curthread)) {
704 need_resched();
705 }
f1d1c3fa 706 } else {
96728c05 707 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 708 }
96728c05 709 lwkt_reltoken(&w->wa_token);
f1d1c3fa 710 } else {
96728c05 711 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
712 }
713 } else {
714 /*
715 * If the wait structure is NULL and we own the thread, there
716 * is no race (since we are in a critical section). If we
717 * do not own the thread there might be a race but the
718 * target cpu will deal with it.
719 */
d0e06f83 720 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa 721 _lwkt_enqueue(td);
57c254db 722 if (td->td_preemptable) {
96728c05 723 td->td_preemptable(td, TDPRI_CRIT);
57c254db
MD
724 } else if (_lwkt_wantresched(td, curthread)) {
725 need_resched();
726 }
f1d1c3fa 727 } else {
96728c05 728 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
729 }
730 }
8ad65e08 731 }
f1d1c3fa 732 crit_exit();
8ad65e08
MD
733}
734
d9eea1a5
MD
735/*
736 * Managed acquisition. This code assumes that the MP lock is held for
737 * the tdallq operation and that the thread has been descheduled from its
738 * original cpu. We also have to wait for the thread to be entirely switched
739 * out on its original cpu (this is usually fast enough that we never loop)
740 * since the LWKT system does not have to hold the MP lock while switching
741 * and the target may have released it before switching.
742 */
a2a5ad0d
MD
743void
744lwkt_acquire(thread_t td)
745{
746 struct globaldata *gd;
cb973d15 747 int ocpu;
a2a5ad0d
MD
748
749 gd = td->td_gd;
750 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5
MD
751 while (td->td_flags & TDF_RUNNING) /* XXX spin */
752 ;
a2a5ad0d 753 if (gd != mycpu) {
cb973d15 754 ocpu = td->td_cpu;
a2a5ad0d
MD
755 crit_enter();
756 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
757 gd = mycpu;
758 td->td_gd = gd;
759 td->td_cpu = gd->gd_cpuid;
760 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
761 crit_exit();
762 }
763}
764
8ad65e08 765/*
f1d1c3fa
MD
766 * Deschedule a thread.
767 *
768 * (non-blocking, self contained on a per cpu basis)
769 */
770void
771lwkt_deschedule_self(void)
772{
773 thread_t td = curthread;
774
775 crit_enter();
776 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa
MD
777 _lwkt_dequeue(td);
778 crit_exit();
779}
780
781/*
782 * Generic deschedule. Descheduling threads other then your own should be
783 * done only in carefully controlled circumstances. Descheduling is
784 * asynchronous.
785 *
786 * This function may block if the cpu has run out of messages.
8ad65e08
MD
787 */
788void
789lwkt_deschedule(thread_t td)
790{
f1d1c3fa
MD
791 crit_enter();
792 if (td == curthread) {
793 _lwkt_dequeue(td);
794 } else {
d0e06f83 795 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
796 _lwkt_dequeue(td);
797 } else {
96728c05 798 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
799 }
800 }
801 crit_exit();
802}
803
4b5f931b
MD
804/*
805 * Set the target thread's priority. This routine does not automatically
806 * switch to a higher priority thread, LWKT threads are not designed for
807 * continuous priority changes. Yield if you want to switch.
808 *
809 * We have to retain the critical section count which uses the high bits
26a0694b
MD
810 * of the td_pri field. The specified priority may also indicate zero or
811 * more critical sections by adding TDPRI_CRIT*N.
4b5f931b
MD
812 */
813void
814lwkt_setpri(thread_t td, int pri)
815{
26a0694b 816 KKASSERT(pri >= 0);
57c254db 817 KKASSERT(td->td_cpu == mycpu->gd_cpuid);
26a0694b
MD
818 crit_enter();
819 if (td->td_flags & TDF_RUNQ) {
820 _lwkt_dequeue(td);
821 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
822 _lwkt_enqueue(td);
823 } else {
824 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
825 }
826 crit_exit();
827}
828
829void
830lwkt_setpri_self(int pri)
831{
832 thread_t td = curthread;
833
4b5f931b
MD
834 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
835 crit_enter();
836 if (td->td_flags & TDF_RUNQ) {
837 _lwkt_dequeue(td);
838 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
839 _lwkt_enqueue(td);
840 } else {
841 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
842 }
843 crit_exit();
844}
845
846struct proc *
847lwkt_preempted_proc(void)
848{
73e4f7b9 849 thread_t td = curthread;
4b5f931b
MD
850 while (td->td_preempted)
851 td = td->td_preempted;
852 return(td->td_proc);
853}
854
855
f1d1c3fa
MD
856/*
857 * This function deschedules the current thread and blocks on the specified
858 * wait queue. We obtain ownership of the wait queue in order to block
859 * on it. A generation number is used to interlock the wait queue in case
860 * it gets signalled while we are blocked waiting on the token.
861 *
862 * Note: alternatively we could dequeue our thread and then message the
863 * target cpu owning the wait queue. YYY implement as sysctl.
864 *
865 * Note: wait queue signals normally ping-pong the cpu as an optimization.
866 */
96728c05
MD
867typedef struct lwkt_gettoken_req {
868 lwkt_token_t tok;
869 int cpu;
870} lwkt_gettoken_req;
871
f1d1c3fa 872void
ae8050a4 873lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
874{
875 thread_t td = curthread;
f1d1c3fa 876
f1d1c3fa 877 lwkt_gettoken(&w->wa_token);
ae8050a4 878 if (w->wa_gen == *gen) {
f1d1c3fa
MD
879 _lwkt_dequeue(td);
880 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
881 ++w->wa_count;
882 td->td_wait = w;
ae8050a4 883 td->td_wmesg = wmesg;
f1d1c3fa 884 lwkt_switch();
8ad65e08 885 }
ae8050a4
MD
886 /* token might be lost, doesn't matter for gen update */
887 *gen = w->wa_gen;
f1d1c3fa
MD
888 lwkt_reltoken(&w->wa_token);
889}
890
891/*
892 * Signal a wait queue. We gain ownership of the wait queue in order to
893 * signal it. Once a thread is removed from the wait queue we have to
894 * deal with the cpu owning the thread.
895 *
896 * Note: alternatively we could message the target cpu owning the wait
897 * queue. YYY implement as sysctl.
898 */
899void
900lwkt_signal(lwkt_wait_t w)
901{
902 thread_t td;
903 int count;
904
905 lwkt_gettoken(&w->wa_token);
906 ++w->wa_gen;
907 count = w->wa_count;
908 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
909 --count;
910 --w->wa_count;
911 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
912 td->td_wait = NULL;
ae8050a4 913 td->td_wmesg = NULL;
d0e06f83 914 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
915 _lwkt_enqueue(td);
916 } else {
96728c05 917 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
918 }
919 lwkt_regettoken(&w->wa_token);
920 }
921 lwkt_reltoken(&w->wa_token);
922}
923
924/*
96728c05 925 * Acquire ownership of a token
f1d1c3fa 926 *
96728c05 927 * Acquire ownership of a token. The token may have spl and/or critical
f1d1c3fa
MD
928 * section side effects, depending on its purpose. These side effects
929 * guarentee that you will maintain ownership of the token as long as you
930 * do not block. If you block you may lose access to the token (but you
931 * must still release it even if you lose your access to it).
932 *
96728c05 933 * YYY for now we use a critical section to prevent IPIs from taking away
a2a5ad0d 934 * a token, but do we really only need to disable IPIs ?
96728c05
MD
935 *
936 * YYY certain tokens could be made to act like mutexes when performance
937 * would be better (e.g. t_cpu == -1). This is not yet implemented.
938 *
a2a5ad0d
MD
939 * YYY the tokens replace 4.x's simplelocks for the most part, but this
940 * means that 4.x does not expect a switch so for now we cannot switch
941 * when waiting for an IPI to be returned.
942 *
943 * YYY If the token is owned by another cpu we may have to send an IPI to
96728c05
MD
944 * it and then block. The IPI causes the token to be given away to the
945 * requesting cpu, unless it has already changed hands. Since only the
946 * current cpu can give away a token it owns we do not need a memory barrier.
a2a5ad0d 947 * This needs serious optimization.
f1d1c3fa 948 */
57c254db
MD
949
950#ifdef SMP
951
96728c05
MD
952static
953void
954lwkt_gettoken_remote(void *arg)
955{
956 lwkt_gettoken_req *req = arg;
957 if (req->tok->t_cpu == mycpu->gd_cpuid) {
a2a5ad0d
MD
958 if (token_debug)
959 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
96728c05 960 req->tok->t_cpu = req->cpu;
a2a5ad0d
MD
961 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
962 /* else set reqcpu to point to current cpu for release */
96728c05
MD
963 }
964}
965
57c254db
MD
966#endif
967
8a8d5d85 968int
f1d1c3fa
MD
969lwkt_gettoken(lwkt_token_t tok)
970{
971 /*
972 * Prevent preemption so the token can't be taken away from us once
973 * we gain ownership of it. Use a synchronous request which might
974 * block. The request will be forwarded as necessary playing catchup
975 * to the token.
976 */
96728c05 977
f1d1c3fa 978 crit_enter();
57c254db 979#ifdef INVARIANTS
a2a5ad0d
MD
980 if (curthread->td_pri > 2000) {
981 curthread->td_pri = 1000;
982 panic("too HIGH!");
57c254db
MD
983 }
984#endif
96728c05 985#ifdef SMP
d0e06f83 986 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
987 struct lwkt_gettoken_req req;
988 int seq;
96728c05
MD
989 int dcpu;
990
991 req.cpu = mycpu->gd_cpuid;
992 req.tok = tok;
993 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d
MD
994 KKASSERT(dcpu >= 0 && dcpu < ncpus);
995 if (token_debug)
996 printf("REQT%d ", dcpu);
96728c05
MD
997 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
998 lwkt_wait_ipiq(dcpu, seq);
a2a5ad0d
MD
999 if (token_debug)
1000 printf("REQR%d ", tok->t_cpu);
f1d1c3fa
MD
1001 }
1002#endif
1003 /*
1004 * leave us in a critical section on return. This will be undone
8a8d5d85 1005 * by lwkt_reltoken(). Bump the generation number.
f1d1c3fa 1006 */
8a8d5d85 1007 return(++tok->t_gen);
f1d1c3fa
MD
1008}
1009
96728c05
MD
1010/*
1011 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1012 * failure.
1013 */
1014int
1015lwkt_trytoken(lwkt_token_t tok)
1016{
1017 crit_enter();
1018#ifdef SMP
1019 if (tok->t_cpu != mycpu->gd_cpuid) {
1020 return(0);
1021 }
1022#endif
1023 /* leave us in the critical section */
1024 ++tok->t_gen;
1025 return(1);
1026}
1027
f1d1c3fa
MD
1028/*
1029 * Release your ownership of a token. Releases must occur in reverse
1030 * order to aquisitions, eventually so priorities can be unwound properly
1031 * like SPLs. At the moment the actual implemention doesn't care.
1032 *
1033 * We can safely hand a token that we own to another cpu without notifying
1034 * it, but once we do we can't get it back without requesting it (unless
1035 * the other cpu hands it back to us before we check).
1036 *
1037 * We might have lost the token, so check that.
1038 */
1039void
1040lwkt_reltoken(lwkt_token_t tok)
1041{
d0e06f83 1042 if (tok->t_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
1043 tok->t_cpu = tok->t_reqcpu;
1044 }
1045 crit_exit();
1046}
1047
1048/*
8a8d5d85
MD
1049 * Reacquire a token that might have been lost and compare and update the
1050 * generation number. 0 is returned if the generation has not changed
1051 * (nobody else obtained the token while we were blocked, on this cpu or
1052 * any other cpu).
1053 *
1054 * This function returns with the token re-held whether the generation
1055 * number changed or not.
1056 */
1057int
1058lwkt_gentoken(lwkt_token_t tok, int *gen)
1059{
1060 if (lwkt_regettoken(tok) == *gen) {
1061 return(0);
1062 } else {
1063 *gen = tok->t_gen;
1064 return(-1);
1065 }
1066}
1067
1068
1069/*
96728c05 1070 * Re-acquire a token that might have been lost. Returns the generation
8a8d5d85 1071 * number of the token.
f1d1c3fa
MD
1072 */
1073int
1074lwkt_regettoken(lwkt_token_t tok)
1075{
96728c05 1076 /* assert we are in a critical section */
d0e06f83 1077 if (tok->t_cpu != mycpu->gd_cpuid) {
96728c05 1078#ifdef SMP
d0e06f83 1079 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1080 struct lwkt_gettoken_req req;
1081 int seq;
96728c05 1082 int dcpu;
57c254db 1083
96728c05
MD
1084 req.cpu = mycpu->gd_cpuid;
1085 req.tok = tok;
1086 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1087 KKASSERT(dcpu >= 0 && dcpu < ncpus);
cb973d15
MD
1088 if (token_debug)
1089 printf("REQT%d ", dcpu);
96728c05
MD
1090 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1091 lwkt_wait_ipiq(dcpu, seq);
cb973d15
MD
1092 if (token_debug)
1093 printf("REQR%d ", tok->t_cpu);
f1d1c3fa 1094 }
f1d1c3fa 1095#endif
96728c05
MD
1096 ++tok->t_gen;
1097 }
8a8d5d85 1098 return(tok->t_gen);
8ad65e08
MD
1099}
1100
72740893
MD
1101void
1102lwkt_inittoken(lwkt_token_t tok)
1103{
1104 /*
1105 * Zero structure and set cpu owner and reqcpu to cpu 0.
1106 */
1107 bzero(tok, sizeof(*tok));
1108}
1109
99df837e
MD
1110/*
1111 * Create a kernel process/thread/whatever. It shares it's address space
1112 * with proc0 - ie: kernel only.
1113 *
1114 * XXX should be renamed to lwkt_create()
8a8d5d85
MD
1115 *
1116 * The thread will be entered with the MP lock held.
99df837e
MD
1117 */
1118int
1119lwkt_create(void (*func)(void *), void *arg,
73e4f7b9 1120 struct thread **tdp, thread_t template, int tdflags,
ef0fdad1 1121 const char *fmt, ...)
99df837e 1122{
73e4f7b9 1123 thread_t td;
99df837e
MD
1124 va_list ap;
1125
a2a5ad0d
MD
1126 td = lwkt_alloc_thread(template);
1127 if (tdp)
1128 *tdp = td;
99df837e 1129 cpu_set_thread_handler(td, kthread_exit, func, arg);
ef0fdad1 1130 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1131#ifdef SMP
1132 td->td_mpcount = 1;
1133#endif
99df837e
MD
1134
1135 /*
1136 * Set up arg0 for 'ps' etc
1137 */
1138 va_start(ap, fmt);
1139 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1140 va_end(ap);
1141
1142 /*
1143 * Schedule the thread to run
1144 */
ef0fdad1
MD
1145 if ((td->td_flags & TDF_STOPREQ) == 0)
1146 lwkt_schedule(td);
1147 else
1148 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1149 return 0;
1150}
1151
1152/*
1153 * Destroy an LWKT thread. Warning! This function is not called when
1154 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1155 * uses a different reaping mechanism.
1156 */
1157void
1158lwkt_exit(void)
1159{
1160 thread_t td = curthread;
1161
1162 if (td->td_flags & TDF_VERBOSE)
1163 printf("kthread %p %s has exited\n", td, td->td_comm);
1164 crit_enter();
1165 lwkt_deschedule_self();
1166 ++mycpu->gd_tdfreecount;
1167 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1168 cpu_thread_exit();
1169}
1170
1171/*
1172 * Create a kernel process/thread/whatever. It shares it's address space
ef0fdad1 1173 * with proc0 - ie: kernel only. 5.x compatible.
99df837e
MD
1174 */
1175int
1176kthread_create(void (*func)(void *), void *arg,
1177 struct thread **tdp, const char *fmt, ...)
1178{
73e4f7b9 1179 thread_t td;
99df837e
MD
1180 va_list ap;
1181
a2a5ad0d
MD
1182 td = lwkt_alloc_thread(NULL);
1183 if (tdp)
1184 *tdp = td;
99df837e
MD
1185 cpu_set_thread_handler(td, kthread_exit, func, arg);
1186 td->td_flags |= TDF_VERBOSE;
8a8d5d85
MD
1187#ifdef SMP
1188 td->td_mpcount = 1;
1189#endif
99df837e
MD
1190
1191 /*
1192 * Set up arg0 for 'ps' etc
1193 */
1194 va_start(ap, fmt);
1195 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1196 va_end(ap);
1197
1198 /*
1199 * Schedule the thread to run
1200 */
1201 lwkt_schedule(td);
1202 return 0;
1203}
1204
26a0694b
MD
1205void
1206crit_panic(void)
1207{
73e4f7b9 1208 thread_t td = curthread;
26a0694b
MD
1209 int lpri = td->td_pri;
1210
1211 td->td_pri = 0;
1212 panic("td_pri is/would-go negative! %p %d", td, lpri);
1213}
1214
99df837e
MD
1215/*
1216 * Destroy an LWKT thread. Warning! This function is not called when
1217 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1218 * uses a different reaping mechanism.
1219 *
1220 * XXX duplicates lwkt_exit()
1221 */
1222void
1223kthread_exit(void)
1224{
1225 lwkt_exit();
1226}
1227
96728c05
MD
1228#ifdef SMP
1229
1230/*
1231 * Send a function execution request to another cpu. The request is queued
1232 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1233 * possible target cpu. The FIFO can be written.
1234 *
1235 * YYY If the FIFO fills up we have to enable interrupts and process the
1236 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1237 * Create a CPU_*() function to do this!
1238 *
1239 * Must be called from a critical section.
1240 */
1241int
1242lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1243{
1244 lwkt_ipiq_t ip;
1245 int windex;
a2a5ad0d 1246 struct globaldata *gd = mycpu;
96728c05 1247
a2a5ad0d 1248 if (dcpu == gd->gd_cpuid) {
96728c05
MD
1249 func(arg);
1250 return(0);
1251 }
cb973d15 1252 crit_enter();
a2a5ad0d
MD
1253 ++gd->gd_intr_nesting_level;
1254#ifdef INVARIANTS
1255 if (gd->gd_intr_nesting_level > 20)
1256 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1257#endif
96728c05
MD
1258 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1259 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1260 ++ipiq_count;
a2a5ad0d 1261 ip = &gd->gd_ipiq[dcpu];
cb973d15
MD
1262
1263 /*
1264 * We always drain before the FIFO becomes full so it should never
1265 * become full. We need to leave enough entries to deal with
1266 * reentrancy.
1267 */
1268 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1269 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1270 ip->ip_func[windex] = func;
1271 ip->ip_arg[windex] = arg;
1272 /* YYY memory barrier */
1273 ++ip->ip_windex;
96728c05
MD
1274 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1275 unsigned int eflags = read_eflags();
1276 cpu_enable_intr();
1277 ++ipiq_fifofull;
cb973d15 1278 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
96728c05
MD
1279 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1280 lwkt_process_ipiq();
1281 }
1282 write_eflags(eflags);
1283 }
a2a5ad0d 1284 --gd->gd_intr_nesting_level;
96728c05 1285 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
cb973d15 1286 crit_exit();
96728c05
MD
1287 return(ip->ip_windex);
1288}
1289
cb973d15
MD
1290/*
1291 * Send a message to several target cpus. Typically used for scheduling.
1292 */
1293void
1294lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1295{
1296 int cpuid;
1297
1298 while (mask) {
1299 cpuid = bsfl(mask);
1300 lwkt_send_ipiq(cpuid, func, arg);
1301 mask &= ~(1 << cpuid);
1302 }
1303}
1304
96728c05
MD
1305/*
1306 * Wait for the remote cpu to finish processing a function.
1307 *
1308 * YYY we have to enable interrupts and process the IPIQ while waiting
1309 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1310 * function to do this! YYY we really should 'block' here.
1311 *
1312 * Must be called from a critical section. Thsi routine may be called
1313 * from an interrupt (for example, if an interrupt wakes a foreign thread
1314 * up).
1315 */
1316void
1317lwkt_wait_ipiq(int dcpu, int seq)
1318{
1319 lwkt_ipiq_t ip;
a2a5ad0d 1320 int maxc = 100000000;
96728c05
MD
1321
1322 if (dcpu != mycpu->gd_cpuid) {
1323 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1324 ip = &mycpu->gd_ipiq[dcpu];
cb973d15 1325 if ((int)(ip->ip_xindex - seq) < 0) {
96728c05
MD
1326 unsigned int eflags = read_eflags();
1327 cpu_enable_intr();
cb973d15 1328 while ((int)(ip->ip_xindex - seq) < 0) {
96728c05 1329 lwkt_process_ipiq();
a2a5ad0d 1330 if (--maxc == 0)
cb973d15 1331 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
a2a5ad0d
MD
1332 if (maxc < -1000000)
1333 panic("LWKT_WAIT_IPIQ");
96728c05
MD
1334 }
1335 write_eflags(eflags);
1336 }
1337 }
1338}
1339
1340/*
1341 * Called from IPI interrupt (like a fast interrupt), which has placed
1342 * us in a critical section. The MP lock may or may not be held.
cb973d15
MD
1343 * May also be called from doreti or splz, or be reentrantly called
1344 * indirectly through the ip_func[] we run.
96728c05
MD
1345 */
1346void
1347lwkt_process_ipiq(void)
1348{
1349 int n;
1350 int cpuid = mycpu->gd_cpuid;
1351
1352 for (n = 0; n < ncpus; ++n) {
1353 lwkt_ipiq_t ip;
1354 int ri;
1355
1356 if (n == cpuid)
1357 continue;
1358 ip = globaldata_find(n)->gd_ipiq;
1359 if (ip == NULL)
1360 continue;
1361 ip = &ip[cpuid];
cb973d15
MD
1362
1363 /*
1364 * Note: xindex is only updated after we are sure the function has
1365 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1366 * function may send an IPI which may block/drain.
1367 */
96728c05
MD
1368 while (ip->ip_rindex != ip->ip_windex) {
1369 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
96728c05 1370 ++ip->ip_rindex;
cb973d15
MD
1371 ip->ip_func[ri](ip->ip_arg[ri]);
1372 /* YYY memory barrier */
1373 ip->ip_xindex = ip->ip_rindex;
96728c05
MD
1374 }
1375 }
1376}
1377
1378#else
1379
1380int
1381lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1382{
1383 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1384 return(0); /* NOT REACHED */
1385}
1386
1387void
1388lwkt_wait_ipiq(int dcpu, int seq)
1389{
1390 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);
1391}
1392
1393#endif