MP Implmentation 4/4: Final cleanup for this stage. Deal with a race
[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 *
d9eea1a5 31 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.21 2003/07/11 17:42:10 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;
26a0694b
MD
110#if 0
111 /*
112 * YYY needs cli/sti protection? gd_reqpri set by interrupt
113 * when made pending. need better mechanism.
114 */
115 if (gd->gd_reqpri < (td->td_pri & TDPRI_MASK))
116 gd->gd_reqpri = (td->td_pri & TDPRI_MASK);
117#endif
f1d1c3fa
MD
118 }
119}
8ad65e08 120
57c254db
MD
121static __inline
122int
123_lwkt_wantresched(thread_t ntd, thread_t cur)
124{
125 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
126}
127
8ad65e08
MD
128/*
129 * LWKTs operate on a per-cpu basis
130 *
73e4f7b9 131 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
132 */
133void
134lwkt_gdinit(struct globaldata *gd)
135{
4b5f931b
MD
136 int i;
137
138 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
139 TAILQ_INIT(&gd->gd_tdrunq[i]);
140 gd->gd_runqmask = 0;
73e4f7b9 141 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
142}
143
7d0bac62
MD
144/*
145 * Initialize a thread wait structure prior to first use.
146 *
147 * NOTE! called from low level boot code, we cannot do anything fancy!
148 */
149void
150lwkt_init_wait(lwkt_wait_t w)
151{
152 TAILQ_INIT(&w->wa_waitq);
153}
154
155/*
156 * Create a new thread. The thread must be associated with a process context
157 * or LWKT start address before it can be scheduled.
0cfcada1
MD
158 *
159 * If you intend to create a thread without a process context this function
160 * does everything except load the startup and switcher function.
7d0bac62
MD
161 */
162thread_t
ef0fdad1 163lwkt_alloc_thread(struct thread *td)
7d0bac62 164{
99df837e 165 void *stack;
ef0fdad1 166 int flags = 0;
7d0bac62 167
ef0fdad1 168 if (td == NULL) {
26a0694b 169 crit_enter();
ef0fdad1
MD
170 if (mycpu->gd_tdfreecount > 0) {
171 --mycpu->gd_tdfreecount;
172 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
d9eea1a5 173 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1
MD
174 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
175 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
176 crit_exit();
177 stack = td->td_kstack;
178 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
179 } else {
180 crit_exit();
181 td = zalloc(thread_zone);
182 td->td_kstack = NULL;
183 flags |= TDF_ALLOCATED_THREAD;
184 }
185 }
186 if ((stack = td->td_kstack) == NULL) {
99df837e 187 stack = (void *)kmem_alloc(kernel_map, UPAGES * PAGE_SIZE);
ef0fdad1 188 flags |= TDF_ALLOCATED_STACK;
99df837e 189 }
26a0694b 190 lwkt_init_thread(td, stack, flags, mycpu);
99df837e 191 return(td);
7d0bac62
MD
192}
193
194/*
195 * Initialize a preexisting thread structure. This function is used by
196 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
197 *
198 * NOTE! called from low level boot code, we cannot do anything fancy!
199 */
200void
26a0694b 201lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
7d0bac62 202{
99df837e
MD
203 bzero(td, sizeof(struct thread));
204 td->td_kstack = stack;
205 td->td_flags |= flags;
26a0694b
MD
206 td->td_gd = gd;
207 td->td_pri = TDPRI_CRIT;
d9eea1a5 208 td->td_cpu = gd->gd_cpuid; /* YYY don't really need this if have td_gd */
99df837e 209 pmap_init_thread(td);
73e4f7b9
MD
210 crit_enter();
211 TAILQ_INSERT_TAIL(&mycpu->gd_tdallq, td, td_allq);
212 crit_exit();
213}
214
215void
216lwkt_set_comm(thread_t td, const char *ctl, ...)
217{
218 va_list va;
219
220 va_start(va, ctl);
221 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
222 va_end(va);
7d0bac62
MD
223}
224
99df837e 225void
73e4f7b9 226lwkt_hold(thread_t td)
99df837e 227{
73e4f7b9
MD
228 ++td->td_refs;
229}
230
231void
232lwkt_rele(thread_t td)
233{
234 KKASSERT(td->td_refs > 0);
235 --td->td_refs;
236}
237
238void
239lwkt_wait_free(thread_t td)
240{
241 while (td->td_refs)
242 tsleep(td, PWAIT, "tdreap", hz);
243}
244
245void
246lwkt_free_thread(thread_t td)
247{
248 struct globaldata *gd = mycpu;
249
d9eea1a5 250 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
251 ("lwkt_free_thread: did not exit! %p", td));
252
253 crit_enter();
73e4f7b9
MD
254 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
255 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
256 (td->td_flags & TDF_ALLOCATED_THREAD)
257 ) {
73e4f7b9
MD
258 ++gd->gd_tdfreecount;
259 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
99df837e
MD
260 crit_exit();
261 } else {
262 crit_exit();
263 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
264 kmem_free(kernel_map,
265 (vm_offset_t)td->td_kstack, UPAGES * PAGE_SIZE);
73e4f7b9 266 /* gd invalid */
99df837e
MD
267 td->td_kstack = NULL;
268 }
269 if (td->td_flags & TDF_ALLOCATED_THREAD)
270 zfree(thread_zone, td);
271 }
272}
273
274
8ad65e08
MD
275/*
276 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
277 * switch to the idlethread. Switching must occur within a critical
278 * section to avoid races with the scheduling queue.
279 *
280 * We always have full control over our cpu's run queue. Other cpus
281 * that wish to manipulate our queue must use the cpu_*msg() calls to
282 * talk to our cpu, so a critical section is all that is needed and
283 * the result is very, very fast thread switching.
284 *
96728c05
MD
285 * The LWKT scheduler uses a fixed priority model and round-robins at
286 * each priority level. User process scheduling is a totally
287 * different beast and LWKT priorities should not be confused with
288 * user process priorities.
f1d1c3fa 289 *
96728c05
MD
290 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
291 * cleans it up. Note that the td_switch() function cannot do anything that
292 * requires the MP lock since the MP lock will have already been setup for
293 * the target thread (not the current thread).
8ad65e08 294 */
96728c05 295
8ad65e08
MD
296void
297lwkt_switch(void)
298{
4b5f931b 299 struct globaldata *gd;
f1d1c3fa 300 thread_t td = curthread;
8ad65e08 301 thread_t ntd;
8a8d5d85
MD
302#ifdef SMP
303 int mpheld;
304#endif
8ad65e08 305
96728c05
MD
306 if (mycpu->gd_intr_nesting_level &&
307 td->td_preempted == NULL && panicstr == NULL
308 ) {
26a0694b 309 panic("lwkt_switch: cannot switch from within an interrupt, yet\n");
96728c05 310 }
ef0fdad1 311
cb973d15
MD
312 /*
313 * Passive release (used to transition from user to kernel mode
314 * when we block or switch rather then when we enter the kernel).
315 * This function is NOT called if we are switching into a preemption
316 * or returning from a preemption. Typically this causes us to lose
317 * our P_CURPROC designation (if we have one) and become a true LWKT
318 * thread, and may also hand P_CURPROC to another process and schedule
319 * its thread.
320 */
321 if (td->td_release)
322 td->td_release(td);
323
f1d1c3fa 324 crit_enter();
4b5f931b 325 ++switch_count;
8a8d5d85
MD
326
327#ifdef SMP
328 /*
329 * td_mpcount cannot be used to determine if we currently hold the
330 * MP lock because get_mplock() will increment it prior to attempting
331 * to get the lock, and switch out if it can't. Look at the actual lock.
332 */
333 mpheld = MP_LOCK_HELD();
334#endif
99df837e
MD
335 if ((ntd = td->td_preempted) != NULL) {
336 /*
337 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
338 * thread. This occurs transparently, whether the preempted thread
339 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
340 * itself).
341 *
342 * We have to setup the MP lock for the original thread after backing
343 * out the adjustment that was made to curthread when the original
344 * was preempted.
99df837e 345 */
26a0694b 346 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 347#ifdef SMP
96728c05
MD
348 if (ntd->td_mpcount && mpheld == 0) {
349 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
350 td, ntd, td->td_mpcount, ntd->td_mpcount);
351 }
8a8d5d85
MD
352 if (ntd->td_mpcount) {
353 td->td_mpcount -= ntd->td_mpcount;
354 KKASSERT(td->td_mpcount >= 0);
355 }
356#endif
26a0694b 357 ntd->td_flags |= TDF_PREEMPT_DONE;
8a8d5d85 358 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 359 } else {
4b5f931b
MD
360 /*
361 * Priority queue / round-robin at each priority. Note that user
362 * processes run at a fixed, low priority and the user process
363 * scheduler deals with interactions between user processes
364 * by scheduling and descheduling them from the LWKT queue as
365 * necessary.
8a8d5d85
MD
366 *
367 * We have to adjust the MP lock for the target thread. If we
368 * need the MP lock and cannot obtain it we try to locate a
369 * thread that does not need the MP lock.
4b5f931b
MD
370 */
371 gd = mycpu;
4b5f931b
MD
372again:
373 if (gd->gd_runqmask) {
374 int nq = bsrl(gd->gd_runqmask);
375 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
376 gd->gd_runqmask &= ~(1 << nq);
377 goto again;
378 }
8a8d5d85
MD
379#ifdef SMP
380 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
381 /*
96728c05
MD
382 * Target needs MP lock and we couldn't get it, try
383 * to locate a thread which does not need the MP lock
384 * to run.
8a8d5d85
MD
385 */
386 u_int32_t rqmask = gd->gd_runqmask;
387 while (rqmask) {
388 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
389 if (ntd->td_mpcount == 0)
390 break;
391 }
392 if (ntd)
393 break;
394 rqmask &= ~(1 << nq);
395 nq = bsrl(rqmask);
396 }
397 if (ntd == NULL) {
a2a5ad0d
MD
398 ntd = &gd->gd_idlethread;
399 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
400 } else {
401 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
402 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
403 }
404 } else {
405 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
406 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
407 }
408#else
4b5f931b
MD
409 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
410 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 411#endif
4b5f931b 412 } else {
a2a5ad0d 413 ntd = &gd->gd_idlethread;
4b5f931b 414 }
f1d1c3fa 415 }
26a0694b
MD
416 KASSERT(ntd->td_pri >= TDPRI_CRIT,
417 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
418
419 /*
420 * Do the actual switch. If the new target does not need the MP lock
421 * and we are holding it, release the MP lock. If the new target requires
422 * the MP lock we have already acquired it for the target.
423 */
424#ifdef SMP
425 if (ntd->td_mpcount == 0 ) {
426 if (MP_LOCK_HELD())
427 cpu_rel_mplock();
428 } else {
429 ASSERT_MP_LOCK_HELD();
430 }
431#endif
8a8d5d85 432 if (td != ntd) {
f1d1c3fa 433 td->td_switch(ntd);
8a8d5d85 434 }
96728c05 435
f1d1c3fa 436 crit_exit();
8ad65e08
MD
437}
438
cb973d15
MD
439/*
440 * Switch if another thread has a higher priority. Do not switch to other
441 * threads at the same priority.
442 */
443void
444lwkt_maybe_switch()
445{
446 struct globaldata *gd = mycpu;
447 struct thread *td = gd->gd_curthread;
448
449 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
450 lwkt_switch();
451 }
452}
453
b68b7282 454/*
96728c05
MD
455 * Request that the target thread preempt the current thread. Preemption
456 * only works under a specific set of conditions:
b68b7282 457 *
96728c05
MD
458 * - We are not preempting ourselves
459 * - The target thread is owned by the current cpu
460 * - We are not currently being preempted
461 * - The target is not currently being preempted
462 * - We are able to satisfy the target's MP lock requirements (if any).
463 *
464 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
465 * this is called via lwkt_schedule() through the td_preemptable callback.
466 * critpri is the managed critical priority that we should ignore in order
467 * to determine whether preemption is possible (aka usually just the crit
468 * priority of lwkt_schedule() itself).
b68b7282 469 *
26a0694b
MD
470 * XXX at the moment we run the target thread in a critical section during
471 * the preemption in order to prevent the target from taking interrupts
472 * that *WE* can't. Preemption is strictly limited to interrupt threads
473 * and interrupt-like threads, outside of a critical section, and the
474 * preempted source thread will be resumed the instant the target blocks
475 * whether or not the source is scheduled (i.e. preemption is supposed to
476 * be as transparent as possible).
4b5f931b 477 *
8a8d5d85
MD
478 * The target thread inherits our MP count (added to its own) for the
479 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
480 * MP lock during the preemption. Therefore, any preempting targets must be
481 * careful in regards to MP assertions. Note that the MP count may be
482 * out of sync with the physical mp_lock. If we preempt we have to preserve
483 * the expected situation.
b68b7282
MD
484 */
485void
96728c05 486lwkt_preempt(thread_t ntd, int critpri)
b68b7282 487{
73e4f7b9 488 thread_t td = curthread;
8a8d5d85
MD
489#ifdef SMP
490 int mpheld;
57c254db 491 int savecnt;
8a8d5d85 492#endif
b68b7282 493
26a0694b 494 /*
96728c05
MD
495 * The caller has put us in a critical section. We can only preempt
496 * if the caller of the caller was not in a critical section (basically
57c254db
MD
497 * a local interrupt), as determined by the 'critpri' parameter. If
498 * we are unable to preempt
96728c05
MD
499 *
500 * YYY The target thread must be in a critical section (else it must
501 * inherit our critical section? I dunno yet).
26a0694b
MD
502 */
503 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 504
cb973d15 505 need_resched();
57c254db
MD
506 if (!_lwkt_wantresched(ntd, td)) {
507 ++preempt_miss;
508 return;
509 }
96728c05
MD
510 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
511 ++preempt_miss;
512 return;
513 }
514#ifdef SMP
515 if (ntd->td_cpu != mycpu->gd_cpuid) {
516 ++preempt_miss;
517 return;
518 }
519#endif
26a0694b
MD
520 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
521 ++preempt_weird;
522 return;
523 }
524 if (ntd->td_preempted) {
4b5f931b 525 ++preempt_hit;
26a0694b 526 return;
b68b7282 527 }
8a8d5d85 528#ifdef SMP
a2a5ad0d
MD
529 /*
530 * note: an interrupt might have occured just as we were transitioning
a5934754
MD
531 * to the MP lock. In this case td_mpcount will be pre-disposed but
532 * not actually synchronized with the actual state of the lock. We
533 * can use it to imply an MP lock requirement for the preemption but
534 * we cannot use it to test whether we hold the MP lock or not.
a2a5ad0d 535 */
a5934754 536 mpheld = MP_LOCK_HELD();
96728c05
MD
537 if (mpheld && td->td_mpcount == 0)
538 panic("lwkt_preempt(): held and no count");
539 savecnt = td->td_mpcount;
8a8d5d85
MD
540 ntd->td_mpcount += td->td_mpcount;
541 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
542 ntd->td_mpcount -= td->td_mpcount;
543 ++preempt_miss;
544 return;
545 }
546#endif
26a0694b
MD
547
548 ++preempt_hit;
549 ntd->td_preempted = td;
550 td->td_flags |= TDF_PREEMPT_LOCK;
551 td->td_switch(ntd);
552 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
553#ifdef SMP
554 KKASSERT(savecnt == td->td_mpcount);
555 if (mpheld == 0 && MP_LOCK_HELD())
556 cpu_rel_mplock();
557 else if (mpheld && !MP_LOCK_HELD())
558 panic("lwkt_preempt(): MP lock was not held through");
559#endif
26a0694b
MD
560 ntd->td_preempted = NULL;
561 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
562}
563
f1d1c3fa
MD
564/*
565 * Yield our thread while higher priority threads are pending. This is
566 * typically called when we leave a critical section but it can be safely
567 * called while we are in a critical section.
568 *
569 * This function will not generally yield to equal priority threads but it
570 * can occur as a side effect. Note that lwkt_switch() is called from
571 * inside the critical section to pervent its own crit_exit() from reentering
572 * lwkt_yield_quick().
573 *
ef0fdad1
MD
574 * gd_reqpri indicates that *something* changed, e.g. an interrupt or softint
575 * came along but was blocked and made pending.
576 *
f1d1c3fa
MD
577 * (self contained on a per cpu basis)
578 */
579void
580lwkt_yield_quick(void)
581{
582 thread_t td = curthread;
ef0fdad1 583
a2a5ad0d
MD
584 /*
585 * gd_reqpri is cleared in splz if the cpl is 0. If we were to clear
586 * it with a non-zero cpl then we might not wind up calling splz after
587 * a task switch when the critical section is exited even though the
588 * new task could accept the interrupt. YYY alternative is to have
589 * lwkt_switch() just call splz unconditionally.
590 *
591 * XXX from crit_exit() only called after last crit section is released.
592 * If called directly will run splz() even if in a critical section.
593 */
ef0fdad1 594 if ((td->td_pri & TDPRI_MASK) < mycpu->gd_reqpri) {
f1d1c3fa
MD
595 splz();
596 }
597
598 /*
599 * YYY enabling will cause wakeup() to task-switch, which really
600 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
601 * preemption and MP without actually doing preemption or MP, because a
602 * lot of code assumes that wakeup() does not block.
f1d1c3fa 603 */
ef0fdad1 604 if (untimely_switch && mycpu->gd_intr_nesting_level == 0) {
f1d1c3fa
MD
605 crit_enter();
606 /*
607 * YYY temporary hacks until we disassociate the userland scheduler
608 * from the LWKT scheduler.
609 */
610 if (td->td_flags & TDF_RUNQ) {
611 lwkt_switch(); /* will not reenter yield function */
612 } else {
613 lwkt_schedule_self(); /* make sure we are scheduled */
614 lwkt_switch(); /* will not reenter yield function */
615 lwkt_deschedule_self(); /* make sure we are descheduled */
616 }
617 crit_exit_noyield();
618 }
f1d1c3fa
MD
619}
620
8ad65e08 621/*
f1d1c3fa
MD
622 * This implements a normal yield which, unlike _quick, will yield to equal
623 * priority threads as well. Note that gd_reqpri tests will be handled by
624 * the crit_exit() call in lwkt_switch().
625 *
626 * (self contained on a per cpu basis)
8ad65e08
MD
627 */
628void
f1d1c3fa 629lwkt_yield(void)
8ad65e08 630{
f1d1c3fa
MD
631 lwkt_schedule_self();
632 lwkt_switch();
633}
634
635/*
636 * Schedule a thread to run. As the current thread we can always safely
637 * schedule ourselves, and a shortcut procedure is provided for that
638 * function.
639 *
640 * (non-blocking, self contained on a per cpu basis)
641 */
642void
643lwkt_schedule_self(void)
644{
645 thread_t td = curthread;
646
647 crit_enter();
648 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa 649 _lwkt_enqueue(td);
26a0694b
MD
650 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
651 panic("SCHED SELF PANIC");
f1d1c3fa 652 crit_exit();
8ad65e08 653}
8ad65e08
MD
654
655/*
f1d1c3fa
MD
656 * Generic schedule. Possibly schedule threads belonging to other cpus and
657 * deal with threads that might be blocked on a wait queue.
658 *
96728c05 659 * YYY this is one of the best places to implement load balancing code.
f1d1c3fa
MD
660 * Load balancing can be accomplished by requesting other sorts of actions
661 * for the thread in question.
8ad65e08
MD
662 */
663void
664lwkt_schedule(thread_t td)
665{
96728c05 666#ifdef INVARIANTS
26a0694b
MD
667 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
668 && td->td_proc->p_stat == SSLEEP
669 ) {
670 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
671 curthread,
672 curthread->td_proc ? curthread->td_proc->p_pid : -1,
673 curthread->td_proc ? curthread->td_proc->p_stat : -1,
674 td,
675 td->td_proc ? curthread->td_proc->p_pid : -1,
676 td->td_proc ? curthread->td_proc->p_stat : -1
677 );
678 panic("SCHED PANIC");
679 }
96728c05 680#endif
f1d1c3fa
MD
681 crit_enter();
682 if (td == curthread) {
683 _lwkt_enqueue(td);
684 } else {
685 lwkt_wait_t w;
686
687 /*
688 * If the thread is on a wait list we have to send our scheduling
689 * request to the owner of the wait structure. Otherwise we send
690 * the scheduling request to the cpu owning the thread. Races
691 * are ok, the target will forward the message as necessary (the
692 * message may chase the thread around before it finally gets
693 * acted upon).
694 *
695 * (remember, wait structures use stable storage)
696 */
697 if ((w = td->td_wait) != NULL) {
96728c05 698 if (lwkt_trytoken(&w->wa_token)) {
f1d1c3fa
MD
699 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
700 --w->wa_count;
701 td->td_wait = NULL;
d0e06f83 702 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa 703 _lwkt_enqueue(td);
57c254db 704 if (td->td_preemptable) {
96728c05 705 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
57c254db
MD
706 } else if (_lwkt_wantresched(td, curthread)) {
707 need_resched();
708 }
f1d1c3fa 709 } else {
96728c05 710 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 711 }
96728c05 712 lwkt_reltoken(&w->wa_token);
f1d1c3fa 713 } else {
96728c05 714 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
715 }
716 } else {
717 /*
718 * If the wait structure is NULL and we own the thread, there
719 * is no race (since we are in a critical section). If we
720 * do not own the thread there might be a race but the
721 * target cpu will deal with it.
722 */
d0e06f83 723 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa 724 _lwkt_enqueue(td);
57c254db 725 if (td->td_preemptable) {
96728c05 726 td->td_preemptable(td, TDPRI_CRIT);
57c254db
MD
727 } else if (_lwkt_wantresched(td, curthread)) {
728 need_resched();
729 }
f1d1c3fa 730 } else {
96728c05 731 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
732 }
733 }
8ad65e08 734 }
f1d1c3fa 735 crit_exit();
8ad65e08
MD
736}
737
d9eea1a5
MD
738/*
739 * Managed acquisition. This code assumes that the MP lock is held for
740 * the tdallq operation and that the thread has been descheduled from its
741 * original cpu. We also have to wait for the thread to be entirely switched
742 * out on its original cpu (this is usually fast enough that we never loop)
743 * since the LWKT system does not have to hold the MP lock while switching
744 * and the target may have released it before switching.
745 */
a2a5ad0d
MD
746void
747lwkt_acquire(thread_t td)
748{
749 struct globaldata *gd;
cb973d15 750 int ocpu;
a2a5ad0d
MD
751
752 gd = td->td_gd;
753 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5
MD
754 while (td->td_flags & TDF_RUNNING) /* XXX spin */
755 ;
a2a5ad0d 756 if (gd != mycpu) {
cb973d15 757 ocpu = td->td_cpu;
a2a5ad0d
MD
758 crit_enter();
759 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
760 gd = mycpu;
761 td->td_gd = gd;
762 td->td_cpu = gd->gd_cpuid;
763 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
764 crit_exit();
765 }
766}
767
8ad65e08 768/*
f1d1c3fa
MD
769 * Deschedule a thread.
770 *
771 * (non-blocking, self contained on a per cpu basis)
772 */
773void
774lwkt_deschedule_self(void)
775{
776 thread_t td = curthread;
777
778 crit_enter();
779 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa
MD
780 _lwkt_dequeue(td);
781 crit_exit();
782}
783
784/*
785 * Generic deschedule. Descheduling threads other then your own should be
786 * done only in carefully controlled circumstances. Descheduling is
787 * asynchronous.
788 *
789 * This function may block if the cpu has run out of messages.
8ad65e08
MD
790 */
791void
792lwkt_deschedule(thread_t td)
793{
f1d1c3fa
MD
794 crit_enter();
795 if (td == curthread) {
796 _lwkt_dequeue(td);
797 } else {
d0e06f83 798 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
799 _lwkt_dequeue(td);
800 } else {
96728c05 801 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
802 }
803 }
804 crit_exit();
805}
806
4b5f931b
MD
807/*
808 * Set the target thread's priority. This routine does not automatically
809 * switch to a higher priority thread, LWKT threads are not designed for
810 * continuous priority changes. Yield if you want to switch.
811 *
812 * We have to retain the critical section count which uses the high bits
26a0694b
MD
813 * of the td_pri field. The specified priority may also indicate zero or
814 * more critical sections by adding TDPRI_CRIT*N.
4b5f931b
MD
815 */
816void
817lwkt_setpri(thread_t td, int pri)
818{
26a0694b 819 KKASSERT(pri >= 0);
57c254db 820 KKASSERT(td->td_cpu == mycpu->gd_cpuid);
26a0694b
MD
821 crit_enter();
822 if (td->td_flags & TDF_RUNQ) {
823 _lwkt_dequeue(td);
824 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
825 _lwkt_enqueue(td);
826 } else {
827 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
828 }
829 crit_exit();
830}
831
832void
833lwkt_setpri_self(int pri)
834{
835 thread_t td = curthread;
836
4b5f931b
MD
837 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
838 crit_enter();
839 if (td->td_flags & TDF_RUNQ) {
840 _lwkt_dequeue(td);
841 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
842 _lwkt_enqueue(td);
843 } else {
844 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
845 }
846 crit_exit();
847}
848
849struct proc *
850lwkt_preempted_proc(void)
851{
73e4f7b9 852 thread_t td = curthread;
4b5f931b
MD
853 while (td->td_preempted)
854 td = td->td_preempted;
855 return(td->td_proc);
856}
857
858
f1d1c3fa
MD
859/*
860 * This function deschedules the current thread and blocks on the specified
861 * wait queue. We obtain ownership of the wait queue in order to block
862 * on it. A generation number is used to interlock the wait queue in case
863 * it gets signalled while we are blocked waiting on the token.
864 *
865 * Note: alternatively we could dequeue our thread and then message the
866 * target cpu owning the wait queue. YYY implement as sysctl.
867 *
868 * Note: wait queue signals normally ping-pong the cpu as an optimization.
869 */
96728c05
MD
870typedef struct lwkt_gettoken_req {
871 lwkt_token_t tok;
872 int cpu;
873} lwkt_gettoken_req;
874
f1d1c3fa 875void
ae8050a4 876lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
877{
878 thread_t td = curthread;
f1d1c3fa 879
f1d1c3fa 880 lwkt_gettoken(&w->wa_token);
ae8050a4 881 if (w->wa_gen == *gen) {
f1d1c3fa
MD
882 _lwkt_dequeue(td);
883 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
884 ++w->wa_count;
885 td->td_wait = w;
ae8050a4 886 td->td_wmesg = wmesg;
f1d1c3fa 887 lwkt_switch();
8ad65e08 888 }
ae8050a4
MD
889 /* token might be lost, doesn't matter for gen update */
890 *gen = w->wa_gen;
f1d1c3fa
MD
891 lwkt_reltoken(&w->wa_token);
892}
893
894/*
895 * Signal a wait queue. We gain ownership of the wait queue in order to
896 * signal it. Once a thread is removed from the wait queue we have to
897 * deal with the cpu owning the thread.
898 *
899 * Note: alternatively we could message the target cpu owning the wait
900 * queue. YYY implement as sysctl.
901 */
902void
903lwkt_signal(lwkt_wait_t w)
904{
905 thread_t td;
906 int count;
907
908 lwkt_gettoken(&w->wa_token);
909 ++w->wa_gen;
910 count = w->wa_count;
911 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
912 --count;
913 --w->wa_count;
914 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
915 td->td_wait = NULL;
ae8050a4 916 td->td_wmesg = NULL;
d0e06f83 917 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
918 _lwkt_enqueue(td);
919 } else {
96728c05 920 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
921 }
922 lwkt_regettoken(&w->wa_token);
923 }
924 lwkt_reltoken(&w->wa_token);
925}
926
927/*
96728c05 928 * Acquire ownership of a token
f1d1c3fa 929 *
96728c05 930 * Acquire ownership of a token. The token may have spl and/or critical
f1d1c3fa
MD
931 * section side effects, depending on its purpose. These side effects
932 * guarentee that you will maintain ownership of the token as long as you
933 * do not block. If you block you may lose access to the token (but you
934 * must still release it even if you lose your access to it).
935 *
96728c05 936 * YYY for now we use a critical section to prevent IPIs from taking away
a2a5ad0d 937 * a token, but do we really only need to disable IPIs ?
96728c05
MD
938 *
939 * YYY certain tokens could be made to act like mutexes when performance
940 * would be better (e.g. t_cpu == -1). This is not yet implemented.
941 *
a2a5ad0d
MD
942 * YYY the tokens replace 4.x's simplelocks for the most part, but this
943 * means that 4.x does not expect a switch so for now we cannot switch
944 * when waiting for an IPI to be returned.
945 *
946 * YYY If the token is owned by another cpu we may have to send an IPI to
96728c05
MD
947 * it and then block. The IPI causes the token to be given away to the
948 * requesting cpu, unless it has already changed hands. Since only the
949 * current cpu can give away a token it owns we do not need a memory barrier.
a2a5ad0d 950 * This needs serious optimization.
f1d1c3fa 951 */
57c254db
MD
952
953#ifdef SMP
954
96728c05
MD
955static
956void
957lwkt_gettoken_remote(void *arg)
958{
959 lwkt_gettoken_req *req = arg;
960 if (req->tok->t_cpu == mycpu->gd_cpuid) {
a2a5ad0d
MD
961 if (token_debug)
962 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
96728c05 963 req->tok->t_cpu = req->cpu;
a2a5ad0d
MD
964 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
965 /* else set reqcpu to point to current cpu for release */
96728c05
MD
966 }
967}
968
57c254db
MD
969#endif
970
8a8d5d85 971int
f1d1c3fa
MD
972lwkt_gettoken(lwkt_token_t tok)
973{
974 /*
975 * Prevent preemption so the token can't be taken away from us once
976 * we gain ownership of it. Use a synchronous request which might
977 * block. The request will be forwarded as necessary playing catchup
978 * to the token.
979 */
96728c05 980
f1d1c3fa 981 crit_enter();
57c254db 982#ifdef INVARIANTS
a2a5ad0d
MD
983 if (curthread->td_pri > 2000) {
984 curthread->td_pri = 1000;
985 panic("too HIGH!");
57c254db
MD
986 }
987#endif
96728c05 988#ifdef SMP
d0e06f83 989 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
990 struct lwkt_gettoken_req req;
991 int seq;
96728c05
MD
992 int dcpu;
993
994 req.cpu = mycpu->gd_cpuid;
995 req.tok = tok;
996 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d
MD
997 KKASSERT(dcpu >= 0 && dcpu < ncpus);
998 if (token_debug)
999 printf("REQT%d ", dcpu);
96728c05
MD
1000 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1001 lwkt_wait_ipiq(dcpu, seq);
a2a5ad0d
MD
1002 if (token_debug)
1003 printf("REQR%d ", tok->t_cpu);
f1d1c3fa
MD
1004 }
1005#endif
1006 /*
1007 * leave us in a critical section on return. This will be undone
8a8d5d85 1008 * by lwkt_reltoken(). Bump the generation number.
f1d1c3fa 1009 */
8a8d5d85 1010 return(++tok->t_gen);
f1d1c3fa
MD
1011}
1012
96728c05
MD
1013/*
1014 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1015 * failure.
1016 */
1017int
1018lwkt_trytoken(lwkt_token_t tok)
1019{
1020 crit_enter();
1021#ifdef SMP
1022 if (tok->t_cpu != mycpu->gd_cpuid) {
1023 return(0);
1024 }
1025#endif
1026 /* leave us in the critical section */
1027 ++tok->t_gen;
1028 return(1);
1029}
1030
f1d1c3fa
MD
1031/*
1032 * Release your ownership of a token. Releases must occur in reverse
1033 * order to aquisitions, eventually so priorities can be unwound properly
1034 * like SPLs. At the moment the actual implemention doesn't care.
1035 *
1036 * We can safely hand a token that we own to another cpu without notifying
1037 * it, but once we do we can't get it back without requesting it (unless
1038 * the other cpu hands it back to us before we check).
1039 *
1040 * We might have lost the token, so check that.
1041 */
1042void
1043lwkt_reltoken(lwkt_token_t tok)
1044{
d0e06f83 1045 if (tok->t_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
1046 tok->t_cpu = tok->t_reqcpu;
1047 }
1048 crit_exit();
1049}
1050
1051/*
8a8d5d85
MD
1052 * Reacquire a token that might have been lost and compare and update the
1053 * generation number. 0 is returned if the generation has not changed
1054 * (nobody else obtained the token while we were blocked, on this cpu or
1055 * any other cpu).
1056 *
1057 * This function returns with the token re-held whether the generation
1058 * number changed or not.
1059 */
1060int
1061lwkt_gentoken(lwkt_token_t tok, int *gen)
1062{
1063 if (lwkt_regettoken(tok) == *gen) {
1064 return(0);
1065 } else {
1066 *gen = tok->t_gen;
1067 return(-1);
1068 }
1069}
1070
1071
1072/*
96728c05 1073 * Re-acquire a token that might have been lost. Returns the generation
8a8d5d85 1074 * number of the token.
f1d1c3fa
MD
1075 */
1076int
1077lwkt_regettoken(lwkt_token_t tok)
1078{
96728c05 1079 /* assert we are in a critical section */
d0e06f83 1080 if (tok->t_cpu != mycpu->gd_cpuid) {
96728c05 1081#ifdef SMP
d0e06f83 1082 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1083 struct lwkt_gettoken_req req;
1084 int seq;
96728c05 1085 int dcpu;
57c254db 1086
96728c05
MD
1087 req.cpu = mycpu->gd_cpuid;
1088 req.tok = tok;
1089 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1090 KKASSERT(dcpu >= 0 && dcpu < ncpus);
cb973d15
MD
1091 if (token_debug)
1092 printf("REQT%d ", dcpu);
96728c05
MD
1093 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1094 lwkt_wait_ipiq(dcpu, seq);
cb973d15
MD
1095 if (token_debug)
1096 printf("REQR%d ", tok->t_cpu);
f1d1c3fa 1097 }
f1d1c3fa 1098#endif
96728c05
MD
1099 ++tok->t_gen;
1100 }
8a8d5d85 1101 return(tok->t_gen);
8ad65e08
MD
1102}
1103
72740893
MD
1104void
1105lwkt_inittoken(lwkt_token_t tok)
1106{
1107 /*
1108 * Zero structure and set cpu owner and reqcpu to cpu 0.
1109 */
1110 bzero(tok, sizeof(*tok));
1111}
1112
99df837e
MD
1113/*
1114 * Create a kernel process/thread/whatever. It shares it's address space
1115 * with proc0 - ie: kernel only.
1116 *
1117 * XXX should be renamed to lwkt_create()
8a8d5d85
MD
1118 *
1119 * The thread will be entered with the MP lock held.
99df837e
MD
1120 */
1121int
1122lwkt_create(void (*func)(void *), void *arg,
73e4f7b9 1123 struct thread **tdp, thread_t template, int tdflags,
ef0fdad1 1124 const char *fmt, ...)
99df837e 1125{
73e4f7b9 1126 thread_t td;
99df837e
MD
1127 va_list ap;
1128
a2a5ad0d
MD
1129 td = lwkt_alloc_thread(template);
1130 if (tdp)
1131 *tdp = td;
99df837e 1132 cpu_set_thread_handler(td, kthread_exit, func, arg);
ef0fdad1 1133 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1134#ifdef SMP
1135 td->td_mpcount = 1;
1136#endif
99df837e
MD
1137
1138 /*
1139 * Set up arg0 for 'ps' etc
1140 */
1141 va_start(ap, fmt);
1142 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1143 va_end(ap);
1144
1145 /*
1146 * Schedule the thread to run
1147 */
ef0fdad1
MD
1148 if ((td->td_flags & TDF_STOPREQ) == 0)
1149 lwkt_schedule(td);
1150 else
1151 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1152 return 0;
1153}
1154
1155/*
1156 * Destroy an LWKT thread. Warning! This function is not called when
1157 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1158 * uses a different reaping mechanism.
1159 */
1160void
1161lwkt_exit(void)
1162{
1163 thread_t td = curthread;
1164
1165 if (td->td_flags & TDF_VERBOSE)
1166 printf("kthread %p %s has exited\n", td, td->td_comm);
1167 crit_enter();
1168 lwkt_deschedule_self();
1169 ++mycpu->gd_tdfreecount;
1170 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1171 cpu_thread_exit();
1172}
1173
1174/*
1175 * Create a kernel process/thread/whatever. It shares it's address space
ef0fdad1 1176 * with proc0 - ie: kernel only. 5.x compatible.
99df837e
MD
1177 */
1178int
1179kthread_create(void (*func)(void *), void *arg,
1180 struct thread **tdp, const char *fmt, ...)
1181{
73e4f7b9 1182 thread_t td;
99df837e
MD
1183 va_list ap;
1184
a2a5ad0d
MD
1185 td = lwkt_alloc_thread(NULL);
1186 if (tdp)
1187 *tdp = td;
99df837e
MD
1188 cpu_set_thread_handler(td, kthread_exit, func, arg);
1189 td->td_flags |= TDF_VERBOSE;
8a8d5d85
MD
1190#ifdef SMP
1191 td->td_mpcount = 1;
1192#endif
99df837e
MD
1193
1194 /*
1195 * Set up arg0 for 'ps' etc
1196 */
1197 va_start(ap, fmt);
1198 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1199 va_end(ap);
1200
1201 /*
1202 * Schedule the thread to run
1203 */
1204 lwkt_schedule(td);
1205 return 0;
1206}
1207
26a0694b
MD
1208void
1209crit_panic(void)
1210{
73e4f7b9 1211 thread_t td = curthread;
26a0694b
MD
1212 int lpri = td->td_pri;
1213
1214 td->td_pri = 0;
1215 panic("td_pri is/would-go negative! %p %d", td, lpri);
1216}
1217
99df837e
MD
1218/*
1219 * Destroy an LWKT thread. Warning! This function is not called when
1220 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1221 * uses a different reaping mechanism.
1222 *
1223 * XXX duplicates lwkt_exit()
1224 */
1225void
1226kthread_exit(void)
1227{
1228 lwkt_exit();
1229}
1230
96728c05
MD
1231#ifdef SMP
1232
1233/*
1234 * Send a function execution request to another cpu. The request is queued
1235 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1236 * possible target cpu. The FIFO can be written.
1237 *
1238 * YYY If the FIFO fills up we have to enable interrupts and process the
1239 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1240 * Create a CPU_*() function to do this!
1241 *
1242 * Must be called from a critical section.
1243 */
1244int
1245lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1246{
1247 lwkt_ipiq_t ip;
1248 int windex;
a2a5ad0d 1249 struct globaldata *gd = mycpu;
96728c05 1250
a2a5ad0d 1251 if (dcpu == gd->gd_cpuid) {
96728c05
MD
1252 func(arg);
1253 return(0);
1254 }
cb973d15 1255 crit_enter();
a2a5ad0d
MD
1256 ++gd->gd_intr_nesting_level;
1257#ifdef INVARIANTS
1258 if (gd->gd_intr_nesting_level > 20)
1259 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1260#endif
96728c05
MD
1261 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1262 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1263 ++ipiq_count;
a2a5ad0d 1264 ip = &gd->gd_ipiq[dcpu];
cb973d15
MD
1265
1266 /*
1267 * We always drain before the FIFO becomes full so it should never
1268 * become full. We need to leave enough entries to deal with
1269 * reentrancy.
1270 */
1271 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1272 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1273 ip->ip_func[windex] = func;
1274 ip->ip_arg[windex] = arg;
1275 /* YYY memory barrier */
1276 ++ip->ip_windex;
96728c05
MD
1277 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1278 unsigned int eflags = read_eflags();
1279 cpu_enable_intr();
1280 ++ipiq_fifofull;
cb973d15 1281 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
96728c05
MD
1282 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1283 lwkt_process_ipiq();
1284 }
1285 write_eflags(eflags);
1286 }
a2a5ad0d 1287 --gd->gd_intr_nesting_level;
96728c05 1288 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
cb973d15 1289 crit_exit();
96728c05
MD
1290 return(ip->ip_windex);
1291}
1292
cb973d15
MD
1293/*
1294 * Send a message to several target cpus. Typically used for scheduling.
1295 */
1296void
1297lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1298{
1299 int cpuid;
1300
1301 while (mask) {
1302 cpuid = bsfl(mask);
1303 lwkt_send_ipiq(cpuid, func, arg);
1304 mask &= ~(1 << cpuid);
1305 }
1306}
1307
96728c05
MD
1308/*
1309 * Wait for the remote cpu to finish processing a function.
1310 *
1311 * YYY we have to enable interrupts and process the IPIQ while waiting
1312 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1313 * function to do this! YYY we really should 'block' here.
1314 *
1315 * Must be called from a critical section. Thsi routine may be called
1316 * from an interrupt (for example, if an interrupt wakes a foreign thread
1317 * up).
1318 */
1319void
1320lwkt_wait_ipiq(int dcpu, int seq)
1321{
1322 lwkt_ipiq_t ip;
a2a5ad0d 1323 int maxc = 100000000;
96728c05
MD
1324
1325 if (dcpu != mycpu->gd_cpuid) {
1326 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1327 ip = &mycpu->gd_ipiq[dcpu];
cb973d15 1328 if ((int)(ip->ip_xindex - seq) < 0) {
96728c05
MD
1329 unsigned int eflags = read_eflags();
1330 cpu_enable_intr();
cb973d15 1331 while ((int)(ip->ip_xindex - seq) < 0) {
96728c05 1332 lwkt_process_ipiq();
a2a5ad0d 1333 if (--maxc == 0)
cb973d15 1334 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
a2a5ad0d
MD
1335 if (maxc < -1000000)
1336 panic("LWKT_WAIT_IPIQ");
96728c05
MD
1337 }
1338 write_eflags(eflags);
1339 }
1340 }
1341}
1342
1343/*
1344 * Called from IPI interrupt (like a fast interrupt), which has placed
1345 * us in a critical section. The MP lock may or may not be held.
cb973d15
MD
1346 * May also be called from doreti or splz, or be reentrantly called
1347 * indirectly through the ip_func[] we run.
96728c05
MD
1348 */
1349void
1350lwkt_process_ipiq(void)
1351{
1352 int n;
1353 int cpuid = mycpu->gd_cpuid;
1354
1355 for (n = 0; n < ncpus; ++n) {
1356 lwkt_ipiq_t ip;
1357 int ri;
1358
1359 if (n == cpuid)
1360 continue;
1361 ip = globaldata_find(n)->gd_ipiq;
1362 if (ip == NULL)
1363 continue;
1364 ip = &ip[cpuid];
cb973d15
MD
1365
1366 /*
1367 * Note: xindex is only updated after we are sure the function has
1368 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1369 * function may send an IPI which may block/drain.
1370 */
96728c05
MD
1371 while (ip->ip_rindex != ip->ip_windex) {
1372 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
96728c05 1373 ++ip->ip_rindex;
cb973d15
MD
1374 ip->ip_func[ri](ip->ip_arg[ri]);
1375 /* YYY memory barrier */
1376 ip->ip_xindex = ip->ip_rindex;
96728c05
MD
1377 }
1378 }
1379}
1380
1381#else
1382
1383int
1384lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1385{
1386 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1387 return(0); /* NOT REACHED */
1388}
1389
1390void
1391lwkt_wait_ipiq(int dcpu, int seq)
1392{
1393 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);
1394}
1395
1396#endif