MP Implmentation 3/4: MAJOR progress on SMP, full userland MP is now working!
[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 *
a2a5ad0d 31 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.18 2003/07/10 04:47:54 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);
173 KASSERT(td != NULL && (td->td_flags & TDF_EXITED),
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;
8a8d5d85 208 td->td_cpu = gd->gd_cpuid; /* YYY don't 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
99df837e
MD
250 KASSERT(td->td_flags & TDF_EXITED,
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
f1d1c3fa 312 crit_enter();
4b5f931b 313 ++switch_count;
8a8d5d85
MD
314
315#ifdef SMP
316 /*
317 * td_mpcount cannot be used to determine if we currently hold the
318 * MP lock because get_mplock() will increment it prior to attempting
319 * to get the lock, and switch out if it can't. Look at the actual lock.
320 */
321 mpheld = MP_LOCK_HELD();
322#endif
99df837e
MD
323 if ((ntd = td->td_preempted) != NULL) {
324 /*
325 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
326 * thread. This occurs transparently, whether the preempted thread
327 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
328 * itself).
329 *
330 * We have to setup the MP lock for the original thread after backing
331 * out the adjustment that was made to curthread when the original
332 * was preempted.
99df837e 333 */
26a0694b 334 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 335#ifdef SMP
96728c05
MD
336 if (ntd->td_mpcount && mpheld == 0) {
337 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
338 td, ntd, td->td_mpcount, ntd->td_mpcount);
339 }
8a8d5d85
MD
340 if (ntd->td_mpcount) {
341 td->td_mpcount -= ntd->td_mpcount;
342 KKASSERT(td->td_mpcount >= 0);
343 }
344#endif
26a0694b 345 ntd->td_flags |= TDF_PREEMPT_DONE;
8a8d5d85 346 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 347 } else {
4b5f931b
MD
348 /*
349 * Priority queue / round-robin at each priority. Note that user
350 * processes run at a fixed, low priority and the user process
351 * scheduler deals with interactions between user processes
352 * by scheduling and descheduling them from the LWKT queue as
353 * necessary.
8a8d5d85
MD
354 *
355 * We have to adjust the MP lock for the target thread. If we
356 * need the MP lock and cannot obtain it we try to locate a
357 * thread that does not need the MP lock.
4b5f931b
MD
358 */
359 gd = mycpu;
4b5f931b
MD
360again:
361 if (gd->gd_runqmask) {
362 int nq = bsrl(gd->gd_runqmask);
363 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
364 gd->gd_runqmask &= ~(1 << nq);
365 goto again;
366 }
8a8d5d85
MD
367#ifdef SMP
368 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
369 /*
96728c05
MD
370 * Target needs MP lock and we couldn't get it, try
371 * to locate a thread which does not need the MP lock
372 * to run.
8a8d5d85
MD
373 */
374 u_int32_t rqmask = gd->gd_runqmask;
375 while (rqmask) {
376 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
377 if (ntd->td_mpcount == 0)
378 break;
379 }
380 if (ntd)
381 break;
382 rqmask &= ~(1 << nq);
383 nq = bsrl(rqmask);
384 }
385 if (ntd == NULL) {
a2a5ad0d
MD
386 ntd = &gd->gd_idlethread;
387 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
388 } else {
389 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
390 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
391 }
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
4b5f931b
MD
397 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
398 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 399#endif
4b5f931b 400 } else {
a2a5ad0d 401 ntd = &gd->gd_idlethread;
4b5f931b 402 }
f1d1c3fa 403 }
26a0694b
MD
404 KASSERT(ntd->td_pri >= TDPRI_CRIT,
405 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85 406
a2a5ad0d
MD
407 /*
408 * Passive release (used to transition from user to kernel mode
409 * when we block or switch rather then when we enter the kernel).
410 * This function is NOT called if we are switching into a preemption
411 * or returning from a preemption.
412 */
413 if (td->td_release)
414 td->td_release(td);
415
8a8d5d85
MD
416 /*
417 * Do the actual switch. If the new target does not need the MP lock
418 * and we are holding it, release the MP lock. If the new target requires
419 * the MP lock we have already acquired it for the target.
420 */
421#ifdef SMP
422 if (ntd->td_mpcount == 0 ) {
423 if (MP_LOCK_HELD())
424 cpu_rel_mplock();
425 } else {
426 ASSERT_MP_LOCK_HELD();
427 }
428#endif
8a8d5d85 429 if (td != ntd) {
f1d1c3fa 430 td->td_switch(ntd);
8a8d5d85 431 }
96728c05 432
f1d1c3fa 433 crit_exit();
8ad65e08
MD
434}
435
b68b7282 436/*
96728c05
MD
437 * Request that the target thread preempt the current thread. Preemption
438 * only works under a specific set of conditions:
b68b7282 439 *
96728c05
MD
440 * - We are not preempting ourselves
441 * - The target thread is owned by the current cpu
442 * - We are not currently being preempted
443 * - The target is not currently being preempted
444 * - We are able to satisfy the target's MP lock requirements (if any).
445 *
446 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
447 * this is called via lwkt_schedule() through the td_preemptable callback.
448 * critpri is the managed critical priority that we should ignore in order
449 * to determine whether preemption is possible (aka usually just the crit
450 * priority of lwkt_schedule() itself).
b68b7282 451 *
26a0694b
MD
452 * XXX at the moment we run the target thread in a critical section during
453 * the preemption in order to prevent the target from taking interrupts
454 * that *WE* can't. Preemption is strictly limited to interrupt threads
455 * and interrupt-like threads, outside of a critical section, and the
456 * preempted source thread will be resumed the instant the target blocks
457 * whether or not the source is scheduled (i.e. preemption is supposed to
458 * be as transparent as possible).
4b5f931b 459 *
8a8d5d85
MD
460 * The target thread inherits our MP count (added to its own) for the
461 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
462 * MP lock during the preemption. Therefore, any preempting targets must be
463 * careful in regards to MP assertions. Note that the MP count may be
464 * out of sync with the physical mp_lock. If we preempt we have to preserve
465 * the expected situation.
b68b7282
MD
466 */
467void
96728c05 468lwkt_preempt(thread_t ntd, int critpri)
b68b7282 469{
73e4f7b9 470 thread_t td = curthread;
8a8d5d85
MD
471#ifdef SMP
472 int mpheld;
57c254db 473 int savecnt;
8a8d5d85 474#endif
b68b7282 475
26a0694b 476 /*
96728c05
MD
477 * The caller has put us in a critical section. We can only preempt
478 * if the caller of the caller was not in a critical section (basically
57c254db
MD
479 * a local interrupt), as determined by the 'critpri' parameter. If
480 * we are unable to preempt
96728c05
MD
481 *
482 * YYY The target thread must be in a critical section (else it must
483 * inherit our critical section? I dunno yet).
26a0694b
MD
484 */
485 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 486
57c254db
MD
487 if (!_lwkt_wantresched(ntd, td)) {
488 ++preempt_miss;
489 return;
490 }
96728c05
MD
491 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
492 ++preempt_miss;
57c254db 493 need_resched();
96728c05
MD
494 return;
495 }
496#ifdef SMP
497 if (ntd->td_cpu != mycpu->gd_cpuid) {
498 ++preempt_miss;
499 return;
500 }
501#endif
26a0694b
MD
502 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
503 ++preempt_weird;
57c254db 504 need_resched();
26a0694b
MD
505 return;
506 }
507 if (ntd->td_preempted) {
4b5f931b 508 ++preempt_hit;
57c254db 509 need_resched();
26a0694b 510 return;
b68b7282 511 }
8a8d5d85 512#ifdef SMP
a2a5ad0d
MD
513 /*
514 * note: an interrupt might have occured just as we were transitioning
515 * to the MP lock, with the lock held but our mpcount still 0. We have
516 * to be sure we restore the same condition when the preemption returns.
517 */
518 mpheld = MP_LOCK_HELD(); /* 0 or 1 */
96728c05
MD
519 if (mpheld && td->td_mpcount == 0)
520 panic("lwkt_preempt(): held and no count");
521 savecnt = td->td_mpcount;
a2a5ad0d 522 td->td_mpcount += mpheld;
8a8d5d85
MD
523 ntd->td_mpcount += td->td_mpcount;
524 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
a2a5ad0d 525 td->td_mpcount -= mpheld;
8a8d5d85
MD
526 ntd->td_mpcount -= td->td_mpcount;
527 ++preempt_miss;
57c254db 528 need_resched();
8a8d5d85
MD
529 return;
530 }
531#endif
26a0694b
MD
532
533 ++preempt_hit;
534 ntd->td_preempted = td;
535 td->td_flags |= TDF_PREEMPT_LOCK;
536 td->td_switch(ntd);
537 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05 538#ifdef SMP
a2a5ad0d 539 td->td_mpcount -= mpheld;
96728c05
MD
540 KKASSERT(savecnt == td->td_mpcount);
541 if (mpheld == 0 && MP_LOCK_HELD())
542 cpu_rel_mplock();
543 else if (mpheld && !MP_LOCK_HELD())
544 panic("lwkt_preempt(): MP lock was not held through");
545#endif
26a0694b
MD
546 ntd->td_preempted = NULL;
547 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
548}
549
f1d1c3fa
MD
550/*
551 * Yield our thread while higher priority threads are pending. This is
552 * typically called when we leave a critical section but it can be safely
553 * called while we are in a critical section.
554 *
555 * This function will not generally yield to equal priority threads but it
556 * can occur as a side effect. Note that lwkt_switch() is called from
557 * inside the critical section to pervent its own crit_exit() from reentering
558 * lwkt_yield_quick().
559 *
ef0fdad1
MD
560 * gd_reqpri indicates that *something* changed, e.g. an interrupt or softint
561 * came along but was blocked and made pending.
562 *
f1d1c3fa
MD
563 * (self contained on a per cpu basis)
564 */
565void
566lwkt_yield_quick(void)
567{
568 thread_t td = curthread;
ef0fdad1 569
a2a5ad0d
MD
570 /*
571 * gd_reqpri is cleared in splz if the cpl is 0. If we were to clear
572 * it with a non-zero cpl then we might not wind up calling splz after
573 * a task switch when the critical section is exited even though the
574 * new task could accept the interrupt. YYY alternative is to have
575 * lwkt_switch() just call splz unconditionally.
576 *
577 * XXX from crit_exit() only called after last crit section is released.
578 * If called directly will run splz() even if in a critical section.
579 */
ef0fdad1 580 if ((td->td_pri & TDPRI_MASK) < mycpu->gd_reqpri) {
f1d1c3fa
MD
581 splz();
582 }
583
584 /*
585 * YYY enabling will cause wakeup() to task-switch, which really
586 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
587 * preemption and MP without actually doing preemption or MP, because a
588 * lot of code assumes that wakeup() does not block.
f1d1c3fa 589 */
ef0fdad1 590 if (untimely_switch && mycpu->gd_intr_nesting_level == 0) {
f1d1c3fa
MD
591 crit_enter();
592 /*
593 * YYY temporary hacks until we disassociate the userland scheduler
594 * from the LWKT scheduler.
595 */
596 if (td->td_flags & TDF_RUNQ) {
597 lwkt_switch(); /* will not reenter yield function */
598 } else {
599 lwkt_schedule_self(); /* make sure we are scheduled */
600 lwkt_switch(); /* will not reenter yield function */
601 lwkt_deschedule_self(); /* make sure we are descheduled */
602 }
603 crit_exit_noyield();
604 }
f1d1c3fa
MD
605}
606
8ad65e08 607/*
f1d1c3fa
MD
608 * This implements a normal yield which, unlike _quick, will yield to equal
609 * priority threads as well. Note that gd_reqpri tests will be handled by
610 * the crit_exit() call in lwkt_switch().
611 *
612 * (self contained on a per cpu basis)
8ad65e08
MD
613 */
614void
f1d1c3fa 615lwkt_yield(void)
8ad65e08 616{
f1d1c3fa
MD
617 lwkt_schedule_self();
618 lwkt_switch();
619}
620
621/*
622 * Schedule a thread to run. As the current thread we can always safely
623 * schedule ourselves, and a shortcut procedure is provided for that
624 * function.
625 *
626 * (non-blocking, self contained on a per cpu basis)
627 */
628void
629lwkt_schedule_self(void)
630{
631 thread_t td = curthread;
632
633 crit_enter();
634 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa 635 _lwkt_enqueue(td);
26a0694b
MD
636 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
637 panic("SCHED SELF PANIC");
f1d1c3fa 638 crit_exit();
8ad65e08 639}
8ad65e08
MD
640
641/*
f1d1c3fa
MD
642 * Generic schedule. Possibly schedule threads belonging to other cpus and
643 * deal with threads that might be blocked on a wait queue.
644 *
96728c05 645 * YYY this is one of the best places to implement load balancing code.
f1d1c3fa
MD
646 * Load balancing can be accomplished by requesting other sorts of actions
647 * for the thread in question.
8ad65e08
MD
648 */
649void
650lwkt_schedule(thread_t td)
651{
96728c05 652#ifdef INVARIANTS
26a0694b
MD
653 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
654 && td->td_proc->p_stat == SSLEEP
655 ) {
656 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
657 curthread,
658 curthread->td_proc ? curthread->td_proc->p_pid : -1,
659 curthread->td_proc ? curthread->td_proc->p_stat : -1,
660 td,
661 td->td_proc ? curthread->td_proc->p_pid : -1,
662 td->td_proc ? curthread->td_proc->p_stat : -1
663 );
664 panic("SCHED PANIC");
665 }
96728c05 666#endif
f1d1c3fa
MD
667 crit_enter();
668 if (td == curthread) {
669 _lwkt_enqueue(td);
670 } else {
671 lwkt_wait_t w;
672
673 /*
674 * If the thread is on a wait list we have to send our scheduling
675 * request to the owner of the wait structure. Otherwise we send
676 * the scheduling request to the cpu owning the thread. Races
677 * are ok, the target will forward the message as necessary (the
678 * message may chase the thread around before it finally gets
679 * acted upon).
680 *
681 * (remember, wait structures use stable storage)
682 */
683 if ((w = td->td_wait) != NULL) {
96728c05 684 if (lwkt_trytoken(&w->wa_token)) {
f1d1c3fa
MD
685 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
686 --w->wa_count;
687 td->td_wait = NULL;
d0e06f83 688 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa 689 _lwkt_enqueue(td);
57c254db 690 if (td->td_preemptable) {
96728c05 691 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
57c254db
MD
692 } else if (_lwkt_wantresched(td, curthread)) {
693 need_resched();
694 }
f1d1c3fa 695 } else {
96728c05 696 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 697 }
96728c05 698 lwkt_reltoken(&w->wa_token);
f1d1c3fa 699 } else {
96728c05 700 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
701 }
702 } else {
703 /*
704 * If the wait structure is NULL and we own the thread, there
705 * is no race (since we are in a critical section). If we
706 * do not own the thread there might be a race but the
707 * target cpu will deal with it.
708 */
d0e06f83 709 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa 710 _lwkt_enqueue(td);
57c254db 711 if (td->td_preemptable) {
96728c05 712 td->td_preemptable(td, TDPRI_CRIT);
57c254db
MD
713 } else if (_lwkt_wantresched(td, curthread)) {
714 need_resched();
715 }
f1d1c3fa 716 } else {
96728c05 717 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
718 }
719 }
8ad65e08 720 }
f1d1c3fa 721 crit_exit();
8ad65e08
MD
722}
723
a2a5ad0d
MD
724void
725lwkt_acquire(thread_t td)
726{
727 struct globaldata *gd;
728
729 gd = td->td_gd;
730 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
731 if (gd != mycpu) {
732 crit_enter();
733 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
734 gd = mycpu;
735 td->td_gd = gd;
736 td->td_cpu = gd->gd_cpuid;
737 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
738 crit_exit();
739 }
740}
741
8ad65e08 742/*
f1d1c3fa
MD
743 * Deschedule a thread.
744 *
745 * (non-blocking, self contained on a per cpu basis)
746 */
747void
748lwkt_deschedule_self(void)
749{
750 thread_t td = curthread;
751
752 crit_enter();
753 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa
MD
754 _lwkt_dequeue(td);
755 crit_exit();
756}
757
758/*
759 * Generic deschedule. Descheduling threads other then your own should be
760 * done only in carefully controlled circumstances. Descheduling is
761 * asynchronous.
762 *
763 * This function may block if the cpu has run out of messages.
8ad65e08
MD
764 */
765void
766lwkt_deschedule(thread_t td)
767{
f1d1c3fa
MD
768 crit_enter();
769 if (td == curthread) {
770 _lwkt_dequeue(td);
771 } else {
d0e06f83 772 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
773 _lwkt_dequeue(td);
774 } else {
96728c05 775 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
776 }
777 }
778 crit_exit();
779}
780
4b5f931b
MD
781/*
782 * Set the target thread's priority. This routine does not automatically
783 * switch to a higher priority thread, LWKT threads are not designed for
784 * continuous priority changes. Yield if you want to switch.
785 *
786 * We have to retain the critical section count which uses the high bits
26a0694b
MD
787 * of the td_pri field. The specified priority may also indicate zero or
788 * more critical sections by adding TDPRI_CRIT*N.
4b5f931b
MD
789 */
790void
791lwkt_setpri(thread_t td, int pri)
792{
26a0694b 793 KKASSERT(pri >= 0);
57c254db 794 KKASSERT(td->td_cpu == mycpu->gd_cpuid);
26a0694b
MD
795 crit_enter();
796 if (td->td_flags & TDF_RUNQ) {
797 _lwkt_dequeue(td);
798 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
799 _lwkt_enqueue(td);
800 } else {
801 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
802 }
803 crit_exit();
804}
805
806void
807lwkt_setpri_self(int pri)
808{
809 thread_t td = curthread;
810
4b5f931b
MD
811 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
812 crit_enter();
813 if (td->td_flags & TDF_RUNQ) {
814 _lwkt_dequeue(td);
815 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
816 _lwkt_enqueue(td);
817 } else {
818 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
819 }
820 crit_exit();
821}
822
823struct proc *
824lwkt_preempted_proc(void)
825{
73e4f7b9 826 thread_t td = curthread;
4b5f931b
MD
827 while (td->td_preempted)
828 td = td->td_preempted;
829 return(td->td_proc);
830}
831
832
f1d1c3fa
MD
833/*
834 * This function deschedules the current thread and blocks on the specified
835 * wait queue. We obtain ownership of the wait queue in order to block
836 * on it. A generation number is used to interlock the wait queue in case
837 * it gets signalled while we are blocked waiting on the token.
838 *
839 * Note: alternatively we could dequeue our thread and then message the
840 * target cpu owning the wait queue. YYY implement as sysctl.
841 *
842 * Note: wait queue signals normally ping-pong the cpu as an optimization.
843 */
96728c05
MD
844typedef struct lwkt_gettoken_req {
845 lwkt_token_t tok;
846 int cpu;
847} lwkt_gettoken_req;
848
f1d1c3fa 849void
ae8050a4 850lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
851{
852 thread_t td = curthread;
f1d1c3fa 853
f1d1c3fa 854 lwkt_gettoken(&w->wa_token);
ae8050a4 855 if (w->wa_gen == *gen) {
f1d1c3fa
MD
856 _lwkt_dequeue(td);
857 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
858 ++w->wa_count;
859 td->td_wait = w;
ae8050a4 860 td->td_wmesg = wmesg;
f1d1c3fa 861 lwkt_switch();
8ad65e08 862 }
ae8050a4
MD
863 /* token might be lost, doesn't matter for gen update */
864 *gen = w->wa_gen;
f1d1c3fa
MD
865 lwkt_reltoken(&w->wa_token);
866}
867
868/*
869 * Signal a wait queue. We gain ownership of the wait queue in order to
870 * signal it. Once a thread is removed from the wait queue we have to
871 * deal with the cpu owning the thread.
872 *
873 * Note: alternatively we could message the target cpu owning the wait
874 * queue. YYY implement as sysctl.
875 */
876void
877lwkt_signal(lwkt_wait_t w)
878{
879 thread_t td;
880 int count;
881
882 lwkt_gettoken(&w->wa_token);
883 ++w->wa_gen;
884 count = w->wa_count;
885 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
886 --count;
887 --w->wa_count;
888 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
889 td->td_wait = NULL;
ae8050a4 890 td->td_wmesg = NULL;
d0e06f83 891 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
892 _lwkt_enqueue(td);
893 } else {
96728c05 894 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
895 }
896 lwkt_regettoken(&w->wa_token);
897 }
898 lwkt_reltoken(&w->wa_token);
899}
900
901/*
96728c05 902 * Acquire ownership of a token
f1d1c3fa 903 *
96728c05 904 * Acquire ownership of a token. The token may have spl and/or critical
f1d1c3fa
MD
905 * section side effects, depending on its purpose. These side effects
906 * guarentee that you will maintain ownership of the token as long as you
907 * do not block. If you block you may lose access to the token (but you
908 * must still release it even if you lose your access to it).
909 *
96728c05 910 * YYY for now we use a critical section to prevent IPIs from taking away
a2a5ad0d 911 * a token, but do we really only need to disable IPIs ?
96728c05
MD
912 *
913 * YYY certain tokens could be made to act like mutexes when performance
914 * would be better (e.g. t_cpu == -1). This is not yet implemented.
915 *
a2a5ad0d
MD
916 * YYY the tokens replace 4.x's simplelocks for the most part, but this
917 * means that 4.x does not expect a switch so for now we cannot switch
918 * when waiting for an IPI to be returned.
919 *
920 * YYY If the token is owned by another cpu we may have to send an IPI to
96728c05
MD
921 * it and then block. The IPI causes the token to be given away to the
922 * requesting cpu, unless it has already changed hands. Since only the
923 * current cpu can give away a token it owns we do not need a memory barrier.
a2a5ad0d 924 * This needs serious optimization.
f1d1c3fa 925 */
57c254db
MD
926
927#ifdef SMP
928
96728c05
MD
929static
930void
931lwkt_gettoken_remote(void *arg)
932{
933 lwkt_gettoken_req *req = arg;
934 if (req->tok->t_cpu == mycpu->gd_cpuid) {
a2a5ad0d
MD
935 if (token_debug)
936 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
96728c05 937 req->tok->t_cpu = req->cpu;
a2a5ad0d
MD
938 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
939 /* else set reqcpu to point to current cpu for release */
96728c05
MD
940 }
941}
942
57c254db
MD
943#endif
944
8a8d5d85 945int
f1d1c3fa
MD
946lwkt_gettoken(lwkt_token_t tok)
947{
948 /*
949 * Prevent preemption so the token can't be taken away from us once
950 * we gain ownership of it. Use a synchronous request which might
951 * block. The request will be forwarded as necessary playing catchup
952 * to the token.
953 */
96728c05 954
f1d1c3fa 955 crit_enter();
57c254db 956#ifdef INVARIANTS
a2a5ad0d
MD
957 if (curthread->td_pri > 2000) {
958 curthread->td_pri = 1000;
959 panic("too HIGH!");
57c254db
MD
960 }
961#endif
96728c05 962#ifdef SMP
d0e06f83 963 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
964 struct lwkt_gettoken_req req;
965 int seq;
96728c05
MD
966 int dcpu;
967
968 req.cpu = mycpu->gd_cpuid;
969 req.tok = tok;
970 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d
MD
971 KKASSERT(dcpu >= 0 && dcpu < ncpus);
972 if (token_debug)
973 printf("REQT%d ", dcpu);
96728c05
MD
974 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
975 lwkt_wait_ipiq(dcpu, seq);
a2a5ad0d
MD
976 if (token_debug)
977 printf("REQR%d ", tok->t_cpu);
f1d1c3fa
MD
978 }
979#endif
980 /*
981 * leave us in a critical section on return. This will be undone
8a8d5d85 982 * by lwkt_reltoken(). Bump the generation number.
f1d1c3fa 983 */
8a8d5d85 984 return(++tok->t_gen);
f1d1c3fa
MD
985}
986
96728c05
MD
987/*
988 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
989 * failure.
990 */
991int
992lwkt_trytoken(lwkt_token_t tok)
993{
994 crit_enter();
995#ifdef SMP
996 if (tok->t_cpu != mycpu->gd_cpuid) {
997 return(0);
998 }
999#endif
1000 /* leave us in the critical section */
1001 ++tok->t_gen;
1002 return(1);
1003}
1004
f1d1c3fa
MD
1005/*
1006 * Release your ownership of a token. Releases must occur in reverse
1007 * order to aquisitions, eventually so priorities can be unwound properly
1008 * like SPLs. At the moment the actual implemention doesn't care.
1009 *
1010 * We can safely hand a token that we own to another cpu without notifying
1011 * it, but once we do we can't get it back without requesting it (unless
1012 * the other cpu hands it back to us before we check).
1013 *
1014 * We might have lost the token, so check that.
1015 */
1016void
1017lwkt_reltoken(lwkt_token_t tok)
1018{
d0e06f83 1019 if (tok->t_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
1020 tok->t_cpu = tok->t_reqcpu;
1021 }
1022 crit_exit();
1023}
1024
1025/*
8a8d5d85
MD
1026 * Reacquire a token that might have been lost and compare and update the
1027 * generation number. 0 is returned if the generation has not changed
1028 * (nobody else obtained the token while we were blocked, on this cpu or
1029 * any other cpu).
1030 *
1031 * This function returns with the token re-held whether the generation
1032 * number changed or not.
1033 */
1034int
1035lwkt_gentoken(lwkt_token_t tok, int *gen)
1036{
1037 if (lwkt_regettoken(tok) == *gen) {
1038 return(0);
1039 } else {
1040 *gen = tok->t_gen;
1041 return(-1);
1042 }
1043}
1044
1045
1046/*
96728c05 1047 * Re-acquire a token that might have been lost. Returns the generation
8a8d5d85 1048 * number of the token.
f1d1c3fa
MD
1049 */
1050int
1051lwkt_regettoken(lwkt_token_t tok)
1052{
96728c05 1053 /* assert we are in a critical section */
d0e06f83 1054 if (tok->t_cpu != mycpu->gd_cpuid) {
96728c05 1055#ifdef SMP
d0e06f83 1056 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1057 struct lwkt_gettoken_req req;
1058 int seq;
96728c05 1059 int dcpu;
57c254db 1060
96728c05
MD
1061 req.cpu = mycpu->gd_cpuid;
1062 req.tok = tok;
1063 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d
MD
1064 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1065 printf("REQT%d ", dcpu);
96728c05
MD
1066 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1067 lwkt_wait_ipiq(dcpu, seq);
a2a5ad0d 1068 printf("REQR%d ", tok->t_cpu);
f1d1c3fa 1069 }
f1d1c3fa 1070#endif
96728c05
MD
1071 ++tok->t_gen;
1072 }
8a8d5d85 1073 return(tok->t_gen);
8ad65e08
MD
1074}
1075
72740893
MD
1076void
1077lwkt_inittoken(lwkt_token_t tok)
1078{
1079 /*
1080 * Zero structure and set cpu owner and reqcpu to cpu 0.
1081 */
1082 bzero(tok, sizeof(*tok));
1083}
1084
99df837e
MD
1085/*
1086 * Create a kernel process/thread/whatever. It shares it's address space
1087 * with proc0 - ie: kernel only.
1088 *
1089 * XXX should be renamed to lwkt_create()
8a8d5d85
MD
1090 *
1091 * The thread will be entered with the MP lock held.
99df837e
MD
1092 */
1093int
1094lwkt_create(void (*func)(void *), void *arg,
73e4f7b9 1095 struct thread **tdp, thread_t template, int tdflags,
ef0fdad1 1096 const char *fmt, ...)
99df837e 1097{
73e4f7b9 1098 thread_t td;
99df837e
MD
1099 va_list ap;
1100
a2a5ad0d
MD
1101 td = lwkt_alloc_thread(template);
1102 if (tdp)
1103 *tdp = td;
99df837e 1104 cpu_set_thread_handler(td, kthread_exit, func, arg);
ef0fdad1 1105 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1106#ifdef SMP
1107 td->td_mpcount = 1;
1108#endif
99df837e
MD
1109
1110 /*
1111 * Set up arg0 for 'ps' etc
1112 */
1113 va_start(ap, fmt);
1114 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1115 va_end(ap);
1116
1117 /*
1118 * Schedule the thread to run
1119 */
ef0fdad1
MD
1120 if ((td->td_flags & TDF_STOPREQ) == 0)
1121 lwkt_schedule(td);
1122 else
1123 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1124 return 0;
1125}
1126
1127/*
1128 * Destroy an LWKT thread. Warning! This function is not called when
1129 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1130 * uses a different reaping mechanism.
1131 */
1132void
1133lwkt_exit(void)
1134{
1135 thread_t td = curthread;
1136
1137 if (td->td_flags & TDF_VERBOSE)
1138 printf("kthread %p %s has exited\n", td, td->td_comm);
1139 crit_enter();
1140 lwkt_deschedule_self();
1141 ++mycpu->gd_tdfreecount;
1142 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1143 cpu_thread_exit();
1144}
1145
1146/*
1147 * Create a kernel process/thread/whatever. It shares it's address space
ef0fdad1 1148 * with proc0 - ie: kernel only. 5.x compatible.
99df837e
MD
1149 */
1150int
1151kthread_create(void (*func)(void *), void *arg,
1152 struct thread **tdp, const char *fmt, ...)
1153{
73e4f7b9 1154 thread_t td;
99df837e
MD
1155 va_list ap;
1156
a2a5ad0d
MD
1157 td = lwkt_alloc_thread(NULL);
1158 if (tdp)
1159 *tdp = td;
99df837e
MD
1160 cpu_set_thread_handler(td, kthread_exit, func, arg);
1161 td->td_flags |= TDF_VERBOSE;
8a8d5d85
MD
1162#ifdef SMP
1163 td->td_mpcount = 1;
1164#endif
99df837e
MD
1165
1166 /*
1167 * Set up arg0 for 'ps' etc
1168 */
1169 va_start(ap, fmt);
1170 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1171 va_end(ap);
1172
1173 /*
1174 * Schedule the thread to run
1175 */
1176 lwkt_schedule(td);
1177 return 0;
1178}
1179
26a0694b
MD
1180void
1181crit_panic(void)
1182{
73e4f7b9 1183 thread_t td = curthread;
26a0694b
MD
1184 int lpri = td->td_pri;
1185
1186 td->td_pri = 0;
1187 panic("td_pri is/would-go negative! %p %d", td, lpri);
1188}
1189
99df837e
MD
1190/*
1191 * Destroy an LWKT thread. Warning! This function is not called when
1192 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1193 * uses a different reaping mechanism.
1194 *
1195 * XXX duplicates lwkt_exit()
1196 */
1197void
1198kthread_exit(void)
1199{
1200 lwkt_exit();
1201}
1202
96728c05
MD
1203#ifdef SMP
1204
1205/*
1206 * Send a function execution request to another cpu. The request is queued
1207 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1208 * possible target cpu. The FIFO can be written.
1209 *
1210 * YYY If the FIFO fills up we have to enable interrupts and process the
1211 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1212 * Create a CPU_*() function to do this!
1213 *
1214 * Must be called from a critical section.
1215 */
1216int
1217lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1218{
1219 lwkt_ipiq_t ip;
1220 int windex;
a2a5ad0d 1221 struct globaldata *gd = mycpu;
96728c05 1222
a2a5ad0d 1223 if (dcpu == gd->gd_cpuid) {
96728c05
MD
1224 func(arg);
1225 return(0);
1226 }
a2a5ad0d
MD
1227 ++gd->gd_intr_nesting_level;
1228#ifdef INVARIANTS
1229 if (gd->gd_intr_nesting_level > 20)
1230 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1231#endif
96728c05
MD
1232 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1233 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1234 ++ipiq_count;
a2a5ad0d 1235 ip = &gd->gd_ipiq[dcpu];
96728c05
MD
1236 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1237 unsigned int eflags = read_eflags();
a2a5ad0d 1238 printf("SEND_IPIQ FIFO FULL\n");
96728c05
MD
1239 cpu_enable_intr();
1240 ++ipiq_fifofull;
1241 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1242 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1243 lwkt_process_ipiq();
1244 }
a2a5ad0d 1245 printf("SEND_IPIQ FIFO GOOD\n");
96728c05
MD
1246 write_eflags(eflags);
1247 }
1248 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1249 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1250 ip->ip_func[windex] = func;
1251 ip->ip_arg[windex] = arg;
1252 /* YYY memory barrier */
1253 ++ip->ip_windex;
a2a5ad0d 1254 --gd->gd_intr_nesting_level;
96728c05
MD
1255 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
1256 return(ip->ip_windex);
1257}
1258
1259/*
1260 * Wait for the remote cpu to finish processing a function.
1261 *
1262 * YYY we have to enable interrupts and process the IPIQ while waiting
1263 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1264 * function to do this! YYY we really should 'block' here.
1265 *
1266 * Must be called from a critical section. Thsi routine may be called
1267 * from an interrupt (for example, if an interrupt wakes a foreign thread
1268 * up).
1269 */
1270void
1271lwkt_wait_ipiq(int dcpu, int seq)
1272{
1273 lwkt_ipiq_t ip;
a2a5ad0d 1274 int maxc = 100000000;
96728c05
MD
1275
1276 if (dcpu != mycpu->gd_cpuid) {
1277 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1278 ip = &mycpu->gd_ipiq[dcpu];
1279 if ((int)(ip->ip_rindex - seq) < 0) {
1280 unsigned int eflags = read_eflags();
1281 cpu_enable_intr();
1282 while ((int)(ip->ip_rindex - seq) < 0) {
1283 lwkt_process_ipiq();
1284#if 0
1285 lwkt_switch(); /* YYY fixme */
1286#endif
a2a5ad0d
MD
1287 if (--maxc == 0)
1288 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_rindex - seq);
1289 if (maxc < -1000000)
1290 panic("LWKT_WAIT_IPIQ");
96728c05
MD
1291 }
1292 write_eflags(eflags);
1293 }
1294 }
1295}
1296
1297/*
1298 * Called from IPI interrupt (like a fast interrupt), which has placed
1299 * us in a critical section. The MP lock may or may not be held.
1300 * May also be called from doreti or splz.
1301 */
1302void
1303lwkt_process_ipiq(void)
1304{
1305 int n;
1306 int cpuid = mycpu->gd_cpuid;
1307
1308 for (n = 0; n < ncpus; ++n) {
1309 lwkt_ipiq_t ip;
1310 int ri;
1311
1312 if (n == cpuid)
1313 continue;
1314 ip = globaldata_find(n)->gd_ipiq;
1315 if (ip == NULL)
1316 continue;
1317 ip = &ip[cpuid];
1318 while (ip->ip_rindex != ip->ip_windex) {
1319 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
1320 ip->ip_func[ri](ip->ip_arg[ri]);
1321 ++ip->ip_rindex;
1322 }
1323 }
1324}
1325
1326#else
1327
1328int
1329lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1330{
1331 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1332 return(0); /* NOT REACHED */
1333}
1334
1335void
1336lwkt_wait_ipiq(int dcpu, int seq)
1337{
1338 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);
1339}
1340
1341#endif