MP Implmentation 3A/4: Fix stupid bug introduced in last commit.
[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 *
a5934754 31 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.19 2003/07/10 18:23:24 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
a5934754
MD
515 * to the MP lock. In this case td_mpcount will be pre-disposed but
516 * not actually synchronized with the actual state of the lock. We
517 * can use it to imply an MP lock requirement for the preemption but
518 * we cannot use it to test whether we hold the MP lock or not.
a2a5ad0d 519 */
a5934754 520 mpheld = MP_LOCK_HELD();
96728c05
MD
521 if (mpheld && td->td_mpcount == 0)
522 panic("lwkt_preempt(): held and no count");
523 savecnt = td->td_mpcount;
8a8d5d85
MD
524 ntd->td_mpcount += td->td_mpcount;
525 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
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
MD
538#ifdef SMP
539 KKASSERT(savecnt == td->td_mpcount);
540 if (mpheld == 0 && MP_LOCK_HELD())
541 cpu_rel_mplock();
542 else if (mpheld && !MP_LOCK_HELD())
543 panic("lwkt_preempt(): MP lock was not held through");
544#endif
26a0694b
MD
545 ntd->td_preempted = NULL;
546 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
547}
548
f1d1c3fa
MD
549/*
550 * Yield our thread while higher priority threads are pending. This is
551 * typically called when we leave a critical section but it can be safely
552 * called while we are in a critical section.
553 *
554 * This function will not generally yield to equal priority threads but it
555 * can occur as a side effect. Note that lwkt_switch() is called from
556 * inside the critical section to pervent its own crit_exit() from reentering
557 * lwkt_yield_quick().
558 *
ef0fdad1
MD
559 * gd_reqpri indicates that *something* changed, e.g. an interrupt or softint
560 * came along but was blocked and made pending.
561 *
f1d1c3fa
MD
562 * (self contained on a per cpu basis)
563 */
564void
565lwkt_yield_quick(void)
566{
567 thread_t td = curthread;
ef0fdad1 568
a2a5ad0d
MD
569 /*
570 * gd_reqpri is cleared in splz if the cpl is 0. If we were to clear
571 * it with a non-zero cpl then we might not wind up calling splz after
572 * a task switch when the critical section is exited even though the
573 * new task could accept the interrupt. YYY alternative is to have
574 * lwkt_switch() just call splz unconditionally.
575 *
576 * XXX from crit_exit() only called after last crit section is released.
577 * If called directly will run splz() even if in a critical section.
578 */
ef0fdad1 579 if ((td->td_pri & TDPRI_MASK) < mycpu->gd_reqpri) {
f1d1c3fa
MD
580 splz();
581 }
582
583 /*
584 * YYY enabling will cause wakeup() to task-switch, which really
585 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
586 * preemption and MP without actually doing preemption or MP, because a
587 * lot of code assumes that wakeup() does not block.
f1d1c3fa 588 */
ef0fdad1 589 if (untimely_switch && mycpu->gd_intr_nesting_level == 0) {
f1d1c3fa
MD
590 crit_enter();
591 /*
592 * YYY temporary hacks until we disassociate the userland scheduler
593 * from the LWKT scheduler.
594 */
595 if (td->td_flags & TDF_RUNQ) {
596 lwkt_switch(); /* will not reenter yield function */
597 } else {
598 lwkt_schedule_self(); /* make sure we are scheduled */
599 lwkt_switch(); /* will not reenter yield function */
600 lwkt_deschedule_self(); /* make sure we are descheduled */
601 }
602 crit_exit_noyield();
603 }
f1d1c3fa
MD
604}
605
8ad65e08 606/*
f1d1c3fa
MD
607 * This implements a normal yield which, unlike _quick, will yield to equal
608 * priority threads as well. Note that gd_reqpri tests will be handled by
609 * the crit_exit() call in lwkt_switch().
610 *
611 * (self contained on a per cpu basis)
8ad65e08
MD
612 */
613void
f1d1c3fa 614lwkt_yield(void)
8ad65e08 615{
f1d1c3fa
MD
616 lwkt_schedule_self();
617 lwkt_switch();
618}
619
620/*
621 * Schedule a thread to run. As the current thread we can always safely
622 * schedule ourselves, and a shortcut procedure is provided for that
623 * function.
624 *
625 * (non-blocking, self contained on a per cpu basis)
626 */
627void
628lwkt_schedule_self(void)
629{
630 thread_t td = curthread;
631
632 crit_enter();
633 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa 634 _lwkt_enqueue(td);
26a0694b
MD
635 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
636 panic("SCHED SELF PANIC");
f1d1c3fa 637 crit_exit();
8ad65e08 638}
8ad65e08
MD
639
640/*
f1d1c3fa
MD
641 * Generic schedule. Possibly schedule threads belonging to other cpus and
642 * deal with threads that might be blocked on a wait queue.
643 *
96728c05 644 * YYY this is one of the best places to implement load balancing code.
f1d1c3fa
MD
645 * Load balancing can be accomplished by requesting other sorts of actions
646 * for the thread in question.
8ad65e08
MD
647 */
648void
649lwkt_schedule(thread_t td)
650{
96728c05 651#ifdef INVARIANTS
26a0694b
MD
652 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
653 && td->td_proc->p_stat == SSLEEP
654 ) {
655 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
656 curthread,
657 curthread->td_proc ? curthread->td_proc->p_pid : -1,
658 curthread->td_proc ? curthread->td_proc->p_stat : -1,
659 td,
660 td->td_proc ? curthread->td_proc->p_pid : -1,
661 td->td_proc ? curthread->td_proc->p_stat : -1
662 );
663 panic("SCHED PANIC");
664 }
96728c05 665#endif
f1d1c3fa
MD
666 crit_enter();
667 if (td == curthread) {
668 _lwkt_enqueue(td);
669 } else {
670 lwkt_wait_t w;
671
672 /*
673 * If the thread is on a wait list we have to send our scheduling
674 * request to the owner of the wait structure. Otherwise we send
675 * the scheduling request to the cpu owning the thread. Races
676 * are ok, the target will forward the message as necessary (the
677 * message may chase the thread around before it finally gets
678 * acted upon).
679 *
680 * (remember, wait structures use stable storage)
681 */
682 if ((w = td->td_wait) != NULL) {
96728c05 683 if (lwkt_trytoken(&w->wa_token)) {
f1d1c3fa
MD
684 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
685 --w->wa_count;
686 td->td_wait = NULL;
d0e06f83 687 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa 688 _lwkt_enqueue(td);
57c254db 689 if (td->td_preemptable) {
96728c05 690 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
57c254db
MD
691 } else if (_lwkt_wantresched(td, curthread)) {
692 need_resched();
693 }
f1d1c3fa 694 } else {
96728c05 695 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 696 }
96728c05 697 lwkt_reltoken(&w->wa_token);
f1d1c3fa 698 } else {
96728c05 699 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
700 }
701 } else {
702 /*
703 * If the wait structure is NULL and we own the thread, there
704 * is no race (since we are in a critical section). If we
705 * do not own the thread there might be a race but the
706 * target cpu will deal with it.
707 */
d0e06f83 708 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa 709 _lwkt_enqueue(td);
57c254db 710 if (td->td_preemptable) {
96728c05 711 td->td_preemptable(td, TDPRI_CRIT);
57c254db
MD
712 } else if (_lwkt_wantresched(td, curthread)) {
713 need_resched();
714 }
f1d1c3fa 715 } else {
96728c05 716 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
717 }
718 }
8ad65e08 719 }
f1d1c3fa 720 crit_exit();
8ad65e08
MD
721}
722
a2a5ad0d
MD
723void
724lwkt_acquire(thread_t td)
725{
726 struct globaldata *gd;
727
728 gd = td->td_gd;
729 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
730 if (gd != mycpu) {
731 crit_enter();
732 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
733 gd = mycpu;
734 td->td_gd = gd;
735 td->td_cpu = gd->gd_cpuid;
736 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
737 crit_exit();
738 }
739}
740
8ad65e08 741/*
f1d1c3fa
MD
742 * Deschedule a thread.
743 *
744 * (non-blocking, self contained on a per cpu basis)
745 */
746void
747lwkt_deschedule_self(void)
748{
749 thread_t td = curthread;
750
751 crit_enter();
752 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa
MD
753 _lwkt_dequeue(td);
754 crit_exit();
755}
756
757/*
758 * Generic deschedule. Descheduling threads other then your own should be
759 * done only in carefully controlled circumstances. Descheduling is
760 * asynchronous.
761 *
762 * This function may block if the cpu has run out of messages.
8ad65e08
MD
763 */
764void
765lwkt_deschedule(thread_t td)
766{
f1d1c3fa
MD
767 crit_enter();
768 if (td == curthread) {
769 _lwkt_dequeue(td);
770 } else {
d0e06f83 771 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
772 _lwkt_dequeue(td);
773 } else {
96728c05 774 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
775 }
776 }
777 crit_exit();
778}
779
4b5f931b
MD
780/*
781 * Set the target thread's priority. This routine does not automatically
782 * switch to a higher priority thread, LWKT threads are not designed for
783 * continuous priority changes. Yield if you want to switch.
784 *
785 * We have to retain the critical section count which uses the high bits
26a0694b
MD
786 * of the td_pri field. The specified priority may also indicate zero or
787 * more critical sections by adding TDPRI_CRIT*N.
4b5f931b
MD
788 */
789void
790lwkt_setpri(thread_t td, int pri)
791{
26a0694b 792 KKASSERT(pri >= 0);
57c254db 793 KKASSERT(td->td_cpu == mycpu->gd_cpuid);
26a0694b
MD
794 crit_enter();
795 if (td->td_flags & TDF_RUNQ) {
796 _lwkt_dequeue(td);
797 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
798 _lwkt_enqueue(td);
799 } else {
800 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
801 }
802 crit_exit();
803}
804
805void
806lwkt_setpri_self(int pri)
807{
808 thread_t td = curthread;
809
4b5f931b
MD
810 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
811 crit_enter();
812 if (td->td_flags & TDF_RUNQ) {
813 _lwkt_dequeue(td);
814 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
815 _lwkt_enqueue(td);
816 } else {
817 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
818 }
819 crit_exit();
820}
821
822struct proc *
823lwkt_preempted_proc(void)
824{
73e4f7b9 825 thread_t td = curthread;
4b5f931b
MD
826 while (td->td_preempted)
827 td = td->td_preempted;
828 return(td->td_proc);
829}
830
831
f1d1c3fa
MD
832/*
833 * This function deschedules the current thread and blocks on the specified
834 * wait queue. We obtain ownership of the wait queue in order to block
835 * on it. A generation number is used to interlock the wait queue in case
836 * it gets signalled while we are blocked waiting on the token.
837 *
838 * Note: alternatively we could dequeue our thread and then message the
839 * target cpu owning the wait queue. YYY implement as sysctl.
840 *
841 * Note: wait queue signals normally ping-pong the cpu as an optimization.
842 */
96728c05
MD
843typedef struct lwkt_gettoken_req {
844 lwkt_token_t tok;
845 int cpu;
846} lwkt_gettoken_req;
847
f1d1c3fa 848void
ae8050a4 849lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
850{
851 thread_t td = curthread;
f1d1c3fa 852
f1d1c3fa 853 lwkt_gettoken(&w->wa_token);
ae8050a4 854 if (w->wa_gen == *gen) {
f1d1c3fa
MD
855 _lwkt_dequeue(td);
856 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
857 ++w->wa_count;
858 td->td_wait = w;
ae8050a4 859 td->td_wmesg = wmesg;
f1d1c3fa 860 lwkt_switch();
8ad65e08 861 }
ae8050a4
MD
862 /* token might be lost, doesn't matter for gen update */
863 *gen = w->wa_gen;
f1d1c3fa
MD
864 lwkt_reltoken(&w->wa_token);
865}
866
867/*
868 * Signal a wait queue. We gain ownership of the wait queue in order to
869 * signal it. Once a thread is removed from the wait queue we have to
870 * deal with the cpu owning the thread.
871 *
872 * Note: alternatively we could message the target cpu owning the wait
873 * queue. YYY implement as sysctl.
874 */
875void
876lwkt_signal(lwkt_wait_t w)
877{
878 thread_t td;
879 int count;
880
881 lwkt_gettoken(&w->wa_token);
882 ++w->wa_gen;
883 count = w->wa_count;
884 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
885 --count;
886 --w->wa_count;
887 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
888 td->td_wait = NULL;
ae8050a4 889 td->td_wmesg = NULL;
d0e06f83 890 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
891 _lwkt_enqueue(td);
892 } else {
96728c05 893 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
894 }
895 lwkt_regettoken(&w->wa_token);
896 }
897 lwkt_reltoken(&w->wa_token);
898}
899
900/*
96728c05 901 * Acquire ownership of a token
f1d1c3fa 902 *
96728c05 903 * Acquire ownership of a token. The token may have spl and/or critical
f1d1c3fa
MD
904 * section side effects, depending on its purpose. These side effects
905 * guarentee that you will maintain ownership of the token as long as you
906 * do not block. If you block you may lose access to the token (but you
907 * must still release it even if you lose your access to it).
908 *
96728c05 909 * YYY for now we use a critical section to prevent IPIs from taking away
a2a5ad0d 910 * a token, but do we really only need to disable IPIs ?
96728c05
MD
911 *
912 * YYY certain tokens could be made to act like mutexes when performance
913 * would be better (e.g. t_cpu == -1). This is not yet implemented.
914 *
a2a5ad0d
MD
915 * YYY the tokens replace 4.x's simplelocks for the most part, but this
916 * means that 4.x does not expect a switch so for now we cannot switch
917 * when waiting for an IPI to be returned.
918 *
919 * YYY If the token is owned by another cpu we may have to send an IPI to
96728c05
MD
920 * it and then block. The IPI causes the token to be given away to the
921 * requesting cpu, unless it has already changed hands. Since only the
922 * current cpu can give away a token it owns we do not need a memory barrier.
a2a5ad0d 923 * This needs serious optimization.
f1d1c3fa 924 */
57c254db
MD
925
926#ifdef SMP
927
96728c05
MD
928static
929void
930lwkt_gettoken_remote(void *arg)
931{
932 lwkt_gettoken_req *req = arg;
933 if (req->tok->t_cpu == mycpu->gd_cpuid) {
a2a5ad0d
MD
934 if (token_debug)
935 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
96728c05 936 req->tok->t_cpu = req->cpu;
a2a5ad0d
MD
937 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
938 /* else set reqcpu to point to current cpu for release */
96728c05
MD
939 }
940}
941
57c254db
MD
942#endif
943
8a8d5d85 944int
f1d1c3fa
MD
945lwkt_gettoken(lwkt_token_t tok)
946{
947 /*
948 * Prevent preemption so the token can't be taken away from us once
949 * we gain ownership of it. Use a synchronous request which might
950 * block. The request will be forwarded as necessary playing catchup
951 * to the token.
952 */
96728c05 953
f1d1c3fa 954 crit_enter();
57c254db 955#ifdef INVARIANTS
a2a5ad0d
MD
956 if (curthread->td_pri > 2000) {
957 curthread->td_pri = 1000;
958 panic("too HIGH!");
57c254db
MD
959 }
960#endif
96728c05 961#ifdef SMP
d0e06f83 962 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
963 struct lwkt_gettoken_req req;
964 int seq;
96728c05
MD
965 int dcpu;
966
967 req.cpu = mycpu->gd_cpuid;
968 req.tok = tok;
969 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d
MD
970 KKASSERT(dcpu >= 0 && dcpu < ncpus);
971 if (token_debug)
972 printf("REQT%d ", dcpu);
96728c05
MD
973 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
974 lwkt_wait_ipiq(dcpu, seq);
a2a5ad0d
MD
975 if (token_debug)
976 printf("REQR%d ", tok->t_cpu);
f1d1c3fa
MD
977 }
978#endif
979 /*
980 * leave us in a critical section on return. This will be undone
8a8d5d85 981 * by lwkt_reltoken(). Bump the generation number.
f1d1c3fa 982 */
8a8d5d85 983 return(++tok->t_gen);
f1d1c3fa
MD
984}
985
96728c05
MD
986/*
987 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
988 * failure.
989 */
990int
991lwkt_trytoken(lwkt_token_t tok)
992{
993 crit_enter();
994#ifdef SMP
995 if (tok->t_cpu != mycpu->gd_cpuid) {
996 return(0);
997 }
998#endif
999 /* leave us in the critical section */
1000 ++tok->t_gen;
1001 return(1);
1002}
1003
f1d1c3fa
MD
1004/*
1005 * Release your ownership of a token. Releases must occur in reverse
1006 * order to aquisitions, eventually so priorities can be unwound properly
1007 * like SPLs. At the moment the actual implemention doesn't care.
1008 *
1009 * We can safely hand a token that we own to another cpu without notifying
1010 * it, but once we do we can't get it back without requesting it (unless
1011 * the other cpu hands it back to us before we check).
1012 *
1013 * We might have lost the token, so check that.
1014 */
1015void
1016lwkt_reltoken(lwkt_token_t tok)
1017{
d0e06f83 1018 if (tok->t_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
1019 tok->t_cpu = tok->t_reqcpu;
1020 }
1021 crit_exit();
1022}
1023
1024/*
8a8d5d85
MD
1025 * Reacquire a token that might have been lost and compare and update the
1026 * generation number. 0 is returned if the generation has not changed
1027 * (nobody else obtained the token while we were blocked, on this cpu or
1028 * any other cpu).
1029 *
1030 * This function returns with the token re-held whether the generation
1031 * number changed or not.
1032 */
1033int
1034lwkt_gentoken(lwkt_token_t tok, int *gen)
1035{
1036 if (lwkt_regettoken(tok) == *gen) {
1037 return(0);
1038 } else {
1039 *gen = tok->t_gen;
1040 return(-1);
1041 }
1042}
1043
1044
1045/*
96728c05 1046 * Re-acquire a token that might have been lost. Returns the generation
8a8d5d85 1047 * number of the token.
f1d1c3fa
MD
1048 */
1049int
1050lwkt_regettoken(lwkt_token_t tok)
1051{
96728c05 1052 /* assert we are in a critical section */
d0e06f83 1053 if (tok->t_cpu != mycpu->gd_cpuid) {
96728c05 1054#ifdef SMP
d0e06f83 1055 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1056 struct lwkt_gettoken_req req;
1057 int seq;
96728c05 1058 int dcpu;
57c254db 1059
96728c05
MD
1060 req.cpu = mycpu->gd_cpuid;
1061 req.tok = tok;
1062 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d
MD
1063 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1064 printf("REQT%d ", dcpu);
96728c05
MD
1065 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1066 lwkt_wait_ipiq(dcpu, seq);
a2a5ad0d 1067 printf("REQR%d ", tok->t_cpu);
f1d1c3fa 1068 }
f1d1c3fa 1069#endif
96728c05
MD
1070 ++tok->t_gen;
1071 }
8a8d5d85 1072 return(tok->t_gen);
8ad65e08
MD
1073}
1074
72740893
MD
1075void
1076lwkt_inittoken(lwkt_token_t tok)
1077{
1078 /*
1079 * Zero structure and set cpu owner and reqcpu to cpu 0.
1080 */
1081 bzero(tok, sizeof(*tok));
1082}
1083
99df837e
MD
1084/*
1085 * Create a kernel process/thread/whatever. It shares it's address space
1086 * with proc0 - ie: kernel only.
1087 *
1088 * XXX should be renamed to lwkt_create()
8a8d5d85
MD
1089 *
1090 * The thread will be entered with the MP lock held.
99df837e
MD
1091 */
1092int
1093lwkt_create(void (*func)(void *), void *arg,
73e4f7b9 1094 struct thread **tdp, thread_t template, int tdflags,
ef0fdad1 1095 const char *fmt, ...)
99df837e 1096{
73e4f7b9 1097 thread_t td;
99df837e
MD
1098 va_list ap;
1099
a2a5ad0d
MD
1100 td = lwkt_alloc_thread(template);
1101 if (tdp)
1102 *tdp = td;
99df837e 1103 cpu_set_thread_handler(td, kthread_exit, func, arg);
ef0fdad1 1104 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1105#ifdef SMP
1106 td->td_mpcount = 1;
1107#endif
99df837e
MD
1108
1109 /*
1110 * Set up arg0 for 'ps' etc
1111 */
1112 va_start(ap, fmt);
1113 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1114 va_end(ap);
1115
1116 /*
1117 * Schedule the thread to run
1118 */
ef0fdad1
MD
1119 if ((td->td_flags & TDF_STOPREQ) == 0)
1120 lwkt_schedule(td);
1121 else
1122 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1123 return 0;
1124}
1125
1126/*
1127 * Destroy an LWKT thread. Warning! This function is not called when
1128 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1129 * uses a different reaping mechanism.
1130 */
1131void
1132lwkt_exit(void)
1133{
1134 thread_t td = curthread;
1135
1136 if (td->td_flags & TDF_VERBOSE)
1137 printf("kthread %p %s has exited\n", td, td->td_comm);
1138 crit_enter();
1139 lwkt_deschedule_self();
1140 ++mycpu->gd_tdfreecount;
1141 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1142 cpu_thread_exit();
1143}
1144
1145/*
1146 * Create a kernel process/thread/whatever. It shares it's address space
ef0fdad1 1147 * with proc0 - ie: kernel only. 5.x compatible.
99df837e
MD
1148 */
1149int
1150kthread_create(void (*func)(void *), void *arg,
1151 struct thread **tdp, const char *fmt, ...)
1152{
73e4f7b9 1153 thread_t td;
99df837e
MD
1154 va_list ap;
1155
a2a5ad0d
MD
1156 td = lwkt_alloc_thread(NULL);
1157 if (tdp)
1158 *tdp = td;
99df837e
MD
1159 cpu_set_thread_handler(td, kthread_exit, func, arg);
1160 td->td_flags |= TDF_VERBOSE;
8a8d5d85
MD
1161#ifdef SMP
1162 td->td_mpcount = 1;
1163#endif
99df837e
MD
1164
1165 /*
1166 * Set up arg0 for 'ps' etc
1167 */
1168 va_start(ap, fmt);
1169 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1170 va_end(ap);
1171
1172 /*
1173 * Schedule the thread to run
1174 */
1175 lwkt_schedule(td);
1176 return 0;
1177}
1178
26a0694b
MD
1179void
1180crit_panic(void)
1181{
73e4f7b9 1182 thread_t td = curthread;
26a0694b
MD
1183 int lpri = td->td_pri;
1184
1185 td->td_pri = 0;
1186 panic("td_pri is/would-go negative! %p %d", td, lpri);
1187}
1188
99df837e
MD
1189/*
1190 * Destroy an LWKT thread. Warning! This function is not called when
1191 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1192 * uses a different reaping mechanism.
1193 *
1194 * XXX duplicates lwkt_exit()
1195 */
1196void
1197kthread_exit(void)
1198{
1199 lwkt_exit();
1200}
1201
96728c05
MD
1202#ifdef SMP
1203
1204/*
1205 * Send a function execution request to another cpu. The request is queued
1206 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1207 * possible target cpu. The FIFO can be written.
1208 *
1209 * YYY If the FIFO fills up we have to enable interrupts and process the
1210 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1211 * Create a CPU_*() function to do this!
1212 *
1213 * Must be called from a critical section.
1214 */
1215int
1216lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1217{
1218 lwkt_ipiq_t ip;
1219 int windex;
a2a5ad0d 1220 struct globaldata *gd = mycpu;
96728c05 1221
a2a5ad0d 1222 if (dcpu == gd->gd_cpuid) {
96728c05
MD
1223 func(arg);
1224 return(0);
1225 }
a2a5ad0d
MD
1226 ++gd->gd_intr_nesting_level;
1227#ifdef INVARIANTS
1228 if (gd->gd_intr_nesting_level > 20)
1229 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1230#endif
96728c05
MD
1231 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1232 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1233 ++ipiq_count;
a2a5ad0d 1234 ip = &gd->gd_ipiq[dcpu];
96728c05
MD
1235 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1236 unsigned int eflags = read_eflags();
a2a5ad0d 1237 printf("SEND_IPIQ FIFO FULL\n");
96728c05
MD
1238 cpu_enable_intr();
1239 ++ipiq_fifofull;
1240 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1241 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1242 lwkt_process_ipiq();
1243 }
a2a5ad0d 1244 printf("SEND_IPIQ FIFO GOOD\n");
96728c05
MD
1245 write_eflags(eflags);
1246 }
1247 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1248 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1249 ip->ip_func[windex] = func;
1250 ip->ip_arg[windex] = arg;
1251 /* YYY memory barrier */
1252 ++ip->ip_windex;
a2a5ad0d 1253 --gd->gd_intr_nesting_level;
96728c05
MD
1254 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
1255 return(ip->ip_windex);
1256}
1257
1258/*
1259 * Wait for the remote cpu to finish processing a function.
1260 *
1261 * YYY we have to enable interrupts and process the IPIQ while waiting
1262 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1263 * function to do this! YYY we really should 'block' here.
1264 *
1265 * Must be called from a critical section. Thsi routine may be called
1266 * from an interrupt (for example, if an interrupt wakes a foreign thread
1267 * up).
1268 */
1269void
1270lwkt_wait_ipiq(int dcpu, int seq)
1271{
1272 lwkt_ipiq_t ip;
a2a5ad0d 1273 int maxc = 100000000;
96728c05
MD
1274
1275 if (dcpu != mycpu->gd_cpuid) {
1276 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1277 ip = &mycpu->gd_ipiq[dcpu];
1278 if ((int)(ip->ip_rindex - seq) < 0) {
1279 unsigned int eflags = read_eflags();
1280 cpu_enable_intr();
1281 while ((int)(ip->ip_rindex - seq) < 0) {
1282 lwkt_process_ipiq();
1283#if 0
1284 lwkt_switch(); /* YYY fixme */
1285#endif
a2a5ad0d
MD
1286 if (--maxc == 0)
1287 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_rindex - seq);
1288 if (maxc < -1000000)
1289 panic("LWKT_WAIT_IPIQ");
96728c05
MD
1290 }
1291 write_eflags(eflags);
1292 }
1293 }
1294}
1295
1296/*
1297 * Called from IPI interrupt (like a fast interrupt), which has placed
1298 * us in a critical section. The MP lock may or may not be held.
1299 * May also be called from doreti or splz.
1300 */
1301void
1302lwkt_process_ipiq(void)
1303{
1304 int n;
1305 int cpuid = mycpu->gd_cpuid;
1306
1307 for (n = 0; n < ncpus; ++n) {
1308 lwkt_ipiq_t ip;
1309 int ri;
1310
1311 if (n == cpuid)
1312 continue;
1313 ip = globaldata_find(n)->gd_ipiq;
1314 if (ip == NULL)
1315 continue;
1316 ip = &ip[cpuid];
1317 while (ip->ip_rindex != ip->ip_windex) {
1318 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
1319 ip->ip_func[ri](ip->ip_arg[ri]);
1320 ++ip->ip_rindex;
1321 }
1322 }
1323}
1324
1325#else
1326
1327int
1328lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1329{
1330 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1331 return(0); /* NOT REACHED */
1332}
1333
1334void
1335lwkt_wait_ipiq(int dcpu, int seq)
1336{
1337 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);
1338}
1339
1340#endif